BACKSCATTER

PYOTR YA. UFIMTSEV

University of California at
Irvine
Irvine, California
The scattering of waves from an object is a diffraction
process that transforms the incident wave into the waves
propagating from the object in all possible directions. The
backscatter, or backscattering, is the scattering of waves
back toward the source of the incident wave. This process
substantially depends on material properties of the scat-
tering object, its shape, size, and spatial orientation
relative to the incident wave, as well as on the frequency
and polarization of the incident wave. One distinguishes
three frequency regions with different physical properties
of scattered waves. They are quasistatic, resonance, and
quasioptical regions.
1. QUASISTATIC SCATTERING
In the quasistatic region (sometimes called the Rayleigh
region), the wavelength l of the incident wave is much
greater than the maximum linear dimension l of the
scattering object (lbl). At a certain time t, the scattered
eld at small distances (r5l) from the object is approxi-
mately the static eld created by dipoles and multipoles
induced by the incident wave in the scattering object at
the same time t. Far from the object, the scattered eld is
an outgoing spherical wave. Its average power ux density
(over one period of oscillations) relative to that of the
incident wave, or radar cross section (RCS), is determined
by the objects volume rather than the shape of the object,
and it is proportional to l
4
. Specically, this dependence
explains the blue color of the cloudless sky during the day.
This color is due to the predominance of blue light
scattered by small inhomogeneities of the air caused by
uctuations of its density. The violet light, which is
shorter in wavelength, undergoes stronger relative scat-
tering and as a result suffers higher attenuation while
propagating through the atmosphere. Analytical expres-
sions for quasistatic RCS of some scattering objects can be
found in Refs. 13. For example, the axial RCS of perfectly
conducting bodies of revolution is determined by the
following approximation (p. 145 of Ref. 1)
s %
4
p
k
4
V
2
.
1
e
t
pt

2
1
where k2p/l is the wavenumber, V is the objects
volume, and t is the characteristic length-to-width ratio
of the object. The quantity t is found for each objects
shape by the allowing the axial dimension of the object to
go to zero so as to obtain the correct result
s
64
9p
k
4
a
6
2
for the circular disk with radius a. Table 8.2 in Ref. 2
(Vol. 2, pp. 558561) contains explicit expressions for RCS
found in this manner for a variety of bodies of revolution.
The rst term in this table (Eq. 8.1-87a on p. 558) contains
a misprint. The letter b should be replaced by h.
In this frequency region, the scattered eld can be
expressed in terms of a convergent series in positive
integer power of the wavenumber k 2p/l. The expansion
coefcients are found from the solution of the recursive
system of boundary value problems in potential theory
(pp. 848856 of Ref. 4). In practice it is possible to nd only
a few rst coefcients. Direct numerical methods are
efcient tools for the solution of quasistatic scattering
problems [57].
2. RESONANCE SCATTERING
In the resonance frequency region, linear dimensions of
scattering objects are comparable to the wavelength of the
incident wave. Eigenoscillations excited by the incident
wave in the scattering object can substantially inuence
the scattering properties. Frequencies of these oscillations
are complex quantities. Their imaginary parts determine
both the internal (thermal) losses inside the object and the
external losses that are due to radiation into the sur-
rounding medium. A major contribution to RCS is given by
the radiation of those eigenoscillations, whose eigenfre-
quency real part and polarization are close to the incident
wave frequency and polarization. If the quality factor of
these oscillations is quite large, the amplitude and the
intensity of the scattered wave sharply increase as the
frequency of the incident wave approaches the real part of
the frequency of eigenoscillations. This phenomenon is
referred to as resonance scattering. It reveals itself, for
example, in scatterings from thin metallic half-wave-
length long wires and narrow strips (pp. 293303 of
Ref. 2). These scatterers are used, in particular, to create
a chaff clutter for radars. Analytical and numerical data
for scattering by thin wires are collected in Chap. 12 of
Ref. 3. Recent analytical results for resistive wires are
presented in Ref. 8.
The resonance scattering from wires can also be ex-
plained as being due to the constructive interference of
multiple current waves arising from the wire ends. This
process is investigated in detail in Refs. 9, 10, and 30. The
total current generated by the incident plane wave in the
perfectly conducting wire (Fig. 1) equals
Jz J
0
z
X
1
n1
J

n
z J

n
z 3
B
481
with the time dependence exp(iot) assumed and sup-
pressed here and later. The term J
0
(z) is the current
induced by the incident wave in the innitely long wire
(NrzrN). The terms J

n
z and J

n
z are multiple
current waves. Waves J

n
z run in the positive z-direction
from the left wire end z l to the right end z l.
Waves J

n
z run in the negative z direction from the right
end z l to the left end z l. The total length of the
wire is L2l. When the wave J

n
z reaches the opposite
end it undergoes diffraction and transforms into the wave
J

n1
z. At the end points of the wire the total current and
its components satisfy the conditions
Jl 0; J

1
l J
0
l; J

1
l J
0
l 4
J

n1
l J

n
l; J

n1
l J

n
l; n1; 2; 3; . . .
5
In thin wires (kao0.2, a is the wire radius), the multiple
current waves are described by the following approximate
expressions
J

2n
z J

1
lckLe
ikL

2n2
ckl ze
ikl z
J

2n1
z J

1
lckLe
ikL

2n1
ckl ze
ikl z
J

2n
z J

1
lckLe
ikL

2n2
ckl ze
iklz
J

2n1
z J

1
lckLe
ikL

2n1
ckl ze
iklz
6
with n1,2,3,y. Function c(kz) is dened in Refs. 10 and
30 as
ckz
2 ln
i
gka
ln
2ikz
gq
E2kze
2ikz
7
where
q ka
2
; g 1:781. . . ; lni ip=2
and
Ex
Z
1
x
e
it
t
dt Cix iSix 8
Functions Ci(x) and Si(x) are the well-tabulated cosine and
sine integrals, respectively. For small arguments (x51),
function E(x) reduces to E(x) ln(gx) ln(i) O(x) and
ensures the equality c(0) 1. Equation (6) shows that all
multiple edge waves starting with secondary waves
(n2,3,4,y) are expressed approximately by the same
function c(x). As a result, the substitution of expressions
(6) into Eq. (3) leads to the geometric series
X
1
n2
J

n
z J

n
z f k; z; l; a
X
1
m0
fckLe
ikL

2
g
m

f k; z; l; a
D
9
which contains the resonance denominator D(ka, kL)
1[c(kL)e
ikL
]
2
. The equation D(ka, kL) 0 denes the
complex resonant frequencies o
res
ck
res
o
0
res
io
00
res
,
where c is the light velocity in vacuum. Due to the
radiation loss, the quantity o
00
res
Imo is always negative
o
00
res
o0. Therefore, for real frequencies (o
00
0), the de-
nominator D(ka, kL) does not vanish. But it acquires
minimum values when the frequency of the incident
wave is close to the real part of the resonant, frequency
o % o
0
res
. This occurs when kLEnp or LEnl/2 with
n1,2,3,y and results in the current resonance. Under
the normal incidence (the direction of the incident wave is
z = l z = + l
z
Figure 1. A thin wire excited by an incident wave. The radius of
the wire is small compared with the wavelength. Such a wire can
support traveling waves arising due to the multiple edge diffrac-
tions. A constructive interference of these waves results in the
resonance behavior of the surface current and scattered eld.
A
(2l)
2
2l

1.0
0.1
0.01
0.001
0.0001
0.5 0 1.0 1.5
= 0.05
= 0.1
2.0 2.5 3.0
2
3
2
5
2
2 3
Figure 2. Integral cross section of thin wires (from Ref. 11). This
quantity has the maximum (resonance) values for wires with the
total length L2lE(2n1)l/2, n0,1,2,3,y. Along such wires
from one end to another, each multiple edge wave acquires the
phase shift of (2n1)p. Due to reection at the end, it acquires an
additional phase shift of p. As a result, this wave becomes
equiphased with all other multiple edge waves. This leads to
the resonance behavior for the current and scattered eld.
482 BACKSCATTER
perpendicular to the wire axis), only the odd resonances
(n1,3,5,y) are realized, due to the symmetry of the
incident eld E
inc
z
z E
inc
z
z. Figure 2, taken from a
classic paper [11], illustrates the resonance behavior of
scattering from thin wires. The incident wave direction
is perpendicular to the wire axis. The quantity A in
Fig. 2 is the total power of the eld scattered in all
directions. The maximum scattering occurs in the direc-
tions perpendicular to the wire and therefore in the
backscattering direction as well. In Fig. 2, the quantity
w 1/[2|ln(ka)|] depends on the wire radius a, 2l L is the
wire length, and akl.
A similar interference of a specular reection with
surface diffracted rays (Fig. 3) explains the backscattering
from perfectly conducting spheres and prolate spheroids
at the upper end of the resonance region (p. 149 of Ref. 2
and pp. 822848 of Ref. 4). However, an important differ-
ence exists between the resonances in scattering from
wires and spheres. The resonance backscattering from
wires is caused by the current resonance in the wires
and it is accompanied by a simultaneous increase of the
scattered eld in other directions. This is a true resonance
effect. The resonance scattering from spheres and spher-
oids is a simple equiphase interference in the single
(backward) direction without the eld increase in other
directions. Additional data for the resonance backscatter-
ing can be found in Ref. 4 (pp. 822848). As in the case of
quasistatic scattering, direct numerical methods are also
efcient for RCS calculations in the resonance frequency
region [57].
3. QUASIOPITCAL SCATTERING
In the quasioptical frequency region, which is often re-
ferred to as the high-frequency region, linear dimensions
of scatterers are much greater than the wavelength of the
incident wave. For example, this occurs in the scattering of
decimeter and centimeter radar waves by such objects as
ships, airplanes, and missiles. In contrast to the quasi-
static and resonance frequency regions, the scatterings by
objects in the quasioptical region are determined mainly
by the objects local properties rather than by their whole
volume.
Large dimensions and complex shapes of scattering
objects allow the existence of various types of scattered
elds. Some of them are illustrated in Fig. 4. Geometric
optics rays and beams (A) reected from the object provide
the main contributions to backscattering. Diffraction of
the incident wave at edges and at lines of curvature
discontinuity or material discontinuity creates edge waves
(B and C), which can be interpreted as diffracted rays.
They represent the second-order contributions [1215].
Diffracted waves arising from corners provide the third-
order contributions [1214]. At the shadow boundary on a
smooth scattering surface, the incident wave excites
creeping waves (D), which propagate along the shadow
side of the object and radiate surface diffracted rays (E).
Due to continuous radiation of these rays, the creeping
waves attenuate exponentially and for this reason the
contribution of surface diffracted rays (E) to backscatter-
ing is small [12,13]. However, for the objects with dimen-
sions comparable to the wavelength, surface diffracted
rays can give appreciable contributions, as it is mentioned
already in the previous section (see also Fig. 3). Diffracted
waves propagating along the scattering object can un-
dergo multiple diffractions at geometric and material
discontinuities and can transform into other types of
waves. This process creates high-order contributions to
backscattering [1215]. A visual description of scattering
from large objects, simple quantitative estimations of
some contributions to backscattering, and relevant refer-
ences are presented in Ref. 16.
High-frequency asymptotic methods are widely used to
predict scatterings in this frequency region. They include
geometric optics (GO) and its extension, geometric theory
of diffraction (GTD); physical optics (PO) and its exten-
sion, physical theory of diffraction (PTD); and various
modications and extensions of GTD and PTD. These
asymptotic techniques are discussed in ELECTROMAGNETIC
WAVE SCATTERING and RADAR CROSS SECTIONS. The present
article supplements these and concentrates mainly on
the physical optics. This method is not so precise as
GTD, PTD, and their extensions, but it allows useful
estimations for the scattered elds in many practical
problems that cannot be treated with other techniques.
1
2
Figure 3. Backscattering from a body of revolution. The total
scattered eld consists of two components. One of them is a
specular reected ray (1); the other is a beam of diffracted rays
(2) radiated by creeping waves traveling along the shadow side of
the scattering body. The equiphase interference of these two
components results in the resonance like increase of the scattered
eld.
B
B
C
C
E
E
E
D
D
A
Figure 4. Backscattering from a convex opaque object. The main
contributor to the scattered eld is beam A reected from the
front planar facet of the object Edge waves B are created at the
edges. Edge waves C are created at the curvature discontinuities.
At the shadow boundary, the incident wave excites creeping
waves D, which propagate along the objects surface and radiate
surface diffracted rays E. Additional creeping waves are excited at
the curvature discontinuities.
BACKSCATTER 483
Direct numerical methods, in their classical forms, are not
efcient in the high-frequency region. Various combina-
tions of these methods with the asymptotic techniques (so-
called hybrid methods) represent a promising direction in
the prediction of high-frequency scattering [17]. Addi-
tional information about numerical, hybrid, and asymp-
totic techniques used for the solution of scattering
problems can be found in Ref. 6,13, and 18 and in the
Further Reading list at the end of this article.
3.1. Geometric Optics Approximation
GO is used for approximate estimations of backscattering
in many practical problems. The basic notion of GO
involves the concept of rays. A ray is an innitely narrow
stream of the wave eld moving with the light velocity
along the lines perpendicular to the phase fronts. These
lines are called ray trajectories. In free space they are
straight lines. Electric and magnetic vectors of the ray
eld are perpendicular to each other and to the direction of
propagation. GO reected rays obey simple rules [19,20]:
(1) the reected ray lies in the incidence plane which
contains the incident ray and the normal to the scattering
surface at the reection point (Fig. 5), (2) the angle of
reection equals the angle of incidence (Snells reection
law), and (3) the power inside an elementary hypothetical
tube conned by neighboring rays is constant.
GO is a good approximation for the eld reected from
large smooth scattering objects. It provides the leading
term in the exact high-frequency asymptotic expansion of
the reected eld. The reection coefcient is found from
the solution of an appropriate canonical problem. For
opaque homogeneous objects, this is the Fresnel reection
coefcient, which determines the amplitude and phase of
plane waves reected from a planar boundary of a semi-
innite homogeneous medium [2, pp. 474479]. For opa-
que objects coated with thin layers, the canonical problem
is the reection of plane waves from an innite planar
layer. This canonical layer is tangential to the scattering
object (Fig. 6). It is homogeneous in the directions parallel
to its surface and has the same material structure in
depth as a real layer at the reection point T. The
canonical layer is placed on the planar boundary of a
homogeneous medium with the same material properties
as a real object at the tangency point. This implies that the
eld on a real coated object is determined exclusively by
its local properties in the vicinity of the reection point.
Nonlocal contributions from various waves propagating
along the object are not treated with this approach.
Creeping and traveling waves [2, pp. 120 and 130] are
examples of such waves.
According to this GO approach, the backscattering RCS
of smooth coated objects equals
sjr0j
2
.
pR
1
R
2
m
2
10
where r(0) is the reection coefcient for the normal
incidence (y 0) and R
1
, R
2
are principal radii of the
curvature of the scattering surface at the reection point
T. In the case of isotropic objects and coatings, the reec-
tion coefcient r(0) does not depend on the incident wave
polarization. Equation (10) is not valid for objects with
planar faces when R
1
R
2
N. It also fails for objects that
contain ruled elements (cylindrical, conical) with a recti-
linear generatrix whose radius of curvature is innite. In
this case, the rays reected by the object form the so-called
reected beams, which undergo the transverse diffusion
while propagating from the object and for this reason lose
their geometrical optics structure in the far zone.
3.2. Physical Optics Approximation
This method goes back to Macdonald [21] and is based on
three concepts: GO, canonical planar layer, and equiva-
lency principle.
The rst step in the physical optics (PO) approximation
is to use GO for the description of elds right on the
scattering surface where GO approximation is still valid.
The second step is to calculate the scattered eld outside
the object using the equivalency principle described in the
following. Tangential components (nE, nH) of electric
and magnetic vectors of the total eld on the scattering
surface (with the external unit normal n) can be inter-
preted as equivalent magnetic and electric currents
j
m
nE V=m
j
e
nH A=m
11
In the PO approach, the equivalent currents are dened
in the GO approximation. The total electromagnetic eld
on the scattering object is considered approximately as the
n
Figure 5. Reection from a planar surface. The reected ray lies
in the plane that contains the incident ray and the unit normal
vector n to the reecting surface.
n

Canonical Layer
Opaque object
with a layer
Figure 6. A scattered eld at the reection point on a coated
scattering object is equal asymptotically (with kR
1,2
-N) to the
eld that would be reected from a tangential layer with the same
material properties. Due to losses, the contributions of rays and
waves propagating along the object (inside the coating) become
small and can be neglected.
484 BACKSCATTER
sum of the GO incident and reected waves (E
GO
E
inc

E
ref
, H
GO
H
inc
H
ref
). Thus, the PO surface currents
are dened as
j
PO
e
nH
GO
j
PO
m
nE
GO
12
This equation denes equivalent currents only on the
illuminated side of the opaque scattering object. On the
shadow side, these currents are assumed to be zero. In the
particular case of perfectly conducting objects, the mag-
netic current does not exist j
PO
m
0 due to the boundary
condition n E0, and the electric current equals
j
PO
e
2nH
inc
according to the GO approximation.
The scattered eld is found by the integration of
equivalent currents over the scattering surface S. Geome-
try of a sample scattering problem is shown in Fig. 7,
where the quantity r is the distance between the integra-
tion (r, y, f) and observation (R, W,j) points. In the far zone
R4kr
2
max
, the scattered eld determined as
E
W
Z
0
H
j
ikZ
0
A
e
W
A
m
j

E
j
Z
0
H
W
ikZ
0
A
e
j
A
m
W

13
A
e;m

1
4p
e
ikR
R
Z
S
j
e;m
e
ikr cos O
dS 14
cos O cos W cos y sin W sin y cosj f 15
Here, E
j,W
is the electric eld intensity (V/m), H
j,y
is the
magnetic eld intensity (A/m); A
e
is the electric potential
vector (A), A
m
is the magnetic potential vector (V), and
Z
0

m
0
=e
0
p
377 O is the impedance of vacuum.
The PO approximation for the scattered eld follows
from Eqs. (13) and (14) when the PO approximation given
by Eq. (12) is used for equivalent surface currents and the
integration region is restricted to the illuminated part of
the scattering surface. The line G shown in Fig. 7 is the
boundary between the illuminated and shadow sides of
the scattering surface S. The PO approach is usually
applied to large convex objects. However, it is also applic-
able to concave objects when the multiple GO reections
are taken into account.
3.2.1. Accuracy of PO. Approximate estimations for
the PO scattered eld [Eq. (13)] can be found by the
application of asymptotic techniques to the integrals [Eq.
(14)] with the PO currents [Eq. (12)]. The rst term of the
asymptotic expansion found in this way is correct for the
elds scattered by smooth convex objects and planar
plates in the specular directions predicted by GO. All
higher-order terms in the PO asymptotic expansion are
incorrect. Only two exceptions exist when PO provides the
exact solution. The rst is the scattering from the innite
perfectly conducting plane. The second is the scattering
from the semiinnite perfectly conducting paraboloid of
revolution illuminated by the plane wave incident along
the symmetry axis (22). In the latter case, the scattered
eld consists of only the GO reected rays. Reference 22
also shows that PO provides the correct second term in the
high-frequency asymptotic expansion for the specular
backscattering from any convex perfectly conducting
bodies of revolution when the incident wave propagates
in the direction parallel to the symmetry axis.
The rst term of the PO asymptotic expansion for the
eld scattered by smooth convex objects in specular direc-
tions represents the GO reected rays [2, pp. 5062].
Therefore, for such objects the PO value of RCS in spec-
ular directions is asymptotically (with k-N) equivalent
to the GO estimation. However, it is well known that GO is
valid only away from the forward direction, i.e., from the
shadow boundary of the incident rays. But PO is more
general than GO and is applicable in the vicinity of this
direction. All known results show that the rst term of the
PO asymptotic expansion for the eld scattered in the
forward direction is correct and leads to the following RCS
for large opaque objects:
s 4p
A
2
l
2
m
2
16
Here, the quantity A is the area of the scattering object
projection on the plane perpendicular to the direction of
the incident wave propagation.
Reference 23 contains similar PO estimations for thin
semitransparent plates. The eld on the plate surface is
dened by complex reection and transmission coefcients
which depend on the incidence angle (g) and polarization
of the incident wave. The incident wave with an arbitrary
linear polarization can be decomposed into two indepen-
dent waves with orthogonal polarizations. A decomposi-
tion is chosen such that either the electric or magnetic
vector of the incident wave is parallel to the plate. In the
rst case, denote the reection and transmission coef-
cients for the electric vector as r
e
(g) and t
e
(g), respectively.
Similar coefcients, r
h
(g) and t
h
(g), for the magnetic vector
describe the plate when the magnetic vector of the
incident wave is parallel to the plate. Reference 2
(pp. 479499) contains instructions for the calculation
of these coefcients. Two directions of scattering are of
n
z
x
y
r

P
S
Figure 7. Schematics of a scattering problem: S is the surface of
the scattering object; the dashed part of this surface (with the
boundary G) is located in the shadow region which is hidden from
the incident rays.
BACKSCATTER 485
greatest interest. The rst is the direction of specular
reection, W pg, and the second is the forward direc-
tion, W g (Fig. 8). According to (23), the PO approxima-
tions of RCS in the specular direction are given by
s
e
p g 4p
A
2
l
2
jr
e
gj
2
s
h
p g 4p
A
2
l
2
jr
h
gj
2
17
and in the forward direction by
s
e
g 4p
A
2
l
2
j1 t
e
gj
2
s
h
g 4p
A
2
l
2
j1 t
h
gj
2
18
where the quantity A is the same as in Eq. (16). This is the
area of the plate projected on the plane perpendicular to
the direction of the incident wave. Equations (17) and (18)
are applicable for planar plates of an arbitrary shape
under the condition Abl
2
. This means that the grazing
angles (gEp/2) cannot be treated with these equations.
Known results for perfectly conducting plates (|r
e,h
(g)|
1, |t
e,h
(g)|0) show that PO estimations given in Eqs.
(17) and (18) are correct. These equations also give the
correct result, s
h
(p/2) 0, for perfectly conducting plates
under the grazing incidence. In this case, the incident
wave does not undergo diffraction because its electric
vector is perpendicular to the plate surface. PO describes
satisfactorily the eld scattered from large conducting
plates not only in the specular and forward directions
corresponding to mainlobes in the directivity pattern, but
also in the directions of neighboring sidelobes. However,
PO fails to predict a eld level in minimums of
the directivity pattern (Figs. 719 and 720 on p. 509 of
Ref. 2) and does not satisfy the reciprocity principle.
The PO currents given by Eq. (12) are discontinuous on
the shadow boundary of a scattering surface. The PO eld
contains spurious waves from such a boundary in the case
of smooth scattering surfaces (Fig. 9). A similar current
discontinuity on scattering objects with edges results in
edge waves. If the scattering edge is visible from the
observation point, such an edge wave does exist. The PO
edge waves coming from invisible edges are spurious
shooting-through waves (Fig. 10). Such shooting-through
waves do not occur in the backscattering direction. All PO
spurious waves can be removed by neglecting the corres-
ponding terms in the asymptotic expansion of the integral
in Eq. (14). For real edge waves, even the rst-order term
of their PO asymptotic expansion is incorrect. This defect
is remedied in PTD by the inclusion of the eld radiated
by the so-called nonuniform currents arising from the
diffraction of the incident wave at edges [14,15,30].
One should emphasize a special role of PO in PTD. PO
is a constitutive part of PTD. Therefore, the POs rst- and
higher-order asymptotic terms are integral parts of the
PTD asymptotic expansions for the total scattered eld.
For example, the terms with coefcients (3d/16a) in the
PTD equations (99) and (100) of (15) are exactly the POs
second-order terms in the asymptotic expansion of the
eld scattered by a perfectly conducting cylinder of nite
length.
3.2.2. Polarization of the PO Scattered Field. In general
cases, components E
W
and E
j
in Eq. (13) have different
amplitudes and phases. This results in the elliptic polar-
ization of the scattered eld even in the case when the
incident wave is linearly polarized. This means that
the electric vector of the scattered eld rotates with
the angular frequency ok c and its endpoint traces
an ellipse. The direction of rotation (clockwise or
z

=
S
=

A
Figure 8. Directions of the forward (W g) and specular (W
p g) scattering from the plate S. The dashed line A denotes the
projection of the plate S.
1
2
Figure 9. The PO current given by Eq. (12) is discontinuous at
the shadow boundary of the scattering surface. This leads to
spurious waves (1 and 2) in the case of a smooth scattering
surface.
1 2
3
Figure 10. A scattered eld is generated by the induced surface
currents. Any approximations for these currents can result in the
appearance of nonphysical components in the scattered eld. In
particular, the PO currents create spurious shooting-through
edge waves (1, 2, and 3) passing through an opaque object.
486 BACKSCATTER
counterclockwise) is determined by the phase shift
between components E
W
and E
j
. The lengths of the
elliptical axes are determined by the amplitudes of these
components. Due to diffraction, the scattered eld can
contain an electric eld component perpendicular to the
incident wave polarization. This phenomenon is known as
depolarization, or cross-polarization.
The PO eld scattered by arbitrary perfectly conduct-
ing objects in the backscattering direction does not contain
the cross-polarized component [2, p. 56]. It is assumed
only that no multiple GO reections occur on the objects
surface. This PO result is correct for scattering objects
with certain symmetry. These are objects with a symmetry
plane parallel to both the electric (or magnetic) vector of
the incident wave and the direction of its propagation.
Each element of such a scattering object may create the
cross-polarized component. But because of the symmetry,
the cross-polarized components from symmetric elements
cancel each other in the backscattering direction (Fig. 11).
A convex smooth body of revolution whose symmetry axis
is parallel to the incident wave direction is a simple
example of such an object. A symmetric plate, illuminated
by the plane wave whose electric (or magnetic) vector and
direction of propagation are parallel to the symmetry
plane, is another example where the backscattered eld
does not contain a cross-polarized component (Fig. 12).
As stated previously, the rst term of the PO
high-frequency asymptotic expansion represents the GO
reected ray. This ray contains the crosspolarized com-
ponent when the electric vector of the incident ray makes
any angle different from 01 and 901 with the incidence
plane at the reection point. This is the case when PO
correctly describes depolarization of the scattered eld in
the high-frequency asymptotic limit (k-N). However, PO
fails to predict all depolarization effects caused by the
diffraction part of equivalent surface currents.
3.2.3. Bistatic RCS. Bistatic RCS determines the power
ux density of electromagnetic waves scattered by the
object in an arbitrary direction. The angle between
the directions to the transmitter and receiver is called
the bistatic angle. Monostatic, or backscattering, RCS is a
particular case of bistatic RCS when the bistatic angle
equals zero. Some PO results for bistatic RCS have
already been discussed. This section addresses the inter-
relationships between bistatic and monostatic RCS.
Using PO, one can prove the following statement:
For perfectly conducting bodies which are sufciently smooth,
in the limit of vanishing wavelength, the bistatic cross section
is equal to the monostatic cross section at the bisector of the
bistatic angle between the direction to the transmitter and
receiver. [1, pp. 157160; 2, p. 11]
There is a simple physical explanation for this result.
As already stated, the rst term of the PO asymptotic
expansion for the eld scattered by smooth objects exactly
equals the GO expression for the reected rays. The
monostatic RCS caused by these rays is given by
Eq. (10). In the case of perfectly conducting objects, this
equation reduces to
s pR
1
R
2
19
It should be noted that this equation is valid also for the
bistatic RCS, which therefore does not depend on the
bistatic angle b 2W (Figs. 5 and 6). This follows directly
from Eqs. (5.32), (6.19), and (6.20), given by Fock [19,
Chap. 8]:
Er
1
2
E0

R
1
R
2
p
e
iks
s
Hr
1
2
H0

R
1
R
2
p
e
iks
s
20
These expressions describe the eld reected by smooth
convex objects at a far distance (sbR
1,2
) from the reec-
tion point for any incidence angle (0rWop/2). In the case
of reection from concave surfaces, the reected eld
acquires the additional phase shift of ( p/2) in passing
through a focus of reected rays. Vectors E(0) and H(0)
E
cr
E
cr
E
inc
E E
S
y
x
Figure 11. Backscattering without depolarization from a sym-
metric perfectly conducting surface S. The incident wave direc-
tion is parallel to the symmetry plane yz. Vectors E
cr
are the
cross-polarized components of the reected eld. Due to the
symmetry, they cancel each other.
E
inc
S
y
x
Figure 12. Backscattering without depolarization from a per-
fectly conducting plate S. The incident wave direction is parallel
to the symmetry plane yz. Cross-polarized components scattered
by the left and right parts of the plate are symmetric and
completely cancel each other.
BACKSCATTER 487
denote the reected eld at the reection point. Expres-
sions (20) clearly show that the GO reected eld really
does not depend on the incidence angle. As a result,
the bistatic RCS does not depend on the bistatic angle
and is the same as the monostatic RCS at the bisector
direction that is perpendicular to the scattering surface
at the reection point. Thus, the cited equivalence be-
tween the bistatic and monostatic RCS is a pure GO effect
and is fullled asymptotically (with k-N) only in the
ray region, away from the shadow boundary behind the
scattering object. It is also clear that this equivalence is
not valid when the scattered eld contains multiple
reected rays arising from concave parts of the scattering
surface. Reference 1 (pp. 160183) presents additional
results for bistatic RCS of some typical objects found using
PO and other approximations.
In the case of coated smooth objects, Eq. (20) leads to
the bistatic RCS
sW jr
e;h
Wj
2
pR
1
R
2
21
where the reection coefcients r
e,h
(W) depend on the
polarization and direction of the incident wave. Therefore,
the asymptotic equivalence between bistatic and mono-
static RCS can be valid only for those bistatic angles
where |r
e,h
(W)||r
e,h
(0)|.
This discussion relates to the situation where the
bistatic scattered eld is produced by a single scattering
source (the reection point). The eld scattered by com-
plex objects can have many sources (scattering centers) on
the objects surface. Figure 13 illustrates the bistatic
scattering from the center located at the point (x
n
, y
n
, z
n
).
The origin of Cartesian coordinates is somewhere inside
the scattering object. The z axis is directed along the
bisector of the bistatic angle b. In the far zone, the bistatic
scattered eld can be represented as the sum of partial
contributions from all scattering centers [4, pp. 983988]:
ub
e
ikR
R
X
n
v
n
e
2ikzn cosb=2
22
Here, u is either the electric or magnetic vector of the total
scattered eld; R is the distance from the origin to the
observation point. Vector v
n
determines the amplitude and
polarization of the wave generated by the nth scattering
center. Suppose that vectors v
n
and the number of scatter-
ing centers are constant inside the 0rbrb
max
. Assume
also that coordinates z
n
of scattering centers do not
depend on the bistatic angle b, while coordinates x
n
and
y
n
can be functions of this angle. Under these conditions,
the monostatic eld scattered in the bisector direction
equals
u0
e
ikR
R
X
n
v
n
e
2ikz
n
23
Comparisons of Eqs. (22) and (23) show that the bistatic
RCS, s(b, k), at the frequency oc k will be equal to the
monostatic RCS, s[0, kcos(b/2)], at the frequency o
c k cos(b/2). This equality requires the additional as-
sumption that each vector v
n
is constant in the frequency
band [c kcos(b/2)rorc k]. The derivation, some appli-
cations, and restrictions of this equivalence relation are
presented in Ref. 4 (pp. 983988). In particular, this
reference notes that this equivalence is not true for the
bistatic scattering from spheres when the bistatic angle
exceeds one degree and the sphere radius is less than 6l.
Before applying this equivalence in practice, we must rst
check carefully that all assumptions made in this scatter-
ing model are really fullled. One can expect that this
approximate model can be reasonable only for small
bistatic angles.
3.2.4. PTD as an Extension of PO. PTD is a natural
extension of PO [14,15,24,30]. In PTD, the PO current
given by Eq. (12) is considered as the uniform component
( j
0
) of the total surface current and is supplemented by
the additional, nonuniform component ( j
1
). In contrast to
the PO current that has the GO origin, the nonuniform
current is caused by diffraction at smooth bendings, sharp
edges, corners, and any other geometric discontinuity and
material inhomogeneity on the scattering surfaces. Creep-
ing and edge current waves are examples of such a
current. The eld generated by the nonuniform current
represents the PTD contribution to the scattered eld.
Exact analytical expressions for nonuniform currents are
not available. Therefore, one has to nd their high-fre-
quency approximations by the solution of appropriate
canonical problems. In this manner, Fock developed spe-
cial functions that describe the nonuniform current on
smooth convex objects in the vicinity of the shadow
boundary [19]. The Sommerfeld solution of the wedge
canonical problem is used for the asymptotic description
of the nonuniform current near perfectly conducting edges
[14,15,24,30]. The concept of uniform and nonuniform
currents plays a key role in PTD and those hybrid
techniques that combine direct numerical methods with
high-frequency asymptotic approximations [6,17,18]. Re-
ference 15 shows that PTD properly denes the leading
term in the high-frequency asymptotic expansions for
primary and multiple edge waves. A close connection
exists between PTD and GTD. The latter automatically
follows from the PTD integrals when they are evaluated
R
x
y
(x , y
n
, z
n
)
R
z
/2
/2
n

Figure 13. Geometry of the bistatic scattering problem. A solid
winding line represents a scattering object with many scattering
centers (x
n
, y
n
, z
n
). The scattering direction forms the angle b
with the direction of the incident wave.
488 BACKSCATTER
by the stationary phase technique [15, pp. 136138]. Some
PTD results are presented in the next section.
4. BACKSCATTERING RCS OF SIMPLE SHAPES
This section contains examples of PO estimations for RCS
of simple objects. Whenever possible, these estimations
are accompanied by more precise PTD counterparts that
include the contributions of primary edge waves gener-
ated by the nonuniform edge currents. Only objects with
symmetry of revolution are considered. All given data are
taken from Refs. 15, 16, and 30. Exact, numerical solu-
tions of scattering problems for bodies of revolution can be
found, for example, in Refs. 6, 18, and 25.
4.1. Semitransparent Disk
The geometry of this scattering problem is shown in Fig. 8.
The backscattering direction is determined by the sphe-
rical coordinates W pg, j p/2. The disk radius is
denoted by the letter a. The incident wave can have either
E or H polarization. In the rst case, the electric vector is
perpendicular to the incidence plane and parallel to the
disk face. The disk properties are described by the reec-
tion and transmission coefcients, r
e
(g), t
e
(g), with respect
to the electric vector. In the case of H polarization, the
magnetic vector of the incident wave is perpendicular to
the incident plane and parallel to the disk face. The
reection and transmission coefcients, r
h
(g), t
h
(g), deter-
mine the magnetic vector on the front (z 0) and rear
(z 0) faces of the disk, respectively. According to
Eq. (67) in (23), the backscattering RCS is given by
s
PO
e
g jr
e
gj
2
pa
2
J
1
2kasing
2
cot
2
g
s
PO
h
g jr
h
gj
2
pa
2
J
1
2ka sing
2
cot
2
g
24
where J
1
(x) is the Bessel function and the incidence angle
is restricted by the values 0rgop/2. For perfectly con-
ducting disks, one should put r
e
(g) 1 and r
h
(g) 1.
Then, in the case of the normal incidence (g 0), Eq. (24)
reduces to
s
PO
e
s
PO
h
pa
2
ka
2
25
Figure 7-24 on p. 514 of Ref. 2 shows that this equation is
in good agreement with the exact results when kaZ5.
Note also that Eq. (18), with Apa
2
cos g, determines the
PO bistatic RCS of this disk for the forward direction
(Wg). PTD estimations for RCS of a perfectly conducting
disk are presented in Chaps. 2 and 5 of Ref. 14. See also
pp. 514521 of Ref. 2. Some important corrections in the
PTD expressions for bistatic scattering from a disk are
given in Ref. 24. Contributions of multiple edge waves to
forward scattering are presented in Ref. 15 (pp. 149151).
4.2. Circular Cone
Geometric parameters of a perfectly conducting cone are
shown in Fig. 14. The incident wave direction is parallel to
the symmetry axis of the cone. The PO backscattering
RCS is given by Eqs. (17.06) and (17.09) in (4) [see also
Eqs. (2.4.7) and Eq. (2.4.10) in Ref. 30]
s
PO
pa
2
.
1
ka
tan
2
a sinkl tanae
ikl

2
26
where the cone length equals l acot a. To clarify the
physics in this equation, we rewrite it as
s
PO
pa
2
i
2ka
tanae
ikl
tana
i
2ka
tana

e
ikl

2
27
The rst term (with exponential e
ikl
) is related to the
wave scattered by the cone tip. Comparison with the exact
solution [3, Fig. 18.15, p. 691] shows that this PO approxi-
mation is quite satisfactory for all cone angles (0rarp/2).
The second term (with the exponential e
ikl
) describes the
edge wave contribution. This PO approximation is incor-
rect. PTD takes into account the additional contribution
from the nonuniform (diffraction) currents located near
the cone edge and provides a more accurate result, given
by Eqs. (17.06) and (17.08) in Ref. 14 [see also Eqs. (2.4.7)
and (2.4.9) in Ref. 30]
s
PTD
pa
2
.
1
ka
tan
2
a sinkl
2
n
sin
p
n
cos
p
n
cos
2a
n
e
ikl

2
28
where n
3
2
a=p. When the cone transforms into the disk
(a-p/2, l-0) the previous expressions reduce to
s
PO
s
PTD
pa
2
ka
2
29
which coincides with Eq. (25).
4.3. Paraboloid
The directrix of a paraboloid is given by the equation
r 2pz where patana (Fig. 15). The length of the
paraboloid equals l a
2
/(2p) (a/2)cot a. The angle a is
formed by the symmetry axis z and the tangent to the
directrix at the point z l. The radius of the paraboloid
base equals a. The incident wave propagates in the
positive direction of the z axis. According to Eq. (18.02)
l
2a

Figure 14. Backscattering from a truncated cone. The base

diameter of the cone (2a) is large compared to the wavelength.
The length of the cone (l) can be arbitrary. In the limiting case l
0, the cone transforms into a disk.
BACKSCATTER 489
in Ref. 14 (see also Sect. 2.5 in Ref. 30), the PO back-
scattering RCS of a perfectly conducting paraboloid equals
s
PO
4pa
2
tan
2
a sin
2
kl 30
This equation can be written in another form as
s
PO
pa
2
tan
2
a
.
je
ikl
e
ikl
j
2
31
which is more convenient for the physical analysis. The
term with the exponential e
ikl
gives the correct contribu-
tion of the specular reection from the paraboloid tip.
The term with the exponential e
ikl
represents the edge
wave contribution and is wrong. PTD includes the addi-
tional contribution from the nonuniform edge currents
and provides the correct result, given by Eq. (18.04) in
Ref. 14 [see also Eq. (2.5.4) in Ref. 30]
s
PTD
pa
2
tana
2
n
sin
p
n
cos
p
n
cos
2a
n
e
2ikl

2
32
where n
3
2
a=p. When the paraboloid transforms
into the disk (a-p/2 and l-0), these expressions reduce
to Eq. (29).
4.4. Truncated Sphere
The geometry of this scattering problem is shown in
Fig. 16. The angle a is formed by the tangent to the sphere
generatrix and the symmetry axis. The sphere radius
equals r a/cos a, where a is the base radius. The length
of the truncated sphere equals l r (1sin a). It is
assumed that lrr. The PO backscattering RCS of a
perfectly conducting sphere equals [14, Eq. (19.05); 30,
Eq. (2.6.2)]
s
PO
pa
2
1
cos a

i
2ka
tana
i
2ka

e
2ikl

2
33
In this equation, the rst two terms represent the specular
reection from the sphere, and both are correct. The third
term (with the exponential e
2ikl
) gives the contribution
from the edge and it is wrong. With kab1, Eq. (33)
simplies to
s
PO
pa
2
1
cos a
tanae
2ikl

2
34
When a 0, the latter gives the RCS of a hemisphere,
s pa
2
. The PTD backscattering RCS is determined by
Eq. (19.12) in Ref. 14 [Eq. (2.6.8) in Ref. 30]
s
PTD
pa
2
1
cos a

2
n
sin
p
n
cos
p
n
cos
2a
n
e
2ikl

2
35
where n
3
2
a=p. When the sphere transforms into the
disk (a-p/2,r-N,l-0), Eqs. (34) and (35) reduce exactly
to Eq. (29).
4.5. Circular Cylinder with Flat Ends
The diameter and length of a perfectly conducting cylinder
are assumed to be large as compared with the wavelength
of the incident wave. PO and PTD estimations for back-
scattering RCS are developed in Chap. 3 of Ref. 14. They
are also presented in Ref. 2 (pp. 308312). PTD asymptotic
expressions for bistatic RCS are given in Ref. 15 (pp.
152154).
5. BACKSCATTERING FROM COMPLEX OBJECTS
AND STEALTH PROBLEMS
Computer codes based on GTD, PTD, and on their hybri-
dizations have been developed for prediction of high-
frequency scattering from complex perfectly conducting
objects. Relevant references can be found in Refs. 16 and
18 and in special issues of Proc. IEEE (1989), IEEE Trans.
Anten. Propag. (1989), and Annales des Telecommunica-
tions (1995), which are mentioned in the Further Reading
list. Note also the XPATCH code (based on the shooting-
and-bouncing ray technique and PTD), which allows the
calculation of backscattering from complex geometries.
Information about this code is published in IEEE Trans.
l

2a
Figure 15. Backscattering from a truncated paraboloid. The
base diameter of the paraboloid (2a) is large compared to the
wavelength. The length of the paraboloid (l) can be arbitrary. In
the limiting case l 0, the paraboloid transforms into a disk.
l
2a

