Turbulence Interactions

110 Notes on Numerical Fluid Mechanics and Multidisciplinary Design (NNFM)
Editors
W. Schröder/Aachen
B.J. Boersma/Delft
K. Fujii/Kanagawa
W. Haase/München
M.A. Leschziner/London
J. Periaux/Paris
S. Pirozzoli/Rome
A. Rizzi/Stockholm
B. Roux/Marseille
Y. Shokin/Novosibirsk
Turbulence and Interactions
Proceedings the TI 2009 Conference
Michel Deville
Thien-Hiep Lê
Pierre Sagaut
(Editors)
ABC
Prof. Michel Deville Prof. Pierre Sagaut
EPFL STI IGM LIN Institut Jean Le Rond d’Alembert
Station 9 Université Pierre et Marie Curie
1015 Lausanne 4 Place Jussieu
Switzerland 75252 Paris cedex 5
E-mail: michel.deville@epfl.ch France
Dr. Thien-Hiep Lê
ONERA
29 Avenue de la Division Leclerc
92322 Chatillon
France
ISBN 978-3-642-14138-6 e-ISBN 978-3-642-14139-3
DOI 10.1007/978-3-642-14139-3
Notes on Numerical Fluid Mechanics

and Multidisciplinary Design ISSN 1612-2909
Library of Congress Control Number: 2010929478

c 2010 Springer-Verlag Berlin Heidelberg
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations
are liable for prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant protective
laws and regulations and therefore free for general use.
Typeset & Cover Design: Scientific Publishing Services Pvt. Ltd., Chennai, India.
Printed on acid-free paper

543210
springer.com
NNFM Editor Addresses
Prof. Dr. Wolfgang Schröder Prof. Dr. Sergio Pirozzoli

(General Editor) Università di Roma “La Sapienza”
RWTH Aachen Dipartimento di Meccanica e Aeronautica
Lehrstuhl für Strömungslehre und Via Eudossiana 18
Aerodynamisches Institut 00184, Roma, Italy
Wüllnerstr. 5a E-mail: sergio.pirozzoli@uniroma1.it
52062 Aachen
Germany Prof. Dr. Jacques Periaux
E-mail: office@aia.rwth-aachen.de 38, Boulevard de Reuilly
F-75012 Paris
Prof. Dr. Kozo Fujii France
Space Transportation Research Division E-mail: jperiaux@free.fr
The Institute of Space
and Astronautical Science Prof. Dr. Arthur Rizzi
3-1-1, Yoshinodai, Sagamihara Department of Aeronautics
Kanagawa, 229-8510 KTH Royal Institute of Technology
Japan Teknikringen 8
E-mail: fujii@flab.eng.isas.jaxa.jp S-10044 Stockholm
Sweden
Dr. Werner Haase E-mail: rizzi@aero.kth.se
Höhenkirchener Str. 19d
D-85662 Hohenbrunn Dr. Bernard Roux
Germany L3M – IMT La Jetée
E-mail: office@haa.se Technopole de Chateau-Gombert
F-13451 Marseille Cedex 20
Prof. Dr. Ernst Heinrich Hirschel France
(Former General Editor) E-mail: broux@l3m.univ-mrs.fr
Herzog-Heinrich-Weg 6
D-85604 Zorneding Prof. Dr. Yurii I. Shokin
Germany Siberian Branch of the
E-mail: e.h.hirschel@t-online.de Russian Academy of Sciences
Institute of Computational
Prof. Dr. Ir. Bendiks Jan Boersma Technologies
Chair of Energytechnology Ac. Lavrentyeva Ave. 6
Delft University of Technology 630090 Novosibirsk
Leeghwaterstraat 44 Russia
2628 CA Delft E-mail: shokin@ict.nsc.ru
The Netherlands
E-mail: b.j.boersma@tudelft.nl
Prof. Dr. Michael A. Leschziner

Imperial College of Science
Technology and Medicine
Aeronautics Department
Prince Consort Road
London SW7 2BY
U.K.
E-mail: mike.leschziner@ic.ac.uk
Preface
The “Turbulence and Interactions 2009” (TI2009) conference was held in Saint-
Luce on the island of La Martinique, France, on May 31-June 5, 2009. The scien-
tific sponsors of the conference were
• DGA
• Ecole Polytechnique Fédérale de Lausanne (EPFL),
• ERCOFTAC : European Research Community on Flow, Turbulence and
Combustion,
• Institut Jean Le Rond d’Alembert, Paris,
• ONERA.
This second TI conference was very successful as it attracted 65 researchers
from 17 countries. The magnificent venue and the beautiful weather helped the
participants to discuss freely and casually, share ideas and projects, and spend
very good times all together.
The organisers were fortunate in obtaining the presence of the following in-
vited speakers: L. Fuchs (KTH, Stockholm and Lund University), J. Jimenez
(Univ. Politecnica Madrid), C.-H. Moeng (NCAR), A. Scotti (University of North
Carolina), L. Shen (Johns Hopkins University) and A.J. Smits (Princeton Univer-
sity). The topics covered by the 62 contributed papers ranged from experimental
results through theory to computations. They represent a snapshot of the state-of-
the-art in turbulence research. The papers of the conference went through the
usual reviewing process and the result is given in this book of Proceedings.
In the present volume, the reader will find the keynote lectures followed by the
contributed talks given in alphabetical order of the first author.
The organizers of the conference would like to acknowledge the support of
EPFL, Université Pierre et Marie Curie, Paris and ONERA. They express their
gratitude for their colleagues of the organizing committee, especially Drs. V. Gle-
ize and M. Terracol, for their help and constant efficiency.
Lausanne, Paris M. O. Deville

December 3, 2009 T. H. Lê
P. Sagaut
Contents
Keynote Lectures
Some Characteristics of Non-Reacting and Reacting Low
Swirl Number Jets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
L. Fuchs
Inner-Outer Interactions in Wall-Bounded Turbulence . . . . . . . 3
Javier Jiménez
Turbulence Interaction with Atmospheric Physical
Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Chin-Hoh Moeng, Jeffrey Weil
LES of Pulsating Turbulent Flows over Smooth and Wavy
Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
A. Scotti, M. Gasser i Rubinat, E. Balaras
Numerical Study of Turbulence–Wave Interaction . . . . . . . . . . . 37
Lian Shen
High Reynolds Number Wall-Bounded Turbulence and a
Proposal for a New Eddy-Based Model . . . . . . . . . . . . . . . . . . . . . . 51
Alexander J. Smits
Regular Papers
PANS Methodology Applied to Elliptic-Relaxation Based
Eddy Viscosity Transport Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Branislav Basara, Siniša Krajnović, Sharath Girimaji
PIV Study of Turbulent Flow in Porous Media . . . . . . . . . . . . . 71
S. Bejatovic, M.F. Tachie, M. Agelinchaab, S.S. Paul
X Contents
A Model for Dissipation: Cascade SDE with Markov

Regime-Switching and Dirichlet Prior . . . . . . . . . . . . . . . . . . . . . . . 79
D. Bernard, A. Tossa, R. Emilion, S.K. Iyer
Wavelet Analysis of the Turbulent LES Data of the
Lid-Driven Cavity Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Roland Bouffanais, Guy Courbebaisse, Laurent Navarro,
Michel O. Deville
A Two-Phase LES Compressible Model for Plasma-Liquid
Jet Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Céline Caruyer, Stéphane Vincent, Erick Meillot,
Jean-Paul Caltagirone
Simulation of a Fluidized Bed Using a Hybrid
Eulerian-Lagrangian Method for Particle Tracking . . . . . . . . . . . 103
Cédric Corre, Jean-Luc Estivalezes, Stéphane Vincent,
Olivier Simonin, Stéphane Glockner
Wavelet-Adapted Sub-grid Scale Models for LES . . . . . . . . . . . . 111
J.A. Denev, C.J. Falconi, J. Fröhlich, H. Bockhorn
Effect of Particle-Particle Collisions on the Spatial

Distribution of Inertial Particles Suspended in
Homogeneous Isotropic Turbulent Flows . . . . . . . . . . . . . . . . . . . . . 119
Pascal Fede, Olivier Simonin
Effect of Near-Wall Componental Modification of
Turbulence on Its Statistical Properties . . . . . . . . . . . . . . . . . . . . . . 127
Bettina Frohnapfel, Yosuke Hasegawa, Nobuhide Kasagi
Large-Eddy Simulation of Transonic Buffet over a
Supercritical Airfoil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
E. Garnier, S. Deck
Large Eddy Simulation of Coherent Structures over Forest
Canopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
K. Gavrilov, G. Accary, D. Morvan, D. Lyubimov, O. Bessonov,
S. Méradji
Toroidal/Poloidal Modes Dynamics in Anisotropic
Turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Fabien S. Godeferd, Alexandre Delache, Claude Cambon
Grid Filter Modeling for Large-Eddy Simulation . . . . . . . . . . . . . 159
Marc A. Habisreutinger, Roland Bouffanais, Michel O. Deville
Contents XI
Pulsating Flow through Porous Media . . . . . . . . . . . . . . . . . . . . . . . 167

Michele Iervolino, Marcello Manna, Andrea Vacca
Thermodynamic Fluctuations Behaviour during a Sheared
Turbulence/Shock Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
S. Jamme, M. Crespo, P. Chassaing
LES and DES Study of Fluid-Particle Dynamics in a
Human Mouth-Throat Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
S.T. Jayaraju, S. Verbanck, C. Lacor
Viscous Drag Reduction with Surface-Embedded Grooves . . . 191
Jovan Jovanović, Bettina Frohnapfel, Antonio Delgado
Study on the Resolution Requirements for DNS in
Turbulent Rayleigh-Bénard Convection . . . . . . . . . . . . . . . . . . . . . . 199
M. Kaczorowski, C. Wagner
On the Role of Coherent Structures in a Lid Driven Cavity
Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Benjamin Kadoch, Emmanuel Leriche, Kai Schneider, Marie Farge
Local versus Nonlocal Processes in Turbulent Flows,
Kinematic Coupling and General Stochastic Processes . . . . . . . 215
Michael Kholmyansky, Vladimir Sabelnikov, Arkady Tsinober
Time-Resolved 3D Simulation of an Aircraft Wing with
Deployed High-Lift System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Thilo Knacke, Frank Thiele
Fluid Mechanics and Heat Transfer in a Channel with
Spherical and Oval Dimples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Nikolai Kornev, Johann Turnow, Egon Hassel, Sergei Isaev,
Frank-Hendrik Wurm
Investigation of the Flow around a Cylinder Plate
Configuration with Respect to Aerodynamic Noise
Generation Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
Michael Kornhaas, Dörte C. Sternel, Michael Schäfer
LES of the Flow around Ahmed Body with Active Flow
Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Siniša Krajnović, Branislav Basara
Enhanced Bubble Migration in Turbulent Channel Flow by
an Acceleration-Dependent Drag Coefficient . . . . . . . . . . . . . . . . . 255
J.G.M. Kuerten, C.W.M. van der Geld, B.J. Geurts
XII Contents
Experimental and Numerical Study of Unsteadiness in

Boundary Layer / Shock Wave Interaction . . . . . . . . . . . . . . . . . . . 263
L. Larchevêque, P. Dupont, E. de Martel, E. Garnier, J.-F. Debiève
Measurement of Particle Accelerations with the Laser
Doppler Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
H. Nobach, M. Kinzel, R. Zimmermann, C. Tropea, E. Bodenschatz
A Novel Numerical Method for Turbulent, Two-Phase
Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
A. Pecenko, J.G.M. Kuerten
Modeling of High Reynolds Number Flows with Solid Body
Rotation or Magnetic Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Annick Pouquet, Julien Baerenzung, Jonathan Pietarila Graham,
Pablo Mininni, Hélène Politano, Yannick Ponty
Direct Numerical Simulation of Buoyancy Driven
Turbulence inside a Cubic Cavity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
R. Puragliesi, A. Dehbi, E. Leriche, A. Soldati, M. Deville
Numerical Simulations of a Massively Separated Reactive
Flow Using a DDES Approach for Turbulence Modelling . . . . 303
Bruno Sainte-Rose, Nicolas Bertier, Sébastien Deck,
Francis Dupoirieux
Particle Dispersion in Large-Eddy Simulations:
Influence of Reynolds Number and of Subgrid Velocity
Deconvolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Maria Vittoria Salvetti, Cristian Marchioli, Alfredo Soldati
Use of Lagrangian Statistics for the Direct Analysis of the
Turbulent Constitutive Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
François G. Schmitt, Ivana Vinkovic
Numerical Simulation of Supersonic Jet Noise with Overset
Grid Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
J. Schulze, J. Sesterhenn
Large Eddy Simulation of Turbulent Jet Flow in Gas
Turbine Combustors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Y. Shimada, B. Thornber, D. Drikakis
Computations of the Flow around a Wind Turbine: Grid
Sensitivity Study and the Influence of Inlet Conditions . . . . . . 345
R.Z. Szasz, L. Fuchs
Stochastic Synchronization of the Wall Turbulence . . . . . . . . . . 353
Sedat Tardu
Contents XIII
Large-Eddy Simulations of an Oblique Shock Impinging

on a Turbulent Boundary Layer: Effect of the Spanwise
Confinement on the Low-Frequency Oscillations . . . . . . . . . . . . . 361
Emile Touber, Neil D. Sandham
Parameter-Free Symmetry-Preserving Regularization
Modelling of Turbulent Natural Convection Flows . . . . . . . . . . . 369
F.X. Trias, R.W.C.P. Verstappen, M. Soria, A. Oliva
An a Priori Study for the Modeling of Subgrid Terms in
Multiphase Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
P. Trontin, S. Vincent, J.L. Estivalezes, J.P. Caltagirone
Computation of Flow in a 3D Diffuser Using a Two-Velocity
Field Hybrid RANS/LES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
J.C. Uribe, A. Revell, C. Moulinec
On the Dynamics of High Reynolds Number Turbulent
Axisymmetric and Plane Separating/Reattaching Flows . . . . . 393
Pierre-Élie Weiss, Sébastien Deck, Jean-Christophe Robinet,
Pierre Sagaut
Numerical Simulation and Statistical Modeling of

Inertial Droplet Coalescence in Homogeneous Isotropic
Turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
Dirk Wunsch, Pascal Fede, Olivier Simonin, Philippe Villedieu
Gas-Phase Mixing in Droplet Arrays . . . . . . . . . . . . . . . . . . . . . . . . 409
M.R.G. Zoby, S. Navarro-Martinez, A. Kronenburg, A.J. Marquis
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417

Some Characteristics of Non-Reacting and
Reacting Low Swirl Number Jets
L. Fuchs
Abstract. The paper considers low swirl turbulent number jets. Swirling jets are
used to stabilize premixed flames in gas turbines. Normally, the swirl number is
large enough to allow vortex break-down and thereby flame stabilization along the
upstream edge of the back-flow bubble. With decreasing swirl the vortex-breakdown
may disappear altogether. However, it has been found that under certain conditions
the flame may be kept at a certain (mean) distance away from the nozzle even with-
out vortex break-down. The mechanism for the flame holding under such conditions
is discussed. The discussion is based upon LES results and some experimental data.
We discuss also the precession of the central core both under non-reacting and react-
ing condition. LES and experimental results show that the precession of the central
core is normally in the same direction as the swirl. However, for certain range of
swirl numbers and at some axial distances one may find precession in the counter
direction. The mechanism for this effect is discussed.
L. Fuchs
Linné Flow Centre, Royal Institute of Technology, Stockholm, Sweden
e-mail: lf@mech.kth.se
M. Deville, T.-H. Lê, and P. Sagaut (Eds.): Turbulence and Interactions, NNFM 110, p. 1.
springerlink.com c Springer-Verlag Berlin Heidelberg 2010
Inner-Outer Interactions in Wall-Bounded
Turbulence
Javier Jiménez
1 Introduction
This paper deals with some of the features that distinguish wall-bounded sheared
turbulence from that in free-shear flows. It concerns itself mostly with the largest
structures at each wall distance, because they are where energy is fed into the fluctu-
ations, and therefore the ones that differ most between the different flows. Because
of the geometric limitations imposed by the wall, the largest scales roughly coin-
cide with the smallest ones in the viscous buffer layer, but the rest of the flow is
characterised, as in most turbulent cases, by a wide range of scales.
Wall-bounded turbulence includes pipes, channels and boundary layers. We re-
strict ourselves to cases with little or no longitudinal pressure gradients, since other-
wise the flow tends to separate and resembles the free-shear case. It was in attached
wall-bounded flows where turbulence was first studied scientifically [16, 6], but
they remain to this day worse understood than homogeneous or free-shear turbu-
lence. That is in part because what is sought in both cases is different. Turbulence
is a multiscale phenomenon. Energy resides in the largest eddies, but it cannot be
dissipated until it is transferred to the smaller scales for viscosity to act. The classi-
cal conceptual framework for that process is the self-similar cascade [35, 28], which
assumes that the transfer is local in scale space, with no significant interactions be-
tween eddies of very different sizes. The resulting model, although now recognised
as only an approximation, describes well the experimental observations in isotropic
turbulence, but also in small-scale turbulence in general. A sketch can be found in
figure 1(a).
The emphasis in shear flows is not on the transfer of energy, but on its production.
Isotropic theory gives no indication of how energy is fed into the cascade. In shear
flows, the energy source is the interaction between the mean velocity gradient and
the average momentum fluxes carried by the velocity fluctuations [39]. In free-shear
Javier Jiménez
School of Aeronautics, U. Politécnica, 28040 Madrid, Spain
e-mail: jimenez@torroja,dmt.upm.es
M. Deville, T.-H. Lê, and P. Sagaut (Eds.): Turbulence and Interactions, NNFM 110, pp. 3–14.
4 J. Jiménez
3
Energy 10
Dissipation
kE(k)
+
10
y
1
(a) 10 (b)
1 3 5 2 3 4
10 10 λ/η 10 10 λx 10
+ 10
Fig. 1 Spectral energy density, kE(k). (a) In isotropic turbulence, as a function of the
isotropic wavelength λ = 2π /|k|. (b) In a turbulent channel [17] with h+ = 2000, plotted
as a function of the streamwise wavelength λx , and of the wall distance y. Shaded contours
are the density of the kinetic energy, kx Euu (kx ). Lines are the spectral density of the surro-
gate dissipation, ν kx Eωω (kx ), where ω are the vorticity fluctuations. At each y the lowest
contour is 0.86 times the local maximum. The horizontal lines, y+ = 80 and y/h = 0.2,
represent conventional logarithmic layer limits. Diagonal is λx = 5y. Arrows indicate the
implied cascades
flows, such as jets or mixing layers, this leads to a large-scale instability of the
mean velocity profile [3], and to large-scale eddies with sizes of the order of the
flow thickness.
The mean velocity profiles of wall-bounded flows are not unstable in the same
way as the free-shear cases, and wall-bounded turbulence is consequently a weaker
phenomenon. While the velocity fluctuations in a jet can easily reach 15-20% of
the mean velocity differences, they rarely exceed 5% in attached boundary layers.
Wall-bounded flows are however of huge technological importance. Roughly half
of the energy spent worldwide to move fluids through pipes and canals, or to move
vehicles through air or through water (20% of the total), is dissipated by turbulence
in the immediate vicinity of the wall.
Wall-bounded flows are also interesting because they force us to face the role of
inhomogeneity. This can be seen in figure 1(b) which is the equivalent of figure 1(a)
for a wall-bounded turbulent flow. Each horizontal section of this figure is equivalent
to the spectra in figure 1(a). The energy is again at large scales, while the dissipative
eddies are smaller. In this case, however, the size of the energy-containing eddies
changes with the distance to the wall, and so does the range of scales over which
the energy has to cascade. The eddies containing most of the energy at one wall
distance are in the midst of the inertial cascade when they are observed farther away
from the wall. The Reynolds number, defined as the scale disparity between energy
and dissipation at some given location, also changes with wall distance. The main
emphasis in wall turbulence is not the inertial energy cascade, but the interplay
between different scales at different distances from the wall.
Inner-Outer Interactions in Wall-Bounded Turbulence 5
Models for wall-bounded turbulence also have to deal with spatial fluxes that are
not present in the homogeneous case. The most important ones are those of momen-
tum. Consider a turbulent channel, driven by a pressure gradient between infinite
parallel planes. Denote by U, V and W the mean velocities along the streamwise,
wall-normal and spanwise directions, x, y and z, and the corresponding fluctuations
by lower-case letters. Streamwise momentum is fed into the channel by the mean
pressure gradient, ∂ x P, which acts over the whole cross section. It is removed only
at the wall, by viscous friction. Momentum has to flow from the centre to the wall,
carried by the Reynolds stress −uv, which resides in eddies of roughly the same
sizes as the energy, and it is clear from figure 1(b) that those sizes change as a func-
tion of the wall distance by as much as the scale of the energy across the inertial
cascade. This implies that momentum is transferred in wall-bounded turbulence by
an extra spatial ‘cascade’. Momentum transport is present in all shear flows, but the
multiscale spatial cascade is characteristic of very inhomogeneous situations, such
as wall turbulence, and complicates the problem considerably.
In this paper we review what is known about the interactions of the different
structures in wall-bounded flows. In section 2 we summarise the present concep-
tual models for both the viscous and the outer regions, and in section 3 we discuss
briefly how the two regions interact with each other, and in particular the question
of whether causality flows from the wall to the outside, or viceversa.
2 The General Organisation of Wall-Bounded Turbulence

The wall-normal variation of the range of the energy cascade divides the flow into
several distinct regions. Wall-bounded turbulence over smooth walls has to be de-
scribed by two sets of scaling parameters [39]. Viscosity is important near the wall,
and the units for length and velocity in that region are constructed with the kine-
matic viscosity ν , and the friction velocity uτ . Magnitudes expressed in wall units
are denoted by + superscripts, and y+ is a Reynolds number for the size of the struc-
tures. It is never large within the viscous layer, which is typically defined at most as
y+ 150 [32], and conventionally divided into a viscous sublayer, y+ 5, where
viscosity is dominant, and a ‘buffer’ layer in which both viscosity and inertial ef-
fects should be taken into account. There is no scale disparity in either region, as
seen in figure 1(b), because most large eddies are excluded by the presence of the
impermeable wall. The energy and the dissipation are at similar sizes.
Away from the wall the velocity also scales with uτ , because the momentum
equation requires that the Reynolds stress, −uv, can only change slowly with y.
This uniform velocity scale is the extra constraint introduced in wall-bounded flows
by the momentum transfer. The lengthscale in the region far from the wall is the
flow thickness h, which in this paper will usually be the semi-channel height.
Between the inner and the outer regions there is an intermediate layer where the
only available lengthscale is the wall distance y. Both the constant velocity scale
across the intermediate region, and the absence of a lengthscale other than y, are
only approximations. It will be seen below that some large-scale eddies of size O(h)
6 J. Jiménez
penetrate to the wall, and that the velocity fluctuations do not scale strictly with uτ ,
even in the viscous sublayer. However, if those approximations are accepted, the
‘logarithmic’ mean velocity profile,
U + = κ −1 log y+ + A, (1)
follows from symmetry arguments. It agrees well with experimental evidence, with
an approximately universal Kármán constant, κ ≈ 0.4, and A ≈ 5 for smooth walls.
The viscous, buffer, and logarithmic layers are the most characteristic features
of wall-bounded flows, and constitute the main difference between them and other
types of turbulence.
2.1 The Buffer Layer

The viscous and buffer layers, although generally very thin, are extremely important
for the flow as a whole. The ratio between the inner and the outer lengthscales is the
friction Reynolds number, h+ , which ranges from 200 for barely turbulent flows, to
5 × 105 for large water pipes. In the latter, the near-wall layer is only about 3 × 10−4
times the pipe radius, but it follows from equation (1) that, even in that case, 40%
of the velocity drop takes place below y+ = 50. Because there is relatively little
net energy transfer among layers, except in the viscous region, those percentages
also apply to the energy dissipation. Turbulence is characterised by the expulsion
towards the small scales of the energy dissipation, away from the large energy-
containing eddies. In the limit of infinite Reynolds number, this is believed to lead
to a non-differentiable velocity field. In wall-bounded flows that separation occurs
not only in the scale space for the velocity fluctuations, but also in the shape of the
mean velocity profile. The singularities are expelled both from the large scales, and
from the centre of the flow towards the walls.
The viscous layers are dominated by coherent streaks of the streamwise veloc-
ity and by quasi-streamwise vortices, which are both the energy-containing and the
dissipative eddies. The streaks are an irregular array of long (x+ ≈ 1000) sinuous
alternating streamwise jets superimposed on the mean shear, with an average span-
wise separation of the order of z+ ≈ 100 [37]. The quasi-streamwise vortices are
slightly tilted away from the wall, and stay in the near-wall region only for x+ ≈ 200.
Several vortices are associated with each streak [23], with a longitudinal spacing of
the order of x+ ≈ 400. Most of them merge into disorganised vorticity outside the
immediate neighbourhood of the wall [36].
It was soon proposed that streaks and vortices are involved in a regeneration cy-
cle in which the vortices are the results of an instability of the streaks [38], while
the streaks are caused by the advection of the mean velocity gradient by the vor-
tices [2, 26]. Both processes have been documented and sharpened by numerical
experiments. For example, disturbing the streaks inhibits the formation of the vor-
tices, but only if it is done between y+ ≈ 10 and y+ ≈ 60 [24], suggesting that it is
predominantly between those two levels that the regeneration cycle works. There is
substantial numerical and analytic work showing that streaks are linearly unstable
to sinuous perturbations associated with inflection points of the distorted velocity

profile, whose eigenfunctions correspond well with the shape and location of the
observed vortices. The implied model is a time-dependent cycle in which streaks
and vortices are created, grow, generate each other, and eventually decay. Reference
[24] discusses models of this type, and gives additional references.
Although the flow in the buffer layer is clearly chaotic, the chaos is not required
to explain the turbulence statistics. Simulations in which the flow is substituted by
an ordered ‘crystal’ of identical ‘minimal’ sets of structures reproduce fairly well the
correct statistics [23]. In a further simplification, that occurred at roughly the same
time as the previous one, nonlinear equilibrium solutions of the three-dimensional
Navier–Stokes equations were obtained numerically in Couette flow, with character-
istics that suggested that they could be useful in a dynamical description of the near-
wall region [31]. Many similar solutions were soon found in other wall-bounded
flows, including limit cycles and heteroclinic connections reminiscent of the cycle
mentioned above. All of them look qualitatively similar [41, 22], and take the form
of a wavy low-velocity streak flanked by a pair of staggered quasi-streamwise vor-
tices of alternating signs, closely resembling the spatially-coherent objects educed
from the near-wall region of true turbulence.
Those solutions were recently reviewed and extended in [22]. It turns out that
they can be classified into ‘upper’ and ‘lower’ branches in terms of their mean
wall shear. The ‘upper’ solutions have relatively weak sinuous streaks flanked by
strong vortices. They consequently have relatively weak streamwise-velocity fluc-
tuations, and strong wall-normal ones, at least when compared to those in the lower
branch. Their mean and fluctuating velocities are reminiscent of experimental tur-
bulence [25, 42], and so are other properties. For example, the range of spanwise
wavelengths in which the nonlinear solutions exist is in the neighbourhood of the
observed spacing of the streaks of the sublayer [22]. ‘Lower’ solutions have stronger
and essentially straight streaks, and much weaker vortices. Their statistics are very
different from turbulence.
The near-wall statistics of full turbulent flows, when compiled over scales cor-
responding to a single streak and to a single vortex pair, are independent of the
Reynolds number, and agree reasonably well with those of the fixed points, although
there is a noticeable contribution from unsteady bursting [22]. When they are com-
piled over much larger boxes, however, the intensity of the fluctuations does not
scale well in wall units, even very near the wall [7]. That effect is due to large outer-
flow velocity fluctuations reaching the wall, and is unrelated to the structures being
considered here.
This is shown in figure 2(a), which contains two-dimensional spectral energy
densities of the streamwise velocity in the buffer layer, kx kz Euu (kx , kz ), displayed
as functions of the streamwise and spanwise wavelengths. The three spectra in the
figure correspond to turbulent channels at different Reynolds numbers. They differ
from each other almost exclusively in the long and wide structures represented in the
upper-right corner of the spectrum, whose sizes are of the order of λx × λz = 10h × h
[8, 21, 17]. The lower-left corner of the spectral plane contains the structures dis-
cussed in this section, which are very approximately universal and local to the
8 J. Jiménez
1
10
3
10
0
10
λ+
z
λz/h
2
10
−1
10
(a) (b)
2 3 4 −1 0 1
10 10 λ+ 10 10 10 λ /h 10
x x
Fig. 2 Two-dimensional spectral energy density of the streamwise velocity, in terms of the
streamwise and spanwise wavelengths. (a) y+ = 15. (b) y/h = 0.2. Numerical channels [8,
11, 17]. , h+ = 547; , 934; , 2003. Spectra are normalised in wall units,
and the two contours for each spectrum are 0.11 and 0.55 times the maximum of the spectrum
for the highest Reynolds number. The heavy diagonal is λz = 0.15λx , and the triangles are
λx = 10h for the three cases
near-wall layer. The larger structures in the top-right corner extend into the log-
arithmic layer, scale in outer units, and correspond approximately to Townsend’s
‘attached eddies’ [40]. An example of their spectra away from the wall is shown in
figure 2(b).
2.2 The Logarithmic Layer

The logarithmic layer is located just above the viscous region, and it is also unique to
wall turbulence. Most of the velocity difference that does not reside in the near-wall
region is concentrated in the logarithmic layer, which extends experimentally up to
y ≈ 0.2h (figure 1b). It follows from the profile in (1) that the velocity difference
above the logarithmic layer is only 20% of the total when h+ = 200, and decreases
logarithmically as the Reynolds number increases. In the limit of infinite Reynolds
number, all the velocity drop is in the logarithmic layer.
This layer is an intrinsically high-Reynolds number phenomenon, which requires
at least that its upper limit should be above the lower one, so that 0.2h+ 150,
and h+ 750. It has been studied experimentally for a long time, but numerical
simulations with even an incipient logarithmic region have only recently become
available [11, 17, 13]. It is much worse understood than the viscous layers, but the
new simulations, together with simultaneous advances in experimental methods,
have greatly improved our knowledge of the kinematics of the structures in this
region, and are beginning to hint at models for their dynamics.
It is important to note at this point that the meaning of the word ‘model’ is prob-
ably different in the logarithmic and in the buffer layers. Near the wall, the local
Reynolds numbers are low, and the structures are smooth and essentially analytic.
It is then possible to speak of ‘objects’, and to write differential equations for their
behaviour. Above the buffer layer both things are harder to do. Since the definition
of the outer layer includes y+ 1, its largest structures have high internal Reynolds
numbers, and are turbulent themselves. There is presumably a cascade connect-
ing those energy-containing structures with the dissipative scales, and their velocity
fields can be expected to have nontrivial algebraic spectra and non-smooth geome-
tries that can only be described statistically. They are ‘eddies’, rather than ‘vortices’,
because turbulent vorticity always resides at the viscous Kolmogorov lengthscale η ,
separated from the energy-containing eddies by a scale ratio O(y+ ).
3/4
While the models for the buffer layer are in the realm of direct numerical sim-
ulations (DNS), the outer layers are the domain of large-eddy simulations (LES).
This of course does not mean that the logarithmic layer can not be DNSed, and it
is almost certain that more direct simulations will be required before this part of
the flow is understood, but we can probably only expect simple models for partial
aspects of the structures involved, rather than full ones including all the flow scales.
The first new information provided by the numerics on the logarithmic layer was
spectral. It had been found experimentally that there are very large scales in the
outer regions of turbulent boundary layers [19, 27], and DNS provided informa-
tion about their two-dimensional spectra, and about their wall-normal correlations
[8, 11]. The longest scales are associated with the streamwise velocity component.
Its spectral density in the logarithmic layer has an elongated shape along the line
λz2 = yλx , while the two other velocity components are more isotropic. When three-
dimensional flow fields became available, it was found that there is a self-similar
hierarchy of compact ejections extending from the outer flow into the buffer layer,
within which the coarse-grained dissipation is more intense than elsewhere [12].
They correspond to the isotropic spectra of the wall-normal velocity. When the flow
is conditionally averaged around them, they are associated with long, conical, low-
velocity regions in the logarithmic layer [12], whose intersection with a y-plane is
parabolic, explaining the quadratic behaviour of the spectrum of u. These structures
are not only statistical constructs. Individual cones are observed as low-momentum
‘ramps’ in streamwise sections of instantaneous flow fields [29].
When the cones reach heights of the order of the flow thickness, they stop grow-
ing, and become cylindrical ‘streaks’ spanning the distance from the central plane
to the wall [8, 11], similar to those of the sublayer, but with spanwise scales of
2 − 3h. They are fully turbulent objects. Neither in simulations nor in experiments
in channels or pipes has it been possible to determine the maximum length of those
‘global modes’, which appear in any case to be longer than 25h [11, 18]. There is
however some evidence that they may be shorter in boundary layers. The overall
arrangement of the ejections and cones is reminiscent of the association of vortices
and streaks in the buffer layer, but at a much larger scale. Their near-wall footprints
are seen in the spectra of the buffer layer as the ‘tails’ in figure 2, and account [17]
for the experimentally-observed Reynolds number dependence of the intensity of
the near-wall velocity fluctuations [7].
Since we saw above that the sublayer streaks originate from the advection of the
mean shear by cross-stream perturbations, which is a linear process, there is some
hope that a linear model could also capture the formation of the outer-layer streaks.
10 J. Jiménez
The mean velocity profile of turbulent channels is linearly stable [34], but it has
been known for some time that even stable flows can lead to large transient energy
amplifications, because the evolution operator of the linearised Navier–Stokes equa-
tions is not self-adjoint [15, 4]. Simple linearised analysis of a uniform shear shows
that the long-time asymptotic state of any localised perturbation is a u-streak, but
it provides no wavelength-selection mechanism. When the analysis is repeated for
nontrivial profiles, such as laminar or turbulent channels, the flow thickness provides
a lengthscale and a wall-normal modal structure. The key modelling assumption to
obtain structures mimicking the turbulent eddies of the logarithmic and outer re-
gions appears to be the use of a y-dependent eddy viscosity similar to that required
to maintain the experimental mean profile [33, 9]. Note that this implies that the
resulting model applies to averaged eddies, rather than to individual structures. It
turns out that there are two sets of wavelengths for which the total energy is most
amplified, with eigenfunctions peaking at the two locations where the viscosity does
not depend on y. Near the wall, where the viscosity is mostly molecular, they have
spanwise wavelengths and eigenfunctions similar to the observed sublayer streaks.
Near the central plane, where νT ≈ uτ h is also roughly uniform, they are large-scale
streaks with spanwise wavelengths of the order of the observed 3h, and wall-normal
eigenfunctions that agree well with the dominant proper orthogonal decomposition
eigenmodes of the streamwise velocity at those wavelengths.
3 The Direction of Causality

We know less about how the ejections are created, but linear analysis also gives
some information on them. In the same way as the linear effect of transverse per-
turbations is to create transient u-streaks, any perturbation of u that is not infinitely
long transfers energy into the transverse velocity components. The same transient-
growth analysis giving the large-scale streaks contains nontrivial amplifications for
v and w, which could in principle feed a linear cycle in which v ejections create
streaks by extracting energy from the mean shear, while the streaks in turn create
ejections. Unfortunately the wavelengths of both processes are different, which is
why the profile is linearly stable. The most amplified u-structures are streaks elon-
gated along x, while the most amplified v and w are roughly isotropic in the wall-
parallel plane. This agrees with the spectral evidence, but means that nonlinearity is
required to match the wavelengths, and to close the cycle. It is however easy to visu-
alise a process by which an ejection creates a strong streak, whose enveloping shear
layer becomes unstable and creates new, shorter ejections. In fact, we have seen that
compact ejections can be identified at all scales in the logarithmic and outer layers,
both numerically and experimentally, and that they are associated with streaks. It
is known, from the analysis of their relative lengths and lifetimes, that the observed
ejections cannot be the origin of the full length of the streak to which they are associ-
ated, and that some causal link from streaks to ejections is also required [9]. In fact,
numerical experiments with ‘minimal’ simulation boxes of the order of the channel
width, several thousand wall units long and wide, show evidence of an ‘outer’ cycle
1 0
10 10
(a) (b)
0
10
y/h
λ /h
−1
10
z
−1
10
−2
−1 0 1
10 −1 0 1
10 10 10 10 10 10
λ /h λz/h
x
Fig. 3 Contour plot of the coherence coefficient of the streamwise velocity fluctuations.
Numerical channel at Reτ = 934 [11]. (a) As a function of both wavelengths at y+ = 20. (b)
Maximum over λz , as a function of the λx and y. The isolines are, –♦ , γ = 0.25(0.1)0.65.
The shaded contours are the premultiplied energy density of the streamwise velocity,
normalised at each wall distance with the fluctuation energy. The horizontal or vertical dashed
lines correspond to λz = 0.5h and λx = 3h
of the largest flow scales, with a characteristic period (tuτ /h ≈ 2.5, t + ≈ 4500), that
appears to be unrelated to any near-wall process [14].
There is clear evidence that this outer structures are coherent, and influence the
wall. Figure 3 shows the correlation coefficient between the temporal and spatial
derivatives of u,
∂ t u ∂ x u
γu = , (2)
[(∂ t u)2 (∂ x u)2 ]1/2
which measures how far is the velocity field from being a coherent wave [10].
Figure 3(a) shows that the only wave-like behaviour in the buffer layer corresponds
to structures that are much larger than the local turbulent cycle, while 3(b) shows
that they are reflections of structures residing near the centre of the channel.
The outer structures clearly modify the inner layer, although most of the mod-
ification is probably trivial. It is clear from figure 3 that the sizes of both sets of
structures are very different. Even at the relatively modest Reynolds number of the
figure the scale ratio is roughly 20, and it increases linearly with the Reynolds num-
ber. The near-wall streaks effectively live in a local boundary layer whose friction
velocity and mean flow direction is defined by the outer structures. For example, one
could think of the local atmospheric surface layer, in which streaks with lengths of
centimetres live within drafts several hundred meters long. When this effect is dis-
counted, most of the effect of the Reynolds number disappears. For example, it was
shown in [22] that, when the velocity fluctuation intensities are computed over local
boxes of the size of a minimal flow unit, with respect to the local mean and nor-
malised with the local friction velocity, their probability density functions become
independent of the Reynolds number. The same is true for larger averaging boxes,
as long as their sizes are scaled in wall units. The direct relation between higher
12 J. Jiménez
local friction velocities and higher intensities reverses away from the wall, but even
there the effect seems to be fairly straightforward. The local friction velocity is just
a reflection of the local wall shear, and whenever it is low it is because the mean
flow has been lifted away from the wall. This implies that there has to be a higher
shear somewhere above, since the bulk velocity is constant, and those high-shear
regions can be expected to give rise to locally higher production and fluctuation in-
tensities. This can be shown to be true, and the correlation coefficient between local
fluctuation intensities and local shear is positive, both at the wall and away from it.
The scenario just described is mostly derived from simulations, and from the
linear analysis of the averaged equations of motion. A different scenario has been
proposed from the observation of experimental flow fields. In it, the basic object
is a hairpin vortex growing from the wall, whose induced velocity creates the low
momentum ramps mentioned above [1]. In that model, which was motivated by the
behaviour of hairpin vortices in the numerical simulation of a particular laminar
velocity profile [43], the hairpins regenerate each other, creating vortex packets that
are responsible for the very long observed streaks [5]. While the two models look
very different at first sight, they can probably be reconciled to a certain extent, with
vortex packets corresponding to the instabilities of the shear layer around the streak.
The main formal difference between the two models is their respective empha-
sis on vortices and eddies, although that might be largely a matter of notation. A
more serious difference is the treatment of the effect of the wall. The ‘numerical’
model emphasises the effect of the local velocity shear rather than the presence of
the wall, and a top-down flow of causality. The ‘experimental’ one appears to re-
quire the formation of the hairpins in the buffer region, and bottom-up causality.
That could again be a matter of notation, but it is more likely due to the reliance of
the experimental model on laminar numerical simulations, using molecular viscos-
ity [43]. There is little question that large structures in turbulence feel the effect of
smaller ones [33]. While the modelling of this randomising effect as a simple eddy
viscosity can be criticised, it should be much closer to reality than the much weaker
molecular dissipation of a laminar environment. When the linear evolution of an
initially compact ejection is analysed using the eddy viscosity mentioned above,
the structures created near the wall do not grow very much, and most ejections
observed at a given wall distance have to be created ‘locally’ [12]. Indeed, numeri-
cal experiments in which the viscous wall cycle is artificially removed, have outer-
flow ejections and streaks that are essentially identical to those above smooth walls
[13, 30]. Experimentally, this is equivalent to the classical observation that the outer
layers in turbulent boundary layers are independent of wall roughness [20].
In summary, there is evidence for at least two regeneration cycles in wall-
bounded turbulence, broadly similar, but acting in the buffer and outer layers re-
spectively. While the former creates the strongest fluctuations, the latter spans the
whole boundary layer, and contains most of the energy. There is clear evidence for
the effect of the outer cycle on the inner one, mostly as a modulation of its intensity.
Essentially, the inner cycle ‘lives’ in the local boundary layers of the outer scales.
There is also evidence of the triggering of the outer cycle by the inner one, although
limited up to now to relatively low Reynolds numbers which leave open the question
of whether that interaction would survives as the range of scales involved gets really
large. Although it would be strange if absolutely no inner-outer interaction exists,
the nature of the trigger is not well understood, and contradicts the evidence that
essentially similar outer regions coexist with very different inner ones, such as over
rough walls.
Acknowledgements. The preparation of this paper was supported in part by the CICYT
grant TRA2006–08226. I am deeply indebted to J.C. del Álamo, O. Flores, S. Hoyas, G.
Kawahara and M.P. Simens for providing most of the data on which this discussion is based.
References
1. Adrian, R.J., Meinhart, C.D., Tomkins, C.D.: Vortex organization in the outer region of
the turbulent boundary layer. J. Fluid Mech. 422, 1–54 (2000)
2. Bakewell, H.P., Lumley, J.L.: Viscous sublayer and adjacent wall region in turbulent pipe
flow. Phys. Fluids 10, 1880–1889 (1967)
3. Brown, G.L., Roshko, A.: On the density effects and large structure in turbulent mixing
layers. J. Fluid Mech. 64, 775–816 (1974)
4. Butler, K.M., Farrell, B.F.: Three-dimensional optimal perturbations in viscous shear
flow. Phys. Fluids A 4, 1637–1650 (1992)
5. Christensen, K.T., Adrian, R.J.: Statistical evidence of hairpin vortex packets in wall
turbulence. J. Fluid Mech. 431, 433–443 (2001)
6. Darcy, H.: Recherches expérimentales rélatives au mouvement de l’eau dans les tuyeaux.
Mém. Savants Etrang. Acad. Sci. Paris 17, 1–268 (1854)
7. de Graaff, D.B., Eaton, J.K.: Reynolds number scaling of the flat-plate turbulent bound-
ary layer. J. Fluid Mech. 422, 319–346 (2000)
8. del Álamo, J.C., Jiménez, J.: Spectra of very large anisotropic scales in turbulent chan-
nels. Phys. Fluids 15, L41–L44 (2003)
9. del Álamo, J.C., Jiménez, J.: Linear energy amplification in turbulent channels. J. Fluid
Mech. 559, 205–213 (2006)
10. del Álamo, J.C., Jiménez, J.: Estimation of turbulent convection velocities and correc-
tions to taylor’s approximation. J. Fluid Mech. (2009) (submitted)
11. del Álamo, J.C., Jiménez, J., Zandonade, P., Moser, R.D.: Scaling of the energy spectra
of turbulent channels. J. Fluid Mech. 500, 135–144 (2004)
12. del Álamo, J.C., Jiménez, J., Zandonade, P., Moser, R.D.: Self-similar vortex clusters in
the logarithmic region. J. Fluid Mech. 561, 329–358 (2006)
13. Flores, O., Jiménez, J.: Effect of wall-boundary disturbances on turbulent channel flows.
J. Fluid Mech. 566, 357–376 (2006)
14. Flores, O., Jiménez, J.: The minimal logarithmic region. In: Proc. Div. Fluid Dyn.,
pp. AE–04. Am. Phys. Soc. (2007)
15. Gustavsson, L.H.: Energy growth of three-dimensional disturbances in plane Poiseuille
flow. J. Fluid Mech. 224, 241–260 (1991)
16. Hagen, G.H.L.: Über den Bewegung des Wassers in engen cylindrischen Röhren.
Poggendorfs Ann. Physik Chemie 46, 423–442 (1839)
17. Hoyas, S., Jiménez, J.: Scaling of the velocity fluctuations in turbulent channels up to
Reτ = 2003. Phys. Fluids 18, 011702 (2006)
18. Hutchins, N., Marusic, I.: Evidence of very long meandering features in the logarithmic
region of turbulent boundary layers. J. Fluid Mech. 579, 467–477 (2007)
14 J. Jiménez
19. Jiménez, J.: The largest scales of turbulence. In: CTR Ann. Res. Briefs, pp. 137–154.
Stanford Univ., Stanford (1998)
20. Jiménez, J.: Turbulent flows over rough walls. Ann. Rev. Fluid Mech. 36, 173–196 (2004)
21. Jiménez, J., del Álamo, J.C., Flores, O.: The large-scale dynamics of near-wall turbu-
lence. J. Fluid Mech. 505, 179–199 (2004)
22. Jiménez, J., Kawahara, G., Simens, M.P., Nagata, M., Shiba, M.: Characterization of
near-wall turbulence in terms of equilibrium and ‘bursting’ solutions. Phys. Fluids 17,
015105 (2005)
23. Jiménez, J., Moin, P.: The minimal flow unit in near-wall turbulence. J. Fluid Mech. 225,
221–240 (1991)
24. Jiménez, J., Pinelli, A.: The autonomous cycle of near wall turbulence. J. Fluid
Mech. 389, 335–359 (1999)
25. Jiménez, J., Simens, M.P.: Low-dimensional dynamics in a turbulent wall flow. J. Fluid
Mech. 435, 81–91 (2001)
26. Kim, H.T., Kline, S.J., Reynolds, W.C.: The production of turbulence near a smooth wall
in a turbulent boundary layer. J. Fluid Mech. 50, 133–160 (1971)
27. Kim, K.C., Adrian, R.J.: Very large-scale motion in the outer layer. Phys. Fluids 11,
417–422 (1999)
28. Kolmogorov, A.N.: The local structure of turbulence in incompressible viscous fluids a
very large Reynolds numbers. Dokl. Akad. Nauk. SSSR 30, 301–305 (1941); Reprinted
in Proc. R. Soc. London. A 434, 9–13 (1991)
29. Meinhart, C.D., Adrian, R.J.: On the existence of uniform momentum zones in a turbu-
lent boundary layer. Phys. Fluids 7, 694–696 (1995)
30. Mizuno, Y., Jiménez, J.: Wall turbulence without walls. In: Proc. Div. Fluid Dyn., pp.
AA–02. Am. Phys. Soc. (2008)
31. Nagata, M.: Three-dimensional finite-amplitude solutions in plane Couette flow:
bifurcation from infinity. J. Fluid Mech. 217, 519–527 (1990)
32. Österlund, J.M., Johansson, A.V., Nagib, H.M., Hites, M.: A note on the overlap region
in turbulent boundary layers. Phys. Fluids 12, 1–4 (2000)
33. Reynolds, W.C., Hussain, A.K.M.F.: The mechanics of an organized wave in turbu-
lent shear flow. Part 3. Theoretical models and comparisons with experiments. J. Fluid
Mech. 54, 263–288 (1972)
34. Reynolds, W.C., Tiederman, W.G.: Stability of turbulent channel flow, with application
to Malkus’ theory. J. Fluid Mech. 27, 253–272 (1967)
35. Richardson, L.F.: The supply of energy from and to atmospheric eddies. Proc. Roy. Soc.
A 97, 354–373 (1920)
36. Robinson, S.K.: Coherent motions in the turbulent boundary layer. Ann. Rev. Fluid
Mech. 23, 601–639 (1991)
37. Smith, C.R., Metzler, S.P.: The characteristics of low speed streaks in the near wall region
of a turbulent boundary layer. J. Fluid Mech. 129, 27–54 (1983)
38. Swearingen, J.D., Blackwelder, R.F.: The growth and breakdown of streamwise vortices
in the presence of a wall. J. Fluid Mech. 182, 255–290 (1987)
39. Tennekes, H., Lumley, J.L.: A first course in turbulence. MIT Press, Cambridge (1972)
40. Townsend, A.A.: The structure of turbulent shear flow, 2nd edn. Cambridge U. Press,
Cambridge (1976)
41. Waleffe, F.: Three-dimensional coherent states in plane shear flows. Phys. Rev. Lett. 81,
4140–4143 (1998)
42. Waleffe, F.: Homotopy of exact coherent structures in plane shear flows. Phys. Fluids 15,
1517–1534 (2003)
43. Zhou, J., Adrian, R.J., Balachandar, S., Kendall, T.M.: Mechanisms for generating
coherent packets of hairpin vortices in channel flow. J. Fluid Mech. 387, 353–396 (1999)
Turbulence Interaction with Atmospheric
Physical Processes
Chin-Hoh Moeng and Jeffrey Weil
Abstract. This article reviews the planetary-boundary-layer (PBL) turbulence and

its interactions with atmospheric processes. We show three examples: turbulence
response to surface heating and cooling over lands, effects of ocean waves, and in-
teractions with radiation and cloud microphysics. We also show how computational
fluid dynamics methods are used to gain fundamental understanding of these inter-
actions mostly under idealized environments. For certain practical applications in
which idealized conditions may not apply, a brute-force method may be needed to
explicitly simulate the turbulence interaction. One way is to nest a large-eddy sim-
ulation domain inside a weather forecast model, and to allow for turbulence feed-
back to other physical processes. This numerical method sounds straightforward
but poses two major problems. We suggest a systematic approach to examine the
problems.
1 Introduction
Turbulence is a difficult scientific problem because of its highly nonlinear nature.
Turbulence poses an even bigger challenge when it interacts with atmospheric phys-
ical processes such as the diurnal heating/cooling cycle, ocean waves, clouds, and
weather events. There are a variety of turbulence regimes in the atmosphere, dif-
fering in scales, characteristics, and the sources of turbulent kinetic energy (TKE).
For example, geostrophic turbulence, which results from nonlinear advection, the
Earth’s rotation and temperature stratification, consists of ”pancake vortices” that
Chin-Hoh Moeng
National center for Atmospheric Research, PO Box 3000, Boulder, CO, 80307-3000, USA
e-mail: moeng@ucar.edu
Jeffrey Weil
Cooperative Institute for Research in Environmental Sciences, University of Colorado,
Boulder, CO 80309
e-mail: weil@ucar.edu
16 C.-H. Moeng and J. Weil
bear more resemblance to two-dimensional turbulence than three-dimensional (3D)

turbulence. Other atmospheric turbulence regimes are more 3D in nature and much
smaller in scale, including clear-air turbulence (due to wind shear near, e.g., the
upper-level jet streams), turbulence inside deep clouds or storms, and turbulence in
the planetary boundary layer (PBL).
In this paper, we will focus only on turbulence in the PBL. The PBL is a distinct
turbulent layer adjacent to the Earth’s surface. It is a layer that directly feels the im-
pact of the Earth’s surface conditions such as roughness, surface heating/cooling,
etc. By definition, the PBL is always turbulent and its turbulent motion is al-
ways three-dimensional and time evolving. The PBL depth varies depending on
the sources and sinks of its TKE. On a calm night, when the wind is weak and the
surface is colder than the air (negative buoyancy force), the PBL may become very
shallow or even collapse.
The PBL is an effective medium for transporting and diffusing heat, moisture
and chemical species. On a bad pollution day, one can clearly see the smog layer,
particularly when it is capped by a strong temperature inversion. Because the major
sources of water, heat and biogeochemical species reside on the Earth’s surface, the
PBL plays a crucial role in the energy, water, and biogeochemical cycles on Earth.
Every weather or climate model requires a turbulence scheme to represent the effect
of the PBL turbulence.
PBL turbulence differs from that of laboratory turbulence in some major ways.
First the Reynolds number of turbulence in the PBL is extremely high, on the or-
der of 109 . The PBL turbulence covers a wide range of scales from millimeters to
kilometers with a velocity scale on the order of few ms−1 . The Reynolds number
is so high that we consider it infinite. Therefore, the molecular viscosity and diffu-
sivity terms are omitted in the governing equations of atmospheric models, includ-
ing LESs (Large Eddy Simulations). Close to the Earth’s surface, the viscous layer
(order of centimeters) is too thin to resolve in any atmospheric model; the first grid
layer often resides in the inertial, logarithmic layer.
Second, the PBL turbulence is almost always affected by buoyancy or temper-
ature stratification, be it positive or negative. There is seldom a purely neutral,
shear-driven PBL outdoors.
Another major difference is the entraining top of the PBL. Unlike wind tunnels
bounded by two walls, one side of the PBL is a free entraining surface. The entrain-
ing interface divides the turbulent air (below) from the non-turbulent air (above).
How PBL turbulence controls entrainment is still poorly understood. Entrainment,
which is not a major topic for traditional turbulence research, is one of the most
difficult–and emphasized—topics in the PBL research.
2 PBL Turbulence and Its Interactions

Turbulence in the PBL is never isotropic and homogeneous due to interactions with
the surface below, the free (non-turbulent) atmosphere above, radiation, or clouds.
These interactions significantly change the behavior of energy-containing eddies
Turbulence Interaction with Atmospheric Physical Processes 17
(i.e., eddies larger than tens or hundreds of meters), and these variations make it
impossible to study or quantify the PBL turbulence with a universal scaling law or
turbulence scheme. Therfore, the PBL is typically classified into several regimes ac-
cording to its interactions or meteorological conditions. In this article we will review
some turbulence interactions, and show how computational fluid dynamics methods,
such as DNS (Direct Numerical Simulation) and LES (Large Eddy Simulation), are
used to study these interactions.
There are two kinds of interactions depending on the time (or spatial) scale of the
physical process that interacts with the turbulence. The turbulence eddy-turnover
time in the PBL is on the order of several mimutes to several tens of minutes de-
pending on the PBL regimes. If the time scale of the physical process is much longer
than that of the turbulence eddy-turnover time scale, turbulence can attain equilibrim
with the physical process before the process changes. This interaction establishes a
statistically quasi-steady turbulence regime with a given process (or meteorological
condition). We will call this an interaction of first kind.
The second kind of interaction is when the time (or spatial) scale of the physical
process is similar to or even shorter than the turbulence scale. Interaction takes place
on a time scale where both the turbulence and the process keep adjusting to each
other. Unless otherwise specified, the ”interaction” mentioned in this article belongs
to the second kind.
2.1 Interaction with Land Processes and Diurnal Cycle

The diurnal cycle of the PBL is a result of the first kind of interaction with land.
When the sun heats the ground, the surface buoyancy flux becomes positive and
convectively drives turbulence in the PBL. This PBL regime is classified as the
convective PBL (CBL) or the unstable PBL. The intensity of convectively driven
turbulence is often very strong so that all conserved variables such as potential tem-
perature, water vapor mixing ratio (no cloud cases) or air pollutants are well mixed
in the CBL. The transport is carried mostly by large turbulent thermals on the size of
the CBL depth (which is on the order of 1 kilometer), and hence is very efficient and
non-local. These thermals carry heat or moisture from the ground to the top of the
CBL within just a few minutes. In the CBL, species originating near the surface can
diffuse upwards faster than those from elevated sources diffuse downward, known
as asymmetric transport. The non-local and asymmetric transport properties in the
CBL are quite distinct from those of shear-driven turbulence.
After sunset, the ground becomes colder than the air, the buoyancy term in the
TKE budget becomes negative, and the PBL decreases abruptly. Shear becomes the
only source of TKE competing with the negative buoyancy force which consumes
turbulence. The turbulence intensity in the stable PBL is therefore much weaker
than that in the CBL.
The term ”very stable PBL” applies to the condition when shear production and
buoyancy consumption of TKE are comparable in magnitude. When buoyancy con-
sumption dominates, turbulence becomes very weak and may collapse. Weak or
no mixing enhances wind shear locally which in turn overcomes the buoyancy
consumption, and turbulence is re-generated. Turbulence in the very stable PBL is

thus intermittent in time and space; its properties, depending strongly on the phys-
ical processes with which it interacts, may be too variable to describe statistically
(see review by [6]). The intermittency and weak TKE also pose a challenge for LES.
Another turbulence regime over land is canopy turbulence where turbulence from
the PBL and a vegetation canopy interact. Turbulent downbursts in the PBL inject
fresh air into the canopy layer generating coherent eddies inside the canopy, which
affects the canopy turbulence. At the same time relatively humid air from the tran-
spiring vegetation is expelled from the canopy and changes the turbulence above. We
will not elaborate on this topic, but refer those who are interested to [4, 5, 14, 11].
2.2 Interaction with Ocean Waves

Turbulence on both sides of the ocean [the PBL above and the oceanic boundary
layer (OBL) below] interacts strongly with surface gravity waves. Surface waves
are driven by surface winds, and depending on the wind forcing, waves form with
different wavelengths and wave heights, propagate at different speeds, and break—
sometimes producing sea spray and bubbles. In turn, waves modify winds in the
PBL and induce elongated streamwise vortices (Langmuir circulations) in the OBL,
thus changing the turbulence. Studies (e.g., [3]) have suggested that hurricane inten-
sity and track are sensitive to the turbulence-wave interaction, which remains poorly
understood and under-studied.
At NCAR, we started out by developing a DNS to simulate turbulent flow over
idealized water waves [17]. Figure 1 shows a sketch of the DNS domain for a Cou-
ette flow over a moving wavy boundary to mimic a shear-driven PBL over a wavy
ocean. This study shows significant influences of the imposed wave on the turbu-
lence statistics; the influence depends on the wave age (c/u∗ where c is the wave
phase speed and u∗ is the friction velocity) and the wave slope (aκ where a is the
wave amplitude and κ the wave number). Figure 2 displays the wave effect on mo-
mentum transport, showing that waves moving slower than a certain value of c/u∗
generate a more negative momentum flux compared to the flat surface case, while
fast moving waves generate the opposite effect. Very close to the surface all moving
waves lead to a more negative momentum flux due to the wave-correlated flux.
The above DNS has been extended to LES to investigate the effects of wave
breaking and Langmuir circulations on the OBL ([18]) and the effects of mov-
ing waves on the PBL ([19]). This LES code is now being extended to include
more realistic waves with multiple wavelengths (personal communication with
Peter Sullivan).
The above studies prescribe the wavy surface and hence provide only a one-
way interaction of waves with turbulence. A future challenge is to allow for the
wavy surface to respond to the turbulence. This two-way interaction study requires
better understanding of how waves respond to turbulence, which is lacking, and a
numerical technique that provides a time-varying surface coordinate.
Fig. 1 Sketch of three-dimensional Couette flow drivn by U0 over a moving wavy surface,
from Sullivan et al (2000)
Fig. 2 Profiles of the vertical momentum flux for various c/u∗ : (a) < uw > /u2∗ , and (b)
(< uw > − < uw > f lat )/u2∗ , from Sullivan et al (2000)
2.3 Interaction with Radiation and Cloud Microphysics

