~

Theoretical and
Computational
Acoustics 2003

This page intentionally left blank

Theoretical and
Computational
Acoustics 2003
Honolulu, Hawaii, USA

11 - 15 August 2003

Editors

Alexandra Tolstoy
ATolstoy Sciences, USA

Yu-Chiung Teng
Femarco, Inc., USA

E C Shang
University of Colorado, Bouldet USA

1:a World -Scientific

-

N E W JERSEY * L O N D O N * SINGAPORE

BElJlNG

SHANGHAI

H O N G KONG

TAIPEI

CHENNAI

Published by
World Scientific Publishing Co. Re. Ltd.
5 Toh Tuck Link, Singapore 596224
USA once: Suite 202, 1060 Main Street, River Edge, NJ 07661
UK ofice: 57 Shelton Street, Covent Garden, London WC2H 9HE

British Library Cataloguing-in-PublicationData

A catalogue record for this book is available from the British Library.

THEORETICAL AND COMPUTATIONAL ACOUSTICS 2003

Copyright 0 2004 by World Scientific Publishing Co. Re. Ltd.

All rights reserved. This book, or parts thereof; may not be reproduced in anyform or by any means, electronic or
mechanical, including photocopying. recording or any information storage and retrieval system now known or to
be invented, without written permissionfrom the Publisher.

For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center,
lnc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from
the publisher.

ISBN 98 1-238-921-0

Printed in Singapore by World Scientific Printers (S) Pte Ltd

Preface

The sixth International Conference on Theoretical and Computational Acoustics (ICTCA)

was held August 11-15, 2003 at the Sheraton Waikiki Hotel in Honolulu, Hawaii. This
meeting was sponsored by the U.S. Naval Undersea Warfare Center (NUWC), the U.S.
Office of Naval Research (ONR), SACLANT Undersea Research Centre (Italy), Yale
University (U.S.A.), Boston University (U.S.A.), Columbia University (U.S.A.),
University of Colorado (U.S.A.), University of Hawaii (U.S.A.), School of Ocean and
Earth Science and Technology (U.S.A.), National Taiwan University (China), Istituto
Nazionale di Oceanografia e di Geofisica Sperimentale (Italy), Hangzhou Applied
Acoustics Research Insitute (China), and the Key Laboratory of Geophysical Exploration
(CNPC, China).
The objective of this conference was to provide a forum for active researchers to
discuss state-of-the-art developments and results in theoretical and computational acoustics
and related topics. It brought together researchers from numerous areas of science to exchange ideas and stimulate future research. The website is located at: www.ictca2003.com.
Approximately one hundred scholars, scientists, and engineers from numerous
countries participated in this event. The presented lectures examined topics in Underwater
Acoustics, Mathematics, Scattering and Diffraction, Seismic Explorations, Boundary
Element Methods, Applications to the Oil Industry, Visualization, Shallow Water
Acoustics, Matched Field Tracking and Processing, Bubbles, Waves in Complex Media,
Seabed Interactions, and Ocean Acoustic Inversion. There were also three special sessions
on the Parabolic Equation (PE) devoted to the work and memory of Professor Fred Tappert
of Miami University (recently deceased).
The conference committee wishes to particularly thank Dr. Ding Lee (NUWC and
Yale University) as the founder and Honorary Chair of the conference. Dr. Lee continues to
organize much of these meetings and to set the tone for their continued success. Other
Honorary Chairs included Profs. John E. Ffows Williams and Martin Schultz. Special Thanks
also go to Prof. Allan Robinson as keynote speaker with the presentation Prediction
systems with data assimilation for coupled ocean science and ocean acoustics.
The special sessions were of particular interest, and we gratefully acknowledge
Drs. Michael Taroudakis, Finn Jensen and Lan Nghiem-Phu for the PE sessions which
included technical as well as personal presentations on Dr. Tappert.Additionally, there were
Drs. G. Herman and J. Carcione, S. Marburg and B. Nolte, Y.C. Teng, K. Lima, W. Carey,
D.B. Maxwell, P.A. Baxley, L. Bjorno and D. Bradley, and G. Seriani with N. Saffari.
We look forward to the seventh meeting scheduled in 2005 for Hangzhou, China,
and the eighth in 2007 for Iraklion, Crete.

This page intentionally left blank

vii

Contents

Preface

Cross hole simulations in elastic formations using off axis sources via BEM
J Antonio and A Tadeu

BEM-based nearfield holography for scattering sound field

X Bao and H Ge

Parabolic equation techniques for propagation and scattering

D C Calvo et al.

16

Seafloor properties and segmentation

G Canepa et al.

29

HHsound transmission and spatial conference in selected shallow water areas:

measurements and theory
W M Carey et al.

38

Fractional derivative modelling of acoustic dissipations obeying arbitrary

frequency power law
W Chen and S Holm

72

Visualization of the energy flow for elastic waves: Comparison and contrast
of conventional vector field vs color coded representation of the Poynting vector
C E Dean and J P Braselton

78

The acoustical Klein-Gordon equation: The direct and inverse problems

B J Forbes and E R Pike

91

Accuracy and efficiency of the multipole Galerkin BEM for acoustics

L Gaul, M Fischer and U Gauger

97

Bottom reflection phase shift parameter inversion from reverberation

and propagation data
H L Ge et al.

107

Acoustic behavior of elastic screens in open and confined spaces

L Godinho and A Tadeu

117

...

Vlll

Near-axial interference effects in long-range propagation in a

range-independent ocean
N S Grigorieva and G M Fridman

123

Investigation of near-axial interference effects for propagation in a

ducted waveguide
N S Grigorieva, G M Fridman and D R Palmer

129

Dynamics of Immiscible Two-Phase Fluid Reservoir Flow

A Hanyga

135

Bottom reflection properties deduced from ambient noise: Simulation

and experiment
C H Harrison and A Baldacci

162

Measurement of the seabed reflection coefficient in shallow water:

A comparison of two techniques
C W Holland and C H Harrison

178

Azimuthal limitation in the parabolic-equation approximation for

three-dimensional underwater acoustic propagation
L-W Hsieh, Y-TLin and C-F Chen

192

Application of FEM/BEM method to interior acoustical field shaping by

Helmholtz resonators
B Irfanoglu and M Caliskan

203

Blind ocean acoustic tomography with source spectrum estimation

S M Jesus and C Soares

211

Projection operator method for solving coupled-mode equations: An application

to separating forward and backward scattered acoustic fields
D P Knobles

221

Molecular dynamics approach to sonoluminescent bubbles

W Lauterbom et al.

233

Revolutionary Influence of the Parabolic Equation Approximation

D Lee

244

A variational approach for geoacoustic inversion using adjoint modeling

of a PE approximation model with non local impedance boundary conditions
J-C Le Gac et al.

254

ix
Visco-acoustic wave equation migration
G Li et al.

264

Time-lapse seismic AVO attributes analysis with variation of oil saturation,

pressure and temperature
J Li, X Chen and L Jin

213

Modeling high-frequency seafloor volume backscatter by shell fragment

distributions
A P Lyons

282

Brain topography of perception of target and non-target acoustic signals

S Lytaev

29 1

Computation of acoustic field on 2D fronts

Part 1. Trajectories
N Maltsev

298

A new technique combining eigenfunction expansions and boundary

elements to solve acoustic radiation problems
J D Maynard

303

Inversion in viscoacoustic media

A Ribodetti and A Hanyga

311

Prediction systems with data assimilation for coupled ocean science and
ocean acoustics
A R Robinson and P F J Lermusiaux

325

Horizontal refraction and the uncoupled azimuth approximation

K B Smith and F D Tappert

343

Broadband propagation results for a shallow water Gaussian canyon test case
using a 3-D parabolic equation model
F Sturm

355

Modeling the pore structure effect on acoustic wave amplitude in

sedimentary rocks
Y-F Sun

366

Sesimic resolution: an old problem but a new challenge for seismic

reservoir characterization
Y-F Sun et al.

375

Wave scattering by a rigid inclusion submerged in a channel bounded by

a sediment layer over a rigid boundary
A Tadeu, L Godinho and J Antonio

384

PE-based methods for treating forward scattering from a rough sea-surface

D J Thomson and G H Brooke

390

Simulated tomographic geoacoustic inversion

A Tolstoy

403

Impacts of bubble plume structures on high-frequency sonar performance

near the ocean surface
M V Trevorrow

414

The analysis of seismic data structure and oil and gas prediction
S Wang and C Lin

424

Mid-frequency bottom backscatter measurements and model/data comparison

Z Wang et al.

44 1

Acoustic propagation through bubbles: an exploration of the 1st and 2nd

moments in various flow conditions
T C Weber et al.

448

Multi-agents systems approach for underwater sound equations solution

V A Zuraisky et al.

464

The modal-vertical-beam (MVB) transmission loss analysis

H F Zhao et al.

475

1
CROSS HOLE SIMULATIONS IN ELASTIC FORMATIONS USING OFF AXIS
SOURCES VIA BEM
JULIETA ANTONIO AND ANTONIOTADEU
Department of Civil Engineering, Polo II, UniversiQ of Coimbra, Portugal
Two fluid-filled boreholes are used, one hosting the source and the other the receivers, to simulate
the cross hole seismic prospecting technique. The problem is formulated in the frequency domain
and solved using the Boundary Element Method (BEM). In this formulation the three-dimensional
solution is obtained as a summation of two-dimensional solutions for different wavenumbers along
the borehole axis. Each two-dimensional problem is obtained by using 2.5D Greens functions for
unbounded media. Responses in the time domain are computed using inverse Fourier transforms,
making use of complex frequencies to avoid the aliasing phenomena.
This model is used to identify the different wave pulses produced for sources and receivers placed on
and off borehole axes. This work also studies the wave propagation dependence on the distance
between boreholes and the type of elastic formation (fast and slow formations). Spectra responses are
computed to help understand the main physical features of the problem solved.

Introduction

Several geophysical and seismic prospecting techniques [8, 101 perform measurements of
the pressure inside a fluid-filled borehole, generated by a source either on the ground
surface or in another borehole. The complexity of the wave patterns recorded at the
receivers depends on the relative participation of the many wave propagation modes, or
normal modes, that may be excited by the source. The ratio between the shear velocity p
of the solid medium and the dilatational wave velocity of the fluid o+ defines two distinct
behaviors for wave propagation - when p>o+ the medium is called a fast formation;
otherwise it is named a slow formation.
Several numerical methods have been applied to obtain the solution of wave
propagation across and within fluid-filled boreholes. The Finite Difference Method [3],
the Boundary Integral approach [ 2 ] , the Boundary Element Method (BEM) [ 5 ] and
Hybrid methods [4] are among the techniques most often used.
In the present work the BEM is applied to model the crosswell seismic prospecting
technique, using two fluid-filled boreholes. The first borehole hosts a monopole source
while the receivers are placed in the second borehole. The full coupling between the solid
and the fluid has been taken into account, in the BEM model without any simplification.
The methodology used extends the studies previously performed by Tadeu, Godinho and
Santos 191. The model has been validated by solving the case of a single borehole, for
which analytical solutions are known.

Problem formulation

Two boreholes placed parallel to the z axis are driven in an unbounded elastic medium
with density p , where a shear wave and a compressional wave propagate with velocity

j? and a , respectively. The boreholes are filled with an inviscid fluid with density pf ,
where the compressional waves propagate with a velocity of af . A dilatational point
source, placed in one of the boreholes at position (xo, yo, zo), oscillating with a

2
frequency w , disturbs the medium, emitting an incident field that can be expressed by the
dilatational potential 4 ,

4 mc
. =

Ae

i-

"( af

r-J(x-xoP+(Y-ro)Z+(z-zo)2

J~~-~o)2+(Y-Yo)2+~z-ZO)2
When the geometry of the problem remains constant along the z direction, the threedimensional solution is frequently calculated as a summation of two-dimensional
problems, following the Fourier transformation of the problem in the z direction. This
summation is rendered discrete if there is a set of virtual sources, equally spaced along the
z direction [l]. The resulting discrete two-dimensional problems are solved for a varying

1
7

sequence of wavenumbers kw = ,-kzm

, with Imkw < 0, where k,, =-2rm is

L

the axial wavenumber, and L is the distance between virtual point sources equally spaced
along z. The incident field is given by

in which H f ) (...) are second Hankel functions of order n. The distance L must be large
enough to avoid contamination of the response by the virtual sources.
The BEM is a numerical technique suited to solve this problem, since it automatically
takes into account the far-field conditions, needing only the discretization of the boundary
of the two boreholes. The application of the method in the frequency domain requires the
evaluation of the integrals along the appropriately discretized boundary of the boreholes
H g ) k r - H(f)
(i, j = 1,2,3)
=J
x, ,n,

H,I')~' H , I ' ) ( ~ ~ , Mc,

CI

-J

fi

( kr~ilni)d~i

CI

In these equations, H t ) ( x k ,x, ,nl ) and G t ) (xk ,x, ) are, respectively, the Green's
tensor for traction and displacement components in the elastic medium, at the point x, in
direction j, originated by a concentrated load acting at the source point x k in direction i;

H ; { ' ( x k , x I , n , ) are the components of the Green's tensor for pressure in the fluid
medium, at the point x, as result of a pressure load placed at the source point x k ;

G ; { ) ( x k , x , ) are the components of the Green's tensor for displacement in the fluid
medium, at point x, in the normal direction, caused by a pressure load acting at the
source point x k ; n, is the unit outward normal for the Lrh boundary segment C, ; the
subscripts i, j = 1, 2 , 3 refer to the normal, tangential and z directions, respectively.
These equations are appropriately transformed from the Cartesian coordinate system
( x, y, z ) using standard vector transformation operators.
Time domain solutions are obtained applying an inverse Fourier transform to the
responses in the frequency domain. The necessary Fourier transformations are performed
by discrete summations over wavenumbers and frequencies, which is equivalent to adding

3
periodic sources at spatial intervals L = 2 a / Ak, (on the z-axis), and temporal intervals
T = 2 x / A w , where Ak, and Aw are the wavenumber and frequency increments,
respectively [l]. Complex frequencies with a small imaginary part of the form
w, = w - iq are used in order to avoid the aliasing phenomena. In the time domain, this
procedure is taken into account by applying an exponential window eql to the response
[7]. The source in the time domain is modeled as a Ricker pulse.

Numerical applications

Simulations are performed to understand the influence of the distance between the two
boreholes and the elastic formation properties in the propagation of different wave types
when the source is placed in one borehole and the receivers are located in a second
borehole. Four receivers equally spaced (10 m) along the z direction are placed on the
axis of the second borehole (receiver line R1, see Fig. 1).

Source

Receiver

a)

b)

Figure 1. Position of the sources (a) and receivers (b).

The examples presented in this work consider fluid-filled boreholes (radius 0.1016
m) located inside a fast or slow formation, as studied by Ellefsen 161, with the properties
listed in Table 1. Frequency and time results are displayed. The amplitude of the time
plots is normalized in relation to the maximum amplitude result presented.
Table 1.Formation properties.

Name
Fast formation
Slow formation
Fluid

3.1

Compressional
wave velocity
a = 4208 m/s

Shear wave
velocity
p = 2656 m / s

Density

a = 2630 m/s

p =1416

p =2250 kg/m3

af =1500

-------

m/s

m/s

p =2140 kg/m3

p r =lo00 kg/m3

Fast formation

Computations are carried out from 40 Hz to 20480 Hz with a frequency increment of 40

Hz, defining the time signatures up to T=1/40=25 ms. The frequency responses are
Fourier transformed to obtain pressures in the time domain. The source is modeled as a
spherical dilatational Ricker pulse with a dominant frequency of 6000 Hz. The ratio of the
wavelength of the incident waves to the length of the boundary elements is kept to a

4
minimum of 28.0. A minimum of 80 boundary elements is used to discretize each
inclusion.
Figure 2 shows the results calculated when the two boreholes are 20.0 m apart. Both the
time responses and the Fourier spectra representations are included. The Fourier spectra is
displayed in a gray scale, varing from white to black with increasing pressure amplitude.
The frequency domain plots reveal the existence of different wave modes propagating in
the dynamic system These modes correspond to the guided and body waves generated by
the excitation source. The modes in the Fourier spectra are identified by a pair of
numbers. The first number of the pair is the azimuth order, which indicates the variation
of the mode with the azimuth, whereas the second is the radial order that provides the
variation of the mode with radial distance.
I

I
z=30rn

n
m

z=20m

z=lOm
0.25 -

z=Om

0.0

I
10

15

20

25

Time (ms)

Figure 2. Spectra and time responses when two boreholes, placed in a fast formation, 20.0 m apart, are excited
by a monopole source placed close to the borehole wall at receiver line R 1.

The time plots show strong pulses related with waves traveling in the solid formation
with P and S body wave velocities (labeled P and S). It can be observed that the
magnitude of the P waves diminishes while the contribution of the S waves grows for
increasing distances between the receiver and the z plane of the source. After the arrival
of these P and S waves the response reveals a periodic behaviour, with a regular interval
between successive pulses, which corresponds to additional resonances of the dynamic
system. The Fourier spectra responses in Figure 2 show an enhanced response at
f l = 13165 Hz (labeled f l ) for low values of k , . For receivers placed on the same
plane of the source ( z = 0.0 m ), this mode creates a periodic behaviour whose periodicity
is expressed by Tl = 1/ fl = 0.076 ms Additionally, the analysis of the time plots shows
that the responses for receivers.placed in the vicinity of z=O.O m are not influenced by the
presence of guided waves. The individual contribution of the various guided modes to the
time responses is not easily distinguishable, given that the group velocities of the different
modes are similar.
Figure 3 presents the results achieved when the boreholes are 10.0 m and 40.0 m
apart. As predictable, the amplitude of the responses is lower when the distance between
the boreholes is greater. The time arrival of the labeled P and S waves corresponds to the
travel path between the two boreholes. The time pressures exhibit features analogous to
those detected when the boreholes are 20.0 m apart. However, the guided waves become
more significant as the distance between boreholes decreases (10.0 m), while they lose
importance when the distance is larger.

5
3.2

Slow formation

Figure 4 shows the time and Fourier spectra results calculated when the two boreholes
20.0 m apart are placed in a slow formation. The computations are performed in the
frequency range [25 Hz, 12800 Hz], with a frequency increment of 25 Hz, leading to time
signatures up to T=1/25=40 ms. The spatial distance between the virtual sources along the
z direction is set to L = 2Ta = 210.4 m . The spherical dilatational source is modeled as a
Ricker pulse with a characteristic frequency of 4000 Hz.

sI
P

.1,

z=30m

z=lOm

z=lOm

0.5
0.0
I
25

10

20

15

10

Time (ms)

15

20

25

Time (ms)

Figure 3. Time responses when two boreholes, placed in a fast formation, are excited by a monopole source
placed close to the borehole wall at Receiver line R2: a) Boreholes 10.0 m apart; b) Boreholes 40.0m apart.

s
12ow

spI

z=30m

lw00

0
a,

+.
._

3 6000

E
f 4000
2000

I
'0

L'
0.1 0.0

10

20

30
40
Kz (radlrn)

50

50

.
5

z=20m

z=lOm

z=Om

'

10

15

20

25

30

35

40

Time (ms)

Figure 4. Spectra and time responses, when two borek es, placed in a slow formation, 20.0m apart, are excited
by a monopole source placed close to the borehole wall, at receiver line R1.

Different wave components, occupying specific sub-domains, according to their

wave velocities, can be observed in the spectra plots. The modes in the Fourier spectra are
once more identified by a pair of numbers. Modes with azimuthal variation are visible in
the responses recorded at the line of receivers R1, placed in the center of the borehole
(Fig. 4). This happens because the axis of the second borehole is not the axisymmetric
axis of the dynamic system
The arrival of waves propagating with P and S body wave velocities in the solid
formation (labeled P and S) is clearly visible in the time plot. An additional pulse is also
distinguishable, particularly for z=30.0 m. This pulse results from waves traveling in the
elastic medium with the S wave velocity, which are critically refracted along the surface
of the second borehole as P waves (labeled in the time plot as SP). As observed for the

6
fast formation, after the arrival of the P and S waves the signal shows the presence of a set
o periodic pulses. The first mode at k , = O corresponds to fi =7839 Hz , which is
labeled in the Fourier spectra as f,.

Conclusions

The results obtained for slow and fast formations and for different borehole separations
allow to observe that the responses contain signals from non-dispersive waves and waves
produced by different modes. The influence of the guided waves in the response is
particularly important for large axial distances. The contribution to the response of each
guided mode is difficult to identify because they have similar group velocities.
References
1. Bouchon, M., Aki, K., Discrete wave-number representation of seismic-source wave
fields. Bulletin of the Seismological Society of America 67 (1977) pp. 259-277.
2. Bouchon, M., Schmitt, D. P., Full wave acoustic logging in an irregular borehole.
Geophysics 54 (1989) pp. 758-765.
3. Cheng, N., Cheng, C. H., Toksoz, M. N., Borehole wave propagation in three
dimensions. J. Acoust. Soc. Am. 97 (1995) pp. 3483-3493.
4. De Hoop, A. T., De Hon, B. P., Kurkjian, A. L., Calculation of transient tube wave
signals in cross-borehole acoustics. J. Acoust. Soc. Am. 95 (1994) pp. 1773-1789.
5. Dong, W., Bouchon, M., Toksoz, M. N., Borehole seismic-source radiation in
layered isotropic and anisotropic media: Boundary element modeling. Geophysics 60
1995) pp. 735-747.
6. Ellefsen, K. J., Elastic wave propagation along a borehole in an anisotropic medium.
Ph.D., Department of Earth and Planetary Sciences, M.I.T, 1990.
7. Kausel, E., Frequency domain analysis of undamped systems. Journal of Engineering
Mechanics, ASCE 118 (1992) pp. 721-734.
8. Krohn, C. E., Crosswell continuity logging using guided seismic waves. The Leading
Edge 11 (1992) pp. 39-45.
9. Tadeu, A., Godinho, L., Santos, P., Wave motion between two fluid filled boreholes
in an elastic medium. Journal of EABE -Engineering Analysis with Boundary
Elements 26 (2002) -.
IXI. 101-117.
10. Toksoz, M. N., Cheng, C. H., Cicerone, R. D., Fracture detection and
characterization from hydrophone vertical seismic profiling data. In EVANS, B . and
WONG, T., (Eds.), Fault mechanics and transport properties of rocks: Academic
Press, 1992, pp. 389-414.
,

BEM-Based nearfield holography for scattering sound field

Bao Xuemei

Ge Huiliang

Hangzhou Applied Acoustics Research Institute, Hangzhou, Zhejiang, People's

Republic of China
P.O.Box 1249, Hangzhou, China, 310012
E-Mail: sklbxm@163.net
Abstract: Near-field acoustic holography (NAH) is an inverse sound transform
method to reconstructed surface pressure and velocity from measured sound
nearfield. Holographic transformation method based on boundary element
method (BEM) is described in this paper. An effective scattering field
separation method is developed and singular value filter technique is used to
improve the reconstructed results. It's shown in this paper that incident field,
scattering field and local surface impedance can be obtained by use of
BEM-based scattering field NAH.

I. Introduction
Nearfield acoustic holography (NAH) is a method used for study of
radiation and scattering problem. In this method, the nearfield complex
pressure field (hologram) is measured for reconstruction of pressure field,
velocity field and sound intensity vector field"'41. The resolution of
reconstruction may be far smaller than wavelength for the high wavenumber
evanescent wave also recorded by hologram.
The work in this paper concentrates on the scattering field holography
when the scatter surface is not conformed to the hologram. The formulation of
BEM-based scattering field NAH is deduced and a scattering field separation is
used. The surface impedance is reconstructed by use of NAH and the
scattering field with arbitrary incidence can be predicted.

11. Formulation on BEM-based scattering field NAH'''

If a scatter is placed in an incidence, then the total field p,(r) can be
expressed as the sum of incidence pressure plnc(r) and scattering field

Where item eJd is omitted for simplification. The scattering pressure

p , (r) obey Helmholtz equation

where G(r,r) =

-jklr-r/

4n(r- rl

is free field Green function and Q is space

angle which is equal 2 n for convex smooth surface.

The integral equation (2) can be solved by boundary element method.
Assuming the surface S is separated as Mdiscrete elements and there are
Nnodes, then the coordinate and pressure of an arbitrary point r on m-th
element can be expressed as the interpolation of the coordinates and values of
the nodes attached on this element. The discrete expression of the integral
equation (2) is:

where { p s H ( r ) } N oisx l the scattering field on the measurement point of

hologram and there are N o measurement point on hologram, N dimensional
vectors (p,},,, and { ~ f } , , are
, ~ ~surface scattering pressure and normal
velocity and N i is the number of inner points.
The extra-determined linear equation group which is combination of
equation (3) and ( 5 ) is:

[;!kl= ]c:.l
D

And the linear equation group which links scattering iield on hologram and
surface velocity can be obtain by combination of equation (4) and equation (6):

{PH,(r)I,",l = [ w l N o x . {v: I N x l
where [ W ]= [El[:]

matrix

'[-I +

[ F ]and

'-I:[

(7)

is the generalized inversion of

[g].

N o must be larger than N to assure equation group (7) have a

unique solution and the solution is:

where [W]:O,, is the inverse or generalized inverse of matrix [W],,,

singular value composition of matrix [ W ]

. The

is

where + represents the complex conjugation transition of the matrix, C is

diagonal matrix which dimension is N x N and its element is the singular
values of matrix [ W I N O X
. The
, generalized inverse of matrix [W],Ox, is

We can reconstruct surface scattering velocity (v,S I N x I from complex

H

pressure data { p , (r)INoxl

by use of equation (9). The surface scattering
pressure can be calculated from surface scattering velocity with equation (3)
and the scattering field outside scatter can be calculated with

P,0-1= [ E l l , ,

[P"I,,l

+ [Fllx,

[v,"l N x l

(11)

111. BEM-based scattering separation technique

Since the measured field on hologram are always total field, a NAH

10

method which directly uses total pressure on hologram is given below.

First, we express the incidence as sum of a set of basic functions, i.e.

n=l

Where

n=l

are basic functions, An are undetermined complex

Y,(r)

coefficients that are separated in to undetermined yn and given w,. For

example, the incidence from a source in half-space z>O can be expressed as

n =1

Where k, and knyare x-direction and y-direction wavenumber correspond to

nth basic function and

,-/,
kz = { -

ki 2 ( k i +kiy)
k i I ( k k +k:y)

(14)

Where k, = o l c is free field wavenumber and c is the sound speed of the

fluid outside the scatter.
The incidence pressure on hologram can be expressed as following
equation with equation (1 3),

b>

{pic}=[HIN~~Nj,,c

Where, the elements of matrix [H]NOxN,nc

are h, = w,Y, (r:)

(15)
and r: is ith

measurement point on hologram. The total pressure on hologram can be

expressed as:

If N o 2 N

+ N,,

,the solution of equation (16) is

11

(17)

Where

([wl

[H])j(llox(N+N,,,)

is

the

generalized

inverse

of

The surface scattering velocity can be reconstructed from total pressure

field on hologram by use of equation (17). The undetermined coefficients can
also be solved and the incident pressure can be calculated with equation (12).
As an example, we consider a rigid sphere in the incidence of point source.
The surface of sphere is divided into 72 elements and there are 62 nodes. The
hologram is two parallel plane out the sphere (as shown in fig. 1 and table 1). A
singular value filter technique [41isused in reconstruction to filter out noise
polluted high wavenumber modes. The relative magnitude error

zs,,,,,,
and

average phase error E ,

are used to evaluate the accuracy of reconstruction, where psiand p r jare

reconstructed pressure and theoretical pressure on ith surface node.

Fig. 1 The scatter and hologram

12

Case

Shape of hologram

Nodes

Two plane ( zo = 0.3A,

128

L, = L ,

Division of
hologram
11*11

=A)

The source is a point source which coordinate is (O,O,l), we use equation (13)
to express the incidence. The wavenumber range in incidence expression is

-k, <k, < k o and - k o <k, < k o . The number of wavernumber sample
pointsis N,,, =10x10=100.

Table 2 Reconstruction results

The reconstructed surface pressure error ESPM and ESPPis given in

table 2. The separated incidence pressure error

E,,

and E,,

are also

given. The reconstructed results with theoretical scattering field hologram are
also given for comparison. It's shown from table 2 that the reconstruction error
with total pressure hologram is a little larger than those obtained with
theoretical scattering pressure.

IV. Reconstruction of surface impedance of scatter and

prediction of scattering field
The incident velocity can be calculated with equation (1 3) and
reconstructed undetermined coefficients:

Where r, is coordination of sth surface node and n, , n s y ,n,, are three

13

components of the normal vector of sth node. The local impedance of surface
node is:

The scattering field can be predicted by use following equation:

Where [ZINxN
is local impedance matrix.
The simulation results of surface impedance reconstruction is given in
Fig.2. The scatter is a impedance sphere which radius is a = 0.22. The local
surface impedance of the sphere is 4 pc . The hologram are two parallel planes
at zH = 0.32 and zH = -0.32. The size of hologram is 2 x 2 and the
space of measurements points is 0.11. The sphere surface is divided as 160
elements and there are 146 nodes on the surface.
The reconstructed impedance when the source is a point source at

z,

= 102/ 3 is given in fig. 2(a). It is shown in fig. 2(a) that the maximum

error is about pc .

(a) single source

(b) 8 sources
Fig. 2 Reconstruction results of surface impedance
During investigation, it's found that the error of reconstructed impedance
can be reduce if we use multi-source to "illuminate" the scatter. The results

14
with 8 sources which location are

x = f l OA / 3, y = +lo2 / 3, z = f l O / 2 / 3

is given in fig. 2(b). It is shown from the figure that the real part of
reconstructed impedance is from 3.7 pc to 4.5 except two top points.
The results of prediction of scattering field in plane wave incidence with
reconstructed impedance (fig. 2 (b)) are given in fig. 3. Its shown from the
figure that the predicted far field direction patterns fit the theoretical results
very well.
Duechon ofplane

270

270

(4
I__

Theoretical results

(b)
0

Prediction

Fig. 3 The direction patterns predicted with reconstructed impedance (the

hologram is measured under 8 sources)

V. Summary
The BEM-base NAH method and BEM-based scattering field separation
method are given in this paper. The surface field of the scatter can be
reconstructed from total pressure measured in the near field of scatter.
Furthermore, the surface impedance can be reconstructed and the scattering
under arbitrary incidence can be predicted. The methods are verified by
simulation.
Reference:
1. J.D.Maynad and E.G.Williams, et. al. Nearfield acoustic holography I:
Theory of generalized holography and development of NAH. J.Acoust.
SOC.Am. 78(4), 1985:1395-1413
Konji
SaijYou, Mirsuhiro Ueda and Whigeru Youshikawa, Application of
2.
generalized near-field acoustical holography to scattering problem. Jpn. J.
Appl. Phys. 33,1991:3120-3135

15

3.

4.

5.

Ming-Te Cheng, J. Adin Mann 111 and Anna Pate. Wave-number domain
separation of the incident and scattered sound field in Cartesian and
cylindrical coordinates. J.Acoust. SOC.Am. 97(4), 1995:2293-2303
W.A.Veronesi and J.D. Maynard. Digital holographic reconstruction of
source with arbitrarily shaped surfaces. J.Acoust. SOC. Am.
85(2), 1989:588-598
Investigation in acoustic holography reconstruction of scattering field of
target, Chinese Journal of Acoustics, 2000, 19(1)

16

PARABOLIC EQUATION TECHNIQUES FOR PROPAGATION AND

SCATTERING
DAVID C. CALVO, MICHAEL D. COLLINS, DALCIO K. DACOL AND
JOSEPH F. LINGEVITCH
Acoustics Division, Code 7145, Naval Research Laboratory, Washington,DC 2037s USA

This contribution first gives an overview of the wide-angle parabolic equation (PE) technique for
solving non-separable wave propagation problems in the ocean and atmospheric science areas. The
principles are described and applications in these areas are briefly discussed. Recent advances made
by the authors in applying the PE technique to the obstacle scattering problem, specifically the socalled on-surface radiation condition approach [Kriegsmann el al. IEEE Trans. Antennas Propagat.
19871, are then discussed.

Introduction

The parabolic equation (PE) method is based on simplifying the two-way elliptic wave
equation to a one-way parabolic equation that is then solved by marching in a preferred
direction. Significant computational savings is obtained when this reduction can be
made. The method was pioneered in the 1940s by Leontovich and Fock in modeling
studies of radio-wave propagation in the atmosphere [ 1,2]. Applications in seismology by
Claerbout [3,4] and in ocean acoustics by Tappert [ 5 ] later appeared in the 1970s marking
the beginning of the widespread use of the parabolic equation method.
The PE technique is especially useful for solving range-dependent (non-separable)
waveguide problems in two dimensions and global-scale propagation problems in three
dimensions. In comparison with calculations using ray theory, no special treatment is
needed near caustics and more accuracy is generally obtained at lower frequencies. The
technique is also computationally less expensive in comparison with coupled-mode
techniques for range-dependent problems [ 6 ] .

The Parabolic Equation Method

For many problems the time-harmonic wave field satisfies the two-way elliptic Helmholtz
equation

whcrc

17
The wavenumber k ( x , z ) , with k, being a reference value, allows for both range (x) and
transverse ( 2 ) dependence of the wave speed profile. Assuming the wave speed varies
gradually in range, (1) can be factored into a product of incoming and outgoing operators:

Neglecting reflected energy, one may work with the simpler one-way parabolic wave
equation:

In numerical applications, boundary conditions in the transverse direction are taken into
account when constructing the discrete approximation of the transverse operator. Three
issues arise in applying the PE: how to march efficiently in range, how to generate the
starting field given some localized source, and how to numerically implement the
(pseudo-differential) square-root operator. These have been studied in detail in the past,
and we will only briefly discuss them here.
The field is propagated by first dividing the domain into a sequence of rangeindependent steps; equation (3) may then be solved exactly over a single range step:

Taking into account (3) and (4) as well as the jump condition across a delta function
source located at a transverse distance z, the starting field may be determined using the
self-starter:

W(X0)

(I

+xy2

Finally, the exponential function of the square-root operator in (4) can be approximated
using a rational function:

where

The coefficients of X have been computed for specific values of the range step using
least-squares or derivative matching techniques taking into account constraints needed for
stability of the marching scheme [7]. The self-starter has been implemented in a similar
way [8]. This technique for marching in range is known as the split-step Pad6 method
and allows large range steps to be taken in practice.

18
It is noted in passing that the majority of studies using the parabolic equation technique
have approximated the square-root function by a Taylor expansion at X = 0:

This expression, with the ikoly term removed, is in fact the Schrodinger equation that is
usually obtained directly from (1) by a carrier-wave expansion about k,. It is most
accurate for narrow-angle propagation angles. The rational approximation of the full
square root, however, is more accurate for problems in which energy propagates at wide
angles relative to the paraxial direction.

3
3.1

Application to Propagation Problems

Ocean acoustics

Because it is accurate for large range steps, the split-step Pad6 method has been
particularly useful for global-scale ocean acoustics problems in which variations of both
the sound speed and bathymetry are encountered. Figure 1 shows results of applying the
PE to simulate propagation in a two-dimensional (range-depth), 18,000 km path
extending from a 57 Hz source at Heard Island in the south Indian Ocean to a listening
station off San Diego, California. This path (nearly a great circle) was determined by ray
theory after confirming that horizontal multi-paths taken by different acoustic modes were
not critically important. Simulated mode interaction, caused mostly by island features off
New Zealand, was found to be in reasonable agreement with received signals [9].
The influence of bathymetry on transmission loss is very important in shallow-water
environments because of the possibility of energy propagation into the sediment. The
conversion of pressure (acoustic) waves into shear (elastic) waves which occurs at the
water-sediment interface, however, requires using the vector elasticity equations which
are not factorable in their primitive form and can lead to numerical instabilities if
evanescent modes are not handled properly. These difficulties have been overcome, and a
parabolic equation coupling an acoustic water column to an elastic sediment has been
developed [lo]. As an example, Figure 2 shows propagation in a sloping beach area and
demonstrates the gradual cut-off of acoustic modes in the water column and the resulting
conversion of energy into shear waves which then propagate into the sediment. A
parabolic equation has also been developed which models the sediment as a poro-elastic
medium using Biot theory [ 111.

19
Mode 1Path

.ModeIPath PE TL

Modal &citation vs Range

Range
Figure 1: Mode 1 path, transmission loss, and modal excitation computed using the parabolic equation method
to simulate the Heard Island experiment.

Figure 2: Acousto-elastic propagation in a sloping beach area. Compressional wave transmission loss is shown
on the left, shear wave transmission loss on the right.

20
3.2

Gravity and acousto-gravity waves

Using techniques similar to those developed for ocean acoustics, wide-angle PEs have
been developed for two-dimensional internal gravity (buoyancy) waves in the ocean and
acousto-gravity waves in the atmosphere. Figure 3 shows the propagation of internal
waves in the ocean in the presence of sloping bathymetry which models a continental
shelf region. The initial condition consists of a realization of the Garret-Munk internalwave spectrum. Interestingly, the reflection angle of the modal rays is independent of the
angle of incidence with the slope, and reflection occurs as if from a flat surface-a
property of internal waves. The presence of tidal generated internal waves in the ocean
can significantly modify the sound speed near shelf-regions and can therefore affect the
coherence of acoustic measurements. This is currently an active area of acoustical
research.
At low (infrasonic) frequencies in the atmosphere, gravity waves and acoustic
waves are coupled. A two-dimension1 acousto-gravity PE has been derived that takes into
account a vertical distribution of the Vaisalla-Brunt frequency and the density [ 121.
Figure 3 also shows the results of propagating a Lamb wave (the lowest order acoustogravity wave mode) over a mountain range. Because of the variable topography, mode
coupling takes place, and a significant wave disturbance can be seen in the upper
atmosphere in the lee of the mountain. Such disturbances can be hazardous to aircraft.

3.3

Atmospheric acoustics in the presence of wind shear

The propagation of sound waves in the atmosphere can be substantially modified by the
presence of variable wind as a function of altitude (wind shear). Recently, a parabolic
equation valid for wide-angle propagation and high-Mach number winds has been

Figure 3: Transmission loss for internal-gravity waves in the ocean over sloping bathymetry (left) and for a
acousto-gravity waves over sloping mountain topography (right).

21

developed for this case [ 131. Derivation of this equation presented challenges due to the
effects of commutators encountered in factoring the Helmholtz equation. A typical
distribution of wind and sound speed is shown in Figure 4. Placing a 0.1 Hz source at an
altitude of 10 km,transmission loss results using this new PE with and without wind are
shown in Figure 5 . The presence of wind is seen to have a dramatic effect.

Wind speed (Mach)

Figure 4: Wind and sound speeds as a function of altitude for the windy-parabolic equation results of Figure 5.

Loss (dB)
tO

10

.,
I .

<O

Figure 5: Transmission loss for a 0.1 Hz source placed 10 km above the ground. Top picture is with wind,
bottom picture is without wind.

22
3.4

Global-scale problems: The adiabatic-mode parabolic equation

For geophysical problems it is often desirable to obtain a synoptic picture of how acoustic
energy is propagated over great distances in any horizontal direction. The adiabatic-mode
PE, which neglects coupling of energy between modes, can be used for this purpose.
Given knowledge of the local eigenvalues of the depth modes, horizontal wave equations
for the mode coefficients are derived. The parabolic approximation is then applied to
these equations [ 141.
Results of placing a 1 Hz acoustic source south-east of the Hawaiian Islands is shown
in Figure 6. Although acoustic intensity is only shown in the water, the computation in
fact marches over both land and sea from the source to the antipode. A close-up of the
figure shows the extent of the diffraction pattern in the wake of the Hawaiian Islands.
Further progress has been made by allowing for mode-coupling and advection due to
variable currents.
The adiabatic mode PE with advection was applied to study the propagation of
energy in the atmosphere of Jupiter upon impact of the fragments of comet ShoemakerLevy 9 in 1995. Focusing of acoustic energy into intense caustics caused by the
latitudinal distribution of winds was predicted [ 151. The parabolic equation method has
provided an effective means for carrying out calculations for these problems that would
otherwise be extremely time consuming using full-wave modeling.

Figure 6: Acoustic intensity of a 1 Hz source computed using the adiabatic mode PE (top) with a close up
showing diffraction around the Hawaiian islands (bottom).

23
4

An Application to Scattering Problems: The On-Surface Radiation

Condition

An interesting application of the PE concept is the so-called on-surface radiation

condition (OSRC). A radiation condition is a parabolic equation applied at the boundary
of a computational domain which expresses that energy should be entirely outgoing.
Kriegsmann et ul. [16] noticed that if a radiation condition were applied directly on the
surface of a compact scatterer then a good estimate of the unknown surface quantity could
be obtained and used to compute the far-field pressure using the Helmholtz integral
equation. The OSRC can therefore be viewed as an approximation of the exact non-local
Dirichlet-to-Neumann surface operator which can be obtained formally from the surface
Helmholtz integral equation.
Recently, Calvo et al. have used rational approximations to create a wide-angle
(relative to the surface normal) version of the OSRC for the two-dimensional obstacle
scattering problem that is more accurate than the narrow-angle OSRC for challenging
cases [17]. Specifically, previous OSRCs have had difficulty in treating the case of endon incidence of a plane wave with a hard ellipse: a case in which creeping waves, which
graze along the body, make an important contribution to the solution. The wide-angle
OSRC takes these effects into account in an approximate way.

4.1

OSRC for Two-dimensional Obstacle Scattering

The basis for the wide-angle OSRC is pseudo-differential representation of the Dirichletto-Neumann map for a circle [ 171:

where

The above result can be obtained from the separation of variables solution after
replacing the index of the Hankel function with the square root of minus the tangential
operator. Considering a circle of radius a , from (9) it is apparent that two parameters are
important: the scattering angle relative to the surface normal and the surface curvature.
For small-curvature, a Taylor expansion of (9) yields the original Kriegsmann et uf.
OSRC

This is a good approximation for narrow scattering angles but becomes non-uniform for
large scattering angles.
To obtain a uniformly valid expression, rational approximations of (9) can be

24
introduced in terms of a scaled tangential operator. Neglecting curvature effects, (9)
tends to the familiar square-root operator seen before

which can also be obtained by factoring the Helmholtz equation in cylindrical coordinates
after neglecting the curvature term. Computations using rational approximations of (1 1)
have indicated, however, that for wide-angle accuracy curvature, even when small, must
be taken into account.
There are two options for proceeding. A multi-term rational approximation of (9)
may be proposed in which the coefficients in the approximation depend on curvature.
This is carried out in Calvo et al. and a four-term rational approximation valid down to
ka=1 is reported there. Here we follow a simpler approach and propose a rational
approximation of (9) in the form

LAE L,(ka)

(1+ av2+ pv4

1 + yV2

where

and

Q, k ~ (0.4043+0.7003i)
+
( k ~ )+
(-0.0727+0.1260i)
~
(k~)-~.*..

v,

The expression
corresponds to the complex zero of (9) which is closest to the real axis
[ 181. This in turn corresponds to the dominant creeping-wave pole excited in scattering
by a hard circular cylinder. By matching (9) and (12) with respect to the zero location
and the slopes at V = 0 and V = 1, the coefficients in (12) are given by

where

C(ka)=

[1- (Q,/ k a ) 2 ]

[1+ (0.1348+0.23341)(ka)-IP]

1
1
L,(ka) = i - -+
2ku 8 ( k ~ ) ~
- * * a .

In line with the work of Kriegsmann et al., this OSRC may be applied to a general
convex, non-circular body if the local arc-length is approximated by the arc-length of its
osculating circle. Accordingly, the replacements

25
are made where n is the outward surface normal, s is the arc-length, and K is the local
curvature. These replacements imply
1

a2

F.

+ (0.4043 + 0 . 7 0 0 3K23
i ) p + (-0.0727 + 0 . 1 2 6 0 9 +.
~ ..
K43

as2

(15)

in (12). If curvature varies slowly with respect to a wavelength, derivatives of curvature

may be neglected.
The wide-angle OSRC has the advantage of being non-local for both hard and soft
boundary conditions: an evaluation of (12) requires solving a boundary-value problem
around the obstacle with periodic boundary conditions because of the differentiation
operator in the denominator. In addition, computational expense scales only linearly with
frequency owing to the use of tri-diagonal matrices whereas the cost of most other
methods scale with the square of the frequency. In the following we compare both wide
and narrow-angle OSRC performance with results obtained using the more
computationally expensive T-matrix and boundary-element methods.
Figure 7 shows the far-field directivity pattern of the scattered field for the case of
end-on incidence of a plane wave with a hard ellipse of relatively large aspect ratio. As
mentioned, given the gradual transition between the lit and shadow regions, creeping
waves are important in these cases, and the physical optics technique (Kirchhoff
approximation) does not provide good results in this frequency range [17]. The
improvement using the wide-angle OSRC is exceptional.

k=20

Hard cllip\c

+
,

d=5

Figure 7: Far-field directivity patterns for acoustic scattering by a hard ellipse as shown above using the
narrow-angle OSRC (10) (left) and the new wide-angle OSRC (right). The aspect ratio is d=5 and the
wavenumber is k = 20.

26
4.2

Three-Dimensional Convex Bodies

For a three-dimensional convex body, the narrow-angle OSRC is given by 1193

where K1 and K2 are the principal curvatures along the orthogonal arc-lengths Sl and S 2 .
The three-dimensional wide-angle OSRC may derived by making the following
replacements in the two-dimensional wide-angle OSRC 1201:

with
K~ = K, cos2 q~

+ K, sin2q ~ ,

where $J is the angle the incident field makes with the principal K1 direction.
The sphere is considered as the first test case, and comparison between the narrow
and wide-angle OSRC and boundary-element method results is shown in Figure 8. The
wide-angle OSRC gives dramatically better surface pressure results in the shadow region
of the sphere, although both narrow and wide-angle OSRCs give reasonably good results
for the directivity. Finally, scattering by a prolate-spheroid is considered and directivity
results are shown in Figure 9. Computations for oblique angles of incidence also yield
good results and will be reported in an upcoming publication.

Other Current and On-Going Research

On the propagation side, advancements have been made in applying the wideangle PE method to seismic waves in layered elastic media with large variations in layer
properties and all types of waves [21]. Recently, advancements in the split-step Pad6
solution have been made to increase efficiency at higher frequencies drawing on attributes
of the split-step Fourier approach. The new split-step PadC-Fourier solution allows an
accurate solution to be obtained while still using large range steps [ 2 2 ] .
Acknowledgment
This work was supported by the Officer of Naval Research. D.C. Calvo was partly
supported by the National Research Council.

27

Total surface pressure

Far-field directivity

Figure 8: Total-surface pressure and far-field directivity patterns for acoustic scattering (ka = 10) by a hard
sphere as shown above using the narrow-angle OSRC (dash) and the new wide-angle OSRC (dash-dot). The
reference solution (solid) is computed using the boundary-element method.

Figure 9: Far-field directivity for scattering by a prolate spheroid (k=20, d=5). Left figure shows narrow-angle
OSRC results: right shows wide-angle results. The reference solution (solid) was computed using the boundaryelement method.

28
REFERENCES
Leontovich M.A. and Fock V.A., Solution of the problem of propagation of electromagnetic waves along
the earths surface by the method of parabolic equation, J. Exp. Theor. Phys. 16 (1946) pp. 557-573.
Fock V.A., Electromagnetic Dijffraction and Propagation Problems (Pergamon, New York, 1965).
Claerbout J.F., Coarse grid calculations of waves in inhomogeneous media with application to delineation
of complicated seismic structure, Geophysics 35 (1970) pp. 407-418.
Claerbout J.F., Fundamentals of Geophysics Data Processing (McGraw-Hill, New York, 1976).
Tappert F.D., The parabolic approximation method. In Wave Propagation and Underwater Acoustics, ed.
by Keller J.B. and Papadakis J.S., Lecture Notes in Physics 70 (Springer, New York, 1977) pp. 224-287.
t61 Jensen F.B., Kuperman W.A., Porter M.B. and Schmidt H., Computational Ocean Acoustics (American
Institute of Physics, New York, 1994), pp. 343-412.
Collins M.D., A split-step Pad6 solution for the parabolic equation method, J. Acoust. Soc. Am. 93 (1993)
pp. 1736-1742.
Collins M.D., A self-starter for the parabolic equation method. J. Acoust. Soc. Am. 92 (1992) pp. 20692074.
[9] McDonald B.E., Collins M.D., Kuperman W.A. and Heaney K.D., Comparison of data and model
predictions for Heard Island acoustic transmissions, J. Acoust. Soc. Am. 96 (1994) pp. 2357-2370.
[ 101 Collins M.D., Higher-order Pad6 approximations for accurate and stable elastic parabolic equations with
application to interface wave propagation, J. Acoust. Sac. A m 89 (1991) pp. 105Ck1057.
[I 11 Lingevitch J.F. and Collins M.D., Wave propagation in range dependent poro-acoustic waveguides, J .
Acoust. SOC.Am. 104 (1998) pp. 783-790.
[12] Lingevitch J.F., Collins M.D. and Siegmann W.L., Parabolic equations for gravity and acousto-gravity
waves, J. Acoust. SOC.Am. 105 (1999) pp. 3049-3056.
[I31 Lingevitch J.F., Collins M.D., Dacol D.K., Drob D.P., Rogers J.C.W. and Siegmann W.L., A wide angle
and high Mach number parabolic equation, J. Acoust. SOC. Am. 111 (2002) pp. 729-734.
[ 141 Collins M.D., The adiabatic mode parabolic equation, J. Acoust. SOC. A m 94 (1993) pp. 223G2278.
[I51 Collins M.D., McDonald B.E., Kuperman W.A. and Siegmann W.L., Jovian Acoustics and Comet
Shoemaker-Levy-9, J. Acousr. Soc. Am. 97 (1995) pp. 2147-2158.
[I61 Kriegsmann G.A., Taflove A. and Uniashankar K.R., A new formulation of electromagnetic wave
scattering using an on-surface radiation boundary condition approach, IEEE Trans. Antennas Propagat.,
35 (1987) pp. 153-161.
[I71 Calvo D.C., Collins M.D. and Dacol D.K., A higher-order on-surface radiation condition derived from an
analytic representation of a Dirichlet-to-Neumann map. IEEE Trans. Antennas Propagar. 51 (2003) pp.
1607-1614.
[I81 Felsen L.B. and Marcuvitz N., Radiation and Scarrering of Waves (Prentice-Hall, Englewood Cliffs, New
Jersey, 1973), p. 691.
[19] Jones D.S., An improved surface radiation condition, IMA J. Appl. Math. 48 (1992) pp. 163-193.
[20] Calvo D.C., A wide-angle on-surface radiation condition for three-dimensional convex bodies, in
preparation.
[21] Jerzak W., Collins M.D., Evans R.B., Lingevitch J.F. and Siegmann W.L., Parabolic equation techniques
for seismic waves, Pure and Applied Geophysics 159 (2003) pp. 1681-1689.
[22] Kiisel E., Espinoza F.A., Siegmann W.L. and Collins M.D., The split-step Pad6-Fourier solution, J.
Acoust. Soc. Am. 112 (2002) pg. 2393.

29
SEAFLOOR PROPERTIES AND SEGMENTATION
G. CANEPA, N. G. PACE, E. POULIQUEN,
P. FRANCHI, R. LOMBARDI, C. SISTI
N A T O Undersea Research Centre
Viale San Bartolomeo 400, I-19138 La Speda, Italy
E-mail: canepa@saclantc.nato.int

T. AKAL
TUBITAK-MAM, Marmara Research Center
Earth and Marine Sciences Research Institute, P.K. 21 Gebze, Kocaeli 41470 Turkey
The interest here is in using acoustic signals t o divide the seabed into regions
that are relevant in the context of MCM operations and extracting those features,
which drive the process. This is termed segmentation followed by classification. A
recently developed seabed segmentation algorithm is presented: SESAM (SEafloor
Segmentation AlgorithM). It is based on the analysis of angular dependence of the
acoustic backscatter. Such information is readily acquired with multibeam systems.
The algorithm will be described together with illustrations of segmentation results
from an area that has been very well characterized. Comparison of segmentations
with ground truth data (grabs, penetrometers, stereo photography, sidescan sonar)
is used to endorse the algorithm and confirm that the identified regions are related
to seabed features of relevance.

Introduction

We present further validation results of the SESAM

supported by
acoustic and non-acoustic data acquired during the CLUTTER'OS cruise, of Marciana Marina, Elba Island, Italy. Data from a 120 beam SIMRAD EM3000 operating a t 300 kHz was used; contemporaneously, ground truth data were acquired
using high frequency sidescan, grabs, expendable bottom penetrometers (XBP)
and stereo-photos. When the segmentation maps produced by SESAM were compared with ground truth data there was good agreement. However, the Marciana
Marina site presented three features that challenged the segmentation algorithm
and are described in the following. Two important aspects of the segmentation
algorithm have been improved. The accuracy of the Backscattering Strength (BS)
data has been improved by calibrating the multibeam echosounder (EM3000) with
respect t o the transmit beam' and then taking into account the way the supplied
echosounder software calculates the BS (the calculation technique changes when the
beam-seafloor incident angle is, approximately, 30"). The algorithm that filters the
processed BS data has also been improved to allow the treatment of a high number
of samples in a relatively short time while concurrently increasing the stability of
the results.
The following word definitions are used:
Segmentation: separation into N categories of the studied area. Each category
corresponds to a particular acoustical and/or geo-acoustical characteristics.

30

m
r-

0
0

m
N

0
Lo

(u
(u

0
0
0
N
0

b
-n

00

2%
0 3
m

.e

2m
Iz

g--o

z>
a
.F s

b-

0
0

vm

0
0

0
0

0
0
0

Figure 1. The segmentation of Marciana Marina obtained using SESAM is shown together with
an artificially illuminated bathymetric map. With the help of ground truth information, it is
possible to obtain a classification map of the site. Blue is used for muddy/silty classes, green for
presence of sea grass while gravel/rocky classes are red. Bathymetric contour lines (yellow) are
indicated.

31
0

Classification: assignment to each category a class name (e.g., clay, silt,

sand, or soft, medium hard , hard, or low clutter, high clutter,
etc).

Characterization: assignment of a physical and quantitative description to

each category or class (e.g., average grain size, RMS roughness, average sound
speed, average density, amount of clutter, etc).

SESAM basics

SESAM uses the variation of BS Versus beam-seafloor incident Angle (BSVA,

Fig. 3) to segment different seafloor regions by comparing the behaviour of the
couple BS/incident angle with a set of typical BSVA variations selected by the operator l>z. The first step of the segmentation process consists of combining the BS
and the beam-seafloor incident angle data from the multibeam data. The BS data
are calibrated, outliers removed, corrected for local slope and gec-referenced. The
BS data are then geographically displayed for the operator to select areas of typical
BSVA characteristics.
The information contained in the BSVA curves is affected by numerous geoacoustic and geometrical parameters:

At a given frequency, water-sediment interface scattering and scattering from

sediment volume heterogeneity compete to produce a complex BSVA, with
frequent dominance of interface scattering at very high frequency.

At normal incidence (0) the BS is strongly dependent on the water seafloor

impedence contrast (2. e. , water-sediment interface dominance).
At very high frequency (e.g., 300kHz) and away from specular, the BSVA
decay is primarily dependent on seafloor roughness: typically the lower the
decay, the higher the RMS roughness.

At 300 kHz, if seafloor roughness is low, a slight specular reflection peak is

observable on the BSVA curve.
It is possible to obtain similar BSVA curves for different geological compositions of
the seafloor: the same impedance can be obtained using mixtures of sand, gravel,
silt and clay; roughness can be modified by the presence of ripples or discrete scatterers. Similarly, identical geological compositions can give different BSVA depending on roughness: this is what happened at Marciana Marian site (see Section 3).
The average seafloor grain diameter, obtained from ground truth grabs, is strongly
correlated with BS at normal incidence (Tab. 1) an important parameter for subsequent classification. Iterations may be required to eliminate categories (e.9. for
a rippled sandy seafloor surveyed along differing tracks) or to add one or more new
categories. In particular, ripple direction seems to influence only the mid-angle part
of the BSVA.
The comparison between the BS/incident angle dataset and the set of typical
BSVA patterns can be performed using different methods. At present a distance

32
criteria is applied by SESAM. Each ping is treated individually using Bayesian
probability. The probability that the ping belongs to a given category is estimated.
Then, the probability that each individual BS belongs to that category is estimated.
The product of the two probability produces a spatially non-filtered segmentation.
A spatial filtering technique that can be applied to the Bayesian probability associating the nearest data points has been implemented to produce more stable
results '.
Experimental validation

This section presents the application of SESAM to data acquired during the CLUTTERO3 cruise. Segmentation is usually performed without a priori knowledge: segmentation and bathymetric maps are produced for comparison with other acoustic
and non-acoustic data acquired at the same site. This allows the improvement of
the segmentation results and the transition to classification and/or characterization.
Figures 1-2 show the segmentation and bathymetric results obtained at the site
with super-imposed coloured dots corresponding to the positions of stereo-photos,
grabs and expendable Bottom Penetrometers (XBP). Details of the ground truth
information are given in Tab. 1. With the help of ground truth information, it is
possible to obtain a classification map of each site. Blue is used for muddy/silty
classes, green for presence of sea-grass and red for gravel/rocky classes.
The Marciana Marina site has three challenging characteristics for the SESAM
segmentation algorithm (see Figs. 1-2):
The zone of silt at longitude between 250-1000 and latitude over 1700, is composed of two unmixed different materials, one of which is never found isolated.
To separate the patch, it is necessary to manually trace the BSVAC over the
data. The graphic user interface was improved to include this feature: for the
patch zone of Fig. 1 the BSVA curve of Fig 4 was obtained.
0

The hard zone found around the point (2200,1200) is composed of rocks over
sandy silt (generated by the mouth of a small stream). The presence of sparse
rocks over sandy silt produces a resulting BSVAC (class 5) which resembles to
other sites BSVAC classified as gravel site (transition between rock and sand).
The longitude meridian 2750 lays over two classes of seafloor: class 3 and 4.
Class 3 is, near shore, a dense Posidonia seafloor class (Fig 3.a). Considering
that dense Posidonia can be found only to z 30m, the seafloor classified as
class 3 at depth greater than 30m there cannot be Posidonia, even if it is
equivalent from an acoustic point of view (at 300 kHz). Stereo photos (Fig 3.c)
and grabs show that this seafloor is sandy-silt (q5 = 0.83 - -3.82) with patches
of sea-grass. On the other hand, class 4 is also similar, from an acoustic point
of view, to class 3 (Figs 3.b and 3.d).

The problem is caused by very similar acoustic response of the silt present at
Marciana Marina (the roughness of which is increased by patches of sea-grass),
and the sea-grass of Figs 3.a and 3.b acoustic response. This is the first site to
present such characteristics: to resolve this kind of uncertainty work is proceeding

pa-tmboe ji deui uoiirejuamSas sq:j

uo prejaaAO a.re (sa[3uBU^) aa^amojiauad uio^oq ajqepuadxa pu^e (sajisnbs) qBjS '(sapjp) so}oi]d
oaja'js :a^is mna^ puno.i i^res ^B UMBJP si sjoqui/Cs ^uajagip I^IM. ssojo v '(mSp altl oo) sS^euii
Suipuodsajjoa aq^ puB (ya[ aq; uo) uoi^uauiSas ^uuiep\[ BU'epaBj^ -g ajnSij

Relative latitude [m]

Relative latitude [m]

34

BSVA curve

Photonr a Dhv

:-20
5
c

-30
-a0

-50

-60

10

20

30
Incident on+

40

50

10

20

40

50

=
",I -10

B
5

-20
-30

E -40

30
Incident angle

-50

-60

-70

10

20
30
Incident angle

40

50

10

20

40

50

-2 -20
-

-30

E -40

30
Incident angle

Figure 3. The photos and the acoustic response of four different seafloor. All the BSVA curves
are very similar but the first and the last are covered with thick sea-grass while the other two
have the similar mean grain size and a high roughness due t o the patch of sea-grass.

35

Figure 4. Mouse constructed BSVA curves for a two distinct zones seafloor.

on segmentation of lower frequency multibeam sonar (EM300 and EM1000 Simrad

sonar, at 30 and 95 kHz respectively).
Except for class 3-4 composition of three different seafloors, the results of the
segmentation are spatially consistent. In particular, they show the patches of class
0 seafloor inside class 1 and the presence of gravel where rock is protruding as
shown by illuminated bathymetry (right of Fig. 1) and sidescan sonar image (right
of Fig. 2). SESAM classification also shows the evolution of soft silt to harder
bottom of class 0 to class 2.
In Fig. 1 the bathymetry is overlaid on the segmentation map: in places the segmentation map is correlated with bathymetry. The dynamics of sedimentation b e
ing dependent on currents and grain size may explain the presence of harderlcoarser
sediments in the shallower areas and softer/finer sediments in the deepest areas.
Also, Posidonia is only found at depths between 12 and 30m.
3.1

Ground truth

A summary table of the class analysis is given in Tab. 1. The geo-acoustic properties and stereo-photos are consistent with the underlying segmentation categories.
Similarly, XBP results (Tab. 1) are in accordance with the segmentation. Stereophotographs were acquired during CLUTTER'O3 by two cameras mounted on a
frame to provide a digital height field3, which allows the quantification of 2D seafloor
roughness. Results from this system are not displayed in this paper but are consistent with the segmentation. Ground truth data within the spatial limits of each

36
Ground truth
No grab

Appearance & Comment

No Photo. The lower BSVA return: soft bottom.

gr 1.4%, sa 36.7%, si 9.1%, cl 52.8%, Silty with some worm activity. BSVA 10 dB
4 =7.79. XBP class 11, 49g
higher than the previous: still soft.
gr O%, sa 57.2%, si 30.5%, cl 12.4%, Silty with intense worm activity. BSVA 5 dB
higher than the previous: silt.
d~=4.63. XBP class 11, 57-94~

gr 0.7-9.1%, sa 27.5-42.6%, si 370.9%, cl 1-48.5%, 4 =4.42-6.42.

Class I, 68-91g
gr O.7%, sa 29%, si 19.4%, cl 50.8%,
I$ =7. XBP class I, 136-146g

Two classes: thick short sea-weed and sandy

silt with patches of sea-weed,
Sand and clay with large stones (12 cm) ; The
resulting seafloor is hard and rough.

Table 1. Summary of the six classes of Fig. 1-2 with corresponding ground truth, stereo photo
interpretation and associated comments. In the table, gr, sa, si, and cl stand for gravel, sand,
silt, clay. 4 = - logz Dm where D , is the mean grain diameter in millimeters. XBP class I to
I11 are, respectively, hard, medium and soft bottom. Penetrometer deceleration (units of g) is an
indicator of seafloor hardness.)

SESAM category have similar properties. This is a point-to-point validation of

SESAM.
3.2 Sidescan images
A comparison between geo-referenced sidescan sonar images acquired during CLUTTER'O3 (using an EdgeTech DF-1000 at 390 kHz) and segmention/classification
map was made (Fig. 2). The SESAM segmentation conforms well to the sidescan
image patterns and provides additional information in areas where the sidescan
image does not display appreciable textural differences: in particular the stronger
return on sidescan signal over class 5, and the lower return where class 1 begins.
Also the correspondence of grey level and class type is coherent over the whole
site. Some black spot is due to shadowing. The comparison between sidescan sonar
images and the SESAM segmentation maps demonstrate the spatial consistency of
the technique.
4

Summary

This work reports further experimental validation of SESAM, a segmentation algorithm, using data acquired by a multibeam echosounder. The methodology, based
on the observation of the variation of the seafloor backscattering strength with the
incident angle, has been improved by following acquisition of further echosounder
system detail and enhancing the filtering applied to the segmentation data. The
results are stable and create potential for further improvement of the algorithm for
classification/characterization. Analysis of lower frequency multibeam sonar may

31
be appropriate for discrimination of sea-grass type and high backscattering seafloor
(gravel versus rock).

Aknowledgment
We would like to thank the captain and the crew of RV Leonard0 for the help and
the nice collaboration during the acquisition of the data during CLUTTER03.

References
1. G. Canepa and N. G. Pace. Seafloor segmentation from multibeam bathymetric sonar. In M. E. Zakharia, editor, Proc. of the fifth European Conf.
on Underwater Acoustics, ECUA 2000, pages 361-7. European Commission,
Lyon, France, July 2000.
2. G. Canepa, E. Pouliquen, N. G . Pace, A. Figoli, and P. Franchi. Validation of
a seafloor segmentation algorithm for multibeam data. In Proc. of the sixth
European Conf. on Underwater Acoustics, ECUA 2002, pages 89-94, Gdansk,
Poland, June 2002.
3. E. Pouliquen, M. Trevorrow, P. Blondel, G. Canepa, F. Cernich, and R. Hollet.
Multi-sensor analysis of the seabed in shallow water areas: overview of the
MAPLE2001 experiment. In Proc. of the sixth European Conf. on Underwater
Acoustics, ECUA 2002, pages 21-29, Gdansk, Poland, June 2002.

38
HHSOUND TRANSMISSION AND SPATIAL COHERENCE IN SELECTED
SHALLOW WATER AREAS: MEASUREMENTS AND THEORY
WILLIAM M. CAREY
Dept. of Aerospace and Mechanical Engineering, College of Engineering, Boston University
Boston, Massachusetts 02215 U.S.A.
JAMES F. LYNCH
APOE, Woods Hole Oceanographic Institution, Woods Hole, Massachusetts 02543 U.S.A.
WILLIAM L. SIEGMANN
Dept. of Mathematical Sciences, Rensselaer Polytechnic Institute, Troy, New York 12180 U.S.A.
ILYA ROZENFELD
Raytheon Electronic Systems, Portsmouth, Rhode Island U.S.A.02871
BRIAN J. SPERRY
Science Applications International Incorporated, Mclean, Virginia U.S.A. 22102
Experiments from several shallow water areas are summarized and coherent sound transmission
results, wavenumber spectra, are compared to range dependent calculations that use oceanographic
and geophysical characteristics from measurements and archives as bounded inputs to the
propagation codes. In general excellent agreement was obtained between the measured and
calculated results for both narrow- and broad- band transmissions between 50 Hz and lkHz to ranges
of 40 km. A relative signal gain, rsg, method for the estimation of horizontal coherence length was
applied to measured rsg results and yielded coherence lengths on the order of 30 A @ 400 Hz at
distances of 40 km. Perturbation theory was applied to the shallow water waveguide under the
condition of adiabatic normal modes and expressions were derived for the phase structure function
that was simplified by the use of Gaussian correlation functions. These analytical results with
estimates of the variances of the environmental variables permitted the estimation of the coherence
function and the rsg. The calculated coherence function and rsg were found to be consistent with
measured rsg and replica correlation results. The fluctuations in the oceanic water volume were
found to be the dominant factor in the loss of coherence.

Introduction

A major question in shallow water acoustics is: how well one can predict the
transmission and coherence characteristics of the propagating acoustic field, amplitude
and phase, in the presence of the natural variability of the surface, ocean volume and
seabed? This question can first be reduced to the prediction of the average intensity as a
function of source receiver depth, frequency, range, sea state, time of year and location.
Second, the prediction of the coherency of the propagating field and the dominant
mechanisms for coherence loss under the same set of conditions stated previously needs
to be quantified.
Coastal margins with depths of between 40 and 200 m have bottom sediment layers
formed by deposition, sandy bottoms. Areas of the Gulf of Mexico, the East Coast of the
United States between Cape Hatteras and Long Island, sections of the Asian Coast, North
Indian Ocean, the Strait of Korea, the East China and Yellow Sea to mention a few, have
these sandy depositional layers that are classified as fast or critical angle bottoms.

39
Exceptions in these areas are the depositional fans of major rivers composed of fine silts
resulting in slow bottoms.
Shallow water areas also have strong seasonal variations, for example winter
produces an upward-refracting or isovelocity sound-speed profile; summer produces a
downward refracting profile; spring produces a profile composed of a mixed layer,
thermocline and an isovelocity layer; and fall produces a profile composed of a mixed
layer, thermocline and negative gradient layer. To first order, one can choose a time of
year that emphasizes either surface, bottom or volume interactions. These basic profiles
are influenced by mesoscale coastal effects such as salinity variations due to mixing of
water masses, river out flows, oceanic eddies, internal waves and solitons. When
comparisons are made between a variety of experiments, these factors must be taken into
consideration, for it is well know that for a large number of variables, basically
deterministic effects can appear to be random.
A number of shallow water experiments (over the past two decades) have examined
the issue of range-dependent variability of the various ocean and seabed processes effects
on properties of the acoustic field, such as the time-range averaged intensity, and phase
coherent effects, the interference pattern of the sound transmission. Additionally, these
variations need to be considered over a large frequency range, 50 Hz to 1 kHz, and thus a
frequency dependent sampling in range and depth is required.
The Gemini experiment [ l ] explored the predictability of the acoustic field in the
presence of seabed variability at 50 and 140 Hz. The predictability of the measured fields
using state of the art propagation codes with high quality geoacoustic information, the
best possible a priori information about the seabed was used in the comparison with
measured results. Since the experiment was performed at a well-characterized site, the
Rubano site [2], where detailed surveys of the bottom had been performed, the
determination of differences between a priori and a posteriori calculations using the
inversion derived geoacoustic models was possible. In general the narrowbandperturbative inversion of the transmission results [3], the broadband perturbative
inversion of Rubano's result, and the synthesized model from archival geophysical
measurements were found to produce qualitatively consistent geoacoustic profiles.
Calculated transmission loss was found to agree with measured results with respect to
level but not with respect to the observed modal interference patterns.
The Hudson Canyon experiment [4] was performed on the New Jersey
continental shelf near Amcor borehole 6010, Amcor-6010, a well sub-bottom surveyed
area. The experiment consisted of along slope and cross slope transmission loss
measurement. The purpose of the experiment was to obtain a high quality transmission
measurement suitable for geoacoustic inversion [5], and for comparisons with
calculations performed with a several range dependent propagation codes. The results
were consistent with expectations at the lower frequencies [6,7] but large differences were
observed at the higher frequencies [8,9]. This experiment was repeated five years later
along the same tracks. When the differences in sound speed profiles were accounted, the
results were consistent at the lower frequencies and agreement was obtained at the higher
frequencies when a nonlinear frequency dependent attenuation was used in the calculation
[lo].
The Area Characterization Tests, ACT 1-111, were conducted at the following
sites: I, the West Coast of Florida [ l l ] , 11, a site on New Jersey shelf and one above the
Hudson canyon, and 111, two sites in the Korean Straits [12]. All produced excellent
agreement between measured and calculated sound transmission results when adequate
environ-mental inputs were available and a nonlinear frequency dependent attenuation
was used with range dependent codes.

40
In the Shallow Water Acoustic Random Medium Experiment, SWARM [ 131, the
continental shelf break experiment in the Nantucket Shoals region of the Mid-Atlantic
Bight named PRIMER [ 141, and the Asian Seas International Acoustics Experiment,
ASIAEX i1.51, the variability and predictability of the water column properties were
examined. In SWARM, the focus was quantifying the role of nonlinear internal waves
and their affect on sound transmission. These waves can strongly scatter sound when
present and are known to produce strong transmission anomalies. The PRIMER
experiment focused on the effects of mesoscale oceanography and the shelfbreak front,
both of which are strong acoustic scatterers. The recent ASIAEX measurements focused
mainly on the very strong nonlinear internal waves found in the South China Sea, similar
to SWARM, but with a far higher intensity of the internal wave field. In these three
experiments, a high degree of variability in sound transmission was observed most likely
due to the internal ocean structures. This variability could not be reduced or subtracted
out by measurements, inverse procedures, or ocean models - the ocean sampling and
inverse data needed were far beyond the measurement capabilities available. These
experiments illustrate that in certain areas and time intervals, the oceanographic weather
can drastically effect sound transmission and consequently our predictive capability under
such conditions.
This paper concentrates on the coherency of sound transmission in selected
shallow water waveguides with fast sandy bottoms, primarily the ACT series. The
emphasis will be on the prediction of the transmission loss and the relative signal gain of
bottom mounted horizontal arrays. Although these are weakly range dependent
environments with a dynamic and variable oceanography the prediction of narrowband
and broadband transmission loss is shown to be possible provided sufficient
environmental information is available. A method of estimating coherence lengths from
relative signal gain measurements is used to compare results from several shallow water
sites on time scales representative of sonar processors. Perturbation theory is used to
provide a means of calculating the coherence length and relative signal gain and is shown
to produce favorable results. The PRIMER experiment also produced matched filter
estimates of the phase variance between separated receivers that is consistent with the
relative signal gain results.

Experimental Sound Transmission Results

The Hudson Canyon Experiment [1-3, 5-81 was conducted near Amcor-6010 on the
continental shelf in 73-m water depth. Calibrated acoustic measurements were made
under calm sea state conditions with a vertical array of 24, 2.5 m spaced, hydrophones
spanning the water column from the water sediment interface to a depth of 50m. The
source was towed at 112 the water depth, radiated tones at four frequencies between 50
and 600 Hz over along- and cross- slope legs between 4 and 24 km. Fig. 2.1 shows the 50
Hz transmission loss versus range, phase of the received signal versus range, and the
horizontal wave number spectra from two hydrophones, one near the bottom and the other
25 m from the bottom. The striking feature observed is the stability of the phase and well
defined modal wavenumber peaks. The associated sound velocity profiles, see [ 1, Fig. 81,
had mixed layer depths between 12-20 m, a variable thermocline, spanning 1% of the
sound speed at depth, to approximately 36m and a roughly isovelocity layer. The
remarkable feature is the coherent characteristic shown in the modal interference and the
wave number spectra.

41

om

range (m)

ais

ai

0.2

horizontal wave number

na hs so,rh.n a.~ r r r nhr u m
mls. 1991. 6 2 1 pm

HC IU

"

0.16

range (m)

o 18

0.2

0.n

0.24

horizontal wave number

Fig. 2.1: The measured transmission loss and residual phase versus range along with the horizontal wavenumber
spectra for hydrophones number 14 and 24 .

The quality of the results is clearly reflected by the modal interference pattern and the
stability of the phase estimates. The geophysical properties from BH 6010 were converted
via Biot and inversions to geoacoustic profiles [2, pages1749-1750 and references 24-28]
consistent with the inversion results of Rajan [7]. These geoacoustic profiles were
simplified into multiple layer representations with gradients. The sensitivity to variation
of input parameters of the intensity levels and wavenumbers was investigated [3-51. In
general, excellent agreement between wide-angle-higher-order parabolic equation
calculated loss [3, Refs.(1,12,13,14,), 61 and measured loss was obtained. Shown in Fig.2
is a comparison between calculations for two simplified geo-acoustic profiles [3] and
measurements for frequencies of 50 Hz and 75Hz (Note that the curves are isospeed result
is displaced by + 15 dB while the experimental result is displaced by -15 dB.). The first
result is from a two-layer isospeed model and shows excellent agreement with intensity
and interference pattern. The second result is from the depth dependent layer model
shown as A on the insert along with the water sound velocity profile. Again agreement
was achieved. This remained true for longer ranges when bathymetry changes and sound
velocity profile changes were used in the calculation to ranges of 20 km [ l , Figs. 19, 201.
Fig.2.1 also shows the estimated horizontal wave number spectra [1, Ref.211 clearly
showing the wavenumbers of the propagating modes.

42
MODELA

-loo1
0

,
I

,
3

RANGE (KM)

a.) 50 Hz Transmission Loss Results

MODEL B

-110

RANGE (KM)

b.) 75 Hz Transmission Loss Results

Figure 2.2: A comparison of calculated transmission loss using two simplified geoacoustic profiles and
measurements. a,) this is the result for a frequency of 50 Hz. and b.) is the result for 75 Hz. [ 3, Figs. 6 , 71

Figure 2.3 shows calculated horizontal wavenumber spectra. Various techniques can
be employed to calculate the spectra; we used the Hankel transform method. The
calculated pressure field at a particular depth z is Hankel transformed in range to produce
the depth dependent Green's function at depth z. Thus peaks in the horizontal
wavenumber spectra correspond to propagating wavenumbers with characteristic
variability due to the subbottom geoacoustic profiles. These variations, caused by
environmental uncertainty, exceed the inherent errors in calculations employing wide
angle codes. The variations in the range dependent bottom properties can produce
variations in horizontal wavenumber on the order of 0.002 for a water wavenumber of
0.208 at 50 Hz . Although small, this change in wavenumber can have dramatic changes
in the modal interference patterns. In the parametric studies of the Hudson canyon
experiment [3-51 a robustness of the ability to calculate the field was observed along with
this type of variability.

43
-0

-35
14751840 18051970

HORIZ. WAVEN

m
9

-40

-45

ABSORBING LAYER

-50
-56

Figure 2.3: Relative intensity level versus range and horizontal wavenumber spectra are shown for four bottom
cz(z) profiles which represent the estimated characteristic subbottom variability. The successive wavenumbers
are shifted by 0.002 ni'.

The sensitivity studies included the variation of the compressional wave speed and
depth dependent attenuation in the upper 5 m of the sedimentary layer since the Amcor6010 geophysical measurements were thought to be representative of actual insitu
conditions at that depth. In particular, an algorithm, based on the work of Stoll [8], was
developed to calculate the porosity versus overburden pressure and consequently the
compressional wave speed versus depth to that measured at 5 m. The inclusion of near
surface velocity gradients shifted the modal wavenumbers and resulted in a change in the
modal interference pattern but little change in the average or peak intensity level versus
range. That is to say environmental characterizations that produce very good agreement
with intensity level versus range may only match the wavenumber spectra for the lower
order modes.
Measured and calculated horizontal wave number spectra versus depth and range are
shown in Fig. 2.4. The high degree of correlation between these two acoustic field results
is striking for the lower order modes but the presence of ghosting in the measured result,
double patterns, for the higher order modes is apparent and absent in the calculated
results. [9] Of particular interest is the bifurcated modal peak near wave number 0.18 in
both the calculated and measured spectra. This feature was shown to be due to sub-watersediment interface channel approximately 2-3 h deep; however the remainder ghosting
was not predicted and was most likely due to unknown range-dependent-subbottom
variability. These results were for 50 Hz; the 75 Hz frequency results were found to
require more detailed environmental information and were more complex.

44

016

014

018

02

0122

HorzontaJ Wavonumbei (m*-l)

a. The measured result at 50Hz

50 Hz Model 2

-----__

014

016
018
0::
Horizontal Wavenumbef (lim)

022

b. The calculated result at 50 Hz.

Figure 2.4: Measured (a) and calculated (b) horizontal wavenumber spectra versus depth showing the effect of a
subsurface channel , the bifurcated modal peak near 0.18. The measurements were made at 24 discrete depths
and interpolated to yield these results. [ 91

Although agreement between calculations and measured results was excellent at 50

and 75Hz, at the higher frequencies the measured result showed considerably more loss
than predicted with a linear frequency dependent sediment attenuation. Rogers [2]
attempted to explain this difference by the inclusion of a near sediment-water-interface
scattering layer but large scattering strengths were required. Howell [ 101 re-examined this
scattering layer in a detailed numerical investigation and found minimum layer thickness
for a specific scattering strength per unit volume and scaled correlation length; however
he observed the importance of an increase attenuation in the near surface layer. The
experiment was repeated extending the frequency range to 1 kHz along the same tracks
used previously. Evans [ 1I] used a two-stage least-squares procedure to quantitatively
examine the results from both experiments over the 50 Hz to 1 kHz range focusing on the
frequency dependent attenuation as a function of depth in the sediment. Calculations
using a parabolic code [6] were performed to compute the range dependent transmission
for both experiments while varying the attenuation frequency dependence. The "best" fit
was for a frequency dependence of the attenuation ( a =alkH,(fllkHz)")
with the 1.57
power between 50 - 1000 Hz . The conclusion was that a geoacoustic model using a

45
nonlinear frequency dependency and a reasonable uncertainty for the sediment volume
attenuation provides a good fit to the measured results.
Nonlinear frequency-dependent-sediment attenuation factors are not a new finding in
shallow water waveguides with medium to hard sandy bottoms. Ingenito [ 121 observed a
nonlinear dependence between 400 and 750 Hz. Beebe [ 131 found that for two sites with
medium to coarse sands factors between 1.57 and 1.83. Rogers and Zhou [14] summarize
several experiments in the Yellow sea and observed that frequency dependent attenuation
factor ranged between 1.6-1.9, and Tappert [15] found in his analysis of the ACT I11
experiment on the West Coast of Florida, Uricks Site, that n=2 was the best fit parameter
consistent with the Biot porous medium model. Chizhik [ 161 and Tattersal [ 171 used the
Biot theory to compute sediment parameters for a fluid bottom, elastic bottom and a fast
field program calculation with a geoacoustic bottom. Calculations using a fast field
program and a multipath expansion code were found to agree with measured transmission
over the 100 Hz to 8 kHz when the Biot bottom was used. That is to say a frequency
dependent attenuation coefficient with n=2 power for frequencies less than 1 kHz and
another dependence above this frequency. Finally in the ACT I11 experiment modeled by
Rozenfeld [I81 was found to require a frequency dependent factor, n=1.8, to describe both
narrowband and broadband transmission loss in the complex Strait of Korea [see [19] for
detailed oceanographic and geoacoustic descriptions].
50- 1WHl

80

I
0

10

15

20

25

10

15

20

25

2W-4CQHz

70

10

15

10

15

20

25

25

60
80

40
BOO-1mHz

60
80
0

10
15
Ranw (a)

20

25

10

15
Range (am)

25

Figure 2.5: A comparison of both narrow band and broadband transmission loss for representative range
dependent profile in the complex Strait of Korea. This figure clearly shows the robustness of the computational
technique and is for a frequency dependent attenuation to the 1.8 power.

The robustness of this result is demonstrated by the finding that when applied to the
other range dependent tracks in the same area excellent agreement between calculated and
measured results was obtained, Fig. 2.5, especially along the coherency experimental
transmission tracks. Calculation of the coherence function by the formalism presented in
this paper requires the assumption of adiabatic-normal-mode propagation or negligible
mode coupling. The strength of mode coupling for this particular range dependent
environment, ACT 111, was determined by use of the calculated horizontal wavenumber
spectrum of the field.
A comparison between the wavenumber spectra at two distinct ranges can be used to
estimate the strength of modal coupling; or the comparison of the spectra for the rangedependent and range-independent environments to estimate the relative importance of

h46
mode coupling. In particular, the examination of single mode propagation shows the
actual coupling to other modes in a range dependent calculation.
To compute the horizontal wavenumber content of the calculated pressure field at
any particular range, the method from [3] was used. First the field is propagated to an
observation range using a range-dependent calculation that has been shown to correctly
calculate the sound transmission. Assuming that propagation from the observation range
continues in an ideal range-independent waveguide over a sufficient range then the
horizontal wavenumber spectrum can be calculated. In the Act I11 experiment extensive
environmental measurements were performed and used to characterize the experimental
area. Although these observations spanned the experimental period, they were not
coincidental with the acoustic measurements [19]. However, when coupled the
oceanography, geoacoustics and bathymetry permit the calculation of the coherent effects
such as the horizontal wavenumber spectrum.

0.465
0.46
F=354Hz
-05

3
._

9
I

l
I

1.445

l
0.5

1.45

1.455 1.46

1.465

1.47

1.475

1.48

1.485

F=WHz

2.48

249

2.5

2.51

2.52

253

Horizontal wavenumbers (lh)

Figure 2.6: Depth averaged horizontal wavenumber spectra at range 25 km with normalized amplitudes for (a)
111, (b) 354,and (c) 604 Hz. The ACT 111 Environment used in these calculations is shown with a mode 1
starting field. Vertical dashed lines show the water wavenumber.

A starting field using the first mode was used in the calculation of the depth averaged
horizontal wavenumber spectra for frequencies 111, 354, 604 Hz and is shown in Fig. 2.6
The vertical dashed lines are the values of maximum horizontal wave number. The
calculations at all frequencies show the modal coupling in this experiment is weak, with
over 90% of the energy remaining in the first mode with the coupling increasing with
frequency. This analysis and other single mode calculations not shown here show for this
region in the Strait of Korea can be characterized as weakly range dependent.
Consequently the use of the coherence theory based on the adiabatic mode approximation
appears to be justifiable .

Experimental Signal Gain Results

The experimental estimation of the transverse coherence was obtained from

measurements in several locations: the West Coast of Florida, the East Coast of New
Jersey and New York, the Strait of Korea and near Cheju-Do Island. The results from

41

these experiments and others will be compared to our most recent experiment; the ACT
I11 is discussed in this section to illustrate the method and the fidelity of the results.
The estimation of the transverse-coherence length may be accomplished by
measurement of the narrowband coherence function, the broadband correlation function
or the coherent signal gain either from a direct measurement with a filled aperture or the
steered beam response from a sparsely filled aperture. The ocean acoustic environment
complicates the measurement of coherence at the low-to-mid frequencies, because a
partly coherent noise background limits the achievable signal-to-noise ratios.
Narrowband coherence estimates are possible, but the require long incoherent averaging
times compared with the time during which the medium is stationary. Motions of the
receiving system and source also produce a dynamic multipath field that can dominate the
coherence function measurement and the statistical measurement bounds are large [ 1-31.
The measured coherence function is thus a strong function of the signal-to-noise ratio,
coherent multipath interference, and furthermore estimation requires a large number of
samples that must be acquired in a time period that is short compared with the stationarity
of the medium and the dynamics of source-receiver motions. [4-51
On the other hand, the measurement of array signal gain (ASG = lOlog(asg)) is
performed with an array of sensors. The asg estimate is based on a larger number of
degrees of freedom, a larger signal-to-noise ratio resulting from the coherent signal
summation, and the spatial filtering of unwanted noise sources. This asg is proportional to
the pairwise coherence function. The relative asg can be related to the coherence length
provided the functional form is known and all other causes of gain reduction, gain
degradation, are eliminated. Signal arrivals can be contained in near broadside beams or
compensated for in the estimation of signal gain because (1) the beam response is
normalized by the average signal over all of the hydrophones comprising the aperture and
(2) the gain of arrivals on individual beams can be summed. The remaining factors which
determine the array signal gain are refractive and scattering effects attributed to the
surface, water sediment interface, volume, and internal wave structure of the sea.
3.1

The experimental method 141

The array signal gain or the array response to a signal with phase randomness is central to
the determination of an array length limit and also provides a method to determine the
transverse horizontal coherence length.. The mean-squared beam response function
depends on two quantities, the phase variance and the phase correlation function. The
array signal gain for a deep-water environment increases as 201og(2Lld) until gain
saturation beyond which the increase is as lOlog(2Lld). In shallow water the theoretical
array signal gain should be determined numerically with a calibrated transmission
algorithm.
The beam pattern for a line array of length L along the x-axis is calculated when the
excitation is taken as a wave with a deterministic component P, exp(ik sinB,x,) and a
phase random component exp(i@(x,)) with the variables x = 2x, I L and -L/2< x, < Ll2,
y r = n L s i n B / d , yo= n L s i n 0 , / A :

(3.1)
The random variable @(x) has the following statistical characterization with mean,
< &) >= 0 ,and variance, < @(.> >= a , and a correlation length, L4 :

48

The statistical expectation values of the beam response magnitude squared, the expected
power, is

The special case of transverse arrivals is important and Eq. (3.3) reduces to the evaluation
of the integral

The integrand in Eq. (3.4) is recognized as the spatial coherence function with a form
determined by the phase correlation function, and shows that the signal gain is
proportional to the coherence function. Randomness in phase with a Gaussian correlation
function produces a spatial coherence function with a quadratic (Gaussian) form. This
quantity, Lh, is the horizontal coherence length. (This expression can easily be discretized
and placed in matrix notation.) The first part of the integrand is recognized as the crosspower spectral density, and the Gaussian exponential is seen to be the spatial coherence
function. . The relationship between relative signal gain, rsg, and coherence length, Lh,
for an aperture La can be shown to be
rsg=r-2J-:(r-lyl)exp(-y
2 )dy , ~ = L , / L , ,.
(3.5)
The mean value the measured signal gain, SGM (dB), standard deviation, SD (dB), and
the coefficient of variation, COV=SD/SGM, can be determined from the cumulative
distribution of the measurement results. If the measurement ensemble has been obtained
for a depth-time-range interval for which the oceanographic conditions have been
stationary, then the mean SGM = lOlog(asgm) is equal to the expectation value,
< asg >= asgm .

SGE = SGM = lolog(< asg >>= 1olog(< lf(o]

> / < P: >>.

(3.6)

The relative signal gain, RSG, is equal to the ratio of the expected gain to the theoretical
gain, RSG = SGM- SGT = lOlog(<asg>/asgt). Since the theoretical signal gain can be
calculated with a calibrated transmission code and the SG measured, then the RSG can be
determined and the signal coherence length, Lh, as a function of frequency and range can
be estimated.

3.2

The signal gain experiments

An experiment was conducted at two sites off the South Korea coast during August and
September 1995 [6, Fig. 11. The objective was to explore effects of shallow water

49
acoustics on array performance under known environmental conditions. An integral part
of these experiments was the measurement of oceanographic, bathymetric, and
geophysical variables as well as utilizing archival survey results. Site 1, a shallow region
with a water depth - 100 m, had a moderate range dependent bathymetry but variable
sound speed profiles due to the Tsushima Current. The experime,nt [6, Fig. 21 had twobottomed- horizontal arrays separated by a distance of some 18.5 km and five reflectors
on a perpendicular bisect spaced at intervals to a distance of 35 km. The bottomed arrays
were in a surveyed area and straightened by tension. Elements locations were determined
using the impulsive sound from imploding light bulbs. These reflectors had an average measured-target strength of 13dB between 100 and 800 Hz.
Signal gain measurements were obtained for the four continuously radiated tones,
CW, as a function of range. This CW signal was processed in a narrow band to increase
the signal to noise ratio. In the shipping dominated noise field the processing trade off
was between the bandwidth, coherent and incoherent integration times and the variations
in the field due to relative source motion. These measurements were also made difficult
by source level limitations and the presence of interfering sources of noise, ships.
Impulsive sources, S, produced large source levels across the band of interest and
because of their short duration, large bandwidth, were processed in broad bands centered
on the frequencies of interest. However, the range coverage was discrete, not continuous,
and the multiple shots were spatially separated. Since the impulsive signal processing and
time-domain beamforming were different than the CW processing; any comparisons must
consider these differences.
CW signals received by the horizontal array were processed by applying a Hann
window, zero filling, a fast Fourier transform (FFT)with 50% overlap, and a linear power
average of five samples. The equivalent noise bandwidth was 0.15 Hz and the signal level
determined by the maximum in a 2 Hz band centered on the transmitted frequency. Noise
estimates were taken as the average of four frequency bins, two on each side of the
expected frequency. These estimates were performed on both hydrophones and
beamformer output. Although measured element positions were used in the beamformer
an uncertainty still exists and is complicated by the relative position of the source to the
normal of the array. Peak FFT beam values were taken from a 20" sector centered on the
source to keep scalloping levels less than 0.2 dB.
The impulsive broadband signals were processed differently. In the broadband
analysis, the energy spectral densities, ESD, for both the signal and reverberation were
determined in multiple bands for each hydrophone. Time domain beamforming was used
and multiple beams were steered toward the source with a 0.1" resolution and the
maximum bean response was chosen. This procedure minimized errors due to scalloping
and source location. This procedure was performed on the reverberation to determine the
beam echo-to-reverberation ratio and the average hydrophone echo-to-reverberation ratio.
The beam echo-to-reverberation ratio divided by the average hydrophone echo-toreverberation ratio provided the estimate of the signal gain and subsequently the RSG.
Because of inherent corrections in the broadband processing and time domain
beamforming, the resulting RSG has a smaller bias correction. The results of both the
narrowband results, (* symbol for ranges between 8 and 12 Km), and the impulsive
source results, ( symbol for ranges between 5-22 Km, and 0 symbol for ranges between
23-46 Km). These results are consistent with a least square fit of the CW as shown by the
line in Fig. 3.1. Although the impulsive results have smaller RSG values between 20 and
70 wavelength apertures since the differences are within a standard deviation, they are
consistent with the CW fit.

50

Figure 3.1. The RSG is plotted as a function of array length in wavelengths for the CW (*, 9-12 Km),
Impulsive ( ,5-22 Km) and impulsive (023-24 JSm).

These experimental measures of CW-RSG were found to decrease with range and
frequency. Nevertheless, the limited number of measurements over a limited range
precludes a definitive conclusion. The impulsive results were found to have a similar
trend. Again, the range resolution and quantity of the measurements preclude quantitative
conclusions concerning the range dependence.
These RSG results pose a quandary as to the fundamental mechanism responsible for
the reduction in RSG. The over riding sound velocity profile, SVP, characteristics which
determine the general sound transmission are described by the presence of a mixed layer
(-20m), variable thermocline ( A C / h D= -0.3to-O.grn/srn), and an isovelocity layer (>
60m) near the water sediment interface. The agreement between the measured and
calculated TL results previously discussed is largely due to the depth of both the CW an S
sources being below this mixed layer. The sonic speed near the sandy bottom was
between 1505 and 1510 m / s ensuring a fast boundary condition with a critical angle. Once
the higher angle propagation energy was absorbed, we observed a cylindrical spreading
with attenuation determined by the bottom. From a transmission viewpoint, the internal
wave and eddy activity results in the large variations observed in the sound velocity
profiles and would certainly have the effect of changing the angular structure near the
bottom; however range dependent calculations can incorporate these effects.
This quandary is complicated because of the relatively flat nature of the bottom and
the small deviation around the local mean. Given we have a fairly flat bottom, small
values of the rms roughness, considerable mid-water variability and a strong mixed layer;
one may suppose that the combination of the mid-water variability and the upper most
layer of the sediment were responsible for the reduction in RSG. Furthermore the
estimated value of coherence length from these RSG values was found to be on the order
of 30 h and implies that with proper processing resolution with active sonar may be
between 1 and 2.
This series of shallow water experiments referred to as the Area Characterization
Tests, ACT, were conducted with bottomed horizontal arrays, explosive and continuous
wave sources in sea sates less than 3. The average values of the transverse coherence
lengths were determined as discussed above from RSG estimates. The coefficient of
variation of the signal-gain measurements was approximately 4-5% determined by the

51
ratio of the measurement standard deviation, dB, and the measured signal gain, dB. Fig.
3.1 showed the ACT 111-Site 1 RSG versus the array length in wavelengths. From the CW
curve the RSG=-2dB at an aperture of 58 h corresponding to the coherence length 0f.z
29 h.
The ACT I experiment was conducted on West Florida Escarpment of the Gulf of
Mexico in water depths between 150-200 m. The environment was weakly range
dependent with a downward refracting profile, a mixed layer depth of 30 m followed by a
rather constant thermocline with a gradient AC 1
= -0.36 m 1 s 1 m . The sound
propagation and coherency at this site was primarily due to the bottom interaction because
the variability in the sound speed profiles was minimal in contrast to the ACT 111. The
estimates of coherence length in the 200-8OOHz band were between 30-32h.
The ACT I1 experiment was conducted north of the Hudson Canyon on the East
Coast of the United States in water depths between 90 and 150 m. The oceanographic
environment was variable due to the interaction of three water masses: the shelf water, the
Hudson River outflow, and an eddy from the Gulf Stream. The sound velocity profiles
featured a mixed-layer depth of 27 m followed by a variable thermocline gradient
AC I ALI = -0.9 to - 1.4 m I s 1 m indicative of internal wave activity. The transmission
loss was strongly affected by the bottom and the coherency resulted from the combined
effects of bottom interaction and water column variability with the coherence length in
the 200-600 Hz band estimated to between 23-25 A.
The ACT 111, site 2 in water depths between 85-105 m, near Cheju-do, was located
on the boundary between the Korean Strait and the Yellow Sea water masses where the
major variable was the tidal mixing. The bottom water appeared to be Yellow Sea cold
water with the intermediate depth waters the result of vertical and lateral mixing of the
Tushima Current and Yangtze River water. This mixing resulted in spatial and temporal
variability in the velocity profile between 20-80 m. The sound-velocity profile at this site
had a mixed-layer 20 m depth and an isovelocity layer below 80 m. The ACT I11 Site 2
bottom was again two layers over the acoustic basement. The first layer approximately 5
m deep is composed of a mixture of mud and sand, silty-clay and sand, over a rough layer
of sand approximately 15 m deep and may well be a slow bottom. On average, a longer
coherence length estimate resulted from site 2; that is, a horizontal coherence length of
38h at a range of 24 km.
The reason for the consistency as well as differences of the estimated transverse
coherence lengths is not understood. These transverse coherence lengths must be
attributed to the variability of the water column, the upper sediment layer, and bottom
properties all of which probably play an interrelated role. These results should be
representative of coastal sites with variable oceanography and sandy I silty-clay bottoms.
Average coherence lengths of L,,= 30h for ranges between 9 and 45 km in the 50 - 800 Hz
band with a decreasing trend with frequency and range should be observed.
A review of SW investigations with comparable environmental conditions can be
found in reference [41 and specific investigations [6-121. A key factor is the SW sound
velocity profile. If the SVP is upward refracting, then the sea state should be the
controlling factor. If the SVP is isovelocity then the critical angle of the bottom and the
sea state should be the main factors. The experiments discussed here had downward
refraction with a near surface mixed layer and consequently water volume and bottom
characteristics were the controlling factors. Wille and Thiele [8] reported results = 18 Llh
at 400 Hz from the North Sea and Scholtz [9] found smaller values. Wasiljeff [lo]
observed similar results from his measurements near Elba in water depths of 115 m with
clay and clay-sand bottoms. Wasiljeff [ 101 compared winter and summer coherence

52
measurements at frequencies greater than 800 Hz and observed large coherence lengths
during the winter experiments and short coherence lengths, 29d @ 800 Hz , at ranges
between 10 and 19 km during the summer downward refraction conditions. In general,
Wasiljeff's measured larger coherence lengths at higher frequencies than those reported in
this work.
The transverse coherence length as a function of range and length is the parameter
limiting the resolution of sonar arrays. We contend that phase randomness introduced in
the acoustic field through scattering as the wave propagates determines the coherence
functional form and the transverse array signal gain. Three shallow water sites with
variable downward refraction were found to have an average coherence length of 30A at
ranges of 40 km for frequencies between 300-400 Hz when Gaussian coherence function
was used. The gain measurements when coupled with a calibrated numerical transmission
code should enable the rank ordering of coherence-loss mechanisms.

Horizontal Coherence

Our first objective is to find a convenient representation for the horizontal coherence.
This coherence function C is the normalized covariance of the solution P to the Helmholtz
equation between two points with horizontal coordinates rl and r2 ,

where z is depth measured positive downward from the sea surface z = 0. Angle brackets
in Eq. (1) represent average over the ensemble of random environmental fluctuations, and
asterisk is complex conjugate. The dependence on radian frequency w is suppressed in P
and later variables.
The environmental mechanisms that are primarily responsible for coherence
degradation in many low-frequency situations are fluctuations in the water column sound
speed and the bottom bathymetry. In the development to follow we focus on these two,
although others including fluctuations at the sea surface, within geoacoustic layers, and at
geoacoustic interfaces can be included at the cost of notational complexity. In addition
the mean environment is taken as range independent, although gradual range dependence
is easy to incorporate. Consequently the water sound speed 7 and bathymetry q are
expressed as mean portions c and H plus perturbations,

q(r)

H + 6h(r).

(4.3)

Assumptions on the statistical properties of the fluctuations are made for simplicity, but
these can be relaxed as well. The perturbation ensembles have zero mean and spatially
homogeneous variances,
<6c>

0 , <6h>

0 , <6c2>

(T,

, <6h2>

oh2,

(4.4)

53
where (3 and o h are standard deviations of the sound speed and bathymetric fluctuations.
Covariances of the sound speed and bathymetry are also assumed homogeneous and
horizontally isotropic,
< 6c (r 1 , z 1 ) 6c (r z ,z 2 ) > = B () (r 1 - r 2 , z 1 , z 2 ) ,

so the horizontal covariance functions B () and B ( h , depend on the difference of

horizontal coordinates. The sound speed covariance need not be vertically homogeneous.

4.1 Adiabatic normal modes and perturbations

For horizontal variations that are sufficiently small and slow, an approximation
for the pressure field can be obtained from the method of horizontal rays and vertical
modes [ 1,2]. Using standard perturbation theory, this expression can be simplified
considerably into the following form [3,4], corresponding to the adiabatic mode solution:

In Eq. (4.6) N is the number of propagating modes, and the formula for the normalization
factor Q, which is proportional to source strength, is not needed here. The amplitude
factor U for mode n is
(4.7)
where r is the horizontal distance measured from the source (at r = 0 ) to a receiver,
cp (z ) is the mode function, and z is the source depth. The mode eigenvalue K is
expressed, analogous to Eqs. (4.2) and (4.3), as a sum of mean and perturbed portions k
and 6k :
Kn(r) = k ,

+ 6k, ( r ) .

(4.8)

An important consequence from perturbation theory is that the perturbation in K only

needs to be retained in the exponential (or modal phase) term in Eq. (4.6) and not in the
modal amplitude term of Eq. (4.7). Similarly the modal attenuation A due to bottom
losses appears only in the exponential. The integration in Eq. (4.6) is along a path
between source and receiver.
The mode eigenvalue is the horizontal component of the local wave number
vector, which has magnitude K , by virtue of the relation

54
where the vertical wave number component for mode n is K , ,. Both K and K ,,can be
expressed similarly to Eq. (4.8),
K(r,z)

k(z) + 6k(r,z),

(4.10)

The requirement of small fluctuations for the validity of Eq. (4.6) is expressed by two
conditions
Oc

/C < < I ,

Oh

k,, << 1,

(4.12)

with similar conditions on derivatives for the requirement of slow fluctuations.

Another important consequence from perturbation theory is a formula for the horizontal
wave number perturbation. For each mode, 6k , is the sum of contributions 6k (') and
6k ( h , from sound speed and bathymetric fluctuations,
6k,

6k.'"

6kn'h' .

(4.13)

Application of the Fredholm Alternative [5] produces the following expressions for these
two contributions:

(4.14)

(4.15)

The bracket notation means the jump in a quantity across the bathymetric interface where
density p and sound speed q are discontinuous:

The boundary conditions require continuity of cp , (H ) and cp (H ) / p . It is interesting

to note that Eq.4.14 was derived by Rajan et a1 [ 61, Shang and Wang [7 3 and Lynch et a1
18 I.

4.2

Wave number contributions

The wave number perturbation expressions are critical because they represent the
mechanisms for environmental influences on the coherence function. Although Eqs.
(4.14) and (4.15) appear somewhat complicated, most of their components can be derived
and interpreted in a straightforward way.
If Eqs. (4.8), (4.10), and (4.1 1) are inserted in Eq. (4.9) and linear terms retained,
we find

55
(4.17)
The mode function normalization condition is

(4.18)
Applying the integral in Eq. (4.18) to Eq. (4.17), and using the formula for 6k obtained
from the definition of K in Eq. (4.9), we obtain

(4.19)
Since the sound speed fluctuation 6c occurs in the water for z < H only, the first
integral in Eq. (4.19) reduces exactly to Eq. (4.14). This result has been useful in other
work, for example inversions in Refs. [6,8].For the second integral, Eq. (4.19), suppose
that the interface variation is downward, that is 6h > 0. The perturbation in k
/ p is
non-zero only in the small interval between H and . A rectangle rule approximation to
the second integral in Eq. (4.19), using the continuity of cp (H ), gives

which is the contribution of the first term in Eq. (4.15). A similar argument applies where
6h < 0.
Therefore, the first term in Eq. (4.17) leads to the entire contribution to 6k .() in
Eq. (4.14), and the second term in Eq. (4.17) produces the first contribution to 6k n ( h ) in
Eq. (4.15). The second term, Eq. (4.15), arises from a perturbation of the mode function,
and is not accounted for in this heuristic calculation..
4.3

Coherence approximations

The coherence function is approximated next for the transverse case. A source
located at (0, 0, z o ) and a receiving array are in the same horizontal plane, and the line
segment (of length R ) from the source to the array center is perpendicular to the array.
We consider two receivers at horizontal positions rl and r2 located a distance d/2 on
either side of the array center. Since the distances from the source to each receiver are
equal, the coherence function C from Eq. (4.1) can be expressed as:

C =

XU,exp(- 2Anr)
n=l

(4.21)

56
Two simplifying assumptions are made to obtain Eq. (4.21). One is that all cross-modal
interference terms have been neglected, since these terms tend to introduce small rapid
oscillations about the function in Eq. (4.21). Another is that the wave number
fluctuations have approximately Gaussian pdf s, so that
<exp(ia)>

exp(-<a2>/2).

(4.22)

The important quantity S , in Eq. (4.21) is the modal phase structure function [9]
for this case and is defined as

(4.23)

Other receiver positions and orientations can be treated (see for instance [lo]) with some
analytical complications.
Since contributions to wave number fluctuations arise from both sound speed
and bathymetry, the modal phase structure function can be expressed as

s,

s.(c)

S p ) .

(4.24)

After interchanging integration and statistical averaging, and for index a replaced by
either c or h, the terms S . ( a ) are

(4.25)

where the wave number covariance functions W

are

Since the sound speed and bathymetry covariances in Eq. (4.5) are horizontally
are as well. They are assumed to have
isotropic, the wave number covariances W
circular symmetry with horizontal correlation lengths L ( a ) that can differ for a = c or h.
Circular symmetry implies that W , , ( a ) is a function of the dimensionless variable
q = I r 1 - r 2 I / L ( a ) , and the components of the modal phase structure function can be
reduced to single integrals of the type

(4.27)
where x = R / L (') is range in correlation lengths. In Eq. (4.27) the function F (q ) from
Ref. [4] i s dimensionless, order one, piecewise continuous, and contributes to the

57
weighting of the wave number covariance integrals. Additional simplifications that have
been used in Eq. (4.27) are

>> 1 , 2 R / d >> 1 ,

(4.28)

which are usually appropriate for applications of interest. The first of Eq.(4.28) is useful
in simplifying the dependence of F(q) and the second of Eq. (4.28) is always satisfied
when the receiving array is far from the source, so that the distance r to any receiver is
close to R.
4.4

Wave number covariances

The final step is to link the wave number covariances to the sound speed and bathymetry
covariances. This allows determination of parameter dependencies of the modal structure
function, and consequently the coherence.
The formula for Fk n ( h ) is rewritten as
Fk,'h'(r) = I , Fh(r),
where I denotes the factor multiplying Fh in Eq. (4.15). Consequently, W ,(
proportional to B ( ) :
W n ' h ) ( q ) = I,'

B'h'(q)

(4.29)
)

is

(4.30)

Since only the properties near the bathymetric interface are relevant for W ,( ) , an
example is obtained from the Pekeris waveguide [ 111. For this case Eqs. (4.15) and
(4.30) along with the Pekeris solutions and eigenvalue relation lead to:
W n ' h ) ( q ) = 1 / ( 4 k n ) 2 ~ ~ n 2 ( ~ ) ~knz2
~o}2{
+ ( P nz ) ( P O / P ~ ) ' ) '
B'h'(q).

(4.31)

In Eq. (4.3 1) p 0 and p b are the constant densities in the water column and bottom,
k,, and P,, are the water and bottom vertical wave numbers (the latter is a decay rate).
Other than the weak dependence on density ratio, the coefficient of B ( h ) in Eq. (4.31)
primarily produces an angular variation depending on mode number.
It is also convenient to rewrite Eq. (4.14) as
(4.32)
where the modal kernel G ,is
(4.33)

58
The approximation in Eq. (4.33) uses constant values p and c 0 for density and sound
speed in the water column. We obtain the following expression for W n ( c ) :

Further simplification of Eq. (4.34) results from reducing the double integral to a single
integral. One approach (see for example [ 121) is to assume that the vertical correlation
length L (') is much smaller than both L I (') and the acoustic wavelength. Then
B ('I(q, z 1, z 2 ) is assumed to a product of 6 (z - z 2 ) and a function B (q ), so that
H

W.("'q)

= Bl(q) J m Z ) d Z .