Figure 16. Backscattering from a truncated sphere. The base

diameter of the sphere (2a) is large compared to the wavelength.
The length of the sphere (l) can be arbitrary. In the limiting case l
0, the sphere transforms into a disk.
490 BACKSCATTER
Anten. Propag. Mag. 36(1):6569 (1994). Computer codes
interfaced with graphical utilities of workstations can
display three-dimensional chromatic views of scattering
centers and magnitudes of their contributions to RCS.
This is the end result of complicated computations. How-
ever, a part of this can be obtained without any computa-
tions. Nature can show us the location of all scattering
centers if we bring a small metallized model of the
scattering object into an anechoic optical chamber and
illuminate the model by the light. Bright shining points
(scattering centers) seen on a scattering object are exactly
those from which the radar waves will be reected toward
the radar, if we look at the object from the light source
direction. (The following text is taken from Ref. 16 and
slightly modied.)
The locations of these points do not depend on the
frequency of incident electromagnetic waves, and they are
determined completely by the location of the light source
(the radar), the observer, and the scattering object. These
shining points obey the Fermat principle. This means that
the path along the ray between the source, the reection
point, and the observer is extremal (minimal or maximal)
in comparison with similar paths corresponding to neigh-
boring points on the objects surface. A more detailed
description of the Fermat principle is presented, for ex-
ample, in Sect. 3.3.2 in Ref. 20.
Waves reected from discrete shining points located on
the smooth parts of the scattering object represent the
usual geometric optics reected rays. Waves reected from
discrete shining points located on edges, tips, and corners
are diffracted rays. The farthest shining points on a
smooth objectthose located on the boundary between
visible and invisible sides of the objectcreate surface
diffracted rays.
As the orientation of the object is changed, the shining
points move along the object. Some of them can merge
with each other and create a brighter point. In this case
our eyes (i.e., the radar) are located on a caustic that is the
envelope of merged rays.
We can also observe bright shining lines and bright
shining spots on the object, which contain an innite
number of continuously distributed shining points. Their
important property is that the optical path through a
shining point from the source to the observer is constant
for all of these points. It is assumed here that the source
and observer are far from the scattering object. All
waves reected from these points reach the observer
with the same phase. From the mathematical point of
view, each such point is a stationary point of the innite
order; the derivatives (of any higher order) of the wave
phase along the shining line (or along the shining spots)
are zero at these points.
Shining spots and lines located on smooth parts of the
scattering surface generate powerful reected beams
(such as those radiated by reector antennas) that repre-
sent the strongest contributors to RCS. Shining edge lines
create edge-diffracted beams whose contributions can be
comparable with those from ordinary reected rays.
It is difcult to model in optics the electromagnetic
properties of realistic scattering surfaces for the radar
frequency band. But the optical modeling can be used to
identify the scattering centers and to control them by an
appropriate shaping of the scattering surface. As it is well
known, one of the basic ideas of the current stealth technol-
ogy is to use an appropriate body shaping and to shift all
reected beams and rays away from the directions to the
radar. See, for example, Refs. 2 and 16, and the radar cross-
section handbooks mentioned in the reading list. Some
interesting details about the development of stealth tech-
nology in the United States are presented in Refs. 2628.
The second idea of stealth technology is traditional: to
use radar absorbing materials (RAMs) and composite
structures in order to reduce the intensity of reected
beams and rays. References 2,16, and 29, and radar
handbooks (mentioned in the reading list) describe funda-
mental concepts used in the design and application of
RAMs. We present here some details taken from Ref. 16.
In order to use RAMs efciently, it is necessary to place an
electric (magnetic) RAM in the region where the average
electric (magnetic) eld is maximal. Location of these
regions in the vicinity of real objects depends on many
factors, such as the radar frequency, geometry, size, and
electrical properties of the object, as well as properties of
materials intended for absorption. Identication of such
regions and optimization of the RAM parameters to mini-
mize RCS is a very complex problem. Its solution is
attainable only in some simple cases. Most of these relate
to absorbing layers on an innite metallic plane. From the
physical point of view such absorbing layers can be
considered as open resonators that can support eigenos-
cillations. Frequencies of eigenoscillations are complex
quantities. Their imaginary part is responsible for the
loss inside the resonator and radiation from the resonator.
It turns out that the minimal reection from such reso-
nators happens when the frequency of an incident wave is
close to the real part of the resonator eigenfrequency.
Note that thin electric RAMs are not efcient when
applied on metallic objects. This is due to the boundary
condition; the tangential component of the electric eld is
very small on the metal surface. On the contrary, magnetic
absorbing materials can be applied directly to the surface
of a metallic object. This is an important advantage of
magnetic materials over electric ones.
However, any RAMs (electric, magnetic, and hybrid)
homogeneous in the direction parallel to the reecting
plate are not efcient for grazing incidence (yE901, in
Fig. 5). In this case, the reection coefcient tends to unity
independently of the incident wave polarization when
y-901. This is a fundamental limitation of ordinary
RAMs. They do not work against grazing incident waves.
That is why ordinary RAMs do not reduce forward scat-
tering. Actually, the RAM terminology is justied only for
incidence angles that are not too far from y 0 and when
the reection coefcient is small enough.
There is another reason why RAMs are not efcient
against the forward scattering. Any large opaque scatter-
ing objects act like a lens focusing the scattered eld in the
forward direction. As shown in Refs. 15, 23, and 30, the
forward scattering can be interpreted as the shadow ra-
diation generated by the black bodies (perfectly absorbing
objects whose reection coefcient equal zero). In Ref. 23
(see also Refs. 15 and 30), it was established that the total
BACKSCATTER 491
power of the shadow radiation in all directions equals one-
half of the total power scattered by a perfectly reecting
body (of the same shape and size as the black body). This
fundamental law means that the total power of the scat-
tered eld cannot be reduced with application of any radar
absorbing coatings more than by a factor of 1/2.
Various geometric and material inhomogeneities on the
scattering surface can partially transform the incident
wave into surface waves propagating along absorbing
layers. This may be used to further reduce the RCS.
However, this idea has two essential defects: (1) any
inhomogeneity creates an additional undesirable scat-
tered eld and (2) it is not a simple problem to design
an absorbing layer that would allow the propagation of
surface waves. To support surface waves with electric
vector parallel to the incidence plane, the surface impe-
dance must be inductive. But the surface impedance must
be capacitive to support surface waves with the electric
vector perpendicular to the incidence plane. This means
that the surface impedance, and therefore the absorbing
layer, must depend on the radar polarization with respect
to the incidence plane. But this plane is different at
different points of the scattering surface and different at
the same point when the scattering object changes its
orientation with respect to radar. It is very difcult and
probably impossible to design such an absorber, especially
against radars with circular polarization. However, for
some chosen orientations of the object and for an appro-
priate polarization of the incident wave, this might not be
a hopeless problem.
Development of efcient hybrid techniques and compu-
ter codes to predict RCS of large complex objects with
realistic materials and research efforts to overcome the
abovementioned physical limitations in RCS reduction
represent challenging problems for future stealth technol-
ogy. One can expect that future advanced computer codes
will contain as necessary constitutive components the
known high-frequency techniques (such as GTD, PTD,
and the uniform theory of diffraction) extended for coated
and composite objects. Diffraction coefcients used in
these techniques can be determined by the numerical
solution of appropriate canonical problems. Direct numer-
ical methods should be used for calculation of scattering
from those elements of the scattering object that cannot be
treated by high-frequency methods. Diffraction interac-
tion between the objects elements handled by high-fre-
quency techniques and by direct numerical methods can
be described by the surface integral equations.
BIBLIOGRAPHY
1. J. W. Crispin Jr. and K. M. Siegel, eds., Methods of Radar
Cross-Section Analysis, Academic Press, New York, 1968.
2. G. T. Ruck, D. E. Barrick, W. D. Stuart, and C. K. Kirchbaum,
Radar Cross-Section Handbook, Vols. 1 and 2, Plenum Press,
New York, 1970.
3. J. J. Bowman, T. B. A. Senior, and P. L. E. Uslenghi, eds.,
Electromagnetic and Acoustic Scattering by Simple Shapes,
Hemisphere Publishing, New York, 1987.
4. P. C. Fritch, ed., Proc. IEEE (Special issue on radar reectiv-
ity) 53(8) (Aug. 1965).
5. M. N. O. Sadiku, Numerical Techniques in Electromagnetics,
CRC Press, Boca Raton, FL, 1992.
6. E. K. Miller, L. Medgyesi-Mitschang, and E. H. Newman,
Computational Electromagnetics, IEEE Press, New York,
1991.
7. P. P. Silvester and G. Pelosi, eds., Finite Elements for Wave
Electromagnetics, IEEE Press, New York, 1994.
8. P. Ya. Umtsev and A. P. Krasnozhen, Scattering from a
straight thin wire resonator, Electromagnetics 12(2):133146
(1992).
9. L. A. Vainshtein, Waves of current in a thin cylindrical
conductor, II. The current in a passive oscillator, and the
radiation of a transmitting antenna, Sov. Phys., Tech. Phys.
4(6):617626 (Dec. 1959).
10. L. A. Vainshtein, Waves of current in a thin cylindrical
conductor, III. Variational method and its application to the
theory of ideal and impedance conductors, Sov. Phys., Tech.
Phys. 6(1):1929 (July 1961).
11. M. A. Leontovich and M. L. Levin, Towards a theory on the
simulation of oscillations in dipole antennas, Zh. Tech. Fiz.
14(9):481506 (1944) (in Russian) (English transl. published
in the following report: K. C. Chen, ed., SAND91-0720, UC-
705, Sandia National Laboratories, Albuquerque, NM and
Livermore, CA, Contract DE-AC04-76DP00789, Jan. 1992).
12. J. B. Keller, Geometrical theory of diffraction, J. Opt. Soc. Am.
52:116130 (1962).
13. R. C. Hansen, ed., Geometrical Theory of Diffraction, IEEE
Press, New York, 1981.
14. P. Ya. Umtsev, Method of Edge Waves in the Physical
Theory of Diffraction, Soviet Radio Publishing House,
Moscow, 1962, pp. 1243, transl. U.S. Air Force, Foreign
Technology Division, Wright-Patterson AFB, Ohio, 1971;
Technical Report AD 733203, DTIC, Cameron Station, Alex-
andria, VA.
15. P. Ya. Umtsev, Elementary edge waves and the physical
theory of diffraction, Electromagnetics 11(2):125160 (1991).
16. P. Ya. Umtsev, Comments on diffraction principles and
limitations of RCS reduction techniques, Proc. IEEE
84(12):18301851 (1996).
17. D. P. Bouche, F. A. Molinet, and R. Mittra, Asymptotic and
hybrid techniques for electromagnetic scattering, Proc. IEEE
81:16581684 (1993).
18. W. R. Stone, ed., Radar Cross Sections of Complex Objects,
IEEE Press, New York, 1990.
19. V. A. Fock, Electromagnetic Diffraction and Propagation
Problems, Pergamon Press, London, 1965.
20. M. Born and E. Wolf, Principles of Optics, Pergamon Press,
New York, 1975.
21. H. M. Macdonald, The effect produced by an obstacle on a
train of electric waves, Phil. Trans. Roy. Soc. Lond. Series A,
Math. Phys. Sci. 212:299337 (1912).
22. C. E. Schensted, Electromagnetic and acoustic scattering by a
semi-innite body of revolution, J. Appl. Phys. 26:306308
(1955).
23. P. Ya. Umtsev, Diffraction of electromagnetic waves at black-
bodies and semi-transparent plates, Radiophys. Quantum
Electron. 11:527538 (1968).
24. P. Ya. Umtsev, Comments on Comparison of three high
frequency diffraction techniques, Proc. IEEE 63:17341737
(1975).
25. R. D. Graglia et al., Electromagnetic scattering for oblique
incidence on impedance bodies of revolution, IEEE Trans.
Anten. Propag. 43(1):1126 (1995).
492 BACKSCATTER
26. M. W. Browne, Two rival designers led the way to stealthy
warplanes, New York Times (Science Times Section), May 14,
1991.
27. S. F. Browne, The secret ship, Popular Sci., (Oct. 1993).
28. B. Rich and L. Janos, Skunk Works, Little, Brown, Boston
New YorkLondon, 1994.
29. K. J. Vinoy and R. M. Jha, Radar Absorbing Materials,
Kluwer Academic Publishers, Boston, 1996.
30. P. Ya. Umtsev, Theory of Edge Diffraction in Electromag-
netics, Tech Science Press, Encino, CA, 2003.
FURTHER READING
This section contains short comments on some related references.
P. C. Fritch, ed., Radar Reectivity, special issue of Proc. IEEE
53(8) (Aug. 1965). [The rst attempt to sum up basic results in
the eld of RCS. Includes a comprehensive subject index, about
1500 titles (pp. 10251064).]
J. W. Crispin Jr. and K. M. Siegel, eds., Methods of Radar Cross
Section Analysis, Academic Press, New York, 1968. [Includes a
short historical survey of high-frequency approximations. Con-
tains results of such approximations for monostatic and bi-
static RCS of simple objects. States a strategy of RCS
calculations for complex objects. Some results are conveniently
summarized in tables. Table 5 on p. 147 contains expressions
for RCS in the Rayleigh region (for objects small in comparison
with wavelength). High-frequency monostatic and bistatic
RCS are given in Tables 7 and 8 on pp. 168, 169, 171.]
G. T. Ruck, D. E. Barrick, W. D. Stuart, and C. K. Kirchbaum,
Radar Cross-Sections Handbook, Plenum Press, New York,
1970. [This is a real encyclopedia of RCS, which includes most
results obtained before 1970. It contains numerous theoretical
and experimental results for both perfectly conducting and
absorbing objects. The physical theory of diffraction (PTD) is
mentioned here as the SommerfeldMacdonald technique.
Equation (5.1-54), presented on p. 351, for the bistatic RCS
of an ellipsoid is incorrect. In the particular case when an
ellipsoid transforms into a sphere, this equation does not
provide the bistatic RCS for the sphere, s pa
2
. Instead it
leads to the wrong quantity s pa
2
/(1cos b)
2
, where b is the
bistatic angle.]
J. J. Bowman, T. B. A. Senior, and P. L. E. Uslenghi, eds.,
Electromagnetic and Acoustic Scattering by Simple Shapes,
Hemisphere Publishing, New York, 1987. (Contains a compre-
hensive collection of theoretical results for RCS of simple
objects which allow the exact solutions of diffraction problems.
Both low-frequency and high-frequency approximations are
presented as well.)
M. Skolnik, ed., Radar Handbook, McGraw-Hill, New York, 1970.
(Contains many results of measurements and calculations for
RCS. Calculations were carried out mostly by the physical
optics approach and GTD. Analytical expressions for RCS are
not given.)
W. R. Stone, ed., Radar Cross Sections of Complex Objects,
special issue of Proc. IEEE 77(5) (May 1989); W. R. Stone,
ed., Radar Cross Sections of Complex Objects, special issue
of IEEE Trans. Anten. Propag. 37(5) (May 1989). [These
two references contain many theoretical results concerning
RCS for complex objects (perfectly conducting objects with
complex shapes and simple objects with complex boundary
conditions).]
W. R. Stone, ed., Radar Cross Sections of Complex Objects, IEEE
Press, New York, 1989. (This book consists of a collection of
articles. It includes expanded versions of about half of the
papers published in two previously mentioned special issues. It
also contains papers written especially for this book and
reprints of some earlier key papers.)
J. M. Bernard, G. Pelosi, and P. Ya. Umtsev, eds., Radar Cross
Sections of Complex Objects, special issue of French journal
Annales des Telecommunications 50(56) (MayJune 1995)
(published in English with abstracts translated into French).
(Contains the asymptotic analysis of RCS for higher-order
curved surfaces, physical theory of slope diffraction, PO and
PTD analysis of trihedral corner reectors, a selective review
of some numerical methods for electromagnetic scattering, and
some other results.)
The following three books contain additional information on RCS.
They include concise descriptions of basic exact and approx-
imate techniques for prediction of RCS, they introduce meth-
ods of RCS enhancement and reduction, and they contain a
large number of calculated and measured data for RCS of
many typical simple and complex objects. The books comple-
ment each other, with emphasis on different aspects in the eld
of RCS.
A. L. Maffett, Topics for a Statistical Description of Radar
Cross Section, Wiley, New York, 1989. (This book treats
the subject of RCS with special emphasis on statistical
aspects and applications. It reects broad interests of the
author: from historical background and perspective through
analytical and numerical methods of RCS calculation and
RCS measurements to elements of detection theory, investiga-
tion of anisotropic layers, and the inverse problem for aniso-
tropic materials with diagonal permittivity and permeability
tensors.)
A. K. Bhattacharyya and D. L. Sengupta, Radar Cross Section
Analysis and Control, Artech House, BostonLondon, 1991.
[The book concentrates on deliberate changes of RCS (en-
hancement and reduction). It contains a useful table (p. 108)
that represents in concise form the comparison of different
methods available for RCS analysis, with discussion of their
advantages, disadvantages, and possible applications. It also
complements other books by inclusion of the Maluzhinets
function, which plays a fundamental role in the theory of
scattering by absorbing objects with sharp edges or with
impedance discontinuities. The computer code for the calcula-
tion of this important function is provided in the appendix of
the book.]
E. F. Knott, J. F. Schaffer, and M. T. Tuley, Radar Cross Section,
2nd ed., Artech House, BostonLondon, 1993. [This book
presents updated material that covers most aspects of RCS:
radar fundamentals, radar detection, RCS prediction, RCS
reduction, radar absorbing materials, and RCS measurements.
Chapter 14 can be especially useful for a brief review. It
contains a table (p. 562) with RCS estimations, as well as the
RCS data presentation formats and data reduction recipes.]
P. Ya. Umtsev, Comments on diffraction principles and
limitations of RCS reduction techniques, Proc. IEEE
84:18301851 (1996). (RCS reduction techniques are discussed
briey from the physical point of view. Attention is concen-
trated on the physical structure of radar waves scattered
from large objects. Possible passive and active techniques to
control and reduce reected beams, rays, and shadowradiation
as well as potential limitations of these techniques are con-
sidered. In particular, it is emphasized that grazing reected
rays and shadow radiation cannot be eliminated by absorbing
materials.)
R. C. Hansen, ed., Geometrical Theory of Diffraction, IEEE Press,
New York, 1981. (This book consists of a collection of articles. It
BACKSCATTER 493
contains key papers on GTD, asymptotic solutions of some
canonical problems, and applications-oriented papers.)
Many scattering objects contain nonmetallic materials, com-
posites, and various layered structures. To simplify the solu-
tion of scattering problems for such objects, it is often practical
to apply approximate boundary conditions. These conditions
are enforced on the external surface of the object and contain
important information about the internal structure of the
scattering object. As a result, this approximation allows one
to substantially reduce the spatial region under investigation.
The two following books present the development and applica-
tions of this approximation technique.
T. B. A. Senior and J. L. Volakis, Approximate Boundary Condi-
tions in Electromagnetics, The Institution of Electrical Engi-
neering, London, 1995.
D. J. Hoppe and Y. Rahmat-Samii, Impedance Boundary Condi-
tion in Electromagnetics, Taylor & Francis, Washington, DC,
1995.
P. Ya. Umtsev, Theory of Edge Diffraction in Electromagnetics,
Tech Science Press, Encino, CA, 2003. [Among the fundamen-
tal results of this theory are the following: (1) The theory of
black bodies that is important for understanding of physics of
diffraction, especially of scattering at large bistatic angles.
This theory also determines the diffraction limit for reduction
of the total scattered power by application of radar absorbing
coatings. (2) PTD estimations of RSC for bodies of revolution.
(3) Extension of PTD for scattering objects with concave ele-
ments. (4) Analysis of diffraction problems (wedge, strips, di-
poles) based on the parabolic equation method. (5)
Investigation of multiple edge diffraction at a single strip and
at an open resonator formed by two strips. Exact and asymp-
totic expressions are found for the current on the strip and for
the far eld that are valid for arbitary directions of incidence
and observation. These expressions allow the calculation of the
current and the far eld with arbitrary high precision].
BANDPASS FILTERS
JAIME RAMIREZ-ANGULO
New Mexico State University
Las Cruces, New Mexico
A bandpass lter is an electrical device that passes the
spectral components of an electrical signal around a cer-
tain frequency f
0
with little or no attenuation, while it
mostly rejects other spectral components (Fig. 1). Band-
pass lters are common building blocks of many electronic
systems. They nd applications in diverse elds such as
communications, audio, instrumentation, and biomedi-
cine. To illustrate a typical application of a bandpass l-
ter, consider the simultaneous transmission of n signals in
one communications channel using frequency multiplex-
ing techniques. The signals can share the same channel
because their individual frequency spectra are shifted so
that they occupy nonoverlapping regions (or bands) of the
frequency spectrum. Frequency modulation is used to
shift the baseband spectrum of each of the signals (cen-
tered originally at zero frequency or dc) to center frequen-
cies f
01
, f
02
, y, f
0n
(Fig. 2). In order to avoid overlapping,
the center frequencies must have a minimum separation
of at least the bandwidth of the signals. Bandpass lters
are used on the receiver side in order to select or lter
out only one of the signals transmitted in the channel. A
process called demodulation is used to shift the spectrum
of the selected signal back to baseband (to a zero center
frequency).
1. IDEAL BANDPASS FILTERS
Ideally, a bandpass lter is characterized by its center fre-
quency f
0
and its bandwidth BW. The bandwidth is dened
in terms of upper and lower passband edge frequencies f
p1
and f
p2
, respectively, and given by BWf
p2
f
p1
. Typical-
ly, f
p1
and f
p2
specify frequency points at which the gain
decreases by a factor

2
p
. The frequency response plot of
Fig. 1 shows the gain A as a function of frequency f of an
ideal bandpass lter. An important parameter of a band-
pass lter is its selectivity factor S
F
, which is expressed by
S
F
f
0
/BW. The selectivity is a measure of the lters abil-
ity to reject frequencies outside its passband. It is also
a relative measure of the reciprocal bandwidth. Highly
selective lters (also denoted narrowband lters) have a
bandwidth that constitutes only a small fraction of f
0
. Nar-
rowband lters with large selectivity factors (e.g., S
F
450)
are commonly required in communication systems. For
ideal bandpass lters, the frequency range is divided into
three regions: the passband region where the gain is ap-
proximately constant ( f
p1
ofof
p2
) and the lower and up-
per stopbands ( fof
p1
and f4f
p2
, respectively), where the
gain is ideally zero or the attenuation or reciprocal gain is
innite. The ideal bandpass lter has sharp transitions
between the passband and the two stopbands.
A
f
f
p1
f
0
f
p2
BW
T( f )
Figure 1. Frequency response of ideal bandpass lter.
f
f
0n
f
03
f
02
f
01
A
Figure 2. Spectrum of n signals in a communication channel.
494 BANDPASS FILTERS
2. PRACTICAL BANDPASS FILTERS
The typical frequency response of a practical passband
lter is shown in Fig. 3. It has a relatively constant gain
possibly with uctuations (ripple) with maximum ampli-
tude A
p
within the passband. The passband edge frequen-
cies f
p1
and f
p2
are dened, in this case, as the minimum
and maximum frequencies with gain, AA
p
. The center
frequency f
0
is commonly the center of geometry of the
passband edge frequencies: f
0

f
p1
f
p2
p
. In a practical l-
ter, there is a gradual transition between the passband
and the upper and lower stopbands. Therefore, besides the
passband and stopband there are also two transition
bands. Two frequencies f
s1
and f
s2
, each denoted lower
and upper stopband edge frequencies, respectively, dene
the stopbands in a practical bandpass lter. Stopband re-
gions are characterized by a small gain with maximum
value, 1/A
s
, where A
s
is the stopband attenuation. Pass-
band ripple A
p
and stopband attenuation A
s
are commonly
specied in a logarithmic unit denoted decibel according to
A
s
(dB) 20 log A
s
. Practical bandpass lters can have
monotonic decreasing gain or equal gain uctuations
(ripple) within both the passband and the stopband.
3. MATHEMATICAL CHARACTERIZATION OF A
BANDPASS FILTER
The simplest bandpass lter is the second-order lter. It is
characterized by the inputoutput relation or transfer
function:
T f
jBWf
f
2
0
jBWf f
2
1
This transfer function is a complex function with magni-
tude and phase. The magnitude (frequency response) ob-
tained from Eq. (1) has the form
jT f j
BWf

f
2
0
f
2

2
BWf
2
q 2
The polezero plot (in the complex frequency plane also
known as the s plane) of a second-order bandpass lter is
shown in Fig. 4. It has a zero at the origin (shown as a
circle) and two complex conjugate poles (shown as marks).
The selectivity is associated with the relative distance of
the poles from the imaginary axis in the s plane. The dis-
tance from the origin to the poles corresponds to the center
frequency in radians, o
0
2pf
0
.
4. BANDPASS FILTER IMPLEMENTATIONS
Depending on the center frequency and selectivity, there
exist several possible implementations of bandpass lters.
In general, the higher the selectivity and the center fre-
quency, the more complex the lter becomes. Two very
important aspects are miniaturization and compatibility
with very large scale integration (VLSI) systems technol-
ogy. These have dictated the evolution of bandpass lters
and of most other electronic circuits. Compatibility refers
to the fact that the lter can be fabricated on a CMOS in-
tegrated circuit as a part of a VLSI system in CMOS tech-
nology. This technology allows for the fabrication of very
complex analog and digital electronic systems with hun-
dreds of thousands or even millions of transistors on a
single integrated circuit. CMOS compatibility is crucial
in order to reduce manufacturing costs and to increase
system reliability. Automatic tuning and low power con-
sumption are also important aspects for bandpass lters.
Tuning consists in the process of adjusting the lters re-
sponse to compensate for unavoidable fabrication toler-
ances in the values of the lter elements. These tolerances
cause the lters response to deviate from the ideal (nom-
inal) response. Low power is important for portable equip-
ment like cellular phones and bioimplantable devices.
4.1. Passive Filters
For many years, most lters were exclusively implement-
ed as passive RLC circuits using resistors, inductors,
and capacitors [1]. Figure 5 shows an example of a reso-
nant RLC circuit used as a bandpass lter. This lter is
characterized by the transfer function of Eq. (1) with
s-plane

Figure 4. Polezero plot of second-order bandpass lter.

A
s
A
p
A
BW
f
f
0
f
p1
f
p2
f
s2
f
s1
T( f )
Figure 3. Practical bandpass lter.
R
v
IN
v
OUT
C
+

L
Figure 5. Second-order RLC bandpass lter.
BANDPASS FILTERS 495
f
0
0
1=2p 1=

LC
p
and BW1/(CR). The transfer function
of this circuit has the polezero plot shown in Fig. 4. Ce-
ramic lters based on bulk electromechanical resonance
phenomena in piezoelectric materials have been used in
the radio frequency range specially for requirements of
high selectivity and accurate center frequency. Inductors
have several problems: (1) they have large physical di-
mensions and cannot be integrated unless their value is
very small, (2) they are temperature-dependent, and (3)
their value is subject to large manufacturing tolerances.
They also have poor characteristics at low frequencies,
and their value tends to change with time. RLC passive
lters cannot be automatically tuned at the manufactur-
ing stage. Additionally, RLC and ceramic lters are not
compatible with CMOS VLSI technology.
4.2. RC Active Filters
In the 1960s, RC active lters using operational ampli-
ers, capacitors, and resistors led to the implementation of
highly selective low-cost lters without inductors (Fig. 6).
These lters are known as RC active lters [2]. They con-
stituted the rst step toward lter miniaturization. RC
active lters were mass-produced in miniaturized form
using thin-lm and/or thick-lm technology. These lters
still required individual tuning. This was automatically
done using computer-controlled laser beams that burn
sections of resistor material to change the resistance and
achieve the desired lter response. Within a short time,
RC active lters replaced most inductor-based lters in
the audio frequency range.
4.3. Monolithic Analog Filters
In the mid-1980s, RC active lters were replaced by fully
integrated switched-capacitor lters [3]. These use only
switches, operational ampliers, and capacitors and can
be fabricated in CMOS VLSI technology (see Fig. 7).
Switched-capacitor lters require no tuning and can op-
erate at center frequencies up to a few hundred kilohertz.
Starting in the mid 1970s, surface acoustic wave (SAW)
ltersanother family of lterswere developed to
implement lters for center frequencies spanning from
10 to several hundred megahertz. These lters are min-
iaturized lters that also require no tuning. They are
based on the controlled propagation of electromechanical
waves on the surface of a piezoelectric crystal with very
small dimensions. SAW lters are not compatible with
CMOS technology. But because of their excellent perfor-
mance characteristics and low cost, they are currently
being used to implement the intermediate frequency lter
of most televisions.
More recently, electronically tunable elements in
CMOS technology have allowed the integrated circuit im-
plementation of active lters for center frequencies up to a
few tens of megahertz. These lters are compatible with
CMOS VLSI technology. Most of these implementations
are based on electronically tunable elements called oper-
ational transconductance ampliers (transconductors or
OTAs) and include on-chip automatic tuning circuitry [4].
The tuning circuit continuously (or periodically) monitors
the lters response and electronically adjusts the gain of
the OTAs to tune the frequency response of the bandpass
lter. This family of lters is known as OTA-C lters (see
Fig. 8). They require no tuning and can be integrated as
part of a VLSI system [5]. Figure 8a shows a conventional
voltage mode lter, while Fig. 8b shows a current mode
lter where input, output, and intermediate variables are
represented by electrical currents rather than voltages
as it is done in conventional voltage mode lters [6]. In
analog signal processing systems, the range above
100 MHz is still based, to a large extent, on passive RLC
lters. At these frequencies, high-quality, small-dimen-
sion inductors are available. A very recent trend for
the implementation of high-frequency high-selectivity
+

v
OUT
v
IN
Figure 7. Second-order switched-capacitor lter.
+
v
OUT
v
IN

Figure 6. Second-order RC active lter.

v
IN
v
out
i
out
i
IN
+
+

+
(a)
(b)

+
+

+
Figure 8. Second-order OTA-C bandpass lter: (a) voltage-mode
implementation; (b) current-mode implementation.
496 BANDPASS FILTERS
integrated analog lters is based on microelectromechan-
ical (MEM) structures, which are based on electromechan-
ical resonance. This approach has shown potential for the
implementation of bandpass lters operating in the radio
frequency range and with very high selectivity factors
similar to those achievable with ceramic and SAW struc-
tures. These MEM-based lters, unlike the SAW lters,
are CMOS VLSIcompatible.
4.4. Digital Filters
Currently, the availability of low-cost digital circuitry in
VLSI systems has made it convenient to replace many
traditional analog functions by digital ones and to process
signals in the digital domain rather than in analog form.
Successful implementation of digital lters for center fre-
quencies up to a few megahertz, as standalone units or as
part of a VLSI system has been achieved. Digital lters
are fully integrated lters that are compatible with CMOS
technology and require no tuning [7]. Their characteristics
(f
0
, BW, or Q) can be easily reprogrammed for many dif-
ferent applications. A digital lter consists basically in
various addition, multiplication, and delay operations ap-
plied to an ordered sequence of numbers that represent
the digitized values of a signal. These digitized values are
obtained by sampling the signal at regular time intervals
(t T, 2 T, 3 T, y) and transforming the sampled values
into binary codes. In digital lters, present values x(n)
x(nT) and past values x(n1) x((n1)T), x(n2)
x((n2)T) . . . of both input and output signals are pro-
cessed. The steps performed on these digitized values lead
to an output sequence y(n), y(n1), y(n2), . . . , which
can be transformed by means of a digital-to-analog
converter into a ltered output signal y(t). A digital lter
is, therefore, an algorithm or sequence of mathematical
steps that relates input and output sequences by means of
multiplication, addition, and delay operations. A digital
lter is mathematically characterized by a difference
equation. There are two types of digital lters: nite im-
pulse response (FIR) lters, where the current value of the
output signal y(n) depends only on the current and past
values of the input signal; and innite impulse response
(IIR) lters, where the current value of the output signal
depends also on past values of the output signal. For
example, the difference equation of a second-order FIR
lter is expressed by
yn a
0
xnT a
1
xn 1 a
2
xn 2 3
while the difference equation of a second-order IIR digital
lter has the form
yn a
0
xn a
1
xn 1 a
2
xn 2
b
1
yn 1 b
2
yn 2
4
where a
0
, a
1
, a
2
, b
1
, and b
2
are multiplying coefcients;
x(n) and y(n), stand for x(t nT) and y(t nT), are the
current values of the input and output signals; and
x(n1), x(n2), y(n1), and y(n2) correspond to the
two previous values of the input and output signals, re-
spectively. The values of the multiplying coefcients and
the sampling frequency, f
s
1/T, determine the selectivity
and center frequency of the digital lter and can be easily
reprogrammed. Figures 9a and 9b illustrate the block
diagram of second-order FIR and IIR lters, respectively.
High-speed digital lters can be implemented using
special-purpose VLSI hardware in the form of digital sig-
nal processors [6] or as relatively low-speed lters using
software in general-purpose digital systems such as com-
puters or microprocessors.
BIBLIOGRAPHY
1. A. B. Williams and F. J. Taylor, Electronic Filter Design Hand-
book: LC Active and Digital Filters, McGraw-Hill, New York,
1988.
2. M. K. Van Valkenburg, Analog Filter Design, Holt, Reinhart,
Winston, Forth Worth, TX, 1982.
3. R. Schaumann, M. S. Ghausi, and K. R. Laker, Design of An-
alog Filters: Passive, Active RC and Switched Capacitors,
Prentice-Hall, Englewood Cliffs, NJ, 1990.
4. J. Silva-Martinez, M. Steyaert, and W. Sansen, High Perfor-
mance CMOS Continuous-time Filters, Kluwer Academic Pub-
lishers, Norwood, MA, 1995.
5. Y. P. Tsividis and J. O. Voorman, Integrated Continuous Time
Filters: Principles Design and Applications, IEEE Press, Pis-
cataway, NJ, 1992.
x(n)
y(n)
+ +
+
+
T T
a
2
a
1
a
0
x(n)
y(n)
T T
a
2
a
1
a
0
(b) (a)
T T
b
1
b
2
+ +

+
+
Figure 9. Second-order digital lters: (a) FIR l-
ter; (b) IIR lters.
BANDPASS FILTERS 497
6. J. Ramirez-Angulo, K. Sanchez-Sinencio, and M. Robinson,
Current mode continuous time lters: two design approaches,
IEEE Trans. Circ. Syst. 39:337341 (1992).
7. R. Higgins, Digital Signal Processing in VLSI, Prentice-Hall,
Englewood Cliffs, NJ, 1990.
BANDSTOP FILTERS
CHIU H. CHOI
University of North Florida
Jacksonville, Florida
A bandstop lter (also known as band-reject, band-
elimination, or notch lter) suppresses a band of frequen-
cies of a signal, leaving intact the low- and high-frequency
bands. A bandstop lter specication can be expressed as
shown in Fig. 1. In the stopband from f
3
Hz to f
4
Hz, the
attenuation is at least A
s
dB. In the passbands (below f
1
Hz and above f
2
Hz), the maximum attenuation is A
p
dB.
The bands from f
1
to f
3
and from f
4
to f
2
are called the
transition bands. The lter requirement is said to be
geometrically symmetric if f
1
f
2
f
3
f
4
.
An approach to designing a circuit (a bandstop lter)
with a frequency response that satises the bandstop
requirements shown in Fig. 1 is described below. It consists
of two steps: the approximation of the requirements by a
transfer function and the synthesis of the transfer function.
In the approximation part of the design process, it is
desirable to nd a transfer function with a frequency
response that satises the bandstop requirements. To
nd that transfer function, rst convert the bandstop
requirements into the normalized lowpass requirements.
For the case that the bandstop requirements are sym-
metric, the corresponding normalized lowpass require-
ments are shown in Fig. 2. The normalized passband
frequency F
p
1 and the passband attenuation is A
p
dB.
The normalized stopband frequency is
F
s

f
2
f
1
f
4
f
3
and the stopband attenuation is A
s
dB.
With such lowpass requirements, we can obtain the
corresponding lowpass transfer function T
LP
(s). (See LOW-
PASS FILTERS for more information about how to obtain the
transfer function.) The bandstop lter transfer function
T
BS
(s) is obtained by making the transformation
T
BS
s T
LP
sj
s
Bs
s
2
2pf
0

2
where B is the bandwidth of the bandstop lter dened
as
B2p f
2
f
1

and f
0
is the center frequency of the bandstop requirement
dened as
f
0

f
1
f
2
p

f
3
f
4
p
To use this method when the requirement is not sym-
metric for the case where f
1
f
2
4f
3
f
4
, we form a more
stringent requirement by either decreasing f
2
or increas-
ing f
4
, so that the symmetrical condition is met. The
bandstop transfer function that satises the new require-
ments must also satisfy the original requirements. In case
f
1
f
2
of
3
f
4
, we either increase f
1
or decrease f
3
and then
apply the same procedure.
A simple example is provided to illustrate this con-
cept. For the bandstop requirements A
s
25 dB, A
p

3.01 dB, f
1
1kHz, f
2
100 kHz, f
3
8kHz, and f
4

12.5 kHz; the corresponding normalized lowpass require-

ments are A
s
25 dB, A
p
3.01 dB, F
p
1, and F
s
22.
Choosing a single-pole Butterworth approximation, the
lowpass transfer function for the normalized lowpass
requirements is
T
LP
s
1
s 1
which meets the stopband requirements easily. The band-
stop transfer function is obtained by the transformation
T
BS
s
1
S1

S
2p100 10
3
110
3
s
s
2
2p10010
3
2p110
3

Frequency (Hz)
f
1
A
p
A
p
A
s
f
3
f
4
f
2
A
t
t
e
n
u
a
t
i
o
n

(
d
B
)
Figure 1. Bandstop lter specication.
Normalized frequency
1
A
p
A
s
A
t
t
e
n
u
a
t
i
o
n

(
d
B
)
Figure 2. Normalized lowpass requirements.
498 BANDSTOP FILTERS
which simplies to
T
BS
s
s
2
3:94810
9
s
2
6:22010
5
s 3:94810
9
Note that the single-pole lowpass function has been
transformed to a two-pole bandstop function. The above
bandstop transfer function is in the so-called biquadratic
form, which is an expression of the form
s
2
a
1
a
2
s
2
b
1
s b
2
1
There are a number of ways to synthesize and biquadratic
function as an active network, such as the Friend biquad
circuit [1,5], the Boctor circuit [2], and the summing four-
amplier biquad circuit [3,5]. The Friend or Boctor circuit
uses one operational amplier. The summing four-ampli-
er biquad circuit is much easier to tune. When a
1
0, the
Bainter circuit can be used. The Bainter circuit [4] is
shown in Fig. 3. For higher performance circuits, see Ref. 5
for the description of different bandstop circuit topologies.
The transfer function is
V
out
s
V
in
s

s
2

R
12
R
11
R
21
R
31
C
1
C
2
s
2

1
R
31
C
1

1
R
32
C
1

s
1
R
22
R
31
C
1
C
2
Comparing with Eq. (1), we obtain
a
1
0; a
2

R
12
R
11
R
21
R
31
C
1
C
2
b
1

1
R
31
C
1

1
R
32
C
1
; b
2

1
R
22
R
31
C
1
C
2
Choose C
1
C
2
1, R
12
/R
11
K, R
31
R
32
. Solving for the
other values, we obtain
R
21

Kb
1
2a
2
; R
22

b
1
2b
2
; R
31

2
b
1
The impedance scaling method can be used to scale the
values of R and C into the practical ranges. In general, a
higher-order bandstop transfer function can be factorized
into a product of biquadratic functions. Each biquadratic
function can be synthesized by using the Bainter or other
circuits. By cascading all the circuits together, the band-
stop lter is realized.
BIBLIOGRAPHY
1. J. J. Friend et al., STAR, An active lter biquad section, IEEE
Trans. Circuits Syst. CAS-22:115121 (1975).
2. S. A. Boctor, Single amplier functionally tunable low-pass
notch lter, IEEE Trans. Circ. Syst. CAS-22:875881 (1975).
3. G. Daryanani, Principles of Active Network Synthesis and
Design, Wiley, New York, 1976.
4. J. R. Bainter, Active lter has stable notch and response can be
regulated, Electronics 115117 (Oct. 1975).
5. G. Moschytz and P. Horn, Active Filter Design Handbook,
Wiley, New York, 1981.
BiCMOS DEVICES AND RF INTEGRATED
CIRCUITS
YUPENG CHEN
XIAOMIN YANG
THOMAS X. WU
University of Central Florida
Orlando, Florida
1. INTRODUCTION
Since the late 1980s, the wireless communication market
has expanded explosively mainly driven by increasing
needs for high-performance, low-cost personal wireless
communication equipments. Worldwide intensive re-
search is being done in process technologies, circuit
designs, system architectures, and communication stan-
dards to meet such demands. Technology advances in sil-
icon germanium (SiGe) heterojunction bipolar transistors
(HBTs) have created innovations in the semiconductor in-
dustry. The RF performance improvement of SiGe HBT by
adding germanium to the base of the device will have a
strong impact on modern radiofrequency integrated cir-
cuit (RFIC) design. Since IBM introduced the rst com-
mercial silicon germanium (SiGe) BiCMOS process in
1997, more and more SiGe products have emerged in the
market, such as cellphones, wireless LAN, global position-
ing system (GPS) ICs, TV turners, hard-disk-drive ICs,
and optical carrier ICs. The estimated worldwide SiGe
market growth is expected to exceed $2.5 billion by the
end of 2006 [1].
In this article, BiCMOS technology development focus-
ing on the most recent achievements of SiGe BiCMOS
technology is reviewed rst. Several important gures of
merit of SiGe device for RFIC designers such as f
T
, f
max
,
BV
CEO
(collectoremitter breakdown voltage with open
base), and NF
min
are discussed in detail. Next, from the
RFIC designers point of view, we begin with the stringent
V
in
R
11
R
21
C
2
R
22
R
31
C
1
R
32