When air inside the PBL becomes saturated, it forms a layer of cloud in the PBL (or
fog if the near-surface air is also saturated). An example is the stratocumulus-topped
PBL (STBL) regime over the east coast of major continents during the Northern-
Hemisphere summer season. This cloud sheet reflects much more sunlight back
to space than the ocean does, while it emits about the same amount of longwave
radiation as the ocean. As a result, the STBL acts to reduce the solar energy into the
Earth system. [12] pointed out that a mere 4% increase in the areal coverage of this
cloud regime could produce a 2-3 K decrease of the global mean temperature, which
is enough to compensate for the global warming due to a doubling of the carbon
dioxide. That is why this cloud regime has brought attention to the geoengineering
community (e.g., [13]).
The cloud amount of the STBL depends strongly on the interactions among
turbulence, radiation and cloud microphysics ([7, 8])—in a time shorter than the
eddy turnover time. This PBL regime is unique in that turbulence is driven mainly
by longwave radiative cooling at the cloud top. (Longwave radiative cooling at
cloud top occurs because cloudy air emits more longwave radiation than non-
cloudy air does, which produces a sharp radiative flux jump right at the cloud top.)
Cooling-from-above can generate a positive buoyancy forcing like heating-from-
below. However, this cooling occurs at a free interface where the entrainment oc-
curring there greatly complicates the problem. Entrainment can warm and cool the
air near the cloud top at the same time. When the inversion air is entrained and
mixes with the cloudy air, the temperature of the mixture can increase solely by
mixing, but it can also decrease due to evaporative cooling of cloud droplets. The
net result of these two competing effects depends on the temperature and moisture
gradients across the STBL top, and this net result can significantly modify the buoy-
ancy production for the TKE. So far we cannot represent the STBL regime properly
in climate models because we do not understand quantatively how these interactions
change the cloud amount and turbulent intensity.
Many research groups worldwide have been studying this PBL regime through
field measurements (e.g., [12, 15]) and LESs. To see if LES is capable of simulating
this complicated PBL regime, ten groups of LES researchers performed an inter-
comparison study of the STBL regime in the mid-1990s ([10]); the result showed
a wide spread of the predicted entrainment rates among the ten LES codes. This
study brought attention to the entrainment-rate issue and motivated several field
campaigns ([15]) and idealized LES studies (e.g., [1]), designed specifically for bet-
ter understanding of the entrainment processes. However, a more recent intercom-
parison study ([16]) suggests that we have gained little improvement in our ability
to simulate the entrainment rate in the STBL. The study suggests that the LES tech-
nique may be too sensitive to the SGS processes near the cloud top to properly rep-
resent the complicated interactions among turbulence, entrainment, radiation and
cloud microphysics. The problem is amplified in certain environments where the
air above the STBL is particularly dry such that the entrained air may destablize
the cloud layer and cause it to break up. As far as turbulence interaction with
atmospheric physical processes is concerned, this PBL regime provides the best
example of the interaction complexity.
3 Turbulence in Complex, Real-World Environments

Most of the PBL LESs have been limited to idealized physical conditions, e.g.,
horizontally uniform or periodic wavy surfaces. Those conditions are ideal for gain-
ing fundamental understanding of turbulence interactions without the real-world
complexity. However, some practical PBL applications require real-world settings.
For example, to select a wind farm site we need to know the local terrain and weather
effects because they can significantly affect the velocity fluctuations at the wind-
turbine height. Air pollution is another example where the local terrain and meteo-
rological conditions can modify the PBL turbulence and influence the likelihood of
an extreme (high) concentration event.
A brute-force method to study turbulence interactions with real-world physical
processes is to explicitly resolve all relevant scales by using a turbulence-resolving
grid over the whole domain that covers all scales of the physical processes. However,
this may become computationally impossible if the largest scale of the physical
processes is much larger than the turbulence scale. One way to solve this problem
is to nest an LES domain for the region of interest inside a weather forecast model.
A typical weather forecast model consists of realistic terrain geometry and realistic
meteorological conditions, but its grid resolution is too coarse to resolve turbulent
motions; it uses a RANS (Reynolds Averaged Navier-Stokes) model to represent the
statistics of PBL turbulence. Thus, a weather model alone is not a proper tool for
studying turbulence interactions.
The idea of nesting an LES inside a weather model is not new. For example, the
study by [2] nests an LES domain over the southern Swiss Alps in a weather model
to simulate PBL over steep, mountainous terrain. Their approach allows for just
one-way (downscaling) interaction where the mesoscale model prediction drives
the LES flow. However, the effect of the resolved turbulence field from LES does
not feedback to the mesoscale prediction.
Two-way nesting allows for two-way interactions between turbulence and
mesoscale processes. Two-way nesting uses the outer-domain mesoscale flow field
to provide the flow conditions along the nest boundaries for the inner-domain LES,
and at the same time uses the spatially averaged LES flow field to overwrite the
outer-domain solution in the overlapped region.
The two-way nesting concept is straightforward, but there are two major issues
that need to be examined. First is the spin-up issue: inside the outer domain where
a RANS model is used, the resolved flow field there is Reynolds averaged. Thus,
the flows along the nest boundaries (from the outer domain) are non-turbulent mo-
tions. Non-turbulent inflow takes time to spin up to fully developed turbulence in
models because of grid discretization. Based on our experience in performing PBL
LESs, the spin-up time from an initial non-turbulent field to a fully developed
turbulent flow is 30 minutes or more depending on the PBL regimes (e.g., [10]).
For a mean wind of 5 ms−1 , the distance for an inflow to spin up to turbulence in-
side the LES domain is at least 9 km, suggesting that the nest LES domain has to
be much larger than 9 km in the downstream direction. The spin-up issue becomes
particularly problematic over complex terrain where the wind directions can change
quickly on time scales comparable to the eddy turnover time, and spin-up regions
may exist in all directions.
The second issue concerns the differences of SGS physics between mesoscale
and LES domains. Turbulent fluxes estimated from any RANS closure model are
unlikely to be the same as those simulated in the LES domain. The difference in
turbulent transport is likely to create a systematic bias of, for example, the mean
wind or temperature field between the two domains. This would generate mean
gradients, and hence artificial circulations, across nest boundaries.
As a start, we investigated these issues in a simple setting. In [9] we designed
an idealized two-way nesting experiment where a fine-grid LES is nested inside
a coarse-grid LES. The same forcing and surface conditions are applied to both
LES domains, but the lateral boundary conditions of the fine-grid LES come from
the coarse-grid LES solution. We then asked if both LESs can generate the same
turbulence statistics even though the scale ranges they resolve are different. The
study shows that for a shear-driven PBL the Smagorinsky-Deardorff type of SGS
scheme leads to a systematic bias in the near-surface wind and the surface stress
(not shown here) between the two LESs. Fig. 3a. compares the resolved (thick
curves) and SGS (thin curves) momentum fluxes, which clearly show that near the
surface the total momentum flux (the sum) is much smaller in the inner domain
(dotted curves) than in the outer domain (solid curves). The bias is due to the grid-
dependent Smagorinsky-Deardorff SGS model: Very close to the surface where SGS
contribution dominates, the outer-domain LES with a larger grid produces a larger
SGS momentum flux than the inner-domain LES. This larger SGS flux cannot be
compensated for by the resolved flux which is nearly zero near the surface. Only
by augmenting a grid-independent, near-wall eddy viscosity in the Smagorinsky-
Deardorff model, were we able to eliminate the bias and produce similar turbulence
statistics in the two LESs as shown in Fig. 3b.
This LES-within-LES experiment provides a first step toward understanding
the two-way nesting issues of matching different grid resolutions with differ-
ent SGS contributions. We intend to extend the LES-within-LES problem to the
LES-within-mesoscale problem, which is a bigger challenge.
4 Summary
Three turbulence-physical processes interactions were discussed. The first exam-
ple showed that the PBL turbulence interacted with the heating and cooling at the
ground which led to a buoyancy-driven PBL in daytime and a shear-driven PBL at
night. The second example showed that the atmospheric (above the ocean surface)
Fig. 3 Comparison of the resolved-scale (thick curves) and SGS (thin curves) momentum-
flux profiles between the outer and the inner domains with (a) the Smagorinsky SGS model
and (b) the two-part SGS model. Solid curves are from the outer domain and dotted curves
from the inner domain
and the oceanic (below the surface) PBLs were both influenced by ocean waves. The
turbulence properties of both PBLs depended strongly on phase speeds, scales, and
heights of the waves. The third example presented a rather sophisticated interaction
of turbulence with clouds and radiation. Near the top of the stratocumulus-topped
PBL, turbulent mixing, entrainment, cloud droplet evaporation, and longwave ra-
diation all interacted simultaneously to modify the cloud amount and the turbulent
intensity.
For some real-world PBL applications, turbulence may depend sensitively on
local terrain and meteorological conditions. For such applications, two-way nesting
of an LES domain inside a weather forecast model may be required to explicitly
simulate turbulence and its interaction with local processes. This brute-force nu-
merical method may be straightforward but great care should be taken in handling
the SGS and the turbulence spin-up problems.
Acknowledgements. The National Center for Atmospheric Research is sponsored by the

National Science Foundation. Jeffrey Weil acknowledges support from the U.S. Army
Research Office under Grant W911NF-04-1-0411.
References
1. Bretherton, C.S. (co-authors): An intercomparison of radiatively driven entrainment
and turbulence in a smoke cloud, as simulated by different numerical models.
Q.J.R.M.S. 125, 391–423 (1999)
2. Chow, F.K., Weigel, A.P., Street, R.L., Rotach, M.W., Xue, M.: High-resolution
large-eddy simulations of flow in a steep Alpine valley, Part I: Methodology, verifica-
tion, and sensitivity experiments. J. Applied Meteorology and Climatology 45, 63–73
(2006)
3. Emanuel, K.: Tropical cyclone energetics and structure. In: Fedorovich, E., Rotunno, R.,
Stevens, B. (eds.) Atmospheric Turbulence and Mesoscale Meteorology, pp. 165–191.
Cambridge University Press, Cambridge (2004)
4. Finnigan, J.J.: Turbulence in waving wheat II: Structure of momentum transfer.
Boundary-Layer Meteorol. 16, 213–236 (1979)
5. Gao, W., Shaw, R.H., Paw, U.K.T.: Observation of organized structure in turbulent flow
within and above a forest canopy. Boundary-Layer Meteorol. 47, 349–377 (1989)
6. Mahrt, L.: Stratified atmospheric boundary layers. Boundary-Layer Meteorology 90,
375–396 (1999)
7. Moeng, C.H.: Large-eddy simulation of a stratus-topped boundary layer. Part I: Structure
and budgets. J. Atmos. Sci. 43, 2886–2900 (1986)
8. Moeng, C.H., Shen, S., Randall, D.A.: Physical processes within the nocturnal stratus-
topped boundary layer. J. Atmos. Sci. 49, 2384–2401 (1992)
9. Moeng, C.H., Dudhia, J., Klemp, J., Sullivan, P.P.: Examining two-way grid nesting
for large eddy simulation of the PBL using the WRF model. Mon. Wea. Rev. 135,
2295–2311 (2007)
10. Moeng, C.H. (co-authors): Simulation of a stratocumulus-topped planetary boundary
layer: Intercomparison among different numerical codes. Bull. Amer. Meteor. Soc. 77,
261–278 (1996)
11. Patton, E.G., Sullivan, P.P., Davis, K.J.: The influence of a forets canopy on top-down
and bottom-up diffusion in the planetary boundary layer. Q.J.R. Meteorol. Soc. 129,
1415–1434 (2003)
12. Randall, D.A., Coakley, J.A., Fairall, C.W., Kropfli, R.A., Lenschow, D.H.: Outlook
for research on subtropical marine stratiform clouds. Bull. Amer. Meteor. Soc. 65,
1290–1301 (1984)
13. Rasch, P.J., Chen, C.-C., Latham, J.: The impact of cloud seeding of marine stratocumu-
lus on the ocean. Geophysical Research Abstracts, 10, EGU2008-A-12211 (2008)
14. Shaw, R.H., Schumnn, U.: Large-eddy simulation of turbulent flow above and within a
forest. Boundary-Layer Meteorol. 61, 47–64 (1992)
15. Stevens, B. (co-authors): Dynamics and chemistry of marine stratocumulus—DYCOMS-
II. Bull. Amer. Meteor. Soc. 84, 579–593 (2003)
16. Stevens, B. (co-authors): Evaluation of large-eddy simulations via observations of
nocturnal marine stratocumulus. Mon. Wea. Rev. 133, 1443–1462 (2005)
17. Sullivan, P.P., McWilliams, J.C., Moeng, C.H.: Simulation of turbulent flow over
idealized water waves. J. Fluid Mech. 404, 47–85 (2000)
18. Sullivan, P.P., McWilliams, J.C., Melville, W.K.: Surface gravity wave effects in
the oceanic boundary layer: Large-eddy simulation with vortex force and stochastic
breakers. J. Fluid Mech. 593, 405–452 (2007)
19. Sullivan, P.P., Edson, J.B., Hristov, T., McWilliams, J.C.: Large eddy simulations and
observations of atmospheric marine boundary layers above non-equilibrium surface
waves. J. Atmos. Sci. 65, 1225–1245 (2008)
LES of Pulsating Turbulent Flows over Smooth
and Wavy Boundaries
A. Scotti, M. Gasser i Rubinat, and E. Balaras
Abstract. Flows driven by a pressure gradient that oscillates periodically around a

non-zero mean (pulsating flows) are found in a variety of geophysical, engineering
and biomedical settings. Moreover, aside from their practical importance, they are
a useful model to understand the more general problem of how unsteadiness affects
the properties of a boundary layer. In this paper, we consider examples of pulsating
flows over smooth and wavy surfaces studied with the aid of LES. For the smooth
case, the surprising result is that the time averaged statistics are marginally if at all
affected by the presence of oscillations (at least in the regime considered of current
dominated flows), whereas the oscillating part is influenced by the underlying steady
turbulence. Introducing waviness of sufficient amplitude to induce flow separation,
at least during part of the cycle, couples the mean to oscillating component much
more tightly, resulting in an increased drag felt by the mean flow. Such enhanced
drag is due to the ejection of large, coherent spanwise vortices that form in the lee
of the ripples, and it has a strong and non trivial dependence on the frequency of the
oscillation.
1 Introduction
In geophysics the problem of oscillating and pulsating flows over wavy surfaces is
of great practical importance, since this kind of boundary layer flow is very common
in the near shore environment. In water shallower than a wavelength, the periodic
pressure signature of the waves extends to the bottom, generating a pulsating bound-
ary layer. The problem is further complicated when, under appropriate conditions,
A. Scotti · M. Gasser i Rubinat
Dept. of Marine Sciences, UNC, Chapel Hill
e-mail: ascotti@unc.edu
E. Balaras
Fischell Dept. of Bioengineering, UMD, College Park
e-mail: balaras@umd.edu
26 A. Scotti, M. Gasser i Rubinat, and E. Balaras
ripple forms, with wavelengths that are of the same order of magnitude of the
orbital oscillations of water particles at the edge of the boundary layer [13]. A de-
tailed understanding of the boundary layer properties is a necessary ingredient to
accurately predict the evolution of natural and man-made structures along the shore-
line. This is a problem of considerable societal interest, for example to evaluate the
effectiveness of beach nourishment projects, which cost from 2150 to 3600 USD
per linear meter of beach treated. Total expenditures in the last two decades along
the US eastern seaboard totaled over 1 billion USD1 .
In an interesting example of cross-discipline convergence, pulsating flows are
also of great interest in the biofluid mechanics of the human circulatory system [21],
with stenosis playing the role of “ripples”. In this paper, we focus on the interaction
(or lack of) between the mean flow and the oscillating part, with particular regard to
how turbulence and oscillations influence the drag experienced by the mean flow.
Much has been written about how turbulent coherent structures contribute to
the transfer of momentum toward a boundary. The canonical view for a steady
and homogenous flow over a smooth surface is that the largest contribution comes
from the generation, evolution and bursting of streamwise vortices, which lift low-
momentum fluid away from the wall and replace it with high-momentum fluid from
further up [17]. Thus, any change that affects the generation-ejection-bursting cycle
should have an effect on the induced drag. An example that has been studied in the
past is a boundary layer in a region of accelerating downstream flow [14]. When
the acceleration exceeds a certain threshold, streamwise vortices are inhibited and
the flow relaminarizes. The behavior is different when the pressure gradient varies
in time. For the specific case of a pulsating pressure gradient, ample experimental
and numerical evidence points to the fact that the oscillations have little effect on
the mean, time averaged flow, over a wide range of frequencies, even though the
oscillating part is strongly affected, and the instantaneous flow may go from quasi-
laminar to fully turbulent [20, 18] within a period. When the boundary is not flat,
the problem becomes more complicated. For purpose of classification, non-smooth
boundaries can be either rough, if the power spectrum of the boundary elevation
extends over a wide range of scales, or wavy, if a single (or a few) wavelengths
dominate the spectrum (for simplicity we exclude the case of a rough boundary su-
perimposed to a wavy one). In the former case, the experimental evidence suggests
that, at least in the steady case, the characteristics of the mean flow are similar to the
flow over a flat wall, provided an appropriate roughness scale is introduced[15, 10].
In the latter case, the periodicity of the boundary is reflected in the structure of
the flow. If the amplitude of the ripples is large enough, the flow separates in the
lees of the obstacles, resulting in an increased drag. This flow has been studied
almost exclusively under steady conditions, both experimentally [8] and numeri-
cally [7, 3]. One of the main differences with respect to the canonical flow over a
flat surface is the appearance of strong streamwise vortices on the upslope of the
ripples, starting immediately after the flow reattaches. A common explanation for
these vortices is that they are due to centrifugal instabilities associated to the curving
1 Data from Duke University Program for the Study of Developed Shorelines.
LES of Pulsating Turbulent Flows over Smooth and Wavy Boundaries 27
streamlines[3]. However, no formal stability calculation has been performed under

realistic conditions to verify whether Görtler vortices can grow so rapidly. In fact,
the problem remains open. Another active area, where intense momentum transfer
can occur is in the shear region capping the recirculation area in the lee of the rip-
ples. Finally, under unsteady forcing conditions, the recirculating bubble detaches
from the ripples during flow reversal, further enhancing the exchange of momen-
tum. Here, we restrict our attention to the case of a sinusoidal wall. From the the-
oretical point of view, the problem has been mostly considered within the context
of Unsteady Reynolds Averaged Equations (URANS), mostly simple one- or two-
equation models. However, there is a mounting body of evidence that single-point
URANS models are not suited but for the lowest order statistics of the pulsating
flows [19, 4], because the flow is dominated by unsteadiness, and when ripples
are present, streamline curvature and flow separation. Even the addition of time-
memory effects in single point model only marginally improves the agreement [16].
It seems that the Boussinesq assumption essentially fails in this type of flows. For
this reason, we first establish the validity of the LES dynamic model [6, 12] for
pulsating flows over smooth surfaces, for which we have good experimental data.
2 Approach
We solve the filtered Navier-Stokes equations
∂ ui ∂ 1 ∂ P ∂ τi j ∂ 2 ui ∂ ui
+ (ui u j ) = − − +ν + fi , = 0, (1)
∂t ∂xj ρ ∂ xi ∂ x j ∂ x j∂ x j ∂ xi
f = ( fs + Uosc ω cos(ω t), 0, 0). (2)
where ui is the large-scale velocity vector, P is the pressure, ρ is the fluid density
and ν is the kinematic viscosity; fi represents an external body force field that drives
the flow, the latter being the sum of a steady and an oscillating component. For flows
over a smooth surface, the domain exmployed is 3π H long in the streamwise direc-
tion, π H long in the spanwise and H in the vertical. No slip boundary conditions
are applied to the upper and lower wall, with periodic conditions along the other
two directions. Spectral discretization is imployed, with Fourier modes in the hor-
izontal directions and Chebychev polynomials in the vertical [18]. The domain in
the rippled case spans two ripple wavelength in the horizontal, and one ripple wave-
length in the vertical. The ripple aspect ratio is 0.10, the same as in Hudson et al.[8].
The domain is discretized with second order finite differences, using 288 × 64 × 128
points. (Figure 1). Within the rectangular domain, the immersed boundary technique
is used to introduce the boundary [1]. In the present implementation the filtering
operator is a top-hat filter in physical space, and is applied implicitly by the finite-
difference operators. In both cases, the effect of the small scales upon the resolved
part of turbulence appears in the subgrid-scale (SGS) stress term, τi j = ui u j − ui u j ,
which is parameterized using the standard dynamic model [6] over the smooth wall.
Over a wavy wall, due to lack of homogeneity, it is not possible to average the eddy
0.8
0.6
z/λ
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
x/λ
Fig. 1 Domain employed when ripples are present. The thick box shows the Control Volume
employed in the analysis of momentum transfer
viscosity over horizontal planes to remove numerically unstable negative values.

Thus, in this case we use the Lagrangian dynamic eddy-viscosity model [12], which
averages the eddy viscosity along Lagrangian trajectories.
3 Results
3.1 Smooth Wall
In addition to the Reynolds u∗ H/ν number based on the friction velocity u∗
calculated from the steady component of the pressure gradient, which in these ex-
periments was fixed at at 350, two additional parameters appear in the problem. The
ratio auc of oscillating to mean centerline velocity and a nondimensional frequency
ω + = ων /u∗2 . An alternative to the latter is the laminar Stokes length, defined as
ls+2 = 2/ω + . When auc < 1 (current dominated flows), experiments [20] show that
the flow is largely controlled by ω + . In the results presented here, auc 0.7 and
0.0004 ≤ ω + ≤ 0.1.
Comparison with highly resolved DNS and experiments show that the dynamic
model, unlike URANS model, is capable of handling the unsteadiness of the flow to
a satisfactory degree of accuracy [18, 19]. With unsteadiness built into the driving
conditions, the appropriate ensemble operator is the combination a phase average in
time and a spatial average over the homogenous directions, which we denote with
. Further, any phase averaged quantity f (z,t) can be subject to harmonic analysis,
that is
∞
f = F(z) + ∑ F n (z) cos(nω t + φ n (z)). (3)
n=1
The numerical experiments show that the amount of variance contained in the higher
harmonics (n > 1) typically follows a bell shaped curve. At high and low driving
frequencies the contribution from higher harmonic is small, whereas a maximum
is reached at intermediate frequencies. This can be explained with the aid of a tur-
bulence penetration length lt . This length measures how far the vorticity gener-
ated by the oscillating pressure acting on the wall propagates in the flow. A simple
dimensional analysis argument, well supported by the numerical experiments [18],

shows that at high frequencies lt+ ls+ , while at low driving frequencies lt+
κ (ls+ )2 , where κ is the von Karman constant. When lt+ < 5, the effect of the os-
cillating part is basically contained within the viscous sublayer, and the turbulence
in the outer flow is simply advected back and forth. As the frequency decreases,
lt increases, and the oscillating part begins to drive turbulence away from its equi-
librium. As the frequency is further reduced, the oscillating vorticity generated at
the wall fills the entire channel in a time short relative to the turbulence adjust-
ment scale. Thus, the flow returns to quasi equilibrium. Figure 2 shows profiles of
Reynolds shear stress at different times during the cycle, showing how unsteadiness
remains confined within 2lt+ from the boundary.
These plots also shows that at intermediate frequencies the flow is near laminar
(during the initial acceleration stage), followed by a sudden switch to turbulent at
the beginning of the decelerating phase.
The surprising result is that the time averaged component of the ensemble
averaged velocity is virtually independendent of lt+ in agreement with the standard
law of the wall, and no change is drag is experienced by the mean flow, despite
the fact that, at intermediate values of lt+ , the flow relaminarizes during part of the
cycle! Laboratory experiments are divided on this issue, with some showing no de-
viation from the standard law-of-the-wall behavior [2], whereas other show a small
departure [20]. This susbtantial decoupling (observed only when the streamwise ve-
locity is averaged in time!), is still unexplained. On the contrary, the modulation
of the mean streamwise velocity, largely contained at the driving frequency, shows
an evolution consistent with the notion that lt+ increases with decreasing frequency,
and approaches a modulated log-layer at very low frequencies (Figure 3).
3.2 Wavy Wall

We show results from a total of 11 numerical experiments (Table 1): 10 with differ-
ent combinations of amplitude, Uosc , and frequency, ω , of the oscillating component
of the forcing, and one with steady conditions, the latter matching the experimen-
tal conditions of Hudson et al.[8] and used as a baseline case. In addition to the
two new parameters already encountered in the smooth wall case, two new nondi-
mensional parameters exist: the ratio of orbital amplitude to the ripple wavelength.
2Uosc /λ ω ≡ a/λ and the ratio of ripple heigth to wavelength, h/λ . We keep the
latter constant at 0.1.
Unlike the smooth wall case, in which the mean flow is essentially unaffected
by the presence of the oscillations[18], here oscillations increase the drag felt
by the mean flow. To better understand how drag is enhanced, we consider the
momentum balance within a Control Volume (CV) extending from the tip of
the ripples to the upper flat wall in the vertical direction, and over the entire
domain in the horizontal direction (see Figure 1). Along the upper boundary,
the only contribution of the momentum flux is due to the stress applied to the
wall τuw . Along the lower boundary, the flux is due to the molecular flux qm =
ν∂ u/∂ z, the turbulent Reynolds Stress RS = (u − u)(w − w) = u
w
, and
8 (a) (b) (c)
10−1 101 103 10−1 101 103 10−1 101 103

+ + +
z z z
Fig. 2 Time profiles of Reynolds Stress profiles as a function of distance from the wall during
a full cycle. The bottom profile is taken at the beginning of the acceleration phase, and the
profiles are equispaced in time. Crosses shows the profile for a steady channel flow. The
dotted line shows the location of the laminar Stokes length, while the dashed is located at
2lt+ . (a) High frequency case, (b) intermediate frequency and (c) low frequency
the stress due to the flux induced by the phase averaged flow (coherent stress)
CS = uw. The latter can be split into the sum of the steady component CSST =
u w and the rectified part CSUNST = CS − CSST. Note that denotes a phase
average, whereas an overbar denotes an average over a period and over a rip-
√ steady forcing conditions, CSUNST = 0. We define the fric-
ple wavelength. Under
tion velocity u∗ ≡ qm + RS + CSST + CSUNST , which we use to define appropriate
25
20
o l+=7
+
* l =14
15 +
+ l =35
U/u*
+
square l =70
10
0
0 1 2
10 10 10
20
15
U/u*
10
0
0 1 2
10 10 10
+
z
Fig. 3 Streamwise velocity profiles. Top: time averaged component; the dashed line shows
the standard law-of-the-wall. Bottom: amplitude of the modulation. Solid and dashed lines
show the laminar Stokes solution at very high and intermediate frequencies
wall units, denoted with the usual + superscript. We found that with this defini-
tion the mean velocity profile above the ripples follows reasonably well the usual
law-of-the-wall,
+ 1
u = log (z − d)/z0 , (4)
k
provided z+ − d + > 100. The offset in the wall position is calculated following
Jackson [9]. Across the experiments d/h 0.6, in good agreement with other
experiments [15] (Figure 4). It is also possible to introduce a drag coefficient
Cd = 2u∗2/Ub2 . Based on how the total momentum flux across the lower boundary
of the CV is partitioned among the different terms, the 10 unsteady experiments fall
quite naturally in three categories: suborbital-ripple flows (a/λ > 1), orbital-ripple
flows (a/λ ∼ 1) and superorbital-ripple flows (a/λ < 1) (see Table 1).
In the steady case, the momentum flux towards the wall is dominated by the
Reynolds stress concentrated along the shear layer which caps the recirculation area
in the lee of the ripples [1]. About 25% of that momentum is recovered via the
coherent stress, due to the stoss-lee asymmetry in the streamlines. In a Clauser plot,
the mean velocity exhibits the downward shift characteristics of flows over rough
surfaces (Figure 4). Assuming an equivalent sand roughness of 70 (the height of
the ripples in wall units), the magnitude Δ U + 8 of the shift is in good agreement
Table 1 Components of the momentum flux at the edge of the ripples, drag coefficient Cd ≡
2
2u∗2 / max u and roughness parameters. a/λ is the ratio of orbital excursion (a = Uosc T /π )
to ripple wavelength λ . The mean bulk velocity Ub is normalized with the values of the bulk
velocity in the steady case, while the components of the stress are normalized with f s H,
where H is the distance from the top of the ripples to the upper wall. Thus, the sum of
the normalized stresses equal one. Positive (negative) values denote outward (inward) flux.
Typical uncertainties are of the order of 5%
Exp. Tinf T1/3U13 T2/3U13 T1U13 T1U26 T2U13 T2U26 T4U13 T4U26 T8U13 T8U26
a/λ – 0.27 0.55 0.82 1.66 1.66 3.31 3.31 6.62 6.62 13.24
Ub 1.00 1.00 0.85 0.78 0.54 0.69 0.64 0.84 0.63 0.89 0.65
τUW 0.19 0.16 0.15 0.15 0.08 0.11 0.10 0.14 0.11 0.13 0.11
qm 0.23 0.17 0.12 0.11 0.05 0.10 0.11 0.17 0.12 0.19 0.13
RS 0.79 0.58 0.34 0.32 0.42 0.56 0.99 0.80 1.02 0.86 1.01
CSST -0.20 -0.10 -0.12 -0.11 0.01 -0.16 -0.20 -0.14 -0.22 -0.14 -0.18
CSUNST – 0.18 0.51 0.55 0.44 0.40 0.01 0.02 -0.02 -0.04 -0.06
Cd 0.013 0.016 0.018 0.025 0.045 0.032 0.032 0.019 0.031 0.017 0.034
z+
0 2.74 3.11 6.67 13.93 44.92 22.42 30.63 8.31 24.90 6.88 21.33
f 1 1.14 2.43 5.08 16.39 8.18 11.18 3.03 9.09 2.51 7.78
ΔU + 7.96 8.27 10.13 11.92 14.78 13.08 13.85 10.67 13.34 10.20 12.96
with Colebrook and White [5]. The effect of the oscillations on the mean steady
flow is to increase the apparent roughness z0 of the bottom. By dimensional analysis
z0 = z0 (a/λ ,Uosc /Ust , h/λ ). In the foregoing analysis, we assume that when h/λ
is large enough that separation occurs during most of the cycle, the relationship
simplifies to z0 = z0 (h/λ ) f (a/λ ,Uosc /Ust ). Not surprisingly, f appears to increase
with increasing values of Uosc /Ust . What is surprising, however, is that for constant
Uosc /Ust , f exhibits a sharp peak near a/λ 1.5. Interestingly, a peak in drag at
exactly the same value of a/λ is found also in purely oscillating flows [11]. In the
rest of the paper, we will investigate what features are responsible for this behavior.
Sub-orbital ripple cases. When Uosc ∼ Ust , the relative partition of the total mo-
mentum flux is almost identical to the steady case. The lion’s share of the interior-
to-wall momentum flux is taken by the Reynolds stress. Since the bulk velocity is
10 ÷ 15% less than the steady case, the equivalent drag RS/Ub2 increases by a 30%.
Increasing the amplitude of the oscillation to twice Ust shifts some of the momen-
tum flux from the molecular term to the Reynolds stress. Also, the bulk velocity is
further reduced to 65% of the steady case, so that the equivalent drag is now almost
three times as large. We have seen that in the steady case the the coherent stress CS
reinjects part of the momentum into the mean flow. Under sub-orbital conditions, the
contribution of CS to the momentum flux is similarly negative (i.e., momentum is
added to the bulk of the flow), though somewhat reduced relative to the steady case.
What is surprising is that the unsteady contribution to CS is negligible in all sub-
orbital cases, regardless of the forcing amplitude. Overall, for sub-orbital flows drag
is dominated by the effects of the small-scale coherent structures generated along
the shear layer that caps the recirculating zones. The large-scale, coherent vortices
that detach from the eddies during each period (see section 3.3) do not contribute to
the net flux.
25
κ−1log z + 5.5
Tinf
T1U26
T2U26
T4U26
T8U26
20
+
〈 u〉 +ΔU
+
15
10
2
10
z+−d+
Fig. 4 Mean velocity profiles above the ripples for selected cases. Velocity profiles are shifted
in the vertical by Δ U + . The straight line is the reference smooth logarithmic profile
Orbital-ripple cases. While suborbital flows are characterized by an enhanced

Reynolds stress, orbital flows show a reduction of the relative role of the Reynolds
stress (still due to shear layer instabilities) in the overall momentum budget. The
slack is picked up by the coherent stress, whose unsteady part contributes as much,
if not more, to draining momentum from the outer region of the flow. The synergis-
tic effect results in a much increased overall drag, and flatter velocity profiles near
the edge of the ripples. For this flows the small-scale coherent structures over the
shear layer add up to the large-scale vortices ejected by the ripples. The behavior of
the roughness height implies a peak in drag when a/λ ∼ 1, with a quick drop on
either side. It thus suggests that this peculiar behavior is due to a resonance between
the oscillating flow and the formation/ejection of the recirculation bubble in the lee
of the ripples.
Superorbital-ripple case. As the frequency increases, keeping Uosc bounded, it is
reasonable to expect that the effect of the oscillations on the mean flow decreases,
and eventually disappear in the limit ω → ∞. In particular, the coherent stress should
decrease. Experiment T1/3U13, undertaken to establish a boundary to the orbital
case, confirms this expectation, but still exhibits a reduced Reynolds stress. The
reduction is though offset by the lower value of CS, and the overall result is only a
slight increase in the drag coefficient.
3.3 Temporal Evolution of the Near Wall Momentum Flux

Reynolds Stress. In this section we contrast the temporal evolution of the Reynolds
stress against the coherent stress. For brevity, we consider only the Uosc = Ust
cases. In the super-orbital case (Figure 5, squares) the Reynolds stress remains
essentially constant over a wave period, smaller than the steady value, because the
large oscillating pressure gradient inhibits almost completely the formation of the
0.02
(a)
0
RS/U2b
−0.02
−0.04
−0.06
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
a/λ=0.27 a/λ=0.55 a/λ=0.82 a/λ=1.66 a/λ=3.31 a/λ=6.62
0.02
0
CS/U2b
−0.02
(b)
−0.04
−0.06
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

phase
Fig. 5 Temporal evolution of the Reynolds Stress u

w

(a) and Coherent Stress uw (b)
over a wave cycle at the upper edge of the ripple. The stresses are averaged over the length of
the ripple. The flow decelerates from 0 to 0.5 and accelerates from 0.5 to 1. The ratio Uosc /Ust
is kept fixed at 1. Both stresses are normalized with the mean bulk velocity. The thick solid
line shows the value for the flow without oscillations
recirculation zone. As the frequency decreases, the Reynolds stress becomes more
and more asymmetric, with large values found during the deceleration stage. Also,
the peak shifts closer to the beginning of the deceleration phase, due to the fact that
the recirculation bubble develops earlier and earlier during this portion of the the
cycle.
Coherent Stress. The evolution of the the coherent stress with increasing driving
frequency is more complex. The asymmetry between the decelerating and the ac-
celerating phase is present only for the orbital and super-orbital ripples. The much
smaller variation of CS in the sub-orbital cases remains nearly symmetric, resulting
in an overall cancellation effect. Also, the peak in CS precedes the peak in RS at
high-frequencies, but lags it as the frequency decreases.
Detachment of the Separated Flow. The striking difference in the behavior of
the coherent stress can be explained considering how the recirculating bubble in
the lee of the ripples, a large-scale spanwise vortical structure, detaches from the
eddies when the flow reverses near the bottom boundary and lifts off from the lower
boundary. In orbital flows (Figure 6, left side), the recirculating vortex is confined
to the lee side of the ripple. As the flow near the bottom reverses, the vortex begins
to lift and is pushed against the lee side. The overall distribution of streamlines
is highly asymmetrical, resulting in a net transport of momentum toward the wall.
In the sub-orbital case (Figure 6, right side), the recirculating vortex occupies the
entire space between crests, and the flow within the bubble is much weaker. When
the near-boundary flow reverses, the vortex lifts off vertically from the wall, nearly
symmetric relative to the trough of the ripples. Thus, the integrated momentum flux
across the vortex is negligible.
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
0.4 0.4
0.2 0.2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
0.4 0.4
y/λ
0.2 0.2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/λ x/λ
Fig. 6 Left panels. Lift-off of the recirculating bubble under orbital conditions. From top to
bottom θ = 0.3, 0.35, 0.4. Right panels, same but for a sub-orbital flow. The thick lines shows
instantaneous streamlines of the phase averaged flow, while the thin lines show isocontour
values of Reynolds stress. Note how in the orbital case the Reynolds stress correlates with the
position of the vortex
4 Conclusions
The flow in pulsating turbulent boundary layers presents interesting challenges from
a theoretical, experimental and computational point of view. In this paper, we have
shown that these flows are within the range of modern numerical tools (LES). We
have contrasted the behavior of pulsating flows over smooth and wavy surfaces, and
found that in the former case the statistics of the time-averaged flow (both first and
second order) are quite insensitive to the presence of the oscillations. This is sur-
prising, and still not quite fully understood. On the contrary, when the boundary
is wavy, the apparent roughness of the boundary is affected by the presence of the
oscillations. In this case, the momentum transfer within a pulsating turbulent flow
over a wavy wall is mediated by eddies covering a wide range of spatial and tempo-
ral scales. The fast, small-scale eddies contributes to the standard Reynolds Stress,
whereas slow and large spanwise eddies that form in the lee of the ripples con-
tribute to the coherent stress. The relative importance is controlled by the driving
frequency and amplitude of the oscillation. The numerical experiments presented
here show that the spanwise vortices exert the largest influence when the orbital
amplitude of the oscillating flow a matches the wavelength λ of the ripples. This
results in a substantially larger drag on the averaged flow. Interestingly, the action
of gravity waves over sandy seabeds forms ripples for which a/λ ∼ 1 [13]. The
link between the present observations and its implication for sediment transport and
bedform evolution are presently being investigated.
Acknowledgements. This work was supported by NSF, grant OCE-0351667.
References
1. Balaras, E.: Modeling complex boundaries using an external force field on fixed cartesian
grids in large-eddy simulation. Comp. Fluids 33, 375–404 (2004)
2. Brereton, G.J., Reynolds, W.C., Jayaraman, R.: Response of a turbulent boundary layer
to sinusoidal free-stream unsteadyness. J. Fluid Mech. 221, 131 (1990)
3. Calhoun, R.J., Street, R.L.: Turbulent flow over a wavy surface: Neutral case. J. Geophys.
REs. 106, 9277–9294 (2001)
4. Chang, Y.S., Scotti, A.: Modeling unsteady turbulent flows over ripples: Reynolds-
averaged Navier-Stokes equations (RANS) versus large-eddy simulation (LES). J.
Geophys. Res. 109, C09,012 (2004)
5. Colebrook, C.F., White, C.M.: Experiments with fluid friction in roughened pipes. Proc.
R. Soc. London Ser. A 161, 367–381 (1937)
6. Germano, M., Piomelli, U., Moin, P., Cabot, W.: A dynamic subgrid-scale eddy viscosity
model. Phys. Fluids A3(7), 1760–1765 (1991)
7. Henn, D., Sykes, R.: Large-eddy simulation of flow over wavy surfaces. J. Fluid
Mech. 383, 75–112 (1999)
8. Hudson, J.D., Dykhno, L., Hanratty, T.J.: Turbulence production in flow over a wavy
wall. Exp. fluids 20, 257–265 (1996)
9. Jackson, P.S.: On the displacement height in the logarithmic velocity profile. J. Fluid
Mech. 111, 15–25 (1981)
10. Jimenez, J.: Turbulent flows over rough walls. Ann. Rev. Fluid Mech. 36, 173–196 (2004)
11. Longuet-Higgins, M.S.: Oscillating flow over steep sand ripples. J. Fluid Mech. 107,
1–35 (1981)
12. Meneveau, C., Lund, T.S., Cabot, W.H.: A lagrangian dynamic subgrid-scale model of
13. Nielsen, P.: Coastal bottom boundary layers and sediment transport, vol. 4, p. 107. World
Scientific, River Egde (1992)
14. Piomelli, U., Balaras, E., Pascarelli, A.: Turbulent structures in accelerating boundary
layers. J. Turb. 1, 1–16 (2000)
15. Raupach, M.R., Antonia, R.A., Rajagopalan, S.: Rough-wall turbulent boundary layers.
Appl. Mech. Rev. 44, 1–25 (1991)
16. Revell, A., Benhamadouche, S., Craft, T., Laurence, D.: A stress-strain lag eddy viscosity
model for unsteady mean flow. Int. J. Heat Fluid Flow 27(5), 821–830 (2006)
17. Robinson, S.K.: Coherent motions in the turbulent boundary layer. Ann. Rev. Fluid
Mech. 23, 601–639 (1991)
18. Scotti, A., Piomelli, U.: Numerical simulation of pulsating turbulent channel flow. Phys.
Fluids 13(5), 1367–1384 (2001)
19. Scotti, A., Piomelli, U.: Turbulence models in pulsating flows. AIAA Journal 40(3),
537–544 (2002)
20. Tardu, S., Binder, G., Blackwelder, R.F.: Turbulent channel flow with large amplitude
velocity oscillations. J. Fluid Mech. 267, 109 (1994)
21. Tullio, M.D.D., Cristallo, A., Balaras, E., Verzicco, R.: Direct numerical simulation of
the pulsatile flow through an aortic bileaflet mechanical heart valve. J. Fluid Mech. 622,
259–290 (2009)
Numerical Study of Turbulence–Wave
Interaction
Lian Shen
Abstract. We develop numerical capabilities of direct numerical simulation and

large-eddy simulation for turbulent flows with waving boundaries, which can be
coupled with nonlinear surface wave simulation, to study the mechanism of
turbulence–wave interaction. Simulation of turbulence in the vicinity of surface
waves with various wave conditions reveals strong dependence of the statistics,
structures, and dynamics of the turbulent flow on wave characteristics including
wave phase, wave age, and wave nonlinearity. Simulation of nonlinear wave evo-
lution provides wave growth quantification in a phase-resolving context, which is
valuable for deterministic wavefield prediction. The results obtained in this study
suggest the importance of two-way coupling between turbulence and waves in their
dynamic evolution.
1 Introduction
The problem of turbulent flows interacting with waves is of interest to many appli-
cations. Examples include wind over water waves in air–sea interaction, turbulence
mixing and transport in the surface wave environment, and in a broad sense, drag
reduction with a waving boundary undergoing fishlike motions. Turbulent flow at
a waving boundary differs substantially from that at a flat wall. In addition to the
complex geometry associated with the wavy surface, the presence of the boundary
waving motion affects the flow field significantly. Our understanding of this type
of flows is limited due to the complexity of the problem. In this study, we develop
numerical capabilities of direct numerical simulation (DNS) and large-eddy sim-
ulation (LES) for turbulent flows with waving boundaries, which can be coupled
Lian Shen
Department of Civil Engineering, Johns Hopkins University, 3400 N. Charles Street,
Baltimore, MD, 21218, USA
e-mail: LianShen@jhu.edu
38 L. Shen
with nonlinear wave simulation. Mechanistic study is performed in order to obtain

a fundamental understanding of wave–turbulence interaction.
2 Numerical Method
In this study, we develop a numerical capability for the simulation of turbulent flows
with no-slip or free-slip wavy surfaces that can undergo arbitrary waving motions
with moderate wave amplitude. Take DNS in the configuration shown in Fig. 1 as an
example. The Cartesian frame is fixed in space, with x, y, and z being the streamwise,
spanwise, and vertical coordinates, respectively. We apply periodic boundary con-
ditions in the streamwise and spanwise directions. The turbulent flow is described
by the incompressible Navier-Stokes equations
∂ ui ∂ ui u j 1 ∂p ∂ 2 ui
+ =− +ν , (1)
∂t ∂xj ρ a ∂ xi ∂ x j∂ x j
∂ ui
=0. (2)
∂ xi
In our simulation, we use a boundary-fitted grid system. As shown in Fig. 1, the
irregular surface-fitted domain in the physical space (t, x, y, z) is transformed to
a rectangular domain in the computational space (τ, ξ, ψ, ζ) with the following
algebraic mapping:
τ=t, (3a)
ξ = x, (3b)
ψ = y, (3c)
z
y
x
z
(x,y,t) y
o
x
H(x,y,t)
H
ζ
ψ
H’(x,y,t) ξ
Fig. 1 Illustration of coordinate transformation. The irregular domain in the physical space
is transformed to a rectangular domain in the computational space by an algebraic mapping
Numerical Study of Turbulence–Wave Interaction 39
z+H z + H + H
ζ= = . (3d)
η+H η
+ H
Here H is the averaged domain height; η(x, y,t) is the top boundary variation; and
H
(x, y,t) is the bottom boundary variation. The total distance variation between the
upper and lower boundaries is η
(x, y,t) = η(x, y,t) + H
(x, y,t).
Note that η
and H
are functions of x, y, and t. By applying chain rule to partial
differentiations in (1) and (2), we obtain fully nonlinear governing equations in the
computational space:
∂ u Ht
− ζηt
∂ u ∂ (uu) Hx
− ζη
x ∂ (uu)
+
+ +
∂τ η +H ∂ζ ∂ξ η +H ∂ζ

∂ (uv) Hy − ζηy ∂ (uv) 1 ∂ (uw)

+ +
+
∂ψ η +H ∂ζ η +H ∂ζ

1 ∂ p Hx
− ζη
x ∂ p
=− −
+ ν∇2 u , (4)
ρa ∂ ξ η +H ∂ζ
∂ v Ht
− ζηt
∂ v ∂ (vu) Hx
− ζη
x ∂ (vu)
+
+ +
∂τ η +H ∂ζ ∂ξ η +H ∂ζ

∂ (vv) Hy − ζηy ∂ (vv) 1 ∂ (vw)

+ +
+
∂ψ η +H ∂ζ η +H ∂ζ

1 ∂ p Hy
− ζη
y ∂ p
=− −
+ ν∇2 v , (5)
ρa ∂ ψ η +H ∂ζ
∂ w Ht
− ζηt
∂ w ∂ (wu) Hx
− ζη
x ∂ (wu)
+
+ +
∂τ η +H ∂ζ ∂ξ η +H ∂ζ

∂ (wv) Hy − ζHy ∂ (wv) 1 ∂ (ww)

+ +
+
∂ψ η +H ∂ζ η +H ∂ζ

1 1 ∂p
=− + ν∇2 w , (6)
ρa η
+ H ∂ ζ
∂ u Hx
− ζη
x ∂ u ∂ v Hy
− ζη
y ∂ v 1 ∂w
+
+ +
+ =0, (7)
∂ξ η +H ∂ζ ∂ψ η + H ∂ ζ η
+ H ∂ ζ
where the Laplacian operator is decomposed into ∇2 = ∇2ξψ + ∇2ζ as:
∂2 ∂2 Hx
− ζη
x ∂ 2 Hy
− ζη
y ∂ 2
∇2ξψ = + + 2 + 2 , (8)
∂ ξ2 ∂ ψ2 η
+ H ∂ ξ∂ ζ η
+ H ∂ ψ∂ ζ

Hxx − ζη
xx + Hyy

− ζη
yy (Hx
− ζη
x )η
x + (Hy
− ζη
y )η
y ∂
∇2ζ = −2
η
+ H (η
+ H)2 ∂ζ
40 L. Shen
1 + (Hx
− ζη
x )2 + (Hy
− ζη
y )2 ∂ 2
+ . (9)
(η
+ H)2 ∂ ζ2
The nonlinear governing equations (4)–(7) are then used in our numerical simula-
tion. For spatial discretization, we use a Fourier series based pseudo-spectral method
in the horizontal directions and a second-order finite-difference scheme on a stag-
gered grid ([2, 12]) in the vertical direction. The time integration of the flow field is
realized through a second-order fractional step scheme, with the incompressibility
obtained by a projection method ([5]).
When the above method is applied to water flows in the upper ocean with surface
waves, we focus on the top wavy boundary shown in Fig. 1. For surface waves of
finite amplitude but without wave breaking, the kinematic and dynamic boundary
conditions at the free surface are:
ηt = w − uηx − vηy , (10)

and
t1 · Σ · nT = 0 , (11a)
t2 · Σ · nT = 0 , (11b)
n · Σ · nT = 0 . (11c)
with
(−ηx , −ηy , 1)
n= , (12)
η2x + η2y + 1
(1, 0, ηx )
t1 = , (13)
η2x + 1
(0, 1, ηy )
t2 = , (14)
η2y + 1
⎡ ⎤
∂u ∂v ∂u ∂u ∂w
⎢ −P + 2ν ∂ x ν + ν +
⎢ ∂x ∂y ∂z ∂x ⎥
⎢ ⎥
⎢ ∂v ∂u ∂v ∂v ∂w ⎥ ⎥
Σ=⎢ν + −P + 2ν ν + ⎥. (15)
⎢ ∂x ∂y ∂y ∂z ∂y ⎥
⎢ ⎥
⎣ ∂u ∂w ∂v ∂w ∂w ⎦
ν + ν + −P + 2ν
∂z ∂x ∂z ∂y ∂z
Here P = p − ρ gη, where g is the gravitational acceleration. The subscripts

associated with η denote partial derivatives.
When the above method is applied to wind over water waves, we focus on
the bottom wavy boundary shown in Fig. 1. The motion of the wavy surface can
be either prescribed or coupled with a wave simulation. In our work, the above
approach based on moving grid fitted to the wave boundary is complemented by

other simulation approaches using level-set method and smooth particle hydrody-
namics, about which the details are not discussed here due to space limitation.
Next, we discuss two simulation examples.
3 Mechanistic Study of Turbulence over Plane Progressive

Waves
To study the problem of wind over water waves, we consider the three-dimensional
turbulent Couette flow over a waving boundary shown in Fig. 2. In this canonical
problem, the flow is driven by a constant shear stress τ at the top boundary. We
consider various surface motions including waving vertically, water-wave orbital
motion, as well as the special case of a fixed wavy wall. In the present work, we
focus on the water wave problem, with sample results shown in Fig. 2. We have
characterized and quantified effects on turbulence by wave phase speed, surface
orbital velocity, wave steepness and nonlinearity, and in particular, wave age c/u∗
that is defined as the ratio between the wave phase speed c and the turbulence friction
velocity u∗ . Representative results are discussed next with the canonical problem of
turbulence over prescribed monochromatic waves with steepness ak = 0.25 (here a
is the wave amplitude and k is the wavenumber) and wave ages c/u∗ = 2, 14, and
25, which correspond to slow, intermediate, and fast waves, respectively.
Figure 3 shows the streamline pattern of the mean flow for turbulence over wa-
ter waves with different wave ages. The height of the critical layer ([8]), where the
mean velocity of turbulence matches the phase speed of the wave, is denoted by the
Fig. 2 3D flow field illustration of turbulence over a monochromatic surface wave.

Streamwise velocity contours are plotted on the vertical planes and pressure contours are
plotted on the wave surface
42 L. Shen
0.5 0.5 0.5

(a) (b) (c)
z/λ
z/λ
z/λ
0 0 0
0 0.5 1 0 0.5 1 0 0.5 1
x/λ x/λ x/λ
Fig. 3 Phase-averaged streamline patterns over plane progressive water waves with steepness
ak = 0.25 and wave ages: a c/u∗ = 2; b c/u∗ = 14; and c c/u∗ = 25. Velocity used for
calculating streamlines is in the wave-following frame, i.e. (u − c, w) is used in the plots.
The dash-dot-dot lines represent the critical layer where u − c = 0
-1.6 1.6
0.5
(a)
z/λ
0
0 0.5 1 1.5 2
0.5
(b)
z/λ
0
0 0.5 1 1.5 2
0.5
(c)
z/λ
0
0 0.5 1 1.5 2
x/λ
Fig. 4 Contours of normalized Reynolds stress −u

w
/u2∗ over plane progressive water
waves with steepness ak = 0.25 and wave ages: a c/u∗ = 2; b c/u∗ = 14; and c c/u∗ = 25
dash-dot-dot line. In the wave-following frame, the critical layer is surrounded by

closed streamlines, known as the “cat’s eyes” ([6]). As the wave age increases, the
height of the critical layer increases.
The distribution of Reynolds stress in Fig. 4 indicates strong dependence on the
wave phase, and this dependence changes drastically with the wave age. For the case
of c/u∗ = 2, the maximum of Reynolds stress lies above the wave trough. Besides
this apparent peak region, there exists a second high Reynolds stress region that
extends from the first to the upward downstream direction over the wave crest. For
cases of c/u∗ = 14 and 25, the positive peak moves to the leeward side of the crest,
while there is a negative peak on the windward side of the crest.
Figure 5 compares instantaneous coherent vortical structures in the near-surface
region between slow (c/u∗ = 2) and fast (c/u∗ = 25) wave cases. The dominant
vortex structures near the wave surface are stretched in the streamwise direction. It
Fig. 5 Snapshot of near-surface coherent vortical structures in instantaneous turbulence field

over plane progressive water waves with steepness ak = 0.25 and wave ages: a c/u∗ = 2 and
b c/u∗ = 25. The vortical structures are represented by the iso-surface of λ2 = −1. Here λ2
is the second largest eigenvalue of the tensor S2 + Ω2 , where S and Ω are the symmetric and
antisymmetric parts of the velocity gradient tensor ∇u
is apparent that features of the coherent vortices are strongly dependent on c/u∗ . In
the fast wave case shown in Fig. 5b, there are vortex sheets lying right above the
wave crests and troughs. Further investigation (not shown here) indicates that these
vortex sheets have vorticity vectors pointing in the spanwise direction, which are
generated by the fast moving wave.
Figure 6 presents a close-up view of the slow wave case shown in Fig 5a. Vor-
tices with positive streamwise vorticity (ωx ) are denoted by grey, while those with
negative ωx are denoted by black. Several typical coherent vortex structures are
identified. Vortices (1) and (3) are quasi-streamwise vortices with positive ωx , while
Fig. 6 Close-up view of the instantaneous vortical structures shown in Fig. 5a. Structures
with grey color are vortices with positive streamwise vorticity; those with black color are
vortices with negative streamwise vorticity. Two types of typical vortical structures are
identified: vortices (1)–(3) are quasi-streamwise vortices; (4)–(6) are reversed horseshoe
vortices
44 L. Shen
Y
X
(4)
(3) (5)
(2)
(1)
Fig. 7 Conditionally-averaged near-surface coherent vortical structures, represented by

iso-surface of λ2 , associated with ejection and sweep for case (ak, c/u∗ ) = (0.25, 2). The
grey and black colors have the same meaning as in Fig. 6. Vortical structures (1)–(4) are
associated with ejection; (5) is associated with sweep
vortex (2) is a quasi-streamwise vortex with negative ωx . As shown in the figure,

these quasi-streamwise vortical structures start from the windward side of a wave
crest, and they extend over the wave crest. Another type of vortex structures is the
horseshoe vortex, e.g. vortices (4)–(6) in Fig. 6. It is found that these horseshoe vor-
tices have their heads located upstream and legs downstream. This is the opposite
of typical horseshoe vortices near a flat wall, of which the heads are downstream of
the legs. Also we found that the heads of these vortices are usually located above
the wave trough. In general, the quasi-streamwise vortices prevail over the entire
surface. The horseshoe vortices, together with some arch vortices and some small
spanwise vortices, concentrate near the wave trough.
In turbulence boundary layer flows, the ejection and sweep events are often re-
lated to near-wall quasi-streamwise vortical structures. We employ the conditional
average method of variable-interval space averaging to educe vortical structures as-
sociated with ejections and sweeps. Figure 7 shows the educed vortical structures
for the case of turbulence over water wave with c/u∗ = 2. The vortical structures
(1)–(4) are associated with ejections, while structure (5) is associated with sweeps.
Structures (2)–(4) are all counter-rotating vortex pairs with upwelling motion in
between. Structures (1) and (5) are horseshoe vortices with the legs upstream and
downstream of their heads, respectively. The results shown above indicate that the
underlying surface waves have strong influences on the type, location, and intensity
of the near-surface coherent vortical structures. Consequently, turbulence transport
and mixing are highly dependent on the phase, amplitude, and speed of the wave
motion.
4 Interaction of Nonlinear Wave Evolution with Turbulence

For the nonlinear wavefield evolution, an efficacious high-order spectral (HOS)
method makes it possible to capture all of the essential nonlinear wave interaction
processes at a reasonable computational cost. The HOS directly simulates evolution
of the surface elevation η and the surface potential Φs , which is defined as the sur-
face value of the velocity potential Φ ([14]). With a perturbation series of Φ with
respect to the wave steepness to the order of M and Taylor series expansions about
the mean water level z = 0,

M M M−m
η ∂ Φ(m)
Φ(x, y, z,t) = ∑ Φ (x, y, z,t) , Φ (x, y,t) = ∑ ∑
(m) s
, (16)
m=1 m=1 =0 ! ∂ z
z=0
and an eigenfunction expansion of each Φ(m) with N modes,

N
∑ Φn
(m)
Φ(m) (x, y, z,t) = (t)Ψn (x, y, z) , z ≤ 0 , (17)
n=1
the kinematic and dynamic free surface boundary conditions are written as
ηt = −∇h η · ∇h Φs + (1 + ∇h η · ∇h η)