(4.35)

A second approach is to assume that the water column sound speed is approximately the
constant c o , and also that B (') is vertically homogeneous so that it depends u = z - z
as well as q. Then using relative center of mass coordinates u and v = (z + z ) / 2, Eq.
(4.34) can be reduced [4] to

where Fo (u ) is an order one function depending on trig functions that arise from the
water eigenfunctions. A third approach is to assume that B (') (q, z z ) is a product of
B 1 (q ) and two components B (z ) and B 0 (z 2 ), which is a useful assumption in
modeling influences of a random field of internal waves. It follows that [ 131

(4.37)
These examples are meant to suggest some, and not necessarily all, of the possibilities
that have been used for environmental covariance modeling.
Once the wave number covariances are estimated, the modal phase structure
function components in Eq. (4.27) can be obtained. For illustration we use the third
approach above, Eq.( 4.37), with the specific Gaussian form for B (q ) = exp ( - q /2 ).
Using this formula and Eq. (4.37) in Eq. (4.27) with a = c , and simplifying the integral
with Eq, (4.28), we obtain

The factor containing the Erffunction in Eq. (4.38) specifies the behavior of the modal
phase structure component S .(') with distance R from source to array center, coherence

59
length L (') of sound speed, and array length d. An analogous factor appears in the
component S n( h , with the choice B ( h ) (q ) = exp ( - q /2 ). Finally, when the array
length d is small compared with L ( a ) , then the Erffunction in Eq. (4.38) is negligible
and the entire modal phase structure Eq. (4.24) becomes

Note the linear dependence on the parameters R and the correlation lengths. The estimate
for the factor I 2 contained in Eq. (4.31) should be useful in Eq. (4.39).
5

Numerical Comparisons

Two experiments provide requisite results to perform calculations with nonacoustic

environmental and geoacoustic parameters for comparison with measured relative array
signal gain results and estimates of horizontal coherence lengths. The first is the
previously discussed ACT I11 experiment that used bottomed-horizontal arrays in the
complex Straits of Korea. [l-31 The second is the Shelf Break Primer experiment
designed to determine the mesoscale oceanographic dynamics on the acoustic
transmission in the 400 Hz frequency region [4-71. This section concentrates on the
results of these investigations with the intent of providing a concise summary,
comparisons and discussion of results.
5.1

The relative signal guin result

The determination of the coherency of the shallow water waveguide using perturbation
theory discussed in the previous section, where the analytical results of several
investigators have been summarized, were applied to the particular case of the ACT I11
experiment. The computational problem is simplified if we assume a Gaussian correlation
function for each of the environmental variations with a standard deviation, 0,and a
correlation length L. For many of the environmental parameters such as the standard
deviations of the sea surface roughness, the water-sediment interface roughness, and the
volume variations due to density and sound speed estimates can be made based on limited
ability to perform environmental measurements over large oceanographic areas. However,
some parameters such as the characteristics of the sub-bottom are simply unknown.
The variation of the sound speed profiles over space and time are shown in Fig.
5.1 where the light curves are the measured profiles, the solid dark curve represents the
mean profile, and the dashed curves the f standard deviations. The mean profile varies
between 1507 and 1530 d s , a maximum deviation about the mean of f 7 d s , and a depth
averaged deviation of ? 4 . 5 d s . These deviations represent a 1.52% change of the sound
speed from surface to the bottom and a 0.4% variation about the depth-averaged profile.
The sea surface variation was estimated to have an average standard deviation 0.8m and
the bottom depth variation an average standard deviation of 0.6m.

60

1505

1510

1515

1520

1525

Sound speed (Ws)

1530

1535

1540

Figure 5.1: Sound speed profiles (light curves) measured at Site 1 along with the mean profile (dark curve) and
k standard deviations (dashed curves).[ 1,3]

The bottom undulates about a mean depth of lOOm with a standard deviation of about
a gently sloping profile of =k 0.6m with the standard deviation of sound speed
fluctuations within the sediment were between 15 and 17Ods. These parameters are
shown in Fig. 5.2 as a simplified analytical model of the ACT I11 waveguide and the
background geoacoustic profile shown in [ 1,2].

Figure 5.2: The stylized environment used in the numerical evaluation of the coherence function. The lighter
lines represent the mean environment while the darker curves represent the superimposed random fluctuations.
The symbol cr represents the standard deviation and L represents a correlation length. The subscripts w, b, s, and
sb denote water sound speed, bathymetry, sea surface, and sub-bottom sound speed fluctuations, respectively.
The superscripts (h) and (v) in this figure represent the horizontal and vertical components. The details of the
mean sub bottom geoacoustic profile can be found in [l, Tables I1 and III].

The coherence function of Eq. 4.21, the phase structure function defined in Eq. 4.27,
Gaussian correlation functions, and the above fluctuations were used in the calculation of
the spatial coherence versus receiver separation in wavelengths as shown in Fig. 5.3.

61

"._,

-.-.-.- .-.-.-.-._._,_._._.

0.21

'I.-.

0.1'

1W

50

150

Receiver Separatim (wavelengths)

Figure 5.3: The calculated coherence function versus receiver separation for three frequencies. The horizontal
line is exp(-1).

The coherence decreases with an exponential characteristics with exp(- 1) crossover

of 4 5 h at a frequency of 604 Hz slightly higher than that estimated from the
measurement. Additional calculations, not shown here, for a equivalent Pekeris
waveguide showed similar trends with coherence versus separation distance as a
monotonic function of frequency. In addition the contributions of each of the sources of
the fluctuations, the sea surface, the volume, and the water sediment interface were
examined at different frequencies from 150 to 1000 Hz, ranges to 50 km, different
correlation scales and fluctuation standard deviation. The results clearly show that higher
and lower modes decorrelate at different rates. Also the higher order modes and
frequencies were found to decorrelate more quickly with range. The volume fluctuations
were found to have a strong effect at the higher frequencies. For the cases studied here the
coherence decreased monotonically with frequency. The sea surface and bottom produced
a small decrease while the volume fluctuations yielded the strongest influence on
coherence.

E.

::

.&

=-15-

Calculations
Measurements LS lil
Cakulations LS fit
-25-

Ib

20

30

40

50

60

Figure 5.4: This figure shows the RSG versus the array length in wave-lengths, Vh.The mean experimental
results are the dots (*) while the diamonds ( * ) represent the calculated values. The solid and dashed lines are the
least squares tits to the experimental and calculated results.

62
In Fig. 5.4 the calculated RSG is shown along with the averaged measured narrow
band results from Section 3. Also shown are the least square fits for both measured and
calculated points accounting for the uncertainty denoted by the error bars. The
calculations were performed using the phase structure function previously discussed
estimate the coherence function and then the results of section 3 to determine RSG. The
largest difference between these two curves is approximately 0.3dB in addition to the
similar slopes or frequency trends. The calculations show less degradation of the signal
consistent with the slightly higher coherence lengths, 4% versus 30h. Thus excellent
agreement between calculated and measured results has been obtained.
5.2

The Mesoscale Result

Environmental results from the New England shelf break PRIMER experiment
(Gawarkiewicz et a1 [51) emphasized dynamics of the ocean mesoscale during a 400 Hz
acoustic transmission experiment (Lynch et al, [6] and Sperry et al, [4,7]). The
experiment was conducted over a 40x40 km area with an average water depth of lOOm
and was surveyed once a day for a week in a radiator grid pattern, 10 km along shelf and
1 km across shelf, using the "Sea Soar" undulating towed CTD [5]. The daily maps of the
measured thermohaline structure permitted estimation of the sound speed variability,
correlation scales and, consequently, effects of mesoscale variability on the modal phase
structure function. The summer horizontal variability suggests a horizontal correlation
scale of about 10 km in both the along-shelf and cross-shelf directions. This unexpected
isotropy could possibly be a result of the large amount of mesoscale turbulence associated
with the front and fronueddy interactions. This implies that a Gaussian sound speed
correlation function in the x and y directions, with decorrelation scale of 10 km, and a
function for the depth variability maybe appropriate (see Eq. (4.38)).
A reasonable and simplified computational model of the experiment is a rangeindependent waveguide of water depth 100 m. The reference summer sound speed profile
is a seven-day average. Figure 5.5 shows the resulting profile along with a geoacoustic
model of the bottom, a 30 m sediment layer with C,= 1700 m/s and p= 1.9 gm/cm3
overlying an acoustic half space with Cb= 1800 m/s and p=1.9g/cm3.
0

100

130

c, = 1800 m / s
p=1.9g/cm3

Figure 5.5. Background water column and bottom sound speed profiles used for calculations. The sound speed
at the surface is 1510 m/s, at the minimum is 1475, and at the bottom is 1480.

63
Figure 5.6 shows the magnitude of the kernel functions, G,(z) ( Eq. (4.33)), for the first
30 modes at 400 Hz calculated with this waveguide model and are (the kernels are
negative semi-definite).

Figure 5.6. The kernel functions, G,(z) (Eq. (4.33)), for the first 30 modes at 400 Hz calculated with the
simplified waveguide model.

The correlation function defined in equation Eq. (4.3, the kernel Eq. (4.33), and the
covariance function Eq. (4.37) when coupled with the assumption of a Gaussian
correlation function yields the following phase structure function (PSF) (see Eq. (4.38)):

The depth function B(z) used was based on an empirical orthogonal function (EOF)
decomposition of the measured sound speed cross correlation matrix. Figure 5.7 shows
the first ten sound-speed perturbation EOFs. Only the first three were used for
calculations since they contain 90% of the variance. The normalization factor, A, in Eq.
(5.1) scales the modeled variance at a reference depth to a value measured with the Sea
Soar.
Substitution of this correlation function into the expressions for auto- and crosscorrelations with the sound speed perturbation variance of 120 ( m / ~ )at~ a depth of 40 m
enabled the calculation of the PSF. Figure 5.8 shows the resulting modal phase difference
function for modes 1 - 20 at 400 Hz as a function of receiver separation, plotted in units
of acoustic wavelength. Several items are worth noting. The shape of each curve is that
of the error function, and the relative amplitudes are controlled by the projection of B(z)
onto the kernel function, G,.

64
EOF modes for & cwanance matrix

-.,0

10

11

Figure 5.7. The first ten EOF's of the measured sound speed field perturbation.

These can be seen by noting that for a fixed mode number and frequency, the
depth integral is a constant, leaving only the range integral contribution (which has the
form of an error function) to vary with receiver separation. A very important point that
Fig. 5.8 does not convey is the high sensitivity of the PSF to the depth-structure of B(').
Defining an acceptable phase variance between two points be less than
((7~/2)'= 2.5) the phase variance versus d curves shown in Fig. 5.8 are less than this
value at 30 wave-lengths (1 12 m) for the low order modes whereas all the modes are less
than this value at 20 wavelengths. Experimental results with arrays in shallow water
indicate that coherence lengths of 20-30h are consistent with these calculations based on
the mesoscale oceanography.
The same calculations may be done for a variety of frequencies and a fixed receiver
separation. This is illustrated in Fig. 5.9. As frequency increases, the higher modes,
which are initially surface-interacting, begin turning below the surface. This is evidenced
by the sharp bend in the curves, most noticeable in the higher modes. The lowest order
modes have a variance less than 2.5 for frequencies less than 550 Hz, a separation
distance of - 37%.
It is possible to include the effects of geoacoustic variability in the bottom in an
analogus manner as in Sec. 4. For this example, a range-independent waveguide was
used with the addition of sub-bottom sound speed perturbations with Gaussian correlation
functions, a 1 km scale in the horizontal and a 10-meter scale in depth. Figure 5.10 shows
the difference in PSFs with and without bottom perturbations. As would be expected, the
PSF with bottom variability shows greater phase variance than that with only water
column variation. The lower frequencies and lower mode numbers that penetrate deeper
into the bottom show more variability.

65
Modal Phase SttuchneFunctionfa4M)Hz

"0

10

20

30
Receiver separation in A u n i t s

40

so

Figure 5.8. Phase difference function for our PRIMER example case. Modes 1-20 are shown as a function of
receiver separation distance, in units of wavelength h.

Modal Phase Stricture Function & 100 m

Figure 5.9. The modal phase structure function versus frequency is shown for modes 1-20 for receivers
separated by 100m.

66

-0M

-0W

-005
t l

IS0

200

25.0

300
350
Frequnry (Hz)

400

450

500

550

600

Figure 5.10. Phase variance for (PSFw-PSFw+b)

versus frequency showing the relative effect of adding in bottom
structure variation to the waveguide.

Assume that range-dependence lies only in the x-direction, that the path separation
between Rland Rz is small, and that one divides the environment into piecewise rangeindependent segments. Then under the adiabatic assumption single-mode coherence
calculations are a straightforward extension insofar as the phase terms of the PSF are
summed for each spatial segment traversed. The inclusion of mode coupling between
segments would require treating various paths between the source and the receivers and is
not included here.
The preceding development of the modal phase structure function, and our
application of that development using environmental data from the PRIMER experiment,
has some very interesting ramifications. In particular, the measured ocean mesoscale
correlation functions led to a transverse horizontal coherence length between 20-30 h
consistent with the observations at a number of other shallow water sites [3] as well as the
RSG results previously discussed. Moreover, this paper has summarized numerical
studies that show the water column interaction appears to dominate the determination the
horizontal coherence length. Given that the higher angle boundary interacting modes are
attenuated far more quickly than those lower order modes which interact mostly with the
ocean volume, the acoustic field at the receiver is more strongly imprinted by the volume
oceanography. The recurrence of estimated transverse coherence lengths of O( -30 h)
implies that the ocean mesoscale correlation length is generally the most important water
column scale for transverse horizontal coherence. However, this implication should be
treated cautiously. For example, work Oba and Finette [8] has shown that very short
acoustic correlation lengths can be produced by energy ducted between coastal solitons.
In this propagation geometry, the fine scale oceanography scales dominate over the larger
ocean scales in determining the characteristics of the acoustic field. In other cases, the
bottom interacting energy might dominate and strongly influence the acoustic correlation
scales. Thus our hypotheses above discuss tendencies rather than absolutely general
rules that work in all geographical areas.

67
6

Summary and Conclusions

This paper has reviewed the results from recent shallow water sound transmission
experiments and has shown that given the requisite environmental information accurate
calculations are possible. In the case of the sandy bottom waveguide previously
discussed, a minimal set of geophysical parameters on the order of a wavelength deep,
bathymetry and sound velocity profiles are required when transient oceanic volume
fluctuations are absent. Under calm sea state conditions with a derived geoacoustic
description of the bottom, low-frequency lower-order modal propagation was accurately
calculated. At these lower frequencies it was also possible to estimate the horizontal wave
number spectra, Hwns. Since the Hwns contains both amplitude and phase information it
is a useful tool in determining the accuracy of a propagation code.
A key result consistent with previous work was that a nonlinear frequency dependent
attenuation is a characteristic of shallow water sandy sediments. This parameter was
required to calculate the narrowband and broadband field necessary for the estimation of
ideal array gain in the waveguide. The relative signal gain based on the difference of the
ideal and measured was used to estimate a horizontal coherence length for a Gaussian
correlation function. Perturbation theory was applied to the shallow water waveguide
under the assumption of adiabatic propagation to yield phase structure functions for the
surface, volume, and bottom. When Gaussian correlations were assumed with variance
and correlation lengths based on measurements, these phase structure functions permitted
the calculation of the field coherence and consequently the relative signal gain. The result
compared favorably with measurement and a coherence length on the order of 30h (3 400
Hz and a range of 40 km was obtained. These calculations were also found consistent
with replica correlation results. The phase structure function analysis further showed that
the volume fluctuations were the dominant factor in the loss of coherence under the
conditions of downward refraction and moderate sea states in the 50 Hz to 1 kHz range.
These coherence results are consistent with shallow water mesoscale turbulence
associated with fronts and frondeddy interactions.
Experiments are required is selected shallow water areas with extensive
oceanographic and geophysical measurements to quantify the results presented here. The
experiments should cover a broader frequency range utilizing narrow- and broad-band
sources, large horizontal and vertical arrays and ranges to 60 km.
7

Acknowledgements

The authors would like to acknowledge the support of the Oceans Acoustics Program
of the Office of Naval Research, ONR 321, and the partial sponsorship from the Defense
Applied Research Projects Agency, DARPA.

References
1.0 References
[ l ] J.F. Lynch, S.D. Rajan, and G.V. Frisk, A Comparison of Broadband and
Narrowband Modal Inversions for Bottom Geoacoustic Properties at a Site near
Corpus Christi, Texas, J. Acoust. Soc. Am., 89(2), 648-665 (1991).
[2.] L.A. Rubano, Acoustic propagation in shallow water over a low-velocity
bottom, J. Acoust. SOC.Am., 67,1608-1613 (1980)

68
[3] S.D. Rajan, J.F. Lynch, and G.V. Frisk, Perturbative Inversion Methods for
Obtaining Bottom Geoacoustic Parameters in Shallow Water, J. Acoust. SOC.
Am., 82(3), 998-1017, (1987).
[4] W. Carey, J. Doutt, R. Evans, and L. Dillman, Shallow-water Sound
Transmission Measurements on the New Jersey Continental Shelf, J. Oceanic
Eng., 20(4), 321-336,(1995).
[5] S. Rajan, J. Doutt, W. Carey, Inversion for the Compressional Wave Speed
Profile of the Bottom from Synthetic Aperture Experiments Conducted in the
Hudson Canyon Area J. Oceanic Eng. 23(3), 174-187, (1998)
[6] R. Cederberg, W. Siegmann, M. Jacobson, and W. Carey, Predictability of
Acoustic Intensity in Shallow Water at Low Frequencies Using Parabolic
Approximation, J. Acoust. SOC.Am., 94(2), 1034-1043, (1993).
[7] R. J. Cederberg, W. L. Siegmann, and W. Carey, Influence of geoacoustic
modeling on predictability of low-frequency propagation in range dependent,
shallow-water environments, J. Acoust. Soc.Am. 97,2754-2766, (1995).
[8] A. K. Rogers, T. Yamamoto, and W. Carey, Experimental investigation of
sediment effect on acoustic wave propagation in the shallow water, J. Acoust.
SOC.Am. 93, 1747-1761 (1992).
[9] K. J. Howell, M. J. Jacobson, W. L. Siegmann, and W. M. Carey, Effects of
a sediment scattering layer in underwater-acoustic studies of intensity sensitivity
and data matching, J. Acoust. SOC.Am. 100, 1971-1980, (1996).
[ 101 R. Evans, W. Carey, Frequency Dependence Of Sediment Attenuation in
Two Low-Frequency Shallow-Water Acoustic Experimental Data Sets, J.
Oceanic Eng. 23(4), 439-447, (1998).
[ 111 F. Tappert, Analysis of ACT 1 Data Using tne UMPE Mode; Arete Report,
ARS-235-020-TR, Arete Assocs., P.O. Box 8050, La Jolla, Ca. 92038, (1993).
[12] I. Rozenfeld , W. Carey, et al, Modeling and Analysis of Sound
Transmission in the Strait Of Korea, J. of Ocean Eng., 26(4), 809-820, (2001).
[13] J.R. Apel, M. Badiey, et al., An Overview of the 1995 SWARM Shallow
Water Internal Wave Acoustic Scattering Experiment, corresponding author on
manuscript, J. Ocean. Eng. 22(3), 465-501, (1997).
[ 141 J. Lynch, A. Newhall, B. Sperry, G. Gawarkiewicz, P. Tyack, P. Abbot and
C.S. Chiu, Spatial and temporal variations in acoustic propagation
characteristics at the New England shelfbreak front. J. Oceanic Eng., 28( l),
129-150 (2003).
[ 151 T. Duda, J. Lynch, A. Newhall, L. Wu, and C.S.Chiu, Fluctuations of 400
Hz sound intensity in the 2001 ASIAEX South China Sea experiment. J.
Oceanic Eng., 29(3), (2004)

2.0 References
[ l ] W. Carey, J. Doutt, R. Evans, and L. Dillman, Shallow-water Sound
Transmission Measurements on the New Jersey Continental Shelf, J. Oceanic
Eng., 20(4), 321-336, (1995).
[2] A. K. Rogers, T. Yamamoto, and W. Carey, Experimental investigation of
sediment effect on acoustic wave propagation in the shallow water, J. Acoust.
SOC.Am. 93, 1747-1761, (1992).
[3] R. Cederberg, W. Siegmann, M. Jacobson, and W. Carey, Predictability of
Acoustic Intensity in Shallow Water at Low Frequencies Using Parabolic
Approximation, J. Acoust. SOC.Am., 94(2), 1034-1043, (1993).

69
[4] R. Cederberg, W. Siegmann, M. Jacobson, W. Carey, " Parameter
Sensitivities in Two-Layer Isospeed Models of Shallow Water Channels", J.
Computational Acoustics, Vol. 1(4), 469-486, ( 1993).
[5]R. J. Cederberg, et al, "Influence of geoacoustic modeling on predictability of
low-frequency propagation in range dependent, shallow-water environments,"J.
Acoust. SOC.Am. 97,2754-2766, (1995).
[6] M. D. Collins, "A split-step Pade' solution for the parabolic equation
method," J. Acoust. SOC. Am. 93, 1736-1742, (1993).
[7] S. Rajan, J. Doutt, W. Carey," Inversion for the Compressional Wave Speed
Profile of the Bottom from Synthetic Aperture Experiments Conducted in the
Hudson Canyon Area", J. of Ocean Eng. 23(3), 174-187, (1998).
[8] R. D. Stoll, Sediment Acoustics, Springer-Verlag, New York, (1989).
[9] R.J. Cederberg, W. M. Carey, W.L. Siegmann, " Modal analysis of
geoacoustic influences on shallow water propagation", J. Oceanic Eng. 22(2),
237-244, (1997).
[ 101K. J. Howell, M. J. Jacobson, W. L. Siegmann, and W. M. Carey, "Effects of
a sediment scattering layer in underwater-acoustic studies of intensity sensitivity
and data matching," J. Acoust. SOC.Am. 100, 1971-1980, (1996).
[ 111 R. Evans, W. Carey, " Frequency Dependence Of Sediment Attenuation in
Two Low-Frequency Shallow-Water Acoustic Experimental Data Sets", J.
Oceanic Eng. 23(4), 439-447, (1998).
[ 121 F. Ingenito, "Measurements of mode attenuation coefficients in shallow
water," J. Acoust. SOC.Am. 53, 58-863, (1973).
[13] J. H. Beebe, S. T. McDaniel, and L. A. Rubano,
Shallow-water
transmission loss prediction using Biot sediment model," J. Acoust. SOC.Am.
71, 1417-1426, (1982).
[14] P. Rogers, J. Zhou, X.Zhang,F. Li, " Seabottom acoustic parameters from
inversion of Yellow Sea experimental data" in Experimental Acoustic Znversion
Methods for Exploration of the Shallow Water Environment , Eds. A. Caiti, J.P.
Hermand et al, Kluwer Acad. Pub. Boston, M a , 219-234, (2000).
[ 151 F. Tappert, "Analysis of ACT 1 Data Using tne UMPE Mode; Arete Report,
ARS-235-020-TR, Arete Assoc., P.O. Box 8050, La Jolla, Ca. 92038, (1993).
[16] D. Chizhik, J. M. Tattersall, Application of Biot Theory to the Study of
Acoustic Reflection from Sediments-Naval Undersea Warfare Center, New
London, Ct., NUWC-NL Technical Report 10,115,(1992)- Avail. DTIC.
[17] J. Tattersall, D. Chizhik, B. Cole, F. DiNapoli, " The effect of frequencydependent bottom reflectivity on transmission loss in shallow water over a sandy
bottom", 125 * Mtg. of the Acoust.Soc. of America, Ottawa, Canada, (1993).
[18] I. Rozenfeld , W. Carey, et al, " Modeling and Analysis of Sound
Transmission in the Strait Of Korea", J.of Ocean Eng., 26(4), 809-820,(2001).
[19] W. Carey, P. Cable, W. Siegmann, J. Lynch and 1. Rozenfeld, "
Measurement of Sound Transmission and Signal Gain in the Complex Strait of
Korea," J. of Oceanic Eng., 27(4), 841-852, (2002).
I'

3.0 References
[ 1.] G. C. Carter, C. H. Knapp, and A. H. Nuttall, " Statistics of the estimate of
the magnitude coherence function," IEEE Trans. on Audio and Electroacoustics,
AU-21(4), pp.388-389, (1973).

70
[2.] G. C. Carter, C. H. Knapp, and A. H. Nuttall, Estimation of the magnitude
squared coherence function via overlapped FFT processing, IEEE Trans. on
Audio and Electroacoustics, AU-21(4), 337-344, (1973).
[3.] E. R. H. Scannell and G. C. Carter, Confidence bounds for magnitudesquared coherence estimates, Proc. IEEE Int. Conf. on Acoustics, Speech, and
Signal Processing, 670-673, (1978).
[4.] W. M. Carey, The determination of signal coherence length based on
signal coherence and gain measurements in deep and shallow water, J. Acoust.
SOC.Am., 102, 831-837, (1998).
[5 .] W. Carey and W. Moseley, Space-Time Processing, EnvironmentalAcoustic Effects, in Progress in Underwater Acoustics, Plenum Press, New
York, 1987, pp. 743-758 (Invited Paper to 12th International Conference for
Acoustics) (Also Published in J. Oceanic Eng., 285-301, (1991)).
[6.] I. Rozenfeld, Analysis of Transmission Loss Signal Gain, and Coherence In
Shallow Water, Dissertation, Rensselaer Polytechnic Institute, Troy, New York,
(Jan. 2001).
[7.] W. M. Carey, P. G. Cable, W. Siegmann, J. F. Lynch, I. Rozenfeld,
Measurement of Sound Transmission and Signal Gain in the Complex Srtait of
Korea, J. of Ocean Eng., 27(4), pp. 841-852, (2002).
[8.] P. Wille and R.Thiele, Transverse horizontal coherence of explosive signals
in shallow water, J. Acoust. SOC.Am. 50 (lpt.2), 348-353, (1971).
[9.] R. Scholz, Horizontal Spatial Coherence Measurements with Explosives
and CW Sources in Shallow Water, in Aspects of Signal Processing. Part I , G.
Tacconi (ed.), D. Reidel Publishing Co., Dordrecht-Holland, 95-108, (1977).
[ 10.1 A.Wasiljeff, Spatial Horizontal Coherence of Acoustical Signals in
Shallow Water, SACLANTCEN SM-68, SACLANT ASW Research Centre,
LaSpezia, Italy, May 1975.
[ 11.1 W.M. Carey, Measurement of down-slope sound propagation from a
shallow source to a deep ocean receiver, J. Acoust. SOC.Am. 79(1), 49-59,
(1986).
[ 12.1 W. M. Carey, I. B. Gereben, and B. A. Brunson, Measurement of sound
propagation downslope to a bottom-limited sound channel , J. Acoust. SOC.
Am. 81(2), 244-257, (1987).

4.0 References
[ 11 H. Weinberg and R. Burridge, Horizontal ray theory for ocean acoustics, J.
Acoust. SOC.Am. 55,63-79, (1974).
[2] L. M. Brekhovskikh and 0. A. Godin, Acoustics of Layered Media 11: Point
Sources and Bounded Beams, Springer-Verlag, (1992).
[ 3 ] R. Zhu and D. Guan, Spatial horizontal coherence of sound in shallow
water, J. Acoust. SOC.Am. 92,956-961, (1992).
[4] I. Rozenfeld, Analysis of Transmission Loss Signal Gain, and Coherence In
Shallow Water, Dissertation, Rensselaer Polytechnic Institute, Troy, New York,
(Jan. 2001).
[5] I. Stakgold, Greens Functions and Boundary Value Problems, 2ndEd., John
Wiley (1997).
[6]S.D. Rajan, J.F. Lynch, and G.V. Frisk, Perturbative Inversion Methods for
Obtaining Bottom Geoacoustic Parameters in Shallow Water, J. Acoust. SOC.
Am., 82(3), 998-1017, (1987).

71
[7] Shang, E.C. and Wang, Y. On the calculation of modal travel time
perturbations. Sov. Phys. Acoust, 37,411-413 (1991).
[8] J. F. Lynch, S. D. Rajan, and G.V. Frisk, A comparison of broadband and
narrowband modal inversions for bottom geoacoustic properties at a site near
Corpus Christi, TX, J. Acoust. SOC.Am. 89,648-654 (1991).
[9] S. M. Flatte, R. Dashen, W. H. Munk, K. M. Watson, and F. Zachariasen,
Sound Transmission Through a Fluctuating Ocean, Cambridge Univ. Press,
Cambridge, U.K., (1979).
[lo] J. F. Lynch, On the use of focused horizontal arrays as mode separation
and source location devices in ocean acoustics. Part I: Theory, 3. Acoust. SOC.
Am. 74, 1406-1416 (1983).
[ l l ] J. L. Worzel, M. Ewing, and C.L. Pekeris, Propagation of Sound in the
Ocean, Acoust. SOC.Amer., (2000).
[12] D. Tang and G. V. Frisk, Spectral parameterization of scattering from a
random ocean bottom, J. Acoust. SOC.Am. 92,2792-2799, (1992).
[13] B. J. Sperry, Analysis of Acoustic Propagation in the Region of the New
England Continental Shelfbreak, Dissertation,, Woods Hole Oceanographic
Institution, Woods Hole, MA, (1999).

5.0 References
[l.] I. Rozenfeld, W. Carey, et al,
Modeling and Analysis of Sound
Transmission in the Strait Of Korea, J. of Ocean Eng., 26(4), 809-820, (2001).
[2.] I. Rozenfeld, Analysis of Transmission Loss Signal Gain, and Coherence In
Shallow Water, Dissertation, Rensselaer Polytechnic Institute, Troy, New York,
(Jan. 2001).
[3.] W. M. Carey, P. G. Cable, W. Siegmann, J. F. Lynch, I. Rozenfeld,
Measurement of Sound Transmission and Signal Gain in the Complex Srtait of
Korea, J. Oceanic Eng., 27(4), pp. 841-852, Oct. 2002
[4.] B. J. Sperry, Analysis of Acoustic Propagation in the Region of the New
England Continental Shelfbreak, Dissertation,, Woods Hole Oceanographic
Institution, Woods Hole, MA, (1999).
[5.] G. Gawarkiewicz, F. Bahr, K. Brink. R.C. Beardsley, M. Caruso, J. Lynch,
and C.S. Chiu, A large amplitude meander of the shelfbreak front during
summer south of New England: Observations from the shelfbreak PRIMER
experiment., in press J. Geophys. Res., (2003).
[6.] J. Lynch, A. Newhall, B. Sperry, G. Gawarkiewicz, P. Tyack, P. Abbot and
C.S.Chiu, Spatial and temporal variations in acoustic propagation characteristics at the New England shelfbreak front. J. Oceanic Eng., 28( l), pp. 129150, (2003).
[7] B. Sperry, J. Lynch, G. Gawarkiewicz, C. Chiu, and A. Newhall,
Characteristics of acoustic propagation to the eastern vertical line receiver
during the summer 1996 New England shelfbreak PRIMER experiment. In
press J. Oceanic Eng., 28(4),719-749, (2003).
[8] R. Oba and S. Finette, Acoustic propagation through anisotropic internal
wave fields: transmission loss, cross-range coherence, and horizontal refraction,
J. Acoust. SOC.Am. 111,769-784,2002.

72
FRACTIONAL DERIVATIVE MODELLING OF ACOUSTIC DISSlPATIONS
OBEYING ARBITRARY FREQUENCY POWER LAW
W. CHEN AND S. HOLM
Simula Research Laboratory, P.O. Box 134, NO-1325 Lysaker, Norway
E-mail: wenc@,simula.no
The standard time-domain partial differential wave equation models of integer orders can describe
merely two extreme cases of ftequency independent and frequency-squared dependent acoustic
dissipations. The otherwisenon-zero and non-square frequency dependency occurring in many cases of
practical interest is thus called the anomalous dissipation.This paper aims to present a brief summary of
our recent works in developing new fractional derivative time-space wave equations to accommodate the
anomalous dissipationsas well as the new definitionsofthe fractional derivatives in space and time. The
Levy stable distributiontheory is also found to provide a mathematical physics interpretationof the [0,2]
magnitude range of acoustic dissipation dependency on kequency.

Introduction

The effect of the dissipation plays a prominent role in acoustic applications, for instance,
seismic wave propagation, acoustic boundary layer, and ultrasound surgery and second
harmonic imaging. Acoustic dissipation typically exhibits E = E,e-ai@)',characterized by an
empirical power law function of fiequency a(w)=

sold' ,Y E [0,2], where E represents the

amplitude of an acoustic field variable such as pressure, z is the traveling distance, and o
denotes angular frequency [I ,2]. cyo andy are media-specific parameters obtained through a
fitting of measured data. It is well known that the standard mathematical modelling
approach based on partial time-space derivatives of integer orders can not accurately reflect
frequency power law dissipation except for two extreme cases: y=0,2. Unfortunately,
O<y<2 exponent cases, called anomalous dissipation, present in most media of practical
interest. For example, sediments and fiactal rock layers have y around 1, different human
tissues have distinct values ofy [3] as shown in Table I .
Table 1. Tissue coefficients of frequency-dependent power law dissipation

cq,(dB/cm/MHz?)
Y

Water

Fat

Duct cancer

structural tissue

0.0022

0.158

0.57

0.87

1.7

1.3

1.5

Recent decades have witnessed great effort to develop a variety of the anomalous
dissipation models such as the adaptive proportional damping model [4], the multiple
relaxation models [5],the frequency decomposition model [6],the Z-transform method [7],
and the Fourier transform method. These modelling approaches, however, have their
respective disadvantages such as inapplicability for nonlinear media and broadband pulse
propagation, computational demanding, requirements of many elusive parameters, and
great mathematical complexity.

13

On the other hand, the fractional time derivative has long been found to be a very
powerful approach to build time-domain lossy wave equations modeling the anomalous
dissipation [8-lo]. The paper aims to summarize our recent advances in developing some
new fractional derivative concepts and corresponding time and space wave equations
[1 1- 141. Among our original technical contributions are
the introduction of the concept of the positive time fractional derivative;
the presentation of the modified Szabo wave equations via the positive time fractional
derivative, where the hyper-singularity of the original Szabo wave equation models for
anomalously attenuative media is eased;
the new definition ofthe fractional Laplacian, which overcomes the hyper-singularity of
the traditional definition;
the development of the linear and nonlinear causal fractional Laplacian wave equations
and the corresponding FEM models for lossy media exhibiting arbitrary frequency
power law dissipation;
the first mathematical physics explanation of [0,2] power dependency of dissipation
coefficient y on frequency via the Levy stable distribution theory.
Positive fractional derivative and modified time-domain Szabo wave models

Szabo [2] developed a new causal lossy time-domain wave equation including arbitrary
fiequency-dependent power law acoustic dissipation, namely,
1 azp 2au
Ap=--+-S,(p)

.,;

atz

c,

where

y=2.

where T i s the gamma function. The above Szabos equation satisfies a general dispersion
equation [2] for arbitrary power y of frequency dependent dissipation
k 2 = (W/C,)~ + i2(~/c,)a,l4
where the wave number k=P+iacorresponds to the Fourier transform of the space variable.
Band aare respectively related to the dispersion and the dissipation. However, the Szabos
equation encounters the severe hyper-singularity and is not clear how to implement the
initial conditions. It is noted that the Szabos models can be recast with the
Riemann-Liouville fractional y-order derivative [1 11. Then, our basic strategy is that the
Riemann-Liouville fractional derivative in the Szabos equation is replaced by the Caputo
fractional derivative while keeping the positivity. Thus, the initial conditions can be easily
prescribed and the hyper-singular integral is regularized inherently. To simplifL the
expression and formalize for potential applications in a broad areas, a new positive time
fractional derivative is introduced [ 1 11

74

whose Fourier transform characterizes the positive operation as follows

FT(DlVIp(l))=Ioy P(w),

O<F2

Note that the fractional derivatives in the Riemann-Liouville sense and the Caputo sense do
not satisfy the above Fourier transform formula. On the other hand, the positive fractional
derivative can be understood a subset of the Caputo fractional derivative since both differ
only in the constant coefficient. Thus, the existence, uniqueness, and well-posedness, and
numerical solution methods of the positive fractional time derivative equation are
essentially the same as those of the Caputo derivative equations. With the help of the
positive fractional time derivative, the modified Szabos equation [ 1 13 is developed

where Dllp = p , Dllp = -a2p/St2 .Note that the modified Szabos equation satisfies the
previous Szabos general dispersion equation but still encounter the hyper-singularity
when y=l .
3
3.I

Fractional Laplacian time-space lossy wave equation

New deJnition ojthe Fructionul Lupluciun

The fiactional Laplacian is also often called the Riesz fractional derivative in terms of the
Riesz potential [15]

where 0 is integration domain, and d represents the dimension. For d=l, s=l leads to a
singularity, which we do not discuss in this study. The traditional definition of the
fractional Laplacian can be stated as [ 121

(5)
which encounters the hyper-singularity d+s>d and uneasy to implement the boundary
conditions. By analogy with the Caputo time derivative, we presented anew definition [12]
r[(d - 2 + s ) p ]
Add dQ(5)
P(X> = -I?
[Aq,(dl= - p r y 222- r[(2 - s)/21
- d(J-z+r
(-

L!Ix

which overcomes the drawbacks of the traditional definition. By using the Green second
identitv. we can establish the link between the two definitions

75
where D and N are respectively the Dirichlet and Neumann boundaq conditions. Note that
the fractional Laplacian is a positive definite operator.
3.2

Linear fractional Laplacian time-space wave equation model

Blackstock [I61 pointed out that the space or space-time modelling of thermoviscous
behavior is often replaced by a pure time operation under the condition that the
thermoviscous term is relatively small. Ochmann and Makarov [lo] fbrther elaborate that
this replacement is impossible in the general case where the interaction between two
oppositely traveling sound waves can not be neglected. The only preference of the time
expression is mostly due to its ease of analysis. It is well known that anomalously
attenuative and dispersive media often establish complicated microstructures in space, and
we also notice that the exponent y is independent of the frequency but dependent on the
medium, the spatial fractional derivative may therefore instead be more suitable as a
modelling approach.
We developed a causal linear time-space wave equation [ 121

which underlies the power law frequency dissipation in the frequency domain. When y=2,
the above equation is equivalent to the thermoviscous wave equation. When y=O, it is
reduced to the standard frequency-independent damped wave equation.
3.3

Finite element numerical model

Let the FEM approximate discretization of a Laplacian operator be expressed by -41* @,

where j? represents the pressure value vector at the discrete nodes. Since the Laplacian is
the positive definite operator, it is well known that its FEM discretization matrix K is also
positive definite. The corresponding FEM formulation of the factional Laplacian of y/2
order is then intuitively obtained by (- A)'/* p 3 K ""Zj [12], where K '/' =
.

@TEA:)"'2@

Accordingly, the FEM discretization of the fractional Laplacian wave equation model is
stated as
Zj,, + 2a,,c:fyKypp,+ c2@ = g ( t )
where the subscript t represents the temporal derivative, and g(t) is due to the source term.
It is obvious that the above semi-discrete FEM model readily takes into account
frequency-dependentviscous effects for a multitude of frequency components (broadband
signal) with empirical coefficients @ and y of the power law dissipation. When y=2, the
semi-discrete model brings out the square frequency dependence and is in fact the classical
Rayleigh proportional damping model. We do not present any numerical verification o f the
present models because of the limited space. For some preliminary experiments, please go
to http://heim.ifi.uio.no/-aichab/Results 0803.html.
4

L&vystable distribution and YE [0,2]

By ignoring the left-side term, the above hyperbolic fractional Laplacian wave equation can
be approximated to the anomalous diffusion equation

76

The fundamental solution of the standard diffusion equation Cv=2) coincides with the
spatial probability density function of the Gaussian normal distribution, while the
fundamental solution of the anomalous difision equation corresponds to the LCvy y-stable
probability distribution, namely, the Cauchy problem solution of the above equation is
where L is the Levy distribution density function. In order to be a stable law, a y-stable
distribution requires the power y to be positive but not greater than 2, i.e., 0 + y I 2 . In
terms of this mathematical physics backdrop, the media having y>2 power law dissipation
are not statistically stable in nature [ 131. It is noted that the LCvy process does not include
y=O, which means that the media obeying absolutely frequency-independent dissipation is
simply an ideal approximation. For acoustic wave propagations, all media exhibit more or
less degree of frequency-dependent dissipation.
y can also be interpreted as the fractal dimension. In terms of geometric shapes, there
are the two kinds of fkactal dimension: process (or function) and trail (or path) fractals [ 171.
The former is commonly known as the H a u s d o e dimension in the literature, while our
analysis [I31 shows that the fractional dimensiony is associated with the trial fractal.
5

Remarks

By using both the fi-actional time derivative and fractional Laplacian, we [14] also
presented a lossy wave equation

where y denotes the viscous constant, s and 77 are a real number. The above equation
conforms the frequency power law dissipation and can be approximated to the generalized
wave-diffusion equation [ 181

It is also worth stressing that the present time and space fractional derivative strategies can
be directly used to develop the corresponding nonlinear lossy wave equation models [ 121
which underlie the frequency power law dissipation.

Acknowledgements

The work reported here is sponsored by Simula Research Laboratory with the project
Mathematical and numerical modeling of medical ultrasound wave propagation.

References
1. He, P., Simulation of ultrasound pulse propagation in lossy media obeying a frequency
power law, IEEE Trans. Ultra. Ferro. Frey. Contr. 45(1) (1998) pp. 114-125.
2. Szabo, T. L., Time domain wave equations for lossy media obeying a frequency power
law,J. Acoust. Soc. Amer. 96(1) (1994) pp. 491-500.

77

3. DAstrous, F.T. and Foster, F.S., Frequency dependence of ultrasound dissipation and
backscatter in breast tissue, Ultrasound in Med. & Biol. 12( 10) (1986) 795-808.
4. Wojcik, G., Mould, J., Lizzi, Jr., F., Abboud, N., Ostromogilsky, M. and Vaughan, D.,
Nonlinear modelling of therapeutic ultrasound. In 1995 IEEE Ultrasonics Symposium
Proceedings, pp. 1617-1622.
5 . Nachnian, A. I., Smith, J . and W a g , R.C., An equation for acoustic propagation in
inhomogeneous media with relaxation losses, J . Acoust. Soc. Amer. 88(3) ( I 990) pp.
1584- 1595.
6. Chen, W, Holm, S., Bounaim, A,, 0degArd, A,, Tveito, A., Frequency decomposition
time-domain model of broadband frequency-dependent absorption. In the 9th
Worluhop of the Finite Element Method in Biomedical Engineering, Biomechanics
and Related Fields, Ulm, Germany (2002) pp. 40-48.
7. Wismer, M. and Ludwig, R., An explicit numerical time domain formulation to
simulate pulsed pressure waves in viscous fluids exhibiting arbitrary frequency power
law dissipation, IEEE. Trans. Ultrason. Ferroelect. Frey. Contr. 42(6) ( 1 995) pp.
1040-1 049.
8. Capiito, M., Linear models of dissipation whose Q is almost frequency independent-11,
Gotph-vs.J. R. utr. Soc. 13 ( 1 967) pp. 529-539.
9 . Baglegy, R. L. and Torvik, P. J., A theoretical basis for the application of fractional
calculus to viscoelasticity, J. Rheol. 27 ( 1 983) pp. 20 1-2 10.
10. Ochmann, M. and Makarov, S., Representation ofthe absorption ofnonlinear waves by
fractional derivative, J. Acoust. SUC.Am. 94(6) ( I 993) pp. 3392-3399.
1 1 . Chen, W. and Holm, S., Modified Szabos wave equation models for lossy media
obeying frequency power law, J. Acoust. Soc. Amer. (accepted) 2003.
12. Chen, W. and Holm, S., Fractional Laplacian time-space models for linear and
nonlinear lossy media exhibiting arbitrary frequency dependency, J. Acoust. Soc. Amer.
(accepted) 2003.
13. Chen. W. and Holm, S., Ltvy stable distribution and [0,2] power dependence of the
absorption coefficient on the frequency, IEEE Trans. Ultrasonics, Ferroelectrics, and
Frequency Control (submitted) 2003.
14. Chen, W. and Holm, S . , Physical interpretation of fractional diffusion-wave equation
via lossy media obeying frequency power law, @reprint), March, 2003.
15. Samko, S. G., Kilbas, A. A., Marichev, 0. I., Fractional Integrals and Derivutive.v:
Theory and Applications (Gordon and Breach Science Publishers, 1987).
16. Blackstock, D. T., Transient solution for sound radiated into a viscous fluid, J. Acoust.
Soc. Am. 41 (1967) pp. 1312-1319.
17. Tatom, F., The relationship between fractional calculus and fractals, Fructu1.v 3(1)
( I 995) pp. 2 17-229.
18. Hanyga, A., Fractional d i h s i o n and wave equations. In conference of Modelli
Matematici e Problemi Analitici per Materiali Speciali, Cortona, 2001,

78

VISUALIZATION OF THE ENERGY FLOW FOR ELASTIC

WAVES: COMPARISON AND CONTRAST OF CONVENTIONAL
VECTOR FIELD VS. COLOR CODED REPRESENTATION OF
THE POYNTING VECTOR
CLEONE. DEAN
Physics Department, P.0.B 8031, Georgia Southern University, Statesboro, Georgia 304618031, USA
E-mail: cdean @GeorpiuSoitthern.edu

JAMES P. BRASELTON
Department of Mathematics and Computer Science, Georgia Southern University, P.O.B. 8093,
Statesboro, Georgia 30460-8093, USA
E-mail: jbraselton@GeorpiaSouthern.edu

Comparison and contrast is made of two complementary methods of visualizing both the timeaveraged and instantaneous elastodynamic Poynting vector for two examples showing elastic
waves: the Rayleigh wave and the Lamb wave. The two methods are the conventional vector field
approach with a set of Poynting vector arrows emanating from a grid of field points and a color
coding scheme whereby hue is used to indicate the direction of the energy flow while intensity of
the color is used to denote the magnitude of the Poynting vector at that point. Each of the two
methods has its successes with some aspects of the visualization and deficiencies in others. The two
methods in combination yield deeper insight into the Poynting vector field.

1 Introduction
The directon of energy flow in ensonified elastic solids and their
surroundings is sometimes counter to physical intuition [1,2]. While it is
possible to plot the undulations of the surface of a scatterer in an attempt to
see the energy flow around the surface this can lead to difficulties. It is
important to remember that the energy flow is not associated with the phase
velocity of the waves, i.e., their individual undulations. Energy flow is
associated with the group velocity of the waves. One must follow packets or
groups of waves around the periphery of the scatterer. This can prove
difficult to do in practice whether with numerical simulation or actual
observation.
The obvious alternative at least so far as theoretical and computational
acoustics is concerned is to plot the elastodynamic Poynting vector field for
the problem in question [3]. The elastodynamic Poynting vector itself has
some interesting properties and is worthy an examination in and of itself. On
the other hand the question arises: how to represent the Poynting vector field
once it has been calculated throughout the region of interest. Two obvious
proposals are as follows:
1. Plot the elastodynamic Poynting vector as a field of arrows at preselected
grid points. The length of the arrow gives the intensity of the energy flow

79

at that point and the direction of the arrow yields the direction of the
energy flow at that point.
2. Plot the field as a set of color values at each selected gridpoint. The color
of the point gives the direction of the energy flow while the intensity of
the energy flow is given by the intensity of the color-dot.
Each of these methods has its own advantages and disadvantages. For some
problems one method gives a clearer picture than the other. For others the
two methods are complementary and together yield the clearest picture of the
energy flow pattern. In this paper both methods of visualization will be used
for two very simple problems, the Rayleigh wave and Rayleigh-Lamb waves,
and then for the problem of a forward guided wave in and around a fluidloaded thin cylindrical shell at resonance.

2 The Elastodynamic Poynting Vector

The elastodynamic Poynting vector has interesting and somewhat poorly
known properties. It is an interesting physical quantity in and of itself [4].

2.1 The Instantaneous Poynting Vector

The instantaneous Poynting vector in a lossless isotropic elastic medium can
be derived [3] as

Sin,,,= -v' . T ,
where 3 is the particle velocity and

f is stress tensor for the medium. This

is a consequence of the Poynting theorem for an elastic medium:

dU
f).lidA + =P ,
dt

where A is the surface area of a small particle in the medium, li is a unit

normal on the surface of the particle, U is the potential energy of the particle,
and P is the power of the wave passing through the particle. Frequently the
time averaged elastodynamic Poynting vector is used, and it is the mean value
of the instantaneous Poynting vector over one period of the wave.
2.2 Thecomplex Poynting Vector
The complex Poynting vector is related to the instantaneous and time
averaged Poynting vectors and can be derived by assuming simple harmonic
time variation for the physical parameters in the problem. Then each physical

80

quantity can be represented as having a complex exponential factor of the

form:

where k is the wavenumber and w is the angular frequency of the wave.

Thus each physical quantity must now be interpreted as the real part of
its complex analog. Also time derivatives can be replaced by multiplication
with - iw or by multiplication by + iw if the quantity is complex cojugated.
These conventions lead to a complex Poynting theorem for harmonic waves
in a lossless isotropic elastic medium of the form:

Here the asterisk denotes complex conjugation, while P,,, is the maximum
power supplied by sources external to the particle, and UKE,,, and fE,rnax
are, respectively, the maximum kinetic and potential energies of the particle.
Thus the complex Poynting vector can be defined as

If we define two real vectors

3,sE R3such that

then we can readily split the complex Poynting vector into real and imaginary
parts.
2.2.1 Real Part
The real part of the complex Poynting vector can be shown to satisfy

where the angular brackets imply an average over one period. In terms of the
complex variables, the instantaneous Poynting vector is now in terms of the
real parts of the complex particle velocity and stress tensor:

The real part of the complex Poynting vector gives the time average of the
instantaneous energy flow and shows the presence of a traveling wave.

81

2.2.2 Imaginary Part

The imaginary part of the complex Poynting vector can be written as

where the magnitude of the imaginary part is, by Eq. (4),

where c = o / k is the phase velocity and u

~and uPE,,,=
~
, are, respectively,
~
~
~
the maximum kinetic and potential energy densities. The unit vector $' is
perpendicular to surfaces of constant uKE,max
- u ~ ~Note
, that
~ for
~ a. simple
plane wave
would vanish. A non-vanishing imaginary part of the complex
Poynting vector shows the presence of a standing wave.
2.3 Geometric Interpretation

If we split the complex amplitudes of the particle velocity and stress tensor
elements into their real and imaginary parts such that

?A]

where each element [To], [T i ], [Go], and [ is real and where 8 = kx - mt .

Then after some algebra the instantaneous Poynting vector can be shown to
be

where

This result should be compared with the somewhat simpler formulas derived
by Mann, et al. for the case of acoustic compressional waves only [5]. The
difference is due to the tensor nature of the elastic problem. To differentiate
our results from those of Mann, et al., that result will be referred to as the
acoustic Poynting vector. The term elastic or elastodynamic Poynting vector
will be used for the current results.
As can be seen from Eq. 12, the figure traced out over time by the tip of
the Poynting vector is an ellipse whose center is the time average of the
vector (See Figs. 1 and 2). Also the tip of the Poynting vector makes two
circuits of the ellipse in one period of the harmonic elastic wave. In the
special case that the vectors Q f i , the vectors Q and 2 form the semimajor

82

and semiminor axes of the ellipse, and the Poynting vector goes around the
ellipse at a constant angular velocity of 2w(See Fig. 1). More generally Q
and 1? are not perpendicular. In that case the Poynting vector still makes two
trips around the ellipse in one period of the wave, but not at uniform angular
velocity. Since the more general case can be derived from the special case
where Q
G by dual shear transformations (or, equivalently, by a velocity
boost transformation as in relativity) followed by circular rotations, it can
be shown that the general elastic Poynting vector traces out sheared
quadrants in equal time intervals (See Fig. 2).

Figure 1. Special case: the Q and i vectors are mutually perpendicular and form the semi-major and
semiminor axes of the ellipse traced by the instantaneous Poynting vector. The tip of the instantaneous
Poynting vector makes two circuits of the ellipse in one time period of the wave at constant angular velocity w.

Figure 2. General case: the Q and i vectors are not mutually perpendicular. The semi-major and semiminor
axes of the ellipse traced by the instantaneous Poynting vector can be found from Q and i? by hyperbolic
rotation or dual shear transforms with circular rotations. The tip of the instantaneous Poynting vector makes
two circuits of the ellipse in one time period of the wave but not at constant angular velocity 4 instead each
sheared quadrant is traversed in equal time.

Since the point of this paper is to contrast two different visualization

methods, the details of the derivations of the solutions to the problems to be

83

shown will be omitted. Suffice it to say that all three examples are solved by
expanding the scalar and vector potentials in appropriate orthogonal
polynomial expansions and applying appropriate boundary conditions [2, 6,
7, 81. The particle deflection is used to solve for the particle velocity, and
appropriate stress tensor elements are also solved for so as to find the
various components of the instantaneous and complex elastodynamic
Poynting vectors on a predefined grid of locations.

3 The Rayleigh Wave

The Rayleigh wave is the simplest case that has a nonvanishing imaginary
part to its complex Poynting vector. Consider a homogeneous, isotropic,
elastic half space fronting a vacuum.

Figure 3. Geometry for Rayleigh wave: the elastic medium is on the left; vacuum is on the right. The y-axis is
out of the page, and the propagation direction is toward +x.

For free waves in the isotropic medium the equation of motion is

where U = displacement, p = mass density, t = time, and p and A are the

so-called L a n k moduli or L a n k coefficients of the material. The general
solution to this wave equation is a combination of longitudinal waves (also
known as compressional waves) and transverse (or shear) waves:
u'= Ue + U f , satisfying Ue = V q and Uf = V x p (where V x Gt = O and
V.Uf =O). The longitudinal and transverse wave speeds are given by
c: = ( A + 2p) / p and c," = p / p, respectively. The scalar and vector
potentials may be written as

respectively, where @ ( Z ) = @,eW and Y(Z) =

contain h e expected
exponential decay. The appropriate boundary equations to obtain a

84

dispersion relation for these so-called Rayleigh waves are expressible in

terms of the elastic stress tensor f and elastic strain tensor E". The elastic
stress tensor with components qj may be defined as
3

-CqjdAj =4,
j=1

where F, is the ith component of the force acting on a surface d i = dA2

with outward normal 2. For sufficiently small strain the strain tensor may be
defined as

Yi

&.. = 'I

-+-

2 dXj

&J)hi

Then Hooke's law for the isotropic medium may be written as

where 9 .u' = TrE". The solution of the problem leads to two simultaneous
equations in @, and Yo.The first equation is

The resulting complex particle velocity has components

and

The nonzero components of the complex stress tensor are

where

85

These results lead to a parametric equation for an ellipse that the Poynting
vector taces out twice per cycle of the incident plane wave:

as previously described. Specifically

= + & [ 2 k d 0 e m - ( k 2-I-p2)Y~ea](-@@,e

Q=-S?,

+ &Yieb)}.C,

(27)

(28)

and

Note that the ellipse is not sheared, but corresponds rather to the case where
Q is perpendicuular to
as in Fig. 1. Furthermore the imaginary part of
the complex Poynting vector is S = R?. Using values for rolled aluminum
[9], the real, time-averaged, elastic Poynting vector is shown in Fig. 4.

z,

Figure 4. The real part of the complex elastodynamic Poynting vector field of the Rayleigh wave for
aluminum shows a time-averaged traveling wave parallel to and just below the surface. The units are
fractions of the Rayleigh wavelength.

Figure 4 shows some of the drawbacks of the conventional grid of arrows in

showing a vector field. The arrows overlap near the surface such that the
field is clearly shown only in the upper region of the figure just below the
surface. Also the arrowheads are drawn even when the intensity of the

86

Poynting vector has dropped to almost zero, e.g., deeper than 0.1 wavelength into the elastic medium, such that the field appears to drop to some
minimum value (the length of the arrowhead) rather than zero.
Figure 5 shows the imaginary part of the complex Poynting vector field.

0.05

0.1

Figure 5 . The imaginary part of the elastodynamic Poynting vector field of the Rayleigh
wave for aluminum shows a standing wave with equipotentials of the differenced in maximum
potential and kinetic energy densities parallel to and just below the service. The units are
fractions of the Rayleigh wavelength.

Again the field vectors overlap, and again the arrowheads persist at depth
within the aluminum half-space, giving a false sense that that the energy flux
density doesnt effectively vanish within one-tenth of a wavelength of the
surface.
Figure 6 shows one frame from a conventional arrow-as-vector
animation of the instantaneous Poynting vector field.

0.05

0.1

Figure 6. This graphic is one frame from an animation of the instantaneous Poynting vector field for
aluminum. Although some vectors protude into the vacuum all energy remains below the surface. The units
are fractions of the Rayleigh wavelength..

This animation suffers from the same problems as the graphs in Figs. 4 and
5, but does make clear, due to the arrowheads, which way the energy is
flowing.
Figure 7 shows one frame from an animation of the instantaneous
Poynting vector field where color gives the direction of the wave according to

87

acolor field while intensity of the energy flow is given by the brightness of
the hue.

005

0 1

Figure 7. This graphic is one frame from an animation of the instantaneous Poyntmg vector field for
aluminum where color gives the direction according to position of the hue on a color wheel and intensity is
given by the brightness of the hue. Blue is basically down the figure and parallel to the surface. The units a r e
fractions of the Rayleigh wavelength.

This scheme has the advantage that the exponential decay of the wave is
shown clearly by the drop-off in the intensities of the colrs depicting the
energy flow. What is not evident in a still frame is that animation of the
energy flow may, due to a sort of stroboscopic effect based on the time
-sampling of the field, can make the field look like it is proceeding up the
figure rather than down it. Furthermore the actual directions of the vector
arrows are settled most clearly only by comparison with the conventional
plot of Fig. 6.
4 Rayleigh-Lamb Waves

Now consider an infinite plate with vacuum on either side of it as shown in

Figure 8.

Figure 8. This shows the geometry for Rayleigh-Lamb waves. The center space is an infinite elastic plate
with Lam6 coefficients h and p fronted on either side by vacuum.

The problem of elastic waves in an infinite plate facing a vacuum on either

side was first solved by Lamb [8]. With the geometry shown in Fig. 8, the
solution follows in a fashion very similar to that for the Rayleigh wave,
except that there are now two surfaces and two sets of boundary conditions.

88

This means that there are multiple possible solutions, the so-called RayleighLamb waves, of varying frequencies. The various frequencies correspond to
various number nodes for standing waves fitting in the width of the plate.
Also there are two kinds of solution: one where the surfaces of the plate
move in and out together - the so-called symmetric solutions - and one
where the surfaces move on one side outward while the other side moves
inward and vice versa - - the so-called antisymmetric solutions. The
conventional approach has been to show contours of the displacement
function. However to show the energy flow it is necessary to plot not
displacements but the elaastodynamic Poynting vector field. Here is a plot
of the real part of the complex elastodynamic Poynting vector field.

Figure 9. Real part of the complex clastodynamic Poynting vector field for the antisymmetric lowest order
mode. The units are fractions of the Rayleigh-Lamb wavelength.

The real part of the complex Poynting vector field shows a traveling wave
whose energy flows in and around the displacement cells. Figure 10 shows
the imaginary part of the complex Poynting vector field.

Figure 10. Imaginary part of the complex elastodynamic Poynting vector field for the antisymmetric lowest
order mode. The units are fractions of the Rayleigh-Lamb wavelength.

The imaginary part of the complex elastodynamic Poynting vector field is

very similar to that of the real part but with a shift in phase evident by
comparing the Figs. 9 and 10. Figure 10 shows the instantaneous
elastodynamic Poynting vector field.

89

Figure 11. One frame of an animation of the instantaneous elastodynamic Poynting vector field for the
antisymmetric lowest order mode with vectors-as-arrows. The units are fractions of the Rayleigh-Lamb
wavelength.

The animation of the instantaneous Poynting vector field with conventional

vectors-as-arrows shows the energy flowing into and out of the conventional
displacement contour cells, but the overall energy flow is downward and
more strongly so at the interstices of the displacement contour cells.
Figure 12 shows the color coded instantaneous elastodynamic Poynting
vector field.

-3

-2

Figure 12. One frame of an animation of the instantaneous elastodynamic Poynting vector field for the
antisymmetric lowest order mode with direction given by the color and magnitude given by intesity. The units
are fractions of the Rayleigh-Lamb wavelength.

The color coded form of the elastodynamic Poynting vector field for the
antisymmetric Lamb-Rayleigh wave of lowest order is confusing. There is
just too much going on in too many directions at once for the color coded
directions to be of much aid in analyzing the energy flow. Also the animation
is capable of giving the incorrect impression that the energy flow is up the
figure rather than down if the field is sampled at the wrong time step. This is
avoided by the arrowheads in the conventional arrows-as-vectors plot of Fig.
11.

90

5 Acknowledgements
We thank the Deparment of Physics and the Department of Mathematics
and Computer Science for their continuing support of our research efforts.

References
Kaduchak, G., Hughes, D. H. and Marston, P. L. Enhancement of the
backscattering of high-frequency tone bursts by thin spherical shells
associated with a backwards wave: Observations and ray approximation.
J . Acoust. SOC.Am. 96 (1994) pp. 3704-37 14.
2. Kaduchak, G. and Marston, P. L. Traveling-wave decomposition of
surface displacements associated with scattering by a cylindrical shell:
Numerical evaluation displaying guided forward and backward wave
properties. J . Acoust. SOC.Am. 98 (1995) pp. 3501-3507.
3. Auld, B. A. Acoustic Fields and Waves in Solids, Vol. 1 (WileyInterscience, New York, 1973), pp. 142-145.
4. Dean, C . E. and Braselton, J. P., A new geometric interpretation of the
elastodynamic Poynting vector, Acoust Lttrs 20 (1996) pp. 5-8.
5. Mann, A., et al. Instantaneous and time-averaged energy transfer in
acoustic fields, J . Acoust. SOC.Am. 82 (1987) pp. 17-30.
6 . Rayleigh, J. W. S. On waves propagated along the plane surface of an
elastic solid, Proc. Lond. Math. SOC. 17 (1887) pp. 4-1 1.
7. Fetter, A. L. and Walecka, J. D. Theoretical Mechanics of Particles and
Continua (McGraw-Hill, New York, 1980), pp. 477-478.
8. Lamb, H. On Waves in an Elastic Plate. Proc. Roy. SOC. London, Ser. 4
93 (1917) 114-128.
9. Anderson, H., Ed. in Chief, Physics Vade Mecum (AIP, New York, 1981)
p. 60.
1.

91

THE ACOUSTICAL KLEIN-GORDON EQUATION: THE DIRECT AND

INVERSE PROBLEMS
BARBARA J FORBES
Phonologica, PO Box 43925, London, NW2 IDJ, UK. Email: forbesG2phonologica.com
E ROY PIKE
Physics Department, Kings College London, Strand, London, WC2R 2LS, UK
The Klein-Gordon equation for one-dimensional duct acoustics is discussed, and dispersive wavemechanical solutions are defined in terms of geometrical potential functions U(x), that give rise
to geometry-dependent dispersions at rapid variations in tract cross-section. It is noted that such
dispersions are not elucidated by other one-dimensional -- cylindrical or conical -- duct models. The
role of the potential function in duct resonance is examined, and an analysis of the infinite step
potential-function type is made that identifies a frequency-dependent pipe end correction with the
wave-mechanical tunnelling depth. The significance of the finite square barrier is subsequently
discussed. It is shown that the duct area function is in a many-to-one mapping with the potential
function, but that the latter maps uniquely to an acoustical output. Since Sturm-Liouville analysis is
not appropriate for inhomogeneous boundary conditions, the exact solution of the Klein-Gordon
equation is achieved through a Greens-function method that refers to the transfer matrix of an
arbitrary string of bamers or wells. The implications of this result for the inverse problem of noninvasive measurement of ducts are considered

The Acoustical Klein-Gordon Equation

Recent work by the authors [l], [2] has been examining the reduced [3], [4] form of the
Webster equation for applications in duct acoustics. We have noted that the second-order
time derivative invokes an analogy with the Klein-Gordon equation of quantum
mechanics, rather than with the Schrodinger. This acoustical Klein-Gordon equation

in terms of the wavefunction, Y (x, t ) , is parameterised by the potential function, U(x),

which is defined as

Under the simplifying assumption of a plane wave and circular symmetry, the potential
function is found to be directly proportional to the tract curvature. It therefore takes a
positive value in a region of expansion, for example, and a negative value in a
constriction. For piecewise-constant potentials, such that U(x)=U,, these potentialfunction types may be referred to as square barriers and wells, respectively, in analogy

92

with the quantum-mechanical case; and the time-independent eigenfunction solutions,

(x),are in terms of a dispersive wavenumber, k/, where

k =

Jkz-r/b

(3)

for k the free-space wavenumber.

By rearranging Eq. (2) as a second-order homogeneous differential equation, it can
be shown that an infinite equivalence set of area function solutions, of arbitrary radius
and slope at the input, can be chosen to correspond to any given potential function. The
relative resonance characteristics of the infinite set may be identical to those of the single
potential function, and since area-function models --- which assume either a cylindrical or
conical discretisation --- also fail to capture dispersive phenomena, it can be concluded
that the latter forms a more accurate and compact basis for the parameterisation of the
duct geometry.

A Potential Function Account of Pipe Resonance

In contrast to cylindrical-segment models, the dispersion at a change in cross section

gives rise to frequency-dependent reflection and transmission coefficients. These may be
used to examine the effect of a curved baffle on pipe resonance.

2.1. The Step Potential

Figure 1 illustrates the case of a straight duct, with input at x = -1, terminated in an
infinitely-long catenoidal baffle for which U(x) takes a constant value.

-1

Figure 1. The step potential, U(x), and a particular solution for the duct radius function, r(x)

For zero transmission, T(k),the reflection coefficient, R ( k ) , is found to be

k + ik
R(k) =k -ik

93
and the pipe resonances fall at frequencies, F,,,
F, =-(27T

- arctan(y,l)

+ ( n+

X ) ~ ,T

n = 0 , 1 , 2 ...

(5)

where

The pipe end correction, Al, is identified with the penetration depth of the wavefunction
into the step potential and, in the low-frequency limit for k2<< Uo, it can be shown that
A1 = llk.

2.2. The Rectangular Barrier

For the case of a positive potential function of finite width and substantial transmission,
simple expressions for the duct resonances are not forthcoming. However, for k2 << Uo,
the specific input impedance of a square barrier is found, to first order, as
(7)

In the low frequency limit (below around 50 Hz), the term (U0A) can be identified with
the value of 271 due to the infinite baffle. In general, a barrier of some l.0x105 m-2height
and lmm width (Figure 2), has impedance characteristics (Figure 3) comparable to the
infinite baffle. It may therefore be implemented as a barrier equivalent to a radiation
impedance [ 2 ] , dealing elegantly with the problem of nonhomogeneous boundary
conditions for which Sturm-Liouvilleanalysis is not strictly appropriate.

Figure 2. Potential-functionprofile of a duct terminated in an end bamer, U ~ l . 0 x 1 0 m-,

5 A=Imm.

94

Figure 4 gives examples from the equivalence set of area functions corresponding to the
potential function profile of Figure 2. for t 17.5 cm. The tract is terminated in an
infinite conical baffle of varying initial angle.
12

- .,-.

._.

. r -lr

-.

.__ -____

?->\-A

-%-K- - -...____..____......_.....
7.-

2o

15

10

Figure 3. Infinite baffle, input resistance Ro (--.--)and reactance XO(---);compared to end barrier, Ro (-)

and

xo (. . .).
60

50

40
N

,-.
5

0
3
.

ol

XI

10

0
'

10

12

14

16

18

20

Axial Distance I em

Figure 4. Examples from the equivalence set of area functions corresponding to Figure 2, for 1=17.5 cm.

The wave-mechanical approximation to a radiation impedance allows the definition of a

Green's function solution, Gf (x~O~O), for the system output in response to a highimpedance source excitation, 6(0), as

where R(k) and T(k) may be obtained by the transfer matrix method for an arbitrary
sequence of square potentials. The Green's function power spectrum for the configuration
of Figure 2 at 1=17.5 cm is shown in Figure 5. The formants fall at 472, 1422, 2382,
3352, and 4330 Hz, that is, at approximately (94 -96)% of the nominal values, and with a
similar spectral slope of around 6 dB per octave in the low-frequency range.

95

Freq./Hz

Figure 5. Greens function power spectrum corresponding to Figure 2.

Inverse Potential Scattering in Duct Acoustics

It is known from the quantum-mechanical literature that the potential function may be
recovered by inverse scattering methods from knowledge of the reflected wave, R(k), in
Figure 2 , but that the solution cannot be obtained from the transmitted wave, T(k), alone.
Further, the literature assumes a source at infinity so that, in a practical inversion, long
lengths of control tubing would be required to separate the reflected wave from source
secondaries (as currently necessary in acoustic pulse reflectometry [ 5 ] ) . Lastly, practical
applications in speech parameterisation demand a reconstruction algorithm that refers to
the acoustical signal radiated from the vocal tract, rather than to the reflected wave. In
response to the questions motivated by these issues and posed by the authors, Aktosun [6]
has recently shown that the Greens function is equivalent to the slope of the Jost solution,
fi (k, 0), from which R(k) may be found by the formula

where, by the Schwarz integral formula,

Any one of several standard methods may then be used to reconstruct the potential from
R(k). The authors are currently examining these theoretical results for applications in the
non-invasive measurement of acoustical ducts.

96
4

Conclusions

A wave-mechanical theoretical framework has been developed for the forward problem of
plane-wave propagation, that shows increased compactness in comparison to areafunction models. Recent work has shown that, whereas the potential function cannot be
reconstructed from the transmitted wave alone, a solution indeed exists from the radiated
Greens function.
References
1.

2.
3.
4.

5.
6.

B. J. Forbes, A Potential-FunctionAnalysis of Speech Acoustics. PhD Thesis,

Kings College London (2000).
B. J. Forbes, E. R. Pike and D. B. Sharp, J. Acoust. SOC.Am. 114(3), 1291
(2003).
A. H. Benade and E. V. Jansson, Acustica 31, 79 (1974).
D. P. Berners, Acoustics and Signal Processing Techniques for Physical
Modelling of Brass Instruments.PhD Thesis, Stanford University, CA (1999).
B. J. Forbes, D. B. Sharp, J. A. Kemp and A. Li, Acustica 89, 743 (2003).
T. Aktosun, IMA Preprint #1926 (May 2003).

97
ACCURACY AND EFFICIENCY OF THE MULTIPOLE
GALERKIN BEM FOR ACOUSTICS
LOTHAR GAUL, MATTHIAS FISCHER AND UTE GAUGER
Institut A fur Mechanik, Uniuersitat Stuttgart
Pfaffenwaldring 9, 70569 Stuttgart, Germany
E-mail: L. Gaul@mecha.uni-stuttgart.de
The fast multipole method is employed to accelerate the Galerkin boundary element method for acoustics. Important parameters as the nearfield size and the
expansion length are discussed and numerical examples are provided. An approximate inverse preconditioner and a preconditioner based on spectral properties of
the boundary integral operators are presented. Both approaches prove suitable to
reduce the iteration count and to improve the efficiency of the method.

1 Indroduction

The application of the Boundary Element Method (BEM) for structural acoustics
suffers from the high computational cost. The arising matrices are dense, thus, requiring O ( N 2 )memory. This cost cannot be handled, even for a moderate number
of unknowns, by standard implementations. The BEM can be accelerated by employing the Fast Multipole Method (FMM) in a multilevel scheme. The key idea is
to form the matrix-vector product of the BEM by approximating the fundamental
solution at some distance from the source point by a series expansion. The resulting complexity of the algorithm is O ( N log2 N ) , and thus, for large-scale problems
much faster than the traditional BEM. The major areas of application for fast BEM
are simulations on fine discretisations and high frequencies. However, both factors
deteriorate the conditioning of the system of equations and the iterative solution
of the ill-conditioned systems demands a high iteration count. Since the FMM
algorithm must be applied in each iteration, the efficiency of the multipole BEM
quickly worsens and the application of preconditioners is mandatory.
In the presented paper, the multipole scheme for Galerkin BEM is outlined.
Considering a numerical example, the proper choice of the nearfield size and the
expansion length is discussed. Both parameters turn out t o be critical factors
for the accuracy and efficiency of the method. Two different preconditioners are
presented. First, an approximate inverse type of preconditioner which reduces the
iteration count at a negligible cost for the application. Second, a preconditioner
based on the spectral properties of the boundary integral operator which guarantees
a condition number independent of the mesh size at a numerical cost equal to a
matrix-vector product of the multipole BEM.
2

Galerkin BEM for Acoustics

Weighting the Helmholtz equation for the acoustic pressure u with the fundamental
solution u* and applying Greens second theorem yields the representation formula

98
The exterior domain Re is the complement to the open set R

c R3 with boundary

r = 80. The unit normal n to I' is defined to point outwards from R.

A boundary integral equation is obtained from Eq. (1) by taking the limit on the smooth
boundary, Re 3 y + r

where (Vau/an)(y) and (Ku)(y) are the single and double layer potential, respectively. For more details on the derivation of boundary element methods, it is
referred to Gaul et a1.l.
A boundary integral formulation for the exterior Neumann problem is obtained
by taking the normal derivative of Eq. (1) on the smooth boundary, Re 3 y -+ I'

Using the boundary condition au(z)/an= q (z ) on

tion is written as

r, the boundary integral equa-

-(D~)(Y)

(K'q)(y)

defining the hypersingular operator (Du)(y) and the adjoint double layer potential
(K'q)(y). Triangularisation of the boundary I' yields an approximate solution uh
on the boundary element space of piecewise linear basis functions (pi. The given
Neumann data is interpolated on the boundary triangularisation with constant basis
functions $k.
N

Testing with isoparametric test functions (pj(y) yields the linear of system of equations

for j = 1.. N , which can be written in matrix notation

99

The hypersingular operator exhibits a strong singularity (1/1z - yI3) and is thus
not an integrable function. Applying Stoke's theorem, it can be reduced to
~ h [ .i1=
i

//

r r

k2nz . 72, vi(-)vj(y)u*(-,

Y) ds, ds,
((nz x

VzCpi(Z)). y.(

x V,Cpj(Y)) u*(-,
Y) dsz ds,

(8)

curlr ( v j )

where curlr(9j) is constant on linear elements, thus, requiring the evaluation of the
single layer potential for linear and constant shape functions.

Multipole Expansion

The reduction of the numerical cost in the multipole BEM is based on the series
expansion of the fundamental solution and a multilevel scheme. The multipole
method was originally developed by Greengard and Rokhlin' for the fast simulation
of large particle fields in physics. The theory of the multipole expansion for acoustics
and its diagonal translation operators was developed as well by Rokhlin3. For the
Helmholtz equation in 3D the multipole expansion reads

9(.)
denotes

the Legendre polynomials and j , (.) the spherical Bessel functions.

Normalised vectors are indicated by (.) = (.)/I . 1. Using the orthonomality of the
Legendre polynomials on the unit sphere S2 and the expansion of spherical waves

one obtains a diagonal form of the multipole expansion

The summation over 1 in Eq. (10) must be truncated, since the Hankel function
diverges for large 1. For the truncated series, summation and integration can be
interchanged and one defines the translation operators
L

D)= C(21+
l)i'h;')(kD)fi(s.

ML(~,

fi) ,

(11)

1=0

which solely depend on the distance vector D and allow the implementation of an
efficient multilevel scheme.
The boundary elements of the triangularisation are assigned to a cluster tree
with levels C = 0 .. m. The cluster on the highest level C = 0 represents a parallelepiped containing all boundary elements. Subsequent child clusters are constructed by bisection of the parent cluster by bisection. The clusters are denoted

100

by Cz where y = 1.. 2e. Two clusters on the same level are in each others nearfield
when the distance D between their centres 22 fulfills the condition

de
D<C-,
2
where de is the diameter of the clusters at level C and c is a suitable constant.
Clusters, whose father clusters fulfill the nearfield condition, but themselves are
not in each others nearfield, form the interaction list.
Using these definitions, the multipole algortihm is formulated for the matrixvector product vl = V h [ l ,k ] W k of the single layer potential evaluated for constant
shape functions. The nearfield part is evaluated directly, wheras for the farfield
part, the multipole method is applied. The double surface integral is substituted
by Gauss quadrature on the elements.

nearfield
j=1
farfield
i=l
where A k , w k , i and x k , i are the Jacobi determinant, Gauss weight and integration
point for element '&, respectively. The terms are defined analogous for the element
71. The expression

can be evaluated efficiently using the multipole method, wheras the remaining operations in Eq. (13) are local at 71. The multipole algorithm consists of the following
steps:
Compute the farfield signature FA(s) on the lowest level C = m for all clusters.
Gk
Tk E

c&

i=l

Translate F,Y(s) to the interaction list using the translation operators M L

interaction list
where D is the distance between z; and the centre of the respective interaction
list cluster.
Shift F;(s) to the centre of the father cluster.
Repeat last two steps upwards until interaction list is empty.
In the downward pass, shift nearfield signature N z ( s ) in interaction lists t o
the child clusters.

101

On the lowest level, recover solution in integration points

yl,j

Finally, the matrix-vector product Eq. (13) is evaluated by summing up the element contributions @(yl,j) and adding the nearfield contribution which is calculated
directly using the standard BEM.
The expansion length L plays a crucial role for the accuracy and the efficiency
of the multipole method. Coifman et al.4 give the semi-empirical fit
Le(kde) = kde

+ pln(kde +

T)

(18)

where the expansion length depends on the wave number k and the cluster diameter
de. The parameter p does not translate directly to the number of accurate digits,
since the error in the truncated multipole expansion does not only depend on the
cluster size but also on the distance between two clusters where the translation
operator is applied. A detailed discussion of the truncation error is given in the
paper by Koc et al.5.
A consequence of the choice Eq. (18) is that the expansion length changes from
level to level and, thus, efficient algorithms for interpolation and filtering are needed
when shifting the expansion to the father and child cluster, respectively. A suitable
scheme based on fast Fourier transforms is proposed by Gyure and Stalzer6.
For the complete algorithm the computational cost can be estimated as
C?(Nlog2 N ) for computations at kh = const what corresponds to the engineering practice of a constant number of elements per wavelength.
4

Preconditioning

For the solution of the system of equations, the General Minimal Residual method
(GMRES) is applied. It proved to perform well for BEM systems in acoustics which
typically are not positive definite. However, convergence decreases significantly
when using restarts, thus, it is required to store all iterates. Using preconditioners
the iteration count can be greatly reduced. This is crucial for the efficiency of
the proposed method concerning memory and computing time, since the multipole
algorithm must be executed at each iteration.
For the preconditioned solution of the system Ax = b, it is replaced by

P - ~ A x= P-'b

(19)

with the preconditioner matrix P . An efficient preconditioner should yield a condition number ts(P-lA) that is constant, i.e. independent of mesh size and wave
number. A second requirement is that the application of P-l can be evaluated
with a numerical cost of O ( N log2 N ) .
4.1

Approximate Inverse Preconditioner

The basic idea of the first preconditioner proposed in the present paper is operator
splitting A = A0 A, where A0 is bounded. It can be shown that the eigenvalues

102

of Ao'A = I Ai1A cluster around 1 which yields an improved convergence of the

iterative solver7.
The application of the operator splitting on the boundary integral operators is
straight forward. For the hypersingular operator, the boundary integral is splitted

-loa

2 *

(Du)(y)=

(x7
an,dny

u(x) ds, -

a2u*(x1y)u(x) ds,

dn,dny

(20)

where I' = ro U f', defining A0 and 2, respectively. The choice of the size of
is
a trade-off between effect and numerical cost of the preconditioner. A two element
layer around the element containing the load point turned out to yield good results
and was employed throughout the numerical examples in Sec. 5.
The operator A0 is sparse, however, the exact application of A;' would be too
expensive. Thus, for the proposed preconditioner, an approximate inverse of Ao
is computed. Instead of finding the minimum of 11I - A0 P-lll, a column oriented
algorithm minimizes I lej - Ao(P-l)jlI, conserving the sparse pattern for P-l.

4.2 Preconditioner based on Spectral Properties

A combination of the boundary integral equation Eq. (2) and the hypersingular
boundary integral equation Eq. (3), yields the Calderon projector. Squaring the
Calderon projector, the relations
1

V D = - I - K 2 and DV = ' I

- (K')2
(21)
4
4
between the boundary integral operators can be derived, which yield the spectral equivalence between the single layer potential and the hypersingular operator.
Steinbach and Wendland' show that

is an efficient preconditioner for the hypersingular operator. For the application of

the preconditioner, two evaluations of the inverse mass matrix M and one evaluation of the single layer potential are required. From there, the numerical cost of
the preconditioned matrix-vector product is about twice the cost of the standard
matrix-vector product.
5

Numerical Results

As a numerical example the sound radiation into air (c = 340 m/s) from an Lshaped domain as depicted in Fig. 1 is considered. The overall dimensions of the
L are 1m x 1m x 1m with an excavation of 0.5 m x 0.5 m x 1m. The particle
velocity on the L-shape due to a monopole source inside the domain is applied as
Neumann boundary condition. The results of the BEM simulations are compared
to the analytical pressure of the monopole source on the boundary of the L-shape.
The Dirichlet error e: is defined as

e2r =

I 1,analytical

- ,BEMI

l2

~j~analyticaljj~ '

(23)

103

Figure 1. Error of matrix-vector product on L-shape.

First the convergence of the multipole expansion is discussed. In Fig. 1, the

error of the matrix-vector product for the hypersingular operator is plotted against
the factor p of Eq. (18), determining the expansion length. The convergence depends significantly on the nearfield size, i.e., the parameter c in Eq. (12). For small
nearfields, the error decreases slowly when increasing the expansion length. The
convergence gets clearly better when the nearfield size is increased. For large expansion length, the series expansion i s not stable, since the Hankel functions in Eq. (10)
diverge. For engineering practice, this limitation doe not play a role, because even
can be achieved.
for very small nearfields, errors of less than
The combination of nearfield size and expansion length defines the accuracy of
the multipole algorithm. Since both parameters increase memory requirements and
computing time, finding an optimal combination is cruical for the efficiency of the
method. In Fig. 2, the total computing time for the acoustic simulation on the Lshape with a discretisation of 2816 elements at a wave number K = 5 rn-l is plotted.
The parameters were chosen t o yield an error of the matrix-vector product of less
than
in Fig. 1. Best results, i.e. the shortest simulation time, were obtained
for a nearfield factor c = 4 a t an expansion length of p = 4. For comparison, the
computing time for a standard BEM simulation is plotted. Even for this small
example problem, the multipole BEM performs clearly more efficiently.
For a series of simulations at kh = const the computing time is presented in
Tab. 1. The number of elements ranges from 704 t o 45056 at a wave number
between 5 m-l and 40 m-l, which corresponds to approximately 10 elements per
wavelength. One notices that the time for the nearfield computation increases
linearly with the number of unknowns, whereas the time per iteration, where the
multipole algorithm is applied, increases slightly faster than linear. In Fig. 3, the
time per iteration is compared t o N log2 N, verifying the predicted complexity.
The results in Tab. 1 were obtained for preconditioning with the single layer
potential. For the investigation of the efficiency of the proposed preconditioners,
iteration counts for simulations without preconditioning, with the approximate inverse preconditioning and with the single layer potential preconditioning are compared in Fig. 4. In the first plot, results are shown for simulations a t a constant

104

BEM

't 6 1 3
.-5
v)

513

3
.
m
.-

414

100

50

250

200

150

computational time I sec

Figure 2. Time for BEM simulation, error of matrix-vector product

elements
704
2816
11264
45056

h
0.125
0.063
0.031
0.016

m
m
m
m

tnear

iterations

5 m-'
10 m-'
20 m-l
40 m-l

16.7s
60.2 s
226 s
1012 s

21
66
163
452

< 0.001.

1 iteration
0.4 s
2.8 s
16 s
90 s

e:
0.09
0.04
0.03
0.08

kh = 0.625= const

10-21

1 o2

1o3

elements

1o4

1o5

Figure 3. Computing time per iteration for kh = const.

wave number k = 5 m-l. Without preconditioning, the iteration count increases

significantly with mesh refinement. Using the approximate inverse preconditioner,
the increase is reduced, whereas the sigle layer potential yields a convergence of
the solver that is independent of the discretisation. The second plot shows the behaviour a t lch = const. Here, all preconditioners show an increase of the iteration
count with the frequency. The single layer potential preconditioner performs best,
but it must be noticed, that the numerical cost of one iteration is about twice as

105
80-

+ no preconditioner
* approximate inverse

300-

kh = 0.625 = const

e
40

302o

8*

16

32
h- I m-l

16

32
h- I m-l

Figure 4. Preconditioned number of iterations for k = const and k h = const.

high as of a standard matrix-vector product. The cost of the approximate inverse

preconditioner is negligible.

Conclusions

In the presented paper, the fast multipole Galerkin BEM for acoustics is presented.
The expansion length and the nearfield size are the most critical parameters for the
accuracy and efficiency of the proposed method. The influence of both parameters
is discussed on the example of sound radiation from an L-shaped domain. For an
efficient application of the fast multipole algorithm, preconditioning of the system
of equations is mandatory. Based on the spectral properties of the hypersingular
operator, the single layer potential proved to efficiently reduce the required number
of iterations.

Acknowledgments
The financial support by the Deutsche Forschungsgemeinschaft DFG in the framework of the collaborative research centre 404 Multifield Problems in Solid and
Fluid Mechanics is gratefully acknowledged.

References
1. L. Gaul, M. Kogl, and M. Wagner. Boundary Element Methods for Engineers
and Scientists. Springer-Verlag, Heidelberg, 2003.
2. L. Greengard and V. Rokhlin. A fast algorithm for particle simulations. Journal of Computational Physics, 73:325-348, 1987.
3. V. Rokhlin. Diagonal forms of translation operators for the Helmholtz equation
in three dimensions. Applied and Computational Harmonic Analysis, 1:82-93,
1993.
4. R. Coifman, V. Rokhlin, and S. Wandzura. The fast multipole method for the

106

5.

6.
7.

8.

wave equation: A pedestrian prescription. I E E E Antennas and Propagation

Magazine, 35:7-12, 1993.
S. Koc, J.M. Song, and W.C. Chew. Error analysis for the numerical evaluation
of the diagonal forms of the scalar spherical additon theorem. S I A M Journal
of Numerical Anlgsis, 36(3):906-921, 1999.
M.F. Gyure and M.A. Stalzer. A prescription for the multilevel Helmholtz
FMM. I E E E Computational Science & Engineering, 5(3):39-47, 1998.
K. Chen and P. Harris. Efficient preconditioners for iterative solution of the
boundary element equations for the three-dimensional Helmholtz equation.
Applied Numerical Mathematics, 36:475-489, 2001.
0. Steinbach and W.L. Wendland. The construction of some efficient preconditioners in the boundary element method. Advances in Computational
Mathematics, 9:191-216, 1998.

107

Bottom reflection phase shift parameter inversion from

reverberation and propagation data
H. L. Ge, H. F. Zhao, X. Y. Gong, E. C. Shang
(State Key Lab. of Oceanic Acoustics, Hangzhou, 310012)
Abstract: An inversion method of bottom reflection phase shift parameter is
described in this paper. First, the reverberation data CSDM combined with
mode-shooting method is used to extract the modal wave number. The result is also
obtained by use of propagation data with similar method. Then, the bottom
reflection phase shift can be estimated by formulation based on the WKB modal
condition.

I. Introduction
The geoacoustical properties of the sea bottom are important parameters for
propagation modeling and model-based underwater signal processing in shallow
water. Since 90's last century, many works have been done on inversion of
geoacoustical properties and matched field process are used in most of the works to
estimate the geoacoustical parameters of sea bottom ",21. In appendix, we will shown
from an example that the inversion results may be non-unique when we try to
estimate bottom density and velocity with matched field processing.
In [ 3 ] , E. C . Shang and his cooperators proposed that the inversion of bottom
inversion may be simplified to reconstruction of the complex reflection coefficient

R ( 8 ) instead of the geoacoustic parameters for the acoustic field prediction purpose.
In their method, the real and the imaginary part of the modal wavenumber are used to
calculate the phase and the magnitude of R ( B ) , respectively.
In this paper, we concentrate on the estimation of the phase of R ( 8 ) and how
to estimate accurately the real part of modal wavenumber. The modal depth functions
are extracted from the CSDM of measured data of a VLA. Then the real part of modal
wavenumber can be estimated by shooting method. The method of extraction of
modal depth functions and estimation of modal wavenumber has been discussed in
[4,5], where the CSDM is constructed from the ambient noise or propagation data of a
towed source. In this paper, we try to use reverberation data or propagation data of a
WBS source instead of the towed source data or ambient noise data.
In section 11, the relation between the phase of R ( 8 ) and real part of modal

108

wavenumber and the method of estimation of modal wavenumber from VLA data are
introduced. Section 111 gives the results obtained by used of the ASIAEX data.
Section IV is the summary and discussion.

I1

Theory of reconstruction of the phase of R ( 6 ) by using of

reverberation or propagation VLA data

A. Relation between the phase of R ( 6 ) and modal wavenumber
The WKB modal c~ndition'~,']for
the m-th normal mode in water:

2 7 , / m d z + B,

+ Bl = 2 m x

Z"

Where,

turning depth

, z,=

0
turning depth

and

The H in eq. (2) and (3) is the depth of the sea. For the modes which modal
wavenumbers are smaller than K ( H ) = o l e ,

which is wavenumber in water at

depth H, the bottom reflection phase (the phase of R ( 6 ) ) can be calculated by:
2,

Qb

= B, = 2 m n -

B, - 2 f J k m d . z

(4)

Z"

It's shown from equation (4) that the bottom reflection phase can be retrieved
if the real part of the modal wavenumber and the SSP in water column are known. In
the following three sections, we discuss how to estimate k,.

B. Extraction mode depth from CSDM of VLA reverberation data

The reverberation field excited by a point source

( 2 , ,0) can

be written as'']:

109

where z, is the depth of the source, Sm,is a scattering kernel relating the incident
mode and the scattered mode, r, is the range of the center of the insonified ring and

q is the random parameter describing the distribution of the medium fluctuation.

If there is an VLA deployed at r=O and the depth of the elements are
z1,z2,. . .,zN , and the range of the center of the insonified ring changes from

rcl to

rcL,then the vector of received reverberation field is:

-

p=YK
where,

and

K=

k,,
>

>

W C 2

K(r,,

I)

(8)

The CSDM of the received field is

--

R = pp' = YKK"Y'

(9)

The SVD of CSDM R is

R= USV
(10)
Just like the discussion in [5], if KK' is a diagonal matrix and the array
spans the whole depth, then equation (9) and equation (10) is equal and
Y=U=V
(1 1)
We will not verify the condition above in this paper and detailed analysis will
be done later. Here we just assume the conditions are satisfied. From the discussion in

110

[5], we assume that the conditions can be also satisfied when we use one r, and
multiple frequencies and the modal depth functions can also be extracted.

C. Extraction mode depth from CSDM of VLA propagation data

The principle of extraction mode depth from CSDM of VLA propagation data is
described in [4] and [5]. If the multiple result of matrix consist of

/fi

elknrJ

with its complex conjugation is approximately unit matrix, then the modal function
can be extracted from CSDM of propagation data just as reverberation data.

D. Estimate the modal wavenumber by shooting

After we extracted the modal depth functions, we can use the SSP and finite
difference equation over the depth to shooting the modal functions and find the modal
wavenumber by seek a global minimum of the function'41

W,)= Crk el)- y, el>I2

(12)

111 Analysis of the ASIAEX data

In this section, we will present results
of processing reverberation data received by
HARRI's 60-element-VLA on Jun 3, 2001.
We used 57 elements of the array and the
spacing of the element used is 1.0m. The
array spanned from 4m to 60m and the sea
depth is 105m. The reverberation data we
used in this paper is recorded at 9:14-9:20,
"
"
"
"
'
liD
Jun 3. The source is lkg WBS. We use the
B
1518 1519 1523 1521 1522 1523 1524 1 5 Z 1526 I527
sound speed
nearest CTD results for shooting and the
SSP is shown in figure
1.
Figure 1 The SSP used in this paper
First, we use multiple center range
data to extract the modal depth functions.
The reverberation data from 6s-10s are divided into 25 0.2s-length segments and
calculate the frequency complex spectrum by FFT. The results at lOOHz are show in
figure 2. Shooting method are used to estimate modal wavenumber and the results are
tried to satisfied the same bottom condition. If the extracted modal depth function that
can not satis@ the same bottom condition then it is eliminated. The estimated modal
wavenumbers are listed in table 1 .
(nY,)

111
0

20

20

1
0.9

0.8
40

-5

40

0.7

0.6

5 60

60

a
P
P)

0.5
0.4

80

80

100

100

0.3
0.2
0.1

Index

* extracted modal depth functions

- shooting results
Figure 2. Extracted modal depth functions at lOOHz and the scalar singular
values. The frequency is lOOHz and multiple c , data are used

10
Index

* extracted modal depth functions

- shooting results
Figure 3. Extracted modal depth functions at 200Hz and the scalar singular
values. Multiple frequency data are used.

112

As a second example, we use multiple frequency data to extract the modal

depth functions and 1s-length data are used. Then we used the complex data from 181
to 220Hz (the central frequency is 200Hz). The extracted modal depth functions and
shooting results are shown in figure 3. The estimated modal wavenumbers are listed
in table 1.
An extraction result of mode depth function from propagation data is given in
Fig.4(a). The data is collected at 11:30 by a 14-element VLA which span from 34.6m
to 86.6m. The SSP is shown in Fig.4(b) and the source is a 38g WBS which is about
30km from the VLA. The estimated modal wavenumber is also listed in tablel. It's
shown from table 1 that the estimated wavenumber is approximately as that extracted
from reverberation data.
mode deoth funclian
1

09.
OB07.
06-

2 0504.

03.

'

'291

0.5

0
6

0.5

* extracted modal depth functions

- shooting results
(a)
Figure 4. Extracted modal depth functions at 200H2, the scalar singular
values ( a ) and the sound speed profile (b). Multiple frequency data are used.
Then we used the estimated modal wavenumbers to reconstruct the phase of
complex reflection coefficient. For small grazing angle, phase shift has approximately
a linear relation with grazing angle and we calculated the parameter P by use of
following equationI7I

Where a, is grazing angle which is given by

113

The calculated parameters are also listed in table 1. We also list P calculated by
use of Hamilton's typical sediment parameters in table 2. It's show from the table 1
that the calculated P 's are consistent. From table 2, it can be seen the bottom is most
likely type 6 .
Table 1. Estimated km and ( D b ' P a , ,a, ,P

less than P extracted from reverberation data. This error may be caused by the error of
the depth of the element. We have found the error of is about 2 if the error of depth of
the element is 1. The relation between the accuracy of the estimated P and the error of
the parameters that include location of elements, sea depth and SSP will be studied
later.
Table 2 Hamilt

IV Error analysis
In [3], the estimated error of (D, caused by the error of k , has been discussed,
it is given by

114

Amb( m ) k S , Akm
where

s, =

2n
km - kmcl

The numerical value of S , calculated by KRAKE"']

is

S , = 7x10-*
The accuracy of k , given by the shooting process is

Akm 1x ~ O - ~
Based on the above estimation, the error bar of our estimation can be concluded
as follows

AP

- f2.0

(20)

V Summary and discussion

In this paper we have presented a method for retrieving the phase of complex
reflection coefficient by use of the reverberation and propagation recorded by VLAs.
First, we extract modal depth function by use of reverberation and propagation data.
Then we estimate modal wavenumber by shooting method. Finally, we use estimated
wavenumber to calculate the phase of complex reflection coefficient. Stable estimated
P values given in Table 1 show that the sediment type corresponding to the 'type 6' of
the Hamilton's classification (See table 2).

APPENDIX: Multi-solution possibility for inversion of bottom

parameters by using matched field process
In the matched field inversion problems, people often try to find the density,
sound speed, attenuation and their profile of sediment. The rule often used for
inversion is find the maximum of the correlation coefficient of the copy field
with measured field p . However, the maximum of the ambiguity figure of the
correlation coefficient is not unique. As a example, we calculate the simulated
measured sound pressure of a 103m length VLA and the spacing of the array is lm.

115

In the calculation, we assume the depth of sea is 105m and a half infinity sea bottom
which density and sound speed are 1.7g/cm3 and 1623ds. The sound speed profile in
the water column used is just same like figure 1. Then we change the density, sound
speed of the sea bottom and calculate the copy field. Figure 5 is the ambiguity figure
of the correlation coefficient of the copy field with measured field. Its shown that the
maximum of the correlation coefficient is on a line and we can not determine the
density and sound speed simultaneously.

95

9
1

.-

85

I
n
c

-0
m

s1

g
a

75

65
1..

1550

1600

1650

1700

Bottom sound speed (m/s)

Figure 5 The ambiguity figure of the correlation coefficient of copy fields with a
simulated measured field

Reference
1. A. Tolstoy, MatchedJield processing for underwater acoustics (World Scientific,
Singapore, 1993)
2. P. A. Baxley , Newel1 0. Booth and W. S. Hodgkiss, Matched- field replica model
optimization and bottom property inversion in shallow water, JASA I07(3),

2000,130 1-1323
3. E. C. Shang, Y. Y. Wang and Shaobo Sun, Inversion in shallow-water using the
WKB modal condition, IEEE J. Ocean. Eng. 2 1(4), 1996,432-439
4. P. Hursky, W. S. Hodgkiss and W. A. Kuperman, Matched field processing with
data-derived modes, JASA 109(4), 2001,1355- 1366
5. T. B. Neilsen and E. K. Westwood, Extraction of acoustic normal mode depth
functions using vertical line array data, JASA 1 1 1(2), 2002,748-756
6. E. C. Shang, T. F. Gao and D. J. Tang, Extraction of modal back-scattering matrix

116

from reverberation data in shallow-water waveguide. Part I-Theory

7. T. C. Wang and E. C. Shang, Underwater Acoustics, Beijing, P.R.O.C.:Science
press, 1981

8. M.Port, The KRAKEN normal mode program, SACLANT, memo SM-245, 1991

117

ACOUSTIC BEHAVIOR OF ELASTIC SCREENS IN OPEN AND CONFINED

SPACES
L. GODINHO AND A. TADEU
Department of Civil Engineering, University of Coimbra, 3030-290 Coimbra, Portugal
E-mail: lgodinho @dec.uc.pt
The Boundary Element Method (BEM) is used to study the sound wave propagation around elastic
and rigid bamer placed over an elastic ground surface and inside a confined space. The presence of
an elastic or rigid ground is simulated using adequate Greens functions, which take into account the
required boundary conditions at the interfaces between either the air, the barrier and the elastic
ground. Using this scheme, the BEM discretization is only required for the elastic barrier. The
method used allows the full interactions occumng at the interfaces between the different media to be
taken into account.
Different simulations are performed assuming the presence of lightweight elastic screens placed over
an elastic or a rigid ground, or inside a confined space. The elastic screen modeled is assumed to
provide the shield for sound generated by spatially sinusoidal harmonic line pressure sources. An air
gap is allowed under the acoustic screen. Results from different simulations are presented and
compared with those obtained for a rigid barrier placed over a rigid ground.

Introduction

Many numerical schemes have been applied to the study of sound scattering around
acoustic screens. Among these, the Boundary Element Method (BEM) has proven to be a
good tool for the simulation of sound wave propagation infinite or semi-infinite media,
since it automatically satisfies the far field radiation conditions. This allows the required
medium discretization to be restricted to the material discontinuities.
In earlier works by Duhamel [l], the BEM was applied to analyze the 3D sound
pressure around acoustic rigid barriers with a 2D geometry placed over a rigid ground. In
that work, the 3D response was computed as a discrete summation of 2D solutions. Later
works have presented algorithms that allow the simulation of absorbing surfaces.
Examples are the works by Duhamel and Sergent [2] and Lacerda et a1 [4].
When analyzing the barrier effect in the protection of sensible buildings, it is also
important to take into account the presence of the surfaces of the building faqade. This
issue has been analyzed in a previous work by the authors of the present paper, computing
the sound pressure level in the vicinity of tall buildings when a 3D pressure source is
excited in the presence of 2D acoustic barriers [3]. In that work, a BEM model was used
to discretize the barrier, while Greens functions based on the method of images to take
into account for the presence of the tall building.
In most of the previous approaches, the numerical models are built assuming that the
acoustic barriers are rigid bodies, either with a perfectly reflecting surface or with
absorbing properties. In some cases, it is important to model the barrier as an elastic body,
taking into account not only the energy that travels around the barrier but also the energy
that travels through it. In the present work, the behavior of movable elastic screens with a
loose connection to the ground, simulating an air gap between the barrier and the ground,
is analyzed. Two different cases are analyzed, corresponding to barriers placed over an
elastic ground or placed inside a space confined by two rigid flat surfaces. The use of
Greens functions that take into account the required boundary conditions at the interface

118
between an elastic ground and the fluid medium or between the rigid ground and ceiling
and the fluid avoids the full discretization of the horizontal surfaces. Thus, only the
surfaces limiting the acoustic screen need to be discretized with boundary elements. These
Green's functions are analytical solutions for the steady state response of a homogeneous
elastic semi-infinite space bounded by a fluid medium or of a fluid layer limited by two
elastic semi-infinite spaces, subjected to a spatially sinusoidal harmonic line load. These
solutions are defined following the process described in [5]and [6].
The remaining part of this article is organized as follows: first, the BEM formulation
is presented, including a brief description of the required Green's functions; then, the
proposed BEM model is applied in the simulation of sound wave propagation around
movable lightweight elastic screens.

BEM Formulation

Figure 1 illustrates the geometry of the model, which assumes the existence of a
submerged elastic inclusion inside a homogeneous fluid bounded either by a flat elastic
ground formation or by two rigid horizontal flat surfaces. The material of the ground
(density p,, shear wave velocity p,, compressional wave velocity a , ) and of the
inclusion (density p 2 , shear wave velocity p, , compressional wave velocity a2) may
differ. The host fluid, with density p f ,allows a compressional wave velocity af. The
system is excited by a spatially sinusoidal harmonic pressure line load along the z
direction, oscillating with a frequency w , located in the fluid at position (xu,y o ) . The
corresponding incident pressure field can be expressed as
ofu"

( m , x , y , k z ) =-iA
-2H t )

(k

a,

, / ( x - ~ ~ ) z + ( y y ~ ) ~ )

-4

with A being the wave amplitude, k , =

the axial wavenumber in the
order n.

z direction and H?)(

with Im ka, 5 0 , i =

J-1 , k,

is

...) are second Hankel functions of

Figure 1. Schematic representation of the BEM models: a) elastic object over an elastic ground; b) elastic
object in a confined space.

119
To establish the BEM equation system, it is necessary to evaluate the following
integrals, defined along the boundary of the elastic inclusion,
(i, j = 1,2,3) H::)" = I ~ ~ i ) ( u , x , , x ~ , n , ) d ~ ,
fit),,= ~H~)(u,x,,x,,n,)dc/
c/

c/

Gt)" = G i s ) ( x~,,, x,)dC,

(i=1,2,3;j=1) G:f)"=IG:i)(~,x,,x,)dC,
c/
c/
where nl is the unit outward normal for the lrh boundary segment C, ; the subscripts
i, j = 1,2,3 denote the normal, tangential and z directions, respectively. In these

x,, x,) are respectively the Green's tensor for

equations, H F ' ( u , x,, x,,n, ) and GF'(u,
traction and displacement components in the elastic medium of the inclusion, at point xI
in direction j caused by a concentrated load acting at the source point x, in direction i .
The required two-and-a-half dimensional fundamental solution (Green's functions) and
stress functions in Cartesian co-ordinates, are those for an unbounded elastic medium [7].
Hif'(w,x,,x,,n,)
are the components of the Green's tensor for pressure in the fluid
are the
medium, at point x, caused by a pressure load placed at point x, ; G~f)(w,x,,xI)
displacements in the fluid medium, at point xI in the normal direction, caused by a
pressure load acting at point x, . The BEM formulation used in this work uses Green's
functions that take into account the presence of the horizontal flat surfaces of the ground
or ceiling. These functions are obtained as a summation of the incident field and the field
originated at those surfaces, expressed by the surface terms, which must verify the
required boundary conditions. At the interface between a solid and a fluid medium, these
conditions are the continuity of normal displacements and stresses, and null tangential
stresses at the interface. In the present BEM formulation, the incident field is the result of
the sum of the source terms (2.5D full-space) and of the surface terms described above.

Analysis of the behavior of elastic screens

The described algorithm was applied in the calculation of the insertion loss provided
by movable lightweight elastic screens subjected to spatially sinusoidal harmonic line
pressure sources. The presence of an air gap between the acoustic screen and the ground
floor is assumed. As a reference for comparison, the results obtained with a rigid barrier
[3] are also presented. To allow a better understanding of the main features of the
responses, the results are also compared with those obtained for an infinite elastic panel
model bounded by two fluid media. These solutions were previously obtained analytically
and can be found in Tadeu et al. [ 5 ] .
The average insertion loss provided by the elastic acoustic barriers has been
computed along a grid of receivers placed on the side of the acoustic screen that does not
contain the source, as in Figures 2a and 2b. For this purpose, the computations are
performed both with and without the presence of the elastic screen, so that the sound level
reduction provided by the barrier can be assessed.

120
3.1

Elastic screen over an elastic ground

In the presented examples, the floor is assumed to be made of concrete, with density
p, = 2 250 kg/m3, and allowing propagation velocities for the S and P wave of

p, = 1415.8 d

s and a, = 2629.8 d s . The elastic screen is assumed to be made of cork

( p2 = 180 kg/m3,

p2 = 204.1 d

s and

a; = 288.7 d s ), with a thickness is 0.15 m . The

host acoustic medium allows a pressure wave velocity of af= 340 m/s and has a density
of pf = 1.22kg/m3. An acoustic source is placed 0.6 m above the ground and 2.75 m
from an acoustic screen of height h , with an air gap of thickness w .
In Figure 3a, the average insertion loss provided by the elastic and rigid screens when
the air gap below the barrier is set at w = 0.05 m is presented. It can be seen that the
average insertion loss given by the elastic screen is in generally lower than that obtained
using the rigid screen. This is particularly true at low frequencies. Figure 3b shows the
results computed when the air gap that separates the screen from the ground is reduced to
w = 0.02 m . As before, the insertion loss for the rigid screen is, in general, higher than
that given by the elastic screen. However, when these results are compared with those
shown in Figure 3a, there is a visible increase in the global insertion loss computed for
both the models. So, the size of the air gap near the floor seem to be a determinant factor
in the final insertion loss provided by the screen. In both examples, the insertion loss
curve of the elastic screen exhibits a set of dips related to the interaction of waves within
the panel. Some of these dips are clearly visible. To better understand this behavior, the
results are compared with the average insertion loss provided by an infinite cork panel
with the same thickness as the acoustic screen, using the model of Figure 3c. The
corresponding insertion loss is displayed in Figure 3d, where a set of dips in insulation can
be seen, occurring at frequencies related to the resonance effect within the elastic barrier
and to the coincidence effect that occurs for waves traveling along the panel. When the
results provided by this model are compared with those obtained from the elastic screen
model, it is observed that the frequency positions of the bigger dips coincide, and this is
especially noticeable for those related to the resonance frequencies.

. . . . . . . .... .. .... . .. ... .

. . -. . . ........... . :.::.....,..
*
* .
. . . . . . . .
<

4 + + +

+I6

,:

$1 + ++ + ++ +hI
0

...
a)

b)

Figure 2. Studied models: a) barrier over an elastic ground; b) bamer inside a confined space.