+
+

R
12
V
out
Figure 3. Bainter circuit.
BiCMOS DEVICES AND RF INTEGRATED CIRCUITS 499
system requirements of typical modern wireless commu-
nication standards. Commercial available technology op-
tions for RFIC are compared according to performance,
power consumption, level of integration, and cost, with
special emphasis on GaAs, RF CMOS, and SiGe BiCMOS.
As a result, SiGe BiCMOS is an optimal device technology
for implementing high-performance, low-power, and high-
integration RFIC products. As important integration
elements in the SiGe BiCMOS process, passive devices
are also explored with respect to their impacts on RFIC
design. This is expected to be followed by the design of key
wireless system building blocks using SiGe BiCMOS
technology. The most recently reported results of SiGe
low-noise ampliers (LNAs), power ampliers (PAs), and
integrated voltage-controlled oscillators (VCOs) are re-
viewed. Finally, future trends of RF integrated system
adopting SiGe BiCMOS technology are discussed with re-
spect to SiGe HBT technology (e.g., SiGe:C), integration
level [e.g., RF SoC (RF system on chip)], and performance
of on-chip passive devices (e.g., substrate transfer tech-
niques).
2. BiCMOS DEVICES
It is well known that bipolar devices offer superior perfor-
mance attributes for analog and RF circuits compared
with plain CMOS [2]: more reliably modeled, less compo-
nent variation, exhibiting better device-to-device match-
ing, lower noise, increased bandwidth, easier matching
to off-chip RF passive components, higher gain for a
given layout size and power consumption level, and
greater transconductance for a given current level. CMOS
devices are also preferred because of their excellent
switching performance with very high OFF-state resistance
and lack of voltage drop when in the ON state. BiCMOS is a
technology combining MOSFETs and bipolar devices into
integrated circuits, which was proposed as early as 1969
[3]. Since then, BiCMOS technology has been applied to
digital logic applications, which utilize the characteristics
of high-speed bipolar and low-power CMOS on the same
chip and RF and mixed-signal applications, which utilize
bipolar RF and analog performance and CMOS digital in-
tegration capability. But as CMOS power supply voltage
scaled down, the performance gap between bipolar and
CMOS in digital logic began to narrow. Consequently,
since the early 1990s there have been very few reports
on developing BiCMOS for digital applications [4]. As
inspired by performance improvement of GaAs technolo-
gy using heterojunction structure, germanium is intro-
duced to the base of silicon BJT to form silicon germanium
(SiGe) heterojunction bipolar transistors (HBTs) with
enhanced RF performance. Since SiGe-based BiCMOS
technology is currently a topic of great interest for
research and wireless products development, our discus-
sions will focus mainly on SiGe BiCMOS technology and
its impact on RFIC design.
Table 1 demonstrates technology milestone for SiGe
BiCMOS early development in IBM, a leader in SiGe
BiCMOS technology. After that IBM introduced the rst
commercial silicon germanium (SiGe) BiCMOS process in
1997.
The essential operational differences between the SiGe
HBT and the Si BJT are best illustrated by considering a
schematic energy band diagram (Fig. 1). Two approaches
are generally applied [7]. In the rst approach, a graded
germanium prole creates an electric eld in the base re-
gion to accelerate the carriers that are diffusing from the
emitter to the collector junction. In the second approach,
the mobility of the minority carriers is increased in the
base and the base resistance is reduced. Both approaches
result in a higher f
T
and lower base resistance.
A scanning electron microscopy (SEM) cross-sectional
view of a 0.2-mm self-aligned selective epitaxial growth
(SEG) SiGe HBT by ultrahigh vacuum/chemical vapor
deposition (UHV/CVD) is shown in Fig. 2 [8], including
an enlarged view of the active region, the key part of the
SiGe HBT.
Device parameters such as f
T
, f
max
, NF
min
, and BV
CEO
are important gures of merit considered by RFIC design-
ers for specific applications. f
T
(short-circuit unity current
Table 1. SiGe BiCMOS Technology Early Development History from IBM Onward
Year Key Technology
1982 UHV/CVD growth technique developed
1986 UHV/CVD epitaxial base silicon transitors fabricated
1987 SiGe based HBTs fabricated using MBE
1988 Graded-baseSiGe polyemitter HBTs demonstrated, utilizing learning of double-polysilicon bipolar transistors
1993 Analog LSI circuit (1gsps, 12bits DAC) demonstrated, utilizing CMOS base processing
1993 SiGe HBTs with f
T
4100GHz fabricated
1994 BiCMOS process in 200-mm CMOS fabricated, aligning processing with CMOS base
1996 RF technology with passive components, models, and design kit
Source: Ref. 5.
e
Ec
Ev
n+ Si
Emitter
n Si
Collector
p SiGe
Base
p Si
Ge
h+
=Eg, Ge(X = Wb)Eg, Ge(X = 0)
Drift Field !
Eg, Ge(X = 0)
Eg, Ge(grade)
Figure 1. Engineered band diagram of SiGe HBT [6].
500 BiCMOS DEVICES AND RF INTEGRATED CIRCUITS
gain frequency) and open emitter) to inuence the dynam-
ic range of device operation. The product of f
T
and BV
CEO
is referred to as the Johnson limit [12], which is an im-
portant gure of merit for high-frequency device. Many
commercial foundries offer SiGe HBTs with several differ-
ent BV
CEO
values by using selective collector implantation
optimized for both high-speed and wide-dynamic-range
applications. We will discuss it later in SiGe RFIC power
amplier design.
Table 2 summarizes major SiGe HBT gures of merit
that are important for RFIC applications. In addition, as
with CMOS technologies, progressive vertical and lateral
scaling of SiGe HBTresults in better performance at lower
power as indicated in Table 2.
Compared with IIIV technologies, SiGe HBTs based
on silicon substrate can inherently integrate with silicon
CMOS, and standard silicon CMOS processes are by far
one of the cheapest IC processing technologies available.
Obviously, it is a cost-effective and processing-compatible
way for SiGe BiCMOS processes to integrate the SiGe
HBT into existing silicon CMOS core processes. Thats ex-
actly the method adopted by successful SiGe BiCMOS
foundries. Historically, two different integration approach-
es have been used in SiGe BiCMOS processes [13]:
base-during-gate (BDG) integration and base-after-gate
(BAG) integration. In the former approach, the HBTs are
fabricated during the CMOS process (e.g., the IBM 0.5-mm
SiGe BiCMOS process). In BAG, the HBTs are fabricated
after the CMOS process to decouple the CMOS thermal
cycle from the bipolar (e.g., IBM 0.25- and 0.18-mm SiGe
BiCMOS) processes. The BAG approach is used in the
most advanced SiGe BiCMOS technology such as 0.25 and
0.18 mm generations because its difcult to fabricate HBT
during a high thermal cycle of CMOS in these generations.
f
max
(maximum unity power gain frequency) are the pa-
rameters determining the frequency bandwidth of the de-
vice. f
T
and f
max
have been significantly improved over
200 GHz in the state-of-art SiGe BiCMOS 0.13-mm gener-
ation as shown in Fig. 3 [9].
In practical design, a good rule for high-quality RFIC
design is f
design
o
1
10
f
T
. According to this rule, when f
T
is
over 200GHz, SiGe HBT can be used for ultra-high-speed
optical network applications that were previously the
realm of IIIV technologies.
The noise gure (NF), a standard metric of system noise
performance, is dened as the degradation of signal-to-
noise ratio (SNR) of an input signal as it passes through
the system; NF is minimized when the device is presented
with an optimized source reactance. A simplied formula
for a SiGe HBT device is given as [6]
NF
min
1
1
b

2g
m
r
b
p

1
b

f
f
T

2
s
1
where b is the device current gain, g
m
is the device trans-
conductance, r
b
is the base resistance, and f is the opera-
tion frequency. From this formula, the key to lower the
noise gure is to reduce r
b
and increase f
T
[6]. Most re-
cently reported SiGe transistors have f
T
> 350GHz [10]
and NF
min
0:4dB at 10 GHz [11].
BV
CEO
refers to the collectoremitter breakdown volt-
age with an open base. This parameter is quoted in device
specications of IC technology together with BV
CBO
(col-
lectorbase breakdown voltage with an Passive elements
such as inductors, capacitors, and resistors are also inte-
grated into a typical SiGe BiCMOS process. Metallization
and back-end processing in IC processes produce several
metal layers utilized to build these passive components. A
Figure 2. SEM cross-sectional view of a self-aligned SEG SiGe
HBT [8].
Table 2. IBM SiGe HBT Development across Four
Generations
Parameter 0.5mm 0.25mm 0.18mm 0.13mm
Beta 100 100 200 400
V
A
(V) 65 75 120 4150
BV
CEO
(V) 3.35 3.35 2.5 1.9
BV
CBO
(V) 10.5 10.5 7.5 6
Peak f
T
(GHz) 47 47 120 210
Peak f
max
(GHz) 65 65 100 185
NF
min
(dB) 0.8 0.8 0.4 o0.3
Source: Refs. 13 and 14.
V
CB
= 1.0 V
0.50 m
0.18 m
0.13 m
f
T
f
T

a
n
d

f
m
a
x
_
U

(
G
H
z
)
f
max
300
250
200
150
100
50
0
1.E-05 1.E-04 1.E-03 1.E-02
Collector current (A)
Figure 3. SiGe HBT f
T
/f
max
improvement across generations [9].
BiCMOS DEVICES AND RF INTEGRATED CIRCUITS 501
SEM cross-sectional view of a SiGe BiCMOS processes
with the passives and metallization is shown in Fig. 4 [8].
High-quality-factor (Q) on-chip passive components
are difcult to acquire in silicon substrate technology
compared with IIIV technology because of high silicon
substrate loss. For example, it has been particularly
challenging to achieve integrated inductors with Q values
of 1520 in the silicon process. Typical Q values for induc-
tors in standard BiCMOS process can maximally reach
about 10 (for inductances 41 nH) [13]. In modern ad-
vanced SiGe BiCMOS processes, two types of methods are
usually adopted to improve the performance of integrated
inductors by decreasing metal series resistance and
reducing substrate loss effects [14]: (1) additional thick
top metal layer, copper metallization, deep-trench isola-
tion, lightly doped substrate and (2) thick dielectric be-
tween inductor and substrate. Improved performance
of passives in the IBM 0.18-mm SiGe BiCMOS process
can be found in Table 3 [15]. We will discuss passive com-
ponents for RFIC design using SiGe BiCMOS technology
in Section 3.3.
3. BiCMOS RADIOFREQUENCY INTEGRATED CIRCUIT
DESIGN
In the late 1980s, the rst CMOS baseband analog and
digital chips for the rst-generation GSM (Group Special
Mobile, now known as Global System Mobile) were fabri-
cated in a 1.2-mm CMOS technology. Together with the
progress of silicon bipolar technology (cutoff frequency f
T
above 10 GHz) and worldwide spread use of mobile digital
phones, a new and exciting area, radiofrequency integrat-
ed circuit (RFIC) technology, was created [7]. RF design
moved from the PCB (printed circuit board) with discrete
components to large-scale integration in silicon technolo-
gy. Until recently, single-chip radio integration capabili-
ties based on SiGe BiCMOS technology have been
demonstrated [9]. A true single-chip radio supporting mul-
tiband, multistandard, and broadband communication
may not just be a dream in the near future.
As stated above, RFIC is a relatively new area in con-
trast to other types of analog or microwave circuits. RFIC
design has borrowed from both analog design techniques
used at low frequencies and high-frequency design tech-
niques used at microwave frequencies. However, RFIC de-
sign as a multidisciplinary area requires knowledge of
many other areas besides integrated circuits, such as
semiconductor device physics, wireless communication
standards, signal and communication theory, microwave
engineering, and system architectures.
From the circuit designer point of view, well-developed
low-frequency analog circuit design and analysis tech-
niques are fundamental knowledge references. Classical
circuit building blocks can be used as prototypes or start-
ing references for RFIC design. S parameters, shielding,
impedance matching, and standing-wave ratios, which are
important microwave system design concepts, are not to-
tally but partially adopted in RFIC design. With the in-
creasing of integration level of RFIC, new RFIC design has
gradually acquired its own features [7]. The 50-O imped-
ance matching culture has been stepped away. SPICE-like
tools and system-level simulation packages including
noise analysis have been adopted for highly integrated
RFIC due to high-frequency substrate crosstalk and par-
asitic coupling.
3.1. Representative Wireless System Requirements
The rapidly growing commercial wireless personal and
data communication market has been the strongest force
driving the rapid development of RFIC technology. Since
the early 1980s, the industry has undergone several gen-
erations of revolutionary changes. Figure 5 [16] schemat-
ically presents the scope of RFIC technologies in terms of
the interplay among commercial wireless communication
applications, available spectrum, and the kinds of elemen-
tal and compound semiconductors likely to be used.
From Fig. 5, we can nd that modern RFIC designers
are confronting with a wide range of system requirements
with respect to different kinds of wireless applications.
Table 4 compares some important specications of several
representative wireless communication standards.
Table 3. 0.18-lm SiGe BiCMOS Passive Device
Performance from IBM Onward
Component
Resistors R
s
(O/square)
a
TCR
b
(ppm/1C)
Subcollector 8.1 1430
n

diffusion 72 1910
p

diffusion 105 1430

p

polysilicon 270 50
p polysilicon 1600 1178
TaN 142 750
Capacitors C
p
(fF/mm
2
) V
CC
d
(5/ 5 ppm/V)
MIM
c
1 o45
MOS 2.6 7500/ 1500
Inductor L (nH) maximum Q at 5 GHz
Alspiral inductor 40.7 18
Varactor tuning range Q at 0.5GHz
CB junction 1.64:1 90
MOS accumulation 3.1:1 300
Source: Ref. 15.
a
Sheet resistance.
b
Temperature coefcient of resistance.
c
Metalinsulatormetal.
d
Voltage coefcient of capacitance.
Figure 4. SEM cross section of a SiGe BiCMOS processes [8].
502 BiCMOS DEVICES AND RF INTEGRATED CIRCUITS
It is clear that application-orientated system specica-
tions vary in frequency, transmitter output power, receiv-
er minimum sensitivity, and modulation schemes. In order
to address such a broad application market, RFIC design-
ers must satisfy constraints of not only performance but
also exibility and cost.
3.2. Technology Trends and Options for RFIC
Semiconductor technologies commercially available for
RFIC design can be categorized as silicon and IIIV com-
pounds according to substrate type. Several semiconduc-
tor devices are based on these materials, including the
silicon bipolar junction transistor (Si BJT), SiGe HBT, RF
CMOS, GaAs metal semiconductor eld-effect transistors
(MESFETs), and GaAs and InP high-electron-mobility
transistors (HEMTs). Each type of device has its own ad-
vantages or disadvantages as shown in Table 5.
Table 5 lists the key gures of merit normally consid-
ered by RFIC designers for various RF application stan-
dards. Obviously, no device technology does well in
everything. Fortunately, different RF components require
different device parameters. For example, the low-noise
amplier (LNA) needs devices with low-noise characteris-
tics, the voltage-controlled oscillator (VCO) employs de-
vices with low 1/f noise, and the power amplier (PA)
requires devices with high output power and high break-
down voltage. As a result, commercial wireless transceiv-
ers typically utilize a combination of different device
technologies.
From the early 1980s, GaAs-based IIIV compound
technology has been the choice for RF and microwave ap-
plications because of their high level of performance for
GHz (gigahertz-range) frequency. Compared with silicon,
they have higher electron mobility and saturation drift
velocity, which means high device speed and a low noise
gure. Because they are wide-bandgap materials, they
also have high breakdown voltage, which is useful in high-
power applications. The semiinsulating property of the
IIIV materials provides high device isolation and low di-
electric loss for high-frequency integrated circuits. This
property also decreases the loss of integrated passive com-
ponents. The increased quality factor of passive elements
can greatly improve the performance of the overall RF
system. However, IIIV technologies also suffer from
inherent problems such as cost, integration level, and
reliability issues. The IIIV material wafer is normally
smaller and more expensive than silicon. In addition,
much effort should be taken to reduce leakage current
and increase uniformity and yield. The overall high cost of
IIIV compounds is a tradeoff for their high performance.
Advances in silicon bipolar technology in the late
1980s, particularly in increasing the cutoff frequency, cre-
ated an epoch in silicon RFIC. Some low-GHz RF front-end
building blocks formerly implemented in GaAs technology
have been replaced by silicon bipolar devices, while GaAs
is still employed for low-noise ampliers, power ampli-
ers, and RF switching, which belong to the high-frequency
area near the antenna. But the vertical diffusion depth of
base and the mobility of carriers in silicon set a limit of
30 GHz cutoff frequency for the bipolar transistor, which
connes this bipolar component to the low-GHz front-end
and analog mixed-signal area [7]. As discussed before,
inspired by the heterobipolar approach in GaAs techno-
logies, the SiGe HBT was introduced as an enhanced
bipolar approach with superior RF performance.
Table 4. Comparison of Wireless Communication Standards
Standard System
Frequency
(MHz)
Data Rate
(mbps)
Peak Transmit
Power (dBm)
Minimum
Sensitivity (dBm)
Modulation
Scheme
DECT
a
Coreless 1800 0.552 24 83 GFSK
GSM Cellular 900 0.014 33 102 GMSK
CDMA Cellular 900/1900 0.014 23 104 BPSK/QPSK
WCDMA Cellular 2000 2 27 117 M-PSK
802.11b Data 2400 11 20 80 PSK-CCK
802.11a/802.11g Data 5200 54 23 82 OFDM
Bluetooth Data 2400 0.723 0 70 FSK
UWB Data 3100B10600 4100
a
Digital Enhanced Cordless Telecommunications.
0.8 GHz 2 GHz 5 GHz 10 GHz 28 GHz 77 GHz 94 GHz
GSM
CDMA
ISM
PDC
GPS
SAT
Radio
DCS
PCS
DECT
CDMA
WLAN
802.11b/g
HomeRF
Bluetooth
SAT
TV
WLAN
802.11a
SAT TV
WLAN
Hyperlink
UWB
LMDS
WLAN
AUTO
RADAR
Contraband
Detection
All Weather
Landing
Si
SiGe
GaAs
InP
Frequency
Figure 5. Commercial wireless communica-
tion applications [16].
BiCMOS DEVICES AND RF INTEGRATED CIRCUITS 503
Several practical RFIC design advantages for SiGe
HBT are summarized as follows [18]:
*
Low DC power consumption
*
High breakdown voltage
*
Improved noise performance with higher f
T
and
lower r
b
*
Superior performance in 1/f noise characteristics over
MOSFETs and MESFETs
*
Higher early voltage V
A
and bV
A
*
Improved linearity over that in SiBJTs
With RF performance comparable to that of GaAs tech-
nology, LNA and PA, high-RF performance demanding
blocks have been designed using SiGe technology rather
than GaAs technology. We will investigate such a trend in
detail later in this article. Besides the abovementioned
high device performance, SiGe HBT can potentially be in-
tegrated with CMOS on the same silicon substrate (SiGe
BiCMOS). This high integration ability creates the possi-
bility of implementing a true single-chip RF system.
In the mid-1990s, at approximately the same time
when SiGe HBT technology emerged, and driven by the
tendency of increasingly shrinking microprocessor size,
the scaling of CMOS resulted in marked improvement in
RF performance at submicrometer feature size. Conse-
quently, CMOS has become a viable option for RFIC de-
sign. Foundries also provide RF CMOS process optimized
for RF applications [19]. But RF CMOS technology
achieves the same performance level using novel circuit
techniques, which may need higher power consumption or
larger chip die size compared with SiGe BiCMOS.
The optimum technology option for RFIC in terms of
performance, cost, and market is the rst line for success-
ful RFIC product developments. But for practical RFIC
designers, this is always complicated by various applica-
tion performance requirements (Table 4) regarding differ-
ent building blocks, various device characteristics
(Table 5), system cost, and market factors. For a whole
RF system, a combination of technologies may achieve op-
timum system performance and cost compromise. In ad-
dition, this multidimensional decision function is strongly
related with the development of device, circuit, and sys-
tem technology. The cost of SiGe BiCMOS technology will
be gradually lowered with its wider applications. Combin-
ing performance, integration, and cost advantages togeth-
er, SiGe BiCMOS should be one of the most promising
technologies for RFIC design.
3.3. RFIC Passive Design in SiGe BiCMOS Technology
Passive components also play important roles in RFIC de-
sign. To reduce board-level components, passive devices
should be integrated on chip as active devices. Integrated
inductors, capacitors, varactors, and resistors have been
developed based on the existing SiGe BiCMOS fabrication
processes while avoiding any significant process altera-
tions.
3.3.1. Inductor. RFIC designs require high-perfor-
mance passive components, particularly the on-chip in-
ductors. Integration of spiral inductors has numerous
challenges. In particular, high Q (inductor quality factor),
high self-resonance frequency, and low cost must be bal-
anced in order to provide a competitive inductor. Circular
and rectangular spirals are two popular shapes for on-chip
inductors. The circular spiral is known to provide some-
what higher Q, while the rectangular version is more
area-efcient [20]. For both of these inductor congura-
tions, the bulk of the spirals are usually implemented in
the topmost available metal, and connection to the center
of the spiral is made with a crossunder of some lower level
of metal.
Losses in inductors are due primarily to currents ow-
ing in the inductors. Substrate and radiation losses also
exist. The quality factor Q and the self-resonance frequen-
cy are the most common gures of merit. Q is used to de-
scribe loss performance of inductors. Many factors can
inuence the Q value, such as operating frequency,
conductor metal conductivity, substrate resistivity, and
geometric characteristics of inductor [21,22]. Q is frequen-
cy-dependent, increasing with the frequency up to a peak
value, and drops at higher frequencies, owing to the fact
that the reactance of input impedance dominated by the
inductance at lower frequency rolls off at higher frequency
because of the parasitic capacitances, which consist of the
overlap capacitance between the spiral and underpass,
Table 5. Comparison of Various Semiconductor Device Technologies
Performance Metrics SiGe HBT Si BJT Si CMOS IIIV MESFET IIIV HBT IIIV HEMT
Frequency response 0 0
1/f and phase noise 0
Broadband noise 0 0
Linearity
Output conductance
Transconductance/area
Power dissipation 0
CMOS integration N/A
Output power density 0
Breakdown voltage 0
Passive performance
IC cost 0 0
Key: excellent; very good; 0 good; fair; poor.
Source: Refs. 6 and 17.
504 BiCMOS DEVICES AND RF INTEGRATED CIRCUITS
fringing capacitance between adjacent metal lines, and
substrate capacitance between metal layer and grounded
substrate.
In practical integrated inductor design, the geometric
characteristics, such as linewidth, line spacing, number of
turns, and inner diameter, should be carefully chosen to
optimize the inductor Q at operating frequencies. The rise
of Q is usually found with increasing linewidth, while
there is no wide variation in inductance. This is because of
the fact that the resistance of inductor decreases with in-
creasing linewidth. The inner diameter also affects the in-
ductor Q. The contribution of innermost turn to the total
inductance is very small. The main effect of the innermost
turn is on resistance. For this reason, hollow spirals are
commonly used in RFapplications to increase the Q factor.
Increasing the number of turns of a given inductor always
yields higher inductance. This increase in inductor design,
as the increase of linewidth, spacing between the turns, or
the inner diameter, may result in larger area. Usually, the
larger the area, the higher the parasitic capacitance, and
the lower the self-resonance frequency. Therefore, it is
important to ensure that the self-resonance frequency is
beyond the operating frequency range.
3.3.2. Capacitor. Besides inductors, capacitors with
high Q at the operating frequency are also essential for
the integration of RF circuits. The design goal is to achieve
both a small series resistance for a high Q and a high ca-
pacitance/area ratio to minimize the area consumption.
Three basic types of capacitors have been developed
in SiGe technology: MOS (polysilicon gated capacitors
on single-crystal silicon), PIP (polysiliconinsulator
polysilicon), and MIM (metalinsulatormetal)/MOM
(metaloxidemetal) capacitors. Each has its own place
for use in different application spaces depending on ca-
pacitance desired and performance at the application fre-
quency.
MOS capacitors are simply built from the gate oxide.
These devices consist of silicided gate polysilicon, thin
gate oxide, and FET well-doped silicon. They have the
highest capacitance density among all the capacitors ow-
ing to the ultrathin oxide. But the Q value of these capac-
itors is relatively poor because of the high resistance
of the doping. PIP capacitors are fabricated in double-
polysilicon BiCMOS processes [23]. The capacitor struc-
ture is formed using p

-doped gate polysilicon as the bot-

tom plate, a deposited oxide layer for the capacitor
dielectric, and silicided extrinsic base polysilicon as the
top electrode. The drawback of MOS and PIP capacitors is
low quality factor (Q) due to high-resistance plates and
capacitive losses at RF frequencies. This limits their use
at RF. They are typically used for power supply bypass
and decoupling. In contrast, MIM or MOM capacitors
can be built using two interconnect layers with a very
small series resistance for high Q. These capacitors are
achieved by using a thin oxide lm to form the capacitors
in place of the thick intermetal dielectrics. MIM or MOM
capacitors are preferred in narrowband applications, such
as resonators, lters, and tank circuits, where high Q is
needed [24].
3.3.3. Varactor. Voltage-variable capacitors (varactors)
are required for tuning circuit designs, such as voltage-
controlled oscillators (VCOs) for on-chip carrier frequency
generation. The key gures of merit for varactors are (1)
tunability (C
max
/C
min
), (2) C
V
linearity for VCO gain vari-
ation, (3) quality factor Q, (4) tolerance, and (5) capaci-
tance density. Two basic types of varactors are supported
by SiGe BiCMOS technology: MOS varactors and junction
varactors.
An accumulation-mode MOS varactor is achieved by
connecting the drain and the source only, while adding the
control voltage on the bulk [25]. The capacitance of this
varactor is high in accumulation and decreases sharply as
the device enters and goes further into depletion. This re-
sults in excellent tunability. However, it has the drawback
of a narrow range across which the capacitance varies
significantly from a low value to a high value. This in turn
could limit the tuning range of the VCOs based on such
varactors. Junction varactors are offered as an alternative
to provide wider capacitance variation range and also
more linearity. The collectorbase (CB) junction varactor
provides wider capacitance variation while maintaining a
relatively high Q.
3.4. Wireless System Key Building Blocks
Figure 6 gives us a general view of RF heterodyne trans-
ceiver system. Some important building blocks that are
necessary in all types of transceiver are listed in Table 6,
with their main features and technology requirements.
During practical RFIC design, several tradeoffs must be
made under product performance and cost constraints.
These tradeoffs have been summarized in the RF design
hexagon as shown in Fig. 7 [26]. We will discuss some
tradeoffs in the design of each building block.
3.4.1. Low-Noise Ampliers (LNAs). The LNA is one of
the key building blocks in the RF system. In the receiving
chain closest to the antenna, the LNA receives a variety of
signals coming from the antenna, and its function is to
boost the signal without adding much noise. Since the
LNA is at the rst stage of receiving chain, its noise per-
formance is the most critical for the entire receiver. Be-
cause of the wide amplitude range of the receiving signal,
low noise and high linearity are both important gures of
merit for LNA design [27]. The third-order input intercept
point (IIP3) is the most common parameter used to mea-
sure circuit linearity by applying the two-tone test. The
noise gure (NF) is often used to measure system noise
performance. Application-dependent system specications
such as minimum detectable signal (MDS) or minimum
sensitivity and spurious-free dynamic range (SFDR) can
be determined from IIP3 and NF [26].
From the perspective of LNA designers, a seemingly
simple structure with a few components means easy de-
sign, but the tradeoffs complicate the design. A matching
network needs to be designed to satisfy both noise and
power matching. The progressive decrease of supply
voltage and maximum allowed power consumption with
an almost constant dynamic range requirement of the
system leads to great challenges in modern LNA design.
BiCMOS DEVICES AND RF INTEGRATED CIRCUITS 505
Additionally, the LNA must provide enough gain to de-
crease the noise contributions from the following stages
and consequently minimize the overall system noise. How-
ever, maximum gain is often limited by system linearity
performance [28].
Noise gure (NF), DC power consumption (P
DC
), power
gain, and IIP3 are the most important design consider-
ations for the LNA. They can be summarized into two g-
ures of merit: gain/(P
DC
NF) and IIP3/P
DC
. The results of
comparing these two parameters for the best reported re-
sults in silicon and GaAs technology at 2 GHz [29] have
indicated the potential advantage of SiGe technology for
LNA, if DC power dissipation is a major consideration.
Table 7 summarizes the most recently reported results
of RF LNA using SiGe BiCMOS technology, demonstrat-
ing a competitive noise gure and linearity compared with
GaAs technology for various wireless applications.
3.4.2. Power Amplier (PA). The power amplier is the
last building block in the transmitter chain of a transceiv-
er system. Its function is to amplify the signal power to a
level that is high enough for signal propagation through
an appropriate antenna into the air. It is one of the most
challenging building blocks in a transceiver system.
The typical gures of merit of a practical power ampli-
er are efciency, linearity, output power, ruggedness, and
power gain. The most commonly used metric for efciency
is power-added efciency (PAE), which indicates how pow-
er drawn from the supply is efciently converted to RF
power delivered to the load. The linearity of PA is usually
specied as an adjacent-channel power ratio (ACPR) in
dBc (decibels with respect to carrier), a measure of the
spectral spillover due to amplier nonlinearity into an
adjacent frequency band by a digitally modulated wave-
form. Ruggedness is the ability to survive power mismatch
conditions at the output using specific DC and RF drive
conditions. Output power is dened by a wireless standard
(see Table 4) and often exhibits a wide range of variation
for mobile applications. Power gain is implicitly set by the
maximum power generated by the preceding stage of
the power amplier together with output power require-
ments [37].
The linearity varies with different signal modulation
schemes employed by different wireless standards (see
Table 6. Features and Requirements for Basic Wireless Building Blocks
Building Blocks Main Features Technology Requirements
LNA Low-noise gure Low NF
min
, low HBT R
b
Gain per stage High early voltage
High linearity High f
T
/f
max
Mixer High linearity High early voltage
High port isolation High f
T
/f
max
Carrier leakage Isolation technologies (DT)
Small mismatch
VCO Low-phase noise High-Q passives farther from carrier
Low-1/f-noise HBT closer to carrier
Tuning range (frequency/V) Varactor tuning range (C/V)
Low gain (K
v
) variation
Synthesizer Low-phase noise Same as VCO
Low reference spurs Small FET asymmetry
Power consumption Small gate length for low supply operation
Chip area Reduced area
PA Power-added efciency Low loss matching and interconnect circuits (high-Q inductor, thick metal)
Gain per stage Maximum available gain HBT
Robustness High-breakdown-voltage HBT
Source: Ref. 15.
0
90
0
90
0
90
Duplexer
LNA
PA
ADC
ADC
DAC
DAC
B
a
s
e

B
a
n
d

P
r
o
c
e
s
s
i
n
g
PLL/
VCO
PLL/
VCO
VGA
VGA
VGA
VGA
Antenna
Down Mixer
Up Mixer
Figure 6. A general RF transceiver system.
506 BiCMOS DEVICES AND RF INTEGRATED CIRCUITS
Table 4). For wireless standards such as GSM, DECT, and
Bluetooth employing constant envelope modulation
schemes, nonlinear power amplier (classes C, D, E, and
F) can be used. For other wireless standards such as
WLAN and WCDMA employing variable envelope modu-
lation schemes, linear power amplier (class AB) is typi-
cally used. Linearity can be a tradeoff with power-added
efciency in a practical power amplier design, where
nonlinear ampliers often achieve higher efciency than
do linear ones [37].
In a practical RF power amplier design several chal-
lenges are presented [29]. The PA must satisfy all the g-
ures of merit discussed above under the constraint of cost
for specific applications; the current trend toward de-
creased power supply (from 5 to 3 V and even lower) has
also made it difcult to maintain the required output pow-
er and efciency due to impedance matching limitations;
power-added efciency should not degrade significantly
with a wide range of output power.
Key device gures of merit for RF power ampliers are
f
max
(bandwidth and high power gain), linearity, and
breakdown voltage. Currently available technologies for
RF commercial power amplier are GaAs, SiGe, and
CMOS. Because of the high linearity, efciency, output
power, and breakdown voltage, GaAs has been the domi-
nant technology for the RF power amplier since the early
1990s. However, GaAs ampliers have the disadvantages
of relatively high cost, low substrate thermal conductivity,
and low integration ability with CMOS technology. More
recent SiGe HBT technology with high f
max
, improved
ruggedness, high linearity, and relatively high breakdown
voltage has made SiGe a technology contender for RF
power amplier applications. As discussed before, SiGe
devices can be engineered to have even higher breakdown
voltage but with the compromise of lower f
max
. In portable
wireless applications, a progressively reduced power sup-
ply makes breakdown voltage less critical, where SiGe can
be a good choice [29]. In addition, SiGe BiCMOS provides
the ability to integrate the power amplier with some re-
lated logic power amplied functions (implemented by
CMOS). High-quality passive components offered by com-
mercial SiGe BiCMOS processes enable the manufacture
of fully integrated RF power ampliers, including on-chip
impedance matching and ltering.
The power amplier is probably one of the last building
blocks to be integrated into a single radio chip. By far, the
only reported real single radio chip including PA is Blue-
tooth [38], which is a relatively low-RF performance de-
manding wireless standard.
More recently a significant amount of research effort
has been taken to design RF power ampliers using
SiGe BiCMOS technology (see Table 8), which have
demonstrated competitive power efciency and linearity
compared with GaAs technology for various wireless
applications.
3.4.3. Integrated Voltage Controlled Oscillator (VCO).
The voltage-controlled oscillator (VCO) is a very impor-
tant building block in the RF system and is often inte-
grated into a phase-locked loop (PLL) or frequency
synthesizer to serve as a frequency reference in a trans-
ceiver system [47]. For example, it is used as a local oscil-
lator together with a mixer to downconvert the received
RF signal to an IF signal. Typical gures of merit for
VCO are frequency of oscillation, phase noise, frequency
tuning range, output power, gain variation, and thermal
stability. Phase noise is perhaps the most important
characteristic of a VCO. Ideally, a VCO output signal is
expected to have all signal power concentrated in one
discrete frequency; however, all practical VCOs suffer
from spectrum impurity. Phase noise is any noise that
changes the frequency or phase of the oscillator signal and
is usually specied in dBc/Hz, meaning noise power in a
1 Hz bandwidth with respect to carrier power.
A good equation for relative phase noise of a VCO can
give us some insight into its function [48]:
1
4Q
2
Do
o
0

2
P
noise
P
carrier
2
Here, Q is the open loop quality factor, Do is the frequency
offset, o
0
is the center frequency, P
noise
is the spectral
density of each noise source, and P
carrier
is the spectral
density of the carrier. From this equation, there are
several rules for decreasing the phase noise: (1) use
high-Q passive components, (2) maximize the oscillation
Table 7. Some Recently Reported SiGe BiCMOS LNAs Compared with GaAs LNAs
Technology Frequency (GHz) Noise Figure (dB) Gain (dB) IIP3 (dBm) Power Dissipation (mW) Ref.
0.5 mmGaAs PHEMT 2.14 1.0 15.0 7.3 25.5 30
0.5 mmSiGe BiCMOS 2.5 1.6 12.0 8 20.625 31
0.5 mmSiGe BiCMOS 2.0 1.15 18.0 15.8 N/A 32
0.5 mm SiGeBiCMOS 1.96 1.9 15.3 7.6 13.5 33
0.2 mmSiGe BiCMOS 23 4.1 21.0 N/A 50 34
0.35mmSiGe BiCMOS 5.26.6 2.3 17.0 7 13 35
0.5 mmSiGe BiCMOS 0.88 1.4 15.7 10.3 11.7 36
Noise Power
Linearity Frequency
Supply
Voltage
Gain
Performance
&
Cost
Figure 7. RFIC design hexagon [26].
BiCMOS DEVICES AND RF INTEGRATED CIRCUITS 507
swing, and (3) minimize the number of active and lossy
passive devices in the oscillation path. Consequently,
high-Q passive components such as inductors and varac-
tors are critical limiting factors for the performance of
VCO. Fully integrated VCO suffers from low-quality
on-chip inductors and varactors. In addition to phase
noise, frequency tuning range and VCO gain variation
parameters also require high-quality passive components.
Because of the lower 1/f noise of SiGe HBT compared
with that of GaAs or Si CMOS, SiGe HBT is preferred
for 1/f noise-sensitive circuits such as VCO. As discussed
previously, more recent advances of SiGe BiCMOS
passive device technology have greatly increased VCO
performance. Table 9 summarizes the most recently
reported VCO design using SiGe BiCMOS and GaAs
technology.
4. FUTURE TRENDS
According to ITRS (International Technology Roadmap for
Semiconductors) 2003 [16], the challenges for future RF
transceivers using BiCMOS technology include the follow-
ing: aggressive scaling of passive elements (capacitors and
inductors); reducing the cost of BiCMOS technology while
improving power and performance tradeoffs; improving
signal isolation as more functions are integrated on a sin-
gle chip; increasing difculty with integrating bipolar de-
vices in aggressively scaled CMOS nodes due to conicting
thermal budget requirements; improving the performance
of active devices to levels that enable reconguration of
the RF transceiver in a software-dened radio, a funda-
mental architectural simplication of the RF transceiver
such as the direct synthesis and/or direct analog-to-digital
conversion of the RF signal; enabling the scaling of bipolar
power supply to reduce power consumption versus CMOS
technology; and integrating new materials in a cost-effec-
tive manner to realize the high-Q inductors and high-den-
sity metalinsulatormetal (MIM) capacitors.
Future challenges can be summarized as follows: SiGe
HBT device performance improvement, RFIC design
technique development, and on-chip passive component
performance and isolation enhancement.
4.1. SiGe HBT Device Performance Improvement
Carbon doping of SiGe HBTs (SiGe:C) is rapidly becoming
the preferred approach for commercial SiGe technologies
as a means to effectively suppress boron outdiffusion dur-
ing fabrication [6], consequently facilitating management
of the overall thermal budget and prole control. Contin-
uous lateral and vertical scaling, novel device structure
design, and improved fabrication techniques will further
enhance SiGe HBT peak f
T
and f
max
. Improved device per-
formance has led to optical and communication applica-
tions at frequencies greater than 20 GHz. The further
increased performance of SiGe HBT also makes it possi-
ble to design a reconguring RF transceiver in a software-
dened radio, which implements most of RF functions in
the digital domain.
4.2. RFIC Design Technology Development
RF system-on-chip (RF SoC) usually means that RF ana-
log/digital circuits are all integrated with memory blocks
Table 9. Some Recently Reported SiGe BiCMOS VCOs Compared with GaAs VCOs
Technology
Frequency
(GHz)
Tuning
Range (GHz) Phase Noise
Power
Supply (V)
Power Dissi
pation (mW) Ref.
GaAs HBT 5 0.290 118dBc/ Hz at 1 MHz 3.5 13.2 49
0.35mmSiGe 4.7 0.509 107dBc/Hz at 1 MHz 3 54 50
0.35mmSiGe 11 0.5 85dBc/Hz at 100kHz 3.3 N/A 51
0.35mmSiGe 47 7.05 104dBc/Hz at 1 MHz 5.5 290 52
0.25mmSiGe 2.3 N/A 102dBc/Hz at 100kHz 3 15.6 53
0.35mmSiGe 20 0.600 85dBc/Hz at 1 MHz 3.6 43.2 54
0.4 mmSiGe 28 4.1 84.2dBc/Hz at 1MHz 5 129 55
0.18mmSiGe 6/9 0.3/0.45 106dBc/Hz at 1 MHz/ 104dBc/Hz at 1 MHz 1.8 19.44 56
Table 8. Some Recently Reported SiGe BiCMOS PAs compared with GaAs PAs
Technology
Frequency
(GHz)
Maximum output
power (dBm) Linearity (ACPR)
Maximum
PAE (%)
Supply
Voltage (V)
Linear
Gain (dB) Ref.
GaAs HBT 1.95 24 42dBc at 5 MHz 36 3.5 48 39
55dBc at 10MHz
SiGe 1.88 30 46dBc at 1.25MHz 41 3.4 23 40
0.35mmSiGe 5 31.5 N/A 24 N/A 20.6 41
SiGe 2.4 27.5 N/A 47 3.3 35 42
0.25mmSiGe 1.95 24 30.7dBc at 5 MHz 30 3.3 24 43
41.3dBc at 10MHz
0.18mmSiGe 2.2 10 N/A 25 1.3 13 44
SiGe 0.824B0.849 28.2 53.6dBc at 885kHz 45 2.5 28 45
60.6dBc at 1980kHz
0.25mmSiGe 5.3 25 N/A 24 2.4 26 46
508 BiCMOS DEVICES AND RF INTEGRATED CIRCUITS
and microprocessors as a single-chip radio. High integra-
tion ability with superior RF performance has made SiGe
BiCMOS a preferred choice for future RF SoC implemen-
tations [18]. IBM has demonstrated a highly integrated
SoC design using 0.18-mm SiGe BiCMOS technology [9]
(see Fig. 8). Multistandard and multiband wireless hand-
sets employing RF SoC technique will not be a dream in
the near future. Future developments of RFIC CAD tech-
niques may change the RFIC design methodology to dig-
ital cell librarylike designs (see Fig. 9) [57]. After picking
up existing RF/analog cell libraries or IPs (intellectual
properties) for different building blocks (e.g., LNA, PA,
VCO) according to various applications (e.g., GSM,
CDMA), future RF system designers can simply assemble
them together to functionalize the whole system. Conse-
quently, RFIC design will be an easy and quick instead of
tedious and complicated process. Available highly accu-
rate RF/analog models that support scaling will enable RF
SoC simulations, including all on-chip coupling parasitic
effects.
4.3. On-Chip Passive Component Performance and Isolation
Enhancement
The performance of active devices has been improved
significantly, as discussed earlier. In contrast, relatively
slow development of passive components will be the key
bottleneck in the future RFIC designs. This is mostly the
result of high RF loss of silicon substrate, which leads to
not only low-Q passive devices but also large coupling
through substrates. A lot of research has been done on
substrate technologies for RF applications. A good review
of modern substrate transfer techniques for RFIC, such
as silicon-on-insulator (SoI), silicon-on-glass (SoG), and
silicon-on-anything (SoA) is given in Ref. 58. Microma-
chining techniques have also been applied to RFIC to
produce high-performance passive devices [59].
5. CONCLUSION
In this article, we have reviewed the advances of SiGe
BiCMOS technologies. SiGe BiCMOS technology has ex-
hibited many advantages over other existing RF semicon-
ductor technologies, such as IIIV compounds and RF
CMOS, which will have a strong impact on future RFIC
design. RFIC designs using SiGe BiCMOS technology are
discussed in terms of passive components and some key
RF building blocks. Future trends of RF integrated system
adopting SiGe BiCMOS technology are also summarized.
Acknowledgements
The authors wish to thank George Remoundos from Agere
Systems (Orlando, FL) and Brittin Kane from Insyte Inc.
(St Petersburg, FL) for providing valuable materials on
this topic.
VCC VCC
VCC
BPF
Mixer
BPF
BPF
LNA
VCO/Prescaler
Power Amp.
(Circuit block of product)
Alumina Sub.
(Product using HIC)
(Cell Example:VCO)
VCO Mixer
LNA
(Product using MMIC)
Digital
Analog
R,L,C
LNA
Mixer
VCO
Cellular
PCS
WLL
ITS
(
C
e
l
l