M M−m N
η ∂ +1
× ∑ ∑ ∑
(m)
Φn (t) +1 Ψn (x, y, z) , (18)
m=1 =0 ! n=1 ∂z z=0
η 1 pa (x, y) 1
Φts = − − ∇h Φs · ∇h Φs + DΦ − − (1 + ∇hη · ∇h η)
Fr2 2 ρw 2
2
M M−m N
η ∂ +1
× ∑ ∑ ∑ Φn (t) ∂ z+1 Ψn (x, y, z)
(m)
. (19)
m=1 =0 ! n=1 z=0
Here, ∇h ≡ ∂ /∂ x + ∂ /∂ y; DΦ is wave dissipation; pa is the atmospheric pressure

applied at the wave surface representing the wind forcing. Equations (18) and (19)
are advanced in time with a fourth-order Runge-Kutta scheme. By using a pseudo-
spectral method with fast Fourier transform in horizontal directions, the HOS
accounts for the nonlinear interactions among all the N wave modes up to the de-
sired perturbation order M in wave steepness. This method requires a computational
cost almost linearly proportional to N and M, and its achieves an exponential con-
vergence rate of the solution with respect to both N and M for moderately steep
wavefields. Complete review of the scheme, validation, and application of HOS is
provided in Chap. 15 of [7].
Previous research revealed that the wind forcing on the large gravity waves is
dominated by the pressure forcing ([1, 10]). With the atmospheric pressure in (19)
given by the turbulence simulation, the HOS simulation advances to next time step.
The new wave surface elevation and surface orbital velocity then provide a new
mapping grid and new Dirichlet boundary condition for the turbulence simulation
to advance in time to the next step. With small time step, this fractional step ap-
proach couples the simulations of the turbulence field and the dynamically evolving
wavefield together. A typical snapshot of the instantaneous turbulent wind field over
the complex wavefield simulated by such a two-way coupling approach is shown in
Fig. 8.
With the above two-way coupling approach, we first study the growth of a
monochromatic wave under the wind pressure forcing. Figure 9 shows the time
46 L. Shen
wind
wave
Fig. 8 Interaction of nonlinear wave evolution with turbulence. In the wind field, contours of
streamwise velocity component are shown on the two vertical planes. The velocity is normal-
ized by the mean velocity at the top of the computational domain. The air pressure contours
are shown on the wave surface and the pressure is normalized by air density and wind friction
velocity. The air domain is lifted up for better visualization
0.2
h 0.1
0
0 200 400 600 800 1000
time
Fig. 9 Growth of height for waves under wind pressure forcing for cases of various c/u∗ :
− − −, c/u∗ = 2; –·–, c/u∗ = 5; ——, c/u∗ = 14
evolution of the wave height for waves with various c/u∗. The result clearly indi-
cates that slow waves (c/u∗ = 2 and 5) grow fast, while the amplitude of interme-
diate wave (c/u∗ = 14) barely increases. The difference in the growth rate of water
waves with various c/u∗ is associated with the distribution of the wave-induced air
pressure. Figure 10 shows the phase-averaged air pressure distribution on the wave
surface. The distribution of the air pressure above water wave is strongly wave phase
dependent, and this dependence varies as the wave age changes. For slow waves, the
2
pa/ρau*
0
-5
0 0.5 1 1.5 2
x/λ
η
Fig. 10 Wind pressure profiles on the surface of waves with various c/u∗ : − − −, c/u∗ = 2;
–·–, c/u∗ = 5; ——, c/u∗ = 14
pressure is high on the windward side of the wave crest and low on the leeward side,
which results in a positive form drag and a net momentum flux from the wind to
the wave. As the wave moves faster, the air pressure distribution becomes almost
symmetric about wave crest. As a consequence, the form drag becomes very small
and the wave barely grows for the case of c/u∗ = 14.
100
80
60
β 40
20
-20
0 5 10 15 20 25
c/u*
Fig. 11 Comparison of wave growth rate parameter β: , measurement data in the literature
compiled by [11]; ×, numerical results by [4, 13]; · · ··, theoretical prediction by [9]; ◦, current
results for monochromatic waves; ——, current result for broadband wave with cp /u∗ = 12.3
We then apply the two-way coupling approach to simulate the interaction of

turbulence with complex broadband wavefield illustrated in Fig. 8. We construct an
48 L. Shen
initial broadband wavefield based on the JONSWAP spectrum ([3]) with the wave
age of the peak wave satisfying cp /u∗ = 12.3. The two-way coupling approach al-
lows us to follow the evolution of the complex wavefield with wave phase resolved,
and to capture the the wave-coherent atmospheric pressure field on the wave surface.
The wave growth rate parameter β with respect to each wave mode is quantified, and
is plotted as a function of c/u∗ for corresponding wave mode together with our re-
sults for the monochromatic wave cases and data in the literature ([4, 9, 11, 13]).
Despite of the large scattering in the measurement data, the comparison in Fig. 11
indicates that our numerical simulation results agree well with the existing data in
the literature. Our simulation result for the broadband wave case also shows that as
the wave mode approaches the short wave end (corresponding to high wavenumber
k and low wave age c/u∗), the values of β depart from those of the monochromatic
cases in our simulation as well as in the literature, suggesting the necessity of the
stochastic model for wave and turbulence simulation near the spectrum tail.
5 Conclusion
In this study we develop numerical capabilities of DNS and LES for turbulent flows
interacting with surface waves, which enable us to perform a simulation based study
on the dynamics of turbulence-wave interaction. It is found that the mean flow, tur-
bulence intensity, Reynolds stress, and vortical structures strongly depend on the
wave motion. Meanwhile, the nonlinear evolution of waves resulting from the in-
teraction with the turbulent flows is elucidated. The results suggest that the dy-
namic, phase-coherent coupling between the turbulence and the wave fields must be
captured in their modeling to improve the accuracy of prediction.
References
1. Donelan, M.A.: Wind-induced growth and attenuation of laborotary waves. In:
Sajjadi, S.G., Thomas, N.H., Hunt, J.C.R. (eds.) Wind-over-Wave Couplings,
pp. 183–194. Clarendon press, Oxford (1999)
2. Harlow, F.H., Welch, J.E.: Numerical calculation of time-dependent viscous incompress-
ible flow of fluid with free surface. Phys. Fluids 8, 2182–2189 (1965)
3. Hasselmann, K., Barnett, T.P., Bouws, E., Carlson, H., Cartwright, D.E., Enke, K.,
Ewing, J.A., Gienapp, H., Hasselmann, D.E., Kruseman, P., Meerburg, A., Mller, P.,
Olbers, D.J., Richter, K., Sell, W., Walden, H.: Measurements of wind-wave growth and
swell decay during the Joint North Sea Wave Project (JONSWAP). Ergnzungsheft zur
Deutschen Hydrographischen Zeitschrift Reihe A(8) (12) (1973)
4. Kihara, N., Hanazaki, H., Mizuya, T., Ueda, H.: Relationship between airflow at the
critical height and momentum transfer to the tranveling waves. Phys. Fluids 19, 015102
(2007)
5. Kim, J., Moin, P.: Application of a fractional-step method to incompressible Navier-
Stokes equations. J. Comput. Phys. 59, 308–323 (1985)
6. Lighthill, M.J.: Physical interpretation of the mathematical theory of wave generation by
wind. J. Fluid Mech. 14, 385–398 (1962)
7. Mei, C.C., Stiassnie, M., Yue, D.K.P.: Theory and Applications of Ocean Surface Waves.
Part 2. Nonlinear Aspects. World Scientific, New Jersey (2005)
8. Miles, J.W.: On the generation of surface waves by shear flows. J. Fluid Mech. 3,
185–204 (1957)
9. Miles, J.W.: Surface-wave generation revisited. J. Fluid Mech. 256, 427–441 (1993)
10. Phillips, O.M.: The Dynamics of The Upper Ocean, 2nd edn. Cambridge University
Press, Cambridge (1977)
11. Plant, W.J.: A relationship between wind stress and wave slope. J. Geophys. Res. 87,
1961–1967 (1982)
12. Shen, L., Zhang, X., Yue, D.K.P., Triantafyllou, M.S.: Turbulent flow over a flexible wall
undergoing a streamwise tranvelling wave motion. J. Fluid Mech. 484, 197–221 (2003)
13. Sullivan, P.P., McWilliams, J.C., Moeng, C.-H.: Simulation of turbulent flow over ideal-
ized water waves. J. Fluid Mech. 404, 47–85 (2000)
14. Zakharov, V.E.: Stability of periodic wave of finite amplitude on the surface of a deep
fluid. J. Appl. Mech. Tech. Phys. 2, 190–194 (1968)
High Reynolds Number Wall-Bounded
Turbulence and a Proposal for a New
Eddy-Based Model
Alexander J. Smits
Abstract. We propose a model for turbulent wall-bounded flows based on new un-
derstanding of the turbulent structure. Specifically, we identify three basic eddy
motions: (1) the Large-Scale Motions (LSMs) which are related to the vortex pack-
ets defined by Head and Bandyopadhyay (1981) and Adrian et al. (2000); (2) the
Very Large-Scale Motions (VLSMs) interpreted by Liu et al. (2001) and Balaku-
mar and Adrian (2007) in terms of a concatenation of the outer layer bulges and
by Monty et al. (2007) in terms of the meandering “superstructures” observed in
pipe, channel and boundary layers; and (3) the streaks associated with longitudinal
vortex-like structures in the near-wall region, as identified by Kline et al. (1967). The
new model maps the attributes of each eddy type in physical space to wavenumber
space. Experimental data are then used to determine the scaling behavior of the three
basic eddy motions in wavenumber space, and the scaling behavior of the Reynolds
stress behavior is recovered b! y integrating over all wavenumbers.
1 Introduction
Here we describe progress in developing a new eddy-based model for turbulent
wall-bounded flows. The model aims to capture the correct scaling behavior of the
turbulence intensities, and other near-wall statistics for arbitrarily large Reynolds
numbers. Such information may lead to a new turbulence model for prediction
purposes.
The model is based on our new understanding of high Reynolds number turbu-
lence. Specifically, for wall-bounded flows, we identify three basic eddy motions:
(1) the Large-Scale Motions (LSMs) that are related to the vortex packets defined by
Head and Bandyopadhyay (1981) and Adrian et al. (2000); (2) the Very Large-Scale
Motions (VLSMs) interpreted by Liu et al. (2001) in terms of the outer layer bulges
Alexander J. Smits
Princeton University, Princeton, NJ 08544, U. S. A.
e-mail: asmits@princeton.edu
52 A.J. Smits
and by Monty et al. (2007) in terms of the meandering “superstructures” observed

in pipe, channel and boundary layers; and (3) the streaks associated with longitudi-
nal vortex-like structures in the near-wall region, as first identified by Kline et al.
(1967). The new model maps the attributes of each eddy type in physical space to
wavenumber space. Experimental data sre then used to determine the scaling behav-
ior of the three basic eddy motions in wavenumber space, and the scaling behavior
of the Reynolds stress behavior is recovered by integrating over all wavenumbers.
The model described here is similar in concept to the attached eddy model origi-
nally proposed for the logarithmic region of wall-bounded flows by Perry, Henbest
and Chong (1986). Scaling arguments were advanced for particular regions of the
spectrum, where low wavenumber motions were assumed to scale on outer layer
variables (uτ and δ ), intermediate wavenumber motions were assumed to be related
to attached eddies (in the sense of Townsend) so that they scaled inversely with the
distance from the wall (y−1 ), and high wavenumber motions were assumed to follow
Kolmogorov scaling (η and υ ). Overlap arguments then defined (for the logarith-
mic part of the velocity profile) a region of k−1 and k−5/3 in the spectrum for the
streamwise component. By integrating the spectrum, scaling laws were derived for
smooth and rough walls (Perry and Li, 1990). By using largely empirical input, the
model was later extended to cover the complete outer layer (by Marusic, Uddin and
Perry, 1997), and to include the near-wall region (by Marusic and Kunkel, 2003).
The success of this hybrid model is clear from the results shown in figure 1.
Fig. 1 Streamwise turbulence intensity measurements: high Reynolds numbers from ASL
(Metzger and Klewicki; Marusic et al.); low Reynolds number data from DeGraaff and Eaton
(2000). Lines are model results from Marusic and Kunkel (2003)
Despite its apparent success, there are a number of compelling reasons to re-
visit this earlier model. First, the model is based on the presence of k−1 and k−5/3
regions in the spectrum. Although the k−5/3 is well established, experiments now
High Reynolds Number Wall-Bounded Turbulence 53
indicate that the k−1 region is only evident at very high Reynolds numbers over a
very limited spatial extent (Nickels et al. 2007). Second, the interactions between
outer layer motions and inner layer motions has become much clearer in recent
years, and the simple division between inner and outer layer scaling that leads to
the k−1 region fails to capture that interaction. Specifically, the region where we
might expect k−1 scaling corresponds to the wavenumbers occupied by the LSMs,
and experiments have clearly shown that although the LSMs appear to behave as
attached motions, they do not scale simply as y−1 . Third, the importance of the
VLSMs was not appreciated until recently. For example, at high Reynolds numbers
the low wavenumber VLSMs contribute up to half of the total energy content of
the streamwise turbulence component. Also, the low wavenumber motions do not
scale with outer layer variables as assumed in the Perry et al. model. Fourth, it has
become clear that the relative importance of LSMs, VLSMs, and superstructures
depends on the nature of the flow: they behave differently in pipes, channels and
boundary layers (Monty et al. 2007, Bailey, Hultmark and Smits 2008), something
that is not captured in the original model. Fifth, the original model does not include
the effects of pressure gradient (although some steps were taken by Perry, Marusic
and Jones 2002), compressibility (although steps were taken by Fernando and Smits
1988, and Dussauge and Smits 1995), or heat transfer.
2 Model Design
To overcome the limitations of the existing model described above, we propose a
new model that captures the correct spectral behavior of wall-bounded turbulence,
and allows prediction of Reynolds stress distributions in wall-bounded flows, and
has the potential to include the effects of pressure gradients, compressibility, and
heat transfer.
A preliminary form for the new model has now been established. Here, the three
eddy motions are modeled using Gaussian distributions in the log of the wave num-
ber, the first notionally describes the spectral content of the VLSMs ( f1 ), the second
describes the spectral content of the LSMs ( f2 ), and the third specifically describes
the spectral content of the near-wall motions ( f3 ). The model is inspired by the form
of the pre-multiplied spectra, which closely resembles a bimodal distribution in the
log of the wavenumber (see, for example, figure 2). Two examples of how the model
functions are used to represent the spectrum are shown in figure 3: one in the loga-
rithmic region where the contribution of f3 is negligible, and one at y+ = 15 where
f3 takes its maximm value. The distributions resemble closely the spectral content
of wavelets in physical space, specifically the Mexican Hat wavelet that mirrors the
“simple” or “schematic” eddy proposed by Tennekes and Lumley (1972). The en-
ergy contained at each wavenumber is taken to be a simple sum over all eddy types,
neglecting nonlinear interactions among eddies. The high wavenumber part of the
spectrum is bounded by the k−5/3 inertial and dissipation ranges, which are modeled
as a single cutoff using a modified Pao spectrum.
54 A.J. Smits
Fig. 2 Pre-multiplied spectra for the streamwise velocity fluctuation in a pipe at ReD = 1.4M,
(Reτ = 27K. Data from Bailey et al. (2008)
Fig. 3 Candidate eddy functions, each Gaussian in log(ky): f 1 : low kR - VSLM (extends
throughout layer); f2 : high kR - LSM (extends throughout layer); f 3 : near-wall - production
(localized over intervals in y+ and y/R). Examples for Reτ = 14K. Left: y/R = 0.1, y+ =
1417; right: y/R = 0.0012, y+ = 15
2.1 Candidate Eddy Functions

The three eddy functions are modeled using Gaussian distributions in the log of the
wave number. Specifically, we have for function 1 (low kR peak):
f1a f1b
f1 = 2 (1)
log(ky)−log(ky1m )
exp log(ky1μ )
where
+ 1.6
f1a = 1 − e−0.032(y ) (2)

y −0.1 y 0.21
f1b = 0.793 − (3)
R R
y 0.65
kym1 = 0.65 0.005 + 0.27 (4)
R
and
ky1μ = 3.2 (5)
Here, ky is the wave number based on distance from the wall, and R is the pipe
radius (or boundary layer thickness).
For function 2 (the high kR peak), we use:
f2a f2b
f2 = (6)
log(ky)−log(ky2m ) 2
exp log(ky2μ )
where
+ 1.6
f2a = 1 − e−0.032(y ) (7)
y 0.35
y −0.053
f2b = 1.111 − 0.8 (8)
R R
y
kym2 = 0.12 + 1.75 (9)
R
and
ky2μ = 13 (10)
For function 3 (the near-wall peak), we use:
f3a
f3 = 2 2 (11)
log(y+ )−log(10)
exp log(2.5) exp log(ky)−log(0.08)
log(4)
where
f3a = 0.975 (12)
It may be seen that f1 and f2 each have a variable amplitude ( f1b and f2b ), a variable
location (ky1m and ky2m ), and a viscous cut-off that reduces their influence to zero
at the wall ( f1a and f2a ). Note that f1b , f2b , kym1 , and kym2 are expressed in outer
scaling (y/R), and f1a and f2a are expressed in inner scaling (y+ ). The function f3
has a fixed position in ky and y+ , and a fixed amplitude. All three functions have a
fixed width.
The amplitudes and locations of f1 and f2 were found by using a wide range
of experimental data on premultiplied spectra by assuming no summation of the
underlying functions (see figures 4 and 5) , and scaled by constant factors for a best
possible fit to the summed spectra.
56 A.J. Smits
Fig. 4 Height of eddy functions. Left: data fit for f1b (V LSM). Right: data fit for f2b (LSM).
Data sources are given in the list of references
Fig. 5 Location of eddy functions. Left: data fit for kym1 (V LSM). Right: data fit for kym2
(LSM). Data sources are given in the list of references
The final component of the model is to limit the spectrum at high and low
wavenumbers. At high wavenumbers, the correct dissipation is ensured by matching
to a modified Pao spectrum:
kφ (k) K0 ε 2/3 −6(kη )4/3

= 2/3 2 e (13)
u2τ k uτ
where K0 = 0.5, and the modification to Pao’s spectrum is to change 1.5 to 6, so that
it gives a more realistic roll-off at high wavenumbers. The dissipation is found by
matching ( f1 + f2 + f3 ) (no smoothing) to this spectrum function,
for y/R < 0.0001, at ky = 0.16

for 0.0001 < y/R < 0.005, at ky = 0.18
y
for 0.005 < y/R < 0.5, at ky = 3.2 + 0.6
R
for y/R > 0.5, at ky = 2 (14)
Matching at low wavenumbers was also ensured by assuming that for kR < 0.1:
kφ (k)
= C0 kR (15)
u2τ
where C0 is found by matching ( f1 + f2 + f3 ) at kR = 0.1. This was a minor step,

and may well be neglected in future versions of the model. The conclusion is that the
functions all scale with ky, not with kR, which is in contrast to the earlier attached
eddy model where a significant portion of the spectrum was assumed to scale with
the outer scale R.
2.2 Turbulence Intensity

The turbulence intensity is given by the integral of functions 1, 2 and 3 up to the low
and high wavenumber limits defined by kR = 0.1 and ky as given by equation 14,
plus the low and high kR contributions. Ignoring these limits and the low and high
kR contributions for now, consider the integral of f1 , f2 and f3 . Note that
∞ ∞

2 a a √ √
e−(t /a) dt
= √ e−t
2 /2
dt = √ 2π = a π (16)
−∞ 2 −∞ 2
Hence, at a fixed y,
∞ ∞
u
2
≈ ( f1 + f2 + f3 ) d(ln k) = ( f1 + f2 + f3 ) d(ln ky)
u2τ −∞ −∞
√
= π f1a f1b log ky1μ + f2a f2b log ky2μ + f3a log 4 ln 10
√ + 1.6
y −0.1 y 0.21
= 0.793 π ln 3.2 1 − e−0.032(y ) −
R R
√ +
y −0.053 y 0.35
+ 1.111 π ln 13 1 − e−0.032(y )
1.6
− 0.8
R R
√
0.975 π ln 4
+ (17)
+ )−log(10) 2
exp log(ylog(2.5)
For y+ ≥ 25,

u
2 √ y −0.1 y 0.21
≈ 0.793 π ln 3.2 −
u2τ R R
y 0.35
√ y −0.053
+ 1.111 π ln 13 − 0.8
R R
√
0.975 π ln 4
+ (18)
+ )−log(10) 2
exp log(ylog(2.5)
For y+ ≥ 100,

u
2 √ y −0.1 y 0.21
≈ 0.793 π ln 3.2 −
u2τ R R
y 0.35
√ y −0.053
+ 1.111 π ln 13 − 0.8 (19)
R R
58 A.J. Smits
2.3 Comparison with the Attached Eddy Model in the Overlap

Region
Compare equation 19 with the model by Perry et al. for the overlap region (y+ ≥ 100,
y/R ≤ 0.1):
u
2 y
2
= B1 − A1 ln − V (y+ ) (20)
uτ R
where A1 = 1.03, B1 = 2.39, and the function V (y+ ) is given by

−0.9 −0.5
V (y+ ) = 5.58 1 − y+ y+ (21)
V (y+ ) is an empirical function that introduces the -5/3 cutoff at high wavenum-
ber (including the dissipation spectrum), and the appropriate blending at lower
wavenumbers. It is a crucial function because it regulates the Reynolds number
dependence of the outer layer profile.
If we use the same viscous cutoff for equation 19 we obtain the complete model
for y+ ≥ 100,

u
2 √ y −0.1 y 0.21
≈ 0.793 π ln 3.2 −
u2τ R R
y 0.35
√ y −0.053
+ 1.111 π ln 13 − 0.8 − V (y+ ) (22)
R R
We can also develop a model for the complete boundary layer by using the same
viscous cutoff with strong damping in the near-wall region, as follows:
u
2 √ y −0.1 y 0.21
−0.032(y+)1.6
= 0.793 π ln 3.2 1 − e −
u2τ R R
√ +
y −0.053 y 0.35
+ 1.111 π ln 13 1 − e−0.032(y )
1.6
− 0.8
R R
√
0.975 π ln 4 + +
−0.003(y ) 1.6
+ 2 − V (y ) 1 − e (23)
log(y+ )−log(10)
exp log(2.5)
2.4 Results
The results are compared with the data of Hutchins and Marusic (2007b) in figure 6,
which show the contour maps of pre-multiplied spectra for the streamwise compo-
nent of the velocity fluctuation in a zero pressure gradient turbulent boundary layer.
The predictions for some pipe flow spectra are shown in figure 7. The predictions
for the turbulence intensity using equation 23 for three different Reynolds numbers
Fig. 6 Contour maps of pre-multiplied spectra for the streamwise component of the velocity
fluctuation in a zero pressure gradient turbulent boundary layer for Reτ = 14K. Top: data
from Hutchins and Marusic (2007b); bottom: model predictions
are given in figure 8, which shows that even at this preliminary stage the new model
performs as well as the original model (figure 1) while capturing the actual spectral
behavior much more closely.
3 Discussion
There are at least two major criticisms that can be directed to the model presented
here. First, the model is derived from a mixture of pipe and boundary layer data.
Data for y/R < 0.05 was derived from boundary layers (especially the atmospheric
surface layer for the very near-wall flow, and for the near-wall peak behavior),
whereas data for y/R > 0.05 was derived from pipe flows. We know there exist
important differences in the spectra, but there are not sufficient data to be able to
60 A.J. Smits
Fig. 7 Pre-multiplied spectrum of the streamwise component of the velocity fluctuation in

a pipe flow at y/R = 0.1, y+ = 1417. Left: model predictions at Reτ = 14K, showing the
contributions from f1 and f2 ( f3 contribution is negligible). Right: data from Bailey et al.
(2008) at Reτ = 27K
Fig. 8 Intensity of the streamwise component of the velocity fluctuation in a zero pressure
gradient turbulent boundary layer. Model predictions for Reτ = 3.3K (Bailey et al., 2008),
14K (Hutchins and Marusic, 2007b), 2.3M (Kunkel and Marusic, 2003)
tailor the model to suit specific flows. For example, the turbulence intensity on the
pipe centerline is part of the model, so it obviously does not agree with boundary
layer data in this region. Second, the model does not reproduce an outer layer peak
in the turbulence intensity, but this peak may be a spurious result due to the limited
spatial resolution of conventional hot wire probes.
We can also make the following observations. First, the low wavenumber motions
that scale on kR occupy only a very small part of the model spectra (almost negligi-
ble in terms of the integral). The spectra are dominated by motions that scale on ky,
implying that the VLSM and LSM peaks are “attached,” although in the outer flow
(y/R > 0.05) they seem to scale slightly better on kR. Second, the location of the
LSM peak in near-wall region coincides closely with the production peak, implying
a causal connection between the production and attached motions. Third, an ex-
tension of the model to the wall-normal component of turbulence seems relatively
straightforward. A preliminary inspection of the wall-normal spectra measured by
Kunkel and Marusic (2003) suggests that it might be based on a single eddy func-
tion corresponding to f2 , although it may be necessary to include some form of the
near-wall function f3 . Finally, it seems promising to investigate the use of stretched
vortex models of the small scales, as suggested by Misra and Pullin (1997), to pro-
vide a simple eddy-based model for the dissipation scales. The stretched-spiral form
proposed by Lundgren (1982) seems particularly appropriate as a substitute for the
modified Pao spectrum adopted in the current model.
Acknowledgements. Financial support for this work was provided through ONR Grant
N00014-09-1-0263 (Ron Joslin). Dr. S. C. C. Bailey helped produce figure 4.
References
1. Adrian, R.C., Meinhart, C.D., Tomkins, C.D.: Vortex organization in the outer region of
the turbulent boundary layer. J. Fluid Mech. 422, 1–54 (2000)
2. Balakumar, B.J., Adrian, R.J.: Large- and very-large-scale motions in channel and
boundary-layer flows. Phil. Trans. R. Soc. A 365, 665–681 (2007)
3. Bailey, S.C.C., Hultmark, M., Smits, A.J., Schultz, M.P.: Azimuthal structure of turbu-
lence in high Reynolds number pipe flow. J. Fluid Mech. 615, 121–138 (2008)
4. Del Álamo, J.C., Jiménez, J., Zandonade, P., Moser, R.D.: Scaling of the energy spectra
of turbulent channels. J. Fluid Mech. 500, 135–144 (2004)
5. Dussauge, J.P., Smits, A.J.: Characteristic scales for energetic eddies in turbulent super-
sonic boundary layers. Experimental Thermal and Fluid Science 14(1) (1997)
6. Fernando, E.M., Smits, A.J.: The kinematics of simple vortex loop arrays, AIAA Paper
88-3657. In: First National Fluid Dynamics Conference, Cincinnati, Ohio, July 24-28
(1988)
7. Ganapathisubramani, B., Longmire, E.K., Marusic, I.: Characteristics of vortex packets
in turbulent boundary layers. J. Fluid Mech. 478, 35–46 (2003)
8. Guala, M., Hommema, S.E., Adrian, R.J.: Large-scale and very-large-scale motions in
turbulent pipe flow. J. Fluid Mech. 554, 521–542 (2006)
9. Head, M.R., Bandyopadhyay, P.R.: New aspects of turbulent boundary-layer structure. J.
Fluid Mech. 107, 297–338 (1981)
10. Hutchins, N., Ganapathisubramani, B., Marusic, I.: Dominant spanwise Fourier modes
and the existence of very large scale coherence in turbulent boundary layers. In: Proc.
15th Australasian Fluid Mechanics Conference, Sydney, Australia (2004)
11. Hutchins, N., Marusic, I.: Evidence of very long meandering features in the logarithmic
region of turbulent boundary layers. J. Fluid Mech. 579, 1–28 (2007a)
12. Hutchins, N., Marusic, I.: Large-scale influences in near-wall turbulence. Phil. Trans. R.
Soc. A 365, 647–664 (2007b)
13. Kim, K.C., Adrian, R.J.: Very large-scale motion in the outer layer. Phys. Fluids 11(2),
417–422 (1999)
14. Kline, S.J., Reynolds, W.C., Schraub, F.A., Runstadler, P.W.: The structure of turbulent
boundary layers. J. Fluid Mech. 30, 741–773 (1967)
62 A.J. Smits
15. Liu, Z., Adrian, R.J., Hanratty, T.J.: Large-scale modes of turbulent channel flow: trans-
port and structure. J. Fluid Mech. 448, 53–80 (2001)
16. Lundgren, T.S.: Strained spiral vortex model for turbulent ne structure. Phys. Fluids 25,
2193–2203 (1982)
17. Marusic, I., Uddin, M., Perry, A.E.: Similarity law for the streamwise turbulence inten-
sity in zero-pressure-gradient turbulent boundary layers. Phys. Fluids 9(12), 3718–3726
(1997)
18. Marusic, I., Kunkel, G.J.: Streamwise turbulence intensity formulation for flat-plate
boundary layers. Phys. Fluids 15(8), 2461–2464 (2003)
19. Meinhart, C.D., Adrian, R.J.: On the existence of uniform momentum zones in a turbu-
lent boundary layer. Phys. Fluids 7(4), 694–696 (1995)
20. Misra, A., Pullin, D.I.: A vortex-based subgrid stress model for large-eddy simulation.
Phys. Fluids 9, 2443–2454 (1997)
21. Monty, J.P., Stewart, J.A., Williams, R.C., Chong, M.S.: Large-scale features in turbulent
pipe and channel flows. J. Fluid Mech. 589, 147–156 (2007)
22. Morrison, J.F., McKeon, B.J., Jiang, W., Smits, A.J.: Scaling of the streamwise velocity
component in turbulent pipe flow. J. Fluid Mech. 508, 99–131 (2004)
23. Nickels, T.B., Marusic, I., Hafez, S.M., Hutchins, N., Chong, M.S.: Some predictions of
the attached eddy model for a high Reynolds number boundary layer Phil. Trans. R. Soc.
Lond. A 365, 807–822 (2007)
24. Perry, A.E., Henbest, S.M., Chong, M.S.: An experimental study of the turbulence struc-
ture in smooth- and rough- wall boundary layers. J. Fluid Mech. 177, 437–466 (1986)
25. Perry, A.E., Li, J.D.: Experimental support for the attached-eddy hypothesis in zero-
pressure- gradient turbulent boundary layers. J. Fluid Mech. 218, 405–438 (1990)
26. Perry, A.E., Marusic, I., Jones, M.B.: On the streamwise evolution of turbulent boundary
layers in arbitrary pressure gradients. J. Fluid Mech. 461, 61–91 (2002)
27. Tennekes, H., Lumley, J.L.: A First Course in Turbulence. The MIT Press, Cambridge
(1972)
28. Tomkins, C.D., Adrian, R.J.: Spanwise structure and scale growth in turbulent boundary
layers. J. Fluid Mech. 490, 37–74 (2003)
29. Tomkins, C.D., Adrian, R.J.: Energetic spanwise modes in the logarithmic layer of a
turbulent boundary layer. J. Fluid Mech. 545, 141–162 (2005)
30. Townsend, A.A.: The Structure of Turbulent Shear Flow, 2nd edn. Cambridge University
Press, Cambridge (1976)
31. Zhou, J., Adrian, R.J., Balachandar, S., Kendall, T.M.: Mechanisms for generating co-
herent packets of hairpin vortices in channel flows. J. Fluid Mech. 387, 353–396 (1999)
PANS Methodology Applied to
Elliptic-Relaxation Based Eddy Viscosity
Transport Model
Branislav Basara, Siniša Krajnović, and Sharath Girimaji
Abstract. The Partially-Averaged Navier-Stokes (PANS) approach is a recently

proposed method which changes seamlessly from the Reynolds-Averaged Navier-
Stokes (RANS) model equations to the direct numerical solution (DNS) of the
Navier-Stokes equations as the unresolved-to-total ratios of kinetic energy and dis-
sipation are varied. Two variants of the PANS model are derived up to now, one
based on the k-ε formulation and the other based on the k-ω formulation. We intro-
duce here another variant which is based on four equation eddy viscosity transport
model, namely ζ -f turbulence model. Benefits of using such near wall model inside
the PANS concept are clearly presented in this paper.
1 Introduction
The Large-Eddy Simulation (LES) has become a very popular method for simu-
lating turbulent flows in recent years. This method is certainly superior to RANS
methods in strongly separated flows, since it simulates directly the large turbulent
structures and models only the influence of the sub-grid scales on the resolved ones.
However, LES is a computationally demanding technique and for this reason it will
not be feasible to compute complex industrial flows at higher Reynolds numbers
Branislav Basara
Advanced Simulation Technologies, AVL List GmbH, Hans-List-Platz 1, 8020 Graz, Austria
e-mail: branislav.basara@avl.com
Siniša Krajnović
Department of Applied Mechanics, Chalmers University of Technology,
SE-41296 Gothenburg
e-mail: sinisa@chalmers.se
Sharath Girimaji
Aerospace Engineering Department, Texas A&M University,
College Station, TX 77843, USA
e-mail: girimaji@aeromail.tamu.edu
64 B. Basara, S. Krajnović, and S. Girimaji
with LES, even in the near future. As an alternative, various hybrid RANS/LES
approaches have been proposed. There are two main groups, namely zonal and
seamless methods. The Partially-Averaged Navier-Stokes (PANS) approach is a re-
cently proposed method by Girimaji et al. [3] and Girimaji [4], which changes seam-
lessly from RANS to the DNS of the Navier-Stokes equations. We introduce here
another variant of the PANS model which is based on four equation eddy viscos-
ity transport model, namely ζ -f turbulence models (Hanjalić et al. [5]), which is
recently derived from more known ν 2 -f model. For comparisons with the variant
of PANS model proposed in this paper, we have used the first formulation of PANS
where the unresolved kinetic energy and dissipation equations are systematically de-
rived from the k-ε model. The parameter which determines the unresolved-to-total
kinetic energy ratio fk is defined based on the grid spacing. The PANS asymptotic
behavior goes smoothly from RANS to DNS with decreasing fk . This parameter is
dynamically adjusted at each point at the end of every time step as done by Basara
et al. [1]. This paper presents comparisons between PANS models based on the k-ε
model and one proposed here which is based on the ζ -f model as well as with LES
calculations (Krajnović [6]).
2 Model Equations
The Partially-Averaged Navier-Stokes (PANS) equations are written in terms of
partially averaged or filtered velocity and pressure fields, thus
∂ Ui ∂ Ui τ (Vi ,V j ) 1 ∂p ∂ 2Ui
+Uj + =− +ν (1)
∂t ∂xj ∂xj ρ ∂ xi ∂ x j∂ x j
where the velocity field is decomposed into two components, the partially filtered
component and the sub- filter component as
Vi = Ui + ui . (2)
The closure for the sub-filter stress can be obtained by using the Boussinesq
approximation as
2
τ (Vi ,V j ) = −2νu Si j + ku δi j (3)
3
where the eddy viscosity of unresolved scales is equal to
ku2
νu = cμ (4)
εu
and the resolved stress tensor is given as

1 ∂ Ui ∂ U j
Si j = + . (5)
2 ∂xj ∂ xi
PANS Methodology Applied to Elliptic-Relaxation 65
The model equations for the unresolved kinetic energy ku and the unresolved
dissipation εu are required to close the system of equations given above, thus

Dku ∂ νu ∂ ku
= (Pu − εu ) + ν+ (6)
Dt ∂xj σku ∂ x j

Dεu εu ε2 ∂ νu ∂ εu
= Cε 1 Pu − Cε∗2 u + ν+ (7)
Dt ku ku ∂ x j σε u ∂ x j
The model coefficients are (see also Girimaji [4])
fk fk2
Cε∗2 = Cε 1 + (Cε 2 − Cε 1 ) ; σku,ε u = σk,ε (8)
fε fε
where the unresolved-to-total ratios of kinetic energy and dissipation are written
respectively as
ku εu
fk = , fε = (9)
k ε
The parameter which determines the unresolved-to-total kinetic energy ratio fk is
defined based on the grid spacing, thus
2
1 Δ 3
fk = √ (10)
cμ Λ
where Δ is the grid cell dimension and Λ is the Kolmogorov scale of turbulence,
while fε was taken to be equal 1. Basara et al. [1] introduced fk as a dynamic pa-
rameter in the computational procedure, changing at each point at the end of every
time step, and then it is used as a fixed value at the same location during the next
time step.
PANS produces the correct production-to-dissipation ratio which changes gradu-
ally from RANS value to DNS value as fk is reduced from 1 to zero. This is analyt-
ically proven in the paper of Girimaji et al. [3] (see also Girimaji [4]). This is clear
evidence that PANS captures all intermediate resolution with precision. It never goes
to a Smagorinsky model, but the problem therein lies with the Smagorinsky model
being a zero-equation closure, whereas PANS is a two-equation closure or it can
be adjusted to four-equation closure as shown below. The near-wall behaviour of
the PANS family of models has been established in Partially Integrated Turbulence
method (PITM) of Chaouat and Schiestel [2] and it can be directly transferred to
PANS.
In this paper, we derived the PANS model from the near-wall RANS model. A
derivation of the PANS ζ -f model can be started by introducing the velocity scale
equation from the RANS v2 -f model, thus

Dv2 v2 ∂ νu ∂ v2
= kf − ε + ν+ (11)
Dt k ∂xj σv2 ∂ x j
We can derive unresolved velocity scale ratio ζu = v2u /ku by using Equation (11),
thus
D f v 2/ f k
Dζu Dvu /ku
2 v2 k f 2 1 Dv2 fv2 v2 Dk
= = = v − (12)
Dt Dt Dt fk k Dt fk k2 Dt
and having in mind that (see Girimaji [4])
ku v2u εu
fk = ; fv2 = ; fε = (13)
k v2 ε
1 εu
Pu − εu = fk (P − ε ) ; ⇒P= (Pu − εu ) + (14)
fk fε
and assuming fε = 1 one can arrive at following equation

Dζu ζu ζu νu ∂ ζu
= fu − Pu + εu (1 − fk ) + ν + (15)
Dt ku ku σζ u ∂ x j
f2
where fu = v
fk f and from the RANS form of the elliptic relaxation function
v2
k f = φ22 − ε b22; b22 = k − 23 , follows
fv2 fv2
k fu = k f= (φ22 − ε b22) (16)
fk fk
Equation (16), together with following relationship
(φ22 − ε b22 )u = fv2 (φ22 − ε b22 ) (17)
lead to
1
fu = (φ22 − ε b22 )u (18)
ku
Therefore, the complete PANS ζ -f model is given by following set of equations
ku2
νu = cμ ζu (19)
εu

Dku ∂ νu ∂ ku
= Pu − εu + ν+ (20)
Dt ∂xj σku ∂ x j

Dεu εu εu2 ∂ νu ∂ εu
= Cε 1 Pu − Cε 2 + ν+ (21)
Dt ku ku ∂ x j σεu ∂ x j

Dζu ζu ζu νu ∂ ζu
= fu − Pu + εu (1 − fk ) + ν + (22)
Dt ku ku σζu ∂ x j

1 P 2
L2 ∇2 fu − fu = c1 + C2
ζu − (23)
Tu ε 3
where L is the length scale, and constants c1 and C2

are taken as proposed by
Hanjalic et al. [5], and Tu is the time scale defined by using unresolved kinetic
energy respectively. The production P in the last equation is obtained from Eq. (14).
It is clear that for fk = 1, the equation for ζu will get its RANS form. Note also that
fε = 1 implies εu = ε .
3 Results and Discussions

Flow past a finite cylinder mounted vertically on a flat plate was computed. Free
stream inlet velocity U0 = 10m/s and diameter D = 0.03m give a Reynolds number
of approximately 2x104. A test section of 24Dx20Dx28D (WxHxL) was used in
simulations, see Fig. 1. The computations are compared with measurements of Park
and Lee [8].
Fig. 1 Computational domain
Fig. 2 Surface pressure coefficient at two locations: z/D=0.05 (left) and z/D=0.5 (right) as
predicted by the PANS k-ε and ζ -f models
Fig. 3 An instantaneous iso-surface of the second invariant of the velocity gradient: LES
fine grid (left) and PANS ζ -f medium grid (right). The iso-surface is colored with velocity
magnitude
Fig. 4 Surface pressure coefficient at different locations: z/D=0.3 (upper left); z/D=0.75 (up-
per right); z/D=0.917 (down left); z/D=0.95 (down right); as predicted by the PANS ζ -f
model and LES
Computations of Krajnović and Basara [7] for this test case, showed limitations
of the PANS based on the k-ε model and the standard-wall function for the accurate
prediction of the separation from the curved surface. Fig. 2 shows predictions of the
surface pressure coefficient at two locations as predicted by the PANS k-ε model
and compared with the PANS ζ -f model on the coarsest mesh which is consisted of
4 million computational cells.
An instantaneous flow from LES on the fine mesh (21 million cells) and PANS ζ -
f on the medium mesh (14 million cells) was presented by iso-surfaces of the second
invariant of the velocity gradient in Fig. 3. The first impression of the resulting
flow is that the both of modeling approaches, despite different meshes used, are
predicting very similar flow structure. As the flow separates along the lateral sides
of the cylinder, a large number of small Kelvin-Helmholtz-like vortices are formed.
The time averaged results for the surface pressure coefficients at four positions
along the cylinder are compared with the experimental data in Fig. 4. PANS ζ -
f models results for the medium grid is almost identical to the fine grid results
obtained by LES.
4 Conclusions
Results presented here, show that the PANS ζ -f model is the step forward in using
a general PANS concept. It is shown that such concept has to be linked with the
appropriate near wall model. The ζ -f model is a good choice due to its simplicity
and accuracy as well. Proposed PANS ζ -f model predicts averaged results on the
coarser grids very well when compared with measurements and LES calculations
on the finest computational grid.
References
1. Basara, B., Krajnović, S., Girimaji, S.: PANS vs. LES for computations of the flow around
a 3D bluff body. In: Proc. of ERCOFTAC 7th Int. Symp. - ETMM7, Lymassol, Cyprus,
vol. 2/3, pp. 548–554 (2008)
2. Chaouat, B., Schiestel, R.: A new partially integrated transport model for subgrid-scale
stresses and dissipation rate for turbulent developing flows. Phys. Fluids 17, 1–19 (2005)
3. Girimaji, S., Jeong, E., Srinivasan, R.: Partially-Averaged Navier-Stokes Model for tur-
bulence: Fixed point analysis and comparison with unsteady Partially Averaged Navier-
Stokes. J. of Applied Mechanics 73, 422–429 (2006)
4. Girimaji, S.: Partially-Averaged Navier-Stokes Model for turbulence: A Reynolds- Aver-
aged Navier-Stokes to Direct Numerical Simulation bridging method. J. of Applied Me-
chanics 73, 413–421 (2006)
5. Hanjalić, K., Popovac, M., Hadziabdic, H.: A robust near-wall elliptic-relaxation eddy-
viscosity turbulence model for CFD. Int. J. of Heat and Fluid Flow 25, 1048–1051 (2004)
6. Krajnović, S.: Large eddy simulation of the flow around a tall finite cylinder. In: Flow,
Turbulence and Combustion (2009) (accepted for publication)
7. Krajnović, S., Basara, B.: Numerical simulations of the flow around a tall finite cylinder
using LES and PANS. In: Proc. of iTi Conference on Turbulence, Bertinoro, Italy (2008)
8. Park, C.-W., Lee, S.-J.: Flow structures around a finite circular cylinder embedded in var-
ious atmospheric boundary layers. Fluid Dynamics Research 30, 197–215 (2002)
PIV Study of Turbulent Flow in Porous Media
S. Bejatovic, M.F. Tachie⋆ , M. Agelinchaab, and S.S. Paul
Abstract. This paper reports on an experimental study of turbulent flow through

model two-dimensional porous media bounded on one side by a solid plane wall
and on the other side by a zone of clear fluid. The porous media comprised of square
arrays of transparent circular acrylic rods that were inserted into holes drilled onto
pairs of removable plates. The removable plates were then inserted into groves made
in the side walls of the test channel. The rod diameter and/or center-to-center spac-
ing between adjacent rods were varied to simulate porosities that ranged from 0.56
to 0.89. A particle image velocimetry technique was used to conduct detailed veloc-
ity measurements in the flow within the porous media and the adjacent clear zone.
From these measurements, the slip velocity and profiles of the spatially averaged
mean velocities and Reynolds stresses were obtained to study the effects of porosity
on the velocity field.
1 Introduction
Fluid flows through porous media occur in diverse engineering and technological
applications such as oil reservoirs and heat exchangers of open cell metal foam.
Depending of the specific application, the flow may be turbulent or laminar while
the porous medium may be unstructured and bounded by a zone of clear fluid or by
another layer of porous medium of different porosity, ε . Over the past decades, a
number of numerical studies of turbulent flows through porous medium were per-
formed using turbulence models of varying complexity. More recently [1], direct
numerical simulation results for turbulent flows in a plane channel through and over
packed beds were reported. Most of the previous studies of turbulent flows through
porous media are reviewed by [1] and will not be repeated.
S. Bejatovic ⋅ M.F. Tachie ⋅ M. Agelinchaab ⋅ S.S. Paul
Dept. of Mechanical & Manufacturing Engineering,
University of Manitoba, Winnipeg, Canada, R3T 5V6
e-mail: tachiemf@cc.umanitoba.ca
⋆
Corresponding author.
springerlink.com © Springer-Verlag Berlin Heidelberg 2010
72 S. Bejatovic et al.
Fig. 1 Schematic of test section (not to scale) and definition of coordinate system
In the recent past, the particle image velocimetry (PIV) technique was used to
conduct whole-field velocity measurements in creeping fluid flows through and over
model porous media. For example, Tachie et al. [2] performed PIV measurements
in the interfacial region of a simple shear flow penetrating model two-dimensional
porous media. The porous media comprised of regular arrays of circular, square,
and equilateral triangular acrylic rods arranged across the flow, and the solid vol-
ume fraction (φ ) was in the range 0.01 ≤ φ ≤ 0.16. It was concluded that the veloc-
ity at the interface between the homogeneous clear fluid and the porous medium
(which is often referred to as the slip velocity, Us ) decreased as φ was increased.
Subsequently, [3,4] conducted PIV measurements in three-dimensional porous me-
dia that comprised of circular rods arranged perpendicular to the main flow. In the
present study, profiles of the mean velocities and Reynolds stresses obtained in tur-
bulent flows through model two-dimensional porous media are reported. The model
porous media comprised of regular arrays of circular acrylic rods arranged across
the main flow (similar to the configuration studied by [2]) and filled 59% of the
channel height. The porosity was varied from ε = 0.56 to 0.89 and the velocity mea-
surements were performed using a PIV system.
2 Experimental Set-Up and Measurement Procedure

Test Section and Porous Media: The experiments were performed in a test channel
of length, L = 2500 mm, internal width, B = 175 mm, and height H = 70 mm. The
test section was inserted into an existing recirculation water channel that is also
2500 mm long but 200 mm wide and 200 mm high. The bottom, top and side walls of
both the test channel and the existing recirculation channel were made of transparent
acrylic plates to facilitate optical access. The circular rods used to model the porous
media span the entire width of the test channel. Figure 1 shows a schematic of
the test section and a portion of the square arrays of circular rods used to model
the porous media. The array of rods (hereafter referred to as the porous medium)
was placed adjacent to the top wall with clear zone of fluid between the porous
medium and the channel bottom wall (Fig. 1a). The streamwise, wall-normal and
lateral directions are, respectively, denoted by x, y, z. The dotted lines located at a
distant h1 from the top wall correspond to the nominal interface between the porous
medium and the clear fluid; h2 = H − h1 is the height of the clear fluid zone adjacent
to the porous medium. As shown in Fig. 1b, the rod diameter and center-to-center
PIV Study of Turbulent Flow in Porous Media 73
Table 1 Summary of porous media and test parameters

√
Test d l ε k h1 ⟨Um ⟩ M f −CZ ⟨Um ⟩/⟨Ub ⟩ ⟨Us ⟩∗
(mm) (mm) (mm) (m/s)
ε56 d4.7 l6.3 4.7 6.3 0.56 0.60 41.5 0.59 0.99 3.07 0.031
ε79 d4.7 l9.0 4.7 9.0 0.79 1.21 41.0 0.47 0.90 2.53 0.073
ε89 d4.7 l12.7 4.7 12.7 0.89 2.48 42.0 0.38 0.72 2.18 0.127
ε60 d3.2 l6.3 3.2 6.3 0.79 0.86 41.0 0.50 0.96 2.76 0.052
ε56 d6.3 l9.0 6.3 9.0 0.60 0.84 41.0 0.52 0.99 2.92 0.061
spacing are designated by d and l, respectively. A second wall-normal axis (Y ) is

defined such that Y =0 is located at the interface and increases toward the bottom
wall.
The rod diameter and/or center-to-center spacing were varied to produce a range
of porosity (ε = 1 − π d 2(4l 2 )). As shown in Table 1, the various test cases are des-
ignated by a notation of the form εi d j lm where i denotes the porosity (%), j denotes
the rod diameter (mm) and m denotes the center-to-center spacing between adjacent
rods (mm). For example, Test ε56 d4.7 l6.3 represents the test case for d =4.7 mm,
l = 6.3 mm and ε = 56%. The rods were inserted into precision holes drilled onto
pairs of removable plates that were then√inserted into grooves made in the side walls
of the test channel. The permeability, k, for each porous medium was calculated
from the following correlation [5]: k = r2 {−ln(1 − ε ) − 1.476 + 2(1 − ε ) − 1.774(1√−
ε )2 + 4.076(1 − ε )3}/8(1 − ε ), where r is the radius of the rods. The values of k
increased from 0.60 mm to 2.48 mm as ε increased from 0.56 to 0.89. The num-
ber of rows × columns of rods varied from 14×4 to 20×5, and the porous medium
covered about 59% of the channel height, i.e. h1 /H ≈ 0.59.
Flow Condition and PIV System: The Reynolds number based on the channel
height and the bulk velocity of the approach flow was fixed at ReH = 16800. The
flow was tripped using equilateral triangular acrylic rods with a nominal size of 4.5
mm taped on both the top and bottom channel walls at x = 50 mm from the inlet
section. The flow was seeded using polymer micro-sphere particles whose mean di-
ameter and specific gravity were, respectively, 5 μ m and 1.05. A 120 mJ/pulse Nd-
YAG laser of 532 nm wavelength was used to illuminate the flow field. The laser
sheet was shot from the bottom of the channel and located at the channel mid-span.
A 12-bit 2048×2048 pixels CCD camera with a 7.4μ m pixel pitch fitted to a 60 mm
Nikkor lens was used to image the flow field. The field of view was approximately
80 mm × 80 mm. The instantaneous images were post-processed using adaptive cor-
relation and moving average validation options of FlowManager developed by Dan-
tec Dynamics Inc. The adaptive correlation uses a multi-pass FFT cross-correlation
algorithm to determine the average particle displacement within the interrogation
area (IA). Details of the PIV system, steps taken to reduce measurement uncertain-
ties and data processing are available in [6]. Preliminary tests were performed to
verify that the number of image pairs used to compute the mean velocity (U) and
Reynolds stresses (u2 , v2 and −uv) is adequate, and also to quantify any effects of
Fig. 2 Spatial averaged mean velocities: ◻ ε56 d4.7 l6.3 ; ∧ ε79 d4.7 l9.0 ; ● ε89 d4.7 l12.7 ; ☆ SM
spatial resolutions on U, u2 , v2 and −uv. For the convergence test, profiles of U, u2 ,

v2 and −uv computed from an ensemble of N = 500, 1000, 1500 and 2000 image
pairs were compared. It was observed that the mean velocity and all the Reynolds
stresses are independent of sample size. The spatial resolution effects were also stud-
ied by comparing profiles of U, u2 , v2 and −uv obtained using interrogation area
(IA) sizes of 32 pixels×32 pixels×50% overlap (hereafter denoted by 32×32×50)
and those from IA of 32×16×50. The spacing between vectors or data points for
the 32×32×50 and 32×16×50 were, respectively, Δ x = 0.67mm × Δ y = 0.67mm and
Δ x = 0.67mm × Δ y = 0.34mm. The results show that U and −uv are independent
of spatial resolution; however, u2 and v2 obtained from 32×16×50 were relatively
higher than those from 32×32×50. Since the differences in u2 and v2 are comparable
to measurement uncertainties, and the data quality from IA of 32×32×50 is better
than that from 32×16× 50, all the data reported hereafter were those obtained from
N = 2000 and IA of 32×32×50 in the region where the flow is periodic.
3 Results and Discussion

The Mean Flow: A number of flow quantities are used to examine the effects
of porosity and permeability on the mean flow through and over the porous me-
dia. These include the spatially averaged streamwise mean velocity (⟨U⟩), wall-
normal mean velocity (⟨V ⟩), the area-averaged bulk velocity (Ub ⟩), the percentage
of flow distributions through the porous medium (M f −PM ) and the adjacent clear
zone (M f −CZ ) and the slip velocity (⟨Us ⟩). The spatial averaging was performed
over a pitch, and the profiles of ⟨U⟩∗ = ⟨U⟩/⟨Um ⟩ and ⟨V ⟩∗ = ⟨V ⟩/⟨Um ⟩ are shown in
Figs. 2a and 2b. It should be noted that 3 to 5 intermediate data points are skipped
in these and subsequent profiles to avoid data congestion. As expected, the refer-
ence SM profile obtained in a channel with both smooth top and bottom walls are
symmetric with respect to Y /h2 = 0.5. Due to higher flow resistance generated by
the porous media compared with the adjacent bottom smooth wall, the velocities
close to the porous media are lower and the profiles are asymmetric. This obser-
vation is akin to prior measurements in a channel with one wall roughened and the
other smooth. Figure 2a also shows that the mean velocity inside the porous medium
(Y /h2 < 0) is strongly dependent on the properties of the porous medium. For exam-
ple, the magnitude of ⟨U⟩∗ decreased from 0.6 to 0.07 as the porosity is reduced
from 89% to 56%.
The wall-normal velocity is practically zero for the reference SM data (Fig. 2b).
For the porous media, on the other hand, ⟨V ⟩ is predominantly negative in the clear
fluid zone with a maximum magnitude of 3% of ⟨Um ⟩. These non-negligible values
of ⟨V ⟩ within the clear zone and the porous medium imply that the production of
Reynolds stresses and the transport of both momentum and turbulent kinetic energy
(k) in flow through and over porous media are quite different from those in generic
fully developed channel flows. The ratio of the maximum velocity to the bulk veloc-
ity for the reference SM test is Um /Ub =1.1. This value is similar to previous results
obtained in a fully developed channel with smooth top and bottom walls. As a result
of the asymmetric resistance in the porous media and on the bottom wall, ⟨Um ⟩/⟨Ub ⟩
values increased approximately two-fold to three-fold (as shown in Table 1).
The portion of the profile in the clear zone and the porous medium were inte-
grated to obtain the mass flux per unit width in the clear zone and within the porous
medium. The mass fractions through the clear zone (M f −CZ ) presented in Table 1
demonstrate that if the rod diameter is fixed and the spacing is de-creased to produce
a medium of lower porosity, the fraction of fluid channeled through the clear fluid
zone increases. For the d = 4.7 mm rods, for example, M f −CZ increased from 72%
to 99% as l decreased from 12.7 mm (ε89 d4.7 l12.7 ) to 6.3 mm (ε56 d4.7 l6.3 ). Similarly,
if the rod spacing is fixed, for example at l =9 mm, and d is increased from 4.7 mm
(ε79 d4.7 l9.0 ) to 6.3 mm (ε60 d6.3 l9.0 ), M f −CZ increased from 90% to 99%. These re-
sults show that, for the present test configurations, a porosity of 0.60 or less would
produce enough flow resistance for 99% of the approach flow to be routed through
the clear fluid zone. For creeping flow through three-dimensional porous media [4],
it was reported that for a porosity of 0.88 or less, 99% or more of the approach flow
was channeled through the clear zone.
Figures 3a and 3b show the variation of U with x between adjacent rods in the
interfacial region. It should be noted that successive Y values (for example, Y−0 and
Y−1 ) in both plots are approximately 0.6 mm apart. The plots show how rapidly
U varies with wall-normal distance in the interfacial region. For each test condi-
tion, the values of U(x) over a pitch at the nominal interface, Y−0 , were used to
compute the slip velocity, ⟨Us ⟩. It is estimated that the uncertainty in determining
⟨Us ⟩ is approximately 5% to 10%. The values of the dimensionless slip velocity
(⟨Us ⟩∗ = ⟨Us ⟩/⟨Umax ⟩) are presented in Table 1. Similar to the mass fraction trends
noted above, ⟨Us ⟩∗ shows a consistent trend with porosity only if the rod diameter is
fixed and the spacing is varied or vice versa. If, however, different combinations of
rod diameter and rod spacing are used to produce identical porosity (e.g.,ε79d4.7 l9.0
and ε79 d3.2 l6.3 ), the porous medium with larger spacing (ε79 d4.7 l9.0 ) produces a
larger dimensionless slip velocity. In contrast, Figs. 4a and 4b demonstrate that the
permeability correlates the mass fractions and dimensionless slip velocities reason-
ably well. The plots demonstrate that the mass fraction in the clear zone (M f −CZ )
Fig. 3 Variation of U with x: ⧫ Y−3 ; ▲ Y−2 ; ● Y−1 ; ⋆ Y0 ; ○ Y1 ; ⊓ Y2 ; △ Y3 ; ☆ Y4
Fig. 4 Variation of mass fraction and slip velocity with permeability
√
decreased linearly with k while the dimensionless slip velocity √ and mass fraction
in the porous medium (M f −PM ) increased nearly linearly with k. The present re-
sults also demonstrate that, at a lower permeability, the slip velocity is only a small
fraction (3%) of the maximum velocity, ⟨Um ⟩.
Reynolds Stresses: The Reynolds stresses normalized by ⟨Um ⟩2 are shown in Fig. 5.
The normal stresses are plotted in Fig. 5a in a staggered format, while the shear
stress is plotted in Fig. 5b. The turbulence levels in the clear fluid zone adjacent to
the bottom smooth wall (Y /h2 > 0.5) are significantly lower than values measured
in the reference SM test. In the region adjacent to the porous media (0 < Y /h2 <
0.5), however, the Reynolds stresses are several times higher than the corresponding
SM values. These relatively higher values are partly explained by the less uniform
velocity distribution (i.e. large values of ∂ U/∂ y) adjacent to the porous medium
(as in Fig. 2a) and the concomitant enhanced production of turbulent kinetic energy
and Reynolds stresses in that region. It is also important to note that as the porosity
decreases, the turbulence level in the region 0 < Y /h2 < 0.5 increases and the location
of the peak value shifts away from the interface the clear zone. For the smooth test,
Fig. 5 Profiles of spatial averaged Reynolds stresses. Symbols are as defined in Fig. 2
the Reynolds shear stress also varied linearly from the peak values on either wall as
should be expected for a fully developed channel flow.
4 Summary and Conclusions

The present results demonstrate that the mean flow and turbulent characteristics are strongly
dependent on the specific properties of the porous media. For the mean flow, it was observed
that 99 percent of the approach flow is channeled through the clear fluid zone if the porosity
is 60 percent or lower. At these lower values of porosity, the slip velocity is only 5 percent
(or less) of the maximum streamwise mean velocity. The results also show that both the mass
fraction and dimensionless slip velocity varied linearly with permeability, irrespective of the
combination of rod diameter and spacing employed. Similar to rough-wall turbulent flows,
the mean velocity profile adjacent to the porous medium is less uniform compared with the
profiles obtained on a smooth wall. Similarly, the Reynolds stresses close to the porous media
are significantly higher than the corresponding smooth wall values.
Acknowledgments
The financial support of this work by the Natural Sciences and Engineering Re-
search Council of Canada through Undergraduate Summer Research Award to SB
and research grants to MFT is gratefully acknowledged.
References
[1] Breugem, W.-P., Boersma, B.J., Uittenbogaard, R.E.: The influence of wall permeability
on turbulent channel flow. J. Fluid Mech. 562, 35–72 (2006)
[2] Tachie, M.F., James, D.F., Currie, I.G.: Velocity measurements of a shear flow penetrating
a porous medium. J. Fluid Mech. 493, 319–343 (2003)
[3] Agelinchaab, M., Tachie, M.F., Ruth, D.W.: Velocity measurement of flow through a
model three-dimensional porous medium. Phy. Fluids 18, 017105, 11 (2006)
[4] Arthur, J., Ruth, D.W., Tachie, M.F.: PIV measurements of flow through a model porous
medium with varying boundary conditions. J. Fluid Mech. (2009) ( in press)
[5] Jackson, G.W., James, D.F.: The permeability of fibrous porous media. Can. J. Chem.
Eng. 64, 364–374 (1986)
[6] Shah, M.K., Tachie, M.F.: Flow relaxation past a transverse square rib in pressure gradi-
ents. AIAA J 46, 1849–1863 (2008)
A Model for Dissipation: Cascade SDE with
Markov Regime-Switching and Dirichlet Prior
D. Bernard, A. Tossa, R. Emilion, and S.K. Iyer
Abstract. Cascade Stochastic Differential Equation (SDE), a continuous time model

for energy dissipation in turbulence, is a generalization of the Yaglom discrete cas-
cade model. We extend this SDE to a model in random environment by assuming
that its two parameters are switched by a continuous time Markov chain whose states
represent the states of the environment. Moreover, a Dirichlet process is placed as a
prior on the space of sample paths of this chain. We propose a Bayesian estimation
method of this model which is tested both on simulated data and on real data of
wind speed measured at the entrance of the mangrove ecosystem in Guadeloupe.
Keywords: Cascade model, Dirichlet process, dissipation, Mangrove, Markov
regime switching, random environment, Stochastic Differential equation.
1 Introduction
The understanding of surface processes during the transition of an air mass through
a natural environment remains a challenge which notably consists in proposing
realistic models in environmental physics of Earth.
D. Bernard
LPAT, Université Antilles-Guyane, B.P. 250 - 97157 Pointe-à-Pitre, Guadeloupe
e-mail: didier.bernard@univ-ag.fr
A. Tossa
CEREMADE, Université Paris Dauphine, 75016 Paris, France
e-mail: adate.tossa@ceremade.dauphine.fr
R. Emilion
MAPMO, Université d’Orléans, B.P 6759 - 45067 Orléans Cedex 2, France
e-mail: richard.emilion@univ-orleans.fr
S.K. Iyer
Department of Mathematics, Indian Institute of Science, Bangalore 560 012 India
e-mail: skiyer@math.iisc.ernet.in
80 D. Bernard et al.
Our domain of study concerns the interactions between an atmospheric flow,

namely the Trade Winds, and a tropical ecosystem, namely the mangrove in Guade-
loupe (West Indies), which can be considered as a coastal barrier and a porous
swept obstacle to the flow. Moreover, this mangrove is also a continuing source
of gaseous compounds generating anaerobic degradation. These gases mixed with a
strong smell provoke an important stench and then are dispersed by winds around
and downstream the vegetal cover. Modelling such a surface process is therefore of
importance in the study of biogeochemical cycles of compounds (such as CO2) that
can influence climate change [4].
Predictive models of atmospheric flows, such as RANS and LES (Brunet, Dupont)
[3, 7, 11], usually deal with heavy deterministic numerical simulation codes. Al-
though well-adapted to the simulation of air flow in the inertial sub-layer within
the canopy, they induce some validation problems as they do not take into account,
for example, the cover turbulence specificity, the swept surfaces heterogeneity and
various terms in the overall equation of dissipation rate [6, 9]. Some fields, such
as the wind speed, are therefore roughly predicted while estimated second moment
profiles, such as flux or kinetic energy, do not match with the observed data. Fur-
thermore, as these models do not take in account the intermittent character of atmo-
spheric flow turbulence, they do not provide any prediction when dealing with high
fluctuations as is the case, for example, in the downstream of the cover [5]. In our
mangrove case, due to sea-surface transitions, even the input parameter profiles are
turbulent and intermittent.
An interesting indicator of turbulent energy production is the dissipation rate
which evaluates the amount of energy transferred at the entrance of the cover and is
of importance in the analysis of diffusion and mixture phenomena.
Celebrated Kolmogorov-Yaglom cascades model of turbulence dissipation has
been recently extended by F. Schmitt and D. Marsan [8], in a continuous-time set-
ting, to a model based on a stochastic differential equation (SDE), called here cascade
SDE, which depends on an intermittency parameter μ and a scale parameter λ .
In the present paper we show that a more realistic model consists in considering
this SDE in random environment, namely with the two above parameters driven by
a continuous-time Markov chain in a finite set representing the regimes of the en-
vironment. Each regime is therefore of random duration and within a regime, the
parameters μ and λ are constant. This setting is known in the litterature as Markov
regime switching.
An other point in our paper concerns a new Bayesian estimation method. As the
sequence of regimes, which is a path of the Markov chain, is unobservable, we pro-
ceed as follows. We place a Dirichlet process as a prior distribution on the path
space and then, simulating a great number of paths with suitable random weights,
we perform a Gibbs sampling algorithm in order to find the Markov chain and the
parameters with maximum likelyhood among a very large number of iterations.
The paper is structured as follows. Section 2 concerns the measurement set-
ting and Section 3 the cascade SDE. In Section 4, we specify our model and we
present its estimation procedure which hinges on some nontrivial computations of
A Model for Dissipation: Cascade SDE 81
conditional distributions. Section 5 deals with the numerical results obtained from
simulated data and wind speed data. We conclude in Section 6.
2 Measurements
Micrometeorological measurements were conducted by D. Bernard and C. d’Alexis
[1] at Arnouville (1661’N, 65,34’W), a mangrove area, located on the East coast of
Guadeloupe island. In April 2007 a meteorological mast was installed at 10m west
from the ocean in order to record vertical and horizontal velocity, wind direction,
temperature, pressure and specific humidity, at different heights.
Wind speed was recorded at 2m by a cup anemometer (A100R, Campbell), at 5m
by a 3D sonic anemometer (CSAT3, Campbell) and at 10m by a 2D sonic anemome-
ter (Gill). Sampling was done at 1Hz and 20Hz respectively. Wind direction was
recorded at 2m and 10m by a windvane (W200P, Campbell) and by the 2D Gill.
Temperature and relative humidity were recorded by a probe (HMP35C, Vaisala)
and barometric pressure by a sensor (PTB101B, Vaisala). All sensors worked
continuously during the monitoring period from April 1rst to August 1rst.
In Guadeloupe island, the referenced meteorological station is ”Le Raizet”. A
brief comparison of wind modulus between the Arnouville station and Le Raizet
one shows that the two signals are quite similar from 8 a.m. to 4 p.m.. But from
4 p.m. to 7 a.m., they appear significantly different. This is due to the fact that
the Arnouville regime is coastal with continuous blowing while Raizet one is
continental (see Brevignon [2]).
The following Fig. 1 put in evidence the intemittent and multifractal character of
the dissipation computed from our measurements.
0.02
0.018
0.016
0.014
0.012
Dissipation
0.01
0.008
0.006
0.004
0.002
0
0 100 200 300 400 500 600 700 800 900 1000
time (1/20s)
Fig. 1 Dissipation process at the entrance of the Guadeloupe mangrove
3 Cascade SDE
F. Schmitt and D. Marsan [10] introduced a stochastic dissipation process ε , with
a continuum of scale levels, which verifies a stochastic differential equation (SDE),
called here cascade SDE, which depends on an intermittency parameter μ and a

scale parameter λ . This SDE is a continuous version of the Yaglom multiplicative
cascade model [12]. Actually the singularity process γλ (t) = log(ελ (t)) satisfies the
SDE:
t
μ
γλ (t) = − ln λ + μ 1/2 (t + 1 − u)−1/2dB(u). (1)
2 t+1−λ
where B(.) is a standard Brownian motion.