3.2

Elastic screen inside a confined space

In a second application, consider an elastic barrier located inside a fluid space confined by
an elastic flat ground a by an elastic flat ceiling. Again, the elastic screen is assumed to be

121

made of cork, with a thickness is 0.15 m and defined by h + w = 3.0 m , with w = 0.05 m ,
while the ground and ceiling are assumed to be made of concrete.
Figure 4 displays the average insertion loss provided by rigid and elastic acoustic
screens located inside confined spaces with heights 4.0m (Figure 4a) and 6.0m (Figure
4b). As a reference for comparison, the insertion loss curves obtained for the elastic
barrier over an elastic ground without modeling the ceiling are also plotted.

10

100

1000 2000

10

Frequency (Hz)

100

1000 2000

Frequency (Hz)

a)

b)

Receivers

214

-0

a=288.7m/s
p 2 0 4 . I nws
2

10

100

1000 2000

Frequency (Hz)

d)
Figure 3. Average insertion loss: a) Elastic versus rigid screens with w = 0.05 m ; b) Elastic versus rigid
screens with w = 0.02 m ; c) model of an infinite elastic panel; d) insertion loss of the infinite cork panel;

For both cases, it is possible to observe that a sequence of insertion loss dips occurs
around specific frequencies. These additional dips are related with the combination of
multiple reflections occurring at the top and bottom surfaces, at f = n x 3 4 0 . 0 / 2 x H .
Thus, the position of the first one is at f = 42.5 Hz when H = 4.0 m , and at
f = 28.33 Hz when H = 6.0 m . Outside these frequency zones, it is also possible to
conclude that the insertion loss registered is, in general, lower than that observed without
the presence of the ceiling. In part, this may be explained by the fact that the energy
portion that is reflected at the ceiling and then travels to the receivers is not affected by
the presence of the barrier, and thus still reaches the receivers without attenuation.
4

Conclusions

In this paper, a BEM model was presented and used in the analysis of sound wave
propagation in the presence of elastic screens placed over an elastic ground or inside a
confined space, subjected to spatially sinusoidal harmonic line pressure sources. The

122
BEM model was found to be efficient, since it only required the surface of the inclusion to
be discretized, while the Greens functions used take the presence of the horizontal flat
surfaces into account.
The results computed for the different situations were compared with those computed
assuming rigid behavior for the barrier. In general, the insertion loss provided by the rigid
screen was higher than that registered for the elastic model. Additionally, the insertion
loss provided by the elastic screen appeared to be highly dependent on the dynamic
behavior of the screen, with insulation dips related to waves travelling within the elastic
material of the barrier. When the barrier was located inside a confined space, additional
insulation dips were visible in the insertion loss curve, which were related to the multiple
reflections occurring between the ground and ceiling surfaces.

--

li

Rigid model
Elastic model

10

100

Frequency(Hz)

a)

1000

10

100

1000

Frequency(Hz)

b)

Figure 4. Average insertion loss for barriers inside confined spaces: a) space defined by H=4.0m; b) space
defined by H=6.0m.

References
1. Duhamel D., Efficient calculation of the three-dimensional sound pressure field
around a noise barrier, J. Sound Vib. 197 (1996) pp. 547-57 1.
2. Duhamel D., Sergent P., Sound propagation over noise barriers with absorbing
ground, J. Sound Vib. 218 (1998) pp. 799-823.
3. Godinho L., Ant6nio J., Tadeu A., 3D sound scattering by rigid barrier in the vicinity
of tall buildings, J. Appl. Acoust. 62 (2001) pp. 1229-1248.
4. Lacerda L., Wrobel L., Power H. and Mansur W., A novel boundary integral
formulation for three-dimensional analysis of thin acoustic barriers over an
impedance plane, J. Acoust. Soc. Am. 104 (1998) pp. 671-678.
5. Tadeu A., Antdnio J., 2.5D Greens functions for elastodynamic problems in layered
acoustic and elastic formations, J. Computer Modeling in Eng. and Sciences, 2
(2001) pp. 477-495.
6. Tadeu A., Godinho L., Scattering of Acoustic Waves by Movable Lightweight Elastic
Screens, Eng. An. Bound. El. 27 (2003) pp. 215-226.
7. Tadeu A., Kausel E., Greens Functions for Two-and-a-half Dimensional
Elastodynamic Problems, J. Eng. Mechanics, ASCE 126 (2000) pp. 1093-1097.

123

NEAR-AXIAL INTERFERENCE EFFECTS IN LONG-RANGE

PROPAGATION IN A RANGE-INDEPENDENT OCEAN
NATALIE S. GRIGORIEVA AND GREGORY M. FRIDMAN
St. Petersburg State Marine Technical University
190008, Lotsmanskaya stx, 3, St. Petersburg, Russia
E-mail: nsgrig CZnata1ie.spb.s~
For ducted propagation in a waveguide when the source and receiver are placed close to the
depth of the waveguide axis, there exist cusp caustics repeatedly along the axis. In neighborhoods of these cusp caustics, the interference of the wave fields that correspond to near-axial
rays occurs. This results in the formation of a coherent structure (the axial wave) that propagates along the waveguide axis like a wave. The goal of this paper is to obtain the integral
representation of the axial wave for an arbitrary deep-water waveguide in long-range acoustic
propagation in a range-independent ocean. The representation has the form of a linear superposition of the solutions of the Helmholtz equation concentrated near the sound-channel axis and
which decrease exponentially outside a strip containing the axis.

1 Introduction
The observed time-of-arrival patterns from a number of long-range acoustic propagation
experiments show early geometrical-like arrivals followed by a crescendo of energy that
propagates along the sound-channel axis and is not resolved into individual arrivals, see,
e.g. [ 6 ] .In these experiments the sound source and receiver are placed close to the depth of
the waveguide axis, and rays emerging from the source at sufficiently small angles intersect
the axis many times forming in its vicinity a large number of caustics with cusp caustics
located repeatedly along the axis. In neighborhoods of cusp caustics, the interference of
the wave fields that correspond to near-axial rays occurs. The goal of this paper is to get the
integral representation for a part of the acoustic field which accumulates all these near-axial
interference effects in long-range propagation in a range-independent ocean.
In the case, when the sound source and receiver are placed at the waveguide axis and
the waveguide is symmetric ( c ( - 2 ) = c ( Z ) , where 2 is the depth coordinate which is
measured relative to the depth of the sound-channel axis), the required integral representation was obtained in [3], where it was shown that the interference of the wave fields which
correspond to near-axial rays leads to a coherent structure propagating along the waveguide axis like a wave. In regions removed from caustics where the geometrical acoustics
can be considered valid, this structure, called the axial wave reduces to a simple wave
associated with the ray path that propagates along the sound-channel axis.
In this paper we construct the integral representation of the axial wave for an arbitrary
deep-water waveguide in a range-independent medium. The sound source is placed at the
waveguide axis and the receiver is located below or above the axis.

2 Geometry of ray paths having small launch angles

For simplicity we consider the situation where the wave field depends only on two spatial
coordinates: the range coordinate X and the depth coordinate 2. It is assumed that the

124

sound speed c( 2)can be expanded in the case of small 2 in the form

c2
c3
c4
.(z)= co + 2 2 + - 2 3 + -2 4 + . .. ,
2!
3!
4!

c2

>o

and when (21

-+ 00, the sound speed tends to a constant c, > cg.
In this section we assume that the sound source and receiver are located at the waveguide axis at points with coordinates (0,O) and ( X ,0) respectively. Comparing the propagation time T2j of a perturbation along a ray path that emerges from the source at an
angle 1c, # 0 and crosses the waveguide axis at the observation point for 2j-th time and the
time To of propagation of a perturbation along the axis (1c, = 0), we obtain that there are
two types of waveguides. If the value

is more than zero, the first to arrive is the perturbation that propagates along the waveguide
axis (the axial wave), and only after it do we get the waves that cross a certain number of
times the sound-channel axis. In contrast, if b < 0 the axial wave will be the last to arrive.
In the first case (b > 0) we will say that the waveguide is strong, and in the second case,
i.e. for b < 0, that it is weak.
It is natural to assume that a wave that crosses the waveguide axis 2 j time can not
be observed independently of the wave that propagates to the observation point along the
waveguide axis if lT2j - TO]< T , where T = 2 r / w is the period of oscillations, w is a
cyclic frequency of the source. We will use the somewhat stronger condition

JTzj- To1 < 00 (

~ 5 ,T). ~ ~

(3)

Here go = 0 ( 1 ) ,0 < 6 < 1/2, ae = (w/co)a is a dimensionless wave number, X, =

X / a is a dimensionless coordinate of the receiver, and a =
is the characteristic
dimension of the waveguide. From the condition (3) it follows that the waves for which
the number of crossings 2 j satisfies the inequalities
for

b>0
(4)

XT
[l-

7r

for b < 0 ,

where u = 2= O( l),interfere with the wave propagating directly along the waveguide axis. As a result, there appears a wave field of the axial wave as a complex interference
wave.

3 Construction of the axial wave

In this section it is presupposed that the sound source has coordinates (0, 0), and the dimensionless coordinates of the receiver are ( z T z,),
,
where z, = 2 / a .
In papers [3,5]the integral representing the exact solution of the reference point source
problem, where the sound speed is given by the expression c( 2 ) = co ( 1 - cr22)-1/2
with
the positive constant cr is transformed in such a way to extract ray summands corresponding
to rays radiated from the source at angles less than a certain angle, the axial wave, and a
term corresponding to a sum of all the rays having launch angles greater than the indicated

125
angle. The integral representation of the axial wave has the form of a linear superposition
of the solutions u:(x,., z,.) of the homogeneous Helmholtz equation which tend to zero as
z, + f o o

where D q ( 6 2,); q = 0, 1 , 2 , . . . are the parabolic cylinder functions, see, e.g. [l].
These solutions are concentrated in a neighborhood of the waveguide axis and decay exponentially outside a strip containing the axis. When ae >> 2(q 1/2), the square root
(1 - 2 ( q 1/2)/ae)lI2 can be expanded in a series.
As it is shown in [3,5],for the reference point source problem, where z, = 0, the axial
wave can be represented by the integral

m exp(3ix/4)

+ 1
1

HI:(X,,O)

00

@Ok>

u0,4,2(x,., 0) u ! L / z ( O 0)
, d7.

(5)

exp(-ix/4)

Here the weight function Oo(7) does not depend on medium properties

r(t)is the gamma-function, and an integer N satisfies the condition

2,. (1 + a(aex,)-0.5+6)
< N 6 5 (1 + o(aex,)-0.5+6) + 1 .

(7)
rr
rr
By analogy with the reference problem, in an arbitrary range-independent medium,
where the sound speed is given by (l), we will construct the axial wave from solutions
u,(X, Z ) , y = 0 , 1 , . . . of the homogeneous Helmholtz equation which are concentrated
near the waveguide axis Z = 0 and decay exponentially with w -+ 03 outside a strip
containing the axis. These solutions can be derived in the form of a product of the exponential function and the parabolic cylinder function, whose arguments are sections of
formal series in powers of w - l l 2 . The coefficients of these series, which are polynomials
in v = w ' / ~ Zcan
, be obtained from a recursive system of partial differential equations.
The general scheme of construction of the solutions u,(X, 2 ) was proposed in [2]. For an
arbitrary range-independentmedium all the required coefficients were found in [4]:

uq(x,.,z,) = exp

+-

i b - l + - x 25) v ; -

'[(

16

+ 241 [(b - 1+ x2)v,"+ 6(y + f ) (b - 1+ 31x2) 4-1;

1 + O ( - -1p ) }

126

Here b is given by (2), x = 2 c 3 6 ( 3 ~ 2 ) - ~ Vr

/ ~=
, 6 Zr.
We will construct the axial wave HN(x,,z,) from the solutions (8) as their linear
superposition. This superposition will have a different form for a strong ( b > 0) and a
weak ( b < 0) waveguide. In the reference problem the waveguide is strong ( a = 1/&,
b = 1, x = 0). For an arbitrary strong waveguide the linear superposition H N ( x ~2,), will
be constructed by analogy with the integral (5)
exp(3ia/4)

03

H N ( x r , 2),

fi

03

u7-1/2(zr,Zr)u T - l / 2 ( 0 7 0 )

dT.

(9)

exp(--ix/l)

In this case the reciprocity principle holds automatically. The weight function O ( T )must
be independent of the medium properties. It can be shown that this requirement is equivalent to the principle of locality which asserts that the high-frequency asymptotic behavior
of a wave field depends only on properties of the medium in the vicinity of the ray. Using (8), we obtain

mexp(3ix/4)

+ & ( 6 - l + < X 52 ) q } x 2

.(T)DT-1/2(/&7)
2

a:
mexp(-i?r/4)

"

+( b - 1+ x2)$ + 67
24ae

If x, is large, the factor exp -iz,T - - b - T in the integrand begins to decrease

2 xr
% 2,
on the upper part of the integration contour only for sufficiently large values of T . For the
error of evaluation of the integral H N to be independent of x,.,it is necessary to select the
weight function O ( T )differently for different values of 2,. We will select an integer N
such that

5
7r (1 + , b ' ~ 2 ( a : z , ) - 0 . 5 + 9

< N 6

xr
7r

(1 + ,b'/2(a:z,)-o.5+6)

+1 ,

(11)

where 0 = 0(1)and 0 < 6 < l / 2 . By virtue of condition (4), if N is even and z, = 0,

the number N is equal to the minimal number of crossings of the waveguide axis by a ray
such that the wave corresponding to this ray is governed by geometrical optics laws and
can be observed independently of the wave propagating directly along the waveguide axis.
On the other hand, when we go over to the reference point source problem ( b -+ 1),the
inequality (1 1) goes over into the inequality (7).
The weight function O(T),which does not depend on parameters of the problem, and
which is therefore the same for all the problems of this class, must be selected in such a

127

0, and z, -+ 0 the integral (10) goes over into the integral (5).
way that for b -+ 1, x
Thus O ( T ) = Oo(7),see (6).
Now let us consider the case of a weak waveguide (b < 0). We will select an integer N
such that

By virtue of condition (4), if N is even and z, = 0, the wave corresponding to a ray that
crosses the waveguide axis N - 1 times will be the last wave that does not interfere with
the wave propagating directly along the waveguide axis and can be observed independently
of the latter. With such a choice of the number N and b < 0, the phase function i(lrN x,) 7 - i bz,7/(2a:) will differ by its sign from the above considered case of a strong
waveguide. For the modulus of the integrand to assume (as before) its maximum value
on the integration contour in a neighborhood of the origin, it is necessary in the case of
a weak waveguide to select the integration contour in the form of a straight line coming
from 00 exp(-3ir/4) to cc exp(in-/4). The weight function will coincide with OO(T)as
before.
0.08

0.06

0.04

0.02

0
-2

-1

Figure 1. Graphs of I H N ( G - , Z ~ ) at
I zT = 0 (solid line) and zTu = -100 m (dashed line) for the Munk
canonical profile.

In Figure 1 graphs of the modulus of the axial wave ~ H N ( X ,z,)]

, are plotted for the
Munk canonical sound-speed profile when a receiver is placed at the waveguide axis (solid
line) or lower the axis z,a = -100 m (dashed line). In this case a = 6651.90 m, b =
-14.8743, and x = 5.12064. The source frequency is equal to 75 Hz that corresponds
to the AET experiment; a: = ( w/ c o ) a = 2089.76. The axial waves are computed in the
interval x, = 2,. lr s, s E [-2,2], where 2, = X,/a = 450.99, X, = 3 . lo6 m.

128

4 Conclusions
In this paper we have proposed a method for obtaining the integral representation of the
axial wave for an arbitrary deep-water waveguide in a range-independent medium. This
representation has the form of a linea superposition of the solutions of the homogeneous
Helmholtz equation which are concentrated in a neighborhood of the waveguide axis and
decrease exponentially outside a strip containing the axis. The weight function is selected
in such a way that the reciprocity principle holds as well as the principle of locality which
allows to use in the limiting case the integral representation of the axial wave obtained for
the reference point source problem.
The expressions for the axial wave are different for a strong waveguide ( b > 0), see (2),
and a weak one ( b < 0). In a strong waveguide, the first to arrive is the axial wave, and only
after it do we get the waves that cross a certain number of times the waveguide axis. In
contrast, in a weak waveguide the axial wave will be the last to arrive. Natural waveguides
(in particular, in the AET experiment) are weak.
The main advantage of the proposed method for obtaining the integral representation
of the axial wave is that all the stages in its construction admit the generalization to the
case of an arbitrary deep-water waveguide in a range-dependent ocean.
5 Acknowledgments
This research is sponsored by the Department of the Navy, Office of Naval Research (ONR).
We are thankful to Dr. D.R. Palmer for useful discussions.
References
1. Bateman H. and Erdklyi A., Higher Transcendental Functions (MC Graw Hill Book
Company, New York, Toronto, 1953).
2. Buldyrev V.S., Asymptotic behavior of solutions of wave equations that are concentrated near the axis of a two-dimensional waveguide in an inhomogeneous medium.
Problemy Mar. Fiz. 3 (1 968), pp. 5-30, in Russian.
3. Buldyrev V.S., The field of a point source in a waveguide. Trudy Mat. Inst. Steklov.
115 (1971), pp. 78-102, in Russian.
4. Grigorieva N.S. and Fridman G.M., Asymptotic behaviour of solutions of the
Helmholtz equation, concentrated near the axis of a deep-water waveguide in a rangeindependent medium. J. Comp. Acousr. (to appear)
5. Grigorieva N.S., Fridman G.M., and Palmer D.R., Investigation of near-axial interference effects for propagation in a ducted waveguide. J. Comp. Acousr. (to appear)
6 . Worcester P.F. el al, A test of basin-scale acoustic thermometry using a large-aperture
vertical array at 3250-km range in the eastern North Pacific Ocean. J. Acousr. SOC.
Am. 105 (1999), pp. 3 185-3200.

129
INVESTIGATION OF NEAR-AXIAL INTERFERENCE EFFECTS
FOR PROPAGATION IN A DUCTED WAVEGUIDE
NATALIE S. GRIGORIEVA AND GREGORY M. FRIDMAN
St. Petersburg State Marine Technical University
190008, Lotsmanskaya stx, 3, St. Petersburg, Russia
E-mail: nsgrig@natalie.spb.su
DAVID R. PALMER
NOAALAtlantic Oceanographic and Meteorological Laboratory
4301 Rickenbacker C s y , Miami, FL 33149, U.S.A.
E-mail: David.R.Palmer@noaa.gov
The observed time-of-arrival patterns from a number of long-range ocean acoustic propagation experiments show early geometrical-like arrivals followed by a crescendo of energy that
propagates along the sound-channel axis and is not resolved into individual arrivals. To describe in a simple model case the interference of near-axial waves which resulted in forming
the so-called axial wave and propose formulas for the axial wave in more general cases, the
two-dimensional reference point source problem for the parabolic index of refraction squared
is investigated. The integral representation for the exact solution is transformed in such a way
to extract ray summands corresponding to rays radiated from the source at angles less than a
certain angle, the axial wave, and a term corresponding to a sum of all the rays having launch
angles greater than the indicated angle. Numerical results for the axial wave are obtained for
parameters corresponding to long-range ocean acoustic propagation experiments.

1 Introduction
In many long-range propagation studies the source and receiver are placed close to the
depth of the waveguide (SOFAR) axis to minimize the interaction of the acoustic field with
the oceans surface and bottom. For these experiments, the most pronounced characteristics
of the time-of-arrival patterns are early geometrical-like arrivals followed by a crescendo
of energy that propagates along the axis, see, e.g. [4]. This late-arriving energy can not be
described using geometrical acoustics, because of the presence of caustics.
For ducted propagation in a waveguide there exist cusp caustics repeatedly along the
axis with a very complicated interference pattern in their neighborhoods. These neighborhoods of interference grow with range and at extremely long distances can even overlap.
As a result, the correct representation of the wave field must include an additional term the axial wave. It describes the wave associated with the ray path that coincides with
the sound-channel axis, as well as the interference between this wave and the waves that
propagate along other ray paths.
In this paper for simplicity we consider the situation where the wave field depends
on two spatial coordinates: the range coordinate, X , and the depth coordinate, 2, which
is measured relative to the depth of the sound-channel axis. Our goal is to construct the
integral representation of the axial wave for the reference point source problem with the
parabolic index of refraction squared and investigate the obtained expression when the
dimensionless distance between the source and receiver is the value of the order of lo3,
that corresponds to the AET experiment. The case when this dimensionless distance is the
value of the order of unity was considered in [2,3].

130

2 Transformation of the exact solution

In the reference point source problem the sound speed is given by the expression c( 2 ) =
(1 - a22)-1/2,
where co and a are positive constants. In terms of dimensionless coordinates x = &
z = 6 2 and a dimensionless wave number a: = w/(&~),
where w is a cyclic frequency, the standard formulation of this problem is as follows: the
functions U j ( 5 ,z ) , j = 1,2, representing the wave fields above and below the depth of the
source respectively must satisfy the wave equation

x,

a2ui a2ui+ c e 2 ( 1 ax2

az2

-$- -

Z2)Uj

=0 ,

j = 1,2,

the conditions

and the radiation condition at infinity. Here 6(x) is the Dirac delta-function, (0, zo) are the
dimensionless coordinates of the source.
Let (zr,
z,) be the dimensionless coordinates of the receiver and 2,. 2 20. Using the
method of separation of variables in (l), ( 2 ) ,we obtain

--M

where Dt (6
z ) is the parabolic cylinder function, see e.g. [ 11,
t = - 1[ z ( l - < )2- l ] .
(4)
2
The contour of integration in ( 3 )must be displaced slightly above the real axis in the second
quadrant and slightly below it in the fourth quadrant.
Below we will examine the case when the source and the observation point are located
at the waveguide axis. Then ( 3 ) takes the form
+-M

--M

where

<

The poles of the integrand in (5) are located at = kern = [l - k ( 2 m l)]12. Representing the wave field by a sum of residues of the integrand at the poles
lying in
the upper half-plane we would get the wave field in the form of a sum of propagating and
decaying modes, but this representation does not correspond to the concept of rays.
To extract from the exact solution (5) the axial wave we will transform the obtained
integral using the transformation proposed in [2]. This transformation is as follows. In the
<-plane, = p + iy, we construct a straight line, see Figure 1

crn

<

y = cot

x2 ( p

3),

(7)

13 1

where x = arg ae, 0 < < x C T / ( N rwith

) , N is an integer. The exact value of N will be
indicated later. Without crossing the poles of the integrand, we can deform the integration
contour L into the contour A0 shown in Figure 1.

Figure 1. Straight line (7) and contours L, XO, A 1 and Xz in the <-plane.

On the contour XO, the ratio (1

progression

+ e int )/ (l

eint) can be expanded in a geometrical

Summands corresponding to unit and the first sum give integrals which vanish since their
integrands do not have singularities in the region to the left as one moves along the integration contour XO. As the result we get for the solution (5) the integral representation

A0

As the next step, let us replace one integral along the contour XO by two integrals taken
over X1 and XZ, see Figure 1, and in the integral taken over A2 we use the expansion

Then the exact solution (9) will be represented as a sum of integrals

M-1

u(G-)=

m=O

SN+~(G)

+ SN+M(G) + HN(G-)

132
where

exp (i.lrtN iaez,<)

d<.
1 - exp(in-t)
A1

Using the method of steepest descent, we obtain that the summand (1 1) describes a wave
that arrives to the observation point after N m intersections of the waveguide axis if

where CT = O(1) and 6 < 0.5 are positive constants. The inequality (14) means that
the correction term specifying the error of the steepest descent method decreases with
increasing IazJ. Then it is evident that the integral (12) describes waves that cross the
waveguide axis N M times and more.
The totality of waves for which the number of crossings of the waveguide axis does
not satisfy the inequality (14) is described by the integral HN(z,). All that waves do not
have geometrical characteristics interfering with the wave propagating directly along the
waveguide axis. Thus, H N ( Z , ) describes the axial wave.

3 Computation of the axial wave

Let us select an integer N as the smallest number satisfying (14). It means that N satisfies
the inequalities
Xr

- (1

+ u~az,~-o~5+6)

B N < 5 (1+ alaez,l-0.5+6)

n-

+ 1.

(15)

In terms of a new variable T = a: (1- <2)/2 the integral H N ( z , ) can be represented in the
form

(16)
where
00

VN(G")= -

exp(3ir/4)

exp (i(rN - x,)T -

R(r-i)

- -7

2 %

+ iexp(in-7) zr

') d r .

(17)

mexp(-ir/4)

Thus the dispersion of the axial wave and the variation of its intensity with distance are
completely determined by the properties of the function VN(z,).

133
Let x: be such that

where No is an integer. For x, = x! - E , E > 0, the number N will be equal to No - 1,

whereas for x, = x: we have N = NO,and in the expansion (10) it becomes smaller
by one term, namely S N ~ - ~Hence,
.
the number N varies in a jump-like manner with
increasing x,.
With the aid of the relation

exp(ilr(N 1 ) ~ ) exp(i7rN.r)
- i exp(i7rN.r)
=i
1 + i exp(im)
1 i exp(i7r.r)

it is easy to see that, to within correction terms, we have

0
HNo(x,) - HN~-~(x:)
=~ ~ 0 - 1 .

(19)

Hence for values x! such that No in (18) is an integer, the wave that crosses NO- 1 times
the waveguide axis will lose its geometrical characteristics, being absorbed by the axial
wave.
Let the axis sound speed, CO, and the spatial scale, 1 1 6 ,be given by
co = 1480.2 d s ,

l/& = 4306.6322 m ,

that corresponds to the AET experiment, see [4]. The source frequency is 75 Hz, so
w = 150.n s-'.
In this case a: = 1371.0665. In Figure 2, ~HN(X,))
is plotted
for the dimensionless distance x,. between the sound source and receiver in the interval
[ I, - 2 ~IT
, 2 ~ 1where
,
2 , = XT&, X, = 3 x 106m, 5, = 696.60. We note that
outside of a neighborhood of the points of caustics cusps x, = lrj, j = 1 , 2 , . . ., this wave
can be approximated by a simple geometrical acoustics formula

where j = [x,/.rr]
is the integer part of x,/lr. In Figure 2, a graph of the function IH;(x,)l
is plotted as well.
4

Conclusions

In this paper we have investigated the exact solution of the two-dimensionalreference point
source problem when the sound source and receiver are located at the waveguide axis.
Using the transformation proposed in [2], we extracted from this solution ray summands
corresponding to rays radiated from the source at angles less than a certain angle, a term
corresponding to a sum of all the rays having launch angles greater than the indicated
angle, and the axial wave. The axial wave describes the wave associated with the ray path
that coincides with the sound-channel axis and the interference between this wave and the
waves that propagate along other ray paths. In regions removed from the caustics, the axial
wave reduces to the wave associated with the ray path coinciding with the sound-channel
axis.

134

0.08

0.06

0.04

0.02

-2

-1

a
Figure 2. Plots of IHjv(rr)I and

lH:(Xr)/

for X r = f r

+ K U ,a E [-2,2].

Acknowledgments
This paper is dedicated to the memory of Professor Frederik D. Tappert who was always
very generous in sharing his knowledge and views regarding sound propagation in the
ocean.
The research is sponsored by the Department of Navy, Office of Naval Research (ONR).
References
1. Bateman H. and Erdtlyi A., Higher Transcendental Functions (MC Graw Hill Book
Company, New York, Toronto, 1953).
2. Buldyrev V.S., The field of a point source in a waveguide. Trudy Mat. Inst. Steklov.
115 (1971), pp. 78-102, in Russian.
3. Buldyrev V.S., Lanin A.I., and Yanson Z.A., Calculation of a field at the axis of a
symmetric waveguide. Voprosy Dynam. Teor: 14 (1974), 84-93, in Russian.
4. Worcester P.F. et al, A test of basin-scale acoustic thermometry using a large-aperture
vertical array at 3250-km range in the eastern North Pacific Ocean. J. Acoust. SOC.
Am. 105 (1999), pp. 3185-3200.

135

DYNAMICS OF IMMISCIBLE TWO-PHASE FLUID RESERVOIR

FLOW
A. HANYGA
Department of Earth Science

University of Bergen, Alligaten 41, Bergen, Norway

A thermodynamic model of porous medium consisting of a solid phase and several
fluid phases is proposed. In contrast to the prevailing trend in current thermodynamics of porous media the the assumption of phase separation is avoided. In
isothermal conditions a single total free energy of the composite thermodynamic
system is assumed. For non-isothermal conditions each phase is endowed with a
free energy and entropy depending on its average temperature, but the constitutive
equations of each phase depend on the deformation and temperature of the other
phases. A general form of drag forces is discussed. The coupling between fluid
pressures is modeled by the Preisach hysteresis. Interphase heat flow is discussed
and the effect of average temperature differences on capillary hysteresis. Equations
for multi-phase flow are derived. For simplicity of the presentation, it is assumed
that each phase consists of a single species. A model of wave propagation is obtained by linearization and neglecting the hysteretic aspect of capillary effects and
the coupling with fluid displacement is demonstrated. The possibility of deriving a
multi-phase flow model and wave propagation equations from a single mathematical model opens a way to connecting seismic modeling with reservoir simulation
in petrophysical applications.

Introduction

An important issue faced while developing a mathematical model of a complex

physical phenomenon is reducing the number of defining functions to a minimum.
Two methods are currently used to this purpose in the theory of multi-phase porous
media: a pseudo-Lagrangian variational framework known as Biot s principle of
virtual dissipation and the Coleman-No11 version of continuum thermodynamics.
Both approaches have an additional advantage over generic phenomenological models or models derived directly from averaging: they establish a duality of kinematic
variables and their conjugate dynamic counterparts. There are several benefits
from such a parametrization of the model state space. In the context of multiphase
porous flows an identification of conjugate variables provides a clue for the choice of
macroscopic variables when they are not obvious, in particular for capillary hysteresis. Of the two approaches the thermodynamical one is less economical because it
requires postulating several independent equations and constitutive functions. On
the other hand, it is impossible to include dissipative phenomena in a Lagrangian
frameworka, unless nonintegrable terms are introduced l y 4 t 5 following the principle of virtual dissipation. In the context of porous media the drag forces and
hysteretic phenomena are the main obstacle to a truly Lagrangian formulation.
The pseudo-Lagrangian theory does not provide any clue about the direction of
dissipative processes, while many advantages of a truly Lagrangian formulation,
such as linking conservation laws to symmetries, are lost in a pseudo-Lagrangian
aFor a class of dissipative media Lagrangian and Hamiltonian theories involving fractional time
derivatives of field variables can be constructed
392.

136
framework. On balance the alternative thermodynamic approach
is
more appropriate for porous media.
In so far as kinematics is concerned, porous media are commonly represented
in terms of generalized mixtures. A generalized mixture consists of several constituents, which are called phases for a reason that will soon become clear. At
every point of the space occupied by generalized mixture all the phases are present
in an amount that is expressed in terms of two independent parameters: mass
density p('") per unit volume of the mixture and volume fraction n(") (the fraction of the total mixture volume occupied by the phase). The true mass density
y(") := p(")/n(")is commonly linked by the constitutive relations to the pressure
in the phase, while p(") appears in the mass balance equations. If one fluid displaces the other at fixed fluid pressures, then y(") remains constant while p(") and
n(') may vary. In the context of a consolidated porous medium consisting of one
solid and several fluids porosity 4 is the complement of the solid volume fraction:
4 = 1 - n ( s ) ,where the superscript "s" stands for solid. Denoting the wetting
and non-wetting fluid by the superscripts "w" and "n" , the saturation is defined
by the formula S = n(")/$. In general each phase is a mixture of several species
(solution of gas in oil or water, a solid matrix consisting of ice and sediment) and
the amount of the species in a phase is represented by its concentration, which can
be represented by mass fraction. Phases can exchange species, e.g. solution of free
gas in oil can be regarded as transfer of gas species between two phases, free gas
and live oil. The rates of such exchanges depend on chemical potentials. Species
transfer between phases is an important aspect of the compositional flow models 13.
For simplicity of presentation we shall assume here that each phase consists of one
species only.
In the single species phase model mass transfer between phases represents local
phase transformations. Non-local phase transformations at the macro scale can be
modeled in terms of moving phase boundaries separating porous medium characterized by different composition.
One of the drawbacks of mixture theories is a proliferation of balance equations.
Whenever an interphase exchange of mass, momentum, moment of momentum or
energy is introduced, the corresponding conservation law is replaced by a set of
partial balance equations, one per each phase, with interphase exchange terms, and
a conservation law which asserts that the exchanged quantities add up to zero. The
interphase exchange terms are assumed to be determined by constitutive equations.
In order to reduce the complexity of the model the simplifying assumption of phase
separation is usually made. According to this assumption, various phases of the
multi-phase porous medium constitute separate thermodynamic subsystems coupled only through mass, momentum and energy exchange terms in the respective
balance equations. In particular the state variables of a phase do not depend on
the states of the other phases. Phase separation is typically assumed in abstract
mixture theories and is often incompatible with the physical properties of more
specific media, in particular with Biot's theory of saturated porous media.
Partial mass and momentum balances are not difficult to formulate by assigning
each phase a mass and a momentum, expressed in terms of densities and velocities,
and it is safe t o assume that the total mass of a species in the mixture and the total
61977710912716t14

137
mixture momentum can be obtained by summing over all the phases. The case of
energy balances is much more complex. Due to the presence of stored and thermal
energy no obvious expression for the partial internal energy or entropy is available.
On account of interactions the total energy is not the sum of the partial energies. At
the microscale level the interaction in porous media is commonly believed to occur
mainly through the interphase interfaces. Consequently, in the generalized mixture
model of porous media
it is often assumed that the interaction problem
can be solved by treating interphase interfaces (and contact lines) as additional
phases endowed with finite mass, energy and even momentum, which is justified
in the case of energy. Associating a finite mass with an interface is justifiable if
adsorption plays an important role. In a mixture model which treats interfaces
as additional phases interface area per unit mixture volume plays the role of a
volume fraction and the formal structure of the model remains unaffected by the
topological peculiarities of the additional phases. Macroscopically, the thickness of
an interface layer corresponds to a few molecular layers (10-100 AA) and therefore
an interface has t o be treated as a surface. The amount of interface in the porous
medium can be expressed in terms of interface area per unit volume, which is
a counterpart of volume fraction, while the extensive variables associated with
interfaces are represented by their superficial densities (per unit interface area).
Perhaps the only difference between volume and interface phases is the fact that
thae latter come into existence only by bringing two volume phases into contact,
which gives the associated extensive thermodynamical variables the appearance of
interaction related. Associating extensive quantities with interfaces does not solve
the interaction problem. Once the interface is considered as a thermodynamic
subsystem endowed with its own set of extensive parameters, the question of its
interaction with the other subsystems can be raised again, leading to an endless
proliferation of hypothetical subsystems. It is preferable to associate a single energy
and entropy with the total system, rather than with its interacting subsystems.
This leads to a single potential for the constitutive equations pertinent for nondissipative phenomena and interaction between phases is automatically built into
the constitutive equations.
Is it possible to avoid separate energy conservation laws for all the phases?
The answer is no if the phases can have different temperatures a t every point. In
this case a separate energy balance must be postulated for each phase in order to
provide the equations for the determination of the temperatures of all the phases.
The temperature of a phase represents an average of the temperature of that phase
in a Representative Elementary Volume.b If the phases are in relative motion then
their locally averaged temperatures are not equal, notwithstanding continuity of
temperature across interface boundaries in the micromodel 14. The difference of
average phase temperatures results in a temperature gradient at the microlevel and
a heat flow across the interphase boundary. The heat flow is proportional to the
difference of the averaged phase temperatures. In the macromodel the average phase
temperature is represented by the temperature of that phase at every point of the
macromodel while interphase heat flow is expressed in terms of an interphase energy
779,10,12

bMore precisely, the averaging process applies to the thermal energy density.

138
exchange at every point. The interphase heat exchange is assumed a constitutive
quantity, in particular it is a function of the phase temperature difference. The
interphase energy exchange appears in the partial energy balances as a source term.
All the energy exchanges, those due to mass and momentum exchanges as well as
the interphase heat exchange sum up to zero. This is the content of the energy
conservation law. As for entropy, its rate of growth exceeds the reversible part
associated with heat supply.
In the case of generalized mixtures it is necessary to complement the usual
system of equations of an ordinary mixture by supplementary equations (closure
equations) corresponding to the additional degrees of freedom represented by the
volume fractions. This problem has been treated in different ways. In the paper of
Goodman and Cowin l5 and other papers influenced by their work, including Ref. 5 ,
a kinetic energy was associated with volume fractions leading to balance equations
reminiscent of other balances. The inertial effects give an oscillatory aspect to the
volume fraction dynamics. The inertial forces in volume fraction equations are
coupled to pressure differences, in particular to the fluid pressure difference 5 .
Alternatively, the volume fractions are interpreted as internal parameters subject to relaxation type constitutive laws 7 . The latter formulation allows for the
dissimilarity between porosity (related to the volume fraction of the solid matrix)
and the fluid volume fractions, which can are affected by fluid displacement. Porosity of a connected solid matrix is not related to the fluid displacement process and
is only affected by the deformation of the matrix. On the other hand, the volume
fractions corresponding to fluids in a three-phase porous medium are commonly
expressed in terms of saturation, which is defined as the ratio of the volume of one
fluid with respect to the pore space. The dynamics of the volume fractions is then
expressed in terms of a relaxation law for the saturation relating the saturation
rate to the difference of the pore pressures of the two fluids. In view of the specific properties of capillary hysteresis this relation is more conveniently expressed
in the inverse form: the saturation is expressed in terms of a hysteresis operator
applied to the history of the excess of the pore pressure difference over its equilibrium value. Capillary hysteresis is a phenomenon involving one solid phase and
two fluid phases, controlled in part by the contact angle hysteresis. Consequently,
in contrast to some papers
we associate one capillary relation with each pair
of fluid phases. In some papers the symmetry between fluids and solids is extended
to drag forces and capillary relations, which are then associated with each pair of
phases l o J z > i l . It takes however at least two fluid phases and one solid phase to
define a capillary relation.
The general form of the hysteretic constitutive relation is suggested by the
residual entropy production inequality while its more specific features have been
established experimentally. The importance of hysteresis for numerical simulation
of hydrocarbon recovery has been demonstrated by Killough 1 7 . A simple model
of capillary hysteresis based on the play hysteresis was proposed in Ref. 2o while a
more realistic model was proposed in Refs l6?l. The above papers base the constitutive equations of capillary hysteresis on continuum thermodynamics. In Ref. l6
it is shown that the Preisach model of hysteresis is consistent with the dissipativity condition derived from the Clausius-Duhem inequality, while in Ref. 21 it is
l 2 7 l o

139

demonstrated that the Preisach model reproduces all the details of the experimentally determined scanning curves and minor hysteresis loops. The last two papers
are otherwise close in spirit t o the speculative models of soil wetting/drying and
adsorption Refs
An important interphase momentum exchange is due to drag forces. The drag
forces have been studied in some detail in several papers, including Refs
We assume that drag forces act between solid and each of the fluids only, although
there are (incomplete) physical models of interfluid drag forces for some porous
media 5 8 . It is then intuitively acceptable t o assume that drag forces in a two-fluid
porous medium are identical with the drag force in a saturated medium. In agreement with Refs 49,5,47, the relation between the drag force (or interphase momentum
exchange) and the relative velocity of the two phases is given by a convolution operator. The high- and low- frequency behavior of the transfer function (or small
time and large time behavior of the convolution kernel) is well-established but there
is some controversy regarding the intermediate frequency range. In Ref. l6 the a
priori condition of dissipativity is discussed for a general drag force.
Equations for wave propagation in porous media saturated by two immiscible
fluids have been previously obtained by an averaging method 63,64 and by variational
principles
Earlier references include Brutsaert 66 and Garg and Nayfeh 67. In
the last references the coupling with fluid displacement and memory effects associated with drag forces and particularly relevant for ultrasonic experiments are
ignored. Assuming that shear stress in the pore fluids is negligible, the number
of slow waves must be equal to the number of fluids. Drag forces determine the
distinctive properties of the slow wave as well as the attenuation of the fast longitudinal and shear wave. Strong damping of high frequencies results in a smoothing
of the wavefields at the wavefronts 6 8 .
Interphase thermal transfer is important if the process brings in contact phases
at different temperatures like in steam flooding. Temperature effects are also important in geothermal reservoirs and in numerous technical applications. Thermal
effects on poroelastic wave propagation in saturated porous media are studied in
several papers, including Ref 6 9 . The additional variables (phase temperatures)
as well as the corresponding equations (partial energy balances) generate an equal
number of additional waves. Ref
demonstrates the effect of thermal dissipation
on low-frequency attenuation. Thermal effects in porous media have been studied
in saturated porous media 70 and in a two-fluid model 8 1 . In multi-fluid flows heat
transfer is affected by both dispersion and interfluid heat transfer 5 8 . A multiphase model of a porous medium involving capillary hysteresis and thermal effects
is briefly discussed in Ref 1 4 .
In Sec. 2 a single-temperature model of a multi-phase porous medium is developed. Thermodynamic restrictions on drag forces and capillary hysteresis are
discussed in Secs 3-4. A model allowing for different phase temperatures and interphase heat flows is discussed in Sec. 5. In Secs 6-7 applications to multi-phase
fluid flow and wave propagation in multi-phase porous media are sketched.
27928,31.

5747y48955.

5165.

I40
2

Balance equations, constitutive equations and the

Clausius-Duhem inequality for a single-temperature immiscible
flow

2.1

Kinematics

The state of the two-fluid immiscible flow is represented by the following set of
variables: the densities of the three phases, the deformation gradient of the solid,
the velocities of the solid matrix v('), d n )and dW)
of the two fluids", the saturation
S of the wetting fluid and the temperature 8. The corresponding equations are three
partial mass balances, three partial momentum balances, an energy conservation
law and the constitutive equations, including the constitutive equations for the
momentum transfer between the solid and each of the fluids (drag forces) and the
capillary relation, which is the closure equation associated with the saturation.
Interphase mass exchanges can be linked by constitutive equations to the chemical
potentials. In non-isothermal conditions the heat flow and the interphase heat
exchange have t o be included. The latter requires a separate energy balance for
each phase with an interphase energy exchange term linked to the difference in
average phase temperatures and other variables.
We now proceed to list the partial balance equations and conservation laws
following the method developoed for ordinary mixtures in Ref. '*. Let v(")denote
the speed of phase a. The associated material time derivative will be denoted by
Da:

for a generic function j ( t,x).

For simplicity we assume that each phase consists of a single species. This
assumption excludes miscible fluids such as solutions of gas in oil and water.
2.2 Partial mass balances and mass conservation

In the miscible flow the mass balances have to be associated with each constituent
and the exchange of a constituent between two phases is a constitutive function of
the chemical potentials. In the immiscible flow a partial mass balances has to be
specified for each phase

or, equivalently,

where p(") denotes the bulk mass density (per unit mixture volume) of phase a and
p(a) denotes a mass transfer to phase a from the other phases, for example due to
=The superscripts "s","w","n" indicate the solid matrix, the wetting fluid and the non-wetting
fluid respectively.

141

a phase transition. The mass conservation law assumes the form

a

If y(a) denotes the true density of phase a and n(")is its volume fraction, then
p(a) =
y("). The true fluid mass densities y ( a ) will appear in the argument of
the constitutive equations.
Let
a

and

The barycentric velocity is defined by the equation

a

and the relative velocity of phase a

,(a)

:= ,(a)

-v

(8)

For a low-porosity material the density of the solid dominates in p and n(")u ( ~ ) ,
a = w, n, can be considered as an approximate filtration velocity of fluid. The exact
filtration velocity of fluid a is n(")[ u(") - u'')']. Let D f := := af /at + v . V f
(barycentric time derivative).
Summing the partial mass balances and using eq. (4)yields the mass conservation equation:
p+pdivv=O

(9)

2.3 Partial momentum balances and momentum conservation

The partial momentum balances are assumed in the usual form

The argument of the divergence operator on the left-hand side is a dyadic. The
vector
represents the supply of phase a momentum density from the other
phases, e.g. the sum of drag forces acting on phase a , while b(") is the external
force density (e.g. gravitational force). T(a)denotes partial stress per unit area
associated with phase a (the true phase a momentum flux density). The momentum
supply associated with mass transfer to phase a is represented by the third term on
the right. The total momentum density balance follows from the action-reaction
law for interphase momentum density exchanges

da)

142

The total momentum density conservation follows from eq. (11) by summing
the partial momentum density balances and applying the identity

for f := C , da)f(") and generic functions f'"), possibly vector- or tensor-valued:

p v = div T

+b

(14)

where the total stress

and b := Cab(").
Identity (13) is easy t o check. Note that our definition of the total stress differs
from Ref. 18.

2.4 Energy balance and entropy production

As announced in the introduction we shall associate a single specific internal energy
E

with the system. It is assumed that the total energy satisfies the energy balance:
= div [TTv]

+ div q +

b(").

+r

where r is an external energy source at every point of the medium, q is the heat
influx density. Like the external loads b("),the arbitrary energy supply r serves
the purpose of removing the constraint of energy conservation. One can handle the
same problem in an alternative way by assuming energy and partial momentum
conservation and accounting for the resulting constraints in the energy production
inequality by means of Lagrange multipliers 1 9 .
A reduced form of the energy balance is obtained by applying the momentum
conservation law (14):

The total kinetic energy density consists of two parts

The intrinsic internal energy density EI is defined by subtracting the second term
on the right-hand side from the internal energy density:
&I := E -

c
a

,(a)

u(aI2/2

(19)

143
It is easy to see that
[,(a) ~ " ~ ( a.U(a)
) ]

Da

[,(a)

(.

+ u(a))]

.u(a)= p(a)

I)'(,

- p(a) D" -u(a)

1
2 + [pia) u ( ~ )
-

: grad v

+ p(a)~

( . Dv
~ 1

The last term drops out upon summing over a. Substituting p(a) D a d a ) from the
partial momentum balance and applying identity eq. (13) with f(") = ~ ( ~ )we~ / 2
have after some straightforward algebra
1
-pD
2

c(~)u ( ~ )= div k

+T :L

with

I denotes here the unit matrix.

Substituting the last identity in the energy balance (17) we have
a

Let q denote the specific entropy of the mixture. The intrinsic free energy is
defined by the equation
+I :=

qe

(25)

Since we shall soon prove that q = 4+1/69,eq. (25) is a Legendre transformation.

It is assumed that the intrinsic free energy, the partial stresses, the heat flux,
the energy flux k and the mass and momentum interphase transfers in three-phase
porous media can be expressed by constitutive functions of the following arguments:
the strain tensor E := FTF - I of the solid (F denotes the deformation gradient),
y("),
S, 6 , g := grade, f(a) := gradpa, h = gradS, f(") := grady(") and
~ ( ~ a1 =
, s,n,w. Note that y(s) = p t ' / [ ( l - 4 ) detF], where p t ) denotes the
bulk density of the solid in the reference configuration. Furthermore the diffusion
velocities u(")are not independent in view of the identity C , da)u(")G 0.
We assume that the entropy production inequality
p7j

2 div (0-'

q)

+~/6'

(26)

is satisfied by every thermodynamic process satisfying the balance equations and

the constitutive equations. Since the external force b(") acting on each phase is
assumed arbitrary, the partial momentum balances do not impose any constraints

144

on the solutions of eq. (26). The energy supply term r can be eliminated from
eq. (26) by using the energy balance. The reduced entropy inequality is obtained
by substituting the energy balance in eq. (26) and substituting the intrisic free
energy EI = $1
87

Following Hanyga l8 we assume that for every collection of smooth function pk'(x)
defined on a neighborhood U of the origin and every smooth d a ) ( t , x ) defined in
[O,T]x U for some T > 0, the partial mass balances

have a local solution in a subset of [O,T]x U. The right-hand side of the last
is a shorthand
equation is determined by the constitutive function for i j ( a ) and
for the set { f ( a ) } a .
Let k be given by the constitutive relation

div k shall be decomposed into a sum of three terms

divk = A

+B +C

with

and S by using the relations p(a) =

and p(") = $ (1 - S ) ?("I.
Furthermore
E = 2FT D F, with D := (L + LT)/2. Substitution of the constitutive relations

+(") can be expressed in terms of divd")

~ ( a- )~ ( a ).f(") - p(") div v(a) p(") = $ ,"y(")

145

in eq. (27) yields the inequality

a=w,n

+ k7(a) . f ( a ) + k e . g + kf(a): VVp(") + k,

: VVf3

+ 8-l

q .g

5 0 (36)

where the subscripts of and k denote derivatives. Since the external loads b(")
are arbitrary, the vectors u(a)can assume arbitrary values and
0. Since
at a fixed time the vectors fc.1, h and g can assume arbitrary values, $h =
=
&a)
E 0. Since VVf3 and VVp(") can be chosen arbitrarily, the matrices kg and
kf(.) are antisymmetric. The fact that 4, L(a),and S can assume arbitrary values
implies the identities

4u(a)

4,

with
$1

7,
s,0)

= 4(E1

(41)

On account of Galilean invariance the function k depends on velocity differences

only. Consequently

da)-

Xk&)= 0
a

and

a=w,n

Eq. (37) shows that partial fluid stresses neednot be isotropic. Introducing the
fluid pressures
p(a) = p y(")

(43)

146
inequality (36) reduces to

[ p & - q5 (p(w) - p'"')]

s+

p?j+)

. .,(a))

/TI@)+

a=w,n

+ C m(a). u ( ~+)

&,-,(a)

. f(")+ (ke + q / O ) . g 5 o (44)

The temperature dependence of the energy flux k results in additional thermal

dissipation. The first term in eq. (44)represents the dissipation associated with
the capillary hysteresis. The equilibrium capillary pressure can be defined by the
formula

p c = -P4s/q5

(45)

Since imbibition occurs spontaneously it is expected that Pc 2 0. The first term

of eq. (44)does not depend on temperature gradient and can be decoupled from
thermal dissipation. The third term represents the power expended by drag forces
and by momentum transfer associated by mass transfer. If the drag forces do not
depend on the temperature gradient then this term also decouples from the thermal
dissipation. The second term of eq. (44)represents the dissipation associated with
mass transfer between phases and the dissipation associated with the motion of
phases in a non-homogeneous distribution of phase mass.
In general many dissipative processes can be coupled in view of the dependence
of the thermodynamic forces on all the thermodynamic fluxes.
3

Drag forces

For sufficiently small Reynolds number Re p v d/mu 5 1, where d is a characteristic linear length associated with the pore, v denotes the avewrage pore velocity
and p is the fluid viscosity, non-linear effects (the Ergun law) can be neglected. In
this approximation the drag forces are determined by the constitutive equations

for generic functions f , g. In order to satisfy momentum conservation it is assumed

that G(",b)z G(b7").
Thermodynamic restrictions on drag forces are met provided -G("lb) is a
matrix-valued function of positive type, i.e.

( t ) ,t 5 T . The conditions ensurfor an arbitrary set of vector-valued functions

ing this property are discussed in Ref. 16. In the time domain they reduce to the

147

requirement of nonnegativity, monotone decrease and convexity of the scalar function -u(") . G("7b)(t)
u ( ~for
) every set of constant vectors u(").For an application
in the context of the constitutive laws for drag forces or viscodynamic operators
proposed in Refs
a criterion for the Fourier transform of -G("7b)is however
more appropriate. Such a criterion can be deduced from the Bochner theorem 56.
If G("ib)(0)is finite then condition (48) is equivalent to the equation
4975,47

cx

[u(a) . G ( a ? b ) ( w )u(b)]

5o

for all real w

(49)

a,b

for all vectors u("). Equivalently, the cosine Fourier transform of -u(") .
G("?b)(t)
u ( ~is
) non-negative for all vectors u("). This condition is reminiscent
of the celebrated Graffi condition in viscoelasticity 5 7 .
The above considerations can be extended to non-integrable kernels G("ib),
in
particular to fractional derivative models of drag forces, by using the SchwartzBochner theorem (Theorem 16.2.6 in Ref. 5 6 ) . The Biot model of a saturated
porous medium involves a fractional derivative od order 1/2 of the relative fluid
velocity 5 0 , which is equivalent to a kernel G ( t ) const x t - 3 / 2 as t + O+. The
necessary and sufficient condition involves the same inequality and the condition of
at most polynomial growth.
The drag forces act between the solid and each of the fluids and can be expressed
in terms of the permeability tensors K ( f )

G ( s > f )=
(~
-IE,f,,(w)
)
K = -K(j)(w)

(50)

The permeability tensors satisfy the dissipativity condition

(51)

~.XK(~)(w)w20

for every vector w.

Interfluid drag forces are occasionally taken into account, cf 5 8 . Some theoretical
models of quasi-static interfluid drag forces have been suggested. They depend on
the detailed description of the flow and therefore are of limited applicability
It is reasonable to assume that the drag forces acting on a fluid in a multiphase medium are the same as in a porous medium saturated with the same fluid.
This assumption is clearly justified in the case of channel fluid flows and BuckleyLeverett wavefronts. Using the notation of Pride et al. 48 for a saturated porous
medium, which includes references to previous work,
59160,61.

. s 'f ) ( w ) = -Pf
G(

wo f(z)/(am Ps - P f )

(52)

with z = w P/wo,and a constant P > 0. A term (pf ( a , - 1)(u(f)- u('))/4from

their expression for the drag force has been included here in the inertial part. In
view of eq. (49) dissipativity is equivalent to Rf(z) 2 0 for all the real z.
Pride et al. 48 discuss five alternative expressions fi, 1 = 1,.. . ,5, for the function
f . The plots of X k / l c ~from their paper show that in the frequency range considered
the inequality is satisfied. For arbitrary real z the inequality Xfi(z) 2 0 can be
checked by direct calculus for 1 = 3, while for 1 = 1 , 2 the identities Xfi(z) =
Xfi(-z), Sfi(z) = -9fi(-z), S z 2 = 2 8 z 9 z are helpful.

148

The matrix-valued kernel G(a3b)(t)

is said to be singular if GcaPb)(t)
as t -+ 0+, for some c > 0, 0 < (Y < 2. It has long been recognized that the
drag forces are given by hereditary constitutive laws which involve singular integral
kernels
In the frequency domain this fact is apparent in the high-freqency
asymptotics, which involves terms 0; ( -iw)-'I2. Partial integro-differential equations with such kernels have long been studied, mainly in the context of viscoelasticity
In this case the singular memory property amounts to the assumption
for small
that the time derivative of the stress relaxation function G'(t)
t
0+, with 0 < a < 1. The main effect on wave propagation is smoothness
of the wavefields at the wavefronts
For very slow processes the frequencies
w >> w o / P and f(z)
f(0) = 1 and the drag force is proportional to the current
value of df)
- JS)
50.
Biot's theory of saturated porous media 51 accounts for an inertial coupling
between partial momentum balances (virtual mass effect). It is related to the
added mass effect in the motion of a solid object in a viscous fluid. The virtual
mass effect can be accomodated by adding a non-dissipative term
4915747.

82762t83.

68983.

to the drag forces, with A(">b)

symmetric and positive definite. For compatibility
with the Eulerian representation assumed in this paper the time derivative represented by the dot can be identified with the time derivative at a fixed point. We
then have

and dissipativity of the drag forces is preserved.

4

Isothermal capillary hysteresis

The entropy production inequality implies that

[p(., - p(w) - P C ( S.,. .)]

s50

(54)

In equilibrium p(") - p(w) = Pc and Pc 2 0 on physical grounds. The excess

pressure difference p := p(") - p(w)- P" drives fluid displacement represented by
the variation of saturation. Eq. (54) determines the direction of the process and,
as usual, is considered as a constraint on the constitutive relation linking excess
pressure to saturation. The current saturation depends on the past history of the
excess pressure and the dependence involves hysteresis.
Several physical mechanisms lead to hysteresis in the fluid flow in a multi-phase
porous medium. The difference between advancing and receding contact angles in
a capillary tube is but one of them. In drainage and imbibition the order in which
different types of pores are filled or emptied is controlled by their diameters and
necks as well as by the relative wetting properties and viscosities of the invading
and resident fluids. The order of invading pores by a fluid is not the reverse of the

149

the emptying order. Trapped oil contributes to the dissymetry between the fluids
in oil-water systems.
If the evolution of the system is expressed in terms of two thermodynamically
conjugate variables, such as p , S, with one of the two variables returning to its initial
value and other variables kept at fixed values, then the other variable also returns
to its initial value. The area enclosed by the loop represents the energy loss during
the cycle. The energy loss tends to a finite value when the time rates the cycle tend
to zero. The relation between the two variables in the zero rate limit is represented
by a hysteresis operator
An additional rate dependent effect can be included
by adding a rate-dependent perturbation 20. Capillary hysteresis in porous media
involves three variables: excess pressure, saturation and relative permeability 39,40.
Vector hysteresis models 46 are needed for a complete description of the capillary
phenomena.
Hanyga l6 has shown that the Preisach hysteresis operator applied to as the
input variable and a rescaled saturation as the output variable satisfies inequality (54), while Hanyga and Seredyriska 21 have showed that the Preisach hysteresis
operator is compatible with experimental data of Morrow and Harris 44. An almost
identical model of hysteresis, known as the independent domain model 24,25,26 7 ha s
been used in modeling wetting and drying of soils
and in the adsorption
theory 32. The independent domain model allows a pore-scale justification of the
macroscopic model. In the independent domain model all the pores can be sorted
according to two critical values A, B of the excess pressure difference. The lower
critical pressure A is defined by the property that for p 5 A the pore is emptied
by the non-wetting fluid and filled by the wetting fluid, while for p 2 B the pore
is immediately invaded by the non-wetting fluid. This assumption is in agreement
with the well-known property of capillary tubes described in the book of Poynting
and Thompson 23: an increase in the pressure of one of the fluids does not cause a
displacement of the meniscus until the excess pressure exceeds a threshold. In the
intermediary period the curvature of the meniscus and the contact angle change
but the contact line does not move. In pore-scale models
the existence of
two critical pressure is associated with the pore bodies and pore throats.
The pore network of a porous medium contains pores with arbitrary critical
pressures in some range. Suppose that the initial saturation in a porous medium is
known and the excess pressure is increased/decreased. At each value of the excess
pressure the upper/lower critical pressure of a group of pores is exceeded, the
corresponding pores are filled/emptied by the non-wetting fluid and the saturation
jumps down/up. The normalized saturation y is defined by the formula
36737.

27,28t31

33734135

where SO,S1 denote the residual saturation and the maximal saturation of the nonwetting fluid, with 0 < SO< S1 < 1. A simple example of a hysteretic constitutive
equation satisfying the entropy production inequality is given by the variational
inequality:

150
where A 5 0 5 B are some constants. The entropy production inequality can be
checked by substituting w = 0 in eq. (57).
A realistic model of capillary hysteresis can be formulated in terms of the
Preisach hysteresis operator. According t o the entropy production inequality the
threshold values A, B for the excess pressure 3 satify the inequalities A 5 0 5 B ,
which is too restrictive for the Preisach model of hysteresis. In Ref.21 the input
variable x of the hysteresis operator is defined in terms of p in such a way that the
threshold values for 2 are restricted only by the inequality a 5 b. The Preisach
operator
represents a continuum of delayed relays (Fig. 1) connected in parallel. The output variable delayed relay has two states y = - l , l . Delayed relays are
parametrized by two real constants a 5 b. The action of a delayed relay is defined
by the following set of rules. If the input variable x < a (x > b) then y = -1
(y = 1). If y = -1, dz/dt > 0 and y = b then y jumps to the upper value 1. If
y = 1, dx/dt < 0 and y = a then y jumps to the lower value -1. The current value
y(t) of the output variable depends on its initial value at a time t o , on the prior
history of the input variable x and on the initial value y(t0) of y:
46736

y(t)=

w;Y(to),

t o , a, b ) ,

t 2 to

(58)
where d ( s ) := x ( t - s ) , 0 < s < t - to, is the history of the auxiliary variable x.
The orientation of the curve representing the delayed relay is uniquely determined
by the entropy production inequality. The saturation and the excess pressure are
thermodynamically conjugate, i.e. the signed area $ p d S 2 0 enclosed by a closed
loop in the (PIS)-plane represents the associated entropy production. A delayed
relay can be imagined t o represent a class of pores that absorb the wetting fluid
a t an excess pressure corresponding to x = b and release it at the excess pressure
correponding t o x = a. Suppose that the distribution of pores sorted according t o
the critical pressure values is given by a positive measure dp(a, b) with the support
in the half-plane a 5 b. The variation of normalized saturation associated with all
the pores is given by the Preisach operator
46936

Since the measure p is positive and the delayed relays satisfying the inequality
a 5 b satisfy the entropy production inequality, the Preisach operator has the same
property 1 6 .
In Ref. 21 it is demonstrated that all the experimentally established properties of
the scanning curves of capillary hysteresis in oil-water systems 44 can be explained
in terms of the Preisach hysteresis operator.
The Preisach operator, like several other hysteresis operators, has the wiping
out property: the current value of y depends only on the past extrema of x ( s ) ,s <
t , and if a new maximum/minimum of x ( s ) is reached then all the lower/higher
previous maxima/minima become irrelevant for the current value of y. Another
important property of the Preisach operator is the closure of minor loops if the
input variable returns to its initial value. In Ref. 21 it is shown that according
to the Preisach model the scanning drainage/imbibition curves always approaches
the primary drainage/imbibition curve as the saturation tends to the residual or
limiting value. If the measure p is not concentrated on the vertical and horizontal

15 1
Y

:b

:a

-1

Figure 1. The delayed relay operator

limiting lines of the ( a ,b ) plane, then the two curves meet in a cusp, as evidenced
in the results of Ref 4 4 . On the other hand a concentration of p on the diagonal
a = b ensures that the scanning curves do not start horizontally in the (z,p)-plane,
which would contradict the observations. The elements of the diagonal correspond
to the reversible delayed relays, which correspond t o the pores release and absorb
the non-wetting fluid at the same value of the excess pressure. The last remark
obviously applies to the delayed relays as well as to the Preisach model loops.
Notwithstanding its success, the independent domain model is an oversimplification of the imbibition and drainage in a pore network. In a porous medium
absorption/release of a fluid by a pore depends on the availability of the fluid in
neighboring connected and on the possibility of absorption of the released fluid by
the pore network. On the experimental side some occasional anomalies reported in
soil drying and in desorption processes are inconsistent with the independent domain model. In order to accomodate such effects the measure p is sometimes allowed
t o depend on saturation and excess pressure (the dependent domain model
or
the nonlinear Preisach operator 46).
29132

Thermal effects in immiscible flows

In hydrocarbon reservoir flows interphase heat transfer is usually neglected. The

Fourier heat flow in the solid skeleton is taken into account but heat flow in the
fluid phases is often neglected in comparison with advective energy transfer. It
is however clear that interphase heat exchanges and heat flow should play an important role in some cases of practical interest, for example in steam flooding,

152

which exploits the effect of heightened temperature on oil viscosity. In multi-phase

porous media significant thermal dispersion asscociated with random variations of
the fluid velocities induces a coupling between the effective heat flow and relative
fluid velocities 5 8 .
The effects of thermal dissipation on wave propagation in saturated media and
the additional waves associated with thermal variables were studied by Pecker &
Deresiewicz 69. Thermodynamic theory of thermal effects in porous media was
later developed for linear poroelastic theory of saturated media in Refs 71,70. The
concept of temperature averaging and phase temperature was discussed at length in
a series of papers by Whitaker, Quintard et al.
and in the book of Kaviany 58.
An in-depth study of thermal conductivity, thermal dispersion and interphase heat
exchanges in multi-phase porous media can be found in Refs
In
contrast to these references we shall develop a more general approach based on
continuum thermodynamics of generalized mixtures. This approach facilitates a
thermodynamically consistent analysis of the heat transfer model.
In order to derive a sufficient number of equations to allow the determination of
temperatures of various phases it is necessary to consider separate energy balances
for all the phases. Consequently it is necessary to associate partial internal energies
and partial entropies with all the phases and account for interphase heat flow in
interphase energy exchange terms. The total interphase energy exchange consists
of an advective term associated with mass exchange, the power of drag forces and
thermal energy exchange driven by the temperature differences. The phase separation assumption need not however be satisfied because the partial energy of a phase
may depend on the states of the other phases, in particular on their deformations.
denote the partial specific internal energy and specific entropy of
Let da)and
phase a. A partial energy balance is assumed in the form (after using the partial
momentum balance to eliminate the kinetic energy)
72779

58,73,74175776780.

a [,(.I

(V("I2/2

at

&'"')I

+ div

(v("l2/2

v ( ~ )=
]

div ( T ( a~) ( " 1 ) + p(") (b(a)+ b(a).

+ q(a) + 8'") (60)

whence (see Appendix)

p(") Da&(a)+ div h(a)
- T ( a :)

= q ( a ) + -(a)
4

(61)

a(")

In eq. (61) q(a) denotes an external energy supply to phase a, while

is the
interphase energy exchange driven by the interphase temperature differences and
therefore determined by a constitutive equation. The energy conservation law requires that

For simplicity we assume that there is no interphase mass exchange. In this case
eq. (11) is satisfied by setting E") = - C , b"', where the sum extends over all the

153

fluids and eq. (62) assumes the form

The momentum exchanges between the solid and the fluids represent the drag
forces.
It would be sufficient to have a single entropy of the mixture in order to formulate
the Clausius-Duhem inequality, but we would then fail to construct the partial
free energies, which are convenient potentials for non-dissipative thermodynamic
variables. We therefore introduce partial specific entropies ~ ( ~ Another
1 .
payoff for
having multiple temperatures is the necessity of several heat flows h(a),one per
each phase. The Clausius-Duhem inequality can now be expressed in the form

where

denotes the partial specific free energy of phase a and

The Clausius-Duhem inequality imposes the following constraints on the constitutive relations l4

as well as the residual inequality

If the interphase transfers of thermal energy are independent of thermal gradients then
a

154

Two special cases will be considered. If the constitutive equations for the energy
transfers
do not involve the relative phase velocities, then energy conservation (63) implies that

c c
q(4 =

5;'"' .

=0

If the constitutive equation for the interphase heat transfer assumes the form

then I d a b )= K ( b a ensures
)
that the first equation (72) is satisfied. Furthermore

b= 1

provided K ( a b )2 0. The last inequality implies that thermal energy is transfered

from phase b to phase a if O(a) < O@).
The coefficients
:= Kyab)O(a) d b ) / A( A is the volumetric density of interface area), known as interfacial convection heat transfer coefficients, have been
measured for some saturated porous media 77. In terms of the pore dimension
scale d = (6 - q5)/A, the Prandtl number Pr = pcp/lcf for the fluid dynamic viscosity p , fluid heat capacity at constant pressure cp and Auid heat conductivity
Icf and the Reynolds number Re = p f v d / p (v = filtration velocity) the experimental values of the interfacial convection heat transfer coefficient lie on the curve
K ( ~ Td/lcfA
~)
= 2 1.1 ReO.6 Pr1l3 for both steady-state and transient regimes. The
~

curve fits the experimental data for 10

( R ~ ' . ~ P I - ' / ~ )5 lo5. An alternative

but similar empirical relation is proposed in Ref. 78. Consequently fast flows are
expected to generate significant interphase heat exchange.
In the second case the heat transfer into phase a is expressed in terms of a
reference temperature OR. The energy conservation (63)and eq. (71)imply that

We shall assume that

ca

# 0 and the thermal energy transfers satisfy (73)then

If
by the equation

and eq. (76)is equivalent to (74).

dR

can be defined

155

Eq. (73) is consistent with the special case of interphase heat exchange considered in Ref. 14. In this context the partial temperature t9(.) is defined by a local
thermal energy average.
Assuming that the term involving the saturation rate can be decoupled from
the remainder of the entropy production

must be satisfied. While our treatment of the non-isothermal case differs significantly from the previous work, eq. (78) has been obtained in several previous
papers, including Refs
6

Multi-phase porous flow

In hydrocarbon reservoir problems the equations of quasistatic multi-phase fluid

flow in a porous medium are usually derived from partial mass balances and the
Darcy law 1 3 . Alternatively, the equations representing the flow can be obtained
from the mass balances proposed in Sec. 2.2 by neglecting the inertial term and
setting the external volume force b equal to the gravitational force. This approach
provides a link between flow models and wave propagation models. The deformation
of the solid and fluid phases as well as thermal effects also neglected. Neglecting
the memory effects the drag force exerted by the solid on a fluid can be expressed
in the form
,(f)

= -K(f)E(f)

(79)

,(f)

:= ,(f) - u(4

(80)

where

Physical considerations lead to the assumption that the permeability tensor K(f)
can be factored into a geometric absolute permeability K and a fluid-dependent
scalar relative permeability kf . The relative permeability exhibits a hysteretic
dependence on saturation 1 3 . This effect could be incorporated by considering
couplings between saturation and drag forces.
Due t o a low capillary number Ca = v q / (v
~ is an average velocity, 17 is the
largest viscosity, T is the typical surface tension of the interface separating the
fluids), one expects that the capillary hysteresis should be significant in quasistatic flow problems, while low frequency content leads to a simplification of the
drag forces. Since y(f) = const, the fluid mass balances reduce to

The right-hand side can represent local phase transitions. In order to derive a
compositional flow model l 3 each phase should consist of several species and mass
balances should satisfy the condition of species conservation, possibly with an account of chemical interactions.

156

Fluid partial momentum balances reduce t o the Darcy law

where z is a unit vector pointing vertically downwards. Additional closure equation

is provided by the condition
n ( f ) = (b = const and the capillary relations. In
the two-fluid case f = n, w
p(") - p ( " ) - P ( S ) = 3c(S)

- ICS,

IC > 0

(83)

where 31. is a hysteresis operator and the second term on the right represents the
rate-dependent perturbation. 3c is an inverse of the Preisach operator introduced
previously. In Ref. 45 it is proved that the inverse of the Preisach operator is a
hysteresis operator. Formulation (83) allows introduction of rate-dependent effects
and is more convenient in numerical applications.
The unknown variables are the volume fractions, the pressures and the velocities.

7 Wave propagation multi-phase porous media

Equations for small-amplitude acoustic wave propagation are obtained essentially
from the momentum balances upon linearization of stress and fluid pressures. In
Biot's theory the inertial effects are determined by a density matrix. In the formulation adopted here each equation initially involves the acceleration of a single
phase. The most important effect is a coupling between wave propagation and fluid
displacement. Consequently we shall neglect the thermal effects and focus on the
coupling of wave motion to saturation.
For a three-phase medium the linearized free energy will be expressed in the
form

(84)
where the index f ranges over the set {n, s}, yLf', So denote the reference values of
the variables y(f),S for linearization and D(ff') = D(f'f). The upper case letters
in eq. (84) refer to the Lagrangean variables. Let w, w ( f )denote the solid and fluid
displacement vectors, 0 = div w, O(f) = div w ( f ) ,3 := S - So. Substituting

in eqs (38,43) with the free energy (84) and F$ = b k wFK,the linearized constitutive relations for stress in the solid and for fluid pressures are

157

where emn := ( w ~ ,~ ~~ , ~ ) / 2 .
Eqs (87) demonstrate a coupling between wave motion and fluid displacement,
represented by saturation, with the saturation gradient acting like a force. The
origin of this term can be traced back to our assumption that the stress, like fluid
pressures, depends on the degree of compression of the fluids, expressed in terms
of y ( f ) , rather than on the fluid mass density per unit mixture volume p ( f ) . This
assumption is in agreement with the result of Drumheller and Bedford 8 5 . The
coupling to the saturation gradient does not appear in other theoretical models 64,65.
Our model allows for a dependence of wave velocities on saturation. Hysteresis
in the phase velocity-saturation relation has been found experimentally 41.
The memory effects introduced by drag forces appear in the form of integrodifferential operators acting on the solid and fluid displacements. Neglecting the
quadratic terms in phase velocities the equations of motion are

.(.)I

* a'"'- G(") * #w) = div [(I - 4)

p ( f )+ ( f+
) G(f)* #)
= -grad [,(f) p ( f ) ] + b ( f ) ,
,w = 4s; ,(n) = - S )
p(s) + - G(")

+ b(S)

(88)

f = n, s

(89)

(90)

where the dot over a symbol denotes the time derivative and G(f) := -G(f+) and
the eq. (83) is also needed for closure. Inertial couplings (a.k.a. virtual mass effects)
can be included following the closing remark of Sec. 3.
The memory effects appearing in eqs (88) can be derived from viscous relaxation
of the pore fluids. In poroacoustics
the effects of thermal relaxation are also
taken into account. Thermal relaxation manifests itself in the frequency dependence
of the elastic moduli in the stress-strain constitutive relations 52. In Ref 6 5 the
capillary relation is replaced by its equilibrium version.
54753

Conclusions

Multi-phase flow equations and seismic wave propagation in multi-phase porous

media can be derived from a single mathematical model by appropriate simplifications. The existence of a common framework ensures consistency of the reservoir
dynamics with its reaction to seismic signals. Consistency with Biot's theory of
saturated porous media can also be ensured provided local partial temperature
differences can be neglected.
For notational simplicity it has been assumed here that each phase consists of
one constituent. The compositional model of oil-water-gas flow can be derived by
assuming that some phases to consist of several components which they exchange

158

according to appropriate constitutive laws. Partial mass balances have to be specified for each constituent in each phase.
Acknowledgments
The research was carried out in the framework of the Petroforsk project 149238/431
"Mathematical Modeling of Seismic Wave Attenuation Based on Physical Models
of Mechanical Rock Behaviour", supported by a grant of the Norwegian Scientific
Council. A travel grant from Norsk Hydro A/S is also acknowledged.
Appendix

A generic balance of a specific partial quantity f ("1 is postulated in the integral

form

where dx denotes the volume, da is the area and n denotes the unit outwatd normal
on the boundary dV of a sufficiently regular set V . The function g(") denotes the flux
of f(") and h(")is the supply of f ("1 at the internal points of V . Taking successively
f(") = 1,d"),
(1/2) v("l2 E(") with the corresponding fluxes g ( a ) = 0, T("),h(")
and supplies p("),b(") + 6'")+ p(") ~ ( " 1 q(")
,
+?$"I
+ (b(")+ 6'"'). we get after
some simplifications

Da p(a) + p ( a ) div ,(a)

p(") Do

= div T(a)+ p(a)

= -P( a )

(b(")

(92)

L(")). v ( a )

(93)

For the energy we obtain the equation

)
div ( T ( av("))

+ p(") (b(4+ p). ,(a) + q(") + q(4 (94)

Using the partial mass balance to eliminate the second term on the left-hand side
we have

p(") D(") 2v(a)

1
2 + .(.I]

div (T(")

-div h(")+p(")

(b(")+ $"I) .v(")+q(")

+ q(4

(95)

Finally, using the partial momentum balance we get

p(") D O E ( " )

: L(") - div h(") f q(") + Ti(")

(96)

The task of constructing a balance for the total energy is more daunting 18. Let
v := C , c(") v(") ( barycentric velocity ), where c(") := p ( " ) / p , p := C , p("). and

159

Df := d f /at + v . V f . The energy conservation law assumes the form '*:

pD iv2

+ &]

= div

(T -

p(") u(")4 u(")

)T

-div h + x p(a)b(a)-v(a)+q
v]

(97)
where u(") := v(") - v . A more convenient reduced energy conservation law is
obtained by defining

and

and

viz.

References

1.
2.
3.
4.
5.

M. A. Biot, Int. J. Solids Struct. 13, 579 (1977).

F. Riewe, Phys. Rev. E 53, 1890 (1996).
0. P. Agrawal, J. Math. Anal. Appl. 272, 368 (2002).
A. Bedford and D. S. Drumheller, Arch. Ratl. Mech. Anal. 68, 37 (1978).
J. G. Berryman and L. Thigpen and R. C. Y. Chin, J. Acoust. SOC.Am. 84,
360 (1988).
6. L. Thigpen and J. G. Berryman, Int. J. Eng. sci. 23, 1203 (1985).
7. S. M. Hassanizadeh and W. G. Gray, Adv. Water Res. 13, 169 (1990).
8. S. M. Hassanizadeh and W. G. Gray, Water Resour. Res. 29, 3389 (1993).
9. W. G. Gray and S. M. Hassanizadeh, Adv. Water Res. 21, 261 (1998).
10. C. Wei and K. K. Muraleetharan, Int. J. Eng. Sci 40, 1833 (2002).
11. C. Wei, Static and Dynamic Behavior of Multiphase Porous Media: Governing Equations and Finite Element ImpZementation,PhD Thesis, University of
Oklahoma, 2001.
12. K. K. Muraleetharan and C. Wei, Int. J. Numer. Anal. Methods in Geomechanics 23, 1579 (1999).
13. M. B. Allen 111, A. Behie and J. A. Trangenstein, Multiphase Flow in Porous
Media, (Springer-Verlag, Berlin, 1988).
14. A. Hanyga and J. F. Lu, Thermal effects in two-fluid porous flow, to appear
in Int. J. Eng. Sci.
15. M. A. Goodman and S. C. Cowin, Arch. Ratl. Mech. Anal. 44, 249 (1972).

160

16. A. Hanyga, Two-fluid porous flow in a single-temperature approximation, to

appear in Int. J. Eng. Sci.
17. J. E. Killough, Trans. Am. Inst. Min. Metall. Petrol. Engng 261,37 (1976).
18. A. Hanyga, Modern Theory of Continuous Media, Polish Scientific Publishers,
2nd ed., Warszawa 1991 (in Polish).
19. Liu Shih-I, Arch. Ratl. Mech. Anal. 46,131 (1971).
20. A. Yu. Beliayev and S. M. Hassanizadeh, Transport in Porous Media 43,487
(2001).
21. A. Hanyga and M. Seredyriska, to appear in the Transport in Porous Media.
22. F. A. L. Dullien, Porous Media. Fluid Transport and Pore Structure, 2nd
edition, Academic Press Inc., San Diego, 1992.
23. J. H. Poynting and J. J. Thomson, A Textbook of Physics - Properties of
Matter (Griffin, London, 1902).
24. D. H. Everett, Trans. Faraday SOC.50, 1077 (1954).
25. D. H. Everett, Trans. Faraday SOC.51, 1551 (1955).
26. J. A. Enderby, Trans. Faraday SOC.51,835 (1955).
27. A. Poulovassilis, Soil Sci. 93,405 (1962).
28. A. Poulovassilis, Soil Sci. 109,5 (1970).
29. A. Poulovassilis and E. E. Childs, Soil Sci. 112,301 (1971).
30. A. Poulovassilis and E. Tzimas, Soil Sci. 120,327 (1975).
31. Y. Mualem, Water Resources Res. 9, 1324 (1973).
32. D. H. Everett, in The Solid-Gas Interface, E. A. Flood, ed., (Marcel Dekker
Inc, New York, 1967).
33. P. C. Reeves, The development of pore-scale network models for the simulation
of capillary pressure-saturation-interfacialarea-relative permeability relationships in multi-fluid porous media. PhD Thesis, Princeton, 1997.
34. P. C. Reeves and M. A. Celia, Water Resour. Res. 32,2345 (1996).
35. R. J. Held and M. A. Celia, Adv. Water Resources 24,325 (2001).
36. A. Visintin, DiBerentiaZ Models of Hysteresis (Springer, Berlin, 1994).
37. M. Brokate and J. Sprekels, Hysteresis and Phase Transitions (Springer-Verlag,
Berlin, 1998).
38. C. T. Miller, G. Christakos, P. T . Imhof, J. F. McBride, J. A. Pedit and J. A.
Trangenstein, Adv. Water Res. 21,77 (1998).
39. J. C. Parker and R. J. Lenhard, Water Resources Res. 23,2187 (1987).
40. R. J. Lenhard and J. C. Parker, Water Resources Res. 23,2197 (1987).
41. R. Knight and R. Nolen-Hoeksema, Geophys. Res. Lett. 17, 1529 (1990).
42. S. M. Hassanizadeh and W. G. Gray, Water Resources Res. 29,3389 (1993).
43. A. Y. Beliayev and R. J. Schotting, Comput. Geosciences 5,345 (2001).
44. N. R. Morrow and C. C. Harris, SOC.Petr. Eng. J. , (1965).
45. F. Bagagiolo and A. Visintin, Zeitschr. fur Anal. u. ihre Anwendungen 19,
977 (2000).
46. I. D. Mayergoyz, Mathematical Models of Hysteresis, (Springer, New York,
1991).
47. S. R. Pride, A. F. Gangi and F. Dale Martin, J. Acoust. SOC.Am. 92,3278
(1992).
48. S. R. Pride, A. F. Gangi and F. Dale Martin, Phys. Rev. B 47, 4964 (1993).

161

49. D. L. Johnson and J. Koplik and R. Dashen, J . Fluid Mech. 176, 379 (1987).
50. A. Hanyga, in Theoretical and Computational Acoustics '97, ed. Yu-Chiung
Teng et al., (World-Scientific, Singapore 1999).
51. M. A. Biot, J. Acoust. SOC.Am. 28, 168 (1956).
52. M. A. Biot, J. appl. Phys. 33, 1482 (1962).
53. D. K. Wilson, J. acoust. SOC.Am 94, 1136 (1992).
54. J. F. Allard, Propagation of Sound in Porous Media, (Elsevier, London, 1993).
55. A. N. Norris, J. Wave-Material Interaction 1, 365 (1986).
56. G. Gripenberg, S.-0. Londen and 0. Staffans, Volterra Integral and Functional
Equations, (Cambridge University Press, Cambridge, 1990).
57. M. Fabrizio and A. Morro, Mathematical Problems in Linear Viscoelasticity
(SIAM, Philadelphia, 1992).
58. M. Kaviany, Principles of Heat Transfer in Porous Media, 2nd ed. (Springer,
New York, 1999).
59. N. K. Tutu, T . Ginsberg and J. C. Chen, in Interfacial Transport Phenomena,
ASME, 37-44.
60. T. Sculenberg and U. Miiller, Int. J. Multiphase Flow 13, 87 (1987).
61. V. X. Tung and V. K. Dhir, Int. J . Multiphase Flow 14, 47 (1988).
62. W. Desch and R. Grimmer, Trans. Amer. Math. Soc. 314, 381 (1989).
63. K. Tuncay and M. Y. Corapcioglu, J . appl. Mech. 64, 313 (1996).
64. K. Tuncay and M. Y. Corapcioglu, J. Geophys. Res. 101, 25149 (1996).
65. J. E. Santos, J. M. Corber6 and J. Douglas, J . acoust. SOC. Am. 87, 1428
(1990).
66. W. Brutsaert, J. geophys. Res. 69, 243 (1964).
67. S. K. Garg and A. H. Nayfeh, J. appl. Phys. 60, 3045 (1986).
68. A. Hanyga and M. Seredyfiska, Quart. Appl. Math. LX, 213 (2002).
69. C. Pecker and H. Deresiewicz, Acta Mechanica 16, 45 (1973).
70. D. F. McTigue, J . Geophys. Res. 91B9, 9533 (1986).
71. V. V. Palciauskas and V. V. Domenico, Water Resources Res. 18, 281 (1982).
72. J. Hager and S. Whitaker, Transport in Porous Media 46, 19 (2002).
73. R. G. carbonelland S. Whitaker, Heat and mass transfer in porous media, in
"Fundamentals of Transport Phenomena in Porous Media", J. Bear and M. Y.
Corapcioglu eds, (Martinus Nijhoff, Dordrecht, 1984), pp. 121-128.
74. M. Quintard and S. Whitaker, Adv. Heat Transfer 23, 269 (1993).
75. M. Quintard, M. Kaviany and S. Whitaker, Adv. Water Res. 20, 77 (1997).
76. F. Zanotti and R. G. Carbonell, Chem. engng Sci. 39, 263 (1984).
77. N. Wakao and S. Kaguei, Heat and Mass Transfer in Packed Beds, Gordon and
Breach, Amsterdam, 1982.
78. S. Whitaker, Fundamental Principles of Heat Transfer (Krieger Publ. Co,
Malabar, 1983).
79. M. Quintard and S. Whitaker, Int. J. Heat Mass Transfer 38, 2779 (1995).
80. S.-H. Chang and J. Slattery, Transport in Porous Media 3, 515 (1988).
81. W.-C. Lo, G. Sposito and E. Majer, Adv. Water Resources 25, 1105 (2002).
82. M. Renardy and W. J . Hrusa and J . A. Nohel, Mathematical Problems in
Viscoelasticity, (Longman Scientific & Technical, Essex 1987).
83. A. Hanyga, Proc. R. SOC.London A 459, 2281 (2003).
84. M. A. Biot, J. Acoust. SOC.Am. 34, 1254 (1962).
85. D. S. Drumheller and A. Bedford, Arch. Rat. Mech. Anal. 73, 257 (1980).

162

BOTTOM REFLECTION PROPERTIES DEDUCED FROM AMBIENT NOISE:

SIMULATION AND EXPERIMENT
CHRIS H. HARRISON, AND ALBERT0 BALDACCI
SACLANT Undersea Research Centre, La Spezia, Italy
E-mail: harrisoizi~sciclantc.nato.int
Distributed noise sources at the sea surface (wind, rain, waves) and distant shipping provide an ideal
plane wave spectrum for investigating geo-acoustic properties. The vertical noise directionality, in
particular the up-down asymmetrical part, is closely related to the bottom properties and can be used
to extract the properties at each angle and frequency. The experimental technique is fairly well
developed having already detected differences between many sites with a moored VLA and spatial
variation with a drifting VLA. This paper investigates performance of the method numerically,
comparing and reconciling ray theory and wave theory approaches. It is shown that reflection loss
interference fringes can be resolved within a frequency-angle window defined by the design
frequency and the angle resolution of the VLA. The size of the window increases with number of
hydrophones, but this does not imply the need for a physically long array. The effects of
superimposing a distant coherent point source (a ship) on the sheet noise source (wind) are modelled
with a wave treatment (OASN), obtaining results qualitatively and quantitatively very similar to
experiment. It is shown that distant ships do not degrade results unless they are loud enough to
interfere with the sidelobes of the up and down steered beams. Tilt of the array seems to be
surprisingly unimportant for tilts less than about 10". This is partly because integration over azimuth
smudges the effect, and partly because the interference fringes are relatively large angle phenomena.
A point source at different azimuths does produce appreciable effects at low angles with array tilt,
however in the low angle region results are considered rather unreliable anyway. By using the rangedependent model RAM the sensitivity of the method to local bottom variability is investigated. It is
shown that the footprint of the VLA is comparable with the water depth, and therefore the technique
can be used for surveying with a drifting VLA.

Introduction

Earlier papers [ 1-41 introduced and demonstrated an experimental technique whereby

bottom reflection loss can be extracted directly from measurements of ambient noise
directionality. The process involves simply dividing the up-going sound by the downgoing sound at each angle, as measured by a vertical array, and in this respect it is not
really an inversion in the usual sense.
In this paper we are neither interested in inversion nor the processing per se, instead
we are interested in simulating the data and the process with a view to exploring the
performance and limitations of the method. Typically when attempting to understand
experimental results one is forced to speculate through lack of knowledge of
retrospectively important quantities. This is particularly so if the quantities are difficult or
expensive to measure and if the only purpose is to eliminate them kom a list of potential
sources of error. In this case it may be much easier to investigate the prospective effect by
numerical simulation.
The particular points that are addressed here are:
0 Foundation of the theory in terms of waves and rays
Effects of uncorrelated noise
Effects of nearby and distant point sources
Effects of strong sources received in sidelobes

163
Effects of array tilt
Finite array length effects
The size of the spatial average or footprint
This list is not intended to be comprehensive and other topics have been considered in
[ 1,4]. In Section 2 we review the simple theory, and then we take the above points in turn
and investigate them each as numerical test cases. The results show that the theory is
sound, the method is surprisingly robust, and that the measurement is local with a
reasonably well-defined footprint.
0

THEORY/BACKGROUND

Noise is measured as a discrete time series on, typically 32, equally spaced vertically
separated hydrophones. Frequency-domain beam-forming with some power averaging
then results in noise intensity vs angle and frequency. Dividing the up-going noise (downsteered beam response) by the down-going noise (up-steered beam response) we then
obtain a power reflection coefficient, and we can map this value fi-om the steer angle at
the array to the angle at the seabed. We can mimic this process in two ways. In the first
we calculate the plane wave response of a layered seabed by adding the coherent response
of all the layers [ 5 ] . We then calculate the noise directionality assuming a sheet of dipole
wind sources, possibly including distant ships as as a sheet of monopoles [ 6 ] .Knowing
the beam pattern B( 6, B ) of the array one can then calculate the beam response A for all
steer angles.
I2

A(B, w ) =

[ r , 2 ~ ( ~ i ,

w ) B(B,el, w ) cos el de

(1)

The derived reflection loss is the up-to-down-ratio which can then be compared with the
original plane wave reflection loss. The mapping from array angle at sound speed c, to
bottom angle at speed CB is simply
QE = acos(cE cos t
9 / c,)
In the second method we can mimic this process conveniently using the variant of
OASES [7] called OASN. This calculates the cross spectral density matrix C,(o)
(at a
set of frequencies a) for a given vertical array in the presence of a surface sheet source
and a stratified environment (water column and sediment layers). We then calculate the
steered beam response from

where the T,, are the steering delays and the a , , are the shading weights.
This estimate of the reflection loss can then be compared with the true reflection loss
as calculated by another variant of OASES called OASR. It calculates the plane wave
reflection coefficient for exactly the same seabed. Thus with both methods we have a selfconsistent numerical method of checking the performance of the proposed noise
inversion technique; the former is ray based and the latter is wave based.
3

Measurement bounds

Some limitations of the technique and resolution are discussed in [1,8]. From these
considerations one expects a frequency-angle window in which there are meaningful

164
results determined by two limits. The upper frequency limit is set by the onset of grating
lobes
fu = 2 f, l(1 -tsin 8)
(3 )
where f, = c12a is the design frequency; related to sound speed c and hydrophone
spacing a. The lower frequency limit is set by beam angle resolution, roughly
fL = ( 2f, I N)lsinB
(4)
where N is the number of hydrophones. Thus at normal incidence (8= 90) the viewable
frequency range is fu I fL = N 1 2 . The separation of fringes at normal incidence is
typically Af
c 12h for a layer of speed c and thickness h, so the thinnest detectable
layers have Af --< fu , i.e. h >- a. By comparing the angle resolution of the beam with the
characteristic angular width of the fringes one can show that angles are resolved in the
pattern provided h --< Ll2. However fine fringes may still be seen at normal incidence for
h > Ll2. In the following subsections we use the first, ray based, approach to investigate
these limits with various layer examples.

3. I

Examples with 16m array, 0.125 m hydrophone spacing, 128 elements

Clearly performance is optimised by having a large number of hydrophones N, but note

that it is not necessary to have a large physical aperture aN. In this example we have a
16m long array with 128 hydrophones and the environment is a low speed sediment layer
over a half space as detailed in Table 1.
12000

12000

10000

10000

E>.

8000

6000

4000

LL

2000

a,

2?

0
0

I
,
>.

8000

6000

a,

E
20

40

4000
2000

80

60

20

Angle (deg)

40

80

60

Angle (deg)
0

5
10
15
Reflection Loss (dB)

Freq= 5960 (Hz)

20

Angle= 70 (deg)
h

25

0
0
._

10

a,

a,

LT

!Y

20

40

60
Angle (deg)

ao

5
-5
0

5000

10000

Frequency (Hz)

Figure 1. Reflection loss with 16m array and 128 hydrophones; low speed example. Clockwise: true loss,
inferred loss, slice vs frequency (dashed line is true, solid is inferred), and slice vs angle.

165
Table 1. Low speed geoacoustic parameters

Thickness (m)

Sound speed

Specific
gravity
1
1.4
1.6

(ds)

Water col.
Sediment
Sub-bottom

1500
1490
1550

co

1
CQ

Attenuation
(dB/h)
0
.14
.14

The top left panel of Fig. 1 shows the unchanged calculated plane wave reflection
loss - true reflection loss. The top right panel shows the result of calculating the noise
directionality using this reflection loss, then beam forming and dividing up beams by
down - inferred reflection loss. One can immediately see the effect of the grating lobe
in the top half and the calculated limit is indicated by the dashed line. At the bottom edge
of the panel one can see the low frequency limit also marked by a calculated dashed line.
Almost all features of the original are retained. The lower two panels are slices through
the inferred reflection loss for constant frequency and constant angle respectively. These
show a more quantitative comparison; for instance, one can see that loss spikes are very
slightly smoothed over. The drop beyond about 80 is caused by proximity to the region
affected by the grating lobe.
12000
10000

8000

>.
0
c 6000
a,

4000

2000
0

Reflection Loss (dB)

65-

20

rn

4?

20
15

cn 15

8
P

4 10
c
.o
+ 5

$ 5
c

-I

c 10
c

0
a,
a,

a,

0
0

20

40
60
Angle (deg)

80

O
-5

5000

10000

Frequency (Hz)

Figure 2. Reflection loss with 16m array and 128 hydrophones; high speed example. Clockwise: true loss,
inferred loss, slice vs frequency (dashed line is true, solid is inferred), and slice vs angle.

A second high resolution case is shown in Fig.2 for the high speed bottom parameters
shown in Table 2.

166

Thickness (m)
Water col.
Sediment
Sub-bottom

Sound speed
(ds)
1500
1600
1700

1
m

Specific
gravity

Attenuation
(dB/h)

1
1.4
1.6

0
.14
.14

In the top right panel one can see that the calculated bounds (dashed lines) include an area
where loss is low (for grazing angles less than critical) but there is nevertheless a
reasonable representation of the true loss. Otherwise the results of the processing are
similar to the low speed case.

3000

3000
h

2
.2000
h

2000

K
a,

;1000
W

sP 1000

LL

li

20

40

80

60

Angle (deg)

Angle (deg)
0

5
10
15
Reflection Loss (dB)

Freq= 1490 (Hz)

20

Angle= 70 (deg)

c 10

5
0

20

40
60
Angle (deg)

80

-5

1000

2000

Frequency

3000

(Hz)

Figure 3 . Reflection loss with 16m array and 32 hydrophones; high speed example. Clockwise: true loss,
inferred loss, slice vs frequency (dashed line is true, solid is inferred), and slice vs angle.

3.2

Examples with 16m array, 0.5 m hydrophone spacing, 32 elements

The experimental array used in [l] had the same length (16m) but fewer hydrophones
than the earlier example (32). The effect on resolution can be seen for the high speed
example in Fig.3. The main point is that the bounds (dashed lines in the top right panel)
are closer together so that we see fewer fringes (note the change in the frequency axis).
Peaks can still be well detected in frequency and angle.
The final example in Fig. 4 shows the finer fringe strucure resulting from a thicker
layer, otherwise the same environmental parameters and array. Now both angle and
frequency resolution are beginning to degrade although the peaks can still be detected.

167

3000

3000
h

3
2000
>.

3
2000
>.

a,

0
C
a,

s!2 1000
0
0

1000

20

40
60
Angle (deg)

80

0
0

5
10
15
Reflection L.oss (dB)

Freq= 1490 (Hz)

20

40

60

80

Angle (deg)
0

"0

40

20

60

Angle (deg)

20

Angle= 70 (deg)

80

-5

'

1000

2000

Frequency

3000

(Hz)

Figure 4. Reflection loss with 16m array and 32 hydrophones; "low speed" example with 4m layer. Clockwise:
true loss, inferred loss, slice vs frequency (dashed line is true, solid is inferred), and slice vs angle.

4. I

Test Cases - Isovelocity Water

Addition of uncorrelated noise

From Eq (2) it can be seen that uncorrelated noise enhances the diagonal of C,, and
results in the same addition to beam response for all steer directions. Thus the up-to-down
ratio U/D becomes (U+a)/(D+a) and is contaminated by the addition. The solution will
tend towards unity (or reflection loss tends to 0 B).As will be seen later, a similar effect
can be had from the side lobes of a dominant noise source such as distant shipping.
4.2

Isovelocity: no absorption

Both analytical and numerical calculations fail when there are no losses at all because of a
two-dimensional version of Olbers' paradox [6].In short, a sheet source carries on
making contributions for ever-increasing ranges in a cylindrical environment if there are
no losses.
Nevertheless it would be desirable to run a case where there is bottom reflection loss
but no absorption in the water. This is difficult to achieve with OASN, but for completely
different reasons, as follows [9]. In OASES and OASN the solution depends on numerical
integration (actually an FFT) ,in wavenumber space in which the resonance of each loss-

168
fiee mode resembles a delta function. To get around the numerical integration problem,
absorption is deliberately introduced before the integration to widen these resonance
features. (Actually the integration is taken along a line slightly displaced fiom the real
axis.) The resulting function of range (in the propagation loss version, OASES) is
compensated for the artificial loss with an exponential growth. However in the noise
model (OASN) it is still possible to offset the line of integration from the real axis, but it
is not possible to compensate with an exponential growth because there is no particular
range assocoated with the noise source, it is a sheet. One option is to increase the FFT
resolution, but there are still practical problems. The alternative is to insert only
uncompensated real absorptions.
hovelocity with volume absorption - sheet source

4.3

4.3.1

Beam response

Figure 5 shows simulated beam response for a 32-element VLA using OASN with
isovelocity water and absorption following [lo]. Parameters are given in Table 3.

dB
60

55
50

45
40

35
30

Frequency (Hz)
Figure 5 Simulated beam response for a 32-element VLA in isovelocity water using OASN. Note that negative
angles are upward pointing beams that collect downgoing noise.
Table 3. OASN/OASR geoacoustic parameters

Thickness (m)

Sound speed
(ds)

Water col.
Sediment
Sub-bottom

130
1.3
a3

1517
1554
1643

Specific
gravity
1
1.88
1.88

Attenuation
(dB/h)
Ref( 10)
0.14
0.14

It is shown in [6] that a predominantly volume-absorbing environment results in

noise directionality being proportional to sin2(8). So Fig. 5 agrees with ray theory since,
even with significant bottom losses, at low enough angles and high enough fiequencies
the dominant loss in a ray cycle is volume absorption. The ray theory of [6] assumes
dipole sources; OASN takes finite depth point sources which act like dipoles if closer

169
than a fraction of a wavelength to the surface or monopoles if deeper. The source depths
in OASN were taken to be one third of a wavelength. In short, both ray and wave theory
predict a 'hole' in the middle which suggests that the distant sources seen experimentally
[l] are not dipoles. In fact a plausible explanation is that they are contributions from
many distant ships whose source depth may be several wavelengths. Typically in noise
studies one takes the source depth for a ship to be 5 metres [ l 11. However this does not
rule out the possibility of deep wind or bubble plume sources.
4.3.2

Reflection Loss

In reality, whether the distant ships or distant wind behave like a dipole or a monopole
makes no difference to this experimental technique; it simply relies on an input noise
directionality with sources on or near the sea surface. Likewise, absorption in the water
column is unimportant. Inserting the geoacoustic parameters of Table 3 into OASR we
obtain the 'ground truth' plane wave reflection loss vs angle and frequency (Fig. 6 ) .
2000
1800

dB
15

1600
h

1400

3 1200

10

5 1000
3
0-

800

LL

600

10

20

30

40

50

60

70

80

DO

Angle (deg)

Figure 6 'Ground truth' plane wave reflection loss from OASR using Table 1 parameters

Features to note in this plot are the main critical angle at 23" (corresponding to the
bottom half-space speed), interference fringes at steeper angles (with frequency
separation depending on layer thickness and depth of modulation depending on
impedance contrasts), and a weak fringe in the angle range between 23" and the sediment
layer critical angle (13").
The up-to-down beam ratio resulting from running OASN for a 16m array (32
hydrophones at 0.5 m spacing) is shown in Fig. 7. Comparing this with the 'ground truth'
(Fig. 6) we see the effects of the poorer angle resolution, and the results are very similar
to the experimental ones of [l]. The thin fringe and the peak loss near 20" are lost, or at
least, smoothed over. These features are the same as those modelled earlier with rays. In
contrast, if the array covered the entire water depth (as demonstrated by the same method
in [ 13 the result would be almost exactly the same as the ground truth. Although this is not
a practical arrangement it demonstrates that the experimental technique really does
measure the plane wave reflection coefficient.

170
2000
1800
15

1600
1400h

E 1200.
x

10

f 10003

!g 800LL

600.

200

400
0

10

20

30

40

50

60

70

80

'

90

u o

Angle (deg)

Figure 7 Up-to-down beam ratio resulting from running OASN for a 32 element array at 0.5 m separation

4.3.3

Isovelocity with volume absorption - distant point source

In this section we simulate sources that cannot be regarded as a spatially slowly varying
sheet, for instance, individual ships or groups of ships in no specific arrangement. A
problem with simulation is that the point source calculation needs to be coherent in order
to retain important propagation features (and some credibility), but this imposes fine
structure in frequency and angle. In this example with point source at 20 km the scale of
the structure is of order 10 Hz and 1".
In reality this would be smoothed out by temporal fluctuations in the propagation,
motion of the ship, finite length of the ship, and depth spread of the ship. It is clear that
the distant ship supplies noise within the critical angle but hardly anything outside.
Therefore we would not expect a complete reflection loss plot with ships alone although
the remaining low angle part ought to be correct. A more interesting result is given by
combining the ship with sheet noise. In Fig. 8 we see a running-frequency-averaged
version of the ship response (source level 150 dB re 1 pPa @ lm, bandwidth equals one
tenth of frequency) combined with the beam response for white sheet noise with a source
level of 65 dB re 1 pPa @ l m per m2. This bears a remarkable resemblance to the
experimental plots in [ 11.
Finally we plot the inferred reflection loss in Fig. 9. As we might expect, there are
chaotic features below 20" and at low frequencies where the ship tends to dominate. Of
course, the proportion of ship to wind is an important quantity for several reasons, the
most important of which is that upward and downward beams may be contaminated by
the sidelobe returns from strong horizontal shipping returns. This weakens the contrast of
the interference fringes. Otherwise the addition of distant shipping has no adverse effects.

171

dB
85
80

75
70

65
60
55
50

500

1000

1500

2000

Frequency (Hz)

Figure 8 Beam response from combination of distant point source and sheet noise source.

2200

15

2000
1800

;
;1600
$1400
1200

10

s 1000

800

600
400
200
0

10

20

30

40
50
Angle (deg)

60

70

80

90

Figure 9 Inferred reflection loss from combination of distant point source and sheet noise source

4.3.4

Isovelocity with volume absorption - nearby point source

For the reasons discussed above pertaining to coherence it is even more difficult to
simulate nearby sources. However our usual main reason for simulating is to be certain
that we have a well defined problem before interpreting the results. In the case of a nearby
source, documented experimental examples exist where we do know the position and
number of ships. Therefore we can use these in preference to simulations in this case.
Here we examine the case of a fishing boat within lOOm of the array followed within
minutes by a very large ship within 1 km or so. Both events were seen and logged on
radar, and the pictures shown below were available in real time. The full sequence is
assimilated most easily as a movie.

172

The main relevant effect of a nearby point source is to provide resolvable eigenrays.
Since generally these are placed asymmetrically about horizontal an instantaneous ratio of
up-to-down no longer works. Nevertheless a time average or a more sophisticated
treatment of the time-varying angles could still reveal the bottom loss. An alternative is to
deliberately degrade the resolution to force eigenray overlap.
vla2002112193833

vla2002112193833

-80
15

40

-40

$ -20
&

10

Ho
0,

2o

40

60
80
0

500

1000

1500

2000

2500

Frequency("ZJ

Moo

10

20

30

40
50
60
Bottom Ang e

70

80

90

Figure 10 Beam response (left) and inferred reflection loss (right) with large ship approaching. "Fingers"
correspond to eigenrays.

The approach of a large ship (close to CPA) is shown in Fig. 10. Five eigenrays can
clearly be seen in the beam response, and several are seen as artefacts in the reflection
loss plot. In this case the data are unusable for about 7 minutes on either side of CPA.
5

5. I

Test Cases - Array Tilt

Tilted array - sheet source

Using conventional beam-forming tilting the array has surprisingle little effect on the
beam response and the inferred reflection loss. A family of curves vs angle and frequency
is shown for tilts of the 16m array of 0", lo, 3", 5", lo", 15" in Fig. 11. Hardly any effect
at all is visible up to 5" in the reflection loss, except for a slight weakening in the intensity
contrast. Changes become more visible at lo", and then finally at 15" there are serious
changes. This is not surprising as the tilt is comparable with the lobe separation in the
beam response. Nevertheless the reflection still retains the main features of the vertical
case.
The reason for this lack of sensitivity is probably that the visual impression one gets
is largely from the high angle interference pattern, and the tilts are relatively small angles.
In addition the interference structure is mainly a function of frequency and changes rather
slowly with angle. Therefore any angle smoothing or smudging associated with the tilt
(remember that an upward tilt in one direction is a downward one in another) has very
little effect.

173
Angle = 60 deg

Frequency = 1200 Hz
12

Figure 11 Inferred reflection loss with sheet noise source for tilts of O", l", 3", 5", lo", 15"; slice in angle (left),
slice in frequency (right).

5.2

Tilted array - combinedpoint and sheet source

In these examples the array tilt is always 5" in the North-Southhertical plane, with the
bottom of the array shifted towards the N. As well as the sheet source there is now a point
source at various bearings but always 20 km from the receiver. Figure 12 shows reflection
loss for source to the N, E, S, W as plots vs angle and frequency.
Angle = 60 deg

Frequency = 1200 Hz

Norlh
East

- - - Souih
---.
west

500

Angle (deg)

1000

1500

DO

Frequency (Hz)

Figure 12 Inferred reflection loss on an array tilted at 5" (bottom towards N) with a sheet noise source and a
point source at 20 km to the North, East, South, and West.

There is no perceptible difference between E and W since the tilt lies in the tangent
plane to the source. The beam response for N and S show differences as one would
expect. Putting the source to the N results in the point source appearing to move
downwards in angle in the beam response (not shown). Conversely moving the source to
the S results in the source contribution moving up. The reflection loss algorithm interprets
this wrongly as increased bottom loss at low angles in the S case. Otherwise the
interference fringe part of the figures remains as good quality as ever.

174

6. I

Test Cases - Range-Dependence

Spatial ffootprint

An interesting question is, where is the reflector whose strength we are measuring? Is it
near the array, or is it some kind of average over great distances? After all, the theory,
whether ray or mode, postulates noise sources out to infinity whose contributions
combine to give the coherence we receive. On the other hand ray theory suggests that it is
the last bounce near the array that makes the up-down difference. A VLA drift experiment
was carried out in April 2002 [I21 and found convincing changes in bottom reflection
properties along a five mile path. To explore the theoretical possibilities we need a rangedependent model in which we can locally change the bottom properties. This is a tall
order because, at the same time we require fidelity at high angles and short ranges. This
rules out discrete spectrum normal mode programs such as C-SNAP[ 131, and it also rules
out many parabolic equation models. However the range-dependent PE model RAM [ 141
appears to be able to handle all three problems.
The hydrophones are modelled as sources, and RAMS marching-in-range solution
contains intensity at all depths including a fraction of a wavelength below the surface
where our noise sources are located. Thus by swapping sources and receivers (as allowed
by reciprocity) RAM provides the correlation matrix at the VLA from which we can
calculate the steered beams as usual.

t
E l

bottom

E2

Figure 13 Three-dimensional geometry for RAM numerical experiment showing incoming path bouncing either
in bottom environment E l (a disk of 2-layer sediment) or environment E2 (a I-layer sediment). By changing the
diameter of the E l disk one expects to deduce bottom properties of either E l or E2 if we believe the ray
hypothesis.

In this numerical experiment, as shown in Fig. 13, we create a variable radius circular
area of seabed beneath the array with three layers as before (environment El), outside
which the bottom is a simple half-space (environment E2), thus providing distinguishable
bottom properties (El creates interference fringes; E2 does not). We use the rangedependent model RAM with swapped sources and receivers to provide the correlation
matrix at the VLA.
Assuming that the ray interpretation is correct we expect to see a transition from E l
to E2 when the steer angle points directly at the edge of the E l disk. Of course diffraction
effects and the finite size of the array will blur this transition. By changing the diameter of

175

the El disk we can find the footprint of the array. This is the diameter where the result
is in clear transition between the entirely E l environment and the entirely E2
environment.