l
i
b
r
a
r
y
)
VCO Mixer
LNA
Figure 9. Cell library for quick RFIC design [57].
Figure 8. Highly integrated SoC solution using SiGe BiCMOS
technology [9].
BiCMOS DEVICES AND RF INTEGRATED CIRCUITS 509
BIBLIOGRAPHY
1. SiGe Advances in the Marketplace, Pathnder Research Inc.,
San Jose, CA, Sept. 2002.
2. D. A. Rich, M. S. Carroll, M. R. Frei, T. G. Ivanov, M. Mas-
trapasqua, S. Moinian, A. S. Chen, C. A. King, E. Harris, J. De
Blauwe, Vuong Hong-Ha, V. Archer, and K. Ng, BiCMOS
technology for mixed-digital, analog, and RF applications,
IEEE Microwave Mag. 3(2):4455 (June 2002).
3. H. C. Lin, J. C. Ho, R. R. Iyer, and K. Kwong, Complementary
MOS-bipolar transistor structure, IEEE Trans. Electron. De-
vices ED-16(11):945951 (Nov. 1969).
4. T. H. Ning, Why BiCMOS and SOI BiCMOS, IBM J. Res.
Devel. 46(2/3):181186 (March/May 2002).
5. S. Subbanna, D. Ahigren, D. Hatame, and B. Meyerson, How
SiGe evolved into a manufacturable semiconductor produc-
tion process, Proc. IEEE Int. Solid-Sate Circuits Conf., 1999,
pp. 6667.
6. J. D. Cressler and G. Liu, silicon-Germanum Heterojunction
Bipolar Transistors, Artech House, Norwood, MA, 2003.
7. J. Sevenhans, The siliconradio decade, IEEE Trans. Micro-
wave Theory Tech. 50(1):235244 (Jan. 2002).
8. K. Washio, SiGe HBT and BiCMOS technologies for optical
transmission and wireless communication systems, IEEE
Trans. Electron. Devices 50(3):656668 (March 2003).
9. A. J. Joseph, J. Dunn, G. Freeman, D. L. Harame, D. Coolba-
ugh, R. Groves, K. J. Stein, R. Volant, S. Subbanna, V. S.
Marangos, S. S. Onge, E. Eshun, P. Cooper, J. B. Johnson,
Rieh Jae-Sung, B. Jagannathan, V. Ramachandran, D. Ah-
lgren, D. Wang, and X. Wang, Product applications and tech-
nology directions with SiGe BiCMOS, IEEE J. Solid-State
Circ. 38(9):14711478 (Sept. 2003).
10. J. S. Rieh, B. Jagannathan, H. Chen, K. T. Schonenberg,
D. Angell, A. Chinthakindi, J. Florkey, F. Golan, D. Green-
berg, S. J. Jeng, M. Khater, F. Pagette, C. Schnabel, P. Smith,
A. Stricker, K. Vaed, R. Volant, R. D. Ahlgren, G. Freeman,
K. Stein, and S. Subbanna, SiGe HBTs with cut-off frequency
of 350GHz, Proc. Int. Electronic Devices Mtg., Dec. 2002,
pp. 771774.
11. D. Greenberg, B. Jagannathan, S. Sweeney, G. Freeman, and
D. Ahlgren, Noise performance of a low base resistance
200GHz SiGe Technology, Proc. Int. Electronic Devices Mtg.,
Dec. 2002, pp. 787790.
12. E. O. Johnson, Physical limitation on frequency and power
parameters of transistors, IEEE Int. Conv. Record 27 (1965).
13. D. L. Harame, D. C. Ahlgren, D. D. Coolbaugh, J. S. Dunn, G.
G. Freeman, J. D. Gillis, R. A. Groves, G. N. Hendersen, R. A.
Johnson, A. J. Joseph, S. Subbanna, A. M. Victor, K. M. Wat-
son, C. S. Webster, and P. J. Zampardi, Current status and
future trends of SiGe BiCMOS technology, IEEE Trans. Elec-
tron. Devices 58(11):25752594 (Nov. 2001).
14. J. S. Dunn, D. C. Ahlgren, D. D. Coolbaugh, N. B. Feilchenfeld,
G. Freeman, D. R. Greenberg, R. A. Groves, F. J. Guarin, Y.
Hammad, A. J. Joseph, L. D. Lanzerotti, S. A. St.Onge, B. A.
Orner, J. S. Rieh, K. J. Stein, S. H. Voldman, P. C. Wang, M. J.
Zierak, S. Subbanna, D. L. Harame, D. A. Herman, Jr., and B.
S. Meyerson, Foundation of RF CMOS and SiGe BiCMOS
technologies, IBM J. Res. Devel. 47(2/3):101138 (March/May
2003).
15. D. L. Harame, A. Joseph, D. Coolbaugh, G. Freeman, K. New-
ton, S. M. Parker, R. Groves, M. Erturk, K. Stein, R. Volant,
C. Dickey, J. Dunn, S. Subbanna, H. Zamat, V. S. Marangos,
M. M. Doherty, O. Schreiber, T. Tanji, D. A. Herman, and
B. S. Meyerson, The emerging role of SiGe BiCMOS
technology in wired and wireless communications, Proc. 4th
Int. IEEE Inter. Caracas Conf. Devices, Circuit and System,
Aruba, April 2002, pp. D052-1D052-17.
16. ITRS 2003 Roadmap (2003) (Online), http://public.itrs.net
17. K. Chang, I. Bahl, and V. Nair, RF and Microwave Circuit
and Component Design for Wireless Systems, Wiley, New
York, 2002.
18. D. Y. C. Lie, Xiaojuen Yuan, L. E. Larson, Y. H. Wang, A. Se-
nior, and J. Mecke, RF-SoC: Low-power single-chip radio
design using Si/Ge BiCMOS technology, Proc. 3rd Int. Conf.
Microwave and Millimeter Wave Technology, Aug. 2002,
pp. 3037.
19. IBM 180nm technology brief (2003) (Online), http://
www.ibm.com.
20. T. Lee, The Design of CMOS Radio-Frequency Integrated Cir-
cuits, Cambridge Univ. Press, Cambridge, UK, 1998.
21. I. Bahl, Lumped Elements for RF and Microwave Circuits,
Artech House, Norwood, MA, 2003.
22. M. Park, S. Lee, C. S. Kim, H. K. Yu, and K. S. Nam, The
detailed analysis of high Q CMOS-compatible microwave spi-
ral inductors in silicon technology, IEEE Trans. Electron. De-
vices 45(9):19531959 (Sept. 1998).
23. A. Pruijmboom, D. Szmyd, R. Brock, R. Wall, N. Morris,
Keng Fong, and F. Jovenin, QUBiC3: A 0.5mm BiCMOS
production technology with f
T
30GHz, f
max
60GHz
and high-quality passive components for wireless tele-
communication applications, Proc. IEEE Bipolar/BiCMOS
Circuits and Technology Meeting, Sept. 1998, pp.
120123.
24. J. N. Burghartz, M. Soyuer, K. A. Jenkins, M. Kies, M. Dolan,
K. J. Stein, J. Malinowski, and D. L. Harame, Integrated RF
components in a SiGe bipolar technology, IEEE J. Solid-State
Circ. 32(9):14401445 (Sept. 1997).
25. P. Andreani and S. Mattisson, Brief papers on the use of MOS
varactors in RFVCOs, IEEE J. Solid-State Circ. 35(6):905
910 (2000).
26. B. Razavi, RF Microelectronics, Prentice-Hall, Upper Saddle
River, NJ, 1998.
27. X. Yang, T. X. Wu, and J. McMacken, Design of LNA at
2.4 GHz using 0.25mm CMOS technology, Microwave Opt.
Technol. Lett. 36(4):270275 (2003).
28. J. Kucera, Low noise amplier design, in J. M. Golio,
and A. M. Shamsuddin, eds., Commercial Wireless Circuits
and Components Handbook, CRC Press, Boca Raton, FL,
2003.
29. L. E. Larson, Integrated circuit technology options for RFICs-
present status and future directions, IEEE J. Solid-State
Circ. 33(3):387399 (March 1998).
30. S. Kumar, M. Vice, H. Morkner, and W. Lam, Enhancement
mode GaAs PHEMT LNA with linearity control (IP3) and
phased matched mitigated bypass switch and differential ac-
tive mixer, IEEE Microwave Symp. Digest, June 2003, Vol. 3,
pp. 15771580.
31. D. Wang, K. Krishnamurthi, S. Gibson, and J. Brunt, A
2.5 GHz low noise high linearity LNA/mixer ICin SiGe BiC-
MOS technology, Proc. IEEE Radio Frequency Integrated Cir-
cuit Symp., May 2001, pp. 249252.
32. Q. Liang, G. Niu, J. D. Cressler, S. Taylor, and D. L Harame,
Geometry and bias current optimization for SiGe HBT cas-
code low-noise ampliers, Proc. IEEE Radio Frequency Inte-
grated Circuit Symp., June 2002, pp. 407410.
33. P. Shah, A 2 GHz low-distortion low-noise two-stage LNA
employing low impedance bias terminations and optimum
510 BiCMOS DEVICES AND RF INTEGRATED CIRCUITS
inter-stage match for linearity, Proc. 26th European Solid-
State Circuits Conf., Sept. 2000.
34. G. Schuppener, T. Harada, and Y. Li, A 23-GHz low-noise
amplier in SiGe heterojunction bipolar technology, Proc.
IEEE Radio Frequency Integrated Circuit Symp., May 2001,
pp. 177180.
35. J. Sadowy, I. Telliez, J. Graffeuil, E. Tournier, L. Escotte, and
R. Plana, Low noise, high linearity, wide bandwidth amplier
using a 0.35mm SiGe BiCMOS for WLAN applications, Proc.
IEEE Radio Frequency Integrated Circuit Symp., June 2002,
pp. 217220.
36. V. Aparin, E. Zeisel, and P. Gazzerro, Highly linear SiGe
BiCMOS LNA and mixer for cellular CDMA/AMPS applica-
tions, Proc. IEEE Radio Frequency Integrated Circuit Symp.,
June 2002, pp. 129132.
37. A. Shirvani and B. A. Wooley, Design and Control of RF Power
Ampliers, Kluwer Academic Publishers, Boston, 2003.
38. F. O. Eynde, J.-J. Schmit, V. Charlier, R. Alexandre, C. Stur-
man, K. Cofn, B. Mollekens, J. Craninckx, S. Terrijn, A.
Monterastelli, S. Beerens, P. Goetschalckx, M. Ingels, D. Joos,
S. Guncer, and A. Pontioglu, A fully-integrated single-chip
SoCfor Bluetooth, Proc. IEEE Int. Solid-State Circuits Conf.,
Feb. 2001, pp. 196197.
39. V. T. S. Vintola, M. J. Matilainen, S. J. K. Kalajo, and E. A.
Jarvinen, Variable-gain power amplier for mobile WCDMA
applications, IEEE Trans. Microwave Theory Tech.
49(12):24642471 (Dec. 2001).
40. X. Zhang, C. Saycocie, S. Munro, and G. Henderson, A SiGe
HBT power amplier with 40% PAE for PCS CDMA applica-
tions, IEEE Microwave Symp. Digest, June 2002, Vol. 2,
pp. 857860.
41. N. Tanzi, J. Dykstra, and K. Hutchinson, A 1-watt doubly
balanced 5GHz ip-chip SiGe power amplier, Proc. IEEE
Radio Frequency Integrated Circuit Symp., June 2003,
pp. 141144.
42. A. Raghavan, D. Heo, M. Maeng, A. Sutono, K. Lim, and J.
Laskar, A 2.4GHz high efciency SiGe HBT power amplier
with high-Q LTCC harmonic suppression lter, IEEE Micro-
wave Symp. Digest, June 2002, Vol. 2, pp. 10191022.
43. I. Rippke, J. Duster, and K. Kornegay, A fully integrated, sin-
gle-chip handset power amplier in SiGe BiCMOS for W-
CDMA applications, Proc. IEEE Radio Frequency Integrated
Circuit Symp., June 2003, pp. 667670.
44. J. A. OSullivan, C. Delabie, K. G. McCarthy, A. Murphy, and
P. J. Murphy, A fully integrated high efciency SiGe HBT
class F power amplier at 2.2GHz, High Frequency Postgrad-
uate Student Colloquium, Sept. 2003, pp. 4851.
45. N. Srirattana, M. S. Qureshi, A. Aude, V. Krishnamurthy, D.
Heo, P. E. Allen, and J. Laskar, SiGe HBT power amplier for
IS-95 CDMA using a novel process, voltage, and temperature
insensitive biasing scheme, Proc. Int. Symp. Circuits and Sys-
tems, May 2003, Vol. 1, pp. I-437I-440.
46. W. Bakalski, W. Simburger, R. Thuringer, A. Vasylyev,
and A. L. Scholtz, A fully integrated 5.3GHz, 2.4 V,
0.3 W SiGe-bipolar power amplier with 50spl Omega
output, Conf. European Solid-State Circuits, Sept. 2003,
pp. 561564.
47. X. Yang, T. X. Wu, and J. McMacken, A 5 GHz fast-switching
CMOS frequency synthesizer, Proc. 2002 IEEE Radio Fre-
quency Integrated Circuit Symp., June 2002, pp. 479482.
48. B. Razavi, Challenges in the design of frequency synthesizers
for wireless communications, Proc. IEEE Custom Integrated
Circuits Conf., May 1997, pp. 395402.
49. Y.Eo, K. Kim, B. Oh, Low noise 5 GHz differential VCO using
InGaP/GaAs HBT technology, IEEE Microwave Wireless Co-
mpon. Lett. 13:259261 (July 2003).
50. Y. G. Kim, H. S. Kim, J. H. Bac, J. H. Oh, and C. W. Kim, Fully
integrated differential VCO with buffer amplier using 0.35
spl mu m SiGe BiCMOS for C-band wireless RF transceiver,
Proc. Radio and Wireless Conf., Aug. 2003, pp. 297300.
51. W. Wong, G. Cibiel, J.-G. Tartarin, E. Tournier, R. Plana, and
O. Llopis, X band BiCMOS SiGe 0.35-mm voltage controlled
oscillator in parallel and reection topology and external
phase noise improvement solution, IEEE Radio Frequency
Integrated Circuit Symp., June 2003, pp. 281284.
52. H. Li, H.-M. Rein, R. Kreienkamp, and W. Klein, 47 GHz VCO
with low phase noise fabricated in a SiGe bipolar production
technology, IEEE Microwave Wireless Compon. Lett.
12(3):7981 (March 2002).
53. D. Wang and X. Wang, The performance comparison of
CMOSvs bipolar VCO in SiGe BiCMOS technology, IEEE Mi-
crowave Symp. Digest, June 2003, Vol. 1, pp. A101A104.
54. K. Ettinger, A. Stelzer, C. G. Diskus, W. Thomann, J. Fenk,
and R. Weigel, Single-chip 20-GHz VCO and frequency divid-
er in SiGe technology, IEEE Microwave Symp. Digest, June
2002, Vol. 2, pp. 835838.
55. S. Hackl, J. Bock, G. Ritzberger, M. Wurzer, and A. L. Scholtz,
A 28-GHz monolithic integrated quadrature oscillator in SiGe
bipolar technology, IEEE J. Solid-State Circ. 38(1):135137
(Jan. 2003).
56. H. Shin, Z. Xu, and M.F. Chang, A 1.8-V 6 9-GHz recongur-
able dual-band quadrature LC VCO in SiGe BiCMOS
technology, IEEE J. Solid-State Circ. 38(6):10281032
(June 2003).
57. Flash communication market overview (online), http://
www.ash-comm.com.
58. R. Dekker, P. G. M Baltus, and H. G. R Maas, Substrate
transfer for RF technologies, IEEE Trans. Electron. Devices,
50(3):747757 (March 2003).
59. J. Y.-C. Chang, A. A. Abidi, and M. Gaitan, Large suspended
inductors on silicon and their use in a 2mm CMOS RF am-
plier, IEEE Electron. Device Lett. 14:246248 (May 1993).
BIOLOGICAL EFFECTS OF RADIOFREQUENCY
ENERGY AS RELATED TO HEALTH AND SAFETY
KENNETH R. FOSTER
University of Pennsylvania
Philadelphia, Pennsylvania
1. BIOLOGICAL EFFECTS OF RADIOFREQUENCY FIELDS
AN ONGOING CONTROVERSY
The biological effects of radiofrequency (RF) and micro-
wave energy is a large and heterogeneous subject that
spills over into biophysics, medicine, and engineering
and more recently, epidemiology, risk assessment, law,
and public policy. The primary scientific literature on bi-
ological effects of RF elds contains several thousand pa-
pers, including many written in the 1960s or before.
This article examines biological effects of RF elds
(3 kHz300GHz) with emphasis on potential hazards,
BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY 511
specifically, effects that a person would perceive as ad-
verse to health and safety. Its approach is tutorial rather
than comprehensive, and its perspective follows that of
major international exposure limits [established by IEEE
and ICNIRP (International Committee on Non-Ionizing
Radiation Protection)]. Its main emphasis will be on un-
equivocal effects that are potentially hazardous, which are
principally thermal in nature. A concluding section will
review major areas of controversy. Issues related to elec-
tromagnetic compatibility, for example, the interaction of
RF elds with medical devices, are excluded from discus-
sion. Additional sources are included in an annotated bib-
liography (see Further Reading section following the
numbered references list).
2. EXPOSURE AND DOSE
In toxicology, exposure is the concentration or a potentially
toxic agent in the external environment of a subject, and
dose is the concentration of the substance in target tis-
sues. For RF elds, the corresponding quantities are eld
levels external to the body, and those induced within body
tissues by the exposure. The bioeffects literature refers to
assessment of both quantities as dosimetry. Exposure is
very different from risk, which is the probability of an ad-
verse effect in an exposed individual and the severity of
the effect.
The coupling between RF elds and the human body
depends on the characteristics of the eld (including fre-
quency, polarization, and eld strength) as well as the
electrical characteristics of the biological target. Because
of the wide range in these characteristics over the fre-
quency range considered here, these coupling properties
are quite diverse.
2.1. Bulk Electrical Properties of Tissue
The bulk electrical properties (dielectric properties) of
tissue determine the propagation characteristics of
RF elds inside tissue; for a review, see Ref. 1. The
bulk properties consist of the permittivity e and con-
ductivity s. It is convenient to dene the complex
conductivity s*:
s

s joee
0
1
where o is the frequency [in radians per second (rad/s)]
and e
0
is the permittivity of space (a constant). The real
and imaginary parts of s* are shown in Fig. 1 for several
representative soft tissues [2].
The dielectric properties of tissues exhibit dispersions
(variations with frequency) due to the charging of cell
membranes, dipolar losses in tissue water, and other ef-
fects. Charging of cellular membranes is chiefly responsi-
ble for the dispersion observed in tissue at megahertz
(MHz) frequencies and below, while dipolar relaxation of
tissue water is the dominant source of dispersion above
B1 GHz. Above B100 MHz, tissue water content is the
chief factor determining the dielectric properties of soft
tissues.
2.2. Quantities Used to Characterize Exposure
Depending on the exposure scenario, different quantities
are useful in characterizing exposure and dose. These in-
clude
1. For assessing external elds:
a. External eld strength E
o
, measured in volts/
meter (electric eld) or amperes per meter (A/m)
(magnetic eld) (used principally at frequencies
below 100 MHz
1
or in the near-eld region of
antennas).
b. Incident power density S, in W/m
2
(used principally
at frequencies above 0.5 GHz or in the far elds of
radiators).
Frequency, Hz
10
2
10
3
10
4
10
3
10
4
10
5
10
6
10
7
10
8
10
9
10
10
10
11
Frequency, Hz
10
3
10
4
10
5
10
6
10
7
10
8
10
9
10
10
10
11
10
1
10
0
10
1
10
2
10
2
10
3
10
4
10
5
10
6
10
1
10
0
10
1
10
2
10
3
Wet skin
Skeletal muscle
Cornea
Brain (White matter)
Wet skin
Skeletal muscle
Cornea
Brain (White matter)
o
c
'
c
o
,

S
/
m
C
o
n
d
u
c
t
i
v
i
t
y

o
,

S
/
m
Figure 1. Dielectric properties of selected tissues: conductivity F and oe
0
e
0
(both in S/m) of
selected tissues at 371C. (Data from Ref. 2.)
1
All frequency ranges in this discussion are approximate.
512 BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY
2. For assessing internal elds:
a. Current density within the body J
i
(in A/m
2
), used
principally in the frequency range up to 10 MHz; a
related measure is internal eld strength E
i
(V/m),
which is related to the current density by Ohms
law
J
i
sE
i
2
b. Total current passed into tissue J., used chiefly to
characterize contact currents, principally in the fre-
quency range up to 100MHz.
c. Internal electrical eld strength E
i
, specifically, the
electric eld induced within body tissues, the in-
duced current density J
i
in A/m
2
.
d. Specific absorption rate (SAR), dened as the power
dissipation in watts per kilogram of tissue. In terms
of the electric eld E
i
in the tissue
SAR
sjE
2
i
j
r
3
where r is the density (kg/m
3
) of the tissue. The SAR can
be expressed as a point value at a given instant in time, a
value averaged over a specied time and distance, or a
value averaged over the whole body of an animal (the
whole-body SAR).
Under most circumstances, the SAR is the preferred
dosimetric quantity for RF elds, particularly over the
frequency range 100kHz10 GHz, and journal editors and
referees typically insist that this information be included
in papers submitted for publication. However, the SAR
becomes less useful at frequencies above 10 GHz, where
the energy is absorbed close to the tissue surface and the
incident power density can be a more appropriate measure
of exposure. The SAR is also less useful below 100 kHz,
where little energy is absorbed by the tissue from the ex-
ternal eld and contact currents become the dominant
source of hazard.
Another parameter is specific-energy absorption (S.A.),
the total absorbed energy in a pulse per kilogram of tissue,
used principally for pulsed elds in the frequency range
300MHz10 GHz.
The SAR can be determined in a number of ways:
*
By measuring the transient rate of temperature in-
crease in the exposed object. A SAR of 1 W/kg pro-
duces a rate of temperature increase of 0.0151C/min
in high-water-content tissues. The rate of tempera-
ture increase can be measured using small tempera-
ture sensors, notably nonperturbing beroptic
probes that are designed specifically for measure-
ments in RF elds. Another approach is thermal
imaging of the surface of the exposed object using
an infrared camera, typically using phantom models
to mimic animals. In either case, it is important to
correct for effects of heat conduction in measure-
ments of the rate of temperature increase.
*
By modeling the exposure situation by computer, to
determine the rate of absorption of the incident en-
ergy in the exposed object.
*
By measuring the internal electric eld in the ex-
posed object with a small probe (taking care to avoid
electrical artifacts).
Additional information may be necessary to characterize
exposure, including the peak eld strength or peak inci-
dent power density (for pulsed elds) or other description
of the modulation characteristics of the eld.
Inadequate dosimetry has been a longstanding
problem in bioeffects studies. Many earlier studies report-
ed only the incident eld intensity, but did not determine
the absorbed power in the exposed animals, and conse-
quently the studies are difcult or impossible to interpret.
Inadequate dosimetry is a particularly noticeable
problem in numerous bioeffects studies in the Russian
scientific literature, which frequently lacks even rudimen-
tary information about the frequency and intensity of
exposure.
A related problem has been the difculty in distin-
guishing effects due to heating of the sample from those
due to the elds themselves, that is, distinguishing ther-
mal from nonthermal effects, due to the difculty in ade-
quately controlling temperature during an experiment.
With careful experimental design, these problems are now
solvable. However, adequate dosimetry and temperature
control remain difcult and expensive to achieve. Typical-
ly, bioeffects studies, particularly those involving exposure
of large animals and humans, require specially designed
facilities and extensive engineering support including in
some cases elaborate computational studies to determine
exposure (Fig. 2).
3. COUPLING CONSIDERATIONS
The coupling between external elds and body tissues
depends on the characteristics of exposure, which vary
greatly in different exposure scenarios, as well on the
geometry and electrical characteristics of the exposed
subject.
There are numerous ways in which human subjects can
be exposed to RF energy. The exposure scenarios can often
be classied in one of three ways, which entail significant
technical differences in determining exposure, described
in the following text.
3.1. Contact Currents
In this exposure scenario, RF currents are passed into the
body when a subject touches a conductive object in the
presence of a eld. Contact currents are a major potential
occupational hazard in some occupational settings. For
example, contact currents induced by touching broadcast
transmission towers or secondary metallic structures
(buildings, guywires, fencing, etc.) near the towers can in-
duce hazardous contact currents.
The magnitude of the contact current is a function of
source impedance and the impedance of the subject to
ground, and is not simply related to the eld strength in
the vicinity of the subject. Contact currents can be mea-
sured by placing an ammeter between the source and the
subject.
BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY 513
3.2. Near-Field or Partial-Body Exposure
In localized regions near RF sources, very high RF elds
can exist, possibly at acutely hazardous levels. Sources
include medical equipment [such as magnetic resonance
imaging (MRI) scanners], industrial equipment (e.g., in-
duction heaters or microwave drying equipment), and ra-
dar and broadcast transmitters. Exposures to RF elds
from such equipment typically occur in the near elds of
RF sources, and to parts of the body. Partial-body exposure
(albeit at thermally innocuous levels) is an everyday oc-
currence to a user of a mobile telephone or other portable
transmitter.
Because of the complex electrical interaction between
the transmitting antenna and the subjects body, exposure
assessment in near-eld exposures is a difcult matter.
Exposure is typically determined using computer model-
ing or physical measurements on phantom models
whose electrical properties approximate those of tissue
at the pertinent frequency [3]. The measurements can in-
clude thermal measurement of SAR or direct measure-
ment of the local eld using small electrical probes
inserted into the model. Figure 3 shows a commercial
apparatus that employs a robot to move a small electric
eld probe about a phantom model of the head, which
has become the most widely used method to measure
SARs produced in the human head by mobile telephone
handsets.
3.3. Far-Field Exposure
In a typical far-eld exposure scenario, plane-wave energy
is incident on large areas of the subjects body. This expo-
sure scenario is important because of public concern about
possible health hazards from broadcast transmitters, as
well as in some occupational exposure settings, for exam-
ple, for workers near radar transmitters.
Far-eld exposure can be modeled by computer or mea-
sured experimentally, and, in comparison to with near-
eld exposures, is comparatively easy to accomplish. Con-
sequently, a large literature exists on the subject (e.g., see
Ref. 4). Initial studies, beginning in the 1960s, modeled
the human body and experimental animals using homo-
geneous spheroids exposed to plane-wave irradiation.
More recent studies have employed detailed anatomical
models of the body with millimeter spatial resolution. A
variety of thermal measurements have also been reported,
most commonly for measurement of whole-body SAR [4].
As with other exposure scenarios, the coupling between
an incident eld and the body depends strongly on the
frequency and other characteristics of the eld. Figure 4
shows the whole-body average SAR in an isolated ellipsoid
modeling an adult human, exposed to plane-wave irradi-
ation of varying polarization. Below B70 MHz, the SAR is
proportional to the square of the frequency, and most of
the energy is deposited near the body surface. This model
exhibits a maximum near 70 MHz due to an electrical res-
onance, reecting the antenna properties of the model.
At higher frequencies, the whole-body SAR decreases
(for a given incident power density) because the energy
is deposited close to the surface. In addition, various parts
of the body exhibit their own resonances; for example, the
human head exhibits an antenna resonance at approxi-
mately 1 GHz. If the model is placed in contact with
Figure 3. Commercial instrument used to measure SAR distri-
bution produced by a mobile telephone in model of a human head.
The robot arm moves an electric eld probe inside a phantom
model lled with a liquid whose electrical properties resemble
those of a tissue. (From Ref. 3 with permission; available at
http://www.dasv4.com/welcome.html.)
Figure 2. Eleanor R. Adair, noted bioeffects investigator [and
immediate past chair of the ICES (formerly known as IEEE
C95.1) standards committee] in microwave anechoic chamber
used for RF bioeffects studies at Brooks Air Force Base, Texas.
514 BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY
ground, the induced currents (particularly those near the
point of contact with ground) are increased as a result of
image effects. Similar resonances occur in animals, but at
higher frequencies reecting the smaller body size, and
with different SAR levels for a given incident exposure
level (Fig. 5).
At sufciently high frequencies, the wavelength of the
incident energy is much smaller than the dimensions of
the model and the exposure approaches the quasioptical
regime. In this case, the distribution of energy within the
body is similar to that produced by a plane wave incident
on an absorbing surface. For a plane wave incident on a
plane surface of tissue, the SAR as a function of depth x
into the tissue is given by
SAR
I
0
T
rL
e
x=L
4
where T is the fraction of energy transmitted in the tissue
(the remaining part of the incident energy being reected)
and L is the energy penetration depth. In terms of the
relative permittivity and conductivity of the tissue (e
0
, s),
these quantities are
T1

p
1

p
1

2
L
c
2oIm

e
0
p

5
where e

e
0
js=2pf e
0
e

e
0
(js/2pfe
0
) and c is the ve-
locity of light. Energy penetration depths of microwaves
into soft tissue such as muscle range from about 1 cm at
1 GHz to a few tenths of a millimeter at 100GHz (Fig. 6).
4. INTERACTION MECHANISMS
The energy of photons of radiofrequency and microwave
elds is several orders of magnitude (at least) below that
needed to disrupt the weakest chemical bonds and create
free radicals. Consequently, RF elds are nonionizing, in
contrast to ionizing radiation such as X rays or ultraviolet
light, which disrupt chemical bonds and create highly re-
active free radicals.
Nevertheless, RF elds can interact with biological sys-
tems, most obviously by exerting forces on charges in tis-
sue [5]. These forces act in the presence of random thermal
agitation, which will overwhelm the effects of the elds if
the forces are too small. The absorbed power in tissue will
also generate heat and increase the local temperature, an
obvious thermal mechanism for biological effects.
4.1. Thermal Mechanisms
Thermal mechanisms arise from the deposition of power
in the biological system (as opposed to nonthermal effects,
which result from direct interaction with the electric or
magnetic elds).
4.1.1. Direct Effects of Temperature Increase. Biochemi-
cal reactions typically vary in rate by about 3%/1C. How-
ever, some biological structures (e.g., thermal receptors in
the human body or infrared receptors in some snake spe-
cies) respond to very small changes in temperature, as lit-
tle as a few hundredths of a degree or less. Thus, even
modest temperature increases, particularly if they persist
A
v
e
r
a
g
e

S
A
R

(
W
/
k
g
)
Man (70 MHz)
Monkey (300 MHz)
Mouse (2450 MHz)
Frequency (MHz)
10
1
10
2
10
3
10
4
10
0
10
1
10
2
10
3
10
1
Figure 5. Whole-body absorption rates for spheroidal models of a
human, a squirrel monkey, and a mouse, exposed to plane-wave
irradiation at 10W/m
2
in the E polarization. (From Ref. 4.)
Frequency (MHz)
10
1
10
2
10
3
10
4
10
5
A
v
e
r
a
g
e

S
A
R

(
W
/
k
g
)
10
1
10
2
10
3
10
4
10
0
H
E
K
Figure 4. Whole-body average SARin a prolate spheroidal model
of man in a plane wave, as a function of frequency of the incident
wave. The semimajor axis is 0.875m, semiminor axis 0.138m,
volume 0.07m
3
. The SAR is given in W/kg, normalized to an in-
cident power density of 10 W/m
2
(from Ref. 4). Three curves are
shown, for E (electric eld vertical), H (magnetic eld vertical),
and K (wave propagating from the top) polarizations.
BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY 515
over appreciable times, can elicit biological effects. Several
classes of thermal effects can be identied.
4.1.2. Effects of Thermal Load on the Body. An addition-
al heat load on the body due to exposure to RF energy can
also elicit effects due to thermoregulatory responses. The
natural rate of power generation in the human body (the
basal metabolic rate) is about 1 watt per kilogram of body
mass. Addition of heat by a RF eld at levels comparable
to or above this rate can produce significant physiological
effects even in the absence of thermally damaging tem-
perature increases anywhere in the body. These changes
can be interpreted as normal thermoregulatory responses,
and may explain, in part, various subtle physiological ef-
fects that have been reported in animals subjected to mod-
erate levels of RF energy exposure. With typical values of
the conductivity of tissue, eld levels inside the body rang-
ing from 100V/m (3kHz) to 30 V/m (1GHz) will produce a
SAR of 1 W/kg.
4.1.3. Effects of Local Temperature Increases. At SAR
levels somewhat above 1 W/kg, depending on the duration
of exposure, rate of cooling of the tissue by bloodow, and
other factors, local heating effects in tissue can become
significant or even damaging. These include a variety of
biological changes and, at sufcient heating levels, frank
tissue damage.
4.1.4. Effects Due to Rate of Temperature Increase. Some
thermal mechanisms depend on the rate of temperature
increase (as opposed to the magnitude of increase). Pulsed
microwave energy will generate acoustic transients in tis-
sue by transient heating of tissue water with correspond-
ing thermal expansion. These transients, under
appropriate exposure conditions, can be perceived by hu-
man subjects as clicks, the so-called microwave auditory
effect. This typically requires that the head of the subject
be exposed to microsecond pulses of RF energy with a car-
rier frequency near 1 GHz, and peak eld intensities
above 10,000W/m
2
. The pulses produce transient increas-
es in tissue temperature of a few microdegrees, at a rate of
about 11C/s. This is sufcient to produce acoustic tran-
sients in the head in excess of 100dB peak sound pressure
[6]. The subject perceives these acoustic transients
through normal hearing mechanisms.
Under more extreme exposure conditions, changes in
membrane potentials and stimulation of cell membranes
can occur, also related to the rates of temperature increase
[7]. For example, mice exposed in the head to intense mi-
crowave pulses, leading to brain temperature increases of
a few tenths of a degree within one second, exhibit a va-
riety of involuntary body movements and other stun
phenomena [8]. The extreme exposure levels needed to
produce such effects are far above those encountered in
realistic environments, and to achieve them generally re-
quires that the animals be placed inside waveguides that
are connected to high-powered military transmitters.
4.2. Nonthermal Mechanisms
Electric elds can exert forces on charges, and magnetic
elds exert torques on magnetic dipoles in biological sys-
tems. This gives rise to several classes of nonthermal
interaction mechanisms. (Whether the mechanisms result
in observable biological effects under realistic exposure
conditions is a different question, however.) Examples are
described below.
4.2.1. Membrane Excitation. The dominant hazard
mechanism for low-frequency currents is shock and other
membrane excitation phenomena. Above a few kilohertz,
the threshold current needed to excite membranes in-
creases rapidly with frequency. With a few exceptions,
over the entire frequency range considered in this article,
the thresholds for producing biologically significant
heating are lower than those for directly exciting cell
Frequency, Hz
10
9
10
10
10
11
10
12
Frequency, Hz
10
9
10
10
10
11
10
12
E
n
e
r
g
y

t
r
a
n
s
m
i
s
s
i
o
n

c
o
e
f
f
i
c
i
e
n
t
0.2
0.4
0.6
0.8
1.0
Skeletal muscle
Wet skin
Cornea
Brain (White matter)
E
n
e
r
g
y

p
e
n
e
t
r
a
t
i
o
n

d
e
p
t
h
,

c
m
10
1
10
2
10
3
10
0
10
1
Skeletal muscle
Wet skin
Cornea
Brain (White matter)
Figure 6. Energy penetration depths and transmission coefcients for several tissues.
516 BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY
membranes (Fig. 7). Consequently, the dominant hazard
mechanisms at RFs are thermal in nature.
4.2.2. Electroporation. Electric elds can create pores
in cell membranes (electroporation) by inducing electrical
breakdown. This requires membrane potentials of B1V,
which in turn requires tissue eld strengths that can ex-
ceed 10
4
10
5
V/m, which would be thermally damaging if
sustained for more than brief periods. However, electrop-
oration can be produced in the absence of thermal damage
using brief pulses of RF elds, typically with carrier fre-
quencies below 1 MHz. Electroporation is considered to be
a factor in some kinds of electrical injury [9].
4.2.3. Field-Induced Forces. Electric elds exert forces
on ions and torques on permanent dipoles, as well as
forces and torques on induced dipoles (dielectrophoretic
forces). Magnetic elds exert torques on magnetic dipoles.
These forces are opposed by viscous drag from the
surrounding medium, and occur in the face of random
thermal agitation.
Dielectrophoretic forces on cells are well documented,
and have important applications in biotechnology. How-
ever, biophysical considerations show that the forces to be
expected on cells from RF elds at normal environmental
levels are far too weak to overcome naturally occurring
thermal agitation.
4.3. Other Proposed Mechanisms
Various scientists, including some prominent physicists,
have suggested mechanisms for biological effects of weak
RF elds. These include various quantum phenomena,
nonlinear effects such as creation of solitons, and kinetic
effects including disruption of bound water and excita-
tion of molecular resonances (to name only a few proposed
mechanisms). These theories remain speculative, and of-
ten are susceptible to criticism, for example, for failing to
consider the effects of thermal agitation or viscous drag or
other dissipative processes [10]. This remains a highly
uneven and, in many places, highly contentious literature.
5. POTENTIAL ADVERSE EFFECTS OF RADIOFREQUENCY
FIELDS IN HUMANS
This subject is wide and, particularly as it relates to pos-
sible chronic health effects in humans, very controversial.
This article rst considers the well-established hazards
of RF energy, which underlie major exposure standards
for electromagnetic elds. It then comments on possible
effects of chronic exposure to RF and microwave elds on
humans.
5.1. Thermal Death
Numerous studies have been conducted, beginning in the
early 1960s, to determine lethal levels of exposures to RF
and microwave energy for animals. The exposure level, in
terms of incident power density, that will kill an animal
varies considerably depending on species, frequency of the
eld, environmental conditions, pretreatment by drugs or
other physiological manipulation, and other factors. In
part this reects the different absorption characteristics of
different species for RF energy, and in part physiological
differences among species, such as different capabilities
for thermoregulation. Prior to death, animals typically
exhibit circulatory failure and other signs of extreme heat
stress, and the thresholds for lethality typically correlate
with an increase in body temperature of several degrees
above normal [11]. Clearly, similar effects would occur in a
human, in the unlikely circumstance that he or she were
exposed at comparable levels.
5.2. Perception and Pain
Thresholds for perception of contact currents have been
measured for humans at several frequencies (Table 1).
At frequencies below B10 kHz, subjects report mild tick-
ling or pricking sensations near the threshold for percep-
tion; at higher frequencies subjects report faint warmth
[12], evidently reecting two different mechanisms
for perception (direct stimulation of membranes at low
frequencies and cutaneous perception of warmth at higher
frequencies).
The thresholds for perception of microwave energy
have been measured by Blick et al. [13] over a wide range
of microwave frequencies. These experiments used brief
(10-s) exposures to microwave energy over an area of
0.024 m
2
on the backs of human volunteers (Table 2)
[13]. The estimated increase in skin temperature at the
threshold for perception is about 0.071C over this entire
frequency range [14]. The threshold for perception de-
creases with increasing frequency because of the shorter
penetration depth into tissue and corresponding increase
in SAR near the skin surface. Thresholds for pain under
similar exposure conditions are approximately 100 times
higher, and correspond to peak temperature increases of
Frequency, Hz
10
0
10
1
10
2
10
3
10
4
10
5
10
6
I
n

s
i
t
u

e
l
e
c
t
r
i
c

f
i
e
l
d

(
V
/
m
)