We propose an extension of this cascade SDE to a model in random environment
by defining a dissipation process subject to regime changes. More precisely we as-
sume that the parameters μ and λ are driven by a continuous time Markov chain
(Xt )t≥0 that takes values in a set S = {1, 2, . . . , M} representing the environment
regimes.
4 Cascade SDE with Regimes and Dirichlet Prior

4.1 Model Specification
In order to take into account environment changes, we assume that the above
parameters λ and μ are themselves stochastic processes, namely governed by a
continuous-time Markov chain.
We consider the following cascade stochastic equation:
t
μX(t)
(t + 1 − u)−1/2dB(u),
1/2
γ (t) = − log(λX(t) ) + μX(t) (2)
2 t+1−λX(t)
where (Bt ) is a standard Brownian motion, X = (X(t))t≥0 is a continuous time

Markov chain taking values in a finite set S = {1, 2, . . ., M} that represents different
regimes of the environment.
μ = (μ1 , . . . , μM ) and λ = (λ1 , . . . , λM ) are random vectors such that their com-
ponents are independent and identically distributed (i.i.d.), that is:
i.i.d.
μi ∼ Γ1 f or i = 1, ..., M,
i.i.d.
λi ∼ Γ2 f or i = 1, ..., M,
Γ1 and Γ2 being arbitrarily fixed distributions.
The meaning of this setting is that during regime i the intermittency parameter
and the scale parameter are constant and equal to μi and λi respectively. If, at a given
time t, the state of the chain is i, that is if X(t) = i, (i ∈ S), then the environment is
in regime i so that μX(t) = μi and λX(t) = λi .
We complete the above model by placing a Dirichlet process (DP) as a prior on
the path space of the Markov chain X. This lead us to a Bayesian estimation of the
model:
⎧
⎨ (X|P, α ) ∼ P
(P|α ) ∼ D(α H) (3)
⎩
(α ) ∼ Gamma(η1 , η2 )
where H is the distribution of a specific Markov chain determined by a transition
rate matrix Q and an initial distribution π0 .
In practice, process (2) is sampled at time t1 , . . . ,tn , providing a finite random
vector γ = (γ1 , . . . γn ) of observations, where γi is the value recorded at time ti . We
prove that the conditional distribution of the vector γ is Gaussian so that we arrive
at the following hierarchical model that we want estimate from the observed data:
⎧
⎪
⎪ (γ |X, μ , λ , P) ∼ Nn (m(X), Σ )
⎪
⎪
⎪
⎪
i.i.d.
μ , ..., μM ∼ Γ1
⎪
⎨ 1 i.i.d.
λ1 , ..., λM ∼ Γ2 (4)
⎪
⎪ (X|P, α ) ∼ P
⎪
⎪
⎪
⎪ (P|α ) ∼ D(α H)
⎪
⎩
(α ) ∼ Gamma(η1 , η2 )
4.2 Conditional Distributions

The estimation procedure of the various parameters of the model hinges on Gibbs
sampling scheme which implementation requires the computation of the following
conditional distributions: (γ |X, μ , λ , p), (μ |γ , λ , X), (λ |γ , μ , X), (X|p, α ), (p|α ).
We proove that:
• (γ |X, μ , λ , p) ∼ Nn (m, Σ )
with m = − 12 μX(1) log(λX(1) ), . . . , − 12 μX(n) log(λX(n) ) and Σ = (σst )s,t=1,...,n
where
s t
σst =(μX (s) μX (t) )1/2COV ( (s+1−u)−1/2 dB(u), (t +1−u)−1/2 dB(u))
s+1−λX(s) t+1−λX(t)
• (X|p, γ , μ , λ ) ∝ ∑Ni=1 p∗i δXi where p∗i = pi fn (γ , μ , λ )

• (μ j |γ , λ j , X, α ) ∼ Γ1 ∝ Nn j (m j , μ j Σ j ).π1
• (λ j |γ , μ j , X, α ) ∼ Γ2 ∝ (γ |λ j , μ j , X, α )π2 .
4.3 Etimation Procedure

The estimation procedure is summarized in the following algorithm. In practice it
will be implemented using the Gibbs sampling technique.
1. Initialization
- Let γ be the vector of observations
- Choose the hyperameters η1 , η2 and N
- Generate α from Gamma(η1, η2 )
- Generate N paths of the Markov chain
- Draw p = p1 . . . pN from stick-breaking(α , N)

- Choose one of the N paths according to p
- Generate μ and λ from their priors
2. Iterations
- Compute fn (γ , μ , λ ) and update pk , k being the index of the chosen path.
- Choose one of the N paths according to p
- For each state j, draw μ j , drawλ j
- Draw α
- Draw a new p
5 Implementation and Numerical Results

5.1 Simulated Data
The present subsection aims at testing the reliability of the model. We perform nu-
merical simulations of the stochastic process γ (t). We consider a model with five
regimes. The associated Markov chain then has five states. We choose the prior of μ
and λ to be independent truncated Gaussian distributions, and simulate the param-
eters μ = (μ1 , ..., μ5 ) and λ = (λ1 , ..., λ5 ), coresponding to the five regimes. Taking
the data γ as input, we estimate the parameters of the model through the algorithm
presented. The values of the last iteration of each run were retained, leading to the
statistics in Table 1 and Table 2.
Table 1 Summary statistics of μ
Regime1 Regime 2 Regime 3 Regime 4 Regime 5

Simulated Values 0.19 0.33 0.36 0.41 0.45
Estimated Values 0.22 0.31 0.38 0.42 0.46
95% Conf. Int.a [0.189, 0.25] [0.285, 0.335] [0.35, 0.40] [0.405, 0.43] [0.44, 0.48]
a Conf. Int.: Confidence Interval
Table 2 Summary statistics of λ

Regime1 Regime 2 Regime 3 Regime 4 Regime 5
Simulated Values 1067 997 1234 1743 1408
Estimated Values 1070 996 1234 1742 1410
95% Conf. Int. [1064, 1075] [992, 1000] [1228, 1240] [1734, 1749] [1403, 1416]
5.2 Wind Speed Data

Considering the histogram of the values of μ (resp. of λ ) over sliding windows, a
truncated Gaussian (resp. a Gamma) distribution is taken as inital prior for μ (resp.
for λ ). The results of our algorithm for 500 Gibbs sampling runs of 25,000 iterations
each, including 3,000 burn-in iterations, are displayed in the following table. We
first ran the algorithm with M = 5 states but the values of the parameters of two of
these states were not significantly different so that we took M = 4 states. We also
evaluated the likelihood (LLH) of the selected path at each of the 25, 000 iterations
of a Gibbs sampling run and we chose the path with maximum likelihood. A sample
of the best paths is presented in Table 3.
Table 3 Likelihood of best paths
Runs 17 22 27 43 65 86 97 108 139 143

path 7 2 5 6 2 2 3 2 2 2
LLH 320 247 145 407 432 120 403 122 197 254
Runs 153 162 170 178 1185 188 194 195 198 200
path 2 6 4 3 7 5 2 4 2 2
LLH 436 83 233 128 107 92 98 218 320 218
Among these paths, the one with the highest likelihood has the following charac-
teristics given in Table 4 and Table 5.
Table 4 Characteritics of the path with highest likelihood
Regime 1 Regime 2 Regime 3 Regime 4

μ 0.3885 0.2703 0.1314 0.2914
λ 1281 1448 1367 1240
% of occupation 17.2% 14% 8.5% 60.3%
Table 5 Sequence of regimes in the highest likelihood path
Regimes 4 1 3 4 2 1
Duration 287 38 51 75 84 65
6 Conclusion
We have proposed a new model for dissipation which extends cascade SDE model
in random medium and in a Bayesian setting. This model can be seen as a complex
hierarchical mixture model. The numerical results obtained here lead us to think that
our mixture model better fit to measurements. It can be thought that such stochastic
modelling of dissipation could be of interest for future predictive models where
dissipation plays a key role.
References
1. Bernard, D., d’Alexis, C.: Experimental study of flow in a forest cover: Application to
the Guadeloupe mangrove ecosystem. Research technical report II, 20 p. (2008)
2. Brevignon: Atlas Climatique de la Guadeloupe 150, 395–409 (2006)
3. Dupont, S., Brunet, Y.: Simulation of turbulent flow in an urban forested park damaged
by a windstorm. Boundary-Layer Meteorology 120, 133–161 (2006)
4. Fluxnet: http://www.fluxnet.ornl.gov/fluxnet/index.cfm
5. Foudhil, H.: Développement d’un modèle numérique de dispersion atmosphérique de
particules à l’échelle d’un paysage hétérogène. Thèse de doctorat, Université de Bor-
deaux 1, Bordeaux, pp. 12–87 (2002)
6. Foudhil, H., Brunet, Y., Caltagirone, J.P.: A k-Model for Atmospheric Flow Over Het-
erogeneous Landscapes. Env. Fluid Mech. 5, 247–265 (2005)
7. Liu, J., Chen, J.M., Black, T.A., Novak, M.D.: k-Modelling of Turbulent Air Flow Down-
wind of a Model Forest Edge. Boundary-Layer Meteorology 77, 21–44 (1996)
8. Schmitt, F.G.: A causal Stochastic multifractal stohastic equation and its statistical prop-
erties. The European Physical Journal B 34, 85–98 (2003)
9. Schmitt, F.G.: About Boussinesq’s turbulent viscosity hypothesis: historical remarks and
a direct evaluation of its validity. C. R.Mécanique 335, 617–627 (2007)
10. Schmitt, F.G., Marsan, D.: Stochastic equations generating continuous multiplicative
cascades. The European Physical Journal B 20, 3–6 (2001)
11. Watanabe, T.: Large-Eddy Simulation of Coherent Turbulence Structures Associated
with Scalar Ramps Over Plant Canopies. Boundary-Layer Meteorology 112, 307–341
(2004)
12. Yaglom, A.M.: The influence of fluctuations in energy dissipation on the shape of turbu-
lent characteristics in the inertial interval. Soviet Physics Doklady 11, 26–29 (1966)
Wavelet Analysis of the Turbulent LES Data of
the Lid-Driven Cavity Flow
Roland Bouffanais, Guy Courbebaisse, Laurent Navarro, and Michel O. Deville
Abstract. Both Fourier and wavelet transforms are performed on data obtained from
large-eddy simulations of the turbulent flow in a lid-driven cubical cavity. The an-
alyzed data or synthetic signals are picked at three specific points inside the cavity
allowing to investigate three regimes over time: laminar, transitional and turbulent.
The main objective of this study is to generate and analyze synthetic signals in or-
der to confirm the correlation between the computed physical quantities and those
expected theoretically.
1 Introduction
The analysis of sampled signals obtained from experiments and direct numerical
simulations (DNS) of turbulent fluid flows through wavelet analysis is now common
Roland Bouffanais
MIT, 77 Massachusetts Avenue, Bldg 5–326, Cambridge, MA 02139
e-mail: bouffana@mit.edu
Guy Courbebaisse
CREATIS-LRMN, INSA-Bâtiment Blaise Pascal, 7 avenue Jean Capelle, F–69621
Villeurbanne cedex, France
e-mail: guy.courbebaisse@creatis.insa-lyon.fr
Laurent Navarro
CIS, École Nationale Supérieure des Mines de Saint-Étienne, 158 Cours Fauriel, F–42023
Saint-Étienne, France
e-mail: navarro@emse.fr
Michel O. Deville
École Polytechnique Fédérale de Lausanne, STI-IGM-LIN, Station 9, CH–1015 Lausanne,
Switzerland
e-mail: michel.deville@epfl.ch
88 R. Bouffanais et al.
practice. Such analysis often provides tremendous insight into the flow behavior
otherwise difficult, if not impossible, to apprehend with more conventional statisti-
cal signal analysis methods (e.g. Fourier transforms). The use of wavelet analysis to
study signals from large-eddy simulations (LES) is not as common, mainly because
of the intrinsic high level of non-physical noise introduced by the subgrid models. In
addition and in comparison with experiments and DNS, LES not only have a reduced
resolution in space but also in time. However, depending on the subgrid model and
the numerical method used, these difficulties may be overcome. In these conditions,
a time-frequency analysis, and in particular, its subdiscipline wavelet analysis could
provide a ‘local’ analysis of transient turbulent events. Such an approach has been
employed and is reported here for the study of the time histories of the pressure and
other fluctuating quantities in the locally-turbulent lid-driven cubical cavity flow.
2 LES of the Lid-Driven Cubical Cavity Flow

Large-eddy simulations of the turbulent flow in a lid-driven cubical cavity have been
carried out at a Reynolds number of 12 000 using the Legendre spectral element
method [2]. Two distinct subgrid-scales models, namely a dynamic Smagorinsky
model and a dynamic mixed model, have been both implemented and used to per-
form long-lasting simulations required by the relevant time scales of the flow. The
resolution in time is high enough so that subgrid modeling is only needed to account
for the reduced resolution in space. Practically, it is usually more cost-effective to
be under-resolved in space than in time. In addition, being under-resolved both in
space and time leads to unavoidable numerical instabilities, e.g. CFL criterion not
respected, etc. All filtering levels make use of explicit filters applied in the physical
space (on an element-by-element approach) and spectral (modal) spaces. The two
subgrid-scales models had been validated and compared to available experimental
and numerical reference results, showing very good agreement. Specific features of
lid-driven cavity flows in the turbulent regime, such as inhomogeneity of turbulence,
turbulence production near the downstream corner eddy, small-scales localization
and helical properties have been investigated and discussed in the large-eddy sim-
ulation framework [2]. Time histories of quantities such as the total energy, total
turbulent kinetic energy or helicity exhibit different evolutions but only after a rel-
atively long transient period. At a Reynolds number of 12 000, the lid-driven cavity
flow is in the locally-turbulent regime and is proved to be highly inhomogeneous in
the secondary-corner regions of the cavity where turbulence production and dissi-
pation are important [2]. The maximum production of turbulence was found to be
located in the downstream-corner-eddy region just above the bottom wall.
3 Spectral Analysis and Wavelet Analysis

The time series of three quantities, namely pressure, local kinetic energy and x-
component of the velocity field, have been extracted from the LES databases
Wavelet Analysis of the Turbulent LES Data of the Lid-Driven Cavity Flow 89
at three distinct locations inside the cavity. Each extraction point has been
purposedly chosen to characterize the three different regimes—laminar, transitional,
turbulent—encountered within the lid-driven cavity flow at Re = 12 000. The Carte-
sian coordinates of these three points, inside the cavity, corresponding to a larmi-
nar, transitional and turbulent region of the flow are respectively given by (0, 0, 0),
(0.65, −0.6, −0.6) and (0.78, −0.94, −0.337) in units of the cavity length[2]. The
groundbreaking work of Kolmogorov [5, 6] highlight the fact that the velocity fluc-
tuations of a turbulent flow can be analyzed and characterized based upon the be-
haviors of the scales of the spatial increment of the Eulerian velocity or the temporal
increment of the Lagrangian velocity. The central postulate of statistical isotropy of
small spatial and temporal scales, used by Kolmogorov in [5] is connected to the
independence of these small scales with respect to the mechanism of injection of
energy which occurs for the large eddies. In addition, the statistical analysis of the
temporal fluctuations of the pressure field allows one to study the vortically-intense
regions of the flow [1]. Given the fact that the pressure field is the solution of a Pois-
son equation (the flow being incompressible), the vortical structures are therefore
directly connected to the rapid changes the temporal signal of the pressure field, see
Fig. 1.
−3
0.1 x 10
2
Velocity Laminar
Pressure Laminar
0.05
0
0
−0.05 −2
−0.1 −4
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200
Time Time
0.2 0.01
Velocity Transition
Pressure Transition
0.1 0.005
0 0
−0.1 −0.005
−0.2 −0.01
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200
Time Time
0.2 0.1
Pressure Turbulent
Velocity Turbulent
0.1 0.05
0 0
−0.1 −0.05
−0.2 −0.1
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200
Time Time
Fig. 1 x-component of the velocity field (left column) and the pressure field (right column)
for the three regimes
A first relevant analysis to give an overall validation of the simulated system

consists in the Fourier transform computation applied to the velocity signal and to
the pressure signal (Fig. 1) at the point of maximum turbulence production within
the cavity. The Fourier transform can be written as

S(ν ) = s(t)e−2iπν t dt, (1)
where s(t) is the analyzed signal, t the time and ν the frequency. The results of
spectral analysis for the intensity (square modulus of the quantity)in a Log-Log
scale are shown in Fig. 2 where a −5/3 (resp. −7/3) slope is observed for the
velocity (resp. pressure) signal. These results which are characteristics of the devel-
oped turbulence region of the cavity flow are in good agreement with those predicted
by Kolmogorov’s statistical theory of the turbulence K41 [5, 9]. Slopes were con-
ventionally estimated using a linear regression method.
A second appropriate analysis consists in the continuous wavelet transform
analysis [7]. The wavelet transform is expressed by

1 t −b
CΨ [s](a, b) = Ψa,b , s = Ψ a,b (t)s(t) dt, with Ψa,b = √ Ψ , (2)
a a
Fig. 2 Fourier transform of the time histories of the velocity and pressure signals at the point
of maximum turbulence production within the cavity
Fig. 3 Mexican hat wavelet
and corresponds to the inner product of the signal s(t) with the successive versions
of the mother wavelet Ψa,b , where a is a real positive parameter, b a real param-
eter, and the overline denotes the complex conjugate. The selected wavelet must
verify the following admissibility condition Ψ (t) dt = 0. In the sequel, all wavelet
analyses are based on the so-called ‘Mexican hat’ wavelet (Fig. 3) expressed by
2π −1/4
√ (1 − t 2)e−t /2 .
2
Ψ (t) = (3)
3
This wavelet is directly connected to the negative normalized second derivative of

a Gaussian (Fig. 3). This wavelet is C∞ and is well localized in the time domain as
well as in the frequency domain. In addition, its two first moments vanish and verify

t qΨ (t) dt = 0, 0 ≤ q < N. (4)
This property is essential in analyzing the singularities within a signal [1].

The continuous wavelet transform CΨ [s](a, b) is now a conventional tool for the
analysis of singularities in a signal at an instant t0 , and hence allows to expand the
concept of ‘singularity exponent’. It is well-known that the wavelet transform, near
a singular point b = t0 , behaves like a power law according to the scale with the
Hölder exponent h(t0 )
|CΨ [s](a,t0 )| ∝ ah(t0 ) , (5)
where the Hölder exponent h(t0 ) is a measure of the strength of the singularity [1]. In
the present framework, given the (inhomogeneous) production of turbulence within
the cavity flow [2], a multi-resolution analysis of some signals extracted from the
LES database is well suited. The results of the wavelet transform computation are
traditionally presented in a graph with a horizontal axis for the time, a vertical one
for the scale a and flooded color contours of the continuous wavelet transform. As
a first step, the continuous wavelet transform of the LES signals is used to reveal
the changing patterns between the three regimes, laminar, transitional and turbulent.
The wavelet analysis of the pressure signal appears in Fig. 4 and in Fig. 5. Figure 4
given in grey levels, reveals the difference of the signal amplitudes considering
the three regimes. Figure 5 highlights the gradual emergence of ‘time-scale’ pat-
terns at large scales. In addition, these patterns grow in size when moving from the
transitional regime to the turbulent one as shown in Fig. 6.
Pressure Laminar log(WT²) Pressure Transition log(WT²) Pressure Turbulent log(WT²)
−5 −5 −5
50 50 50
100 −10 100 −10 100 −10

Scale
Scale
Scale
150 150 150

−15 −15 −15
200 200 200
250 −20 250 −20 250 −20

0 1 2 3 0 1 2 3 0 1 2 3
Time Time Time
Fig. 4 Logarithm of the square of the continuous wavelet transform of the pressure field at
the three locations of the cavity corresponding to laminar, transitional and turbulent regimes,
from left to right respectively
To quantify the turbulence, a powerful yet straightforward technique consists in

implementing the wavelet transform modulus maxima method [1]. More specifi-
cally, one aims at obtaining the skeleton defined by lines of maxima of the wavelet
transform modulus, which is calculated for the pressure signal in both laminar and
turbulent regimes. The results are shown in Fig. 7 where filaments appear in both
Pressure Laminar CWT −3 Pressure Transition CWT Pressure Turbulent CWT

x 10
4
50 2 50 50
0.01 0.1
100 0 100 100
Scale
Scale
Scale
0 0
150 −2 150 150
−0.01 −0.1
200 −4 200 200
250 −6 250 250 −0.2

1 2 3 1 2 3 1 2 3
Time Time Time
Fig. 5 Continuous wavelet transform of the pressure field at the three locations of the cavity
corresponding to laminar, transitional and turbulent regimes, from left to right respectively.
Red (resp. blue) represents positive (resp. negative) contour levels
Fig. 6 3D wavelet transform of the pressure for the laminar (Left), the transitional (Center)
and the turbulent (Right) regimes
Skeleton of Wavelet Transform

Pressure signal − Laminar Skeleton of Wavelet Transform
7 Pressure signal − Turbulent
7
7.5 7.5
8
8
8.5
log2(scale)
log2(scale)
8.5 9
9.5
9
10
9.5
10.5
11
10 0 100 200 300 400 500 600 700 800 900 1000
0 100 200 300 400 500 600 700 800 900 1000
Time
Time
Fig. 7 Top row: skeleton of the wavelet transform of the pressure field for both the laminar
(left column) and turbulent (right column) regimes. Bottom row: amplitude of the wavelet
transform along the longest filaments of the skeleton
cases. The longest filaments are collected and represented in Fig. 7 (bottom row)
where the scale is on the horizontal axis and the magnitude of the filaments on the
vertical one. The laminar filaments appear almost constant in amplitude, whereas
for the ones for the turbulent regime highlight a −2/3 slope in the large scales. This
−2/3 slope corresponds to very large depressions associated with the turbulence
burst occurring when a pair of counter-rotating vortices is produced. This pair of
counter-rotating vortices has been identified as the highly-coherent vortical struc-
ture responsible for the production of turbulence [2]. Herein, it is identified by very
strong singularities characterized by a negative Hölder exponent.
The above results for both the transitional and turbulent regions of the flow pro-
vide structures possessing a clear fractal signature, which is also related to the high
level of singularities in the Hölder spectra.
4 Towards a Spectrum of Singularities

To take into account all singularities in the LES signal, one has to resort to a
multi-fractal model and hence study the spectrum of Hölder exponent, also known
as ‘spectrum of singularities’ [9]. A common technique consists in applying the
wavelet transform modulus maxima method (WTMM). Unlike the ‘traditional’ tech-
nique employed in Sec. 3, the WTMM allows one to estimate the spectrum of sin-
gularities without accounting for the singularities of negative Hölder exponent [1].
Considering a multi-fractal formalism [9], the K41 theory [5] leads to a statis-
tical homogeneous velocity field characterized by only one Hölder exponent when
the Reynolds number tends to infinity. For small scales, the properties of scale in-
variance of the Navier-Stokes equations are statistically preserved. But when con-
sidering intermittency phenomena in fluid flow, the hypothesis of homogeneity falls
because of these small scales . An improvement of the earlier statistical theory of
homogeneous and isotropic turbulence K41 [5] is given by the KO62 [6, 8] theory
which allows to consider intermittent phenomena. Similarity assumptions (statisti-
cal relations between local fluctuations of the velocity field and fluctuations of the
dissipative energy field) in the case of isotropic turbulence can be understood within
the multi-fractal framework [1, 3, 4, 7]. However, the computation of the spectrum
of singularities requires signals of several hundred thousand samples, which are un-
fortunately not available at the present time. On-going computations should be able
to feed the system in order to achieve higher statistical sampling and hence calculate
the spectrum of singularities D(h) by a multi-fractal formalism based on the wavelet
transform modulus maxima method [1].
5 Conclusions
In the present work, wavelet analysis and other methods derived from it are em-
ployed to treat LES signals and to characterize the locally turbulent flow in a lid-
driven cubical cavity. Despite the inherently low space and time resolution of the
LES signals, the wavelet analysis reveals some very interesting features through
its Hölder exponent in a multi-fractal framework. Comparisons of these results for

three different locations corresponding to three different turbulent regimes show the
effectiveness of this approach for a locally inhomogeneous and anisotropic turbulent
flow such as the one in the lid-driven cavity at a Reynolds number of 12 000. This
first step allows to consider a more general and systematic investigation and charac-
terization of physical quantities computed through LES in other flow configurations.
Comparisons with a similar analysis of the same flow but obtained by DNS [10] is
now considered.
References
1. Arneodo, A., Argoul, F., Bacry, E., Elezgaray, J., Muzy, J.F.: Fractales, Ondelettes et
Turbulence: de l’ADN aux croissances cristallines. Diderot Edn. (1995)
2. Bouffanais, R., Deville, M.O., Leriche, E.: Large-eddy simulation of the flow in a lid-
driven cubical cavity. Phys. Fluids 19, Art. 055108 (2007)
3. Carmona, R., Hwang, W.L., Torrésani, B.: Practical time-frequency analysis. Academic
Press, London (1997)
4. Jaffard, S.: Multifractal formalism for functions. Part 1 & 2. SIAM J. of Math.
Anal. 28(4), 944–998 (1997)
5. Kolmogorov, A.N.: The local structure of turbulence in incompressible viscous fluid for
very large Reynolds numbers. C. R. Acad. Sci. USSR 30, 301, 299–303 (1941)
6. Kolmogorov, A.N.: A refinement of previous hypotheses concerning the local struc-
ture of turbulence in a viscous incompressible fluid at high Reynolds number. J. Fluid
Mech. 13, 82–85 (1962)
7. Mallat, S.: A wavelet tour of signal processing, 2nd edn. Academic Press, London (1999)
8. Oboukhov, O.M.: Some specific features of atmospheric turbulence J. Fluid Mech. 13,
77 (1962)
9. Frisch, U.: Turbulence: The legacy of A.N. Kolmogorov. Cambridge University Press,
Cambridge (1995)
10. Leriche, E.: Direct numerical simulation in a lid-driven cavity at high Reynolds number
by a Chebyshev spectral method. J. Sci. Comput. 27(1-3), 335–345 (2006)
A Two-Phase LES Compressible Model for
Plasma-Liquid Jet Interaction
Céline Caruyer1, Stéphane Vincent2 , Erick Meillot1 , and Jean-Paul Caltagirone2
Abstract. The numerical simulation of the interaction between a plasma flow and a
liquid jet is important for understanding and predicting the physical parameters in-
volved in plasma spraying processes. This work proposes an original model for deal-
ing with three-dimensional and unsteady turbulent interactions between a plasma
flow and a liquid water jet. A compressible model, based on augmented Lagrangian,
Large Eddy Simulation (LES) turbulence modeling and Volume of Fluid (VOF) ap-
proaches, capable of managing incompressible two-phase flows as well as turbulent
compressible motions is presented.
1 Introduction
The plasma spray process is a widely used technique for producing surface coatings
for various industrial applications. To decrease the deposit thickness, nanometric
particles are injected in the plasma jet with a liquid precursor. According to the
distance of the injector, a continue or split liquid jet penetrates the plasma flame.
In the literature, few models concerning the introduction of a liquid phase into a
plasma flow exist but all belonging to the class of Lagrangian models which are
based on the tracking of particles in the plasma flow [9].
For the plasma modelling, a lot of works are based on statistical continuum ap-
proaches of Reynolds Averaged Navier-Stokes model type [13][12][11]. A deter-
ministic turbulent Large Eddy Simulation approach, which solves the large scales
of the turbulence and models the small dissipative characteristics of the motion is
used in this work. This method allows to reach unsteady information. Simulating
the interaction between a liquid jet and a plasma flow with an unsteady model is a
recent practice [19].
1 CEA-DAM Le Ripault, BP 16, 37260 Monts, France
2 Tranferts Ecoulements Fluides Energétique (TREFLE), UMR CNRS 8508, ENSCPB,
Université Bdx1
96 C. Caruyer et al.
Compressible effects must be taken into account in the motion equations because
of the plasma flow Mach number range (0.3 to 0.7). Few works exist in the literature
on 1-fluid models which are compressible, multifluid and made of separated phases
[2][14][15][26]. The incompressible single fluid described by Kataoka [7] and
Scardovelli et al.[17] is adapted for compressible flows (Vincent et al., under sub-
mission [3]) in order to manage compressible plasma motions and incompressible
liquid jet features of the two-phase flow.
2 Numerical Model
2.1 Modelling
The motion equations dedicated to a deterministic description of the compressible
two-phase flow are represented by a generalization of the incompressible 1-fluid
model [7] [17] with a new mass conservation equation which is reformulated as
an equation for the pressure. The details of this specific aspect of the model are
submitted for publication [3]. The low-Mach compressible one-fluid model reads:
⎧
⎪ ∂p 1
⎪
⎪ + ∇·u = 0
⎪
⎪ ∂t χT
⎪
⎪
⎪
⎪
⎨
∂u τ
ρ + (u · ∇)u = −∇ p − ∇ · u + ρ g + ∇ · μ
∇u + ∇t u + FT S (1)
⎪
⎪ ∂t χT
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩ ∂C
+ (u · ∇)C = 0
∂t
where u is the local velocity, p the pressure, t the time, g the gravitational accelera-
tion, χT = ρ1 ddρp the adiabatic compressibility, τ a characteristic time of compressible
effects linked to inertial time scales of the problem, C the local liquid volume frac-
tion, μ
the dynamic viscosity, ρ the density and FT S the surface tension force.
The thermal exchanges are taken into account by employing the energy equation
formulated in terms of temperature:
∂T
ρC p ( + (u · ∇)T ) = ∇ · (
λ ∇T ) + Φray (2)
∂t
with T the temperature, C p the specific heat, λ the conductivity, Φray a volume
energy density term representative of radiative effects.
The advection and diffusion of the the different species (plasma and air) concen-
tration is described by the following equation:
∂ ψi i ∇ψi )
+ (u · ∇)ψi = ∇ · (D (3)
∂t
A Two-Phase LES Compressible Model for Plasma-Liquid Jet Interaction 97
where ψi et D i are respectively the concentration and the diffusion coefficient of

species i. The local thermo-dynamic charateristics ρ, μ, are dependent on
λ , Cp , D
local temperature T, concentration ψi and the phase function C.
The Large Eddy Simulation (LES) is based on a concept of scale separation
where the large scales, which drive the flow dynamics, are solved explicitly, whereas
the small scales are modelled thanks to a so-called subgrid-scale model (SGS). Sim-
ulations can be run on coarse grids which cannot take into account the small dissipa-
tive scales of the flow. The separation between the resolved and the modeled scales
is mathematically formalized by applying a convolution filter to the Navier-Stokes
equations, which introduces the notion of cut-off length. For two-phase flows, ad-
ditional subgrids terms exist[23]. However, as no closure models are proposed to
evaluate the new subgrid terms, a simpler one- fluid model for LES has been used
in this work as a first approach, with the assumption that the subgrid terms resulting
from the filtering of two phase terms are considered to be negligible.
The Smagorinsky model, in which the eddy viscosity, based on the resolved
scales, is defined as:
νT (x,t) = (CS Δ )2 |S| (4)
where |S| = 2Si j Si j with Si j is the strain rate tensor and CS is the Smagorinsky
constant [18]. In our work, a value of 0.2 is used.
The Mixed Scale model [16], which is derived by taking a weighted geometric
average of the previously detailed Smagorinsky model and of the Turbulent Kinetic
Energy (TKE) approach defined by:
1
νT (x,t) = CT KE Δ (q2SGS (x,t)) 2 (5)
The Mixed Scale model exhibits a triple dependency on the large and small struc-
tures of the resolved field as a function of the cut-off length.[16]
In order to improve the previously detailed subgrid-scale models, a sensor based
on local information of the flow [4] is used. This sensor, or selection function, is
related to the local angular fluctuations of the vorticity ω such as:
νTS (x,t) = νT (x,t) × f (w,t) (6)
The dynamic viscosity μ

is the sum of the molecular μm and the turbulent viscosities
μT = νT × ρ :

μ
= μm + μT (7)
For the energy and transport equations, turbulent models based on Schmidt and
Prandtl numbers are used to take into account the LES subgrid effects on tempera-
tures and species [20].
2.2 Numerical Methods

The simulation tool is based on the Computational Fluid Dynamic library Aquilon,
developed at the TREFLE laboratory. The Navier-Stokes equations are discretized
by an implicit finite volume method on a staggered mesh. An iterative bi-conjugate
gradient stabilized BICG-Stab II [24], preconditioned under a Modified and In-
complete LU method [6] is implemented. The spatial discretizations are centered
schemes. The compressible mass and momentum equations are discretized in time
with an augmented Lagrangian-like method [22]. The numerical methods have been
detailed and validated in previous articles [22][20][21].
2.3 Physical Parameters

For the single phase argon plasma flow, the calculation domain is a box of dimension
85mm × 100mm × 100mm. The calculation grid contains 486720 cells. The mesh is
refined in the central zone where the plasma is. The unsteady time step is equal to
10−6 s. The plasma torch is not explicitly simulated. The temperature, concentration
of Ar (equal to 1) and velocity outlet conditions are enforced at the boundary of the
calculation domain with the following temperature and velocity contours:
⎧ r nT
⎪
⎪ T (r) = (T (0) − T (R)) 1 − + T (R)
⎨ R
n (8)
⎪
⎪
⎩ u(r) = u(0) 1 − r V
R
where r is the radial position. The parameters T(0)=12913, u(0)=1092, nT = 2.3 et
nV = 1.4 are determined from experimental data, mass flow and nominal power of
the torch [25]. The tabulated characteristics in air and Ar plasma are given in
Figure 1. These data are obtained by polynomial extrapolation of calculations and
measurements realized at the French Atomic Energy Commission CEA-Ripault.
3 Single Phase LES Plasma Simulation

The numerical simulation of a high velocity, high temperature argon plasma jet is-
suing into air is investigated in order to validate the compressible LES numerical
model. Four LES models are tested: Smagorinsky, Smagorinsky with the selection
function, Mixed Scale and Mixed Scale with the selection function. Figure 2 (left)
compares the mean temperature for the four LES models along the centerline axis
of the plasma to the measurements of Fincke et al. [5].
The use of the Mixed Scale model seems to give results closer to experiments in
comparison with the Smagorinsky model, which is in agreement with the literature.
For both models, the simulations with the selection function give better results. The
selection function allows to adapt the turbulent viscosity according to the level of the
flow turbulence. In Figure 2 (right), the turbulent viscosity is too important without
Ar Ar
5 Air 1.5 Air
4
Conductivity [W/m.K]
Density [kg/m3]
1
3
2
0.5
0
5000 10000 1500 0 5000 10000 15000
Temperature [K] Temperature [K]
Ar
Air
Ar
20000 Air
0.00025
Dynamic viscosity [Pa.s]
Specific heat [J/kg.K]
0.0002 15000
0.00015
10000
0.0001
5000
5E-05
0
5000 10000 15000 5000 10000 15000
Temperature [K] Temperature [K]
Fig. 1 Tabulations of the characteristics for Air and Argon
the selection function in the zone where the flow is not turbulent, i.e. at the center
of the plasma.
4 Multiphase Plasma Simulation

The injection of water droplets (diameter 150 μ m) in a Ar/H2 plasma flow is
considered, corresponding to a fragmented jet penetrating a plasma motion. The
calculation domain is a box of dimension 0.5mm*0.4mm*0.3mm with a mesh of
125 × 125 × 100 cells. The droplet is injected with a velocity of 30m.s−1 vertically
downwards. The velocity of the cross flow, corresponding to peripheral velocity
conditions in the plasma, is included between 0 and 250m.s−1 . The time step used
for the calculations is 10−8 s. In this case, thermal phenomena are not taken into
ρ f U02 d0
account. The Weber number is We = = 11 with ρ f the fluid density, U0
σ
Smagorinsky
Smagorinsky
Mixed Scale
Mixed Scale
Smagorinsky with selection
Smagorinsky with selection
Mixed Scale with selection
1 Mixed Scale with selection
Experiments
0.00035
Normalized mean temperature
0.8 0.0003
Turbulent viscosity [m2/s]

0.00025
0.6
0.0002
0.00015
0.4
0.0001
0.2 5E-05
0.02 0.04 0.06 0.08 -0.04 -0.02 0 0.02 0.04

X[m] r [m]
Fig. 2 Comparison between four different LES models with experiments [5]. Left: Mean
normalized temperature along the centerline axis of the plasma. Right: Turbulent viscosity in
the radial direction at x=40mm.
Fig. 3 Water droplet injected in a plasma flow at different times: 10−6 s, 5.10−6 s, 9.10−6 s
and 14.10−6 s
the relative velocity, d0 the drop diameter and σ the surface tension. This Weber
number corresponds to the bag breakup regime.
Figure 3 presents preliminary results iso-surface of the water concentration
C=0.5. First, the droplet evolves from its initial spherical shape into an oblate
spheroid. Then a bag, attached to a ring, forms and finally breaks up forming small
fragments. A typical shape of bag break-up is observed, in concordance with the
Weber number [8].
5 Conclusions and Perspectives

A compressible Eulerian model is proposed for the numerical simulation of plasma
flows interacting with a liquid jet. The turbulent character of the flow can be mod-
eled thanks to a Large Eddy Simulation approach. First, LES Mixed Scale and
Smagorinky models are tested on a single-phase argon plasma flow with and without
the selection function. Results are better with a selection function. A good agree-
ment is obtained with regard to mean temperature as compared to the measurements
of Fincke et al. [5].
Then the interaction of a liquid drop in a Ar/H2 plasma is simulated. A lo-
cal study of droplet shape is carried out to determine, in the future, the appropri-
ate macroscopic fragmentation models useful for small scale two phase modelling
[10] [1].
Future works will be devoted to an extension of the compressible model in the
case of heat transfer and compressible flows. Thermal effects must be added to the
mass conservation equation written in terms of pressure equation. Moreover, spe-
cific subgrid models related to jet fragmentation and phase change will be added to
the Eulerian models.
Acknowledgments
The authors wish to thank the Aquitaine Regional Council for the financial support dedicated
to a 256-processor cluster investment, located in the TREFLE laboratory. We are grateful for
access to the computational facilities of the french CINES (National computing center for
higher education) under project number x2009026115.
References
1. Basu, S., Jordan, E.H., Cetegen, B.K.: Fluid mechanics and heat transfer of liquid pre-
cursor droplets injected into high-temperature plasmas. J. Therm. Spray Tech. 17, 60–72
(2008)
2. Caiden, R., Fedkiw, R.P., Anderson, C.: A numerical method for two phase flow con-
sisting of separate compressible and incompressible regions. J. Comput. Phys. 166, 1–27
(2001)
3. Caltagirone, J.P., Vincent, S., Caruyer, C.: A multiphase compressible model for the
simulation of multiphase flows. Physics of Fluids (under submission)
4. David, E.: Modélisation des écoulements compressibles et hypersoniques. PhD thesis,
Institut National Polytechnique de Grenoble (1993)
5. Fincke, J.R., Crawford, D.M., Snyder, S.C., Swank, W.D., Haggard, D.C., Williamson,
R.L.: Entrainment in high-velocity, high-temperature plasma jets. Part I: experimental
results. Int. J. Heat Transfer 46, 4201–4213 (2003)
6. Gustafsson, I.: On first and second order symmetric factorization methods for the solu-
tion of elliptic difference equations. Chalmers, University of Technology 1 (1978)
7. Kataoka, I.: Local instant formulation of two-phase flow. Int. J. Multiph. Flow 12, 745–
758 (1986)
8. Liu, Z., Reitz, R.D.: An analysis of the distorsion and breakup mechanisms of high speed
liquid drops. Int. J. Multiphase Flow 23, 631–650 (1997)
9. Marchand, C., Vardelle, A., Mariaux, G., Lefort, P.: Modelling of the plasma spray pro-
cess with liquid feedstock injection. Surf. Coat. Technol. 202, 4458–4464 (2008)
10. Marchand, C., Vardelle, A., Mariaux, G., Lefort, P.: Modelling of the plasma spray pro-
cess with liquid feedstock injection. Surf. Coat. Technol. 202, 4458–4464 (2008)
11. Mariaux, G., Vardelle, A.: 3-D time-dependent modelling of the plasma spray process.
part I: flow modelling. Int. J. Therm. Sci. 44, 357–366 (2005)
12. Meillot, E., Guenadou, D.: Thermal plasma flow modeling: A simple model for gas heat-
ing and acceleration. Plasma Chem. Plasma Process. 24, 217–238 (2004)
13. Meillot, E., Guenadou, D., Bourgeois, C.: Three-dimension and transient D.C. plasma
flow modeling. Plasma Chem. Plasma Process. 28, 69–84 (2007)
14. Nourgaliev, R.R., Dinh, T.N., Theofanous, T.G.: Adaptive characteristics-based matching
for compressible multifluid dynamics. J. Comput. Phys. 213, 500–529 (2006)
15. Nourgaliev, R.R., Theofanous, T.G.: High-fidelity interface tracking in compressible
flows: Unlimited anchored adaptive level set. J. Comput. Phys. 224, 836–866 (2007)
16. Sagaut, P.: Large eddy simulation for incompressible flows - An introduction. Springer,
Heidelberg (1998)
17. Scardovelli, R., Zaleski, S.: Direct numerical simulation of free-surface and interfacial
flow. Ann. Rev. Fluid Mech. 31, 567–603 (1999)
18. Smagorinsky, J.: General circulation experiments with the primitive equations. I: The
basic experiments. Month. Weath. Rev. 91(3), 99–165 (1963)
19. Trelles, J.P., Pfender, E., Heberlein, J.: Multiscale finite element modeling of arc dynam-
ics in a d.c. plasma torch. Plasma Chem. Plasma Process. 26, 557–575 (2006)
20. Vincent, S., Balmigere, G., Caruyer, C., Meillot, E., Caltagirone, J.: Contribution to the
modeling of the interaction between a plasma flow and a liquid jet. Surf. Coat. Technol.
(2008), doi:10.1016/j.surfcoat.2008.11.009
21. Vincent, S., Caltagirone, J.P.: Efficient solving method for unsteady incompressible flow
problems. Int. J. Num. Meth. Fluids 30, 795–811 (1999)
22. Vincent, S., Caltagirone, J.P., Lubin, P., Randrianarivelo, T.N.: An adaptative augmented
lagrangian method for three-dimensional multimaterial flows. Comput. Fluids 33, 1273–
1289 (2004)
23. Vincent, S., Larocque, J., Lacanette, D., Toutan, A., Lubin, P., Sagaut, P.: Numerical
simulation of phase separation and a priori two-phase les filtering. Comput. Fluids 37,
898–906 (2008)
24. Van Der Vorst, H.A.: A fast and smoothly converging variant of bi-cg for the solution of
non-symmetric linear systems. J. Sci. Stat. Comput. 44, 631–644 (1992)
25. Williamson, R.L., Fincke, J.R., Crawford, D.M., Snyder, S.C., Swank, W.D., Haggard,
D.C.: Entrainment in high-velocity, high-temperature plasma jets. Part II: computational
results and comparison to experiment. Int. J. Heat Transfer 46, 4215–4228 (2003)
26. Yabe, T., Yuan, P.Y.: Unified numerical procedure for compressible and incompressible
flow. J. of The Physical Society of Japan 60, 2105–2108 (1991)
Simulation of a Fluidized Bed Using a Hybrid
Eulerian-Lagrangian Method for Particle
Tracking
Cédric Corre, Jean-Luc Estivalezes, Stéphane Vincent, Olivier Simonin, and

Stéphane Glockner
Abstract. The characterisation of fluidized beds still requires specific investigation

for understanding and modelling the local coupling between the dispersed phase and
the carrier fluid. The aim of this work is to simulate this type of unsteady particle
laden flows via Direct Numerical Simulations in order to provide a local and instan-
taneous description of particle flow interactions and to extract statistical parameters
useful for large scale models. A fluidized bed has been studied experimentally by
Aguilar Corona ([1]). In this laboratory experiment, 3D tracking of a single bed
particle provided Lagrangian properties of the discrete phase motion, while 2D PIV
was used to characterize the flow of the continuum phase. This fluidized bed has
been simulated during nine seconds in order to compare experimental and numeri-
cal results and to obtain some data that experimental studies can’t give.
1 Introduction
Particles and processes involving particles are of paramount importance in the
chemical and applied industries. Fluidized beds are widely employed in industrial
C. Corre · J.L. Estivalezes
ONERA/DMAE, 2 avenue Edouard Belin, BP 74055, 31055 Toulouse, France
e-mail: ccorre@onera.fr;Jean-Luc.Estivalezes@onera.fr
S. Vincent · S. Glockner
Université de Bordeaux, TREFLE-ENSCPB, UMR 8508 CNRS, 16 avenue Pey-Berland,
33607 Pessac Cedex, France
e-mail: vincent@enscpb.fr;glockner@enscpb.fr
O. Simonin
Université de Toulouse; INPT, UPS; IMFT (Institut de Mécanique des Fluides de
Toulouse); Allée Camille Soula, F-31400 Toulouse, France
CNRS ; IMFT ; F-31400 Toulouse, France
e-mail: simonin@imft.fr
104 C. Corre et al.
operations, ranging from the pharmaceutical and food industry, to processes such as
catalytic cracking of petroleum, combustion and biomass gasification. The success
of fluidized bed processes is due to their excellent heat and mass transfer character-
istics. Although rather simple in its conception, the application of a fluidized process
still faces some challenges: a sound understanding of the mechanisms governing the
complex flow phenomena involved in a fluidized bed still remains an open technical
and scientific issue. The simulation of a fluidized bed is a possible research way to
access to local and unsteady flow informations by using a Direct Numerical Simu-
lation (DNS). Indeed, the DNS mesh grid is smaller than particle diameter and fur-
thermore, no explicit models are used to take into account phenomenons occuring
in a fluidized bed. Previous DNS of fluidization of spheres have been performed
using the method of distributed Lagrange multipliers (Pan [7]). The originality
of the numerical methods is based on a generalization of the one-fluid formulation of
the incompressible Navier Stokes equations for particulate flows. The presence of
the particles is modelled with a fictitious domain method. The particle tracking is ob-
tained via an original hybrid Eulerian-Lagrangian Volume Of Fluid (VOF) method.
Particles are managed in a Lagrangian manner whereas the local solid fraction is de-
scribed on an Eulerian way. At each time step, the particle characteristics are stored
and statistics are performed on the particle-particle and fluid-particle interactions.
2 Numerical Methods
Modelling incompressible fluids and two-phase flows can be achieved by the Navier-
Stokes equations in their 1-fluid formulation. This fluid formalism is extended to the
particle fluid two phase flows as:
ρ ∂∂ ut + ∇ · (ρ u ⊗ u) = −∇P + ∇ · γ̄¯ ∇·u = 0 (1)
∂C
ρ = f (C) μ = f (C) ∂t + u · ∇C = 0 (2)
where u is the velocity, P the pressure, t the time, ρ the density and γ̄¯ the viscous
stress tensor. In the case of particle flows, the surface tension term is not considered.
The local solid fraction C is equal to 1 in the particles and 0 elsewhere. The particle
fluid interface is defined as C = 0.5.
2.1 Fictitious Domain Approach and Penalty Method

The code used, named Thétis and developped in TREFLE laboratory, is based on a
1-fluid formulation in which the presence of the particles is modelled by a fictitious
domain method and so the presence of the particles is simulated by an immersed
boundary method. It is particularly adapted to the resolution of conservation equa-
tions on a fixed Eulerian grid. The code is based on an original method for treating,
with the same set of equations, the carrier fluid and solid particles. The idea is to
split the stress tensor in the Navier-Stokes equations into four parts, associated to
Simulation of a Fluidized Bed Using a Hybrid Eulerian-Lagrangian Method 105
incompressibility, elongation, pure shearing and rotation respectively (Caltagirone

[3]). Thus, the viscous stress tensor can be written as:
¯ + κ Γ̄¯ (u) + ζ Θ̄¯ (u) − η Ω̄
γ̄¯ = (−P + λ ∇ · u)Id ¯ (u) (3)
In the fluid zones, the artificial viscosity λ will be chosen such as the term λ ∇·u will
be predominant. In the same way, in the solid zones, the different artificial viscosities
κ , ζ and η are chosen in order to have the term κ Γ̄¯ (u) + ζ Θ̄¯ (u) − η Ω̄¯ (u) domi-
nant in Navier Stokes equations for the cells belonging to a particle. Consequently,
with these differents choices of the artificial viscosities, the incompressibilty of the
flow is guaranteed and the non deformation of particles is maintained. Pratically,
the one fluid viscosity is defined as: μ = (1 −C) × μL +C × μP with μP > 103 × μL
(Randrianarivelo [8], Vincent [9, 10]). The decoupling between pressure and ve-
locity is realised via a scalar projection method. The discretization is carried by an
implicit finite volume on fixed Cartesian staggered grids. This Eulerian numerical
approach is commonly used for its efficiency in tracking incompressible particulate
flows that does not require remeshing. The space derivatives of the Navier-Stokes
equations are discretized with a second order centred scheme and times derivatives
are approximated by an Euler scheme of first order. The resolution of Navier Stokes
equations is achieved by using the Hypre library and parallelised iterative solvers
such as Bi conjugate gradient stabilised under a parasails preconditionning.
2.2 Tracking of Particles and Collision Model

Tracking particles is obtained via a hybrid Eulerian-Lagrangian method. Indeed,
a color function C is used on the Eulerian grid and furthermore, each particle is
located by the coordinates of its mass center. Navier Stokes equations are resolved
and velocity vectors are calculated in each cell. Each particle velocity is defined by
averaging the velocities in the cells belonging to the particle. Finally, each particle
can be advected by a standard advection equation:
−− → − → → −
Xin+1 = Xin + Vi Δ t (4)
−→ →
−
where Xin is the position vector at the instant n and Vi the velocity vector of the
particle i. Eq. (4) replaces Eq. (2) as corresponds to its Lagrangian form. In this
way, the new position of each particle is known. Then, knowing the analytic shape
of the particles (particles are spherical), their shape can be projected onto the Eule-
rian mesh to rebuild the Eulerian solid fraction C. Concerning a cell which contains
solid and fluid zones, the calculation of C is obtained by adding control points, ie a
local mesh refinement is implemented in the cells where ∇C = 0, as represented in
figure 1, and then C in this cell is equal to the ratio between the number of control
points belonging to a solid zone and the total number of additional points. At this
instant, the color function is known over the whole computational domain and the
processus can cycle for the next iteration.
106 C. Corre et al.
point outside
Xj
25
C(M) = 144
M
D Xi
D
R R + ΔX
point inside
Fig. 1 Calculation of C Fig. 2 Soft spheres Fig. 3 Virtual spheres
In order to take into account interactions between solid particles and to avoid
numerical agregation as a two particle interface share the same Eulerian cell, a
collision model has to be implemented. Two kinds of collision models can be dis-
tinguished: soft sphere and hard sphere. In this simulation, a soft sphere model is
adopted. However, as the tangential component is not considered, the soft sphere
assumption corresponds to a linear spring model (Cundall [4]). The force between
two particles i and j is expressed as:
−−→
Xi X j !
Fi/ j = −K × D × !!−−→! (5)
!Xi X j !
where D is the overlapping length as represented in figure 2 and K is a constant

of stiffness. The hybrid Eulerian-Lagrangian method for particle tracking does not
allow overlapping of two particles. Therefore, virtual particles as represented in
figure 3 have to be introduced. In order to reproduce correcty the rebound between
two particles, some tests have shown that the selection of a virtual radius equal to
R + Δ x and a constant of stiffness equal to 2.5 kg.s−2 is convenient. If the value of K
is too small, real particles will overlap. On the contrary, if K is too large, the duration
of overlapping of virtual !particles will be below the time step. The overlapping
−−→!
length D is equal to D = !Xi X j ! − 2 × (R + Δ x) and the force is activated only if
D is negative. At the end of each iteration, the velocity of each particle is predicted
for the next iteration and the likely position of each particle is estimated for the
next iteration. Then, tests on the possible collisions between every pair of particles
are calculated with these likely positions. If a collision is dectected, the force is
calculated and added as a volume source term in the right hand side of the Navier
Stokes equations for the cells belonging to the particles going to have a collision.
For each pair of particles, the force is different as the value of D is different too.
Consequently, at each calculation step, the collision force is implicitly taken into
account in the momentum equation.
3 Results
The simulated case is a fluidized bed studied experimentally by Aguilar Corona
([1]). In this bed, there are 2133 solid particles of Pyrex with a density ρP equal
to 2230 kg.m−3 . The fluid is a solution of KSCN with a density ρL equal to

1400 kg.m−3 and a dynamic viscosity μL equal to 3.8 10−3 Pa.s. The flow occurs
in a cylinder of radius Rcyl = 4 cm and height H = 64 cm. A fluidisation velocity
equal to 0.12 m.s−1 is imposed on the lower boundary of the computation domain.
Initially, particles are spread into the cylinder on a height of 48 cm. The particle
diameter is 6 mm so the terminal velocity for these particles is equal to 0.24 m.s−1
and so the Reynolds number based on this velocity equal to 530. For this case, a par-
allepipedic box of dimensions 8 cm × 8 cm × 64 cm with 100 × 100 × 800 grid mesh
is considered. The cylinder is represented with a Darcy penalty method (Khadra [5]).
The rebound of particles on the walls of the cylindrical tank is treated as a rebound
between two particles by inserting a virtual cylindrical tank of radius Rcyl − Δ x.
3.1 Vertical Profile of Volume Fraction and Pair Distribution

Function g(r)
The figure 4 shows vertical profile of the solid volume fraction α . It can be noticed
that the height of the bed tends to be constant. Moreover, distribution of particles
over the height of the cylinder seems to be non homogeneous. Also, an increase of
volume fraction is observed at the bottom of the bed, that can be explained by the ac-
cumulation of particles near to the wall at the bottom of the bed due to recirculation
loops.
0.64 2.5
0.56 t=7s
Vertical coordinate H
t=8s 2 t=7s
0.48 t=8s
t=9s
0.4 t=9s
1.5
g(r)
0.32
0.24 1
0.16
0.5
0.08
0
0 0.05 0.1 0.15 0.2 0
αvol 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4
r/R
Fig. 4 Vertical volume fraction profile, Fig. 5 Pair distribution function g(r) av-
averaged between t=5s and t=9s eraged between t=5s and t=9s according
to dimensionless radius
The pair distribution function g(r) is the conditional probability density function
of finding the mass center of a particle at a distance r knowing the mass center of
another particle is at the coordinate origin. Thus, g(r) provides a measure of local
spatial ordering in a fluid. As we can see in the figure 5, the maximum of g(r) is
obtained for a distance r equal to 2 (R + Δ x), which corresponds to two times the
radius of virtual spheres, where the particles collide. The value of the maximum is
108 C. Corre et al.
0.3 0.3
0.14 0.44 0.14 0.44
0.28 0.28
0.12 0.42 0.12 0.42
0.26 0.26
0.1 0.4 0.1 0.4
0.24 0.24
0.08 0.38 0.08 0.38
0.22 0.22
0.06 0.36 0.06 0.36
0.2 0.2
0.04 0.34 0.04 0.34
0.18 0.18
0.02 0.32 0.02 0.32
0.16 0.16
0 0.3 0 0.3
0 0.02 0.04 0 0.02 0.04 0 0.02 0.04 0 0.02 0.04 0 0.02 0.04 0 0.02 0.04
r r r r r r
Fig. 6 Average local velocities in coordinates Fig. 7 Streamlines of local average velocity
(r,H) (average between t=8s and t=9s); (average between t=8s and t=9s) ;
left: 0 ≤ H ≤ 0.15m ; center: 0.15 ≤ H ≤ left: 0 ≤ H ≤ 0.15m ; center: 0.15 ≤ H ≤
0.30m right: 0.30 ≤ H ≤ 0.45m 0.30m right: 0.30 ≤ H ≤ 0.45m
important because this value is used in the statistical approaches to evaluate collision
terms. This value, named g0 depends on the volume fraction as described by the
correlation of Lun and Savage ([6]):
−2.5αmax
α
g0 = 1 − αmax (6)
In our numerical case, for α = 0.11, the value of g0 is 2.6 as can be seen in figure
5. If we put into practice the correlation of Lun and Savage with α = 0.11 and
αmax = 0.64, we find : g0 = 1.35. With the same correlation and with a volume
fraction based on virtual particles, we obtain α = 0.16 and g0 = 1.58.
3.2 Average Local Velocities and Local Variances