6.2

Isovelocity at twofrequencies

Figure 14 is a slice through reflection loss at one frequency (600 Hz). The two thick black
lines correspond to extremely large bottom disks of respectively environments El and E2;
E l can be recognised by its resonance peak, and E2 is a simple Rayleigh shape. In
between are coloured lines corresponding to different radius El environment, and the red
crosses on each line mark the steer angle to the disk edge, as seen from the centre of the
array. (As a cross-check for identifying the coloured lines and their red crosses, the
shortest radii go with the largest angles). If we inspect each line in turn we find that for
angles greater than that of the cross (angles nearer the vertical) the curve tends to the E l
thick line. Conversely for shallower angles the curve tends to the E2 thick line. This
demonstrates with a full wave model that the reflection loss we deduce really is that of the
area within about one fraction of a ray cycle from the array (order of magnitude, one
water depth). A corollary is that a drifting VLA could be used with this technique to
survey the

- 40m

i?

---- 50m
........... 60m

.................

-._._ 70m

u)
u)

- 80m

---.90m

, ,,, ,, , ,

CI

c)

_,_._

1OOm
160m
200m
El
E2

---.

rn

,................. -

............. .-

Angle (degrees)
Figure 14. Inferred bottom loss at 600 Hz (dash-dot-coded lines) for various sizes of bottom disk with El
properties (three-layer, giving resonance peak), outside which there is E2 properties (two-layer, standard
Rayleigh). Crosses indicate the angle for each coloured line at which we expect a rough transition from E2 to E l
(thick solid and dashed black lines).

Refracting cases can also be done by this method and some were included in [4].

176

Conclusions

Plane wave seabed reflection properties can be inferred from measurements of ambient
noise directionality by comparing the strengths of up- and down-steered beams on a VLA.
This paper investigated performance and limitations of the method by using numerical
simulation and paying particular attention to phenomena that were difficult to investigate
experimentally.
By using a simple ray-based layer model and simulating the effects of beam-forming
then forming the up-to-down-ratio it was possible to demonstrate the measurement
bounds. Examples were given for the case of a 16m array with either 128 or 32 equally
spaced hydrophones. The useable domain is strictly a function of frequency and angle, but
roughly there is a frequency window whose size is determined by number of
hydrophones; the more hydrophones, the clearer and more complete the interference
pattern, and therefore the more certain the geoacoustic interpretation (with or without
numerical inversion).
For the sake of credibility simulations of sheet noise sources were also done with the
wave model OASES (and OASR). Results demonstrate that with a long array (in
isovelocity water) the up-down-ratio (in dB) is, to all intents and purposes, identical to the
plane wave reflection loss. The similarity naturally degrades as the array becomes
smaller, but the main effect, as already stated, is to cut out low and high frequencies. The
wave treatment also shows that in the chosen frequency range the noise spatial coherence
is effectively independent of depth (Toeplitz) and the finite array size is not a problem.
Simulations of a sheet source combined with a point source were capable of
reproducing the marked near horizontal maximum in the noise field. One could also see
that variability of about 1dB in noise directionality is quite tolerable at large angles where
there are interference fringes but it results in spurious fluctuations in the low loss, low
angle region of inferred reflection loss.
Although nearby point sources could be modelled, experimental examples of noise
measurements with well documented nearby ships already exist. These showed that
unmodified processing leads to a degraded solution because of the resolvable eigenray
arrivals. Nevertheless slightly more sophisticated processing could counter this angular
effect although very loud sources are still a problem for beam-forming.
The effect of array tilt was investigated with sheet and point sources. Because of the
azimuthal symmetry of a sheet source, near horizontal arrivals, for example, are spread
into the up and down beams of the tilted VLA. However it is the nature of the interference
patterns to vary slowly in angle and rapidly in frequency, so beam angle spreading has
only a small effect on the pattern. For tilt angles up to about 10' (and conventional beamforming significant effects were only seen in the vicinity of the critical angle. By 15'
there were serious differences in level although the main features of the pattern still exist.
In the case of a distant point source the azimuthal direction of tilt becomes important.
This has a serious effect on the very low angle response but hardly any effect on the
steeper interference fringes for the obvious reason that the (distant) point source
contributes nothing to these angles.
The question of the size of the spatial 'footprint' was addressed by running the model
RAM with an environment in which the bottom properties of a circular area beneath the
VLA differed from elsewhere. By changing the size of this disk and recognising the
environment from the reflection loss character one could find the transition size (i.e.
footprint of the VLA) for each angle. These findings concur with spatial variations seen in
recent VLA drift experiments.

177

Under the conditions that there is a reasonable wind noise to ship noise intensity
ratio, none of these factors appear to cause any serious problems with the technique above
10" or so. Still the main problem is that distant shipping may provide a very strong near
horizontal noise contribution which (though not important per se) can contaminate the
steered beams through their side-lobes if winds are weak. The exact transition point has
not been investigated numerically because the threshold is easy to see experimentally
with a variable wind.

References
1. Harrison C. H., and Simons D. G., Geoacoustic Inversion of Ambient Noise: A
Simple Method, J. Acoust. Soc. Am., 112, (2002), 1377-1389.

2. Harrison C. H., and Simons D. G., Geoacoustic inversion of ambient noise: a simple
method, Conference on 'Acoustical Oceanography', Southampton, UK, 9-12 April
2001, Ed. T. G. Leighton, Proc. Inst.ofAcoust., 23, (2001), 91-98.
3. Harrison C. H., Bottom reflection properties by inversion of ambient noise, Proc.
ECUA, Gdansk, Poland, (June 2002).
4. Harrison C. H., and Baldacci A., Bottom reflection properties deduced from ambient
noise: simulation of a processing technique, SACLANT Memo Report SM-392,
November 2002, SACLANT Undersea Research Centre, La Spezia, Italy.
5. Jensen, F.B., Kuperman, W.A., Porter, M.B., and Schmidt,H., Computational ocean
acoustics, AIP, NY, 1994.
6. Harrison C. H., Formulas for ambient noise level and coherence, J. Acoust. SOC.Am.,
99, (1996), 2055-2066.
7. Schmidt H., OASES User's Guide and Reference Manual, Dept. of Ocean
Engineering, M.I.T., (1999).
8. Holland C.W., and Harrison C.H., Measurement of the Seabed Reflection Coefficient
in Shallow Water: A comparison of two techniques, Proc. 6" ICTCA, Hawaii,
(2003), This Volume.
9. Schmidt H., MIT, private communication.
10. Skretting A. and Leroy C. C., Sound attenuation between 200 Hz and 10 kHz, J
Acoust Soc Am, 49, (1970), 276-282.
11. Breeding J. E., Pflug L. A., Bradley M., Hebert M., Wooten M., RAND1 3.1 User's
Guide, Naval Research Laboratory Report No. NFUJMW7 17694-7552, (1994).
12. Harrison C.H., Geoacoustic inversion of ambient noise, J Acoust SOC Am, 112,
(2002), 2282.
13. Ferla M. C., Porter M. B. and Jensen F. B., C-SNAP: coupled SACLANTCEN
normal mode propagation loss model, SM-274, SACLANT Undersea Research
Centre, La Spezia, Italy, 1993.
14. Collins M. D., User's Guide for RAM Versions 1.0 and l.Op, NRL Report (1996).

178

MEASUREMENT OF THE SEABED REFLECTION COEFFICIENT IN SHALLOW

WATER: A COMPARISON OF TWO TECHNIQUES
CHARLES W. HOLLAND
The Pennsylvania State University Applied Research Lab, State College, PA 16804 USA
e-mail: holland-cw@psu.edu

CHRIS H. HARRISON
SACLANT Undersea Research Centre, La Spezia, Italy
e-mail: harrison @saclantc.nato.int
The seabed reflection coefficient is an important quantity required for the prediction of acoustic
propagation and reverberation in shallow water and bottom-limited regions of the worlds oceans.
The angular and frequency dependence of the reflectivity also contains significant information about
the seabed geoacoustic properties. Two reflection measurement techniques have been recently
developed that share the common feature of spatially averaging over a relatively small region (of
order the water depth). These local techniques have the ability to probe the spatial variability of
the seabed to much finer scales than traditional methods. One of the techniques employs a towed
broadband source and a single hydrophone receiver, the other technique employs ambient noise as
the source and a vertical array of receivers. Reflection measurements at nearly co-located shallow
water sites in the north Atlantic and the Straits of Sicily show good agreement despite completely
different experimental approaches. The characteristics and limitations of each technique, indicate that
they are mutually complementary.

Introduction

In continental shelf environments, acoustic energy that travels more than 1 water depth or
so generally interacts with the seabed and/or the sea surface. For frequencies of interest
in this paper (100-5000 Hz), the losses due to seabed reflection are generally larger than
at the sea surface hence are often the limiting environmental factor for shallow water
propagation. Thus, a key property that must be known in the prediction of propagation or
reverberation is the seabed reflection coefficient.
Traditionally, the reflection coefficient has been estimated from long-range (tens of
km) propagation or reverberation measurements (e.g., see [l-31 as a small sample of a
large body of literature). The advantage of the long-range methods is that large areas can
be covered rapidly. However, the disadvantage is that the estimates of the reflection
coefficients are spatially averaged over many kilometers so that probing seabed
variability is difficult. Also the long-range methods suffer from oceanographic variability
that add significant uncertainty to the estimates (e.g., [4]).
Recently, there has been a burgeoning interest in single bounce interactions with the
seabed. Single interaction measurements have the advantage of avoiding large-scale
spatial averaging and mitigating the unwanted effects of ocean variability. In this paper,
two shallow water seabed reflection measurement techniques are reviewed. One
technique [5-81 employs single bounce reflection measurements from a broadband source
(referred to henceforth as BB); the other [9-121 deduces reflection loss from ambient
noise (henceforth referred to as AN). Comparison of the assumptions and limitations of
the techniques indicates that they can be employed in complementary situations.

179
Reflection coefficients can be employed as inputs to some acoustic models (e.g.,
[13]) or used to obtain estimates of the sediment geoacoustic properties [e.g., 5,9] for
predicting propagation or aid in analysis of single interaction scattering measurements
[6]. The focus of this paper is to compare the reflection coefficient itself and the
underlying assumptions in the two measurement techniques.
2

Broadband Source and single hydrophone receiver (BB)

The geometry of the BB measurement technique (see Fig. la) employs a fixed single
receiver and a moving broadband source. The pulse width is 0.1 ms for a usable
bandwidth of about 10 kHz (see Figure lb). The source is towed nominally at 4 knots and
the ping rate is 1 pulse per second leading to an angular resolution that is highest at low
grazing angles. At angles below 30", the angular resolution is typically better than 0.5"
(see Figure 2a).
The magnitude of the pressure reflection coefficient can be defined as:

where p r is the received pressure from the bottom reflected path at range x and p o is the
received level of a bottom reflected path at x for a perfectly reflecting halfspace (see
Figure 1 for measurement geometry). The data were processed with the following
equation [8]

where "/o is the transmission factor from source to receiver along the bottom reflecting
path including the effects of spreading, refraction, and absorption, the surface multi-path
and bottom reflection (which for a perfectly reflecting bottom is unity), yd is the
transmission factor from source to receiver along the direct path including the effects of
spreading, refraction, and absorption, and multi-path, and q d is the amplitude of the direct
path at height h. The finite range interval of the measurements (-2 m) means that in
practice q d is interpolated rather than measured directly.

0.8

0.6

l i m e (ms)

Figure 1. a) Geometry of BB reflection technique, b) source pulse at normal incidence

180

15

30

45
9

60

75

90

(O)

."

15

30

45

e,

60

75

90

Figure 2. Reflection angles for a typical experiment a) angle distribution, b) standard deviation of angle
estimate.

2. I

Error Analysis

Error analysis of Eq ( 2 ) shows that the dominant contribution to the error is the variability
of the source amplitude due to non-constant drag forces on the source plate and
catamaran. A typical standard deviation of the reflection loss (-20 loglolRI)is k0.5-1 dB
depending on frequency and experiment geometry. Angle resolution can be traded
against variance, e.g., averaging over 1" window can reduce the standard deviation by
about one half. Equations for the standard deviation associated with the angle estimates
are provided in [S]. Figure 2 shows a typical standard deviation associated with the angle
estimate.
The short pulse-length employed in this technique allows reflection coefficients to be
computed at individual layers [7] or layer groups [5] in addition to the traditional total
energy reflection coefficient, which includes all significant reflectedhefracted
contributions. Time windowing at various layers is advantageous when extracting
sediment geoacoustic information since it effectively increases the amount of independent
data. For data presented here, the reflection coefficients are total energy, commensurate
with those from the ambient noise method.

2.2

Spherical Wave Reflection

The fact that the technique uses a point source and receiver means that the measured
reflection is a spherical wave quantity, which differs from a plane-wave quantity in two
ways. First, the wavefront curvature: Brekovskikh [ 141 gives the saddle point
approximation for the reflection of a point source from a plane boundary separating two
half-spaces as:

V (8)= R (8)- i [R'(tl

)+ R'(t?)cot(8)]/ (2kr)

(3)

where R(B} is the plane wave reflection coefficient (' indicates derivative), k is the
acoustic wavenumber and r is the total length of the bottom reflected path. Inspection of
Eq(3) indicates that the effect of wavefront curvature is small when total path length of
the bottom reflection is large relative to a wavelength (for all BB experiments kr is of
O( 10' -lo5) and when the plane wave reflection coefficient is slowly varying in angle.

181
The second effect of a point source on the reflected field is that sub-bottom paths
arrive at different angles (higher angles) than the path from the water sediment interface,
which is used as the reference angle 8 (see Fig 3a). Since the reflection data are used to
extract information of the sediment geoacoustics, this is not an issue because the
inversion model is based on a spherical wave decomposition. How different the
measurement is from a plane-wave quantity (see Fig 3b) depends on many factors
including source, receiver height, range, sediment layer thickness, sound speed profile in
the water and the sediment. Frequency/absorption are also factors since the layer depths
that contribute to the total reflection become smaller as frequency increases. Figure 3c
shows the angular difference in the paths for various layer depths as a function of grazing
angle for typical geometry. The impact that the various path angles have on the final
result is also ultimately linked to the angular dependence of the reflectivity. In angular
regions where the reflection changes rapidly with angle (i.e., near a critical angle), the
departure from a plane wave result will be larger than regions where the reflection
changes slowly. In summary, the three factors that determine the similarity between the
spherical and plane wave reflection coefficient are: 1) kr (affects wavefront curvature), 2)
how rapidly R varies with angle (affects wavefront curvature), and 3) how thick the
sediments are relative to source-receiver height.

Effect of Sub-bottom layers

- - 2mdepth
- 5mdepih

4.

10m depth

OI

Grazing Angle (deg)

Figure 3. Reflection path cartoon for a) spherical reflection (BB), b) plane wave reflection (AN) and c) angle
difference between sub-bottom and water-sediment interface reflected paths for typical BB geometry.

In order to gain insight into the differences between a spherical and plane-wave
reflection coefficient, several theoretical predictions are provided. Figure 4 shows a
comparison of the full-wave solution of a point source above a plane-layered media (as
calculated by OASP module of OASES [15]) as compared to the plane-wave solution.
The environment consists of two layers above a halfspace (Table 1). OASES generated
synthetic time series at 30m intervals for a source and receiver at 128 m and 20 m above

182
the seabed respectively. These time series were then processed with the same software
employed for the measured data. The resulting reflection coefficients are almost identical
to the plane wave results (except at the lowest frequencies, where the integration time
used to process the data was less than a wavelength) since by the 3 conditions listed in the
preceding paragraph: kr >>1, R varies with relatively slowly with angle and the sediment
thickness is small compared to the source-receiver height.
When sub-bottom contributions from deeper structure in the sediment are important,
differences may become more pronounced. Figure 5 shows the spherical reflection
coefficient for the geoacoustic model of Table 2. There are differences at low angles and
low frequencies as compared with the plane wave result due to angular steepening as in
Fig 3a. The differences displayed between the plane wave and spherical reflection
coefficient simply indicate the reason why spherical reflection coefficient models are used
to interpret the BB data. The differences are also instructive since AN measures the plane
wave and not the spherical response of the sediment. The negative reflection loss for
spherical reflection (e.g. -20, Fig 5) are well-known and does not violate conservation of
energy.

201
15

6300 Hz

10000 HZ

--i

15anSk
30(deg)45

60

Figure 4. Comparison of the full-wave spherical reflection coefficient (blue dot), the saddle point
approximation Ekq ( 3 ) (red-dashed line) and the plane wave reflection coefficient (black line) for the geoacoustic
model in Table 1 . Above 1600 Hz, the saddle point approximation and the plane-wave results were identical and
are not shown. The reflection coefficients are averaged over a 1/3 octave band.
Table 1. Seabed geoacoustic model for reflection model comparisons (see Fig 4)

Thickness
(m)
0.8

Sound Speed

Density

1511

Attenuation
(dBlmkHz)
0

1475

0.092

1.35

W S )

WCC)

1.o

0.7

1800

0.28

-_

1600

0.3 13

1.8

183

Attenuation

Density

1500

(dB/m/kHz)
0

Wee)
1.o

2.5
3.0

1600

0.088

1.4

1700

0.088

1.6

_-

1800

0.083

1.6

Thickness
(m)

Sound Speed
(ds)

800 Hz

1wO Hz

1250 Hz

1800 Hz

1 5 3 0 4 5 8 0 7 5

(dw)

Figure 5. Comparison of the spherical reflection coefficients for a source/receiver depths of 127115111 (black
line), and the plane wave reflection coefficient (green line) for the geoacoustic model in Table 2. The results
are not frequency averaged in order to show the effect of angular steepening, i.e., the spherical reflection
coefficient peaks generally occur at lower angles.

The reflection process occurs over a region of the seabed near the specular point in the
plane of the source and receiver. That region is defined by the first Fresnel zone, which is
an ellipse centered on the specular. The size of the first Fresnei zone is a function of
frequency, the source/receiver geometry and the sound speed profile. For isovelocity
conditions and a typical geometry, the frequency and angle dependence is shown in Fig.
4. The Fresnel zone represents the physical area over which the main contribution to the
field at the receiver takes place.

183

-1001

'

100

200

300

400

500

800

700

Range (rn)

Figure 6. Fresnel zones at 100 Hz (dashed) 1000 Hz (solid) and 10 kHz (filled) for a source and receiver height
above the seabed 120 m and 20m respectively. The sound speed profile is isovelocity; grazing angles are shown.

Ambient noise and vertical receive array (VLA): AN technique

Generally the ambient noise from a sheet source, such as wind, or from distant sources
such as ships behaves locally as an angular spectrum of plane waves. At each angle the
ratio of upward going flux to downward going flux is exactly the seabed reflection
coefficient. Therefore a vertical array with appropriate angle resolution can measure the
plane wave reflection coefficient by comparing upward and downward beams. One can
construct more complicated arguments involving multipath rays [9] or wave theoretic
treatments to show that an annulus of seabed contributes at each angle, and so the
technique measures a single interaction with the seabed within a few water depths of the
array [lo]. Therefore a buoyed array can be used as a survey tool by allowing it to drift
[ 10,111. The cylindrical symmetry means that each angle incorporates spatial averaging
around the 360" annulus.
Strengths and weaknesses were discussed in [9] and sensitivity to various
phenomena, such as array tilt, was evaluated numerically in [12] and extended in [17].
The main strength is that with no more than beam-forming processing one can obtain
reflection loss versus angle and frequency, as shown in Fig 7. Frequency limits are set by
the onset of grating lobes in the beam pattern above the design frequency and the beam
resolution at low frequency. Clearly the upward and downward beams are
indistinguishable for elevation angles comparable with the beam resolution, but at low
frequencies this angle becomes so large that the up-down ratio tends to unity. Both these
limits are distinguished in Fig 7 as anomalously low loss regions below 200Hz and above
about 1500Hz.
Reflection loss is of the most direct interest in naval applications inside the critical
angle. However intensity ratios at low angles, even if angles are resolvable, are not very
reliable because they are so close to one. This technique is better at measuring the higher
angle interference fringes from which one can deduce layer separations, sound speeds and
densities [9].
The beam angle resolution At), for an array of length L at grazing angle 8 and
frequency f in water of sound speed c is AOB = c / ( L f cos8) . This can also be stated in
terms of the design frequencyf o and number of hydrophones N as AOB = 2f , l(Nf cos8) .
At the design frequency (1500Hz) of a 32 element array the horizontal beam width is
3.58". Interestingly, despite the aforementioned low frequency mechanism whereby the
up-down ratio tends to unity, the poor resolution does not blur pictures such as Fig 7. The
reason for this is that there is a relation between the beam resolution and the fringe
separation. The fringe spacing for a pair of horizontal boundaries separated by h is

185
AOF = { c / ( 2 h f cos8)](tan8, / tan8) where 8 is the ray angle in the layer. For most
angles where fringes exist (i.e. above the critical angle) AOB is proportional to A8F and
furthermore an array of length L can resolve the fringes from multiple layers provided 2h
< L. Poor angle resolution affects the measurement where the reflection coefficient varies
rapidly (e.g., close to the critical angle) where
may be significantly smaller than 8. The
upper frequency limit is fo (= 1500Hz) and the lower limit is set by A& - 1, i.e.
f = 2 f 0 / N , in this case f 100Hz. Thus layer thicknesses are detectable between W2,
(2h < L), and the hydrophone spacing a (2h > 2WN, h > a). Layers thicker than L12 can
also be detected as finely spaced fringes but only in the endfire beam. The effect of angle
smoothing by the beamforming can be seen in Fig 8.
Important features of the reflection loss include the critical angle, fringe spacing, and
the depth of modulation. From these one can obtain, almost by inspection 193,
respectively sound speed, layer thickness and density. However there are potential
systematic errors in the depth of modulation if there is a dominant directional sound
source. This may happen with distant ships and a weak wind, as follows. When forming
up and down beams, sidelobes pick up the strong horizontal contribution and swamp the
wind noise in the steered direction. Because up and down beams are affected in the same
way the ratio tends to unity and contrast is lost. A partial solution to this problem is
adaptive beam forming. Experimentally it is worthwhile keeping track of the wind speed
and the resulting ratio. The deduced reflection loss is reliable when it is independent of
wind speed variation. Ray analysis is simplified by assuming that the array is small
compared with the size of features in the sound speed profile. However the analysis using
OASES in [12] and [17] assumed a finite array size. The experimental reflection loss
measurements made in [9] appear to tolerate the assumption of a short (point) array. The
effect of refraction is simply to map the measured quantities into different angles
according to Snells law. There are three important sound speed depths: sea surface, the
array, and the seabed; sound speeds: cs, c,, cb. The sound speed ratio c,/c, determines
whether the angle range at which there are rays connecting the array and the surface
(essential to receive surface noise). The sound speed ratio cdc, determines the mapping of
bottom angle to angle seen at the array.

10

20

30

40

50

60

70

80

90

Bottom Angle

Figure 7. Example of reflection loss interference fringes vs. angle and frequency derived from the average of
about 5 minutes worth of ambient noise.

186

Comparison of the two approaches

In this section, the two reflection measurement techniques are compared, first pertaining
to measurement and processing considerations, second via simulations, and third via colocated measurements.

4.I

Measurement and processing considerations

While many of the differences and similarities in the techniques were discussed in
preceding sections and can be gleaned from the references [5-121, Table 3 provides a
summary with some additional details. The main differences between the measurements
are that BB is spherical wave measurement and integrates over the first Fresnel zone, AN
is a plane-wave measurement and integrates over the beamwidth in the vertical plane and
360" in azimuth.
Table 3. Comparison of reflection techniques, BB and AN

4.2

Reflection simulation

In order to show how the plane-wave and spherical reflection coefficients differ and also
how each measurement samples the reflection coefficient, we compare simulations of the
two techniques for the same environment. Narrow band cases are considered for

187

simplicity. Since both AN and BB use -50 Hz bandwidth the frequency averaging is
comparable between the two methods.
The plane wave result [ 161 (applicable to AN) is shown in Fig 8, corresponding to the
environment in Table 2. To estimate how the AN beam-forming affects the measurement,
one first calculates the noise directionality with the assumed bottom loss and reasonable
assumptions about surface losses and distant shipping [17]. This is then convolved with
the beam pattern and the up-to-down-ratio calculated. This solution then includes all the
effects of resolution, grating lobes, etc. The angular averaging of AN is a function of the
array characteristics; in Fig 8 the array characteristics are the same as that used in the
measurements (32 elements and 0.5 m spacing).
The spherical wave result (applicable to BB) is shown in Fig 9, also corresponding to
the environment in Table 4, with source and receiver are 127 and 15m above the seabed
respectively. The model (OASES corrected by the direct and bottom reflected path to
yield V ( 0 ) ) was run at a resolution of every 10 cm; the simulated BB corresponds to
samples every 2 m (which corresponds to typical tow speeds and pulse repetition rates).
For this environment the plane wave and spherical wave are clearly different, however,
what is important is the fidelity with which each technique approximates the theoretical
value since both techniques rely on the theoretical value in estimation of the geoacoustic
properties.
500 Hz

630 Hr

800 Hz

15

30 45
Angb (-1

80 75

Figure 8. Plane wave reflection predictions (red line) and AN measurement simulation (blue line) for a 32
element array with 0.5m inter-element spacing and the environment of Table 2.

188
500 Hr

800 Hr

'
I

1000 Hz

1250 Hr

I .

1 5 3 0 4 5 6 0 7 5
(deg)

Figure 9. Spherical wave reflection predictions (red line) and BB measurement simulation (blue line) for
environment in Table 2. source and receiver are 127 and 15m above the seabed.

4.3

Rejlection measurements

Seabed reflection measurements were conducted on the Malta Plateau (south of Sicily)
during 3 experiments: SCARAB98 (April 1998), Malta99 (May 1999), and Boundary02
(April 2002). Co-located BB and AN measurements exist at 3 locations (see Fig lo), from
west to east termed Site 10, Site 1, and Site 2.
The Malta Plateau can be divided into two physiographic regions: the Ragusa Ridge,
a lineated rocky outcrop forming a spine between Sicily and Malta (roughly defined by
depths shallower than -1 10 m) and the area west of the Ragusa Ridge. On the map, the
Ragusa ridge is visible just east of Site 10 (while the ridge does indeed connect Malta and
Sicily, high resolution bathymetric data were only available from about 36" 17' to 36" 28'
N). Nearly the entire area west of the ridge is covered with a high-porosity silty-clay
sediment whose thickness is a function of distance from its riverine source on Sicily [6].
Site 10 lies just west of the flanks of the Ragusa Ridge (see Fig lo). Four separate
BB measurements were made to explore the azimuth dependence of the seabed reflection.
Along two of the azimuths a critical angle of 22' is observed, while along the other two
azimuths, the critical angle is 26" (see Fig 11). The two azimuths with the higher critical
angle are the closest to the ridge. Our hypothesis is that coarser sediments (i.e., with
higher sound speed and therefore higher critical angle) close to the ridge are due to
erosion from the ridge flanks.
A comparison of BB and AN at this site is shown in Figure 12. The dashed and
dash-dotted lines are the AN for conventional and adaptive beam forming respectively.
The solid lines are BB for a single (red) and a 3 azimuth average (black). One of the
azimuths had a poor signal-to-noise ratio and was not used. In general, the measurements
look quite similar. The structure in BB at low angles is real and due to interference from

189

discrete layer packets. The differences between BB and AN at low angles are real and
largely due to the difference in spherical vs plane-wave measurements and difference in
angle resolution. At 1600 Hz at high angles, there is large discrepancy in reflection loss;
but the AN data above 45" are known to be spatially aliased (i.e., due to the finite aperture
of the array).

14.85

14.0

Longitude (E)

14.05

Figure 10. Location of BB (4red lines) and AN (diamond) measurements on the Malta Plateau at Site 10. The
measurements are -1 km from the flanks of the Ragusa Ridge which trends north-northeast. Inset shows the
regional setting with the Sicily to the North and Malta (Gozo) to the southwest. Measurement locations are
indicated by the squares from west to east Sites 10, 1, and 2. Water depths are in meters.
1800 Hr

a
.
-I
m

G d w Ansb (W

Figure 11. BB Reflection measurements as a function of azimuth, the blue and cyan lines are closest to the ridge
("+ridge") and show a higher critical angle (i.e., sediment sound speed).

190

15

1250 Hz

1600 Hz
1

15

g10

m
5

0
0 15 30 45 60 75 90
Grazing Angle (deg)

0 15 30 45 80 75 90
Grazing Angle (deg)

0 15 30 45 80 75 90
Grazing Angle (deg)

Figure 12. Seabed reflection measurements near the Ragusa Ridge, BB single azimuth (solid red), BB averaged
over 3 azimuths (black), AN (blue dashed) and AN with adaptive beam forming (cyan chain dashed).

Conclusions

Two complementary methods for measuring local reflection coefficient were presented,
the BB method requires a broadband source but a simple single hydrophone receiver. AN
exploits ambient noise from wind as its noise source but needs a vertical line array
receiver. The most significant differences between the two techniques are that BB
measures a spherical wave reflection coefficient spatially averaged over a Fresnel zone
(typically a few degrees in azimuth at low angles), while AN measures the plane wave
reflection coefficient spatially averaged over vertical angles depending on the array
characteristics and over 360" in azimuth. Neither measurement is suitable for direct input
to a propagation model such as [13], since BB is a spherical coefficient (in general not
equal to the plane wave reflection coefficient) and AN measurements below the critical
angle are not reliable.
Measured reflection coefficients from the two methods are similar when the upper
few meters of seabed are important to the reflection process and when either the
environment is independent of azimuth or BB is spatially averaged.

Acknowledgements

The authors gratefully acknowledge the support of the NATO SACLANT Undersea
Research Centre and the Office of Naval Research.

191
References
1. Rubano, L.A., Acoustic propagation in shallow water over a low-velocity bottom J.
Acoust. Soc. Am. 67, 1608-1613, 1980.
2. Frisk G. V. and J. F. Lynch, Shallow water waveguide characterization using the
Hankel transform, 76, J. Acoust. Soc. Am., 205-216, 1984.
3. Collins M.D. and W. A. Kuperman, H. Schmidt , Nonlinear inversion of sediment
parameters in a range-dependent region of the Gulf of Mexico, J. Acoust. Soc. Am.,
92, p. 2770-2783, 1992.
4. Siderius M., Nielsen P.L., Sellschopp J., Snellen S. and D. Simons, Experimental
study of geo-acoustic inversion uncertainty due to ocean sound-speed fluctuations,
J. Acoust. SOC.,110,769-781,2001.
5. Holland, C.W., and J. Osler High resolution geoacoustic inversion in shallow water:
A joint time and frequency domain technique, J. Acoust. Soc. Am., 107, 1263-1279,
2000.
6. Holland, C.W., Shallow water coupled scattering and reflection measurements,
IEEE J. of Ocean Eng., 27,454-470,2002.
7. Holland, C.W., Geoacoustic inversion for fine-grained sediments, J. Acoust. Soc.
Am., 111, 1560-1564, 2002.
8. Holland, C.W., Seabed Reflection Measurement Uncertainty, JASA, in press.
9. Harrison, C.H., and Simons, D.G.,Geoacoustic inversion of ambient noise: a simple
method, J. Acoust. Soc. Am., 112, 1377-1389,2002.
10. Harrison, C.H., and Baldacci, A.,Bottom reflection properties by inversion of
ambient noise, Proc. 6 European Conference on Underwater Acoustics, ECUA,
Gdansk, Poland, June 2002.
11. Harrison, C.H., Geoacoustic inversion of ambient noise, J. Acoust. Soc. Am., 112,
No. 5 , Pt.2, Nov 2002.
12. Harrison, C.H., Bottom reflection properties deduced from ambient noise:
simulation of a processing technique, Saclant Undersea Research Centre
Memorandum No. SM-392, November 2002.
13. Weinberg H. and R.E. Keenan, Gaussian ray bundles for modeling high-frequency
propagation loss under shallow-water conditions, J. Acoust. Soc. Am., 100, 142 11431,1996.
14. Brekovskik, L., Waves in Layered Media, Academic Press, London, pg 232-246,
1980.
15. Schmidt H., OASES Users Guide and Reference Manual, Dept. of Ocean
Engineering, M.I.T., (1999).
16. Jensen, F.B., Kuperman, W.A., Porter, M.B., and Schmidt, H., Computational ocean
acoustics, AIP, NY, 1994.
17. Harrison, C.H., Formulas for ambient noise level and coherence, J. Acoust. Soc.
Am., 99,2055-2066,1996.

192
AZIMUTHAL LIMITATION IN THE PARABOLIC-EQUATION
APPROXIMATION FOR THREE-DIMENSIONAL UNDERWATER ACOUSTIC
PROPAGATION
LI-WEN HSIEH, YING-TSONG LIN, AND CHI-FANG CHEN
Department of Engineering Science and Ocean Engineering, National Taiwan University,
No. I , Sec. 4, Roosevelt Rd., Taipei, Taiwan 106, R.O.C.
E-mail: chifnnP@,ntu.edu.tw
The azimuthal limitation of the Parabolic-Equation (PE) approximation in solving the threedimensional (3D) wave equation in cylindrical coordinate ( I , 0, *) has been studied in this paper.
Typically, three-dimensional problems are dealt with an N by 2D (Nx2D) approximation, which
treats a 3D field as a fan-like composition of many two-dimensional vertical slices ( r-z plane in
cylindrical coordinate) regardless of the 8 -coupling terms. However, the Nx2D approximation can
only handle problems with weak 3D effects. To deal with problems possessing strong 3D effects, the
8 -coupling terms have to be considered in PE approximation. Nevertheless, the azimuthal
limitation in the 3D PE approximation is not defined as well as the vertical angle limitation in the 2D
PE approximation. Hence, the theoretical derivation estimating the azimuthal limitation is put forth
in this work.

Introduction

In the last decade, the 3D effects of underwater acoustic propagation have come to
attention of the acoustic community [ 11. Typically, 3D problems are dealt with an Nx2D
approximation, which treats a 3D field as a fan-like composition of many r-z plane
without the 6-coupling terms, thus only weak 3D effects that refer to the horizontal
refraction are considered. In order to deal with problems likely to possess strong 3D
effects, the 6 -coupling terms are required in PE approximation [2]. When the PE
approximation is introduced to solve the underwater acoustic wave propagation problem,
there are many approaches to apply the rational function approximations to the squareroot operator. The approximations pose limitations in the propagation angles. Twodimensional wide angle capability of PE in vertical direction has been discussed by Lee et
al., [3]. A comparison of angular limitations of some rational function approximations is
given in Ref. [4]. Phase errors in parabolic approximations were analyzed with normalmode theory [5] and also investigated by Tappert et al., [6]. However, these results only
considered the vertical propagation angle. In some 3D PE models, the approximations
include both vertical and azimuthal expansions, yet only the vertical wide angle capability
is taken in to account. Chen et al., [7] modified FOR3D [2, 81 to include wide-angle
capability in azimuthal direction. This model is used in this paper to discuss the
propagation angle limitation in azimuthal direction. Munk et al., [9] investigated the
refraction of sound by islands and seamounts using the concept of vertical mode and
horizontal ray. They clearly showed the refraction of mode-rays on horizontal plane, thus
in this paper we took this problem as our numerical test case.
This paper is organized as follows. We begin with the derivation of the modal
decomposition of Helmholtz equation and illustration of the propagation angles in
horizontal plane in Sec. 2. The narrow and wide angle PE approximations are presented in
Sec. 3 along with the phase error estimation that demonstrates the azimuthal propagation
angle limitation. Numerical simulations of propagation refracted by a seamount are

193
presented in Sec. 4 using Nx2D, 3D Narrow-, and Wide-Angle schemes. Concluding
remarks will be given in Sec. 5.
2

Modal Decomposition in Horizontal Plane

The 3D Helmholtz equation in cylindrical coordinates is

a 2 p 1 ap a 2 p 1 d 2 p
-+--+++--+kk,2n2(r,8,z)p
=0,
(2.1)
dr2 r dr dz
r2 ae2
where p ( r 7 8 , z ) is the spatial portion of the acoustic pressure field,
n ( r ,8,z ) = co1c ( r ,8,z ) is the index of refraction, co is a reference sound speed,

c ( r , 8 , z ) is the sound speed, and ko is the reference wave number such that

ko = 2~ f / c 0 ,with f being the source frequency. Since the physics on the horizontal
plane is of our interests in this study, we may neglect the z plane and consider it as a 2D
problem where the spatial variables are r and 8.Thus the Helmholtz equation written in
polar coordinates ( r ,8)can be deduced to the far-field wave equation of the form

1
urr+2ikour+ r z u , + k ~ [ n 2 ( r , 8 ) - l ] u = 0 ,

(2.2)

by letting

P ( r ,8 ) = 24 (y , e ) v ( r ),
where

is the zeroth-order Hankel function of the first kind for the outgoing wave representation.
Equation (2.2) may be solved by the method of separation of variables,

u =R(r)O(B).

(2.5)

By substituting Eq. (2.5) into Eq. (2.2), we can derive

r 2 -+2ikor
Rrr
R

-+ko(n
4
2
R

-1)r 2 =--,0,

which may be separated into two differential equations with

constant,

0,

i2 being the separation

+ i20= 0

(2.7)

and

i2
R, +2ikoR,+[kk,2(n2- l ) - l ] R
r
A simple solution can satisfy Eq. (2.7),
@ = p.

For Eq. (2.8), we can assume its solution of a simple form

= 0.

(2.8)

194

R = eim"

(2.10)

With this assumption, Eq. (2.8) can be rewritten as

K 1

m' +2k,m+ ki + y - k i n 2 =O,

(2.1 1)

1'
kir' .

(2.12)

whose solution is

rn = -ko k k0{

+ (n'

-1)

--

By means of Eqs. (2.4), (2.5), (2.9), and (2.10), we can write the solution of the
propagating wave as
(2.13)
From the above, we can also deduce the dispersion relation on the horizontal plane being

k 2 = kr2+ k j ,

(2.14)

k, = m + k,,

(2.15)

k, =I.

(2.16)

where the radial wavenumber is

and the azimuthal wavenumber is

Fig. 1. x - Y and Y - 8 coordinate systems on a horizontal plane.

Fig. 1 shows the relation between X - Y and r - 0 coordinates. The components
of wave number in r and directions, respectively, are

(2.17)
(2.18)

where is the bending angle of the acoustic ray, i.e., the propagation angle.
If we consider only the outgoing wave and rewrite Eq. (2.12) as

m = -ko
where

+ kOd-,

(2.19)

195

(2.20)
Taking the square of Eq. (2.19) results in

(in + ko) = k,(n2 - p),

(2.21)

kin = k = (in + k0) ikip.

(2.22)

or
From Eqs, (2.15), (2.16), (2.18), and (2.22), it yields

k
=-sin{=n.sin{,
k0

(2.23)

p = n2sin2 5.

(2.24)

or

PE Approximation and Error Estimation in Horizontal Plane

Equation (2.2) represents a two-way propagating wave and can be expressed in an

operator form

{-$ + 2ika a +
-

ar

1 a2
r 36 a

-- k2 [ n 2 (r,6)-1]

(3.1)

If only the outgoing wave is considered, it results in the operator form as

ur = ika(-l

+J

m ) u ,

(3.2)

where

x = (n -1)

(3.3)

and
(3.4)
There are several ways to expand the square root in Eq. (3.2) [4]. In the following we
will present the azimuthally narrow and wide angle expansions [2, 81.
In Eq. (3.2), considering azimuthally narrow angle, one can write

1
1
1
4 G E Y = 1+ -x
- -x2+ -Y + o(x3,Y).

(3.5)

Applying Eq. (3.5) to Eq. (2.19), we have

m+ko = ko{[l+-(n
1
2
In Eq. (3.2), when considering azimuthally wide angle, one can write

1
1
1
1
m = 1+ -x
- - x+ - Y -- Y + o(x,
Y ).

With Eq. (3.7), Eq. (2.19) can be written as

(3.7)

196

(3.8)
Next we follow McDaniel [5] by considering a single normal mode propagation in
the PE approximation and comparing the results with the solution of Helmholtz equation
(2.13).
We know fiom the solution of the Helmholtz equation (2.13) that the correct modal

phase is exp[ i ( m k , ) r ] ,or exp (ikrr) from Eq. (2.15). Also, from Eqs. (2.19) and
(2.24), the exact phase velocity relates to propagation
angles as
.~
~

m + k,

= k, = k,J1

+ (n'

- 1) - n2 sin2

5.

(3.9)

The same result also can be directly derived from Eq. (2.17)

k, = k.cos5 = k,nJ-.

(3.10)
Comparing Eq. (3.9) to the rational expansion of the square-root operator in PE
approximations reveals the inherent phase error. In the following, we will examine the
phase errors in the narrow angle and the wide angle approximations. Meanwhile, a case
referring to an Nx2D calculation in 3D problem is also included to present the effect
when the @-coupling is omitted. For simplifying the results, we choose k, = 1 and
consider the property of the square-root approximation

In'

1 << 1,

(3.11)

i.e., let n' = 1. Then the exact modal phase in Eq. (3.9) now reduces to

m + k , = k r =Jl-sin2{.
(3.12)
With Eq. (3.11) and k, = 1,the narrow angle approximation Eq. (3.6) is reduced to
m + k , =l--p=l--sin
1
1
2 5.
(3.13)
2
2
With Eq. (3.1 1) and k, = 1,the wide angle approximation Eq. (3.8) is reduced to
1
1
1
1
m + k, = 1- - p - -p2 = 1- -sin2 5 - - sin44.
(3.14)
2
8
2
8
In an Nx2D calculation, the @ -coupling is omitted, i.e., the operator Y is deduced to
zero in Eq. (3.4) and the square-root expansion. The phase is thus
(3.15)
With Eq. (3.1 1) and

ko = 1,the above equation is reduced to

m + k , = 1.

(3.16)
Comparing the phases for the various approximation Eq. (3.13), (3.14), and (3.16) to
the exact form Eq. (3.12), we plotted the absolute values of phase errors as a function of
propagation angle in Fig. 2. In the upper graph the acceptable error has been set to 0.002,
as indicated by the horizontal dashed line. The approximate angle limitations to Nx2D,
narrow- and wide-angle approximations can be seen as 3.5", 20.5" , and 33". In the
lower graph, where the vertical scale is expanded ten times, when the error tolerance is set

197

to 0.0002 which corresponds to TL difference about 1 dI3 [3], the limitations go to 1",
11.5" , and 22" . The angular limitations for both narrow-angle and wide-angle
approximations are about the same as those in vertical directions [3,4].

0.009
0.008

0.007

0.006 -

t
f

0.005 -

3
3

0.004-

3.5'
0.003

,
.I

I-

I
1
I

I
I
I

I
I

I,

--- -

I <

20
25
Angle 5 (deg)

30

35

40

I
I
I
I

I
I

I
1

0.9 -

I
I

0.8 -

0.7-

i
!

I
I
I
I
I
I
I
I

0.5-

0.3-

0.1

;I/

-di
Fig. 2.

I
I
I
I

I
I
1
1

I
I
I

:
;

1
1
I
I

3 0.4 -

.j

I
1

4-

I
I

i?
W 0.6-

lo3

'r

No 0 coupling
NarrowAngle
Wide Angle

__

I
I
I

10

,'

,
:

,;
'

I
I
I
I
I

:
I

_-

I
I
I

t
I

1
I
I

I
I . ^

I,

- "..

No 0 coupling

NarrowAngle

I
I

Wide Angle
I

Phase error vs. azimuthal propagation angle for no 8 -coupling, narrow angle, and wide angle
approximations. (a) Nominal beamwidths for acceptable phase error less than 0.01. (b) Nominal
beamwidths for acceptablephase error less than 0.001,

198
4

Numerical Results of Seamount Problem

In this section we consider the seamount problem [9]. All the following numerical results
have been obtained running FOR3D and its azimuthal wide angle version.
4.1

Description of the Test Case

The environment consists of a non-attenuating isovelocity water column over two

velocity varying layers, each is of IOOM thick, and an isovelocity half-space bottom. The
density in the first layer is 1.5 gkm3 whereas in the below it changes to 1.8 g/cm3. The
absorption coefficient in the first layer is a constant of 0.5 dB/ h whereas in the below it
becomes depth-dependent function by an additional rate at 0.05 dl31 h N. The seamount
is located in a waveguide with 700 M isobath. The summit of the seamount is 17.25 M
deep and the slope ,Ld is 1/10 or about 5.71" with respect to the horizon. The source is
placed 7 KM from the center of the seamount. The maximum range of computation is 18
KM. Details are referred to Fig. 3.

OM

700M
800M
900M
1OOOM

1950MfS

I
I

1.8glcm3

Fig. 3. Schematic drawing of three-dimensional penetrable seamount problem.

A modal starting field is placed at the center of the computation region, right in front
of the seamount, where the water depth is 700 M. Source frequency is 25 Hz. Mode 1 and
mode 5 are used as the starting field in two sub-cases. The simulation is carried out over
360 degrees. TL results and modal amplitude of Nx2D, 3D Narrow-angle, and 3D Wideangle solutions are both investigated and compared. In all calculations, we set A r = 6 M,
A z = 1 M,and A 8 = 1".

199

4.2

Transmission Loss and Comparisons

In Fig. 4, it shows TL computed by FOR3D wide-angle version and comparisons to

Nx2D and narrow-angle solutions at specific depth of 10 M. The left column in the figure
presents the results from mode 1 starting field whereas the right column presents the
results from mode 5 starting field.
For a 25 Hz source put in a waveguide with free surface and a rigid bottom, there is a
barrier depth [9] where the phase velocity is infinite and the waves cannot propagate
beyond this point. Under such circumstance, with reference sound speed being 1500 WS,
the critical depth will be 15 M and 135 M for mode 1 and mode 5, respectively. The
seamount in this problem generates several shadow bands at certain azimuths, also some
black holes at several locations, which can be seen in the first row of Fig. 4. From the
TL differences plots, the second and third rows of Fig. 4,we can see the most differences
occur at such locations and along the edge of such bands where the TL values become
much higher. In the middle row of Fig. 4, it is obviously that such differences arise like a
phase-shift in azimuthal direction, producing the 3 0 effects, or the 0 -coupling effects.
The differences between 3DWide-angle and Nx2D solutions can be up to more than 10
dE?.As for the differences between wide- and narrow-angle solutions, shown in the
bottom row, some differences up to 1 dE3 or so still tell such effects. However, the scale
seems to be much less, say, one order of magnitude on dl3 scale.
Comparing the results from mode 5 to mode 1 starting field, the shadow zone along
the crest line becomes clearer and wider, and the caustics of sound energy are repelled so
that the shadow bands are spread out. Furthermore, there is an envelop apparently
dividing the sound field into two phases that in the inner circle the sound caustics and the
alternate shadow bands are bending toward the seamount, whereas beyond the envelop
they are somewhat spread and forward in different azimuths. This is clearer in mode 5
results where the TL differences are also more widely seen indicating more serious
azimuthal coupling effects.
4.3

Modal Amplitudes and Comparisons

In Fig. 5 , it shows modal amplitude computed by FOR3D wide-angle version and

comparisons to Nx2D and narrow-angle solutions. The left column in the figure presents
the results from mode 1 starting field whereas the right column presents the results from
mode 5 starting field.
The depth of the summit of the seamount being 17.25 M (c.f. Fig. 3) is very close to
the critical depth for mode 1, therefore it can be seen the shadow zone for mode 1 starts to
show up at about 7 KM &om the source, the position of the summit of the seamount. For
mode 5, the shadow zone begins at about 6 KM (5822.5 M) and the radius corresponding
to the critical depth is about 1 KM (1 177.5 M) to the center of the seamount, thus the
shadow zone for mode 5 is wider than the area for mode 1.
From the second and third row showing the differences of modal amplitudes, we find
that most differences occur at the transition region, i.e., between the starting mode
dominant field and the shadow zone. For mode 5 such transition region becomes wider so
that differences can be found in broader area. The differences plots also indicate the 0 coupling effects as a phase-shift in azimuth direction.

200

130

120

110

1W

93

70

-0

8
10
Range (KMI

<s

12

3DW.- W D . TL wrl~k4mdcDm= 10M

10

8
6

2
0

2
4

5
5

'-

-0

10

12

11

16

18

10

RangelMIl

3DW. -3DN. TLWrflcMtB, dCC(h i10 M

1
16

08

14

06

04

12

02
10

-02
6
0 4
'

06

*
0-0
0

I
'
'
?

10

Range (KM)

12

14

18

0 8

>

J ,'I8

-0

0
10
Range (KM)

12

11

16

18

Fig. 4. From top to bottom, transmission loss of 3DWide-angle computation (horizontal slices at constant depth
10 M), TL difference between 3DWide-angle and Nx2D solutions, and TL difference between
3DWide-angle and 3DNarrow-angle solutions for Mode 1 starting field on the left column, and for
Mode 5 starting field on the right column.

20 1
3DW Mode 5

in

I8

16

07

12
06

05

10

S n
04
6

03
02

-0

10

12

11

16

01

n
"

n
-0

R a w IW

3DW - N S D Mode 1

05

in

Ok

16

8
10
Range (KMI

12

14

18

)OW. N-QDModcS
05
04

03
02
01

0
4 1

-02
0 3

4 5
Range (KM)

X10

3DW-10NMod~5

4
35

16

I4

25
12

2
15

g 10

t n

0 5
1

10

12

I4

16

R a w (KMI

Fig. 5. From top to bottom, modal amplitude of 3DWide-angle computation, modal amplitude difference
between 3DWide-angle and Nx2D solutions, and modal amplitude difference between 3DWide-angle
and 3DNarrow-angle solutions for Mode 1 starting field on the left column, and for Mode 5 starting
field on the right column.

202
5

Concluding Remarks

In this paper, the azimuthal angular limitation of PE approximation has been proposed for
3D one-way wave propagation. The angle limits exist in PE approximations of the squareroot operators Eqs. (3.5) and (3.7). The limits are about the same for both azimuthal and
vertical directions since the approximations are similar. From the plots of results it has
been found that the limits for Nx2D, 3DNarrow-, and Wide-angle approximations are
3.5, 2 0 5 , and 33, respectively, when the error is set as 0.002. The limits become lo,
11.5,and 22 when the tolerance is set as 0.0002. The numerical results of Mode 1 and
Mode 5 starting field are presented to show the azimuthal coupling effects. The
differences between wide- and narrow-angle can be manifested by making A0 smaller so
that the numerical scheme can catch the details [lo] and reveal the effects as expected.
Acknowledgement

The authors are very grateful to Dr. Ding Lees invaluable help and comments on 3D
propagation; also to Dr. M. Taroudakis and Dr. F. Jensen for their invitation to
ICTCA 2003 PE special session. This work is supported by National Science Council
of Republic of China.
References

1. A. Tolstoy, 3-D propagation issues and models, J. Comp. Acoust. 4 (3) (1996),
243-271.
2. D. Lee and M. H. Schultz, Numerical Ocean Acoustic Propagation in Three
Dimensions (World Scientific, Singapore, 1995).
3. D. Lee and S. T. McDaniel, Ocean Acoustic Propagation by Finite Difference
Methods (Pergamon Press, Oxford, 1988).
4. F. B. Jensen, W. A. Kuperman, M. B. Porter, and H. Schmidt, Computational Ocean
Acoustics (AIP Press, New York, 1994).
5. S. T. McDaniel, Propagation of normal mode in the parabolic approximation, J.
Acoust. SOC.Am., 57 (2) (1975), 307-3 1 1.
6. F. D. Tappert and M. G. Grown, Asymptotic phase errors in parabolic
approximations to the one-way Helmholtz equation, J. Acoust. SOC.Am., 99 (3)
(1996), 1405-1413.
7. C.-F. Chen, Y.-T. Lin, and D. Lee, A three-dimensional azimuthal wide-angle
model for the parabolic wave equation, J. Comp. Acoust., 7 (4) (1 999), 269-286.
8. G. Botseas, D. Lee, and D. King, FOR3D: A computer model for solving the LSS
three-dimensional wide angle wave equation, U.S. Naval Underwater Systems
Center TR No. 7943, 1987.
9. W. H. Mu& and F. Zachariasen, Refraction of Sound by Islands and Seamounts,J.
Atmosph. Ocean. Tech., 8 (1 99 l), 554-574.
10. F. Sturm and John A. Fawcett, On the use of higher-order azimuthal schemes in 3-D
PE modeling, J. Acoust. SOC.Am., 113 (6) (2003), 3134-3145.

203
APPLICATION OF FEM/BEM METHOD TO INTERIOR ACOUSTICAL FIELD
SHAPING BY HELMHOLTZ RESONATORS
BULENT IRFANOGLU, MEHMET CALISKAN
Dept. of Mech. Eng., Middle East Technical Universiv, Inonu Boulevard, 06531, Ankara, TURKEY
E-mail: bulentirfanogIu@yahoo.com, caliskan@metu.edu. tr
Finite element and direct boundary element method is used to analyze sound field within cavities.
Modification of cavity response due to its coupling to Helmholtz resonators is investigated. Sound
field driven by harmonic acoustical excitations in the low frequency range is considered. Influence of
bounding surface elasticity on the coupled system response due to sound structure interaction is
illustrated. Effectiveness of Helmholtz resonators for cases of coupling to either acoustical modes or
structural modes or both is explored. Furthermore, admittance boundary conditions in order to model
sound absorptive treatment and structural damping are taken into consideration as well.

Introduction

Modification of sound field within enclosures in lower audio frequency range is sought
for different reasons. Whether intention is to attenuate or to amplify certain frequencies,
acoustic resonators may be utilized. Specifically, if adequate, Helmholtz resonators are
popular means for such purposes because they do not require energy input for their
operation owing to their passive nature. Elasticity of bounding surface of the enclosure
may complicate the acoustic field due to its interaction with the interior sound field and
may also be viewed as another form of coupled oscillator.
Previous researchers had employed single andor multi-modal techniques to analyze
sound field modification within enclosures due to interaction with Helmholtz resonators
and elastic bounding surfaces [l-41. In this study, finite element and boundary element
method is used to explore acoustic field modification, by considering several sound
structure interaction case studies. Influence of dimensional, spatial variations of resonator
and elastic bounding surface with differing dissipative conditions, on interaction problem
is illustrated.
2

FEM/BEM Method and Implementation

In applying FEWBEM method to the coupled problem, acoustic field is modeled using
direct boundary element method and elastic structure is modeled by finite element
method. Matrix form of boundary element and finite element field equations are given in
Eq. (1) and Eq. (2), respectively.

Hp=Gv+q
Mu+Cu+Ku=f
where H is pressure influence matrix, G is velocity influence matrix, p is acoustic
pressure vector, v is acoustic particle velocity, q is excitation vector due to acoustic point
source, M is mass matrix, C is damping matrix, K is stiffness matrix, u is structural
displacement vector and f is load vector due to cavity acoustic pressure.

204

On structural surface where interaction of each medium takes place, continuity of

normal displacements and equilibrium of forces must be satisfied as given in Eqs. (3-4).

v=TDu
f=TFp

(3)
(4)

where TDand TFare coupling matrices for displacements and loads, respectively [5].
Assuming locally reacting surface conditions, absorptive boundary surfaces can be
modeled using expression given in Eq. (5) 163.

v=vS+Yp

(5)

where vs is the bounding surface normal velocity, Y is diagonal admittance matrix.

In case studies presented in next section, which involve Helmholtz resonators, multidomain boundary element formulation is utilized to model the acoustic field. Hence,
acoustic cavity and Helmholtz resonator are treated as separate domains. For each
domain, boundary element equations are obtained. Equations of each domain are coupled
using force equilibrium and displacement compatibility conditions at cavity and resonator
interface. Reduced form of resulting equations for the configuration shown in Fig. 1 is
given in Eq. (6).

where

Hll = [Hf: (HFl -GFl TDD-' TF)]

D=-m2 M + ( l + j q ) K
A

H
2
2

= H22 -G22

y22

where, in above equations, subscripts 11 and 22 are associated with first (enclosure) and
second (Helmholtz resonator) domain respectively and subscripts 12 and 21 are
associated with the interface between two domains. Superscripts R and E are associated
with rigid and elastic parts of the bounding surface in first domain, respectively, D is
dynamic stiffness matrix and ? is structural loss factor.
3

Case Studies

Case studies involve three main groups of parametric study for sound structure interaction
problems considered. These are, enclosure and Helmholtz resonator coupling, enclosure
and elastic structure coupling, and combined enclosure, Helmholtz resonator and elastic
structure coupling. Hence, combined study investigates the effects of the first and second
group, as illustrated in Fig. 1. Cavity dimensions are 1.36 m, 1.13 m, 0.97 m, in x-, y- and
z-axis directions, respectively. Acoustic medium is air (?=1.21 kg/m3, c=340 d s ) . In case

205
studies, effect of dimensional change of resonators and elastic plate structures, sound
absorptive treatment, structural damping and location effectiveness of resonators and
plate structures are investigated. In each case study, Helmholtz resonator and/or elastic
plates first flexural mode is tuned to 125 Hz, the first axial mode of rigid-walled cavity in
x-direction. Excitation of coupled system is by means of harmonic point source located
near origin (0.01, 0.01, 0.01) and sound pressure level response at an interior surface
point near another comer (1.3033, 1.13,0.060625) is obtained.
/Enclosure
Helmholtz
resonator

Elastic

Figure 1. Acoustic cavity coupled to Helmholtz resonator and elastic plate

The Helmholtz resonator consists of cubical resonator cavity, and a rectangular prism
neck having square cross section is symmetrically located with respect to the resonator
cavity. Elastic structure is a square brass plate (E=104 GPa, ?=0.37, ?=8500 kg/m3), with
assumed simply supported boundary conditions. Dimensional information of Helmholtz
resonators and plates used in case studies are given in Table 1.
Table 1. Helmholtz resonator and plate dimensions

Results and Discussion

Response of rigid wall cavity together with its modification due to resonators H1 to H5
are illustrated in Fig. 2 and modification of rigid wall cavity response due to resonators
H2, H6 and H7 is illustrated in Fig. 3. For these case studies, all resonators are centrally
located on the surface at x = 1.36 m. Influence of location for resonator H2 along x
direction is illustrated in Fig. 4. Assuming, linear dependency of admittance on frequency

206
93

70

60

50

v, 40
a

10

Frequency (Hz)

Figure 2. Enclosure and resonator coupling for different resonator volumes and neck lengths

Figure 3. Enclosure and resonator coupling for different resonator cross-sectional area and neck lengths

f, effects of sound absorptive treatment applied on to the neck interior surface of H2 for
admittance amplitudes of zero (no treatment), f/lOOOO and 2f/10000 are shown in Fig. 5
when the resonator is centrally located on the surface where x = 1.36 m. Modification of
rigid wall cavity response due to its coupling to elastic plate P1 located centrally on the
surface where x = 0, without damping and with different structural loss factors are
illustrated in Fig. 6. Modification of rigid wall cavity response due to its coupling to
elastic plates P1, P2 and P3 located centrally on the surface where x = 0 are illustrated in

207

Figure 4. Enclosure and resonator coupling for different resonator locations

Figure 5. Enclosure and resonator coupling for different absorptive treatment on resonator neck surface

Fig. 7. Influence of position of plate P1 along x direction when centrally located at (LJ2,
Ly/2, LJ, (2/3 L,Ly/2, LJ and (5/6 L,, LJ2,.LZ) is illustrated in Fig. 8 (L,=1.36, Ly=1.13,
L2=0.97). Coupled response of the cavity in the presence of Helmholtz resonator H2
located centrally on the surface at x = 1.36 m and elastic plate P1 located centrally on the
surface at x = zero with and without dissipative treatments is illustrated in Fig. 9.
Illustrated results of case studies are obtained by using constant rectangular boundary
elements and conforming four-node, twenty-four degree-of-freedom rectangular finite

208

10 -

Figure 6. Enclosure and elastic plate coupling with different structural loss factors

Figure 7. Enclosure and elastic plate coupling with different plate edge lengths

elements. Convergence of in-vacuo structural and rigid wall acoustic modes in frequency
range of interest is attained with 0.1% accuracy. BEM solutions introduce inherently
small amount of damping. In this study the equivalent damping ratio is calculated to be
around 0.0008 for enclosure mode at 125 Hz, using half power points.
Introduction of Helmholtz resonators may result two new, coupled frequencies for
which separation between them is proportional to square root of resonator volume I
enclosure volume ratio [ 11. Moderate to large attenuation (25 dB) of sound pressure level

209

Figure 8. Enclosure and elastic plate coupling for different plate locations

Figure 9. Enclosure coupling to both Helmholtz resonator and elastic plate

at original enclosure mode is attainable, while latter may amplify sound pressure level at
new, coupled frequencies. Increased level of absorptive treatment to the neck interior
further attenuates sound pressure level at coupled modes. With vanishing effect when
placed at a pressure node, resonator effectiveness increases as moved to pressure
antinodes. Introduction of plates may yield similarly, two new coupled frequencies for
which separation between these modes is proportional to square root of cavity effective

210
stiffness / plate mass ratio [3]. Increased level of structural damping yields moderate
amount of sound pressure level attenuation at all modes compared to very large amount of
attenuation at original enclosure mode without any reduction at coupled new modes in the
case of no structural damping. Spatial effectiveness is similar to the case of Helmholtz
resonator. Introducing Helmholtz resonators and plates together may yield at most three
new, coupled frequencies. In this combined configuration, individual effects of Helmholtz
resonator and plate reestablish with inter modal coupling to each other and enclosure
response.

Conclusions

Several aspects of shaping sound field within an enclosure due to its coupling to
Helmholtz resonators introduced as coupled spaces to an enclosure or simply supported
plates introduced as elastic boundary surfaces or both are illustrated. Observations are in
agreement to that of previous researchers related work using modal techniques [ 1-41.
Besides, in these previous studies Helmholtz resonators are treated as point sources
(sinks). However, in this study, effects of coupling, positional dependence and energy
dissipation due to Helmholtz resonators and plates are illustrated by means of FEMBEM
approach. Both single domain and multi domain BEM formulation are utilized, while the
latter yielded preferable results in case of complicated geometries. Parametric
investigation to sound structure interaction within enclosures is provided to illustrate the
importance and effectiveness of proper tuning. It is demonstrated that, if there exists a
fixed frequency excitation at a mode higher suppression of noise can be attained by
deploying a tuned elastic plate. In the case of variable frequency excitation a tuned
Helmholtz resonator to this mode poses better solution due to presence of two side peaks
in the plate case.
References

1 . Fahy F. J., Schofield C., A note on the interaction between a Helmholtz resonator and
an acoustic mode of an enclosure. J.S.V., 72 (1980) pp. 365-378.
2. Cummings A., The effects of a resonator array on the sound field in a cavity. J.S.V.,
154 (1992) pp. 25-44.
3. Dowel1 E.H., Gorman G.F., Smith D.A., Acoustoelasticity: general theory, acoustic
natural modes and forced response to sinusoidal excitation, including comparisons
with experiment. J.S.V., 52 (1977) pp. 519-542.
4. Gonenc-Sorguc A., Shaping of acoustical fields inside acoustoelastically coupled
enclosed spaces. (METU, Ph.D. Thesis, 1997).
5. Irfanoglu B., Caliskan M., Gonenc-Sorguc A., Analysis of structurally coupled
acoustic fields by acoustoelasticity and integrated FEM-BEM approaches. Proc. of
the gthInt. Congr. on Sound and Vibration, (Orlando, USA, 2002).
6. Suzuki S., Maruyama S., Ido H., Boundary element analysis of cavity noise problems
with complicated boundary conditions. J.S.V. 130 (1989), pp. 79-91.

211

BLIND OCEAN ACOUSTIC TOMOGRAPHY WITH SOURCE

SPECTRUM ESTIMATION
S.M. JESUS AND C. SOARES
SiPLAB-FCT, University of Algarve, Campus de Gambelas,
PT-8000-117 Faro, Portugal
E-mail: { sjesus, csoares} @ualg.pt
Classic acoustic tomography uses controlled sound sources to probe the ocean for
its physical properties. Instead, passive ocean acoustic tomography aims at using
natural noise sources, such as wind induced noise, wave noise, or shipping noise
with the scope of inverting for the ocean and/or bottom geophysical properties.
Most studies found in the literature make use of ambient noise and sea surface wind
generated noise, to invert bottom parameters in shallow water regions. Recently,
another approach used shipping noise as illuminating signals to invert for water
column parameters [Jesus et al., Conference on Acoustic Variability, Lerici (Italy),
September 20021. In that work, a focalization process was used t o simultaneous
invert known geometrical and unknown environmental parameters. In particular
it was shown that known geometrical parameters such as source range and depth,
and receiving array geometry, could be used as focus and out of focus indicators.
During the focus periods, estimated water column parameters favorably compared
to independently measured values. One of the difficulties found with the shipping noise was the low received power and the difficulty to determine a sufficient
number of stable frequencies. In the present work, the received signal is used to
deconvolve the source power, and thus obtain a full-spectrum weighting function
for optimum frequency combination during the focalization process. Results obtained in the same ship noise data set have shown an certain improvement where
a stable localization and inversion could be seen throughout the run.

Introduction

Ocean acoustic tomography (OAT) was introduced nearly 30 years ago by W. Munk
but, to date, it has not became a standard technique in oceanic observation. Although, synoptic ocean temperature profiling in near real time appears extremely
appealing for any oceanographer, OAT suffers from at least two main practical
drawbacks. One is that in order to have a true synoptic ocean observation, the
quantity of equipment in terms of source/receiver pairs is generally quite prohibitive
in terms of cost, deployment needs and management or, at least, comparable to
that required for classic oceanographic surveys. The other is that once the tomographic equipment has been installed the only information available (after inversion)
are space integral temperature profiles between the several source/receiver pairs.
Therefore, it is, in general, difficult to convince an oceanographer non adicted to
acoustics, of the interest of using OAT instead of usual ocean sampling techniques.
As opposed to classic (active) OAT, in passive tomography natural noise sources
in the ocean are used as input for inversion. Various forms of passive tomography
have been proposed in the literature among which those that attempt to invert
only for the bottom parameters in shallow regions making use of ambient noise
directionalityl , sea surface wind generated n0ise~3~
or even aircraft noise4. Recently,
passive OAT with shipping noise has been proposed also for estimating water column parameters5 or the full environment characterization6i7. Both studies provided

212
meaningful water column SSP results but in the former a priori knowledge for all
model parameters but the EOF coefficients was used while in the later a full-model
inversion was performed throughout the data run. Such an inverse problem, where
both the input signal and the propagation channel are unknown, is termed a blind
deconvolution problem, and is common in areas such as wireless communications,
geophysics and astronomy. This served as motivation to call the technique proposed
in6t7 as Blind Ocean Acoustic Tomography (BOAT). In BOAT, a focalization technique, similar to that proposed by Collinss for source localization, was implemented
to simultaneously invert known geometrical and unknown environmental parameters. In particular, it was shown that known geometrical parameters such as source
range, source depth, and receiving array geometry, could be used as in focus and
out of focus indicators, in order to preclude convergence to false inverse solutions
exhibiting high model fit values, the so called equivalent environmental models. Although the risk of converging to equivalent models always exists it was found that
during the in focus periods estimated water column parameters favorably compared
to independently measured values. This was particularly true when using assumed
unknown deterministic signals, in a first attempt6, while for the shipping noise in7
the low received power and the difficulty to determine enough stable frequencies
destroyed the result during several portions of the run. An attempt to improve the
situation in the case of the ship noise inversion is presented in this paper, where the
received signals are used to deconvolve the emitted source power, and thus obtain
a full-spectrum weighting function for optimum frequency combination during the
focalization process. This procedure avoids the requirement for an adhoc search of
source frequencies, thus making the process fully automatic and near optimal in the
sense of the minimum variance power spectrum estimation. The results obtained
in the same ship noise data set have shown some degree of improvement, where a
much more stable source localization and inversion could be seen throughout the
run than in the previous results.

Experimental setup and data description

The data set shown in this paper was recorded during the INTIFANTEOOsea trial
that took place in October 2000, near the town of Setubal, approximately 50 km
south from Lisbon, in Portugal. An overall description of the sea trial can be found
in9, whereas in this paper the interest will be focused only on Event 6. During this
Event the signals received at the 16-hydrophone vertical line array (VLA) consisted
on the noise radiated by the research vessel NRP D. Carlos I, cruising over a mildly
range-dependent area up to 3.3 km range from the VLA. The NRP D. Carlos I is
a 68 m overall length hydrographic ship with a gross displacement of 2800 tons.
Her propulsion is obtained from a double helical diesel-electric engine with a total
shaft power of 800 HP, attaining a maximum speed of 11 kn. It should be noted
that NRP D. Carlos I was originally built for acoustic surveying so she is supposed
to be a rather quiet ship. During Event 6, NRP D. Carlos I performed a triple
bow shaped pattern at approximate ranges of 1.2, 2.2 and 3.3 km from the VLA as
shown in figure 1.
A detailed bathymetry of the area was not available, but approximate bathy-

213
Depth (rn)

124

126

128

130

13

Militar Long. (km)

Figure 1. INTIFANTEOO sea trial Event 6 and site bathymetry. XBT casts locations are marked
with X and ULVA denotes the V L A location.

metric profiles were made along both the NW and the NE legs (see figure 1).
Therefore acoustic propagation between the ship and the VLA is assumed to be
slightly downslope range-dependent to the NE, and progressively becoming rangeindependent, at 120 m water depth, to the NW. The maximum range-dependence
is obtained for the 3.5 km range bow, with a maximum water depth difference of
20 m at the NE leg. A number of XBT casts were made during the sea trial a t
various times and locations as marked by the X signs on figure 1.
Ships speed and heading, as obtained from GPS, is shown in figure 2, plots (a)
and (b), respectively. It can be seen that mean ship speed was about 9 kn with
several abrupt drops to 7 kn during the sharp ship turns along the trajectory. As it
will be seen in the sequel, these speed drops will have a determinant impact in the
ship radiated noise in terms of power spectrum and frequency extent in the useful
band.
Acoustic signals received at the VLA exhibit a tipical exponentially decreasing
mean power spectra, characteristic of shipping noise with, however, a few strong
lines at 250, 260 and 359 Hz and a coloured noise spectra in the band 500 to 700.
On a preliminary analysis the maximum power frequency bins were extracted from
a 8 s sliding window that gave a time varying number of bins betweenll and 4 as
shown in7, and which is extremely poor for environmental inversion purposes in
such a challenging environment. A second attempt, presented in this paper, steems
from the idea that frequency information and source power may not be being used
correctly in this case where the source signal is highly fluctuating both in time and
frequency with a low signal-tenoise ratio (SNR), causing difficulties in selecting
the right useful frequencies.

214

14
S
-12

C l O-0
a, 8 a,

8 6.n
4r
20291.09

291.1

291.11

291.12

291.14

291.15

(b)

400,

0'
291.09

291.13

291.1

291.11

291.16

291.12

291.13

291.14

291.15

291.16

Julian date
Figure 2. Event 6: GPS measured ship speed (a) and ship heading (b).

Multichannel source spectrum deconvolution

Let us assume that the L-sensor array-received signal at frequency w can be written
aslo
Y(w1Qo) = a(w)h(w,& > s ( w )

+ u(w)

(1)

where y = [y(zl),y(z2>,. . . y ( z ~ ) and

] ~ similar definitions hold for h and u, the
replica model vector and the additive observation noise vector, respectively; s(w) is
the source spectrum at frequency w and 8, is a vector with the relevant parameters
under estimation. The noise process u is assumed to be uncorrelated from sensor
to sensor and with random factor a!. Note that random factor (Y = la[exp(jq5) is
supposed to represent the channel random variations on the emitted signal between
source and receiver and it is assumed space invariant but frequency dependent. For
the design of optimal estimators it is useful to consider that a is zero-mean and
Gaussian distributed.
Using data model (l),the broadband inchoerent Bartlett processor has the
following form''

where Q is the test parameter vector seeking the true value do, h(wk,Q)is the
replica model vector taken at frequency W k and for test parameter 0, c y y ( w k ) is
the data covariance matrix at frequency W k l K is the number of frequencies] H(8)

215

is a matrix formed with all replica channel vectors h along the main diagonal and
s(wk) is the source amplitude at frequency wk, and s is the source vector with entries
s(wk)a(wk)for all K frequency bins. Equation (2) is optimum if the noise is spatially
uncorrelated and the signal cross frequency correlations are zero. Another source
of non-optimality is that in practice, the source emitted power l s ( u k ) 1 2 is unknown
and a flat source spectrum is often assumed, leading to an equally weighted form
of (2) where ls(wk)I2 = 1. This is a suboptimum processor, that is as far from
the optimum case as the source power spectrum is non flat and the cross-frequency
data correlations are different from zero which is seldom the case in practice. It
is believed that this is the case for the data set of Event 6 where the ship noise
spectrum is highly time-varying and non flat in general. The source power estimates
.?(wk) where obtained with a Maximum Likelihood estimator (MLE) conditioned
on the environmental parameter 8, of the form

The basic idea is to reach a better approximation of the source spectrum in the
band used for inversion, thus relaxing the problem of frequency selection: if a low
source power frequency is selected at a given time, it will be correctly assigned
with a low source power estimate. Even in this case, computational limitations lead
to the necessity of reducing the frequency sampling, thus another criterion was used
to select the spectral components with smaller variance (in a given time frame).
This is based on the assumption that the frequency bins with higher variance are
more likely to contain only ambient and electronic noise. Thus if the received signal
variance is estimated by

t=l

where yyl(w, t) is the received signal on hydrophone k in time window snapshot t

at frequency w, T is the total number of time snapshots and py(w) is an estimate
of the mean of y1 in the same data window. The frequency components with the
lower variance are those that maximize the functional
L
(5)
where the summation is calculated over the L hydrophones. As an example, applying this criteria to a 16 s duration data window at julian time 291.125 gave the
results shown in figure 3: spectrogram in (a) and minimum variance selection in
(b).
4

4.1

Baseline model and data inversion

The baseline model

An important step towards a successful data inversion relies on the choice of a

suitable environmental model. There was no extensive oceanographic or geoacoustic

216

Figure 3. INTIFANTEOO sea trial, event 6, 16 s data window for hydrophone 8 at Julian time
291.125: spectrogram (a) and selected frequency bins for inversion wing the minimum variance
criterion (b).

survey concerning the area of Event 6. Therefore, as in a previous work6, identical

generic assumptions based on archival data were adopted, giving rise to a rangeindependent baseline model with a 119 m depth waveguide with a mean downward
refracting profile described by the two first EOFs drawn from a series of XBTs
taken at different locations in space and time (see X signs in figure 1). Geoacoustic
characteristics were empirically drawn from geological tables where the bottom was
formed by a fine sand 4 m thick sediment layer with a constant compressional
speed c,, a compressional attenuation a,=0.8 dB/X and a density ps=1.9 g/cm3
over a half space subbottom characterized by C b , (Yb = 0.8 dB/X and f b = 1.9
g/cm3. A focalization procedure carried out in the range independent area of the
NW leg gave the estimates c, = 1556 m/s and c b = 1657 m/s for the compressional
speeds of the sediment and bottom, respectively. These values were included in the
baseline model for inverting the data of Event 6.

4.2 Data inversion

Data inversion was carried by a Matched-Field based optimization procedure using
the C-SNAP forward modell and a Genetic Algorithm developed by Fassbender12.
The objective function is that given in (2) with the source amplitudes from (3). The
covariance matrices were computed from 16 s duration time observations that were
divided into segments of 1 s resulting in an average of 16 outer products at each
frequency. The search parameters, respective search intervals and discretization
steps are shown in table 1. Each inversion was carried out using three independent
populations of 80 individuals each and 40 generations. The crossover and mutation
probability were respectively set to 0.7 and 0.008. Each population was initialized
using the solution obtained in the previous inversion: 30% of the individuals were
initialized randomly in an interval with 10% of the search interval length with
center at the best individual obtained in the previous iteration. This modified
procedure lead to a faster convergence to the solution since the ship was moving

217

Symbol

Unit
m/s
m/s
km
m
m
rad

QIl

a2

sr
sd
rd

Search interval
min
max
20
-20
20
-20
3.5
0.5
10
1
95
85
-0.03
0.03

Steps
64
64

128
32
32
32

Table 1. Focalization parameters and search intervals: E O F l ( a l ) , E O F 2 (az), source range

(sr), source depth (sd), receiver depth (rd), VLA tilt (8)

in radial paths around the array and the environment was changing slowly over
time. The inverted parameters can be divided in three groups: in the first group

1
08

206

so
02

291 112 291 117

291 123 291 128

JulianTime

291 134

lo

291 112 291 117

291 123 291 128

JulianTirne

291 134

Figure 4. Focalization results for Event 6 - leading parameters with source amplitude estimation:
Bartlett power (a), source range (b)[the dashed lane i s the G P S measured source-receiver range]
and source depth (c); without source amplitude estimationrBartlett power (a), source range (b)[the
dashed line is the GPS measured source-receiver range] and source depth (c).

there are the leading geometric parameters such as source range and source depth;
the second group has the least important parameters, that were included in the
search to allow model adaptation to the data (focalization), such as array tilt,

218

sensor depth and other bottom parameters (held fixed during Event 6) and in the
last group there are the desired output parameters as the EOF coefficients ai,that
provide the final result. The results obtained for the leading geometric parameters
plus the Bartlett power are shown in figure 4 and the reconstructed temperature
profiles, with the estimated EOF coefficients, in figure 5. When comparing the

1 8
175

20

17
16.5

40

16

155

60

15

2i

145

80

14
135

100

..,
13

291.11

291.115

291.12 291.125
Julian Time

291.13

291.135

8 L . G .

18
17.5
17
16.5
16
15.5
15
14.5
14
13.5
13
291.11

291.115

291.12 2 9 1 . 1 2 5
Julian Time

291.13

291.135

12.5

Figure 5. Focalization results f o r Event 6 - output temperature profile estimate: with source
amplitude estimation (a) and without source amplitude estimation (b).

results obtained with source amplitude estimation (a) - (c) and those obtained
without amplitude estimation (d) - (f) the following differences can be noted: i)
no estimation drops exist anymore on the source-receiver range estimate as the
ship turns while the estimated curve closely follows the GPS source range curve
[plot (b)]; ii) the Bartlett power curve that, to some extent indicates model fit,
is more stable over time in case (a) than in case (d) with still a few low values
concentrated at the ship turn points; iii) these drops at the ship turning points are
accompanied with mis-estimated source depth in curve (c) where values drop well
below the expected estimate for a surface ship. So, a close inspection of these curves

219
can let us expect slightly better environmental estimate specialy after the larger
bow run, with only two exceptions: between times 291.119 and 291.122 and at
a short interruption at time 291.132. Indeed, these expectations are verified when
inspecting figure 5, in plot (a) using the source amplitude estimation procedure and
in plot (b), without source amplitude estimation as obtained in7, for comparison.

Conclusion

Ocean tomography with sound sources of opportunity has been a longely sought
dream for acoustic oceanographers. Preliminary results were obtained in7 using
a technique called BOAT, in a ship radiated noise data set collected off Portugal
during the INTIFANTEOO sea trial. As opposed to other passive inversion techniques, that use analytical or partial model inversions, BOAT uses a full model
parameter focalization, including environmental unknown as well as geometric partially known parameters. The coherence of the geometric parameter estimates was
used as indicator of the convergence of the inversion to a focused estimate of the
environment. It was found that this procedure performed relatively well in part of
the data set when the ship was steeming at full speed in a benign environment. At
low speed, during ship turns and when there was a significant mismatch between
the true and the assumed environment, there was a clear loss of convergence and
the estimates were erratic. These portions of the run were clearly indicated as out
of focus by the geometric parameter estimates. One of the problems associated
with these dropouts was a lack of frequency power support of the received ship
noise at the receiving array. This paper presented an alternative method for providing a coherent estimate of the source radiated power through a multichannel
deconvolution maximum likelihodd (ML) based approach. The results show that
inserting the ML estimated source power in the Bartlett power estimator, used as
objective function in the inversion search algorithm, clearly provided an adaptive
frequency weighting function that gave superior results than the previous method.
In particular, it was found that the dropouts in the source range estimates have
disapeared and the model fit is more stable than previously. It might be difficult
to see a net improvement on the final estimated temperature profiles due to the
still highly variable thermocline on a obviously too short run time of less than one
hour. Therefore the proposed technique can be seen as an improvement to BOAT
changing adhoc frequency selection into a fully automated procedure in presence
of low or highly frequency and/or amplitude dispersed power sources. Validation
of the BOAT technique, as a whole, requires longer runs of hours and days where
it might be able to evaluate longer term oceanographic variations trends, before a
final opinion can be formulated.

Acknowledgments
This work was supported by FCT, Portugal, under projects INTIMATE
(2/2.1/MAR/1698/95), ATOMS (PDCTM/P/MAR/15296/1999)
and TOMPACO. The authors are also in debt to the crew of NRP D. Carlos I of the Instituto
Hidrogriifico, Lisbon, Portugal that made the sea trial successful.

220

References
1. Carbone N.M., Deane G.B. and Buckingham M.J., Estimating the compressional and shear wave speeds of a shallow water seabed from the vertical coher-

ence of ambient noise in the water column, J. Acoust. SOC.America, 103(2),

801-813, (1998).
2. Harrison C.H. and Baldacci A., Bottom Reflection Properties by Inversion
of Ambient Noise, Proc. Sixth of European Conf. on Underwater Acoust.,
ECUA 02, pp. 471-476, Gdansk, Poland, June (2002).
3. Harrison C.H., The Influence of Noise and Coherence Fluctuations on a new
Geo-Acoustic Inversion Technique, Impact of Littoral Environment Variability
on Acoustic Predictions and Sonar Performance, ed. Pace and Jensen, pp.139146, Kluwer, September (2002).
4. Buckingham M.J. and Giddens E.M. and Pompa J.B. and Simonet F. and Hahn
T.R., A Light Aircraft As a Source of Sound for Performing Ge-Acoustic Inversion ofthe Sea Bed, Proc. of Sixth European Conf. on Underwater Acoust.,
ECUA02, pp. 465-470, Gdansk, Poland, June (2002).
5. De Marinis E., Gasparini O., Picco P., Jesus S.M., Crise A. and Salon S.,
l 1 Passive Ocean Acoustic Tomography: theory and experiment, Proc. of Sixth
European Conf. on Underwater Acoust., ECUA02, pp. 497-502, Gdansk,
Poland, June, (2002).
6. Jesus S.M., Coares C., Onofre J . and Picco P., Blind Ocean Acoustic Tomography: Experimental results on the INTIFANTEOO data set, to appear in Proc.
of European Conf. Underwater Acoust., pp.9-18, Gdansk, Poland, June (2002).
7. Jesus S.M., Soares C., Onofre J , Coelho E. and Picco P., Experimental testing
of the Blind Ocean Acoustic Tomography concept, Impact of Littoral Environment Variability on Acoustic Predictions and Sonar Performance, ed. Pace
and Jensen, pp.433-440, Kluwer, September (2002).
8. Collins M.D. and Kuperman W.A., Focalization: Environmental focusing and
source localization, J. Acoust. SOC.America, 90(3), 1410-1422, (1991).
9. Jesus S.M, Coelho E., Onofre J, Picco P, Soares C. and Lopes C., The INTIFANTEOO sea trial: preliminary source localization and ocean tomography
data analysis, Proc. of the MTS/IEEE Oceans 2001, Honolulu, Hawai, USA,
November (2001).
10. Soares C. and Jesus S.M., Broadband Matched-Field Processing: coherent and
incoherent approaches, Journal of the Acoustical Society of America, 113(5),
pp. 2587-2598, (2003).
11. Ferla C.M., Porter M.B. and Jensen F.B., C-SNAP: Coupled SACLANTCEN
normal mode propagation loss model, SACLANTCEN Undersea Research Center, Memorandum SM-274, La Spezia, Italy, (1993).
12. Fassbender T. , Erweiterte Genetische Algorithmen zur globalen Optimierung
multimodaler Funktionen, Diplomarbeit, Ruhr-Universitat Bochum, (1995)

22 1
Projection Operator Method for Solving Coupled-Mode Equations: An
Application to Separating Forward and Backward Scattered Acoustic
Fields
D. P. Knobles
Applied Research Laboratories, The University of Texas at Austin, P. 0. Box 8029,
Austin,
T X r 8 r i 3 - 8 0 ~ 9 ,USA
E-mail: knobles@arlut.utexas.edu
Examined is the concept of a nonlocal effective interaction applied to scattering problems in underwater acoustics. A projection operator approach previously introduced in
nuclear theory is employed t o place the basic coupled mode integral equations into an
alternate form. The dimension of the original coupled equations is reduced by half, to
the dimension of the forward propagating modes, by a channel elimination process. The
reduction in dimension occurs at the expense of the introduction of a nonlocal optical
potential that couples the forward and backward propagating subspaces. A perturbation
method leads t o a multiple scattering representation of the effective interaction for the
coupling. It is demonstrated that it is possible t o generate a consistent methodology t o
separate the forward and backward propagating fields and to form a forward propagating field that includes perturbative corrections because of the coupling of the backward
propagating field.

Introduction

Solving coupled-mode equations (or coupled-channel equations) presents a difficult

problem in computational physics. It is well known that the computation time t o
solve coupled-channel equations is proportional t o N 3 , the cube of the number of
channels [l].The interest in the coupled-channel problem is created largely because
wave phenomena (classical and quantum) within complex systems can generally
be described by solutions to inhomogeneous coupled integro-differential equations.
One method of solving coupled equations involves the transformation of the basic
coupled-channel equations into an auxiliary set of integral equations that have properties allowing the Lanczos method t o be employed for their solution [2,3]. Once
the solution for the auxiliary equations is obtained, the original solution is recovered from the transformation. Instead of N 3 , the computation time is proportional
t o N 2 K where K << N , and K is the number of Lanczos vectors required for
convergence. Even though this new approach has significantly improved the ability t o efficiently solve the large dimension coupled-channel problem, solving such
equations for typical physical problems remains difficult.
Linked t o the coupled-channel issue is the problem in mathematical physics of
how t o construct transformations t o the Helmholtz equation that separate forward
and backward scattering components of the field. For example, an approximate
approach was recently introduced that utilized a transformation to decouple forward and backward scattering components of an acoustic field for the case where
backward scattering was weak [4].The idea was t o apply the Foldy-Wouthuysen
transformation [5-71, well known in relativistic quantum mechanics to decouple the
particle and antiparticle states, to decouple term by term the forward and backward

222

scattering components of the acoustic field. The Foldy-Wouthuysen transformation

to the Helmholtz equation, carried out to first order within the weak coupling limit,
produced a correction term (called the Foldy term) t o the parabolic wave equation approximation [4]. Although the concept of applying the Foldy-Wouthuysen
transformation to the classical problem is novel in many aspects, one feature that
requires additional attention is a consistent approach to separating the forward and
backward scattered field components within a perturbative scheme, even when the
effects of backscattering are not weak.
This paper focuses on the solution to coupled differential equations by testing a
mathematical method to separate the forward and backward acoustic fields in a nonseparable waveguide. In contrast to the Foldy-Wouthuysen method, the projection
operator approach does not seek to decouple the forward and backward scattered
fields, but rather to allow for a meaningful separation of the two components and
to construct corrections to the forward propagating field in a consistent manner.
The idea is to first divide the channels into two subspaces, P and Q, where the
P and Q channels represent the forward and backward propagating components of
the field, respectively. The second step is to then reduce the whole coupled-channel
problem to that of the P-channels by eliminating the Q-channels. The effects of
the Q-channel are then treated as an effective interaction in the P-channels with
a perturbation method. In this way, the problem can be potentially solved in a
very efficient manner. Even though the idea of dimension reduction was originally
introduced as a means of solving the nuclear many-body problem, its relevance
transcends large sections of mathematics and physics. For example, for practical
applications it is generally necessary to solve the quantum many-body problem in
a model or restricted basis space. This requirement leads to the introduction of
an effective interaction within the model space. The reduction of the dimension
makes a computation feasible if one can find an efficient method to accurately
handle the effective interaction. Important examples of effective interactions include
the nucleon-nucleon interaction or G-matrix inside a nucleus [8-121 for shell model
calculations, and the nuclear nonlocal optical potential [13-161.
The remainder of this paper is organized such that Section 2 presents the
methodology for the projection operator approach. Section 3 demonstrates an example applied to scattering in underwater acoustics. Section 4 presents a summary.
2

METHODOLOGY

The theoretical formulation presented in this paper uses the example of the 2-D
scalar Helmholtz acoustic wave equation for inhomogeneous media and boundary
conditions. The inhomogeneity in the x-coordinate results in a backward propagating component to the field. The first part of this section defines the 2-D inhomogeneous wave equation and introduces an eigenfunction expansion. This leads to a
set of coupled differential equations. Next, a projector operator method is used to
reformulate the basic coupled equations into a form that includes a nonlocal effective interaction. The integral form of these equations is then introduced. A series
expansion is then introduced for the effective interaction that allows a meaningful
separation of the forward and backward components of the field and a consistent

223
method to construct the forward propagating channels that includes the effects of
coupling t o the backward propagating channels.
For the case of Cartesian coordinates in two dimensions (x and z ) with an
infinite line source in the y direction, the inhomogeneous Helmholtz wave equation
for the pressure P is
a2P

+ +-azz + k2(x,z ) P = -S(x

ax2

a2p

- xo)b(z - zo)

where k = 27rf c(z)-' and (x0,zo) is the position of the source. The acoustic
frequency and the sound speed are represented by f and c, respectively.
The method of partial separation is employed for Eq. (1) by expanding P in
terms of amplitudes Rn as

This leads to a Sturm-Liouville-like problem for depth eigenfunctions 4n, where

d2

-4n
dz2

+ (k2 - k;)&

=0

(3)

The kn are the x-dependent wave number eigenvalues belonging to qhn where
= Snm .

(4)

Using the orthonormal properties of the q5n in Eq. (4),one can obtain a set of
inhomogeneous coupled equations for the x-dependent amplitudes [17]:

where

224

where H(x)is the depth of the waveguide at x and p1 and p2 are the density just
above and below the boundary at H. The notation H+ and H- represent points
infinitesimally above and below the boundary H ,respectively.
For convenience, the coupled equations in Eq. (5) are expressed in vectoroperator form as
&~R)=-~p)-C~R).
(7)
HOis a diagonal matrix where

I p) and I R) are column vectors for the source and x-dependent amplitudes, respectively. C,, is the coupling matrix having elements
Cmn(2)= Am, +am,

+ (2Bmn + &n)-

d
dx

(9)

Recognizing that Hclis the negative of the free Green's function matrix operator,
Go, one converts (see, for example [19]) the set of coupled differential equations in
Eq. (7) into a set of coupled integral-differential equations as

I R) = G O I p) + G O C I R ) .

(10)

In the current work it is important to note that Go explicitly contains, without

ambiguity, both the forward and backward free propagators. The reason is that the
inverse of Hi' has both forward and backward propagators; however, they are not
coupled.
We now present an idea, first introduced in nuclear theory to treat scattering problems involving resonances, for constructing an effective interaction in the
coupled-channel equations. Following Feshbach [16], we introduce the projection
operators P and Q, which are chosen so that

p I R ) =I

p+

Rp),Q

I R ) =I RQ)

Q = 1 , p 2 = p , Q == Q , P Q = Q p = 0 .

(11)

For the time being, the formalism is presented in a general form; however, later the
P and Q channels are identified as the forward and backward propagating channels,
respectively. Operating P and Q on Eq. (10) and using the above operator relations
for P and Q, we have

HPPI R P )= -CPQ I RQ)- CPp I R p ) - I P P )

HQQI RQ)= -CQPI R P )-CQQI RQ)- I PQ)
where

(12)
(13)

225
Solving for the Q channel in Eq. (13) yields

Substituting this expression into Eq. (12) gives

or, solving for 1 R p ) , one obtains

where

The above process is known as channel elimination. One may view I p ) and V p p as
an effective source and interaction in the P-space, respectively, resulting from the
elimination of the Q-channels.
There are several important properties of V p p to note at this point. The first
term in V p p (0th order term) is simply the coupling matrix in the P-subspace.
The second term results from the elimination of RQ and represents coupling of the
P-channel t o the Q-channel at the coordinate x', followed by propagation in the
Q-channel, which is then followed by coupling from the Q-channel back to the Pchannel at the coordinate x. This type of interaction is known as nonlocal. Since
in general the Q - P and Q - Q coupling and the Q-space propagator are complex,
the nonlocal operator in V p p acts as an absorptive potential; i.e., it describes a
loss of flux from the P-channels into the Q-space channels. In electrodynamics and
acoustics, we may view the interaction of P-channel modes with Q-channel modes
as propagation of the P-channel modes through an effective nonlocal potential produced through channel elimination of the Q-space modes. Another property of V p p
is that it depends on the particular division of the total space into the P- and Qsubspaces. Thus, although one has reduced the dimension of the basic equations,
the complexity of their full solution has been replaced by a nonlocal effective interaction, which depends on the details of the definition of the P - and Q-subspaces.
The operator (HQQ CQQ+ i&)-'
represents the exact Green's function for the
Q-space, which in general is difficult to compute; it requires the full solution of the
coupled equations in the Q-space.
The most straightforward means of producing a perturbation expansion for the
potential is by means of an expansion of the Q-space propagator in the form of a
geometric series. The Q-space propagator is expanded as

1
1
1
CQQ-+-CQQ-CQQ-HQQ+~QQ+~&
HQQ HQQ
HQQ HQQ
HQQ
1

HQQ

... . (19)

226
By defining

one observes that the integral equation for the Q-space Greens function is of the
same form as the integral equation for the exact Greens function. Clearly, from
Eq. (19), the effective interaction in the P-channel in Eq. (18) can now be expanded
as
1

Vpp = Cpp - CPQ-CQP

HQQ

1
1
+ CpQ-CQQ-CQp

HQQ

HQQ

1
1
-~PQ-CQQ-CQQHQQCQP
1

HQQ

HQQ

+ ...

The theoretical development presented thus far has been general. As discussed,
the projection operator method depends on the choice of the separation of the
modal spectrum into the P and Q channels. This paper examines the relative contributions of the forward and backward scattering channels of the acoustic field for
non-separable problems. The basic integral equations are restructured in a manner
that separates the forward and backward scattered channels of the x-dependent amplitudes, using the projection operator method. The P and Q channels are identified
as forward (+) and backward (-) scattering channels, respectively. The forward and
backward propagating Greens function operators are defined as

respectively. Also, C++ = C+- = C-+ = C-- = C. One can express the free
Greens function in Eq. (10) in coordinate form for the nth mode, G:, as
G:(z,z)

= --g,(Z)g:(x)Z

--g;(x)g,+(x)d

>z
<x ,

(23)

where g$ and g; are independent solutions to Eq. (5) in the absence of coupling,
that satisfy the radiation boundary condition at x=+m and X=-CQ, respectively,
and W is the Wronskian. G$ and G i are diagonal matrices whose nth components
are integral operators, G$,, and Gin, respectively. Acting on a function f(x), the
components of G& and Gin are
G&f(X) = --

227

respectively. Further, one defines the mode amplitudes, and source functions as

R, = R,++ R;
Pn = P: i- P,

(25)

which leads to the coupled equations for I R+) and I R-),

I R+) = G,+ I p+) +G,+ c I R+)+G; c I R-),

1 R-)= G, 1 p - ) +G, C I R+)+G, C 1 R-) ,
or in coordinate form

(26)
Employing Eq. (21) one obtains,

I R)

+ G i C G i C + GiCGiCGGC + ...)G i I p+)

- GiC(G; + G i C G i -+ G i C G i C G i + ...)C 1 R)

= G$l p+) - G$(C

+ G i C I R)

(27)

The Lanczos method[2,17,20] is then used to efficiently solve for I R+),and I R-)
is obtained from I R+) from Eq. (14). Equation (27) is a key result of the current
research. It gives a consistent method whereby a one-way propagation model can
be modified to include backscattering effects. The second and higher order terms
represent energy lost to backscattering from the imaginary part of the effective
interaction.
Finally, from Eq. (2), one can define the following pressure fields:

where RL represents the solution to Eq. (26) when R; is neglected. P+ may be

interpreted as the forward propagating part of the acoustic field, but including the
effects of coupling with the backward propagating field.

228

Applications

This section presents a numerical example of the projection operator method applied
to the scattering of acoustic waves in a 2-D waveguide. Of interest is t o examine
P, Pf, P+, and P - . The top boundary of the waveguide is a pressure release
surface and the bottom boundary is defined by a rigid boundary condition. The
layer between the boundaries is a perfect fluid with a sound speed of 1500 m/s and
a density of 1.0 g/cc. This layer has a variable thickness with a hill of a height
12.5 m above the flat portions of the waveguide (a thickness of 200 m) followed
by an additional hill, separated from the first hill by 300 m. For the calculation
presented in this manuscript the acoustic frequency is 10 Hz or equivalently the
acoustic wavelength is 150 m. The depth and range of the infinite line source are
50 m and -150 m, respectively.
Figure 1 and Fig. 2 show the results of the calculation employing the projection operator method. The pressure is presented in terms of transmission loss, i.e.
TL = 2010g,,(PP*). Superimposed on Figs. 1-2 is the bathymetry of the two-hill
waveguide. Figure l a shows the result for P . It was found that the perturbation
series for the effective interaction required only three terms for convergence such
that the addition of higher order terms did not cause a change in the TL greater
than 0.1 dB at any point in the range-depth plane. Figure l b shows the result
for P f . Clearly there are major differences between P and Pf, since much of the
structure in Fig. l a results from the interference between forward and backward
propagating components of the field. Figure l c shows the result for P where only
one term is included from the expansion for the effective interaction. Although
there are some differences with Fig. l a , they are minor, indicating that the effect
of backscattering is dominated by the first Born term in the effective interaction.
Fig. I d shows the result for P+ that has a single correction term from the effective
interaction. One observes that the difference between the forward scattered field
without any corrections and the forward scattered field with the correction due t o
the backward scattered channels is small. Figure 2 examines the TL for a depth
of 50 m far P, P+, Pf, and P-. An interesting area of comparison is for ranges
beyond the second of the two hills. This is a region where the backscatterd field
vanishes. Specifically one observes the TL for Pf is lower (greater intensity) than
the case for P. One may ascribe this t o the fact that the effective potential acts as
an absorptive potential where energy flux is taken from the forward channels. This
causes the overall loss for P t o be greater on the right hand side of the scatterers
as compared t o P f .
4

Summary

In conclusion, a channel elimination approach has been introduced t o solve the

Helmholtz equation for non-separable problems. The approach was applied t o the
case of scattering of sound waves in a 2-D waveguide. The projection operator
approach was used t o separate the forward and backward scattered channels of the
field in a perturbative manner. An interesting aspect of the current work is that it
gives a systematic approach to computing the forward propagating channels without

229
attempting t o decouple them from the backward propagating channels.

Acknowledgments
The author wishes t o acknowledge Dr. Takeshi Udagawa for very helpful suggestions
and discussions. Also, the author expresses his appreciation to the late Dr. Fred
Tappert for his helpful comment on the relationship of the projection operators t o
the problem of channel separation. Finally the author expresses his appreciation
t o Dr. Thomas Yudichak for a critical reading of the manuscript. This work was
supported by the Office of Naval Research.

References
1. T. Tamura, Analysis of the scattering of nuclear particles by collective nuclei
in terms of the coupled-channel calculation, Rev. Mod. Phys. 37, 679
(1965).
2. T. Udagawa and B. T. Kim, Random phase approximation calculations of
nuclear response in the continuum, Phys. Rev. C 40, 2271 (1989).
3. D. P. Knobles and T. Udagawa, LContinuumrandom phase approximation
calculations using a finite range interaction, Nucl. Phys. A533, 189 (1991).
4. D. Wurmser, G. J. Orris, and R. Dashen, Application of the FoldyWouthuysen transformation of the reduced wave equation in range-dependent
environments, J. Acoust. SOC.Am. 101, 1309 (1997).
5. L. L. Foldy and S. A. Wouthuysen, On the Dirac theory of spin particles
and its non-relativistic limit, Phys. Rev. 78, 29 (1950).
6. J. D. Bjorken and S. D. Drell, Relativistic Quantum Mechanics (McGraw-Hill,
New York, San Francisco, 1964).
7. P. Ring, Relativistic effects in low energy nuclear structure, Heavy Ion
Physics 10 1999.
8. P. J. Ellis and E. Osnes, An introductory guide t o effective operators in
nuclei, Rev. of Mod. Phys. 49, 777 (1977).
9. P. Ring and and P. Schuck, The Nuclear Many-Body Problem (SpringerVerlag, New York, 1980).
10. J. Goldstone, Derivation of the Brueckner many-body theory, Proc. Roy.
SOC.(London) A239, 267 (1957).
11. B. D. Day, Elements of the Brueckner-Goldstone theory of nuclear matter,
Rev. of Mod. Phys. 39, 719 (1967).
12. B. Barrett, An introduction t o effective interactions and operators in nuclear
physics, Czech. J. Phy. 49, 1 (1999).
13. N. C. Francis and K. M. Watson, Phys. Rev. 92, 291 (1953).
14. A. K. Kerman, H. McManus, and R. M. Thaler, Ann. Phys. 8, 551 (1959).
15. H. Feshbach, Unified theory of nuclear reactions, Ann. Phys. 5 , 357 (1959).
16. H. Feshbach, Unified theory of nuclear reactions 11,Ann. Phys. 19, 287
(1962).
17. D. P. Knobles, Solutions of coupled-mode equations with a large dimension
in underwater acoustics, J. Acoust. SOC.Am. 96, 1741 (1994).

230
18. J. Fawcett, A derivation of the differential equations of coupled-mode propagation, J. Acoust. SOC.Am., 92, 290 (1992).
19. P. Morse and H. Feshbach, Methods of Theoretical Physics (McGraw-Hill, New
York, 1953).
20. D. P. Knobles, S. A. Stotts, R. A. Koch, and T. Udagawa, Integral equation
coupled mode approach applied to internal wave problems, J. Comp. Acoust.
9, 149 (2001).

23 1

Forward with no corrections

Forward+Backward components

50

50

E
I

E
100

100

8 150

a8 150
200

200

100 200 300 400 500

Range - m

100 200 300 400 500

Range - m

Forward + Backward with single corrections Forward with single correction

50

50

E
100
3

[3 150

100

3 150

200

200
0

100 200 300 400 500

Range - m

Range - m

-30

100 200 300 400 500

-20

-10

Transmission loss - dB
Figure 1: Transmission loss versus range and depth for total solution compared to
approximate cases.

232

Forward only
Exact = corrected forward + backscatter
Corrected forward

E
2

--

I+

-3 0
0

200
400
Range - m

600

Figure 2: Various components of acoustic field for two-hill example.

233
MOLECULAR DYNAMICS APPROACH T O SONOLUMINESCENT
BUBBLES
W. LAUTERBORN, T. KURZ, B. METTEN, R. GEISLER AND D. SCHANZ
Drittes Physikalisches Institut, Universitat Gottingen, Biirgerstrasse 42-44,
0-37073 Gottingen, Germany
Sonoluminescence is a phenomenon where sound in a liquid is transformed into
light via one or many bubbles set into strong oscillations by the sound field. The
processes inside the bubble leading to light emission are difficult to monitor, in
fact, only indirect conclusions can be drawn from the shock waves radiated into
the liquid, from the light emission itself and from chemical species that may be
formed and found later in the liquid. The interest in sonoluminescence is high as
there are speculations whether, by special arrangements, the processes inside the
bubbles may be enhanced up to nuclear reactions. Theoretical considerations and
calculations of what may go on inside an oscillating bubble are one approach to
arrive at more knowledge about the interior processes. We study the interior of
collapsing bubbles by molecular dynamics simulations with up to several million
particles. The chosen bubble and acoustic driving parameters correspond to typical
conditions of single bubble sonoluminescence. Bubbles containing a noble gas, noble gas mixtures and gas-vapor mixtures are considered. Mass and heat diffusion,
heat conduction across the bubble wall, condensation of vapor at the gas-liquid
interface, and dissociation reactions are taken into account. Typical examples of
the evolution of temperature and density within the bubble are presented.

Introduction

Bubbles widely appear in nature. Particularly large amounts of them are found
in breaking waves in the ocean, for instance. They are often detrimental when
appearing on ship propellers and in turbomachinery where they may initiate erosion
processes destructing vital parts. In technical applications they are used in chemical
reactors, for instance, when gases and liquids are t o be mixed. Bubbles can also be
generated acoustically by irradiating a liquid with sound. These bubbles have found
an application in ultrasonic cleaning. The collapse of acoustically excited bubbles
in a liquid may generate very high temperatures and pressures in the interior of the
bubble. The extreme conditions in the bubble trigger chemical reactions' and give
rise t o the phenomenon of sonoluminescence, i.e., the light emission of acoustically
driven b u b b l e s 1 , 2 . Sonoluminescence is long known as a faint glow of cavitating
bubbles in insonicated liquids (MBSL - multi-bubble sonoluminescence) and has
attracted considerable attention in the past decade after the discovery of single
bubble sonoluminescene (SBSL), the light emission of single, acoustically trapped
cavitation bubbles3. Sonoluminescene light is composed of very short pulses that
are emitted at the very moment of bubble collapse. Thus, they are very likely of
thermal origin. Measurements of MBSL indicate maximum temperatures of about
5000 - 6000 K and optical spectra featuring emission bands. On the other hand,
analysis of SBSL light gives smooth spectra without pronounced features that are
compatible with blackbody or bremsstrahlung emission and temperatures in excess
of 10000 K. Currently, experiments do not permit to measure the spatiotemporal
dynamics inside a collapsed bubble, because the bubble then is very small (about

234

Figure 1. Sonoluminescent single bubble in a spherical container.

1 pm) and not resolvable by visible light microscopy. Sophisticated numerical

simulations have to be used to study the physics taking place in the bubble4i5
and to set up predictions that can be tested by experiment. In this endeavor,
continuum fluid dynamics (CFD) models prevail. They work well in situations
where the Knudsen number of the flow is small compared with unity. In this work,
a different numerical approach is taken: molecular dynamics (MD) s i m u l a t i ~ n
A typical SBSL bubble filled, e.g., with argon, has an equilibrium radius & of the
order of 4 pm t o 10 pm, which implies that about lo9 to lo1' atoms will occupy the
bubble. For smaller bubbles the particle number now comes into reach of realistic
MD simulations with present-day computers. As a first step, one to several million
particles have been used in the scaled-down MD simulations of bubble collapse
presented in this paper.

235
1

60,

10

20

30

40

50

50

10

0
0

time [p]

Figure 2. Radius-time curve of a trapped bubble in a water-glycerine mixture derived from

photographic observations. A numerically calculated curve (GiImore model) is superimposed on the experimental data points (open circles). The calculation is based on the
following parameters: driving frequency fo = 21.4 kHz, ambient pressure po = 100 kPa,
driving pressure po = 132 kPa, vapor pressure p,, = 0, equilibrium radius R, = 8.1 pm,
density of the liquid p = 1000 kg/m3, viscosity p = 0.0018 Ns/m2 (measured) and surface
tension D = 0.0725 N/m. The gas within the bubble is assumed t o obey the adiabatic
equation of state for an ideal gas with K = 1.2.

Sonoluminescence

Before going into details of the molecular dynamics calculations some experimental
results on sonoluminescent bubbles will be given. Figure 1 shows a single bubble in
a spherical container driven by two piezoelectric disks at frequencies in the range
of 20 to 25 kHz. The bubble is emitting a short flash of faint light each period
of the sound field when undergoing a strong collapse. The tiny bright spot in
the middle of the flask is the light from the bubble made visible through a long
exposure photograph. The corresponding bubble oscillation during one period of
the driving acoustic field is given in Fig. 2 as compared to a numerical model of
spherical bubble oscillationlo. The characteristic oscillation of a sonoluminescing
bubble with a strong collapse and decaying rebounds is observed. The radius at
rest of the bubble is only 8.1 pm, but it expands under the pressure oscillation of
the sound field to more than 55 pm. A photographic sequence of a similar bubble
is given in Fig. 3. The large excursions and the quickly decaying oscillations can
be seen in this stroboscopically taken sequence.