10
3
10
2
10
1
10
0
10
1
10
2
I
n

s
i
t
u

c
u
r
r
e
n
t

d
e
n
s
i
t
y

(
A
/
m
2
)
10
2
10
1
10
0
10
1
10
2
Cardiac fiber
Nerve fiber in brain
Phosphenes
SAR = 10 W/kg
Figure 7. Estimated thresholds for different stimulatory effects
(excitation of 10-mm nerve bers in the brain, excitation of cardiac
bers, production of phosphenes, together with current level for a
SAR of 1 W/kg.
BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY 517
several degrees Celsius at the surface of the skin [15].
These thresholds clearly reect thermal phenomena.
5.3. Burns
Burns and other thermal damage are clear potential
hazards of RF and microwave energy. Typically, human
tissues can tolerate temperatures up to B431C for pro-
longed periods without damage; progressively higher tem-
peratures will cause thermal damage in progressively
shorter times [16].
While RF burns are an obvious potential hazard, few
such injuries are reported. The exposures are very painful,
causing the victim to withdraw from exposure before ther-
mal injury occurs. Serious injuries generally require ex-
treme exposure levels in which damage occurs quickly, or
circumstances in which the victim is unable to voluntarily
leave the area of exposure. There is a scattering of reports
of burns produced by microwave ovens whose interlock
mechanisms (which turn off the microwave power when
the door is opened) have failed, and of burns to patients
from medical treatments involving RF and microwave en-
ergy. There have been at least two reported cases of in-
fants being placed deliberately in microwave ovens and
suffering severe burns [17].
5.4. Other Adverse Thermal Effects
Many thermal effects have been demonstrated in animals,
and analogous effects can be anticipated in humans as
well. Deleterious effects include the following:
1. Cataracts. These are an established hazard of
microwave energy, and have been reported occasionally
in humans exposed to high intensity microwaves. Catar-
acts can be produced in animals at high exposure levels,
generally 41000W/m
2
, where they are associated with
temperature increases of several degrees in the eye. Such
exposure levels are similar to those capable of producing
skin burns. Occasional claims of cataract from low-level
microwave exposure have not been substantiated by ani-
mal studies and remain controversial [18].
2. Adverse Reproductive Effects. Development abnor-
malities have been convincingly demonstrated in animals
exposed to microwave energy at near-lethal levels. These
effects are associated with significant increases in body
temperature, which is itself a well-established cause of
birth defects [19,20]. Heating the testes to microwave en-
ergy to induce temporary sterility has been explored in
China as a means of contraception. Some men have re-
ceived more than 100 treatments, each raising the surface
temperature of the scrotum to 40421C. No significant ab-
normalities were reported in a follow-up study in these
men after cessation of treatment [21].
3. Behavioral Disruption. This has been observed in
several species of animals exposed to microwaves at
whole-body exposures of 46 W/kg, at several frequencies
above 100 MHz [22]. In such studies, the animals are
trained to carry out a task (e.g., pressing a lever to obtain
food pellets) and exposed to microwave energy. At suf-
cient exposure levels, the animals stop performing the as-
signed task and begin a different behavior, typically one
associated with thermoregulation (in rats, spreading sali-
va on the tail). The threshold SAR for behavioral disrup-
tion is approximately 4 W/kg in several species, and does
not vary greatly with frequency (Table 3). The effect is re-
versible; that is, no lasting effects are observed in the an-
imals after the exposure is terminated. Consequently it is
not considered deleterious, although exposures that are
sufcient to produce behavioral disruption are undoubt-
edly close to those that would be thermally injurious if
sustained for sufcient lengths of time.
5.5. Mobile Phones and Health
The hazards discussed above are principally (if not exclu-
sively) thermal in nature, and require high exposure lev-
els. Much public and legal controversy has concerned the
possibility of hazards from use of mobile telephones or
Table 2. Thresholds for Perception of Microwave Energy by Humans
Frequency (GHz)
Threshold for Perception
(10s Exposure, 0.024m
2
Exposure Area) (W/m
2
)
Approximate Energy
Penetration Depth (m)
Approximate Temperature
Increase at Skin Surface
(1C) at Threshold
2.45 630 0.01 0.06
7.5 195 0.003 0.05
10 196 0.002 0.07
35 88 410
4
0.08
94 45 110
4
0.05
Source: Adapted from Ref. 33.
Table 1. Ranges of Threshold Currents for Indirect Effects, Including Children, Women, and Men
Threshold Current (mA) at Frequency
Indirect Effect 50/60Hz 1 kHz 100kHz 1MHz
Touch perception 0.20.4 0.40.8 2540 2540
Pain on nger contact 0.91.8 1.63.3 3355 2850
Painful shock/let-go threshold 816 1224 112224 Not determined
Severe shock/breathing difculty 1223 2141 160320 Not determined
Source: From Ref. 25.
518 BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY
living near cellphone base stations. The exposure levels
encountered in these two situations are very different, but
both are far below any threshold for a thermal hazard.
Certainly the most visible public issue has been the
possibility that use of mobile telephones can cause brain
cancer. Since the issue was rst raised (by a nonscientist on
a U.S. television show in 1992), several epidemiology stud-
ies have been completed, together with several long-term
animal studies designed as cancer assays. The results so
far have failed to identify a link between use of mobile
telephones and cancer. A review of the literature published
in 1999 found the evidence for a causal relationship be-
tween RF radiation from cellphones and cancer to be weak
to nonexistent [23]. However, given the long time that it
takes brain tumors to develop (years to decades between
initiation and clinical diagnosis), the studies completed to
date have only limited ability to address the question of
possible long-term effects from use of mobile telephones.
A number of studies have been conducted to address
health issues other than cancer. Many of them have re-
ported effects of using mobile telephones on neurological
function, response time, sleep, and other functions in hu-
man subjects. However, the effects have generally been
insignificant, with no clear health significance. Reviews of
the issue by a number of expert committees have generally
concluded that no convincing evidence exists for a hazard
from mobile phones or their base stations, but also that
numerous open questions still remain.
6. TWO MAJOR WESTERN EXPOSURE LIMITS
Government and nongovernment agencies throughout
the world have established guidelines for human expo-
sure to electromagnetic elds. The rationale for two
major Western limits [Institute of Electrical and Electron-
ics Engineers (IEEE) standard C95.1-1999 [24] and the
International Commission on Non-Ionizing Radiation Pro-
tection (ICNIRP) guideline [25]] is now considered (see
also Table 4). The IEEE committee has recently been
renamed International Committee on Electromagnetic
Safety (ICES).
The standards are complex, and readers should refer to
the original documents for authoritative statements of the
limits. They have similar rationales but differ in many
particulars. ICNIRP limits include basic restrictions,
which are limits on absorbed power (or other measure of
internal eld) based directly on adverse health effects, and
reference levels, which are limits on external eld
strength designed to ensure that the basic restrictions
are not exceeded.
Each standard has two tiers: for controlled and un-
controlled environments (IEEE) or for occupational
groups and the general public (ICNIRP). The controlled
limits in IEEE C95.1 are intended for environments in
which the exposed subjects are aware of the potential for
exposure, and correspond roughly to occupational expo-
sure guidelines in ICNIRP.
Table 4. Comparison of ICNIRP and IEEE Limits for Contact Current
A. ICNIRP Reference Levels for Time-Varying Contact Currents from Conductive Objects
Exposure Characteristics Frequency Range Maximum Contact Current (mA)
Occupational exposure r2.5kHz 1.0
2.5100kHz 0.4f
a
100kHz110MHz 40
General public exposure r2.5kHz 0.5
2.5100kHz 0.2f
a
100kHz110MHz 20
B. IEEE Induced and Contact Radiofrequency Currents
b
Maximum Current (mA)
Frequency Range Through Both Feet Through Each Foot Contact
3100kHz 0.9f
a
0.45f
a
0.45f
a
100kHz100MHz 90 45 45
a
Frequency in kilohertz.
b
Limits for uncontrolled environments; limits for controlled environments are higher by a factor of B2.
Table 3. Thresholds for Behavioral Disruption in Different Animals
Species and Conditions
225MHz
(Continuous-Wave) 1.3 GHz (Pulsed)
2.45GHz
(Continuous-Wave) 5.8 GHz (Pulsed)
Norwegian rat power density SAR 10mW/cm
2
(2.5W/kg) 128mW/cm
2
(5.0W/kg) 20mW/cm
2
(4.9W/kg)
Squirrel monkey power density SAR 45mW/cm
2
(4.5 W/kg) 40mW/cm
2
(7.2W/kg)
Rhesus monkey power density SAR 8 mW/cm
2
(3.2W/kg) 57mW/cm
2
(4.5W/kg) 67mW/cm
2
(4.7 W/kg) 140mW/cm
2
(8.4W/kg)
Source: Adapted from Ref. 34.
BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY 519
6.1. Limits to Whole-Body SAR
Both the IEEE and ICNIRP limits are designed to limit
the whole-body average SAR to 0.4W/kg (controlled or
occupational) or 0.08 W/kg (uncontrolled or general pub-
lic limits). The rationale, as described in ICNIRP is [25]
Established biological and health effects in the frequency
range from 10MHz to a few GHz are consistent with respons-
es to a body temperature rise of more than 11C. This level of
temperature increase results from exposure of individuals un-
der moderate environmental conditions to a whole-body SAR
of approximately 4Wkg
1
for about 30 minutes. Awhole-body
average SAR of 0.4Wkg
1
has therefore been chosen as the
restriction that provides adequate protection for occupational
exposure. An additional safety factor of 5 is introduced for ex-
posure of the public, giving an average whole-body SAR limit
of 0.08Wkg
1
.
The IEEE and ICNIRP limits are compared for one ex-
posure situation (general public for ICNIRP, uncontrolled
environments for IEEE) in Fig. 8. The antenna resonance
in the human at 70 MHz accounts for the lower limits near
that frequency in both limits.
6.2. Limits to Partial-Body SAR
For the head and trunk, ICNIRP limits for partial body
exposures are 2W/kg (general public) and 10 W/kg (occu-
pational), averaged over any 10g mass of contiguous tis-
sue. Limits elsewhere in the body are a factor of 2 higher.
The corresponding limits in the present IEEE limit are 1.6
and 8 W/kg, averaged over 1g of tissue (but is anticipated
to change with a new edition of the limit to be released in
2003).
6.3. Contact Currents
The limits in these two standards are compared in Fig. 8.
The ICNIRP limits are somewhat more restrictive, but
both sets of limits are clearly rather similar.
7. CONTROVERSIES AND UNRESOLVED ISSUES
The literature on biological effects and health hazards of
RF energy is complex and frequently inconsistent and
controversial even among scientists. Several areas of con-
troversy are discussed below.
7.1. Many Reported Biological Effects of RF Energy
The scientific literature abounds with reports of biological
effects of RF energy, including some at low exposure levels
that the investigators considered nonthermal. The stud-
ies vary greatly in biological endpoint, exposure level,
quality, and relevance to health and safety. Many of the
studies are controversial, due to technical criticisms of the
studies or the inability of other studies to conrm the ini-
tial ndings [26].
Faced with an inconsistent and incomplete literature
(the normal state of affairs in environmental health
issues), health agencies typically employ a weight of
evidence approach, using panels of experts to evaluate
critically all relevant information. These panels normally
give little weight to studies that are technically decient
or whose relevance to health is unclear. Health agencies
typically point to the incompleteness of the scientific
literature and the existence of open questions. For
example, a 1999 review by the Royal Society of Canada
concluded [27]
Scientific studies performed to date suggest that exposure to
low intensity non-thermal RF elds do not impair the health of
humans or animals. However, the existing scientific evidence
is incomplete, and inadequate to rule out the possibility that
these non-thermal biological effects could lead to adverse
health effects. Moreover, without an understanding of how
low energy RF elds cause these biological effects, it is difcult
to establish safety limits for non-thermal exposures.
7.2. Reports of Adverse Health Effects of RF Energy at Low
Exposure Levels
A major factor in the public controversy about the safety of
RF energy is the existence of a scattering of reports in the
medical literature of associations between adverse health
outcomes and presumed exposure to RF energy. For ex-
ample, there is one report of clusters of testicular cancer
cases among police ofcers using radar guns [28]. One in-
vestigator has reported an increase in cancer mortality in
Polish soldiers whose job classications suggest exposure
to RF radiation, compared to other soldiers used as con-
trols [29]. A more recent epidemiology study by a French
investigator reported an association between nonspecific
health symptoms and residence near a mobile base
station [30].
Perhaps even more perplexing, the medical literature
of the former Soviet Union and its former Warsaw Pact
Frequency, MHz
10
1
10
2
10
3
10
4
10
5
10
6
E
x
p
o
s
u
r
e

l
i
m
i
t
,

g
e
n
e
r
a
l

p
u
b
l
i
c
/
u
n
c
o
n
t
r
o
l
l
e
d
,

W
/
m
2
1
10
100
IEEE
ICNIRP
Figure 8. Comparison of ICNIRP and IEEE limits for general
population/uncontrolled environments. Limits for occupational
groups/controlled environments are generally 5 times higher
than for the general public/uncontrolled environments.
520 BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY
allies contains numerous reports of adverse health effects
in people exposed to RF energy. For example, Russian
physicians have identied a neurasthenic syndrome
among workers in radio equipment factories, which is
characterized by diverse and nonspecific symptoms such
as fatigue, anxiety, headache, and impotence [31]. The ex-
posures to these subjects was undocumented, but almost
certainly the levels of RF energy were far below Western
exposure limits.
So far, these reports have carried little weight in West-
ern assessments of possible health risks of RF energy, in
part because of technical deciencies in the studies and in
part because of lack of independent conrming evidence.
However, these reports would certainly tend to convince
many nonscientists that health hazards do exist from low-
level exposures to RF energy.
7.3. Large Variations in Exposure Limits to RF Energy
throughout the World
The underlying philosophy in setting exposure limits to
RFand microwave energy varies greatly in different coun-
tries, and there is a corresponding large variation in ex-
posure limits. These limits can be classied as science-
based (in that they were designed to avoid identied
hazards, on the basis of detailed analysis of the relevant
scientific literature) and precautionary (not designed to
exclude particular hazards). The limits for general public
exposure to 2000MHz energy (similar to that emitted by
many mobile telephone base stations) are compared in
Table 5.
7.4. Science-Based Limits
One group of science-based limits consists of the IEEE and
ICNIRP limits discussed above.
Avery different group of science-based limits is in effect
in Russia and most of its former Warsaw Pact allies. These
limits are far lower (as much as 100 times, if the limits are
expressed in terms of incident power density) than the
ICNIRP and the IEEE limits (as well as those of many
other Western countries). The Russian standards clearly
reect a conviction that prolonged exposure to low-level
RF elds causes health problems of many sorts. It has so
far proved impossible to reconcile these two sets of limits,
in part because of the brief nature of the Russian reports
(may of which lack sufcient technical detail to permit ex-
pert evaluation) and in part because of technical limita-
tions of the studies (which seldom provide information
about exposure) [32]. This situation goes back for many
years and is not likely to be resolved anytime soon.
7.5. Precautionary Limits
More recently, Italy, Switzerland, and a few other coun-
tries have adopted strict limits for RF exposure that are
even below those of Russia. To a large extent, these limits
reect public fears about the safety of mobile telephone
masts (whose exposure levels are invariably far below
ICNIRP and other Western limits). These limits are not
science-based, but are justied by the precautionary prin-
ciple, and thus are designed to be highly restrictive due to
uncertainty about the safety of RF energy at low exposure
levels.
BIBLIOGRAPHY
1. K. R. Foster and H. P. Schwan, Dielectric properties of tis-
suesa review, in C. Polk and E. Postow, eds., Handbook of
Biological Effects of Electromagnetic Radiation, 2nd ed., CRC
press, Boca Raton, FL, 1995.
2. S. Gabriel, R. W. Lau, and C. Gabriel, The dielectric proper-
ties of biological tissues. 2. Measurements in the frequency
range 10 Hz to 20 GHz, Phys. Med. Biol. 41:22512269 (1996).
3. T. Schmid, O. Egger, and N. Kuster, Automated E-eld scan-
ning system for dosimetric assessments, IEEE Trans. Micro-
wave Theory Tech. 44:105113 (1996).
4. C. H. Durney, H. Massoudi, and M. F. Iskander, Radiofre-
quency Radiation Dosimetry Handbook, 4th ed., Brooks Air
Table 5. Comparison of Five Different Exposure Limits for RF Energy at 2000MHz
a
.
Country
Limit for General Public
Exposure to RF Fields (2000MHz)
for Extended Periods of
Exposure (W/m
2
) (Applies
to Far-Field Exposure,
Extended Duration) Basis
ICNIRP (adopted in numerous countries worldwide) 10 Science-based
U.S. Federal Communications Commission Bulletin 65,
Evaluating Compliance with FCC Guidelines for Human Exposure
to Radiofrequency Electromagnetic Fields, Washington, DC, 1997.
10 Science-based
China: UDC 614.898.5GB 917588 0.1 Science-based
Russia: Sanitary Norms and Regulations 2.2.4/2.1.8.055-96 0.1 Science-based
Switzerland: Ordinance on Protection from Nonionising
Radiation (NISV) of Dec. 23, 1999
0.1 Precautionary
Typical maximum exposure from cellular base station mounted on
50-m tower (assuming a total effective radiated power of 2500W
in each sector, summed over all channels)
0.01
a
This 2000-MHz limit is similar to that used by many cellular telephones throughout the world. The limits are for long-term exposure to the general pop-
ulation.
.......................................................................................................................................................................................................
BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY 521
Force Base, San Antonio, TX, Report USAFSAM-TR-85-73;
available online at http://www.brooks.af.mil/AFRL/HED/
hedr/reports/handbook/home.html.
5. H. P. Schwan and K. R. Foster, RF-eld interactions with bi-
ological systems: Electrical properties and biophysical mech-
anisms, Proc. IEEE 68:104113 (1980).
6. K. R. Foster and E. D. Finch, Microwave hearing: Evidence
for thermoacoustic auditory stimulation by pulsed micro-
waves, Science 185:256258 (1974).
7. F. S. Barnes, Cell membrane temperature rate sensitivity
predicted from the Nernst equation, Bioelectromagnetics
5:113115 (1984).
8. H. Wachtel, D. Brown, and H. Bassen, Critical durations of
pulse microwave exposures that evoke body movements, Proc.
Bioelectromagnetics Soc. 12th Annual Meeting, June 1014,
1990, San Antonio, TX, p. 55.
9. R. C. Lee, D. Zhang, and J. Hannig, Biophysical injury mech-
anisms in electrical shock trauma, Ann. Rev. Biomed. Eng.
2:477509 (2000).
10. R. K. Adair, Biophysical limits on athermal effects of RF and
microwave radiation, Bioelectromagnetics 24:3948 (2003).
11. S. M. Michaelson and J. C. Lin, Biological Effects and Health
Implications of Radiofrequency Radiation, Plenum Press,
New York, 1987.
12. I. Chatterjee, D. Wu, and O. P. Gandhi, Human body imped-
ance and threshold currents for perception and pain for
contact hazard analysis in the VLF-MF band, IEEE Trans.
Biomed. Eng. 33:486494 (1986).
13. D. W. Blick, E. R. Adair, W. D. Hurt, C. J. Sherry, T. J. Walters,
and J. H. Merritt, Thresholds of microwave-evoked warmth sen-
sations in human skin, Bioelectromagnetics 18:403409 (1997).
14. P. J. Riu, K. R. Foster, D. W. Blick, and E. R. Adair, A thermal
model for human thresholds of microwave-evoked warmth
sensations, Bioelectromagnetics 18:578583 (1997).
15. T. J. Walters, D. W. Blick, L. R. Johnson, E. R Adair, and K. R.
Foster, Heating and pain sensation produced in human skin
by millimeter waves: Comparison to a simple thermal model,
Health Phys. 78:259267 (2000).
16. A. J. Welsh, Laser irradiation of tissue, in A. Schitzer and R.
C. Eberhart, eds., Heat Transfer in Medicine and Biology, Vol.
2, Plenum Press, New York, 1985, p. 167.
17. R. C. Alexander, J. A. Surrell, and S. D. Cohle, Microwave
oven burns to children: An unusual manifestation of child
abuse, Pediatrics 79:255260 (1987).
18. S. F. Cleary, Microwave cataractogenesis, Proc. IEEE 68:49
55 (1980).
19. P. S. Nawrot, D. I. McRee, and R. E. Staples, Effects of 2.45
GHz microwave radiation on embryofetal development in
mice, Teratology 24:303314 (1981).
20. R. P. Jensh, Behavioral teratologic studies using microwave
radiation: Is there an increased risk from exposure to cellular
phones and microwave ovens? Reprod. Toxicol. 11:601611
(1997).
21. Y. H. Liu, X. M. Li, R. P. Zou, and F. B. Li, Biopsies of human
testes receiving multiple microwave irradiationan histological
and ultramicroscopical study, J. Bioelectric. 10:213230 (1991).
22. J. A. DAndrea, Microwave radiation absorption: Behavioral
effects, Health Phys. 61:2940 (1991).
23. J. E. Moulder, L. S. Erdreich, R. S. Malyapa, J. Merritt, W. F.
Pickard, and Vijayalaxmi, Cell phones and cancer: What is
the evidence for a connection? Rad. Res. 151:513531 (1999);
available online at http://www.radres.org/rare_151_05_
0513.pdf.
24. IEEE Std C95.1-1991, 1999 ed., IEEE Standard for
Safety Levels with Respect to Human Exposure to Radio-
frequency Electromagnetic Fields, 3 kHz to 300 GHz (avail-
able from the Institute of Electrical and Electronics Engi-
neers, Inc., Service Center, 445 Hoes Lane, Piscataway, NJ
08854-1331).
25. International Commission on Non-Ionizing Radiation Protec-
tion, ICNIRP guidelines: Guidelines for limiting exposure to
time-varying electric, magnetic, and electromagnetic elds
(up to 300 GHz), Health Phys. 74:494522 (1998).
26. K. R. Foster and W. F. Pickard, Microwaves: The risks of risk
research, Nature 330:531532 (1987).
27. Royal Society of Canada, A Review of the Potential Health
Risks of Radiofrequency Fields from Wireless Telecommuni-
cation Devices, Ottawa, Canada, 1999, available online at
www.rsc.ca/english/RFreport.html.
28. R. L. Davis and F. K. Mosto, Cluster of testicular cancer in
police ofcers exposed to hand-held radar, Am. J. Indust. Med.
24:231233 (1993).
29. S. Szmigielski, Cancer morbidity in subjects occupationally
exposed to high frequency (radiofrequency and microwave)
electromagnetic radiation, Sci. Total Environ. 180:917
(1996).
30. R. Santini, P. Santini, J. M. Danze, P. Le Ruz, and M. Seigne,
Investigation on the health of people living near mobile tele-
phone relay stations: Incidence according to distance and sex,
Pathol. Biol. (Paris) 50:369373 (2002).
31. N. J. Roberts and S. M. Michaelson, Epidemiological
studies of human exposures to radiofrequency radiation. A
critical review, Int. Arch. Occup. Environ. Health. 56:169178
(1985).
32. M. H. Repacholi, Radiofrequency electromagnetic-eld expo-
sure standards, IEEE Eng. Med. Biol. 6:1821 (1987).
33. P. J. Riu, K. R. Foster, D. W. Blick, and E. R. Adair, A thermal
model for human thresholds of microwave-evoked warmth
sensations. Bioelectromagnetics 18:578583 (1997).
34. J. Osepchuk and R. C. Petersen, Safety standards for
exposure to RF electromagnetic elds, IEEE Microwave
Mag. 5769 (June 2001).
FURTHER READING
A vast amount of material is available about health and safety
issues related to RF energy, much available on the Internet.
Below is a selection of sources for further reading.
BOOKS
C. Polk and E. Postow, eds., Handbook of Biological Effects of
Electromagnetic Fields, 2nd ed., CRC Press, Boca Raton, FL,
1996. (A collection of review chapters including biological
effects of RF energy, dosimetry, and electrical properties
of tissue.)
S. M. Michaelson and J. C. Lin, Biological Effects and Health Im-
plications of Radiofrequency Radiation, Plenum Press, New
York, 1987 (a now-dated review, but with extensive discussion
of earlier literature).
O. P. Gandhi, ed., Biological Effects and Medical Applications of
Electromagnetic Energy, Prentice-Hall, Englewood Cliffs, NJ,
1990. (Several excellent review articles on bioeffects, dosime-
try, medical applications of RF energy.)
522 BIOLOGICAL EFFECTS OF RADIOFREQUENCY ENERGY AS RELATED TO HEALTH AND SAFETY
TUTORIALS AND REVIEWS
M. H. Repacholi, Low-level exposure to radiofrequency elds:
Health effects and research needs, Bioelectromagnetics 19:
119 (1998). (Recent review of research questions related to
biological effects of RF energy.)
C.-K. Chou, H. Bassen, J. Osepchuk, Q. Balzano, R. Petersen, M.
Meltz, R. Cleveland, J. C. Lin, and L. Heynick, Radiofrequency
electromagnetic exposure: tutorial review on experimental
dosimetry, Bioelectromagnetics 17:195208 (1996). (Excellent
review of dosimetry.)
K. A. Hossmann and D. M. Hermann, Effects of electromagnetic
radiation of mobile phones on the central nervous system,
Bioelectromagnetics 24:4962 (2003). (Review of possible neu-
rologic effects of RF energy, with reference to mobile phones.)
E. Litvak, K. R. Foster, and M. H. Repacholi, Health and safety
implications of exposure to electromagnetic elds in the fre-
quency range 300 Hz10 MHz, Bioelectromagnetics 23:6882
(2002). (One of the few available reviews of biological effects of
RF energy below 10 MHz.)
J. D. Boice and J. K. McLaughlin, Epidemiological studies of cel-
lular telephones and cancer riska review, Swedish Radiation
Protection Authority, Stockholm, 2002; available online at
http://www.ssi.se/ssi_rapporter/pdf/ssi_rapp_2002_16.pdf.
EXPERT REPORTS
These reports of expert groups summarize the current status of
knowledge about biological effects of RF energy, largely from
the perspective of possible health risks of mobile phones and
their base stations.
Health Council of the Netherlands, Mobile Telephones; an Eval-
uation of Health Effects, Health Council of the Netherlands,
The Hague, 2002, publication 2002; available online at http://
www.gr.nl/.
Independent Expert Group on Mobile Phones, Mobile Phones and
Health, UK, May 2000; available online at http://www.ie-
gmp.org.uk/index.htm.
New Zealand Ministry for the Environment, National Guidelines
for Managing the Effects of Radiofrequency Transmitters, Dec
2000; available online at http://www.mfe.govt.nz/publica-
tions/rma/radio-freq-guidelines-dec00.pdf.
The Royal Society of Canada, A Review of the Potential Health
Risks of Radiofrequency Fields from Wireless Telecommunica-
tion Devices, Ottawa, March 1999; available online at http://
www.rsc.ca/english/RFreport.html.
INTERNET SITES
Bioelectromagnetics Society, (the major professional society con-
cerned with studies of possible health and safety issues related
to RF energy); report available online at http://bioelectromag-
netics.org.
FAQ (frequently asked question) sheets related to electromagnet-
ic elds and health by John Moulder at The Medical College of
Wisconsin offers comprehensive and up-to-date reviews about
mobile phones and health and other issues; available online at
www.mcw.edu/gcrc/cop/cell-phone-health-faq/toc.html.
IEEE Committee on Man and Radiation (COMAR) prepares
technical information statements on issues related to health
and safety of electromagnetic elds; available online at http://
ewh.ieee.org/soc/embs/comar/.
International Commission on Non-Ionizing Radiation Protection
(ICNIRP) is an international agency that establishes exposure
guidelines for electromagnetic elds; information available on-
line at www.icnirp.de.
World Health Organization EMF Project offers an extensive array
of information including fact sheets about health and safety
issues related to RF energy, comparison of international expo-
sure limits; available online at www.who.int/peh-emf/.
BOUNDARY-VALUE PROBLEMS
BENJAMIN BEKER
GEORGE COKKINIDES
MYUNG JIN KONG
University of South Carolina
Columbia, South Carolina
Many problems in electrical engineering require solution
of integral or differential equations that describe physical
phenomena. The choice as to whether integral or differ-
ential equations are used to formulate and solve specic
problems depends on many factors, whose discussion is
beyond the scope of this article. This article strictly
concentrates on the use of the nite-difference method
(FDM) in the numerical analysis of boundary value pro-
blems associated with primarily static and to some extent
quasistatic electromagnetic (EM) elds. Although all ex-
amples presented herein deal with EM-related engineer-
ing applications, some numerical aspects of FDM are also
covered. Since there is a wealth of literature in numerical
and applied mathematics about the FDM, little will be
said about the theoretical aspects of nite differencing.
Such issues as the proof of existence or convergence of the
numerical solution will not be covered, while the appro-
priate references will be provided to the interested reader.
Instead, the coverage of FDM will deal with the details
about implementation of numerical algorithms, compact
storage schemes for large sparse matrices, the use of open
boundary truncation, and efcient handling of inhomoge-
neous and anisotiopic materials.
The emphasis will be placed on the applications of
FDM to three-dimensional boundary value problems
involving objects with arbitrary geometric shapes
that are composed of complex dielectric materials. Exam-
ples of such problems include modeling of discrete passive
electronic components, semiconductor devices and their
packages, and crosstalk in multiconductor transmission
lines. When the wavelength of operation is larger than
the largest geometric dimensions of the object that is
to be modeled, static or quasistatic formulation of the
problem is appropriate. In electrostatics, for example,
this implies that the differential equation governing the
physics (voltage distribution) is of Laplace type for the
source-free environment and of Poisson type in regions
containing sources. The solution of such second-order
partial-differential equations can be readily obtained
using the FDM.
BOUNDARY-VALUE PROBLEMS 523
Although the application of FDM to homogeneous
materials is simple, complexities arise as soon as inhomo-
geneity and anisotropy are introduced. The following
discussion will provide essential details on how to over-
come any potential difculties in adapting FDM to
boundary value problems involving such materials. The
analytical presentation will be supplemented with abun-
dant illustrations that demonstrate how to implement the
theory in practice. Several examples will also be provided
to show the complexity of problems that can be solved by
using FDM.
1. BRIEF HISTORY OF FINITE DIFFERENCING IN
ELECTROMAGNETICS
The utility of the numerical solution to partial-differential
equations (or PDEs) utilizing nite-difference approxima-
tion to partial derivatives was recognized early [1]. Im-
provements to the initial iterative solution methods,
discussed in Ref. 1, by using relaxation were subsequently
introduced [2,3]. However, before digital computers be-
came available, the applications of the FDM to the solu-
tion of practical boundary value problems was a tedious
and often impractical task. This was especially true when
a high level of accuracy was required.
With the advent of digital computers, numerical solu-
tion of PDEs became practical. They were soon applied to
various problems in electrostatics and quasistatics such as
in the analysis of microstrip transmission lines [4] among
many others. The FDM found quick acceptance in the
solution of boundary value problems within regions of
nite extent, and efforts were initiated to extend their
applicability to open region problems as well [5].
As a point of departure, it is interesting to note that
some of the earliest attempts to obtain the solution to
practical boundary value problems in electrostatics in-
volved experimental methods. They included the electro-
lytic tank approach and resistance network analog
technique [6], among others, to simulate the nite-differ-
ence approximations to PDEs.
Finally, it should be mentioned that in addition to the
application of FDM to static and quasistatic problems, the
FDM was also adapted for use in the solution of dynamic,
full-wave EM problems in time and frequency domains.
Most notably, the use of nite differencing was proposed
for the solution to Maxwells curl equations in the time
domain [7]. Since then, a tremendous amount of work on
the nite-difference time-domain (or FDTD) approach was
carried out in diverse areas of electromagnetics. This
includes antennas and radiation, scattering, microwave
integrated circuits, and optics. The interested reader can
consult the authoritative work in Ref. 8, as well as other
articles on eigenvalue and related problems in this en-
cyclopedia, for further details and additional references.
2. ENGINEERING BASICS OF FINITE DIFFERENCING
It is best to introduce the FDM for the solution of engi-
neering problems, which deal with static and quasistatic
electromagnetic elds, by way of example. Today, just
about every elementary text in electromagneticsas well
as newer, more numerically focused introductory EM text-
bookswill have a discussion on FDM (e.g., pp. 241246 of
Ref. 9 and Sect. 4.4 of Ref. 10) and its use in electrostatics,
magnetostatics, waveguides, and resonant cavities. Re-
gardless, it will be benecial to briey go over the basics
of electrostatics for the sake of completeness and to provide
a starting point for generalizing FDM for practical use.
3. GOVERNING EQUATIONS OF ELECTROSTATICS
The analysis of electromagnetic phenomena has its roots
in the experimental observations made by Michael Fara-
day. These observations were cast into mathematical form
by James Clerk Maxwell in 1873 and veried experimen-
tally by Heinrich Hertz 25 years later. When reduced to
electrostatics, they state that the electric eld at every
point in space within a homogeneous medium obeys the
following differential equations [9]
r
~
EE0 1
r
.
e
0
e
r
~
EEr
v
2
where r
v
is the volumetric charge density, e
0
(E8.854
10
12
F/m) is the permittivity of freespace, and e
r
is the
relative dielectric constant. The use of the vector identity
rrf0 in Eq. (1) allows for the electric eld intensity,
~
EE
(V/m), to be expressed in terms of the scalar potential
~
EE rf. When this is substituted for the electric eld in
Eq. (2), a second-order PDE for the potential f is obtained
r
.
e
r
rf r
v
=e
0
3
which is known as the Poisson equation. If the region of
space, where the solution for the potential is sought, is
source-free and the dielectric is homogeneous (i.e., e
r
is
constant everywhere), Eq. (3) reduces to the Laplace
equation r
2
f0.
The solution to the Laplace equation can be obtained in
several ways. Depending on the geometry of the problem,
the solution can be found analytically or numerically using
integral or differential equations. In either case, the goal is
to determine the electric eld in space due to the presence
of charged conductors, given that the voltage on their
surface is known. For example, if the boundaries of the
charged conductors are simple shapes, such as a rectan-
gular box, circular cylinder, or sphere, then the boundary
conditions (constant voltage on the surface) can be easily
enforced and the solution can be obtained analytically. On
the other hand, when the charged object has an irregular
shape, the Laplace equation cannot be solved analytically
and numerical methods must be used instead.
The choice as to whether integral or differential equa-
tion formulation is used to determine the potential de-
pends heavily on the geometry of the boundary value
problem. For example, if the charged object is embedded
within homogeneous medium of innite extent, integral
equations are the preferable choice. They embody the
boundary conditions on the potential f at innity and
524 BOUNDARY-VALUE PROBLEMS
reduce the numerical effort to nding the charge density
on the surface of the conductor (see Sect. 5.2 or 4.3 of Ref. 9
or 10 for further details).
On the other hand, when the boundary value problem
includes inhomogeneous dielectrics (i.e., e
r
varying from
point to point in space), the surface integral equation
methods are no longer applicable (or are impractical).
Instead, such problems can be formulated using volu-
metric PDE solvers such as the FDM.
It is important to note that similar considerations (to
those stated above) also apply to the solution of the
Poisson equation. In this case, in addition to the integra-
tion over the conductor boundaries, integration over the
actual sources (charge density) must also be performed.
The presence of the sources has the same effects on PDEs
and FDM, as their effects must be taken into account at all
points in space where they exist.
3.1. Direct Discretization of Governing Equation
To illustrate the utility and limitations of FDM and to
introduce two different ways of deriving the numerical
algorithm, consider the geometry shown in Fig. 1. Note
that for the sake of clarity and simplicity, the initial
discussion will be restricted to two dimensions.
The innitely long, perfectly conducting circular cylinder
in Fig. 1 is embedded between two dielectrics. To determine
the potential everywhere in space, given that the voltage on
the cylinder surface is V
0
, the FDM will be used. There are
two approaches that might be taken to develop the FDM
algorithm. One approach would be to solve Laplace (or
Poisson) equations in each region of uniform dielectric
and enforce the boundary conditions at the interface be-
tween them. The other route would involve development of
a general volumetric algorithm, which would be valid at
every point in space, including the interface between the
dielectrics. This would involve seeking the solution of a
single, Laplace-type (or Poisson) differential equation:
r
.
e
r
x; yrf e
r
x; y
@
2
f
@x
2

@
2
f
@y
2

@e
r
@x
@f
@x

@e
r
@y
@f
@y

rx; y
e
0
8
>
>
<
>
>
:
4
which is valid everywhere, except for the surface of the
conductor.
The numerical approach to solving Eq. (4) starts with
the approximation to partial derivatives using nite dif-
ferences. This requires some form of discretization for the
space (area or volume in two or three dimensions) where
the potential is to be computed. The numerical solution of
the PDE will lead to the values of the potential at a nite
number of points within the discretized space. Figure 2
shows one possible discretization scheme for the cylinder
in Fig. 1 and its surroundings. The points form a two-
dimensional (2D) grid and need not be uniformly spaced.
Note that the gridpoints, where the potential is to be
computed, have to be dened all along the gridlines
to allow for properly approximating the derivatives in
Eq. (4). In other words, the gridlines cannot abruptly
terminate or become discontinuous within the grid.
Using nite differences, the rst-order derivative at
any point in the grid can be approximated as follows
@f
@ x

i; j
%
f
I
1
; J
f
I1; j

h
i
h
i1
=2

f
i 1; j
f
i; j
2

f
i; j
f
i1; j
2
!
1
h
i
h
i1
=2

f
i 1; j
f
i1; j
h
i
h
i1
5
with the help of intermediate points I, J (darkened circles
in Fig. 2). The approximation for the second derivatives
can be obtained in a similar manner and is given by
@
@x
@f
@ x
!
i; j
%
@f
@ x

I 1; j

@f
@ x

I1; j
h
i
h
i1
=2

f
i 1; j
f
i; j
h
i

f
i; j
f
i1; j
h
i1
!

1
h
i
h
i1
=2
6
y
x
V
0
a
c
1
c
1
Figure 1. Charged circular cylinder embedded between two
different dielectrics.
j
h
i 1
h
j 1
h
j
i , j
I , J 1
h
i
i
i
j
a
Grid cell
c
1
c
1
Figure 2. Staircase approximation to boundary of circular
cylinder and notation for grid dimensions.
BOUNDARY-VALUE PROBLEMS 525
Once all derivatives in Eq. (4) are replaced with their
respective nite-difference approximations and all similar
terms are grouped together, the discrete version of Eq. (4)
takes on the following form:
f
i; j
%
1
Y
i; j
Y
i 1
f
i 1; j
Y
i1
f
i1; j
Y
j 1
f
i; j 1
Y
j1
f
i; j1

0
r
i; j
r
i1; j
r
i; j1
r
i1; j1

e
0
8
>
>
<
>
>
:
7
In this equation, the Y factors contain the material para-
meters and distances between various adjacent grid-
points. They are expressed below in a compact form:
Y
i1

2
h
i
h
i1

2
e
i; j1
e
i; j

3h
i
h
i1
h
i
h
i
h
i1
h
i1
0
B
B
B
B
@
1
C
C
C
C
A
8
>
>
>
>
<
>
>
>
>
:
e
i1; j1
e
i1; j

h
i1
h
i
h
i
h
i
3h
i1
h
i1
0
B
B
B
B
@
1
C
C
C
C
A
9
>
>
>
>
=
>
>
>
>
;
8
Y
j1

2
h
j
h
j1

2
e
i1; j
e
i; j

3h
j
h
j1
h
j
h
j
h
j1
h
j1
0
B
B
B
B
B
@
1
C
C
C
C
C
A
8
>
>
>
>
>
<
>
>
>
>
>
:
e
i1; j1
e
i; j1

h
j1
h
j
h
j
h
j
3h
j1
h
j1
0
B
B
B
B
B
@
1
C
C
C
C
C
A
9
>
>
>
>
>
=
>
>
>
>
>
;
9
Y
i; j
Y
i 1; j
Y
i1; j
Y
i; j 1
Y
i; j1
10
It is important to add that in deriving these equations, a
particular convention for associating the medium para-
meters to individual grid cells was employed. Specically,
it was assumed that the medium parameter values of the
entire grid cell were associated with (or assigned to) the
lower left corner of that cell. For example, e
i,j
is assumed to
be constant over the shaded grid cell area shown in Fig. 2,
while e
i, j 1
is constant over the hatched area, which is
directly below.
Observe what are the consequences of converting the
continuous PDE given in Eq. (4) to its approximate dis-
crete form stated in Eq. (7). First, the boundary value
problem over the continuous space, shown in Fig. 1, was
mapped onto a discrete grid (see Fig. 2). Clearly, if the
number of gridpoints increases, then spacing between
them will become smaller. This provides a better approx-
imation to the actual continuous problem. In fact, in the
limit as the number of gridpoint reaches innity, the
discrete and continuous problems become identical.
In addition to illustrating the mapping of a contin-
uous problem to its discrete analog, Fig. 2 also clearly
demonstrates one of FDMs undesirable artifacts. Note
that objects with smooth surfaces are replaced with a
staircased approximation. Obviously, this approxima-
tion can be improved by reducing the discretization grid
spacing. However, this will increase the number of points
where the potential has to be calculated, thus increasing
the computational complexity of the problem. One way to
overcome this is to use a nonuniform discretization, as
depicted in Fig. 2. Specically, ner discretization can be
used in the region near the smooth surface of the cylinder
to better approximate its shape, followed by gradually
increasing the gridpoint spacing between the cylinder and
grid truncation boundary.
At this point, it is also appropriate to add that other,
more rigorous methods have been proposed for incorpor-
ating curved surfaces into the nite-difference type of
algorithms. They are based on special-purpose differen-
cing schemes, which are derived by recasting the same
PDEs into their equivalent integral forms. They exploit
the surface or contour integration and are used to replace
the regular differencing algorithm on the curved surfaces
or contours of smooth objects. This approach was already
implemented for the solution of dynamic full-wave pro-
blems [11] and could be adapted to electrostatic boundary
value problems as well.
Finally, Eq. (7) also shows that the potential at any
point in space, which is source-free, is a weighted average
of the potentials at the neighboring points only. This is
typical of PDEs, because they represent physical phenom-
ena only locallythat is, in the immediate vicinity of the
point of interest. As will be shown later, one way to
propagate the local information through the grid is to
use an iterative scheme. In this scheme, the known
potential, such as V
0
at the surface of the conducting
cylinder in Fig. 1, is carried throughout the discretized
space by stepping through all the points in the grid. The
iterations are continued until the change in the potential
within the grid is very small.
3.2. Indirect Discretization of Governing Equation
As shown in the previous section, appropriate nite-
difference approximations were required for the rst-
and second-order derivatives in order to convert Eq. (4)
to a discrete form. Intermediate points (I, J) were used
midway between the regular grid nodes for obtaining
average values of the potential, its rst derivatives, and
dielectric constants to facilitate the derivation of the nal
update equation for the potential. This can be avoided and
an alternative, but equally valid nite-differencing
scheme to that given in Eq. (7), which for the sake of
brevity is restricted to the Laplace equation only, can be
obtained. The rst step is to simplify Eq. (4) by recasting it
into an integrodifferential form. To achieve this, Eq. (4)
should be integrated over a volume that completely en-
closes any one of the grid nodes. This will be referred to as
the volume of the unit cell (V
U
), which is bounded by
526 BOUNDARY-VALUE PROBLEMS
surface S
U
(see Fig. 3). Stokes theorem is applied to
replace the volume integration by a surface integral:
Z
V
U
r
.
e
r
x; yrf dv
I
S
U
e
r
x; yrf
.
^ nnds

I
S
U
e
r
x; y
@f
@n
ds 0
11
where ^ nn is the unit vector, normal to S
U
and pointing out
of it.
To illustrate this indirect discretization procedure in
two dimensions, consider surface S
U, i, j
(i.e., just a con-
tour) shown in Fig. 3, which completely encloses gridpoint
i, j. The integral in Eq. (11) reduces to four terms, each
corresponding to one of the faces of S
U,i, j
. For example, the
integral over the right edge (or face) of S
U,i, j
can be
approximated as
f
i 1; j
f
i; j
h
i
h
j1
2
e
i; j1

h
j
2
e
i; j

12
When the remaining integrals are evaluated and the like
terms are grouped together, the nal form of the indirect
FDM algorithm is obtained. This algorithm is identical in
form to that given earlier in Eq. (7). However, the weight-
ing Y factors are different from those appearing in Eqs. (8)
and (9). Their complete expressions are given by
Y
i 1
i 1
& '
h
j1
e
i; j 1
i 1; j 1
( )
h
j
e
i; j
i 1; j
( )
0
B
B
B
@
1
C
C
C
A
1
2h
i
i 1
( )
13
where Y
i, j
is the sum of all other Ys, the same as before
[see Eq. (10)].
It should be added that this approach has been sug-
gested several times in the literaturefor example, in
Refs. 12 and 13. Therein, Eq. (11) was specically used to
enforce the boundary conditions at the interface between
different dielectrics only in order to connect FDM algo-
rithms based on the Laplace equation for the homoge-
neous regions. However, there is no reason not to use
Eq. (11) at every point in space, especially if the boundary
value problem involves inhomogeneous media. This form
of FDM scheme is completely analogous to that presented
in the previous section. In fact, it is a little simpler to
derive and involves fewer arithmetic operations.
3.3. Numerical Implementation in Two Dimensions
There are several important numerical issues that must
be addressed prior to implementing FDM on the computer.
Questions such as how to terminate the grid away from
the region of interest and which form of the FDM algo-
rithm to choose must be answered rst. The following
discussion provides some simple answers, postponing the
more detailed treatment until later.
3.3.1. Simplistic Grid Boundary Truncation. Clearly,
since even todays computers do not have innite re-
sources, the computational volume (or space) must some-
how be terminated (see Fig. 4). The simplest approach is to
place the truncation boundary far away from the region of
interest and to set the potential on it equal to zero. This
approach is valid as long as the truncation boundary is
placed far enough not to interact with the charged objects
within, as for example the staircased cylinder shown in
Fig. 4. The downside of this approach is that it leads to
large computational volumes, thereby requiring unneces-
sarily high computer resources. This problem can be
partly overcome by using a nonuniform grid, with pro-
gressively increasing spacing from the cylinder toward the
truncation boundary. It should be added that there are
other ways to simulate the open-boundary conditions,
which is an advanced topic and will be discussed later.
3.3.2. Iteration-Based Algorithm. Several iteration me-
thods can be applied to solve Eq. (7), each leading to
different convergence rates (i.e., how quickly an accepta-
ble solution is obtained). A complete discussion of this
h
i 1
h
j 1
h
j
h
i
i
j
S
U
V
U
Figure 3. Closed surface completely enclosing a grid node.
V
0
j = 3
j = 2
j = 1
i = 1 i = 3
82
71
60
54
48
42 45
91
92
B
Grid truncation boundary
34
12 13
1 2 3 A
23
36 37
i = 2
c
1

c
2
Figure 4. Complete discretized geometry and computational
space for cylinder in Fig. 2.
BOUNDARY-VALUE PROBLEMS 527
topic, as well as of the accuracy of the numerical approx-
imations in FDM, appears in Ref. 14 and will not be
repeated here. The interested reader may also nd Ref. 15
quite useful, because it covers such topics in more rigorous
detail and includes a comprehensive discussion on the
proof of the existence of the nite-differencing solution
to PDEs. However, for the sake of brevity, this article
deals with the most popular and widely used approach,
which is called successive overrelaxation (SOR) (see,
e.g., Ref. 14).
SOR is based on Eq. (7), which is rearranged as
f
p1
i; j
% f
p
i; j