In order to realise reliable statistics, the axisymmetry of the problem is taken into
consideration and consequently the study of average local velocities is done in cylin-
drical coordinates (r,H). A new mesh is created (6 × 48), and in each cell of this new
mesh, the average velocity vector on particle velocity vector is calculated between
the time t = 5s and t = 9s. The mesh size and time interval considered allows us to
obtain converged results. These average vectors are represented in figure 6. It can be
noticed that particles tend to go up near the axis of symmetry and to go down near
the wall. This behaviour is all the more obvious with the steamlines represented in
figure 7. However, in comparaison with experimental results (Aguilar Corona ([1]),
same order of magnitude on velocities and variances are found as it can be seen in
figures 8 and 9. Also, it can be observed that variance of local average velocities
is larger on z velocity component. Furthermore, variances on z velocity component
are two times larger near the wall than in the center of the cylindrical tank.
0.3 00.0.0
0.0
0.002
0.0
0.0 0.0.00
00 0032
01
02
3
01 0.0025
05
0.0
0.0 .00
5
5
2
5
0.0
0
0.002
0.00
0.14 0.44
02 2
01
5
0.00
0.14 0.44
02
01
5
0.0
02
0.0
5
0.0
15
0.0
01
02
01 0.0
0.0
0.0005
25
01 0.00
0.28 0.001
0.00
0.28
0.001
3
0.002
0.002
0.001
0.003
0.002
0.0
5
01 0.001
0.001
0.003
0.0 0.00
01
0.0015
02
5
0.0 3
03
0.12 0.42
5
01
0.0
0.0
05
0.12 0.42
0.0 0.0015
0.0
0.0
0.00
01
0.001
00
0.00
0.00
0.0
5
1
02
0.003
0.00025
02
0.001
3
15
5
3
0.26
0.00
0.001
0.26
0.0
01
5
0.0
01
02
0.00
0.0
0.0
0.00
1
0.003
0.0
0.001
0.00
25
0.1 0.4
01
0 0.0025
0.1 0.4
0.0015
0.0
0.0
2
01
02
01
1 0.00
0.0
0.00 0.0
0.001
0.0
5
3
5
01
0.0
02
0.0005
01 0.0
0.24 15
0.0
0
0.24 0.0
0.0
5 0 0215
01
5
5
01
0.0
0.001
.0 0.
0.0
0
1
5
0.0
0300
01 0.003
0.00
0.00
0.001
0.00
0.0
02
0.08 0.38 05 0.0 0.00
0.08 0.38 00.0.0
02
25
0.0005
1
15 25
0.00
0.00
0.002 0015
15
0001
5 02 5
0.00
0.0 0.0 0.0 001
0.0.0
0.22 00
0.22 0.
0.002
02
0.0
0.00.0 0.0
0.00
0.0
3
5
0.0
5
02
0202
03
02
03
0.0 0.0
0.0025 15
0.0
0.0 0.0
005 1 0.0
0.06 0.36 0.0
01 01
0.0
0 3 03
01
0.06 0.00 0.36
0.0
0.00
02
5
03
0.00
5
0.003
01
25
5
0.00
0.0
0.0
0.001
03
00
0.2 0.0 0
25
0.001
0.2 0.00 01 .00
0.0
0.0
5
02
0.00
0.00 5
5 2
0.0
0.002
02
0.001
01
0.00
0.00 05 5
05
15
2
1 0.00
0.00
1
0.0 0.0
0.04 0.34 0.04 0.0005 01 2
0.34 0.0 02
0.0
0.0
5
01
01
00
0.00
0.0
0.0
5
0.0025
5
0.003
0.0
3
02
0.0
05
01
0.00
0.18
15
0.00 01
5
01
0.18 0.0 5
0.002
0215
5
01 0.00115
0.0015 001
0.000
.00
15 0.00
0.00
00.0
0.0025
0.00
0.002
0.002
0.02 0.32 0.0015
0.
5
0.02 0.32
1
005
5
0.00 01 0.0
0.0 01
05
0.0
05 15 02
0.0
0.00 0.0 0.0015
0.00
05
02
0.00
0.16
0.001
0.0005 0.16 0.0
0.0
0.00 0
03
01
0.0015
0.0025
5
0.001
0.001
15.0
0.00
0 0.3 0 0.3 00
5 3
01
00
25
00
0 0.02 0.04 0 0.02 0.04 0 0.02 0.04
5
0 0.02 0.04 0 0.02 0.04 0 0.02 0.04
r r r r r r
" #
Fig. 8 uz (average between t=8s and t=9s) Fig. 9 u
z 2 (average between t=8s and
left: 0 ≤ H ≤ 0.15m ; center: 0.15 ≤ H ≤ t=9s) left: 0 ≤ H ≤ 0.15m ; center: 0.15 ≤
0.30m right: 0.30 ≤ H ≤ 0.45m H ≤ 0.30m right: 0.30 ≤ H ≤ 0.45m
4 Conclusion
The simulation of a fluidized bed has been presented with fictitious domain methods.
The results are interesting owing to the fact that our results will be compared to
experimental results (Aguilar Corona[1]). Moreover, this simulation allows to better
understand the distribution of particles in a turbulent flow. Particularly, thanks to our
DNS approach, the pair correlation function has been possible to be estimated for
the first time. This function is used in statistical approaches to specify collision
terms.
Acknowledgements. This work has been supported by the French ANR, STI NT05-1 43573.
Moreover, we thank the Midi Pyrénées region for providing the funding for this PhD thesis.
We are grateful for access to the computational facilities of the french CINES (National
computing center for higher education) under project number x2009026115.
References
1. Aguilar Corona, A.: Agitation de particules dans un lit fluidise liquide. Etude
expérimentale. PhD Thesis, Université de Toulouse, INPT (2008)
2. Angot, P., Bruneau, C.-H.B., Fabrie, P.: A penalization method to take into account ob-
stacles in incompressible viscous flows. Numerische Mathematik 81(4), 497–552 (1999)
3. Caltagirone, J.P., Vincent, S.: Tensorial penalisation method for solving Navier-Stokes
equations. Compte rendu de l’académie des sciences, Série II 329, 607–613 (2001)
4. Cundall, P.A., Strack, O.D.L.: A discrete numerical model for granular assemblies.
Géotechnique 29(1), 47–65 (1979)
5. Khadra, K., Angot, P., Parneix, S., Caltagirone, J.P.: Fictitious Domain Approach for
Numerical Modelling of Navier-Stokes Equations. International Journal for Numerical
Methods in Fluids 34(8), 651–684 (2000)
110 C. Corre et al.
6. Lun, C.K.K., Savage, S.B.: The effect of an impact velocity dependent coefficient of
restitution on stresses developed by sheared granular materials. Acta Mechanica 63, 15–
44 (1986)
7. Pan, T.W., Joseph, D.D., Bai, R., Glowinski, R., Sarin, V.: Fluidization of 1204 spheres:
simulation and experiment. Journal of Fluid Mechanics 451, 169–191 (2002)
8. Randrianarivelo, T.N., Pianet, G., Vincent, S., Caltagirone, J.P.: Numerical modelling of
solid particle motion using a new penalty method. International Journal for Numerical
Methods in Fluids 47, 1245–1251 (2005)
9. Vincent, S., Caltagirone, J.P., Lubin, P., Randrianarivelo, T.N.: An adaptative aug-
mented lagrangian method for three-dimensional multimaterial flows. Computers and
Fluids 33(10), 1273–1289 (2004)
10. Vincent, S., Randrianarivelo, T.N., Pianet, G., Caltagirone, J.P.: Local penalty methods
for flows interacting with moving solids at high Reynolds numbers. Computers and Flu-
ids 36, 902–913 (2007)
Wavelet-Adapted Sub-grid Scale Models for LES
J.A. Denev, C.J. Falconi, J. Fröhlich, and H. Bockhorn
Abstract. The paper presents two novel turbulent models for LES based on a
wavelet decomposition. This approach, denoted WALES, is simple and easy to im-
plement. Tests on a number of flows using grids with different resolution near walls
show that the models exhibit the same quality as the Smagorinsky model without the
need of wall functions or near-wall damping. In the paper the basic wavelet frame-
work and two such models are described in detail. Physical benefits of the models
due to the use of wavelets are discussed. Results obtained with the models are com-
pared to those using the Smagorinsky model, to experimental results and to results
from Direct Numerical Simulations. The agreement achieved is generally good.
1 Introduction
Wavelets cosntitute a mathematical tool designed to describe multi-scale physical
effects and hence appear as a natural way for modelling turbulence and mixing phe-
nomena. One possible way to apply the wavelets in turbulence modelling is to use
them as a criterion for adaptive grid refinement to automatically detect all energetic
coherent flow structures. This method, known as Coherent Vortex Simulation (CVS)
[1] resolves all energetic turbulent structures down to the Kolmogorov scale. Due
to this, CVS is still expensive for engineering applications typically featuring com-
plex geometry with high Reynolds and/or Schmidt numbers. Another wavelet-based
method, called SCALES, explores the idea to solve only for the most energetic co-
herent eddies while modelling the effect of the less energetic (and unresolved) mo-
tions [2]. It applies a thresholding wavelet filter to perform dynamic grid adaptation
J.A. Denev · C.J. Falconi · H. Bockhorn
Institute for Technical Chemistry und Polymer Chemistry, University of Karlsruhe (TH),
Kaiserstraße 12, 76128 Karlsruhe, Germany
J. Fröhlich
Institute of Fluid Mechanics, Technical University of Dresden,
George-Bähr-Straße 3c, 01062 Dresden, Germany
112 J.A. Denev et al.
with respect to the energetic structures, thus being more ”complete” than classical
Large Eddy Simulation (LES). In the present work a different approach is followed
which embeds wavelets in the framework of classical LES methods. The wavelet
details, obtained after wavelet filtering of the velocity components on a LES-like
numerical grid, are directly used for modelling of the subgrid-scale stress tensor.
We named this approach Wavelet-Adapted LES (WALES).
The code into which the models were implemented is the incompressible Finite
Volume code LESOCC2 for LES on block-structured hexahedral curvilinear grids.
The code is fully parallelized and the implementation of the WALES models is
parallel as well. Temporal and spatial schemes are of second order accuracy. More
details are given, e.g., in [3].
2 Wavelet Approach to SGS Modelling and Numerical Issues

With LES, the sub-grid term τi j = ui u j − ui u j , where the overbar represents the
filtering, is to be represented by a subgrid-scale (SGS) model. Two widely used
classes of such models are the eddy-viscosity models and the scale-similarity mod-
els. Therefore in the present study wavelet-based SGS models have been designed
and tested in the spirit of both of these approaches. Both models are based on the
wavelet details of the velocities in order to model the SGS stress tensor. The method
is genuinely three-dimensional, but will be outlined here in the one-dimensional set-
ting for simplicity using Fig.1. Starting point is the fine-grid solution on grid level
k. Then values on the next coarser level k − 1 are obtained by some averaging proce-
dure (restriction). In the present case, the average of the two neighbours is used; in
three dimensions the averaging includes eight neighbour cells of the fine grid. Then,
a prolongation of the coarse-grid solution to the fine grid is performed according to
s
u(xk2i−1 )| prol = u(xk−1
i ) + ∑ γl [u(xk−1
i+l ) − u(xi−l )]
k−1
l=1
s
u(xk2i )| prol = u(xk−1
i ) − ∑ γl [u(xk−1
i+l ) − u(xi−l )]
k−1
(1)
l=1
Fine Fine
x2ki 1 x2ki
Fine Fine
xik11 xik 1 xik11

Coarse Coarse Coarse
Fig. 1 Arrangement of hierarchical grid points in relation to a Finite Volume scheme in 1D,
illustrating the wavelet decomposition algorithm used for the WALES approach
Wavelet-Adapted Sub-grid Scale Models for LES 113
In the present work one neighbour on the coarse grid is used (s = 1) with γ1 =
−1/8 and the multiresolution accuracy being r = 2s + 1 = 3 [4]. The chosen value
for s = 1 presents a compromise between the accuracy of the approach and the
required amount of information exchange for the parallel algorithm at the block-
boundaries of the sub-domains. In the simplest case of Haar wavelets [5] with s = 0
the prolongated value is equal to the coarse-grid value and no information needs to
be exchanged at the block boundaries in this step, provided an even number of cells
is used.
The fine-scale details of the signal are computed as the difference between the
original values on the fine grid and the prolongated values:
d{u(xk2i−1 )} = u(xk2i−1 ) − u(xk2i−1)| prol

d{u(xk2i )} = u(xk2i ) − u(xk2i)| prol (2)
The above algorithm decomposes of the original signal u(xki ), i = 0, ..., 2k − 1 on

level k into a coarser signal u(xk−1
i ), i = 0, ..., 2k−1 − 1 on level k − 1 and details
d{u(xi )}, i = 0, ..., 2 − 1. The corresponding expressions in three dimensions are
k k
lengthy and available elsewhere [6]. In the following the indices of the details will
be omitted. In the present work the decomposition is applied in a two-grid fashion,
but the extension to more levels is immediate. As the LES-grids usually feature good
quality with low stretching and skewness, geometrical factors are not included for
simplicity and efficiency, as is common practice for the filtering applied with the
dynamic procedure. Similarly, the prolongation step assumes zero values beyond
walls for simplicity.
By construction, wavelets fulfil moment conditions, i.e. the coarse-grid solution
reproduces exactly polynomials up to a certain degree, here up to degree r − 1.
Details in (2) then vanish. Hence, with the present choice s = 1, constant, linear
and parabolic velocity profiles do not produce any contribution in the details. SGS
models based on d{u} consequently turn off automatically in well-resolved laminar
flows. That was already explored in [3] for a parabolic velocity profile of a transi-
tional jet into a crossflow.
The above decomposition then is applied to all three velocity components ui . Two
WALES models are proposed here, an eddy-viscosity type and a scale-similarity
type model. The former is termed WALES-E and based on τimod j = −2νt Si j (νt being
the turbulent viscosity and Si j the deformation tensor) with

νt = CWALES−E Δ d{ui }d{ui } , (3)
where the model constant is equal to 0.02. This model bares some similarity with
the structure function model but generalizes it as the order can be chosen freely.
The second model is termed scale-similarity WALES model (WALES-S) and
reads
τi j = −CWALES−S,i j d{u
i u
j } − d{u
i}d{u
j } , (4)
with CWALES−S,i j = 0.2 + 0.2δi, j , determined by calibration. Note that in this model
the fluctuating velocity values u
i = ui − ui and not the instantaneous ones are used
(here Reynolds-averaged quantities are denoted by ...).
Below, the results obtained with the above models are compared with those em-
ploying the Smagorinsky model (SM) using CS = 0.1. The latter was used with van
Driest damping near walls for fine grids and a Werner-Wengle wall function for
coarse grids. The present WALES models were not altered near walls.
3 Results for Plane Channel Flow

The channel flow was calculated on a grid with 147000 inner cells and the parame-
ters sumarized in Tab. 1. Fig.2(left) and Fig.2(center) show comparisons of the two
WALES models for Reτ = 395. Better agreement is achieved for the shear stress
u
v
in Fig.2(left) predicted by the WALES-E model, while the SM matches better
the peak of the streamwise velocity fluctuations u
u
. A simulation for the same
Reynolds number was conducted with a finer wall-resolving grid and the WALES-
E model, yielding only very slight improvement. For the WALES-S model good
agreement for the Reynolds stresses u
v
, w
w
and v
v
is achieved, but the
overprediction of the maximum of u
u
is stronger than with WALES-E.
Fig.2(right) shows results using the WALES-E model and the SM in the channel
flow with scalar transport. Both models show good agreement with the DNS data
of [7]. Similar qualitative and quantitative agreement with DNS apllies also for the
predicted scalar transport with the WALES-S model at Reτ = 180 (not shown).
Table 1 shows that the friction velocity Uτ at the wall is considerably better pre-
dicted by the two WALES models than by the Smagorinsky model.
Table 1 Cases calculated, grid parameters and results for the friction velocity Uτ . The relative
error of Uτ is calculated with the reference DNS data from Moser et al. (1999) (0.0632 for
Reτ = 180 and 0.0570 for Reτ = 395)
Model Reτ Δ x+ Δ y+
1 Δ z+
1 Uτ Relative error
for Uτ in [%]
SM 180 17.8 1.3 11.9 0.0588 7.0

WALES-E 180 17.8 1.3 11.9 0.0618 2.3
WALES-S 180 17.8 1.3 11.9 0.0632 0.0
SM 395 39.2 2.9 26.1 0.0514 9.8
WALES-E 395 39.2 2.9 26.1 0.0553 3.0
WALES-S 395 39.2 2.9 26.1 0.0549 3.6
4 Results for the Flow in a Model Combustor

The non-reactive flow in the TECFLAM / TYPE II combustor was experimen-
tally investigated by means of LDA in [8]. The configuration consists of an an-
nular swirling jet entering a cylindrical combustion chamber. As a representation
of the entire swirl-generating device is too costly, stationary top-hat profiles for the
radial and azimuthal velocity component were imposed at the circumferencial in-
flow boundary. This aproach was validated for a very similar configuration in
[9] and does not affect the results as turbulence readily develops in the annulus
upstream of the jet exit.
A block-structured grid was used with 6.2 million hexahedral cells and 620
blocks. 96 grid points were used in the azimuthal direction. Fig.3 shows details
of the grid.
The outer diameter of the jet is D0 = 60mm and the inner diameter is d0 = 30mm.
A massflow rate of m = 0.066363kg/s yielding a bulk velocity of Ub = 39.51m/s.
The Reynolds number based on these quantities is Reb = 159, 758. The theoretical
swirl number S0 , defined as the ratio of angular momentum to axial momentum flux
normalized by the outer radius R0 = D0 /2, was S0 = 0.9. The simulations were run
for several time units tb = D0 /Ub to eliminate the effects of the initial conditions,
before statistical quantities were collected over about 1000 tb .
Fig. 4 shows the calculations with the WALES-E model and compares them to the
SM and LDA measurements. The WALES-S model was found numerically unstable
0.035 <uú´>DNS-MKM 0.035 <uú´>DNS-MKM 0.025

<w´w´>DNS-MKM <w´w´>DNS-MKM -<uć´>DNS
-<u´v´>DNS-MKM -<u´v´>DNS-MKM -<uć´>SMAG.
0.03 0.03 -<uć´>WALES
<uú´>SMAG. <uú´>SMAG.
<uú´>, <w´w´>, -<u´v´>
<uú´>, <w´w´>, -<u´v´>
0.02 <vć´>DNS
<w´w´>SMAG. <w´w´>SMAG.
0.025 -<u´v´>SMAG. 0.025 -<u´v´>SMAG. <vć´>SMAG.
-<uć´>, <vć´>
<uú´>WALES <uú´>WALES <vć´>WALES

<w´w´>WALES <w´w´>WALES 0.015
0.02 0.02
-<u´v´>WALES -<u´v´>WALES
0.015 0.015
0.01
0.01 0.01
0.005
0.005 0.005
0 0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 1
y y y
Fig. 2 Results for the plain channel. Left: WALES-E model, Re = 395; Middle: WALES-
S model, Re = 395; Right: Passive scalar turbulent quantities for the WALES-E model,
Re = 180
4
2 0.3
2
1.5 0.15
z/D 0
y/D 0
0
z/D 0
1 0
0.5 -2 -0.15
0 -4 -0.3
-1.5 -1 -0.5 0 -4 -2 0 2 4 -0.3 -0.15 0 0.15 0.3
x/D 0 y/D 0 y/D 0
Fig. 3 Computational grid for TECFLAM burner. Left: y − x plane near the jet exit, every 2nd
grid line is shown. Middle: Cut in the y − z plane of the entire computational domain. Right:
Zoom around the x-axis
1 WALES 0.3 WALES 1 WALES

SMAG. SMAG. SMAG.
0.8 LDA LDA 0.8 LDA
0.2
0.6 0.6
<w>/Ub
<u>/Ub
<v>/Ub
0.1
0.4 0.4
0.2 0 0.2
0 -0.1 0
-0.2 -0.2
-0.2
0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4
y/D 0 y/D 0 y/D 0
0.5 WALES 0.5 WALES 0.5

SMAG. SMAG.
LDA LDA WALES
0.4 0.4 0.4
SMAG.
LDA
Wrms/Ub
Urms/Ub
Vrms/Ub
0.3 0.3 0.3
0.2 0.2 0.2
0.1 0.1 0.1
0 0 0
0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4
y/D 0 y/D 0 y/D 0
0.4 WALES 0.2 WALES 0.3 WALES

SMAG. SMAG. SMAG.
0.3 LDA LDA LDA
0.2 0.1 0.2
<w>/Ub
<u>/Ub
<v>/Ub
0.1
0 0.1
0
-0.1
-0.2 0
-0.1
0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4
y/D 0 y/D 0 y/D 0
0.25 WALES 0.25 WALES 0.25 WALES

SMAG. SMAG. SMAG.
0.2 LDA 0.2 LDA 0.2 LDA
Wrms/Ub
Urms/Ub
Vrms/Ub
0.15 0.15 0.15
0.1 0.1 0.1
0.05 0.05 0.05
0 0 0
0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4 0.5 1 1.5 2 2.5 3 3.5 4
y/D 0 y/D 0 y/D 0
Fig. 4 Results for the isothermal flow in the TECFLAM combustor. First row: Mean axial,
radial and azimuthal velocity profiles, from left to right, at x/D0 = 0.0833. Second row:
corresponding fluctuations at x/D0 = 0.0833. Third row: Mean axial, radial and azimuthal
velocity profiles, at x/D0 = 1.5. Fourth row: corresponding fluctuations at x/D0 = 1.5
for a high-Reynolds number periodic-hill flow and therefore the WALES-E model
(denoted WALES in the figure) was preferred for the present flow configuration.
For lack of space, only few profiles can be shown here. The global agreement
between experiments and the simulations is very good at the inlet indicating that
the top-hat profile applied is adequate. The WALES-E model and the SM show an
overall similar performace with exception of the time-mean radial velocity < v > at
x/D0 = 0.0833. Here, the peak of the radial profile of < v > is matched better with
the SM, while the WALES-E model shows a better agreement near the axis. Further
downstream the WALES-E model and the SM show very much the same results for
mean flow as well as fluctuations. The same conclusion applies also for quantities
like e.g. the turbulent viscosity, calculated by the two models.
5 Conclusions
An approach to SGS modelling for LES based on a wavelet decomposition was
presented in two variants. The results obtained with two of these WALES models
were compared with the Smagorinsky model, experimental data and results from
Direct Numerical Simulations. The models feature simplicity and are quite easy to
implement. For the flows considered, the present WALES models yield the same
quality of results as the Smagorinsky model without the need of wall functions or
wall damping.
Acknowledgements. Computations have been performed on the national super computer

HP XC4000 at the High Performance Computing Centrer Stuttgart (HLRS) under the grant
with acronym DNS-JET.
References
[1] Farge, M.: Wavelet transforms and their applications to turbulence. Ann. Rev. of Fluid
Mech. 24, 395–457 (1992)
[2] De Stefano, G., Vasilyev, O.V., Goldstein, D.E.: Localized dynamic kinetic-energy-based
models for stochastic coherent adaptive large eddy simulation. Physics of Fluids 20, 045–
102 (2008)
[3] Denev, J.A., Froehlich, J., Falconi, C.J., Bockhorn, H.: Direct Numerical Simulation,
Analysis and Modelling of Mixing Processes in a Round Jet in Crossflow. To appear in
Springer series on Heat and Mass Transfer (2009)
[4] Harten, A.: Discrete multi-resolution analysis and generalized wavelets. J. Appl. Num.
Math. 12, 153–193 (1993)
[5] Haar, A.: Zur Theorie der orthogonalen Funktionensysteme. Math. Ann. 69, 331–371
(1910)
[6] Tsigulin, A., Bockhorn, H., Roussel, O., Schneider, K.: A conservative fully adaptive
multiresolution algorithm for parabolic PDEs. J. Comput. Phys. 188, 493–523 (2003)
[7] Schwertfirm, F., Manhart, M.: ADM–modelling aproach for semi–direct numerical simu-
lation of turbulent mixing and mass transport. In: Humphrey, J.A.C., Gatski, T.B., Eaton,
J.K., Kasagi, N., Friedrich, R., Leschziner, M.A. (eds.) Procs. of 4th Int. Symp. on Turbu-
lence and Shear Flow Phenomena, Williamsburg, June 27-29, vol. 2, pp. 823–828 (2005)
[8] Schmittel, P., Günther, B., Lenze, B., Leuckel, W., Bockhorn, H.: Turbulent swirling
flames: Experimental investigation of the flow field and formation of nitrogen oxide.
Symposium (International) on Combustion 28(1), 303–309 (2000)
[9] Garcı́a-Villalba, M., Fröhlich, J., Rodi, W.: Identification and analysis of coherent struc-
tures in the near field of a turbulent unconfined annular swirling jet using large eddy
simulation. Phys. Fluids 18, 055–103 (2006)
Effect of Particle-Particle Collisions on the
Spatial Distribution of Inertial Particles
Suspended in Homogeneous Isotropic Turbulent
Flows
Pascal Fede and Olivier Simonin
Abstract. Preferential concentration of solid inertial colliding particles suspended in

homogeneous isotropic turbulence is numerically investigated using Direct Numeri-
cal Simulation (DNS) coupled with Discrete Particle Simulation (DPS). The results
show that the preferential concentration is decreasing when the collision frequency
increases. This effect is found enhanced for non-elastic particle-particle collisions.
1 Introduction
Turbulent particle laden flows are found in a wide range of industrial applica-
tions such as catalytic chemical process, pollutant transport and deposition or di-
rect fuel injection in IC engine. The present study is focusing on the instantaneous
spatial particle distribution in homogeneous turbulence and its modification by
particle-particle collision.
Understanding and modeling of particle spatial distribution is challenging
because it may strongly modify the particulate phase statistical properties. Con-
sidering droplet coalescence, the preferential concentration modifies locally the co-
alescence rate and consequently the particle diameter distribution. In addition, the
particle distribution may affect the gas turbulence modulation resulting from the
momentum transfer between the gas and the particles.
It has been shown [7, 2] that inertial particles suspended in homogeneous isotropic
turbulent flows may concentrate in low-vorticity regions of the turbulence. This
phenomenon, called preferential concentration, results roughly from the competi-
tion between the drag force and centrifugal forces. It takes place for particles with
a response time to the drag force in the range of the characteristic time scales
of the turbulence (from Kolmogorov time microscale to Lagrangian integral time
Pascal Fede · Olivier Simonin
Université de Toulouse; INPT, UPS; IMFT; 31400 Toulouse, France
CNRS; Institut de Mécanique des Fluides de Toulouse; FR-31400 Toulouse, France
e-mail: fede@imft.fr,simonin@imft.fr
120 P. Fede and O. Simonin
macroscale). In the literature, a few papers [5, 9] investigated the effect of prefer-
ential concentration on the particle-particle collision rate. In the present paper we
take the opposite point of view because we study the modification of preferential
concentration by the collisions.
2 Numerical Methods
The fluid flow in this study is governed by the 3-dimensional incompressible
Navier-Stokes equations where the turbulence modulation by the particles is not
taken into account because the solid mass fraction is negligible. The computa-
tional domain is a cubic box of length Lb = 0.128m with periodic boundary con-
ditions. Statistically steady turbulence is obtained with a stochastic spectral forc-
ing proposed by Eswaran & Pope[1]. The forcing is accomplished in Fourier
space using a stochastic force added on a given range of wavenumber: [2k0 , 6k0 ]
(where k0 = 2π /Lb is the first computed wavenumber). This parameter leads to
energy containing eddies ten time smaller than the computational domain. The
Reynolds number is about ReL f = 112 and kmax .ηK = 1.1 ensuring a good spa-
tial resolution of the smallest turbulent eddies (where kmax is the largest com-
puted wavenumber and ηK the Kolmogorov length scale). Relevant turbulence
properties are summarized in Table 1.
The dispersed phase is composed of N p solid spherical identical particles.
Assuming a large particle to fluid density ratio and neglecting the gravity, the forces
acting on the particles are reduced to the drag force. Then the particle motion
governing equations are:
dx p du p u p − u f @p
= up and =− (1)
dt dt τp
Table 1 Turbulence properties
Unit
Grid points N = 1283

Simulation box length (m) Lb = 0.128
Fluid density (kg.m−3 ) ρ f = 1.17
Viscosity (m2 .s−1 ) ν f = 1.47 × 10−5
Reynolds number ReL f = 112
Fluid kinetic energy (m2 .s−2 ) q2f = 30.6 × 10−3
Fluid dissipation (m2 .s−3 ) ε f = 150.3 × 10−3
Integral longitudinal length scale L f /Lb = 0.09
L f /Lg = 1.05
Kolmogorov length scale ηK /Lb = 0.032
Eulerian integral time scale (s) τ Ef = 79.3 × 10−3
Lagrangian integral time scale (s) τ tf /τ Ef = 0.751
Effect of Particle-Particle Collisions on the Spatial Distribution of Inertial Particles 121
where x p and u p are the particle position and velocity. The particle response time,
τ p , is defined by
ρp 4 dp 1
τp = , (2)
ρ f 3 Cd |u p − u f @p |
where d p is the particle diameter, ρ p the particle density and Cd the drag coefficient
given by Schiller & Nauman’s correlation [6]
24
Cd = 1 + 0.15Re0.687
p , (3)
Re p
with the particle Reynolds number Re p = d p |u p − u f @p |/ν f . u f @p (t) is the fluid

velocity at the particle position (locally undisturbed by the particle) also called the
fluid velocity seen by the particle. As the two-way coupling is not taken into account,
the undisturbed fluid velocity is the one given by the DNS and evaluated at the
particle position by an interpolation scheme. Equation (1) are time-advanced using
a second order Runge & Kutta scheme with the same time step than the DNS solver.
For the hereafter
" # analysis let us introducing the mean particle relaxation time as
1/τ Ffp = 1/τ p with . the particle ensemble average operator.
The numerical treatment of collisions is performed using an algorithm devel-
oped originally by Hopkins & Louge [4] for rapid granular flow. The algorithm is
based on a grid detecting all particle pairs that have overlapped during the previous
time step. Considering two particles with velocities before collision u p and uq , the
particle velocities after collision are given by
1 1
u∗p = u p + (1 + ec)(w pq .k)k and u∗q = uq − (1 + ec)(w pq .k)k (4)
2 2
where w pq = uq − u p is the relative velocity of colliding particles, k the normal unit
vector linking the particle centers and ec the normal restitution coefficient.
3 Preferential Concentration Characterization

The particle segregation is analyzed with the global particle accumulation
parameter, Σ p , defined as:
1 Np
Σp = (σ − σPois) with λ= (5)
λ Ncell
where σ is the standard deviation of f (C), the particle-concentration Probability

Density Function (PDF) measured in DNS, and σPois the standard deviation of the
Poisson distribution. The function f (C) is computed on a cartesian uniform grid
containing Ncell cells, of size Δcell . The parameter λ is the mean number of particles
in a cell for a random uniform particle distribution. Figure 1 shows the particle con-
centration PDF for fluid elements and solid inertial particles measured from DNS.
0.2
0.15
f (C)
0.1
0.05
0
0 5 10 15 20 25 30
C
Fig. 1 Particle concentration Probability Density Function. The symbols correspond to the
DNS results for fluid elements (diamonds) and inertial particles with a Stokes number
τ Ffp /τ tf = 2.27 (squares). The lines are the Poisson distribution given by (6)
0.4
0.3
α p = 0.75 %
α p = 0.5 %
Σp
0.2 α p = 0.25 %
α p = 0.1 %
α p = 0.075 %
α p = 0.05 %
0.1
0
0 0.5 1 1.5 2
Δcell
Fig. 2 Particle accumulation Σ p with respect to cell size and the particle volume fraction
(α p ). Statistic computed with inertial particles with a turbulent Stokes number τ Ffp /τ tf = 2.27
As expected the fluid element spatial distribution is uniform and the concentration
PDF is well predicted by the Poisson distribution given by:
e− λ λ C
fPois (C) = (6)
C!
In contrast, Fig. 1 shows that the concentration PDF of inertial particles differs from
the Poisson distribution indicating particle clustering.
As shown by Fessler et al. [2] and by Fig. 2 the particle accumulation Σ p depends
on the cell size Δcell . The Figure 2 shows that Σ p has a maximum for a specific grid
size. As proposed by Février et al.[3] we take this value as the relevant statistic to
0.8
0.6
Σ̃ p
0.4
0.2
0 -2
τ Ffp /τ tf
-1 0 1
10 10 10 10
Fig. 3 Evolution of particle accumulation, Σ̃ p , with respect to the turbulent Stokes number.
White symbols: α p = 0.05%, black symbols: α p = 0.75%, +: DNS from Février et al.[3] for
ReL f = 110 and ×: LES from Février et al.[3] for ReL f = 700
0.8
0.6
Σ̃ p
0.4
0.2
0 -2 -1 0 1
10 10 τ Ffp /τcol 10 10
Fig. 4 Particle accumulation Σ̃ p with respect to the collision frequency. The collision
timescale τcol is proportional to the inverse of collision frequency
study the preferential concentration. In the following, Σ̃ p is the maximum of particle

accumulation and Δ̃cell the corresponding grid size.

Figure 3 shows the particle accumulation with respect to the turbulent Stokes num-
ber. For small Stokes number the particles behavior tends to the one of fluid ele-
ments. As fluid elements are uniformly distributed in homogeneous isotropic tur-
bulence, the particle accumulation is decreasing. Between both limiting cases, the
competition between the drag and the centrifugal force (resulting from the local
particle-fluid eddy interaction) leads the particles to accumulate in low-vorticity re-
gion [7, 8, 2].
1.6
1.4
1.2
Δ̃cell /L f
1
0.8
0.6
0.4
0.2
0 -2 -1 0 1
10 10 10 10
τ Ffp /τcol
Fig. 5 Optimum cell size scale Δ̃ cell with respect to the collision frequency
1.3
1.2
Σ̃ p /Σ̃ p (ec = 1)
1.1
0.9 -2 -1 0 1
10 10 τ Ffp /τcol 10 10
Fig. 6 Normalized particle accumulation Σ̃ p of St = 2.2 with respect to the collision

frequency and normal restitution coefficient. In this case the particles have a Stokes
number τ Ffp /τ tf = 2.27. Solid line: ec = 0.95, dashed line: ec = 0.90 and dotted line: ec = 0.80
The Figure 3 shows the global particle accumulation with respect to the turbulent
Stokes number. For a given Stokes number, the particle accumulation is increasing
when the solid volume fraction increases. It suggests that the particle-particle colli-
sions lead to increase the preferential concentration effect. This is confirmed by Fig.
4 showing the particle accumulation with respect to the collision frequency. The ex-
planation of this trend is that the collisions reduce the particle mean free path and
the relative dispersion between two separate particles.
The Figure 5 shows the evolution of the optimal cell size Δ̃cell with respect to
the turbulent Stokes number. This statistic gives information of the cluster size. The
maximum of Σ p should appears when a few number of cells contain a large part of
the full number of particles. Then the optimum cell size, namely the one given the
maximum of Σ p , should be nearly equal to the diameter of the clusters (assuming a
cylindrical shape of the cluster). The Figure 5 shows that Δ̃cell is decreasing when
the collision frequency increases. It suggests that the collisions reduce the cluster
size.
The Figure 6 shows the effect of non-elastic particle collisions on the preferential
concentration. The particle accumulation is enhanced when the normal restitution
coefficient is decreasing. This effect looks similar to particle clustering formation in
dry granular flows due to the dissipation of the particle-particle relative velocity by
inelastic collisions.
5 Conclusions
Direct numerical simulation coupled with Lagrangian particle tracking have been
carried out in order to investigate the effect of particle-particle collisions on the par-
ticles spatial distribution. The preferential concentration is found increasing when
the collision frequency is increasing by the reduction of the particle mean free path.
The cluster characteristic length scale has been found decreasing when the collision
frequency is increasing. When the particles are not elastic the particle accumulation
is also found increasing.
Acknowledgements. Numerical simulations were performed on NEC-SX8 supercomputer.

The CPU time has been provided by the Institut de Développement et des Ressources en
Informatique Scientifique (IDRIS) in the frame of the Project 91066.
References
1. Eswaran, V., Pope, S.B.: An examination of forcing in direct numerical simulations of
turbulence. Computers & Fluids 16, 257–278 (1988)
2. Fessler, J.R., Kulick, J.D., Eaton, J.K.: Preferential concentration of heavy particles in a
turbulent channel flow. Phys. Fluids 6, 3742–3749 (1994)
3. Février, P., Simonin, O., Legendre, D.: Particle dispersion and preferential concentration
dependence on turbulent reynolds number from direct numerical simulation and large
eddy simulation of isotropic homogeneous turbulence. In: Proc. 4th Int. Conference on
Multiphase Flow, ICFM 2001 (2001)
4. Hopkins, M., Louge, M.: Inelastic microstructure in rapid granular flows of smooth disks.
Phys. of Fluids 3, 47–57 (1991)
5. Reade, W.C., Collins, L.R.: Effect of preferential concentration on turbulent collisions
rates. Phys. of Fluids 12, 2530–2540 (2000)
6. Schiller, L., Nauman, A.: A drag coefficient correlation. V.D.I. Zeitung 77, 318–320
(1935)
7. Squires, K.D., Eaton, J.K.: Preferential concentration of particles by turbulence. Physics
of Fluids A: Fluid Dynamics 3(5), 1169–1178 (1991)
8. Wang, L.P., Maxey, M.R.: Settling velocity and concentration distribution of heavy par-
ticles in homogeneous isotropic turbulence. J. Fluid Mech. 256, 27–68 (1993)
9. Wang, L.P., Wexler, A.S., Zhou, Y.: Statistical mechanical description and modelling of
turbulent collision of inertial particles. J. Fluid Mech. 415, 117–153 (2000)
Effect of Near-Wall Componental Modification
of Turbulence on Its Statistical Properties
Bettina Frohnapfel, Yosuke Hasegawa, and Nobuhide Kasagi
Abstract. Turbulence control techniques are of great economical and ecological

interest. In the present work a fundamental study is carried out in which body forces
are introduced in the near-wall region of a turbulent channel flow and thus modify
the near-wall behavior. It is investigated how these forces, which selectively act
on one of the velocity components, modify near-wall turbulence and its statistical
properties with the goal to extract properties that can directly be linked to the skin
friction drag. The alignment between the principal axis of the Reynolds stress tensor
and the mean flow direction is identified as an interesting quantity in this respect.
1 Procedure
We carry out direct numerical simulations of a fully developed channel flow with
a constant flow rate. The bulk Reynolds number based on the channel height is
Reb = 4460 and the corresponding friction Reynolds number for the uncontrolled
case is given by Reτ = 150. As a control input for modification of the near-wall
velocity fluctuation we consider a feedback body force, b f j , which is introduced
into the Navier-Stokes equations. The modified momentum equations ( j = 1, 2, 3)
thus read:
∂ u+j ∂ u+j ∂ p+ ∂ 2 u+j f (x+

+ 2) + +
+ u = − + + u − U , (1)
∂ t+ i
∂ x+
i ∂ x+j ∂ x+
i ∂ xi
+
$Φ
+
%&
j j
'
bf+
j
Bettina Frohnapfel
Center of Smart Interfaces, Technical University of Darmstadt, Petersenstr. 32,
64287 Darmstadt, Germany
e-mail: frohnapfel@csi.tu-darmstadt.de
Yosuke Hasegawa · Nobuhide Kasagi
Department of Mechanical Engineering, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku,
Tokyo 113-8656, Japan
e-mail: hasegawa/kasagi@thtlab.t.u-tokyo.ac.jp
128 B. Frohnapfel, Y. Hasegawa, and N. Kasagi
where the subscript ()+ indicates normalization with the kinematic viscosity, ν , and
the wall shear velocity, uτ , of the uncontrolled channel flow. The velocity compo-
nents in streamwise, wall-normal and spanwise direction are given by u1 , u2 , u3 and
u, v, w, respectively and the decomposition into mean and fluctuating components is
defined as: u j = U j + u
j . The last term on the right hand side of equation (1) repre-
sents the body force which is proportional to the corresponding velocity fluctuation.
The relaxation time constant,Φ + , determines the strength and type of the forcing,
and f (x+ 2 ) is a step function which determines the region where the forcing is ap-
plied on top and bottom wall: f (x+ + + +
2 ≤ y f ) = 1, f (x2 > y f ) = 0. The control input
can thus be varied by modifying the relaxation time constant,Φ + , and the forcing
layer thickness, y+f . For positive values of Φ + the introduced body force acts to en-
hance the velocity fluctuations in the corresponding direction; for negative values
of Φ + the introduced body force acts as a damping force. In the present study the
focus is placed on the influence of the forcing orientation. Therefore, the forcing
layer thickness is set to a constant value of y+f = 10.

Table 1 shows the obtained drag reduction for different exemplary test cases in
which a body force according to equation (1) is introduced within the forcing layer.
The results show that damping of the wall-normal velocity fluctuation, v
, and the
spanwise fluctuation, w
, leads to drag reduction (DR). The DR obtained with a
fixed value of Φ + is higher for w-forcing. This result is not surprising since the
body force depends on the strength of the velocity fluctuation itself and similar re-
sults have been reported by [5] who employed an exponentially decaying damping
force in the near-wall region. With increasing strength of the w
-damping force, the
theoretical maximum of DR (Iwamoto et al., 2005) is realized for (Φ + )−1 < −1 [1].
Table 1 Drag reduction (DR) obtained for channel flows in which different forcing is applied
within a forcing layer thickness of y+f = 10. The forcing strength and direction are given by
(Φ + )−1
forcing (Φ + )−1 DR
u’-damping -1/15 −6%

v’-damping -10/15 20%
w’-damping -10/15 42%
u’-increase 2/150 2%
(weak)
u’-increase 5/150 −8%
(strong)
v’-increase 1/15 −8%
w’-increase 5/150 −25%
Effect of Near-Wall Componental Modification of Turbulence 129
101
uj,rms
w‘-damping v‘-damping u‘-damping

10-4
w‘-increase v‘-increase u‘-increase

10-2 x*2 102
Fig. 1 Near-wall behavior of the rms values (u∗rms , v∗rms , w∗rms from top to bottom in each plot)
for the controlled flow (open symbols) in comparison to the uncontrolled channel flow (solid
symbols). The subscript ()∗ indicates that all values are normalized with the inner variables of
the corresponding flow. The plot for u’-increase contains results for both, weak (black open
symbols) and strong (grey solid symbols) forcing
This amount of DR can not be achieved with increasing the strength of v’-damping
up to (Φ + )−1 = −1000/15, which corresponds to the maximum forcing strength
considered in the present investigation.
For positive values of b f2 and b f3 , which increase v
and w
, respectively, drag
increase is observed. Again, a body force acting on the w-component (b f3 ) has a
more pronounced effect than b f2 with similar values of (Φ + )−1 . The general simi-
larity between modifications of the v- and w-component can be explained as a result
of streamwise vortices in the near-wall region which are damped or enhanced due
to the introduced body forces.
For body forces acting in the streamwise direction, b f1 , drag increase is observed
for damping of the streamwise fluctuations, u
. This is in contrast to the results of
[5] who reported slight DR for an exponentially decaying damping force in stream-
wise direction. In the present investigation, weak enhancement of the streamwise
velocity fluctuations leads to a slight DR, while stronger forcing enhances the skin
friction drag. Figure 2 shows the near-wall behavior of the velocity fluctuations of
the controlled flow in comparison to the uncontrolled channel flow. It can be seen
that strong body forces in spanwise or wall-normal direction significantly modify
the other near-wall components, too.
In [1] it was demonstrated that, for a variety of different flow control techniques,
DR is associated with increased anisotropy of turbulence in the near-wall region.
controlled
IIa flow
0.5 uncontrolled
flow
IIa=aijaji
IIIa=aijajkaki
-0.05 0.05 0.15 IIIa 0.25
Fig. 2 Anisotropy-invariant mapping for the uncontrolled flow and one of the controlled
(w’-damping, DR=42%) flows. The second invariant, IIa , captures the magnitude of
anisotropy and arrows mark the value of IIa at the wall
When the near-wall behavior of the present controlled flows is analyzed in the
anisotropy-invariant map (see figure 2), an increase of the second invariant, IIa , of
the anisotropy tensor, ai j [6]:
u
i u
j 1
ai j = − δi j , , IIa = ai j a ji , (2)
q2 3
in the near-wall region is observed for all drag reduced cases. This result is in agree-
ment with the above mentioned prior findings. It is interesting to check whether
a reversed trend in the near-wall turbulence anisotropy can be found for drag in-
creased flows. In the present investigation, the anisotropy in the near-wall region
is decreased for all drag increased flows, expect for the case of strong streamwise
forcing for which a slight increase of IIa is observed.
In general, the magnitude changes of the near-wall anisotropy cannot be related
quantitatively to the Reynolds shear stress, −u
v
, which forms the turbulent con-
tribution to the skin friction drag in a fully developed turbulent channel flow [3]:
1
12
cf = + 12 2(1 − y)(−u
v
)dy (3)
Reb 0
Further information about −u

v
can be found when considering the transformation
of the Reynolds stress tensor to its diagonal form which forms the basis for the
invariant analysis:
sin 2a
- u ' v' = - (EV 2 - EV 1)
2
8 0.40
channel, Ret=150
7 channel, Ret=100 0.35
6
w’-damping:
0.30
(?+)-1= -1/15
5 (?+)-1= -10/15 0.25
4 0.20
3 0.15
2 0.10
1 0.05
0 0.00
0 50 100 x*2 150 0 50 100 x*2 150
Fig. 3 Contributions to the Reynolds shear stress in terms of the eigenvalues of the Reynolds
stress tensor and the misalignment of the principle axis with the channel coordinate system.
Normalization is based on the inner variable of each flow
⎛ ⎞⎛

⎞⎛ ⎞ ⎛ ⎞
cos α sin α 0 uu uv 0 cos α − sin α 0 EV 1 0 0
⎝− sin α cos α 0⎠⎝u
v
v
v
0 ⎠⎝ sin α cos α 0⎠ = ⎝ 0 EV 2 0 ⎠(4)
0 0 1 0 0 w
w
0 0 1 0 0 EV 3
It is possible to express −u
v
in terms of the angle α of this rotation and the differ-
ence of the first and second eigenvalues, EV , of the Reynolds tensor:
sin2α
−u
v
= −(EV 2 − EV1) . (5)
2
Figure 3 shows the change in both terms of this identity for two w
-damping cases
in comparison with the uncontrolled channel flow. In order to distinguish between
Reynolds number effects (for the controlled flows Reτ = 130 and Reτ = 115, re-
spectively) and the influence of the applied control, channel flow data of Reτ = 100
is also included. It is obvious that the reduction of the angle α , i.e. the reduced
misalignment between the principal axis and mean flow direction, in the near-wall
region is the main reason for the reduction of the Reynolds shear stress and thus
DR, while the reduction of α in the outer flow region can be attributed to the de-
crease in Reynolds number. The same trends are found for v’-damping. For weak
u’-enhancement, only DR = 2.2% is achieved and the corresponding changes in
the quantities of equation (5) are small but they suggest the same conclusion: the
reduction of α can be identified as the source for the reduced Reynolds shear stress.
The drag increase for u’-damping, v’-increase and w’-increase is related to a
general increase of α while (EV 2 − EV 1) is reduced. Especially for u
-damping a
drastic increase of 0.5sin2α is observed in the forcing layer. It is accompanied by

a strong increase of the correlation coefficient for −u
v
in this region. This fact
explains the increase of the Reynolds shear stress in spite of damping of the stream-
wise velocity fluctuations u
.
For strong enhancement of the streamwise velocity fluctuations, an increase of
IIa is observed while the skin friction drag is increased. This result suggests that the
commonly observed trend of increased near-wall anisotropy for drag reduced flows
cannot be identified as a sufficient condition for DR. The analysis in terms of eigen-
values and alignment can shed some light into this point of discussion. Figure 4
shows the split-up of the corresponding Reynolds shear stress into its contributions
according to equation (5). In the near-wall region a reduction of the α -term in com-
parison to the uncontrolled channel flow can be observed. Although the Reynolds
number effect would correspond to an increase of α in the entire flow domain, it
remains identical to the one for an uncontrolled channel flow in the region of ap-
proximately 30 < y+ < 100 before shifting to higher values around the channel
center. In spite of the α -reduction in the near-wall region −u
v
increases due to
a large increase of (EV 2 − EV 1). This term can further be split up into the turbu-
lent kinetic energy, k = 0.5q2, of the flow field and the difference between first and
second eigenvalue of the anisotropy tensor, ai j , according to:
(EV 2 − EV1) = q2 (EV 2(ai j ) − EV 1(ai j )) (6)
The split up reveals that the large increase in the eigenvalue difference (EV 2 − EV 1)
and thus the increase of −u
v
can be related to the increase of turbulent kinetic en-
ergy in the flow field due to the enhancement of u
. This behavior suggests that the
high turbulent kinetic energy, induced by the u
-enhancing body force, suppresses
the drag reducing potential of the improved alignment in the near-wall region. In this
respect it should be noted that the energy input required to change v
and w
is gen-
erally small compared to the pumping power of the channel flow while significantly
higher energy input is needed for changes of u
.
In summary, the results of the present investigation suggest that flow control that
acts on the w’-component is most effective. Modifications of the v’-component
yield similar results but tend to be less effective resulting in lower gain, i.e. en-
ergy savings per energy input for the DR cases. Direct modifications of the stream-
wise velocity fluctuation yield the surprising result that drag increase is found for a
reduction of the turbulent kinetic energy in the flow field and vice versa (for small
increase of u
). Since these modifications require high energy inputs, however, they
do not seem a smart target for flow control. The misalignment angle, α , between
principle axis of the Reynolds stress tensor and the mean flow direction of the chan-
nel is introduced as an interesting quantity for flow control. At this point the present
results suggest that a reduction of α in the near-wall region which is achieved with
little energy input will result in drag reduction. Much further work will be needed
to clarify whether this fact can be used in the design of practical flow control tech-
niques. Simulations of w
-damping at Reτ = 300 and 450 show that the drag reduc-
tion rate achievable with near-wall componental modifications exhibits almost no
sin 2a
- u ' v' = - (EV 2 - EV 1)
2
10 0.4
channel, Ret=150
u’ -increase
(F+)-1=5/150
5 0.2
0 0
0 50 100 x*2 150 0 50 100 x*2 150
Fig. 4 Contributions to the Reynolds shear stress in terms of the eigenvalues of the Reynolds
stress tensor and the misalignment of the principle axis with the channel coordinate system.
Normalization is based on the inner variable of each flow
Reynolds number dependency, at least in the investigated regime, and confirm that
the reduction of the misalignment angle α is a key quantity for the reduction of the
Reynolds shear stress, −u
v
.
References
1. Frohnapfel, B., Hasegawa, Y., Kasagi, N.: Flow Control by Turbulence State Modifica-
tions in the Near-Wall Region. In: XXII Int. Cong. Theoretical and Applied Mechanics,
Adelaide, Australia (2008)
2. Frohnapfel, B., Lammers, P., Jovanović, J., Durst, F.: Interpretation of the mechanism
associated with turbulent drag reduction in terms of anisotropy invariants. J. Fluid.
Mech. 577, 457–466 (2007)
3. Fukagata, K., Iwamoto, K., Kasagi, N.: Contribution of Reynolds stress distribution to the
skin friction in wall-bounded flows. Phys. Fluids 14, L73–L76 (2002)
4. Iwamoto, K., Fukagata, K., Kasagi, N., Suzuki, Y.: Friction drag reduction achievable
by near-wall turbulence manipulation at high Reynolds number. Phys. Fluids 17, 011702
(2005)
5. Lee, C., Kim, J.: Control of the viscous sublayer for drag reduction. Phys Fluids 14, 2523–
2529 (2002)
6. Lumley, J.L., Newman, G.R.: The return to isotropy of homogeneous turbulence. J. Fluid.
Mech. 82, 161–178 (1977)
Large-Eddy Simulation of Transonic Buffet over
a Supercritical Airfoil
E. Garnier and S. Deck
1 Introduction
The transonic buffet is an aerodynamic phenomenon that results in a large-scale
self-sustained periodic motion of the shock over the surface of the airfoil. The time
scale associated to this motion is much slower than the one of the wall bounded
turbulence. It is then an appropriate case for URANS approaches and first attempts
with these methods have been reasonably successful in reproducing the mean fea-
tures of such flows. Nevertheless, as shown by Thiery and Coustols [2] results are
very sensitive to the turbulence model. Moreover, with some models, it is neces-
sary to increase the angle of attack with respect to experiment to obtain an unsteady
flow. Furthermore, they have evidenced a significant sensitivity of the results to the
confinement due to the wind tunnel walls. The first hybrid RANS/LES computation
on this configuration was performed by Deck [1] who has demonstrated that zonal
DES (ZDES) generally improves the results with respect to URANS computations
carried out with the Spalart-Allmaras model. In particular, the spectral content of the
pressure fluctuations in the separated zone is much more closer to the experimental
data with ZDES than with URANS. In this latter computation, the shock/boundary
layer interaction was treated in RANS mode and one of the purpose of the present
study is to assess the improvement that may result from a fully turbulent treatment
of the boundary layer on the suction side of the airfoil by means of LES.
More generally, the main objective of this study is to assess the capabilities
of LES to capture the buffet phenomenon. The large amount of data provided by
these simulations could then support the progress in the physical understanding
E. Garnier
ONERA, Applied Aerodynamics Department, 8 rue des Vertugadins, 92190 Meudon, France
e-mail: eric.garnier@onera.fr
S. Deck
ONERA, Applied Aerodynamics Department, 8 rue des Vertugadins, 92190 Meudon, France
e-mail: sebastien.deck@onera.fr
136 E. Garnier and S. Deck
of such flows. The validation of the computation is performed against the very
comprehensive experiment performed at ONERA by Jacquin et al. [3] which was
also used by Thiery and Coustols [2] and Deck [1].
2 Description of the Computation

The supercritical OAT15A airfoil was computed in the same flow conditions than in
the experiment by Jacquin et al.[3]. This airfoil has a chord of 230 mm and a relative
thickness of 12.3 %. Its angle of attack is equal 3.5 degrees. The free-stream Mach
number was set to 0.73 and the Reynolds number based on the chord length is equal
to 3 106.
The flow solver is the structured multiblock code FLU3M developed at ONERA.
It is second-order accurate in space and time. The numerical scheme dedicated to the
computation of the convective fluxes is based on a Roe scheme which was modified
to adapt locally its dissipation using the Ducros et al. sensor [4]. The dissipation
level of the scheme can be tuned by an additional modification of the sensor [5].
The Selective Mixed Scales Model has been chosen for this study [6]. The time step
has been imposed to 3.10−7s in order to ensure the convergence of the subiterative
process of the temporal implicit scheme using 5 subiterations.
In order to limit the required computational effort, the flow is computed in RANS
mode on the pressure side of the airfoil and in LES mode on the suction side and
in the wake. Moreover, RANS zones are treated in 2D. The grid refinement criteria
commonly used in LES of attached flows are satisfied (Δ x+ ≈ 50 in the longitudinal
direction, Δ z+ ≈ 20 in the spanwise direction and Δ y+ min ≈ 1 in the wall-normal di-
rection). Despite the zonal treatment of the flow, 20.8 millions of cells are necessary
to compute a domain width of only 3.65 % of chord in the grid A (Nz = 140). The
span and consequently the number of points were doubled to construct the grid B
(Nz = 280). This may be insufficient but the grid size results from a compromise
with the long integration time required to capture few buffeting periods. Two com-
putations denoted B1 and B2 have been carried out with the grid B. The first one is
based on a high dissipation set of parameters defining the modified Ducros sensor
whereas the second one corresponds to low dissipation parameters.
3 Mean Field Analysis

After a transient of 2 periods, the flow has been averaged over only one period of
the buffet phenomenon for the case A. The span was then doubled to generate the
grid B and, after a transient of one period, the statistics were collected over seven
additional periods to generate the results denoted B1. The dissipation level was then
diminished (B2 case) and after a one period transient, statistics were collected over
one period. Figure 1 presents an isovalue of the Q criterion colored by the lon-
gitudinal velocity. The separation occurs after the location of the shock identified
by one isovalue of the pressure (in purple). On this snapshot which corresponds to a
Large-Eddy Simulation of Transonic Buffet over a Supercritical Airfoil 137
Fig. 1 An isovalue of the Q criteria colored by the longitudinal velocity and one isovalue of
the pressure (purple) to mark the shock location (B1 case)
situation where the shock moves downstream, the flow is separated under the lambda
shock and near the trailing edge.
Figure 2 (left) shows the averaged pressure distribution on the airfoil. For both
cases A and B1, the buffet zone is shifted downstream by 7 % of chord with re-
spect to the experiment whereas the agreement is sligtly better in the case B2. This
demonstrates some influence of the numerical dissipation on the mean shock posi-
tion. In preamble to the discussion concerning the pressure fluctuations presented in
Fig. 2 (right), it is worth noticing that in the experiment the signal is collected over
2300 periods whereas at best 7 periods are available in the computation. In order to
permit fair comparisons between computation and experiment, the variability (min-
imum and maximum values) of the experimental pressure distributions averaged
over 7 periods is also reported in Fig. 2 (right). It is observed that the forward excur-
sion of the shock can significantly differ from its long term averaged value, its aft
excursion being more repeatable. The shock movement is far from being exactly pe-
riodic in the experiment and pressure fluctuations distribution averaged over such a
short time in the simulation can only be expected to lie between the upper and lower
bounds of the 7 periods averaged experimental distribution. Some improvement is
registered in the case B2 with respect to other cases. A correlation between the up-
stream excursion of the shock and the maximum level of fluctuation is evidenced,
a upstream movement of the shock giving a smaller level of pressure fluctuations.
A small part of this error (about 7 % at x/c = 0.9) can be attributed to the fact that
pressure signals were acquired up to a 5 kHz cut-off frequency but the rest of the
overestimation remains to be explained. The use of a doubled span (grid B) signifi-
cantly reduces the fluctuations near the trailing edge with respect to the original grid.
The analysis of the instantaneous fields obtained on grid A has evidenced that this
overestimation was due to the presence of intense two-dimensional coherent struc-
tures developing when the flow separates from the shock up to the trailing edge.
The span of the grid B allows the three-dimensionalisation of these structures which
limits their intensity and subsequently the wall pressure fluctuations.
The profiles of averaged and fluctuating longitudinal velocity at x/c=0.35 are
plotted in Fig. 3 for cases B1 and B2. These data evidence that upstream of the
Fig. 2 Averaged pressure coefficient distribution (left) and rms pressure distribution (right)
interaction the velocity field is well estimated, the agreement with the LDA (Laser
Doppler Anenometry) measurements being almost perfect for the B1 case. This re-
sult is far from being trivial since the flow undergoes a numerically forced transition
at the same station than in the experiment (x/c = 0.07). In the B2 case, the transition
occurs instantaneously at x/c = 0.07 whereas it takes some distance in the B1 case
(in Fig. 2 (right), the level of pressure fluctuations of the B1 case reaches the one of
the B2 case at x/c = 0.12). This leads to a fuller velocity profile in the B1 case than
in the B2 case, the former case being in fair agreement with the experiment both for
the mean and fluctuating velocity field. This fuller velocity profile is consistent with
a downstream shifted shock position for case B1 evidenced in Fig. 2. This suggests
that a possible improvement on the shock location deduced from pressure sensors
might be done at the expense of the agreement with the LDA measurements in the
velocity profiles upstream of the shock.
Fig. 3 Mean longitudinal velocity (left) and longitudinal velocity fluctuations (right) profiles
at x/c=0.35 (Grid B)
Downstream from the interaction (at x/c=0.75), one can observe in Fig. 4 that
the agreement of the simulations with both averaged and fluctuating longitudinal
velocity profiles is more than satisfactory. A better agreement with LDA data is
registered in the first third of the boundary layer for the B2 case than for the B1 case.
The contrary is observed above this limit. It is however worthwhile to notice that
between x/c=0.4 and x/c=0.6, experimental and numerical velocity profiles differ
significantly since the shock is not located at the correct mean position.
Fig. 4 Mean longitudinal velocity (left) and longitudinal velocity fluctuations (right) profiles
at x/c=0.75 (Grid B)
The signal duration being too short to undertake a physical analysis of the flow
with the B2 case results, the following discussion is based on the results of the B1
case.
4 Spectral Analysis
Due to the short duration of the LES simulations, an auto-regressive (AR) model
method has been used to compute the Power Spectral Density of the pressure. In-
deed, this method is well adapted to study short data that are known to consist of si-
nusoids in white noise[9]. The AR parameters are obtained with Burg’s method[10].
The pressure spectrum for x/c = 0.9 is compared to experiment in Fig. 5. The oc-
currence of strong harmonic peaks highlights the periodic nature of the motion.
On the experimental side, the main peak at 69 Hz represents the frequency of the
self-sustained motion of the shock over the airfoil. A slightly higher frequency near
72 Hz is found in the computation. The two first harmonics of the main peak are
also correctly captured numerically but additional harmonics are also observed in
the computation. This illustrates the fact that the periodicity of the computation is
stronger than in the experiment.
5 Space and Time Scales