236

Figure 3. Photographic series of a sonoluminescing buble driven at 21.4 kHz with a sound
field amplitude of 132 kPa. The frame size is 160 pm x 160 pm. The frame interval is

500 ns.

The molecular dynamics model

A bubble is modelled as a spherical container of varying radius R(t)holding a large

number N (about lo6 or more) particles. These molecules are imagined as hard
spheres that interact by elastic or inelastic collision upon direct contact; they have
no long range interaction. A molecular species i is defined by the particles mass
mi and diameter di. Molecules with rotational degrees of freedom are represented
by spinning spheres which can also exchange angular momentum. The number of
particles may change by mass transfer at the bubble wall (evaporation and condensation) and by dissociation and recombination in reactive collisions. All physical
quantities are scaled to dimensionless form for the numerical calculations. The
simulation of a hard-sphere gas can effectively be performed by an event-driven
algorithm6 because between collisions the particles propagate in rectilinear motion
- gravity being neglected - and the time of the next collision can be calculated
easily. The computer code keeps a calendar of all possible collisions (and further
relevant events) and advances the time evolution from one event to the next.
The liquid is described as a continuous medium with a spherically symmet-

237
ric flow field which is consistent with the assumed spherical shape of the bubble
wall. The bubbles radial dynamics is calculated by an extended Rayleigh-Plesset
equation that includes liquid viscosity and damping by acoustic radiation:

3 .
2

RR + -R2 = - ( p ~ - p a )
PL

R
+ -@(R,
PLCL

t ) - &,),

Here,

2a

R
R

PL = p ( R ,t ) - - - 4 p ~ ~ -

is the liquid pressure at the bubble wall, p~ is the density of the liquid (here, water
with p~ = 996.6 kg/m3), c ~ = 1 4 8 1m/s is the Eulerian speed of sound, v = 8.569
lo- m2/s is the kinematic viscosity, and a = 72.75
N/m is the surface tension
of the liquid, p ( R ,t ) is the pressure in the bubble at the wall and is calculated by
averaging over momentum transfer from wall collisions of the model particles. The
liquid pressure far off the bubble,

pa = PO

+ pa ( t )= PO - pdO sin Wdt,

(3)

is composed of the static pressure po = 1 atm and a sinusoidal acoustic driving

pressure of amplitude pdo = 1.3 atm and frequency fd = wd/2 = 26.5 kHz. Unless
ptherwise noted all of the forthcoming results were obtained for a bubble of rest
radius & = 4.5 pm and the above given numerical parameter values. The rules of
interaction of the model particles with the bubble wall define the mechanical and
thermal boundary conditions. In our simulations, different implementations of these
rules have been used: (i) in a reflective collision the particles velocity component
normal to the wall is reversed and increased by the amount of the wall velocity; (ii) a
thermal boundary layer with quadratic temperature profile is assumed. It yields the
wall temperature according to the heat power flowing through the bubble surface,
calculated from particle impacts in a self-consistent way. Adiabatic and isothermal
compression of the bubble contents are special cases in this formulation.
The macroscopic quantities (density, pressure, velocity, temperature) are obtained
by averaging over a number of cells of fixed volume (concentric shells) over a certain
time interval. Therefore, the spatial resolution of averaged quantities decreases
towards the bubble center. The MD simulation of a bubble collapse is started a t
maximum bubble expansion. As initial condition, the particles are distributed on
a regular grid in the volume with random velocities according to a Maxwellian
distribution. It turns out that the initial regular arrangement is randomized after
a few collision times. Note that as the initial bubble radius & and density are
prescribed the particle mass is fixed, and the molecules have to be interpreted
as quasiparticles corresponding to a coarsegrained model of the real situation.
Diffusion processes may thus be overestimated by our calculations.
To calculate the sonoluminescene light power, a simple Bremsstrahlung model is
assumed, where the degree of ionization (needed for the power calculation) is derived
from the temperature of the medium by means of the Saha equation. For more
details on the implementation of the MD method, in particular, on the modelling
of condensation and of chemical reactions, the reader is referred to7.

238

65

0 60

-$5

0555
0 50:
0 45

05

US.

<fl0>=519 ke/m3

05

040%

$
\

t = + 28

35

:;:lN

0 30

O0

0 15
-0

+
,

01000 052
0 00

-0.5
r

0.0
Pm

0.5

-05

-0.5

t = + 40 p s , <p>=584 kg/m3

-05

00

00
ILm

05

t = + 66

DS.

-05

0.0
~m

<u>=753 k e / m 3

00
Il.m

05

- .

05

/
t=+106

DS.

<~>=998
ke/rng

0.5

% 0.0
,
N

-0.5

-0.5

0.0
Pm

0.5

-0.5

0.0
r

0.5

~l.m

Figure 4. Two-dimensional temperature distribution (averaged over 100 x 100 cells) of a

spherical Argon bubble with reflective boundary for different times during bubble collapse.

Results

Figure 4 gives an example of the temperature distribution in a two-dimensional

cross-section of a collapsing bubble composed of Argon gas with reflective boundary conditions and a total of lo6 particles. An inward-traveling compression wave

239
focuses at the center and yields a core temperature of more than lo5 K. Due to
neglect of energy-consuming effects in this calculation (heat conduction into the
liquid, condensation and dissociation) the maximum temperature is overestimated.
The wave is reflected at the center and forms an outgoing compression wave heading towards the bubble wall.

lo5
Ft
P= 1.4 atm
P= 1.3 atm
P= 1.2 atm

Figure 5. Dependence of the temperature distribution in the bubble at collapse on the

driving pressure amplitude for three different driving pressure amplitudes.

The temperature in the bubble strongly depends on the pressure amplitude

of the driving. Neglecting possible instabilities of the bubble surface and energy
consuming chemical reactions in the inside very high temperatures may indeed be
reached. Figure 5 shows the temperature inside a bubble at maximum compression
at the bubble center for three different pressure amplitudes of 1.2 atm, 1.3 atm and
1.4 atm. It is seen that the temperature in the center strongly rises with the driving
pressure and approaches a million Kelvin at only 1.4 atm driving pressure. Below
it is shown, confirming CFD results, that just the chemical reactions due to the
unavoidable water vapor content significantly lower the maximum temperature.
In general, a bubble will contain several sorts of particles, for example, a mixture of noble gases in single bubble sonoluminescence, or different species originating from sonochemical reactions (sonolysis). MD runs not detailed here reveal, as
do CFD simulations", that species may be segregated in the bubble according to
their molar mass. The process has a bearing on shock wave formation, chemical
reactions in the bubble, transport of chemical species out of the bubble, and on
sonoluminescence light characteristics.
Calculations were also done with Argon bubbles containing water vapor. When the

240

dissociation of water is taken into account as needed beca.use of the high temperatures inside the bubble, the central temperature in the bubble is lowered. A further
strong lowering of the temperature is obtained when not only dissociation of the
water molecules, but also chemical reactions of the dissociation products are taken
into account.

\
-I
I

+-,

a,

a,

0.0

0.2
Radius r

0.4

0.6

pm

Figure 6. Temperature profiles of a spherical bubble containing a mixture of Argon and

water vapor with dissociation only (- - -) and including chemical reactions with 10 species

(-1.

Figure 6 shows temperature profiles at two moments around maximum compression for an Argon-vapor bubble with and without chemical reactions. The
temperature in the central core of the bubble is lowered significantly by the endothermal processes of chemical species formation (for instance of OH radicals).
Figure 7 gives the evolution of the temperature distribution inside the bubble
near collapse. Actually the last two nanoseconds before collapse are shown. The
vertical axis is the radial axis from the bubble center and the color gives the temperature in Kelvin according to the color bar scale. The upper solid line in the
figure is the bubble wall. The blue part is water. It is again seen that the temperature forms a hot spot in the center of the bubble. Also, the OH radicals form
in the interior of the bubble (Fig. 8) and thus are not likely to be found in the
liquid in large amounts, unless the bubble gets unstable upon collapse and delivers
its contents to the outside.

24 1

x lo4

E3.

Figure 7. The distribution of temperature inside the bubble in dependence on time. The
color signifies the amount of the quantity according to the bar aside the figure in K. The
upper line in the figure is the bubble wall.

Conclusions

The molecular dynamics method is a versatile and powerful, yet computationally

demanding tool for investigating the dynamics within small collapsing bubbles.
With further improvement and extension, it can help to better understand the
physics and chemistry within cavitating bubbles even in complex situations, that
prevail in applications such as ultrasonic processing or sonochemistry. Molecular
dynamics calculations may also be useful for calculating the behavior of coated
bubbles with drugs in their interior when acoustically loaded for drug delivery.

References
1. L.A. Crum, Sonoluminescence, sonochemistry, and sonophysics , J. Acoust.
SOC. Am. 95, 559 (1994); K.S. Suslick, Sonochemistry: A physical perspective, in: W. Lauterborn, T . Kurz, (eds.), Proc. 15th Int. Symp. on Nonlinear
Acoustics, Melville, New York: AIP, 2000, pp. 95 - 104
2. B.P. Barber, R.A. Hiller, R. Lofstedt, S.J. Putterman, Defining the unknowns

242

0.8

0.6

E=L

0.4

0.2

Figure 8. The distribution of OH radicals (density) inside the bubble in dependence on

time. The color signifies the amount of the quantity according t o the bar aside the figure
in natural logarithmic scale in kg/m3. The upper line in the figure is the bubble wall.

3.

4.

5.
6.

7.

of sonoluminescence, Phys. Rep. 281,65 (1997); D. Hammer, L. Frommhold,

Topical review sonoluminescence: how bubbles glow, J. Mod. Optics 48, 239
(2001); T. Lepoint, F. Lepoint-Mullie, An introduction to sonoluminescence,
in: T.J. Mason (ed.), Advances in Sonochemistry, JAI Press Inc., Stamford,
CT, 1999, Vol. 5, pp. 1 - 108
D.F. Gaitan, L.A. Crum, Observation of sonoluminescence from a single,
stable cavitation bubble in a water/glycerine mixture, in: M.F. Hamilton,
D.T. Blackstock (ed.), Proc. 12th Int. Symp. on Nonlinear Acoustics, London:
Elsevier, 1990, pp. 459 - 463
W.C. Moss, D.B. Clarke, D.A. Young, Calculated Pulse Widths and Spectra
of a Single Sonoluminescing Bubble, Science 276, 1398 (1997)
K. Yasui, Temperature in multibubble sonoluminescence , J . Chem. Phys.
115, 2893 (2001)
D.C. Rapaport, The Art of Molecular Dynamics Simulation, Cambridge University Press, Cambridge, 1997
B. Metten, Molekulardynamik-Simulationen zur Sonolumineszenz, Der Andere

243
Verlag, Osnabriick, 2001 (in German)
8. B. Metten, W. Lauterborn, Molecular Dynamics Approach to SingleBubble
Sonoluminescence, in: W. Lauterborn, T. Kurz, (ed.), Proc. 15th Int. Symp.
on Nonlinear Acoustics, Melville, New York: AIP, 2000, pp. 429 - 432
9. T. Kurz, W. Lauterborn, B. Metten, Molecular dynamics approach to bubble dynamics and sonoluminescence , Proc. 17th Int. Congr. on Acoustics
(ICA17), Rome 2001, CD Vol. I
10. W. Lauterborn, T. Kurz, R. Mettin, and C.D. Ohl, Experimental and theoretical bubble dynamics , Advances in Chemical Physics 110, 295 (1999)
11. B.D. Storey, A.J. Szeri, Mixture segregation within sonoluminescence bubbles,, J. Fluid Mech. 396, 203 (1999)

244

REVOLUTIONARY INFLUENCE OF THE

PARABOLIC EQUATION APPROXIMATION
This paper is dedicated in memory of Frederick D. Tappert
Ding Lee
Naval Undersea Warfare Center
Newport, RI, USA
Dept. of Computer Science
Yale University, New Haven, CT, USA
ABSTRACT
It is extremely complicated and impractical to develop a tool which can solve all
acoustics problems in one model. The cost of this effort is very expensive both in
labor and in time. A practical approach is t o solve a class of acoustics problems
efficiently and then extends the solution to solve a more general class of problems.
A class of acoustic propagation problems involves long-range, low-frequency under
range-dependent environments in fluid medium drew interest of acoustic scientists
and engineers. It is desirable to have a model which can efficiently solve the above
class of problems. To attack this class of problems, Frederick Tappert introduced
the Parabolic Equation Approximation Method. This paper reviews what contributions Tappert has made for the formulation of the representative Parabolic Equation
(PE), then, the Split-step Fourier algorithm to solve the PE. The introduction of the
P E not only can it solve the above class of problems effectively but also influence
the progress of developing numerious models to a variety of realistic problems in the
acoustics community. This paper is confined to state Tapperts contribution in the
P E development to the acoustics community. A description of the original development of the PE approximation is outlined along with the solution by the Split-step
Fourier algorithm. Then, the vital influence of the P E approximation to the acoustics
community is discussed.

INTRODUCTION

Lenotovich and Fock [ll,161 introduced the PE to solve electromagnetic problems

way back to 1946. In late 1960, the P E had been used for a variety of applications

245
such as the study of beam propagation problems in radom media which consist of
three types of waves: radio waves (radar), acoustics waves (sonar), and optical waves
(laser). The P E had been used for seismic wave propagation [4, 81. These geophysical
applications have been successful [ 5 , 6, 81. Aimed at developing an efficient model to
solve long-range, low-frequency acoustics problems under the range-dependent environment, Frederick D. Tappert introduced the application of P E to solve this class
of problms. First, a representative PE was derived. Tapperts PE is recognized as
the Standard PE. Then, a Spli-step Fourier algorithm was developed to solve the
Standard PE. The formulation of the Standard P E was an important contribution;
the development of the Split-step Fourier algorithm was a very useful and efficient
for solving the Standard PE. Immediately following the above developments, computer programs were written to implement the Spli-step Fourier algorithm to solve
the Standard PE. A computer code labelled AESD (Acoustic Environmental Support
Detachment of U.S. Office of Naval Research) was developed by Brock [l]. There
was another P E computer, developed at SCALANT Research Center by Jensen and
Krol [13]. This code was labelled PAREQ which was a direct implementation of the
Split-step Fourier algorithm to solve the Standard PE.
Hereafter, the PE development attracted acoustic scientists and engineers and
stimulated their interest to use the Standard P E and available Split-step PE code
for real applications. In addition, Tapperts PE development influenced the scientific community to deveope different PEs other than the Standard P E with solutions
given by a variety of numerical algorithms; a fairly comprehensive report in the whole
19th century can be found in Ref.[30]. This paper focused on the prime contribution
by Frederick D. Tappert. Other contributions, related to the Standard PE, such as
modifie and/or expanded Standard PE, solutions, simulations, and applications are
summarized; Reference [30]also summarizes such progress.
In this entire context, the PE mainly means the Standard PE. Because of the
vital influence of the Standard PE, there are a number of equations which are not
formulated explicitly in the Parabolic Equation format but they are PE-like where
the P E is a special case; these equations are all classified in the category of PE.
This paper begins with the basic Standard P E development aong with the Splitstep Fourier solution algorithm. The advantages and limitations are outlined. The
Standard PE drew a lot of attentions and interests. Early influence encouraged the
use of the Standard PE for realistic applications. iFrom these applications, scientists
and engineers experienced the need of additional PE capabilities to handle a variety
of applications. This influenced the acoustic community to develop PE-like representative equations and a number of numerical solutions other than the Split-step
Fourier algorithm.

246

This paper reports only a part of the P E contributions and its influences. A more
comprehensive extended paper is in preparation and to be reported a t a later date.

BASIC DEVELOPMENT

This section involes three parts: The theoretical development of the PE, the solution
introced t o solve the PE, and a listing of advantages and limitations.

2.1

Theoretical Development

The PE approximation decomposes an elliptic wave equation into two equations

through the choice of an arbitrary constant reference wavenumber. One resulting
equation, by means of the far-field elliptic approximation, gives the Parabolic wave
Equation, the PE recognized by the acoustic community.
The basic P E development begins with the 2-dimensional (range and depth)
Helmholtz equation
1

47.7. + -4,
r

422

+ k 2 ( r ,Z ) 4 = 0,

(1)

where 4 ( r , z ) is the wave field, IE(T,z) = W

is the wavenumber, w = 27r times the
c(r,z)
source frequency, and C ( T , Z ) is the sound speed. Substituting ~ ( T , z =
) U(T,Z)V(T)
into Eq.(l) and apply the far-field approximation, kor 2 1, gives

u,,+ 2iko21,

+ u,,+ k,2(n2(r1

Z)

- 1). = 0.

(2)

The term 12ikou,) dominates the term Ju,,J. Therefore. Eq.(2) can be rearranged in
the standard form:
iko
2
a2
21, = -(n2(r, Z ) - 1).
--= 0.
(3)
2
2ko d Z 2

Equation ( 3 ) is recognized as the Standard Parabolic Equation, originally introduced by Fred Tappert.
Equation 3 can also expressed in an operator form: [17, 19, 301
U,

= i(A(r,Z )

+ B ( z ) ) u= 0,

where
k0

A(T,Z ) = -2( n 2 ( r , Z )

and

B(Z) =

1 a2
2k0 d Z 2 .

1)

(4)

241

2.2

Split-step Fourier Algorithm

Tappert[lS] introduced to solve the Standard P E which, at

u(r0+ AT)= e

ik Av(n2-1)

F-1e*22k0Fu(ro,

(T

+ AT),is

z),

(5)

where the F is the integral Fourier tranform and s is the integration variable.
In the second chapter of Ref.[36] a number of fractional spliting schemes were
discussed. The Split-step Fourier algorithm appears to be only one splitting which is
a special case of multiple fractional splittings.

2.3

Advantages and Limitations

All advantages and limitations are applied to the Standard PE and its Split-step
Fourier algorithm solution.

Advantages
1. The Standard P E along with the Split-step Fourier solution gives the
speedy solution.
2. After the P E formulation, the original boundary-value problem becomes
an initial-boundary-value problem. The far end boundary condition can be
assumed to the Sommefeld radiation condition; therefore there is no need
for the far end vertical bpoundary condition. This allows the solution to
march forward; thus, saves tremendously the memory location for very
long-range propagation problems.
3. save memory

4. artificial bottom very good for deep water problems

Limitations
Limitations come from two sources, limitations related to the Standard PE and limitations related to the Spli-step Fourier Algorithm.
1. far field
2. one way, no backscattering

3. narrow-angle
4. two-dimensional problems

248
5. the selection of depth mesh points in using the Split-step algorithm requires
the number of mesh points has to be a power of two.
6. the treatment of realistic bottom by the Spli-step algorithm.

Due to the vital fluences, the above limitations have been relaxed gradually but not
restricted to Standard PE along with the Split-step Fourier Algorithm.

249

Immediate Effect

The P E development to the acoustic community has been quite influencial from the
application of the Stardard P E with Split-step Fourier algorithm to the formulation
of different forms of P E and the use many different numerical techniques to solve different PE. A fairly comprehensive report of various progress can be found in Ref. [30].

NORDA Workshop I An Improvements - Coming relaticely not too far from the start of a period of
early PE computer codes, Brock et al. [2] modified the sound speed to improve the
accuracy of the P E technique.
A Seismic Applications - Corone et al.[S] discussed the split-step technique. They
discussed the application of a bremmer type series solution to the three-dimensional
reduced wave equation. The series is obtained by iterating gneralizations of an integral equations of the Bellman-Kalaba type. The lower-order term is the solution of
parabolic approximation to the reduced wave equation. Their derivations are based on
the spliting a solution to the reduced wave eauation into "upward" and "downward"
components.

Vital Influence

The influence of the Standard P E is so big which encouraged scientists and engineers to develop not only a variety of various acoustics PE models which offer useful
capabilities, not only stimulated a number of efficient numerical solutions but also
stimulated many different P E applications. these influences are summarized below.

The Wide-Angle P E - There are problems whose solutions are inaccurate if they
are solved by the standard P E due the inappropriate treatment of the angle of propagation. That is why the Standard P E was, also, called the "Narrow-ange PE".
Thus, a wide-angle P E is needed. Discussions have been given in A Benchmark
Test Example, ...ref ... Wide-angle Developments, and An Estimation of the
Propagation Angle. The size of the angle depends upon the required accuracy.
Commonly recognized estimates for the narrow-angle is 5 23"and for the wide-angle
is 5 40".
The Interface PE - In 1981 Homer Bucker presented an interface problem that
cannot be solved by any existing Standard P E model at that time. Lee-McDaniel,
then, developed a technique to handle the fluid-fluid interface. The development was
using the matching conditions on the interface t o construct a representative PE on
the interface; their development was based on the Standard PE but later it was mod-

250
ified to handle the wide-angle propagation. Lee-McDaniels interface wide-angle was
based on a finite difference technique. With respect to the use of the Standard P E
along with the Split-step Fourier algorithm, Thomsn and Smith [26, 331 developed a
wide-angle Split-step aogorithm for the PE.

The Reference Wave Number, ko - The selection of ko is arbitrary but not unique.
Pierce analyzed since the sound speed may change with the range, the ko has to be
updated after every range step is completed. Pierce introduced the following formula
for updating the ko:

Conclusive Remarks

LFrom what has been reported in this paper, it is seen that the introduction of the P E
stimulared a great deal of research interest. This stimulation influenced the acoustic
community t o develope PE-like equations and t o introduce various methods t o solve
these PE-loke equations. Thus, the whole acoutic community has been benefitted by
this contribution. While we experience the influence of the PE, Frederick D. Tappert
must be remembered.

5.1

DEDICATION

This paper is presented in and in recognition of Frederick D. Tappert in recognition

of his his influencial P E contribution to the acoustics community.

ACKNOWLEDGMENTS
This research was supported jointly by the Department of Computer Science of Yale
University, the U S . Naval Undersea Warfare Center Independent Research project.
The author would like to thank Prof.Er-Chang Shang and Dr. David Palmer for their
valuable technical discussions.

25 I

References
[l]Brock,H.K., The Parabolic Equation Model, AESD TN-75-07, ONR, Arlington, VA, 1975

[2] Brock,H.K., R.H.Bucha1, and C.W.Spoffors, Modifying the Sound Speed File
t o Improve the Accuracy of the Parabolic Equation Technique, AESD Tech.
Memo, 1975.

[3] Chan,T.F., LShen, and D.Lee, Difference Schemes for the Parabolic Wave
Equation in Ocean Acoustics, in Computational Ocean Acoustics, eds.
M.H.Schultz and D.Lee, Pergamon Press, Oxford, 1985, 747-754.
[4] Claerbout,J, Coarse Grid Calculations of Waves in Inhomogeneous Media with
Applications to Delineation of Complicated Seismic Structure,, Geophys, 35,
1970, 407-418.
[5] Claerbout,J, Numerical Holography, in Acoustical Holography, 3, ed.
F.Metherel1, Pleum Press, New York, 1971, 273-283.
[6] Claerbout,J, Fundamentals of Geophysics Data Processing,, McGraw Hill, New
York, 1976.
[7] Collins,M.D., Generalization of the Split-step Pade Solution, J. Acoust.
Am.,96(1), 1994, 382-385.

[8] Corones,J.B., B.DeFacio, and R.J.Krueger, Parabolic Approximation to the

Time-independent Elastic Wave equation, J . Math. Phys.,23 (4), 1982, 577586.
[9] Davis,J.A., D.White, and R.C.Cavanagh, NORDA Parabolic Equation Workshop, NORDA TN-143, 1981.
[lo] Dozier,L.B., PERUSE: A numerical Treatment of Rough Surface Scattering
for the Parabolic Wave Equation, J. Acoust. Am.,75(5),1984, 1415-1432.
[ll]Fock, V.A. Electromagnetic Diffraction and Propagation Problems, Pergamon

Press, 1965.
[12] Holmes,E.S. and L.A.Gainey, The Navy Standard Parabolic Equation Model,
Broadband PE, nad P E Workshop 11, in Proc. PE Workshop 11, eds.
S.A.Ching-Bing, D.King, J.A.Davis, and R.Evans, U.S. Naval Research Laboratory, 1993, 175-213.
[13] Jensen,F. and H.Kro1, The Use of the Parabolic equation Method in Sound
Propagation Modeling, SACLANTCEN Memo SM-72, 1975.

252

[14] Godiq0.A. and V.Mokhov, Small-angle Parabolic Approximation for Sound

Field in a Moving Fluid, Sou. Phys. Acoust.,38, 1992, 243-246.
[15] Lan,N, K.B.Smith, and F.D.Tappert, Fast PE, Slow PE, Your PE, MiPE: what
are the real issue?, in Proc. PE Eoekshop 11, eds. S.A.Ching-Bing, D.King,
J.A.Davis, and R.Evans, U S . Naval Research Laboratory, 1993, 249-264.
[16] Lenotovich,M. and V. A. Fock, Zh Eksp. Tero. Fiz., J. Phys. USSR 10, 1946,
13-24.
[17] DiNapoli,F.R. and R.L.Deavenport, Numerical Models for Underwater Acoustic Propagation, in Topics in Current Physics, ed. J.A. DeSanto, SpringerVerlag, Berlin, 1979.
[18] Fishman,L. and S. Wales, A Fast, Filtered, Fourier-transform Marching Algorithm for Wide-angle, One-way Propagation, in Computationa Acoustics- Wave
Propagation, ed. D.Lee, R.L.Sternberg, and M.H.Schult, North-Holland, Amsterdam, 1988, 369-385.
[19] Tappert, F.D., The Parabolic Equation Approximation Method, in Wave
Propagation and Underwater Acoustics, ed. J.B.Keller and J.S.Papadakis, Lecture Notes in Physics, 7 0 , Springer-Verlag, Heidelberg, 1977, 224-287.
[20] Tappert,F.D. and R.H.Hardin, in A Synopsis of the AESD Workshop on Acoustic Modeling by Non-ray Tracing Techniques,, C.W.Spofford, TN-73-05, Arlington, VA, 1973.
[all Hardin,R.H. and F.D.Tappert,in SIAM Review, (chronicle) 15, 1973, 423.
[22] Tappert, F.D., in SIAM Review, (chronicle) 16,1974, 140.

[23] Tappert, F.D.,, J . Acoust. Am., 55, 1974, 534.

[24] Tappert,F.D. and R.H.Hardin, in Eighth Proc. Inter. Congress on Acoustics, 2,
London, 1974, 534.
[25] Tappert,F.D., D.Lee, and H.Weinberg, The Hybrid Parabolic equation, in
Recent Progress in the Development and Application of the Parabolic equation,
ed. P.D.Scully-Power and D.Lee, U.S. Naval Underwater Systems Center T D
NO. 7145, 1984, 6-1 to 6-14.
[26] Thomson,D.J. and N.R.Chapman, A Wide-angle Split Step Algorithm for the
Parabolic Equation, J . Acoust. Am.,74, 1983, 1848-1854.
[27] Lan,N. and F.D.Tappert, Modeling of Reciprocity in the Time Domain Using
the Parabolic Equation Method, J . Acoust. Am.,78(1), 1985, 164-171.

253

[28] Lan,N., S.C.Daubin, and F.D.Tappert, A High-speed, Compact, and Interactive Parabolic Equation Solution Generator PESOGEN, J . Acoust. SOC.
Am.,75, 1984, S1.
[29] Lan,N. and F.D.Tappert, Parabolic Equation Modeling of the Effects of Ocean
Currents on Sound Transmission and Reciprocity in the Time Domain, J.
Acoust. SOC.Am.,72(2), 1985, 1768-1780.
[30] Lee,D., Allan D. Pierce, and E.C. Shang, Parabolic Equation Development in
the Twentieth Century, J. Comp. Acoust., 8(4), 2000.
[31] Lee,D. and S.T.McDanie1, Ocean Acoustic Propagation by Finite Difference
Methods, Pergamon Press, Oxford, 1988.

[32] Lee,D. and M.H.Schultz, Numerixal Ocean Acoustic Propagation in Three Dimensions, World Scientific Pub. Co., Singapore, 1995.

[33] Scully-Power,P.D. and D.Lee (eds), RECENT PROGRESS IN THE DEVELOPMENT AND APPLICATION OF THE PARABOLIC EQUATION, U.S.
Naval Underwater Systems Center T R No. 7145, 1984.
[34] Smith,K.B. and F.D.Tappert, The University of Miama Parabolic Equation
Model, Version 1, MPL, TM No. 432, 1993.
[35] Smith, K.B., Modeling the Effect of Azimuthal Coupling on Acoustic Propagation in the Presence of 3-dimensional, Rough Ocean Interfaces Using the
Parabolic Approximation, in Proc. Theo. and Comp. Acoustics 95, eds. D.Lee,
Y-H Pao, M.H.Schultz, and Y-C Teng, World Scientific Pub. Co., Singapore,
1996, 115-131.
[36] Yanenko,N.N., The Method of Fractional Steps, Springer-Verlag, New York,
1971.
[37] Yevick,D. and D.J.Thomson, Split-step/finite difference and Split-step/ Lanzos algorithms for Solving Alternative High-order Parabolic Equations, J.
Acoust. SOC.Am.,96(1), 1994, 396-405.

254

A VARIATIONAL APPROACH FOR GEOACOUSTIC INVERSION USING

ADJOINT MODELING OF A PE APPROXIMATION MODEL WITH NON
LOCAL IMPEDANCE BOUNDARY CONDITIONS
J.-C. LE GAC AND Y. STEPHAN
EPSHOM, Centre Militaire d'Octanographie, 13, rue du Chatellier, B.P.30316, 29603 Brest
Ckdex, France - E-mail: legac@shom$r
M. ASCH, P. HELLUY
Universitk de Toulon et du Var, ISITV, Avenue Georges Pompidou, B.P.56, 83162 La Valette du
Var Ckdex, France - E-mail: mark.asch@univ-tlnJir
J.-P. HERMAND
Royal Netherlands Naval College, P.O. Box 10000, 1780 CA Den Helder, The Netherlands
E-mail:jhermand@ulb.ac. be
The adjoint model method of control theory is known to give accurate and efficient data assimilation
processes in oceanography and meteorology. However, it has rarely been applied in underwater
acoustics for inversion purposes. Based on the back-propagation of the mismatch between
observations and their predictions, the adjoint model can produce the corrections to the associated
direct forward model input parameters. In this paper, the adjoint of a parabolic equation propagation
model with non local impedance boundary conditions at the water sediment interface is used in order
to determine an acoustically equivalent representation of the seabed. The bottom is represented by
these boundary conditions that play the role of the control parameter.

Introduction

Bottom properties are essential in shallow water acoustics for prediction of the
transmission losses encountered in sonar applications. In the last decade inversion
methods based on signal processing approaches (matched field processing, inversion of
broadband signals) have been applied extensively to estimate geoacoustic features of the
bottom (see [1,2] for a survey of the state of the art). The optimization techniques that
have so far been used within these methods are mostly perturbative approaches based on
the iterative adjustment of the input parameters of a forward model. The forward model is
run for a great number of input parameter values until the mismatch between the
observations (like the spatial structure of the complex pressure field or the temporal
structure of the propagated signal) and their predictions is low. This often results in highly
computationally intensive methods.
In this paper, we present an alternative technique based on a variational approach
coming from the control theory. This technique aims to solve optimization problems
based on partial differential equations. The theory is detailed in [3,4]. Variational methods
by adjoint modeling are tools that have so far been developed in the framework of inverse
modeling of physical systems, in particular in the domains of geophysics and molecular
physics. Their use has been generalized especially to meteorology and oceanography for
the purposes of data assimilation [5-71, model adjustment [8], data inversion [9-10,ll-121
and sensitivity studies [13]. Variational methods have been recently used for acoustic
inversion 114-15,22-231. In 1141, we proposed an optimal control approach based on the
use of a parabolic equation (PE) propagation model in which a local impedance condition
at the water-sediment interface plays the role of the control parameter. The geoacoustic

255
concept that was used was that of the acoustically equivalent medium. Rather than
seeking the geoacoustic parameters in the physical parameter space, we computed the sea
bottom conditions in a way that was not directly comparable with a ground truth model
but that was sufficient to predict the field of acoustic transmission losses. These first
developments and a feasibility study allowed us to demonstrate the controllability of the
PE approximation suggesting the variational approach as a rather promising technique for
geoacoustic inversion. However, the use of local impedance conditions is rather limited
(though realistic in the framework of aerial acoustics [16]), since they can only be used
for local reacting media. For underwater acoustics, a downgoing radiation condition must
be imposed on the transmitted component of the field since the seabed is penetrable to
sound waves, especially at low frequencies. In PE models, this can be done by appending
an absorbing layer [I?] or alternatively by applying a non local boundary condition
(NLBC). This last method that was first formulated and implemented in a finite difference
code by Papadakis [ 181 is rather attractive. Since then, numerous different forms of the
NLBC concept have been proposed (see [ 191 for some references).
In this paper, we apply the NLBC formulation described in [ 191 to generate realistic
boundary conditions and then adopt a generalized form of these NLBC conditions for the
rest of the study. In section 2, the inverse continuous physical problem is presented. We
define the associated optimal control problem and compute the analytical form of the
gradient of the cost function by using the adjoint model of a PE propagation model. The
use of the adjoint model and the gradient of a cost function within a gradient based
optimization loop is then presented. In section 3, we present some numerical simulations
showing the data assimilation of acoustical data for geoacoustic inversion purposes.
2

The continuous problem

The variational approach being new in the framework of geoacoustic inversion, we adopt
a rather simple monochromatic matched field processing inverse technique. The
acoustical data we thus use (or the observations) are the spatial structure of the complex
pressure field measured along a vertical line array.

2. I

The direct forward model

The direct forward model is the standard PE approximation [20]. The model
recognized as the Standard PE is given for a harmonic point source of time dependence
exp(-iot) in cylindrical coordinates ( Y , z ) independent of the azimuthal angle 8 :

au a2u

2ik0-+7+k0(n
Jz

aY

Here

U(Y,Z)

-l)u=O.

is the envelope of the acoustic pressure field P ( r ,z ) ,

reference wave number and

ko = U/cois a

n(r,z ) = co/ c ( Y , z) is the index of refraction. The pressure

is assumed to take the form of

P ( r ,z ) = U ( Y , Z ) .lfi1)(k0r)where Hi1)is a Hankel

function of the first kind and U ( Y , z ) is supposed to be slowly varying with range.
In order to obtain a well-posed initial boundary value problem, we add an initial
pressure field ~ ( 0z ,) given by a classical Gaussian source. The boundary conditions at
the sea surface

z =0

are Dirichlet conditions (perfectly reflecting interface).

256

For the boundary conditions at the water-sediment interface z = H , we adopt the

NLBC formulation proposed by Yevick and Thomson [19]. In that paper, the authors
proposed a procedure for obtaining NLBC's directly from the z-space Crank-Nicolson
solvers for both the narrow-angle standard and the wide-angle Claerbout PE's. The
calculation of the wave field u(r,z ) at range r Ar along the water-sediment interface

is obtained in terms of the known field at the previously calculated range values from 0
to r by expanding the approximate vertical wave number operator for the downgoing
-ha,
field in powers of the translation operator T = e
.
Define the z-space vertical wave number operator as:

where V 2 = 4i/koArand nb is the index of refraction of the bottom. The authors obtain
the following impedance boundary condition:

[$-

i$ro]u(r

+ Ar,H ) = 0,

(3)

where p is the density, the subscript b indicates the bottom, and the subscript w indicates
the sea water. The numerical implementation of this NLBC requires that the operator
is Taylor-expanded in T ,which can be finally expressed as:

ro

)
where p = ( p w / p bkoJ-

and go,, are coefficients coming from the Taylor

expansion.
The above model was shown to give accurate simulations of the acoustic pressure
field for homogeneous semi-infinite fluid or solid half-spaces. In the following study, this
reference model is used to generate realistic NLBC's related to some given geoacoustic
parameters. However, due to the difficulty to directly apply the variational approach to
this reference model, we chose to generalize Eq.(4) in the following form:

For the numerical results presented in section 3, the initial and the true NLBC
parameters and F are first calculated with the reference model to obtain simulations
that represent realistic environments of underwater acoustics.
As a synthesis, the continuous direct model can be written as:

257

au a2u

2 i k 0 - + 7 + k , 2 ( n 2 -1)u = O
at-

az

Vr,VzE [O,R]*[O,H]

u=o

r = O , V ~ [o,H]
E
z=O,VrE [O,R]

-- iP(r)u = F ( r )

z=H,VrE [O,R]

u = S(z,z,)

au

az

(6)

where R is the distance between the source and the vertical line array on which the
acoustic pressure field is measured, and S(Z,z,) is the acoustic source at z = z , .~
2.2

The inverse problem formulation

The control procedure consists of finding the optimal controls

[Pop,,Fop,] and the

Fop,) which minimize a cost criterion

corresponding optimal field uopj= u(Popj,

J ( P , F ) based on the standard matched field processing between a measured field,

ud( R ,Z ) , measured on the vertical line array at a given range R and the corresponding
predicted field, u ( P , F ; R, z ) , obtained from the direct model Eq.(6):

The problem is to find:

min{J(P,F ) ,

[P,FI'

Gad

(8)

where Gadis the set of admissible controls. We note that Eq.(7) is the simplest possible
cost function and that we can add regularization terms to it.

= [ P , F ] , the concatenation of both control parameters and

$ = [b,f ] , the concatenation of some perturbations of the control parameters. The
necessary condition for the existence of a minimum is given by the following theorem.
Let us define

Theorem 1: If J attains a (local) minimum at a point l,Y* E Gad, then

V$ E Gad &J(y.;$) = 0 , where

directional Gdteaux derivative of

J(ly+t$)-J(Y)
4) = lim
,-to

dl(l+f/;

the

is

J at the point y along the direction $ .

Following the same developments as in [14], we start by calculating the variation of

258

where g is the function defined by:

We introduce the real-valued scalar product: (u,V) = %e(u?),

where ? is the complex conjugate of v .
Assuming that

J'(v,(b) exists for

all (band is continuous and linear in (b, the

gradient vJ((b) can be defined as the linear form:

The optimization problem Eq.(8) cannot be resolved directly by using

J'(v,(b) = 0 .

vJ((b)from J'(ly,(b)based on Eq.(12). The variation can

The idea is thus to identify

be written in the following form:

J'(v,(b)= g'(t)l

1=0

with

We then have to write down the conditions satisfied by w = du 1dt , In order to do

this, we write the two systems satisfied by
and
t(b) . They are:

u(v) u(v+

2ik0 -+ -+ ki (n' - l)u(v) = 0 b'r,Vz E [0, R]* [0, H ]

and

az2

ar

= S(Z, z,>

r=O,b'zE [O,H]
Z = O , b ' r E [O,R]
z = H,b'rE [O,R]

259

By subtracting Eq.(l5) from Eq.(14), dividing the result by t and letting t tend to
zero, the system satisfied by W , called the tangent linear model (TLM) is given by:
r

aw aZw

2ik, -+ 7+ k,2(n2 - 1)w = o

ar az
w(0,z ) = 0
w(r,O)= 0

Vr,VzE [0,R]*[0,H ]
VZ [O,H]
V r E [0, R ]

aw(r,H , - ip w(r,H ) - ib u(r,H ) =

V r E [O,R]

az

The aim is to use the TLM to evaluate

(16)

J ' ( v , @ ) .By introducing a function

p ( r , z ), and by using the TLM, vJ(@) can be identified directly from Eq.(12). The
new variable p ( r ,z ) , called the adjoint variable, satisfies the following system (that is,
of the same degree of complexity as the TLM satisfied by

w:

aP + 7
J 2 P + ki (n' - 1)p = 0 Vr,Vz E [0,R]* [0,H ]
2ik, ar

az

p ( R , z )= u -ud
p(r,O) = 0

<

'

az + iiip(r,H ) = 0
(

V z E [0,H ]
V r E [o, R J
Vr E [0,R ]

This model, called the adjoint model is running backwards in r from r = R to

r = 0 . It allows us to calculate J ' . Taking the scalar product of the TLM Eq.(16) with
the adjoint variable p and integrating by parts:

we finally find, after some analytical developments, that the derivative of J with respect
to @ is given by:

260

(V,0) = [
2fr

(b,up) f

r.**H\

F/

(/,!>}

2JL J ''="V

(19)

F/

This latest equation can be reformulated in terms of the gradient of the cost function:

up

VJ =

2k0
ip
T)'->

(20)

r,z=H

where p is the solution of the adjoint equation obtained from the application of the
theorem.
2.3

The optimization procedure

By using the successive resolution of both the forward direct model and the backward
adjoint model, we can therefore calculate the exact gradient of the cost function relative to
the control parameter lff(r) = [j8(r), F(r)J. Once the gradient of the cost function VJ
is known, we can seek a (local) minimum of J ( \ j f ) . The global optimization procedure is
summarized in Fig. 1. In order to accelerate the convergence, we use a classical conjugate
gradient method of Polack-Ribiere type rather than the simplest method of steepest
descent. The algorithm that is used is described in [21] and proceeds in two steps. First,
the direction of minimization is computed following the Polack-Ribiere algorithm that
exploits the gradient of the cost function Eq.(20). Then the step length along the
minimization direction is computed thanks to a soft linesearch algorithm that applies the
Wolfe criteria. The interested reader is referred to [21] for a precise description of the
implemented algorithm.
Direct model
(from 0 to K)
Not converged

d
VJ@,F)

(from R to 0)

Converged
(P,F)op,=

(p,F)

Adjoint model

uopt = u

Figure 1. Optimization algorithm - J is minimized until some convergence criteria are attained.

U-Ud/r-R

26 1
3

Numerical results

The numerical results that are presented in this section rely on the finite-difference
implementation of the direct and the adjoint models presented above, and of the
associated discrete version of the gradient of the cost function. Before applying the
inversion procedure, a gradient check (not shown here) was performed.
For a frequency of 500 Hz,we present the results obtained with an isovelocity sound
speed profile in the water column of 1500 m s']. The true non local impedance boundary
conditions were initially calculated for a sandy reflecting bottom with a compression
speed of 1600 m s-', an attenuation of 0.5 dB/h and a density of 1.8 g ~ r n ' ~The
. initial
ones were obtained with a soft sediment (mud) characterized by a compression speed of
1505 m s-', the other parameters remaining unchanged. No more than 5 or 6 iterations of
the optimization algorithm were needed for the algorithm to converge. However, due to
the drastic convergence criteria we adopted for our calculations, we reached a maximum
of 10 iterations. The corresponding computations, performed on a high-end PC (Pentium
IV, 2 GHz), required around 20 minutes for a complete inversion. By optimizing the
convergence criteria, this could be reduced by a factor of two.
Figure 2 gives an overall representation of the reconstruction of the field in the entire
range-depth domain. Although the initial field (top) was quite different from the true field
(middle), the field calculated after inversion (bottom) agrees very well with the true one.
The improvement of the acoustic pressure filed is drastic, the initial error for ranges
greater than 400 m being reduced from 30-40% to only 2% approximately.
UI"H

k
3

0
03
50

02
0 1

100
100

200

300

400

-::,3

600

700

800

900

03
50

0 2

0 1
100

200

300

400

22-v-

600

700

800

900

03

0 2
0 1
0

Figure 2. Visualization of the amplitude of the initial (top), true (middle) and inverted (bottom) u fields for
j%OO Hz. After the assimilation process, the inverted and the true fields are nearly identical.

262

0
0 221

go
-0 2
1

.
200

1
400

I
600

800

lo00

r(m)

Figure 3. Comparison of the initial, the true and the inverted control parameters - (a)(b) Inversion of the real and
imaginary parts of the p parameter - (c)(d) Inversion of the real and imaginary parts of the F parameter.

Figure 3 is devoted to the representation of the control parameters. We can observe

that the accuracy of the F parameter is excellent after the inversion process. On the
contrary, the p parameter is badly determined, which is certainly due to the ill-posed
nature of the inverse problem. The use of a wider angle PE approximation is susceptible
to improve our results since the angular limitation of the standard PE approximation
imposes that the chosen Gaussian source has no interaction with the bottom near the
source. This assumption is under investigation.
4

Conclusions

We have generalized the variational approach that was presented in [14]. Here we treat
the case of a PE propagation model constrained by NLBC's to simulate realistic
environments in underwater acoustics. The obtained results are promising. They show
that for a reasonable computational cost punctual acoustical data can be assimilated (here,
vertical array data) which results in a global and drastic improvement of the prediction of
complex pressure fields. The other central concept underlying our variational approach is
that of the acoustically equivalent medium. This results in a rather abstract but concise
representation of the geoacoustic properties of the seabed via its non local boundary
conditions. These NLBC's are not easily related to the classical geoacoustic parameters of
the bottom but are able to simulate the same acoustical behavior, at least for the simple
but realistic environments considered in our simulations. In sonar applications, one of the
main objectives is to be able to predict the acoustical losses in the water column. For this
purpose the concept of acoustically equivalent medium is relevant and NLBC's are
therefore good potential candidates as equivalent models. The presented method is rather
new and should be studied in order to improve the generality of the approach. Further
works should be involved in the use of sparse linear arrays and wider angle PE
approximations or other propagation models. The application of the method will also be
studied in the case of more complex seabed properties (multi-layered sub-bottoms). More
importantly, and following the last tendencies in geoacoustic inversion, multi-tone and
broadband signals will be considered in order to physically constrain the inverse problem
and to aim at a unique inverted solution.

263
References
1. Chapman R., Tolstoy A,, Benchmarking geoacoustic inversion methods - Special Issue, J. Comp. Acoust.
6, N01&2 (1998) 289 pages.
Hermand J.-P., Broad-band geoacoustic inversion in shallow water from waveguide impulse response
measurements on a single hydrophone: Theory and experimental results, IEEE J.Ocean.Eng. 24( I ) (1999)
pp. 41-66.
3. Lions J.-L., Contr6le optimal des systtmes gouvernes par des tquations a m dtrivtes partielles (Dunod,
Paris, 1968). In French.
4. Lions J.-L., Contr6labilitt exacte, perturbations et stabilisation des systbmes distributs, Vol. 1 (Masson,
Paris, 1988). In French.
5 . Talagrand O., Courtier P., Variational assimilation of meteorological observations with the adjoint
vorticity equation Part I, Quaterly J. Royal Meteo. Soc. 113 (1 987) pp. 131 1-1328.
6. Courtier P., Talagrand O., Variational assimilation of meteorological observations with the adjoint
vorticity equation - Part 11, Quaterly J. Royal Meteo. Soc. 113 (1987) pp.1329-1347.
7 . Leredde Y., Lellouche J-M., Devenon J-L., Dekeyser I., On initial boundary conditions and viscosity
coefficient control for Burgers' equation, Int. J. Num. Meth. Fluids 28 (1998) pp.113-128.
8. Schroter J., Variational assimilation of GEOSAT data into an eddy-resolving model of the Gulf-Stream
extension area, J. Geophys. Ocean. 23 (1992) pp.925-953.
9. Plessix R-E., De Roeck Y-H., Chavent G., Waveform inversion of reflection seismic data for kinematic
parameters by local optimization, SIAMJ. Scient. Comp. 20 (1999) pp. 1033-1052.
10. Jameson A,, Aerodynamic design via control theory, J. Scient. Computing, 3 (1 988) pp.233-260.
1 1. Le Dimet F.X., Talagrand O., Variational algorithms for analysis and assimilation of meteorological
observation: Theoretical aspects, Tellus 38 (1 985) pp.309-322.
12. Belmiloudi A,, Brossier F., A control method for assimilation of surface data in a linearized NavierStokes-type problem related to oceanography, SIAM J. Control. Optim. 35 (1997) pp.2183-2197.
13. Cacucci D., Sensitivity theory for nonlinear systems. Part I: Non linear functional analysis approach, J.
Math. Physics 22 (1981) pp.2794-2812.
14. Asch M., Le Gac J-C., Helluy P., An adjoint method for geoacoustic inversion, Invited Lecture, In Proc.
PICOF'O2 2"" Conference on Inverse Problems, Control and Shape Optimization, Carthage, Tunisia
(2002).
15. Hursky P., Porter M., Comuelle B.D., Hodgkiss W.S., Kuperman W.A., Adjoint-assisted inversion for
shallow water environment parameters, In Impact of Littoral Environmental Variability on Acoustic
Predictions and Sonar Performance, eds. N.G. Pace and F.B. Jensen, Kluwer Academic Publishers (2002)
pp. 449-456.
16. Robertson J.S., Siegmann W.L., Jacobson M., Low frequency sound propagation modeling over a locally
reacting boundary with the parabolic approximation, J. Acoust. Soc. Am. 98(2) (1995) pp. 1 130-1 137.
17. Jensen F.B., Kuperman W.A., Porter M.B., Schmidt H., Computational Ocean Acoustics (AIP Series in
Modem Acoustics and Signal Processing, New York, 1994)
18. Papadakis J.S., Impedance formulation of the bottom boundary condition for the parabolic equation model
in underwater acoustics, in J.A. Davis, D. White and R.C. Cavanagh, "NORDA parabolic equation
workshop, 31 march - 3 April 198I': Tech. Note 143, Naval Ocean Research and Development Activity,
NSTL Station. MS. 1982. availablefrom NTIS, No.. AD-I21 932 (1981) p.83.
19. Yevick D., Thomson D.J., Non local boundary conditions for finite difference parabolic equation solvers,
J. Acoust. Soc. Am. 106(1) (1999) pp.143-150.
20. Tappert F.D., The parabolic approximation method, in Wave Propagation in Underwafer Acoustics, eds.
J.B. Keller and J.S. Papadakis, (Spinger-Verlag, New-York, 1977) pp.224-287.
2 1. Frandsen P.E., Jonasson K., Nielsen H.B., Tingleff O., Unconstrained optimization, Lecture Note, IMMLEC-2, Technical University of Denmark (1999). Online version available at the site

2.

22.
23.

<http://www.imm.dtu.dk/-hbn/software.html>.
Asch M., Le Gac J-C., Adjoint modeling for geoacoustic inversion - 1: PE approximation with local
impedance boundary conditions, Submitted to J. Acoust. Soc. Am. (July 2003).
Le Gac J-C., Asch M., Adjoint modeling for geoacoustic inversion - 11: PE approximation with non local
impedance boundary conditions, Submitted to J. Acoust. SOC.Am. (July 2003).

264

Visco-acoustic Wave Equation Migration

Li Guofa,MouYongGuang,Wang Changchqand Liang Qi
(Petroleum University of Beijing, China)

Abstract
Conventional migration methods have been being limited in its imaging ability due to without
consideration of absorption and dispersion of subsurface medium, Thus it is difficult for them to
very precisely image subsurface structures. Also the amplitude of image can not correctly indicate
.;the changes of lithological and physical properties. In this paper, seismic absorption and
dispersion is considered when a migration extrapolation operator is computed, and also
compensated during the wave extrapolation. extrapolation operator is obtained by a nonlinear
optimization along with a linear optimization. With this method, the operator can be of stability
and accuracy. The processing on numerical models and field seismic data show that, in
comparison with traditional acoustic wave migration algorithms, the migration with absorption
compensation algorithm can get more correct and accurate subsurface image.

Key words: migration, absorption, compensation, dispersion, extrapolation, optimization

INTRODUCTION
The classical elastic wave theory is based on the Hookes law precisely, and is suitable to
depict and describe seismic wave propagation in the elastic medium. But both the experimental
results on earth core and the practical results of exploration and development in oilfields show that
the earth media express more or less the non-elastic characteristic, namely absorption and
dispersion. Especially when the media contain gas or oil, the visco-elastic behavior is more
obvious. The visco-elasticity can cause the energy attenuation, frequency dispersion and phase
distortion of seismic wave, and some times lead the seismic processing and interpretation based on
the elastic wave theory into a risky trap.
Along with the development of migration technology and the improvement of exploration
accuracy, our expectation for migration is getting higher too. Through migration, now we not only
just want to get accurate subsurface image as we did before, but also want the image amplitude to
be able to more correctly indicate the change of lithological and physical property of subsurface
structures. AVO, as a key technology to detect subsurface lithological attributes, has been
successfully used in some areas. But it also failed and caused AVO traps in other areas,
especially when the diffraction of seismic wave is strong. In this case, the reliability of AVO will
greatly be lowered. Study show that this is mainly related to a un-valid re-distribution of reflection
energy and diffraction energy by migration. This is the main reason why amplitude-persevered
migration technologies begin to catch a common interest within the industry in recent years. Since
the major factors of changing amplitude include dispersion, transmission and absorption, so a true
amplitude-preserved migration should include the consideration of seismic absorption and
dispersion in the viso-elastic media.
Early visco-absorption compensation technology, i.e. Inverse Q filter that is still widely used
within the industry, is was developed in ID way. There are different Inverse Q filter methods

265
accomplished by Biche1(1998), Gelius( 1987), Varela( 1993), Hargreves(l994) and Santwlar(1996).
But, seismic visco-acoustic attenuation is not only a function of time, but also of wavelet
frequency, velocity field, ray-path and etc. Only using inverse Q filter as a way to compensate the
visco-elastic absorption and dispersion is not enough, it is obviously unable to fully recover the
original characters of a seismic wave field.
Visco-acoustic migration is obtained using pseudo-spectral algorithm (Dai and West, 1994), and
it made visco-elastic compensation method extended from ID to 2D. Mittet and Sollie(1995)
introduced a visco-elastic pre-stack migration in f-x domain using explicit convolution filter, an
numerical analyses on the sensitivity of migration to macro Q models was also presented.
Visco-acoustic extrapolation operator can be solved by use of frequency, velocity and
absorption coefficient. This i s a typical nonlinear optimization problem. The solution is combined
with linear optimization in this paper, thus the computation accuracy and efficiency is efficiently
improved. In order to improve the computation efficiency of 3D migration and enhance the
applicability of the algorithm in field data, McCellan transformation was introduced to the explicit
migration algorithm which uses a table-driven scheme, and downward continuation of a 3D
wave-field is realized with a 1D visco-acoustic extrapolation operator. In this way, the
computation efficiency of 3D migration is greatly increased.
The analyses on the error of visco-acoustic extrapolation operator were done. Also the
migration method presented in this paper is applied to both numerical models and field seismic
data, in comparison with acoustic wave migration, visco-acoustic migration gets a better image.

THEORY AND REALIZATION

Standard linear visco-acoustic wave equation can be expressed as

Where

k
f is complex modulus, p

is density.

The complex velocity is defined as

analytic solution of equation (1) is

The complex wave-number of

k(o)

is defined as

k(m)= m/v(m)= k , (a)+ik,( m )

r:

(4)

The phase-shift operator can be expressed as

&~+4
)'-k:+k:bz

D(k, , k , , k xA, )=
Let

)z-(k;+k;&

For a down-going wave

For an up-going wave

(5)

exmM
,e,ma represent the maximum angle in x-direction and y-direction, respectively. We

266
define four sub-spaces

a,,Q2,, ,a3as follows

Q2

Ro:{-nlAx5kx<nlAx,-n/Ay5ky S n l A y )

L2

62

Now, the design of optimum convolution extrapolation operator can be expressed as follows:
Given

k, ,k,

and a attenuation function

size of (2Lx l@Lv

+ 1)

y(kx,k y )

in

Qz,to find a filter w(i,j )

which is

, symmetrical to the origin point, and Its response w ( k x , k y )in

(kx,ky)domain is nearest to D ( k x , k y )in

a,, nearest to

y ( k z , k y ) xD(kx, k y) in

a,

and/W(k,,ky] 4 l i n Q,
Due to the symmetry of 2D extrapolation operator w(i,j), its Fourier transformation in

(kx,ky) domain can be expressed as

Where

sij = 1/2

The amplitude error

respectively as:

i>l.and.j>l

(i = 0.and.j 2 l).or.(i 2 1.and.j = 0)

l/4

i = 0.and.j = 0

and the phase error & between

w and

are expressed

So the solution to an optimum convolution extrapolation operator is equivalent to the following

267
constraint optimization problem.
Target function:

J(w)=

jj(d + k ) d k x d k y
Ql+%

Constraint:

Obviously it is a nonlinear problem. There are many algorithms and ready-made software
packages which can be used to solve the problem, but the iterative efficiency and the final
accuracy is dependent on the initial. To improve the efficiency and the accuracy, we firstly
approximate the equation (12) as a linear problem; then we use the solution of linear problem as
an initial value of the nonlinear equation (12). In doing so, the computation efficiency and
accuracy of the equation (12) is improved.
The nonlinear equation (12) can be approximated as the following problem.

Where

T,

is a positive real number that is less than 1. The solution of the equation (1 3) is:

W, ,W , are real and imaginary parts of convolution extrapolation operator

w , respectively.

Where:

Where

D,,D,

are the real and imaginary parts of ideal phase shift operator D , respectively.

268
We take the solution of the equation (14) as an initial value of the constrained optimization
equation (12) and do the iterations to find the final the convolution extrapolation operator that
satisfies the precision requirement and has a minimum size.
Since the extrapolation operators are only the function of Q and ratio of frequency to velocity,
so in order to enhance the migration efficiency, an operator-table was calculated out before
migration. During the migration the operator can be directly accessed from operator-table
according to frequency, velocity and Q.
Though it is of higher precision to use the 2D operator for 3D wave-field extrapolation,, but it
causes a lower efficiency, especially for 3D pre-stack migration. In order to increase the migration
efficiency while maintaining the precision, we introduce the McClellan transformation to perform
the 3D migration in a similar way we do in the acoustic wave 3D depth migration (Hale, 1991).

TESTING ON SYNTHETICMODEL AND FIELD SEISMIC DATA

To understand the approximation degree between the ideal visco-acoustic extrapolation operator
and the calculated visco-acoustic operator, we conduct a numerical error analysis of which the
basic parameters are: frequency40Hz, velocity-3000m/s, trace interval--20m, maximum
angle-75
, extrapolation s t e p 3 0 m , operator length-20
trace. Here we can calculate the
~

wave-number of evanescent wave as

k r d = 2 @ / ~= 1.6755161

Figure l a is the amplitude spectrum of ideal extrapolation operator, Figure l b is amplitude

spectrum of practical extrapolation operator. We require the two spectra are close enough when

kx < k:

, and the amplitude of the practical operator is less than 1 when

kx > k: ,

Figure

l c is the error between the two spectra, from which you can see the error is less than one
thousandth when the reflection angle is less than 7 0 ~.. Figure 2a is the phase spectrum of the ideal
extrapolation operator, Figure 2b is the phase spectrum of the practical extrapolation operator.
Figure 2c is the error between the two phase spectra,. From here we can see that the error is less

than one thousandth when the reflection angle is less than 7 0 .

Figure 3 is the synthetic model test for visco-acoustic post-stack migration. Figure 3a is a
designed geological model. Figure 3b is the visco-acoustic response of figure 3a, Figure 3c is the
migration result of figure 3b using acoustic wave migration. Comparing with figure 3a, there are
serious errors in energy, frequency, and contacting relationship of layers, Especially for two
dipping reflections, the waveform and energy are distorted. Figure 3d is the migration result of
figure 3b using visco-acoustic migration. From this figure we can see the visco-acoustic method
have compensated absorption and dispersion and corrected waveform distortion.
Figure 4 is the synthetic model test for visco-acoustic pre-stack depth migration. Figure 4a is
the geological model, in which the main parameters are listed. Designed trace interval is 20m, shot
interval 40m, minimum receiver offset 20m, trace number per shot 150, sample rate 4ms,
recording length 3s, the source wavelet is Ricker wavelet with a dominant frequency of 22Hz.
Figure 4b is a common-shot record at cdp80 obtained by using visco-acoustic modeling. Figure 4c
is the acoustic pre-stack depth migration result, Since the influence of absorption and dispersion
are not eliminated, the result is far from designed model. Figure 4d is the visco-acoustic pre-stack
depth migration result, in which the reflection energy, frequency and phase have been well

recovered. Figure 5 is test on 3D field data of DaGang Oilfield. Figure 5a is acoustic post-stack
migration result. Figure 5b is the visco-acoustic post-stack migration result, we can see that
visco-acoustic migration greatly improves the whole imaging quality compared with acoustic
migration.

CONCLUSION
Due to the absorption and dispersion it is difficult to get image of high quality by use of
conventional acoustic migration methods. In the paper, We presents a visco-acoustic migration
method for the purpose of elimination the influences of viscosity. The tests on synthetic model and
field seismic data show that visco-acoustic migration can improve imaging quality greatly, and
will play important roles in detailed seismic exploration.

REFRENCES
(1) Mittet R., prestack depth migration with compensation for absorption and Dispersion.,
Geophysics, 1995,60(4)
( 2 ) Sollie,R.,and Mittet R.,

prestack depth migration: sensitivity to Macro Absorption model.,

1994,64" SEG meeting

( 3 ) Jose M. carcione, Visco-acoustic wave propagation simulation in the earth, Geophysics,
1988,53(5)
( 4 ) Hale D.,

Stable explicit depth extrapolation of seismic wave fields, Geophysics, 1991,56,

1770-1777

(5) Holberg O., Towards optimum one-way wave propagation, Geophys. Prosp., 1988,
36,99-114
( 6 ) N.,D.,Hargreaves, Inverse Q filtering by Fourier transformation, Geophsics, 1991,56(4)
( 7 ) S., H., Bicket, Plane wave Q deconvolution. Geophysics,1985,50(9)

( 8 ) Dave Hale, 3-D depth migration via McClellan transformations, Geophysics,l991,56( 11)
(9) Samuel H.Gray, True-amplitude seismic migration: a comparison of three approaches,
Geophysics , 1997,62(3)
(10) Blacquiere,G., 3D table-driven migration, Geophys. Prosp., 37, 925-958

270

110

1.M

1 .m

1.0

095

05

0.85

041

Ill

Ill

4.0 .25 -21 .15 -1.0-05 0.0 0.5 1.0 15 2.0 2.5 31:

-3.0 -2.5 -2.0 -15 -1.0-05 0 0 05 1.0 1.5 2.0 2.5 3.0

XI

KX

(b)

05

.30 -2.5 -20 -1.5 ~1.0-0.5 0.0 0.5 1.0 1.5 2.0 2.5

3.0

kx

4.0 25 -2.0-1.5 -1.045 0.0 0.5 1.0 15 2.0 2.5 3.1

XX

-3.0.2.6 m2.0.1.6 -1.04.6 0.0 0.6 1.0 1.6 2.0 2.6 3.0
KX

Fig. 1 (left): error analysi of amplitude

spectrum of extrapolation operators.

Fig. 2 (right): error analysi of phase

spectrum of extrapolation operators.

(a) amplitude spectrum of ideal extrapolation

(a) phase spectrum of ideal extrapolation

operator
(b) amplitude spectrum of practical extrapolation

operator
(c) amplitude error between the two operators

operator
(b) phase spectrum of practical extrapolation

operator
(c) phase error between the two operators

27 I

60

70
80
90

LOO

110
120

130

Fig. 3a: synthetic geological model in time domain

600
700

800

900
,000
1.00

1200

1300

Fig.3b: visco-acoustic response of Fig.3a

~.

Fig.3~:acoustic migration result of Fig. 3b

Fig.3d: visco-acoustic migration result of Fig. 3b

272

Fig 4a: synthetic geological model

Fig 4b: co-on-shot

record at cdp80
with visco-acoustic modeling

Fig 4c: acoustic pre-stack depth

migration result

Fig 4d: visco-acoustic pre-stack depth

migration result

Fig 5 : comparision between acoustic post-stack migration (a) and visco-acoustic

post-stack migration (b)

273

Time-lapse Seismic AVO Attributes Analysis with Variation of Oil

Saturation, Pressure and Temperature
Jingye Li*, Xiaohong Chen and Long Jin
(Key Lab of Geophysics, University of Petroleum, Beijing, 102249, P.R China)

Summary
Time-lapse seismic survey has been found one of effective reservoir monitoring techniques. And
time-lapse seismic AVO (Amplitude Versus Offset) attributes analysis has great potential in
distinguishing different reservoir parameters changes. In this paper, through Gassmans equation, velocity
empirical formula related to effective pressure, clay content and porosity, and Bound Averaging Method,
the changes of velocity and density are calculated with the variation of oil saturation, pressure and
temperature. By Zoeppritz equations, corresponding P-P wave and P-SV converted wave AVO curves are
obtained, then the relations between AVO curves changes and variation of reservoir parameters are
analyzed. Time-lapse seismic P-P wave and P-SV converted wave AVO curves have different sensitivities
to different reservoir parameters variation, so time-lapse seismic differences caused by different
parameters variations can be distinguished according to this point, which has great sense in studying
time-lapse seismic feasibility and guiding time-lapse seismic data interpretation.
Introduction

Time-lapse seismic reservoir monitoring technique has been successfully used in monitoring the oil
saturation change in production, but it is known that the variations of reservoir parameters, such as oil
saturation, pressure and temperature, and so on can all cause measurable seismic amplitude change,
which has brought great difficulty in time-lapse seismic data interpretation and possibly cause
interpretation error. In the same time, different reservoir parameters variations affect the seismic
amplitude in different ways and the post-stacked seismic amplitude difference is only the final result after
all processing, so fully understanding and particular characterization of how different reservoir
parameters affect the impedance and seismic amplitude variation is rather important in better reservoir
monitoring.
Time-lapse seismic AVO attributes analysis, as a preferable method, has great potential in
distinguishing different reservoir parameters changes. Studies in laboratory and in-situ have showed S
wave velocity and P wave velocity have different responses to reservoir oil saturation, effective pressure
and temperature changes respectively. Time-lapse seismic AVO attributes have different change likewise.
Through forward modeling, the changes of time-lapse seismic AVO attributes are obtained with the
variations of oil saturation, pressure and temperature, then variation laws are followed, which will be
helpful for actual time-lapse seismic feasibility study and data interpretation.
AVO Theory

It is well known that when the plane simple harmonic longitudinal wave beats in the interface of two
semi-infinite elastic media, the wave will not only be reflected, but will penetrate the second medium and

274
transmit in it, that is, reflection and transmission coexist. Both waves contain P wave and S wave. P wave
angle of reflection 6, and transmission anglee, , S wave angle of reflection

and transmission angle

A, are satisfied the Snell law, that is:

sine,
-1

sine, - s i n 4 - sinil,

a,

PI

The P wave velocities of medium 1 and 2 are

(1)

P2

a,,a, respectively, the S wave velocities of medium

Zoeppritz equations provide the relation between reflection and transmission amplitudes when the
incidence wave is P wave.

A, cose, - B, s i n 4

+ A, cosB, + B, sin 2, = A, cose,

A, sin 0, + B, cos A, - A, sin 8, + B, cos A, = -A, sin 0,

(2a)
(2b)

A,Z, cos24 - B,W, sin2il, - A,Z, cos2i12- B2W2sin2i12 = -A,Z, cos24

A,y,W, sin28, + B,W, cos24

+ A,y,W,

sin28, - B,W, c0s2i12 = A,y,W, sin28,

(2d)

and

The reflection P wave and transmission P wave amplitudes are A,,A, respectively. The amplitude of
incidence P wave is A,. The amplitudes of the reflection S wave and transmission S wave are B, ,B,
respectively. p, is the density of the media, and Z , ,W, are wave impedances. Study on time-lapse
seismic AVO different responses to the variation of oil saturation, pressure and temperature will be
conducted by solving Zoeppritz equations. Thus they are the theory basis of seismic AVO.
Time-lapse Seismic AVO Responses to Oil Saturation Variation
The oil saturation will decrease because of fluids replacement for water drive or gas drive reservoirs,
which will affect the seismic velocity. The velocity change can be calculated using Gassmanns equation
(Gassmann, 1951). When one reservoir has different oil saturation, Gassmanns equation can be written
as following:

275

k, is mineral bulk modulus. And kd, is dry rock bulk modulus. k,, ,k,,
fluid type 1 and 2 respectively. k,,
respectively.

,k,,

are the bulk modulus of

are the bulk modulus of rock filling with fluid type 1 and 2

is rock porosity. When the pore fluid consists of different fluid components, the bulk

modulus of the mixed fluids can be calculated according to following Wood equation (wood, 1955):

kf = l / ( C x , / k , )

(5)

k, is bulk modulus of fluid components in pore space, x, is its corresponding bulky percent.
Rock density will change when the reservoir oil saturation has changed, and this change can be
calculated from the following equation (Domenico, 1976):

Po = 4Ss,P,+ 4(1- S,)P, + (1

-@IPS

(6)

po is density of rock saturated with fluids, and p,,p,,ps are the densities of crude oil, water and
rock frame respectively. So is oil saturation.
In order to analyze the responses of time-lapse seismic AVO attributes with variation of oil saturation,
AVO curves of the reservoir describing as Table 1 are calculated in condition of water drive and gas drive
respectively. The velocity of P wave in the shale above the reservoir, Vp, is 3400m/s, the velocity of S
wave, Vs, 1750m/s, and the density is 2.35g/cc.
Table 1 Reservoir parameters

Components of rock
Quarter

Feldspar

Clay

Original

Porosity

Density

Fluid

Burden

Density

Reservoir

Proportion

oil

ofrock

ofoil

pressure

pressure

of

temperature

of gas to

specific

dcc

MPa

MPa

formation

oil in

gravity

saturation

water
dcc
70%

27%

3%

80%

30%

0.762

0.984

Through Voigt-Reuuss-Hill average method (Hill, 1952), the elastic moduli of rock framework can be
calculated. And in the calculation, the other reservoir parameters are assumed stable except oil saturation
during the reservoir development by injecting water or gas. Figure 1 shows the AVO variation curves of
P-P reflected wave and P-SV converted wave with the variation of oil saturation in water-injecting
development. Figure 2 indicates the AVO curves of P-P reflected wave and P-SV converted wave
respectively with the variation of the oil saturation during the gas-injecting (C02) development.

276

oi I sat urat i on20%

\
I

'

'

'

o l o P 3 l 4 3 E o a
incident anglee(")

0 1 O a 3 3 4 l 9 )
incident angle0

e)

(a)
(b)
Figure 1 (a) P-P wave AVO curves in the case of water drive with the variation of oil saturation.
P-SV converted wave AVO curves in the case of water drive with the variation of oil saturation.

0 1 0 a ) J ) 4 1 9 ) 6 )
i nci dent angl e8 (" )
(a)
(b)
Figure 2 (a) P-P wave AVO curves in the case of gas drive with the variation of oil saturation
P-SV converted wave AVO curves in the case of gas drive with the variation of oil saturation

(b)

Based on the above figures, the conclusion can be followed that the variation of P-P wave AVO cume is
more sensible than that of P-SV converted wave AVO curve whether in condition of water drive or gas
drive, and the variation trend of P-P wave AVO curves with variation of oil saturation in water drive is
contrary to that of P-P wave AVO curves with variation oil saturation in gas drive, and same to P-SV
converted wave AVO curves. The calculation of reservoir wave velocity indicates that, the P wave

211
velocity has increased while the S wave velocity has decreased in condition of water drive, whereas, the P
wave velocity has decreased, while the S wave velocity has increased in gas drive.
Time-lapse Seismic AVO Responses to Pressure Variation

During reservoir development, oil extraction can cause decrease of the pore fluid pressure while water
drive or gas drive will increase pore fluid pressure. Many previous studies indicate that the change of
effective pressure can affect seismic velocity. According to the study of Eberhat-Phillips etc (1989), there
are empirical function relationships as follows between S wave velocity, P wave velocity and reservoir
effective pressure, rock porosity and clay content:

V,

= 5.77 - 6.944 - 1.73fi

+ 0.446(&

100

- e-0.167pe
)

(7)

V, = 3 . 7 0 - 4 . 9 4 4 - 1 . 5 7 f i + 0 . 3 6 1 ( ~ - e - 0 . ' 6 7 p e )
100

is the clay content in rock. And

(8)

p , ,whose unit is Mpa, is reservoir effective pressure, and its value

is the difference of reservoir burden pressure and pore fluid pressure.

The fluid seismic velocity and its density will change with the variation of fluid pressure. According to
Batzle and Wang's studies (1992), there are relationships as following between oil density and pore fluid
pressure, temperature, and between velocity and pore fluid pressure, temperature:

pp = po+ (0.00277P -1.71 x

V = 2096(-

2.6 - po

- 3.7T

P 3 ) ( p 0-1.15)2

+ 3.49 x 104P

+ 4.64P + 0.0115[4.12(1.08p;'

- 1)'I2 - 1]TP

(9)

(10)

po is the density of crude oil under condition of one standard atmospheric pressure and 15.6

c.

is

fluid pressure and T is reservoir temperature. Similarly, there are relationships between density and
velocity of formation water and fluid pressure, temperature:

+ S(0.668 + 0.44s + 1x 10"[300P - 2400PS +

T(8O + 3T 3300s - 13P + 47PS)I)

pB = p,

(11)

V,

= V,

+S(1170-9.6T+0.055T2 -8.5x109T3 +2.6P-O.O029TP

(12)

-0.0476P2)+S'~5(780-10P+0.16P2)-820S2

v, are the density and velocity of pure water respectively. And S

p, ,

is salinity of formation water,

and the unit is ppm/lOOOOOO.

To study the responses of time-lapse seismic AVO curves to pore fluid pressure variation, the other
parameters of the reservoir described in Table 1 are assumed stable except pore fluid pressure. Then
according to the above relationships, the density and the velocities in the reservoir can be obtained with
the variation of pore fluid pressure. Figure 3 shows the AVO curves of P-P reflected wave and P-SV

27 8
converted wave respectively with the variation of the reservoir pore fluid pressure.

a@-

incident angle8

e)

fluidpressure20&

incident angl eQ )

Figure 3 (a) P-P wave AVO curves with the variation of pore fluid pressure. (b) P-SV converted wave
AVO curve with the variation of pore fluid pressure
Figure 3 indicates that P-SV converted wave AVO curve is more sensible to pressure variation than P-P
wave AVO curve, especially when the incident angle is more than 30" . With the increase of incidence
angle, the difference of P-P wave reflection coeficients under different pore fluid pressure becomes
smaller. These laws are different or contrary to that in the case of oil saturation variation. The calculation
of reservoir seismic velocities shows that the S wave and P wave velocities increase simultaneously with
the decrease of pore fluid pressure (or with the increase of effective pressure).

Time-lapse Seismic AVO Responses to Temperature Variation

The effects of small temperature variation to velocity and density of reservoir fluids can be calculated
from the equations (9), (lo), (11) and (12). The effect of small reservoir temperature variation to rock
skeleton modulus can be neglected since it is so small. However, P wave and S wave velocities in rocks
saturated with heavy oil can be affected strongly by the temperature variation, and P wave and S wave
velocities in rocks saturated with heavy oil decrease drastically upon increase in temperature due to a
transition of heavy oil from a solid state to liquid state and a decrease of viscosity with temperature
increase.
Gassmann's equation has been extensively applied for calculation of velocities in rocks saturated with
fluid under the assumption that the shear modulus of pore fluid is zero, that is, the shear modulus of dry
rock is equal to that of rock saturated with fluid. However, the heavy oil has bigger shear modulus and
bulk modulus when the temperature is lower than the melting point, and has smaller bulk modulus and
nearly zero shear modulus when the temperature is higher than the melting point for the transition from
solid to liquid state. In the case, the Bound Averaging Method (Marion and Nur, 1991) can be applied to
calculate the elastic moduli of rocks saturated with heavy oil.

279
For a given rock, the upper and lower bounds of its elastic moduli carries information regarding the
stiffness of the pore space, that is closeness to the upper bound may correspond to a rock with stiffness of
the pore space whereas, proximity to the lower bound may be representative of a rock whose porosity is
uncemented grain contacts. Therefore the elastic moduli of rock can be described by averaging bounds
using the following relation:

K=K-+w(K'-K-)
G = G - + A(G'

(13)

G-)

(14)

K ' , K - are the upper and lower bounds of rock bulk modulus respectively. G',G-are the upper and
lower bounds of rock shear modulus respectively. w and
are the weighting factors, which take
values from 0 to I , and are measures of pore stiffness. For a given rock, upper and lower bounds will be
dependent on the elastic moduli of pore-filling material, but the type of pore-filling material will not
affect the rigidity of the pore structure. Thus, the weighting factors will not be affected by the type of
pore-filling material and is assumed to be invariant to a given rock. Hence, the elastic moduli of rock
saturated with pore-filling material of type 2 can be estimated from the measured elastic moduli of same
rock saturated with pore-material of type I , and the equation can be written as following.

K, =K,+(K,-K,)(

GI = GI- + (Go - G;)(-

K: - Kl)-I
K,' - K ,
G: -G,G,' -G,

(16)

)-I

According to the studies of Hashin and Shtriman (1963), the upper and lower bulk modulus bounds of
rock can be calculated using the following functions:

KP>
1-I
4
K , +-G,
3

K' = K ,

+ (1- @ ) ( K , - K p ) ( l+ @(&

K- =K,

+ (1 - @ ) ( K , - K,)(l+

(17)

KP>
)-I
4
K , +-G,
3

@(%

(18)

And the upper and lower shear modulus bounds of rock can be calculated using the following functions:

G' = G,

+ (1 - @)(G, - G,)(l+

@(G, - G, 1
G, ( 9 K , + 8G,
Gp + 6 ( K , + 2G,)

1-I
1

(19)

K ,G , are the bulk and shear modulus of pore-filling material respectively. K , ,G, are the bulk and

280
shear modulus of mineral grain respectively, and 0 is the porosity of rock. According to equation (15),
(16), (17) and (18), the elastic moduli of rock saturated with any pore-filling material can be calculated.
To study the time-lapse seismic AVO response with the variation of temperature in the stream-drive
heavy oil reservoir, simulation is conducted for the reservoir characterized in Table 2. The P wave
velocity in the shale above the reservoir, Vp, is 3400m/s, the S wave velocity, Vs, 1750ds and the
density is 2.35gkc.
Table 2 Heavy oil reservoir parameters

Components of rock

quarter

feldspar

clay

Original

Before solidification (72'

oil

point of

saturation

heavy

of rock

oil

Bulk

Shear

Density

modulus

modulus

g/cc

GPa

GPa

2.18

After solidification ( 5 5 '

"C
70%

27%

3%

100%

65

3Ooh

0.97

*,solidification

before heavy oi I
sol idificat i on

.-t.

8 QB
.5 Q10.
c

P QP
t i = .

beforeheavyoil
sol idif icat ion

.-

E 414

?!?

QB

\\

44

afterheavyoil
',,solidification,'

'*.-

'I

/
-*#

0%

- c m l ~ , , l ' l ' , ~

Figure 4

0 1 0 a ) J ) 4 ) 9 )
incident anglee f')

(a) P-P wave AVO curves before heavy oil solidification ( 7 2 C ) and after heavy oil

solidification (55 'C ). (b) P-SV converted wave AVO curves before heavy oil solidification (72
and after heavy oil solidification (55

c)

c ).

Figure 4 indicates that the temperature variation in heavy oil reservoir can affect P-P wave AVO

28 1
curve and P-SV converted wave AVO curves drastically due to the transition of heavy oil from solid state
to liquid state, and the P wave and S wave velocities in rock saturated with heavy oil both decrease in big
proportion.
Conclusion

Time-lapse seismic P-P wave AVO curves and P-SV wave curves have different responses to the
variation of oil saturation, pore fluid pressure and temperature, that is, the P-P wave AVO curves have
better sensitivity to the oil saturation variation, the P-SV wave AVO curves better sensitivity to the pore
fluid pressure (effective pressure), and P-P wave and P-SV wave AVO curves change drastically for
heavy oil melting with the temperature variation in stream drive reservoir. Thus, according to the
different changes of time-lapse seismic AVO curves, the seismic amplitude differences caused by the
variation of oil saturation or effective pressure can be distinguished, and the range affected by stream can
be defined in heavy oil reservoir. The quantitative analysis of time-lapse seismic AVO variation in some
oil field can be conducted by simulation using logging data, seismic data and production data, and so on,
which is important to time-lapse seismic feasibility study and interpretation.
Acknowledgments

We thank National Natural Science Function of China for the imbursement to the project (Grant No.
40174037). We also thank Ms. Zhang for his support of translation of the paper.
Reference

Sung, H, Y., Seongsik, Y., et al, 2000, Effects of pressure and fluid saturation changes on time-lapse
Expanded Abstracts of 70thAnnualInternet SEG Mtg.
AVO response
2 Ying, Z., Laurence, R., 2002, Time-lapse well log analysis, fluid substitution and AVO
SEG
International Exposition and 72"dAnnual Meeting.
3 Telford, W. M., Geldart, L. P., et al, 1990, Applied Geophysics Second Edition. New York:
Cambridge University Press, 155-156.
4 Gassmann, F., 1951, Elastic waves through a packing of spheres, Geophysics, 16: 673-685.
5 Wood, A. M., 1955, A textbook of sound. The MacMillan. Cvo., New York, 360-360.
6 Domenico, S. N., 1976, Effect of brine-gas mixture on velocity in an unconsolidated sand reservoir.
Geophysics, 41(5): 882-894.
7 Zhang, G.J., and Hu, T.Y, 2002, Seismic wave AVO and formation lithology analysis. Oil
Geophysical Prospecting, 37 (6): 579-584.
8 Hill, R., 1952, The elastic behavior of crystalline aggregate, Pro. Phys. SOC,London, A65: 349-354.
9 Eberhart-Phillips, D., Han, D-H., and Zoback, M. D., 1989, Empirical relationships among seismic
velocity, effective pressure, porosity, and clay content in sandstone. Geophysics, 54: 82-89.
10 Batzle, M., and Wang, Z., 1992, Seismic properties of pore fluids. Geophysics, 57(11): 1396-1408.
11 Marion, D., and Nur, A,, 1991, Pore-filling material and its effect on velocity in rocks, Geophysics,
56: 225-230.
12 Hashin, Z., and Shtrikman, S., 1963, A variational approach to the elastic behavior of multiphase
materials. J, Mech. Phys. Solid, 11: 127-140.
1

u].

u].

282
MODELING HIGH-FREQUENCYSEAFLOOR VOLUME BACKSCATTER BY
SHELL FRAGMENT DISTRIBUTIONS
A.P. LYONS
Penn State University Applied Research Lab. State College, PA.
E-mail: ap12 @psu.edu
There has been much recent activity on developing and testing high-frequency seafloor volume
scattering models and on the related high-resolution characterization of seafloor volumes. This paper
will address another plausible volume scattering mechanism that has not yet received much attention
- distributions of broken shell fragments in sediment (commonly referred to as shell hash). The shell
fragments are modeled as non-aggregating spherical scatterers, and the spatial distribution of the
shell pieces is determined using the Percus-Yevick packing factor. Computer simulations of the
multiple scattering and a single scattering model are used to study the power backscattered by shell
hash sediments as a function of the volume of scatteren and the frequency of the incident wave (10100 kHz). Parameter values for simulations are obtained by stereological analysis of x-ray computed
tomography scans of sediment cores.

Introduction

In addition to rough interface scattering, a complete description of backscattering from

the ocean bottom must include scattering from the volume beneath the water-sediment
interface. The volume scattering component might include particle scattering andor
continuum scattering. The sediment volume matrix can be thought of as a random
continuum whose properties vary randomly and continuously in space. The continuous
property variations of the sediment continuum can be caused by small scale layering of
the subbottom or variations in sediment water content. Discrete scatterers are random
distributions of particles within the sediment volume such as bubbles, buried rocks or
shells, and marine organisms or their burrows within the sediment. A considerable
amount of work has been reported concerning acoustic sediment continuum scattering,
but more limited progress has been made in the study of scattering from discrete
scatterers within the sediment such as shell fragments, common in many shallow water
areas.
In the next section, theoretical and numerical techniques for predicting volume
backscatter from seafloor environments in the frequency range from 10-100 kHz are
examined. Specifically, this section discusses the adaptation of multiple scattering
techniques to scattering from dense shell hash sediments and a comparison of this model
to a single scattering model. The third section will consist of a discussion of ground truth
data collected as part of the Coastal Benthic Boundary Layer experiments in Panama
City, Florida. Discussed in this section will be the use of CT scans and stereological
methods to determine the distribution of shell pieces in Panama City sediments. The third
section will also include numerical simulations from seafloor scattering models, the
parameters of which have been constrained by ground truth information obtained in the
shell hash filled environment. Addressed specifically is an examination of the effect of
scattering by a lognormal distribution of shell pieces on the total scattering response of
the sandy sediments of Panama City. The simulations are useful in isolating the physical
scattering mechanisms that dominate scattering and in examining the effectiveness of the
characterization and modeling.

283
2

Scattering from discrete shell pieces in sediment

This section will address two different ways of thinking about discrete scatterers in
sediments. A multiple scattering model for distributions of discrete scatterers (shell
pieces in our case) will be compared to a single scatter formulation. Comparisons will be
made between simulations from each model in order to examine the applicability of the
multiple scattering model. If the seafloor shell piece distribution is relatively tenuous,
then single scatter theory should be adequate to explain many observations. In the single
scatter approximation, the incident wave is essentially unchanged as it propagates through
the scattering medium, and the particles are assumed to interact with the incident pressure
field only once, i.e., all double and multiple scattering are assumed negligible. But, for
high concentrations of shell pieces, the multiply scattered waves must be included.
Wave scattering in random media is usually classified into two different categories:
continuum scattering and discrete scattering. The two kinds of scattering theories were
developed independently and are quite different. In the theory of continuum scattering,
the backscattered waves are expressed in terms of the correlation or power spectral
density functions of compressibility variations, yK = ( K - K~)/$,and density variations,
yp= (p - pJ/p, in the medium where po is the spatial average density in the random
medium and
is the spatial average compressibility of the medium. Weak property
variations are usually considered for the continuum formulation. Applied in this
derivation are the assumptions that the scattering volume is in the far field and the
variations in magnitude of the density and compressibility fluctuations are sufficiently
small for the Born approximation to hold, i.e. the incident plane wave pressure is
substituted for the total pressure. The scattering wave number, K = k (i -8), describes
the spatial periodicity of the density or compressibility variations. The scattering wave
number is related to the angular frequency of the acoustic wave and the mean sound speed
and is equal to 2wlc for backscattering.
It is assumed that the size of the scattering volume is greater than the correlation
distance in the medium. Noting that the Fourier transform of the correlation function,
Bdr), is the spectral density (a statement of the Wiener-Khinchin theorem), the
differential cross section can be expressed in terms of the spectral density function of the
medium variations, SLK), for a medium that is statistically homogeneous and isotropic
(for more detail see [ 13) as

where

S y ( K ) = G1j B y ( r ) e x p ( i K . r ) d V .
V

Eq. ( 2 ) states that once a form for the correlation function, Bdr), is specified a
spectral density can be found which can be used in Eq. (1) to give the scattering cross
section per unit volume for the inhomogeneous medium. The derivation of continuum
scattering expressions given above provides a clue for an alternative method of evaluating
discrete scattering effects. This alternative method requires a proper correlation function
that can describe particle distributions and the sharp gradients in properties at the
boundaries of particles in a discrete random medium. In the following, only spherical

284
particles of one size are considered and the fluctuation term,

= y, (r)+y, (r)cos(B,P),

including variations of both compressibility and density is used. The complete derivation
of the correlation function for this type of discrete particle fluctuation term is given in
detail in [2]. Only the relevant results will be given below.

2. I

Correlation Function

As shown in [2] an expression for the correlation function, Bg(r-r),can be obtained by

series expanding the pair distribution function around a volume concentration, = 0 and
truncating the terms with order higher than one in the series. The Fourier transform of the
correlation function, as in the continuum derivation yields an expression for the power
spectral density function:

x,

For spherical particles with radius, a, explicit forms for SXk), S@(k), and S,,(k) have
been derived as [2]:
S, ( k ) = ~127T
[sin(ka)-(ka)cos(ka)],

(4)

where
Go ( k ) = --[sin(2ka)-(2ku)cos(2ka)],
3
k3a3

(7)

G, (k) = -[-24(ka)sin(4ka)+(24k2a2
3
-6)cos(4ku)+(8k3u3
4k6a6
+ 12ku)sin (2ka)+ (20k4a4+ 12ka + 6)cos (2ku)]

(8)

and

and are related to the Fourier transform of the pair correlation function. Equations (4) (8) constitute the bases for computing the scattering contribution from particles including
multiple scattering.
Using Eq. (3) and Eq. (1) the scattering cross sections for distributions of
particles can now be computed. For these calculations, particles are assumed much
smaller than an acoustic wavelength so that the Rayleigh scattering approximation is
valid. In the Rayleigh approximation, the total field inside a small particle is replaced by
a uniform field [l]. Yang [2] has shown that in the Rayleigh approximation and for an
isotropic,random medium the model as developed is identical with the result of Twersky
131.

285
2.2

Scattering Theory Comparisons

The validity of the continuum model was tested by comparison with the traditional theory
of independent scatter from small spherical non-resonant particles in the Rayleigh
approximation, 2 >> a, where a is the particle radius. This was done in order to define
when a multiple scattering model might be necessary, i.e. at what concentration of shell
particles do we need to abandon simple single scattering in favor of a multiple scattering
model. This information will be used later with real sediment shell distributions in order
to assess the importance of multiple scattering. For a non-resonant sphere the
backscattering cross section is given by:

As in multiple scattering, this equation describes scattering due to compressibility

differences and density differences between the particle (defined with subscript p ) and
surrounding medium (defined with subscript s). Total scattering from N particles is then
given by the sum of the individual particles, ptotal = N ps if the particles scatter
incoherently.
In the following simulations the background sediment density, ps, was set to 2100
kg/m3, that of a sand, the particle (calcium carbonate shell piece) density, p,,, to 2700
kglm3, the sediment compressibility, &, to 1.428 x 10.'' Pa-', and the particle
compressibility, K,,,to 3.125 x lo-'' Pa". A single size particle of 1.5 mm was used in the
simulations. For the frequency range of interest this size particle has a ka < 1, so that the
Rayleigh approximation should be valid. The effect of changing volume concentration
holding the frequency constant at 75 kHz is shown on Fig. 1 for comparison. The two
theories start to diverge almost immediately, at a volume concentration of about 0.02
where the results from single scattering theory become increasingly larger than those
from the multiple scattering theory. The decreased scattering cross section of the multiple
scattering theory estimates can be attributed to the effect of assuming the particle
positions are correlated at the higher volume concentrations. When examining real
sediment shell hash concentrations, the single scattering model will be assumed valid if
the volume concentration of shell pieces is approximately 0.01 or less.

286
0.12

0.1
h

4
Y
7

0.08
.c
0

a,
(0
v)

0.06
0
n,

c
._
L

B
d

0.04

cn

0.02

0 '

Volume Concentration
Figure 1. Comparison of single scattering and multiple scattering estimates of volume scattering cross section
versus volume concentration.

The Coastal Benthic Boundary Layer Panama City, Florida experiment

During August of 1993, the Naval Research Laboratory (NRL) Coastal Benthic Boundary
Layer Special Research Program (CBBLSRP) conducted experiments in the northeastern
Gulf of Mexico 23 nmi southeast of Panama City, Florida [4). The seafloor in this area
consists of a hard sandy bottom that includes many shell fragments (shell hash) so that the
two possible contributors to acoustic scattering in this region were volume scattering from
the discrete shell pieces and roughness scattering from the water-sediment interface.
Interface scattering is almost always assumed to dominate in sandy seafloor environments
and comparisons between scattering models and experimental data confirmed this for
seafloor scattering at 40 kHz at the Panama City CBBL site [5]. This is for two reasons:
little energy enters the bottom because of the high impedance of the sandy sediment, and
a sandy seafloor is usually rougher than a silty seafloor [5]. Sediment ground truth cores
were taken in the same location as the acoustic measurements in this experiment and
provided an excellent opportunity to examine this assumption for a sandy shell filled
sediment at higher frequencies.

3.I

High Resolution Characterization of Sediments

A very important parameter that is needed to describe the sediment structure of Panama
City in which there are dispersed shell fragments in a sand matrix is the number of shell
fragments per unit volume, n. This is not something that can be directly determined with
fragile shell fragments, as a sieving process to determine size distribution has the

287
possibility of breaking larger fragments into smaller pieces thereby distorting the
estimate. The method of stereological examination of cores that had been x-ray computed
tomography (CT) scanned was chosen to estimate the shell size distribution in the cores.
This is possible because of the extremely thin horizontal resolution of the scanner, 0.25
mm for a vertical slice of data, and because the X-ray CT data can be directly related to
values of density [6,7].
Before a stereological analysis was carried out a threshold was determined to
separate shells from matrix. This threshold was set at a density of 2.4 g/cm3.
Thresholding is possible because of the strong gradient in density due to a shell particle;
the sandy sediment has a density of about 2.1 g/cm3 while the shell pieces have a density
of 2.7 g/cm3. Fig. 2 shows a density image from one of two sub cores of box core 413
taken in the Panama City sediments during the CBBL experiment. The shell pieces are
evident as bright white in the colors of the sediment background. Once the threshold was
set, the image was converted into a binary format with values of one being the shell
pieces and values of zero being the background sediment matrix. In order to increase the
confidence in the measured size distribution, six vertical planes from each of two sub
cores of core 413 were examined. The size distribution of feature profiles was obtained
from the binary images by using the public domain NIH Image program.
0 .o

20.0
40.0
60.0

80.0
100.0

120.0
140.0

0.0 25.0 50.0 75.0

Across Core (rnrn)
1 .o
2.0
Density (grn/crn3)
Figure 2. Density image of core 413. Shell pieces are white in this image.

Fig. 3 shows the distribution with the size axis being the log of the particle radius.
The shape of this curve corresponds closely to a lognormal distribution (solid line). The
lognormal distribution often occurs in particle sizes [8]. The function

288
will be used in calculations of scattering strength when integrating over the particle size
distribution. In this equation A is the offset, p is the peak value, is the mean and 0is
the standard deviation.

.5

Figure 3. Calculated distribution of core 413 shell piece radius and associated lognormal curve fit.

3.2

Seafloor Scattering Simulations

Integration of the function in (10) over the range of particle radii seen in core 413 yields
over 17,000 shell fragments per cubic meter of sediment. While this may seem a very
large number, the volume concentration of shell pieces in core 413 is actually less than 1
percent. The criterion established in the preceding section was that the volume
concentration exceeds approximately 1 percent before multiple scattering effects became
important for shell scattering in sandy sediment. Thus, in calculations of scattering
strength the single scattering model of Eq. (9) will be used along with the distribution
given by Eq. (10). The seafloor interface roughness scattering model used in the
following comparisons is a perturbation approximation surface scattering model that has
been used with success in previous scattering experiments [5]. The model and
assumptions used in their derivation are discussed in [9,10]. The required parameters and
their values used in the interface scattering model are given in Jackson, et al. [5] taken
from the Panama City site at the time of the CBBL experiment.
A simulation of the total backscattering strength versus grazing angle for a sandy
shell filled sediment is presented in Fig. 4. The solid line on the figure represents the total
scattering strength, the dash-dotted line the component due to interface roughness, and the
dotted line shows volume scattering component due to the shell piece distribution. The
top panel in the figure shows the simulation predictions for a frequency of 20 kHz while
the lower panel sho'ws predictions for 100 kHz. Volume scattering due to the shell pieces

289
is very low at 20 kHz and contributes very little to the total scattering strength. At 100
kHz and above the critical angle (approximately 30"), scattering from the shell hash
contributes up to 4 dB, an amount that should be observable in experimental data. Below
the critical angle, the volume scattering component is extremely small and doesn't
appreciably contribute to the total scattering strength due to the lack of penetration for the
range of frequencies studied in this paper.

g -20

10

15

20

25
30
Grazing Angle (deg)

35

40

Figure 4. Predicted volume (dotted line), interface (dash-dotted line) and total scattering strength (solid line)
versus grazing angle at 20 kHz (top) and 100 kHz (bottom).

Summary and Conclusions

Scattering from a distribution of shell material buried in acoustically hard sediments was
put forth as a mechanism for the volume component of total backscattering strength at the
CBBL experimental site near Panama City, Florida. Comparisons between single and
multiple scattering models show that multiple scattering from shells in a sandy sediment
can be neglected in favor of simpler single scattering theory for volume concentrations
less than 1 percent. The actual shell hash size distribution for the CBBL site was obtained
by stereological examination of x-ray computed tomography data sets of seafloor cores
taken at the study site and was determined to be approximately log-normal. The volume
concentration of shell pieces in the sediment ground truth cores was found to be less than
1 percent, justifying the use of a single scattering model instead of a more complicated
multiple scattering model. A simulation of seafloor backscattering was carried out that
included a volume scattering component due to the measured distribution of shell pieces.
When compared to predictions of interface scattering strength using parameter values
obtained from the same area, volume scattering from buried shell pieces increased the
scattered level above predictions based solely on roughness scattering. Simulations

290
carried out with the shell hash scattering model suggest that, above the critical angle,
scattering from the shell pieces may dominate the scattering from the hard, rough watersediment interface.

Acknowledgements

Thomas Orsi obtained the x-ray CT measurement data and performed the conversion to
density values for the sediment core analyzed in this work. The Office of Naval Research
under Grant N00014-01-1-0352 supported this work.

References

1. Ishimaru, A., Wave Propagation and Scattering in Random Media (Academic Press,
New York, 1978).

2. Yang, C.C., Evaluating particle scattering effects by using turbulence scattering

theories, IEEE Tran. on Ant. and Prop. AP-34 (1986) pp. 1189-1195.
3. Twersky, V., Multiple scattering of waves by periodic and random distributions. In
Electromagnetic Scattering, ed. by Piergiorgio L.E. Uslenghi (Academic Press, New
York, 1978) pp. 221-251.
4. Richardson, M.D., Investigating the coastal benthic boundary layer, EOS Trans. AGU
75 (1994) pp. 201-206.
5. Jackson, D.R, Briggs, K.B., Williams, K.L. and Richardson, M.D., Tests of models
for high-frequency seafloor backscatter, IEEE J. Ocean. Eng. 21 (1996) pp. 458-470.
6. Orsi, T.H., Computed Tomography of Macrostructure and Physical Property
Variability of Seafloor Sediments, (Ph.D. dissertation, Texas A&M University,
1994).
7. Orsi, T.H., A.L. Anderson and Lyons, A.P., X-Ray Tomographic Analysis of
Sediment Macrostructure in Eckernforde Bay, Western Baltic Sea, Geo-Marine
Letters 16 (1996) pp. 232-239.
8. Russ, J. C., Practical Stereology (Plenum Press, New York, 1986).
9. Jackson, D.R. and Briggs, K.B., High frequency bottom backscattering: Roughness
versus sediment volume scattering, J. Acoust. SOC.Am. 92 (1992) pp. 962-977.
10. Lyons, A.P., Anderson, A.L. and Dwan, F.S., Acoustic scattering from the seafloor:
Modeling and data comparison J. Acoust. SOC.Am. 95 (1994) pp. 2441-2451.

29 1

BRAIN TOPOGRAPHY OF PERCEPTION OF TARGET AND

NON-TARGET ACOUSTIC SIGNALS
SERGUEI LYTAEV
S.M.Kirov Military Medical Academy, St.Perersburg, Russia
lvtaev serzei@mail.rii
This work relates to Department of the Navy Grant N00014-03-1-4012 issued by the Office of Naval
Research International Field Office. The United States Government has a royalty-free license throughout the
world in all copyrightable material contained herein.
In this paper the relationships between of middle- and long-latency acoustical evoked potentials on model of
selective attention are estimated. Topographical brain mapping in this situation is model for approximation of
direct and back influences in neural networks at processing the acoustic information. The mechanisms of
perception of target and non-target acoustic signals in the brain on various epoch of the analysis are estimated.

Introduction

During last more than 100 years human brain is classical model of the black
box. We can estimate parameters of acoustic signals acting to the brain. On the other
side it is possible to register some target data of the analysis of the acoustic information.
The answer reactions of the brain, namely, auditory evoked potentials (AEPs) (eventrelated potentials) are the target data. The modern methods of brain mapping allow to
investigate the spatial-temporary characteristics of reactions of the brain in reply to
acoustic signals, and also to estimate mechanisms of processing of the acoustic
information [3, 8, 10, 181.
The topical organization of acoustical system testifies about multichannel transfer
of the information to higher levels of the brain. There are specialized channels connected
with local acoustical fields of the receptors. These channels are capable to transform a
plenty of primary signs of the acoustic signal [13, 17, 201. Thus, the formation of
selective reactions of neural networks on the complex acoustic signals is caused by
interaction of neurons of the different channels.
If in visual system the functioning of the neural networks - detectors and
mechanisms of the invariant identification of images does not cause doubts, in acoustical
system these mechanisms dont work. Now two basic variants of coding of complex
signals are considered. First of them assumes localization of detectors in others not
sensory (not acoustical) structures of the brain. The second way (not necessarily
alternative) deciphers the complex acoustic information on the basis of integrative activity
of the neurons [9, 121.
The researches of mechanisms of the early filtration of acoustical signals acting
in the brain, allow to speak about participation of the brainstem and of the undercortex
formations in selection of the information. Researches of the short-latency and middlelatency AEPs specify an opportunity of maintenance of mechanisms of the selective
attention by switching nucleuses of the brainstem [4,6, 19,221. The representations about
back regulation of the bottom formations from the party brain cortex are actively
developed [ 18,21,221.
The purpose of research is on the background study of brain topography of
middle- and long-latency acoustic evoked potentials to establish features of mechanisms
of perception of the target information on various intervals of time of an acoustic signal.

292

Methods

26 healthy examinees (man, age 20-22 years, which are taking place training on a
hydroacoustics specialty) with normal parameters of thresholds of absolute acoustical
sensitivity are surveyed.
Registration of AEPs was carried out with the help of the computer neurosystems
Brain surveyor and Viking IV according to the International requirements [ 13. As the
showed (deviant) the stimulus were signals with a level of sound pressure (LSP) 70 dB,
frequency of tone 1000 Hz and duration 70 ms. They were binaural submitted through
headphones TDH 39P. The interval between stimuli was 1000 ms. AEPs were averaged
on 50 accumulations. As target (relevant) signals were used sound ones with frequency of
tone 2000 Hz (LSP and duration same). They were binaural submitted through
loudspeaker of the phonostimulator Sanei. An interval between target stimuli was 3000
ms.
In the first series 50 rhythmic signals through headphones were submitted to the
examinee, and in the second ones - on a background of these 50 (deviant) stimuli 17
targets (relevant) stimuli were showed. It was necessary, according to the instruction, to
count up stimuli with the subsequent verbal answer. At all 26 examinees by amount of
target signals the answers were correct.
AEPs were analyzed by the data of the topographical brain mapping and by the
estimation of the spatial - temporary characteristics of most stable waves of the AEPs [ 18,
191 in seven sites - (P4, P3, C4, Cz, C3, F4 and F3). For the choice of significant
informatively parameters was used the step-wise discriminant analysis. An estimation of
parameters carried out by criterion of F-statistics (F > 4.0).

Results

The data of the registration of the brainstem (short-latency) AEPs (BSAEPs) show,
that the acoustic signal is involved in processing by structures of the brain through 3 ms
from the moment of its presentation. Usually on this analyzed epoch (3-10 ms) authors
allocate 6-8 components [ l l , 151, in our case - 6 waves (fig. 1).
Figure 1. The components of the BSAEPs, IVI. Epoch of the analysis - 10 ms; on line of
ordinate - in 1 cm 1 uV.

Topographical mapping of the brain carried out on analyzed epoch 10-400 ms.
Through 16-18 ms from the moment of presentation of signal the greatest difference of
potentials at target stimulation is marked in frontal areas of the left hemisphere, while in
occipital and parietal sites the insignificant symmetric activation (fig. 2a, b) is marked.

293
Among middle-latency AEPs (MLAEPs) most stable in both tests are 4 waves - P16,
N30,P40and N60, They appropriate according to the classification of T.Picton et al. [ 151 to
components Po, Na, Pa and Nb. The amplitude-temporary characteristics of the complex
Pa/Nb dont differ in variants of testing (Fc4.0, fig. 3) significantly. Some features are
observed in formation of waves Po and Na at relevant stimulation. The amplitude of Po in
parietal (P3 and P4), and also in left hemispheric C3 and F3 sites (F > 4.0) grows
significantly.
The similar tendency is kept and on the time interval of formation of a wave Na in
left P3 and F3 (F > 4.0) points of registration. In conditions of submission of the relevant
signals the tendency to decrease of peak latency (PL) of wave Na is observed, that is
shown in sites P4 and O2significantly (fig. 3). The brain mapping (fig. 2b; 26 ms) shows
distinct symmetry with the maximal activation of occipital-parietal areas, to a lesser
degree of central departments and, practically, absence of electrical dipoles in frontal
areas of the cortex.
A

Figure 2. The consecutive brain maps at marks of AEPs: a - non-target, b - target stimulation.
Epoch of the analysis - 400 ms. Figures - time, ms.

The importance of positive components of AEPs, settling down above zero line
further is resulted only, as statistically significant of differences among negative waves
on epoch 100-400 ms it was not marked.
Figure 3. The average meanings of amplitudes and of
peak latencies of components of AEPs at non-target
(continuous) and target (dotted line) stimulation.
Badges (+) and (*) - significant (F > 4.0) accordingly
for amplitudes and PL. P4, P3, ..., 73 - sites on
system 10/20.

As a rule, the positive wave of AEPs

with PL 90-110 ms has the maximal
amplitude among all fluctuations of
acoustical EPs and the most stable display.
In the present research the amplitudetemporary
characteristics
of
this
component have the most expressed
distinctions practically in all areas of a
brain. They have two tendencies. First, PL
decreases in test with a relevant signal in
back and central areas of the brain.
Secondly, in all sites (F > 4.0) the
amplitude of component Pg0 (fig. 3) is
reduced. The marks of these waves gives
in both cases svmmetric maDs with more

294
appreciable activation of the cortex in a situation without target signals (fig. 2a, b).
PL of wave PI40 is significant higher (F > 4.0) in all sites in conditions of deviant
stimulation. At action of the relevant signals the increase of amplitude Pi40 in C3 and Cz
(F > 4.0) is marked, in other areas of the brain distinction are statistically insignificant
(fig. 3).
The temporary parameters of late waves (200-400 ms) are leveled in both tests.
Only PL of a wave P250is significant lower at action of the relevant signals (F > 4.0).
Simultaneously generally the amplitude of this component in conditions of a target task is
reduced. Significant differences are observed in left P3 and C3 sites and in vertex (fig. 3).
The marked facts are reflected in brain mapping (fig. 2a, 220 ms; fig. 2b, 226 ms). At
performance of the target task the hemispheric symmetry of processes of excitation pays
attention appreciable.
More expressed symmetric excitation in test with relevant signals is observed at
mapping of the component with the latent period 300-350 ms. The analysis of the spatial temporary characteristics in this task marks significant simplification of amplitude of this
wave in all sites, except for parietal (F > 4.0; fig. 3).

4 Discussion

2D and 3D+t models of information processes of the brain in modern researches

are submitted by the mapping of the evoked potentials [3, 8, 191, by dipole localization
method (DLM) [ 5 , 12, 17, 201 and by the method of construction of the 3-channel
Lissajous' trajectory (3-CLT) [7, 141. DLM represents algorithm of account of
localization and direction of equivalent dipole for any temporary point on multichannel
record from the surface of the head. The method 3-CLT assumes the usage of dipole
model [5, 121, however, diagram differs from MDL. 3-CLT represents a method of
definition of the greater information about neural generators ensuring occurrence of the
brainstem EPs, in comparison with multichannel methods of registration. Such analysis is
applicable to neural networks, where the groups of neurons, located in the certain
distance-spatial direction (sensory systems, generators of a alpha-rhythm etc.) are
selectively activated [7, 141.
The following classification of generators of negative waves BSAEPs now is
considered as the most accepted [ l l , 12, 161: I and I1 of waves - acoustical nerve
(fluctuation of potential in a snail and fluctuation at an input of an acoustical nerve in
cochlear nucleus accordingly), I11 a wave - level of trapeze body, IV and V of a wave level of lateral lemniscuses and its dorsal nucleus, and wave VI - level of a back hill. The
positive fluctuations of potential between waves I1 and 111, I11 and IV, V and VI are
presumably generated at a level of cochlear nucleuses, of medial nucleus of top olive and
back hill accordingly.
Some researchers believe, that the effects of change of activity of structures of an
acoustical input are not connected to processes of attention [6]. At the same time, there
are available data, received at registration of electrocochleargramm about significant (in
10 times more, than at registration of BSAEPs of a far field) increase of amplitude of the
first wave of BSAEPs at maintenance of tasks of selective attention [18]. On the basis of
the carried out researches a line of the authors states the assumption about possible
centrifugal influence on brainstem structure, which, as the consequence, finds reflection
in the spatial - temporary characteristics of BSAEPs [22].