O
Y
i; j
Y
i 1
f
p
i 1; j
Y
i1
f
p1
i1; j
Y
j 1
f
p
i; j 1
Y
j1
f
p1
i; j1
Y
i; j
f
p
i; j

1 Of
p
i; j

O
Y
i; j
Y
i 1
f
p
i 1; j
Y
i1
f
p1
i1; j
Y
j 1
f
p
i; j 1
Y
j1
f
p1
i; j1

14
In this equation, superscripts p and p1 respectively
denote the present and previous iteration steps and O is
the so-called overrelaxation factor, whose value can vary
from 1 to 2. Note that O accelerates the change in the
potential from one iteration to the next at any point in the
grid. It can be a constant throughout the entire relaxation
(solution) process or be varying according to some heur-
istic scheme. For example, it was found that the overall
rate of convergence is improved by setting O near 1.8 at
the start of the iteration process and gradually reducing it
to 1.2 with the numbers of iterations.
At this point, the only remaining task is to dene the
appropriate criteria for terminating the FDM algorithm.
Although there are rigorous ways of selecting the termina-
tion criterion, the following approach, which seems to work
quite well, is presented instead. As stated below, it is based
on calculating the change in the potential between succes-
sive iterations at every point in the grid and comparing the
maximum value to the (user-selectable) error criterion:
ERR
max

X
Nx
i 1
X
Ny
j 1
maxjf
p1
i; j
f
p
i; j
j 15
Although it is simple, the redeeming feature of this
approach is that the error is computed globally within
the grid, rather than within a particular single grid node.
The danger in monitoring the convergence of the algorithm
at a single node may lead to premature termination or to
an unnecessarily prolonged execution.
Now that an error on which the algorithm termination
criterion is based has been dened, the iteration process
can be initiated. Note that there are several ways to
march through the grid. Specically, the updating of
the potential may be started from point A, as shown in
Fig. 4, and end at point B, or vice versa. If the algorithm
works in this manner, the solution will tend, to be
articially biased toward one region of the grid, and
the potential is more converged in regions where the
iteration starts. The obvious way to avoid this is to change
the direction of the marching process after every few
iterations. As a result, the potential will be updated
throughout the grid uniformly and will converge at the
same rate. Note from Fig. 4 that the potential only needs
to be computed at the internal points of the grid, since the
potential at the outer boundary is (for now) assumed to be
zero. Moreover, it should also be evident that the potential
at the surface of the cylinder, as well as inside it, is known
(V
0
) and need not be updated during the iteration process.
3.3.3. Matrix-Based Algorithm. Implementation of the
nite-difference algorithms is not restricted to relaxation
techniques only. The solution of Eq. (4) for the electrostatic
potential can also be obtained using matrix methods.
To illustrate this, the FDM approximation to Eq. (4)
namely, Eq. (7)will be rewritten as
Y
i 1
f
i 1; j
Y
i1
f
i1; j
Y
j 1
f
i; j 1
Y
j1
f
Y
i; j
f
i; j
% 0
16
This equation must be enforced at every internal point in
the grid, except at the surface of and internal to the
conductors, where the potential is known (V
0
). For the
particular example of the cylinder shown in Fig. 4, these
points are numbered 1 through 92. This implies that there
are 92 unknowns, which must be determined. To accom-
plish this, Eq. (16) must be enforced at 92 locations in the
grid, leading to a system of 92 linear equations that must
be solved simultaneously.
To demonstrate how the equations are set up, consider
nodes 1, 36, and 37 in detail. At node 1 (where i j 2),
Eq. (16) reduces to
Y
i
3
f
12
Y
j
3
f
2
Y
2;2
f
1
0 17
where the fact that the potential at the outer boundary
nodes (i, j) (1,2) and (2,1) is zero was taken into account
and superscripts i and j on Y were introduced as a
reminder whether they correspond to Y
i71
or Y
j71
. In
addition, the potential at nodes (i, j) (2,2), (3,2), and (2,3)
was also relabeled as f
1
, f
2
, and f
12
, respectively. Simi-
larly, at nodes 36 (i, j 4,5) and 37 (i, j 5,5), Eq. (16)
becomes
Y
i
5
f
37
Y
i
3
f
35
Y
j
6
f
44
Y
j
4
f
25
Y
4;5
f
36
0 18
and
Y
i
4
f
36
Y
j
4
f
26
Y
5;5
f
37
Y
i
6
V
0
Y
j
6
V
0
V
37
19
where [in Eq. (19)] the known quantities (the potentials on
the surface of the cylinder at nodes 5,6 and 6,5) were
moved to the right-hand side.
Similar equations can be obtained at the remaining
free grid nodes where the potential is to be determined.
Once Eq. (16) has been enforced everywhere within the
grid, the resulting set of equations can be combined into
the following matrix form:
528 BOUNDARY-VALUE PROBLEMS
which can be written more compactly as
Yf V
0
21
Clearly, the coefcient matrix of this system of equations
is very sparse, containing few nonzero elements. In
fact, for boundary value problems in two dimensions
with isotropic dielectrics, there will be at most ve
nonzero terms in a single row of the matrix. Although
standard direct matrix solution methods, such as Gauss
inversion of LU decomposition and backsubstitution,
can be applied to obtain the solution to Eq. (20), they
are wasteful of computer resources. In addition to per-
forming many unnecessary numerical operations with
zeros during the solution process, a large amount of
memory has to be allocated for storing the coefcient
matrix. This can be avoided by exploiting the sparsity of
the coefcient matrix, using well-known sparse matrix
storage techniques, and taking advantage of specialized
sparse matrix algorithms for direct [16] or iterative
[1719] solution methods.
Since general-purpose solution techniques for sparse
linear equation systems are well documented, such as in
Refs. 1619, they will not be discussed here. Instead, the
discussion will focus on implementation issues specic to
FDM. In particular, issues related to the efcient con-
struction of the [Y] matrix in Eq. (20) and to the sparsity
coding scheme are emphasized.
In the process of assembling [Y], as well as in the
postprocessing computations such as in calculating the E
eld, it is necessary to quickly identify the appropriate
entries f
k
within the vector [f], given their locations in
the grid (i, j). Such searching operations are repeated
many times, as each element of [Y] is stored in its
appropriate location. Note that the construction of [Y], in
large systems (5005000 equations), may take as much
CPU time as the solution itself. Thus, optimization of the
search for index locations is important.
One approach to quickly nd a specic number in
an array of N numbers is based on the well-known
bisection search algorithm (Sect. 3.4 in Ref. 17). This
algorithm assumes that the numbers in the array are
arranged in an ascending order and requires, at most,
log
2
(N) comparisons to locate a particular number in
the array. In order to apply this method, a mapping
that assigns a unique code to each allowable combina-
tion of the grid coordinates i, j is dened. One such
mapping is
code i
.
N
y
j 22
where N
y
is the total number of gridpoints along the j
direction. The implementation of this algorithm starts by
dening two integer arrays: CODE and INDEX. The array
CODE holds the identication codes [computed from
Eq. (22)], while INDEX contains the corresponding value
of k. Both CODE and INDEX are sorted together so that
the elements of CODE are rearranged to be in an ascend-
ing order. Once generated and properly sorted, these
Y
2;2
Y
j
3
0
.
0 Y
i
3
0
.
0 0
. .
0 0 0 Y
j
4
0
.
0
0 0
. .
0 0 0 0 Y
j
4
0
.
2
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
4
Y
j
3
Y
4;5
Y
i
5
0
.
0 Y
j
6
0
. .
0
0 Y
i
4
Y
5;5
0
. .
0
. . .
0
3
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
5
f
1
f
2
.
.
.
f
12
.
f
25
f
26
.
.
f
35
f
36
f
37
.
.
.
f
44
.
f
92
2
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
4
3
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
5

0
.
.
0
V
37
.
0
2
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
4
3
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
5
20
BOUNDARY-VALUE PROBLEMS 529
arrays can be used to nd the index k (to identify f
k
) for
grid coordinates (i, j) in the following manner:
1. Given i and j, compute code i N
y
j.
2. Find the array index m, such that code CODE(m),
using the bisection search algorithm.
3. Look up k using k INDEX(m), thereby identifying
the appropriate f
k
, given i, j.
The criteria for selecting a particular sparsity coding
scheme for the matrix [Y] are (1) the minimization of
storage requirements and (2) optimization of matrix op-
erationsin particular, multiplication and LU factoring.
One very efcient scheme is based on storing [Y] using
four one-dimensional arrays:
1. Real array DIAG(i) diagonal entry of row i
2. Real array OFFD(i) ith nonzero off-diagonal entry
(scanned by rows)
3. Integer array IROW(i) index of rst nonzero off-
diagonal entry of row i
4. Integer array ICOL(i) column number of ith non-
zero off-diagonal entry (scanned by rows)
Assuming a system of N equations, the arrays DIAG,
IROW, OFFD, and ICOL have N, N1, 4N, and 4N
entries, respectively. Therefore, the total memory required
to store a sparsity-coded matrix [Y] is approximately 40N
bytes (assuming 32-bit storage for both real and integer
numbers). On the other hand, 4N
2
bytes would be needed
to store the full form of [Y]. For example, in a system with
1000 equations, the full storage mode requires 4 mega-
bytes (MB), while the sparsity-coded matrix occupies only
40 kilobytes (KB) of computer memory.
Perhaps the most important feature of sparsity coding
is the efciency with which multiplication and other
matrix operations can be performed. This is best illu-
strated by a sample FORTRAN code needed to multiply
a matrix stored in this mode, by a vector B(i):
DO I 1; N
CI DIAGI BI
DO J IROWI; IROWI 1 1
CI CI OFFDJ BICOLJ
ENDDO
ENDDO
23
This double loop involves 5N multiplications and 4N
additions, without the need of search and compare opera-
tions. To perform the same operation using the brute-force,
full-storage approach would require N
2
multiplications
and N
2
additions. Thus, for a system of 1000 unknowns,
the sparsity-based method is at least 200 times faster
than the full-storage approach in performing matrix
multiplication.
To solve Eq. (20), [Y] can be inverted and the inverse
multiplied by [V
0
]. However, for sparse systems, complete
matrix inversion should be avoided. The reason is that, in
most cases, the inverse of a sparse matrix is full, for which
the advantages of sparsity coding cannot be exploited. The
solution of sparsity coded linear systems is typically
obtained by using the LU decomposition, since usually
the L and U factors are sparse. Note that the sparsity of
the L and Ufactor matrices can be signicantly affected by
the ordering of the grid nodes (i.e., in which sequence [f]
was lled). Several very successful node ordering schemes
that are associated with the analysis of electrical networks
were reported for the solution of sparse matrix equations
(see Ref. 16). Unfortunately, the grid node connectivity in
typical FDM problems is such that the L and U factor
matrices are considerably fuller than the original matrix
[Y], even if the nodes are optimally ordered. Therefore,
direct solution techniques are not as attractive for use in
FDM as they are for large network problems.
Unlike the direct solution methods, there are iterative
techniques for solving matrix equations, which do not
require LU factoring. One of them is the conjugate-gra-
dient method [1719].
This method uses a sequence of matrix/vector multi-
plications, which can be performed very efciently using
the sparsity coding scheme described above.
3.3.4. Convergence. Given the FDM equations in ma-
trix form, either direct [16] or iterative methods [1719]
such as the conjugate-gradient method (CGM) can be
readily applied. It is important to point out that CGM-
type algorithms are considerably faster than direct solu-
tion, provided a good initial guess is used. One simple
approach is to assume that the potential is zero every-
where, but on the surface of the conductor, and let this be
an initial guess to start the CGM algorithm. For this
initial guess, the convergence is very poor and the solution
takes a long time. To improve the initial guess, several
iterations of the SOR-based FDM algorithm can be per-
formed to calculate the potential everywhere within the
grid. It was found that for many practical problems, 1015
iterations provide a very good initial guess for CGM.
From the performance point of view, the speed of CGM
was most noticeable when compared to the SOR-based
algorithm. In many problems, CGM was found to be an
order of magnitude faster than SOR. In all fairness to
SOR, its implementation, as described above, can be
improved considerably by using the so-called multigrid/
multilevel acceleration [20,21]. The idea behind this
method is to perform the iterations over coarse and ne
grids alternatively, where the coarse gridpoints also coin-
cide with and are a part of the ne grid. This means that
iterations are rst performed over a coarse grid, then
interpolated to the ne grid and iterated over the ne grid.
More complex multigrid schemes involve several layers of
grids with different levels of discretization, with the
iterations being performed interchangeably on all grids.
Finally, it should be pointed out that theoretical aspects
of convergence for algorithms discussed thus far are well
documented and are outside the main scope of this article.
The interested reader should consult Ref. 15, 18, or 19
for detailed mathematical treatment and assessment of
convergence.
530 BOUNDARY-VALUE PROBLEMS
4. ADVANCED TOPICS
4.1. Open-Boundary Truncation
If the electrostatic boundary value problem consists of
charged conductors in a region of innite extent, then the
simplest approach to truncate the computational (or FDM)
boundary is with an equipotential wall of zero voltage.
This has the advantage of being easy to implement, but
leads to erroneous solution, especially if the truncation
boundary is too close to the charged conductors. On the
other hand, placing it too far from the region of interest
may result in unacceptably large computational volume,
which will require large computational resources.
Although some early attempts to overcome such dif-
culties [5] provided the initial groundwork, rigorous ab-
sorbing boundary truncation operators were more
recently introduced [2224] for dynamic problems, which
can be modied for electrostatics. They are based on
deriving mathematical operators that help simulate the
behavior of the potential on a virtual boundary truncation
surface, which is placed close to charged conductors (see
Fig. 5). In essence, these operators provide the means for
numerically approximating the proper behavior of the
potential at innity within a computational volume of
nite extent.
Such absorbing boundary conditions (ABCs) are based
on the fact that the potential due to any 3D charge
distribution is inversely proportional to the distance mea-
sured from it. Consider an arbitrary collection of charged
conductors shown in Fig. 5. Although it is located in free
unbounded space, a ctitious surface will be placed
around it, totally enclosing all conductors. If this surface
is far away from the charged conductor systemnamely, if
~rr is much greater than ~rr
0
then the dominant radial
variation of potential, f, will be given by
1
j~rr
~
r
0
r
0
j
!
1
r
24
If the ctitious boundary is moved closer toward the
conductor assembly, then the potential will also include
additional terms with higher inverse powers of r. These
terms will contribute to the magnitude of the potential
more signicantly than those with lower inverse powers of
r, as r becomes small.
The absorbing boundary conditions emphasize the
effect of leading (dominant) radial terms on the magnitude
of the potential evaluated on the ctitious (boundary
truncation) surface. The ABCs provide the proper analytic
means to annihilate the nonessential terms, instead of
simply neglecting their contribution. Numerically, this
can be achieved by using the so-called absorbing boundary
truncation operators.
In general, absorbing boundary operators can be of any
order. For example, as shown in Ref. 23, the rst- and
second-order operators in 3D have the following forms:
B
1
u
@u
@r

u
r
O
1
r
3

!0 as r !1 25
B
2
u
@
@r

3
r

@u
@r

u
r

O
1
r
5

!0 as r !1 26
where u is the scalar electric potential function, f, that
satises the Laplace equation, and r j~rrj is the radial
distance measured from the coordinate origin (see Fig. 5).
Since FDM is based on the iterative solution to the La-
place equation, small increases in the overall lattice
(discretized 3D space whose planes are 2D grids) size do
not slow the algorithm down signicantly, nor do they
require an excessive amount of additional computer mem-
ory. As a result, from a practical standpoint, the ctitious
boundary truncation surface need not be placed too close
to the region of interest, therefore not requiring the use of
high-order ABC operators in order to simulate proper
behavior of the potential at lattice boundaries accurately.
Consequently, in practice, it is sufcient to use the rst-
order operator, B
1
, to model open boundaries. Previous
numerical studies suggest that this choice is indeed
adequate for many engineering problems [25].
To be useful for geometries that mostly conform to
rectangular coordinates, the absorption operator B
1
,
when expressed in Cartesian coordinates, takes on the
following form
@u
@x
%
u
x

y
x
@u
@y

z
x
@u
@z

27
@u
@y
%
u
y

x
y
@u
@x

z
y
@u
@z

28
@u
@z
%
u
z

x
z
@u
@x

y
z
@u
@y

29
where 8 signs correspond to the outward pointing unit
normal vectors ^ nn ^ xx; ^ yy; ^ zz for operators in Eqs. (27),
(28), and (29), respectively.
The nite-difference approximations to the preceding
equations that have been employed in the open boundary
z
n = r
r
y
x
/ /
r
R = r r
Charged conductor
system
Fictitious outer
boundary
Figure 5. Virtual surface used for boundary truncation.
BOUNDARY-VALUE PROBLEMS 531
FDM algorithm are given by
u
i 1;j;k
u
i1; j; k

h
i
h
i1

x
i; j; k
x
ref

u
i; j; k

y
i; j; k
y
ref
u
i; j 1;k
u
i; j1;k

h
j
h
j1

z
i; j;k
z
ref
u
i; j; k 1
u
i; j; k1

h
k
h
k1

!
30
u
i; j 1;k
u
i; j1;k

h
j
h
j1

y
i; j; k
y
ref

u
i; j; k

x
i; j; k
x
ref
u
i 1; j; k
u
i1; j; k

h
i
h
i1

z
i; j; k
z
ref
u
i; j; k1
u
i; j; k1

h
k
h
k1

!
31
u
i; j; k 1
u
i; j; k1

h
k
h
k1

z
i; j; k
z
ref

u
i; j; k

x
i; j; k
x
ref
u
i 1; j; k
u
i1; j; k

h
i
h
i1

y
i; j; k
y
ref
u
i; j 1;k
u
i; j1;k

h
j
h
j1

!
32
where (x, y, z)
ref
are the x,y,z components of a vector
pointing (referring) to the geometric center of the charged
conductor assembly, with other quantities that appear in
Eqs. (30)(32) shown in Fig. 6. It is important to add that
the (x, y, z)
i, j,k
(x, y, z)
ref
terms are the x, y, z components
of a vector from the truncation boundary to the geometric
center of the charged conductor system.
Notice that Fig. 6 graphically illustrates the FDM
implementation of the open-boundary truncation on lat-
tice faces aligned along the xz plane. On this plane, the
normal is in the y direction, for which Eq. (31) is the FDM
equivalent of the rst-order absorbing boundary operator
in Cartesian coordinates. Similarly, Eqs. (32) and (30) are
used to simulate the open boundary on xy and yz faces of
the lattice, respectively.
When reduced to two dimensions, the absorption op-
erator, B
1
, in Cartesian coordinates, takes on the form
given below:
@u
@x
%
1
x

y
x
@u
@y

33
@u
@y
%
1
y

@u
@x

34
The discrete versions of these equations can be written as
u
i 1;j
u
i1;j

h
i
h
i1

x
i; j
x
ref

&
u
i; j

y
i; j
y
ref
u
i; j 1
u
i; j1

h
j
h
j1

!'
35a
u
i; j 1
u
i; j1

h
j
h
j1

y
i; j
y
ref

&
u
i; j

x
i; j
x
ref
u
i 1; j
u
i1;j

h
i
h
i1

!'
35b
where, as before, 8 correspond to the outward pointing
unit normal vectors ^ nn ^ xx; ^ yy for operators in Eqs. (33)
and (34) or (35a) and (35b), respectively. The points x
i, j
and y
i, j
denote those points in the grid that are located one
cell away from the truncation boundary, while x
ref
and y
ref
correspond to the center of the cylinder in Figs. 2 and 4.
Finally, another approach to open-boundary trunca-
tion, which is worth mentioning, involves the regular
nite-difference scheme supplemented by the use of elec-
trostatic surface equivalence [26]. A virtual surface S
v
is
dened near the actual grid truncation boundary. The
electrostatic potential due to charged objects, enclosed
within S
v
, is computed using the regular FDM algorithm.
Subsequently, it is used to calculate the surface charge
density and surface magnetic current, which are propor-
tional to the normal and tangential components of the
electric eld on S
v
.
Once the equivalent sources are known, the potential
between the virtual surface and the grid truncation
boundary can be readily calculated (for details, see Ref.
26). This procedure is repeated every iteration, and since
the potential on the virtual surface is estimated correctly,
it produces a physical value of the potential on the
truncation boundary. As demonstrated in Ref. 26, this
approach leads to very accurate results in boundary value
problems with charged conductors embedded in open
regions.
It is vastly superior to simply using the grounded
conductor to terminate the computational space.
h
k +1
h
i +1
h
i
h
k
h
j
(i , j, k + 1)
(i , j, k)
Point on the lattice
truncation boundary
z
y
h
j +1
x
(i 1, j, k)
(i + 1, j, k)
(i , j + 1, k)
(i , j 1, k)
(i , j , k 1)
Figure 6. Detail of FDM lattice near the boundary truncation
surface.
532 BOUNDARY-VALUE PROBLEMS
4.2. Inclusion of Dielectric Anisotropy
4.2.1. Network Analog Approach. Many materials such
as printed-circuit board (PCB) and microwave circuit
substrates, which are commonly used in electrical engi-
neering, exhibit anisotropic behavior. The electrical prop-
erties of these materials vary with direction and have to be
described by a tensor instead of a single scalar quantity.
This section will examine how the anisotropy affects the
FDM and how the algorithm must be changed to accom-
modate the solution of 3D problems containing such
materials. The theoretical development presented below
is a generalization of that available in Ref. 27 and is
restricted to linear anisotropic dielectrics only.
In an attempt to provide a more intuitive interpreta-
tion to the abstract nature of the FDM algorithm, an
equivalent-circuit model will be used for linear inhomoge-
neous, anisotropic regions. This approach is called resis-
tance network analog [6]. It was initially proposed for
approximating the solution of the Laplace equation in
two dimensions experimentally, with a network of physi-
cal resistors whose values could be adjusted to correspond
to the weighting factors [e.g., the Ys in Eq. (7)] that appear
in the FDM algorithm. Since its introduction, the resis-
tance network approach has been implemented numeri-
cally in the analysis of (1) homogeneous dielectrics in 3D
(28) and (2) simple biaxial anisotropic materials (described
by diagonal permitivitty tensors) in 2D [29].
Since the resistance network analog gives a physical
interpretation to FDM, the discretized versions of the
Laplace equations for anisotropic media will be recast
into this form. As the details of FDM were described
earlier, only the key steps in developing the two-dimen-
sional model are summarized below. Moreover, for the
sake of brevity, the discussion of the three-dimensional
case will be limited to the nal equations and their
pictorial interpretation.
The Laplace equation for boundary value problems
involving inhomogeneous and anisotropic dielectrics in
three dimensions is given by
r
.
e
0
e
r
x; y; z
.
rfx; y; z 0 36
In this equation, [e
r
] stands for the relative dielectric
tensor and is dened as
e
e
xx
e
xy
e
xz
e
yx
e
yy
e
yz
e
zx
e
zy
e
zz
2
6
6
4
3
7
7
5
37
Since the material properties need not be homogeneous in
the region of interest, the elements of [e
r
] are assumed to
be functions of position. The dielectric is assumed to
occupy only part of the modeling (computational) space,
and its properties may vary from point to point. When
Eq. (37) is substituted into Eq. (36) and rewritten in a
matrix form as
@
@x
@
@y
@
@z
!
.
e
xx
@f
@x
e
xy
@f
@y
e
xz
@f
@z
e
yx
@f
@x
e
yy
@f
@y
e
yz
@f
@z
e
zx
@f
@x
e
zy
@f
@y
e
zz
@f
@z
2
6
6
6
6
6
6
6
6
4
3
7
7
7
7
7
7
7
7
5
0 38
it provides the starting point for developing the corres-
ponding FDM algorithm.
After eliminating the z-dependent terms and fully ex-
panding the equation above by following the notation used
throughout this article, the nite-difference approximation
for the potential at every nodal point in a 2Dgrid is given by
f
p1
i; j
1 Of
p
i; j

O
Y
i; j

f
p
i 1; j
Y
i 1
f
p1
i1; j
Y
i1

f
p
i; j 1
Y
j 1
f
p1
i; j1
Y
j1

f
p
i 1; j 1
Y
i 1; j 1
f
p1
i1; j1
Y
i1; j1

f
p
i1; j 1
Y
i1; j 1
f
p1
i 1; j1
Y
i 1; j1

8
>
>
>
>
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
>
>
>
>
:
9
>
>
>
>
>
>
>
>
>
>
>
>
=
>
>
>
>
>
>
>
>
>
>
>
>
;
39
where
Y
i1

e
yy
i; j1
e
yy
i; j
e
yy
i1; j1
e
yy
i1; j

0
B
@
1
C
A
1
h
i

2
h
i
h
i1

2
h
i
h
i1

2
h
j
h
j1

e
zy
i; j
e
zy
i1; j

e
zy
i1; j1
e
zy
i; j1

40
Y
j1

e
zz
i1; j
e
zz
i; j

e
zz
i1; j1
e
zz
i; j1

0
B
@
1
C
A
1
h
j

2
h
j
h
j1

2
h
i
h
i1

2
h
j
h
j1

e
yz
i; j1
e
yz
i; j

e
yz
i1; j1
e
yz
i1; j

41
Y
i1; j1
Y
i1; j1

2
h
i
h
i1

2
h
j
h
j1

e
yz
i; j
e
yz
i1; j
e
yz
i1; j1
e
yz
i; j1
e
zy
i; j
e
zy
i1; j
e
zy
i1; j1
e
zy
i; j1

42
Y
i; j
Y
i 1
Y
i1
Y
j 1
Y
j1
43
Note that, unlike the treatment of isotropic dielectrics,
the permittivity of each cell is now described by a tensor
BOUNDARY-VALUE PROBLEMS 533
(see Fig. 7). In addition, the presence of the anisotropy
is responsible for added coupling between the voltage f
i, j
and voltages f
i71, j71
(actually all four combinations of the
subscripts).
The symbols Y in Eq. (39) can be interpreted as
admittances representing the electrical link between
the gridpoint voltages. The resulting equivalent network
for Eq. (39) can thus be represented pictorially as shown in
Fig. 8.
Similarly, after fully expanding Eq. (36) in three
dimensions, the following nite-difference approximation
for the potential at every nodal point in a 3D lattice can be
obtained:
f
pI
i; j; k
1 Of
p
i; j; k

Of
new
Y
i; j; k
44
where f
new
is dened as
f
new
f
p
i 1; j; k
Y
i 1
Y
A
1
f
p1
i1; j; k
Y
i1
Y
A
1

f
p
i; j 1; k
Y
j 1
Y
A
2
f
p1
i; j1; k
Y
j1
Y
A
2

f
p
i; j; k1
Y
k1
Y
A
3
f
p1
i; j; k1
Y
k1
Y
A
3

Y
A
4
f
p
i 1; j 1; k
f
p1
i1; j1; k
f
p
i1; j 1; k
f
p
i 1; j1; k

Y
A
5
f
p
i 1; j; k1
f
p1
i1; j; k1
f
p
i1; j; k1
f
p
i 1; j; k1

Y
A
6
f
p
i; j 1; k1
f
p1
i; j1; k1
f
p
i; j 1; k1
f
p
i; j1; k 1

45
and
Y
i; j; k
Y
i 1
Y
i1
Y
j 1
Y
j1
Y
k1
Y
k1
The Y terms appearing in Eqs. (44) and (45) are given by
Y
i1

2
h
i
h
i1
T
1; xx
1
h
i1

2
h
i
h
i1

T
2; xx
1
h
i1

2
h
i
h
i1
!
46
Y
i 1

2
h
i
h
i1
T
1; xx
1
h
i

2
h
i
h
i1

T
2; xx
1
h
i

2
h
i
h
i1
!
47
Y
j1

2
h
j
h
j1
T
1; yy
1
h
j1

2
h
j
h
j1

T
2; yy
1
h
j1

2
h
j
h
j1
!
48
Y
j 1

2
h
j
h
j1
T
1; yy
1
h
j

2
h
j
h
j1

T
2; yy
1
h
j

2
h
j
h
j1
!
49
Y
k1

2
h
k
h
k1
T
1; zz
1
h
k1

2
h
k
h
k1

T
2; zz
1
h
k1

2
h
k
h
k1
!
50
Y
A
1

2
h
j
h
j1

2
h
i
h
i1

T
1; yx
T
2; yx

2
h
k
h
k1

2
h
i
h
i1

T
1; zx
T
2; zx

51
Y
A
2

2
h
j
h
j1

2
h
i
h
i1

T
1; xy
T
2; xy

2
h
k
h
k1

2
h
j
h
j1

T
1; zy
T
2; zy

52
Y
A
3

2
h
k
h
k1

2
h
i
h
i1

T
1; xz
T
2; xz

2
h
k
h
k1

2
h
j
h
j1

T
1; yz
T
2; yz

53
[c
i 1
,
j
]
[c
i 1
,
j 1
] [c
i
,
j 1
]
[c
i
,
j
]
i , j
i + 1, j 1
i , j + 1

i 1, j + 1
i 1, j 1 i , j 1
i 1, j
i + 1, j + 1
i + 1, j
h
j 1
h
j
h
j
h
i 1
Figure 7. Detail with FDM cell for anisotropic medium in two
dimensions.
i 1, j 1 i + 1, j 1 i , j 1
i , j + 1
i 1, j
i 1, j + 1 i + 1, j + 1
i + 1, j
h
j 1
h
j
h
i
Y
j +1
Y
i +1
Y
j 1
Y
i 1
Y
i 1, j + 1
Y
i +1, j + 1
Y
i 1, j 1
Y
i +1, j 1
h
i 1
Figure 8. Network analog for 2D FDM algorithm at gridpoint i, j
for arbitrary anisotropic medium.
534 BOUNDARY-VALUE PROBLEMS
Y
A
4
2
2
h
j
h
j1

2
h
i
h
i1

T
1; xy
T
2; xy
8

T
1; yx
T
2; yx
8

54
Y
A
5
2
2
h
k
h
k1

2
h
i
h
i1

T
1; xz
T
2; xz
8

T
1; zx
T
2; zx
8

55
Y
A
6
2
2
h
j
h
j1

2
h
k
h
k1

T
1; yz
T
2; yz
8

T
1; zy
T
2; zy
8

56
with the T terms having the following forms:
T
1; xx
e
xx
i; j1; k1
e
xx
i; j; k1
e
xx
i; j1; k
e
xx
i; j; k
57a
T
2; xx
e
xx
i1; j1; k1
e
xx
i1; j; k1
e
xx
i1; j1; k
e
xx
i1; j; k
57b
T
1; yy
e
yy
i1; j; k1
e
yy
i; j; k1
e
yy
i1; j; k
e
yy
i; j; k
58a
T
2; yy
e
yy
i1; j1; k1
e
yy
i; j1; k1
e
yy
i1; j1; k
e
yy
i; j1; k
58b
T
1; zz
e
zz
i1; j1; k
e
zz
i; j1; k
e
zz
i1; j; k
e
zz
i; j; k
59a
T
2; zz
e
zz
i1; j1; k1
e
zz
i; j1; k1
e
zz
i1; j; k1
e
zz
i; j; k1
59b
T
1; xy
e
xy
i; j1; k1
e
xy
i; j; k1
e
xy
i; j1; k
e
xy
i; j; k
60a
T
2; xy
e
xy
i1; j1; k1
e
xy
i1; j; k1
e
xy
i1; j1; k
e
xy
i1; j; k
60b
T
1; xz
e
xz
i; j1; k1
e
xz
i; j; k1
e
xz
i; j1; k
e
xz
i; j; k
61a
T
2; xz
e
xz
i1; j1; k1
e
xz
i1; j; k1
e
xz
i1; j1; k
e
xz
i1; j; k
61b
T
1; yx
e
yx
i1; j; k1
e
yx
i; j; k1
e
yx
i1; j; k
e
yx
i; j; k
62a
T
2; yx
e
yx
i1; j1; k1
e
yx
i; j1; k1
e
yx
i1; j1; k
e
yx
i; j1; k
62b
T
1;yz
e
yz
i1; j; k1
e
yz
i; j; k1
e
yz
i1; j; k
e
yz
i; j; k
63a
T
2; yz
e
yz
i1; j1; k1
e
yz
i; j1; k1
e
yz
i1; j1; k
e
yz
i; j1; k
63b
T
1;zx
e
zx
i1; j1; k
e
zx
i; j1; k
e
zx
i1; j; k
e
zx
i; j; k
64a
T
2; zx
e
zx
i1; j1; k1
e
zx
i; j1; k1
e
zx
i1; j; k1
e
zx
i; j; k1
64b
T
1; zy
e
zy
i1; j1; k
e
zy
i; j1; k
e
zy
i1; j; k
e
zy
i; j; k
65a
T
2; zy
e
zy
i1; j1; k1
e
zy
i; j1; k1
e
zy
i1; j; k1
e
zy
i; j; k1
65b
Note that Eq. (45) has a similar interpretation as its 2D
counterpart Eq. (39). It can also be represented by an
equivalent network, whose diagonal terms are shown in
Fig. 9. For clarity, the off-diagonal terms, which provide
the connections of f
i, j,k
to the voltages at the remaining
nodes in Eq. (45), are shown separately in Fig. 10.
4.2.2. Coordinate Transformation Approach. Coordinate
transformations can be used to simplify the solution to
electrostatic boundary value problems. Such transforma-
tions can reduce the complexity arising from complicated
geometry or from the presence of anisotropic materials. In
general, these methods utilize coordinate transformation
to map complex geometries or material properties into
simpler ones, through a specic relationship that links
each point in the original and transformed problems,
respectively.
One class of coordinate transformations, known as
conformal mapping, is based on modifying the original
complex geometry to one for which an analytic solution is
available. This technique requires extensive mathemati-
cal expertise in order to identify an appropriate coordinate
transformation function. Its applications are limited to a
few specic geometrical shapes for which such functions
exist. Furthermore, the applications are restricted to two-
dimensional problems. Even though this technique can be
very powerful, it is usually rather tedious and thus it is
considered beyond the scope of this article. The interested
reader can refer to Ref. 30, among others, for further
details.
i + 1, j , k
i 1, j ,k
i , j , k +1
i , j +1, k
i , j , k
i , j , k 1
Y
j +1
+Y
2
A
Y
j 1
Y
1
A
Y
k 1
Y
3
A
Y
j 1
Y
2
A
Y
j +1
+Y
1
A
Y
k +1
+Y
1
A
i , j 1, k
Figure 9. Network analog for 3D FDM algorithm at gridpoint i, j
for anisotropic dielectric with diagonal permittivity tensor.
i + 1, j 1, k 1 i + 1, j + 1, k 1
i + 1, j, k 1
i 1, j 1, k + 1 i 1, k + 1
i 1, j +1, k 1
i 1, j +1, k
i 1 , j +1, k + 1
i, j +1, k 1
i + 1, j 1, k + 1
i, j 1, k + 1
i + 1, j 1, k
Figure 10. Network analog for 3D FDM algorithm at gridpoint i,
j for arbitrary anisotropic dielectric.
BOUNDARY-VALUE PROBLEMS 535
The second class of coordinate transformations reduces
the complexity of the FDM formulation in problems invol-
ving anisotropic materials. As described in the previous
section, the discretization of the Laplace equation in
anisotropic regions [Eq. (36)] is considerably more compli-
cated than the corresponding procedure for isotropic
media [Eq. (7)]. However, it can be shown that a sequence
of rotation and scaling transformations can convert any
symmetric permittivity tensor into an identity matrix (i.e.,
free space). As a result, the FDM solution of the Laplace
equation in the transformed coordinate system is consid-
erably simplied, since the anisotropic dielectric is elimi-
nated.
To illustrate the concept, this technique will be demon-
strated with two-dimensional examples. In 2D (no z
dependence is assumed) the Laplace equation can be
written as
r
.
e
r
x; yrf 0 66
where
e
r

e
xx
e
xy
e
yx
e
yy
" #
67
If the principal (crystal or major) axes of the dielectric are
aligned with the coordinate system of the geometry, then
the off-diagonal terms vanish. Otherwise, [e
r
] is a full
symmetric matrix. In this case, any linear coordinate
transformation of the form
x
0
y
0
" #
A
x
y
" #
68
(where [A] is a 22 nonsingular matrix of constant
coefcients) also transforms the permittivity tensor as
follows:
e
0
A
1
e
r
A 69
Next, consider the structure shown in Fig. 11a. It consists
of a perfect conductor (metal) embedded in an anisotropic
dielectric, all enclosed within a rectangular conducting
shell. The eld within the rectangular shell must be
determined given the potentials on all conductors. In
this example, [e
r
] is assumed to be diagonal:
e
r

e
xx
0
0 e
yy
" #
70
By scaling the coordinates with
A
1=

e
xx
p
0
0 1=

e
yy
p
" #
71
the permittivity can be transformed into an identity
matrix. The geometry of the structure is deformed as
shown in Fig. 11b, with the corresponding rectangular
discretization grid depicted in Fig. 11c. Note that the
locations of the unknown potential variables are marked
by white dots, while the conducting boundaries are repre-
sented by known potentials and their locations are de-
noted by black dots. The potential in the transformed
boundary value problem can now be computed by applying
the FDM algorithm, which is specialized for free space,
since [e
r
] is an identity matrix.
Once the potential is computed everywhere, other
quantities of interest, such as the E eld and charge,
can be calculated next. However, to correctly evaluate the
(c)
(b)
(a)
y
y
y
Major axis
y
Metal
Anisotropic dielectric
Equivalent isotropic dielectric
x
x
x
x
(d)
Figure 11. Graphical representation of coordinate transforma-
tion for homogeneous anisotropic dielectric with diagonal permit-
tivity tensor.
536 BOUNDARY-VALUE PROBLEMS
required space derivatives, transformation back to origi-
nal coordinates is required, as illustrated in Fig. 11d. Note
that in spite of the resulting simplications, this method is
limited to cases where the entire computational space is
occupied by a single homogeneous anisotropic dielectric.
In general, when the principal (or major) axes of the
permittivity are arbitrarily orientated with respect to the
coordinate axes of the geometry, [e
r
] is a full symmetric
matrix. In this case, [e
r
] can be diagonalized by an
orthonormal coordinate transformation. Specically, there
exists a rotation matrix of the form
A
cos y sin y
sin y cos y
" #
72
such that the product
e
0
A
T
e
r
A 73
is a diagonal matrix. The angle y is dened as the angle by
which the coordinate system should be rotated to align it
with the major axes of the dielectric.
Consider the structure shown in Fig. 12a, which is
enclosed in a metallic shell. However, in this example the
nonconducting region of interest includes both free space
and an anisotropic dielectric. Furthermore, the major axis
of [e
r
] is at 301 with respect to that of the structure. The
effect of rotating the coordinates by y 301 leads to a
geometry shown in Fig. 12b. In the transformed coordi-
nate system, the major axis of the permittivity is horizon-
tal and [e
r
] is a diagonal matrix. Observe that this
transformation does not affect the dielectric properties of
the free-space region (or of any other isotropic dielectrics,
if present). However, the subsequent scaling operation for
transforming the properties of the anisotropic region to
free space is not useful. Such transformation also changes
the properties of the original free-space region to those
exhibiting anisotropic characteristics. Regardless of this
limitation, the coordinate rotation alone considerably
simplies the FDM algorithm of Eq. (45) to
f
new
f
p
i 1; j; k
Y
i 1
Y
A
1
f
p1
i1; j; k
Y
i1
Y
A
1

f
p
i; j 1; k
Y
j 1
Y
A
2
f
p1
i; j1; k
Y
j1
Y
A
2

74
where all z-dependent (or k) terms have been removed.
Without the rotation, the permittivity is characterized
by Eq. (67). Under such conditions, the corresponding
FDM update equation includes four additional potential
variables, as shown below:
f
new
f
p
i 1; j; k
Y
i 1
Y
A
1
f
p1
i1; j; k
Y
i1
Y
A
1

f
p
i; j 1; k
Y
j 1
Y
A
2
f
p1
i; j1; k
Y
j1
Y
A
2

Y
A
4
f
p
i 1; j 1; k
f
p1
i1; j1; k

f
p
i1; j 1; k
f
p
i 1; j1; k

75
The simplication resulting from coordinate rotation in
three dimensions is even more signicant. In the general
y
Major axis
Metal
Free space
Dielectric
x
x
y
x
(b)
(a)
y
(c)
x
y
(d)