Once, the main statistical and spectral features of the flow have been found, it is
worthwhile to study the kinematics of these pressure waves. To this end, let us
consider the fluctuating pressure at different stations. The two-point two-time cor-

1 ,t)P (x1 +Δ ξ ,t−τ )

relation coefficient: R (Δ ξ , τ ) = √P (x

2
√
2 establishes the correlation
(P (x1 )) (P (x1 +Δ ξ ))
Fig. 5 PSD of pressure fluctuations
between two signals located at abscissa x1 et x1 + Δ ξ and separated by a time delay

τ . The convection velocity can be obtained as the slope of the linear fitting of the
Δ ξ versus τmax (τmax represents the delay where the correlation coefficient reached
its maximum), as illustrated in Fig. 6.
Fig. 6 Propagation velocities obtained by a least square fitting of the linear relation between
the separation distance Δ ξ and time delay τ (filled symbol: exp, solid line: upper side of the
airfoil, dashed line: upper side of the airfoil)
On the upper-side of the airfoil, a downstream propagation velocity equal

6.11 10−3U∞ is clearly identified for the LES and appears to be slightly lower than
in the experiment. On the lower side of the airfoil, a forward motion at velocity
0.341 10−3U∞ is evidenced. The latter velocity is close to the upstream travelling
acoustic waves on the lower side of the airfoil.
6 Discussion
To assess the frequency of the motion, Lee[7] proposed that the period of the shock
oscillation should agree with the time it takes for a disturbance to propagate from
the shock to the trailing edge added to the time needed for an upstream moving
wave to reach the shock from the trailing edge. A simplified model has been used in
c−xs
reference [1] to assess the total duration to complete such a loop: T = vdownstream +
c−xs
|vupstream | where c is the chord and xs is the mean location of the shock wave. xS can
be obtained by noting the first abscissa where the skewness of pressure fluctuations
is zero. One gets (xs /c)LES = 0.52 while (xs /c)exp = 0.45. The velocity of upstream-
travelling acoustic waves is vupstream = a(M − 1) where a is the local speed of sound
in the field outside the separated area. With M = 0.8 and a = 330 m/s, the Lee’s
equation gives f = 1/T ≈ 110 Hz which is higher than the frequency FLES ≈ 70 Hz.
More recently, Crouch et al.[8] advocated that transonic buffet results from global
instability where the unsteadiness is characterized by phase-locked oscillations of
the shock and the separated shear layer. Within this scenario, the region downstream
from the shock is not the only region contributing to the feedback loop. Indeed, an
upstream travelling acoustic motion has been highlighted on the lower surface of
the airfoil (see Fig. 6). A deeper investigation of these phenomena will follow the
present work.
Acknowledgements. This work has been partly sponsored by the French National Research
Agency (project ANR-07-CIS7-009-04).
References
1. Deck, S.: AIAA J. 43, 1556–1566 (2005)
2. Thiery, M., Coustols, E.: Flow, Turbulence and Combustion 74, 331–354 (2004)
3. Jacquin, L., Molton, P., Deck, S., Maury, B., Soulevant, D.: AIAA paper 2005-4902
(2005)
4. Ducros, F., Ferrand, V., Nicoud, F., Weber, C., Darracq, D., Gacherieu, C., Poinsot, T.: J.
Comput. Phys. 152, 517–549 (1999)
5. Garnier, E.: Stimulated Detached Eddy Simulation of three-dimensional shock/boundary
layer interaction (to be published in Shock waves)
6. Lenormand, E., Sagaut, P., Ta Phuoc, L., Comte, P.: AIAA J. 38, 1340–1350 (2000)
7. Lee, B.H.K.: Aeronautical Journal, 143–152 (1990)
8. Crouch, J.D., Garbaruk, A., Magidov, D., Jacquin, L.: Proceedings of IUTAM confer-
ence, Corfou, GREECE (2008)
9. Trapier, S., Duveau, P., Deck, S.: AIAA J. 44, 2354–2365 (2006)
10. Burg, J.P.: In Modern Spectrum Analysis. In: Childers, D.G. (ed.), pp. 34–41. IEEE
Press, New-York (1978)
Large Eddy Simulation of Coherent Structures
over Forest Canopy
K. Gavrilov, G. Accary, D. Morvan, D. Lyubimov, O. Bessonov, and S. Méradji
Abstract. This paper deals with the numerical simulation (using a LES approach)
of the interaction between an atmospheric boundary layer (ABL) and a canopy, rep-
resenting a forest cover. This problem was studied for a homogeneous configuration
representing the situation encountered above a continuous forest cover, and a hetero-
geneous configuration representing the situation similar to an edge or a clearing in
a forest. The numerical results, reproduced correctly all the main characteristics of
this flow, as reported in the literature: the formation of a first generation of coherent
structures aligned transversally from the wind flow direction, the reorganisation and
the deformation of these vortex tubes to horse shoe structures. The results obtained,
introducing a discontinuity in the canopy (reproducing a clearing or a fuel break
in a forest), were compared with experimental data collected in a wind tunnel. The
results confirmed the existence of a strong turbulence activity inside the canopy at a
distance equal to 8 times the height of the canopy, referenced in the literature as an
Enhance Gust Zone (EGZ) characterized by a local peak of the skewness factor.
1 Introduction
The studies concerning the interactions between turbulence and canopy can inter-
est numerous environmental problems such as the evaluation of the heat and mass
K. Gavrilov · D. Lyubimov
Department of Theoretical Physics, Perm State University, Perm Russia
G. Accary
Université Saint Esprit de Kaslik, Jounieh Lebano
D. Morvan
UNIMECA, Technopôle de Château Gombert, 60 rue Joliot Curie 13453 Marseille cedex 13
France
e-mail: dominique.morvan@univmed.fr
O. Bessonov
Institute for Problem in Mechanics RAS, Moscow Russia
S. Méradji
M2P2 CNRS-Université dAix-Marseille, Marseille France
Corresponding author
144 K. Gavrilov et al.
transfer between the atmospheric boundary layer (ABL) and a forest, the effects
of wind gusts on trees plantations, the effects of wind on the behaviour of wild-
fires. Compared to a standard boundary flow, the structure of the flow above a forest
canopy presents more similarities with a mixing layer [1, 2]. The drag force induced
by the elements constituting the canopy (foliage, branches, twigs) slows down the
wind flow inside the vegetation layer. The resulting action of this force, induces a
significant shear of the free wind flow above the canopy, in a similar manner than the
situation observed in a mixing layer. For dense canopies, the flow inside the canopy
is structured in two zones [5]:
• a wake zone located just near the ground, where the streamwise velocity
profile is nearly uniform and mainly affected by the micro-wakes generated by
the elements (foliage, branches) composing the vegetation
• an exchange zone, located at the upper layer of the canopy, dominated by strong
shear effects and quick exchanges with the external flow
To observe the transition between a standard atmospheric boundary layer flow
(ABL) and a mixing layer (ML) type flow, the drag force induced by the canopy
must be larger than a threshold value. A critical parameter can be defined to express
the ABL/ML transition, using the product between the average drag coefficient Cd ,
with the Leaf Area Index (LAI). The transition is observed when this physical pa-
rameter is larger than a critical value nearly equal to 0.1 (Cd × LAI > 0.1) [4]. The
shear layer induced by the presence of the canopy upon the wind flow promotes the
development of a Kelvin-Helmholtz (KH) instability, characterized by the formation
large scale three dimensional structures with a horse shoe shape [2, 3]. The devel-
opment of such large scale structures, are responsible of the penetration of coherent
gusts inside the canopy, inducing longitudinal wave motions across the vegetation
called ”Honami” [4].
Many authors highlighted that the wavelength of these coherent structure - ΛX
(streamwise direction), was correlated to the shear length scale LS , defined as the
ratio between the average streamwise velocity with the vertical shear rate (LS =
dU/dz) evaluated at the top of the canopy z = H [6, 7].
In the present study, we demonstrated some characteristics stages of the coherent
turbulent structures formation, using large eddy simulation (LES) carried out for
continuous and discontinuous forest canopy.
2 Physical and Mathematical Model

Classically in a LES formulation, the equations of motion (momentum, continu-
ity) governing the atmospheric flow inside and above the canopy were filtered to
eliminate small scale perturbations [8]:

∂ ∂ ∂ p ∂ ∂ ui ∂ u j
[!t] (ρ ui ) + (ρ ui u j ) = − + μL + +
∂t ∂xj ∂ xi ∂xj ∂xj ∂ xi
∂ τi j ∂ u j
+ ρ aLCd (ui u j )1/2 ui , = 0, (1)
∂xj ∂xj
Large Eddy Simulation of Coherent Structures over Forest Canopy 145
Fig. 1 Snapshot of the vertical velocity component w calculated above a continuous canopy
(on the right)
where the brackets , denote the filtered variables: ui is the velocity component in
the xi -direction, p is the pressure, ρ is the air density, τi j is the subgrid-scale (SGS)
flux of momentum. The final term on the right hand side of Eq. (1) represents the
sink of momentum due to the presence of the vegetation. For this term, we intro-
duced two parameters: the leaf area density aL and the drag coefficient Cd . Because
of the actions resulting from the micro-wakes induced by the elements constitut-
ing the vegetation (foliage, twigs), the effective drag coefficient measured in-situ
inside a canopy, was smaller than the value which can be measured individually
for one particle in a wind tunnel. This phenomena is known in the literature, as the
shelter effect [1]. In agreement with experimental measurements, the average drag
coefficient Cd was chosen to be constant and equal to 0.15.
As adopted classicaly by the ABL community, the SGS momentum flux in Eq.
(1) is evaluated using the eddy viscosity concept calculated from the SGS kinetics
energy k and a length scale L = (Δ xΔ yΔ z)1/3 (where Δ x, Δ y, Δ z are the grid sizes
in the three directions of space) [8, 9, 10]. The SGS kinetics energy is predicted
solving the following transport equation [8]:

∂ ∂ ∂ ∂k
(ρ k) + (ρ u j k) = μT +
∂t ∂xj ∂xj ∂xj
∂ ui k3/2
τi j − ρ Cε − 2aL ρ Cd (ui u j )1/2 k. (2)
∂xj L
For the dissipation term (proportional to k2/3 ), we introduced two parameters: the
constant model Cε = 0.93 and one more time the SGS length scale L. The final term
in Eq. (2) represents the sink for SGS kinetics energy due to the rapid dissipation
of wake turbulence in the lee of elements constituting the vegetation layer (foliage,
twigs, branches). The set of transport equations (1), (2) were solved using a implicit
Finite Volumes method, using a QUICK third order scheme (with a flux limiting
strategy) for the convective terms and a second order central difference scheme for
the diffusive terms.

A first set of numerical simulations was performed for a continuous canopy. The
vertical structure of the vegetation (especially the distribution of the leaf area
Fig. 2 Turbulent structures (Q-isosurfaces) above a homogeneous forest canopy computed at

different times
density) was chosen to reproduce a pine forest (Jack pines), often encountered in
boreal forest, for which a large number of experimental campaigns were already re-
ported in the literature. This type of canopy is generally composed with two species
Black Spruce (height ∼ 10 m) and Jack Pine (height ∼ 20 m), the Leaf Area In-
dex (LAI) was here equal to 2. The dimensions of the computational domain were:
LX = 200 m, LY = 80 m - horizontal directions parallel and perpendicular to the wind
direction, LZ = 60 m = 3H - vertical direction. The boundary conditions imposed at
the limit of the domain were periodic at the vertical inlet and outlet faces and on
the two lateral faces. At the top of the domain we imposed a streamwise velocity
component equal to 3.8 m/s.
A snapshot of the distribution of the vertical component u3 = w is shown in
Figure 1. This result highlights clearly the formation of transverse coherent struc-
tures. The average streamlines allowed to extract the wavelength of these structures
ΛX , which was approximately equal to 8 times the shear length scale LS , in good
agreement with the previous results proposed by [6]. This first step for the organi-
zation of the flow, resulted from the development a KH instability, promoted by the
shear of the wind flow at the interface between the top of the canopy and the free
6 6 6
5 5 5
4 4 4
z/h z/h z/h
3 3 3
2 2 2
1 1 1
0 0 0
0 1 2 0 1 2 0 1 2
U/Uref U/Uref U/Uref
6 6 6
5 5 5
4 4 4
z/h z/h z/h
3 3 3
2 2 2
1 1 1
0 0 0
0 0.4 0 0.3 0 0.4
σU/Uref σU/Uref σU/Uref
Fig. 3 Vertical profiles of streamwise wind velocity U and standard deviation σU at various
distances downwind from the leading edge (x/H = −8.5, 2.1, 10.6), compared with experi-
mental data [12]
zone located just above (see Figure 2, top on the left). The second step was marked
by the clumping of these initial structures, into larger transverse rolls, connected
by regions of highly sheared flow (see Figure 2, top on the right). The third step is
the development of a secondary transverse instability, transforming the alignment of
transverse rolls, into horseshoe shape structures (see Figure 2, bottom).
Then the configuration was extended to study the modifications induced by the
creation of a discontinuity inside the forest cover. This particular configuration can
represent the situation encountered at a forest edge or in a clearing. In this case
the height of the canopy H was reduced to 7.5 m, the dimension of the clearing was
6 6 6
5 5 5
4 4 4
z/h z/h z/h
3 3 3
2 2 2
1 1 1
0 0 0
-1 SkU 0 1 -1 SkU 0 1 -1 0 SkU 1 2
Fig. 4 Vertical profiles of streamwise velocity skewness SkU at various distances downwind
from the leading edge (x/H = −8.5, 2.1, 10.6), compared with experimental data [12]
fixed to 21.3H, the origin (x = 0) was chosen at the exit side of the clearing, in such
a way that the point x/H = −8.5 was located inside the clearing, and the two other
positions x/H = 2.1 and 10.6 were located in the region newly cover by the forest.
The results shown in Figure 3 and 4, represent the distribution of average streamwise
velocity U, the standard deviation σU and the skewness factor SkU , calculated for
these three positions (x/H = −8.5, 2.1, 10.6), all the data were reduced using the
canopy height H and a reference velocity Ure f evaluated approximatively at the
middle of the clearing, 2H above the ground level.
As shown in Figure 3 and 4, the comparison theory/experience seems to be more
satisfacting for U, and σU , the agreement is less positive for SkU . For the profile
crossing the canopy (x/H = 2.1 and 10.6) the effects of the drag force induced by
the presence of the vegetation were correctly reproduced (in shape and intensity) for
all these statistical moments:
• The modules of U and σU for the streamwise component of the velocity vec-
tor were significantly reduced inside the canopy. We noticed an inflection point
for U at the top of the canopy (z/H = 1), favorising the development of a KH
instability.
• We noticed the existence of a second local maximum on the velocity profile
located near the clearing (x/H = 2.1), in the region located between the ground
level and the crown based height (CBH), where the vegetation was less dense.
Even if the agreement with experimental data for the skewness factor SkU was less
satisfactory (this high order moment, needs probably a smaller mesh size), the gene-
nal tendancy was correctly reproduced, i.e. positive values for the region located
inside the canopy (z/H ≤ 1), resulting from the incursion of gusts coming from the
free flow above the canopy [13].
4 Conclusion
The interaction between an atmospheric boundary layer (ABL) and a forest canopy
was simulated using a LES approach. The numerical results obtained above a con-
tinuous canopy confirmed that this flow presented numerous similarities with a mix-
ing layer. The visualisation of large scale coherent structures, highlighted how the
shear induced by the presence of the vegetation upon the free part of the flow, con-
tributed to the organization of the flow. The numerical results were compared with
experimental data obtained both for a homogeneous and an heterogeneous canopy.
Acknowledgements. The authors thank the European Commission for financial support in
the context of the Integrated Project Fireparadox (6th EU FP).
References
1. Kaimal, J.C., Finnigan, J.J.: Atmospheric Boundary Layer Flows. Oxford University
Press, Oxford (1994)
2. Finnigan, J.: Turbulence in plant canopies. Annu. Rev. Fluid Mech. 32, 519–571 (2000)
3. Marshall, B.J., Wood, C.J., Gardiner, B.A., Belcher, R.E.: Conditional sampling of
canopy gusts. Boundary-Layer Meteorol 102, 225–251 (2002)
4. Ghisalberti, M., Nepf, H.M.: Mixing layers and coherent structures in vegetated aquatic
flows. J. of Geophys. Res. 107(2), 1–11 (2002)
5. Ghisalberti, M., Nepf, H.M.: The structure of the shear layer in flows over rigid and
flexible canopies. Environmental Fluid. Mech. 6, 277–301 (2006)
6. Raupach, M.R., Finnigan, J.J., Brunet, Y.: Coherent eddies and turbulence in vegetation
canopies: the mixing-layer analog. Boundary-Layer Meteorol. 78, 351–382 (1996)
7. Brunet, Y., Irvin, M.R.: The Control of Coherent eddies in vegetation canopies: stream-
wise structure spacing, canopy shear scale and atmospheric stability. Boundary-Layer
Meteorol. 94, 139–163 (2000)
8. Watanabe, T.: Large-eddy simulation of coherent turbulence structures associated with
scalar ramps over plant canopies. Boundary-Layer Meteorol. 112, 307–341 (2004)
9. Moeng, C.H.: A large eddy simulation model for the study of planetary boundary-layer
turbulence. J. Atmos. Sci. 41, 2052–2062 (1984)
10. Patton, E.G., Shaw, R.H., Judd, M.J., Raupach, M.R.: Large-eddy simulation of wind-
break flow. Boundary-Layer Meteorol. 87, 275–306 (1998)
11. Accary, G., Bessonov, O., Fougère, D., Méradji, S., Morvan, D.: Optimized parallel ap-
proach for 3D modelling of foresr fire behaviour. In: Malyshkin, V.E. (ed.) PaCT 2007.
LNCS, vol. 4671, pp. 96–102. Springer, Heidelberg (2007)
12. Raupach, M.R., Bradley, E.F., Ghadiri, H.: A wind tunnel investigation into aerodynamic
effect of forest clearings on the nesting of abbott’s Boody on Christmas Island. Internal
report, CSIRO Centre for environmental Mechanics, Canberra, p. 21 (1987)
13. Shen, S., Leclerc, M.Y.: Modelling the turbulence structure in the canopy layer. Agricul-
tural and Forest Meteorol. 87, 3–25 (1997)
Toroidal/Poloidal Modes Dynamics in
Anisotropic Turbulence
Fabien S. Godeferd, Alexandre Delache, and Claude Cambon
1 Context
Isotropic turbulence receives a continuous effort for an increasingly refined
description, but complex effects modify the dynamics of turbulence, and are poorly
understood. Instances of distorted turbulence by external body forces are present
throughout natural and industrial flows, as in geophysical flows submitted to the
Earth’s rotation, and to density or temperature stratification. We focus here on the
effects of stable stratification and solid body rotation on the dynamics and structure
of homogeneous turbulence. We perform high resolution Direct Numerical Simula-
tions (DNS), to characterize the 3D structure of anisotropic turbulence and its sta-
tistical properties. Vertical structures appear in rotating turbulence, or a layering in
stably stratified turbulence, depending on the rotation rate and the density gradient,
parameters that are varied in our simulations (see [8]).
2 Numerical Approach
The divergence free and Navier-Stokes equations with body forces along with the
equation for buoyancy are as follows:
∂u 1 1
− ∇2 u = −∇(p + u · u )+ u × ∇ × u + 2Ω ez × u + beez
∂ t Re 2 $ %& ' $ %& ' $%&'
NL rotation stratification
Fabien S. Godeferd
LMFA, Université de Lyon, École Centrale de Lyon, France
e-mail: fabien.godeferd@ec-lyon.fr
Claude Cambon
LMFA UMR 5509, Université de Lyon, École Centrale de Lyon, France
Alexandre Delache
LAMUSE, Universit Jean Monnet, Saint-Étienne, France
152 F.S. Godeferd, A. Delache, and C. Cambon
∇·u = 0
∂b
− κ ∇2b = − (uu · ∇)b − N 2
$%&'w
∂t $ %& '
NL stratification
with N the Brunt-Vaisala frequency that contains the magnitude of the vertical
stable density gradient. u is the fluctuating velocity field, p the fluctuating pres-
sure divided by a reference density ρ0 and b the buoyancy term. Under Boussinesq
approximation, b is proportional to the gravitational acceleration g and to either the
potential temperature (for a gas) or the density fluctuation (for a liquid, for instance
b = gρ
/ρ0 ). The vertical direction, which bears both the buoyancy force and the
system vorticity, is referred to e z , with w = u · e z .
In the simulations, we set the Prandtl number Pr = 1. The above equations are
written in a frame rotating at constant angular velocity Ω corresponding to a Coriolis
parameter f = 2Ω . The Coriolis force appears in the Navier-Stokes equation as well
as the buoyancy force. The latter introduces the coupling with the buoyancy equation
that contains a source term related to the vertical velocity.
Assuming spatial homogeneity and space periodicity in the three directions,
the above equations are Fourier-transformed and solved using a classical pseudo-
spectral method with full de-aliasing and second order time scheme (see for instance
[10] or [11]. The computation is performed on a massively parallel computer at a
resolution of 10243 grid points.
The thus-obtained Eulerian velocity field is used for following the trajectory
of fluid particles. This provides Lagrangian statistics on the displacement of the
particles, the Lagrangian velocity distribution and the acceleration distribution.
The characterization of the anisotropy introduced by the presence of the Corio-
lis and buoyancy forces is therefore done both in the Eulerian and the Lagrangian
frameworks.
We start by simulating a high Reynolds number freely decaying isotropic turbu-
lent flow from initial random conditions. This field is important, since it provides the
start-up conditions for the anisotropic rotating stratified cases. Energetic constraints
and numerical resolution constraints are imposed by the available possible number
of grid points (here 10243), and the anticipation of the development of turbulence
submitted to the Coriolis and buoyancy forces, with an inhibited energy cascade
with respect to isotropic turbulence dynamics. The thus obtained isotropic turbulent
field is then impulsively submitted to rotation and stable stratification, without ad-
ditional forcing, with varying parameters: a case with a ratio f /N of rotation over
stratification of order 1/10, relevant for the Earths geophysical flows at mesoscales,
a rotation dominant case f /N > 1, and a non dispersive case with f /N = 1. These
cases correspond to low values of the Froude and Rossby numbers, typically of the
order of 0.01 to 0.1, corresponding to the ratios of buoyancy (or rotating) timescale
to the turbulence timescale.
Toroidal/Poloidal Modes Dynamics in Anisotropic Turbulence 153
3 Results
The incompressible velocity fields are decomposed in toroidal and poloidal modes.
This corresponds to a purely geometric decomposition in spectral space, thanks to
the orthogonality of the Fourier component of the velocity field with the correspond-
ing wave-vector, as shown on figure 1. In the purely stratified case, we identify
the role of the toroidal transfers that are responsible for the creation of directional
anisotropy, by lack of counterbalance by the poloidal modes that are phase scram-
bled, due to their wavy nature. In mixed stratified/rotating cases, the toroidal trans-
fers are also of mixed vortex/wave nature, but the specific role of the toroidal mode
is still well identified.
eP
z
eT
y
k
x
Fig. 1 Toroidal-poloidal decomposition (also known as Craya decomposition) of the spectral

velocity vector ûu, orthogonal to the wave-vector k from the incompressibility condition k · ûu =
0. The two components of velocity are along the unit vectors e P and e T
Directional spectra of the toroidal and poloidal modes are computed in the var-
ious parametric cases, as well as the directional potential energy spectrum, and are
shown along with the spherically averaged corresponding spectra, on figure 2.
The first row of figure 2 thus shows the power law slope obtained in the inertial
range for the four parametric cases. In the isotropic case, and in the non rotating
stratified case, or with small rotation (left three plots of first line), the inertial range
scaling is that of classical isotropic turbulence with Kolmogorov scaling E(k)
k−5/3 . The rightmost plot of first line on figure 2 corresponds to a larger rotation
rate that yields f /N > 1. In that case, as also observed in purely rotating turbulence,
the inertial range scaling corresponds to a steeper slope closer to k−3 , which is also
observed in models of inertial wave turbulence by Bellet et al. (2006) [2], in which
it is shown to differ from the classical 2D one, also expected as k−3 , because of a
different prefactor, and a marginally relevant inverse cascade.
4 Directional Spectra in the Fast Rotating Case

The isotropic turbulence directional spectra plotted on the leftmost column of
figure 2 show that, as expected, no dependence on the orientation θ of the wave-
vector is present. Note that the corresponding energy density spectra are scaled as
for the spherically averaged ones, for better comparison. Neither the poloidal or
toroidal spectra are anisotropic. There is of course no potential energy in this case.
The rapidly rotating case corresponds to the rightmost column of figures 2. The
physical meaning of the poloidal/toroidal decomposition in that case is a splitting
between inertial waves mode that have no particular difference other than the di-
rection of their propagation (see an interpretation of inertial waves triadic transfers
in [12], who calls them helical modes). Correspondingly, the toroidal and poloidal
spectra are not expected to be different. Their plots show that no strong anisotropy
is reflected in these spectra. Compared with the results at lower resolution presented
in Cambon et al. (1997) [4] or by Liechtenstein et al. (2005) [8], the anisotropy
seems to be smaller. One could explain this with three reasons: (a) rotation acts at
this stage of the turbulence decay over smaller scales, as hinted by the slight di-
rectional dependence in the small scales of the toroidal spectrum; (b) the transition
towards structures elongated along the rotation axis — like Taylor columns — has
not taken up yet (strong cyclonic eddies observed directly in experiments such as
those of McEwan (1976) [9], Jacquin et al. (1990) [7] — reflected by the dramatic
growth of a specific integral length scale — , not to mention LES by Bartello et al.
(1994) [1] and DNS by Liechtenstein et al. (2005) [8]), and the simulation should
be run for a longer time to reach this stage; (c) the imposed Brunt-Vaisala frequency
is of a magnitude close to the rotation frequency for this choice of parameter, and
the two anisotropic forces almost annihilate each other’s anisotropic effect.
5 Directional Spectra in the Strongly Stratified Cases

The two innermost middle columns of figure 2 show the poloidal and toroidal
energy spectra for the purely stratified case and the strongly stratified case with
weak rotation, respectively from left to right.
In both cases, the poloidal and toroidal spectra exhibit a characteristic large scale
anisotropy, which is expected from the effects of the buoyancy force, as already
observed by Godeferd & Staquet (2003) [5]. The closer the spectral direction to the
gravity axis, the higher the corresponding energy spectrum, with a one decade ratio
between the vertical and the horizontal spectra. Although the anisotropy decreases
in the small scales, the small scale anisotropy is still significant, and is larger in the
toroidal spectra than in the poloidal ones. The potential energy spectrum has an even
larger smallest scales anisotropy, showing the difference in the active scalar field
structure with respect to the velocity field, even at the chosen Prandtl number Pr =
1. Small scale anisotropy was also observed in previous DNS at lower Reynolds
numbers, which was assumed to be the cause of this non universality of the small
scales. It appears that, when increasing the Reynolds number, return to isotropy of
the small scales is indeed evidenced. Larger resolution simulations will confirm this
trend, but the present simulations correspond to a Reynolds number of the same
order of magnitude as in grid turbulence experiments with stratification, in which
similar magnitudes of anisotropy velocity tensor gradient are observed (see e.g. [6]).
One however would expect that at much higher Reynolds numbers, there will
still be a range of Rossby or Froude numbers at which the smallest scales will still
appear to be anisotropic, although whether the corresponding values of the rotation
and stratification parameters will be realistic is still a pending question, which is
rediscussed in section 7.
Isotropic Stratified f /N = 0.05 f /N = 1.15

1 1 1 1
sum poloide and toroide energy sum poloide and toroide energy sum poloide and toroide energy sum poloide and toroide energy
average spectral energy average spectral energy average spectral energy average spectral energy
k^(-5/3) k^(-5/3) k^(-5/3) k^(-5/3)
0.1 0.1 0.1 0.1
0.01 0.01 0.01 0.01
0.001 0.001 0.001 0.001
0.0001 0.0001 0.0001 0.0001
1e-05 1e-05 1e-05 1e-05
1e-06 1e-06 1e-06 1e-06

1 10 100 1000 1 10 100 1000 1 10 100 1000 1 10 100 1000
1 1 1 1
O2 O2 O2 O2
O3 O3 O3 O3
O4 O4 O4 O4
O5 O5 O5 O5
O6 O6 O6 O6
O7 O7 O7 O7
0.1 0.1 0.1 0.1
0.01 0.01 0.01 0.01
0.001 0.001 0.001 0.001
0.0001 0.0001 0.0001 0.0001
1e-05 1e-05 1e-05 1e-05

1 10 100 1000 1 10 100 1000 1 10 100 1000 1 10 100 1000
1 1 1 1
O2 O2 O2 O2
O3 O3 O3 O3
O4 O4 O4 O4
O5 O5 O5 O5
O6 O6 O6 O6
O7 O7 O7 O7
0.1 0.1 0.1 0.1
0.01 0.01 0.01 0.01
0.001 0.001 0.001 0.001
0.0001 0.0001 0.0001 0.0001
1e-05 1e-05 1e-05 1e-05

1 10 100 1000 1 10 100 1000 1 10 100 1000 1 10 100 1000
10 O2 10 O2 10 O2
O3 O3 O3
O4 O4 O4
O5 O5 O5
O6 O6 O6
O7 O7 O7
1 1 1
0.1 0.1 0.1
0.01 0.01 0.01
0.001 0.001 0.001

1 10 100 1000 1 10 100 1000 1 10 100 1000
Fig. 2 Energy spectra for the four different cases with the same initial Reynolds number,
from left to right the columns correspond to the following cases: isotropic, purely stratified
N = 11, stratified N = 11 with small rotation and f = 0.55, stratified N = 11 with stronger
rotation f = 12.6. From top to bottom, each row represents: the spherically averaged energy
spectrum E(k); the spectrum of the toroidal component of the velocity field E tor (k, θ ), re-
taining the angular dependence, that is averaged over angular sectors of spectral spheres; the
spectrum of the poloidal component of the velocity field E pol (k, θ ) with the same kind of
averaging; the potential energy spectrum E pot (k, θ )
6 Lagrangian Structure of Anisotropic Turbulence

From the alteration of the shapes of the energy spectra by stratification or ro-
tation, one deduces that the dynamics of energy exchange is also modified with
1 1
dU/dt dU/dt
dV/dt dV/dt
dW/dt dW/dt
0.1 0.1
0.01 0.01
0.001 0.001
PDF acceleration
PDF acceleration
0.0001 0.0001
1e-05 1e-05
1e-06 1e-06
1e-07 1e-07
1e-08
isotropic 1e-08
stratified
1e-09 1e-09
-20 -15 -10 -5 0 5 10 15 20 25 -25 -20 -15 -10 -5 0 5 10 15 20
1 1
dU/dt dU/dt
dV/dt dV/dt
dW/dt dW/dt
0.1 0.1
0.01 0.01
0.001 0.001
PDF acceleration
PDF acceleration
0.0001 0.0001
1e-05 1e-05
1e-06 1e-06
1e-07 1e-07
1e-08
stratified 1e-08
stratified
1e-09 weakly rotation 1e-09 rotation
-20 -15 -10 -5 0 5 10 15 20 25 -15 -10 -5 0 5 10 15
components of acceleration vector components of acceleration vector
Fig. 3 Pdf of Lagrangian acceleration components a normalized by a rms (t)
respect to the isotropic case, and therefore that the structure of the flow will appear
to have a different intermittency. This is clear in physical space, when observing
iso-vorticity surfaces in the three-dimensional domain, in which elongated struc-
tures appear, along the rotation axis when rotation is dominant (see the analysis of
the vorticity statistics in [3]), or orthogonal to the axis of gravity when stratification
is dominant.
The intermediate case with stratification and rotation equally balanced is quite
interesting, since it exhibits a kinetic energy spectrum that does not depend on the
wave-vector orientation, but has a modified slope with respect to the classical k−5/3
isotropic law. The modified dynamics is then responsible for a much different spatial
distribution of the flow vortices, unlike the well-known “worm”-like structures.
The intermittency is characterized here by the probability density functions of
the normalized Lagrangian acceleration, which we have computed in all the para-
metric cases, shown on figure 3. The normalization is here performed by the rms
acceleration at each time, to compensate for the decay of turbulence energy.
In the isotropic case, a classical shape of the pdf is observed, the same for all
the components of the acceleration vector. Exponential tails are observed as well as
large values of the acceleration.
In the strongly stratified cases (top right figure and bottom left figure), the peak
of the distribution is slightly thicker, so that the fit by a Gaussian distribution would
lead to a somewhat larger rms value. The shape of the acceleration pdf tails is mod-
ified in consequence. This could be a trace of stronger acceleration regions in be-
tween the sheared layers of stratified turbulence. However, no significant anisotropy
appears. A slight asymmetry is hinted, but more thorough investigation is needed to
decide whether it could be only due to the statistics sampling.
Finally, in the strongly rotating case (bottom right figure), a different distribution
is observed. The largest accelerations are smaller than in the isotropic or in the
stratification dominant cases, and, in the larger acceleration values, one could detect
a difference of the distribution between the component of acceleration along the axis
of rotation or orthogonal to it.
Overall, the Lagrangian anisotropy exhibit on figure 3 is insignificant with respect
to the spectral Eulerian anisotropy of figure 2. Since the acceleration is essentially
related to the pressure, it is expected to be less directly influenced than the Eule-
rian velocity by the anisotropic external forces. Moreover, being a derivative of the
velocity, the acceleration vector represents smaller scales, and we have seen above
that there seems to be a return to isotropy of the small scales.
7 Future Work
In the above, we have briefly presented some features of rotating stratified turbu-
lence, characterized by Eulerian spectral statistics and Lagrangian statistics. The
relation between the Eulerian structure of turbulence and the Lagrangian distribu-
tion has been investigated in previous works in the isotropic case, but it appears that
anisotropic turbulence deserves a specific investigation. We have here considered
for instance the effect of the Coriolis force or a buoyancy effect in stably stratified
fluid, but instances of such anisotropic structuring can also be found in other con-
texts, e.g. in a conducting fluid placed within a background magnetic field, by means
of the Lorentz force.
Another issue is the whereabouts of the very large anisotropy we have observed
in the strongly stratified cases, when the Reynolds number increases, of concern
for geophysical applications. Concerning spectral scaling, a clear conclusion of this
work is that even though one could observed Kolmogorov-like scaled spherically
averaged spectra of kinetic energy, a refined description of directional dependent
energy spectra shows that this scaling is by no means universally applicable. This
consideration is of special importance when comparing to experimentally measured
spectra, which very often correspond to the choice of a particular single orientation
of the probes.
The problem of re-isotropization of small scales is a very complex one. In stably-
stratified turbulence, it is suggested that re-isotropization
is possible for scales
smaller than an Oszmidov length scale, or LO ∼ ε /N 3 , where ε is the dissipation.
A similar scale can be defined for rotating turbulence, just replacing N by 2Ω . The
role of an Oszmidov-type scale is consistent with some recent studies using the con-
cept of ‘critical balance’, which was suggested in MHD turbulence, but these studies
ignore the detailed description and dynamics of anisotropy and its specificity in the
different cases (rotating, stratified, MHD, mixed. . . ). At least in rotating turbulence,
it is clear that anisotropy, quantified by the scattering of angle-dependent spectra,
begin to increase with increasing wavenumber, so that a very large intermediate
spectral zone, in which the anisotropy can decrease, should separate the anisotropic
scales from an hypothetic ‘isotropized tail’ for spectra. Such a configuration must
involve a very large spectral domain, with conditions of very high Reynolds number
Re and very small Rossby number Ro outside the scope of present DNS. In addition,
a condition such as moderate (not too small) ReRo2 is probably required, in analogy
with stratified turbulence, in which the corresponding parameter (using the Froude
number Fr) ReFr2 is known for a long date by oceanographers as the ‘buoyancy
efficiency parameter’. Finally, the problem of isotropy in the dissipative range
remains open.
Acknowledgments
The simulations presented in this work were performed at IDRIS (France), thanks to computer
time allocated through grant number 071433.
References
1. Bartello, P., Métais, O., Lesieur, M.: Coherent structures in rotating three-dimensional
2. Bellet, F., Godeferd, F.S., Scott, J., Cambon, C.: Wave turbulence modelling in rapidly
rotating flows. J. Fluid Mech. 562, 83–121 (2006)
3. van Bokhoven, L., Cambon, C., Liechtenstein, L., Godeferd, F., Clercx, H.: Refined vor-
ticity statistics of decaying rotating three-dimensional turbulence. J. of Turbulence 9(6)
(2008)
4. Cambon, C., Mansour, N.N., Godeferd, F.S.: Energy transfer in rotating turbulence. J.
Fluid Mech. 337, 303–332 (1997)
5. Godeferd, F.S., Staquet, C.: Statistical modelling and direct numerical simulations of
decaying stably-stratified turbulence: Part 2: Large scales and small scales anisotropy. J.
Fluid Mech. 486, 115–150 (2003)
6. Itsweire, E.C., Helland, K.N., Van Atta, C.W.: The evolution of grid-generated tubulence
in a stably stratified fluid. J. Fluid Mech. 162, 299–338 (1986)
7. Jacquin, L., Leuchter, O., Cambon, C., Mathieu, J.: Homogeneous turbulence in the pres-
ence of rotation. J. Fluid Mech. 220, 1–52 (1990)
8. Liechtenstein, L., Godeferd, F., Cambon, C.: Nonlinear formation of structures in rotat-
ing stratified turbulence. J. of Turb. 6, 1–18 (2005)
9. McEwan, A.: Angular momentum diffusion and the initiation of cyclones. Nature 260,
126–128 (1976)
10. Rogallo, R.S.: Numerical experiments in homogeneous turbulence. Tech. Rep. 81315,
NASA Technical Memorandum (1981)
11. Vincent, A., Meneguzzi, M.: The spatial structure and statistical properties of homoge-
neous turbulence. J. Fluid Mech. 225, 1–20 (1991)
12. Waleffe, F.: Inertial transfer in the helical decomposition. Phys. Fluid A(5), 677–685
(1993)
Grid Filter Modeling for Large-Eddy Simulation
Marc A. Habisreutinger, Roland Bouffanais, and Michel O. Deville
Abstract. An interpretation to the use of deconvolution models when used in

implicitly filtered large-eddy simulations as a way to approximate the projective
grid filter is given. Consequently, a new category of subgrid models, the grid filter
models, is defined. This approach gives a theoretical justification to the use of de-
convolution models without explicit filtering of the solution and explains how the
use of such models can be effective in this context.
This viewpoint also allows to consider a new way of designing the convolution
filter which has to approximate the grid filter and therefore a new way of improving
such subgrid models. In this framework, a general technique for the approximation
of the grid filter associated with any function-based numerical method is proposed.
The resulting subgrid model is parameterless, only depends on the mesh used for
the large-eddy simulation which is a priori known and vanishes locally if the flow is
not turbulent, thereby ensuring the consistency of the model with the Navier–Stokes
equations.
1 Introduction
In this work, the focus is put on large-eddy simulation (LES) based on deconvolu-
tion subgrid models which aim at a partial recovery of the full velocity field from
Marc A. Habisreutinger
École Polytechnique Fédérale de Lausanne, STI–IGM–LIN, Station 9, CH–1015 Lausanne
e-mail: marc-antoine.habisreutinger@epfl.ch
Roland Bouffanais
Massachusetts Institute of Technology, Vortical Flow Research Laboratory,
Cambridge MA 02139
e-mail: bouffana@mit.edu
Michel O. Deville
École Polytechnique Fédérale de Lausanne, STI–IGM–LIN, Station 9, CH–1015 Lausanne
160 M.A. Habisreutinger, R. Bouffanais, and M.O. Deville
its filtered counterpart by attempting to invert the filtering operator applied to the
Navier–Stokes equations.
Deconvolution models are purely algorithmic as they only rely on the definition
of the filter without the need to resort to any physical modeling. As noted by
Domaradzki and Adams [1], “since the need for the physical models is removed
this approach seems to be more promising than the classical models. However, this
promise is not fulfilled if the effects of numerical discretization are not accounted
for”. Moreover, for a given filter, its inverse may be analytically determined which
theoretically allows to express the full velocity field as a function of the filtered
field. Therefore, solving the filtered Navier–Stokes equations should in principle de-
liver as much information as the solution of the standard non-filtered Navier–Stokes
equations. This apparent paradox is resolved by acknowledging the unavoidable and
irreversible effects of the implicit filter associated with the numerical discretiza-
tion. This point has long been noted by LES practioners, e.g. by the earlier work
of Zhou et al. [2], and more recently by Langford and Moser [3], Domaradzki and
Loh [4], and Winckelmans et al. [5]. Nevertheless, the discretization effects are very
often considered for the simple case of node-based methods but very few studies for
function-based methods—e.g. spectral and finite/spectral element methods—are re-
ported in the literature. Consequently, many of the arguments commonly expressed
about grid filtering, are neither valid nor applicable when a function-based method
is used as in the present study using the spectral element method (SEM).
When no explicit filter is applied but only the implicit grid filter is considered,
deconvolution models can actively contribute to subgrid modeling but in a very dif-
ferent way than the one associated with LES relying on explicit filtering techniques.
As a first step, an interpretation to the use of deconvolution models when used with
implicit filtering as a way to approximate the projective grid filter is given. Conse-
quently, a new category of subgrid models, the grid filter models (GFM) is defined.
The GFM approach gives a theoretical justification to the use of deconvolution mod-
els without explicit filtering of the solution and explains how the use of such models
can be effective in this context.
This viewpoint also allows to consider a new way of designing the convolution
filter which has to approximate the grid filter and therefore a new way of improving
such subgrid models. In this framework, a general technique for the approximation
of the grid filter associated with any function-based numerical method is proposed,
which addresses the previous note of Domaradzki and Adams.
2 Governing Equations
In the case of isothermal flows of Newtonian incompressible fluids, the LES
governing equations for the filtered quantities denoted by an overbar, obtained by
applying a convolution filter G to the Navier–Stokes equations, read

∂ ui ∂ ∂p ∂ ∂ ui ∂ u j ∂ τi j
+ (ui u j ) = − +ν + − , (1)
∂t ∂xj ∂ xi ∂ x j ∂ x j ∂ xi ∂xj
Grid Filter Modeling for Large-Eddy Simulation 161
∂uj
= 0, (2)
∂xj
the filtered velocity field u = G u satisfying the divergence-free condition (2)

through the filtered reduced pressure field p. The components of the subgrid
tensor τ are given by
τi j = ui u j − ui u j , (3)
and ν is the kinematic viscosity. The closure of the filtered momentum equation (1)
requires τ to be expressed in terms of the filtered field which reflects the subgrid
scales modeling and the interaction among all space scales of the solution.
The numerical method treats equations (1) and (2) within the weak Galerkin
formulation framework. The spatial discretization relies on the SEM, see the
monograph by Deville et al. [6] and Habisreutinger et al. [7] for full details.
3 Subgrid Modeling
The problem of subgrid modeling consists in taking into account the interaction
between resolved and subgrid scales which is represented by the additional sub-
grid term in the filtered momentum equation (1) and thereby expressing the subgrid
contribution as a function of the resolved quantities.
3.1 Deconvolution Models

The deconvolution approach aims at reconstructing the unfiltered fields from the
filtered ones, the subgrid modes being not modeled but reconstructed using an
ad hoc mathematical procedure. Borrowing the notation adopted in Habisreutinger
et al. [7], the filtered Navier–Stokes momentum equation is written formally as
∂u
+ f(u) = [f, G ](u), (4)
∂t
where [f, G ](u) is the subgrid commutator, which only retains the nonlinear and
non-commutating terms. This equation is strictly equivalent to (1) with
∇ · τ.
[f, G ](u) = −∇ (5)
The exact subgrid contribution appears as a function of the non-filtered field, which
is not computed when performing a LES. This field being unknown, the idea is to
approximate it using a deconvolution procedure. For instance, Stolz & Adams [8]
proposed the following procedure
u u∗ = QN u = (QN ◦ G ) u, (6)
where QN is an Nth-order approximation of the inverse of the filter G such as

N
QN ◦ G = I + O(Δ ), Δ being the filter cutoff length and I the identity filtering
operator. The subgrid term is then approximated as
[f, G ](u) [f, G ](QN u). (7)
3.2 Grid Filter Models

The previous developments do not take explicitly into account the major role played
by the grid filter in LES. Indeed, the filter G to consider in practical simulations is
a composition of the convolution filter L and the projective grid filter, referred to
as P [9, 10] and represented by a hat in the sequel, such that
G = (L ◦ P) . (8)
Considering the grid filter in the filtered Navier–Stokes equations leads to
∂u
+ f(u) = [f, (L ◦ P)](u). (9)
∂t
In the sequel, the focus is put on the particular case where implicit grid filtering is the
only effective filter. The objective is to give a theoretical interpretation allowing to
use a deconvolution model in this framework. When no LES filter L is explicitly
applied, the filtered Navier–Stokes equations reduce to
∂ û
+ f(û) = [f, P](u), (10)
∂t
where û is the grid filtered velocity, or in other words the part of the velocity field
that can be resolved by the grid used to perform the LES. In this framework, subgrid
modeling based on deconvolution models requires the subgrid commutator of equa-
tion (10) to be expressed solely in terms of the known projected velocity û, which
formally reads
[f, P](u) = [f, P](P −1 û), (11)
where u = P −1 û is the formal inverse grid filtering operation to be devised. It is
worth noticing that due to its implicit nature, the grid filter P entirely depends on
the numerical method used to discretize in space the Navier–Stokes equations.
The problem with equation(11) is that P −1 does not exist, this filter being a
projector. The central idea introduced in Habisreutinger [11] and subsequently in
Bouffanais [12] is then to approximate P with an invertible filter reproducing as
closely as possible its effect
P M , (12)
in order to have
[f, P](u) [f, M ](M −1 û) = [f, M ](u• ), (13)

with u• = M −1 û. For such subgrid models, modeling only requires to build M
in order to achieve the best approximation of P, as stated in equation (12). The
problem is exactly solved if P = M , which is not achievable since P applies
on an infinite spectrum while M only applies to computable wave numbers [1].
Moreover, M must be invertible while P is projective. The approximation (12)
being the only required modeling, the subgrid models arising from this new ap-
proach are referred to as grid filter models. The governing equations to be solved in
the GFM framework are obtained by combining equation (10) with approximation
(13), and read
∂ û
+ f(û) = [f, M ](u• ). (14)
∂t
3.3 Grid Filter Modeling

As described in section 3.2, the GFM approach only needs to build an approximation
of the grid filter in order to model the effect of the subgrid scales. This approxima-
tion can be performed in many different ways and from many different viewpoints
which opens a full field of research.
In this work, the grid filter is approximated through a transfer matrix in a modal
basis which was proposed in the p-version of finite elements and first used by Boyd
[13] as a filtering technique. The transfer matrix is build in oder to preserve C0 -
continuity across the elements, to be invertible and to approximate the grid filter. Its
construction relies on an exact determination of the projection operation from the
continuous space onto the LES mesh and on a statistical approximation of the un-
known subgrid field which is based on a reciprocal probabilistic existence between
every pair of modes of the spectrum. Consequently, the resulting subgrid model is
parameterless, only depends on the LES mesh which is a priori known and vanishes
locally if the flow is not turbulent, thereby ensuring the consistency of the model
with the Navier–Stokes equations.
4 Lid-Driven Cavity Flow Simulation

A DNS of the flow in a lid-driven cubical cavity performed at Reynolds number of
12’000 with a Chebyshev collocation method due to Leriche and Gavrilakis [14]
is taken as the reference solution to validate this new modeling approach. Subgrid
modeling in the case of a flow with coexisting laminar, transitional and turbulent
zones such as the lid-driven cubical cavity flow represents a challenging problem.
As the flow is confined and recirculating, any under- or over-dissipative character
of the subgrid model can be clearly identified. Moreover, the very low dissipation
and dispersion induced by SEM allow a pertinent analysis of the energetic action
induced by any subgrid model, which is not feasible in the framework of low-order
numerical methods. The coupling of the lid-driven cubical cavity flow problem with
the SEM builds therefore a well suited framework to analyze the accuracy of the
newly defined subgrid model.
1 1 1
0.5 0.5 0.5

y/h
y/h
y/h
0 0 0
-0.5 -0.5 -0.5
-1 -1 -1
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x/h x/h x/h
Fig. 1 Contours of û◦x û◦y from −0.0007 to 0.0065 in U02 units, by increments of 0.0002 in
the midplane z/h = 0. DNS (left), GFM (center), ILES (right). Dashed contours correspond
to negative levels. The cavity lid y/h = 1, moves at U0 in the x-direction
In an attempt to provide a comprehensive assessment of the performances of the

GFM approach, the determination of the Reynolds stress tensor components has
been envisaged as a challenge in the framework of the lid-driven cubical cavity
flow. The Reynolds statistical decomposition u = u + u◦ , introduces the average
value denoted into brackets and its fluctuating part u◦ . The statistics for all LES are
based on a sampling approximately 10 times smaller than the one of the DNS, more
precisely 400 samples are collected over 80 time units.
Figure 1 displays the low-amplitude cross term û◦x û◦y in the symmetry plane
of the cavity. The contours for Implicit LES (ILES) show that implicit modeling is
totally inoperative in the SEM framework and highlight the need for explicit subgrid
modeling as well as the under-resolution of the flow. The contours for GFM are
relatively close to the DNS ones taking into account the reduced sampling and the
low amplitude of the field as compared to the lid velocity U0 = 1. This allows to
asses the validity of this new modeling approach. However, further investigations
and researches are required to improve or modify the grid filter modeling strategy
briefly described in section 3.3, and to evaluate the relative performance of this
models as compared to existing ones.
5 Conclusion and Perspectives

In conclusion, the GFM approach gives a theoretical justification to the use of de-
convolution models without explicit filtering of the solution and explains how the
use of such models works in this context. This viewpoint allows to consider a new
way of designing the convolution filter which has to approximate the grid filter and
therefore a new way of improving such subgrid models.
The results obtained so far demonstrate the validity of this modeling approach
but further investigations are required to improve the grid filter modeling and to
compare this models with existing ones.
Since the need for the physical models is removed, the validity of the GFM ap-
proach extends beyond the limited scope of incompressible Newtonian fluid flows
considered in this article. For example, LES of compressible and viscoelastic fluid
flows can also be envisaged.
Acknowledgments
This research is being partially funded by a Swiss National Science Foundation Grant (No.
200020-112085), whose support is gratefully acknowledged.
References
1. Domaradzki, J.A., Adams, N.A.: Direct modelling of subgrid scales of turbulence in
large-eddy simulations. J. of Turbulence 3(24)
2. Zhou, Y., Hossain, M., Vahala, G.: A critical-look at the use of filters in large-eddy
simulation. Phys. Lett. A 139, 330–332 (1989)
3. Langford, J.A., Moser, R.D.: Optimal LES formulations for isotropic turbulence. J. Fluid
Mech. 398, 321–346 (1999)
4. Domaradzki, J.A., Loh, K.C.: The subgrid-scale estimation model in the physical space
representation. Phys. Fluids 11, 2330–2342 (1999)
5. Winckelmans, G.S., Wray, A.A., Vasilyev, O.V., Jeanmart, H.: Explicit-filtering large-
eddy simulation using the tensor-diffusivity model supplemented by a dynamic
Smagorinsky term. Phys. Fluids 13, 1385–1403 (2001)
6. Deville, M.O., Fischer, P.F., Mund, E.H.: High-order methods for incompressible fluid
flow. Cambridge University Press, Cambridge (2002)
7. Habisreutinger, M.A., Bouffanais, R., Leriche, E., Deville, M.O.: A coupled approximate
deconvolution and dynamic mixed scale model for large-eddy simulation. J. Comput.
Phys. 224, 241–266 (2007)
8. Stolz, S., Adams, N.A.: An approximate deconvolution procedure for large-eddy simu-
lation. Phys. Fluids 11, 1699–1701 (1999)
9. Winckelmans, G.S., Jeanmart, H.: Assessment of some models for LES without/with ex-
plicit filtering, Geurts, Friedrich and Métais Edition. In: Direct and large-eddy simulation
IV, pp. 55–66. Kluwer, Dordrecht (2001)
10. Gullbrand, J., Chow, F.K.: The effect of numerical errors and turbulence models in large-
eddy simulations of channel flow, with and without explicit filtering. J. Fluid Mech. 495,
323–341 (2003)
11. Habisreutinger, M.A.: Large-eddy simulation of turbulent flows by approximate decon-
volution models using spectral element method. Master’s thesis, École Polytechnique
Fédérale de Lausanne, Section de Génie Mécanique (2006)
12. Bouffanais, R.: Simulation of shear-driven flows: transition with a free surface and con-
fined turbulence. Ph.D. thesis, no. 3837, École Polytechnique Fédérale de Lausanne
(2007)
13. Boyd, J.P.: Two comments on filtering (artificial viscosity) for Chebyshev and Leg-
endre spectral and spectral element methods: Preserving boundary conditions and
interpretation of the filter as a diffusion. J. Comput. Phys. 143, 283–288 (1998)
14. Leriche, E., Gavrilakis, S.: Direct numerical simulation of the flow in the lid-driven
cubical cavity. Phys. Fluids 12, 1363–1376 (2000)
Pulsating Flow through Porous Media
Michele Iervolino, Marcello Manna, and Andrea Vacca
Abstract. The present work investigates the response of a porous media to an

unsteady forcing resulting from the superposition of an harmonic component to a
mean one. The analysis is carried out both in terms of global parameters and local
fields obtained processing data from numerical solution of the Navier–Stokes equa-
tions at pore level performed with a spectrally accurate multi–domain algorithm.
1 Introduction
The transport of fluid through porous media is encountered in several groundwater,
marine, chemical and mechanical engineering applications. In many circumstances
the relevant global flow parameters may be predicted through the Darcy’s law,
i.e. assuming steady flow conditions and negligible inertial effects. However, in
some of the above applications both unsteadiness and considerable inertia may
be encountered. While many efforts have been devoted to the description of the
purely oscillatory flow [13, 14, 10, 3, 9, 11], the pulsating one has been far less
investigated.
Starting from experimental evidences and accounting for the acceleration of
the fluid in the global momentum equation, several corrections to the Darcy–
Forchheimer’s law have been proposed [8, 6, 2]. The first attempt to extend the
analysis also at the pore level is due to Graham & Higdon [5] who considered two–
dimensional periodic media consisting of constricted channels or cylinder arrays
Michele Iervolino · Andrea Vacca
Dipartimento di Ingegneria Civile, Seconda Universitá di Napoli, Via Roma, 29, 81031
Aversa (Ce), Italy
e-mail: michele.iervolino@unina2.it,vacca@unina.it
Marcello Manna
Dipartimento di Ingegneria Meccanica per l’Energetica, Universitá di Napoli ”Federico II”,
Via Claudio, 21, 80125 Napoli, Italy
e-mail: manna@unina.it
168 M. Iervolino, M. Manna, and A. Vacca
with high solid volume fraction (φ ). In [5] it has been shown that an oscillatory
forcing larger than the mean one may produce a reduction of the time–averaged
flow rate up to 40% of the steady state value. The present paper analyses the pul-
sating flow field in a two-dimensional porous medium whose base element consists
of square cylinders with moderate solid volume fraction. The analysis is carried out
solving the unsteady Navier-Stokes equations at the pore scale.
2 Flow Problem and Computational Setup

The problem under investigation is the pulsating flow through a periodic array of
l × l square cylinders spaced by 2h (l = 3h, φ = 0.36, see figure 1a) driven by a
harmonically time varying pressure gradient Δ p/L:
Δ p Δ p Δ posc
= + cos (ω̃ t˜) . (1)
L L L
The time mean value Δ p/L, the amplitude Δ posc /L, and the pulsation ω̃ = 2π /T̃ ,
T̃ being the dimensional oscillating period, have been chosen to span at least one
order of magnitude of the governing dimensionless parameters. Among the possible
triplets of the relevant dimensionless parameters we shall use the following ones:
Δ p h3 ρ Δ posc h3 ρ h
Re = , Re osc = , χ=
, (2)
L μ2 L μ2 δ

in which μ and ρ are the fluid viscosity and density and δ = 2μ /ρ ω̃ is the Stokes
layer thickness. Table 1 summarizes the investigated space parameters.
Table 1 Run matrix
Re 32 64 128
Reosc 32 64 96 128 32 64 128 32 128
χmin 0.50 0.32 0.50
χmax 12.94 12.94 12.94
The analysis is carried out solving at the pore scale the unsteady 2–D Navier–
Stokes equations with the spectral algorithm developed and validated in [7]. The
computational domain Ω , shown with dashed line in figure 1a, consists of a square
region with dimensions 10h × 10h. No-slip boundary conditions are imposed at the
solid walls, while periodicity is enforced on the boundaries of Ω . Viscous terms are
discretised implicitly (Crank-Nicolson); the convective operator and the source term
(1) are instead treated with the explicit Adams-Bashforth scheme.
The time discretization is carried out with the second order projection scheme of
Van Kan [15]. The three resulting elliptic equations, two for the velocity components
and one for the pressure, are solved by a weak Legendre multi–domain algorithm.
Pulsating Flow through Porous Media 169
The computational domain has been discretized with 77 patched sub–domains,

each of which with 15 × 15 Legendre modes. Resolution adequacy was assessed
through a grid convergence study, carried out increasing the number of the Leg-
endre modes up to 19 × 19. The flow variables have been made dimensionless as-
suming L = h and U = h2 Δ p/ (Lμ ) as length and velocity characteristic scales,
respectively. Independently on the values of the triplet Re, Reosc , χ , all the com-
putations, after an initial transient, reached a periodic behavior with dimensionless
period T = T̃ L /U . Under the above conditions, based on the local velocity field u,
the spatially average fluid velocity (or seepage velocity) U and its time mean value
U have been evaluated respectively as:
T
1 1
U(t) = |u(t)| d Ω , U= U(t) dt (3)
Ωf Ωf T 0
where Ω f = (1 − φ ) Ω represents the volume occupied by the fluid.