295
Data about exact localization of separate components of MLAEPs (10-60 ms) are
more inconsistent. The most settled point of view is considered, that MLAEPs occur on
an interval of time between early components and late cortical cognitive waves [15, 161.
According to it MLAEPs concern both to entrance transmitters of acoustical system, and
to processes of subcortical delay [3], and also to the answers of primary acoustical cortex
[ 181. Results of experiments on study of early mechanisms of selective attention can serve
as the confirmation for the benefit of neural nature of MLAEPs [6, 221. On the basis of
the received data the authors have made the conclusion about an accessory of components
of MLAEPs to mechanisms of selective attention.
The increase of amplitude of MLAEPs on an interval of time 15-40 ms is
registered in our researches at action of target signals is observable, mainly, in parietal
both left temporal and frontal areas of cortex. This fact is regarded from positions of back
cortifugal influences of these departments on brainstem structure in maintenance of
mechanisms of selective attention.
The fact, that the processes of selective attention can render influence on sensory
transmission in peripheral departments of the acoustical system at an animal, more than
30 years back for the first time were marked. Subsequently were received the authentic
proofs for the benefit of modulation of selective stimulation at a level of eighth pair of
cranial nerves and of switching nucleuses of a trunk [18]. Also the registration of the
single evoked activity in primary acoustical cortex of the cats and monkeys has allowed
to find out the selective control above patterns of neural unloading as functions of
relevant stimulus.
The data about of corticofugal modulation of structures of an acoustical input at the
man have appeared much later. The messages on registration MLAEPs at mixed
stimulation by visual and acoustical signals specify increase of their importance on an
interval of time 20-50 ms [6]. The increase of probability of modulation of subcortical
and of cortical flows by means of back corticofugal mechanisms is shown. Such
assumption proves to be true by the data, that the conductivity through sensory-specific
circuits of some brainstem reflexes can be modulated by attention selectively [6].
In last 5-8 years are received the interesting data about the origin of waves on epoch
of the analysis 20-30 ms. These waves correspond to high-frequency activity of the brain
- 40 Hz. Is established, that these oscillations coordinate activity of acoustical system
down to braking of acoustical cortex. However these oscillations are under modulation of
deep acoustical structures [2, 211.
If the changes of MLAEPs on an interval of time 15-35 ms according to our data
have authentic character, the spatial - temporary parameters P40and N60practically do not
differ in both tests. These facts are regarded from the point of view of identity of
mechanisms ensuring transformation of acoustical signals of different complexity. It is
probably also the absence of influence of higher formations on this interval of time on
processes of attention. The brain mapping shows symmetric distribution of dipoles at
marks of PL 62 ms, mainly, in forward departments of a brain.
The distinct symmetry in maps of the brain is marked during the marks of a
component PWIOO
in both cases. At the same time the appreciable reduction of this wave
in conditions of presentation of a target task in all departments of a brain is traced.
Frequently components of AEPs, which are generated in acoustical cortex on an
interval of time 70-100 ms, are designated by concept of mismatch negativity [18, 221.
From the point of view of the authors, who allocate them, these waves reflect automatic,
not connected with function of selective attention process of comparison of physical signs
of sound stimulus to a trace of a standard signal stored during 5-10 seconds in sensory or
I

echo-memory. This wave develops during a deviation of physical signs of stimulus from
the characteristics of a trace of repeatedly showed standard stimulus.
The considered mechanism in many respects explains the results, received by us,
where the amplitude of a component PWlmis much higher at deviant stimulation. It is
obvious, that this wave reflects automatic, not connected with function of selective
attention process of comparison of physical attributes of sound stimulus with a trace of a
standard signal in echo-memory. Summation of a selected signal in such conditions is
more adequate. It is specified also by results of brain mapping, where the areas with the
maximal difference of potentials symmetrically occupy practically all forward-central
departments of the brain.
The late positive wave with PL about 250 ms in central and left parietal sites in
conditions of a target task is reduced, but at the same time is characterized by more
expressed by hemispheric symmetry on the data of mapping. The similar tendency with
more expressed symmetry of topomaps is observed and for a wave P350.However, here in
everything, except for of parietal areas, the increase of amplitude in a task with a relevant
signal is observed in comparison with simple rhythmic stimulation.
As a rule, the majority of the researchers specify the increase of amplitude P300in
tasks connected to the decision of psychophysiological loadings [16, 201. At the same
time, there are data and about a possible reduction of long-latency AEPs both in the
relation P ~ wand
, for a wave Nlso in situations with target signals [18]. The reduction of
amplitude of component P300the authors connect to the phenomena deactivation, which
reflects P3W after of phase nonspecific activation. We thing that during the estimation of
late cognitive fluctuations of AEPs it is necessary to take into account all complex of
these waves. The changes of spatial - temporary characteristics of these waves are defined
not only by task, but also by technical characteristics of stimulation.

Acknowledgements

I thank Office of Naval Research International Field Office for support in

performance and representation of this work.

References
1. American ElectroencephalographicSociety. Guidelines for clinical evoked
potentials studies, J. ClidVeurophysiol. 1 (1984) pp. 3-54.
2. Barth D.S., MacDonald K.D., Thalamic modulation of high-frequency oscillating
potentials in auditory cortex, Nature 383 (1996) pp. 78-81.
3. Deiber M.P., Ibanez V., Fischer C . et al., Sequential mapping favours the hypothesis
of distinct generation for Na and Pa middle latency auditory evoked potentials, EEG
Clin.Neurophysiol. 71 (1988) pp. 187-197.
4. Giard M.-H., Perrin F., Pernier J., Bouchet P., Brain generators implicated in the
processing of auditory stimulus deviance: a topographic event-related potential study,
Psychophysiol. 27 (1990) pp. 627-640.
5. Grandory F., Dipole localization methods (DLM) and auditory evoked brainstem
potentials, Rev.Luryngol. 105 (1984) pp. 171-178.
6 . Hackley S.A., Woldorff M., Hillyard S.A., Cross-modal selective attention effects on
retinal, myogenic, brainstem and cerebral evoked potentials, Psychophysiol. 27

297
(1990) pp. 195-208.
7. Jewett D.L., Martin W.H., Sininger G.S. et al. The 3-channel Lissajous' trajectory
of the auditory brain-stem response, EEG Clin.Neurophysio1. 68 (1987) pp. 321-414.
8. Khil'ko V.A., Lytaev S.A., Ostreiko L.M., Clinical Physiological Significance of
Intraoperative Evoked Potentials Monitoring, Human Physiology. 28 (2002) pp. 617624.
9. Lytaev S.A., Shostak V.I., Image Perception and Noise Stability in Auditory
System, Human Physiology 17 (1991) pp. 38-44.
10. Lytaev S., Shevchenko S., VEPs and AEPs Mapping of Occlusive Lesions in
Cerebral Vessels, Ann. New York Acad. Sci., New York Acad. Press 821 (1997)
pp. 524-528.
11. Matthies C., Samii M., Management of vestibular schwannomas (acoustic neuromas):
the value of neurophysiology for evaluation and prediction of auditory function in 420
cases, Neurosurgery 40 (1997) pp. 919-929.
12. Moore J.K. The human auditory brain stem as a generator of auditory evoked
potentials, Hear.Res. 29 (1987) pp. 33.
13. Pantev C., Hoke M., Lehnertz K. et al., Tonotopic organization of human auditory
cortex revealed by transient auditory evoked magnetic fields, EEG Clin.
Neurophysiol. 69 (1988) pp. 160-170.
14. Paquereau J., Ingrand P., Marilland A., Wave I11 of brainstem auditory evoked
potentials analysed both with 3-channel Lissajous' trajectory and dipole localization
method, EEG Clin.Neurophysiol.80 (1991) pp. 298-302.
15. Picton T.W., Hillyard S.A., Krausz H.I. et al., Human auditory evoked potentials.
1. Evaluation of components, EEG Clin. Neurophysiol. 36 (1 974) pp. 179-190.
16. Picton T.W., Human Event-Related Potentials. Handbook of EEG and Clin.
Neurophysiol. Revised Ser. Amsterdam, Elsevier 3 (1988) 556 p.
17. Scherg M., Von Cramon D., Evoked dipole source potentials of the human auditory
cortex, EEG Clin. Neurophysiol. 65 (1986) pp. 344-360.
18. Shostak V., Lytaev S., Golubeva L., Topography of Afferent and Efferent Flows in
Auditory Selective Attention, Neurosci. & Behav. Physiol. 25 (1994) pp. 665-673.
19. Shostak V.I., Lytaev S.A., Bisaga G.N., The Nonspecific Factors of Formation of a
Complex Pa/Nb of Middle Latency Auditory Evoked Potentials, Human Physiology
19 (1993) pp. 7 1-80.
20. Snyders A.Z., Dipole source localization in the study of EP generators: a critique,
EEG Clin.Neurophysiol.80 (1991) pp. 321-325.
21. Sukov W., Barth D.S., Cellular Mechanisms of Thalamically Evoked Gamma
Oscillations in Auditory Cortex, J.Neurophysio1. 85 (2001) pp. 1235-1245.
22. Woldorff M.G., Hackley S.A., Hillyard S.A., The effects of channel-selective
attention on the mismatch negativity wave elicited by deviant tones,
Psychophysiol. 28 (1991) pp. 30-42.

298
COMPUTATION OF ACOUSTIC FIELD ON 2D FRONTS
PART 1. TRAJECTORIES
NICK MALTSEV

ABSTRACT.Traditional methods of wave field computation include modes,

rays and parabolic approximation. This report presents description of new
approach, based on computation of wave fronts and corresponding phases and
amplitudes of the field on them. Final result can be interpreted as sum of raylike arrivals, computed while running fronts intersect receiver location, and
contain arrival times and angles. Introduction of involutes of the wave fronts
clarify the structure of the field. Examples of application of this method are
presented.

Keywords: Rays, fronts, asymptotic expansion, curvature.

1. RAYEQUATIONS

In 2D ray equations are [l]:

dr
= c(r)t,
dr
dB
dr
t = icosB+ksinB,n= -isinB+kcosB

- = -(Vc,n),

(1.2)
(1.3)

r(0) = ro,O(O) = a
where Q: is a launch angle in the source. Parameters a ,T are ray coordinates. Lame
parameters for our ray coordinates are

h , = J G ; = c , h , = J w = J

(1.4)
so

Wave front defined as a curve r

= TO

or

x = ~ ( 7 0a, ) ,z = z(r0,a ) ,0 = O(70, a )

(1.6)
Vector tangent to the front is

(1.7)

ra(rO,a)=ix,+jz,

= n J = J(-isinO+kcosB)

so

x, = -JsinB,z, = JcosB,O = 8,
(1.8)
To compute the curvature of the front rc we need

(1.9)

x,,

= -J, sin 0 - J Q cos 0, z,,

= J, cos B - J O sin 0

299
then
(1.10)
and radius of the curvature of the front
(1.11)
We also calculate

= J-~(x,x,, z,z,,)
=
87
-sinO(J(Vc,n)cosO - csinO0) +cosO(J(Vc,n)sinO+ccosO0) = c 0
-

so

(1.12)

R(O,T)

CJ
JT

2. FRONT
EQUATIONS
Consider wave fronts as an involutes[2] of some curve C, or

r ( 0 , ~=
) C(CY,T)
+R(a,~)t(a,~)

(2.1)

where C = iX, kZ, is some unknown center, R radius of front curvature and t
- unit tangent vector along the ray. Differentiating last equation with respect to a
we obtain
r, = n J = C, +Rat+RnO,
or

c, + R a t = 0

(2.2)

Differentiating equation (2.1) with respect to

(2.3)
or

r, = t c = C,

we have

+ R,t + RnO,

C, (R, - c)t - Rn(Vc, n)

(2.4)
In free space, when c = const
(2.5)

C,+(R,-c)t=0,C,=0,R,

=0

=c,

where the last equation is an illustration of Huygens principle. When we have media
with constant speed of sound between some boundaries, last equations obviously
produce location of "imaginary" or "reflected" sources, C, which generate field
propagating with speed c. So method of reflections appears to be an intrinsic part
of ray equations. In general case

0
C, = R(-lt
0

+ (VC,n ) n )

300
3. EQUATIONS
T O SOLVE

IN

2D

Finally, collecting all equations which determine rays and fronts we have

= c cos
z , = csin8
0, = c, sin 8 - c, cos 8
J, = c 0

x,

@,=JA+OB
C , = R ( ( R A B)t (Vc, n)n)
R, = c - R(RA+ B)
A = - c,, sin' e + 2c, cos 6 sin o - c,, cos2 8
B = c, cos0 + c, sin 6
These equations are solved numerically, using approximation of the media described
in details in [ 2 ] .
(3.1)

4. EXAMPLES
4.1. Pekeris waveguide. First example displays wave fronts in the Pekeris waveguide. Since speed in the bottom is greater than in the waveguide, fronts run for
bigger distance at the same time.
4.2. Transparent cylinder, 1. This example displays wave fronts dropping on
the transparent cylinder. Since speed in the surrounding space is greater than in
the cylinder, distance between fronts inside cylinder is greater than outside. There
is a shadow zone inside, since field is pushed out, by the same reason.

o m

1500 m

1. Wave fronts in a Pekeris waveguide. Speed in the

FIGURE
waveguide 1500 y , density 1
speed in the half space 1700
s , density 1.5
source depth 100 m

3,

3,

30 1

1500 m

FIGURE
2. Wave fronts dropping on transparent cylinder. Speed
in the surrounding space is 1500 ,: density 1 3,speed inside
the cylinder is 2000 ,: density 1.5 &, sorce deprth 100 m.

4.3. Transparent cylinder, 2. This example displays wave fronts dropping on

the transparent cylinder. Since speed in the surrounding space is smaller than in
the cylinder, distance between fronts inside cylinder is smaller than outside. There
is no shadow inside, since field is attracted to the lower speed.

302

1500 m

FIGURE
3. Wave fronts dropping on transparent cylinder. Speed
in the surrounding space is 1500 y , density 1
speed inside
the cylinder is 500 y , density 0.5
sorce deprth 100 m

3,

5,

5. ACKNOWLEDGMENTS
Author is sincerely grateful to Dr. Alexandra Tolstoy for support and discussions and Dr. Finn Jensen, for presenting materials on benchmark probblem of
propagation, which is a subject of the next part of this paper.
REFERENCES
[l] L.M. Brekhovskih, Waves in layered media, Moscow (1965).
[2] N. Maltsev, Journal of computational acoustics, Vol. 9, No. 1 (2001)

169-182

SUPERCONDUCTOR
TECHNOLOGIES,,
969 WESTMAUDEAVENUE,SUNNYVALE,
CA, 94085
E-mail address: nmaltsev(0sv.suptech. corn

303

A NEW TECHNIQUE COMBINING EIGENFUNCTION

EXPANSIONS AND BOUNDARY ELEMENTS TO SOLVE
ACOUSTIC RADIATION PROBLEMS
J. D. MAYNARD
T h e Pennsylvania State University, University Park, PA 16802 USA
E-mail: maynard@phys.psu. edu
A central problem in noise control involves the calculation of sound radiation and
the holographic reconstruction of sources. A standard method for such calculations
employs boundary elements and the surface Helmholtz integral equation (SHIE).
However, difficulties occur a t frequencies for which a congruent pressure-release
boundary would have interior resonances. In such cases the SHIE does not have
a unique solution and the matrix used in the boundary element method is singular. This talk will present a new technique which employs both eigenfunction
expansions (typically using spherical wavefunctions) and boundary element matrices. This new method avoids problems with internal resonances and singular
matrices, does not require adding internal CHIEF points, and permits fast and
accurate holographic reconstructions for arbitrarily shaped sources. Advantageous
operations termed iterative deepening, which automatically discards unnecessary eigenfunctions and detects any symmetries in the problem, and forward
refinement with parametric relaxation, which guarantees convergence to a correct and accurate solution, will be presented. Application of this new method to
inverse propagation will be discussed.

Introduction

For the calculation of the sound radiation from vibrating structures, the boundary
element method (BEM) applied t o the surface Helmholtz integral equation (SHIE)
has become a common standard. Research continue^^>^ with the goal of improving
the speed of calculation for cases involving a large number of elements ( N > 1000);
BEM calculations require the time intensive procedure of inverting a complex N x N
matrix, which in some cases may be singular. Recently attention has been directed
toward revisiting older methods employing eigenfunction expansions, with particular interest in spherical wavefunctions (SWF)4y5. While the first such methods
used the orthogonality of the eigenfunctions on a spherical surface, more recent
methods employ fits to data on arbitrarily shaped surface^^?^. A difficulty with
such fits t o non-spherical surfaces, as will be discussed in this paper, is that the fits
are limited t o relatively low orders of the functions, with the result that outwardly
propagated wave fields, and more particularly inversely propagated wavefields, are
rather inaccurately calculated. In this paper we present a new method, based on
spherical wavefunctions, iterative deepening and forward refinement using the
surface Helmholtz integral equation, which has the following significant advantages:
a) The method requires the inversion of only one N x M matrix, with M
N/10,
resulting in large time savings for N > 1000. b) The method works even at frequencies for which there are internal resonances (when the SHIE does not have a
unique solution), and does not require the use of internal points as in the combined
Helmholtz integral formulation (CHIEF) methodg. For the singular cases it is usually not necessary t o calculate extra functions as in the Burton-Miller method. c)

304

The method can produce accurate results (at the level of a few percent) for inverse
as well as forward propagation calculations; even the inverse propagation requires
only the single N x M matrix inversion. The key to this method lies in overcoming
the limitation on the number of spherical wavefunctions which can be used to fit
the data for non-spherical surfaces. The origin of the limitation may be understood
as follows:
We consider experimental data for a sound field measured on a surface, e.g.
the surface normal velocity measured on a vibrating surface, or the sound pressure
measured on a hologram surface. The data is measured at discrete points, and
some assumption must be made concerning the behavior of the field on the continuous surface between the measured points. The only reasonable assumption is
that the field varies slowly between the measured points; for example, it should be
the case that the actual field could be accurately represented by the measured data
points and quadratic shape functions. One of the reasons for the great success of
the BEM is that it takes advantage of the assumption that fields vary according to
shape functions in evaluating integrals for the SHIE over the measured data. By
contrast, it is in satisfying this assumption, for non-spherical surfaces, that spherical wavefunction fits are of limited value. For non-spherical surfaces, high order
functions would be required simply to conform to the location of the data points
(e.g. fitting relatively sharp corners) in addition to fitting the values at the data
points. The result is that while many high order functions could well fit the data at
the measured points, the field so represented between the data points would vary
rapidly, not satisfying the assumption that the actual field varies slowly. An important point is that this problem cannot be solved by simply having a high density of
data points; in this case the fit with the spherical wavefunctions may require a very
large number of terms, each with very small coefficients, and such coefficients may
be too difficult to determine in practice. Mathematically, representing a function
with a complete set of eigenfunctions requires an infinite sum, and the sum may
converge too slowly for numerical implementation with a truncated series. To avoid
unlikely rapidly varying features resulting from high order terms in a fit with spherical wavefunctions, fits must be limited to low orders. Such low order fits provide
a correspondingly inaccurate representation of the shape and value of the field on
the surface, in error by 20% to 100%.
In our new method for solving forward and inverse radiation problems for nonspherical sources, fitting with spherical wavefunctions is a critical part of the solution, but the fit is not the final solution, as will be explained in this paper. Our
research for non-spherical shapes, such as the muffler and engine shown in Figure 1, and for values of ka (wavenumber times characteristic source size) up to 25,
shows that this method reproduces the results of a standard BEM calculation for
forward propagation to better than three percent.

305

Figure 1. Non-spherical shapes, designated a muffler77and an engine , which

have been studied with the new method for calculating sound fields. Starting with
N N 1000 points for a boundary condition, sound fields can be quickly calculated
by inverting a single N x M matrix, with M N N/10. There are no problems with
singular matrices and no necessity for selecting interior CHIEF points.
2

Spherical Wavefunctions

Separating variables in spherical coordinates yields the spherical wave functions.

For finite size sources, one has the boundary condition that solutions be outgoing
waves at infinity. Thus for the spherical wavefunctions (assuming a time dependence
exp ( i w t ) ) we use

where

The j , and yl are the spherical Bessel functions, and the Pim are the associated
are similar to the spherical harmonics, but with
Legendre polynomials. The
the la dependence removed; the 4 dependence is included explicitly in Eq. (1).
Any solution to a radiation problem may be written as a linear combination of
these functions:

vm

A necessary field is the normal component of the particle velocity at a surface given
by points Fs. We define a surface field which incorporates the acoustic impedance
pc as w (Fs)= pcfi (Fs) . t7(Fs),where fi (Fs) is the unit normal at the surface point
Fs. In terms of the eigenfunctions aim,we have
1

where

@im(Fs)= (l/k)fi (Fs). a@,,( F s ) .

306

Low Order Fits with Eigenfunctions

For the forward propagation problem, we assume that the normal component of the
particle velocity field is known at a finite set of discrete points Fi with i = 1 , 2 , 3 ,...N .
We simplify notation by writing u (Fi) = ui,and similarly for other functions of ri.
We simplify notation by letting the one integer p index the two subscripts 1 and
m : p = [l ( p ) ,m ( p ) ]. We truncate t o a finite number of basis functions, so that
p = 1 , 2 , 3 ,...M . Finally, we use the convention that an index repeated in a term
indicates a sum over that index. Now we can rewrite Eq. (4)as vi = @:,A,. ui
is a complex vector of length N, A, is a complex vector of length M , and ,@: is
an N x M matrix. We shall later find that M
N/10, so that the matrices are
not square. The method of Singular Value Decomposition (SVD) allows one t o find
the inverse of such matrices. This is equivalent t o adjusting the coefficients A,
to least-squares fit the boundary data vi. Once the A, have been determined from
the surface normal velocity ui,then the surface pressure may be determined with
p ( 6 )= pi = @i,A,. In holography, it is usually the surface pressure on a hologram
surface which is measured. In this case the equation for pi is inverted to find A,,
from which ui may be calculated.
What has been discussed above is not new; it is essentially the Rayleigh-Ritz
method, discussed in basic texts on mathematical methods4, implemented with
contemporary computer algorithms. Our technique evolved from the method of
Koopmann12 and was at first referred t o as the equivalent source m e t h ~ d .
Other researchers had used similar methods before our implementation5i6. Wul4?l5
has extensively refined and extended the technique. What is new in our current
method is the procedure of forward refinement with parametric relaxation.
The application of iterative deepening for solving boundary value problems is
also new.
As discussed earlier, the discrete nature of data points on a non-spherical surface
limits the order (and the number of basis functions M ) which may be found by leastsquares fitting the data. This limit causes the vectors vi and pi t o be relatively
inaccurate representations of the actual sound fields. The method for determining
the optimal order for least-square fits to the surface data, as well as the method
of forward refinement for converting wi and pi t o accurate versions, employs the
BEM matrices for the SHIE. These are discussed next.
N

The BEM Matrices for the Surface Helmholtz Integral Equation

The surface Helmholtz integral equation (SHIE) for the sound pressure p (Fi) is
p (fi) =

fl (Ti)

(//

(4.
f,G

(6;
fs))

p (F,) dS

//

(-ZlcG (C;F,)) w (7,)d S )

(5)
wherefl(7) = j ~ i i . f , G o ( r ; r , ) d S ,G(r;F,) = exp(-ilc]r-fs 1)/47r I F - F s I
, Go (r; F,) = 1 / 4 ~I r- F, 1, and i i . V, ia a normal derivative at the point F,.
Taking the derivative of this equation t o obtain a similar integral equation for
the surface normal velocity u (6)
results in singular integrands. These singularities
may be removed by subtracting kernels with similar singularities and by having

307
factors which vanish at the singular points. In the boundary element method, the
surface is divided into elements, with a number of nodes on their peripheries. The
integrals over the elements are accomplished using shape functions and/or constant
elements (with the field over the entire element taking on the value of the field a t
the center of the element). When this is done the SHIE and its derivative become
matrix equations:
p a. - A ,ajpj
. .

+ Bijvj

(6)

The matrices A , B , C, D are integrals of the shape functions and/or various kernels
(G, etc.) over individual elements, and the vectors p and u are the sound fields
evaluated at the element nodes or centers.
5

Forward Refinement, Iterative Deepening, and Parametric

Relaxat ion

To describe forward refinement, as well as the procedures of iterative deepening

and parametric relaxation, the case of forward propagation, where u is known and
p is t o be determined, will be used as an example. A simple version of forward
refinement with iterative deepening will be discussed first, and then this procedure
itself will be refined with parametric relaxation. To clarify the procedure, we
re-write Eq. (6) as
P,~= A z 3 p F

+ E,

(8)

where Ea = B,,v,. The basic procedure is as follows: a) Start with a basis set
of SWF @ l m using 1 = 0 through l,,
where l,
is a low order (e.g. 3). b)
Use least-squares fitting of the known va with the low order basis set to find the
relatively inaccurate fields wSWf and p S W f .c) Set pan = p S w f .d) Using pan and
2
Eq. (8), calculate pout. Also calculate S = C, (pi - p,Out) . e) Set pan = pout, and
iterate by returning to step d). Terminate the iteration if S starts t o increase. f )
Save the best value of S for this l,
as Sl, and increase .1,
However, before
returning to step b) for an outer iteration loop, remove functions from the basis set
which had relatively low values of the coefficients in the fit for uswf. This reduces
the value of M and decreases the time required to invert the N x M matrix in the
least-squares fit. This outer iteration loop, in which l,,
is increased, continues
until Sl starts t o increase.
The term iterative deepening refers to the pruning of the basis set and the
reduction of M for the next value of l,
in step f); the term is taken from a
procedure in computer chess, where good moves found in a shallow search of a
tree of possible moves are used t o optimize a search to a greater depth16. Fkom the
minimum of Sl, the optimum value of l,,
is found, usually resulting in M
N/10.
Step d) is referred to as forward refinement because the calculation is done in
a forward direction with the original matrices, rather than their inverses. This
procedure so far is the same as solving the matrix equation Eq. (6) by iteration2,
N

308
using p S w f as a seed; however, the procedure becomes significantly different with
the addition of parametric relaxation, discussed next.
If the iteration in step d) were undertaken as described, it is likely that the
iteration would not converge to the desired solution. Parametric relaxation is used
t o guarantee the desired convergence. First of all, general convergence could be improved by not using pZn = pout in step e), but instead using ptn = (pout y e w
/2, )
where Pew
was the previous version of pan which produced the current version
of pout Further convergence t o the desired solution may be obtained by using
ptn = [pout (1 - p)pP'""
pp""f] / 2 with the parameter /3 between 0 and 1.
Even though p S w f is relatively inaccurate, it still gives the iteration an excellent
idea of where it should be heading. In any case, convergence t o the desired solution
is further guaranteed by replacing E, = Bt3u3with E: = B,, (1- a ) vj auiWf ,
with the parameter a between 0 and 1. Consider the case when a = ,i3 = 1; this
corresponds t o starting with p S W f and uSwf in Eq. (6). Since p S W f and u S W fare
linear combinations of spherical wavefunctions, they exactly satisfy the SHIE, and
we have Si = S = 0 exactly, clearly the global minimum. If a and/or p are changed
slightly, then the global minimum will shift slightly, but the previous solution will
not be far off, so that the iteration procedure should bring it to the new global
minimum of S. Thus by slowly relaxing a and to zero, the iterative solution will
track the global minimum of S , and converge to the desired solution. This slow
variation of a and p from 1 t o 0 is the parametric relaxation procedure. In the
current implementation of the computer program, ,Ll is set at 1 during the iterative
deepening process (finding the optimum Imax) and a is relaxed from 1 to 0 in steps
of 0.2. Once the optimum l,
is found, p is relaxed from 0.9 to 0 in steps of 0.1,
and a t each value of p, a is relaxed as before. It is important t o note that in order
to find the accurate version of the unknown surface field, it is crucial t o know p S W f
and u S W fwhile
;
they may be relatively inaccurate versions of the actual wave fields,
they are still approximations of the actual wave fields which are exact solutions of
the SHIE, and thus serve to initially locate the global minimum of S.
For holographic reconstruction, where p is measured on a hologram surface, the
SWF and forward refinement are used as above to find an accurate version of the
surface normal velocity v ; the only difference is that the roles of Eq. (6) and Eq.
(7) are reversed.

Inverse P r o p a g a t i o n (Holographic R e c o n s t r u c t i o n )

In holographic sound source reconstruction, the sound pressure p h measured on

a hologram surface T;L, with ph = p ( f h ) , must be used t o determine the normal
velocity on the surface of a source, us = u ( F s ) . Here it will be assumed that the
hologram measurements are taken in the nearfield, very close t o the source; in our
current research we assume measurements are taken on a congruent surface a t a
distance 1 f h - Fs 1- 0.1 I r', 1 .
The first step is to use the normal derivative surface Helmholz integral equation
(NDSHIE, Eq. (7)) and p h t o determine the normal velocity uh a t the hologram
surface. This can be done with a boundary element method; while the Dirichlet

309
problem is more difficult than the Neumann one (SHIE) for the BEM, one at least
has an analytical equation which can be solved with the BEM. For the remainder
of the inverse problem, one does not have an analytical equation t o address.
The next step is t o fit ph and V h with spherical wavefunctions. With the coefficients of this fit one can evaluate SWF fields on the hologram surface and on the
source surface: p L w f , vLwf and p i w f , v,f. These fields are then used t o seed an
iteration proceedure as follows:
a) The hologram fields at individual points, phi and V h i , are transferred back to
corresponding points on the source surface with

pii = (p:yf /pi:>

phi

and

Vii

V sswf
i /ViYf)

(9)

This simple proceedure alone can result in a 10% - 30% reconstruction of the source.
At a few points the denominators in the equations may be zero, and a t these points
one uses the equations
p,i = ( p , f / v ~ ~ fvhi
)

Or

Vii

(V,yf

/ P L Y f ) phi

(10)

In the rare case any of these ratios are also singular, one simply uses the average
of non-singular ratios a t nearby points.
b) The fields pLi and vii may be refined with Eqs. (6) and (7), t o ensure that
they obey the surface Helmholtz integral equation.
c) The next step is t o use the Helmholtz integral equation on the source surface
to propagate pii and vii out t o the hologram surface, t o obtain the fields pLi and
VLi.

d) One now loops back t o step a), but with the SWF fields replaced with the
new fields pAil vLi and pLil v:~, and the iteration continues.
The iteration is monitored with I phi - p i i 12, and is terminated when this
starts t o increase. It may be possible to reduce the growth of unwanted features by
smoothing the iteration fields by adding some fraction of the SWF fields (as with
the parameter p in the iterative solution of the SHIE). However, we have not yet
implemented this feature in our current computer code.

References
1. T. W. Wu, Boundary Element Acoustics: Fundamentals and Computer Codes,

(WIT Press, South Hampton, Boston, 2000).

2. S. Amini, Chen Ke, and P. J. Harris, Iterative solutions of boundary value
equations for the exterior Helmholtz problem, J. Sound Vib. 112, 257-262
(1990).
3. Steffen Marburg, Stefan Scheider, and Has-Juergen Hardtke, Iterative solution techniques for boundary element method (BEM) equations, J. Acoust
SOC. Am. 112 (pt. 2) 2381 (2002); Steffen Marburg and Stefan Schneider,
Performance of iterative solvers for acoustic problems, preprint (200).
4. Philip M. Morse and Herman Feshbach, Methods of Th,eoretical Physics,
(McGraw-Hill, New York, 1953).
5. W. Williams, N. G. Parke, D. A. Moran, and Charles H. Sherman, Acoustic
radiation from a finite cylinder, J. Acoust. SOC.Am. 36,2316-2322 (1964).

310
6. Y-C. Chao, An implicit least-squares method for the inverse problem of acoustic radiation, J. Acoust. SOC.Am. 81, 1288-1292 (1987).
7. J. D. Maynard, Nearfield acoustic holography and arbitrarily shaped sources,
International Meeting on Acoustical Imaging, in J. Imagerie Acoustique (SFA,
GAIE, Lyon, France, 1994) pp 25-28.
8. J. D. Maynard, Nearfield acoustic holography: a review, in Proceedings of
InterNoise 2001 (2001).
9. H. A. Schenk, Improved integral formulation for acoustic radiation problems,
J. Acoust. SOC.Am. 44, 41-58 (1968).
10. A. J. Burton and G. F. Miller, The application of internal equation methods
to the numerical solution of some exterior boundary-value problems, Proc.
Roy. SOC.Lond. A 323,201-210 (1971).
11. W. H. Press, S. A. Teukolsky, W. T. Vettering, and B. P. Flannery, Numerical
Recipes (Cambridge University Press, Cambridge, 1992).
12. G. H. Koopmann, L. Song, and J. B. Fahnline, A method for computing
acoustic fields based on the principle of wave superposition, J. Acoust. SOC.
Amer. 86, 2433 (1989).
13. Y. Huang, Computer Techniques f o r Three-Dimensional Source Radiation,
Ph.D. thesis, The Pennsylvania State University, (1990).
14. Z. Wang and S. F . Wu, Helmholtz equation-least-squares method for reconstructing the acoustic pressure field, J. Acoust. SOC. Am. 91, 2020-2032
(1997).
15. V. Isakov and S. F. Wu, On theory and application of HELS method in inverse
acoustics, Inverse Problem 18, 1147-1159 (2002).
16. Chess Skill in Man and Machine, ed. Peter W. Frey (Springer Verlag, New
York, 1978).

311

INVERSION IN VISCOACOUSTIC MEDIA

A. RIBODETTI
IRD-GEOSCIENCES A ZUR UMR6526, Observatoire OcCanologique d e Vdlefranche,
B P 48, Quai de la Dame 06235, Villefranche-sur-mer Cedex France

A. HANYGA
Department of Geosciences, University of Bergen,
Allkgaten 41, NO5007 Bergen, Norway
An adaptation of a method of viscoacoustic inversion developed by Ribodetti et
al. (2000) is applied to the power law attenuation model. The model is used
in inversion of real data from different acquisition geometries, which removes the
restriction to the circular acquisition of the previous work of Ribodetti et al. (2000).
We present an application of 2.5-D asymptotic viscoacoustic inversion to ultrasonic
waves scattered by a lava cylinder immersed in water. It is demonstrated that
the rheological parameters (velocity, attenuation and a related to the frequency
dependence of the relaxation functions) of rock samples can be determined by this
method.

1 Introduction
During the last decade, the interest in the attenuation of seismic waves has increased. This interest was motivated by the improvements of the characterization
and monitoring of hydrocarbon reservoirs and, more generally, of the Earth models which can be obtained when accounting for the anelasticity of rocks 2 . The
improvements concern the lithologic description of the earth, the analysis of its
physical state and the degree of saturation of the rocks. Different techniques have
been used in field and in laboratory to study the attenuation of acoustic waves
propagating through rocks 3 .
In a viscoacoustic medium the seismic imaging and inversion methods
based
on Beylkins paper fail because of the frequency dependence of the phase of the
wave field. Lack of symmetry with respect to frequency sign reversals imposes serious restrictions on the acquisition geometry. This aspect of the problem was succesfully dealt with in the paper of Ribodetti et al. 7, where the constant (frequency
independent) Q parameter was applied t o represent the attenuative properties of
the medium. The constant Q models violate causality. We show that this drawback can be overcome by using a three-parameter class of power law attenuation
models *,, which satisfy the causality and dissipativity requirements lo. The replacement of the constant Q model by the power law model allows a separation
of variables in the inversion method of Ribodetti et a1 and a one-step inversion
procedure.
In this paper we combine the method of viscoacoustic inversion developed by
Ribodetti et al. with the power law model parameterization. Power law models
belong t o the class of singular memory models discussed in Ref 12. Solutions of the
associated initial value problems are infinitely smooth at the wavefronts. Since the
solution vanishes ahead of the wavefront, an initial singular delta-spiked pulse sent
by the source immediately spreads out and its maximum lags behind the wavefront.
415

312
Since the energy travels with a delay behind the wavefront, the effective speed of
the seismic signal is lower than the infinite-frequency limit of the propagation speed
associated with the wavefront. Consequently the usual imaging methods, based on
the assumption that the signal travels with a wavefront, do not apply. As shown
in Ref. 1 2 , for a given source, receiver and a singly scattered signal the location of
the scatterer lies in a thin layer above the isochron determined from travel-time
considerations.
It is worth noting that the singular memory is not an artefact of our choice of
model parametrization. Singular memory is a common property of a wide range
of materials, ranging from polymers l4 to complex porous materials 15. Power law
attenuation can be considered as a general asymptotic property of wave field phase
functions in attenuative media.
In an attenuative medium the phase of a signal consists of an imaginary part,
which is an odd function of frequency, and a negative real part, which is an even
function of frequency. The Beylkin single-step inversion method 6 , commonly used
in elastic imaging, implicitly assumes that the phase is an odd function of frequency.
An extension of Beylkins imaging method, adapted for dealing with viscoacoustic
media was recently developed by Ribodetti and applied to a non-causal constantQ model. It takes into account lack of symmetry between positive and negative
frequencies in attenuative media. In causal hereditary models the appearance of an
additional imaginary dispersive phase due associated with the attenuation factor
causes some additional complications in the transformation from the frequency variable to the composite wavenumber vector length, as required by Beylkins method.
We present an application of 2.5-D asymptotic viscoacoustic diffraction tomography to ultrasonic data recorded during a physical-scaled laboratory experiment.
This scaled experiment was used to test the reliability of our method when applied
to a real dataset to estimate the attenuation. Preliminary results shows the possibility to extract the rheological parameters of a lava sample and suggested that
the method could be used as a tool for rock properties estimation in laboratory.
2

Theoretical background

The method we use is cast in the general frame of least-squares inverse theory. It
however exploits an asymptotic approximation of the forward operator in order to
obtain a one-step explicit inversion formula 6,4. The attenuation is modeled by
a power law an the medium is assumed to vary smoothly in space. The forward
modeling is based on the Born approximation. An asymptotic ray tracing method
is used to calculate travel time, amplitude and attenuation between source, receiver
and scattering points. The inverse problem is solved by an optimization approach
based on the iterative least-squares minimization of the mismatch between the
observed and the computed scattered wavefields.
We begin the description of the inversion method with the description of the
power law models of viscoacoustic wave propagation. A scalar model of a singlemode wave propagation with a power law attenuation can be represented by the
frequency-domain expression
C ( u , x) oc f ( w ) exp (iuT(x) - (-iw)aA(x))

313
with A 2 0, 0 < Q < 1. f(w)denotes the Fourier spectrum of the source signal
and is analytic in the upper half of the complex plane. It is easy to see that the
inverse Fourier transform u(t,x) of G(w, x) vanishes for t > T provided the function
f ( w ) is analytic in the upper complex half plane l 3 and satisfies the partial integrodifferential equation
utt

+ K ( t ,x)) * utt = c2 v2u + d ( t ) d(x);

u(0,x) = U t ( 0 , x) = 0

(2)

where

K ( ~ , x=) [2(t/T)-ff/r(i - a ) + (t/7)1-2ar(2 - 2 4 1 / T

(3)

More generally, we shall consider viscoacoustic models defined by eq. (2) with

K ( t , x )= 2 4 x )

( t / T )ff--l

(4)

r(Q)

with a(x), b(x) close t o unity and Q S ( Y O .

Eq. (1) implies that complex-valued wave propagation velocity for the model is
given by the formula
v ( w ) = c/q5(-iw)
while the quality factor

Q-l = 4~C O S ( T Q / 2 ) ( W T ) f f - 1
where an arbitrary constant T > 0 of dimension [TIhas been introduced for purely
dimensional.
For inversion we choose a reference medium for which eq. (2) has an explicit
solution 1 3 :

c2 = c; := K O / ~ O KO
,
> 0 , po > 0 and approximate the target medium by a
perturbation of the reference medium in the class defined by eqs (2),(4). The
solution of eqs (2-3) is calculated in the Born approximation with respect to the
reference medium.
Let Go(t,x) denote the Green function for the background medium, i.e. the
solution of eq. (2). The perturbed medium is characterized by the memory kernel
K ( t ,x) = Ko(t) bK(t,x), and the parameter K(X) = KO + ~ K ( x ) . The corresponding Green function satisfies the equation

Gtt

+ K * Gtt = ( l / p ) V . [KVG]+ b ( t )b(x);

u(O,X) = ~ t ( 0X), = 0

(6)

or, equivalently

Gtt KO* Gtt - (l/po)V. [KVG]= d ( t ) b(x)-bK * Gtt 6p-I V . [K VG] (l/p) V . [ d VG]
~
u(0,x) = U t ( 0 , X ) = 0

(7)

3 14

The Green function G(t, x) of eq. (7) can be expressed in terms of Go@,x):

G(y, X,ti - t o ) = G o ( ~X,

7 ti - t o ) +

Go(Y,Y,t ) * [ - dK(t7 Y) * Gtt(~7X,t)+

+ p i 1 V . [SK(Y)VG(Y,X,t ) ]]

It=tl-tO

d ~ ( 8 )

where the asterisk denotes the time convolution,

S K ( t , x ) = 2[a(x) - I]*

+ [b(x)- 11=+
2 [ln(t/T) - (t/T)OO-l liicao)] ( a - ao)/r(ao) (9)

and $(z) = dl nr ( z ) / d z.
In the Born approximation

+ p i i V . [ ~ K ( Y )VGO(Y,X,t ) ]] dy

(10)

In the frequency-domain the Born approximation assumes the form

G ( w , y , x ) = G ( w , y , x ) + J~(w,Y,x)[w2~(w7z7Y)d~(w7Y)+

+dp-l(y)V. [KO(Y)VG(Y,x,a ) ]+po(y)-l v . [WY)VGO(w,Y,41 ] It=tl--to dY

(11)
The phase of the Green function (we will describe below) can be decomposed
in an imaginary part and a real part

+ i@r

(12)

= wT(x) sgn(w)lwTla sin(.rra/2) A(x)

(13)

@ = iwT(x) - (-iwT) A(x) = @R

with
@I

T denotes the travel time, i.e. the arrival time of the wavefront. While constructing
the inversion operator the slowly varying factor exp(@R) will be included in the
amplitude, while the imaginary phase will be used in the construction of the inverse
operator.
Let x(<),z(<) denote a set of sources and receivers parameterized by [ varying
over a region U c R2 of the plane. The scatterer coordinates are denoted by y
(Figure 1).
The total imaginary phase of a singly scattered wave is equals
@IT(Y) = w

[ T ( x ,Y) + T ( z ,Y)] + sgn(w)(wT)a sin(ra/2) M x ,Y) + A ( z ,Y)I

(15)

315
The Green function of the reference medium can be expressed in the form
G:o(y,x; w ) = D(x,y) exp[i@~],
where D(x, y ) = B(x, y) C(x,y ), B(x,y) denotes
the geometric divergence factor (= 1/47r]x - y J in a homogeneous background
medium) and C(x,y) = exp(@R(x,y)).
After an asymptotic evaluation of the Laplacian the Born approximation can
be expressed in a compact form
~ G ( zX,, W ) := G ~ ( zX,, W ) - Go(z,X, W ) =
with

g(z,X, W ; y) M h ( y ) dy

~ ( x , z , w ; Y )= D(Y,x)D(z,Y)
~XP(~~@IT(Y))

(16)

(17)

and

6a
0 if the value of a is kept fixed; in this case the third components of the
vectors M, bm drop out. We denote by G the integral operator (Gf)(x,z) =
9(x,z; Y) M f(Y) dY.
The final inversion formula has the following form

.r

6m = (Gt W G)- GtW 6G

(20)

where 6G(<) = GobS(z(<),x(<))- Gc,(z(<),x(<)),where Gobs(z(<),x(<)) are the

observed data and Go(z(<),x(<))are ray+Born synthetics. In this expression, after
applying a change of coordinates as in Ref with the Jacobian denoted by I , the
operator Y := Gt W G with the weight W = D(y , x(<))-lD(y, x(<))- I ( w ,<) Z
becomes
2%

1 lrn
d<

d w n , w , Y> 9(r,w , Y)z (R + s>

where Mt, M = R S , R is a matrix involving the highest powers of w , while S

contains lower powers of w . If R is invertible and Z = R-l, then

where U denotes the unit matrix. Consequently,

2%

+ z s]= 2 ( 2 4 3 u qY- y) -I-x

dkek(Y-Y) [u

(21)

where X is a pseudodifferential operator of a negative order. X maps elements of

the model space t o smoother objects while the first term of eq. (21) yields a sharp
image of the model.

316
We shall choose R in such a way that it contains all the powers of w which tend
to w4 when a -+ 1. The Hermitean matrix R is obtained by expanding Mt M in
powers of w:
Rll = 4W4-2ffoT-2ffo
(22)

R22 = pP2[w4( V T T ) ~ 4

T~~~ (VTT
~

. +VAT)^+
~

~ 70
a +( V T~T ) ~
VTT
o
. VAT]
2 sin(.rrcuo/2)~
= -41~14-2ao ?-2ffo
R13 =
[ln(wr) - i.rr/2]

R23 =

= 2 ( - i ~ ) ~ - ~rPaO
O ( V T T ) [ln(rw)
~
- i7r/2]/p
R33 = 4 1 ~ 1 [ln2(.rw)
~ - ~ ~ ~.rr2/4] r P 2 = o

(24)
~
(25)
(26)
(27)

(28)
The matrix R is invertible, in contrast to the constant Q case. In the latter case R is
in general a rank one matrix 7, except for the circular acquisition, where the positive
and negative frequencies can be treated separately. with TT = T(x,y) + T(y,z),
AT = A(x,y) A(Y,z), V = V,.

Application to the laboratory data: scattering by a lava specimen.

A cylindrical Neapolitan Yellow Tuff sample (pyroclastic rock) from a drilling experiment performed near the Chiaiano (Campi Flegrei, Italy) of height 0.5 m and
diameter 0.06 m was immersed in the center of the water tank equipped with
computer-based control and data acquisition systems (National Instruments system Is).
The source and receiver were placed on a horizontal plane orthogonal to the
sample at a depth 0.67 m below the water level and rotated synchronously around
the axis of the cylinder at a distance of 0.469 m (Fig 2).
The azimuthal offset between the source and the receiver was 15", the initial
azimuth of the source was 7.5' and the azimuth step was 5'. The central frequency
of the source is around 100 kHz. The diffracted wavefield is presented in Fig. 3.
The main arrivals between 0.5 ms - 0.6 ms are respectively the reflections from the
external and internal edge of the rock sample. The ray-Born synthetics obtained
after the first iteration using the perturbation in Fig. 4 , 5 and 6, are plotted alogside
the observed traces. A good fit is obtained. The tomographic models for 6n, 6a
and 6a are shown in part (a) of Figs 4, 5 and 6.
In the post-processing stage traces extracted from the tomographic image for
different azimuths (centered at the source of the specimen) are matched by a synthetic trace obtained by convolving the source wavelet with a boxcar. The source
wavelet is estimated by stacking several traces centered on the direct arrival. The
height of the boxcar represents the estimated amplitude of the perturbation Sn, 6a
or 60. The width of the boxcar represents the estimated width of the specimen.
Before convolution the boxcar is transformed t o the time domain by scaling the independent variable with the background propagation velocity. The convolved signal
is then transformed back t o the space by an inverse scaling transformation. After
matching the traces for all the selected azimuths, the best fitting boxcar and its

3 17
convolution with the source signal are used to build 2-D syntheric images, denoted
as the impulsional and convolutional models.
The best fitting convolved model shown in Part (b) of Figs 4-6 is obtained by
convolution with the source wavelet shown in Fig. 4(c), while the model shown
in Part (d) of Figs 4-6 is obtained by convolution with the best fitting impulsional model. The influences of an imperfect acquisition geometry are indicated
by arrows. The best fitting convolved model shown in Part (b) is obtained by
convolution with the source wavelet shown in Fig. 4(c), while the model shown in
Part (d) is obtained by convolution with the best fitting impulsional model. The
influences of an imperfect acquisition geometry are indicated by arrows. Because of
the ray-Born approximation and of the constant background model assumed during
inversion, only the physical properties along the contour of the rock sample can be
estimated. The location of the sample and its dimensions are recovered successfully
(see Part (c) of Figs 4, 5 and 6). The shape of the recovered section is acceptable on all the three images. Nevertheless, we note that the recovered shape of
the cylinder is slightly oval. This may result from a combination of inaccuracies
in the experimental setup. On the Sa and Sa synthetic images the post-processing
failed locally t o fit the exact shape due to the sensitivity of the parameters to the
amplitude variations related to some local high heterogeneity of the sample.
The absolute quantitative estimation of the propagation parameter SK and attenuation parameters Sa and Sa is obtained during the post-processing of the tomographic models. The 12 misfit between the tomographic model and the best
fitting convolved model is also determined. The misfit for Sn presented in Fig. 4(e)
shows a minimum at 3.908106N/m2, corresponding to a velocity perturbation
Sv = 1318 m/s, then the velocity of the tuff sample is vtuff = vg Sw x 3800 m/s;
for ha the misfit presents a minimum at 1.68910-4, corresponding to a SQ around
-209880 estimated at the central frequency of 100 kHz. The attenuation Q of the
sample is Qtuff = QO + SQ x 120. The estimates of the velocity and the Q factor
are in agreement with results obtained by Vanorio 17. For the parameter Sa (Fig. 6
(d)) the minimum of the misfit function is obtained at Sa = 0.15 corresponding
to "tuff = a0 Sa = 0.5 0.15 = 0.65. The estimated value of a is attribuable
to wave attenuation in a porous water saturated sample. A similar value was obtained by Hanyga and Rok l1 from the Gurevich-Lopatnikov model of fast wave
attenuation in a thin-layered porous medium, but there is not enough evidence for
validating this model for the lava specimen in our experiment.

Conclusions

We have obtained an extension of the viscoacoustic inversion method developed by

Ribodetti et al. t o the power law model class. The power law model class is more
realistic than the constant-Q models because dispersion and absorption effects are
taken into account. In addition the power law models respect the causality principle.
In comparison with the constant Q parameterization the power law attenuation
avoids the invertibility problems and does not require circular aquisition.
Choosing water as a reference medium with a = 0.5, the new inversion scheme
runs successfully on ultrasonic laboratory data. The choice of a power law medium

3 18

in the inversion procedure yields a good fit between the data and the synthetics
in only one iteration as compared with six iterations in the constant-Q case and
twelve iterations in the non-attenuative case.
Our results of inversion of ultrasonic laboratory data for a lava sample show
that the adopted method can be useful for estimating rock properties in laboratory
experiments, in particular the parameter (Y governs the frequency dependence of
the relaxation functions and can be an interestng indicator of the degree of porosity
of the material.
Acknowledgments

The research was made possible by a grant of PA1 and the Norwegian Royal Society
in the framework of the program Aurore of Franco-Norwegian scientific exchanges
2002-2003. Authors thank H.P. Valero, G. Saracco, F. Conil and D. Gibert for the
ultrasonic data collected in the water tank at the University of Rennes, France.
References

1. R. Sheriff, Geophys. Prosp. 23, 125 (1975).

2. M. Randall, Phys. Earth Planet. Int. 12, 1 (1976).
3. N. Toksoz and D. Johnson, 1981, Seismic wave attenuation (SOC. Expl. Geophys., Tulsa, 1981).
4. S. Jin, R. Madariaga, J. Virieux and G. Lambark, Geophys. J. Int. 108, 575
(1992).
5. P. Thierry, Migration/inversion 3-D en profondeur B amplitude prkservke: application aux donnkes de sismique rkflexion avant sommation, Ph.D. thesis
(Universih de Paris VII, Paris, 1997).
6. G. Beylkin, J. Math. Phys. 26, 99 (1985).
7. A. Ribodetti, S. Operto, J. Virieux, G. Lambark, H. P. Valkro and D. Gilbert,
Geophys. J. Int. 140, 324 (2000).
8. T. L. Szabo, J. Acoust. SOC.Am. 96, 491 (1994).
9. T. L. Szabo, J. Acoust. SOC.Am. 97, 14 (1995).
10. A. Hanyga and M. Seredyriska, Power law attenuation in acoustic and isotropic
anelastic media, to appear in Geophys. J. Int..
11. A. Hanyga and V. E. Rok, J. Acoust. SOC.Am. 107, 2965 (2000).
12. A. Hanyga and M. Seredybska, Geophys. J . Int. 137, 319 (1999).
13. A. Hanyga & M. Seredyriska, Quart. Appl. Math. LX, 213 (2002).
14. R. Metzler and T. F. Nonnenmacher, Int. J. Plasticity 19, 941 (2003).
15. A. Hanyga, in Theoretical and Computational Acoustics '97, Yu-Chiung Teng
ed. (World-Scientific, Singapore 1999).
16. H. P. Valero, Endoscopie sismique, Ph.D. thesis (Institut de Physique du Globe
de Paris, Paris, 1997).
17. T. Vanorio, M. Prasad, D. Patella and A. Nur, Geophys. J. Int. 149, 22
(2002).

319

source

receiver

Figure 1. Source, receiver and scattering points varying over a region U C

W2 of the plane

3 20

Figure 2. Acquisition geometry

32 1

322

Distance (m)
0.05 0.10 0.15 0.20 0.25 0.30

Distance (m)

recovered 8~

best fitting convolved 6~

c)

I
0

0.5

Distance (m)
0.05 0.10 0.15 0.20 0.25 0.30

'1.0

1.5

Time

2.0

-5
x104

m i s f 1t

best fitting impulsional6~model

Figure 4. Inversion for the lava sample. (a) Tomographic model for 6 ~ (b)
;
the best fitting
convolved model results of the post-processing sequence applied to (a). The source wavelet is
plotted in (c); (d) the best fitting impulsional model showing the geometry and location of contour
of the section of the rock sample; (e) misfit function representing the perturbations of 6~ and the
minimum.

323

b)

a)

Distance (m)

Distance (m)

best fitting convolved 6a

recovered 6a

d)
Distance Im)

best fitting impulsional &amodel

Figure 5. Inversion for the lava sample. (a) Tomographic model for Sa; (b) the best fitting
convolved model from the post-processing sequence applied to (a); (c) the best fitting impulsional
model; (d) the misfit function representing the perturbations of ba and the minimum.

324

b)

a)
Distance Irn)

recovered 6a

Distance (m)
0.05 0.10 0.15 0.20 0.25 0.30

best fitting convolved 6a

d)
-0.25

Distance (rn)
0.05 0.10 0.15 0.20 0.25 0.30

m i s f i t

best fitting impulsional6a model

Figure 6. Inversion for the lava sample. (a) Tomographic model for ba; (b) the best fitting convolved model results of the post-processing sequence applied to (a); (c) the best fitting impulsional
model; (d) the misfit function representing the perturbations of 601 and the minimum.

325
PREDICTION SYSTEMS WITH DATA ASSIMILATION
FOR COUPLED OCEAN SCIENCE AND OCEAN ACOUSTICS

ALLAN R. ROBINSON
PIERRE F.J. LERMUSIAUX
Division of Applied Sciences, Harvard University, 29 Oxford Street
Cambridge, MA 02138, USA
Ocean science and ocean acoustics today are engaged in coupled interdisciplinary research on both
fundamental dynamics and applications. In this context interdisciplinary data assimilation, which
melds observations and fundamental dynamical models for field and parameter estimation is
emerging as a novel and powerful methodology, but computational demands present challenging
constraints which need to be overcome. These ideas are developed within the concept of an
interdisciplinary system for assessing sonar system performance. An end-to-end system, which
couples meteorology-physical oceanography-geoacoustics-oceanacoustics-bottom-noise-target-sonar
data and models, is used to estimate uncertainties and their transfers and feedbacks. The approach to
interdisciplinary data assimilation for this system importantly involves a full, interdisciplinary state
vector and error covariance matrix. An idealized end-to-end system example is presented based
upon the Shelfbreak PRIMER experiment in the Middle Atlantic Bight. Uncertainties in the physics
are transferred to the acoustics and to a passive sonar using fully coupled physical and acoustical
data assimilation.

Introduction

Interdisciplinary ocean science today now involves research on coupled physical,

biological, chemical, sedimentological, acoustical and optical processes. Research
advances in interdisciplinary ocean science have led to the emergence of new dynamical
concepts in which non-linear interdisciplinary processes are now known to occur on
multiple interactive scales in space and time with bi-directional feedbacks. Such
processes importantly can be dominated by strong sporadic events that are intermittent in
space and time. Understanding specific non-linear dynamics of known events and
identifying important additional as yet unknown multi-scale interactive processes provides
a framework for realistic representation and prediction of the interdisciplinary coastal
ocean [ 121.
Ocean prediction for science and operational applications has now been initiated on
global, basin and regional scales [13]. A system approach that synthesizes theory, data
and numerical computations is essential for efficient and rapid progress in
interdisciplinary ocean science and prediction [ 171. The concept of Ocean Observing and
Prediction Systems for field and parameter estimations has recently been crystallized with
three major components: i) an observational network: a suite of platforms and sensors for
specific tasks; ii) a suite of interdisciplinary dynamical models; and, iii) data management,
analysis and, importantly, data assimilation schemes. Systems are modular, based on
distributed information providing shareable, scalable, flexible and efficient workflow.
Data assimilation is a modern methodology of relating natural data and dynamical
models [16]. The general dynamics of a model is combined or melded with a set of
observations. All dynamical models are to some extent approximate, and all data sets are

326
finite and to some extent limited by error bounds. The purpose of data assimilation is to
provide estimates of nature that are better estimates than can be obtained by using only the
observational data or the dynamical model. There are a number of specific approaches to
data assimilation that are suitable for estimation of the state of nature, including natural
parameters, and for evaluation of the dynamical approximations. Generally data
assimilation demands large computational resources and efficient computational
algorithms are essential. Interdisciplinary data assimilation can contribute powerfully to
understanding and modeling physical-acoustical-biologicalprocesses and is essential for
ocean field prediction and parameter estimation [ 151. Model-model, data-data and datamodel compatibilities are essential and dedicated interdisciplinary research is needed.

Figure 1. Density, fluorescence and acoustic backscatter across the Alneira-Oran Front in the Mediterranean
Sea (from [4]).

To illustrate these ideas, Figure 1 presents a set of simultaneous physical, optical and
acoustical measurements obtained in the region of the Almeira-Oran Front in the
Mediterranean Sea. The purpose of these measurements was to study coupled biologicalphysical frontal dynamical processes. Hydrographic measurements directly provide the
sea-water density distribution, fluorescence is a proxy for phytoplankton density and
acoustic backscatter is a proxy for zooplankton density. The density front separates
Atlantic water from Mediterranean waters. Drawn down phytoplankton are concentrated at
the front and further details are found in [4, 51. The methodological approach that can
best exploit the dynamical information in such measurements is data assimilation. A
single multi-variate state vector for the physical, biological, optical and acoustical state
variables should be defined. The error covariance matrix that determines the relative
weights of data and dynamics in the melded estimates should include the covariances
among all the physical, biological, optical and acoustical variables. Importantly, in
addition to backscatter data, acoustical propagation data can enhance interdisciplinary
dynamical studies and ocean prediction efforts [2].

3 27

The theme of this presentation is that coupled ocean scientific and ocean acoustical
data assimilation presents novel and powerful opportunities and challenges for theoretical
and especially computational acoustics. In the next section we will introduce the
particular interdisciplinary system (physical/geologicaVcoustical/signal-processing/sonarsystem) to be studied. Section 3 further develops data assimilation methods, Section 4
applies the system and the method to the Shelfbreak PRIMER data set and Section 5
concludes.

End-to-End System Concept

Advanced sonar performance prediction requires end-to-end scientific systems: ocean

physics, bottom geophysics, geoacoustics, underwater acoustics, sonar systems and signal
processing. The littoral environment can be highly variable on multiple scales in space
and time, and sonar performance is affected by these inherent variabilities. Uncertainties
arise in estimates of oceanic and acoustic fields from imperfect measurements (data
errors), imperfect models (model errors), and environmental variabilities not explicitly
known. A conceptual basis has been developed to achieve the following: i) generic
methods to efficiently characterize, parameterize, and prioritize system variabilities and
uncertainties arising from regional scales and processes; ii) error, variability and
uncertainty models for the end-to-end system and it's components to address forward and
backward transfer of uncertainties; and, iii) transfers of uncertainties from the acoustic
environment to the sonar and its signal processing in order to effectively characterize and
understand sonar performance and predictions. In order to accomplish these objectives, an
end-to-end system approach is necessary [ 131.

Figure 2. Schematic diagram of the end-to-end system from a model point of view (from [13]).

Figure 2 schematizes the end-to-end system from the model point of view, where
models are used to represent each of the coupled dynamics (boxes) and also the linkages
to observation systems (circles). An effort was made to make the diagram exact but as
simple as possible. The diagram illustrates the forward transfer of information, including
uncertainties, in terms of observed, processed and model data (dots on arrows) and
products and applications (diamond). The system concept encompasses the interactions
and transfers of information with feedback from: i) observing systems, the information
being meteorological-physical oceanographic-acoustical-bottom-noise-target-sonar
data,

328
and, ii) coupled dynamical models, the information being physical-acoustical-bottomnoise-sonar state variables and parameters.
Specific applications require the consideration of a variety of specific end-to-end
systems. Note the backward pointing dotted arrows coming from the sonar model which
indicate that the specifics of the sonar system determines e.g. the acoustic propagation
calculations required. To specify an application requires the choice of: i) observational
circles and model boxes to be included; ii) data types and sampling schemes; and, iii)
appropriate form of the dynamical models to be used. Some of the wide range of
approximate dynamics useful for various oceanic purposes are shown in Table 1.
Table 1. Coupled (Dynamical) Models and Outputs.

Acoustic
Propagation

Types
Non-hydrostatic models (PDE, x,y,z,t)
Primitive-Eqn. models (PDE, x,y,z,t)
Quasi-geostrophic, shallow-water
models
Objective maps, balance eqn. (thermalwind)
0 Feature models
Parabolic-Eqn. models (x,y.z,t/f)
(Coupled)-Normal-Modeparaboliceqn. (x,z,f)
0 Wave number eqn. models (x,z,f:
OASIS)
Ray-tracing models (CASS)

IRcverberalion

0
0

and signal
processing

I Target
3

Full-field TL (pressure
p , phase j >
Modal decomposition of
p field
Processed series: arrival
strut., travel times, etc.
CW I Broadband TL
Scattering strengths

Surface, volume and bottom scattering

Noise

outputs
T, S,velocity fields and
parameters, C field
Dynamical balances

Hamilton model, Sediment flux models

(G&G),etc
Statisticallstochasticmodels fit-to-data
Wenz diagram
Empirical modelslrule of thumbs

Sonar equations fit)

Detection, localization, classification
and tracking models and their
inversions
MeasuredEmpirical

Wave-speed, density and

attenuation coefficients

f-dependent ambient
noise (f;x,y,z,t): due to
sea-surface, shipping,
biologics
SNR, SIR, SE, FOM
Beamforming, spectral
analyses outputs
(timelfrequency
domains)
SL, TS for active

InterdisciplinaryData Assimilation

Data assimilation provides a powerful methodology for state and parameter estimation via
the melding of data and dynamics. It makes feasible such estimates on a substantial and

329
sustainable basis. The general process is schematized in Figure 3. Sensor data are linked
to state variables and parameters and transformed as appropriate for the dynamical model
via measurement models. Dynamics interpolates and extrapolates the data. Dynamical
linkages among all the state variables and parameters allows all of them to be estimated
from observations of some of them (i.e., those more accessible to existing techniques and
prevailing conditions). Error estimation and error models play a crucial role. Using data
assimilation schemes, data and dynamics are melded, often with weights inversely related
to their relative errors. The melding is based on an assimilation criterion involving a cost
or penalty function. The final estimates should agree with the observations and
measurements within data error bounds and should satisfy the dynamical model within
model error bounds. There are many important feedbacks in the generally highly nonlinear
ocean observing and prediction system schematized in Figure 3, which illustrates the
system concept and three feedbacks. Prediction provides the opportunity of efficient
sampling schemes adapted to real-time structures, events, and errors. Data collected for
assimilation also used for ongoing validation can identify model deficiencies and lead to
model improvement, including the adaptation of the approximate models in real-time [ 151.
OBSERVATIONAL NETWORK
SAMPLING SCHEME
SENSOR DATA
AND DATA ERRORS

STATE AND PARAMETER

ESTIMATES AND ERRORS

SV: STATE VARIABLE

P PARAMETER
OBSERVED
M MEASURED
U UNOBSERVEDOR UNMEASURED
E. ERROR
Mm DYNAMICAL LINKAGES

Figure 3. Illustration of a generic Ocean Observing and Prediction System with data assimilation.

Mathematically, the generic data assimilation problem is represented in terms of:

Dvnamical models
d$; + v . V $ ; d r - V . ( K i V $ ; ) d t = B , ( $,,... $;,... $,)dr+dq,

(i= l - - n )
(1 = 1,..., p )

( j = 1,..., n)

330

In the dynamical models, $; is a generic state variable (e.g. i = u, v, T, ... , Zoo, ... p),
t is time, 2 the velocity vector and K,a diffusivity. The first term on the left is local time
change at a point, the second advection and the third diffusion. On the right, the term Bi
represents sources and sinks of $i; the stochastic forcings dqi the model uncertainties.
Model parameters (diffusivities, bottom attenuations, etc.), I l l = { KI, RI ...}, are also
represented by an equation with stochastic forcings d&, where Ciare functionals that
describe the possible deterministic evolution of parameters with time and space. The state
variables $; are related to the data yj, e.g. yj = ( XBTj, SSHj, Fluo, etc.}, via measurement
models, also with stochastic forcings E) The assimilation or melding criterion involves the
minimization of a functional J of the stochastic or error forcings dq;, d<; and E ~ and
,
of
their a priori statistical properties or weights denoted by q,,, qc and qE, subject to the
constraints the other equations. The three sets of equations and assimilation criterion
define the assimilation problem (Fig. 3). For state estimation, we refer to the estimates just
before (forecast) and just after (nowcast) data assimilation as a priori and a posteriori
respectively. For parameter estimation, a priori and a posteriori refer to parameter values
at the beginning and at the conclusion of the optimization. Data residuals or data-model
misfits refer to the differences between the data and model-estimated values of the data.
For the coupled physical-acoustical assimilation problem of interest here, the coupled
discrete vector X , associated with the continuous vector $, is the combined X = [X, XoJ,
where the physics is represented by Xo = [T S, U,V, WJ and the acoustics is represented
by X A= [Pressure@),Phase (q)].The coupled error covariance for the state defined by

= ~ { (-kX t ) ( i - X t ) T )

corresponds to P = PAA P A 0
[POA P O j '

The error covariance P is of paramount importance for the coupled assimilation, here the
minimum error variance update defined by,

where X. is the a priori estimate and X + the a posteriori estimate. The matrix H
corresponds to the coupled measurement models and the matrix R is the coupled data
error covariance.

33 1
Specific classes of data assimilation schemes are listed in Table 2. Specific schemes are
overviewed in [16].
Table 2. Classes of Data Assimilation Schemes

Class
(Filtering and
Smoothing)

Calculus of
Variations
(Smoothing)
Optimization Theory
(Direct locaUgloba1
smoothing)

Examples
1. Direct Insertion, Blending,
Nudging
2. Optimal interpolation
3. Kalman filter/smoother
4. Bayesian estimation (FokkerPlank equations)
5. EnsembleMonte-Carlo
methods
6 . Error-subspacerneduced-order
methods: Square-root filters,
e.g. SEEK
7. Error Subspace Statistical
Estimation (ESSE): 5 and 6
1. Adjoint methods (with
descent)
2. Generalized inverse (e.g.
representer method with
descent)
1. Descent methods (Conjugate
gradient, Quasi-Newton, etc)
2. Simulated annealing, Genetic
algorithms
Combinations of the above

[mportant Properties
1 . Linear
2. Linear, Least Squares
(LS)
3. Linear,LS
4. Non-linear, Non-LS
5. Non-linear, LS/NonLS
6. (Non)-Linear,LS

7. Non-linear, LS/NonLS
1. Linear, LS

2. Linear.LS

1. Linear, LS/Non-LS

2. Non-linear, LS/NonLS

The ocean observing and prediction system utilized at Harvard is the Harvard Ocean
Prediction System (HOPS), which is schematized in Figure 4a. The data assimilation
scheme associated with HOPS, Error Subspace Statistical Estimation (ESSE) is pictured
in Figure 4b. HOPS is an integrated system of data analysis and assimilation schemes,
and a suite of coupled interdisciplinary (physical, acoustical, optical, biogeochemicalecosystem) dynamical models [ 141. This system was developed for producing
interdisciplinary oceanic field estimates that include effective and efficient data
assimilation, dynamically consistent model initialization, multi-scale nesting, and modeldriven adaptive sampling with feedbacks. HOPS employs a primitive equation (PE)
physical dynamical circulation model. Boundary layers (top and bottom) and isopycnal
and diapycnal turbulence are modeled through process parameterization and scaledependent filters. Multiple sigma vertical coordinates are calibrated for accurate
modeling of steep topography. Multiple two-way nests are an existing option for the
horizontal grids.

332

RVARD OCEAN PREDICTION SYSTEM HOPS

r-

Figure 4. Schematic of (a) the Harvard Ocean Prediction System (HOPS) and (b) the Error Subspace
Statistical Estimation (ESSE) procedure.

ESSE is a four-dimensional multivariate estimation scheme for physicalbiogeochemical-acoustical fields and parameters that aims to capture and forecast the
dominant uncertainties, i.e. the error subspace, and assimilate all relevant data to control
and reduce errors [6,7,10]. Instead of characterizing and capturing all uncertainties, ESSE
focuses on the uncertainties that matter. The sub-optimal truncation of the error space is
itself optimal. For example, if a variance criterion is used to combine data and dynamics,
the error subspace is then defined by the ordered eigen-decomposition of the error
covariance. The error subspace is initialized by decomposition on multiple scales and
evolved in time by an ensemble of model iterations. Initial conditions are first perturbed
using random combinations of the initial error principal components. For each perturbed
initial conditions, the nonlinear dynamical model (HOPS), forced stochastically to
represent various model errors, is integrated until the next data time. These Monte-Carlo
integrations are carried out in parallel until the size of the ensemble is large enough to
describe most of the error variance forecast. This is assessed by a convergence criterion.
Once the error forecast has converged, the available data are assimilated, presently based
on a singular value decomposition of the minimum error variance update. After data
assimilation, the posterior data residuals are computed and utilized to correct the posterior
error estimates, i.e. adaptive learning of the dominant errors is carried out. Using error
estimates, adaptive sampling plans are determined using schemes that are consistent with
the data assimilation scheme itself. Ultimately, the ESSE smoothing can be used to correct
the past estimates, including initial conditions and boundary forcings, based on the future
data-model misfits and their error estimates. ESSE has been developed for and applied to
both fundamental research and real-time operations. It has been used in several regions of
the world's ocean for varied purposes, including error forecasting, adaptive sampling,
dynamical studies, model improvements, 3D objective analyses and predictability studies.
Current ESSE developments focus on a fully distributed and scalable architecture,
interdisciplinary parameter estimation, adaptive sampling and adaptive modeling.
In the HOPSBSSE methodology, the real-time initialization of the dominant error
covariance decomposition is evaluated under the assumption that dominant uncertainties
are missing or that there is uncertain variability in the initial state, e.g., smaller mesoscale

333
variability. Important issues include the fact that some state variables are not observed
and that the uncertain variability is multi-scale. The evaluation approach is multi-variate,
multi-scale and three-dimensional, the "observed portions" are directly specified and
eigen-decomposed from differences between the intial state and data, andor from a
statistical model fit to these differences, and for the "non-observed portions" the
"observed" portions are kept fixed and the "non-observed portions" are computed from an
ensemble of numerical (stochastic) dynamical simulations.

Shelfbreak PRIMER Example

In this section we present an example to illustrate three major points: i) end-to-end

uncertainty transfer through an idealized system; ii) coupled physical and acoustical data
assimilation; and, iii) the careful simulation of the physical environment necessary for
realistic acoustic propagation studies.
The Middle Atlantic Bight (MAB) shelfbreak marks a dramatic change, not only in
water depth, but also in the dynamics of the waters that lie on either side. The shelf is
about 100 km wide, extending from Cape Hatteras to Canada. The shelfbreak, which
refers to the first rapid change in depth that occurs between the coastal and deep ocean, is
near the lOOm isobath. The main oceanographic feature in the MAB is a mesoscale front
of temperature, salinity and hence sound speed, separating the shelf and slope water
masses (Figure 5a). The shelf-water to the north is cold and fresh while the slope-water to
the south is warm and salty. Located near the shelfbreak, this front is usually tilted in the
opposite direction of the bottom slope (Figure 5b).

Figure 5. (a) Map of the Middle Atlantic Bight shelfbreak front, (b) Vertical sections of temperature, salinity,
zonal velocity and vertical velocity across the front.

During July and August of 1996, data were collected in the MAB south of New
England, as part of the ONR Shelfbreak PRIMER Experiment [ 111. The main objective
was to study the influence of oceanographic variability on the propagation of sound from
the slope to the shelf. Intensive in situ measurements were carried out in a 45 km by 30
km domain between the 85 m and 500 m isobaths. The measurements consisted of
temperature, salinity, velocity, chlorophyll, bioluminescence and acoustic transmissions.
The physics considered here are the mesoscale dynamics of the Middle Atlantic Bight
shelfbreak front, including remote influences from the shelf, slope and deep ocean. The
acoustics is the transmission of low-frequency sound from the continental slope, through

334
the shelfbreak front, onto the shelf. These dynamics, and also the model parameters, data
assimilated in the physical model, and acoustical-physical uncertainties are described in
[S]. The coupled assimilation via ESSE is illustrated for the 3D physical fields and 2D
transmission loss along an actual Shelfbreak-PRIMERacoustic path (Figure 6).
Simulation Experiment Acoustic Paths
40.45

-71 5

-71

Longitude

-705

Figure 6. Acoustic paths considered overlaid on bathymetry.

The present example of the end-to-end system using PRIMER data involves the
physical, acoustical and passive system components of Figure 2. The method adopted to
characterize sonar performance involved the concept of Predictive Probability of
Detection (PPD). Figure 7 illustrates conceptually the PPD, details of which are given in
[l]. The system-based environmental PDF was derived by a comparison of model
predictions with system data. A histogram of the differences between the data and the
acoustic model was fit with an appropriate distribution to yield the PDF. This PDF
represents the uncertainty in the computational modeling process, typically small, and the
inherent variability of the environment not contained in the model inputs, which typically
is larger. The PPD is a prediction of the system performance versus range. Rather than use
a single range value (e.g. range-of-the-day or range-of-the-moment), the PPD
provides the system operator with a probabilistic representation of the system
performance. The operator can thus use this information to operate the system more
effectively, and can make more informed decisions on search, risk, and expenditure of
assets.

30

Figure 7. Example Predictive Probability of Detection (PPD)

335
In this example, the problem is to start with physical environmental data, transfer it
through the idealized end-to-end system and compute a depth and range dependent PPD
from first principals. A novel approach is used in which coupled physical-acoustical data
assimilation method is used in Transmission Loss (TL) estimation. The methodology
involves a coupled physical-acoustical identical-twin experiment [9] and is ESSE based.

Methodology: Uncertainty prediction, transfer, processing and reduction

ESSE is first utilized to predict uncertainties in the ocean physics, transfer them to the
acoustics and compute acoustics uncertainties [ 101. The approach is based on dominant
error principal components and on ensembles of Monte-Carlo simulations. Specifically,
the dominant physical uncertainties are initialized using a multi-scale, principal
component decomposition of the initial error covariance whose parameters are a function
of oceanic data and dynamical models. An ensemble of initial physical conditions for the
ocean physics primitive-equation model of HOPS is then computed, in accord with this
initial statistics or error subspace. For each initial realization, the ocean physics model is
integrated forward in time, which produces an ensemble of physical field forecasts,
providing environmental uncertainties for the volume around the acoustic region of
interest. The different realizations of the sound speed field forecast, i.e. the sound speed
ensemble, are then interpolated onto the vertical planes of acoustic interests and fed into
the NPS coupled-mode sound propagation model. For a specific vertical plane, running
the acoustics model for each sound-speed realization produces an ensemble of predicted
acoustic wavefields. The resulting coupled physical-acoustical error subspace captures the
dominant uncertainties in each physical and acoustical field, but also captures their
(cross)-covariances and other coupled statistics.
Presently, the main acoustic variable of interest is the transmission loss (TL) because
the ultimate objective is to provide uncertainty estimates for sonar systems. To obtain TL
uncertainties for a broadband sonar, the above continuous-wave uncertainties for a singlefrequency forecast TL are processed. A variable-width running-range average is applied to
the ESSE ensemble of single-frequency TL realizations. This produces an estimate of the
uncertainties in the broadband TL term of a passive sonar equation.
Once physical and/or acoustical data are available, they are combined with the model
estimates via ESSE data assimilation so as to reduce errors and improve predictions.
Presently, the physical and acoustical data are assimilated such that the total error variance
in the error subspace is minimized. After assimilation, the a posteriori acoustical variables
and corresponding (reduced) uncertainties are available. By variable-width running-range
average, the a posteriori broadband TL and its uncertainties are computed. Ultimately, the
coupled assimilation reduces uncertainties in the sonar predictions. This coupled
uncertainty estimation and data assimilation approach is now illustrated.
Twin-experiments
The approach is exemplified in identical-twin experiments using PRIMER data [ 113 In
such an experiment, the data are synthetic, i.e. they are output from a simulation that
defines the truth, and the model used to create data is identical to the one used for
prediction. In the present case, the simulation defining the truth assimilates some real data
to be closer to reality. Goals of such twin-experiments are to study the assimilation in an
ideal situation and to find out if the a posteriori fields become close to the known true
fields. Importantly, all computations are made at resolutions and on domains appropriate
for each field. The acoustic 2D-space resolution is here much finer than the 3D-space

336
resolution of the ocean physics. In addition to the possible different scales within each
discipline, there are multiple scales because of the interdisciplinary nature of the problem.

Figure 8. (a) Bottom slope used in the model, (b) optimized model levels, (c) surface temperature for PRIMER
domain.

To achieve numerical predictions of the ocean physics at accuracies suited for useful
acoustical computations, substantial efforts were necessary. This included compiling all
data sets collected over the large-scale New England shelfbreak region relevant to the
(sub)-mesoscale physics. Bathymetric data sets were collected and combined in accord
with their uncertainties and biases. Numerical research (i.e. minimize numerical errors due
to steep topographies/pressure gradient, non-convergence issues, pycnocline resolution,
etc) was carried-out to allow the ocean physics model (HOPS) to run on accurate
topographies, with almost no smoothing of the bathymetry. This was necessary for the
acoustics and is illustrated on Figure 8 by the bottom slope (Fig. 8a) and optimised model
levels along a cross-slope section (Fig 8b). The physical representations of boundary
influences (e.g. Hurricane Bertha) and their numerics (e.g. OBC) were also improved. The
latest numerical ocean simulation is illustrated in Fig. lc. Of interest are: the impacts of a
large meander of shelf water intruding slope waters at latitudes further south than usual,
the corresponding upstream presence of a large slope eddy (confirmed by SST) that is
pinching off from the shelfbreak front, and the strong effect of Hurricane Bertha in setting
up the basis of the overall regional internal circulation.
In the present coupled assimilation via ESSE example, the synthetic representation of
the true ocean is a 4D (x, y, z, t) ocean physics model simulation that assimilates real
physical data. After 5 days in this simulation, a snapshot of the true ocean is taken and
the corresponding true sound-speed field input to the 2D (r, z) acoustical couplednormal-mode model. Running the acoustical model then leads to the true transmission
loss (TL) field on day 5. Different synthetic physical and acoustical data were then
coarsely sampled from this true physical-acoustical ocean. These different data sets
were then successfully assimilated using ESSE. As a technical aside, the order of the

331
assimilation (ocean physics before acoustics, or vice-versa) does not matter and sequential
processing of observations is utilized.

-E-

1w

50

1520

50

5
4

-100

100

1510

1500

e: 200

1490

250

1480

m0

%15

El50

8 zoo
250

300
0

10

20

30

2
1
10

20

30

Mean of C1 and statistics of error estimate for C1

Figure 9. Ensemble mean, standard deviation, skewness and kurtosis of sound speed section

The uncertainties in the ocean physics along one of the PRIMER acoustical path (Fig.
6) are illustrated on Figure 9. An ESSE ensemble of 79 members was used to compute this
prior error estimate and the ensemble mean and statistical moments of the deviations from
this mean are shown. Of course, the 79 simulations were independent from the physicalacoustical realization that defines the truth. The mean sound-speed section show the
shelbreak front, surface summer thermocline and cold pool (sound channel) on the shelf.
The standard deviation is maximum around 35 m depth, in accord with independent data
(Glen Gawarkiewicz, personal com.), and has an amplitude of 5 d s e c . This depth is where
the front surfaces within the middle of the thermocline, near the largest hydrographic
gradients and largest internal velocities (shelfbreak front jet). At the front, the skewness is
close to 0 and kurtosis close to 3 (relatively Gaussian), except near the surface and bottom
boundary layers. Just away from the front, uncertainties are less Gaussian.
Coupled assimilation results for the sound-speed and continuous-wave TL fields are
illustrated on Figures 10-11. In this example, the physical data are coarse temperature and
salinity measurements sampled in the synthetic true ocean: 2 CTD profiles are taken
over 35 km across the shelfbreak front, along the PRIMER path (Fig. 6). The 224 Hz
source is at 300 m depth on that path (bottom left corner). The acoustical data are
synthetic towed-receiver TL data, measuring the relative sound intensity received at 224
Hz along that path. The TL observations are made at constant 70 m depth, every 50 m
from a range of 150 m to almost the receiver (about 35 km from the source). These are
sub-sampled data since the (r, z) acoustic grid resolution is 5 m by 5 m.
The sound-speed (C) residuals before assimilation (prior field residuals), after
assimilation of the TL data, after assimilation of both the TL and C data (posterior field
residuals), and the error standard deviation estimate for this posterior sound-speed field
are on Fig. 10. The true TL, prior TL (i.e. the mean or forecast), TL after assimilation of
TL data and TL after assimilation of both TL and C data (posterior TL) are shown in Fig.
1 1. Even though the sub-sampled data are limited, the posterior C and TL are substantially
closer to the true C and TL than the priors. The improvements of both the TL and soundspeed fields due to the assimilation of the TL data alone, and their subsequent

338
improvements due to the assimilation of the two sound-speed profiles are clearly visible.
On Fig. 10, it is interesting to see that the coupled assimilation improves the larger scales
of the TL field, the higher loss in the surface thermocline and low loss in the sound
channel over the shelf, but also some of the smaller scales (nulls, etc). Similarly, the
impact of the sound-speed data (two CTDs) shows the effects of the mainly mesoscale
ocean-ocean correlations and multi-scale ocean-acoustic correlations.

Figure 10. Sound-speed residuals before assimilation, after assimilation of the TL data, after assimilation of
both the TL and C data, and error standard deviation estimate for this posterior sound-speed field

Figure 1 1 . True TL, prior TL,TL after assimilation of TL data and TL after assimilation of both TL and C data

The posterior ESSE ensemble properties (error variance, covariances, etc)

importantly estimate the uncertainty reduction as a result of the coupled data assimilation.
This is illustrated on Fig. 1Id. The impact of the 2 CTD profiles (on each side of the
positive residuals pancake) is clearly visible. The posterior residuals (Fig. 1lc) and
posterior error estimates (Fig. 1 Id) agree on average, but the posterior error estimates are
likely too small near the bottom of the summer thermocline (e.g. see residual pancake at
about 2.5 db and posterior error at 0.6db). This is likely due to the small ensemble size.
The error covariances utilized in the coupled assimilation are illustrated on Figure 12
within the acoustic vertical section. The ESSE estimate of the covariance/correlation
function between TL at a range I = 3km and depth z = 30m with the other TL values (top)
and with the C field (bottom) are plotted. These fields correspond to a row of the coupled
TL-C covariance/correlation matrix. The covariances (Fig 12a) are dimensional; the

339
correlations are non-dimensional (Fig 12b). Looking at the covariances, for that (r, Z)
location, a positive change in the TL term at (I;z) is linked to a reduction of the TL term
in the acoustic waveguide (cold pool) over the shelf and to a warming (higher sound
speed) of the upper thermocline above the shelfbreak front and slight cooling in the
bottom boundary layer (e.g. a warming of the slope water and/or steepening of the front).
Should a TL measurement be made at this (I;Z) location, the coupled ESSE assimilation
would influence the TL and C fields over whole section in accordance with these patterns;
e.g. increase TL in red regions (more loss), reduce TL in blue regions (less loss). Similar
comments can be made for the correlation fields (Fig, 12b). In general, correlation
estimates tend to over-estimate the influences at large-scales and covariances are utilized
for data assimilation.

Figure 12. ESSE estimate of the (a) covariance, (b) correlation, function between TL at a range r = 3km and
depth z = 30m with the other TL values (top) and with the C field (bottom).

en
70

I0
20
Rewe (rm)

30

10
Range L

20
W

30

Figure 13. Four realizations of the variable-width running-range average TL field

Broadband TL and impact of DA on the TL term of a passive sonar PPD

To simulate the transfer uncertainties to a broadband sonar system (TL term in a sonar
equation), the ensemble of single-frequency (224Hz) TL realizations are processed, using
a variable-width running-range average (Figure 13). This is because a variable-width
running-range average in space is approximately equivalent to a frequency average around
a central frequency as a real sonar does. Four realizations (out of the 79 computed) of the
broadband TL estimate are shown on Fig. 13. As expected, the length scales increase with

340
range. The different realizations (ensemble members) have of course different modal
couplings, reflecting the differences in the sound-speed realizations.
To illustrate the impact of the coupled assimilation of physical and acoustical data on
the performance of a simulated sonar predictions, the prior and posterior histograms of
deviations from the mean broadband TL (i.e., the error PDF estimates) as a function of
range and depth are shown on Figure 14. The estimated prior PDFs are found to be depth
and range dependent (Fig. 14a). Near the depth (55 m) of the main wave-guide, the
predicted error standard deviation is relatively constant with range and relatively large,
around 3 to 4 (db). Above (30 m) and below (85 m), standard deviations tend to decrease
with range (down to 2 db), leading to a higher PDF peak. After assimilation (Fig. 14b), the
uncertainties are reduced to f ldb and are more Gaussian at all depths.

*%._

I_-

Figure 14. PDF estimates of broadband TL as a function of range and depth, along a PRIMER acoustic path:
(a) Prior PDF (predicted by ESSE); (b) Posterior PDF (after ESSE assimilation).

Conclusion

Transferring and forecasting uncertainties from the physics, through the acoustics, using
ESSE, and processing these dominant error estimates to obtain TL uncertainties for a
broadband sonar equation has been demonstrated for an idealized problem using
Shelfbreak PRIMER data. ESSE nonlinear coupled assimilation recovered the fine-scale
TL structures (10-100m) and the mesoscale ocean physics (1Okm) from coarse TL data
and/or coarse C data. Broadband TL uncertainties were predicted to be range and depth
dependent and coupled data assimilation sharpened and homogenized the broadband
PDFs.
We are entering a new era of fully interdisciplinary ocean science and ocean
acoustics. Ocean prediction systems for science, operations and management are now
being utilized. These systems involve interdisciplinary estimation of state variables and
error fields via multivariate physical-biological-acoustical data assimilation. This
provides novel and challenging opportunities for theoretical and computational acoustics.

34 I

Acknowledgements

The authors are grateful to Prof. Ching-Sang Chiu (Naval Postgraduate School) for his
acoustic propagation contributions to the example. They thank Dr. Phillip Abbot (Oasis,
Inc.), Dr. Patrick J. Haley, Jr. and Wayne G. Leslie (Harvard University) for scientific
input. We are indebted to Wayne G. Leslie for preparation of the manuscript. This work
was supported by the Office of Naval Research under grants NOOO14-97-1-0239, N0001402- 1-0989 and NO00 14-00-1-0771.
References
1. Abbot P. and Dyer I., Sonar Performance Predictions Based on Environmental
Variability, In Acoustic Variability, 2002. N.G. Pace and F.B. Jensen (Eds.),
SACLANTCEN. Kluwer Academic Press (2002) pp. 61 1-618.
2. Elisseeff P., Schmidt H., and Xu W., Ocean Acoustic Tomography as a Data
Assimilation Problem, IEEE J. Ocean. Engin. 27 (2002) pp. 275-282.
3. Evangelinos C., Chang R., Lermusiaux P.F.J. and Patrikalakis N., Rapid Real-Time
Interdisciplinary Ocean Forecasting using Adaptive Sampling and Adaptive Modeling
and Legacy Codes: Component Encapsulation using XML, In Computational
Science - ICCS 2003, International Conference on Computational Science 2003,
P.M.A. Sloot, D. Abramson, A. Bogdanov, J.J. Dongarra, A. Zomaya, Y. Gorbachev
(Eds.). Springer, Lecture Notes in Computer Science, 2660 (2003) pp. 375-384.
4. Fielding S., Crisp N., Allen J.T., Hartman M.C., Rabe B., and Roe H.S.J., Mesoscale
subduction at the Almeria-Oran front Part 2: Biophysical interactions, J. Mar. Sys. 30
(2001) pp. 287-304.
5. Griffiths G., Fielding S., and Roe H.S.J., Biological-Physical-Acoustical Interactions,
In THE SEA: Biological-Physical Interactions in the Sea, Volume 12 (A.R.
Robinson, J.J. McCarthy and B.J. Rothschild, eds.), John Wiley and Sons, NY,
(2002) pp. 441-474.
6. Lermusiaux P.F.J., Data assimilation via Error Subspace Statistical Estimation. Part
11: Middle Atlantic Bight shelfbreak front simulations and ESSE validation. Month.
Weather Rev., 127(7), (1999) pp. 1408-1432.
7. Lermusiaux P.F.J. and Robinson A.R., Data assimilation via Error Subspace
Statistical Estimation. Part I: Theory and schemes. Month. Weather Rev., 127(7),
(1999) pp. 1385-1407.
8. Lermusiaux P.F.J. and Chiu C.-S., Four-dimensional data assimilation for coupled
physical-acoustical fields. In Acoustic Variability, 2002. N.G. Pace and F.B. Jensen
(Eds.), SACLANTCEN. Kluwer Academic Press (2002) pp. 4 17-424.
9. Lermusiaux P.F.J., Chiu C.-S. and Robinson A.R.., Modeling Uncertainties in the
Prediction of the Acoustic Wavefield in a Shelfbreak Environment. Proceedings of
the 5th ICTCA, May 21-25, 2001, in Theoretical and Computational Acoustics
2001, E.-C. Shang, Q. Li and T.F. Gao (Eds.), World Scientific Publishing Co.,
(2002) pp. 191-200.
10. Lermusiaux P.F.J., Robinson A.R., Haley P.J. Jr. and Leslie W.G., Advanced
Interdisciplinary Data Assimilation: Filtering and Smoothing via Error Subspace
Statistical Estimation. Proceedings of Oceans 2002 IEEE/MITS Conference, (2002)
pp. 795-802.

342
11. Lynch J.F., Gawarkiewicz G.G., Chiu C.-S., Pickart R., Miller J.H., Smith K.B.,
Robinson A., Brink K., Beardsley R., Sperry B. and Potty G., Shelfbreak PRIMER An integrated acoustic and oceanographic field study in the mid-Atlantic Bight, In
Shallow-Water Acoustics (R. Zhang and J. Zhou, Editors), China Ocean Press, (1997)
pp. 205-212.
12. McCarthy J.J., Robinson, A.R., and Rothschild B.J., Biological-Physical Interactions
in the Sea: Emergent Findings and New Directions, In THE SEA: Biological-Physical
Interactions in the Sea, Volume 12 (A.R. Robinson, J.J. McCarthy and B.J.
Rothschild, eds.), John Wiley and Sons, NY, (2002) pp. 1-17.
13. Pinardi N. and Woods J. (Eds.), Ocean Forecasting: Conceptual Basis and
Applications, Springer Verlag, Berlin, (2002) 472 pp.
14. Robinson A.R., Forecasting and Simulating Coastal Ocean Processes and
Variabilities with the Harvard Ocean Prediction System, Coastal Ocean Prediction
(C.N.K. Mooers, ed.), AGU Coastal and Estuarine Studies Series, American
Geophysical Union, (1999) pp. 77- 100.
15. Robinson A.R. and Lermusiaux P.F.J., Data assimilation for modeling and predicting
coupled physical-biological interactions in the sea, In THE SEA: Biological-Physical
Interactions in the Sea, Volume 12 (A.R. Robinson, J.J. McCarthy and B.J.
Rothschild, eds.), John Wiley and Sons, NY, (2002) pp. 475-536.
16. Robinson A.R., Lermusiaux P.F.J., and Sloan N.Q., Data Assimilation. In THE SEA:
The Global Coastal Ocean, Volume 10: Processes and Methods (K.H. Brink and A.R.
Robinson, eds.), John Wiley and Sons, NY, (1998) pp. 541-594.
17. Robinson A.R., McCarthy J.J., and Rothschild, B.J., Interdisciplinary Ocean Science
is Evolving and a System Approach is Essential, J. Mar. Sys. 22 (1999) pp. 231-239.
18. Robinson A.R., Abbot P., Lermusiaux P.F.J. and Dillman L., Transfer of uncertainties
through physical-acoustical-sonar end-to-end systems: A conceptual basis. Acoustic
Variability 2002. N.G. Pace and F.B. Jensen (Editors), Saclantcen. Kluwer
Academic Press, (2002) pp. 603-610.

343
HORIZONTAL REFRACTION AND THE UNCOUPLED AZIMUTH
APPROXIMATION
KEVIN B . SMITH
Depamnent of Physics, Naval Postgraduate School, Monterey, CA 93943
E-mail: kbsrnith@nps.navy.mil

Division of

FREDERICK D. TAPPERT
Applied Marine Physics, RSMAS, University ofMiami, Miami, FL 33149

Over much of the last decade, the majority of numerical simulations designed to predict the influence
of three-dimensional (3-D) acoustic propagation effects have relied upon 3-D ray-based algorithms.
Advances in modem computational platforms have now made it possible to calculate the 3-D
acoustic field using full-wave models. However, such complete treatments of the propagation
problem can still be highly inefficient. To obtain accurate predictions of the coupling between
azimuths, a large number of radial bearings must be computed. In contrast, a low-resolution 3-D
acoustic field can be generated by computing the two-dimensional, uncoupled azimuth solution along
N bearings, a procedure known as the Nx2-D method. It is important, therefore, to understand the
significance of 3-D propagation effects due to azimuthal coupling. One effect which has been used
to distinguish 3-D propagation is the horizontal refraction characterized by deflections of the
horizontal arrival angle at a distant receiver. As will be shown, such effects are not unique to 3-D
models but can arise in uncoupled, Nx2-D predictions within 3-D variable environments.
Comparisons of horizontal arrival angles calculated using a 3-D parabolic equation model and its
corresponding Nx2-D uncoupled azimuth version in a shallow water, ridge-like environment have
been made. The results show that most of the refracted structure can be predicted by the uncoupled
azimuth approximation. A proper treatment of the ray equations in the uncoupled azimuth
approximation is also shown to contain an azimuthal deflection angle which would produce an
apparent horizontal refraction in a 2-D calculation. The analysis from such a derivation is discussed
in the context of the full-wave results.

Introduction

Observations of three-dimensional (3-D) acoustic propagation effects for both long-range

basin scale propagation [ 11 and shorter range littoral zone propagation [2] have generated
some concern over the significance of such effects in general environments. Numerous
efforts have been made to include the influence of such 3-D propagation effects into
standard underwater acoustic propagation models [3,4,5,6,7]. However, such models are
generally inefficient and require substantial computing power and memory. Adequate
knowledge of the 3-D environment is also difficult to obtain, and the significance of any
predicted 3-D effects may be overshadowed by the uncertainty in the environmental
description [8]. The question then is to what degree do 3-D propagation effects alter the
received field, and how may this be observed as a distinct feature in the data? Of interest
here will be the requirements of acoustic propagation models necessary to predict these
effects.
One obvious approach would be to search for the effects of horizontal refraction,
specifically the deflection of horizontal arrival angles. This would appear to be an
exclusively 3-D propagation effect since a two-dimensional (2-D) model would be
expected to predict only energy propagating directly outward along radials. For example,
a typical 2-D ray code by definition will only permit rays to travel along a single radial

344
without allowing for any azimuthal deflections. The effect of the 3-D propagation is then
presumably the amount of azimuthal deflection away from the initial radial.
Using ray-based acoustics, it has been shown for some ideal environments [9] that
this horizontal refraction of the acoustic energy can strip individual modes from the field
by deflecting them downslope. The results from this analytical work have been
confirmed in tank-scale experiments [lo]. Similar results have also been used to
benchmark some full-wave, 3-D numerical models [11,12]. All of the full-wave analysis,
however, was for continuous-wave (CW), tonal transmissions.
In order to analyze the full-wave propagation in terms of horizontal arrival angles, it
is useful to calculate the transmission of broadband pulses. Standard beamforming
techniques may then be applied to determine horizontal angles and azimuthal deflections.
In order to compare the results of 3-D models with 2-D predictions, one must also
calculate the 2-D propagation along identical radials using an Nx2-D model. Such a
model merely computes the 2-D, uncoupled azimuth (UNCA) approximation to the 3-D
field along the various radials of the 3-D environment.
The purpose of this paper is to examine the differences between the results of a fully
3-D calculation and the Nx2-D UNCA approximation to the wave equation in an
environment which should produce significant azimuthal coupling. Specifically, we will
examine the horizontal deflection of the acoustic propagation as observed by standard
beamforming techniques in the horizontal plane. A theoretical treatment of the
corresponding ray equations will also be presented. As will be shown, much of the effect
of horizontal refraction can be predicted within the framework of the UNCA
approximation.

Three-dimensional Parabolic Equation Model

The acoustic propagation model used in this work is a full-wave model based on the 3-D
parabolic equation (PE) approximation [ 131 to the Helmholtz equation. The complex
acoustic pressure P is assumed to be time-harmonic at frequencyf such that
p ( r ,z , p,t ) = p ( r , z , p)e-*/ ,

(1)

where

and the outgoing solution to the Helmholtz equation approximately satisfies an equation
of parabolic form

Pulse propagation can then be easily modeled by Fourier synthesizing multiple CW

solutions.
The operator Q,, is the well-known PE square-root operator

Q,, = ( l + p + + + ~, ) ~
where

(4)

345

C
2 x f , n(r, z,p) =
, c,, , and c ( r ,z , p) are the reference
A
The variables k, =CO
c(r, z , p)
acoustic wavenumber, the acoustic index of refraction, the reference sound speed, and the
sound speed in the environment, respectively. The influence of density contrasts can also
be incorporated into an effective index of refraction term, but will not be described here.
Note that in the absence of range-dependence (which implies, therefore, no azimuthaldependence), Eq. (3) is exact.
Because the Q,, operator is a pseudo-differential operator and cannot be explicitly
applied directly, some form of algebraic approximation is required. There exists a vast
literature of work on approximations to the corresponding 2-D operator [ 13,14,15,16]
which invoke the uncoupled azimuth (UNCA) approximation,

Three-dimensional environments may still be treated in an Nx2-D fashion where the

UNCA approximation is utilized to predict the propagation along individual bearings,
each of which is defined by a unique index of refrzction, n(r, z , p) .
The model used here employs a two-dimensional (depth and azimuth) split-step
Fourier [17] algorithm to march the solution of the field outward in range. To improve
the efficiency of the calculation, a limited azimuthal aperture approach [12] is utilized.
The specific PE approximation employed is the wide-angle PE [18] in depth, and the
standard PE in azimuth. A wide-angle approximation in azimuth could also be employed,
but requires substantially more computational memory and is not expected to affect the
results significantly.
In order to benchmark the 3-D code and contrast the solution to its 2-D counterpart, a
standard 3-D wedge environment was employed [ 1I]. As depicted in Fig. 1, the wedge
varies from a depth of 380 m to a depth of 20 m with a slope of 2.86'. The water column
is isospeed with sound speed 1500 d s overlying a sediment bottom of sound speed 1700
m/s and density 1.5 g/cm3. This environment extends infinitely in a direction
perpendicular to the plane of the paper. The source is placed halfway up the slope at a
depth of 200 m.
SalIce

peition

c, 8= 1 5 O O 0d s

_ _ _ _ _ _

cb = 1700mls

pb = 1.5 g/cm

Figure 1. Schematic diagram of 3-D wedge environment cross section.

For this benchmarking exercise, a CW source of frequency 25 Hz was defined which

excited only the first mode of the waveguide at the 200 m depth. A transmission loss
(TL) calculation was made at a depth of 36 m along the 200 m isobath in the cross-slope
direction. A 2-D calculation of TL along this isobath should only exhibit cylindrical

346
spreading. However, 3-D propagation is expected to generate refracted paths downslope.
Along the isobath, we expect to encounter the interference structure of such paths initially
launched upslope. Figure 2 compares the results of such TL, calculations.

65

'

Range (krn)

Figure 2. Mode 1 transmission loss calculated at 36 m depth along 200 m isobath in the cross-slope direction.
Dashed curve is 2-D calculation, solid curve is 3-D calculation using 2048 radials over full 360" azimuthal
aperture.

As expected, the 2-D calculation produces a smooth curve corresponding to lologr,

or cylindrical spreading. The 3-D calculation shows the interference structure between
various paths initially launched upslope which are refracted downslope across the 200 m
isobath at various ranges along the radial. These results show good agreement with
accepted benchmark results [ 111. For this environment and source characteristics, it was
determined that an azimuthal sampling of approximately 0.2" is required to accurately
calculate the influence of azimuthal coupling out to 25 km. Coarser sampling results in
curves which appear more like the 2-D results and do not properly predict the range of
mode cut-off.
Two additional points should be noted about these results. First, although the 3-D
horizontal refraction is obvious, a range-average of the 3-D TL curve would look very
similar to the 2-D result. This suggests that in a real experiment, such distinctions may be
difficult. However, a second important point is the distinction between 2-D and 3-D at
ranges beyond what are displayed here. For this mode, all of the initially upslope energy
has been either refracted downslope across the 200 m isobath or stripped away at ranges
beyond about 30 km. Thus, for ranges beyond this, the 3-D propagation shows
effectively no acoustic energy along the radial for this mode. This is in sharp contrast to
the 2-D prediction which continues the lOlogr drop-off for all ranges.

347

Geometrical Limit of Uncoupled Azimuth Approximation

For the purposes of providing a theoretical argument for the numerical observations made
later, we shall now consider the corresponding ray equations of the parabolic equation
when azimuthal coupling is included and when it is neglected. Although all subsequent
calculations will be based on the full-wave PE model previously described, this ray
analysis will help us interpret the final results.
To simplify the ray treatment presented here, we shall use the first order, standard
parabolic approximation,

where

Defining the PE field function v ( r , z , v) according to

the resulting 3-D parabolic equation is then

To obtain the corresponding ray equations, we substitute the following geometrical

acoustics expression into Eq. (10):

v(r,z , p) = a ( r , z , p)exp [ik,s(r, z , PI].

(11)

Here, a is the ray amplitude and s is the phase function. Because we are only interested
in the ray trajectories, we will not consider the influence on the amplitude. For large k, ,
the leading term in the asymptotic expansion produces the parabolic eikonal equation,

From classical mechanics, we see that we can define the momenta that are
canonically conjugate to the ray position variables z and p as

respectively. The dynamical equations governing the evolution of the ray trajectories are
then
dz
--P,
dr

d v - Pz
dr r2

348
and correspondingly

au

dPz dr

dpm -_--

az

dr

au
ap

(15)

We must now interpret the two abstract momenta in terms of physical angles of
propagation. In the standard parabolic equation, the vertical grazing angle of a ray is
dz
tane=e=-,
dr

where IBI << 1. Thus, the momentum conjugate to depth is

as

p,=@=-.

aZ

Similarly, for this first order PE approximation, the azimuthal deflection angle can be
defined as
@ = raa
-,
ar

where

141<< 1. The momentum conjugate to azimuth is then defined by

or

Figure 3 displays the physical interpretation of these ray angles. The thick curve is a
ray trajectory launched at bearing p,, . The thin straight line is a radial at bearing p . The
intersection out at the maximum range is the ray location at position ( r , p ) . The
azimuthal deflection angle, given by Eq. (18), is indicated. A receiver at the position
(r,p) that is capable of measuring bearing (e.g., a horizontal array) would observe the
apparent bearing p = p + 4 for this ray. Thus 4 may be regarded as the bearing error
when compared to the environment with no lateral variability.
In terms of the physical angles defined above, the dynamic ray equations may now be
written

-dz= B ,
dr

dP-4
dr r

and

Thus, the bearing error of a ray observed at a particular depth and bearing at range R is
then

349

(23)

Figure 3. Schematic representation of ray horizontal deflection angle.

Let us now consider the corresponding ray equations for the UNCA approximation.
We begin by defining an appropriate solution in the geometric limit,

where now both the ray amplitude and phase are consistent with the uncoupled azimuth
approximation. The parabolic eikonal equation is then

As before, the vertical plane grazing angle is

and the azimuthal deflection angle is

qj

"

1 ds,
=--.
rdy,

We find that, since 4" remains finite if U ( r ,z , q ) depends on bearing , horizontal

refraction due to lateral variability remains even when the UNCA approximation is made.
This was first pointed out by Perkins and Baer. [4] The dynamicai ray equations are now

350
and

Although the bearing is constant in this case, the azimuthal deflection angle is not.
Specifically,

which maintains much of the influence of the horizontal refraction along the bearing.
4

3-D Ridge and Apparent 2-D Horizontal Refraction

In what follows, we shall limit our calculations to those predicted by the full-wave
parabolic equation model described in Section I. Specifically, we now introduce a threedimensional ridge-like structure with extremely large slopes. A diagram depicting the
cross section is presented in Fig. 4. The excessively large slope of +I- 20" on either side
of the ridge was chosen to maximize the influence of down-slope refraction. The
minimum depth at the peak of the ridge was chosen to be 200 m with a maximum depth
on either side of 2000 m. The water column is isospeed with a sound speed of 1500 m/s,
and the bottom has sound speed, density, and attenuation of 1700 d s , 1.2 g/cm3, and 0.01
dB/km/Hz, respectively. This environment extends infinitely in a direction perpendicular
to the plane of the paper.
position

A
a003 m

'I
= 0.01 dB/kmMz
Figure 4. Schematic diagram of 3-D ridge environment cross-section.

A broadband, omnidirectional, point source is used to simulate the propagation of a

pulse within the environment. The source location is directly over the ridge peak at a
depth of 100 m, and has a center frequency of 100 Hz with a 100 Hz bandwidth. The
solution to the acoustic field is stored roughly every 2 km out to a maximum range of 20
km for all depths and all computed radials. For this environment and source
characteristics, a minimum azimuthal sampling of about 0.2" is required out to the
maximum range of 20 km.
Standard plane-wave beamforming techniques are applied at ranges of 10 and 20 km
for the central 11.25" of computed radials (such that 0" is directly along the ridge) at the

35 1
source depth of 100 m. This corresponds to the central 64 radials of the calculation and
gives a total horizontal aperture of nearly 2000 m and 4000 m, with an inter-element
spacing of about 30 m and 60 m, at the ranges of 10 km and 20 km,respectively.
The results of beamforming the data from both the Nx2-D and 3-D calculations are
displayed in Fig. 5 (10 km) and Fig. 6 (20 km). In both cases, the two calculations
produce nearly identical results with respect to the predicted arrival angle structures.
However, the Nx2-D calculation contains considerably more diffuse energy across all
angles after the initial arrivals than the 3-D data. This is not surprising, since most of the
Nx2-D energy drops off roughly as cylindrical spreading without any additional loss due
to horizontal refraction. In general, both Nx2-D and 3-D calculations predict horizontal
refraction of the acoustic propagation. The arrival structures indicate that the earliest
arrivals are along the shortest path, directly along the isobath of the ridge peak. Later
arrivals trail off symmetrically to higher angles, indicating the energy refracted
downslope by the ridge.
Upon close inspection, it might be argued that the Nx2-D refraction angles are not as
large as the 3-D results. The ridge has a very large (20') slope, and so it is perhaps not
surprising that the Nx2-D calculation was unable to capture the total effect. However,
such large slopes are not typical of Ocean environments, and more realistic features would
probably be adequately modeled with the Nx2-D approach. It should also be noted that
the poor resolution of the arrivals is due to the relatively small number of frequencies that
could computed in a timely manner on a standard computer platform.