Figure 12. Graphical representation
of coordinate transformation for inho-
mogeneous anisotropic dielectric with
diagonal permittivity tensor.
BOUNDARY-VALUE PROBLEMS 537
case, the full FDM algorithm [Eq. (45)] contains 18 terms,
while in the rotated coordinates the new equation has
only 6.
Next, a rectangular discretization grid is constructed
for the transformed geometry as shown in Fig. 12c, with
the unknown potential represented by white dots and
conducting boundaries denoted by black dots. As can be
seen, the rotation complicates the assignment (or deni-
tion) of the boundary nodes. In general, a ner discretiza-
tion may be required to approximate the metal boundaries
more accurately.
Once the potential eld is computed, the transforma-
tion back to the original coordinates is performed by
applying the inverse rotation [A]
T
, as illustrated in
Fig. 11d. Note that in the original coordinate system, the
grid is rotated and, as such, complicates the computation
of electric eld. In addition to the required coordinate
mapping, this method is also limited to boundary value
problems containing only one type of anisotropic dielec-
tric, though any number of isotropic dielectric regions may
be present.
The examples above illustrate that coordinate trans-
formations are benecial in solving a narrow class of
electrostatic problems. Undoubtedly, considerable compu-
tational savings can be achieved in the calculation of the
potential using FDM. However, the computational over-
head associated with the pre- and postprocessing can be
signicant, since the geometry is usually complicated by
such transformations.
5. SAMPLE NUMERICAL RESULTS
To illustrate the versatility of FDM in solving engineering
problems that involve arbitrary geometries and inhomo-
geneous materials, consider the cross section of the micro-
wave eld-effect transistor (FET) shown in Fig. 13. Note
that this device is composed of many different materials,
each of different thickness and cross-sectional prole. The
FET is drawn to scale, with the 1mm thickness of the
buffer layer serving as a reference. FDM can be used to
calculate the potential and eld distribution throughout
the entire cross section of the FET. This information can
be used by the designer to investigate such effect as
material breakdown near the metallic electrodes. In addi-
tion, the computed eld information can be used to
determine the parasitic capacitance matrix of the struc-
ture, which can be used to improve the circuit model of
this device and is very important in digital circuit design.
Finally, it should be noted that the losses associated with
the silicon can also be computed using FDM as shown in
Eq. (25).
It should be added that in addition to displaying the
potential distribution over the cross section of the FET,
Fig. 13 also illustrates the implementation of open-bound-
ary truncation operators. Since the device is located in an
open boundary environment, it was necessary to arti-
cially truncate the computation space (or 2D grid). Note
that, as demonstrated in Ref. 25, only the rst-order
operator was sufcient to obtain accurate representation
of the potential in the vicinity of the electrodes as well as
near the boundary truncation surface.
A sample with three-dimensional geometry that can be
easily analyzed with the FDM is shown in Fig. 14. The
insulator in the multilayer ceramic capacitor is assumed
to be anisotropic barium titanate dielectric, which is
commonly used in such components. The permittivity
tensor is diagonal and its elements are e
xx
1540, e
yy

290, and e
zz
1640. To demonstrate the effect of anisotropy
on this passive electrical component, its capacitance was
calculated as a function of the misalignment angle be-
tween the crystal axes of the insulator and the geometry of
the structure (see Fig. 15).
For the misalignment angle (or rotation of axes) in the
yz plane, the capacitance of this structure was computed.
The results of the computations are plotted in Fig. 16.
Note that the capacitance varies considerably with the
rotation angle. Such information is invaluable to a de-
signer, since the goal of the design is to maximize the
capacitance for the given dimensions of the structure.
V
gs
= 0.75 V
V
ds
= 2.75 V
SiO
2
Si substrate
Ground plane
GaAs
Buffer layer
Source
Drain
Gate
SiO
2
Figure 13. Equipotential map of DC-biased micro-
wave FET. [From B. Beker and G. Cokkinides, Com-
puter-aided quasi-static analysis of coplanar
transmission lines for microwave integrated circuits
using the nite difference method, Int. J. MIMICAE
4(1):111119. Copyright r 1994, Wiley.]
538 BOUNDARY-VALUE PROBLEMS
The preceding examples are intended to demonstrate
the applicability of FDM to the solution of practical
engineering boundary value problems. FDM has been
used extensively in analysis of other practical problems.
The interested reader can nd additional examples where
FDM was used in Refs. 3137.
6. SUMMARY
Since the strengths and weaknesses of FDM were men-
tioned throughout this article, as were the details dealing
with the derivation and numerical implementation of this
method, they need not be repeated. However, the reader
should realize that FDM is best suited for boundary value
problems with complex geometries and arbitrary material
composition. The complexity of the problem is the primary
motivating factor for investing the effort into developing a
general-purpose volumetric analysis tool.
Acknowledgments
The authors wish to express their sincere thanks to many
members of the technical staff at AVX Corporation for
initiating, supporting, and critiquing the development and
implementation of many concepts presented in this arti-
cle, as well as for suggesting practical uses of FDM. Many
thanks also go to Dr. Deepak Jatkar for his help in
extending FDM to general anisotropic materials.
BIBLIOGRAPHY
1. H. Liebmann, Sitzungsber. Bayer. Akad. Munchen 385 (1918).
2. R. V. Southwell, Relaxation Methods in Engineering Science,
Clarendon Press, Oxford, UK, 1940.
3. R. V. Southwell, Relaxation Methods in Theoretical Physics,
Clarendon Press, Oxford, UK, 1946.
4. H. E. Green, The numerical solution of some important
transmission line problems, IEEE Trans. Microwave Theory
Tech. 17(9):676692 (1969).
5. F. Sandy and J. Sage, Use of nite difference approximation to
partial differential equations for problems having boundaries
at innity, IEEE Trans. Microwave Theory Tech. 19(5):484
486 (1975).
6. G. Liebmann, Solution to partial differential equations with
resistance network analogue, Br. J. Appl. Phys. 1(4):92103
(1950).
7. K. S. Yee, Numerical solution of initial boundary value
problems involving Maxwells equations in isotropic media,
IEEE Trans. Anten. Propag. 14(3):302307 (1966).
8. A. Taove, Computational Electrodynamics: The Finite-Dif-
ference Time-Domain Method, Artech House, Boston, 1995.
9. N. N. Rao, Elements of Engineering Electromagnetics, 4th ed.,
Prentice-Hall, Englewood Cliffs, NJ, 1994.
10. S. R. H. Hoole and P. R. P. Hoole, A Modern Short Course in
Engineering Electromagnetics, Oxford Univ. Press, New York,
1996.
11. T. G. Jurgens, A. Taove, K. Umashankar, and T. G. Moore,
Finite-difference time-domain modeling of curved surfaces,
IEEE Trans. Anten. Propag. 40(4):357366 (1992).
12. M. Naghed and I. Wolf, Equivalent capacitances of coplanar
waveguide discontinuities and interdigitated capacitors using
three-dimensional nite difference method, IEEE Trans.
Microwave Theory Tech. 38(12):18081815 (1990).
13. M. F. Iskander, Electromagnetic Fields & Waves, Prentice-
Hall, Englewood Cliffs, NJ, 1992, Sect. 4.8.
14. R. Haberman, Elementary Applied Partial Differential Equa-
tions with Fourier Series and Boundary Value Problems,
Prentice-Hall, Englewood Cliffs, NJ, 1983, Chap. 13.
15. L. Lapidus and G. H. Pinder, Numerical Solutions of Partial
Differential Equations in Science and Engineering, Wiley,
New York, 1982.
L
t
= 3.06
L
e
= 2.67
W
e
= 1.03
W
t
= 1.61
H = 0.42
Figure 14. Geometry of a multilayer ceramic chip capacitor. All
dimensions are in millimeters.
y
z
x
z
y
0
Figure 15. Denition of rotation angle for anisotropic insulator.
0
0 10 20 30 40 50 80 100
5
10
15
20
C
a
p
a
c
i
t
a
n
c
e

(
n
F
)
Rotation angle (degrees)
25
30
35
40
45
90 70 60
Figure 16. Capacitance of multilayer chip capacitor as a function
of rotation angle of the insulator.
BOUNDARY-VALUE PROBLEMS 539
16. W. F. Tinney and J. W. Walker, Direct solution of sparse
network equations by optimally ordered triangular factoriza-
tion, IEEE Proc. 55(11) (1967).
17. W. T. Press, B. P. Flanney, S. A. Teukolsky, and W. T.
Vetterling, Numerical Recipes: The Art of Scientic Comput-
ing, 2nd ed., Cambridge Univ. Press, Cambridge, UK, 1992,
Section 2.7.
18. Y. Saad, Iterative Methods for Sparse Linear Systems, PWS
Publishing, Boston, 1996.
19. G. H. Golub and C. F. Van Loan, Matrix Computations, 3rd
ed., Johns Hopkins Univ. Press, Baltimore, 1996, Chap. 10.
20. R. E. Philips and F. W. Schmidt, Multigrid techniques for the
numerical solution of the diffusion equation, Num. Heat
Transfer 7:251268 (1984).
21. J. H. Smith, K. M. Steer, T. F. Miller, and S. J. Fonash,
Numerical modeling of two-dimensional device structures
using Brandts multilevel acceleration scheme: Application
to Poissons equation, IEEE Trans. Comput. Aided Design
Integrated Circ. Syst. 10(6):822824 (1991).
22. A. Kherib, A. B. Kouki, and R. Mittra, Higher order asymp-
totic absorbing boundary conditions for the nite element
modeling of two-dimensional transmission line structures,
IEEE Trans. Microwave Theory Tech. 38(10):14331438
(1990).
23. A. Kherib, A. B. Kouki, and R. Mittra, Asymptotic absorbing
boundary conditions for the nite element analysis of three-
dimensional transmission line discontinuities, IEEE Trans.
Microwave Theory Tech. 38(10):14271432 (1990).
24. R. K. Gordon and H. Fook, A nite difference approach that
employs an asymptotic boundary condition on a rectangular
outer boundary for modeling two-dimensional transmission
line structures. IEEE Trans. Microwave Theory Tech.
41(8):12801286 (1993).
25. B. Beker and G. Cokkinides, Computer-aided analysis of
coplanar transmission lines for monolithic integrated circuits
using the nite difference method, Int. J. MIMICAE 4(1):111
119 (1994).
26. T. L. Simpson, Open-boundary relaxation, Microwave Opt.
Technol. Lett. 5(12):627633 (1992).
27. D. Jatkar, Numerical Analysis of Second Order Effects in SAW
Filters, Ph.D. dissertation, Univ. South Carolina, Columbia,
SC, 1996, Chap. 3.
28. S. R. Hoole, Computer-Aided Analysis and Design of Electro-
magnetic Devices, Elsevier, New York, 1989.
29. V. K. Tripathi and R. J. Bucolo, A simple network analog
approach for the quasi-static characteristics of general, lossy,
anisotropic, layered structures, IEEE Trans. Microwave The-
ory Tech. 33(12):14581464 (1985).
30. R. E. Collin, Foundations for Microwave Engineering, 2nd ed.,
McGraw-Hill, New York, 1992, App. III.
31. B. Beker, G. Cokkinides, and A. Templeton, Analysis of
microwave capacitors and IC packages, IEEE Trans. Micro-
wave Theory Tech. 42(9):17591764 (1994).
32. D. Jatkar and B. Beker, FDM analysis of multilayer-multi-
conductor structures with applications to PCBs, IEEE Trans.
Comput. Pack. Manuf. Technol. 18(3):532536 (1995).
33. D. Jatkar and B. Beker, Effects of package parasitics on the
performance of SAW lters, IEEE Trans. Ultrason. Ferro-
electric Freq. Control 43(6):11871194 (1996).
34. G. Cokkinides, B. Beker, and A. Templeton, Direct computa-
tion of capacitance in integrated passive components contain-
ing oating conductors. IEEE Trans. Comput. Pack. Manuf.
Technol. 20(2):123128 (1997).
35. B. Beker, G. Cokkinides, and A. Agrawal, Electrical modeling
of CBGA packages, Proc. IEEE Electronics Computer Tech-
nology Conf., 1995, pp. 251254.
36. G. Cokkinides, B. Beker, and A. Templeton, Cross-talk ana-
lysis using the oating conductor model, Proc. ISHM-96, Int.
Microelectronics Society Symp., 1996, pp. 511516.
37. B. Beker and T. Hirsch, Numerical and experimental model-
ing of high speed cables and interconnects, Proc. IEEE
Electronics Computer Technology Conf., 1997, pp. 898904.
BURIED OBJECT DETECTION
AHMET SERDAR TU

RK
ALI KO

KSAL HOCAOGLU
Information Technologies
Research Institute
Gebze, Kocaeli, Turkey
1. INTRODUCTION
The subsurface detection and identication of buried geo-
logic and synthetic (human-produced) structures is cur-
rently an important research area and a progressive
technological concept on the global scale. The scope of
the problem is very complex, and aspects of the topic de-
pend on wide-area demands from the military to the com-
mercial necessities. Some examples are the localization of
the underground pipes, buried mines, and archeological or
anthropological artifacts. The variety of unknown false
and undesired targets under the ground complicates the
object identication task. In addition, the medium in-
volved is usually lossy and inhomogeneous. The size,
geometry, constitution (dielectric, metallic, explosive,
etc.), and depth of the target object and the characteris-
tics of the medium (wet soil, sand, etc.) are the principal
parameters for the detector designs. As yet, there is un-
fortunately still no single method or unique system that
can detect all kinds of objects at arbitrary size, structure,
depth, and soil. Therefore, a few numbers of the conve-
nient sensor technologies are considered regarding to the
specific situations. These sensors may have common, sim-
ilar, or different capabilities for the detection of the desired
targets.
This article is an introduction to buried-object detec-
tion and attempts to compile and present the available
methods to readers who are potentially involved in
physics, geophysics, civil engineering, archeology, or
electrical engineering, regardless of their educational
or academic background levels. On this scope, different
sensor technologies and radar systems mainly based on
the electromagnetic, acoustic, infrared, or chemical
methods are examined with the cons and pros such
as operating principles, strengths, limitations, and
feasibilities. Furthermore, some convenient multisensor
approaches for wide-range applications are discussed
to attain the best detection performance level in each
specific case.
540 BURIED OBJECT DETECTION
2. SENSOR TYPES
The main sensors for buried-object detection are based on
the electromagnetic, acoustic, seismic, and optical tech-
nologies. The electromagnetic sensors are divided into
two groups: electromagnetic induction (EMI) and ground
penetrating radar (GPR). The EMI or metal detector uses
low-frequency electromagnetic elds to induce the eddy
currents in metal components of buried objects. It is a
very mature, popular, and relatively inexpensive sensor
technology performed in a wide range of environments
for conducting target detection. The main limitation of
the EMI sensor is the possible metallic clutter of the
medium, especially for the low-conductivity or small-
metallic-content objects. The GPR is another widely used
and established electromagnetic sensor type. The operat-
ing principle is the reection and backscattering of the
radiofrequency (RF) waves from buried metallic or dielec-
tric objects. GPR detects all anomalies under the ground,
even if nonmetal. The most important problems are the
clutter due to the unwanted objects in soil and the buried
targets that have constitutive parameters similar to those
of soil. The acoustic and seismic sensors benet from the
sound or seismic wave reection from the object princi-
ples. Although they are not completely established sensor
technologies, such detector systems can have low false-
alarm rates. Nevertheless, deep buried objects, rough
surfaces, and vegetation/mineral-covered or frozen ground
present the main difculties in detection. Optical sensors
use the infrared and hyper spectral band to distinguish
the target from the temperature and light reectance dif-
ferences. If the heat absorbing capacity of the object is
much more different from that of the soil, the sensor per-
formance will be very satisfactory. The infrared detector
can quickly scan wide areas from high altitudes. However,
this is not a suitable sensor for deep-buried-object detection.
Beyond the main sensor technologies mentioned above,
a few additional detector systems can be examined either
for main sensor performance improvement or some spe-
cific applications such as explosive-materials detection.
For instance, subsurface microwave tomography uses op-
erating principles similar to those of GPR and achieves
the high-resolution imaging of shallow buried objects,
electrical impedance tomography determines the electri-
cal conductivity distribution of the surface, and the X-ray
sensor images the buried objects with X-ray radiation.
Moreover, the systems based on electrochemical, piezo-
electric, nucleonic, and biological methods have proved
useful just for the explosive vapor detection. Detailed dis-
cussions of all sensor types are presented in the following
sections.
3. GROUND-PENETRATING RADAR
3.1. Overview
The terms ground probing, subsurface detecting, and sur-
face penetrating refer to a wide range of electromagnetic
methods designed for the detection and identication of
buried artifacts or structures beneath the surface, gener-
ally termed ground penetrating radar (GPR). GPR uses
the electromagnetic wave propagation and scattering
principles to locate, quantitatively identify, and image
the variations in electrical and magnetic properties under
the ground. It may be performed from Earths surface,
from a land vehicle, or from an aircraft and has high-
resolution in subsurface imaging equal to that of any geo-
physical method, approaching the centimeter range under
the right conditions. Since GPR senses subsurface electri-
cal inhomogeneities, the detection performance of the sub-
surface features such as depth, orientation, water density,
size, and shape are related to the contrast in electrical and
magnetic properties. The detection performance can be
further improved by quantitative feature interpretation
through modeling. The GPR technology is largely
application-oriented. Implementation of its hardware
and software determines the overall system performance
characteristics. The detection range can vary from a few
centimeters to tens meters. Metallic or dielectric struc-
tures can be detected. Identication of small objects even
in the centimetric range is possible. Any metallic object
and dielectric discontinuity can be detected by the GPR
system in principle. The target may be a long, thin, cylin-
drical, or spherical object or a planar soil layer and can be
classied according to its physical geometry and electrical
constitution. All these performances strongly depend on
the soil properties and the radar parameters, which are
mainly operational frequency, transmitter power, system
dynamic range, and electromagnetic wave polarization
[1,4,7,13].
Two kinds of GPR systemsthe impulse GPR and
swept-frequency GPRare employed in most current
applications. The impulse GPR uses time-domain
electromagnetic pulse radiation and scattering principles.
Table 1. Radar Cross Section Formulas of Basic Target Geometries
Object Aspect Direction RCS (dBsm) Shape Parameters
Sphere Omnidirectional pR
2
Sphere radius R (m)
Cylinder Broadside
2pad
2
l
Cylinder radius a (m)
Cylinder length d (m)
Flat plate Normal
4pw
2
h
2
l
2
Plate width w (m)
Plate length h (m)
Corner reector, dihedral Symmetry axis
8pA
2
l
2
Reector plate area A (m
2
)
Wavelength l (m)
Corner reector, rectangular trihedral Symmetry axis
4pL
4
3l
2
Sidelength L (m)
BURIED OBJECT DETECTION 541
The shape of the impulse signal determines the
frequency bandwidth. The object detection is done by
means of the amplitude and time delay of the received
pulse. In the case of the swept-frequency GPR, the
operational band is clearly dened by the designer.
The electromagnetic wave scattering problem is solved
in the frequency domain. The target information is
obtained from the amplitude and the phase of the received
signal [1].
The GPR is not very different from conventional radar.
The main differences are the short-range concept, near-
eld analysis, and lossy inhomogeneous propagation me-
dium. The detection technique is usually based on the
backscattered radiation signal from the underground ob-
ject, but forward scattering can also be employed to yield
target information in specific cases. Since this is a short-
range radar, the clutter is the main problem for target
detection and identication. The sources of the clutter are
the antenna ringing, transmitter/receiver (T/R) coupling,
the ground surface reection, and unwanted (spurious)
backscattering from other buried structures or soil inho-
mogeneities. The preventive measures such as proper
antenna matching, shielding enclosure design, and adap-
tive methods for processing the signal should be taken to
improve radar performance.
Selection of the center frequency, the bandwidth, and
the power of the GPR system are the key factors in design.
The electromagnetic wave reection from the ground
surface and the attenuation under the ground dramati-
cally increase with the increasing frequency and the
soil conductivity as given in Table 2. As a rule, material
that has a high conductivity will have a large attenuation.
Thus GPR object probing in sand, gravel, dry soil, and
freshwater is much easier than in clay, wet, mineral
soils, and salty water. Since the ground attenuation
and reection parameters indicate power loss and clutter,
lower frequencies are used to detect deeply buried
objects. Nevertheless, higher frequencies are needed for
better resolution and detailed echo to determine small
objects [4]. Thus, the ultrawideband (UWB) GPR is
mostly preferred in order to benet from both low and
high frequencies. UWB antenna design is a significant
area in terms of GPR performance. Transmitter and
receiver antennas are designed to comply with the
existing hardware to radiate and couple the signal into
the ground efciently [3].
The object-oriented design procedure and critical pa-
rameters for GPR systems are discussed in Section 3.2.
More specifically, suitable broadband antennas, appropri-
ate continuous-wave (CW) or impulse signal generator de-
sign, low-noise RF receiver units, and signal processing
techniques are discussed.
3.2. System Design Procedure
Ground penetrating radar has a enormously wide appli-
cation area from underground structure exploration to
small-object detection just beneath the surface. The oper-
ational principle consists in generation of impulse or ra-
diofrequency (RF) power by the signal source, radiation of
the electromagnetic wave by the transmitter antenna,
characterization of the soil and airEarth interface, re-
ception of the backscattered electromagnetic wave from
the target object by the receiver antenna, and nally,
sensing, digitizing, and processing the received RF signal
by the receiver (see Fig. 1). The critical parameters such
as soil type and target properties strongly affect the de-
tection performance. The main goal of the design is to
provide application-oriented information and to illustrate
the technical options available for the operator or design-
er. To operate the GPR successfully, the following require-
ments must be satised:
1. Sufciently high power RF or impulse signal gener-
ator to attain the required penetration depth
2. Suitable antenna to couple the electromagnetic
wave into the ground efciently with sufcient
bandwidth, high-gain, narrowbeam, and low-input
reection
3. Proper frequency bandwidth to obtain a sufciently
large scattered signal from the smallest buried
object
4. Appropriate receiver hardware to achieve an
adequate signal-to-noise (S/N) and signal to clutter
ratios for the best target detection performance
5. Adaptive signal processing techniques to distinguish
and classify the buried objects
Table 2. Propagation Losses at 100MHz and 1GHz in Different Soils
Material Relative Dielectric Constant Conductivity (mS/m) Loss at 100MHz (dB/m) Loss at 1GHz (dB/m)
Air 1 0 0 0
Asphalt, dry 24 110 215 20150
Asphalt, wet 612 10100 520 50200
Clay 240 11000 5300 503000
Concrete 420 1100 125 10250
Freshwater 81 0.1 0.1 1
Rock 410 0.11 0.011 0.110
Sand, dry 46 0.00011 0.012 0.120
Seawater 81 4 1000 10000
Snow 812 0.0010.01 0.12 120
Sandy soil 430 0.110 0.15 150
Loamy soil 420 0.1100 0.560 5600
Clayey soil 415 0.11000 0.3100 31000
542 BURIED OBJECT DETECTION
The electrical properties of the medium (soil) are impor-
tant parameters in GPR design. The soil in concept
is quasihomogeneous, and its electrical behavior may be
dened by its constitutional parameters, namely, dielec-
tric permittivity (e), magnetic permeability (m), and
electrical conductivity (s). Here, the relative dielectric
and the magnetic constants (e
r
, m
r
) designate the effective
wavelength, and the soil conductivity determines the
propagation loss. The variability of both material param-
eters and local geologic conditions cause great difculty
in accurate prediction of propagation behavior. All of these
parameters are generally the functions of frequency as
well. For instance, RF propagation and ground surface
reectivity losses sharply increases after 3GHz and in
many cases, this fact physically limits the operational
bandwidth. Nevertheless, when the buried objects
decrease in size, the simulation results show that the
radar cross section (RCS) values are insufcient before
the resonance frequency region (see Fig. 3). As is well
known, the RCS of a target object represents the objects
electromagnetic wave backscattering level. Since the
RCS depends on the electrical sizes (measurements with
respect to wavelength) of the object and the effective
wavelength is inversely proportional to the square root
of the soil dielectric constant, the GPR frequency can be
shifted downward in high dielectric mediums (e.g., in wet
soils). For example, the operating frequency must be at
least 2 GHz to obtain sufciently powerful backscattered
signal from a 5-cm-diameter cylindrical target buried in
dry soil. However, if this object were buried in wet soil, the
frequency bandwidth about 1 GHz would be enough for
the desired backscattering level. The range resolution in
depth (DR
z
) is another important performance parameter
for GPR design. The thickness of the depth-range layers
on the GPR screen is directly related to the range resolu-
tion (see Fig. 14). For impulse radars, the bandwidth (BW)
expressed in (1) is inverse proportional to the downrange
resolution
BW
c
2
.
DR
z
.

e
r
p 1
where c is the free-space speed of light and e
r
is the
relative permittivity of the soil. If it is presumed that
e
r
9 and DR
z
2 cm, the required bandwidth will be
2.5 GHz.
The radar equation given in (2) determines the detec-
tion range for a given target RCS in different parametric
Timing &
synchronize
Controller
Display
Low noise
amplifier
Receiver
antenna
Object
Ground
Transmitter
antenna
Analog to
digital
converter
Sample-hold
or
demodulator
Signal
proccesing
Signal
generator
& power
amplifier
Microprocessor
Figure 1. Block diagram of GPR operating
system.
BURIED OBJECT DETECTION 543
cases, such as power, frequency, soil type, and antenna
structure (see conguration in Fig. 2)
P
r

P
t
G
t
4pR
2

2
A
e
RCS
L
2
LdB L
a
dB L
r1
dB L
s1
dB L
s2
dB L
r2
dB
3
where
P
r
received power (W)
P
t
transmitter power (W)
G
t
transmitter antenna gain
R target detection range (m)
A
e
receiver antenna effective area (m
2
)
RCSradar cross-section of buried object (m
2
)
L total propagation loss (except free space loss)
L
a
transmitter and receiver antenna efciency loss
L
r1
surface reectivity loss (air to ground)
L
r2
surface reectivity loss (ground to air)
L
s1,2
soil attenuation losses
The transmitter antenna gain (G
t
) depends on the an-
tenna type, geometry, and frequency. The gain at given
frequency band is taken for stepped frequency GPR, while
the peak gain (or the average gain) must be considered for
an impulse GPR. Typical gains can be from minus a few
decibels over isotropic to 410 dBi. The receiver effective
area is expressed in terms of the antenna gain (4). The
antenna efciencies are generally related to the radiation
losses due to the resistive loadings and impedance mis-
matching at the structures. Such losses are typically ex-
pressed in a few decibels:
A
e

G
r
l
2
4p
4
The RCS of the target depends on the object geometry and
structure. However, the model formulas for canonical
geometries given at Table 1 and the numerical simulation
plots for 5-cm-diameter cylindrical dielectric target shown
in Fig. 3 are presented to express an authors opinion [18].
The reectivity losses between airground and ground
air layers are subject to the problem of electromagnetic
wave scattering from a dielectric surface. Too many
parameters, such as antennaground surface and tar
getground surface distances, dielectric constant of soil,
antenna and target geometries, antenna radiation and
target scattering waveshapes at surface, and tilt angles of
antenna and object, are included in this formulation. In
this case, some useful approximations are usually pro-
posed for easy-to-calculate coherent estimations [1]. For
example, the airground reection loss is given by (5) pre-
suming that the radiated wave on the surface is a uni-
formly distributed plane wave. Another generic method
based on the transmission-line approach in Eq. (6) that
uses the characteristic impedances of the air (Z
air
) and the
soil (Z
soil
) is expressed as
L
r1
20 log
10

e
0
p
cos y
1

e
r
p
cos y
2

e
0
p
cos y
1

e
r
p
cos y
2

dB; see Fig: 2
5
L
r1
20 log
10
2Z
soil
Z
soil
Z
air

dB 6
where
Z
air
377O and
Z
soil

m
0
m
r
e
0
e
r
r cos
d
2
j sin
d
2

1 tan
2
d
1=4
7
Typical values for many soil materials, Z
soil
is about 100O,
hence L
r1
E810dB and L
r2
% 5dB. Finally, soil attenua-
tion losses are dened as
L
s1;2
8:686
.
2pf R
0

m
0
m
r
e
0
e
r
2

1 tan
2
d
q
1
& '
s
8
Transmitter
antenna
Object
RCS
Ground
Receiver
antenna
Air
P
t
L
r1
L
s1
L
s2
G
r
, A
e
L
r 2
0
2
0
1
0
1
c
0
,

j
0
c
r
,

j
r
,
G
t
P
r
o
Figure 2. Propagation modeling of GPR
system.
544 BURIED OBJECT DETECTION
where d is the soil propagation path loss in nepers per
meter (1 N8.686 dB) in soil, f is the operating frequency,
and R
0
is the pathlength in the soil. Table 2 exhibits the RF
propagation losses in some soils at 100MHz and 1 GHz.
As a general rule, the system noise threshold deter-
mines the minimum detectible signal (MDS) level. Detec-
tion of the target is possible if the received power
backscattered from the object is sufciently higher than
both noise and clutter. The MDS can be calculated by (9)
using the system frequency bandwidth (BW), the receiver
noise gure (NF), and the required signal-to-noise ratio
(SNR)
MDSk
.
T
0
.
BW
.
NF
.
SNR 9
where k is Boltzmanns constant and T
0
is room temper-
ature in kelvins (kT
0
D4 10
21
W/Hz). To obtain a re-
ceived power, Pr, that exceeds this value, the MDS willnot
sufce for detection when clutter dominates the signal.
The sources of the clutter are mostly the undesired cou-
pling signals from transmitter to receiver, the unwanted
backscattering from other buried objects, and the reec-
tion from the ground surface. For this reason, minimum
SNR requirement of the system should be accurately es-
timated on the design to guarantee the physical radar de-
tection range.
Design Example. Suppose that we want to detect a di-
electric cylindrical object of 10 cm diameter that is buried
up to R100cm. The system operating frequency is cho-
sen 1 GHz. The soil is dry and loamy (e
r
4). The trans-
mitter antenna has 3dBi gain, and the receiver antenna
effective area is A
e
0.0012 m
2
. Using (2)(9), Table 2, and
Fig. 3, we nd that the two-way reection losses are to-
tally taken as 15 dB, the soil attenuation loss is 10 dB, the
target has an RCS of 10 dBm
2
, and the overall power
loss (dynamic range) is calculated as B94 dB. If we
assume that the amplier of the GPR receiver has a noise
gure (NF) of 3dB over 1 GHz bandwidth (BW), the MDS
for the specied receiver characteristics yields 8
10
10
W or 91 dBw, considering a signal-to-noise ratio
(SNR) of 20. Under the circumstances, the peak power of
the transmitter must be at least 3dBw or approximately
2 W. For a 50-O-impulse GPR system, this corresponds to
10 V peak voltage of the transmitter output.
3.3. Signal Sources and Modulation Methods
Many kinds of the modulation techniques can be employed
for ground penetrating radar systems. The most frequent-
ly used GPR systems are the impulse radar that generates
short pulses or impulses in the category of amplitude
modulation (AM) and the swept or stepped-frequency ra-
dar that uses frequency modulation (FM) followed by a
synthesized pulse.
Impulse GPR is commercially the most popular system
preferred for a wide range of applications. The operational
principle is based on the impulse signal generation at re-
quired peak power and frequency bandwidth. The signal
sources generate monocycle or monopulse waveforms that
can last for a few hundred picoseconds to a few nanosec-
onds. The impulse waveforms are generally Gaussian in
shape and their frequency bandwidths start from a few
hundred MHz up to a few GHz depending on the applica-
tion (Fig. 6). The main difference between the monocycle
and monopulse is in the frequency-domain behavior. The
monopulse signal highly contains DC or low-frequency
components. Nevertheless, the monopulse waveform has
some physical advantages for time-domain radar signal
transmission and radiation.
Frequency-modulated continuous-wave (FMCW) and
stepped-frequency GPR systems are used when the tar-
gets of interest are shallow and an operational frequency
of Z1GHz has to be maintained. These systems are easier
to design compared to the wideband impulse radars. Such
swept- and stepped-frequency radars have the advantage
of greater selection and control of the operational frequen-
cy bandwidth. This feature provides better radiation ef-
ciency with high-power generators and high-gain
antennas, lower noise level at narrower bands, and con-
15
10
5
0
10
15
20
25
30
35
5
10 cm diamater dielectric cylindrical object
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Frequency (GHz)
R
C
S

(
d
B
s
m
)
R
C
S

(
d
B
s
m
)
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
30
25
20
15
10
5
0
5
10
15
12 cm diameter metallic cylindrical object
Frequency (GHz)
Figure 3. RCS behaviour of different cylindrical objects in air.
BURIED OBJECT DETECTION 545
sequently wider dynamic range in comparison to the im-
pulse radar.
3.3.1. Impulse Generator Design. There are two major
concerns in an impulse generator used in GPR applica-
tions: (1) the pulseshape and plusewidth and (2) pulse
amplitude. In many cases, both of these parameters are
highly critical. The pulseshape and plusewidth imply
broadband characteristics, conducive to the detection of
buried small objects. On the other hand, pulse amplitude
implies system power related to the operational detection
depth of the target in soil.
The shape of the pulse signal rigorously affects the
operational frequency bandwidth of the GPR system. For
example, a steplike pulse has a large amount of spectral
energy at lower frequencies with the spectrum falling off
as inverse frequency (1/f ), while an impulselike pulse has
a at spectrum over the bandwidth. In every case, the
high-frequency content of the pulse signal is limited by
either the step risetime (T
s
) or the impulse duration (T
d
).
For ideal impulse waveforms, the rst null (f
0
) in the spec-
trum occurs at 1/T
d
. The spectral power distribution of the
impulse decreases rapidly beyond f
0
and typically follows
an equation of the sinc function form. A monocycle signal
can be analyzed as a pulse AM modulation of a RF sine-
wave carrier. Thus, the spectrum of the monocycle impulse
is shifted upward in the frequency domain with respect to
the risetime and centered on the carrier frequency that
corresponds to 1/T
d
. Generation of these various wave-
forms normally begins with a very fast-risetime-edge,
steplike pulse. Then one takes the rst derivative of the
step risetime to obtain an impulse signal. To obtain a
monocycle, the rst derivative of an impulse or the second
derivative of a step pulse is taken. In practice, this process
can be implemented by adding a proper short-circuit-ter-
minated transmission line to the impulse generator out-
put. The UWB antenna and the wideband amplier must
be appropriately designed to avoid bandpass ltering,
which can cause multicycling [4].
Various high-power switching transistors or FETs can
be used to acquire the typical GPR pulses of a few hundred
picoseconds to a few nanoseconds in duration. Avalanche
transistors are often used to generate peak power levels
from 50 W to several kilowatts. For the 310-GHz UWB
band, selection of available devices is much smaller. Step
recovery diodes (SRDs) can generate 50200ps edges with
amplitudes of several volts [2].
The following is given as a monocycle impulse circuit
design procedure especially for high-frequency and high-
amplitude GPR applications:
*
A fast risetime and high-amplitude square-wave gen-
erator
*
Impulse generator with step recovery diodes
*
Wideband impedance matching with resistors and
high-speed Schottky diodes
*
Impulse bandwidth improvement by short-circuit
shunt stubs
*
Wideband MMIC amplier (if the signal power is
insufcient)
The monocycle pulse generator shown in Fig. 4 is
developed using transmission-line stubs realized by
rigid cables, SRDs, and a Schottky diode, a MESFET
and a monolithic microwave integrated circuit (MMIC)
amplier. At the input, a square-wave signal with a rep-
etition frequency of 400kHz is applied by a microproces-
sor. The risetime of the signal is shortened with a
comparator circuit. The signal proceeds to the impulse
generator, composed of SRDs in active mode [8]. Operating
in active mode enable the designer to shape the impulse
by adjusting the DC bias current drawing from the
DC source. The MESFET is used particularly for the
circuit isolation and wideband impedance matching.
Finally, the rst stub, which is at the gate of the ME-
SFET, is used to generate a very short impulse. It is noted
that, due to the MESFET buffer, the antenna cannot
load back the circuit in an undesirable way. Figures 5
and 6 show the measured output monocycle pulse in
time domain and in frequency domain, respectively. The
measured waveform has less than 1.5ns pulse duration
in time domain, and its frequency domain bandwidth is
approximately 2GHz.
The design procedure of the avalanche-mode mono-
cycle impulse generator is similar to that of the SRD
circuit. The main difference is that an avalanche tran-
sistor is used instead of the short recovery diodes.
The transistor is operated in avalanche breakdown
mode used with switching by discharging the stored
energy into the short transmission lines. These are
output loads, which also include the transmitter antenna
input impedance. The time-domain impulse signal
output and its frequency-domain transformation of an
avalanche-mode impulse generator are presented in
Figs. 7 and 8.
3.3.2. FMCWand Stepped-Frequency Radar. The FMCW
radar system transmits a sweeping RF carrier signal con-
trolled by a voltage-controlled oscillator (VCO) over a se-
lected frequency bandwidth. The operational frequency
band can be selected by considering mainly the maximum
depth and the minimum size of the buried object. The
backscattered signal from the target object is mixed with a
sample of the transmitted waveform and the results in a
difference frequency, called the intermediate frequency
(IF), is derived from an I/Q (in-phase/quadrature) mixer
pair. The IFoutput contains the amplitude and phase data
of the received signal. The phase information yields the
range of the target since it is correlated with time delay.
Only if there is a target on the scope will the IF signal be
produced. If the variation in transmitter frequency is a
linear function of time, then a target return will occur at a
time T
r
given by
T
r

2R
u
10
where R is the range in meters and u is the velocity of
electromagnetic wave on the propagation path (air,
ground, or both) in meters per second.
Since the FMCW radar essentially measures the phase
of the IF signal, which is directly related to target range, it
546 BURIED OBJECT DETECTION
requires a high degree of linearity of frequency sweep with
time to avoid spectral widening of the IF, which indicates
the system range resolution degradation. In practice,
the sweeping nonlinearity levels should keep below 0.1%
[1]. The amplier and the antenna gain responses over
the frequency band must also be linear for an appropri-
ate system design. A time-domain equivalent to the
impulse radar can be acquired by the inverse Fourier
transformation.
Stepped-frequency radar is a continuous wave system
that is sequentially repetitive at different frequencies as
shown in Fig. 9. This kind of radar radiates a sequence of
N number of carrier frequency steps and the amplitude
and the phase of the received signal is stored. Downcon-
version to IF signal is similar in principle to that
for FMCW radar. The complex Fourier transformation is
applied to obtain the time-domain response of the reected
or backscattered target signal. The range resolution is
dened as
DR
u
2NDf
11
where DR is the range resolution and Df is the frequency
increment of every step.
For both FMCW and stepped-frequency radar, it is ev-
ident that the thermal noise level is much lower than that
of the receiver of the time-domain impulse radar. Never-
theless, the discrimination of the target backscattering
signal from the ambiguous clutter at the receiver unit is a
more difcult problem due to the futility of simple anti-
aliasing lters. Therefore, some calibrating methods such
as iterative range gating are usually employed for such
systems.
3.4. GPR Antennas
The performance of GPR is highly dependent on the prop-
er design of the transmitter and receiver (T/R) antenna
pair. In practice, the transmitted GPR signal has
broadband characteristics, which may vary from 10 MHz
to 5 GHz for impulse system and from 1 to 8GHz for
stepped/swept-frequency systems. Therefore, wideband T/
R antennas must be designed to reach the largest dynamic
range, best focused illumination area, lowest T/R antenna
coupling, reduced clutter, and uniformly shaped impulse
C1
L1
Input
C2
C3
Antenna
Amplifier
D1
SRD
SRD
SRD
Stub
Stub
Schottky diode
Mesfet
D2 D3 D4
R1
Figure 4. Schematic of the monocycle pulse generator.
: 1.44ns
@: 560ps
C2 PkPk
3.64 V
500mV
M 1.00ns Ch2 440mV
2
Ch2
T
Figure 5. Measured output monocycle pulse in time domain.
Log
10
dB/
# ATN
20 dB
WA SB
SC FC
Corr
Start 100 MHz
If BW 1.0 MHz AVG BW 3 MHz SWP 56.0 msec
STOP 2.900 GHz
Ref 30.0 dBm
Figure 6. Frequency-domain transformation of the output mono-
cycle pulse.
BURIED OBJECT DETECTION 547
radiation [3]. These antennas should satisfy the following
electrical characteristics over the operational frequency
band such as high directivity gain, low input reection
level, suppressed side/backlobes, convenient polarization
with respect to the target shape, high T/R shielding, and
linear phase response (for impulse GPR). Moreover, some
physical restrictions, such as lightweight requirements for
movable handheld GPR, can also rigorously restrict the
convenient antenna types. A useful reference to the gen-
eral case of antennas in matter is given by King and Smith
[9]. Further considerations in the selection of the suitable
antenna types are subject to the types of targets and
radar systems. On this scope, most popular ultrawide-
band (UWB) antenna models, such as dipole, bowtie, spi-
ral, and TEM horn, are proposed for the impulse GPR
system. For FMCW and stepped-frequency GPR, wide-
band-selective octave horns, dielectric lled open-ended
waveguides, cavity-backed planar spirals, and pencil-
beam dish antennas (only in certain specific cases) may
be preferred because of their higher gain characteristics
[1,6,11,12,18].
The simplest GPR antenna is an electrically small di-
pole element. It is a linearly polarized antenna that has
far-eld intensity proportional to its electrical length.
Thus, the antenna gain and the radiation efciency
can be very small at lower frequencies. In order to avoid
performance degradation, resonance dipoles such as half-
wave are usually preferred for impulse GPR. Although the
dipoles have a physical advantage especially for handheld
GPR systems because of their small size, lightweight and
easy-to-fabricate structures, and their frequency-depen-
dent input impedance and radiation gain characteristics.
Some dielectric and absorber loading techniques are
usually employed to improve antenna frequency band-
width [10]. The bowtie conguration is similar to the
dipole wideband antenna frequently used for pulse radia-
tion (Fig. 10c). The bowtie exhibit shows higher gain and
wider band performance than does the simple dipole. For
a circular object, the bowtie is a suitable antenna type
since it produces linear polarization and has relatively
good linear phase response over the wideband [3,10]. If the
target is noncircular, with respect to electromagnetic wave
scattering principles, circularly polarized planar antennas
such as spirals may be preferred. The spiral is based on
which Archimedian and logarithmic models, is theoreti-
cally a frequency-independent antenna (see Fig. 10). But
in practice, both are bandlimited with respect to the feed
radius for the upper frequency and the armlength for the
lower frequency [3]. The frequency-domain responses of
bowtie and spiral antennas are presented in Fig. 11.
When a wider band, lower-voltage standing-wave ratio
(VSWR), higher gain, and narrower beamwidth are
required, the TEM horn is one of the most promising
antenna types for impulse GPR systems. Such a structure
consists of a pair of triangular or circular slice-shaped
conductors forming a V-dipole structure. It is character-
ized by L, d, a, and y parameters, which correspond to
the length of the antenna, the feedpoint gap, the conductor
plate angle, and the elevation angle, respectively
Ch3
3
T
10.0 V
1.4 V
C3 PkPk
48.4 V
: 1.38ns
@: 940ps
M 1.00ns Ch3
Figure 7. Time-domain signal output of the avalanche-mode im-
pulse generator.
Ref 43.0 dBm
Log
5
dB/
ATN
10 dB
WA SB
SC FC
Corr
Start 100 MHz
If BW 1.0 MHz AVG BW 3 MHz SWP 56.4 msec
Stop 2.921 GHz
Figure 8. Frequency-domain behavior of the avalanche-mode
impulse generator signal.