Steady–state preliminary computations were carried out in order to define the
upper bound of applicability of the Darcy’s law, i.e. negligibility of the inertial ef-
fects, along with the limiting pressure gradient above which steady flow conditions
cannot take place (Re ∼ 300). Figure 1b shows the computed dimensionless steady
seepage velocity Ust versus the Reynolds number, i.e. Ust = f (Re). The indepen-
dence of Ust on the Reynolds number confirms the validity of the Darcy’s law up to
Re ∼ 1 [1, 4, 12]. Based on the steady results, the quasi–steady approximation U qst
of the period–averaged seepage velocity has been computed as:
T
1
U qst = f (Re + Reosccos (ω t)) dt (4)
T 0
a) b)
0.04
Ust
l 0.03
2h
0.02
Δp
0.01
10-1 100 101 102

Re
Fig. 1 Sketch of the geometry of the porous medium (a); seepage velocity Vs steady pressure
gradient (b)
a) b)
-1
1.00 10
χchi0.5
= 0.50
-3
χchi1
= 1.00
U/Ust
10 χchi1.42
= 1.42
__
χchi3.54
Φ
= 3.54
χchi5.01
-5
10 = 5.01
0.75 χchi12.94
= 12.94
-7
Rerosc==332 10
Rerosc==2.5
64
Rerosc==296 10-9
Rerosc==1.5
128
0.50 10-11 0 1 2
1 10 10 10 10
χ f/fT
Fig. 2 Mean seepage velocity as a function of the relative Stokes layer (Figure 2a): dashed
lines denote the quasi-steady approximation (U = U qst ). Figure 2b depicts seepage velocity
power spectra made dimensionless with U ; f T = 1/T denotes the fundamental frequency of
the oscillation
a) b)
0.06 0.16
0.00 0.00
Fig. 3 Shaded plot of total velocity for period–averaged flow field with streamlines
superposed (Re = 32 and Reosc = 32): a) χ = 12.94, b) χ = 0.50
3 Results
The effects of the pulsating forcing onto the period-averaged seepage velocity is
enlightened in figure 2a, for the Re = 32 case. Four values of the amplitude of the
oscillating pressure gradient have been considered, namely Reosc = 32, 64, 96, 128,
so that the ratio between the steady component and the unsteady one varies between
1 and 4. Similar results have been obtained also for the Re = 64 and Re = 128 cases
(results not shown).
a) b)
0.001
0.5 0.001
0.5
c) d)
0.001
0.5 0.001
0.5
Fig. 4 Velocity vector plot of the oscillating component u(τ ) (Re = 32, Reosc = 32, χ =
12.94): a) τ = 1/8T ; b) τ = 3/8T ; c) τ = 5/8T ; d) τ = 7/8T
Independently on the Reosc value, the oscillatory forcing reduces the period-
averaged seepage velocity with respect to the steady value. The larger reductions
occur for the smaller values of χ . Moreover this effect grows with the amplitude of
the oscillation (for Reosc = 32: U/Ust = 0.92; Reosc = 128: U/Ust = 0.56). Increas-
ing the oscillation frequency at a constant Reosc value, the ratio U/Ust gradually
grows and tends to the unity in the high frequency regime (for χ > 4). Figure 2b
depicts the temporal power spectrum of U for the Re = 32 and Reosc = 32 case, for
different values of the χ parameter. In the high frequency regime the power associ-
ated to the fundamental frequency largely overwhelms all the other ones, while in
the low frequency regime a fuller spectrum has been found.
The main features of the flow field at pore level are shown in figures 3-5. Figure
3 reports the period–averaged velocities u for the Re = 32 and Reosc = 32 case, with
χ = 12.94 (figure 3a) and χ = 0.50 (figure 3b). The corresponding oscillating veloc-
u(τ ) = u(τ )− u, at the phases τ = 1/8T, 3/8T, 5/8T, 7/8T, are reported in
ity fields
figures 4 and 5, respectively. In the high frequency regime (see figure 4) the oscillat-
ing flow field is characterized by a plug flow between the grains, with high vorticity
a) b)
0.1
50 0.1
50
c) d)
0.1
50 0.1
50
Fig. 5 Velocity vector plot of the oscillating component u(τ ) (Re = 32, Reosc = 32, χ =
0.50): a) τ = 1/8T ; b) τ = 3/8T ; c) τ = 5/8T ; d) τ = 7/8T
values concentrated only at the neighborhood of the solid boundaries. Moreover, the
oscillating flow field shows strong similarities with the one computed neglecting the
non–linear terms, i.e. Stokes flow, whereas the mean field resembles the correspond-
ing steady solution of Navier-Stokes equations forced with Δ p/L. In contrast, in the
low frequency regime (χ < 1, see figures 3b and 5) the period–averaged flow field
strongly differs from the corresponding steady one. The oscillating component con-
siderably deviates from the Stokes solution, exhibiting large separation zones in the
deceleration phase.
The results of figure 2 can be therefore explained on the basis of the flow fields
at pore level. In the high frequency regime, the unsteady term dominates over the
convective ones and the flow is mainly governed by the Stokes equations. Indeed
in this regime the modifications of period–averaged seepage velocity caused by the
oscillating pressure gradient vanishes. In contrast, in the low frequency regime the
non–linear term cannot be neglected, determining a strong interaction between the
steady and the oscillating components, with a noticeable energy transfer toward the
smaller temporal scales, characterised by frequencies one order of magnitude higher

than the fundamental one.
4 Conclusion
Pulsating flows through a porous medium whose base element consists of square
cylinders have been investigated processing the two–dimensional field data obtained
by numerical solution of Navier–Stokes equations. A spectrally accurate multi–
domain solver has employed. Both high and low frequency regimes have been char-
acterized, analyzing the steady and phase locked mean velocities. Power spectra of
the seepage velocity show that the transition from the low to the high frequency
regimes occurs smoothly.
References
1. Amaral Souto, H.P., Moyne, C.: Dispersion in two-dimensional periodic porous media.
Part I. Hydrodynamics. Phys. Fluids A 9(8), 2243–2252 (1997)
2. Burcharth, H.F., Andersen, O.H.: On the one-dimensional steady and unsteady porous
flow equations. Coastal Engineering 24, 233–257 (1995)
3. Champman, A.M., Higdon, J.J.L.: Oscillatory Stokes flow in periodic porous. Phys. Flu-
ids A 4, 2099–2116 (1992)
4. Firdaouss, M., Guermond, J.-L., Le Quéré, P.: Nonlinear corrections to Darcy’s law at
low Reynolds numbers. J. Fluid Mech. 343, 331–350 (1997)
5. Graham, D.R., Higdon, J.J.L.: Oscillatory forcing of flow through porous media. Part 2.
Unsteady flow. J. Fluid Mech. 465, 237–260 (2002)
6. Hannoura, A.A., McCorquodale, J.A.: Virtual mass of coarse granular media. J. Waterw.
Port Coastal Ocean Div. ASCE 104(WW2), 191–200 (1978)
7. Manna, M., Vacca, A., Deville, M.O.: Preconditioned spectral multi-domain discretiza-
tion of the incompressible Navier–Stokes equations. J. Comput. Phys. 201, 204–223
(2004)
8. Madsen, O.S.: Wave transmission through porous structures. J. Waterw. Harbours
Coastal Eng. Div. ASCE 100(WW3), 169–188 (1974)
9. Nishimura, T.: Oscillatory flows and mass transfer within asymmetric and symmetric
channels with sinusoidal wavy walls. Heat Mass Transfer 30, 269–278 (1995)
10. Ralph, M.E.: Oscillatory flows in wavy-walled tubes. J. Fluid Mech. 168, 515–540
(1986)
11. Roberts, E.P.L., Mackley, M.R.: The development of asymmetry and period doubling for
oscillatory flow in baffled channels. J. Fluid Mech. 328, 19–48 (1996)
12. Rojas, S., Koplik, J.: Nonlinear flow in porous media. Phys. Rev. E 58(4), 4776–4782
(1998)
13. Sobey, I.J.: On flow through furrowed channels. Part 1. Calculated flow patterns. J. Fluid
Mech. 96, 1–26 (1980)
14. Stephanoff, I.J., Sobey, I.J., Bellhouse, B.J.: On flow through furrowed channels. Part 2.
Observed flow patterns. J. Fluid Mech. 96, 27–32 (1980)
15. Van Kan, J.: A second order accurate pressure correction scheme for viscous incompress-
ible flow. SIAM J. Sci. Stat. Comput. 7, 870–891 (1986)
Thermodynamic Fluctuations Behaviour during
a Sheared Turbulence/Shock Interaction
S. Jamme, M. Crespo, and P. Chassaing
Abstract. Direct Numerical Simulation is used to study the mechanisms underlying

the production of turbulent density fluctuations in a sheared turbulent flow with uni-
form mean velocity gradient and non-uniform mean density and temperature gradi-
ents. The coupling between the production mechanism of the Reynolds stresses and
the one of the fluctuating density is investigated through the budgets of several rel-
evant turbulent quantities. An interaction of this kind of turbulent flow with a shock
wave is then considered in order to elucidate the effect of the shock on the previous
mechanisms. The influence of the sign of the upstream correlation u
1 T
is finally
reported.
1 Introduction
The thermodynamic properties of an isotropic turbulent flow interacting with a
shock wave has been investigated in several previous works (see e.g. Jamme et al. [2]
for a review). The presence of the entropy mode (cf. Kovasznay [3]) in the upstream
turbulence has been shown to strongly influence the amplification mechanism across
the shock compared to the case of a pure solenoidal upstream turbulence (see
Mahesh et al. [5] and Fabre et al. [1]). In all these works, the mean upstream flow
was uniform, leading to a quite simple picture of the thermodynamic fluctuations.
In the present study, we consider a sheared turbulent flow interacting with a shock
wave, with a uniform (constant) mean velocity gradient and a non-uniform mean
density (and temperature) gradient. The turbulent flow is thus not homogeneous in
the tranverse direction of the shock wave, and more complicated processes occur for
the production of density fluctuations. The purpose of the present work is to inves-
tigate using Direct Numerical Simulation (DNS) the turbulent density production
S. Jamme · M. Crespo · P. Chassaing
Université de Toulouse, ISAE, 10 avenue Edouard Belin, 31055 Toulouse, France
e-mail: stephane.jamme@isae.fr,matthieu.crespo@isae.fr,
patrick.chassaing@isae.fr
176 S. Jamme, M. Crespo, and P. Chassaing
mechanism upstream of the shock, and to describe how this mechanism is affected
during the interaction with the shock wave.
2 Numerical Method
We solve the full three-dimensional Navier-Stokes equations in non-dimensional
conservative form using a finite difference approach. The inviscid part is resolved
using a fifth-order Weighted Essentially Non-Oscillatory scheme [7]. Viscous terms
are computed using a sixth-order accurate compact scheme [4], and a third-order
Runge Kutta algorithm [8] is used to advance in time.
Equations are solved on a cubic domain of size 2π in the three directions (cf.
figure 1) and a grid of 256x128x128 points is used. The mean flow is aligned with
x1 . Periodic conditions are specified in the x3 direction, and non-reflecting boundary
conditions of Poinsot & Lele [6] along with a sponge layer are used for the top and
bottom boundaries as well as for the outflow where the flow is subsonic. At the
beginning of the calculation, a plane shock wave at Mach number M1 is specified
in the middle of the computational domain; the flow is steady on each side of the
shock, satisfying the Rankine-Hugoniot relations.
2π
mean flow
2π
supersonic subsonic
inflow outflow
M1 > 1 M1 < 1
T1
p
1 2π
U1
x2 mean shock
+ position
turbulent data
u’1 u’2 u’3 p’ T ’ x1
x3
Fig. 1 Flow configuration
At each time step, velocity, pressure, temperature, and density fields are specified
at the inflow. These fields are superpositions of a supersonic mean flow and turbu-
lent fluctuations (denoted further by a prime) in velocity, pressure, temperature, and
density. The mean velocity at the inflow varies linearly across streamlines while the
mean pressure is uniform. The mean temperature and density vary such as the mean
Mach number is uniform:
2
U 1 (x2 )=U0 +S(x2 −x2min ), U 2 =U 3 = 0, P(x2 ) = 1/(γ Mr2 ), T (x2 ) = Mr2U 1 /M12 , (1)
where the overbar denotes the conventional Reynolds average. The shear stress mag-
nitude is controlled by the parameter S where S = ∂ U 1 /∂ x2 . Turbulent fluctuations
are then superposed onto the mean upstream flow and advected through the inflow
boundary using Taylor’s hypothesis. These turbulent data come from preliminary
Thermodynamic Fluctuations Behaviour 177
calculations of temporally evolving sheared turbulence so that the anisotropy of the

velocity field used in the inflow plane is typical of a turbulent shear flow.

A reference DNS (STSI1) is conducted with the following values of the numerical
∗ ∗ ∗ ∗
parameters: Rer = ρr μur∗Lr = 94, Mr = uc∗r = 0.1, Pr = 0.7, where (·)∗r refers to a di-
r r
mensional reference variable. The mean Mach number is fixed to M1 = 1.5, and the
turbulence parameters in the inflow plane are the following : Reλ = Rer λ rms ν = 47,

q2 q u
i u
i
2 = 1.5, Mt = = = 0.173. The mean velocity gradient equals S = 1.5,
c c
with U0 = 15. The presence of a density and temperature gradient in the mean flow
leads to non-isentropic thermodynamic fluctuations on both sides of the shock. Tem-
perature and density fluctuations (entropy mode) dominate (see figure 2a), and the
entropy fluctuations are correlated with the velocity field such that the correlation
between u
1 and T
is positive. A second DNS (STSI2) is also investigated with
the same parameters as STSI1 except that u
1 and T
correlate negatively as in a
compressible turbulent boundary layer.
(a) (b)
0.25 1
0.2
0.5
0.15
0
0.1
-0.5
0.05
0 -1
5 10 15 20 5 10 15 20
k0 x1 k0 x1
Fig. 2 STSI1 : Streamwise evolution (x2 = L2 /2) of the thermodynamic fluctuations (a):
(—–) prms /P ; (− − −) ρrms /ρ ; (− · −) Trms /T ; and of the density-velocity correlations (b):
ρ u
1 ρ u
2 ρ u
3
(—–) ; (− − −) ; (− · −) .
ρ
2 u
2
1 ρ
2 u
2
2 ρ
2 u
2
3
3.1 Analysis of the Reference Case (STSI1)

We first focus on the production mechanism of the fluctuating density before the
interaction. Investigation of the budget equation of ρ
2 (2) shows that a positive
density-velocity correlation ρ u
2 (see figure 2b) together with a negative mean
density gradient Sρ = ∂ ρ /∂ x2 produce a positive source term −ρ u
2Sρ (main
contribution of term IIb). Figure 3a shows that this term is equilibrated by the
correlation between fluctuating density and fluctuating dilatation (IV ), leading to

a nearly constant positive level of ρ
2 usptream of the shock (cf. figure 2a).

U 1 ∂ ρ
2 2
2 ∂ U 1 2 ∂ρ ∂ρ 1 ∂ ρ
2 u
1 ∂ ρ
2 u
2 2ρ
∂ u
j 1
2 ∂ u
j
=− ρ − ρ u
1 + ρ u
2 − + − ρ − ρ (2)
ρ
2 ∂ x1 ρ
2 ∂ x1 ρ
2 ∂ x1 ∂ x2 ρ
2 ∂ x 1 ∂ x 2 ρ
2 ∂ x j ρ
2 ∂xj
$ %& ' $ %& '$ %& '$ %& '$ %& '$ %& '
(I) (IIa) (IIb) (III) (IV ) (V )
(a) (b)
15
10
10
5
5
0 0
-5
-5
-10
-10
-15
5 6 7 8 9 10 11 13 13.5 14 14.5 15 15.5 16
k0 x1 k0 x1
Fig. 3 STSI1 : budget of ρ

2 upstream (a) and downstream (b) of the shock. (◦ ◦ ◦ ◦ ◦) (I);
(· · · · ·) (IIa); () (IIb); (—–) (III); (− − −) (IV ); (− · − · −) (V )
Investigation of the budget equation of ρ u

2 (3, with i = 2) shows that a positive

source term −u22 Sρ (main contribution of term IIb) is working. This production
term is dominated by the correlation between fluctuating density and the instan-
taneous pressure gradient (negative term V ), leading to a slight decrease of ρ u
2
usptream of the shock. Figure 2b also displays an upstream negative value of ρ u
1
and a slight decrease of this quantity. This behaviour can be explained through the
budget equation of ρ u
1 (3, with i = 1) which displays two negative production term
−ρ u
S and −u,

2 u S (main contributions of terms IIa and IIb respectively). The

1 2 ρ
production mechanism of the fluctuating density is thus coupled to the well-known
production mechanism of the Reynolds stresses in the usptream sheared turbulent
flow. This coupling may finally be summarized in the sketch presented in figure 4.
∂ u ∂ u
1 1 1
∂ ρ u
iU ∂U ∂U
∂ ρ
∂ ρ
∂ ui ρ
∂ P ρ
∂ τi j
δi1 −u − u
u u
= −ρ u
1 δi1 − ρ u
2 i u1 i u2 −ρ i 1 − ρ i 2 +ρ u j − +
∂ x1 ∂ x1 ∂ x2 ∂ x1 ∂ x2 ∂ x1 ∂ x2 ∂xj ρ ∂ xi ρ ∂ x j
$ %& ' $ %& '$ %& '$ %& ' $ %& ' $ %& ' $ %& '
(I) (IIa) (IIb) (III) (IV ) (V ) (V I)
(3)
In the second part of the analysis, we show how the interaction with the shock
affects the above described mechanisms. We focus on the near field behaviour of the
statistics behind the shock wave (i.e. for 13 ≤ k0 x1 ≤ 15). As one can see in figure 2,
thermodynamic fluctuations are clearly amplified inside the shock zone, but a rapid
decrease of pressure and density fluctuations is observed in the near field behind
u12 ρ u
1 ρ
2
−u,

1 u2 S
Πidj u, −u, −ρ u
2 Sρ

1 u2 1 u2 Sρ
−u22 S

u32 u22
−u22 Sρ ρ u
2
Production mechanism Production mechanism

of velocity fluctuations of density fluctuations
Fig. 4 Illustration of the coupling between the production mechanisms of turbulent velocity
and density fluctuations upstream of the shock wave
the shock wave. Previous studies [2] have shown that the acoustic field generated
behind the shock has both a decaying and a propagating contribution. The rapid
decrease of the thermodynamic fluctuations is governed by pressure-dilatation or
density-dilatation correlations [term IV of equation (2)] in the near field (see figure
3b) where an energy transfer between decaying acoustic energy and kinetic energy
takes place. All these trends are essentially linear since they can be well reproduced
using inviscid linear theory (LIA : Linear Interaction Analysis) in the isotropic situ-
ation [2]. The presence of a mean shear upstream of the shock wave does not seem
to modify the well-known linear features of the interaction. ρ u
2 and ρ u
3 are poorly
affected by the interaction, whereas ρ u
1 increases significantly and even changes
sign (see figure 2b). The budgets of ρ u
1 and ρ u
2 [cf. equation (3)] in the near field
behind the shock are presented in figure 5. They show that the correlation between
density and the instantaneous pressure gradient (V ) together with turbulent diffusion
(III) are the dominant terms just behind the shock, and that they have an opposite
effect. The correlation between fluctuating velocity and fluctuating strain rate (IV )
also displays a significant influence. The production term of ρ u
(−u,

u S , main
1 1 2 ρ
contribution of term IIb) is nearly zero downstream of the shock because the level of
u,

u has changed from an upstream value of -0.32 to a downstream value of -0.03

1 2

during the interaction (cf. figure 6a). The production of ρ u

2 (−u22 Sρ ) is slightly

higher downstream since both u22 and |Sρ | are amplified during the interaction (am-
plification factors of 1.34 and 1.85 respectively), but this term is not high enough to
overcome the correlation between density and the instantaneous pressure gradient
(V ).
(a) (b)
5 5
4 4
3 3
2 2
1 1
0 0
-1 -1
-2 -2
-3 -3
-4 -4
-5 -5
13 14 15 16 13 14 15 16
k0 x1 k0 x1
Fig. 5 STSI1 : budgets of ρ u

1 (a) and ρ u
2 (b) in the near field downstream of the shock.
(◦ ◦ ◦ ◦ ◦) (I); (· · · · ·) (IIa); (+ + + + +) (IIb); (—–)(III); (− · − · −) (IV ); (× × × × ×) (V );
() (V I)
(a) (b)
0.6 0.6
0.4 0.4
0.2 0.2
0 0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
5 10 15 20 5 10 15 20
k0 x1 k0 x1
Fig. 6 Streamwise evolution (x2 = L2 /2) of the off-diagonal Reynolds stresses. STSI1 (a);

; (− − −) u,
STSI2 (b). (—–) u,
1 2

; (− · −) u,
1 3

2 3
3.2 Influence of the Sign of the Upstream Fluctuating

Velocity-Temperature Correlation
The aim of this section is to describe how the different mechanisms previously re-
ported are affected when the upstream turbulent sheared flow displays a negative
correlation between u
1 and T
. This is done by comparing the results of run STSI2
with the reference case STSI1. This negative value of u
1 T
(which leads to ρ u
1 > 0),
is artificially imposed in the developped turbulent fields that feed the domain of the
interaction run. They thus not correspond to a natural (physical) state of the turbulent
flow given the direction of the mean gradients. This explains why ρ u
1 decreases and
ρ u
2 increases on both sides of the shock (see figure 7b). The flow evolves towards a
state where the sign of the density-velocity correlations is compatible with the mean
gradients. Figure 7a also displays lower upstream values for ρrms and Trms in the
present case (STSI2) than in the reference situation (STSI1). This behaviour can be
explained through the mechanisms described in figure 4 where ρ u
2 was shown to
play a crucial part in the production of ρ

2 . In the first portion of the computational
domain (upstream of the shock wave), ρ u
2 is rather small so that the production
mechanism of ρ
2 is low. However, downstream of the shock, the density-velocity
correlation is high enough to bring about an increase of the density (and tempera-
ture) fluctuations (see figures 7a and 7b). It can also be noticed that the turbulent
kinetic energy is more amplified across the shock wave for case STSI2 than for case
STSI1 (amplification factors of 1.58 and 1.21 respectively). The same influence of
upstream entropy fluctuations with a negative correlation between u
1 and T
has
been reported in the isotropic situation with a uniform mean upstream flow both by
DNS and LIA: see Jamme et al. [2]. Finally, u,
is less affected by the interaction

1 2
in case STSI2 than in case STSI1 (cf. figure 6).
(a) (b)
0.25 1
0.2
0.5
0.15
0
0.1
-0.5
0.05
0 -1
5 10 15 20 5 10 15 20
k0 x1 k0 x1
Fig. 7 STSI2 : Streamwise evolution (x2 = L2 /2) of the thermodynamic fluctuations (a):
(—–) prms /P ; (− − −) ρrms /ρ ; (− · −) Trms /T ; and of the density-velocity correlations (b):
ρ u
1 ρ u
2 ρ u
3
(—–) ; (− − −) ; (− · −)
ρ
2 u
2
1 ρ
2 u
2
2 ρ
2 u
2
3
References
1. Fabre, D., Jacquin, L., Sesterhenn, J.: Linear interaction of a cylindrical entropy spot with
a shock. Phys. Fluids 13, 2403–2422 (2001)
2. Jamme, S., Cazalbou, J.-B., Torrès, F., Chassaing, P.: Direct numerical simulation of the
interaction of a shock wave and various types of isotropic turbulence. Flow, Turbulence
and Combustion 68, 227–268 (2002)
3. Kovasznay, L.S.J.: Turbulence in supersonic flows. J. of the Aero. Sci. 20, 657–682 (1953)
4. Lele, S.K.: Compact finite difference schemes with spectral-like resolution. J. Comp.
Phys. 103, 16–42 (1992)
5. Mahesh, K., Lele, S.K., Moin, P.: The influence of entropy fluctuations on the interaction
of turbulence with a shock wave. J. Fluid Mech. 334, 353–379 (1997)
6. Poinsot, T.J., Lele, S.K.: Boundary conditions for direct simulations of compressible vis-
cous reacting flows. J. Comp. Phys. 101, 104–129 (1992)
7. Ponziani, D., Pirozzoli, S., Grasso, F.: Development of optimized Weighted-ENO schemes
for multiscale compressible flows. Int. J. Numer. Meth. in Fluids 42, 953–977 (2003)
8. Shu, C.W., Osher, S.: Efficient implementation of Essentially Non-Oscillatory
shock-capturing schemes. J. Comp. Phys. 77, 439–471 (1988)
LES and DES Study of Fluid-Particle Dynamics
in a Human Mouth-Throat Geometry
S.T. Jayaraju, S. Verbanck, and C. Lacor
Abstract. A CT based simplified upper human airway model was created by

preserving all critical geometrical features. The fluid flow at a normal breathing flow
rate of 30 l/min is numerically studied employing RANS, DES and LES methods.
The complex flow patterns with skewed velocity profiles and flow separations are
discussed for the LES model. The deposition efficiency and the deposition patterns
for the particle diameters 2, 4, 6, 8 and 10 μ m are presented. For particle diameters
in the respirable range, LES and DES showed considerable improvement over the
RANS model, however, for the particles above 5 μ m, RANS performs as good as
LES/DES. The frozen LES method for particle tracking consistently underestimated
the deposition of bigger particles.
1 Introduction
Inhaled medication is generally the preferred method of drug administration to the
lung for the first-line therapy of asthma and chronic obstructive pulmonary diseases.
For an efficient treatment, the inhaled aerosol particles need to bypass the complex
airway morphology and reach the alveolar zone of the respiratory tract where it
eventually gets absorbed. The complexity of extrathoracic pathway which involves
bends, sudden cross-sectional changes, branching bronchi and non-symmetry of the
geometry generally results in more than half of the inhaled medication deposit in
S.T. Jayaraju
Dept. Mechanical Engineering, Vrije Universiteit Brussel, Pleinlaan 2, 1050 Brussel,
Belgium
e-mail: santhosh.jayaraju@vub.ac.be
S. Verbanck
Respiratory Division, Academic hospital Brussel, Laarbeeklaan 101, 1090 Brussel, Belgium
C. Lacor
Dept. Mechanical Engineering, Vrije Universiteit Brussel, Pleinlaan 2, 1050 Brussel,
Belgium
184 S.T. Jayaraju, S. Verbanck, and C. Lacor
the pathway before reaching the lungs [2, 4, 3]. A lot of literature is focused on
studying, both experimentally and numerically, the behavior of aerosol particles in
the human airways with a final aim of efficient drug delivery to the lungs.
Most of the numerical work so far employs Reynolds Averaged Navier Stokes
(RANS) simulations, mainly using two equation turbulence models like k − ε and
k − ω for the fluid phase coupled with Eddy Interaction Model (EIM) for the parti-
cle phase. Stapleton et al. [9] studied the suitability of k − ε turbulence model and
concluded that the accurate particle deposition prediction requires a more accurate
turbulence model. Following this study, Matida et al. [5] investigated standard k − ε
model and standard k − ω model to see that the deposition results with k − ω model
were comparatively better, but never close to the desired experimental curve fit of
Stahlhofen et al. [8]. The reason was found to be the assumption of isotropy and
once the anisotropy effects next to the walls were taken care of, a better overall
prediction of particle deposition were obtained. However, there still remained a sig-
nificant over-prediction in deposition at smaller Stokes numbers, and the particles in
these Stokes number range are the ones generally used in inhaled medication. With
the upper airway flow being transitional, the use of models which are basically de-
veloped for turbulent flows may always result in poor prediction. Reviewing the pre-
vious works, Breuer et al. [2] and Matida et al. [4] also conclude that the Reynolds
Averaged Navier Stokes (RANS) as well as Reynolds Stress Model (RSM) does
not capture relevant features of the flow and highlighted the need to switch towards
Large Eddy Simulation (LES). In [2] LES results for a 90 degree bend are presented.
Matida et al. [4] consider an idealized mouth-throat region without trachea or bifur-
cations. The geometry used in [3] is comparable to the one considered here but the
experimental results, used for comparison, are on a slightly different geometry. In
the present paper, simulations and experimental results are obtained on identical
geometries. In addition, apart from LES, also Detached Eddy Simulation (DES) is
considered.
To have a complete overview of available numerical methods, all the modeling
methods, namely the Reynolds Averaged Navier Stokes (RANS), Detached Eddy
Simulation (DES) and Large Eddy Simulation (LES) have been tested in the present
work. In case of RANS, the most widely used SST k − ω turbulence model is em-
ployed. DES is based on Spalart-Allmaras model for the near-wall region. In case
of LES, two subgrid scale models, namely the Smagorinsky-Lilly and the WALE
model are tested. The particle phase is computed in truly unsteady mode, both for
DES and LES. The frozen-flux method proposed by Matida et al. [4] is also tested.
All the results are compared with the experimental data to analyze their reliability.
2 Quality Control
RANS : The suitability of SST k-ω model for simulating transitional flows in
upper human airways has been reported in detail by Matida et al. [5]. Approx-
imately 800,000 cells are used to grid the domain. Increasing the mesh size did
not have any influence while comparing the cross-sectional velocities in the larynx
LES and DES Study of Fluid-Particle Dynamics 185
region. Around 15,000 particles for each particle diameter are injected at the
inlet face assuming uniform surface area distribution and initial velocity set equal to
that of fluid velocity. Increasing the number of particles didnot have any consider-
able influence on the total airway wall deposition. Lagrangian particles with density
ρ p =912 kg/m3 and diameters 2, 4, 6, 8 and 10 μ m are considered. At the considered
flow rate of 30 l/min this corresponds to Stokes numbers in the range of 0.1210−2
to 2.410−2.
DES and LES : The converged steady state solution based on SST k-ω model is
perturbed by adding random fluctuations and is used as an initial solution for faster
convergence of LES/DES computations. To get rid of any possible initial condition
effects, 3 through flow cycles were performed before starting the time-averaging.
The time-step to advance the flow is chosen such that the CFL number in the entire
domain is less than 1. The typical time step was 1 × 10−5 sec on 1.9 × 106 mesh
points. To be certain regarding the mesh independence of the solution, an additional
simulation was performed with the same time step but on 2.9 × 106 cells and essen-
tially no difference was found in the average velocity magnitude. Obtaining a time
independent averaged solution took approximately 3-4 through flow cycles. For the
DES simulation, 1.2 × 106 grid points were used and looking at the very good agree-
ment of DES results with that of LES, we infer that 1.2 × 106 cells would suffice. For
LES, each through flow cycle took approximately 160 hours while running parallel
on four AMD Opteron 2.4 MHz dual-core processors.
Once the time-independent average solution is obtained, Lagrangian particles are
introduced in to the domain with their corresponding velocity set equal to the fluid
velocity at the inlet. A total of 25,000 particles for five different particle diameters
(2, 4, 6, 8, 10 μ m) are uniformly injected over a time span of one through flow
cycle. Jin et al. [3] used LES to study the particle deposition in human respiratory
tract including the triple bifurcation and report using a total of 10,000 particles to
track three different particle diameters (1, 5, 6.5 μ m). The breathing intensity of 30,
60 and 90 l/min were considered. Matida et al. [4] also report the adequacy of using
10,000 particles for three particle diameters (2.5, 3.7 and 5 μ m) with their frozen
LES study in an idealized mouth at a flow rate of 32.2 l/min. Considering these
previous investigations, tracking a total of 25,000 particles to predict the deposition
of five different particle diameters at 30 l/min is presumed to be sufficient. 4 to 5
through flow cycles were required to get all the particles to either deposit on the
wall or reach the outlet. Once the unsteady tracking is finished, the flow solution is
used for the frozen LES simulations where the particles are injected in the frozen
instantaneous velocity field and tracked as in RANS, but without the EIM.
As the work of Armenio, [1] indicates that, for well resolved LES, the SGS effects
on the particle motion may be neglected, this has not been considered in the present
study. More recently some models have been developed in the literature to include
these effects, e.g. Shotorban and Mashayek, [7] and Salvetti et al., [6]. The inclusion
of such model will be the subject of future work.
Fig. 1 Comparison of normalized 2 component velocity magnitude. (a) Five millime-

ters above epiglottis; (b,c,d) One, two and three tracheal diameters downstream of larynx
respectively
Fig. 2 Comparison of normalized 2 component kinetic energy. (a,b,c) One, two and three
tracheal diameters downstream of larynx respectively

3.1 Fluid Phase
The quality of LES, DES and RANS are assessed by comparing the two-component
normalized velocity magnitude and kinetic-energy profiles with that of experiments
at different cross-sections (Fig. 1, 2). In LES, both Smagorinsky and WALE model
Fig. 3 Left: Time averaged mid-plane velocity magnitude and corresponding streamlines;
Right: Time averaged velocity magnitude (above 50% of maximum velocity in the airway
model) and corresponding secondary velocity vector lines at six different cross-sections
perform equally good in all four cross-sections, except for a slightly better prediction
by Smagorinsky model at the pharynx region on the posterior side (Fig. 1(a)). It is
very interesting to note that DES performs as good as LES in all four cross-sections.
The best available RANS model, namely the SST k − ω model does a very poor job
when compared to LES/DES at all four cross-sections. Fig. 2 shows the normalized-
kinetic-energy comparison where k − ω model clearly over-estimates the turbulence
levels in the tracheal region where the laryngeal jet plays an important role in the
amplification of kinetic energy. This over-estimation of turbulence levels may also
result in unrealistic tracheal deposition of inhaled aerosol particles. It is clear that
k − ω model performs very poor compared to LES/DES.
The time-averaged flow patterns and the secondary flow structures are presented
in Fig. 3 for the best performing LES Smagorinsky model. The six cross-sectional
views show velocity magnitudes that are above 50% of the maximum velocity in the
airway model (> 2.6 m/s). The flow entering through the mouth piece, impinges on
the tongue and takes a bend upwards. As it continues to move forward, it accelerates
in the middle part of the mouth due to reduction in cross-sectional area. As can be
seen in section A1-A2, velocity in most of the cross-section is above 2.6 m/s. Also,
two distinct recirculation zones are seen which may play a part in the regional mouth
deposition of aerosols. Towards the end of the mouth, the flow takes a downward
turn and enters the pharynx in the form of a jet which undergoes an expansion due
to increase in cross-sectional area. Consequently, the velocity is reduced and com-
plex secondary motions are set as shown in slice B1-B2. The discrepancies between
experiments and the CFD predictions is highest in the pharynx region (Fig. 1(a))
due to this complex behavior of the flow. Soon after the epiglottis region, the flow
again accelerates due to reduction in cross-sectional area and we can clearly see
high velocity zone on the anterior side of section C1-C2. At the end of the pharynx,
a step on the anterior side guides the flow towards the posterior side of the trachea in
Fig. 4 Left: Simulated total deposition as a function of Stokes and Reynolds number. The
solid line represents the experimental best fit curve. In case of RANS, ’+’ represents turbulent
tracking i.e. considering EIM; ’×’ represents mean flow tracking i.e. without EIM. Right:
Simulated deposition values in three model subparts
the form of a laryngeal jet. As a result of this laryngeal jet, two distinct recirculation
zones originate at the posterior side at section E1-E2 and move towards the center as
the flow moves further downstream (section F1-F2). These secondary motions may
have a pronounced effect on the tracheal deposition of aerosols.
3.2 Particle Phase

In Fig. 4, for the 2 and 4 μ m particles, LES and DES show particle depositions
that are much closer to the experimental curve than those obtained with RANS
k − ω (without near-wall anisotropic correction), while for the 8 and 10 μ m par-
ticles, RANS, LES and DES perform equally well. Alternatively, RANS k − ω with
mean flow tracking, i.e. without EIM, consistently underestimates deposition for all
particle diameters greater than 2 μ m. The same is true for the frozen LES method,
possibly because the dispersive effects on the particles are under-estimated.
In order to better understand the discrepancy in particle deposition between LES
(Smagorinsky) and RANS (k − ω ), deposition in the three model sub-parts are
shown for the five particle diameters (Fig. 4). It is seen that RANS overestimates
larynx/trachea deposition, showing relatively greater discrepancy with LES for the
smaller particles. This was seen to be due to relative over-prediction of velocity as
well as kinetic energy in the larynx-trachea region by RANS model.
Considering that 5 μ m is generally referred to as the upper limit of the respirable
range for inhalation drugs (represented by the dash-dotted line in Fig. 4), our find-
ings suggest that in the mouth-throat geometry, the prediction of medication aerosol
deposition inhaled at normal flow rates were more accurate for LES and DES than
for the RANS k − ω model considered here. For the description of particle transport
with diameters above 5 μ m (e.g., in the upper range of air pollutant particle dis-
tributions) or for small diameters but inhaled at greater inspiratory flows (e.g., dry
powder inhalers), RANS with its vastly lower computational requirements suffices
to adequately predict aerosol deposition.
4 Conclusions
1. RANS, LES as well as DES simulations were carried out by placing high atten-
tion on the quality control of all three methods. For the fluid phase, both LES and
DES perform superior to RANS. The kinetic energy in case of RANS is clearly
over-estimated in all three sections of the trachea, which might lead to unrealistic
depositions in this region.
2. For particles below 5 μ m, which is generally referred to as the upper limit of
respirable range for inhalation drugs, RANS completely over-estimates the de-
position and hence LES or DES is inevitable for inhalation drug study.
3. RANS is as good as LES/DES and hence the ideal method for simulating parti-
cles above 5 μ m, for example, the study of inhaled air pollutants.
Acknowledgments
Part of this research was funded by the VUB Research Council in the framework of a con-
certed research action (GOA) and this funding is gratefully acknowledged.
References
[1] Armenio, V., Piomelli, U., Fiorotto, V.: Effect of the subgrid scales on particle motion.
Physics of Fluids 11, 3030–3042 (1999)
[2] Breuer, M., Baytekin, H.T., Matida, E.A.: Prediction of aerosol deposition in 90 degree
bends using les and an efficient lagrangian tracking method. Journal of Aerosol Sci-
ence 37, 1407–1428 (2006)
[3] Jin, H.H., Fan, J.R., Zeng, M.J., Cen, K.F.: Large eddy simulation of inhaled particle
deposition within the human upper respiratory tract. Journal of Aerosol Science 38,
257–268 (2007)
[4] Matida, E.A., Finlay, W.H., Breuer, M., Lange, C.F.: Improving prediction of aerosol
deposition in an idealized mouth using large eddy simulation. Journal of Aerosol
Medicine 19, 290–300 (2006)
[5] Matida, E.A., Finlay, W.H., Lange, C.F., Grgic, B.: Improved numerical simulation of
aerosol deposition in an idealized mouth-throat. Journal of Aerosol Science 35, 1–19
(2004)
[6] Salvetti, M.V., Marchioli, C., Soldati, A.: Lagrangian tracking of particles in large-eddy
simulation with fractal interpolation. In: Conf. Proc. TI 2006, Porquerolles (2006)
[7] Shotorban, B., Mashaye, F.: Modeling subgrid-scale effects on particles by approximate
deconvolution. Physics of Fluids 17, 081701 (2005)
[8] Stahlhofen, W., Rudolf, G., James, A.C.: Intercomparison of experimental regional
aerosol deposition data. Journal of Aerosol Medicine 2, 285–308 (1989)
[9] Stapleton, K.W., Guentsch, E., Hoskinson, M.K., Finlay, W.H.: On the suitability of k − ε
turbulence modeling for aerosol deposition in the mouth and throat: A comparison with
experiment. Journal of Aerosol Science 31, 739–749 (2000)
Viscous Drag Reduction with
Surface-Embedded Grooves
Jovan Jovanović, Bettina Frohnapfel, and Antonio Delgado
Abstract. Turbulent drag reduction in wall-bounded flows is investigated

experimentally by considering the dynamic effects provoked by large variation of
anisotropy in the velocity fluctuations. Deductions based on the analysis of near-
wall turbulence lead to the design of the grooved surface topology, for which it
is demonstrated experimentally that it can produce a maximum drag reduction of
DR 25%. The drag reduction effect persisted in a narrow range of flow velocities
and for the reported experimental conditions corresponds to groove dimensions of
about 0.8 viscous length-scale.
1 Introduction
A significant portion of the drag which counteracts the motion of a body through
a fluid is generated in the thin viscous region close to the solid boundary where
the flow is nearly always turbulent. The viscous contribution to the total drag
amounts to about 50% on commercial aircraft, 90% on underwater vehicles and
almost 100% for pipe and channel flows [5]. If we consider current trends in the
crude oil prices, worldwide expansion of public transport and initiatives for reduced
Jovan Jovanović
Institute of Fluid Mechanics, Friedrich-Alexander University Erlangen-Nürnberg,
Cauerstr. 4, 91058 Erlangen, Germany
e-mail: jovan.jovanovic@lstm.uni-erlangen.de
Bettina Frohnapfel
Center of Smart Interfaces, Darmstadt University of Technology, Petersenstraße 32,
64287 Darmstadt, Germany
e-mail: frohnapfel@csi.tu-darmstadt.de
Antonio Delgado
Institute of Fluid Mechanics, Friedrich-Alexander University Erlangen-Nürnberg,
Cauerstr. 4, 91058 Erlangen, Germany
e-mail: antonio.delgado@lstm.uni-erlangen.de
192 J. Jovanović, B. Frohnapfel, and A. Delgado
II a
one-component turbulence
two-component
turbulence
wall
0.4
isotropic
two-component
turbulence
axisymmetric turbulence
channel centerline
-0.1 0 0.1 III a 0.3

isotropic
turbulence
Fig. 1 Anisotropy-invariant map according to Lumley [6] with the trajectory of a turbulent
channel flow (dotted line). The arrow ”wall” marks the position of the turbulence state at
the wall. A movement of this point towards the upper right-hand corner (one-component
turbulence) along the two-component line marks increasing anisotropy of turbulence as it is
commonly found for a variety of different drag reduced flows
pollution of the environment, the beneficial outcomes of reducing viscous drag and
therefore energy consumption are obvious. The reduction in skin-friction drag can
be achieved by the addition of long chain polymers, surfactants, micro-bubbles and
small particles to the flow or by modifications of the solid boundary by V-shape or
blade like structures which are directly inserted in the wall surface [1].
2 The Mechanism of Turbulent Drag Reduction

The origin of turbulence in wall-bounded flows can be analyzed by looking into
the evolution of anisotropy in the turbulent stresses ui u j , which can be quan-
tified using the anisotropy tensor [6] defined as ai j = ui u j /q2 − 1/3δi j (where
q2 = us us ) and its scalar invariants IIa = ai j a ji and IIIa = ai j a jk aki . A plot of IIa ver-
sus IIIa for axisymmetric turbulence, IIa = 3/2(4/3|IIIa|)2/3 , and two-component
turbulence, IIa = 2/9 + 2IIIa, defines the anisotropy-invariant map shown in figure
1 which bounds all physically realizable turbulence [6]. The two curves in this figure
represent axisymmetric turbulence. Along the straight line resides two-component
Viscous Drag Reduction with Surface-Embedded Grooves 193
turbulence. The limiting states of turbulence are located at the corner points on the
right- and left-hand sides of the anisotropy-invariant map and correspond to one-
component turbulence and isotropic two-component turbulence, respectively.
In [4] and [3] it was shown that turbulent drag reduction is associated with an
increase of the anisotropy of turbulence in the near-wall region for a number of dif-
ferent flow control techniques. In this region, where the continuity equation and a
no-slip boundary condition at the wall dictate turbulence to be two-componental,
the turbulence anisotropy is directly proportional to the ratio between streamwise
(u21 ) and spanwise (u23 ) stress components: with the streamwise component gen-
erally larger than the spanwise one. The increase of turbulence anisotropy, which
can be captured in the anisotropy-invariant map as shown in figure 1, is most nat-
urally achieved with a suppression of the spanwise stress components which cor-
responds to forcing turbulence to satisfy axisymmetry simultaneously at large and
small scales very close to the wall with invariance under rotation about the axis
aligned with the mean flow [4]. For this special situation theoretical considerations
and numerical simulations show that the turbulent dissipation rate must vanish at the
wall, εwall → 0 [4]. This fundamental deduction implies that as long as turbulence
assume the one-component state at the wall the energy of turbulence (k = 1/2q2)
cannot be amplified since k grows as k → (εwall /ν )x22 /2 as the wall is approached,
x2 → 0.
cf =12/ Reb
cf =0.073 Reb-0.25
0.1
average total dissipation rate
average direct dissipation rate
cf = tw/(0.5rUb2) average turbulent dissipation rate
0.01
0.001
100 1000 Reb 10000 100000
Fig. 2 The skin friction coefficient, c f , which can be evaluated based on the average total
energy dissipation rate, Φ , and its split up into the contributions from the average direct and
turbulent dissipation rates based on DNS data at different Reynolds numbers. Withincreasing
Reynolds number the contribution of the average turbulent dissipation rate, ε = V1 V ε dV , in-
creases such that significant drag reduction can be expected when a reduction of ε is realized
In a turbulent wall-bounded flow the averaged total energy dissipation rate, Φ =

1
V V Φ dV , is given by:
⎡ I II ⎤
& '$ % & '$

%

1 ⎢⎢ ∂U i ∂U j ∂U i ∂ ui ∂ u j ∂ ui ⎥⎥ Aw τwUb
⎢ν + + ν + ⎥ dV = . (1)
V V⎣ ∂xj ∂ xi ∂ x j ∂ x j ∂ xi ∂ x j ⎦ ρV
$ %& '
Φ
It is composed of the direct dissipation (I) and the turbulent dissipation, ε (II), and
can be evaluated from the work done against the wall shears stress, τw , per unit
mass of the working fluid, ρ V , where Aw is the wetted surface area and Ub is the
bulk velocity. At high Reynolds numbers the largest contribution to Φ , and thus
to the wall shear stress as shown in figure 2, is due to the turbulent dissipation,
ε , which reaches a maximum value at the wall in uncontrolled flows. Under the
circumstances of increased turbulence anisotropy towards the one-component limit
as described above, the reduction of the turbulent dissipation at the wall, εwall , will
lead to the minimization of Φ and thus to significant drag reduction [4].
3 Experimental Investigations
3.1 Grooved Surface Topology

A surface topology with grooves embedded in the wall and aligned in the streamwise
direction as shown in figure 3 was designed. Inside a groove, the velocity fluctua-
tions are suppressed not only in the normal but also in the spanwise direction due
to the side walls and therefore it is expected that turbulence in the groove will tend
toward the one-component limit at the wall.
A grooved surface pattern was produced by milling grooves into flat metal plates
and subsequent polishing of the plates. Groove dimensions of h ≈ 150 μ m with
a separation of 2h ≈ 300 μ m were produced in metal plates of the dimensions
1.5 m × 0.3 m which served as replacement for the smooth top and bottom walls
of a two-dimensional channel flow facility shown in figure 4. Preliminary numerical
simulations showed that grooves with such dimensions and separation were most
promising [1].
3.2 Channel Test Rig

The channel flow facility is driven by a centrifugal blower. The air stream is guided
through a well-designed settling chamber to ensure uniform flow condition at the
channel inlet. On the suction side of the blower, a Venturi nozzle was located, which
allowed the determination of the volume flow rate. The channel itself is L = 3.5 m
long and B = 300 mm wide and can be adjusted to heights of H = 25, 35 and 41 mm.
wall
x2 0
wall
x2 0
grooves h
2h
Fig. 3 Inside the grooves the flow is forced towards the one-component limit. Sketched are
trajectories across the anisotropy-invariant maps (dotted lines) for each portion of the surface
At the channel inlet trips of different heights were installed to promote turbulence
development and to ensure fully developed flow conditions and reproducible results
at low Reynolds number. After the trip, a length of 0.7 m is provided for the flow
development, which corresponds to x1 /H = 28.0, 20.0 and17.0, respectively. The
following 2.8 m of the channel comprise the test section which is equipped with
14 pressure taps of 300 μ m diameter at intervals of 200 mm located on both of the
channel side walls.
temperature 0.2
hot wire T1
smooth surface grooved surface T2
trip d
3500
1
CTA 2 13 pressure scanner
3 out
4
amplifier
Höntzsch
pressure transducer
- +
integrating DV
DISA
out
E~Dp
55D31
out:
integrating DV
E,e
channel cross section
H1 = 0.025
H
H2 = 0.035
0.3
Fig. 4 Experimental set up and instrumentation used for drag reduction and turbulence
measurements
The test section is divided into two parts of equal length. In each part of the test
section, the top and bottom walls are interchangeable so that smooth and grooved
channel walls can be installed. To evaluate the drag reduction performance of the
surface grooves, the pressure drop in a channel with and without grooves needs to
be compared. To ensure identical operating conditions for both cases, an in line set-
up is chosen in which one part of the test section has smooth walls while the second
part is equipped with grooved walls. Based on pressure drop measurements in the
subsequent parts of the test section, drag reduction is evaluated according to:
(Δ p/Δ l)grooved
DR = 1 − . (2)
(Δ p/Δ l)smooth
The grooved walls can either be inserted at the beginning or the end of the channel
test section. In preliminary measurements it was established that the arrangement of
grooved top and bottom walls followed by smooth ones yields the more conservative
estimate of drag reduction. Therefore, measurements are carried out in this set-up.
Further details on the measurement facility and measurement accuracy can be found
in [2].
3.3 Experimental Results

Figure 5 shows all drag reduction data obtained. These measurements were car-
ried out in a Reynolds number range of Reb ≈ 1000-45000. Over a wide range of
+3%
-3%
Fig. 5 Drag reduction versus dimensionless groove dimension for different heights, H, of the
channel test section. The error-bar indicate estimate of measurement accuracy
Reynolds number the measured drag reduction deduced from pressure drop mea-
surements is within the limit of the measurement accuracy which is estimated to
be within the range of ±3%. However, in very narrow range of flow velocities for
three channel heights significant drag reduction of up to 25% are detected. The drag
reduction data collapse if plotted versus the dimensionless groove size h+ = huτ /ν
and show a common peak for a dimensionless groove size of h+ ≈ 0.8. For grooves
with h ≈ 150 μ m the corresponding Reynolds number where drag reduction occurs
is in the range of Reb = 2300 − 3200.
4 Conclusions
The performance of the grooved surface with respect to drag reduction was tested
in an experimental facility in which the pressure drop over a channel test section
was measured. The first part of the test section was equipped with a grooved surface
and the second part had smooth channel walls. In experiments performed at low
Reynolds numbers, drag reduction up to 25% was measured, which is significantly
higher than any values reported for flow control with surface modifications.
Acknowlegdement
This work was sponsored by grant Jo 240/5-3 from the Deutsche Forschungsgemeinschaft.
The authors gratefully acknowledge this support.
References
1. Frohnapfel, B.: Flow control of near-wall turbulence. PhD thesis, University Erlangen-
Nuremberg. Shaker Verlag, Aachen (2007)
2. Frohnapfel, B., Jovanović, J., Delgado, A.: Experimental investigation of turbulent drag
reduction by surface embedded grooves. J. Fluid Mech. 590, 107–116 (2007)
3. Frohnapfel, B., Lammers, P., Jovanović, J., Durst, F.: Interpretation of the mechanism
associated with turbulent drag reduction in terms of anisotropy invariants. J. Fluid
Mech. 577, 457–466 (2007)
4. Jovanović, J., Hillerbrand, R.: On the peculiar property of the velocity fluctuations in
wall-bounded flows. Thermal Science 9, 3–12 (2005)
5. Lee, K.H., Cortelezzi, L., Kim, J., Speyer, J.: Application of reduced-order controller to
turbulent flows for drag reduction. Phys. Fluids 13, 1321 (2001)
6. Lumley, J.L.: Computational modelling of turbulent flows. Adv. Appl. Mech. 18, 123–176
(1978)
Study on the Resolution Requirements for DNS
in Turbulent Rayleigh-Bénard Convection
M. Kaczorowski and C. Wagner
1 Introduction
In fundamentel research the geometrically simple Rayleigh-Bénard experiment is
often chosen to investigate the turbulent heat exchange between a thermally driven
fluid and a hot bottom and a cold top wall, respectively. The characteristic parame-
ter of natural convection is the Rayleigh number Ra = α̂ ĝĤ 3 ΔT̂ /(ν̂ κ̂ ), which is a
non-dimensional measure of the buoyancy and diffusive forces acting on the fluid.
Here α̂ , ν̂ and κ̂ denote the thermal expansion coefficient, kinematic viscosity and
thermal diffusivity, respectively. ΔT̂ is the vertical temperature gradient between the
two horizontal plates, Ĥ is their vertical distance and ĝ the gravitational accelera-
tion. Values noted with a hat (.̂) are dimensional quantities, whereas values without
are non-dimensionalised.
For moderate to high Rayleigh numbers the required resolution for the grid
spacing h of a volume element in the core can be approximated using the estimate
hVi
≤ 1, (1)
πηk
with the Kolmogorov length scale
√
π Pr
ηk = , (2)
Γ ((Nu − 1)Ra)1/4
M. Kaczorowski
German Aerospace Centre (DLR) – Institute of Aerodynamics and Flow Technology,
Bunsenstr. 10, D-37073 Göttingen, Germany
e-mail: matthias.kaczorowski@dlr.de
C. Wagner
German Aerospace Centre (DLR) – Institute of Aerodynamics and Flow Technology,
Bunsenstr. 10, D-37073 Göttingen, Germany
e-mail: claus.wagner@dlr.de
M. Deville, T.-H. Lê, and P. Sagaut (Eds.): Turbulence and Interactions, NNFM 110, pp. 199–205.
200 M. Kaczorowski and C. Wagner
where Nu = q/qcond is the Nusselt number, where q is the heat flux and qcond the
heat flux due to conduction. Γ = Ŵ /Ĥ is the aspect ratio of the geometry, where W
is the width of the container.
Based on the analysis of higher order correlations Kerr [6] found that the ratio of
Kolmogorov scales and grid spacing should be within a factor of 2. At sufficiently
high Rayleigh numbers the flow in the core region of Rayleigh-Bénard convection
can be considered to behave like an isotropic passive scalar field, since the effects
of buoyancy in this region become negligible (see e.g. Kaczorowski & Wagner [5]).
The resolution requirements of the boundary layers are, however, more difficult to
estimate, since the flow is strongly non-isotropic in this region. The strong temper-
ature and velocity gradients associated with the thermal plumes play an important
role and need to be resolved.
Grötzbach [4] studied the resolution requirements for direct numerical simula-
tions (DNS) of turbulent Rayleigh-Bénard convection (RBC) in a fluid layer with
periodic boundary conditions in both horizontal directions for Ra up to 3.81 × 105
(Pr = 0.7) and resolved the near wall region with 1, 3 and 5 grid points. He con-
cluded that the thermal boundary layer should be resolved with at least 3 grid points
in order to get the correct Nu and rms-values of the temperature and velocity fields.
However, as the availability of computational resources has been continuously in-
creasing over the past years, turbulence reseach by means of DNS has advanced far
into the turbulent regime. At present the majority of simulations are performed for
Rayleigh numbers between 107 and 1011, see e.g. Emran & Schumacher [3], Verz-
icco & Sreenivasan [10], Shishkina & Wagner [9]. The highest Rayleigh number
achieved at present is 1014 (see Amati et al. [1]).
Verzicco & Sreenivasan [10] conducted a mesh refinement check for a simulation
at Ra = 2 × 1010 with either 6, 12 or 33 nodes within the thermal boundary layer
and evaluated the effect of the different resolution on Nu, stating that it is sufficient
to resolve the thermal boundary layer with 8 grid points. However, they did not
investigate the effect on the small scale statistics of the flow like e.g. skewness or
kurtosis of the temperature profile, in order to examine the impact on the dynamics
of the flow. In the present study we compare the influence of the resolution of the
boundary layers on global heat transport and the dynamics of the flow.
2 Numerical Procedure
The numerical simulations are performed solving the non-dimensional incompress-
ible governing equations which involves the Boussinesq approximation. The non-
dimensionalisation of the governing equations is carried out using xi = x̂i /Ŵ ,
ui = ûi /(α̂ ĝΔT̂ Ĥ)1/2 , θ = (T̂ − T̂0 )/ΔT̂ , p = p̂/(ρ̂ α̂ ĝĤ ΔT̂ ) and t = tˆ(α̂ ĝΔT̂ Ĥ)1/2 /Ŵ .
The set of equations is given by
∂ ui
=0 (3)
∂ xi
Resolution Requirements for DNS in Turbulent Rayleigh-Bénard Convection 201
∂ ui ∂ ui ∂ p Pr ∂ 2 ui
+ uj + = 3 +θ δ3i (4)
∂t ∂ x j ∂ xi (Γ Ra)1/2 ∂ x2j
∂θ ∂θ 1 ∂ 2θ
+ ui = , (5)
∂t ∂ xi (Γ RaPr) ∂ x2i
3 1/2
where ui are the velocity components in i = x, y, z direction, p and θ are the pressure
and temperature, respectively. The simulations are rectangular cell with periodic
boundary conditions in the longitudinal direction and adiabatic lateral and isother-
mal top and bottom walls. No-slip and impermeability conditions are applied to all
walls and the grid spacing is gradually reduced towards the rigid walls, so that the
computational domain is discretised using non-equidistant meshes. The volume bal-
ance procedure by Schumann et al. [7] is used for the integration over the fluid cells
and the solution is evolved in time by means of the explicit Euler-Leapfrog scheme.
Spatial derivatives and cell face velocities are approximated by piecewise integrated
fourth order accurate polynomials, where the velocity components are stored on
staggered grids which are described in detail by Shishkina & Wagner [8]. The ve-
locity pressure coupling is carried out through the projection method by Chorin [2]
which requires the solution of a Poisson equation. Here, a direct solver is employed
using a cyclic reduction algorithm. The system of equations is decoupled using a
fast Fourier transform in the homogeneous direction.
3 Results
Based on the results of our numerical simulations [5] we investigate RBC for three
different Rayleigh numbers Ra = 3.5 × 105 , 3.5 × 107 and 2.31 × 108 . For each
case (Ra) we construct grids that have the same resolution in the bulk, but different
resolutions in the boundary layers. The coarsest resolution of the thermal boundary
layer has approximately 5 grid points and the finest resolution 10 or more grid points
within the boundary layer. Details of the grids are provided in Table 1.
Table 1 Grid parameters: grid spacing in vertical (z), horizontal (x) and periodic (y) direction;
Nδθ denotes the number of grid points within the thermal boundary layer
Δ xc Δ zc Δ x|wall Δ z|wall Δy
Ra Run Nx Ny Nz Nδθ
10−3 10−3 10−3 10−3 10−3
3.5 × 105 R511 64 128 64 11 26.4 26.4 4.62 4.62 39.1
R57 64 128 54 7 26.4 25.7 4.62 9.23 39.1
R55 64 128 48 5 26.4 26.4 4.62 12.84 39.1
3.5 × 107 R78 152 512 152 8 10.3 10.3 2.34 2.34 9.77
R75 152 512 132 5 10.3 10.7 2.34 3.58 9.77
R74 152 512 120 4 10.3 10.8 2.34 4.76 9.77
2.31 × 108 R89 256 1024 210 9 8.42 8.57 0.508 1.133 4.88
R85 256 1024 176 5 8.42 8.57 0.508 2.235 4.88
1.05
δθ
Nu/NumaxN
0.95
0 Nδθ 20
Fig. 1 Relative changes of Nusselt number over number of grid points Nδθ within the thermal
boundary layer for Ra = 3.5 × 105 (), Ra = 3.5 × 107 () and Ra = 2.31 × 108 (•). The
dashed line indicates more than 1% error from the Nu obtained with the highest resolution
0 0.2
(a) (b)
θrms
θ̄
0.5 0
10−3 z 10−1 10−3 z 10−1
2 12
(c) (d)
(∂ θ /∂ z)rms
θS
−1.5 0
10−3 z 10−1 0 z 0.08
Fig. 2 Profiles averaged in time and periodic direction of (a) temperature, (b) rms values of
temperature and (c) skewness of temperature for Ra = 3.5 × 105 ; R55 -·-, R57 - -, R511 —
Figure 1 illustrates the influence of the number of grid points in the thermal
boundary layer on the mean Nusselt number averaged in time and over both hor-
izontal surfaces. It is found that the Nusselt numbers of all simulations, irrespective
of their boundary layer resolution, are within 1% accuracy of the simulation with the
highest resolution. Consequently, for our fourth order accurate solver it is sufficient
to resolve the thermal boundary layer with approximately 5 grid points in order to
get the satisfactory results for the mean Nusselt number.
Results of the statistical analysis obtained through averaging in time and in the
periodic direction are provided in Figs. 2, 3 and 4. It can be seen that the resolu-
tion of the boundary layers does not have a significant impact on the statistics of
the temperature field. The most significant difference can be observed for the rms
values of the wall normal temperature gradient. With increasing wall distance the
behaviour of (∂ θ /∂ z)rms is not captured appropriately. This is a measure of the tem-
poral behaviour of the heat transfer and should therefore be reflected in the higher
order moments of the temperature statistics. However, we find that these differences
occur only in the vicinity of the walls and do not have a measureable effect on
0 0.2
(a) (b)
θrms
θ̄
0.5 0
10−3 z 10−1 10−3 z 10−1
2 12
(c) (d)
(∂ θ /∂ z)rms
θS
−1.5 0
10−3 z 10−1 0 z 0.04
temperature and (c) skewness of temperature for Ra = 3.5 × 107 ; R74 -·-, R75 - -, R78 —
0 0.2
(a) (b)
θrms
θ̄
0.5 0
10−3 z 10−1 10−3 z 10−1
temperature and (c) skewness of temperature for Ra = 2.31 × 108 ; R85 -·-, R89 - -
0.15 0.15
(a) (b)
k
0 0
10−3 z 10−1 10−3 z 10−1
Fig. 5 Profiles averaged in time and periodic direction of turbulent kinetic energy k for (a)
Ra = 3.5 × 105 and (b) Ra = 3.5 × 107 . Representation of lines like in Figs. 2 and 3 for (a)
and (b), respectively
the temperature profiles or the Nusselt number. The differences in magnitude of