Summary

Using a full-wave, three-dimensional acoustic propagation model and its two-dimensional

counterpart, the influence of a sloping bathymetry has been investigated. In particular,
horizontal refraction as measured by a horizontal receiving array has been modeled. By
propagating a broadband pulse over a ridge-like structure with a 20' slope, both Nx2-D
and 3-D calculations exhibited horizontal refraction when standard beamforming
techniques were applied.
A ray theoretic approach was also provided. It was shown that the ray equations
appropriate for the uncoupled azimuth approximation still maintain much of the effect of
horizontal refraction. Thus, a proper two-dimensional calculation can still predict such
effects. This suggests that the efficiency gained by 2-D models is not necessarily at the
expense of the computation of 3-D effects.
In contrast, the downslope interference effects observed due to horizontal refraction
were predicted only by a 3-D model incorporating azimuthal coupling. The increase in
attenuation with range due to horizontal refraction was also unique to the fully 3-D
calculation. However, it was found that the correct prediction of these effects required
fine azimuthal sampling of the environment, which significantly increased the
computational load.
In summary, the influence of azimuthal coupling on energy propagation, and most
notably mode cutoff due to horizontal refraction, appear to require full 3-D modeling for
accurate predictions. However, such effects tend to require either large slopes or low
frequencies. In realistic ocean environments with uncertain 3-D environmental
variability, and typical sonar operating frequencies exciting many modes in shallow
water, the prediction or observation of these effects may be extremely difficult. Still, one
may hope to observe 3-D propagation by measurements of horizontal refraction.

352

7.2

7.3

7.4

7.5

7.6

7.2

7.3

7.4
Time (sec)

7.5

7.6

Figure 5. Beamformed data for broadband signal propagating over 20" ridge environment at a range of 10 h.
Upper plot shows results of Nx2-D (UNCA) calculation, lower plot displays 3-D results with influence of fully
coupled azimuthal propagation. Grayscale values are in units of dB re 1 m.

13.2

13.25

13.3

13.35

i32

1325

133

1335

13.4

13.45

13.5

13.55

13.6

134
1345
Time (sec]

135

1355

136

. --

I LU

Figure 6. Beamformed data for broadband signal propagating over 20" ridge environment at a range of 20 h.
Upper plot shows results of Nr.2-D (UNCA) calculation, lower plot displays 3-D results with influence of fully
coupled azimuthal propagatioii. Grayscale values are in units of dB re 1 m.

353
It has been shown both theoretically and numerically that such horizontal refraction
effects can be obtained from Nx2-D models. This is significant since such models are
much more efficient than their corresponding full 3-D implementations. This is, in part,
due to the fact that the environment need not be sampled as finely in azimuth, which is
more consistent with actual measurements of an experimental area. Such coarse sampling
could then more appropriately be used with a proper Nx2-D calcuation to estimate
horizontal refraction in a computationally efficient manner.

Acknowledgements

This work was supported by the Office of Naval Research, Code 3210A.

References
1. Heaney, K. D., Kuperman, W. A., and McDonald, B. E. (1991). Perth-Bermuda
sound propagation (1960): Adiabatic mode interpretation, J. Acoust. SOC.Am. 90,
2586-2594.
2. Doolittle, R., Tolstoy, A., and Buckingham, M. J. (1988). Experimental
confirmation of horizontal refraction of cw acoustic radiation from a point source in a
wedge-shaped ocean environment, J. Acoust. SOC.Am. 83,2117-2125.
3. Baer, R. N. (1981). Propagation through a three-dimensional eddy including effects
on an array, J. Acoust. SOC.Am. 69, 70-75.
4. Perkins, J. S. and Baer, R. N. (1982). An approximation to the three-dimensional
parabolic-equation method for acoustic propagation, J. Acoust. SOC.Am. 72, 5 15522.
5. Siegmann, W. L., Kriegsmann, G. A., and Lee, D. (1985). A wide-angle threedimensional parabolic wave equation, J. Acoust. SOC.Am. 78, 659-664.
6. Jones, R.M., Riley, J.P., Georges, T.M. (1986). HARPO: A Versatile ThreeDimensional Hamiltonian Ray Tracing Program for Acoustic Waves in an Ocean
with Irregular Bottom, Wave Propagation Laboratory, National Oceanic and
Atmospheric Administration, Boulder, CO, 457 pp.
7. Smith, K. B. (1996) Modeling the effects of azimuthal coupling on acoustic
propagation in the presence of 3-dimensional, rough ocean interfaces using the
parabolic approximation, in Special Issue of Theoretical and Computational
Acoustics, eds. Lee et al., 115-131 (World Scientific Publishing Co., 1996).
8. Smith, K.B., Miller, C.W., DAgostino, A., Sperry, B., Miller, J.H., and Potty, G.R.
(2002). Three-dimensional propagation effects near the Mid-Atlantic Bight, J.
Acoust. SOC.Am. 112,373-376.
9. Deane, G. B. and Buckingham, M. J. (1993). An analysis of the three-dimensional
sound field in a penetrable wedge with a stratified fluid or elastic basement, J.
Acoust. SOC.Am. 93, 1319-1328.
10. Glegg, S. A. L., Deane, G. B., and House, I. G. (1993). Comparison between theory
and model scale measurements of three-dimensional sound propagation in a shear
supporting penetrable wedge, J. Acoust. SOC.Am. 94, 2334-2342.
11. Fawcett, J. A. (1993). Modeling three-dimensional propagation in an oceanic wedge
using parabolic equation methods, J. Acoust. SOC.Am. 93, 2627-2632.
12. Smith, K. B. (1999). A three-dimensional propagation algorithm using finite
azimuthal aperture, J. Acoust. SOC.Am. 106, 3231-3239.

354
13. Tappert, F. D. (1977). The parabolic equation method, in Wave Propagation and
Underwater Acoustics, ed. J. B. Keller and J. Papadakis, 224- 286 (Springer-Verlag,
New York, 1977).
14. Chin-Bing, S. A., King, D. B., Davis, J. A. and Evans, R. B. (editors), PE Workshop
11: Proceedings of the Second Parabolic Equation Workshop, Naval Research
Laboratory NRL/BE/7181-93-0001 (US Government Printing Office), 1993, pp. 4955.
15. Smith, K.B. (2001). Convergence, stability, and variability of shallow water
acoustic predictions using a split-step Fourier parabolic equation model, J. Comp.
Acoust. 9, 243-285.
16. Jensen, F. B., Kuperman, W. A., Porter, M. B., and Schmidt, H. (1994).
Computational Ocean Acoustics (AIP, New York, 1994), Chaps. 5-6,271-412.
17. Hardin, R. H. and Tappert, F. D. (1973). Applications of the split-step Fourier
method to the numerical solution of nonlinear and variable coefficient wave
equations, SIAM Rev. 15, 423.
18. Thomson, D. J. and Chapman, N. R. (1983). A wide-angle split-step algorithm for
the parabolic equation, J. Acoust. SOC.Am. 74, 1848-1854.

355

BROADBAND PROPAGATION RESULTS FOR A SHALLOW

WATER GAUSSIAN CANYON TEST CASE USING A 3-D
PARABOLIC EQUATION MODEL

F. STURM
Institut National des Sciences Applique'es de Lyon, UMR CNRS 5514 and 5585,
Biit. Le'onard de Vinci, 20 auenue A . Einstein, 69621 Villeurbanne cedes, fiance.
E-mail: frederic.sturmQinsa-lyon.fr
The propagation of a broadband pulse in a three-dimensional canyon-type waveguide is studied numerically. The acoustic problem considered is a variant (though
very similar) of the 3-D gaussian canyon test case devised for the SWAM'99 Shallow Water Acoustic Modeling Workshop. The numerical method used to solve
this 4-D acoustic problem is based on Fourier synthesis technique. The calculations in 3-D relative to each discrete frequency are carried out using the fully 3-D
parabolic equation based model 3DWAPE recently modified to include wide angle
paraxial approximation in azimuth. A source pulse centered at 25 Hz is considered.
First, the 3-D results corresponding to a 25Hz-CW point source are presented and
compared with the predictions of 3-D adiabatic mode theory. Then, the acoustic
problem is solved considering the broadband source pulse. The received signals
on distinct vertical arrays placed along the canyon axis are analyzed and exhibit
multiple mode arrivals of the propagating signal.

Introduction

It has been demonstrated both experimentally1 and numerically2 that in some

particular oceanic environments, the horizontal refraction of the propagating sound
waves can not be neglected and leads, far from the source, to typical 3-D effect^.^
Only fully three-dimensional models can predict such 3-D effects (a model is said
3-D on condition it allows for azimuthal coupling of the propagating field, otherwise
it is said 2-D, Nx2-D or pseudo-3-D). Among the existing 3-D codes available in
the underwater acoustics ~ o m m u n i t y , 4 * ~
parabolic
~ ~ ~ ~ J ?equation (PE) based models
are largely used since they are efficient for solving complex sound propagation
problems in various environment^.^ The main drawback of the 3-D codes is that
they are computationally intensive (a 3-D code is at least two orders of magnitude
slower than any Nx2-D code). However, the high performance of recent computers
allows now to run a 3-D PE code with reasonable CPU times. The calculation in
4-D (ie., three space dimensions plus time) using Fourier synthesis is thus possible
(at least at low frequencies) and the question of broadband signal dispersion in
general 3-D waveguides can be addressed.1
The aim of this paper is to study the propagation of a broadband pulse in a
three-dimensional canyon-type waveguide. The numerical method used to solve the
P D acoustic problem is based on Fourier synthesis of frequency-domain solutions.
The calculations in 3-D are done using the fully three-dimensional parabolic equation based code 3DWAPE.l' The model is presented in section 2 and its wide angle
capability in depth and azimuth is discussed. In section 3, the acoustic problem
is described. It corresponds to a variant of the original SWAM'99 3-D gaussian
canyon test case.12 The modifications of the original test case have been made such

356
that only three propagating modes now exist at the source position (in its original form, eleven propagating modes were present) which makes easier the modal
structure analysis of the received signals. Results corresponding to both CW calculations and time series are presented and discussed: the 3-D results corresponding
to a 25Hz-CW point source are f i s t analyzed and thoroughly compared with the
prediction of 3-D adiabatic mode theory, showing good agreement between the two
approaches. Then, the acoustic problem is solved considering the broadband source
pulse. The received signals on two distinct vertical arrays placed along the canyon
axis are analyzed and exhibit multiple mode arrivals of the propagating signal.
2

Mathematical Modeling

We consider a multi-layered waveguide composed of one water layer overlying one

(or several) fluid sedimental layer(s). The geometry of each layer is fully threedimensional. Cylindrical coordinates are used, with z being the depth below the
ocean surface, 8 being the azimuthal angle and r being the horizontal range, related
to the Cartesian coordinates by x = r cos 8 and y = r sin 8. An isotropic broadband
point source S is placed at range r = 0 and at depth z = zs. We solve the acoustic
wave equation

where P = P ( r , 8 , z ; t ) is the acoustic pressure as a function of the three space

variables r , 8, z plus the time variable t , and where S ( t ) is the time-dependence
of the point source S . The terms c and p represent, respectively, the varying (in
space) sound speed and the density term (constant on each layer). In order to
reduce to three the dimension of the previous 4 D equation, we proceed by Fourier
synthesis. Let p denote the Fourier transform of the time-domain acoustic pressure
P , defined by
A

P(r,8, z ; w ) =

['" P (r,8, z ; t)eiwtdt,

J--00

where w = 27r f . Then taking the Fourier transform of Eq. (1)leads to the frequencydomain (or 3-D Helmholtz) equation

where ka is the complex (to account for lossy layers) wavenumber and where S^(w)
is the source spectrum defined by

S(w) =

1,
+-O0

S(t)eiUtdt.

The angular frequency w is now treated as a parameter in Eq. (3). The complexvalued field = ?(r, 8, z ; w ) is sought as a function of only the three space variables
r , 8 and z. The acoustic pressure 5 is assumed to satisfy a pressure-release condition at the horizontal ocean surface, an outgoing radiation condition at infinity, a

357
2x-periodicity condition in azimuth and appropriate boundary conditions-at each
sedimental interfaces. We then transform the frequency-domain solution P to the
time-domain using the following inverse Fourier transform

+w

P(r,8,z;t)= 2n

-,-&

P(r,8,z ; w)e-iwt dw,

(4)

with the time-domain acoustic pressure P satisfying the initial 4-D wave equation (1). Eq. (3) is solved numerically for selected discrete frequencies within the
frequency band of interest, using a PE appToach. The cylindrical spreading is
handled by factorizing the acoustic pressure P as

where Hi1) denotes the zeroth-order Hankel function of the first kind and ko =
w/cref with cref a reference sound speed. Assuming that r P 2 approximately commutes with d / d r for r >> 0, Eq. (3) can be factorized and assuming only outward
propagation in range, the 3-D problem based on Eq. (3) is reduced to an initialand boundary-value problem. We seek a complex-valued function G = G(r,8,z; w ) ,
which represents the acoustic field for ro 5 T 5 rma, 0 5 8 5 27r and 0 5 z 5 Z m U ,
and which satisfies

a0)

and G(r = ro, 8, z ; w ) = 8'O)(8,z ; w ) where

denotes the initial outgoing field at
r = 7-0 and where Zis the identity operator and

p a I d
1 d2
Y = -X = (ni - 1)Z+ --(--),
ki dz p dz
kir2 a82 '
where n, denotes the complex index of refraction. The operator Y handles the
azimuthal diffraction term. Assuming that 3-D effects are sufficiently gradual, one
can write

where
and bk,np for 1 5 k 5 np are real or complex valued Pad6 ~0efficients.l~
By neglecting the last term in ( 6 ) , Eq. (7) is replaced by

The approximate model given by Eq. (7) accounts for refraction effects which are
higher in depth than in azimuth. The 3-D P E model obtained is called 3DWAPE.I'
Eq. (7) is solved numerically using a splitting method. The partial derivative with
respect to the depth present in X is approximated using a finite element scheme.
The partial derivative with respect to the azimuth present in Y is discretized using
an efficient higher-order accurate finite difference scheme which allows both a significant reduction of the azimuthal sampling and a speed-up of the computation.
Notice that neglecting Y would lead to a Nx2-D (or pseudo-3-D) model which

358
would be unable to predict the horizontal refraction of the energy. In the next
section, instead of the linear approximation in azimuth present in Eq. (6), we will
use the following higher-order pad6 series expansion in a z i m ~ t h : ~

b,,m, = cos2(&)
for 1 5 k
sin2(*),
where a,,m, = l2m,+l
that this approximation is now in O(yzm,+l)but still in O ( X y ) .
3

5 m,,. Notice

Numerical simulations

3.1 Description of the three-dimensional canyon waveguide

In this section, we describe the 3-D canyon test case. It consists of an isovelocity
water layer of sound speed c, = 1500 m/s and density pw = 1 g/cm3, overlying
a lossy homogeneous halfspace sedimental layer of sound speed C&d = 1700 m/s,
density psed = 1.5 g/cm3 and attenuation (Ysed = 0.1 dB/X. The parametrization of
the water/sediment interface is given by : Zsed(P,8) = h s e d ( Z ) with 2 = r cos8 and
y = P sin 8 and h s e d ( 2 ) = ho hl exp (- x 2 / u 2 ) , where the parameters ho, hl and
0 are expressed in meters and where
= 4hl. The values ho = 200 m, hl = 500 m

Figure 1. Geometry of the 3-D canyon waveguide. The point source S is located midway across
the canyon. The two vertical arrays A1 and A2 and the source S lie in the same vertical plane
(corresponding to 8 = 90 deg) for which the water depth is constant.

and u = 2000 m correspond to the values used for the SWAM99 canyon test case.12
Here we use ho = 20 m, hl = 180 m and (T = 720 m. In both case, the maximum
slope of the bathymetry in the z-direction (to which we will refer to as the acrosscanyon direction) is approximately 12.1 deg at z = ./&. An isotropic point source
is located at Cartesian point S = (zs = 0, ys = 0, zs = 40 m). Its time dependence

359

005

IMAGINARY PART

,
I

I
I

I
I

I
,

I
4 f f i - !
0 5 1

40Hz

I-

I
"

I
"

2 5 3 0 3
FREQUENCY W )

Figure 2. Hanning-weighted four-period sine pulse (left subplot) and spectrum (right subplots).

is a Hanning-weighted four-period sine wave (cf. Fig. 2) given by

- (1- cos(wct/4)) x sin(w,t) if 0 5 t 5 4/fc

0 if t > 4/fc

(8)

where wc = 27rfc and fc = 25 Hz. The source pulse is centered at frequency

fc = 25 Hz with a bandwidth of 40 Hz. We consider two distinct vertical arrays
labelled A1 and A2 placed along the canyon axis ( i . e . , 6 = 90 deg) respectively
at ranges r A l = 16 km and r ~ =
2 20 km from the source S. Each of the vertical
arrays is composed of 19 elements evenly spaced in depth between 10 m and 190 m.
3.2 CW-point source modeling
Since the source pulse is centered at fc = 25 Hz, we first analyze the acoustic
problem at that specific frequency. The results presented hereafter are carried out
using Ar = 10 m, AZ = 2 m and an eighth-order finite difference (FD) azimuthal
scheme which requires only 2880 discrete points in azimuth ( i .e, an azimuthal step
size of 1/8th of a degree). Notice that using a second order FD azimuthal scheme
requires an azimuthal step size of 1/64th of a degree.14 We use an artificial absorbing
= 600 m with a depth-dependent attenuation
layer from &bs = 450 m to ,z
coefficient (Yabs(Z) given by aabs (Z) = (Ysed gabs (Z - dabs) for Zabs z 5 ,z
with
gabs = 0.05 (dB/X)/m. The maximum computation range is r,,
= 20 km and the
reference sound speed is Cref = 1500 m/s. Both Nx2-D and 3-D calculations are
generated using a fifth-order Pad6 2 approximation in depth (i.e., np = 2). For
the 3-D calculation, we use a third-order Pad6 1 approximation in azimuth ( i e . ,
mp = 1). Gray-scale images of the transmission loss (TL) (horizontal slices at
z = 35 m) corresponding to Nx2-D and 3-D calculations are shown in Fig. 3. Due
to the geometrical symmetry of the problem about the x-direction, both Nx2-D

360

10

12

14

16

18

20

Figure 3. Transmission loss (horizontal slices at z = 35 m) corresponding to Nx2-D (upper

subplot) and 3-D (lower subplot) calculations. Both solutions have been obtained using 3DWAPE.

and 3-D solutions are displayed as a function of range and azimuth in a limited
azimuthal sector. By comparing the two subplots of Fig. 3, we observe noticeable
differences in both fields. The effects of azimuthal coupling are evident. When
azimuthal coupling is taken into account in the calculation (which is not the case
for any Nx2-D model), the acoustic energy is horizontally refracted by the sidewalls
of the canyon and gets channeled in the y-direction along the canyon axis. Notice
that both solutions are symmetric about the along-canyon direction.
We compare in Fig. 4 the 3-D solutions corresponding to different paraxial
approximations in azimuth. The dashed curve corresponds to the narrow-angle
Tapperts approximation in azimuth, the solid curve to a third-order Pad6 1 a p
proximation in azimuth (mp = 1) and the dashed-dotted curve to a fifth-order
Pad6 2 approximation in azimuth (mp = 2). Also shown in the same plot in bold
solid line is the Nx2-D solution. The comparison of any of the 3-D solutions with
the N x 2-D solution confirms an enhancement of the acoustic level when azimuthal
coupling is handled, due to the focussing of the acoustic energy along the canyon
axis. We see that the 3-D solutions corresponding to mp = 1 and mp = 2 agree
extremely well. On the other hand, the validity of the narrow angle approximation
in azimuth could be discussed for the present test case. In order to characterize
the 3-D effects present in the canyon, we now excite individually each of the three
propagating modes at the source location and propagate it outward in range. For
each modal initialization, the reference sound speed is selected equal to the phase
speed of the mth mode excited at the source range. We display in Fig. 5 (left
column) the gray-scale TL plots (horizontal slices at constant depth z = 35 m)
obtained initializing the 3-D model SDWAPE with mode 1, mode 2 and mode 3. We
also display in the same Fig. 5 (right column) the modal ray diagrams (derived
from 3-D adiabatic mode theory15) for each omnidirectional modal initialization.

36 1
RECEIVER DEPTH = 35 m

- AZIMUTH = 90deg

--------

3-0- Pad6 2 /Tappert

3-D

_ _ - _ _ -

- Pad6 2 I Pad6 1

3-D - Pad6 2 I Pad6 2

Nx2-D

-40

-0

-50

::

3
+ -60
-70

-80

-90

8
10
12
ALONG CANYON (km)

14

16

18

20

Figure 4. Transmission loss curves at I = 35 m and 0 = 90 deg (along canyon) obtained running
JDWAPE with different paraxial approximations in azimuth.

The modal ray paths have been calculated following the method given by Fawcett.6
By comparing the left and right columns of Fig. 5, the solutions obtained using
the PE approach and the adiabatic modal ray theory are satisfactorily in good
agreement. The effects of the 3-D varying bathymetry on the different modal propagations are now evident. We observe that modes 1, 2 and 3 are present at receiver
arrays A1 and A2 (no shadow zone along the canyon axis). We also observe that
both receiver arrays A1 and A2 lie in the multiple mode arrival area of each of the
three propagating modes. Anticipating the analysis presented in the next section,
we thus predict multiple arrivals (not necessarily distinct in time) for each of the
three propagating modes on both receiver arrays A1 and A2. On the contrary, any
azimuthally uncoupled model should predict only one single arrival for each of the
three propagating modes.
3.3 Broadband modeling

Let turn now to the analysis of the acoustic problem involving the broadband source
signal given by Eq. (8). The signal arrivals on vertical arrays A1 and A2 computed
with the fully three-dimensional PE model 3DWAPE are displayed in the fourth
columns of Figs. 6 and 7. For comparison, we also display on the same figures the
received signals obtained using a N x 2-D ( 2 . e. azimuthally uncoupled) model. The
signals are plotted intentionally as stacked time series us. depth which is helpful
when analyzing the modal structure of the signal arrivals. The Nx2-D results (see
fifth columns of Figs. 6 and 7) clearly show, as expected, that the propagating signal
splits up in three distincts wave packets which correspond to the signals carried by
the three propagating modes of the waveguide. Now, due to the focussing of the
acoustic energy by the sidewals of the canyon, the 3-D results are much more

362
HORIZONTALSUCE- RECENER DEPTH = 35 rn

0''

10

10

12

12

14

14

16

16

18

18

10

12

14

16

18

20

F2
rIw2'

'

'

'

'

'

'

'

'

12

20

XI

i-lt
8-21

-2

8
10
12
ALONGCANYON &n)

14

16

18

20

MODEI'

'

10

'

'

'

14

16

'

'

'

8
10
12
ALONGCAWON (km)

14

16

'

1
1

18

20

18

XI

Figure 5. Left column: transmission loss (horizontal slices at constant depth I = 35 m) corresponding to 3-D calculations (with 3DWAPE) for different modal initializations; right column:
corresponding modal ray diagrams. From top to bottom: mode 1, mode 2 and mode 3.

complicated (see fourth columns of Figs. 6 and 7). Though the first wave packet
can clearly be attributed to the first arrival of mode 1 (its amplitude being low near
z = 0, increasing with depth toward mid-depth, then decreasing toward 200 m),
the rest of the received signals do not clearly show any modal structure.
One way to discriminate the multiple arrivals of one mode from the multiple
arrivals of the two other modes consists in initializing (in the Fourier domain) the
PE model by only one single mode (if present) at each single discrete frequency,
The results obtained intializing separately by mode 1, by mode 2 and by mode 3 are
displayed in the first, second and third columns of Figs. 6 and 7. It is also helpful
to isolate from the modal ray diagrams given in the right column of Fig. 5 the
modal ray paths that connect the source 5' and one of the receiver arrays A1 and
A2. These modal eigenrays are displayed in Fig. 8. They confirm multiple arrivals
for each of the three propagating modes on receiver arrays A1 and A2. It is also
interesting to notice that multiple arrivals of the same mode can be well separated
in time. For instance, this is the case for the two distinct arrivals of mode 1 on A1
(see first column of Fig. 6) and for the two distinct (though very close) arrivals of
mode 2 on A1 (see second column of Fig. 6). This is also the case for the three
distinct arrivals of mode 1 on A2 (see first column of Fig. 7). On the contrary, the
multiple arrivals of mode 3 on A1 (see third column of Fig. 6) are very close in
time and the wave packets observed correspond to the merger of the three different

363
3-D

0,

- MODE 1

[---------1o,

3-D

- MODE 2

3-D - MODE 3

N~ 2 - D

3-D

10

-t

-t

m
0

m
0

0
to

0
to

r
0
.

Figure 6. Stacked time series vs. depth corresponding to vertical array A1 (placed along the
canyon axis at r = 16 km) obtained using a 3-D computation.

arrivals of the same mode.

4

Concluding remarks

In this paper, we have studieL the propagation of a broadbanc. pulse in a 3-D canyon
waveguide. The analysis of the received signals has clearly shown the presence of
multiple arrivals for each of the three propagating modes on the two receiver arrays
placed along the canyon axis. The multiple arrivals for each of the propagating
modes is due to the focussing of the acoustic energy by the sidewalls of the canyon.
The situation is different from what can be observed for the 3-D wedge-shaped
waveguide for which a shadow zone is present across-slope for each mode.16J0
References
1. R. Doolittle, A. Tolstoy and M.J. Buckingham, "Experimental confirmation of
horizontal refraction of CW acoustic radiation from a point source in a wedge-

364
3-D

- MODE 1

1 z1 -

3-D - MODE 2

3-D - MODE 3

3-D

N x 2-D

Figure 7. Stacked time series us. depth corresponding to vertical array A2 (placed along the
canyon axis at T = 20 km) obtained using a 3-D computation.

shaped ocean environment, J. Acoust. SOC.Am. 83,2117-2125 (1988).

2. M. J. Buckingham, Theory of three-dimensional acoustic propagation in a
wedgelike ocean with a penetrable bottom, J . Acoust. SOC.Am. 82, 198-210
(1987).
3. A. Tolstoy, 3-D propagation issues and models, J . Comput. Acoust. 4,
243-271 (1996).
4. D. Lee, G. Botseas and W.L. Siegmann, Examination of three-dimensionnal
effects using a propagation model with azimuth-coupling capability (FORSD),
J. Acoust. SOC.Am. 91, 3192-3202 (1992).
5. G.H. Brooke, D.J. Thomson and G.R. Ebbeson, PECan: A Canadian
Parabolic Equation Model for Underwater Sound Propagation, J. Comput.
Acoust. 9, 69-100 (2001).
6. J.A. Fawcett, Modeling three-dimensional propagation in an oceanic wedge
using parabolic equation methods, J. Acoust. SOC.Am. 93,2627-2632 (1993).
7. K.B. Smith, A three-dimensional propagation algorithm using finite azimuthal
aperture, J. Acoust. SOC.Am. 106, 3231-3239 (1999).

365
TOP VIEW

TCQVIEW

. .
.

10

12

14

16

18

1 0 1 2 1 4 1 6 1 8 2 0

10

12

1 0 1 2 1 4 1 6 1 8 2 0

-1

p'0-2

-2
0

a
2

fiWMW)

20'$

14

16

18

20

M WW
C
N (h)

Figure 8. Modal ray paths that connect the source S and receiver array A1 (left column) or
receiver array A2 (right column). For each path, 40 denotes the angle that the modal ray makes
initially at the source (40 = Odeg points along the canyon).

a.
9.
10.
11.
12.

13.
14.
15.
16.

M.D. Collins and S.A. Ching-Bing, "A three-dimensional parabolic equation

model that includes the effects of rough boundaries", J . Acoust. SOC.Am. 87,
11041109 (1990).
D. Lee, A.D. Pierce and Er-Chang Shang, "Parabolic equation development in
recent twentieth century", J. Comput. Acoust. 8 , 527-637 (2000).
F. Sturm, "Examination of Signal Dispersion in a 3-D Wedge-Shaped Waveguide using 3DWAPE", Acust. Acta Acust. 88, 714717 (2002).
F. Sturm and J.A. Fawcett, "On the use of higher-order azimuthal schemes in
3-D PE modeling", J. Acoust. SOC.Am. 113 (6), 31343145 (2003).
A. Tolstoy, K. Smith and N. Maltsev, "The SWAM'99 Workshop - an
Overview", J. Comput. Acoust. 9, 69-100 (2001).
M.D. Collins, "Higher order Pad6 approximations for accurate and stable elastic parabolic equations with application to interface wave propagation" , J.
A C O U SSOC.
~ . Am. 89, 1050-1057 (1991).
J.I. Arvelo, A.P. Rosenberg, "Three-dimensional effects on sound propagation
and matched-field processor", J. Comput. Acoust. 9, 17-39 (2001).
H. Weinberg and R. Burridge, "Horizontal ray-theory for ocean acoustics", J.
Acoust. SOC.Am. 5 5 , 63-79 (1974).
E.K. Westwood, "Broadband modeling of the three-dimensional penetrable
wedge", J. Acoust. SOC.Am. 92 (4),2212-2222 (1992).

366

MODELING THE PORE STRUCTURE EFFECT ON ACOUSTIC

WAVE AMPLITUDE IN SEDIMENTARY ROCKS
WE-FENG SUN
Lamont Doherty Earth Observatory, Columbia University
Palisades, NY l O 9 6 ~USA
,
E-mail: sunyf@ldeo.columbia.edu
Acoustic wave propagation is sensitive to many rock properties including fluid content, porosity, and pore structure, among others. Pore structure can be one of the
important parameters in controlling both seismic wave velocity and permeability
in sandstones and carbonate rocks. For a given porosity of two similar rocks with
different pore structures, their acoustic wave speeds can differ 2 km/s and permeability can span nearly six orders of magnitude from 0.01 mD to 20 D in both
sandstones and limestones. In this paper, we introduce a two-parameter velocity
model and define three pore structure types (PST) to quantify the pore structure
effects on elastic properties of sedimentary rocks. These three PSTs have their
distinct characteristics on synthetic shot gathers. This study indicates that it may
be feasible to use this new concept and method t o detect pore structure variations
in reservoir rocks from field seismic data. This study also helps explain why AVO
in some cases fails for fluid detection: pore structure effect on seismic waves can
mask all the fluid effects, especially in carbonate rocks.

Introduction

One of the major tasks in seismic inversion and reservoir characterization is to

estimate three important reservoir properties: porosity, fluid saturation, and permeability, which has been conventionally the ultimate goals of well logging. No
matter what approach (deterministic or statistical) is used, a 3-dimensional porosity volume has been a key product of seismic data mining, using simple or excluding altogether rock physics models. Next, a permeability volume is produced
using simple poroperm relationships. Through up-scaling, these seismic inversion
products are further fed to reservoir flow simulation models which usually predict
results against in-situ production data in both geometrical distribution and physical
and hydraulic properties.
Among the problems causing these poor predictions are low seismic resolution
and data processing accuracy, and a poor understanding of the fundamental links
between seismic properties and rock properties. Simple one-parameter models of
porosity-velocity and porosity-permeability relations are far from a good approximation of reality. In very complex ways, both elastic properties and hydraulic
properties depend on mineralogy, fluid content, porosity, pore structure, pressure
and temperature. In this paper, we study the importance of pore structure affecting both seismic wave velocity and permeability and illustrate a method to analyze
seismic shot gathers using AVO analysis to estimate pore structure variations in
reservoir rocks.

361
2

Pore Structure, Velocity, and Permeability

Biots theory of wave propagation in porous media and its various extensions have
been increasingly used in seismic reservoir studies. However, the effects of fractures
and cracks in natural rocks have been largely ignored in his theory, whereas other effective medium theories consider only noninteracting cracks. Sun (1994) developed
a topological characterization of structural media that provides a representation of
the internal structure of a fractured porous medium at the microscopic scale and
investigated the general mechanics and thermodynamics of fractured porous media.
One of major developments in the mathematical theory of fractured porous media
(Sun, 1994) is to quantity directly the profound effects of pore structure and its
connectivity on the elastic properties of rocks.
For two rocks of the same porosity and saturation, their acoustic wave speeds
can differ 2 km/s and permeability can span nearly six orders, resulting from their
difference in pore structures and pore size (e.g., Wang et al., 1991; Anselmetti and
Eberli, 1993; Sun, 1994; Basan et al., 1997). Nur and Simmons (1969) and Gregory (1976) report velocity measurements on dry Bedford limestone (porosity 12%)
and Leuders limestone (porosity 17%). As in Fig. 1, compressional wave velocity
(Vp) at zero differential pressure in the Leuders limestone (Fig. lb) is more than 1
km/s higher than that in the Bedford limestone (Fig. la), even though its porosity is 5% greater. Sun and Goldberg (1997) estimate pore structure changes with
pressure using these data to explain such seemingly abnormal velocity data. The
Bedford limestone (Fig. lc) has a greater number of small fractures and cracks
and it is populated with cracks of aspect ratio nearly 4 times smaller than in the
Leuders limestone (Fig. Id). In Fig. 2a and 2b, the velocity measurements and
estimated aspect ratio variations with pressure on a high-porosity (21.7%) Gulf
Coast sandstone shows also a similar strong effect of pore structure on controlling
the compressional wave velocity of the rock. King et al. (1995) report permeability
measurements under hydrostatic pressures while velocity measurements are measured on a dry sandstone of a porosity 13%with clay content less than 3% (Fig. 2c
and 2d). Similar measurements are performed after aligned cracks are introduced
to the same sample after applying independently controlled triaxial stresses. King
et al. found a good correlation between the crack density changes with pressures
and the velocity and permeability variations. These results demonstrate that pore
structure plays an important role in controlling both elastic and hydraulic properties of both sandstones and carbonate rocks. For oil-saturated reservoir rocks,
similar pore structure effects on acoustic velocity should also be detectable depending on the pore structure type and in-situ reservoir pressure. Fig. 3 shows these
effects on velocity measurements on water-saturated clean sandstones reported by
Han et al., (1986).

Pore Structure Type and AVO signature

The ultrasonic velocity changes caused by pore structure variations are usually
quite greater than those allowed by seismic dispersions. It is thus expected that
the pore structure variations should have appreciable characteristics on seismic

368
5.0

'

'

'

4.5

-2
3>

120

4.0

100

3.5

porosity 12%

3 3.0

80

-.- - - - - - - -

2.5

60
40

2.0
1.5
l.Oi-' I
0
50

"

100

'

150

' - 1

200

50

Differential Pressure (MPa)

(a)

4.5

200

150

100

DifferentialPressure (MPa)

(d
40.0

37.5

'3i 4.0

35.0 -

limestone
porosity 17%

32.5
30.0 27.5 :

20

40

60

80

100

25.0
0

Differential Pressure (MPa)

'

"

"

"

'

40
60
80
DifferentialPressure (MPa)

20

100

(4

(b)

Figure 1. Acoustic wave velocity and estimated pore structure parameter (aspect ratio a) versus
differential pressure for two dry limestones.

wave propagation.
Suppose that the fluid content and saturation and the matrix mineralogical composition are known for a given reservoir. The heterogeneity and property variation
of the reservoir can be ascribed to the changes in both porosity and pore structure
(Sun, 2000; Sun et al., 2002). If we define an effective pore structure factor f to
characterize the combined effects of crack density and pore shape, wave velocity
( V )can then be approximated as a tweparameter function of the porosity (4) and
this effective pore structure factor, (f):

v = V ( 4 ,f )The velocity variation can then be described as

dV
dV
6V = -64 + -6f.
84
df
It means that for a given porosity and 64 = 0, the velocity variation is quantified by
a variation in f . Although porosity 4 and f could be mathematically independent,
the distributions of pore structure factor f in naturally occurring rocks are likely

369
4.0

Gulf coast sandstone

4-0-

---

=
-

20

1.5

4.5

40

80

60

100

Differential Pressure (MPa)

Differential Pressure (MPa)

(4

(b)

160

Sandstone. porosity 13%

With aligned crack

130
2.0
1.5

20
30
40
Effective Pressure (MPa)

10

50

120

(d

20
30
40
Effective Pressure (MPa)

10

50

(d)

Figure 2. Acoustic wave velocity, estimated pore structure parameter (aspect ratio a) and permeability versus effective pressure for dry sandstones.

related to the porosity distribution statistically, depending on the formation history

and diagenesis. That is, we assume

= fo(4),

(3)
where the averaging process is for each porosity value. We can then classify rocks
into three possible pore structure types (PST):
PSTl: f < fo - Afo;
PST2: fo - Afo < f < fo+Afo;
PST3: fo

+ Afo < f ,

where Afo is a tolerance value for fo for a given porosity, which can be set to be
proportional to fo as Afo = cfo, and c is a constant.
To illustrate the concept of effective pore structure type (PST) and its characteristics on seismic wave propagation, we use a simple two-layer model consisting
of a limestone reservoir of a given porosity 30% underlying a thick shale layer, as

370

5.6 I
5.5

5.4

5.3

5.2
5.1

5.0
\

hL
v

4.9

4.8

4.7

4.6

4.5

4.4

4.3
4.2

'

10
20
30
40
Effective Pressure (MPa)

50

Figure 3. Acoustic wave velocity versus effective pressure, for water-saturated clean sandstones
with different porosities. (From Han et al., 1986.)

shown in Fig. 4. Table 1 lists the typical seismic wave parameters for both the
encasing shale and the limestone reservoir with different PST and saturation for
six model cases.
We solve the complete Zoeppritz equation for six cases of a plane P-wave incident
on the shale/limestone interface with model parameters given in Table 1. Fig. 5
shows the P-wave reflection coefficients versus angle of incidence. Dotted curves
w l , w2, and w3 are for the water-saturated reservoir having a different PST1,
PST2, and PST3, respectively. Solid curves gl, g2, and g3 are for the dry or gassaturated reservoir having a different PST1, PST2, and PST3, respectively. Fig.
6 shows the synthetic seismogram gathers for these six cases, generated using a
Ricker wavelet of 25 Hz convolved with the calculated reflection coefficients. Fig. 5
shows that for carbonate reservoirs with a variation of pore structures, conventional
AVO analysis can hardly succeed for any fluid detection purposes. Such AVO fluid

37 I

source
receiver receiver

receiver

receiver

water-, or gas-saturated limestone of different pore structure types with a given porosity
30%. Detailed physical parameters are given in Table 1 .

Figure 4. Schematic diagram of a shale/limestone interface with illustrative seismic ray paths
from a P-wave source. Reflected converted shear wave and transmitted wave rays are not shown.

detection is possible only where the pore structure type is PST1. In this case gassaturated reservoir reveals a complete polarity reversal for the entire range of angle
of incidence with no critical refraction. Where the reservoir has a pore structure
type PST2 or PSTS, it becomes almost impossible to see the fluid effects on seismic
waves. However, it is evident that these three PST types have very distinctive
signatures on their critical angles of incidence (Fig. 5 ) , combined with abnormal
amplitude increase near critical refraction (Fig. 6). As shown in Figure 7, finiteelement modeling (FEM) of two-dimensional elastic wave propagation in these six
cases is also used to verify the PST concept. The direct waves are subtracted from
the FEM seismograms to show the reflected waves. The FEM results agree very
well with the analytical Zoeppritz solution.
Table 1. Physical and elastic properties of the synthetic model

w-limestone: water- or oil-saturated limestone;

g-limestone: dry or gas-saturated limestone.

372
1.o

0.9
0.8
0.7

I \

0.6

-1

0.5

0.4
0.3
0.2
0.1

0.0
0.1

0.2

0.3
0.4

0.5
0.6

0.7
0.8
0.9

1.o

45

50

55

60

65

70

75

80

85

Angle of incidence
Figure 5 . Zoeppritz P-wave reflection coefficients for a shale/limestone interface for three possible
pore structure types PST1, PST2, and PST3 of a limestone reservoir with a given porosity of
30%. Solid curves (gl, g2, g3) are for a gas-saturated reservoir, whereas the dotted curves (wl,
w2, w3) for a water-saturated reservoir.

Concluding Remarks

We introduced a two-parameter velocity model and defined the concept of effective

pore structure type (PST). For a proven reservoir, the concept of PST can be used
as a spectrometer in the AVO domain to detect the rock pore structure type,
which will be useful to improve the accuracy of porosity inversion and possibly to
derive permeability from seismic data. These concepts and methods could help
explain the frequent failures of AVO for hydrocarbon identification, epecially in
carbonate reservoirs. This method could be readily tested using long offset 3D
seismic shot gathers or CMP gathers.

313
wl

w2

w3

e3

Figure 6. P-wave synthetic seismograms generated using the Zoeppritz equation for a
shale/limestone model. See text for details.

wl

w2

w3

el
"

e2

e3

Figure 7. P-wave synthetic seismograms generated using the finite element method for a
shale/limestone model. See text for details.

374

Acknowledgments
The author thanks J. L. Massaferro and G. P. Eberli for helpful discussions. He
is grateful to Dr. Y. -C. Teng, Fermaco, Inc., New York, for the finite-element
modeling work. This work was supported partially by the Doherty Foundation,
National Science Foundation, and Shell Oil Company. LDEO contribution number
6518.

References
1. Anselmetti, F.S., and G.P. Eberli, 1993. Controls on sonic velocity in carbonates. Pure Appl. Geophys., 141, 287-323.
2. Basan, P.B., B.D. Lowden, P.R. Whattler, and J.J. Attard, 1997. Pore-size
data in petrophysics: a perspective on the measurement of pore geometry, in
Lovell, M. A. & P. K. Harvey (eds.), Development in Petrophysics, Geological
Society Special Publication No. 122, pp 47-67.
3. Gregory, A.R. 1976. Fluid saturation effects on dynamic elastic properties of
sedimentary rocks. Geophysics, 41, 895-921.
4. Han, D.-H., A. Nur, and D. Morgan, 1986. Effects of porosity and clay content
on wave velocities in sandstones. Geophysics, 51, 2093-2107.
5. King, M. K., Chaudhry, N. A. & A. Shakeel 1995. Experimental ultrasonic
velocities and permeability for sandstones with aligned cracks. Int. J. Rock
Mech. Min. Sci. & Geomech. Abstr., 32, 155-163.
6. Nur, A. & Simmons, G. 1969. The effect of saturation on velocity in low
porosity rock. Earth and Planet Science Letters, 7, 183-193.
7. Sun, Y.F. 1994. On the Foundations of the Dynamical Theory of Fractured
Porous Media and the Gravity Variations Caused by Dilatancies, Ph.D. dissertation, Columbia University.
8. Sun, Y .F. 2000. Core-log-seismic integration in hemipelagic marine sediments
on the eastern flank of the Juan de Fuca Ridge, I n Fisher, A., Davis, E. E., and
Escutia, C., (Eds.) Proceedings of Ocean Drilling Program, Scientific Results,
168: Texas A&M, College Station TX, 21-35.
9. Sun, Y.F., and D. Goldberg, 1997. Effects of aspect ratio changes on wave
velocities in fractured rocks, SEG Expanded Abstract, 67, 925-928.
10. Sun, Y.F., J.L. Massaferro, G. Eberli, and Y.-C. Teng, 2002. Quantifying
the effects of pore structure and fluid saturation on acoustic wave velocity in
carbonates, I n E.-C. Shang, Qihu Li, and T.F. G m (Eds.), Theoretical and
Computational Acoustics 2001, World Scientific Publishing Co., 335-347.
11. Wang, Z., W.K. Hirsche, and G. Sedgwick, 1991. Seismic velocities in carbonate rocks. J. Canadian Petro. Tech., 30, 112-122.

375

SEISMIC RESOLUTION: AN OLD PROBLEM BUT A NEW

CHALLENGE FOR SEISMIC RESERVOIR CHARACTERIZATION
YUE-FENG SUN
Lamont-Doherty Earth Observatoiy, Columbia University
Palisades, NY I0964 USA

YU-CHTITNG TENG
Fermarco, Inc.
New York, Ny 10025, USA

CUIXIA REN, LICHENG LI, AND HUANTONG LU

Xinjiang Oi! Company
Urumqi, Xinjiang, China

Application of reflection seismology in the oil and gas industry has been gradually shifting its role
from exploration to reservoir characterization and production. Nowadays accurate geometrical
imaging of geological structures is a prerequisite for any successful drilling. The detailed spatial
distribution of pore fluid composition and lithological and hydrological properties such as
permeability is demanded from true amplitude seismic inversion for critical reservoir production
planning and management. Because most of the production reservoirs lie 1000 to 6000 meters below
surface on land or below ocean floor offshore, penetrable seismic frequency is limited to 5-50 Hz
due to energy attenuation caused by natural dissipation and scattering, whereas the seismic wave
speed in and around the reservoirs is usually elevated due to natural compaction and diagenesis.
These factors combine to put severe natural limits on obtainable vertical and horizontal resolution of
mapping the geometrical bodies of heterogeneous reservoirs and delineating the compositional,
mechanical, and hydrological properties within. Considering these natural limits and the critical
reservoir scale of fluid flow of 0.5 meter, seismic resolution problem of about a century old will
become a central issue of reflection seismology in a decade to come. As a case study, we report our
investigation on this important problem using synthetic theoretical model, field geological model,
and 3D field seismic data. It is concluded that seismic resolution can be enhanced two to three times
through rigorous amplitude-preserving data processing.

Introduction

Over the last 30 years, one of the great challenges in exploration of the Junggar
basin of China has been to delineate accurately prospective thin-bed sand layers. The
Junggar basin was formed in the interior of the Asian plate in continental collision, which
holds a great potential for oil and gas reserves (Zhang, 1991). Over geological times, it
developed many stratigraphically complex anticlines with complicated fault systems.
Unconformities and stratigraphic traps are abundant in the basin. Oil-producing
sandstone reservoirs in Junggar basin are usually deeply buried and thinly bedded,

376

relative to the seismic wavelengths used in exploration. Seismic frequency penetrable to

a great depth is limited to 5-50 Hz,due to energy attenuation caused by dissipation and
scattering. The seismic wave speed in and around the reservoirs is usually elevated, due
to natural compaction and diagenesis. These factors combine to put severe limits on the
vertical and lateral resolution of mapping the geometrical bodies of heterogeneous
reservoirs, not to mention extracting from seismic data the reservoir properties within,
such as porosity, saturation, and permeability. Considering these natural limits and the
critical reservoir scale of fluid flow of 0.5 meter, seismic resolution in Junggar basin
remains to be a key problem.
In the past three decades, we studied many methods for both pre-stack and poststack resolution enhancement, together with constant improvement on field acquisition
techniques. We report our recent seismic resolution studies using synthetic theoretical
model, geological model, and 3D seismic data. The high resolution method used for
analysis is a method based on a neural network implementation of solving a Volterra
integral equation of the first kind, derived from a plane-wave solution of poroviscoelasticity. We find that seismic resolution of post-stack data can usually be
improved two to three times, through rigorous amplitude-preserved data processing.
However, three problems need to be further addressed: (1) how to improve deterministic
high-resolution methods to handle seismic data with relatively low signal-to-noise ratio;
(2) what is the physical reason that the improvement of seismic resolution is always
much better on synthetic data than on field data; and (3) large-scale structures are still
best interpreted on conventional seismic volume. Special techniques for interpretation of
high-resolution seismic data need to be developed. In this report, we briefly summarize
some of our successes and the typical problems encountered, to illustrate the challenges
ahead.
2

Method

To illustrate the problems in high-resolution analysis and interpretation, we use one high
resolution processing method among others, which is a neural network implementation of
solving a Volterra integral equation of the first kind, derived from a plane-wave solution
of poro-viscoelasticity. The following summarizes its two key aspects.

2.1

Physical wavelet

The effect of the viscoelastic properties of rocks on seismic wave propagation is

important in changing waveforms, including amplitude and phase. It causes the highfrequency content of seismic energy to decrease dramatically with increasing propagation
distance. Based on a plane-wave solution of poro-viscoelasticity, a Volterra integral
equation of the first kind is derived for a seismic trace modely(t) (Sun et al., 1994):
I

~ ( =4fr(r)w(r - 7, f ( r ) ,Q<r>,
+(r))d~
0

377
where r is the reflectivity function and w is the time-variant wavelet. The wavelet is
parameterized by a dominant frequency5 quality factor Q, and porosity 4, instead of
making any assumptions on the wavelet phase and the white-noise nature of reflectivity
series. The last three factors directly measure the viscoelastic effect of the rock on
changing the overall shape and duration of the propagating wavelet (Sun et al., 1992).

2.2

Deterministic Approach

The four functions, r,J; Q, and 4, are estimated using a neural network approach (Yan
and Li, 1993; Yan et al., 1996; Yan and Li, 1997), to minimize the following objective
function:

>

where is a seismic trace and T is the total record length. In this report, we use only the
estimated reflectivity function for seismic resolution analysis.

3
3. I

Interpretation and analysis

Synthetic Theoretical Models

The first exercise is to test the algorithm on synthetic models consisting of horizontal
layers of reflection coefficients. As an example, a third-party team is asked to generate a
seismic section of synthetic seismograms of a dominant frequency 25 Hz, using an ultrafine alternating sandshale model with added noise of less than 10%. Highres estimates
are obtained without knowing the model, the waveform and the synthetic method
(portions shown in Fig. 1). Two wedges of opposite polarity are much more clearly
resolved on the Highres section than on the original synthetic seismogram. One reflection
event overlying the wedges appeared on the original synthetic seismogram is clearly
resolved into a nine-layer model with alternating polarity changes. However the resolved
layers are not spikes. Highres method has thus a great resolving power but it is still an
approximation, though a better one. As the signal-to-noise ratio decreases, the Highres
result becomes less accurate, like other similar methods.
3.2

Geological Models

A realistic geological model of a production field is constructed using well controls. It

has many fine structures such as pinch outs, unconfomities, stratigraphic traps, small
faults, anticlines, and thin-bed sandshale layers. Commercial standard software is used
to generate synthetic seismograms with different dominant frequencies, using the
geological impedance model with colored noise being added. A portion of Highres
section obtained by processing a 25 Hz synthetic seismic model is shown in Fig. 2, in
comparison with a 60 Hz synthetic model. The Highres section appears quite similar to

378

the high frequency 60 Hz synthetic model. Spectral analysis also shows that the Highres
inverted data have much broader frequency spectrum than the 60 Hz synthetic model,
which is also evident on the seismic sections. The fault systems in the geological model
are quite clear on the Highres inversion profile.
3.3

3 0 Seismic Volumes

The Highres method has been applied to several 3D seismic volumes from different oil
fields with mixed results. Detailed analysis shows that some of the noise reduction
modules and the P-wave approximation module prior to Highres processing can have
severe destructive effects on the high frequency contents of the seismic signals.
However, these adverse effects are not evident on low-frequency seismic sections. They
can only be seen by applying high-pass filter on the data. Fig. 3 shows a comparison
between the two spectra before and after Highres from the F5 field in the Junggar basin.
The dominant frequency after Highres is about 55 Hz with a nearly flat spectrum from 25
to 85 Hz, whereas the dominant frequency before Highres is about 20 Hz. Frequency
components higher than 85 Hz and less than the Nyquist frequency are recoverable after
Highres but they, however, seem to contain some artifacts, depending on different data
sets. After Highres processing, faults and other fine structures are easy to interpret. As
shown in Fig. 4, a Highres profile reveals very distinct fault systems, whereas the
original seismic volume processed in 1996 shows only the diffraction patterns at the fault
positions. These fault systems among others revealed by Highres processing are
confirmed by later processing in 1999. Highres-processed data also result in a better logseismic match as shown in Fig. 5. Before Highres processing, it was predicted that an oil
layer in Jurassic sand drilled in the downdip F5 well could have a better prospective in its
updip direction (Fig. 6). Interpretation of Highres data, however, indicated that the updip
and the downdip of the one target layer appeared on low-resolution seismic data are
actually two separate sand bodies. Later drilling in the updip direction resulted in a dry
well, F12.

Conclusions

Our study shows that seismic resolution of post-stack data can be improved two to three
times, through rigorous amplitude-preserved data processing. However, one of the
common problems still remains that all high-resolution methods requires high signal-tonoise ratio of both the synthetic data and the post-stack seismic data. Secondly, we still
cannot answer quantitatively and scientifically why the improvement of seismic
resolution is always much better on synthetic data than on field data sets. The third
problem, but not the last, is that gross structural features are much easier to interpret

379

using the conventional seismic volume than the high-resolution volume. Special
techniques for interpretation of high-resolution seismic data need to be developed.

Acknowledgments
We are grateful to Mr. Yi Zongh and Mr. Jia Mingchen who provided the theoretical
model and the field synthetic models respectively. LDEO contribution number 65 19.

References
1. Zhang, G.J. (Ed.), 1991, Petroleum Geology of China: Junggar Basin, Xinjiang
Artistic Photography Publishing House.
2. Sun, Y.F., Kuo, J.T., and Teng, Y.C., 1994, Effects of porosity on seismic
attenuation, Journal of Computational Acoustics, 2, 53-69.
3. Sun, Y.F., Kuo, J.T., and Teng, Y.C., 1992, An estimation of seismic attenuation
and temporal resolution, AGU Transaction, Vol. 73, No. 43, p593.
4. Yan, L.L., and Li, C.J., 1993, Robot arm control using recurrent neural to identify
robots inverse dynamics, Transactions of the North America Manufacturing
Rsearch Institute of SME, Vol. XXI, 319-325.
5. Yan, L.L., Li, C.J., and Huang, T.Y., 1996, Function space BFGS quasi-Newton
based learning algorithm for time-varying recurrent neural networks, Journal of
Dynamic Systems, Measurement, and Control, Vol. 118, 132-137.
6. Yan, L.L., and Li, C.J., 1997, Robot learning control based on recurrent neural
network inverse model, Journal of Robot System, 199-212.

380

Figure 1. Comparison between theoretical model and Highres result.

38 1

Figure 2. Comparison of synthetic models of a production field with Highres results. Upper: 25 Hz synthetic
model; Middle: Highres results of the 25 Hz synthetic model; Lower: 60 Hz synthetic model.

382

Figure 3. Comparison between the amplitude spectra before and after Highres.

Figure 4.Highres resolving power of fault systems. Left: a profile processed in 1996; Middle: Results showing
the clear fault systems after Highres performed on the left profile. Right: the same profile processed in 1999
confirming the Highres resolved fault systems.

383

Figure 5 . Highres results compared with the synthetic seismogram generated using log data, showing very good
agreement between the two.

Figure 6 . Highres resolving power of thin reservoir beds. Left: low-resolution seismic data used to predict and
define the prospective F12 well in the updip direction of a target layer. Right: Highres results showing that the
updip and the downdip parts of the previously interpreted target layer are actually two separate sand layer
systems.

384
WAVE SCATTERING BY A RIGID INCLUSION SUBMERGED IN A CHANNEL
BOUNDED BY A SEDIMENT LAYER OVER A RIGID BOUNDARY
ANTONIOTADEU, Lufs GODINHO AND JULIETA ANTONIO
Department of Civil Engineering, University of Coimbra, Portugal
E-mail: tadeu@dec.uc.pt
The paper uses the Boundary Element Method (BEM) to compute the response of rigid inclusions
subjected to spatially sinusoidal harmonic line pressure loads when these inclusions are submerged
in a channel bounded by a sediment layer, which is over a rigid boundary. The formulation uses
2.5D Greens functions for the steady state response of the layered media (fluid and solid sediment),
defined in the frequency domain. The use of these Greens functions in the BEM formulation
eliminates the need to discretize the free horizontal interfaces, and this contributes to the efficiency
of the BEM model.
Time responses have been computed for different scenarios, allowing the identification of features
that can be used to detect the presence and location of inclusions submerged in a fluid channel. The
synthetic results are compared with those given when the fluid channel is on a rigid seabed, in order
to evaluate the importance of the mechanical properties of the channel floor on the final pressure
field.

Introduction

The cases of waveguides bounded by a rigid or sedimentary floor with submerged or

buried objects have been studied using a variety of techniques.
Lim et al. [3] studied the scattering by a buried object in underwater sediments
insonified by a bounded acoustic beam, comparing theoretical with experimental results.
They used a numerical model based on a full-wave transition matrix implementation of the
Helmoltz equation for a scattering geometry consisting of an upper fluid half-space
containing the source, a lower fluid half-space containing the elastic object, and an
interfacial fluid layer between the two half-spaces. A spectral formulation for 3D object
scattering in layered fluid media has been described by Makris [4].This formulation is
valid for receivers and sources placed sufficiently far from the object, so that the multiple
scattering between the object and the waveguide can be neglected and the scattered field
can be expressed as a linear function of the objects plane-wave scattering function.
Dawson et al. [2] used a boundary integral equation method to compute the scattering of
underwater sound from compact rigid deformations of an oceanic waveguide surface. The
method involves the use of a Greens function that allows a sound speed variation with
depth.
Most of the previous approaches neglect the influence of important factors such as the
shear rigidity of the sediments and details of the incident field. The present work is
intended to contribute for a better understanding of this type of problem by modeling the
case of a fluid waveguide bounded by an elastic layer over a rigid formation.
The problem will be solved using a Boundary Element formulation in the frequency
domain. In the present paper the horizontal waveguide free surface, the solid fluid
interface and the surface of the rigid formation below the elastic layer are not discretized
since Greens functions are formulated to account for the required boundary conditions at
these interfaces. These analytical Greens functions are derived using the solid
displacement potentials presented by the authors [6} when developing the Greens

385
functions for a harmonic (steady state) line load, with a sinusoidally varying amplitude in
the third dimension, placed in an infinite medium. The fluid potential is obtained via a
technique similar to that used for the solid potentials.

The 2.5D Green' functions for layered media

In this section, the Green's functions that allow the analytical computation of the pressure
field inside a free fluid-filled channel over an elastic layer founded on a rigid surface is
presented. When the system is subjected at point ( x o , y o ) to a spatially sinusoidal
harmonic pressure line load along the z direction, the direct incident pressure field can be
expressed as

in which A is the wave amplitude, i = A,k

being the frequency, a

af

( Imk,,

< 0), with w

the pressure wave propagation velocity in the host fluid, k z the

spatial wavenumber along the z direction, and H p ) ( ...) are second Hankel functions of
order n.
Consider a homogeneous fluid layer of infinite extent, with height h , bounded at the
top by a free surface and at the bottom by a flat homogeneous elastic layer, with thickness
h2 , placed over a rigid surface. The fields originated at the free surface, at the solid-fluid
interface and at the rigid bottom surface can be obtained as a function of solid
displacement and fluid pressure potentials. The amplitudes of these potentials are defined
so as to verify the required boundary conditions: null pressures at the free surface, the
continuity of normal displacements and stresses and null tangential stresses at the solidfluid interface, and null displacements at the rigid bottom surface.
Solid displacement potentials at the intelface between solid and fluid

with E cl = e-iyny ,

=e*"."

Solid displacement potentials at the rigid bottom

n=+N

n=+N
@ 2 = Ea

(Eb2Dn)Ed

; v,"= E a k z

v Y2= O ; v l = Ea

+Gn)

n=-N

n=-N

Ed

n=+N

n=-N

where E, =

1
2P1W2L,

, v, =

Jk,:_k,z_k,l with Im(vn ) I 0 , ~/n= gk,:_k,2_k,2

386
the velocities for P (pressure) waves and S (shear) waves, respectively,
Lam6 constants,

and pl are the

2n
p1 is the mass density of the elastic medium, k , = - n ,
LX

-e-ik,(x-xo)
d -

, E c 2 = e- i ~ n l y - h , l and

G2=e-i~nl~l

Fluid pressure potential on the solid-fluid interface

Fluid pressure potential on the free surface

and k p , =@/af.A , , B , , C , , D , , Fn , G , , Hf and If are as yet unknown

coefficients to be determined from the appropriate boundary conditions, so that the field
produced simultaneously by the source and surface terms should produce u f = 0 , u: = 0
and u," = O at y = h 2 , O ; ~ = O { ~ = O , o;z = o { ~= 0 , ofy=oYYand uk =uyf at
y = 0 , and

= 0 at y = -h

. Imposing the stated boundary conditions for each value

of n , a system of eight equations in the eight unknown constants is found.

The final expression for the pressure field in the fluid medium is obtained in the
following form,

These Green's functions were verified against Boundary Element models requiring
the full discretization of the flat boundaries (not illustrated).

BEM formulation

A Boundary Element Method (BEM) formulation used here employs Green's functions
that take into- account the presence of the horizontal interfaces between the different
media, as defined in section 2. Therefore, only the boundaries of the inclusion need to be
discretized by boundary elements. The BEM solution requires the evaluation of the
integral equations along the appropriately discretized boundary of the inclusion. In the
case of a rigid inclusion, if the boundary is discretized into N straight boundary elements,
with one nodal point in the middle of each element, the following equation can be defined:

387
with T::)

being the nodal pressures in the fluid and Gi:

=IGl:(@,Xk

,xl)dC, ,

Cl

where

is the unit outward normal for the I f h boundary segment C , . Gif)(@,xk,+)

are

the components of the Greens tensor for displacement in the fluid medium, at point

Xi

in

the normal direction, caused by a pressure load acting at the source point xk .

Time domain solutions

Signals in the time domain are obtained by applying an inverse Fourier transformation in
o,using a source with temporal variation given by a Ricker pulse. To prevent the socalled aliasing phenomenon, the separation between virtual sources must be large enough
so that their contribution arrives at times later than T . To achieve this goal, complex
frequencies with an imaginary part of the form w, = w - i q (with q =0.7Aw) are used
[ 1, 51. This technique allows the contribution of the virtual sources to be considerably
diminished. In the time domain, the effect of complex frequencies must be taken into
account by rescaling the response by a factor eq .

Numerical Application

In the examples presented, a fluid waveguide 20.0m deep and filled with water
( p f =1000Kg/m3 , af =l5OOm/s) is assumed to be illuminated by a harmonic
pressure line load placed 0.5m from its bottom. A physical model is simulated,
corresponding to a waveguide limited from below by an elastic layer 10.0m thick placed
over a rigid surface. The properties of the elastic ground correspond to those of a
sedimentary material, with a density p , =1590Kg/m3 , and allowing P and S wave
propagation velocities of a, = 1643m/s and 0, = 526m/s . Inside the waveguide, a
circular cylindrical inclusion of radius R is modeled using constant boundary elements,
whose lengths are at least 12 times smaller than the wavelength of the incident pressure
waves. A minimum of 20 elements is used. A line pressure load placed 0.5m away from
the ground illuminates these systems, and pressure variations are computed at two lines of
receivers, one of them placed vertically (line l), 0.5m away from the source, and the other
horizontally (line 2), 0.25m away from the solid ground. A schematic representation of
this geometry is presented in Figure 1.
The response is computed for a full set of frequencies ranging from 10.OHz to
1280.OHz, with an increment of 10.OHz. Time responses are then computed for a Ricker
pulse with a characteristic frequency of 400.OHz.
Figure 2a illustrates the behavior registered at the two lines of receivers when the
floor of the waveguide is modeled as a rigid surface and for an inclusion defined by
R=l .Om. The first signal arriving at the receivers corresponds to theincident and first
reflected pulses. As expected, the arrival of this pulse occurs at the beginning of the time
window for receivers placed close to the source, while it arrives at successively later times
as receivers placed further away are considered.

388
The pulse generated by the source hits the top and bottom surfaces of the waveguide,
suffering successive reflections that arrive at both lines of receivers forming regular
patterns. However, the presence of other pulses is evident, and can be clearly
distinguished in both plots. These are caused by the reflection and diffraction of pressure
waves by the rigid surface of the submerged object, which then suffer multiple reflections
on the rigid and free surfaces of the waveguide, appearing in the time responses as a
second regular pattern with a significant but lower amplitude. As an example, for lower
placed receivers in line 1, the arrival time of the first reflection of this type can be
determined as t = 26.7/1500.0 = 17.8ms.Right after the first arrivals coming from the
rigid object, another pulse is visible, resulting from a pulse reflected on the solid ground
after being first reflected by the rigid scatterer. Higher order reflections also become
visible at the two lines of receivers, forming successive patterns of lower amplitude. They
correspond to further interactions that occur between the submerged object and the
surfaces of the waveguide.

Receivers line 1

Figure 1. Geometry of the model.

Figure 2b illustrates the computed results when the fluid channel is limited below by
an elastic layer placed over a rigid surface. At the first line of receivers, the multiple
reflections occurring between the different boundaries of the domain make it much more
difficult to identify the presence of the scatterer. Although some traces of its presence can
still be found at the line 1, it becomes much easier to identify it if the response at the
second line of receivers is analysed. At that line, the receivers placed at the same side of
the inclusion as the source register the arrival of a significant pulse that corresponds to the
first reflection coming from the inclusion, allowing its position to be clearly defined.
6

Discussion

The present paper presented a numerical study of the pressure field generated by
spatially-sinusoidal harmonic pressure line sources placed in the vicinity of an inclusion
submerged in a fluid channel. A frequency domain Boundary Element Method
formulation was used, in conjunction with appropriate Greens functions. These Greens
functions take into account the presence of a sediment layer placed over a rigid surface.
Time solutions are subsequently computed by means of an inverse Fourier transformation.
The proposed implementation was found to be efficient, allowing the successful
simulation of pressure wave propagation inside fluid-filled waveguides.

389
The findings described show that the presence of an elastic layer placed over a rigid
surface produced a complex wavefield, with many visible interactions between the
different interfaces and the submerged object.
20

30

20

--E

El0

10

25

50

75

time(ms)

25

ar

50

75

50

75

time(rns)

20

30

20

-E

El0

10

25

50
time(rns)

75

hr

25
time(rns)

figure 2: Time response registered when R=l .Om: a) waveguide with rigid floor; b) waveguide with elastic
loor placed over a rigid surface.

References
1. Bouchon M., Aki K., Time-domain transient elastodynamic analysis of 3D solids by
BEM. Znt. J. Numer. Methods in Eng. 26 (1977) pp. 1709-1728.
2. Dawson, T.W., Fawcett, J.A., A boundary integral equation method for acoustic
scattering in a waveguide with nonplanar surfaces, J. Acoust. SOC. Am. 87 (1990)
pp. 11 10-1125.
3 . Lim R., Lopes J.L., Hackman R.H., Todoroff D.G., Acoustics scattering in an
inhomogeneous waveguide, J. Acoust. SOC.Am. 93 (1993) pp. 1762-1782.
4. Makris N.C., A spectral approach to 3D object scattering in layered media applied to
scattering from submerged spheres, J. Acoust. SOC.Am. 104 (1998) pp. 2105-21 13.
5. Phinney R., Theoretical calculation of the spectrum of first arrivals in layered elastic
mediums, J. Geophys. Res. 70 (1965) pp. 5 107-5123.
6. Tadeu A., Antdnio J., 2.5D Greens functions for elastodynamic problems in layered
acoustic and elastic formations, J. Computer Modeling in Eng. and Sciences 2 (2001)
pp. 477-495.

390

PE-BASED METHODS FOR TREATING FORWARD SCATTERING

FROM A ROUGH SEA-SURFACE
DAVID J. THOMSON
Defence R&D Canada-Atlantic, PO Box 1012, Dartmouth, NS, Canada B2Y 327
E-mail: david.thomson@drdc-rddc.gc.ca
GARY H. BROOKE
Anteon Corporation, 947 Fort Street, Suite 8A, Victoria, BC, Canada V8V 3K3
E-mail: gbrooke@pacijiccoast.net
We describe three procedures for incorporating the effects of a rough sea-surface
into propagation models based on one-way parabolic equation (PE) approximations. In the first procedure, the vertical extent of the computational cell adjacent
to the rough surface is allowed to vary with range. This non-uniform grid approach
can readily be incorporated into existing finite-difference P E solvers and is suitable
when the rms roughness is a small fraction of a wavelength. In the second procedure, a low-impedance air-layer backing is appended to the rough surface and the
original rough-surface scattering problem is replaced by one involving scattering
from an internal rough interface between water and air. In this instance, larger
values of rms roughness can be accommodated. Moreover, the use of an air-layer
backing can be incorporated into both finite-difference and split-step Fourier P E
solvers. Our third procedure is appropriate for treating forward-scattering from a
stochastically-rough sea-surface. In this case, the classic Kirchhoff reflection coefficient is combined with a nonlocal boundary condition formalism that can be
implemented in the context of a finite-difference P E algorithm.

Introduction

In this paper, we describe three PE-based procedures for modeling the coherent
scattering of a wavefield that has interacted with a rough sea-surface. Two of
the procedures are appropriate when the profile of the rough surface is known
deterministically. The motivation for this aspect of the paper was provided by
the particular rough-surface configuration that appeared as Test Case 1 of the
Reverberation and Scattering Workshop held during 2-6 May, 1994 in Gulfport,
MS. We compare our results for this test case, which were computed using standard
finite-difference PE algorithms, to reference solutions obtained using a boundary
integral equation method.> In contrast to these approaches, we also describe a
third PE-based method for modeling the coherent forward-scattering from a rough
surface whose roughness profile is only known stochastically.
The earliest work on incorporating rough-surface scattering from deterministic boundaries into models based on the parabolic approximation focused on the
narrow-angle, split-step Fourier algorithm. Dozier3 for example, applied conformal
mapping techniques in order to locally flatten the sea surface before each range
step. In contrast, Tappert and Nghiem-Phu4 and later Thorsos et al.5 modified the
complex P E image field to satisfy the pressure-release boundary condition on each
sloping segment of the surface. Both of these approaches have been extended to
wider-angle, split-step Fourier and higher-order Pad6 finite-difference7PEs. More
recently, Evans applied a surface-flattening transformation to handle boundary

39 1
roughness while Rosenbergg embedded the rough surface within a non-equispaced
computational grid and explicitly enforced the pressure-release boundary condition
along the surface profile by zeroing the field above it, an idea that originated with
the rotated P E concept of Collins" and later adapted by Collins et
t o allow
the sea surface t o intrude into the computational grid. Both of these approaches
were implemented in the context of a high-order PadBbased P E formalism. l2
Other methods for introducing the effects of surface roughness into PE codes
have adopted a stochastic viewpoint. For example, Moore-Head et al.I3 applied
a localized DFT near the surface t o determine the vertical wavenumber spectrum
there and then made use of ray theory concepts to map wavenumbers into equivalent
grazing angles. In this manner, attenuation could be introduced using a simple lossversus-angle curve for a given sea-surface roughness. S ~ h n e i d e r 'examined
~
some
of the numerical approximations that underlie this approach and demonstrated its
capability for modeling both forward- and backscattered fields in practical applications. Stochastic losses due t o roughness can also be accommodated through
the introduction of special boundary conditions. Collins and Chin-Bing15 used
an approximate expansion of the plane-wave reflection coefficient at low angles t o
model rough-surface effects in terms of a local boundary condition while Brooke
and Thornson'' made use of a nonlocal, impedance-type boundary condition t o
account for these coherent scattering losses-an approach due t o Ma~fie1d.l~
In this paper, we describe two techniques that can be introduced into existing P E codes with little programming effort18 when compared t o the methods
referred to above. Our first technique is applicable only to finite-difference-based
PE's and involves a straightforward adjustment of the top row of coefficients of the
corresponding tri-diagonal matrix representation of the propagator. Essentially,
a non-uniform grid is allowed at the top of the P E computational mesh in order
to track the variable-height surface as a function of range. In our more general
second technique, we simply replace the vacuum region above the sea-surface with
a low-impedance fluid, e.g., air. The large impedance mismatch between air and
water results in nearly perfect reflection of the field of a water-borne source and,
for numerical purposes, allows the rough-surface interface t o be treated as an internal boundary. This approach can be easily adapted t o any existing P E code and
solution procedure. Finally, for treating scattering from a stochastic rough surface,
we describe a third technique based on a straightforward extension of a nonlocal
boundary condition that has been incorporated into some finite-difference PE's.
The rest of the paper is organized as follows. First, we review our procedure
for setting up our finite-difference P E code and describe how it can be modified to
handle a range-dependent, variable-depth mesh spacing a t the top of the computational grid. Next, we simple append an air-layer backing t o the rough surface to
transform the given problem of scattering from a variable-depth, pressure-release
boundary t o scattering from an irregular internal interface (involving water and
air). We compare our PE-based numerical results with reference boundary integral
equation solutions generated for the deterministic rough-surface, scattering configuration introduced a t the 1994 Reverberation and Scattering Workshop. Finally,
we use a nonlocal boundary condition to determine the coherent scattering from
a stochastically-rough boundary and apply it t o propagation in a shallow-water

392
waveguide, comparing our PE results t o those generated using the wavenumber
integration code SAFAR1.l'
2

Theory

2.1 PE Background
For a range-independent acoustic medium, the 2D Helmholtz equation in rectangular coordinates (r,z ) for the pressure p can be decomposed into outgoing and
incoming components with respect to the horizontal range r.20,21
The outgoing component can be recovered from the reduced field $ = p exp(-ikor)fi by marching
the formal solution to the one-way wave equation,

$(r

+ Ar, z ) = exp { -26

-t iW-}

$(., z ) .

(1)

Here z is measured positively down, ko = W / C O is a reference wavenumber and we

have set b = k o A r . The 2D operator X is given by

where N = n(1 Za), n = C O / C is the refractive index and p, c and a denote the
density, sound speed and absorption, respectively. In the case of weakly rangedependent media, Eq. (1) can readily be modified22 t o take energy-conservation
approximately into account.
Efficient computational schemes for advancing the reduced field $ are obtained
by making approximations to the exponential propagator. In the context of finitedifferences, by combining the M-term partial fraction expansion of the argument
of the propagator23 with Cayley's second-order accurate approximation of the exponential leads to the higher-order PE

{c
M

exp 26

m=l

A,X
1 BmX

} fi
~

m=l

1+&X
1 +cmx

amX

=l+Cl + b m X

(3)

m= 1

where c$ = B, f $ibAm. To ensure an accurate representation of the propagator

for X < -1, complex-valued {AmlB,} are introduced by rotating the branch cut
of J T X . into the lower h a l f - ~ l a n eBy
. ~ ~inspection, it is evident that b, = c; in
Eq. ( 3 ) . The am's can be found numerically by solving the linear system that results
by evaluating the right-hand side a t the J zeroes, X = -(.A)-'.
Substituting
Eq. ( 3 ) into Eq. (1) yields
M

$(r + ~ r l z=) $ ( ~ , z+)

where each

$m

C $m(r +
m=l

AT,Z)

(4)

satisfies

(1

+ b m X ) $m(r + Ar, Z) = a m X $ ( r , z).

(5)

Because each gm depends on the same total field $ on the right-hand side, the
solution of Eq. (5) for m = 1,.. . , M can proceed in parallel.

393
Numerical solutions to Eq. (5) are computed on a discrete grid in T and z .
At each grid point, values of c, p and LY need to be specified. These values are
often based on interpolation from a given set of environmental profiles provided
at irregularly-spaced locations. Nominally, the grid is terminated above by a free
surface at z = 0 and below by an absorbing layer overlying a free surface at z =
z m a . We used a perfectly matched layer (PML)25 to suppress any reflections
from the base of the numerical grid. A 20-point PML was implemented via a zdependent imaginary stretching of the grid point spacing by setting the complex
distance between grid points Z k and zk+l to Az; = [l ia(zk)]Az. To reduce
discretization-induced reflections within the PML, a tapered a-profile of the form
a ( z k ) = 5 (1 - c o s ( k ~ / 2 0 )was
~ used.
In terms of the depth vector z = [ z l ,2 2 , . . . , ZJ]* where [. .
denotes transpose
and z j = j A z , j = 1 , 2 , . . . , J , we approximate X ( z ) $ ( z ) by

.IT

X(Z)$(Z)

L ( z ) $ ( z - A z ) + D ( z ) $ ( z ) + U(Z)+(Z+ A Z ) ,

(6)

where the lower ( L ) ,upper (V), and diagonal ( D ) matrix entries are given by

L(z)= YP-(Z),

V ( z )= Y P + ( Z ) ,

D ( z ) = N 2 ( 4 - 1 - L(Z) - V ( z ) ,

(7)

) ~ p+ denotes the density combinations

respectively. Here y = l / ( k ~ A zand

This finite-difference representation transforms Eq. (5) into a tri-diagonal system

of equations for each qrnthat can be solved efficiently by double recursion.
2.2

Non-unaform Depth Grid

In most applications of the PE to underwater sound propagation, the sea-surface

is taken to be flat and pressure-release, i.e., $(r,O) = 0. Moreover, both finitedifference and split-step marching algorithms are usually restricted to an equispaced
depth-mesh with spacing Az. In the present work, the rough surface z = ( ( T ) ,
defined by its vertical displacement from the plane z = 0, is taken to have zero
mean ((C) = 0). For finite-difference representations, it is possible to use a nonuniform depth-grid that adjusts with range t o the local displacement of the irregular
boundary. The simplest approach, then, one that does not require the number of
depth-grid points to change, is to modify only the grid-spacing of the cell adjacent
to the surface.
Figure ( 1 )shows a schematic of the geometric configuration for the rough-surface
scattering problem. The insert illustrates how the surface is represented in PE
calculations by a sequence of range-independent sections. We arrange for the shallowest interior mesh point of the PE grid t o lie AZ below the deepest portion of
the surface which we take to be the plane z = 0. This amounts to shifting the
rough surface upwards by the distance H = max{<}. At the range position I, the
uppermost mesh point lies along the plane z = - h A z where h 2 0. Along this 1~
cally flat surface, the pressure-release boundary condition requires $(r, - h A z ) = 0.
To derive a variable-spaced difference formulation for the PE in the vicinity of the

394
pressure-release boundary

....
.

z=-h&

z=O

z=Az

z=2Az

Figure 1. Configuration for the variable-depth grid method.

boundary there, we expand the field in a Taylor series to O(Az2) about the level
z = Az t o obtain

$(r, -hAz)
$(r,2Az)

$(r, Az) - (1 h)Az$(r, Az) f 4(1 h)2Az2$(r, Az),

$(r,A t )

where the prime denotes

$(r, Az)

+ Az$(r, Az) + ~ A z 2 $ ( rAz),

,

a/&.

Solving this system for

h-$(r, -hAz) - (h-

$(T,

(9)

Az) yields

+ h + ) $ ( ~Az)
, + h++(r, ~ A z )

(10)
AZ2
where the coefficients h+ are given by h+ = 2/(h + 2) and h- = h + / ( h + 1).
Note that at range position 11, where h + 0, these coefficients reduce to the usual
centered-difference values for the operator d 2 / d z 2 .If we consider the density to be
uniform near z = 0, then it is seen that the effects of a variable-height surface can
be accommodated simply by replacing the variable-density coefficients in Eq. (8)
along the top row with the corresponding h-weighted ones.
There are two obvious limitations of this method. First, by construction, there
are no grid points lying between min{<} < z < H and, consequently, no knowledge
of the field in this region. Second, the accuracy of the method is expected to degrade
when the displacement of the rough surface exceeds some significant fraction of an
acoustic wavelength. That is, for a given rough surface, the accuracy should improve
as the frequency is reduced.
7

2.3 Air-layer Backing

To overcome the restrictions of the variable-depth grid method, an alternative approach for treating the effect of a rough surface is proposed. It is based on the
well-known observation that the large impedance drop across the sea/air interface

395
pressurerelease boundary
I

..

Figure 2. Configuration for the air-layer backing method.

(M 2 .
results in nearly perfect, out-of-phase reflection of sound for a waterborne source (see, e.g., Chapman et a1.26). Consequently, most numerical realizations of underwater sound propagation idealize the sea-surface as a pressure-release
boundary. This observation suggests that the rough-surface scattering problem be
modified by appending a low-impedance air-layer t o the region above the boundary.
In this way, scattering by an external pressure-release boundary is replaced with
scattering by an internal fluidlfluid interface across which the usual boundary conditions on the acoustic field apply. The underwater reflection coefficient associated
with the sealair boundary should be close t o -1. The configuration associated with
this approach is illustrated in Fig. (2) for the same rough-surface scenario considered previously. With the addition of an air layer, the standard equispaced P E
computational grid and solution procedures can be applied to the modified scattering problem. It is only necessary t o modify the geoacoustic-profile information that
is supplied to the P E code. Moreover, it is worthwhile pointing out that with this
approach, values of the field between the keels of the rough surface are available a t
the resolution of the computational grid.

2.4

Nonlocal Boundary Condition

The approximate (though effective) treatment of the downgoing radiation condition

using a perfectly matched layer absorber can be replaced with a nonlocal boundary

396
condition (NLBC) that exactly transforms the original semi-infinite PE problem
t o an equivalent one in a bounded domain.28 In addition t o handling radiation
into (and scattering from) a half-space, NLBCs can also be derived t o handle the
coherent scattering losses due t o a rough sea-surface.lG
To see this, consider a thin, homogeneous layer of thickness d at the ocean
surface with N ( d f0) = NO and p(d 0) = po. Within this region, Eq. (1) satisfies

where we used the low-order approximation t o the square-root operator that leads
t o the original, low-angle PE.20 Application of the Laplace transform defined by
Q(s, z ) =
$(r, z ) exp(-st) dt to Eq. (11) yields

sr

a2q(s,
dZ2

).

+ r2(s)Q(s,
z ) = 0,

where y(s) = k o d N 2 - 1 - 2s/ik0 is the vertical wavenumber. Across the plane

z = d, continuity of impedance iwG requires

where we have expressed G(s) at z = d - 0 in terms of the surface reflection coefficient R ( s ) . In the context of the Kirchhoff approximation, the reflection coefficient
for a rough free surface with Gaussian statistics is given by1Y2l
~ ( s=
) -exp [--2a2y(s)] ,

(14)

where a is the rms roughness along z = 0. Substituting Eq. (14) into Eq. (13) gives
as d 4 0,

G ( s )= ipo tanh [ a 2 y 2 ( s ) ]M ipoa 2 y(s).

Y(S)

Using Eq. (15), rearranging and taking inverse Laplace transforms of Eq. (13) determines the nonlocal Dirichlet-to-Neumann (DtN) mapping
--

exp(ibt)

dt
az
m and b = i k o (No2 - 1). Although this NLBC is most relevant
1

where B =
t o low-angle propagation, it can be applied t o many shallow-water situations in the
far field since the steep-angle energy is usually scattered into the bottom where it
is absorbed. It is also worthwhile noting that while Eq. (16) implies that this loworder form of NLBC is mismatched when combined with the higher-order, multiterm propagators of Eq. (5), there is numerical evidence to support their use in
cases of practical interest.28 A numerical procedure for implementing an NLBC
similar t o Eq. (16) is given el~ewhere.~

391

3 Numerical Results
3.1 R & S Workshop Test Case 1

Test Case 1 of the Reverberation and Scattering Workshop involved forward scattering from a rough sea surface that was constructed t o be consistent with a 1D
Pierson-Moskowitz spectrum for a wind speed of 15 m s-'. The total surface length
was 750 m. The lossless half-space was assumed to be an isovelocity medium of
1500 m s-'. The field incident on the rough surface was specified in the form of a
single-frequency, Gaussian-tapered plane wave having a mean upward grazing angle
of 8 = -loo. At the left endpoint of the surface, the form of this incident field is
given by

p ( z ) = exp [-(z

Z O ) ~ / W ~ i k z sin

e] ,

(17)

where k = 2xf/1500, zo = 66.12 m, and w = 27.55 m. Numerical reference

solutions for the forward-scattered field, given as a function of depth for 0 < z <
200 m at the right endpoint of the surface, were obtained using an integral equatioif
method'i2 for frequencies f = 100 Hz and f = 400 Hz.
The PE results for this problem were obtained using djtPEl, a ~ r e - P E C a n ~ ~
version of a code based on Eq. (5) with M = 4, ko = 27r f /1500 s-l, Ar = 0.375 m,
AZ = 0.25 m a t f = 100 Hz and AZ = 0.125 m at f = 400 Hz. Figure (3) and
Fig. (4) show the comparisons between the reference solutions for the variable-grid
P E method while Fig. (5) and Fig. (6) show similar comparisons for the air-backed
layer P E method. Note that in these comparisons IpIJ;' versus z is shown on a linear
scale a t the range r = 0.75 km. It is evident that the variable-depth grid method
can easily handle the scattering from the rough surface at the lower frequency
but cannot reproduce the forward-scattered field structure at the higher frequency
where the peak excursion of the rough surface is of the order of a wavelength. On
the other hand, the comparisons in Fig. (5) and Fig. (6) show that the simple
expedient of backing the rough-surface boundary with a 20-m layer of air of density
0.00129 g ~ m - sound
~ , speed 340 m s-' and absorption 1 dB X-' yields accurate
forward-scattered P E solutions at both frequencies. To fully gauge the influence of
this specific rough surface on the forward-scattered energy, full-field intensity plots
versus range and depth are shown for f = 400 Hz in Fig. (7) for a flat surface and
in Fig. (8) for the rough surface. For the flat surface, the Gaussian beam undergoes
specular reflection that exhibits interference between the upgoing and downgoing
components of the beam. In contrast, the main effect of the rough surface is t o
scatter the energy into non-specular directions.

3.2 Modified NORDA Test Case 3

When the characteristics of the rough-surface profile are only known statistically,
the scattering losses in the coherent component of the propagating fields can be determined by incorporating a nonlocal boundary condition into the parabolic equation formalism. To examine this feature, we consider the shallow-water Pekeris
waveguide that was introduced at a 1981 P E Workshop as Test Case 3.30The region 0 < z < 100 m contains a homogeneous fluid with sound speed 1500 m s-l

398
0

20

_I

40 60

- 80I

flM)-

120140160

180-

m-

Figure 3. Variable-depth grid PE versus reference solution for f = 100 Hz.

60

la/
180-

-2-

__---

2M)-=%?-.0
0.1

---___
-- - --

0.2

0.3

1-

0.4

0.5

0.6

0.7

0.8

Ampliude. Ipl

Figure 4. Variable-depth grid PE versus reference solution for f = 400 Hz.

and density 1 g cm-3 that overlays a fluid half-space with constant sound speed,
density and absorption of 1590 m sP1, 1.2 g cm-3 and 0.5 dB X - l , respectively.
The 250-Hz source frequency gives rise t o 11 propagating modes in this waveguide.
Both the source and the receiver are located at mid-depth (50 m) in the upper
layer. We compute predictions of transmission loss for both a flat sea surface and
one characterized by a zero-mean, Gaussian sea surface with an rms roughness of
4 m. For this range-independent example, we compare our results to reference
results obtained using SAFARI," a well-known spectral model that includes the
capability for modelling rough b o u n d a r i e ~ . ~ ~
The transmission loss (-10 log,, lpI2) results of applying PECan27 with nonlocal
boundary conditions to the Pekeris waveguide are displayed in Fig. (9) for the
flat-surface case and in Fig. (10) for the rough-surface case. For both PECan
calculations, we used M = 4, AZ = 0.125 m, A r = 5 m and applied a perfectly
matched layer within 117.5 < z < 120 m. The flat-surface PECan results shown
in Fig. (9) are indistinguishable from the SAFARI results. For the rough-surface

399

Figure 5. PE with air-layer backing versus reference solution for f = 100 Hz.

cn

I-/

2--

:I

0:2

0:s

014

0:s

0:6

o:,

0.8

Arrpldude. Ipl

Figure 6. PE with air-layer backing versus reference solution for f = 400 Hz.

problem, PECan used the NLBC of Eq. (16) applied along z = 0 m. The agreement
with the SAFARI prediction is observed t o be good, especially beyond a range of
5 km where the low-angle NLBC is expected to be'most accurate. At this frequency,
the 4-m rms surface roughness clearly has a considerable effect on the propagating
waves. Due to the stripping of the higher-order modes, the transmission loss curves
are considerably smoother and the overall losses have increased by several dB.
4

Conclusions

Two methods for incorporating a deterministic rough free surface into parabolic
equation models were described and applied t o the particular rough-surface configuration that was offered for numerical consideration at the Reverberation and
Scattering Workshop, held during 2-6 May 1994 in Gulfport, MS. The first method,
suitable only for finite-difference-based P E models, introduced a variable-depth grid
element adjacent t o the rough surface t o track its displacement as a function of

400

Figure 7. Intensity level versus range and depth for the flat-surface reflection at f = 400 Hz.

Figure 8. Intensity level versus range and depth for the rough-surface reflection at f = 400 Hz.

range. This method is limited to rough surfaces whose height variation is sufficiently small relative to an acoustic wavelength. To overcome this restriction, a
second method was introduced that involved appending an air-layer backing to the
rough surface. In this case, scattering from an external, pressure-release boundary
is replaced with scattering from an internal, water/air interface. This approach is
applicable to both finite-difference and split-step Fourier solution procedures. For
the air-layer method, the finite-difference solutions at both frequencies were in excellent agreement with reference solutions based on integral equation techniques.
For treating forward scattering from a stochastic rough-surface, a third method was
described that is based on the concept of evaluating a nonlocal boundary condition
applied along z = 0 that incorporates the scattering due to the rough surface itself.
Within the context of Kirchhoff scattering, the formal integral boundary condition
was approximated and implemented for propagation in a shallow-water waveguide
with a given rms surface roughness. The use of such a nonlocal boundary condition
with a PE was shown to agree well with solutions generated using a well-known

40 1
40

50

T60

rn

p.

f 70
2

80

90

Range qinl

Figure 9. Transmission loss comparisons at 250 Hz for the flat-surface ( u = 0 m) Pekeris waveguide. The source and receiver depths are 50 m.

901

I
2

Range Urn]

10

Figure 10. Transmission loss comparisons at 250 Hz for the rough-surface ( a = 4 m) Pekeris
waveguide. The source and receiver depths are 50 m.

benchmark code for layered media.

Acknowledgments
We are grateful to Eric Thorsos for providing the 100-Hz reference solution for the
rough-surface scattering problem in response to our request prior to the R & S
Workshop. This allowed us to evaluate the dependence of the variable-grid method
on source wavelength, and inspired us to examine the air-backed layer approach.

References
1. E.I. Thorsos, J. Acoust. SOC.Am. 88, 335 (1990)
2. E.I. Thorsos, Boundary integral equation solution for test case I, presented

402

at the Reverberation and Scattering Workshop, 2-6 May 1994, Gulfport, MS.
3. L.B. Dozier, J. Acoust. SOC.Am. 75,1415 (1984)
4. F.D. Tappert and L. Nghiem-Phu, J. Acoust. SOC. Am. Suppl. 1 77, SlOl
(1985)
5. E.I. Thorsos et al., J. Acoust. SOC.Am. Suppl. 1 84,S220 (1988)
6. K.B. Smith and F.D. Tappert, The University of Miami Parabolic Equation Model, Marine Physical Laboratory, Scripps Institution of Oceanography,
MPL Tech. Memo. 432, May 1993.
7. G.V. Norton et al., J. Acoust. SOC.Am. 97,2173 (1995)
8. R.B. Evans, J. Acoust. SOC.Am. 104,2167 (1998)
9. A.P. Rosenberg, J. Acoust. SOC.Am. 105,144 (1999)
10. M.D. Collins, J. Acoust. SOC.Am. 87,1035 (1990)
11. M.D. Collins et al., J. Acoust. SOC.Am. 97,2767 (1995)
12. M.D. Collins, J. Acoust. SOC.Am. 93,1736 (1993)
13. M.E. Moore-Head et al., J. Acoust. SOC.Am. 86,247 (1989)
14. H.G. Schneider, J. Acoust. SOC.Am. 93,770 (1993)
15. M.D. Collins and S.A. Chin-Bing, J. Acoust. SOC.Am. 87,1104 1990)
16. G.H. Brooke and D.J. Thomson, Wave Motion 31,117 (2000)
17. M.E. Mayfield, Nonreflective boundary conditions for Schroedinge s equation,
Ph.D thesis, Univ. of Rhode Island, Kingston, RI, 1989.
18. D.J. Thomson et al., PE Approximations for Scattering from a Rough Surface, Defence Research Establishment Pacific, Victoria, BC, Canada, Tech.
Memo. 95-21, March (1995).
19. H. Schmidt, SAFARI Seimo-Acoustic Fast field Algorithm for RangeIndependent environments, SACLANT Undersea Research Centre, San Bartolomeo, Italy, Report SR-113, September (1988).
20. F.D. Tappert, The parabolic approximation method, in Wave Propagation
and Underwater Acoustics, eds. J.B. Keller and J.S. Papadakis (Springer, New
York, 1977), pp. 224-287.
21. F.B. Jensen et al., Computational Ocean Acoustics (AIP Press, New York,
1994).
22. M.D. Collins and E.K. Westwood, J. Acoust. SOC.Am. 89,1068 (1991)
23. A. Bamberger et al., SIAM J. Appl. Math. 48,129 (1988)
24. F.A. Milinazzo et al., J. Acoust. SOC.Am. 101,760 (1997)
25. D. Yevick and D.J. Thomson, J. Acoust. SOC.Am. 107,1226 (2000)
26. D.M.F. Chapman et al., J. Acoust. SOC.Am. 91,1904 (1992)
27. G.H. Brooke et al., J. Comput. Acoust. 9,69 (2001)
28. J.S. Papadakis, J. Comput. Acoust. 2 , 83 (1994)
29. D.J. Thomson and M.E. Mayfield, J. Comput. Acoust. 2, 113 (1994)
30. J.A. Davis et al., Cavanagh, NORDA Parabolic Equation Workshop, 31
March-3 April, 1981, Naval Ocean Research and Development Activity, NSTL
Station, MS, Tech. Note 143, September (1982).
31. W.A. Kuperman, J. Acoust. SOC.Am. 58,365 (1975)

403

Simulated tomographic geoacoustic inversion

A. Tolstoy
ATolstoy Sciences
8610 Battailles Court
Annandale VA 22003
phone: (703) 978-5801 fax: (703) 978-1319 email: atolstoy@ieee.org
August 11 2003
Abstract

It is well known that geoacoustic inversion in shallow water is difficult - even for
a single, range-independent slice between a source and receiver array. This paper will
discuss simulated results which show the application of a tomographic inversion method
over a number of quite variable areas (each of which involves over 100 range-dependent
slices) to determine water depth. The final tomographic (matrix) inversions giving the
3-D range dependence are extraordinarily simple, fast, and accurate - more so than for
the inversion of an individual path. However, the overall quality of the results depends
on the preliminary inversion accuracies (individual slices) which in turn depend on array
parameters (how many phones, how densely packed, where distributed), source parameters (frequencies and distribution), and on the geoacoustic parameters to be determined
(here we consider only water depth). Additionally, adiabatic normal modes have been
assumed for the derivation of the method but are not required for its application. The
implications for this approach include an ability to rapidly and non-invasively estimate
bottom parameters over a 3-D area of interest.

Introduction

The history of deep ocean tomography, for the estimation of sound-speed profiles, is nearly 25 years
old (Munk and Wunsch, '79). It was been demonstrated experimentally to work quite well (Behringer
et al., '82), continues to evolve (Chiu, '92), and has become an accepted tool for oceanographers
studying ocean dynamics (Cornuelle et al., '85; Desaubies et al., '90; Munk et al., '95).
More recently, shallow water geoacoustic estimation has become of interest. This interest has included single path range-independent inversions (Tolstoy et al., '98 and the papers of that JCA
special issue) as well as single path range-dependent inversions (Chapman et al., special issues to
appear in JOE '03). However, there have been very few efforts to provide an overall, volumetric
image of geoacoustic parameters - nearly all have involved at best a collection of non-inter-related
single paths (Badiey et al., '94).
Perhaps the only technique currently used for the tomographic inversion of geoacoustic parameters
and producing a volumetric image of those overlapping parameters, is the one to be discussed here.

404
This method has been under development over the past few years (Tolstoy, '00; '01) and is quite
different from the more traditional time-domain approach begun by Munk et al.

2 Method
The method is based on the Matched Field Processing (MFP, see Tolstoy, '93) tomographic approach
originally developed for deep ocean tomography (Tolstoy, '92; Tolstoy et al., '91), but it is now
applied for the estimation of shallow water bottom properties. It was developed assuming adiabatic
normal modes (Jensen et al., '94)- but - this assumption is not necessary for successful inversion,
as we shall soon see. Advantages of the method include:
It is independent of the unknown parameter (water depth, bottom sound-speeds, densities,
attenuations, etc.);
it is independent of the number of parameters per grid cell (region discretization);
it is independent of the inter-relationships between parameters (such as sound-speeds versus
densities);
it is extremely fast (after individual path inversions have been performed);
more resources (more sources, more arrays) result in improved inversion resolution.
Of course, there are also disadvantages to the method as well including:
Multiple receiving arrays (preferably vertical) are required;
Multiple low (but multi-frequency) sources are required;
Preliminary path inversions (sources to arrays) are necessary and must be high quality;
Parameters at the source and array locations are assumed known or pre-determined.

A key component of the method is the gridding of the region of interest into cells. These cells do
not need to be uniform although all analysis to date has assumed such. We note that the more cells
a region is divided into, the more resources (sources and arrays) will be necessary - or - one would
have fewer paths per cell and a degraded (larger) condition number A (the smaller value the better;
see Tolstoy, '02). In Fig. 1 we see a typical grid (10x10) and resource placement (36 sources, 3
arrays). This diagram will be used for most of the results to follow. We note that there are at most
22 paths in a cell, at least 5 per cell, and A x lo3.

A Variety of Environments

For the results to follow we shall consider three (3) types of shallow water environments: continental
shelf, Gaussian feature (hill), and a combination (slope plus Gaussian depression). We shall be
inverting for the estimation of water depth which will range from 100 to 250 m (for the combination
environment) or from 100 to 500 m (for the slope alone or a Gaussian hill alone). The path ranges
will vary from a bit over 1.0 km to nearly 14.0 km.

405
Distribution of sources and arrays

m
3

I?* *

00

10

Figure 1: Typical gridding (10x10) of region of interest with 36 sources (indicated by * I s ) distributed around
the region perimeter and 3 arrays (indicated by the circles) in the region interior. The array placements
were optimized as in Tolstoy, 92.

3.1

Combination

In Fig. 2 we see the simulated shallow water environment representative of a Gaussian dip at the
lower left with an upslope at the upper right. In Fig. 2 left we have the true continuous water
depth variation while in Fig. 2 right we see the gridded discrete water depth representation which we
will attempt to duplicate via the tomographic inversion using the gridding and resource distribution
of Fig. 1.
If we assume the approximate water depth values (for the true grid) at the source and array grid
squares, i.e., 39 cells known, plus an average intermediate water depth per source-array path (these
we can compute exactly since the environment is simulated), and we then perform the associated
matrix inversion for the method (61 unknowns), we arrive at the water depths shown in Fig. 3
where the true values (left figure) are seen to be quite close to those of the estimated values (right
figure). The rms error is seen to be 3.3 m with a maximum error of 8.6 m. We see that the key
features (depths) of the region have been well estimated.
If we were to add a fourth array at x=2.6 km, y = 4.6 km, then the rms error would reduce to 2.9
m and A would be approximately 200. Thus, the results would improve to within an rms error of
less than 2% (out of 150m).
In Fig. 4 we see the true continuous environment (upper left), a selected source-array path
(upper right, S8, R3), the depths for the selected path (middle) with the solid dots indicating the
continuously varying water depth (true) versus dashed line for the assumed 3 segmented water depth
for the path (nearly accurate depths at the source and array with an average depth D in between.
Finally, in the bottom of the figure we see the MFP results at various indicated frequencies (21-29
Hz) for a variety of possible average depths for the mid-path water depth. The true field seen at
the array was computed at each frequency using RAMGEO (Collins, 94), and the MFP comparison

406
Topography
10
24

9
8

28

m
6
IM

4
1M

3
2

14

1 8

x (W

Wgid

Figure 2: Combination region (Gaussian dip plus continental slope). The true continuous environment
(left) and the discrete true environment (right) to be estimated.

TRUE (T)

max T-e

estimated (e)

m, min T-e
r m s T-e = 3 . 3 m

= 8.63

= -8.34

Figure 3: Combination region (Gaussian dip plus continental slope). The true environment (left) and the
estimated environment (right) are quite close with errors as indicated. The gridding and resource distribution
are as in Fig. 1.

407
field was also computed using RAMGEO but for only a 3 segment bottom (true at source and
array plus intermediate average D). We note that a multi-frequency sum of MFP results is required
to achieve a good estimate of D. We find that multi-frequencies are critical to the inversion of
nearly all bottom parameters (Tolstoy et al., 98; Chapman et al., 03) - for each source-to-array
path. Clearly, a difficult issue for the method is the successful preliminary inversion necessary for
each individual path. For some paths it is quite difficult to estimate D as shown in Fig. 5.

3.2

Continental Shelf

For this next environment we again assume nearly true depth values at the sources and arrays
and invert using the correct D (average) depth values per path with the resource distribution as in
Fig. 1. The true environment is a slope with the bottom depths ranging from a deepest depth of
500 m at the lower left corner of the region (left Fig. 6) to a shallowest depth of 100 m at the upper
right corner. Inverting for the 61 unknown depths we find the estimates shown in the right portion
of Fig. 6. We see that again we have captured the key elements of the environment. The rms error
is 10.8 m (2.7 % of 400 m). If we again add the fourth array we reduce this to 6.7 m.
In Fig. 7 we see the components of an individual path for this region (as in Figs. 4 and 5 ) . However,
at the bottom of the figure we see the 2-dimensional MFP search results at a single frequency (25
Hz). That is, we have had to search over an average bottom slope I9 as well as over D . A search over
only D will not find an accurate value even with frequency averaging. We see in this figure that there
is an area of values for D and I9 (correct D = 339m) for this frequency. Usually, multi-frequencies
will again be necessary. If we return to the combination scenario of the previous example, we find
that for the difficult paths (such as S1-R3 of Fig. 5), we can get a good D estimate by including I9
in the MFP search (not shown). Thus, we find that the preliminary individual path inversions to
find D are likely to involve searches over more than D alone and at a number frequencies.

3.3

Gaussian hills

In Fig. 8 we see single Gaussian hills (rising from 500 m to 250 m) placed in upper corners of the
10x10 region. The components represent a true hill (left) versus a tomographically estimated hill
(right). The top elements assume a hill placed in the left corner while the lower portions assume
a hill of the right region corner. The upper figures actually give the best inversion result (single
hills were also placed in the lower corners) while the lower figures give the worst results. Clearly
both sample environments, while sensitive to placement of feature versus resources, are respectably
reproduced.

Noise

Finally, results were generated adding increasing levels of uniform random noise to the average D
values. We found, as expected, that inversion results degraded with increasing noise. The results
were still quite acceptable for the combination environment with 1%noise (rms error goes from 3.3
m to 5.1 m) while 5% noise result in still recognizable features but with large rms topograpy errors
(23.3m). Thus, the inversion process is stable, at least up to a point.

408

Towaraohv
10

*
#

*
6
- *
8 5
* *
7

tI

I
I

*
*

58-r3
1.1

1
0.9
0.8
0.7

0.6
0.5
0.4

0.3

0.2
0.1
OlbO.

120

140

160

180

200

220

240

D-a.Z

Figure 4: Components of an individual path for the Combination region. In the upper left portion of
the figure we see the true environment. In the upper right we see the indicated path t o be analyzed. In
the middle of the figure we see the bottom topography for the selected path with the dots indicating the
continuous true bottom and the dashed line showing the 3 segmented bottom with the average D to be
estimated via MFP. In the bottom of the figure we see the MFP efforts to find D as a function of frequency
and note that the single frequencies suffer from sidelobes (alternate D values) while the incoherent sum does
peak near the correct D (dashed vertical line at 198.4 m).

409
Distribution of sources and arrays
10

9
8
7

5.5

4
3
2

I
0

S1, R 3

loo

Sl-R3
1.1
1
0.9

-g

0.8

0.7
0.6

0.5
0.4
0.3

0.2
0.1
0
100

1 2 0

140

160

1 8 0

200

220

240

D-a"e

Figure 5: Same a s Fig. 4 but for a different path. In the bottom of the figure we see the MFP efforts
to find D as a function of frequency and note that the single frequencies suffer from sidelobes (alternate D
values) while the incoherent sum peaks with a low value near D (220.3 m).

410

TRUE
.

estimated (e)

(T)

max T-e

19.90 m, min T-e

rms T-e = 10.8 m

= -31.73

Figure 6: Continental shelf region. The "true" environment (left) and the estimated environment (right)
are quite close with errors as indicated. The gridding and resource distribution are as in Fig. 1.

Conclusions

We find that the major MFP issue for this method involves the preliminary path inversions which
are subsequently input to the tomographic method. These path estimates are sensitive to the array
characteristics as well as to frequency selections. Other such issues (not discussed here) relate to the
parameters of interest themselves (including their inter-relationships) and propagation modeling.
We conclude that the tomographic inversion - assuming high quality values of D for each path input
(preliminary inversion results) - shows excellent feature reproduction for a variety of environments
(including a continental slope, Gaussian hills, and a combination slope and Gaussian dip). The
results are sensitive to resource placement (both via the condition number A and via the array
location versus feature location) and to noise (modeled as a random source of error in D).We note
that adiabatic normal modes while used to derive the theory of the method are not required for
successful inversion results. All propagation modeling here was done using a P E model (RAMGEO)
allowing for mode coupling and high angle propagation.

Acknowledgements

The author would like to thank ONR for its continued support for this effort.

41 1
Distributionolsourcesand arrays

** *

-x

* * * * * *

MV
1
09

08

07

06

05

04

03

02

0 1

25

Hz

Figure 7: Components of individual path for the Continental shelf region. In the upper left portion of
the figure we see the true environment. In the upper right we see the indicated path to be analyzed. In
the middle of the figure we see the bottom topography for the selected path with the dots indicating the
continuous true bottom and the dashed line showing the 3 segmented bottom with the average D to be
estimated via MFP. In the bottom of the figure we see the MFP efforts to find D and 6 (at one frequency).
A search over only D will not find an accurate value even with frequency averaging. We see also that there
is an area (slash) of high MFP values for D and 6.

412

Gaussian Hills
estimated ( e )

TRUE ( T )

60

7
WI

310

3
xa

1
o

max T - e

m, m i n T - e
= 4.2 m

= -11.6

r m s T-e

= 63.2

m, min T - e

= -38.7

= 13.20

10

9
8
7

-6
5 5
>I

3
2
1

max T - e

r m s T-e

= 20.5

Figure 8: Tomographic results for selected Gaussian hills. Each left figure represents a truehill placed in
an upper region quadrant while the right figures represent the corresponding tomographic inversion estimates
for the hill.

413

7 References
1. Badiey, M., I. Jaya, and A. H.-D. Cheng, Shallow-water acoustic/geoacoustic experiments
at the New Jersey Atlantic Generating Station site, J. Acoustic. SOC. Am. 96, 3593-3604
(1994).
2. Behringer, D., T. Birdsall, M. Brown, B. Cornuelle, R. Heinmiller, R. Knox, K. Metzer, W.
Munk, J. Spiesberger, R. Spindel, D. Webb, P. Worcester, and C. Wunsch, A demonstration
of ocean acoustic tomography, Nature 299, 121-125 (1982).
3. Chiu, C.S. (1992), Acoustical Oceanography: Comparison of Ocean Tomography Methods,
J. Acoust. SOC.Am. 92, 2324.
4. Chapman, R.N., et al. (2003), J. Ocean. Eng..
5. Collins, M.D. (1994), Generalization of the split-step Pade solution, J. Acoust. SOC. Am.
96, 382-385.
6. Cornuelle, B., C. Wunsch, D. Behringer, T. Birdsall, M. Brown, R. Heinmiller, R. Knox, K.
Metzger, W. Munk, J. Spiesberger, R. Spindel, D. Webb, and P. Worcester, Tomographic
maps of the ocean mesoscale. Part 1: Pure acoustics, J. Phys. Ocean. 15, 133-152 (1985).
7. Desaubies, Y., A. Tarantola, and J. Zinn-Justin (editors), Les Houches 1988 Session L, oceanographic and geophysical tomography (Elsevier Pub, 1990).