Controller
Sweeping
synchronizer
Microprocessor
Correlator &
discriminator
Display
Data
processing
I/Q
demodulator
RF signal
generator
Mixer
Transmitter
antenna
Low noise
amplifier
Power
amplifier
Recevier
antenna
Figure 9. Block diagram of swept/stepped-frequency radar.
548 BURIED OBJECT DETECTION
(Fig. 10d). A TEM horn generally shows bandpass lter-
like antenna gain behavior over a large bandwidth. In
practice, its electrical dimensions strongly limit the lower
operation frequency. Thus, dielectric-lling techniques are
applied mostly to decrease the lower cutoff frequency due
to certain physical restrictions in GPR applications [11].
In this case, the operational band of the antenna can be
shifted downward in the frequency domain but the fre-
quency bandwidth probably cannot be broadened. The
partial-dielectric-loaded TEM (PDTEM) horn structure
is proposed by Turk to achieve low VSWR (voltage stand-
ing-wave ratio) and high directivity gain with stable per-
formance over the wideband [6]. The PDTEM structure is
modeled as a microstrip line, and its geometry is designed
to match the antenna output impedance to the feed source
impedance (nominal, 50 O) by decreasing the segment
characteristic impedances along the antenna line. For
this purpose, using multiangular dielectric prole loading
and arranging the segment widths of the plate wings
properly, one can implement an efcient geometric design
for the TEM horn (see Fig. 10f ). This approach is quite
effective for higher frequencies where the electrical length
A
r
m

w
i
d
t
h

(
c
m
)
Arm length (cm)
(a)
(c)
(e)
(b)
(d)
(f)
L L
L
Plate
Resistive
sheet
Feed
gap
Absorber
coating
c
r
c
o
r2
r1
fi
r
o
r=a.fi
1 2 3 4 5 6 7 8 9
0
1
2
3
1
2
3
r=e
a

r2

: :
:
0

o
Figure 10. Wideband GPR antenna types: (a) Two-armed Archimedian spiral; (b) two-armed log-
arithmic spiral; (c) triangular plate bowtie; (d) triangular plate TEM horn; (e) grating model TEM
horn arm (top view); (f) PDTEM horn geometry (side view).
BURIED OBJECT DETECTION 549
of the antenna line is sufciently long to transform output
impedance. Nevertheless, the antenna output impedance
strongly determines the matching performance rather
than the characteristic impedance. An extensional resis-
tive aperture loading is used to keep VSWR levels reason-
able at lower frequencies. Finally, the feed gap is coated by
a piece of low-conductivity absorber to improve the overall
VSWR performance. The grating model arm structure is
the metal-reduced version of the PDTEM horn antenna,
which is recommended for multisensor (GPR combined
with metal detector) systems (see Fig. 10e). Typical input
reection and antenna gain behaviors of TEM, PDTEM,
and rigid horns are shown in Figs. 12 and 13.
Impedance matching is a critical step to improve the
antenna gain and to reduce time-domain ringing effects.
Since most of the GPR impulse signals are generated by
avalanche-mode transistors that produce impedance-de-
pendent pulseshapes, the variability of the antenna im-
pedance over the wide operational band must be stabilized
by a matching circuit design. In addition, the antenna
structure should be balanced to avoid structural imbal-
ances using a wideband balun (balanced-to-unbalanced)
transformer that has a sufciently short risetime [3].
Standard-gain-horn, rigid-horn, dielectric-loaded wave-
guide, and dish antennas can also be considered for
FMCW and stepped-frequency GPR systems. All of these
models are high-gain and narrowbeam band-selective an-
tennas that are used from L band to X band (from B1 to
B10 GHz). The typical gain of the standard horn antenna
is B15 dBi over a 21 frequency bandwidth. To increase
S2:b2
REF 0.0 dB
10.0 dB/
18.355 dB
log MAG
log MAG
5 1
Scale
10.0 dB/div
1
2
3
4
5
Start 0.010000000 GHz
Stop 1.500000000 GHz
log MAG
5
Source 1 Power
20.0 dBm
Start 0.010000000 GHz
Stop 1.500000000 GHz
3
2
1
4
M
M
M
5
1>
1>
hp
hp
hp
hp
log MAG
Scale
10.0 dB/div
3
2
4
1
1>
Start 0.010000000 GHz
Stop 1.500000000 GHz
5
Start 0.010000000 GHz
Stop 1.500000000 GHz
4
3
2 1
1>
S2:b2
REF 0.0 dB
10.0 dB/
31.266 dB
S2:b2
REF 0.0 dB
10.0 dB/
27.315 dB
S2:b2
REF 0.0 dB
10.0 dB/
60.262 dB
(a) (b)
(c) (d)
Figure 11. Measured frequency-domain S
21
responses of the planar antennas for 60cm dry soil:
(a) logarithmic spiral (r
1
0.5 cm, r
2h
5.5cm, r
2v
7.3 cm); (b) Archimedian spiral (r
1
0.5cm, r
2
8cm); (c) bowtie (a 601, d0.5cm, L7.3 cm)cross-polarization; (d) bowtiecopolarization.
550 BURIED OBJECT DETECTION
the antenna directivity gain with a narrower beamwidth,
a model using proper dielectric loadings is available [12].
For multiband GPR operations, the rigid horn that yields
B10 dBi gain over 101 frequency bandwidth can be cho-
sen (Fig. 13). When the highest gain and the narrowest
beamwidth are highly necessary and the physical size of
the antenna is not restricted (i.e., as in vehicle-mounted
forward-looking systems), the dish antenna is an appro-
priate solution because of its high gain up to 30 dBi and
pencil-beam characteristics. The electrical size of the dish
aperture is directly related with the directivity gain and
HPBW as follows [18]
G20log
D
l

9:94 dBi 12
HPBW
180l
pD
degrees 13
where G is the antenna gain for 100% efciency and D is
the diameter of the circular dish.
3.5. Signal Processing Techniques
The development of advanced processing algorithms
for GPR data is very important to achieve high perfor-
mance for the detection and identication of objects.
The signal processing methods are based on interpreting
the backscattered signal parameters; these are the
amplitude and the phase of the received signal for contin-
uous or stepped-frequency GPR, and the time-domain
shape of the received signal for impulse GPR. In both
cases, the amplitude represents how much power is re-
turned back from the target. The impulsive oscillations or
the change in phase indicates the depth range of the
target. For CW GPR, the phase difference between the
transmitted and received signals indicates the depth of
the target. The phase of the electromagnetic wave changes
when propagating over the path proportional to its effec-
tive wavelength. Thus, every wavelength distance corre-
sponds to 3601 phase difference, but it is not easy to nd
the target range if the medium is inhomogeneous
and cluttered. At this stage, advanced signal processing
algorithms are needed to estimate the parameters adap-
tively from the raw data collected from inhomogeneous
regions [1].
In the case of impulse radar, the received pulse signals
are collected in the scanning direction. The A-scan and
B-scan raw data series are obtained as shown in Fig. 14.
A-scan data represent the received impulse signals for any
scanning step. The B-scan plot is simply a two-dimension-
al (2D) representation of consecutive A scans in the scan-
ning direction. Here, the x and y axes correspond to the
scanning direction and the propagation time, respectively.
Time also means the range, since the electromagnetic
SWR
hp
S
11 CH 1
PRm
Cor
1
Start 100 000 000 GHz Stop 4.000 000 000 GHz
2.000 000 000 GHz
1 / Ref 1 1:1.0842
Figure 12. VSWR plot of PDTEM horn (L10 cm, a 301,
y 801).
15
10
5
0
5
10
15
20
25
1 2 3 4 5 6 7
Frequency (GHz)
A
n
t
e
n
n
a


G
a
i
n


(
d
B
i
)
Ridged Horn
TEM30
PDTEM30
Figure 13. UWB antenna gain performances
of GPR antennas.
BURIED OBJECT DETECTION 551
wave has a velocity of
u
3 10
8

e
r
m
r
p m=s 14
where e
r
and m
r
respectively are the relative constitutional
dielectric and magnetic coefcients of the propagation me-
dium. Thus, if any anomaly occurs under the ground, the
signal amplitude changes at the timescale to express the
related depth range. When the soil structure is not well
known, inhomogeneous and there are many unwanted
buried objects, detection and discrimination of the target
will be a difcult problem. Most of the GPR processing al-
gorithms are based on background estimation, differenti-
ated power density analysis, and subband processing of
GPR data [1,5]. The main goal is to nd the target em-
phasizing the anomalies under the ground. For back-
ground estimation and removal, an average A-scan
dataset is calculated from multiple A-scan data that do
not contain any signal from a target, and then this aver-
age is taken as a reference. The other A-scan data are cal-
culated relatively. The magnitude of the GPR data after
the background removal process is computed (see Fig. 15).
Further processing can be done in several frequency
bands separately. The advantage of subband processing
is that it allows one to take into account the different tar-
get responses and clutter behaviors in different frequency
regions. For example, the average RCS values of the prob-
able target objects can be calculated for every subband,
and some correlation algorithms can be estimated to dis-
criminate and classify the detected objects.
4. ELECTROMAGNETIC INDUCTION (EMI) SENSOR
4.1. Overview
Electromagnetic induction (EMI), which is the basis for
the metal detectors, is one of several sensor modalities
widely deployed for subsurface buried-object detection and
identication. The basic technology is very mature. The
EMI sensors have been used for landmine detection in
World War I, and EMI metal detectors were further de-
veloped during World War II. Afterward, usage of EMI
sensors was also directed to detect conducting objects in
other application areas such as mineral exploration, non-
destructive body testing, archeological investigation, and
security enforcement.
A-Scan data
A
m
p
l
i
t
u
d
e

(
m
V
)
D
e
p
t
h

r
a
n
g
e
2
0
50
100
150
200
250
1.5
1
0.5
0
150 200 250 100 50
Time scale (depth range)
Scanning direction
0
0 10 20 30 40 50 60 70 80 90 100
0.5
1
1.5
2
10
4
Figure 14. Measured A-scan and B-scan impulse GPR raw data.
50
100
150
200
250
10 20 30 40
Scanning direction
Pre-processed GPR B-scan data
50 60 70 80 90 100
D
e
p
t
h

r
a
n
g
e
x
8
7
6
D
i
f
f
e
r
e
n
t
i
a
l

p
o
w
e
r

d
e
n
s
i
t
y
5
4
3
2
1
0
10 20 30 40
Scanning direction
50 60 70 80 90 100 0
10
5
Figure 15. Processed A-scan and B-scan impulse GPR data.
552 BURIED OBJECT DETECTION
The operating principle of the EMI sensor is to trans-
mit a time-varying primary electromagnetic eld that in-
duces currents in any metallic object that the primary
eld penetrates. When the primary eld is abruptly
turned off, eddy currents in the metallic object produce a
secondary electromagnetic eld that is measured by the
receiver. The employed frequency range is generally lim-
ited to a few tens of kHz [13]. EMI sensors usually consist
of a pair of transmitter and receiver coils (see Fig. 16). An
electrical current owing in the transmitter coil of the
wire produces the primary eld, and the electromagnetic
waveform is often either a broadband pulse or a continu-
ous wideband. The transmitted eld induces a secondary
current in the earth (ground) as well as in any buried
conducting objects. This secondary current is usually de-
tected by sensing the voltage induced in the same or an-
other coil of wire, called the receiver coil. In the case of
pulsed excitation, the transmit waveform is quenched
quickly and the receiving coil measures the decaying sec-
ondary eld that has been induced in the earth (ground)
and subsurface objects. In the case of wideband excitation,
the receiving coil is placed within the magnetic cavity so
that it senses only the weak secondary eld radiated by
the earth and buried objects [14,16,17].
4.2. Operational Theory
The detection performance of the EMI sensor depends
mainly on the system parameters, namely, the primary
magnetic eld level created by the transmitter coil, the
target range, the eddy-current inducement capability of
the target object, and the receiver coil probe characteris-
tics. Assuming that the elds are time-dependent and the
coil position is stationary @ S
*
=@t 0, the electrical cur-
rent/magnetic eld and the magnetic eld/electrical volt-
age transformation ratios of the transmitter and receiver
coils are calculated by the related Maxwell equations giv-
en in (15) and (16):
rE m
@H
@t
)
I
c
Edl m
Z
S
@H
@t
dS
)V
AB
i2pf mNHS cos j
15
rHJe
@E
@t
)
I
c
Hdl
Z
S
JdSe
Z
S
@E
@t
dS
)
I
c
Hdl I
16
where E is the electric eld, H is the magnetic eld, f is the
frequency, N is the number of turns, S is the coil surface
area of the coil, J is the current density, I is the coil cur-
rent, j is the angle between the magnetic eld and the coil
surface normal vectors, e; m is the dielectric permittivity
and magnetic permeability of the coil material and V
AB
is
the induced voltage (see Fig. 17).
Equation (15) denotes that the coil probe sensitivity
increases with the operation frequency, the number of coil
turns, the magnetic permeability of the coil ferrite, and
the effective area of the coil. The physical shapes of the
coils are also important for the sensitivity in terms of the
coils effective area and the eld pattern. The eld ampli-
tude is inverse proportional to the range. Since EMI sen-
sors use very low frequencies, the ground surface
reectivity and the soil conductivity losses are very slight
values. Nevertheless, the specific soil types that contain
high conductive and magnetic molecules or particulars
can cause some attenuation, and remarkable clutter levels
can occur. The eddy currents set up in an object, and hence
the induced eld, depend on the size, shape, and compo-
sition of the object. Because of this, different objects gen-
erally have different EMI responses. The basic issues are
then how much information about an object can be in-
ferred from its EMI response, and whether that informa-
tion can be used to distinguish between target and clutter.
The objects that contain much more metallic elements will
naturally produce more powerful eddy currents and sec-
ondary elds on the receiver coil, but the shape of the me-
tallic part in the sense of the angular position with respect
to the primary electromagnetic eld vector will be an im-
portant parameter for detection sensitivity, as well.
4.3. System Performance
The system detection capability of buried metallic
objects is generally a well-known parametric problem as
mentioned above. It is possible to detect any small
amount of metallic or metal-like object under the dynam-
ic range limitations of the system. The range equation
is not as complicated as that of a GPR. For instance, the
soil attenuation and reection loss parameters are not
strong enough to affect system performance; even in most
cases, the size of the metallic content of the object is the
most effective system parameter. The experimental re-
sults presented in Ref. 14 show that the measured volt-
ages (the receiver voltage, after eliminating from the
transmitterreceiver coupling voltage) at the receiver
Transmitter coil
Ground
I
Metallic object
Receiver coil
H
1

H
2
Figure 16. Electromagnetic induction system.
N
S S
A
B
V
AB
A
B
V
AB
I
I
I
H
H
H
H
(a) (b)
Figure 17. Schematic representations of operational theories of
the receiver (a) and transmitter (b) coils.
BURIED OBJECT DETECTION 553
coil are 3000, 450, and 50 mV for the 15-mm-, 8-mm-, and
5-mm-diameter spherical aluminum objects, respectively.
These gures decrease by half for the same size of iron
objects.
The concept of EMI sensor capability involves not only
detection of metal objects, but also discrimination of the
targets of interest from clutter. The traditional detection
technique is made based on the secondary eld energy
calculation method for the decision regarding the presence
or absence of an object. In some physical applications, this
approach can lead to excessively high false-alarm rates
due to the presence of the other metal pieces or granules in
soil. Several modications of traditional EMI sensors are
considered in order to facilitate the target discrimination
and classication. Wideband sensor operation is one of the
most promising methods. In this way, the frequency-de-
pendent response of the induced eld by a buried conduct-
ing object is used to distinguish the target from the clutter
[15]. Another way is to use a simple phenomenological
model that describes the measured time-domain wave-
form as a weighted sum of decaying exponentials to pro-
vide an accurate discrimination model for EMI sensors
[17]. The measured EMI response can be modeled as a
sum of decaying exponential signals whose characteristic
decay rates are intrinsic to the interrogated target. Since
the decay rates are not dependent on target/sensor orien-
tation, decay rate estimation has been proposed as an ef-
fective and robust method for target identication by Ho
et al. [17]. In general, the decay rates associated with me-
tallic objects are slower than that of the earth (ground), so
there is more energy in the received signal when a metal-
lic object is present under the surface of the earth. This
simple phenomenology allows very basic signal processing
to be employed; for example, either an energy detector or
the overall amplitude of the signal in a given time gate
may be used when the goal is to detect any metallic sub-
surface object. Nevertheless, such processing can be the
source of many false alarms in highly cluttered media.
5. ACOUSTIC AND SEISMIC SENSOR
5.1. Overview
Most technologies for the detection of shallow buried ob-
jects are electromagnetic methods. They measure the con-
trast in ferrous content, electrical conductivity, or
dielectric constant between the object and the soil in
which they are buried. Plastic objects have no or little
such contrast because of their nonconductive constitu-
tions. They have dielectric constants that are low, very
close to the dielectric constant of dry soil. This makes
them very difcult to detect by these technologies. Their
acoustic compliance, on the other hand, makes them eas-
ier to detect by seismic and acoustic technologies. These
methods search for buried objects by causing them to vi-
brate by introducing sound or seismic waves into the
ground. As the materials with different properties vibrate
differently, the pattern of ground motion leads to the de-
tection and possibly the identication of the buried object.
Seismic sensors provide a means of sensing the me-
chanical properties of buried objects remotely. Seismic
technologies are commonly employed in exploration of
oil. The equipment used in such applications is designed
to nd layered geology or use eld techniques too expen-
sive to apply. Geophones provide an inexpensive solution
to record seismic data and thus sense buried objects. The
main difculty of this approach is to generate clean Ray-
leigh waves. Also, the geophones have to contact the
ground to perform the measurements. This is not desir-
able for detecting objects containing explosives.
Seismic motion can also be generated by an airborne
acoustic wave. A loudspeaker above the ground can be
used for this purpose. A phenomenon called acoustic-to-
seismic coupling occurs when an airborne acoustic wave is
incident at the ground surface. This term refers to the
coupling of acoustic energy into the ground as seismic mo-
tion. Acoustic-to-seismic coupling causes particles to vi-
brate on the soil surface. More recent techniques use a
laser Doppler vibrometer (LDV) to measure and study the
spatial pattern of particle velocity amplitude of the sur-
face. Avariety of sensors such as radars, microphones, and
ultrasonic devices have also been tried to detect the vi-
brations and the backscattered sound.
There are three main approaches for buried-object de-
tection using acoustics:
1. The Linear Acoustic Technique. This technique has
proved to be an extremely accurate technology for
locating buried landmines [19]. In this technique,
loudspeakers insonify broadband acoustical noise
over the soil and a laser Doppler velocimeter
(LDV) equipped with XY scanning mirrors is used
to detect increased soil vibration across a scan re-
gion.
2. The Nonlinear Acoustic Technique. This technique
uses a single speaker to drive two tonal excitations
to provide the airborne sound to produce A/S cou-
pling [20]. An accelerometer is used to measure the
surface acceleration.
3. Excitation of Elastic Waves and Surface Displace-
ment Measurement. The main objective of this ap-
proach is to excite elastic waves in the soil and then
measure the surface displacements [21]. It uses an
electrodynamic shaker to excite elastic waves. As
the waves propagate through a scan region, the sur-
face displacements are measured using radar.
5.2. Linear Acoustic Techniques
Acousticseismic systems proposed by Sabatier et al. have
shown great promise in detecting low metal mines [19,22].
In these systems, a loudspeaker above the ground surface
insonies the target region at the surface. Acoustic energy
is coupled into the ground, producing vibrations. The ve-
locity of the ground surface resulting from these vibra-
tions is measured with a laser Doppler vibrometer (LDV)
at several positions over a rectangular grid. The output of
the system at a single spatial position is a collection of
complex velocities at a set of frequencies, typically within
the range from B100 to 300 Hz for large objects and with-
in 1001000 Hz for small objects. The data collected over a
554 BURIED OBJECT DETECTION
region form a three-dimensional (3D) complex image. This
3D image can be regarded as a stack of two-dimensional
(2D) images, referred to as spectral images. These spectral
images can be viewed as the displacement energy of the
soil above the buried object at various frequencies. The
analysis of these images leads to detection and identica-
tion of the buried object [23,24].
There are three different versions of the linear acoustic
systems proposed by Sabatier et al:
1. The rst version uses a stop-and-stare data collec-
tion scheme [19,22]. The LDV collects data staring at a
single spatial position and then stares at a position close to
the previous one. A region of interest is thus scanned by
stopping and staring. The data are collected with some
predened spatial resolution, which obviously depends on
the size of the object to be detected. As a rule, the spatial
resolution is selected such that a minimum of three equal-
ly spaced positions is scanned along the major axis of the
object to increase detection rate. The scantime of this type
of data collection scheme may be too slow for some appli-
cations such as mine detection, where a faster scanning
speed is always desirable.
2. A faster version of the linear acoustic system scans a
region without stopping [25]. A laser beam continuously
sweeps over an area by moving back and forth as shown in
Fig. 18. The spatial resolution in downtrack can be ad-
justed by factoring in the size of the buried object. How-
ever, the cross-track resolution depends on the imaging
techniques employed as well as the sweeping speed. Un-
like the previous systems, spectral images are not readily
available. Transforming the time-domain data into spec-
tral images is a critical step in this second approach. The
scanning speed of an area by the sweeping beam approach
is faster than that of the stop-and-stare approach.
3. A much faster scanning speed is achieved using an
array of LDVs placed on a moving platform [26]. Scanning
of the area is performed by moving the platform. The
LDVs on the platform are positioned such that they point
to uniquely spaced positions in cross-track direction (see
Fig. 19). When the platform is moved in the downtrack
direction, each LDV scan the region along the path. Since
this system uses more than one LDV, a normalization
problem is likely to occur due to nonuniqueness of the
LDVs.
5.3. Nonlinear Acoustic Techniques
Nonlinear acoustic technique proposed by Donskoy em-
ploys dual-frequency ( f
1
and f
2
) acoustic signal transmit-
ted toward a buried object [20]. It relies on reception of a
signal generated on the soilburied object interface. This
signal has frequencies different from the initially radiated
signal. The dual-frequency sound causes an object to ra-
diate a signal at the intermodulation frequencies ( f
1
7f
2
).
Thus, the presence of the difference frequency in the ra-
diated signal leads to detection of the object. Some solid
object such as bricks, steel disks, tree roots, and rocks do
not radiate this new signal with the difference frequency.
This makes it easier to differentiate certain objects from
debris.
Dynamic behavior of the buried object is strongly de-
pendent on the stiffness of the soil and the buried object.
The stiffness of solid objects such as rocks and shrapnel is
usually much higher than that of the soil in which they
are buried. Their burial of depth also affects their stiffness.
The stiffness increases as the depth of the object increases.
When the stiffness of the object is much higher than that
of the soil, the nonlinear motion is not observable; rather,
it is a linear motion. Therefore, the nonlinear acoustic
technique is not sensitive to such solid objects and deeply
buried objects. Donskoy explains this by modeling the
buried object as a massspring system [20]. This model
is discussed in Section 5.5.1.
5.4. Simultaneous Use of Elastic and Electromagnetic Waves
The detection system proposed by Scott et al. simulta-
neously uses elastic and electromagnetic waves [21]. Elas-
tic surface waves are generated using an elastic wave
transducer, which is an electrodynamic shaker in direct
contact with the soil (see Fig. 20). The shaker is excited
Crosstrack
D
o
w
n

t
r
a
c
k
Figure 18. Sweeping scheme for laser measurement of the ve-
locity of ground vibration.
Figure 19. LDV array on a moving platform.
BURIED OBJECT DETECTION 555
with a differentiated Gaussian pulse with a center fre-
quency of 400Hz. The waves propagate through the scan
region. A radar-based sensor is used to measure the sur-
face displacements of the soil throughout the scan region.
The displacements of the surface are measured as a func-
tion of time and position. The sensitivity of the radar is
typically around 1 nm. As in the linear and nonlinear
acoustic techniques, the technique relies on the exible
mechanical structure of the object to be detected. If the
object is solid and inexible, it will not exhibit resonant
response. The surface displacements of the soil are then
will not likely yield to the detection of the object. Flexible
objects such as plastic pipes can be detected using this
technology. The main advantage of this method is that
it can be used even if the surface is covered with some
vegetation. Electrical arc source and air acoustic source
have also been investigated in place of the electrodynamic
shaker.
5.5. Detection Algorithms
An important step in object detection by acoustic/seismic
technology is to process the acquired data. The processing
steps for detecting an object are dependent on the
acquired data, sensor type, soil type, target type, and
environmental conditions. For example, the acquired
data can be in either time or frequency domain, and this
may require applying different processing steps to prepare
for further processing. Figure 21 shows a sensor output in
the time domain. Each row represents the time-domain
signals acquired by sweeping a region on the ground along
the scan direction. The location of the buried object is
shown by the dotted line in the same gure. A number of
imaging techniques, such as short-time fast Fourier trans-
form (FFT) and the YuleWalker method, can be applied to
obtain spectral images. In some cases, on the other hand,
the sensor readily provides the spectral images as shown
in Fig. 22. In the sample spectral images shown in the
gure, the buried object is located at the lower left of the
8 12 grid with a spatial resolution of 2 cm. If multiple
sensors are used to collect the signals at each row in Figs.
21 and 22, a processing step to equalize the channels may
be necessary if the sensors are not unique. This step is
skipped if a single sensor is used to scan the same region
by stopping and staring, as shown in Fig. 18. It is, there-
fore, necessary to design detection algorithms specific to
the detection technique. Some examples of automated
methods for detecting and discriminating buried objects
from clutter objects are discussed below.
5.5.1. Model-Based Detection Techniques. Mathemati-
cal models are often used for detection and identication
of objects. For buried objects, the depth and the physical
properties of the objects and the soil are the common pa-
rameters of the models. Unfortunately, some of these pa-
rameters such as the depths of the objects are not known
or sometimes their estimation is too time-consuming and
impractical. These problems render the model useless
even if a perfect model is available. Also, it is obvious
that a general model for any type of buried object cannot
Scanning
Electrodynamic
shaker
Elastic surface
waves
Air
Soil
Object
Electromagnetic waves
S S
N
Radar
Figure 20. Use of elastic and electromagnetic waves for object
detection.
Figure 21. LDV measurements of an acoustic/seismic object de-
tection system producing time-domain signals.
90 Hz
100 Hz
110 Hz
120 Hz
Figure 22. LDV measurements of an acoustic/seismic object de-
tection system producing spectral images.
556 BURIED OBJECT DETECTION
be given. Therefore, each model must be specific to each
problem.
A mathematical model for a buried mine using coupled
damped harmonic oscillators was developed by Donskoy to
interpret acoustic measurements taken on the soil surface
[20]. On the basis of this model, it can be shown that the
transfer function of a buried mine is in the following form:
Hs
Vs
Fs
K
s
3
n
1
s
2
n
2
s
s
4
p
1
s
3
p
2
s
2
p
3
s p
4
17
where V is the velocity of the vibration on the surface of
the ground and F is the driving acoustic force at a partic-
ular spatial location [27]. The coefcients are somewhat
complicated functions of the model parameters. They are
very difcult to estimate and are not given here. Assuming
that F(s) is nearly constant over all frequencies, the fre-
quency spectrum of the velocity of the ground surface vi-
bration is proportional to |H(jo)| and it is what is
actually measured by the LDV:
jVjoj %
jo
3
n
1
jo
2
n
2
jo
jo
4
p
1
jo
3
p
2
jo
2
p
3
jo p
4

18
A sample Bode diagram of the transfer function in (17) is
shown in Fig. 23. This system has four poles and three
transmission zeros. As the depth of the mine increases, the
leftmost peak shifts to the left and the rightmost peak
shifts to the right.
It can be easily veried that an autoregressive moving-
average (ARMA) process with order (4,4) has the same
form as shown in (17). This suggests that the acoustic
measurements, namely, the velocity of the vibration of the
ground surface, have frequency spectra that can be mod-
eled by an ARMA process with order (4,4). This leads to
pole analysis for detection of the buried objects [27].
Observe that the change in the phase is dramatic at the
same frequency of the leftmost peak of the magnitude plot
in the Bode diagram in Fig. 23. Another approach uses
this property to extract feature based on the phase infor-
mation [28]. It significantly lowers false-alarm rates at
given detection probabilities.
5.5.2. Shape-Based Detection. A common approach for
object detection is to use shape information [23,24]. For
buried objects, the spectral images usually provide good
shape information. In Fig. 24, for example, the shape of
the object shown in the spectral images at various bands is
quite round. Thus, the roundness can be used as a feature.
Shape-based detection algorithms usually have a pre-
processing step that may include background subtraction,
normalization, and thresholding. Connected component
analysis usually follows this step. Features such as eccen-
tricity, minor-axis length, major-axis length, area, and
compactness are found useful for discriminating different
objects [23]. Hocaog lu et al. used independent component
analysis to extract additional features to enhance the de-
tection rate [24]. Similar features can also be applicable to
the images representing the basis vectors of independent
components.
6. OPTICAL DETECTORS
Detection of buried objects by optical sensing is regarded
as one of the most promising techniques [29,30]. Its main
advantage is that passive sensors are employed, which is
very important especially in military applications. A com-
mon practice is to use several sensors (cameras) sensitive
at different wavelengths. The spectrally reach reection
region (from 0.4 to 2.5 mm) is available for only daylight
detection. The detection concepts using the midwave in-
frared (35 mm) and the longwave infrared (812 mm) are
usually emphasized for a day/night system.
There are two main approaches for optical detection;
the sensor can detect either (1) the thermal signature of
the buried object itself or (2) a disturbance to the surface
soil caused by the process of burying an object. The rst
approach uses the thermal contrast governed by the
alternation of temperature over day and night and it
is referred as the thermal detection method. As buried
objects heat up at a rate different from that for soil, a
thermal signature can be obtained to detect them. The
latter is referred as the nonthermal detection method. The
principal rationale behind the nonthermal detection
method is that the spectral properties of the surface
layer are different from those of the subsurface soil.
The act of burying an object will bring some subsurface
materials such as SiO
2
content of rocks and soils to the
surface. Thus, the spectral properties of the soil above
the buried object will be different from the spectral prop-
erties of the soil around the same region. The contrast
between the spectral properties will lead to the detection
of a freshly buried object exploiting the 89.4 mm Rests-
trahlen feature, which refers to absorption of energy as a
function of silica content. The emissivity decreases more
Bode diagram
Frequency (rad/sec)
M
a
g
n
i
t
u
d
e

(
d
B
)
P
h
a
s
e

(
d
e
g
)
90
45
60
40
20
0
20
0
45
90
10
1
10
0
10
1
10
2
Figure 23. A sample Bode diagram of the transfer function in
Eq. (17).
BURIED OBJECT DETECTION 557
and also shifts as wavelengths become longer and as silica
content increases.
Soil contains both large and small particles. In fact,
many of the larger particles are usually coated with small-
er particles. The smaller particles in the surface layer, on
the other hand, tend to be cleared from the larger particles
by wind and rain. This weathering process over time caus-
es the clean large particle to dominate the surface layer,
and thus changing the spectral property of the soil at the
surface layer. When the soil is disturbed, the chemical
composition of the surface will change. The weathering
process over time will bring the surface to its original
state. Therefore, this technique may not be suitable to de-
tect objects that were buried a long time ago.
Both of these methods have extreme variability in
performance. Their performances depend heavily on
environmental conditions such as the humidity and the
temperature. Figure 24 shows images of two buried ob-
jects in the thermal band, obtained by a thermal camera
with an operating wavelength of 7.513 mm. The objects
are 10 cm deep and they have circular shape with a diam-
eter of 15cm. The best contrast between the areas near
the objects and the surrounding areas is obtained at the
warmest time of the day. But, the objects lose heat more
quickly at night. Thus, they show up as dark spots in the
imagery, instead of the bright spots. Laser illumination or
microwave radiation can be used to induce these differen-
tial temperature proles.
7. OTHER SENSORS
In this section, the auxiliary sensor types generally used
for the specific application areas or to improve detector
performance, will briefly be presented with their operat-
ing principles, current capabilities, limitations, and fur-
ther improvement potentials. Most of these sensor
technologies, such as electrical impedance tomography,
X-ray imaging, and explosive-material detectors, are still
at the stage of prototyping or technical research [13].
Electrical impedance tomography (EIT) is a low-cost
and basic measurement technique employed to obtain the
conductivity distribution of the ground surface. The de-
tection principle is based on the discrimination of the
shallow buried objects that have electrical resistance dif-
ferent from that of soil. The ground surface is probed by
many numbers of low-current electrodes to image the sur-
face resistivity with a required resolution [31]. The detec-
tion capability of EIT is limited to the probing depth,
which is usually less than 20 cm. It is also not convenient
for quick-time wide-area scanning operations.
X-ray scattering is one of the highest-resolution buried-
object detector technologies. The pencil beam of the X rays
is emitted to the scanning area, and the backscattered
rays are detected by the sensor [32]. The performance de-
pend essentially on the X-ray absorbing characteristics of
the target object. For instance, organic materials, soil lay-
ers, and explosives have less absorbing and more scatter-
ing characteristics than do metals. Solid objects are more
easily distinguished than are clutters. Since higher X-ray
energy sources are needed to penetrate deeper, this sen-
sors are usually suitable for shallow buried targets.
Subsurface microwave tomography is actually a type of
electromagnetic sensor. The operating principle and hard-
ware units of the diffraction tomography system are sim-
ilar to those of stepped-frequency GPR. However, such
systems generally use higher frequencies up to millimeter
bands and their detection process is based on solution of
the inverse scattering problem. The sweeping microwave
signal is radiated to the buried object, and the diffracted
wave from the target is received. Wideband antennas,
mostly dielectric-loaded horns, are employed for wave
transmitting and receiving. Then, I/Q data signatures
of the buried target are processed using some image
reconstruction algorithms such as BornKirchhoff approx-
imations. The diffraction tomography sensor can yield
high-resolution subsurface mapping and target imaging
[33]. Nevertheless, the detection depth is less then 10 cm,
due to the strong soil attenuation at high frequencies, and
they are not still convenient for real-time detection since
the scanning time is usually very long.
In addition to these detector types, some biological,
electrochemical, and piezoelectric sensor systems are em-
ployed for specific applications such as explosive-material
detection, which may be very important especially for mil-
itary applications (i.e., mine detection). Biological smell
sensors can detect the vapors or the gases produced by the
explosive components of objects. The electrochemical de-
tector measures the changes in polymer electrical resis-
tance of the explosive vapors. The piezoelectric sensor
uses the resonant frequency shifting of various materials
when exposed to the explosive vapors [13].
BIBLIOGRAPHY
1. D. J. Daniels, Surface Penetrating Radar, IEE Radar, Sonar,
Navigation and Avionics Series 6, London, 1996.
2. Picosecond Pulse Labs Application Notes, www.picosec-
ond.com.
3. A. S. Turk and B. Sen, Ultra-wide-band antenna designs for
GPR impulse radar systems, IEEE EMC Symp. Records, May
2003.
4. D. S. A. Sahinkaya and A. S. Turk, UWB GPR for detection
and identication of buried small objects, SPIE Proc.
5410:174184 (2004).
5. M. Sezgin, F. Kurugollu, I. Tas-delen, and S. Ozturk, Real time
detection of buried objects by using GPR, SPIE Proc.
5415:447455 (2004).
6. A. S. Turk, Ultra wide band TEM horn design for ground
penetrating impulse radar system, Microwave Opt. Technol.
Lett. 41(6):333336 (June 2004).
(a) (b) (c)
Figure 24. IR imagery taken at different times of day: (a) 8:00
A.M.; (b) 2:00 P.M.; (c) 3:00 P.M.
558 BURIED OBJECT DETECTION
7. L. C. Chan, D. L. Moffat, and L. Peters, A characterization of
subsurface radar targets, Proc. IEEE 67(7):9911000 (July
1979).
8. J. S. Lee and C. Nguyen, Novel low-cost ultra-wideband,
ultra-short-pulse transmitter with MESFET impulse-shaping
circuitry for reduced distortion and improved pulse repetition
rate, IEEE Microwave Wireless Compon. Lett. 11(5) (May
2001).
9. R. W. P. King and G. S. Smith, Antennas in Matter, MIT Press,
1981.
10. K. L. Schlager, G. S. Smith, and J. G. Maloney, Optimization
of bow-tie antenna for pulse radiation, IEEE Trans. Anten.
Propag. 42(7):975982 (July 1994).
11. A. G. Yaravoy, A. D. Schukin, I. V. Kaploun, and L. P. Lightart,
The dielectric wedge antenna, IEEE Trans. Anten. Propag.
50(10):14601471 (Oct. 2002).
12. V. P. Chumachenko and A. S. Turk, Radiation characteristics
of wide-angle H-plane sectoral horn loaded with dielectric of
multiangular shape, Int. J. Electron. 88(1):91101 (Jan.
2001).
13. Alternatives for Landmine Detection, RAND Science and
Technology Policy, RAND Corp., Santa Monica, CA, 2003.
14. S. Yamazaki, H. Nakane, and A. Tanaka, Basic analysis of a
metal detector, Proc. IEEE Instrumentation and Measure-
ment Technology Conf. Budapest, Hungary, May 2123, 2001.
15. P. Gao, L. Collins, N. Geng, L. Carin, D. Keiswetter, and I. J.
Won, Classication of landmine-like metal targets using
wideband electromagnetic induction, IEEE Trans. Geosci. Re-
mote Sens. 38(3):13521361 (May 2000).
16. T. H. Bell, B. J. Barrow, and J. T. Miller, Subsurface discrim-
ination using electromagnetic induction sensors, IEEE Trans.
Geosci. Remote Sens. 39(6):12861293 (June 2001).
17. K. C. Ho, L. M. Collins, L. G. Huettel, and P. D. Gader, Dis-
crimination mode processing for EMI and GPR sensors for
hand-held land mine detection, IEEE Trans. Geosci. Remote
Sens. 42(1):249263 (Jan. 2004).
18. M. I. Skolnik, Radar Handbook, McGraw-Hill, New York,
1970.
19. J. M. Sabatier and N. Xiang, An investigation of a system that
uses acoustic-to-seismic coupling to detect buried ant-tank
mines, IEEE Trans. Geosci. Remote Sens. 39:11461154
(2001).
20. D. M. Donskoy, Nonlinear vibro-acoustic technique for land-
mine detection, SPIE Proc. 3392:211217 (1998).
21. W. R. Scott, G. D. Larson, J. S. Martin, and P. H. Rogers,
Seismic/electromagnetic system for landmine detection, J. Ac-
oust. Soc. Am. 107(5)(Part 2):28 (May 2000).
22. J. M. Sabatier and C. J. Hickey, Acoustic-to-seismic transfer
function at the surface of a layered outdoor ground, Proc.
SPIE 4038:633644 (2000).
23. A. K. Hocaog lu, P. Gader, J. Keller, and B. Nelson,
Anti-personnel landmine detection and discrimination using
acoustic data, Subsurf. Sens. Technol. Appl. 3(2):7593 (April
2002).
24. A. K. Hocaog lu and P. Gader, Continuous processing of acous-
tics data for landmine detection, Proc. SPIE 4742:654664
(Detection and remediation technologies for mines and mine
like targets VII) (April 2002).
25. N. Xiang and J. M. Sabatier, Acoustic-to-seismic landmine
detection using a continuously scanning laser Doppler vibro-
meter, Proc. SPIE 5089:591595 (Detection and remediation
technologies for mines and minelike targets VIII) (April
2003).
26. R. D. Burgett, M. R. Bradley, M. Duncan, J. Melton, A. K. Lal,
V. Aranchuk, C. F. Hess, J. M. Sabatier, and N. Xiang, Mobile
mounted laser Doppler vibrometer array for acoustic land-
mine detection, Proc. SPIE 5089:665672 (Detection and re-
mediation technologies for mines and minelike targets VIII)
(April 2003).
27. A. K. Hocaog lu, P. D. Gader, and G. X. Ritter, Acoustic/Seismic
imaging using spectral estimation for landmine detection,
Proc. Int. Conf. Requirements and Technologies for the Detec-
tion, Removal and Neutralization of Landmines and UXO,
Brussels, Belgium, 2004, Vol. 2, pp. 489495.
28. T. Wang, J. Keller, P. D. Gader, and A. K. Hocaoglu, Phase
signatures in acoustic-seismic landmine detection, Radio Sci.
39(4):RS4S02 (2004).
29. J. Salisbury, L. Walter, N. Vergo, and D. DAria, Infrared
(2.12.5mm) Spectra of Minerals, Johns Hopkins Univ. Press,
Baltimore, 1991.
30. S. Hong, T. W. Miller, H. Tobin, B. Borchers, J. M. Hendrickx,
H. A. Lensen, P. B. Schwering, and B. A. Baertlein, Impact of
soil water content on landmine detection using radar and
thermal infrared sensors, Proc. SPIE 4394:409416 (Detec-
tion and remediation technologies for mines and minelike
targets VI) (Oct. 2001).
31. A. Wexler, Electrical impedance imaging in two and
three dimensions, Clin. Phys. Physiol. Meas. (Suppl. A):
2933 (1988).
32. B. Towe and A. Jacobs, X-ray backscatter imaging, IEEE
Trans. Biomed. Eng. BME-28:646654 (1981).
33. A. A. Vertiy and S. P. Gavrilov, Modeling of microwave images
of buried cylindrical objects, Int. J. Infrared Millimeter Waves
19(9):12011220 (1998).
BURIED OBJECT DETECTION 559