8. F.B. Jensen, W.A. Kuperman, M.B. Porter and H. Schmidt (1994), Computational Ocean
Acoustics (Am. Inst. Physics, New York).
9. Munk, W.H. and C. Wunsch, Ocean acoustic tomography: a scheme for large scale monitoring, Deep Sea Res. 26A, 123-161 (1979).
10. Munk, W.H., P. Worcester, and C. Wunsch, Ocean acoustic tomography (Cambridge University
Press, 1995).
11. A. Tolstoy Tomographic inversion on multiple receivers/arrays from multiple sources for the
estimation of shallow water bottom properties in Inverse Problems in Underwater Acoustics
(Taroudakis and Makrakis, ed., Springer), pgs. 37-46, 01.
12. A. Tolstoy Tomographic inversion for geoacoustic parameters in shallow water J. Computat.
Acoust., Vol. (8) No. 2, pg. 285-293, 2000.
13. A. Tolstoy (1993), Matched Field Processing in Underwater Acoustics, World Scientific Publishing.
14. A. Tolstoy, Linearization of the matched field processing approach to acoustic tomography,
J. Acoust. SOC.Am. 91(2) p.781-787, 1992.
15. A. Tolstoy, N.R. Chapman, and G. Brooke (1998), Workshop97: Benchmarking for geoacoustic inversion in shallow water, J. Computat. Acoust., 6 (1&2), pp.1-28,
16. A. Tolstoy, 0. Diachok, and L.N. Frazer, Acoustic tomography via matched field processing,
J. Acoust. SOC.Am. 89 (3) p.1119-1127, 1991.

414

IMPACTS OF BUBBLE PLUME STRUCTURES ON

HIGH-FREQUENCY SONAR PERFORMANCE
NEAR THE OCEAN SURFACE
MARK V. TREVORROW
Defence Research & Development Canada -Atlantic, PO Box 1012,
Dartmouth, Nova Scotia, B2Y 327 Canada
E-mail: Mark.Trevorrow@drdc-rddc.gc.ca
Open-ocean observations have shown the frequent occurrence of bubbles organized into vertical,
plume-like structures. A review of recent inverted echo-sounder observation of bubble plume
densities and depth scales is presented. Using standard high-frequency (HF)models for bubble
scattering and absorption, along with recent acoustic resonator measurements of bubble size-spectra,
simple sonar performance models can be developed. These models show that bubble plume
structures induce both significant spatial variations in the reverberation level and path-integrated
extinction loss anomalies relative to predictions from uniform bubble layer models. Results are
presented modelling the impacts of bubble plume structures on the performance of a nearhorizontally oriented HF sonar close to the ocean surface.

Introduction

High-resolution, high-frequency (HF, > 1OkHz) sonars are useful for applications such as
torpedo and mine detection, obstacle avoidance, harbour surveillance, or fisheries surveys.
However, in the near-surface ocean the performance of HF sonars can be strongly limited
by acoustic scattering from aggregations of air bubbles. Breaking-wave induced bubbles
typically begin to appear above 6 - 10 knots wind speed [ 1,9,10], and thus are a common
occurrence in the ocean. In the immediate vicinity of a breaking wave the bubble densities
can be very high, with total void fraction up to roughly 1% [9]. However once injected
into the near-surface ocean these bubbles rise to the surface, dissolve, coalesce and are
dispersed by surface waves, turbulence, Langmuir circulation, and other circulation
phenomena. The bubble density and depth of penetration generally increase as the wind
speed increases, however there are significant spatial variations due to the patchiness of
wave breaking and near-surface circulations.
Most previous HF sonar studies in the near-surface zone, while acknowledging the
cloud- or plume-like structure of near-surface bubble clouds, have usually chosen to
describe the bubble effects in a time-averaged, horizontally-uniform sense [2,12]. One
exception is a recent HF modeling effort by Norton et al. [ l l ] which confirmed the
importance of the bubble plume structures on near-surface propagation. Similarly, both
McDonald [6] and Henyey [3] utilized bubble plume structures in their modeling of lowfrequency (< 2 kHz)surface reverberation.
The focus of this work is to summarize data on bubble spatial structures and then to
assess the significance of these bubble structures for the operation of forward-looking HF
sonars in the near-surface ocean. This study draws upon data from an 81-day deployment
of a 200-kHz IES in the open Pacific ocean during the spring of 1996. In the open ocean,
autonomous Inverted Echo-Sounder (IES) devices have been used previously for
investigation of bubble clouds [ 13,151. IES systems have the advantage of remotely
sensing bubble properties while positioned safely away from the destructive motions of

415
the ocean surface. The bubble plume measurements and acoustic reverberation model
summarized herein are presented in greater detail in a recent paper by the author [ 141.

2
2. I

Summary of Oceanic Bubble Plume Observations

Instrumentation and Location

A self-contained 200 kHz IES was deployed and operated continuously from March 01
until May 20, 1996. The mooring location was Ocean Station Papa (OSP, N50" W145'),
situated approximately 1400 km west of Vancouver Island in the NE sub-arctic Pacific.
The IES was deployed at the top of a sub-surface mooring at a nominal depth of 43 m in a
total water depth of roughly 4100 m.
This 200-Hz IES was a prototype instrument designed for long duration, selfcontained deployments. The instrument utilized an externally mounted 4.6" conical-beam
transducer. The IES transmitted a 0.5-ms pulse at 1 s intervals, then computed a 3-ping
average to reduce overall storage requirements. Backscatter amplitude was digitized with
8-bit resolution (48 dB dynamic range) then averaged into 21.5-cm depth bins spanning
10 to 53 m (upward) range. No time-varying gain was used, which turned out to be
beneficial for this geometry where the bubble scattering strength increased with range.
Prior to deployment the instrument was acoustically calibrated. Then, using the standard
sonar equation the raw IES amplitude was converted to volumetric backscatter crosssection, s,(z) in m-', where z is depth, or the deciBel equivalent S,(Z). The receiver gain
was configured as a compromise between resolving both meso-zooplankton and nearsurface bubbles, thus its dynamic range spanned s,(~)
from -85 to -30 dB (re 1 me'). The
relatively high gain and limited dynamic range of the IES receiver imposed an upper limit
to bubble measurements. The A/D conversion clipping level of 255 counts corresponded
to S, near -28 dB at the ocean surface. Thus, correct measurements of the ocean surface
echo and denser portions of the near-surface bubble plumes were impossible. Also, near
the maximum S,(Z) the extinction cross-section was near 0.04 m-' (bubble loss = 0.17
dB.m-'), making extinction loss corrections negligible for the short vertical paths.
This 81-day time-series of 200-kHz IES data covered a range of conditions spanning
the late winter and spring of 1996. Data from Ocean Weather Buoy 46004, located
roughly 650 km to the east, showed wind speeds ranging up to 20 m.s-' and significant
wave heights from 1 to 8 m. Three-month average values of wind speed and significant
wave height were 7.5 mX' and 2.8 m, respectively. It should be noted that this average
wind speed significantly exceeds the generally accepted threshold for breaking wave
activity of roughly 3 m.i' [ 1,101, implying that on average white-capping and sub-surface
bubbles should be present. Unfortunately, the distance between the weather buoy and OSP
made detailed correlations of bubble penetration and surface weather impossible.

2.2

Bubble Plume Observations

The IES data consists of a continuous time-series of 3-s averaged volume scatter strength
vs. range profiles extending upward toward the ocean surface. The first stage of
processing involves extracting the range to the ocean surface within each profile and using
this as reference. This allows averaging of data records and display in depth vs. time
coordinates, as shown in Figure 1. This 2-hour echogram is typical of the stronger bubble
plume penetration events observed with this IES, but is not extreme. The echogram shows

416
a large number of individual bubble plumes extending to depths of 5 to 20 m, separated by
regions of lower backscatter (herein called null zones) which sometimes extend close to
the surface. The contrast in backscatter level between the plume and null zone profiles
reaches up to 40 dB. Detailed examination of the profiles within a given plume or null
zone indicated that volume scatter strength generally decreased linearly with depth,
equivalent to an exponential decrease in backscatter cross-section. Typically 10 to 40
plumes per hour were observed, with duration in the beam of between 20 to 90 s.
However, without a specific measurement of surface current (none was available on this
moorine). little more can be concluded from the plume durations.

Time (minutes)
Figure 1. Volume scatter strength (dB re 1 m-') vs. depth and time for a 2-hour period starting 0600UT March
7'h, 1996 at Ocean Station Papa.

To quantify the bubble penetration depth, a threshold of S, = -50 dB (re 1 m-')was

used to define the bottom of the bubble layer. This threshold effectively separates bubble
and zooplankton scattering contributions, and at scattering levels below -50 dB the
acoustic extinction and total void fraction (< 2 x lo-') are negligible. Assuming the linear
decrease in bubble scatter strength to hold in general, the bubble penetration depth,
extrapolated surface backscatter cross-section, s,(O), and e-folding scale were then
calculated for each 3-s data record in the entire data set. An example distribution of
bubble penetration depths is as shown in Figure 2. This data was taken from a 3-day
period where the bubble penetration depth statistics were reasonably stationary and the
bubble conditions close to the 3-month average. The overall bubble penetration depth
distribution is positively skewed (skewness = 0.73) and is consistent (reduced Chi-squared
= 0.18) with a log-normal distribution.
The bubble depth parameters also showed an interesting inverse relation between the
bubble densities and e-folding depth scales, as shown by Figure 3. In this case the
extrapolated s,(O) is taken as a proxy for surface bubble density (the two are proportional).

417
The extrapolated s,(O) values are truncated at 10 m-', as this corresponds to a void fraction
near 2 x lo", comparable to densities found under breaking waves. The figure shows that
deeper plumes tend to be associated with low surface bubble densities, whereas the higher
bubble densities have much shorter e-folding scales. Extreme bubble densities (s,(O) > 1.O
m-') were rare, occurring in less than 1% of this data set, and were generally associated
with e-folding depth scales c 0.6 m. The overall mean s,(O) and e-folding scales for this
data period were 0.23 m-l and 0.60 m, respectively.
0'41
OSP: YD 84.5 to 87.5,28645 x 3s records
J

1
.

0.3-

c
._

UJ

c
0
0
1 . -

E 0.2R

ChiWDoF = 0.18

9
a
0.1

0.0

10

Bubble Penetration Depth (m)

Figure 2. PDF of bubble penetration depths, with best-tit log-normal distribution, for OSP IES data covering a
3-day period starting 1200UT March 231d, 1996. Distribution has mean = 2.96 m, standard deviation = 1.41 m,
and maximum value of 10.55 m.

T
4

e-folding depth scale

Figure 3. Relation between extrapolated surface backscatter and e-folding depth scale for OSP JES bubble data.
Same data period as Fig. 2.

418

3
3.I

High-Frequency Bubble Models and Sonar Performance

Bubble Scattering and Extinction Models

To proceed from the 200-kHz IES backscatter measurements to more general sonar
performance predictions, a form for the bubble size distribution must be determined.
Oceanic micro-bubbles have a density vs. size spectrum, n(a) where a is radius
(normalized to 1 pm radius bin), typically decreasing rapidly from a peak near 20 pm
radius [4,7,9,15]. After injection by breaking waves, the bubble spectral densities are
dependent on the interplay of processes such as bubble dissolution, buoyancy, turbulence,
fractionation, and coalescence. Within the deeper near-equilibrium bubble plumes, recent
measurements with broad-band (5 to 200 kHz) acoustic resonator devices showed sizespectral variations close to an exponential decrease with radius. Specifically, Vagle &
Farmer [15] reported n(a) = 1.6~10' exp[-a / 34pml for a > 20 pm. The acoustic
resonator measurements generally show higher densities of smaller bubbles (a c 50 pm)
than older photographic techniques (e.g. [41). These high densities of smaller bubbles
have been confirmed with coincident 200 and 400 kHz IES measurements [15].
Most previous HF acoustic reverberation models assumed a horizontally-uniform
bubble layer, with a vertical exponential depth-scale and surface bubble density dependent
only on wind speed [ 1,21. However, the IES measurements discussed here clearly show
that the bubble concentrations and e-folding depth scales are temporally and spatially
variable. The model proposed here keeps the Farmer-Vagle size-spectral shape and the
observed exponential decay with depth, but uses surface bubble density, no, and e-folding
depth, d, parameters extracted from the IES data on a per record (3 s ) basis. The observed
temporal variations in no and d can then be converted to spatial variations by assuming a
small (20 cm.s-') but realistic horizontal current (see [ 131).
Using well-known models for bubble scattering [ 81, the extinction and backscattering
cross-section per unit volume, seu> and s,u> (units of m-'), can be calculated as integrals
over the product of the bubble size spectra with the acoustic scattering cross-section for a
single bubble. For HF sonars (particularly >lo0 kHz) both the resonant and non-resonant
scattering contributions are important, thus the limits for numerical integration are taken
as 1 and 500 pm. The extinction cross-section determines an extra transmission loss,
TLbub(f,z)
= 4.341 s,(f, z ) d / ,(in dB), where 1 is the distance along the acoustic path.
path

This loss term will be significant for the long, near-horizontal paths through the denser
portions of the bubble plumes. Note that for a single frequency and with a specific bubble
size-spectral shape, these integrals can be evaluated to produce relations between s,(z),
s,(z), and no(z).For example, at 200 kHz and z = 0 the ratio of s, to no is 9.21 x lo6 m-2(re
1 pm bin). Conversion of the 200-kHz IES extrapolated s,(O) data to no then allows bubble
calculations to be performed at other frequencies.
For near-horizontal sonar performance in near-surface bubble layer, the horizontal
and vertical variations of s,, s,, and sonar beam-pattern must be included together in the
calculation of the total target strength of the bubble layer at any given range. This is done
by combining several terms of the sonar equation into a vertical integral for apparent
backscatter cross-section.

419

where c is sound speed, z is the pulse length, x is horizontal range, cp is the horizontal
beam-angle, and D(0) is the transducer vertical directivity function vs. angle 8 from the
transducer axis, evaluated as a function of range and depth given the sonar geometry. The
above assumes slant range r >> sonar depth, so that r = x. Contributions from surfacereflected paths are neglected because of the large bubble extinction losses encountered
within the uppermost meter. The sonar equation is then used to convert this apparent
cross-section to a reverberation level, RL, as a function of range, i.e.
RL(x)=SL+lOlog,,[o]-40log,,[x]-2.a.~,
where SL is the sonar source level and a
is the acoustic absorption in seawater.
An alternate formulation for bubble layer reverberation based on a semi-empirical
approach has been advocated by Dahl [I]. This model features a horizontally-uniform
bubble layer, coherent summation of direct and surface-reflected multi-paths, a resonant
bubble scattering approximation, and includes both backscatter and extinction effects.
Dahl's model uses a empirical quantity pv,which is related to the depth-integrated density
of resonant-sized bubbles and has been fitted to a large set of surface backscatter data as a
function of frequency (up to 100 kHz) and wind-speed.
Perhaps the most demanding and structure-sensitive application is that of an HF
obstacle avoidance sonar oriented near-horizontally to search for a near-surface object
(perhaps a whale or floating mine). The H F sonar is modeled with a 1" horizontal by 30"
vertical beam-pattern oriented horizontally from a depth of 5 m (as from the bow of a
vessel). In this geometry the grazing angles vary from about 27" at close range to 1" at the
maximum range (300 m). The transmit source level is taken as 210 dB (re 1 pPa at 1 m)
with a pulse length T = 500 ps. Initially, a frequency of 100 lcHz is chosen as being typical
of a compact HF obstacle avoidance sonar. In all these models the complicating effects of
surface waves and near-surface sound-speed (including bubble-induced sound speed
anomalies) variations are purposely ignored. This is not because such effects are
insignificant, but rather because for the purposes of contrasting the effects of bubble
spatial structures against horizontally-uniform models, the inclusion of these effects would
unnecessarily complicate the picture. In principle the effects of surface waves and soundspeed anomalies could be incorporated into the above model by replacing straight-line
paths with ray-tracing results.

3.2

H F Sonar Performance

Clearly, the modeled sonar performance will be dependent on the bubble parameters
chosen. To be realistic, model results will be examined in detail using data from a period
of average bubble activity. The horizontally-uniform bubble layer case was generated
using parameters extracted from the one-hour averaged S,(z) profile. The spatially-varying
bubble plume case was generated using the measured no and d values from individual 3-s
records from the same period, converted to horizontal distance using a current magnitude
of 20 cmX'. Reverberation vs. range curves from 20 separate bubble-plume realizations
@ings) were calculated, each with start time advanced by 50 records (150 s) relative to the
previous. Finally, 20-ping averages of the reverberation and target signal-to-reverberation
ratio were calculated.
Figure 4 shows the comparison of predicted RL at 100 kHz, along with the Dahl
reverberation prediction using wind speed = 7.4 mi' (pv = 0.054). Firstly, note the

420

agreement between the Dahl prediction and the horizontally-uniform layer case, lending
confidence in the accuracy of the bubble reverberation model outlined herein.
Additionally, the 20-ping averaged reverberation curve is in close agreement with both the
horizontally-uniform prediction and the Dahl model, providing a useful consistency check.
Figure 4 includes one arbitrarily-chosen single ping. Although the overall rangedependence of the single-ping plume RL curve is similar to the horizontally-uniform
cases, there are large variations that correlate strongly with the bubble plume structures. In
particular, the modeled RL near plume maxima is typically 3 to 8 dB higher than predicted
by the horizontally-uniform models, whereas within the null zones, where there is strong
acoustic shadowing, the plume-case RL drops by up to roughly 20 dB. Moreover, the
anomalies in RL increase with range, due to the decreasing grazing-angle inducing greater
shadowing effects. These variations are important as they imply the visibility of a target
on a single-ping basis will be strongly dependent on its location relative to the bubble
plumes. In many instances the plume-case RL is much better (lower) than predicted by
horizontally-uniform models, however reverberation is only one part of sonar
performance.
r

. . . . . . . . . . . . . . . . . . . . . . . . . .

120

110
100

- - - Dahl model

50

40

20

40

40

'

60

80

single-ping

100

120

-D-

140

160

180

20-ping average

200

220

240

260

280

300

-horizontally-uniform

30

-B
K

a:
v,

20
10
O
-10
-20
-30

20

40

60

80

100

120

140

160

180

200

220

240

260

280

300

Horizontal Range (rn)

Figure 4. Comparison between horizontally-uniform and single-ping predictions of reverberation level (RL)
and signal to reverberation ratio (SRR) at 100 kHz.Bubble conditions taken 1700-2100UT March 24*, 1996.
Top plot shows bubble penetration depth corresponding to the single ping.

Figure 4 also shows a comparison between the horizontally-uniform and plume cases
for the signal to reverberation ratio (SRR) of a TS = -15 dB target located at 1.0 m depth
for a given horizontal range. This TS is a generic value for objects like sea-mines or small
whales. Clearly, although the overall range-dependence is similar, the single-ping plume
result shows important range-dependent features. For example, within the 7-m deep plume
and its accompanying shadow-region near 120 m range, the SRR drops by roughly 45 dB.
This drop-out is due to high extinction loss within the plume. Conversely, at some ranges

42 I
the SRR is higher than predicted by the horizontally-uniform model. Clearly, the details of
the SRR anomalies will change from ping to ping, and this variability is significant.
Since it is the ping-to-ping variations in RL and SRR that impose limitations on sonar
performance, it is useful to consider the statistical distributions of these parameters.
Another model run was performed, evaluating RL and SRR at 5 discrete ranges over 400
separate bubble plume realizations. From this the probability density of SRR was
calculated, as shown in Figure 5(a). The SRR distributions vary with horizontal range,
concentrated near +20 dB at 50 m range, but increasing in spread to cover the range -40
to +30 dB by 250 m range. Since there is no other built-in source of variability, in the
absence of bubble plumes these distributions would tend towards delta-functions at all
ranges. Additionally, the Probability of False Alarm (PFA) can be calculated from the
Cumulative Density Function (CDF)of normalized reverberation amplitude, A, using PFA
= 1 - CDF [5]. For comparison, fits to a Weibull distribution (PFA = exp[-aAq) are
shown for the closest and farthest ranges. The simplest statistical models predict that the
reverberation amplitudes should follow a Rayleigh distribution [5], i.e. a Weibull
distribution with a = 1 and p = 2. However, conditions for Rayleigh statistics, specifically
a randomized superposition of echoes from a large number of scatterers, were not built
into these simple reverberation models. As shown in Figure 5(b) the PFA curves show
significant differences to Rayleigh, being more compact ( p > 2) at close range and more
spread (p c 2) at greater range. Other similar simulations showed that the bubble-induced
variability decreases as the sonar and/or target move to greater depth.
Horizontal Range:
t 5 0 r n
-A- 150rn
t250m

0.0

0.5

1.o

-~-lOOrn
+ZOOIT

1.5

2.0

2.5

Normalized Reverberation Amplitude

3.0

3.5

4.0

Figure 5. For 400 synthetic pings using bubble plume data from 1700-2100UT March 24" at 100 k& using
sonar geometry as in Fig. 4 (a) probability density of signal-to-reverberationratio vs. range, and (b) probability
of false alarm compared to Weibull distributions.

Summary Discussions

Bubble plumes were consistently observed with a 200-kHz IES in the near-surface ocean,
with typical penetration depths from 2 to 15 m and maximum penetrations up to 25 m.

422
Bubble penetrations in excess of 3 m occurred over 61% of this late-winter to spring
period. This depth is significant as it is below the reach of direct breaking-wave injection,
which is generally confined to the upper meter [9], and thus other mechanisms
(turbulence, convection, Langmuir circulation) must be responsible for this deeper
advection. Generally, the deeper portions of these bubble plumes were rather tenuous in
terms of bubble densities, with corresponding volumetric backscatter strengths in the
range -60 dB to -30 dB (re 1 m-') and void fractions in the range 10.'' to lo-'. However,
the contrast between bubble plume and accompanying null zone backscatter at vertical
incidence was observed to reach up to 40 dB. Within the near-surface bubble layers the
bubble backscatter cross-section was observed to decrease exponentially with depth, with
e-folding scales from roughly 0.2 to 3.0 m. Under conditions with relatively constant
bubble activity over a 3-day period, the penetration depths were shown to follow a lognormal distribution.
While similar observations of bubbles in plume-like structures in the near-surface
ocean have been reported previously, the significance of these structures on forwardlooking HF sonar performance has been largely overlooked. Considerable work has been
performed in the measurement and modeling of HF surface reverberation at low grazing
angles [1,10], with the general goal of predicting the (temporally or spatially) averaged
reverberation level vs. grazing angle and wind speed. In this study a simple model for
bubble-induced reverberation using time-averaged bubble density profiles was able to
match closely the previous model results. This agreement is remarkable given the
differences in modeling assumptions. Additionally, ping-averaged reverberation and target
signal-excess predictions derived from actual bubble plume structures matched the
predictions based on horizontally-uniform bubble density profiles. This suggests that the
average sonar performance can be realistically predicted using horizontally-uniform
models. However, active sonar detection occurs in the context of the actual, not timeaveraged, bubble spatial structures and thus it is important to understand the ping-to-ping
variability and how it limits sonar performance.
When the observed bubble plume structures were used in the model, important
deviations in sing le-ping reverberation level and target signal to reverberation ratio were
found. As expected, the reverberation peaks were coincident with the bubble plumes, and
reverberation drop-outs occurred in shadow zones behind the plumes. Probability of False
Alarm predictions based on the reverberation were found to follow Weibull distributions,
with increasing spread at longer range. As these reverberation curves provide a
background for target detection, this suggests that target visibility, even for targets below
the bubble plumes, will be sensitive to the location of the target relative to the bubble
plume structures. When the effects of bubble extinction loss are included for near-surface
targets, the situation becomes even more complicated. In the case of a target at 1 m depth,
anomalies up to +30 dB were found in single-ping target SRR with respect to horizontallyuniform predictions. The models also predict that SRR distributions will increase in spread
at longer range.

Acknowledgements

The IES development and moorings program were supported during 1994-98 by Fisheries
and Oceans Canada and the Canadian Joint Global Ocean Flux Study. The author is
grateful for the support and guidance of Dr. David Farmer, then of the Institute of Ocean
Sciences, during the instrumentation development and mooring program.

423
References
1. Dahl, P., The contribution of bubbles to high-frequency sea surface backscatter: A
24-h time series of field measurements. J. Acoust. SOC.Am. 113 (2003) pp. 769-780.
2. Hall, M., A comprehensive model of wind-generated bubbles in the ocean and
predictions of the effects on sound propagation at frequencies up to 40kHz. J. Acoust.
SOC. Am. 86 (1989) pp. 1103-1 117.
3. Henyey, F., Acoustic scattering from ocean microbubble plumes in the 100 Hz to 2
kHz region. J. Acoust. SOC.Am. 90 (1991) pp. 399-405.
4. Johnson, B., and Cooke, R., Bubble populations and spectra in coastal waters: a
photographic approach. J. Geophys. Res. 84 (1979) pp. 3761-3766.
5. Lyons, A., and Abraham, D., Statistical characterization of high-frequency shallowwater seafloor backscatter. J. Acousr. SOC.Am. 106 (1999) pp. 1307-1315.
6. McDonald, B. E., Echoes from vertically striated subresonant bubble clouds: a model
for ocean surface reverberation. J. Acoust. SOC.Am. 89 (1991) pp. 617-622.
7. Medwin, H., and Breitz, N., Ambient and transient bubble spectral densities in
quiescent seas and under spilling breakers. J. Geophys. Res. 94 (1989) pp. 1275112759.
8. Medwin, H., and Clay, C. Fundamentals of Acoustical Oceanography (Academic
Press, San Diego, 1998).
9. Monahan, E., and Lu, M., Acoustically relevant bubble assemblages and their
dependence on meteorological parameters. IEEE J. Oceanic Eng. 15 (1990) pp. 340349.
10. Nutzel, B., Herwig, H., Koenigs, P., and Monti, J., Acoustic backscatter
measurements in the North Sea: 3-18 kHz. J. Acoust. SOC. Am. 95 (1994) pp. 24882494.
11. Norton, G., Novarini, J., and Keiffer, R, Modeling the propagation from a
horizontally directed high-frequency source in shallow water in the presence of
bubble clouds and sea surface roughness. J. Acousr. SOC. Am. 103 (1998) pp. 32563267.
12. Novarini, J. and Bruno, D., Effects of the sub-surface bubble layer on sound
propagation. J. Acousr. SOC.Am. 72 (1982) pp. 5 10-514.
13. Thorpe, S., Measurements with an automatically recording inverted echo-sounder,
ARIES and the bubble clouds. J. Phys. Oceanogr. 16 (1986) pp. 1462-1478.
14. Trevorrow, M., Measurements of near-surface bubble plumes in the open ocean with
implications f6r high-frequency sonar performance. J. Acoust. SOC. Am. (2003 in
press).
15. Vagle, S., and Farmer, D., A comparison of four methods for bubble size and void
fraction measurements. IEEE J. Oceanic Eng. 23 (1998) pp. 21 1-222.

424

The analysis of seismic data structure and oil and gas

prediction
Wang, Shangxu + Lin, Changrong ++

Abstract
In the present paper, a new concept called numerical structure of seismic data is
introduced and the difference between numerical structure and numerical value of
seismic data is explained. Our study shows that the numerical seismic structure is
closely related to oil and gas-bearing reservoir, so it is very useful for a geologist or a
geophysicist to precisely interpret the oil-bearing layers from the seismic data. This
technology can be applied to any exploration or production stage.
This new method has been tested on a series of exploratory or development wells
and proved to be reliable in China. Hydrocarbon-detection by this new method for 39
exploration wells on 25 structures indicates a success ratio of over 80 percent. Four
large oiVgas field were found in China Sea via application of the method. The new
method can be applied for oil prediction for:
0 depositional environment of reservoirs with marine facies, delta, or non-marine
facies (including fluvial facies, lacustrine facies);
0 sedimentary rocks of reservoirs that are non-marine clastic rocks and carbonate
rock;
0 burial depths range from 300m to 7000m, and the minimum thickness of these
reservoirs is over 8 m (main frequency is about 50 Hz).
Keywords: Hydrocarbon prediction,
Hydrocarbon-bearing strata
Seismic data structure and data value, Seismic facies variation
Success ratio
University of Petroleum, Key Laboratory of Geophysical Prospecting ,CNPC ,Beijing, 102249
Production Institute of CNOOC Research Center, Beijng,lOll49

+
+t

1. Introduction
A system consists of many factors. If factors of the system are determined, the
relationship between factors is clear, the structure of the system is clear and the
principle of action of the system is clear, then such system is called as White System.
If the information of the system is completely absent, then such system is called as
Black System. The system between the Black and White one, that is part information
is known and part information is unknown, such system is called as Grey System.
Since during the process of the petroleum seismic exploration one may obtain the

425
primary understanding on the geological object due to previous geological and
geophysical study. The geophysical theory already has created the basic relations
between the geological targets and seismic data. However, due to complexity of the
geological problems, the complexity of rock media and properties of contained fluids,
recent seismic wave theory could not thoroughly and clearly describe the propagation
regularity of seismic waves in the rocks. Consequently, in the system of seismic
exploration the geological object, the relation between geological target and seismic
data are partially clear. As well as the principle of reciprocal actions between systems
is not clear. Therefore, the seismic exploration system is a Grey System.
Recent paper takes the seismic exploration system as a Grey System. The
purpose of using the theory of Grey System is to analyze the amplitude characteristics
of the seismic reflection waves. Hence, the seismic gas reservoir prediction technique
is formed. The basic point of view is as follows: The reservoir is only exists at local
area of the geological stratum. The seismic reflection waves mainly come from the
interface, which is not a gas bearing one. The regularity of such reflection is a leading
one. The reflection of gas bearing reservoir has an individual character. Between both
gas bearing and non gas bearing reflection regularities there exists difference. The
Grey Prediction model, created by receipted seismic amplitude, satisfies non gas
bearing reflection regularity. The use of such model will relatively accurately predict
the amplitude variation of non gas bearing seismic reflection. Whereas the doing
prediction for amplitude of seismic reflection of gas reservoir with this manner will
create relatively large error in comparison with the actual seismic reflection amplitude.
The existence of prediction error by Grey modeling indicates the existence of gas
reservoir. Basing on such a basic point of view, recent paper has create a method of
seismic amplitude Grey modeling, a method of prediction error analysis. Therefore, a
new technique of prediction for reservoir, using seismic Grey modeling, is formed and
has been tested and verified by actual data.
1.1 Data generation
The purpose of treatment of Grey values by theory of Grey System is to find the
regularity among the data by the method of data processing, rather than to find the
probability distribution or statistical regularity. The data, located in the row of the
initial data, being treated by some data processing manner, are called as data
generation. The data generation is one of the important manners for generation of the
data in the theory of Grey system. The data generation in recent paper consists of
following steps.
1.2 Acuuumulated Generation Operation(AG0)

AGO is a common generation in the Grey system modeling and is a one of the
methods of whitening in the Grey process, since from AGO a development tendency

426
of accumulation process of some Grey values may be observed and a situation of
some regularity in characterization of data scattering and in containing integral
characteristics may be explicitly displayed. For the amplitude series Z'O) the AGO
series is

z(I)

"z,

(1)

,z2(1) ,...,

2(1)

"-1,

z"'],

The Grey theory proves, that in AGO generation there exists a Grey exponential
rule, e.g. the smooth discrete function after doing AGO obtains exponential form.
Since Z'') is a series X'') after triple sampling interpolation, the given series
obtains discrete smooth character. Therefore,
satisfies Grey exponential rule.

1.3 Inverse Acuuumulated Generation Operation(IAG0)

The reverse process of AGO is called IAGO. For series Z(O) and Z(I)
Z:O)

4,)

Its IAGO is written as Z(O)= IAGO(Z"')

2. Grey modeling
The Grey modeling is establishment of a differential equation by rows of initial
data after generation. The given equation is called as the Grey model due to its non
uniqueness and approximate property. The given model has a general form

dZ
dt

-+a2

=u,

where a,u are parameters, which require determination.

Its differential form is
2:)

- z:",

+ a(zf) + z:",) / 2 = u

S i n c e z r ' = z ~ ) - z ~ ! , , k = 2 , 3,...,n-l,n,then
zy

+ u(2y' + z 3 / 2 = u

Assuming, that series

equation ( 2 ) . Substituting

satisfies equation ( 1 ) , at the same time certainly

z(I) into equation ( 2 ) ,one may solve parameters a and

427
u by least square method.

--

where

y=

[zy

z(0)

... z (o)
,

The analytical solution of differential equation ( 1 ) will be

U

Z = ( Z ( 0 )- -)e-"'
a

U
+-

The discrete form of the solution will be

*(I)
k+l

= (zl(l)--)e-"

+--,k=1,2 ,..., n-1

a

.(I)
=[z1

*(I)
z2

...

Let
z*(l)

z n

...

Z(O)should approach to Z(O),but with some error. The absolute error is Iz'!? -

and the relative error is

From point of view of theory the obtaining of model parameters a and u is
*(O)
mainly influenced by non gas reservoir reflection amplitude. z mainly reflects the
existence of the prediction errors of Grey model, which indicate the relatively large
change of the reflection amplitude, dealing with non gas reservoir.
3. Test and verification by actual data
Since the traditional seismic "BrightIDim Spot" method was developed in oil
exploration in 1940's, many other traditional methods such as AVO, Pattern

Recognition and Neural Network, etc., were also introduced. These traditional
approaches are mainly based on seismic data value[l]. However, they have the
following weaknesses:
(1). The traditional methods need other information from geology, well logging,
and etc.These methods cannot be directly used to perform the prediction only based
on seismic information.
(2) The traditional methods cannot provide a quantitative prediction result.

428
(3) The traditional methods bring much uncertainty since data value is instable.
A well-known example would be Bright spots, which might, or might not mean a
hydrocarbon possibility. Therefore, the prediction result using these methods can be
very inaccurate and inconsistent.
Therefore, a direct, quantitative and accurate method for hydrocarbon prediction
is needed.
4.

The characteristics of seismic data structure

In the paper the characteristics of seismic data structure means characteristics of
wave form of discrete data points of each seismic trace, displayed by time sequence
(Fig.5). The basic peculiarity is to study the seismic data structure and is not to study
the values of seismic data (i.e. Seismic facies variation). It transfonns independent
discrete data into continuous predictable data by generating the gray digits (GD)[2],
Finally it aims at reach goal for petroleum prediction . Consequently, it is not a
mathematic statistical method, it is a continuous dynamic prediction process of
petroleum recognition of whole area data body, beginning with study of creating 1-D
GM model seismic trace and ending by 2-D gray cross-correlation . Finally a
continuous dynamic prediction process of 3-D data body for petroleum recognition of
whole area will be reached.
The determination of petroleum bearing strata by seismic data structure has
following advantages: simplicity, promptness, high stability, wide applicability and
high accuracy.

5. Study on characteristics of seismic data structure and relation with petroleum.

5.1 Matives

The bulk information volume of seismic profiles is tremendous. It contains at

least two aspects: data value and data structure. The data values by principle of
difference information (i.e. principle of seismic facies variation) are only a displayed
part of total information. Another part is data structure. Since the data structure is
relatively stable, it can be used for petroleum prediction, which is based on the
difference between physical property of oil and gas bearing reservoir sandstones and
that of surrounding rocks and on the difference of fluid properties. Due to such
difference the seismic facies will change, when seismic wave ( P wave) penetrates oil
and gas bearing strata and, seismic data structure will change as well.

5.2 Graphic Solution

Fig. 1 is a migrated seismic section in black-white after stacking of main profile

429
ILN316 in Penhu oil field. On lower right comer there is a thick gas bearing stratum,
called P4 horizon with thickness of 20m , penetrated by well W2 on the F-2 fault
block (Total thickness of sand body is 40m). The wave peak corresponds to the top of
gas stratum. The middle part, discovered by well W1, is main block of the gas field.
Fan-2 fault is the boundary fault between them.

04

Fig. 1 Seismic Section of Line ILN3 16(Black and White)

Fig.2 is bipolar section of main profile ILN313. Its central part represents
distribution of P4 gas bearing horizon on the Fan-2 fault block. The area of
distribution is bounded by fault F2 from left side and by boundary point CDP884 of
this line from right side. The depth of gas-water boundary is 3124m(2448ms); CDP
distribution range is 830-883 and time interval 2422-2448ms. The top of gas stratum
P1 corresponds to the wave peak and displayed by red color, whereas the wave
through--by black color. The locations of two profiles are shown in Fig. 1 and Fig.2.
Fig.3 is a map of sand body thickness of gas horizon P4 at walls W3-W6 of
Fan-2 fault block. It can be seen, that the thickness of gas horizon at the highest point
of the top can reach 45m.The intersection of two lines(main line ILN3 13 and liaison
line XLN884) is the point of gas-water boundary.
Fig.4 is a laterally and horizontal enlarged map of characteristics of seismic
waves form at 4 adjacent CDP points in the vicinity of gas-water boundary. The wave
forms of peak between CDP points are completely separated due to sufficient lateral
enlargement between traces. Therefore, it can clearly display the difference of the
wave forms at the time 2445ms from CDP883 to CDP884. Here A2 at time 2448ms of

430

Fig.2 Seismic Section of Line ILN313(Color)

Fig.3 Map of Sand Body

Thickness of Gas Horizon P4
at Wells W2-W6

Fig.4 Map of waveform variation

characteristics o f A l andA2 at line
ILN3 13

CDP884 is outside of the gas-water boundary. Whereas A1 at 2445ms of CDP883

and CDP882 is within the limit of gas horizon. Such a detail change can not be
differentiated on the facies variation maps Fig.1 and Fig.3 or on their inversion
sections (including hydrocarbon detection facies variation maps), because their

43 1
characteristics of amplitude of wave peaks are stacked on each other and form strong
amplitude anomaly, which can not show the change of the wave form.
Fig.5 is a map of two seismic traces, enlarged and extracted independently and
simultaneously from same seismic lines as from Fig.2. The horizontal axis is
reflection amplitude and the vertical one is t Otime. The table y l and y2, located at
right side from seismic trace, are respectively corresponding amplitudes of real t 0
time of CDP882 and CDP884 through SEGY transformation with sampling rate 2ms.
It can be seen from comparison between wave forms, that the wave forms at time
interval 2430ms to 2445ms are different. The left wave form at 2440ms is through,
whereas right--peak. It shows the difference in data structure.

-am -am -tmo

iom

am

JOO

Fig.5 Map of variation characteristics of data structures

(wave forms) at different CDP point of line ILN3 13

5.3 Analysis
The impossibility of lateral response of strong amplitude event to change from
gas bearing strata to water bearing strata draws following inspiration: seismic profile
in black-white (Fig. 1) can not dispose detailed characteristics of wave form changes
due to stacking of wave peaks on each other. Therefore strong pick amplitude reflects
only change characteristics between whole sand body and its surrounding rocks
(shale). It can not display the change of characteristics inside the gas or water bearing

432
strata. The bipolar colored display section (Fig.2) enhances optical dynamic range,
and help interpreter to judge the effect of amplitude anomaly, but it adds some
uncertainties in recognition of oil-gas-water relations. Whereas the phenomenon of
change process of wave forms A1 and A2 from CDP882 to CDP884 may reconfirm
the interpretation results of lateral change with the same sand strata , especially, the
sand body of alluvial deposits under the tidal actions, which have stronger lateral
change.
Fig.5 is a sketch of wave from qualitative description to qualitative interpretation
form Fig.4. It sufficiently explains, that different peak wave form characteristics
correspond to existence of different data structures, and vice versa, different data
structure certainly corresponds to different external wave form change . Fig.6 is an
analysis map of mathematic calculation trend process of different seismic structures
for realization of recognition of Fig.5. Yl(882)and Y2(884) are obtained after gray
data transformation of measured data points in fig5. In the map the dotted line (Y 1)

3500
3000

&')

(Grey Model I ing)

-+&')=U

dt

2500
2000
T

4:

0. 2
S3W

1371. h

R2 = 0.958

1500

1000
500

0
2420

R 2 = 0.8574

2430

2440

2450

2460

2470

2480

Time
Fig.6 The map for tendancy analysis of
different seismic data structures
and solid line (Y2) are the results, obtained by method of polynomial prediction.
Triangles (A) and solid circle point ( 0 ) indicate the results, obtained by gray
modeling method. The comparison and analysis of different methods, such as
recursion classification, trend extrapolation, exponential smoothing, gray prediction
etc, show that method of gray prediction most efficiently reflects the characteristics of

433
seismic data structures. The method of gray prediction will be described in other
paper.
The above mentioned results of analysis indicate, that if enlarge seismic trace in
lateral and vertical directions, that is, treat seismic trace as a well-logging curve (the
meanings are different), convert SEGY format into ASCII format and introduce gray
prediction model, then it is of significance to study the relationship between seismic
data structure and petroleum bearing.
6. Application of Seismic Data Structure Method (2 Cases)
Following are two cases histories, Dongfangll-11 gas field and Ledongll5-11

gas field. Describing the prediction effect, obtained at well Dongfang 11- 11- 1 ,
Dongfang 11- 11-2 and well Ledong 115 - 11- 1.
7.1 Gas field DF11-11
Well Dongfang 11- 11- 1 was first one, drilled in November 1991. Before
drilling a gas bearing stratum was predicted below depth 2000m (T23) by traditional
method (see Fig.7), Whereas the gas bearing stratum, predicted by our proposed
method, was at time1.31 s, corresponding to depth 1300m. This gas bearing stratum
were thickening to both sides up to 40-80m (see Fig.8, Fig.9) . The real drilling
discovered a gas bearing stratum with thickness of 6Sm, interpreted by well logging
at the depth 1326.5m. The thinning of thickness at this place was mainly explained by
reason, that the well was drilled at place, where thickness of gas bearing stratum was
minimal. On the Seismic Section this place was displayed as a gap (see Fig. 8) . Later
the second well-well Dongfang 11- 11-2 was dnlled in July 1992. Initially this
well was designed and expected to meet 2-nd, 3-rd, 4-th and 5-th sets oftarget
sandstone strata respectively at depth interval of 1275- 1297.8m (22.8m thick),
1318.5- 1362.4m(43.9m thick), 1414.2- 1507m (92.8m thick) and 1538- 1627.2m
(89.2m thick) . The prediction before dnlling, made by proposed method was
expecting to meet two gas bearing strata A and B (see Fig.10, Fig.11) ,which
correspond respectively to 0.95s and 1.28s seismic section second target stratum of
designed well. The real dnlling meets second ( B ) stratum at depth interval of 1284
-1296m (1.28s) (Fig.11). Probe testing gave natural gas of 33X104m3 per day.
Whereas the upper stratum A was met by drilling at depth interval of 922-926m. The
lithology is siltstone with thickness of 4m. A gas anomaly (value of gas-logging from
1% to 2.4% ) is discovered. The time difference of acoustic waves in the sonic log
has a leap (see Fig, 12) . The lithology of upper and lower surrounding rocks became
fine mudstone. On the section of tracing integration there appears a sign, which is

434

Fig.7 The Dredicted reservoir column of well Donnfang: 11-11-1

Fig.8

Seismic section passing through well Dongfang 11-11- 1

analogous to the second stratum. The probe testing for gas was not undertaken,

435
because it was not included within the designed target stratum. Afterwards, the forth
and fifth wells also met gas bearing strata, which confirmed the accuracy of previous
prediction. Finally a large gas field with reserve of 1000X108m3was confirmed.

DF11-11-1

Y3489

d NE

1.6

1.7

Fig.9 Sketch map of reservoir prediction section of

Y3485

DF11-11-2

Fig. 10 Seismic section passing through well Dongfang 11-11-2

436

Y3485
DF 11-11-2

I
I
I

Fig. 11 Sketch map of reservoir prediction section of

well Dong.fang. 11-11-2

Fig.12 Section of met reservoir of well Dongfang 11-11-2 by dnlling

437
7.2 Gas field LD115-11

The fist exploration well--well Ledong 115- 11- 1 began to be drilled on

February the second 1993 on the structure of this gas field, and was completed March
the 29-th. The depth of completed borehole is 2541m at Huangyin group. Before
dnlling a prediction by other methods was made: inferred major target strata are
located between T10 and T20. Since the structure Dongfang 11- 11 is located not far
from recent structure, in process of prediction the stratum of Dongfang 11- 11 was
considered as well. Whereas our prediction with proposed method shows, that the
depth of gas bearing stratum of Ledong 115 - 11 is shallower than that of Dongfang
11-11. It is located between TIO-TI1. There are two gas bearing strata at time
1.38s and 1.48s ( see Fig. 13, Fig. 14) . The real drilling has discovered two gas bearing
strata at depth interval of 1418.9-1429.3 and 1547-1553.2m with total thickness of
16.8m which coincide with result of prediction interpretation. The lithology is mudy
siltsone with porosity of 21.0-26.5%, mud content of 14.7-30.4% and water
saturation of 51.4-60.6%. The probe testing (The total thickness of intervals 14171429m and 1540-1557m was 29m and was put for testing together) gave daily gas
production of 73X104m3. Free current gave daily production of 200x104m3, which
confirms the accuracy of prediction. A large gas field with reserve of 300X108m3 is
finally fixed.

1. Case # I : Gas field DDFll-11

Methods

Depth

Thickness

Reserve

2000-3500m

NIA

NIA

1300m

40-80m

1000x 1o

First Well

1326.5

6.5m

no DST

Second Well

1284-1296.5m

12.5m

Traditional method
Data Structure

h3

Method
33 x
104m3(DST)
Third Well

1320-1360.4m

40.4m

185X
104m3(DST)

Fourth

1324.2-1407m

72.8m

300 X
104m3(DST)

Total

1250-1450m

40-80m

Iooo x 1o h 3

438
2. Case #2: Gas field LLD115-11

I Methods
I
1 Traditional method 1
Data Structure

DeDth
2000-2500m

I
1

Thickness

Reserve

NIA

NIA

1400-1700m

15-30m

1418.9-1429.3m

16.8m

I
1

300 x 1o h 3

Method
First Well

73 x

1547-1553.2m

104m3(DST)

Second Well

1540-1567m

27m

Third Well

1407-1429m

22m

108X
1O4m3(DST)

75 x
104m3(DST)

Fourth

Total

91-3617

1647-1653.2m

1400-1700m

6.2m

15-30m

LD115-11-1

Fig. 13 Seismic section passing through well

no DST

300X108m3

-NE

439

LDI 15-11-1
Line: 91-3617

Fig.14

Sketch map of reservoir prediction section of well Ledongl15-11-1

7. Conclusion
The characteristics of seismic data structures have been used for research on
petroleum bearing strata since 1990. Up to 1999 the research results had been used for
hydrocarbon prediction before drilling in 39 boreholes of 25 structures. The benefit of
prediction may reach over 80%. Hence the economical efficiency is tremendous.
Since it is a continuous dynamic process of hydrocarbon recognition, so it may
be applied for long distance prediction. The prediction can be made in the new area,
where lack of drilling information, but do exist high quality of seismic data. The
prediction of shallow petroleum reservoirs of DF11-11 large gas field in 1991 was a
successful case history.
The areas of application are geological settings of terrestrial clastic rocks with
deposition environment of marine (deltaic transition zone) and terrestrial (alluvial,
lacustrian) facies. Depth of research is from 800m-5000m. The target of prediction is
sandstone horizon with thickness of more than 10m (main frequency around 50 Hz).
The algorithm for gray modeling (GM) is partially computerized. But the
geological factors ,such as system dividing and relative concept of numerical data can
not be fully computerized up to now. These are key techniques of the method, which
expect further solution.

440

References
1.Liu Wenlin, Seismic technique for development of oil and gas fields, Beijing,
Petroleum Press, 1996 (in Chinese)
2. Den Julong, et al, The new developments in the research on gray systems, Wuhan,
Press of Central China university of Seance and technology, 1996 (in Chinese)
3. A.R.Brown, Interpretation of 3-D Seismic data, Beijing, Petroleum press, 1996
(Chinese translation )
4.Lin Changrong, et al, Numerical Structure of Seismic Data and Hydrocarbon
Prediction, China Offshore Oil and Gas(Geology), Vol. 14, No.6, Dec.2000

44 1

Mid-frequency bottom backscatter measurements

and modeydata comparison
Zhongkang Wang, Xiufeng Lin, Xianyi Gong
State Key Laboratory of Ocean Acoustics,
Hangzhou Applied Acoustics Research Institute,
Hangzhou, P. 0. Box 1249, Zhejiang, 310012,China
and
Dajun Tang
Applied Physics Laboratory, University of Washington,Seattle, WA
ABSTRACT
The bottom backscattering results mainly from the roughness of seafloor and heterogeneity of
sub-bottom. An experiment was conducted in ASIAEX ECS 2001 site using a 32-element
3.5kHz reverberation array deployed from Melville to measure the bottom backscattering
strength. The direct, bottom-reflected, surface-reflected, bottom-surface reflected,
surface-bottom reflected and surface-bottom-surface reflected waves were found from the
analyzed data. Through beam forming bottom backscattering strength changing with grazing
angle was obtained. ModeVdata comparison with known environmental measurements shows
that bottom roughness is the dominant scattering cause for the ASIAEX ECS site.
1. INTRODUCTION:
The scope of this paper is to describe a direct-path bottom backscatter measurement technique
and compare the bottom backscattering strength with model estimated strength. Through
model/data comparison with known environmental measurements, it shows that bottom
roughness is the dominant scattering cause for the ASIAEX ECS site.

2. EXPERIMENT DESIGN
The experiment design was dictated by the requirement to conduct the measurement in
shallow water(water depths of less than 150m), over the bandwidth(3000-4000Hz) with
greatest interest on low to intermediate grazing angles(0-45 deg) that control sonar system
performance.
Perhaps the most significant challenge of such an experiment in shallow water is to avoid or
control multi-paths. Figure 1 shows the various paths. A vertical receive array and source can
be used to control these types of multi-paths. Short pulse lengths are required in order to
resolve angular dependence of the scattering function. Figure 2 depicts the system employed
in this study. Short pluses from the source located near the bottom of the array provided a
repeated and stable source. The types of short pluses used in the experiment were 3.5kHz and
4.0kHz PCW signal. The duration of the pulse was 100ms, 20ms, IOms, 5ms, 2ms and Ims.
The weight of the source(100 kg in air) provides enough ballast to keep the array straight
when the current shear is small. And the used steel cable from the winch stabilizes the array
from rotational forces.
2.1 Source

442
I

FIG. 1 Bottom scattering multipath

a)bottom-reflected b)surface-reflected c) bottom-surface reflected d)surface-bottom reflected
e)surface-bottom-surface reflected f) bottom-surface-bottom reflected

-134m

FIG. 2 Experiment geometry used for measureing bottom backscatter.

To get the backscattering strength we need to know the vertical index of the source. The
source is horizontal omni-directional. The vertical beam pattern can be modeled as the
following formula:

I:[-

B, CP, = B, * e
(1)
Where Bo is the source pressure at horizontal angle, Bis the grazing angle, w is a constant
dictating the width of the beam. Before the experiment, the source level and the parameter w
is got through tank calibration. For there is 6dB lower below or up the source than in the
horizontal direction, so w is l t J 2 J m . Figure 3 shows the source vertical index. There
is another proof of this 6 dB. For 31 channels recorded direct-path sound they fit the

20*1og10(R) spread law good. For example in edata042.dat. It is 1.17m between the
source and the first channel. So the first channel pressure level is the following formula:

191.3dB- 6dB - 20* loglO(l.17) = 184dB

443
It is 7.59m between the source and the 31 channel. So the 31 channel pressure level is the
following formula:

191.3dB-6dB -20*10g10(1.17+ 30*0.214)= 167.7dB

These two results fit the measured pressure level at lstand 31'' channel good. Figure 4 shows
the direct-path pressure level at 1'' and 3lStChannel.

FIGA Direct-path pressure level

FIG.3 Source vertical index
2.2 Array
The 32 elements vertical array is spaced by 21.4cm, the wavelength of 3.5kHz underwater
sound. The sensitivity of each hydrophone including the preamplifier and 180m length cable
is measured in the tank. Mean sensitivity is -184.4dB at 3.51cHz, the standard deviation of
sensitivity is 0.5dB. Each hydrophone is omni-directional. Figure.5 shows the picture of the
array. The analog signals from the hydrophones through a band-pass amplifier were recorded
in a digital tape recorder. Figure.6 shows the recorded reverberation data of first 16 channels.
Detail of the data can be seen from the zoom of channel 30 in figure.7. For the estimated
arrival time, we can find the reflected sound from bottom, surface. bottom-surface.
surface-bottom, surface-bottom-surface.
3.5 kHz,5.1 rns reverberarion data

0 15

01

0 05

0
160

200

220

240

260

280

300

lime (ms)

FIGS Picture of the 32 elements vertical array Fig.6 Recorded reverberation data of 16

444
channels
3.5 k H r . 5 . 1 ms reverberation data.30th Channel

1701

160

s! 150
E?
ZZ

140

2
3

130

CZ 1 2 0
110

Fig.7 The recorded reverberation data by the 30th hydrophone

3. DATA PROCESSING
Theoretically the backscattering strength can be extracted from one hydrophone data. But for
the high background noise especially at low grazing angle, it is hard to got the true
backscattering strength by only one hydrophone. Through beam forming, background noise
can be suppressed. Through beam-time series analysis backscattering strength can be
extracted.
3.1 Beam-time series analysis
Data can be beamformed using a plane wave frequency-domain beamformer. Frequency bins
were selected to centre at the centre frequency with the bands equal the reciprocal of pulse
duration. Beams were spaced 1 deg from upward to downward. Every time 31 channels
128-point FFT were done with 32-point data new. Then frequency-domain beamforming was
done using selected frequency bins. Figure.8 shows the beam-time series result. Vertical
arrival angle versus time for bottom backscattering can easily be seen.

-20

-30
10

-40

-50

-60

-70

-80

.-"

Elapsed Time (ms)

-90

445
FIG.8 Beam-time series

FIG.9 Geometry of bottom backscattering

FIG. 10 Estimated backscattering strength

(red:view from 0 s;blue:view from 0 ),

3.2 Backscattering strength

Through peak-pick bottom backscattering strength versus grazing angle can be extracted
from figure.8. This result can be looked as the result being extracted from one hydrophone
which is centred at the array. Figure.9 shows the one source and one hydrophone geometry of
bottom backscattering. We can get the backscattering strength formula as below"]:

Where (IP(t)l') is the pressure level at hydrophone. Where Aft) is the bottom area which
influence the receive when the pulse width is

A = 2mAr =

T ,A(?) can be

written as:

case

(3)

Where Zo is the source intensity in the horizontal direction. R is the distance from the source
to the incidence point. For bi-static case there will be R, and R,.

B ( t ) is the vertical index of the source as shown in fomula 1. The time t can be calculated to
the grazing angle 8 .
Figure.10 shows the estimated backscattering strength at 3.5kHz. There are two curves, red
one curve is got from the view of 8, , blue one curve is got from the view of 8,.

4.MODELDATA COMPARISON
From material [ 5 ] the theoretical bi-static backscattering strength is given by

o(e,,e,) =k;lz12w(k,(Coses+case,))

446
Where

k, = 2@/c, is the water wavenumber, k , = 2n(1+i6)/c2 is the complex

wavenumber of the bottom with 6 being frequency independent for a

linear-frequency-dependent attenuation coefficient. Where 7 is a function of material
properties and grazing angles 8, and 8, relative to the source at z, and receiver at
&,respectively:

1
s i d r+(l--)(l+V,)(l+V,)cod, cod,

(6)

Where p = p 2 / p l r ~ = k 2 /=(c,/c2)(1+i6).
k,
The quantitities V, and V, are the plane wave reflection coefficients corresponding to the
source and receiver grazing angles, respectively:

V=

p sin8 -JK2

(7)

p sin 8 + JiGZ

Where W is the two-dimensional power spectrum of the rough interface, its spectrum obeys a
power-law f0r[21[31[41:

The parameters in (8) are: y2 = 2.68,w,= 4.29 *

Figure.11 shows the modeVdata comparison at 3.5kHz. In the model we use the parameter as
follows: c, = 1520m/s,c, = 1600m/s,p, =l.O,p, = 1.8,6 = 0.01 .ModeVdata comparison
shows they fit well.

Fig. 11 ModeVdata comparison at 3.5kHz of backscattering strength

447
5 . SUMMARY
ModeVdata comparison with known environmental measurements show that bottom
roughness is the dominant scattering cause for the ASIAEX ECS site. The critical angle effect
on the scattering strength at about 18 degrees is clearly demonstrated.
Because the interference of the ships self noise, beam-forming is necessary for extracting the
scattering strength. The use of a vertical array is a necessary tool.

REFERENCES:
[1]A P Lyons and A L Adderson, Acoustic scattering from the seafloor: Modeling and data
comparison, JASA 95
[2]Dajun Tang, Thomas H.Orsi,Three-dimensinal density spectra of sandy
sediments,J.A.S.A107(4),April2000
[3]Dajun Tang,Simulation of shallow-water reverberation based on physical
parameters,ECUA 2000
[4]Dajun Tang,Thomas H.Orsi, Three-dimensinal density spectra of shallow water
sediments,ECUA 2000
[5]Dajun Tang ,Numerical Simulation of Scattering from a Rough Sea Bottom Measured on a
Vertical Line Array, to be published at J.A.S.A 2003

448
ACOUSTIC PROPAGATION THROUGH BUBBLES: AN EXPLORATION OF
THE lST
AND 2m MOMENTS IN VARIOUS FLOW CONDITIONS
T.C. WEBER, D. L. BRADLEY, A. P. LYONS
Penn State University/Applied Research Lab, State College, PA.
E-mail: tcwl41 @psu.edu
L.BJ0RN0
UltraTech Holding, Stendiget 19, DK-2630 Taastrup, Denmark
The classic theory for the propagation of acoustic waves in bubbly fluids, which is the 1'' moment of the
multiple scattering solution, is well known. An integral equation describing the 2" moment can also be
formulated using the multiple scattering approach, although solutions to this equation are difficult and
typically require approximations. In this paper, the multiple scattering approach is used to derive a
solution to the 2"dmoment in a slightly different fashion. This is done by first considering the solution to
the 2"d moment for the single scatteren only, following with the solution for the double scatteren, and
eventually generalizing to the complete solution. The solutions to both the 1" and 2" moments require no
knowledge of any flow parameters, although they do require averaging over all bubble configurations.
The time required for enough bubble configurations to be observed forms the basis for an indirect link
between the multiple scattering theory and the flow. This relationship between the statistics of the
propagation through bubbles and the flow in which the bubbles exist is explored in a laboratory
environment. Multi-frequency attenuation measurements are inverted to estimate the bubble population
distribution, which in turn is used to predict the 2" moment. The predicted 2"dmoment is then compared
to the measured 2* moment. The time scales associated with the convergence of the measurements to the
prediction is then estimated for the flow conditions investigated, providing an estimate of the mixing time
of the flow. Although these time scales are specific to the laboratory studies described herein, it is hoped
that they will serve to help illustrate the relationship between the multiple scattering theory of bubbles and
the flow in which they reside.

Introduction

The classic theory describing the average soundspeed and absorption of acoustic waves
propagating through bubbly liquids can be derived by either viewing the bubbly liquid as
a continuum [4] or as a multiple scattering environment [5]. Both viewpoints lead to the
same result, which has been shown to match experimental data quite well [4]:

k i = k: +4&
where

The quantity k, describes the mixture (effective) wavenumber, k, describes the

wavenumber in the bubble-free liquid, a is the bubble radius, fb is the resonance
frequency of the bubble, f is the ensonification frequency, and 6 describes the bubble
damping. 7he quantity g is the scattering coefficient of the bubble, which is dependent on
the parameters describing the bubble and frequency. The probability density function for
the location, r, and size, a, of a particular bubble is given by the quantity n(r,a)/N, where

449

N is the total number of bubbles. From the standpoint of the multiple scattering theory,
(1) is derived from an equation describing the observed acoustic pressure due to some
particular configuration of bubbles of radius, a, at positions, r, and averaging the observed
pressures over all configurations of bubbles weighted by the probability that a particular
configuration of bubbles will exist. Equation (1), then, is not expected to characterize the
propagation through bubbles for a particular realization of bubbles, but rather
characterizes the average propagation (the 1" moment) that would be observed for all
possible configurations of bubbles. Spatial structure in the bubble population is
accounted for by the theory as long as it exists on average: the bubble density, n, is a
function of position. On the other hand, temporal variability within some volume that
has, on average, a stationary bubble density is not addressed by the theory, other then to
imply that, regardless of the character of the temporal variability, the long-term statistics
will not change.
The multiple scattering theory predicts the statistics of propagation through bubbles,
and so a more complete characterization of the expected acoustic observations would
include higher order statistics. An equation describing the 2"dmoment of the acoustic
observations can be readily formulated in the multiple scattering approach, as was done
by Foldy [5],although the solution to the integral equation describing the 2ndmoment is
difficult. Approximations to the 2"dmoment have been used in the past, for example [3,
61. As an alternative to solving the integral equation, successive orders of multiple
scattering (i.e. single scattering, single plus double scattering, and so on) can be
investigated. This methodology leads to a solution for the 2ndmoment which is no longer
an integral equation and is much simpler to solve. The derivation is presented in the
following section. Similar to the 1" moment, the 2"dmoment addresses only stationary
spatial structure.
Although there appears to be no direct link between the multiple scattering theory
describing propagation through bubbles and the flow in which they exist (other then the
flow's effect on the average bubble distribution), an indirect link exists in the time
required for the bubbles to be mixed through enough bubble configurations in order to
match the theory. Since the different statistics described by the multiple scattering theory
may converge at different rates, it would be desirable to examine the pdf rather then a
single statistic. A useful metric would then be based on the hypothesis that the samples
have adequately explored an expected pdf.
If the acoustic observations are
al,a2,...,a,,,...aN.I,aN,then the sample pdf would be found from the first n observations and
the expected pdf would be found from all N observations. The minimum number of
independent observations required for the two p d f s to be considered identical at some
acceptable confidence level could then be found. Multiplying this minimum number of
samples by the correlation time for the observations would provide a measure of the
minimum time required for the flow to mix the bubbles into enough different
configurations to meet the assumptions inherent in the multiple scattering theory (namely,
that the acoustic observations are averaged over all bubble configurations). Here, this
will be termed the mixing time.
With solutions to both the 1'' and 2"dmoments of the bubble propagation problem in
hand, this indirect link is explored in a laboratory environment. Forward propagation
experiments were conducted in a small tank filled with bubble-laden water. The
consistency between the Is' and
moments was investigated by transmitting many
pulses at a high rate between source and receiver in order to measure the frequency

450
dependent attenuation. The attenuation measurements were then inverted to find the
bubble density as a function of radius using the iterative method described by Caruthers
et. a1 [2]. With a known bubble distribution, it was then possible to predict the 2nd
moment of the received pulses using the theory developed in the following solution,
which was found to match the observed 2"dmoment, providing an indication that enough
observations had been made to apply the multiple scattering theory (at least for the lstand
2ndmoments). The mixing time is also estimated using a chi-square goodness-of-fit test
to compare the sample pdf from subsets of the acoustic observations to the pdf from the
entire ensemble of observations. Although the mixing times observed here are specific to
the laboratory studies described in this paper, it is hoped that they will serve to help
illustrate some aspects of the general problem of linking the multiple scattering theory of
bubbles to the flows in which bubbles exist.

Multiple scattering of acoustic waves in bubbles: the Znd moment

In addition to the classic theory, the multiple scattering approach can be used to
investigate the higher order moments. Following Foldy [5], the exact solution for the
pressure observed at some field point, r, in the presence of a bubble population described
by bubbles located at positions rj is

j'+ j

where po is the bubble-free incident pressure (that which would be present in the absence
of any bubbles), g, is the scattering coefficient for the f h bubble, E is the free space
Green's function, and p' is the incident pressure at the f h bubble which includes the
bubble-free incident pressure as well as contributions from all other bubbles. Equation
(2) describes the observed pressure at r that is valid for only one realization of bubble
positions. It is more useful to examine the statistical nature of p ( r ) , and Foldy did this to
derive the classic theory given by (1) for the first moment. Foldy also laid the foundation
for an approach to the second moment, and although this will not be used here, much of
his general problem formulation will be used. The second moment will be explored by
considering the multiple orders of scattering separately, starting with the single scatter
approximation, then the double scatter approximation, and so on.
The single scatter approximation is made by replacingp' with Po in (2):

where the term pSIis used to indicate the single-scattered pressure field. The 1'' moment
is found by taking the expected value of ( 3 ) , which contains 2N random variables (i.e the
bubble positions and radii). The joint bubble probability density function is given by
n(r,a)A? It is assumed that the bubble positions, r, and radii, a, are independent of each
other, and further it is assumed that the bubble PDF for each bubble is identical, meaning

45 1
that the sum of N integrals may be replaced by the factor N , resulting in the 1" moment
for the single scatter approximation:

(4)

= P o (I
+)
(Psl

(4)

The squared pressure amplitude may be written

P1 (r)P;(r)= Po(.)P:

Po(dP:1 ( r ) + P,*(r)Psl(r)+ PSI

(I)+

(dP;h)

(5)

The 2ndmoment is the expected value of (3,

and the variance is found by subtracting the squared mean pressure from (6):

The first term on the right-hand side of (7) can be written as two integrals, with the first
containing the N magnitude squared terms and the second containing the N(N-1)cross
terms. Assuming that N is very large, the two integrals can be written

v v

where the quantity H , which is proportional to the scattering cross section per unit volume
[ 5 ] ,is defined as

The second term on the right-hand side of (7) is the product of two integrals and can be
written

v v

which is identical to the second term in (8). Thus, the single scatter approximation to the
variance can be written

452

The same methodology can be used to improve the approximation by including the
double scattering terms. Starting with a single realization of the bubble cloud, the
pressure measured at a field point, r, can be written as the sum of the incident pressure,
the single-scattered pressure, and the double-scattered pressure:
r

+ ( P s A r ) ) P m + ( P d (r)P;l (4)+(PSl ( r ) P : k ) )

(13)

(d)+( P s 2 ( d P : 2 (4)

+ ( P s 2 ( r ) ) d ( r ) + Psz (r)Ps*l

Again, the variance is found by subtracting the magnitude squared mean pressure from
(13):

The first term on the right hand side is the variance for the single scatter approximation,
given by (10). The second term is the difference between

453

and

v v

which reduces to a single integral:

( P ~ ~ ( T ) P I ~ ( I ) ) - ( P ~ ~ ( T ) ) ( P J: ~H(k~j)I)P=o ( r j Y P : l ( r j

) ) I ~ G 7 rI'drj
j

(17)

The third term in (14) is simply the complex conjugate of (17). The fourth term in (14) is
reduced in a similar fashion, where

and since

it can be written
(Ps2(r)P:z(r))-(Ps2(r))(~:2(r))=

JH(rj i ( P I l ( r j ))[lEC.,rj 12drj

(20)

Finally, equations (lo), (17), and (20) can be used in (14) to find a simplified equation for
the double scatter approximation to the variance:

Equation (4)can be re-arranged and substituted for the mean single-scattered pressure in
the bracketed expression in (21), which then reduces to

454

Without going into the details of each step, this procedure can be carried out ad
infiniturn, a process made easier by recognizing that,
(Psfl)=(Pn)-(Pfl-*)

(23)

with the result at each step given by

For very large n, the (n-1)Ih approximation to the 1'' moment will be almost identical to
the nrhapproximation, and so an approximation to the general result to the variance can be
written

Equation (25) states that the mean square pressure amplitude observed at some field point
is equal to the squared mean pressure amplitude at that point plus contributions that
represent the incoherent sum of the mean pressure field scattered off each bubble.
It should be pointed out that the integral equation describing the 2"d moment
developed by Foldy in (51 is similar in form, but differs in two ways. Foldy uses the
mean square pressure amplitude instead of the squared mean pressure amplitude inside
the integral in (25). He also uses a Green's function that is identical in form to E, but that
propagates with the effective medium wavenumber k , rather then the bubble-free
wavenumber k,. In Foldy's solution he is summing the individual contributions of the
entire field incoherently at each bubble, as he points out, and then propagating that sum to
the field point, r, through the effective medium. The first difference would seem to cause
an overestimate of the 2"dmoment by ignoring any interference between scattered waves
in the incident field at each bubble. On the other hand, the second difference would tend
to reduce this overestimate by attenuating the wave as it propagated through the effective
medium. Irregardless of the net effect of these differences, an advantage of (25) is that it
is not an integral equation.
An approximation suggested by Foldy [ 5 ] and used by Carstensen and Foldy in
experimental work [3] was equivalent to replacing the mean square pressure amplitude
with the squared mean pressure inside the integral, which is the first term in a series
solution to the integral equation described by Foldy's exact result. If the solution
presented above (25) is correct, this would be double counting. That is to say, it is the
coherent sum of scattered signals from each bubble propagating through the bubble-free
medium that results in the attenuation. To now assume that the scattered signals are
propagating through the effective medium in order to arrive at an estimate for the mean
squared pressure would seem inconsistent.

455
3

Laboratory Measurements a the 2"dMoment

Laboratory measurements were conducted with the goal of exploring the consistency
between the 1'' and 2"d moment under a variety of conditions. The general process was to
record many transmissions through a bubble cloud over a wide range of frequencies to
measure the attenuation due to the bubbles. The attenuation measurements were then
inverted to find the bubble population using the method described by Caruthers et. al. 121.
Using this bubble population estimate, the variance was then predicted using the theory
described in the previous section and compared to the variance of the measured data.
The laboratory setup used to conduct the measurements is depicted in Figure 1.
Bubbles were generated using several cylindrical pieces of porous ceramic connected to a
pressurized air supply. The bubble generators were placed in two 4" vertical tubes near
the comers of the tank, allowing the larger bubbles to rise to the surface without
becoming entrained in the flow. Smaller bubbles drifted out of openings near the bottom
of each tube, and were entrained in the flow produced by a mixer which was connected to
a variable speed DC motor. Acoustic signals were transmitted from a 1 inch circular
piston through the bubble field to an identical circular piston hydrophone, separated by a
distance of 0.38m, amplified and filtered, and recorded on a 500 ksample/s 12-bit data
acquisition system. The entrained bubble population was dependent on the voltage
applied to the DC motor of the mixer (i.e. on the mixer speed), with higher mixing speeds
entraining a larger population of bubbles. The average velocity magnitude associated
with the three mixing speeds used for this experiment was measured using an acoustic
doppler velocimeter, and found to be 4 c d s , 9 c d s , and 12 c d s .
Gated cw pulses were transmitted at ten frequencies (45 kHz to 235 W z in 20 lrHz
steps) chosen to best characterize the frequency dependent bubble attenuation curve
within the limits of the measurement equipment. A constant bandwidth of 5 kHz was
used throughout the experiment so that the ensonified volume was dependent only on the
transducer beam patterns. The data was taken at a very high ping rate (40 Hz) for 25
separate 10 second intervals, where each ping contained the ten cw-bursts corresponding

Figure 1. Laboratory setup used to make acoustic measurements in bubbles.

456

14.5'

5
tim (seconds)

I
9

10

Figure 2. An example of the data used to calculate the 1'' and 2" moments (the amplitudes of 400 pings at 85
kHz for the second highest mixing speed).

to each frequency. Each direct path arrival was time gated, weighted with a Hanning
window, and then Fourier transformed. The amplitude of the FFT bin corresponding to
the pulse frequency was used to create a data set of amplitude versus frequency for the
different mixing conditions. The hydrophone output was recorded for seven different
conditions: no bubbles present with the mixer turned off, no bubbles present at each of
the three mixing speeds, and with bubbles resent at the same mixer speeds. An example
of the data used to calculate the lstand 2" moments corresponding to a frequency of 85
kHz and an average velocity magnitude of 9 c d s is shown in Figure 3.

In order to both place confidence intervals on the variance estimate and to estimate
the mixing time, only uncorrelated observations were used in the analysis. The
autocorrelation function was estimated at each frequency for each of the 10 second (400
sample) data sets. No strong frequency dependence was found, and the average
autocorrelation function is shown in Figure 3. Correlation times were estimated from the
average autocorrelation functions by integrating the functions up to the point where they
equaled a value of 0.1, resulting in correlation times of 0.32, 0.29, 0.21 seconds in order
of ascending mixing speed. Every 20mdata sample was used in the subsequent analysis
in order to make certain that the samples were uncorrelated.
The ratio of the mean bubble data to the mean no-bubble data was used to calculate
the frequency dependent effective medium absorption coefficient which was then inverted
for the bubble population:

where

is the mean of the bubble data,

is the mean of the no bubble data, and L is

457
average of all data

0.8

0.6

0.4

0.2

Figure 3. The estimated autocorrelation function of the acoustic pressure amplitudes for the three mixing
conditions.

the distance between the projector and the transmitter. The absorption coefficient, a6,has
units of nepers per m when L is specified in meters. It should be noted that the
experimental design has restricted the estimate of the absorption coefficient to be
independent of position, meaning that the average bubble density is also assumed to be
constant at least within the ensonfication area of the source.

To convert the attenuation measurements to the bubble population the iterative

approach described by Caruthers et al. [ 2 ] was used. This method begins by
approximating the bubble population using the resonant bubble approximation with the
measured absorption data, then calculates the absorption due to the approximated
population using the full theory described by (1). Following the terminology used in [ 2 ]
and using RBA to describe the resonant bubble approximation operator and FT to
describe the operator that uses the formal theory to calculate the effective medium
attenuation caused by some bubble population, this first step can be written

The resonant bubble approximation is then applied to a, and subtracted from twice no in
order to generate the 1'' order result (the result of the first iteration)
ni = RBA[a,]
n, = 2n, - n i = 2RBA[a6]-RBA[aO]

The second order approximation follows the same procedure, but this time using the
difference between the measured absorption and the approximated absorption, a
.:

458
a, =a,-a,
noe = RBA[a,]
a
m=

noel

nie = RBA[a,,]
n2 = no - 2n,, - ni, = RBA[a,]- 2RBA[a,]- RBA[a,,]
This process can be continued to any number of iterations; here 10 iterations were found
to provide the best match between the measured attenuation and the attenuation predicted
from applying the formal theory describing the effective medium to the predicted bubble
population.
The mean attenuation data and the results of the iterative process are shown for three
different mixing speeds in Figure 4. It can be seen that as the mixing speed increases,
more and larger bubbles are entrained in the flow. The void fractions corresponding to
the three bubbles distributions are similar: 0.85xIO-', 2 . 4 ~ 1 0 .and
~ ~ 2.8~10.'. These are
not large void fractions, although it should be pointed out that there are virtually no
bubbles with a radius greater then 8 0 ~ .

1
'

-lo6

I 7
......................
..
.....................
...........................................

iiiiiii_:iiiiiiiiii_:~

.:::;
1
.....
iL.
cc1 0 .....

.E lo+
10'

<

1o2

bubble radius (pm)

Figure 4. The mean attenuation data (top) and the corresponding predictions of the bubble size distribution
(bottom) as a function of mixing speed.

459
The frequency dependent normalized variance in the amplitude data is plotted in
Figure 5, where the normalized variance is defined as

For each mixing speed, two curves are shown in Figure 5. The thin line is the normalized
variance of the data collected at each mixing speed. The thick dashed line is the
normalized variance of the data with bubbles minus the normalized variance of the data
without bubbles, where it is assumed that the process leading to the variance in the nobubble data is uncorrelated to the variance caused by the presence of bubbles. The
variance in the no-bubble data is typically much less then the variance in the bubble data,
giving a good indication that the observed fluctuations in the amplitude data are due to the
bubbles themselves.

The data in Figure 5 can be compared to the normalized variance that would be
predicted from the bubble populations shown in Figure 4 using (25). To do this, the
predicted incident pressure at some field point r must include the effect of the projector
beam-pattern, PI, and the Green's function describing the propagation from r to the
hydrophone must be weighted by the hydrophone beam-pattern, P 2 . Thus, the volume
integral is restricted to the overlapping region describing the beam-patterns of the source
and receiver. Using the geometry defined in Figure 6, the variance can be written

where

--I-:

./ $

...i

,\...
'\,

.......................

1 --. p

~I --.

..........,.
j --'

4cds
1 2c cd ds s

"..................

60

80

100

120
140
160
frequency (Wz)

180

200

22r

Figure 5. Normalized variance (thin line is normalized variance of bubble data, thick dashed line is normalized
variance of bubble data minus normalized variance of no-bubble data at the same mixing speed).

460

dV--drpypz,

....

........

..........*..,I.....

%..

"........9
.........................
I

........

..............

:<-- - - _ _ - - - -_"-- - - I

r ......................................... :>@)

>;

Figure 6. Measurement geometry showing the relevant distances and angles of an differential scattering
volume.

The squared mean pressure in (29) accounts for the hydrophone beam-pattern, and can be
written

where a describes the attenuation due to bubbles. Equation (30) is solved numerically,
normalized by the squared mean pressure observed at r, and then compared to the
measured normalized variance in Figure 7. Note again that it was assumed that the
average bubble population was not changing over the ensonfied area, allowing the
quantity H to be removed from the integral. Both H and a are estimated from the
inversion of the measured attenuation data.
The acoustic observations are not wildly varying - the largest standard deviation is
less then 5% - but the agreement between the prediction for the variance and the data is
good. The error bars at each data point in Figure 7 describe a 95% confidence interval for
the measured variance. The prediction is generally within this confidence interval,
lending support to the derived solution, although the measurement is consistently higher
then the prediction at the higher frequencies. The source of this apparent bias is not
known. It might be expected that the match between data and prediction would worsen at
the higher frequencies due to an unaccounted for misalignment of projector and receiver,
thereby reducing the ensonified volume in the experiment, but this would result in a
prediction that would be higher then that observed rather then lower. The small size of
the tank used for this experiment increases the risk of unwanted scattering from the mixer
and tank walls, but this effect should be more problematic at the lower frequencies. It is
perhaps more likely that the number of small bubbles was incorrectly estimated in the
inversion process due to the lack of data at and above the resonance frequency associated
with peak of the bubble distribution (12.5 p).

46 1
x

lo4

40

60

80

100

120
140
160
frequency (kHz)

180

200

220

240

Figum 7. Comparison between the predicted normalized variance and the measured normalized variance for
each of the three mixing conditions.

Mixing time

The multiple scattering theory describing the statistics of the observed pressure amplitude
requires no knowledge of the flow in which the bubbles reside. It only requires
knowledge of the average bubble density, and inherent in the comparison of data to theory
is the assumption that an adequate number of bubble configurations have been observed.
It might be expected, however, that the length of time required for an adequate number of
bubble configurations to be observed is a flow dependent parameter. From the standpoint
of gaining insight into the link between the multiple scattering theory and the flow it
would be desirable to have a metric describing how long it takes for the observations to
adequately explore the PDF describing the statistics of acoustic pressure amplitude.
One possible metric, which is the one that will be used here, is the chi-square
goodness-of-fit test. This test determines whether an observed probability density
function agrees with an expected probability density function at some specified
confidence level [ 11. No precision is specified in this test, it either passes or fails at the
given confidence level. The metric is found as follows. The entire data set (at a
particular frequency and mixing condition) of independent samples is used to establish an
expected probability density function, which is considered to be normally distributed with
the sample mean and sample variance. A chi-square goodness-of-fit test is then
performed for, say, the first 50 independent samples, and the total discrepancy is
compared to the chi-square value. This process is repeated for 51 samples, then 52
samples, and so on. For each test, the degrees of freedom is fixed (i.e. the same number
of histogram bins are used). The metric describing the mixing time for a particular flow

462
condition is defined as the number of samples required to pass the chi-square goodnessof-fit test multiplied by the correlation times previously discussed.

To improve the overall result, the assumption is made that for a particular mixing
condition, the chi-square goodness-of-fit test converges on average in a manner
independent of the frequency of the acoustic signal. Accordingly, this process is done for
each frequency for a given mixing condition, and the results are then averaged. The
results for the three mixing conditions are shown in Figure 8. The number of degrees of
freedom was identical for the three mixing speeds, and corresponding chi-square value
was 6.25. Although somewhat noisy, the plots in Figure 8 give a rough estimate for the
number of samples required to pass the goodness-of-fit test. In order of increasing mixing
speed, the required number of samples are 40, 40, and 35. Considering the large
confidence interval shown in Figure 8, these estimates are approximately the same.
Multiplying this result by the corresponding correlation length for each mixing speed
gives approximate mixing times of 13, 12, and 8 seconds, again in order of increasing
mixing speed. The differences between the three mixing times is based almost entirely on
the difference in correlation time.

1816.

16-

fi,
,I

14-

i !,
.,, ...

12-

12:

j ;,

- 0
--I

# samples

14.

20

40 60 80
#samples

Figure 8. 'Ihe total squared discrepancy in the chi-squared goodness-of-fit test for the three mixing conditions,
with mixing velocities increasing from left to right. The hypothesis that the sample probability density function
matches the expected probability density function is true for a total squared discrepancy less then 6.25. The
dotted lines describe the standard deviation in the estimate of total squared discrepancy.

Discussion

The multiple scattering theory describing propagation through bubbles requires no

knowledge of the characteristics of the flow in which the bubbles exist, provided that
enough independent samples are collected in order for the statistics to converge (which
requires a minimum sampling duration greater then the so-called mixing time). At the
same time, bubbles do not exist independent of the flow in which they reside, and so a

463
relationship between the flow and the acoustic observations is expected. There are
different spatial and temporal scales on which this relationship can be explored. Thorpe,
for instance, explored the link between turbulent diffusion and acoustic backscatter
observations of a diffusing bubble concentration [7]. This does not necessarily require
the establishment of any link between the multiple scattering theory and the flow,
however, since the evolving backscatter observations can be viewed as a diffusion of the
average bubble concentrations. In this paper, two possible flow dependent time scales
associated with propagation through homogenized bubble distributions were explored: a
correlation time and a mixing time. In this particular experiment, it was found that the
difference in mixing time between the three conditions explored was due only to the
difference in correlation time describing the fluctuation acoustic observations, which was
found to decrease with increasing flow intensity. These are relatively tenuous
relationships between the flow and the acoustic observations, however, and are not
represented specifically in the multiple scattering theory (but are inherently included in
the statistical derivation). A more demanding relationship between the flow and the
acoustic observations might be expected for spatial patchiness in the instantaneous
bubble concentration, even if the average bubble concentration is independent of space. It
is expected that an instantaneous and localized high (or low) concentration of bubbles
would exhibit a higher variance then a homogenous mixture with the same concentration,
although the multiple scattering theory would predict the same result for both cases
because it is only a function of the average bubble concentration. Evidently, the
experiments described in this paper were conducted with relatively homogenous bubble
clouds. Future work will be directed towards the problem of patchiness.

Acknowledgements

This work has been supported by ONR Code 321US. The first author is supported by a
National Defense Science and Engineering Graduate Fellowship.

References
1. Bendat, J.S. and A. G. Piersol, Random Data Analysis and Measurement Procedures,
3rdEd., John Wiley & Sons, Inc., 2000.
2. Caruthers, J. W., P. A. Elmore, J.C. Novarini, and R. R. Goodman, An iterative
approach for approximating bubble distributions from attenuation measurements. J.
Acoust. SOC.Am. 105 (1999) pp. 185-189.
3. Carstensen, E. L., and L. L. Foldy, Propagation through a liquid containing bubbles.
J. Acoust. SOC.A m 19 (1947) p ~ 481-501.
.
4. Commander, K. W. and A. Prospenetti, Linear pressure waves in bubbly liquids:
Comparison between theory and experiments. J. Acoust. SOC.Am. 85 (1989) pp. 732746.
5. Foldy, L. L., The multiple scattering of waves, I. General theory of isotropic
scattering by randomly distributed scatterers. Phys. Rev. 67 (1945) pp. 107-119.
6. Medwin, H. and C. S . Clay, Fundamentals of Acoustical Oceanography, Academic
Press, 1998.
7. Thorpe, S. A., On the determination of K , in the near-surface ocean from acoustic
measurements of bubbles. J. Phys. Ocean. 14 (1984) pp. 855-863.

464

MULTI-AGENTS SYSTEMS APPROACH FOR UNDERWATER SOUND

EQUATIONS SOLUTION

VICTOR A. ZARAISKY, VASILY V. POPOVICH, RAFAEL M. YUSUPOV.

St. Petersburg Institute for Informatics and Automation of the Russian Academy of Sciences
(SPIIRAS) 39, 14 Linia, VO St. Petersburg, Russia 199178
Tel. +7 812 3288071
e-mail: povovicli @mad.iias.spb.su
Problems of underwater acoustic modeling using multi-agents systems approach are considered in
the paper. It is well known that calculation and modeling of an acoustic field are rather difficult
problems accounting for a lot of factors, and some of them are not numerical values or variables of
binary logic of yes or no type; first and foremost the above factors are region type and
parameters (e.g., deep sea, shallow water and other), some parameters of a sound source and sound
receiver, a set of available computer and analytical models etc. At that as a rule the data
incompleteness as well as their fast variability in time and space are specific for the considered
problems.
In many cases the use of Conventional programming approaches, like object-oriented approach, leads
to the development of numerous rather complex computer models for the underwater sound area, and
their common disadvantage is that these products could be used only by the developers.
The present paper considers an approach to the acoustic field modeling based on decomposition of
underwater sound equations into a set of intelligent constituents - agents. The proposed approach
includes the following stages:
development of an ontology of the subject area;
decomposition of underwater sound equations into a set of intelligent agents;
development of a structure of each intelligent agent and its interface;
development of a system of interaction between agents;
development of underwater sound equations computer model based on a multi agent distributed
system.

1
Introduction
In a popular research publication on underwater acoustic [ 13 the principles of underwater
sound from the viewpoint of the engineer and the practical scientist were summarized. It
was denoted in the material of the book that this area of investigation ...lies squarely in
the middle of the spectrum-between theory at one end and sonar technology at the other

Ill.
The framework of the current paper is the computational sonar equations
development for computer practical use. An idea is rather close to that described in [I].
The main problems are to collect all approaches and equations, to make links between
different equations and account for properties of the medium. The medium properties
influence upon the selection of analytical approaches and computable algorithms.
Moreover, the medium properties are changing in 3D space and in time. So, it is quite
difficult to build a good collection of computer sound equations models using
conventional techniques like the object-oriented approach.
The goal of our paper is a multi-agents approach development for underwater sound
equations modeling on PC. The first step of the approach is ontology of the subject area
development.
2
Development of an ontology of the subject area
Ontology of the subject area is the frame for integrity, consistency and interaction
between distributed intelligent agents. In other words ontology is a specification of the

465

objects, concepts, and relationships in the area of subject domain. First of all main entities
of subject domain should be selected. In general there is a list of the main entities as
follows:
- surface of the water;
- medium (sound channel in the sea);
- bottom;
- sonar system;
target;
- mathematical approaches (equations).
The upper level of the entities interaction is shown in Fig. 1.

Sonar
System

Medium (sound channel in the sea)

Target

Figure 1.Upper level of the entities interaction.

Let us examine the above selected entities in detail.

The Sea and its boundaries (surface, bottom), form a remarkably complex medium for the
propagation of sound, see Fig.2.
Parameter of transmission losses express the magnitude of one phenomenon of many
associated with sound propagation in the sea. It has the following sources:
1. spreading loss (a geometrical effect representing the regular weakening of
sound signal as it spreads outward from the source);
2. attenuation loss (includes some effects as follows: absorption, scattering,
diffraction and refraction);
3. sound scattering layers;
4. boundaries.
An example of velocity distribution is shown in Fig.3.
It is difficult to develop a useful equations or experimental data for transmission losses
(TL) for a common case. As a rule, it is necessary to consider at least the following three
phenomena.
A. Shallow water. For this case the following parameters should be taken into
account.
1.TL for three ranges. Range can be calculated through Marshall and Shulkin
parameter:

466
1

H = 0.58(h + h, )T , where:
h- depth of the sea;
k, thickness of isothermal laver.

Figure 2. An example of medium complexity for the Arctic Ocean

There exist the following three cases:

1) r l H ;
2 ) H l r l H ;
3) r 1 8 H .
2.Distribution of sound velocity as a function of depth.
3.Depth of sonar system.
4.Depth of the target.
5.Type of bottom.
6.Depth of the sea.
7.Surface of bottom type.
B. Deep sea. For this case the following parameters should be accounted for.
1.Upward sound channel.
2.Negative gradient of sound velocity.
3.Deep sound channel.
4.Caustics and Convergence Zones.
C. Variable channel along the range of propagation. For this case the following
parameters should be taken into account.
1.Propagation range should be not shorter than 1000 km.
2.Coastal wedge.
3.Front zones.

467

Range, km
Figure 3. An example of velocity distribution.

Sonar System. It is undoubtedly a complex system, and let us consider only

external things affecting SS. The most important of them are:
1. Ebbs, flows and rough sea.
2. Seismic processes.
3. Turbulences in ocean.
4. Waves on the surface.
5. Noise in shallow water.
Surface of the sea. It can be divided in two groups: smooth and not smooth.
Target. We do not account for the target acoustic properties. In common case some
groups of frequency have a meaning for this case as follows.
1 .Ultrasound.
2.Sound.
3.Infrasound.
4.Seismic.
Mathematical methods. The selected methods for acoustics fields calculation are
given below.
I. Wave Theory.
The following calculation (accurate) methods have been selected:
- DAlembers method;
- Fouriers method.

468
The following calculation (approximate) methods have been selected: Ventcel - Kramer
-Brilluen.
11. Ray Theory.
- Focus factor;
- Averaged law of loss;
- Parabolic equations;
- Ray invariant, adiabatic approximation.
The paper limits only allow to show the upper level of the problem. The authors did their
best to consider under this angle the complex interaction between above selected entities.
A user case diagram is shown in Fig. 4.
Indeed, for most of practical cases a traditional approach (centralized method of acoustic
field model development) is a way to the impasse. It is possible to select certain reasons
for that as follows:
- Difficulties in a development of such a complex centralized algorithm when
so many properties should be accounted for in accordance with the applied
field;
- Complexity of support and change of centralized algorithms;
- Calculation models should be dynamically rebuild depending on a lot of
factors such as: end user, kind of receiver and source of sound, goals of
modeling, etc.;
- Limitation of computer recourses for some kind of algorithms;
- Some data for special algorithms can be not available for the end user;
- Difficulty in having a complete input data for the end user (vertical and
horizontal profiles of sound velocity, bottom etc. for some region of sea).
By the way, complexity exists not only as a fact that some data are not
available but also as a limitation of computer resources for storage,
restoration and statistical processing of data.
3
Decomposition of underwater sound equations into a set of intelligent agents
Following Urick R.J [ 11 the active and passive-sonar equations can be given as:
The active-sonar equation: SL-2TL+TS=RL+DT;
The passive-sonar equation: SL-TL=NL-DI+DT,
(1)
where:
- SL - Projector Source Level;
- TL - Transmission Loss;
NL - Self-Noise Level;
- DI - Receiving Directivity Index;
- DT - Detection Threshold.
The object-oriented analysis of (1) has shown that first of all it is needed to make an
object-oriented decomposition of the applied field to the base class. Main classes of
passive sonar equation (1) are:
Sonar System, Source of Sound and Medium. From this point of view parts of equation
(1) can be collected as follows.
1.Equipment (Sonar System):
- Projector Source Level (SL);
- Self-Noise Level (NL);
- Receiving Directivity Index (DI);
Detection Threshold (DT).

469
- ______

0 b s e rve r,

---.

I Targets

- _ _
Deep S e a,,
Shallow water,

c
-

1
I

j...

,/

R a y Theory,
W a v e Theory,

Sound
P ro pa g a t io n ,
i
- - *Range
..
.
I
Estimation,
Applied t a s k s
...
..
/

---

User

Figure 4. User case diagram.

2.Medium:
- Transmission Loss (TL);
Reverberation Level (RL);
Ambient-Noise Level (NL).
3.Target:
Target Strength (TS);
- Target Source Level (SL).
A set of classes as a Class Diagram is presented in Fig. 3. Additional properties can help
to convert each class into Intelligent Agent.

470

Figure 5. Classes diagram

Development of a structure for every intelligent agent and its interface

Currently we have a set of Agent term definitions. Let us discuss a weak definition of
the Agent 131. In comparison with the object-oriented approach let us add one additional
property pro-activity to the term class. Pro-activity is a possibility for the Agent to
make initiative decisions, and there exists a principal difference between Class and
Agent. The latest realization of Unified Modeling Language (from UML 2.0) provides for
a specification of the active class. From this point of view the difference between Class
and Agent is rather relative.
From our point of view the principal difference between Agent and Class structure is an
availability of two special methods for the Agent. The first one is Interaction - a
possibility to interplay with other Agents in local and global computer area. The
Interaction can be initiated as a self generation and as a request from other Agents. The
second one is InlImplementation - knowledge system realization according to the expert
systems methods. As a rule it is possible to correct knowledge of Agent using special
rules that are very close to one of natural languages, say, Russian or English. It is possible
to correct knowledge base without recompilation of the computer project. End user or
expert can easily change knowledge base. This feature is useful for real experiments in
the sea. Well done multi-agents system allows to link new agents or replace some other
agents in the work process. Such kind of technology is already developed for distributed
systems like DCOM and CORBA. The main idea of these technologies is an Interface
idea. Interface is an external representation of Agent for interaction with computer area.
Let us distinguish two types of Interface - system and visual. State1Activity diagram of
Agent System Interface is shown in Fig.6.; the State/Activity diagram of Agent Visual
Interface is given in Fig.7.

47 1

Figure 6.System Interface of agent.

'

Request of
U s e r ...

&. '

Run process

. -

..

Figure 7.Visual Interface of agent.

5
Development of a system for interaction between agents
A basis for the Agents' interaction system is a behavior of each Agent. Each Agent
interacts with a defined set of Agents. To support a flexible behavior of Agent the use of
logic deduction system could be a good solution. For this it is possible to use a machine
logic deduction, for example, CLIPS (htto://www.gh~.net/clius/CLIPS.html).
If CLIPS is
selected it is easy to change behavior rules during the work process without a complete or
partial recompilation of computer project. Moreover, it is possible to write rules in
natural human language.
Fig. 8 presents a model of behavior of Target Source Level (SL) Agent. The behavior
depends on Target's changes.

472
Target% Deep

Target's Velocity

Targetb SL

Figure 8. Model of SL Agent behavior.

Behavior of more complex Agent is shown in Fig.9.

Figure 9. Model of behavior of Transmission Loss Agent.

6
Development of underwater sound equations computer model based on a
multi agent distributed system
A model of underwater sound equation is shown graphically in FiglO.

Figure 10. Underwater sound equation model.

Every part of the model should be realized as an Agent; it also should be noted that Fig.
10 shows are classes of Agents and not Agents, since as it was mentioned above each
Agent depends on different conditions as follows:
- kinds of Sonar System and Target;
- type of Medium;
available mathematical methods and their computer realization.
For the above described case one can see the main advantage of multi-agents approach:
automatically created chain of Agents depends on a set of conditions in defined time and
in defined point of the sea.
In Fig.11 interaction diagram of selected (auto selected) Agents is shown. If the
above selected condition changed a matching set of such kind diagrams would be
developed.

473
Preliminary consequences of the computer development of underwater sound
equations as a multi-agents system.
1. Developer tries to develop a set of Agents that can form a computer model of
underwater sound equations rather than the model itself.
2. Ideal multi-agent system exists without a general Agent or direct control of
developer or end user.
3. Agents should interact with each other and can make decisions independently,
according to knowledge base (rules base).
4. Multi-agent system is not a big computer model like as flock of birds is not a big
bird, and a traffic jam is not a set of cars.

I
I

I
I

I
I

I
I

I
I

I
I

Figure 11. Interaction diagram of Agents.

Conclusion
In the process of developing a multi-agent system approach for underwater acoustic
application some general comments were generated. Multi-agents technology can be
successful when it is needed to simplify modeling problems in underwater acoustic. It is
possible to join different research directions within the frame of computer model of
acoustic field. Benefits for end user are evident. It is also easy to combine different
computer models during computer experiments. Similarly to computer technologies like
those supported by, e.g., Microsoft Word, the computer model of acoustic field can be
created combining different computer applications including graphical editors, electronic
worksheets etc. It is possible to automatically collect a chain of Agents.

474
Multi-agents system technology is a way to the higher level technology - GRID
computing. Certain reasons exist why to use the proposed approach, and one the main
motivations is that it can accumulate significant resources required for some problems
modeling when the need for storing and manipulating multi-dimensional input data raises
as well as providing for a considerable computational resource for some mathematical
approaches like Wave theory, and other.
8
Acknowledgements
Authors extend their appreciation to ONRIFO for funding the participation in ICTCA
2003.

References
1. Robert J. Urick Principles of underwater sound for engineers. - 1975, 1967 by
McGraw-Hill, Inc. 384pp.
2. Michael Wooldridge Intelligent Agents //Multiagent Systems: a modern approach to
distributed artifical intelligence/ edited by Gerhard Weiss. - 1999 Massachusetts
Institute of Technology. Pp.27-79.
3. FIPA 2000 Specifications. Geneva, Switzerland, Foundation for Intelligent Physical
Agents, 2002. http://www.fiipa.org/rcuositorv/lipa2OOO.h~ml.
4. K. V. Avilov, Pseudodifferential Parabolic Equations of Sound Propagation in the
Slowly Range-Dependent Ocean and Their Numerical Solutions, Akust. Zh., Vol. 41,
no. 1, 1995, pp. 5-12.
5 . K.V. Avilov and O.E. Popov, The Calculation of the Signal from a Broadband Point
Source Arbitrarily Moving in the Range-Dependent Ocean, in Ocean Acoustics,
Proceedings of the L.M. Brekhovskikh Symposium, Moscow, GEOS, 1998.
httu://www.module.ru/.

475

The modal-vertical-beam(MVB) transmission loss analysis

H.F.Zhao, H.L.Ge, X.Y.Gong
(State Key Lab. Of Oceanic Acoustics, Hangzhou, 3 10012)
E. C. Shang*
(Scripps Institution of Oceanography, La Jolla, California, 92093-070 1)
Abstract
The group-averaged modal characters (attenuation coefficient, back-scattering
coefficient, ...) are important for higher frequency normal mode field modeling. In
this paper, the method of extracting the group-averaged modal attenuation coefficient
from the Modal-Vertical-Beam (MVB) analysis is discussed. The beamformer output
spectrum is consisted of two parts: one is the modal energy weighted sum and
another is the intermode interference sum. If the intermode term can be ignored,
group-averaged-modal attenuation coefficient can be obtained from the beamformer
output at two different ranges. Numerical simulations have been conducted to
investigate the effective approaches for reducing the intermode term. Transmission
loss data collected by a 57m VLA during the 200 1 ASIAEX is also performed.
* Now, Guest professor of the State Key Lab. of Ocean Acoustics
1 Introduction
Under the WKB approximation, the mode depth function at the receiver location
can be written as a sum of upgoing and downgoing plane waves. A vertical plane
wave beamformer acts as a spatial bandpass filter, passing only those vertical wave
numbers in some band, corresponding to a given group of adjacent modes. The
beamformer output spectrum is consisted of two parts: one is a sum of weighted
square of normal modes, another is the intermode sum. If the intermode sum can be
ignored, the beamformer output spectrum of two different ranges can be used to
calculate propagation loss, and furthermore extract beam-averaged modal attenuation
coefficient. The beam-averaged modal attenuation coefficient is important for bottom
back-scattering matrix extraction from reverberation data[ 11. In this paper, we
propose a method to extract beam-averaged modal transmission loss from the
propagation data.
2 TheoryFrame
At ranges greater than several water depths, the underwater acoustic pressure
field from a single omnidirectional point source can be expressed as a finite sum of
normal modes. Then the complex pressure in a range-independent waveguide can be

476

written as:

n =O

where, a, (r) =

d~

27T

, r is range, z, are depth of point receivers, zs is depth of

source, w is circular frequency of single tone, k, are modal horizontal

wavenumbers, <pn(z) are mode depth functions, and 6, are modal attenuation
coefficients. Under the WKB approximation, the mode depth function at the receiver
location can be written as a sum of upgoing and downgoing plane waves. A vertical
plane wave beamformer acts as a spatial bandpass filter, passing only those vertical
wave numbers in the band of

7 L Ay / 2 , corresponding to a given group of adjacent

modes Ek Am / 2 . For an ideal spatial filter response, the beamformer output can
be written asf2]:
-

m+hl2

b(r,y,z,,a) = - ~a,(r)p,(z,)exp[-i(ot - k n r + n / 4 ) - 6 , r ]

(2)

m-AmI2

The beamformer output spectrum is:

B(r, Y, ,2,

4 = x a : (rh:(2,) eW-26,r)
n

m
n#m

= B, +B,

the summation in Equation (3) are done over those mode in the Am band centered
on E. Akn,is the difference in mode wave numbers. The beamformer output
spectrum is consisted of two parts: one (B1) is a sum of weighed square of normal
modes, the other (B2) is the intermode sum. If the intermode sum can be ignored, the
beamformer output spectrum is written as:
n

We define the beam-averaged modal attenuation as follow:

477

Then beam-averaged modal attenuation coefficient 6, of E group can be

extracted by calculating the propagation loss for two ranges:

6,

log(tl(r, r2 1
2G2 -r1)
'
Y

3 Numerical simulation
We use the KRAKEN normal mode code to calculate the mode fieldc3].We
consider a Pekeris waveguide with water depth H=103m, CO=1519m/s, Cb=1623m/s,
P b=1.72, a b=0.49dB/km Hz, f=1390Hz. The 103m VLA is covered all depth of
water column with lm space of adjacent element. We have 65 modes, and the critical
grazing angle is 20.6" . We divide modes into 4 groups, the corresponding center
grazing angle, modes in every group, theoretical value of 6 ,and calculated value of
6 are shown in Table 1. From Table 1, we can find 6 ,of group 1 can not be
estimated correctly. This is due to the effect of intermode interference. Figure 1 and
Figure 2 show the effects of intermode interference of single tone 1390Hz, solid line
is the value B2/B1 vs. range. From Fig.1 and Fig.2, effects of intermode interference
can not be ignored, then 6 calculated by Eq. (5) is error. How can we reduce the
effects of intermode interference? One method is averaged the beamformer output
spectrum with f i n a narrow band to smooth the second item in Eq.(3).
Table 1

Theoretical value and calculated value of 6, in Pekeris waveguide

(f=1390Hz)
Group
of
modes

Theory value

group

Center
grazing
angle

1
2
3
4

2.3
6.8"
11.5"
16.2"

1-8
9-16
17-24
25-32

3.367554e-6
2.051041e-5
5.064872e-5
9.020034e-5

of 6,

Calculated value of 6,
r 1= 1 Okm,
r I= 1Okm,
r2=20km,
r2=20km,
zs=20m
zs=40m
-5.030152e-6
3.158571 e-6
1S07090e-5
3.249568e-5
3.641855e-5
5.89848 le-5
9.085626e-5
8.014695e-5

478

Table 2

Theory value and calculated value of 6 , in Pekeris waveguide

(f=l340- 1440Hz)

Fig. 1

Fig.2 Effect of intermode interference

Effect of intermode interference

(f=1390Hz, zs=40m)

-041
05

'

'
15

'

'
'
25
3
ralcle(m)

(f=1390Hz, zs=20m)

'
35

'
4

'

I t

J o
5

x lo4

rWea(m)

104

Figure 3 and Figure 4 show the effects of intermode interference of narrow band
1340-1440Hz. The value B2/B1 is reduced effectively. Table 2 gives the calculated

479
value of 6 ,calculated with narrow band 1340-1440Hz. 6 ,of 4 groups are extracted
correctly for all three cases.
Table 3

Theoretical value and calculated value of 6, in simulated shallow

Center
grazing
angle

Group
of
modes

Theory value

group

1
2
3
4

2.3"
6.8"
11.5"
16.2"

1-8
9- 16
17-24
25-32

6.01 1310e-6
2.272945e-5
3.60 1585e-5
8.115852e-5

of 6,

Calculated value of 6,
r 1=1O h ,
r l = 1Okm,
r2=20km,
r2=20km,
zs=40m,
zs=4Om,
103mVLA
57mVLA
2.193981e-5
1.630264e-5
3.075 102e-5
4.356382e-5
5.283 101e-5
6.074829e-5
8.317109e-5
1.071509e-4

sound speed(rris)

Fig.5 SVP of ECS

Figure 5 draws the SVP of ECS in June, 6, 2001. Input the SVP and parameters
of sediment described in section 3, we can acquire the simulated 103m VLA data and
57m VLA data. The 57m VLA covers from 3m to 60m of underwater. The extracted
6 of this two VLA are displayed in label 3. All the calculated 6 are larger than
the theoretical value. This is due to the short VLA covered the upper half part of the
water column is not able to capture the lower modes properly and effected by higher
modes by leakage.
4 ECS Real data analysis
The 57m VLA data are acquired from ECS acoustical experiment which was

480
conducted from June 2 to 7,2001 in the site centered at E 126O49.02' and N 29O40.52' .
The suspended 60-element, 57m VLA covers from 3m to 60m. The array's tilt angle
is measured by the depthhilt sensors for array shape and tilt angle estimations, and
signals received by VLA and measured sensors are simultaneously recorded by the
multichannel recording system.
The projecting array with the different depths deployed from transmitting vessel,
which moves to the positions being 5, 10, 20 and 40km away from the receiving
vessel, radiates signals with different forms, for example, CW, PCW, LFM, and PRN.
In the paper only signals of 1100-2000Hz PRN and at range 5km and lOkm , depth
20m and 40m are used to estimate beam-averaged 6 ,.
1

1-

09

09-

08

08-

07

a7 -

06

06 n

05

%05-

04

04.

03

O0 l3 -

02
01.

01

-100

-80

-60

-10

-m

20

40

60

en

im

.Pm
mde(deg6-31

amwe(degee~

Fig.6 Vertical beam pattern with

Fig.7 Vertical beam pattern with

20m source depth

40m source depth

The vertical beam pattern of VLA with narrow band 1340- 1440Hz at 5km range
(solid line) and lOkm range (dash line) is given in Figure 6 and Figure 7. The source
depth is 20m in Figure 6, and 40m in Figure 7. The vertical beamforming performed
with 10s long data, and beam-averaged 6 ,are the average of 20 snapshots. The
estimated 6 are shown in Table 4.
The estimated value of 6, in real shallow waveguide (f=1340-1440Hz)

Table 4

group

1
2
3
4

Center
grazing
angle
2.3
6.8"
11.5"
16.2"

Group of
modes

1-8
9-16
17-24
25-32

Estimated value of 6,
rl=5km,r2=1Okm,
rl=5km,r2=1Okm,
zs=20m
zs=40m
1.195583e-4
2.785 835e-5
2.4472 15e-4
1.224623e-4
1.853596e-4
1.175435e-4
1.934406e-4
2.953209e-4

48 1

Summary
(1) If the intermode sum can be ignored, the beamformer output spectrum of
two different ranges can be used to extract beam-averaged modal attenuation
coefficient.
(2) When vertical beamformer is performed in a narrow band not in single tone,
the intermode sum can be reduced effectively, as proved by the numerical examples.
Reference
[ 13 E.C.Shang, T.F.Gao and D.J.Tang, Extraction of Modal Back-Scattering Matrix
from Reverberation Data in Shallow-water Waveguide, 6'h ECUA , June 2002

[2] G.L.D.Spain, W.A.Kuperman and J.J.Murray, Matchless Field Processing in

Shallow Water, 2000 IEEE
[3] M.B.Porter, The Kraken Normal Mode Program, AD-A252409