θ |rms for the lowest resolution of Ra = 3.5 × 105 appear to be spurious since the
same resolution of the thermal boundary layer at Ra = 3.5 × 107 seems to be suf-
ficient to resolve the profile of θ |rms appropriately. Therefore, the relatively large
grid spacing at the wall for R55 in contrast to R57 is possibly responsible for these
differences.
The analysis of the turbulent kinetic energy reveals that the velocity fluctuations
are significantly more sensitive to the resolution of the boundary layer. A clear de-
pendency on the resolution can be observed from Fig. 5 showing that the differences
for Ra = 3.5 × 105 are relatively small, but increase significantly when the Rayleigh
number is increased to 3.5 × 107. At this Ra it can be seen that with Nδθ = 4 grid
spacing in the near wall region is not sufficiently small to resolve the turbulent ki-
netic energy.
4 Conclusions
It is concluded from the above analysis that for the range of Rayleigh numbers
investigated the mean Nusselt number as well as the higher order moments of the
temperature field do not show a significant dependency on the resolution of the
boundary layer when increasing the number of grid points from Nδθ = 5 to Nδθ ≈
10 or more. In contrast the dynamics of the velocity field, however, reflect a clear
dependency. In order resolve the turbulent kinetic energy within the boundary layers
Nδθ 5 is required. This choice seems to be appropriate for Ra 108 as well.
Further analyses are required in order to investigate the impact of the boundary
layer resolution on the dynamics of the velocity field, and hence the dynamics of
the large scale structures.
References
[1] Amati, G., Koal, K., Massaioli, F., Sreenivasan, K.R., Verziccio, R.: Turbulent thermal
convection at high Rayleigh numbers for a Boussinesq fluid of constant Prandtl number.
Phys. Fluids 17(121701) (2005)
[2] Chorin, A.J.: Numerical solution of the Navier-Stokes equations. Mathematics of Com-
putations 22, 745–762 (1968)
[3] Emran, M.S., Schumacher, J.: Fine-scale statistics of the temperature derivatives in con-
vective turbulence. J. Fluid Mech. 611, 13–34 (2008)
[4] Grötzbach, G.: Spatial resolution requirements for direct numerical simulation of
Rayleigh-Bénard convection. J. Comp. Phys. 49, 241–264 (1983)
[5] Kaczorowski, M., Wagner, C.: Analysis of the thermal plumes in turbulent Rayleigh-
Bénard convection based on well-resolved numerical simulations. J. Fluid Mech. 618,
89–112 (2009)
[6] Kerr, R.M.: Higher-order derivative correlations and the alignment of small-scale struc-
tures in isotropic numerical turbulence. J. Fluid Mech. 153, 31–58 (1985)
[7] Schumann, U., Grötzbach, G., Kleiser, L.: Direct numerical simulations of turbulence.
In: Prediction methods for turbulent flows. VKI-lecture series 1979, vol. 2. Von Kármán
Institute for Fluid Dynamics, Brussels (1979)
[8] Shishkina, O., Wagner, C.: Boundary and interior layers in turbulent ther-
mal convection in cylindrical containers. Int. J. Sci. Comp. Math. 1(2/3/4),
360–373 (2007)
[9] Shishkina, O., Wagner, C.: Analysis of sheet-like thermal plumes in turbulent Rayleigh–
Bénard convection. J. Fluid Mech. 599, 383–404 (2008)
[10] Verzicco, R., Sreenivasan, K.R.: A comparison of turbulent thermal convection between
conditions of constant temperature and constant heat flux. J. Fluid Mech. 595, 203–219
(2008)
On the Role of Coherent Structures in a Lid
Driven Cavity Flow
Benjamin Kadoch, Emmanuel Leriche, Kai Schneider, and Marie Farge
Abstract. The coherent vortex extraction (CVE) is a technique based on the non-
linear filtering of the vorticity field projected onto an orthonormal wavelet basis.
The coherent vortices of the flow are reconstructed from few strong wavelet coeffi-
cients, while the incoherent background flow corresponds to the majority of weak
wavelet coefficients. Here CVE is applied to a lid driven cavity flow. Only 2.3% of
wavelet coefficients are necessary to capture the coherent structures and contains al-
most all the enstrophy. The incoherent flow, which is the remaining, is structureless
and noise-like. The results show that lid driven cavity flows are characterized by the
presence of pronounced coherent structures.
1 Introduction
This project aims to investigate the coherent structures appearing in the so-called
Lid Driven Cavity (LDC) flow. The LDC consists of a 3D cubic fully confined
Benjamin Kadoch
M2P2–UMR 6181 CNRS & CMI, Universités d’Aix-Marseille, 39, rue F. Joliot- Curie,
13453 Marseille Cedex 13, France
e-mail: kadoch@L3m.univ-mrs.fr
Emmanuel Leriche
Faculté des Sciences et Techniques, Université Jean-Monnet, rue du Docteur Paul Michelon
23, Saint-Étienne, F-42023, France
e-mail: emmanuel.leriche@univ-st-etienne.fr
Kai Schneider
M2P2–UMR 6181 CNRS & CMI, Universités d’Aix-Marseille, 39, rue F. Joliot- Curie,
13453 Marseille Cedex 13, France
e-mail: kschneid@cmi.univ-mrs.fr
Marie Farge
LMD–CNRS, ENS, 24 rue Lhomond, 75231 Paris Cedex 5, France
e-mail: farge@lmd.ens.fr
208 B. Kadoch et al.
domain, where the flow is generated by the motion of one of the wall’s cavity, see
the flow domain and naming conventions given in Fig. 1. The fluid enclosed in the
cavity is assumed to be incompressible, viscous, Newtonian and homogeneous. The
LDC is an archetype for internal recirculation flows and a benchmark for methods
of direct numerical simulations. By opposition with the two-dimensional LDC flow
which has become a standard benchmark problem for testing the accuracy of new
numerical methods, the three-dimensional lid-driven cavity flow has received much
less attention in the literature [9]. Besides its theoretical interest, the lid-driven cav-
ity is a simplification of several engineering situations such as slots on the walls of
heat exchangers or coating devices used in the paper industry for producing high
grade paper and photographic film. Even though the problem statement appears
fairly simple, the flow phenomena encountered within such systems are many and
poorly understood. From a physical point of view, several well known flow phe-
nomena appear in the LDC like shear flow, boundary layers, eddies and core vortex,
and, with the occurrence of curved streamlines, a contribution due to the associated
(Görtler) instability mechanism to the dynamics of the flow, Taylor-Görtler-like vor-
tices. Less well known aspects of this flow concern the presence of streaks near the
moving wall, of jet impingement at the walls, and of corner spiraling vortices and
the coexistence of adjacent regions of laminar and turbulent flow within the cavity
[6, 7, 2, 5].
2 Numerical Simulation and Coherent Vortex Extraction

Direct numerical simulation (DNS) –which involves the solution of the transient,
non-linear, three dimensional Navier-Stokes equations without any modeling, and
thus provides a complete description of a turbulent flow in space and time (from the
largest turbulence generating eddies, down to the smallest dissipative Kolmogorov
scales)– is performed at Reynolds numbers (based on the maximum velocity on
the lid, and the cavity width), between 1.2 104 and 2.2 104 . An analysis based on
the mean Navier-Stokes equation and the Reynolds stresses budgets is provided in
[6, 7]. The DNS use a spectral method with up to 5 millions Chebyshev collocation
points, which enable the detailed representation of all dynamically significant scales
of motion. The agreement with the available experimental data of Prasad et al.[8] is
fairly good but the comparison is limited to the mid-plane.
To study the role of coherent structures in such a flow, the coherent vortex ex-
traction (CVE) is applied at a given time. The CVE method is based on denoising
theory [3] and does not assume any pattern for the coherent structures. First, the
wavelet coefficients of the vorticity ω = ∇ × u are computed using the fast wavelet
ω λ . Thus, we
transform. A thresholding is then applied to the wavelet coefficients ω̃
split the field into two contributions, the wavelet coefficients of the coherent part
defined by -
ω λ if |ω̃
ω̃ ω λ | > ε = 4/3Z ln N
ωλC =
ω̃ (1)
0 else
On the Role of Coherent Structures in a Lid Driven Cavity Flow 209
Top lid
Side wall
q
ll
ll
wa
wa
am
am
Bottom wall
tre
re
s
st
wn
Up
Do
Side wall
-q
-h -h h
Y Z
Fig. 1 Sketch of the flow domain Ω
and those of the incoherent part as the remainder. The indices .C and .I denote the
coherent and incoherent part, respectively, Z = 1/2ω , ω is the enstrophy and N
the resolution. A first threshold is calculated from the total field then we apply the
thresholding. The next step consists in splitting the total field with a threshold cal-
culated from the incoherent field. Thus, this process is iterated with the threshold
based on the new coherent field [1]. In this study this thresholding is applied with
three iterations. The coherent ω C and incoherent ω I vorticity is reconstructed using
the fast inverse wavelet transform. The two fields thus obtained are orthogonal by
construction. As a consequence the separation of the total enstrophy is Z = ZC + ZI
and ω C + ω I = ω .
3 Results
3.1 Vorticity
In the following we use the coherent vortex extraction to filter the vorticity field
of a lid driven cavity flow corresponding to a Reynolds number Re = 1.2 104 . The
DNS data computed with a Chebyshev collocation method on a non-equidistant grid
with resolution 1293 has been reinterpolated on an equidistant grid with resolution
N = 2563. Table 1 shows that almost all the enstrophy is retained in the coherent
part, while the enstrophy of the incoherent part is negligible. The compression rate
is hereby 2.26%N, corresponding to the number of wavelet coefficients retained in
the coherent part. The compression rate is slightly increased with respect to homo-
geneous isotropic turbulence at similar resolution ∼ 3% [4]. Moreover, the skewness
and flatness of the total and coherent vorticity are similar.
Table 1 Enstrophy Z, skewness, flatness and compression rate of the total, coherent and
incoherent vorticity
Enstrophy Z %Z skewness flatness compression

rate
total: 3.59 100 32.2 2200 100%
coherent: 3.59 99.99 32.2 2200 2.26%
incoherent: 1.28 10−5 3.6 10−4 −0.0432 16 97.74%
In Fig. 2, the PDFs of total, coherent and incoherent vorticity are plotted. We
remark that the total and coherent vorticity are nearly the same, while the incoher-
ent part has a strongly reduced variance. As shown in the inset, the latter presents a
shape close to an exponential behavior.
101
total
100 coherent
incoherent
10-1
1021
10-2 100
10
10-1
10-3 10-2
10-3
10-4
10-4 10-5
-0.06 0 0.06
10-5
10-6
10-7
-50 0 50 100 150
Fig. 2 PDFs of total, coherent and incoherent vorticity. Inset: PDF of incoherent vorticity
Figure 3 shows isosurfaces of vorticity modulus for the total, coherent and in-
coherent parts. The structures are retained in the coherent part. The incoherent part
exhibits a structureless flow like a background noise, in almost all the domain except
near the walls where some small sheets structures appear.
3.2 Velocity
To reconstruct the velocity fields, which are solenoidal, from the vorticity fields,
the irrotational part has to be taken into account. Knowing that the velocity u can
be decomposed into rotational and irrotational components (Helmholtz decomposi-
tion) u = u s + u p = ∇ × ψ + ∇ φ where ψ is the potential vector and φ is the scalar
potential which is harmonic (∇2 φ = 0 by taking into account the boundary condi-
tions, see fig. 1). Inverting the curl operator only allows to reconstruct its rotational
component. To obtain the irrotational velocity we proceed as follows:
(a) |ω |
(b) |ω c | (c) |ω i |
Fig. 3 Visualization of the modulus of the total, coherent and incoherent vorticities. The
isosurfaces used are 10/9σ 2 ∼ 2 for the total and coherent part and 1/120σ 2 ∼ 0.015 for the
incoherent part, where σ 2 is the variance of the total flow
• A Poisson equation is solved for each component of the potential vector imposing
homogeneous Dirichlet boundary conditions, i.e. ∇2 ψ = −ω with ψ |∂ Ω =0 and
where ∇2 denotes the Laplace operator.
• The rotational velocity is computed using u s = ∇ × Δ −1 ω . A Fourier spectral
Poisson solver is used supposing the periodicity of the vorticity fields. Then the
curl is applied to compute the velocity from ψ .
• The irrotational velocity is deduced from the velocity: u p = u − us .
• The coherent rotational velocity is then u s coh = ∇ × Δ −1 ω coh .
• The coherent velocity is the sum of the coherent rotational velocity and the
irrotational velocity u coh = u s coh + u p .
• The incoherent velocity is equal to u inc = u s inc .
Almost all the energy of the flow is retained in the coherent contribution as
illustrated in Table 2. The energy of the incoherent part could be also neglected,
like for the incoherent enstrophy.
Table 2 Energy E, skewness, flatness and compression rate of the total, coherent and
incoherent velocity
Energy E %E skewness flatness
total: 7.73 10−3 100 2.83 58.4

coherent: 7.73 10−3 100 2.83 58.4
incoherent: 3.49 10−9 4.15 10−5 −0.18 10−2 5.58
The PDFs of the total, coherent and incoherent velocity are plotted in Fig. 4.
Likewise the vorticity, the total and coherent velocity PDFs are fairly similar, and
the PDF of incoherent part is close to an exponential and exhibits also a strong
reduced variance.
104 total
106 103 coherent
1021 incoherent
100
104 10-1
10-2
10
102 -0.0005 0 0.0005
100
10-2
10-4
10-6
-1 -0.8-0.6-0.4-0.2 0 0.2 0.4 0.6 0.8 1
Fig. 4 PDFs of total, coherent and incoherent velocity. Inset: PDF of incoherent velocity
2.36 101
2.34
100
2.32
10-1
compression rate
2.3
Enstrophy
2.28 10-2 total

coherent
2.26 10 -3 incoherent
2.24
10-4
2.22
2.2 10-5
2.18 10-6
0 200 400 600 800 1000 0 200 400 600 800 1000
time time
Fig. 5 Time evolution of compression rate (left) and time evolution of total, coherent and
incoherent energy (right)
Table 3 Mean and variance of enstrophy Z and compression rate of the total, coherent and
incoherent vorticity
Enstrophy Z %Z compression
rate
mean
total: 3.56 100 100%
coherent: 3.56 99.99 2.26%
incoherent: 1.33 10−5 3.7 10−4 97.74%
variance
total: 1.2 10−3 0
coherent: 1.2 10−3 8.96 10−4 %
incoherent: 2.68 10−12 8.96 10−4 %
3.3 Application of CVE to the Flow Evolution

To check the efficiency of CVE and the variability of the results for different time
steps, we apply CVE to a set of 106 snapshots . Fig. 5 shows the time evolution of the
compression rate and the time evolution of total, coherent and incoherent enstrophy
contributions. The mean and variance values of compression rate and total, coherent
and incoherent enstrophy are summarized in Table 3. It shows that the statistics
are well converged and it confirms that the results and conclusions obtained in the
previous section.
4 Conclusion
In conclusion, the coherent vortex extraction was applied to DNS data of a lid driven
cavity flow. A detailled study is performed at one time. The total and coherent vor-
ticities yield almost the same behavior, while the incoherent vorticity is negligi-
ble. As a consequence, the coherent structures are responsible for this kind of flow.
Moreover the CVE is slightly more efficient than in homogeneous isotropic turbu-
lence. Indeed the compression rate shows that less than 3% of wavelet coefficients
are necessary to represent the coherent structures. These results motivate the use of
Coherent Vortex Simulation (CVS) to model the time evolution of lid driven cavity
flows.
Acknowledgements. This work was supported by the ANR project ’M2TFP’, the
DFG-CNRS Research Program ’LES and CVS of Complex Flows’. We thank Michael
Schlegel from TU Berlin, for fruitful discussion on the reconstruction of the velocity field.
References
1. Azzalini, A., Farge, M., Schneider, K.: Nonlinear wavelet thresholding: A recursive
method to determine the optimal denoising threshold. Appl. Comput. Harmon. Anal. 18,
177 (2005)
2. Bouffanais, R., Deville, M.O., Leriche, E.: Large-Eddy Simulation of the flow in a Lid-
Driven Cubical Cavity. Physics of Fluids 19(5), 055108 (2007)
3. Donoho, D., Johnstone, I.: Ideal spatial adaptation via wavelet shrink-age. Biometrika 81,
425 (1994)
4. Farge, M., Pellegrino, G., Schneider, K.: Coherent vortex extraction in 3D turbulent flows
using orthogonal wavelets. Phys. Rev. Lett. 97, 054501 (2001)
5. Habisreutinger, M.A., Bouffanais, R., Leriche, E., Deville, M.O.: A Coupled Approximate
Deconvolution and Dynamic Mixed Scale Model for Large-Eddy Simulation. Journal of
Computational Physics 224(1), 241–266 (2007)
6. Leriche, E., Gavrilakis, S.: Direct Numerical Simulation of the Flow in a Lid-Driven Cu-
bical Cavity. Physics of Fluids 12(6), 1363–1376 (2000)
7. Leriche, E.: Direct Numerical Simulation in a Lid-Driven Cubical Cavity at High
Reynolds Number by a Chebyshev Spectral Method. Journal of Scientific Comput-
ing 27(1-3), 335–345 (2006)
8. Prasad, A.K., Koseff, J.R.: Reynolds number and end-wall effects on a lid-driven cavity
flow. Physics of Fluids 1(2), 208–218 (1989)
9. Shankar, P.N., Deshpande, M.D.: Fluid Mechanics in the driven cavity. Annu. Rev. Fluid
Mech. 2000 32, 93–136 (2000)
Local versus Nonlocal Processes in Turbulent
Flows, Kinematic Coupling and General
Stochastic Processes
Michael Kholmyansky, Vladimir Sabelnikov, and Arkady Tsinober
Abstract. In this paper we show that the role of kinematic relationships in the issue
of nonlocality goes far beyond their use in the nonlocal interpretation of the Kol-
mogorov 4/5 law and applicable also to general stochastic processes, unrelated to
the N-S equation. We put special emphasis on this aspect pointing to a large number
of such relations for the structure functions expressed via terms all of which have the
form of correlations between large- and small-scale quantities, and giving examples
of their experimental verification at large Reynolds numbers in field and airborne
experiments.
1 Introduction
The intent of this paper is to consider the relation of local versus nonlocal processes
in turbulent flows. The traditional view is that the former ones are dominating. How-
ever, it is noteworthy that neither the original derivation of the Kolmogorov 2/3 and
4/5 laws (for second- and third-order structure functions, respectively) nor all the
subsequent derivations of 4/5 law use the assumption on locality of interactions
and the existence of cascade (also sweeping decorrelation hypothesis). Moreover,
it has been demonstrated [3] that contrary to frequent claims on locality of interac-
tions and similar things, the 4/5 law points to an important aspect of nonlocality of
M. Kholmyansky
Tel-Aviv University, Tel-Aviv 69978, Israel
e-mail: kholm@eng.tau.ac.il
V. Sabelnikov
ONERA, Fort de Palaiseau, Chemin de la Huniére, 91761 Palaiseau, France
e-mail: vladimir.sabelnikov@onera.fr
A. Tsinober
Tel-Aviv University, Tel-Aviv 69978, Israel
e-mail: tsinober@eng.tau.ac.il
216 M. Kholmyansky, V. Sabelnikov, and A. Tsinober
turbulent flows understood as direct and bidirectional interaction of large and small
scales1 . It appears that in the nonlocality interpretation of the Kolmogorov law an
essential role is played by purely-kinematic coupling constraints.
2 Definitions and Kinematic Relations of Second and Third

Orders
Let us consider in the homogeneous isotropic turbulence two spatial points separated
by a distance r and denote the longitudinal components of the velocities at these
points by u1 = u(x) and u2 = u(x + r). Following the traditional viewpoint [6], we
keep up with the convention that the two-point velocity differences Δ u = (u2 − u1 ),
defined by small-scale motions, are local characteristics, and the velocities at one
point u1 (or u2 ), defined by large-scale motions, are nonlocal characteristics. The
interactions between small-scale motions are interpreted as local ones, and between
small- and large-scale motions – as nonlocal. The mixed moments < ul1 (Δ u)m > de-
termine the statistical connection of local and nonlocal characteristics. Angle brack-
ets denote averaging. Let us consider (as in [7]) first the mixed moments of the
second u1 Δ u and third order u21 Δ u, u1 (Δ u)2 . Substituting Δ u = (u2 − u1 ) in the
definitions of these moments, we have
< u1 Δ u >=< u1 u2 > − < u21 >= BLL (r)− < u21 > , (1)
< u21 Δ u >=< u21 u2 > − < u31 >= BLL,L (r) , (2)
< u1 (Δ u)2 >=< u1 u22 > −2 < u21 u2 > + < u31 >= −3BLL,L (r) , (3)
where BLL (r) and BLL,L (r) are longitudinal moments of second and third order,
respectively. We took into account that in isotropic turbulence < u31 >= 0, <
u21 u2 >= − < u1 u22 >. Similar equations are obtained for the moments < u2 Δ u >,
< u22 Δ u > and < u2 (Δ u)2 >.
Unless there is no turbulence, BLL (r) = 0 and BLL,L (r) = 0, and we immediately
conclude from (1)–(3) that Δ u and u1 are statistically dependent. In other terms,
there is nonlocality of interactions. What is important, this nonlocality is the conse-
quence of purely-kinematic coupling constrains. We will demonstrate now that the
local interactions are intrinsically non-separable from the nonlocal ones. Indeed,
the following relations exist between the second- and third-order structure functions
BLL (r) and BLL,L (r) [6]:
< (Δ u)2 >=< u22 > −2 < u1 u2 > + < u21 >= 2(BLL (r)− < u21 >) , (4)
< (Δ u)3 >=< u32 > −3 < u1 u22 > +3 < u21 u2 > + < u31 >= 6BLL,L (r) . (5)
Comparing the equations (1)–(3) with (4)–(5), we obtain
1 On other aspects/manifestations of nonlocality in turbulence, see [8, 9, 10].
Local versus Nonlocal Processes in Turbulent Flows 217
< (Δ u)2 >= −2 < u1 Δ u > , (6)
< (Δ u)3 >= 6 < u21 Δ u > , (7)

< (Δ u)3 >= −2 < u1 (Δ u)2 > . (8)
It follows from (6)–(8) that the structure functions of second and third orders
(relations for higher-order structure functions are considered below), in contrast to
traditional viewpoint, are determined not only by local interactions but by nonlocal
too. We conclude that, though in some limited sense one can speak about the sep-
aration of the local and nonlocal contributions, generally such a separation seems
meaningless since ”what is local is also nonlocal”. For the scales in inertial range
the third-order structure function is given by Kolmogorov 4/5 law [5]
4
< (Δ u)3 >= − < ε > r , (9)
5
where < ε > – is the mean dissipation. Combining (7), (8) with (9), we find for the
mixed moments of the third order in the inertial range
1 2
< u21 Δ u >= < (Δ u)3 >= − < ε > r , (10)
6 15
1 2
< u1 (Δ u)2 >= − < (Δ u)3 >= < ε > r . (11)
2 5
The full list of kinematic relations for homogeneous, isotropic and non-homogeneous
turbulence is given in the Appendices in [4].
3 Hosokawa Paper: Local Versus Nonlocal Contributions

Hosokawa [2], instead of velocity at one point u1 , considered as a large-scale
(nonlocal) quantity symmetrical two-point velocity half-sum u+ = (u1 + u2 )/2 =
u1 + Δ u/2. He obtained the following relation, valid in the inertial range
< u2+ u >=< ε > r/30 , (12)
where u = (u2 − u1 )/2 ≡ Δ u/2. The relation (12) is a consequence of the Kol-
mogorov 4/5 law and a purely kinematic relation which is valid under isotropy
assumption:
< u3 >= −3 < u2+ u > , (13)
which is a clear indication of the absence of statistical independence between u+
and u , i.e., between small- and large-scales motions.
We would like to emphasize the following aspect. As it was formulated above in
the Sect. 2, the separating the local and nonlocal contributions generally seems im-
possible, and it depends, of course, on the convention what characteristics are taken
as local and nonlocal. In this sense, to our opinion, it is more logical to consider
the convention, that one-point velocity is considered as nonlocal (large-scale)

characteristic. Indeed, though the version of the 4/5 law, as expressed in (12), clearly
points to the nonlocality, it contains, because u+ is a two-point quantity, both non-
local and local contributions. This can be seen looking at correlations of a differ-
ent kind, involving u1 (or equivalently u2 ), which is one-point quantity, and u ,
following the approach in [7]. Namely, we have for the square of u+
u2+ = u21 + 2u1u + u2 . (14)
Multiplying both sides of the last identity by u , we get
u2+ u = u21 u + 2u1u2 + u3 . (15)
The first two terms in (15), which are due to nonlocal interactions (in the frame
of our convention), after averaging become (using relationships (7), (8) and the
definition of u )
2
< u21 u > +2 < u1 u2 >= <ε >r. (16)
15
The third term in (15), which reflects the local interactions, is u3 , after averaging it
gives
1 1
< u3 >= < (Δ u)3 >= − < ε > r . (17)
8 10
The sum of the two leads to (12). It is noteworthy that the contributions from
nonlocal and local effects are of opposite signs, i.e., the nonlocal effects strongly
reduce the local ones. Note that the above interpretation is possible due to the use
of simple algebra before the averaging, i.e., not directly with < u2+ u >, but rather
with the terms in the relation for u2+ u .
4 Experimental Facilities and Related Matters

The experiments were preformed with a measurement system, developed by the
group of Prof. Tsinober, described in detail in the recent paper [1]. It consists of the
multi-hot-wire probe (Fig. 1) connected to the anemometer channels, signal normal-
ization device (sample-and-hold modules and anti-aliasing filters), data acquisition
and calibration unit. The probe is built of five similar arrays. Each calibrated array
allows to obtain three velocity components “at a point”. The differences between the
properly chosen arrays give the tensor of the spatial velocity derivatives (without in-
voking of Taylor hypothesis), temporal derivatives can be obtained from the differ-
ences between the sequential samples. The Taylor micro-scale Reynolds numbers,
Reλ , for the experiments are shown in Table 1. At Kfar Glikson measurement sta-
tion, Israel, the measurements were performed from a mast of 10 m height (Fig. 2a,
the corresponding data are marked “102”). At Sils-Maria, Switzerland, a lifting ma-
chine was used that allowed to reach various heights from about 1 to 10 m (Fig. 2c,
the two runs from this site are marked “SNM11” and “SNM12”). The airborne ex-
periment, Fig. 2b, was based on a Falcon research aircraft of the DLR, Germany. The
Fig. 1 The multi-hot/cold-wire probe. a - Assembled probe. b - Micro-photograph of the tip

of the probe. c - Tip of individual hot-wire array. d - Schematic of the position of the arrays
1–4 relative to the central array 0
Table 1 The Taylor micro-scale Reynolds numbers, Reλ , for the experiments
Experiment 102 SNM12 SNM11 Falcon
Reλ · 10−3 10.7 5.9 3.4 1.6
Fig. 2 Kfar Glikson measurement station, Israel, the probe on the mast (a). Airborne experi-
ment, Germany, the probe in the flight (b). Sils-Maria experiment, Switzerland, the probe on
the lifting machine (c)
airborne data are marked “Falcon”. For this experiment a special probe-mounting
device was designed, permitting to expose the probe to the atmosphere and to get it
back to the cabin during the flight, without breaking the hermeticity of the aircraft.
It was impossible to use the calibration unit onboard the aircraft. The calibration
in this experiment was performed by comparison of the properly averaged recorded
data with the synchronous data from the navigation system of the aircraft. After the
hot-wire data were transformed into the velocity components, the components of
the aircraft velocity vector were subtracted from them, thus giving the components
of the wind velocity.
5 Validation of Kinematic Relations and Their Role

Direct experimental evidence for the validity of the relation (12) at very large
Reynolds numbers obtained in field and airborne experiments [1] was presented
in [3]. The 4/5 law and its equivalent, as displayed by the relation (12), both nor-
malized on < ε > r, are shown in Fig. 3.
Fig. 3 Conventional 4/5 law (a). Verification of Eq. (12) (b)
The distance r is calculated via Taylor hypothesis. It is seen that both hold for
about 2.5 decades for the field experiments and more than for 3.5 decades in the air-
borne experiment. It is remarkable that the relation (12) holds much better than the
4/5 law, especially in the case of lower quality data, as in the airborne experiment.
The reason is due to the fact that Eq. (12) is linear in velocity increment u , whereas
the 4/5 law is cubic in u .
As mentioned, the role of kinematic relations in the issue of nonlocality goes
far beyond their use in the nonlocal interpretation of the Kolmogorov 4/5 law. The
Fig. 4 Verification of relation (18). a — for u1 , b — for u2 . The main figures are plotted in
limited range of r/η to show details; the plots in the full range of r/η are at the insets. The
results are for the run SNM11, for other runs they look similar
structure functions < (Δ u)n > are expressed via terms all of which have the form of
correlations between large- and small-scale quantities. In other words, in the absence
of nonlocal interactions – as manifested by correlations between large scale (veloc-
ity) and small scale (velocity increments) – all structure functions vanish. Hence the
utmost dynamical importance of purely kinematic relations. In Fig. 4 examples of
experimental verification of kinematic relations
(Δ u)n = −2u1(Δ u)n−1 = 2u2 (Δ u)n−1 , n ≥ 2 (18)
are shown. The asymmetric versions are chosen to emphasize the nonlocal aspects.
Acknowledgements. The experimental results, used in the present report, were obtained in
the research works supported in parts by the US–Israel Binational Science Foundation (BSF),
grant No. 2002264, the Israel Science Foundation (ISF), founded by the Israel Academy of
Sciences and Humanities, grant No. 34/02, the German–Israel Science Foundation for Scien-
tific Research and Development (GIF), grant No. I-541-132.08/97 and the Vice President for
Research of ETH, Zurich.
References
1. Gulitski, G., Kholmyansky, M., Kinzelbach, W., Lüthi, B., Tsinober, A., Yorish, S.:
Velocity and temperature derivatives in high-Reynolds-number turbulent flows in the
atmospheric surface layer. Part 1. Facilities, methods and some general results. J. Fluid
Mech. 589, 57–81 (2007)
2. Hosokawa, I.: A Paradox concerning the refined similarity hypothesis of Kolmogorov
for isotropic turbulence. Prog. Theor. Phys. 118, 169–173 (2007)
3. Kholmyansky, M., Tsinober, A.: Kolmogorov 4/5 law, nonlocality, and sweeping decor-
relation hypothesis. Phys. Fluids 20, 041704/1–4 (2008)
4. Kholmyansky, M., Sabelnikov, V., Tsinober, A.: New developments in field ex-
periments in ASL: Kolmogorov 4/5 law and nonlocality. In: 18th AMS Sym-
posium on Boundary Layers and Turbulence, Stockholm, June 9–12 (2008),
http://ams.confex.com/ams/pdfpapers/139408.pdf
5. Kolmogorov, A.N.: Dissipation of energy in locally isotropic turbulence. Dokl. Acad.
Nauk SSSR 32, 19–21 (1941); for English translation see: Selected works of A. N. Kol-
mogorov, I, Tikhomirov, V.M. (ed.), pp. 324–327. Kluwer, Dordrecht (1991)
6. Monin, A.S., Yaglom, A.M.: Statistical Fluid Mechanics. In: Lumley, J. (ed.), vol. 2.
MIT Press, Cambridge (1975)
7. Sabelnikov, V.: Obtained a number of kinematic relations between Δ u = (u2 − u1 ),
u1 = u(x) and u2 = u(x + r): Two presentations made in Laboratoire de mecanique des
fluids et d’acoustique, Ecole centrale de Lyon: 1) Large Reynolds-Number Asymptotics
of Karman–Howarth Equation, May 26; 2) Kolmogorov’s local isotropy turbulence the-
ory: state-of-the art, June 27 (1994)
8. Tsinober, A.: An Informal Introduction to Turbulence. Kluwer, Dordrecht (2001)
9. Tsinober, A.: Nonlocality in turbulence. In: Donnelly, R.J., Winen, W.F., Barenghi, C.
(eds.) Quantized Vortex Dynamics and Superfluid Turbulence, pp. 389–395. Springer,
New York (2001)
10. Tsinober, A.: Nonlocality in turbulence. In: Gyr, A., Kinzelbach, W. (eds.) Sedimenta-
tion and Sediment Transport: At the Crossroads of Physics and Engineering, pp. 11–22.
Kluwer Academic, Dordrecht (2003)
Time-Resolved 3D Simulation of an Aircraft
Wing with Deployed High-Lift System
Thilo Knacke and Frank Thiele
Abstract. The turbulent flow around a high-lift configuration consisting of slat,

main element and flap is simulated at a Reynolds number of 1.7 × 106 with an im-
plicit finite-volume based numerical method. The 3D unsteady motion in separated
flow regions is resolved on a 25 million volume mesh employing the recent Delayed
Detached-Eddy Simulation (DDES) approach [12]. Compressible calculations and
the use of non-reflecting boundary conditions enable sound radiation to be captured
in the simulation. The presented results cover the first step in a two-step approach
towards the prediction of noise emitted into the acoustic farfield and provide insight
into the complex flow dynamics in the slat region. The computed pressure distri-
butions, statistics and spectra exhibit good agreement with findings from NASA
Langley Research Center (LaRC) [2, 5].
1 Introduction
The reduction of airframe noise is an important issue for aircraft manufacturers to
reduce noise exposure caused by the continuous growth in air traffic. A good un-
derstanding of the underlying noise generation mechanisms is essential for the de-
velopment of noise reduction strategies. Insight can be obtained from numerical
simulations, which provide detailed flow and sound field information. A three-
element high-lift configuration is investigated here under landing approach condi-
tions using a finite-volume based numerical method. Because deployed slats have
Thilo Knacke
Technische Universität Berlin, Institut für Strömungsmechanik und Technische Akustik,
Müller-Breslau-Strasse 8, D–10623 Berlin, Germany
e-mail: thilo.knacke@cfd.tu-berlin.de
Frank Thiele
Technische Universität Berlin, Institut für Strömungsmechanik und Technische Akustik,
Müller-Breslau-Strasse 8, D–10623 Berlin, Germany
e-mail: frank.thiele@cfd.tu-berlin.de
224 T. Knacke and F. Thiele
been identified as one of the most prominent airframe noise contributors [4], the
main focus of this simulation is on slat noise generation.
While small scale turbulence can be modelled in the setup considered, it is
important to resolve the large scale turbulent structures in the simulation, as these
are largely responsible for the broadband sound generation. It can be anticipated
that this vortical motion will be captured correctly using the DDES approach [12].
Although the free-stream Mach number is low at landing approach, compressible
calculations are conducted to capture mean flow compressibility effects in the slat
gap and to incorporate sound radiation in the simulation.
2 Numerical Method and Setup

An implicit pressure based solver with a fully conservative approximation of the
Reynolds-averaged Navier-Stokes equations (RANS) is employed in the simula-
tion. The code, named “ELAN”, is based on curvilinear coordinates and uses
cell-centered co-located storage arrangement on semi-block-structured grids for all
quantities [14]. The 1-equation Strain-Adaptive Linear Spalart-Allmaras (SALSA)
model [11] is used both as turbulence model in RANS regions and as subgrid model
in Large-Eddy-Simulation (LES) mode. The CDES parameter was set to 0.6, as cal-
ibrated against the decay of isotropic turbulence [10]. In order to best address the
inconsistency in the demands posed by RANS and LES on the numerical scheme
for the convective fluxes, the blending function of Travin et al. [13] is employed
to assure an appropriate switch between a higher order TVD-scheme and central-
differencing. The diffusive fluxes are approximated using a second order central
scheme and for time discretisation second order backward-differencing is applied.
The continuity equation is conserved by the SIMPLE-algorithm whereby the de-
coupling of pressure and velocity is prevented through a generalised Rhie & Chow
interpolation suitable for unsteady flows [9]. The code is parallelised via domain
decomposition and the data interchange between processors is realised through the
standardised MPI-library.
The high-lift model investigated is the three-element MD 30P30N configuration.
The same configuration was previously examined in experimental and numerical
studies at NASA Langley Research Center (LaRC) [2, 5]. The testcase parameters
and geometrical settings are summarised in Tables 1 and 2. The trailing edges (TE)
of the main element and of the flap are sharp and the slat has a blunt TE with a finite
thickness of ΔTE = 6.1 × 10−3 cslat .
The structured computational mesh enclosing the high-lift configuration con-
sists of some 25 million control volumes, which were arranged in 622 blocks and
distributed for 240 CPUs, whereby an even load balance was achieved. A spanwise
cut of the 3D mesh is displayed in Fig. 4. No wall functions are applied and the
boundary layers are resolved with at least 25 points in the wall normal direction
with y+ kept below 1 on all elements. The slat cove area, being the focus region
of the simulation, contains more than 35% of all mesh points. To capture the down-
stream evolution of the slat wake a fine spatial resolution is also used above the main
Time-Resolved 3D Simulation of an Aircraft Wing 225
Table 1 Flow parameters Table 2 Geometrical settings

Rea = 1.7 × 106 Slat Flap
Ma = 0.17
u∞ = 56.0 m/s Chorda 15 % 30 %
AoA = 4.0 ◦ Gapa 2.95 % 1.27 %
Tu = 0.09 % Overlapa −2.5 % 0.25 %
cref = 0.457 m Deflection 30 ◦ 30 ◦
a Relative to stowed chord cref
element and flap. The 3D grid has a maximum of 40 cells in the spanwise
direction with a Δz spacing similar to Δx and Δy in the slat cove region, leading to a
spanwise extent of approximately 0.22 cslat . The spanwise resolution is successively
reduced for the outer regions of the computational domain and reaches a minimum
of 10 cells at the inflow and outflow boundaries. Sound waves and convective dis-
turbances leave the computational domain without major reflections owing to the
non-reflecting boundary conditions based on those of Bogey & Bailly [1]. Periodic-
ity is assumed in the spanwise direction. The vortex shedding occurring at the slat
TE is most demanding for the global timestep. A physical time step of 1 × 10−6 s
has proved sufficient in preceding 2D simulations [7] and is also used for the time-
accurate 3D simulation here. Roughly 3 orders of magnitude residual reduction
are reached within 9 outer iteration cycles per timestep. A steady RANS solution
obtained with the SALSA model was used as initial condition for the DDES.
3 DDES Results
The sampled data and the time-average cover the last 30 000 timesteps correspond-
ing to 30 × 10−3 s or 30 periods of a 1 kHz signal from an overall computed time
span of 60 × 10−3 s. Defining a slat flow through time as tslat
= c /u the eval-
slat ∞
uated data cover approximately 24 such units. Spanwise averaging was applied to
both time-averaged data and spectra.
Figure 1 shows time-averaged surface pressures for each single element, whereby
the slat and flap were rotated such that the abscissae coincide with the chord of
each element. The DDES results are compared to simulation data from LaRC [2]
which were obtained for the same configuration at a slightly larger angle of attack
(AoA = 5.5◦ ). In accordance with the 1.5◦ AoA difference, small deviations in sur-
face pressures are perceivable on all elements, however the characteristics of the
particular distributions are quite similar. The decrease in AoA causes a reduction
in slat loading but an increased loading on the flap because separation over the aft
portion of its upper surface is seen in the AoA = 5.5◦ case only.
In Fig. 2 time-averaged streamlines reveal the mean flow topology in the slat
region. The large separation zone present on the slat pressure side is bounded by
a free shear layer, measuring smax ≈ 0.8 cslat arc length from separation point to
reattachment position (consistent with [2]). Note that three locations in the shear
x [in] x [in] x [in]

0 1 2 0 2 4 6 8 10 12 14 0 1 2 3 4 5
-5
-1 -3
-4 AoA 4.0 deg. AoA 4.0 deg.

AoA 5.5 deg. AoA 5.5 deg.
-0.5 -2
-3
cp
0 -2 -1
-1
0.5 0
AoA 4.0 deg.
0
AoA 5.5 deg.
1 1 1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/cslat x/cmain x/cflap
Fig. 1 Time-averaged surface pressure distributions for slat, main element and flap. Results
for the AoA = 5.5 deg case (dashed lines) were taken from Choudhari and Khorrami [2]
|u|/u∞
0 0.25 0.5 0.75 1
0.5
rake
s/smax = 0.20 |u|/u∞
s/smax = 0.44 urms/u∞
s/smax = 0.65 0.4 vrms/u∞
slat TE wake wrms/u∞
0.3
d/cslat
0.2
0.1
0
0 0.05 0.1 0.15 0.2
urms/u∞
Fig. 2 Time-averaged streamlines with point probe and rake markers (left) and wall-normal
velocity profiles along the rake (right) from the slat region
layer, another 1.1 ΔTE downstream of the slat TE and a wall normal rake through the
center of the recirculation bubble are marked in the left hand plot. Time-averaged
mean and fluctuation velocity profiles along the rake are displayed on the right.
Peak intensities of the resolved velocity fluctuations are observed in the shear layer
at d/cslat ≈ 0.31, whereby the vertical component makes the most significant contri-
bution to the turbulent kinetic energy (TKE). Consistent with the findings of Deck
[3] on a similar configuration, the vrms levels exhibit a local minimum in the outer
part of the recirculation bubble, then reach a local maximum in the inner part before
the turbulence is finally damped due to wall proximity.
In Fig. 3 vertical velocity fluctuations from the 4 probe locations marked in
Fig. 2 are investigated. The left plot compares two-point correlations between DDES
and the hybrid RANS/DNS approach from LaRC [2], whereby the spanwise ex-
tent of the LaRC grid Lz was chosen to normalise the spanwise probe displacement
Δz in both cases. Although the DDES domain is smaller by a factor of about 0.6
the correlation decay indicates that it is sufficiently large to capture 3D turbulence
Choudhari and Khorrami, grid S2 DDES

FFT setlength 1024, Δf ≈ 97.6 Hz, Hanning window
1
s/smax = 0.20 10
-5
0.75 s/smax = 0.44

s/smax = 0.65
slat TE wake
0.5
PSD(v/u∞) [1/Hz]
R vv
-6
10
0.25
0 s/smax = 0.20
10-7 s/smax = 0.44
s/smax = 0.65
-0.25 slat TE wake
AoA 5.5 deg AoA 4.0 deg
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3

Δz/Lz 100 f cslat/u∞
101
Fig. 3 Spanwise correlation coefficients Rvv (Δz) = v

(z) v
(z + Δz) /(vrms (z) vrms (z + Δz))
(left) and spectra (right) of vertical velocity fluctuation v
= v − v, with denoting en-
semble averaging, at selected locations in the slat free shear layer and slat TE wake (left plot
contains data from [2])
Fig. 4 Cut-outs from the block-structured mesh (every 4th grid line shown)
in the slat-cove region. Comparing both sides of the plot the initial decorrelation
trends appear similar, however the DDES correlation coefficients settle to smaller
absolute values for large spanwise probe displacements. The location downstream
of the slat TE exhibits significant correlation only for very small Δz. The power-
spectral-density (PSD) plotted on the right shows increased intensities in a narrow
band around Stcslat = f cslat /u∞ = 36 for this position, corresponding to a ΔTE based
Strouhal number of StTE = 0.22. This frequency peak relates to vortex shedding be-
hind the blunt slat TE and soon vanishes further downstream. A similar effect can be
seen in the spectra obtained at the three shear layer positions. At the position closest
to the slat cusp (s/smax = 0.2) a local maximum around Stcslat = 9 is apparent in the
spectrum, which is shifted to lower frequencies and damped further downstream.
In Fig. 5 time-averaged contours of spanwise vorticity and TKE are displayed.
The latter are based on resolved u and v fluctuations in the center plot and recon-
structed from the modelled turbulent viscosity field in the right hand plot. Except
for a small part of the early shear layer and attached boundary layers, which are
Fig. 5 Time-averaged contours of spanwise vorticity (left) and turbulent kinetic energy from
resolved u and v fluctuations (center) and turbulence model (right) in the slat cove region
treated in RANS mode, the contribution of the turbulence model is small compared
to the resolved fluctuation energy, a result of the DDES operating in LES mode. Both
spanwise vorticity and TKE2D distribution exhibit remarkable similarities with PIV
measurement data and 3D simulation results from LaRC [2, 5].
In Fig. 6 instantaneous snapshots reveal the complexity of the recirculating mo-
tion in the slat region. The left hand plot also shows that the ratio between modelled
and kinematic viscosity does not exceed the value of 10 in the recirculation region.
The right hand plot displays resolved 3D turbulent structures. Spanwise rollers and
rib vortices as described by König et al. [8] are observable in the free shear layer.
Fig. 6 Snapshots from the slat cove region showing turbulent to kinematic viscosity ratio
(left) and 3D vortical structures (isosurface of λ2 -criterion [6]) coloured with local Mach
number (right)
In Fig. 7 both sound waves and turbulent flow regions become visible in the
“Schlieren” like visualisation. Three epicenters of approximately circular wave pat-
terns can be identified by close observation, which are located at the trailing edges
of the slat, main element and flap.
Fig. 7 Snapshot with “Schlieren”-like visualisation (|∇ρ

|) of turbulent flow regions and
emanating sound waves
4 Conclusion and Future Work

The turbulent flow around a three-component high-lift configuration was simulated
using a hybrid RANS/LES approach. The unsteadiness of the slat shear layer and
the slat wake was successfully resolved and the results show encouraging agreement
with data from LaRC simulations and PIV measurements [2, 5]. Sound emission
from trailing edges was evident in the simulation and the acquired unsteady flow
field data will be employed for a subsequent farfield noise prediction in future work.
Acknowledgements. The work presented in this paper was kindly supported by Airbus,
which is gratefully acknowledged by the authors. The authors would also like to express their
appreciation to Mr. Mehdi R. Khorrami from NASA Langley Research Center for provision
of the profile coordinates and settings. The simulation was performed using 80 × 103 CPUh
on the National Supercomputer HLRB II (SGI Altix 4700) at the Leibniz-Rechenzentrum
(LRZ), Munich, Germany.
References
1. Bogey, C., Bailly, C.: Three-dimensional non-reflective boundary conditions for acoustic
simulations: far-field formulation and validation test cases. Acta Acustica united with
Acustica 88, 463–471 (2002)
2. Choudhari, M.M., Khorrami, M.R.: Effect of three-dimensional shear-layer structures on
slat cove unsteadiness. AIAA Journal 45(9), 2174–2186 (2007)
3. Deck, S.: Zonal-detached eddy simulation of the flow around a high-lift configuration.
AIAA Journal 43(11), 2372–2384 (2005)
4. Dobrzynski, W., Pott-Pollenske, M.: Slat noise source studies for farfield noise predic-
tion. AIAA Paper 2001-2158 (2001)
5. Jenkins, L.N., Khorrami, M.R., Choudhari, M.M.: Characterization of unsteady flow
structures near leading-edge slat: Part I. PIV measurements. AIAA Paper 2004-2801
(2004)
6. Jeong, J., Hussain, F.: On the identification of a vortex. J. Fluid Mech. 285, 69–94 (1995)
7. Knacke, T., Thiele, F.: Slat noise reduction using steady suction. AIAA Paper 2008-17
(2008)
8. König, D., Schröder, W., Meinke, M.: Large-eddy simulation of the flow around a
high-lift airfoil configuration. In: Proc. of 5th Intern. Symp. on Turb. and Shear Flow
Phenom., Munich, vol. 2, pp. 885–890 (2007)
9. Lien, F.S., Leschziner, M.: A general non-orthogonal collocated finite volume algorithm
for turbulent flow at all speeds incorporating second-moment turbulence-transport clo-
sure, Part 1: Computational implementation. Comput. Methods Appl. Mech. Engrg. 114,
123–148 (1994)
10. Mockett, C.: A comprehensive study of detached-eddy simulation. Ph.D. thesis,
Technische Universität Berlin (2009)
11. Rung, T., Bunge, U., Schatz, M., Thiele, F.: Restatement of the Spalart-Allmaras
Eddy Viscosity Model in Strain-Adaptive Formulation. AIAA Journal 41(7), 1396–1399
(2003)
12. Spalart, P.R., Deck, S., Shur, M.L., Squires, K.D., Strelets, M.K., Travin, A.: A new
version of detached-eddy simulation, resistant to ambiguous grid densities. Theor.
Comput. Fluid Dyn. 20, 181–195 (2006)
13. Travin, A., Shur, M., Strelets, M., Spalart, P.R.: Physical and numerical upgrades in the
Detached Eddy Simulation of complex turbulent flows. In: Friederich, R., Rodi, W. (eds.)
Advances in LES of Complex Flows. Fluid Mechanics and it’s Applications, vol. 65,
pp. 239–254 (2002)
14. Xue, L.: Entwicklung eines effizienten parallelen Lösungsalgorithmus zur dreidi-
mensionalen Simulation komplexer turbulenter Strömungen. Ph.D. thesis, Technische
Universität Berlin (1998)
Fluid Mechanics and Heat Transfer in a
Channel with Spherical and Oval Dimples
Nikolai Kornev, Johann Turnow, Egon Hassel, Sergei Isaev, and

Frank-Hendrik Wurm
Abstract. Vortex mechanism of heat transfer enhancement in a narrow channel

with dimples has been investigated numerically using unsteady Reynolds averaged
Navier Stokes equations (URANS SST and SAS) and Large Eddy Simulations
(LES). The flow separation results in a formation of vortex structures which signifi-
cantly enhance the heat transfer on dimpled surfaces conducted by a small increase
of the pressure loss. The vortex structures and the flow are sufficiently unsteady. The
vortex structure inside of the dimple changes steadily its orientation causing the long
period oscillations with opposite-of-phase motion. The heat transfer enhancement
is caused mostly by the amplification of convection. The effect of the wetted area
increase is sufficiently smaller.
1 Introduction
Vortex enhancement of heat transfer caused by turbulators is currently used in
modern heat exchangers. In case of using dimples as turbulators the hydrodynamic
and thermodynamic properties of exchangers are favourable in comparison with
other heat transfer enhancement methods like ribs and fins. Especially with respect
to the pressure loss this innovative cooling method shows major advantage as com-
pared with existing conventional methods. Basically the physics of the flow inside
the dimple is complicated and it is still not completely understood. Main attention
of papers published in the literature has been paid to the heat and mass transfer
effects averaged in time whereas the unsteady processes and their role in the heat
N. Kornev · J. Turnow · E. Hassel · S. Isaev
University of Rostock, Dept. Technical Thermodynamics, Albert Einstein Str. 2,
18059 Rostock, Germany
e-mail: nikolai.kornev@uni-rostock.de
F.H. Wurm
WILO SE, Nortkirchenstrae 100, 44263 Dortmund, Germany
e-mail: Frank-Hendrik.Wurm@wilo.com
232 N. Kornev et al.
transfer enhancement have not been thoroughly investigated. The structure of

vortices created within dimples remains not quite clear especially when the flow
is highly unsteady at large depth to diameter ratio of the dimple (see [1]- [4]). Since
the form of the vortex has a strong impact on the heat transfer the deep understand-
ing of physics inside of the dimple is important for the further improvement of the
heat exchanger efficiency.
2 Numerical Model
A system of transient URANS equations is written in a conservative form for the
increments of the governing variables (velocity and pressure) in curvilinear non-
orthogonal coordinates. The system is discretised by finite-volume method using
multi-block overlapping grids. Convective fluxes are discretised by second order
upwind schemes. Pressure correction technique SIMPLEC is applied, in which ten-
tative velocity field for frozen pressure is determined followed by pressure correc-
tion to satisfy continuity equation. The LES is based on the finite-volume method
applied to filtered continuity, Navier-Stokes and temperature transport equations.
The dynamic mixed subgrid stress model (DMM) is used both for the velocity field
and the temperature. The filtered transport equations are solved on a non staggered
Cartesian grid, the discretisation in space and time of the transported and trans-
porting quantities at the cell faces is of the second order using central differential
scheme. The molecular Prandtl number Pr was set to 0.7, whereas in LES the turbu-
lent viscosity μt and the Prandtl number Prt are determined dynamically within the
DMM. The Reynolds number based on the averaged bulk velocity Ub and dimple
print diameter D was equal to ReD = 40000. The ratio of dimple depth to dimple
print diameter is t/D = 0.26 whereas the ratio of channel height to dimple diameter
H/D is 0.326. The coordinate system is shown in Fig. 1.
3 Results
3.1 Validation
Numerical results are compared with experimental LDV (Laser Doppler Velocime-
try) data published in [2] for a smooth channel with a single dimple. Fig. 1 shows
the profile of normalized mean velocity ux obtained from simulations and experi-
ments along the y-axis at x/H = 0.0 and z/H = 0.0 at the center of the dimple. The
data are normalized by the maximum velocity determined in the center of the chan-
nel flow at y = H/2, z = 0 far upstream from the dimple. Numerical results show
good agreement with experimental data. Small discrepancy with measurements can
be observed in the near wall region inside of the cavity. The most probable reason
for the discrepancy between theory and experiment in this region might be typical
LDV measurements problems in close proximity to the wall. A discrepancy between
numerics and measurements for fluctuations (Fig. 1) is observed in location of the
maximum of the u-rms value. Simulations predict the largest fluctuations exactly
Fluid Mechanics and Heat Transfer in a Channel 233
on the dividing surface whereas the measured maximum of u-rms lies inside the
dimple. However, it should be noted that the numerical results seem to be reason-
able. Indeed, the maximum pulsations occur in the area of the strongest velocity
gradients which is in accordance with well tried turbulence closure models.
Fig. 1 Mean velocity profile (left) and the normal stress (right) across the channel height at
x/H = 0.0 obtained from URANS SST(line), LES (cross) and experiments [2] (circle)
3.2 Flow Structures

The distributions of mean velocities presented in Fig. 1 shows clearly the creation
of a large recirculation zone within the dimple which occupies about 90% of the
cavity. Concerning the vortex structures appearing within the dimple there is a clear
discrepancy between URANS and LES results. The URANS based on the SST tur-
bulence model predicts a very interesting phenomenon. Depending on the depth to
diameter ratio t/D the flow experiences a bifurcation [5]. At small t/D the flow is
symmetric (Fig 2 a) and b)). At t/D ∼ 0.2 the flow experiences a topological change
and becomes asymmetric with the creation of a vortex mono structure inclined to the
mean flow direction at the angle of approximately α = ±45◦ . The choice between
α = +45◦ and α = −45◦ depends on the numerical scheme and the grid used. In
both cases the flow is sufficiently steady and the position of the vortex structure
is not changed. Application of the Scale Adaptive Simulation (SAS) brings small
corrections. The vortex structure oscillates around a middle position with a low
frequency and the flow becomes slightly unsteady.
LES results also reveal the formation of asymmetric structures but in contrast
to URANS they are sufficiently unsteady. Since instantaneous flow snapshots from
LES look very chaotically the analysis is performed for characteristics averaged
locally in time. Fig. 3 shows streamlines averaged within three sec of real time
at different time instants. Streamlines indicate the asymmetrical vortex structures
inclined to the mean flow. The coherent vortex structure is switched nearly period-
ically from the position α = +45◦ to the position α = −45◦. However, the flow
averaged over a sufficiently long time period is nearly symmetric. In this context
LES results and qualitative measurement observations contradict to URANS ones
which predict the asymmetry even in the averaged flow pictures. LES reveals the
generation of unsteady asymmetric monocore vortex structures with a predominant
transversal direction which are responsible for the appearance of organized self-
sustained oscillations. This interesting physical phenomenon of the dimple flow has
been qualitatively described in experiments [2], [3] and [4]. For instance, Terekhov
et al. [2] observed visually transversal oscillations of the flow inside the dimple us-
ing hydrogene bubbles. These oscillations have low-frequency and high-frequency
components. The presence of two dominating modes can be detected in the time his-
tory of the longitudinal velocity recorded from LES and statistically evaluated at two
spatial points with coordinates x/H = 1.0, z/H = +0.66 (point 1) and x/H = 1.0,
z/H = −0.66 (point 2) (Fig.4). A clear peak in the frequency spectrum is observed
around f = 0.48Hz for the point 1 and f = 0.23Hz for the point 2. These results
are in accordance with measurements of Terekhov et al. [2] who observed oscilla-
tions with frequencies varying from 0.05Hz to 0.2Hz. The long period oscillations
are caused by the vortex structure switching. The second dominating frequency is
observed at ∼ 10Hz.
The next physical phenomenon revealed from LES is the presence of a sort of
opposition-of-phase of the flow which can be seen from Fig 4. Once the flow is
accelerated at the point z/H = −0.66 it is decelerated or even changes the direc-
tion at the point z/H = 0.66. This effect is due to oscillation of the vortex mono
structure which is inclined both in horizontal and vertical directions. The vortex in-
let is located deeply inside of the dimple whereas the outlet is above the dimple
edge downstream. Such a vortex induces both positive and negative longitudinal
velocities depending on the point location. At the position α = +45◦ (Fig. 3a) the
point z/H = +0.66 lies above the vortex core which induces the positive veloc-
ity resulting in the flow acceleration. The point z/H = −0.66 can be located either
under the vortex axis, above or close to it. In the first case we have negative in-
duced velocity, in the second case the induced velocity is lowered. The situation is
changed when the vortex position is switched. The flow acceleration takes place at
z/H = −0.66 whereas the flow at z/H = 0.66 is decelerated. In both cases the prod-
uct of pulsations at both points is negative. As a result the autocorrelation function
Ruu (η ) = u
x (x, y, z)u
x (x, y, z + η )/u
x (x, y, z)2 calculated at these two points changes
its sign in transversal direction (Fig. 4).
3.3 Heat Transfer

A frequently asked question concerning dimpled heat exchangers is about the
physical mechanisms behind the heat transfer enhancement. Indeed, the heat ex-
change can be augmented by two effects. First, forming a spherical depression in
a plate surface results in an increase of the heat exchanger surface. Second, forma-
tion of vortex structures amplifies the convection and heat transfer. The experience
shows, that the second effect proves to be stronger.
For the case of a single dimple with t/D = 0.26 the increase of the wetted area is
about 20 per cent. The heat flux from the dimple surface is only about 10 per cent
Fig. 2 Bifurcation of vortex structure depending on the depth to diameter ratio t/D.
a) t/D = 0.1, b) t/D = 0.14 and c) t/D = 0.18
U/Ub U/Ub
0.5 0.5
0 0
(a) (b)
Fig. 3 Vortex structures obtained from LES velocity field averaged over three sec realtime at
different time instants
larger than that from the projected area on the flat plate. At a glance, the larger effect
could be attained just by an enlargement of the surface and without dimples. How-
ever, this conclusion is wrong since the heat transfer on the flat surface around the
dimple is not taken into account. Indeed, the most contribution to the heat transfer
enhancement is made by a flat surface behind the dimple. The heat flux from the
flat surface rectangle 1.5D × 1.5D behind the dimple increases also by 10 per cent
comparing with the flat plate case. Therefore, the total increase of the heat flux from
the surface consisting of the dimple and this rectangle is about 10 per cent whereas
the increase of the surface wetted area is five per cent. Therefore, the contributions
of two mechanisms are similar for the considered case. The reason of the heat trans-
fer increase behind the dimple is amplification of convection caused by vortices
0.8 1
z/H = -0.66
z/H = +0.66
0.8
0.6
0.6
0.4
0.4
ux /U0
0.2 0.2
Ruu
0
0
-0.2
-0.2
-0.4
z/H = -0.66
z/H = +0.66
-0.4 -0.6
0 1 2 3 4 5 6 7 8 -1.5 -1 -0.5 0 0.5 1 1.5
t[s] z/H
Fig. 4 Time history of longitudinal velocity ux recorded at two symmetric points at x/H =
0, y/H = 0, z/H = ±0.66 (left) and the autocorrelations Ruu at these points versus z/H (right)
generated inside of the dimple and shed downstream. Their effects is sufficiently
larger in dimpled packages. Indeed, each dimple and adjacent area are in the vortex
wake of preceding dimples. The convection becomes the dominating mechanism of
the heat exchange enhancement. As a result, the heat exchange can be up to 140
per cent larger than on the flat plate [6] whereas the increase of the wetted area in
typical dimple packages is about only 10 per cent.
Numerical analysis shows that the creation of asymmetry is a positive effect with
respect to the heat transfer enhancement. Artificial enforcement of the symmetry
in simulations results in the decrease of the heat exchange. Asymmetric vortices
are proved to be more efficient than symmetric ones. This fact gained from the
numerical analysis was used to design of a novel dimple form which consists of two
semi spherical dimples with a horizontal cylindrical insert of the length D [7]. The
axis of the oval dimple is inclined to the mean flow direction at the angle of about
60◦ . At this angle the ratio between the heat enhancement and the hydraulic loss is
maximum. The vortex structure arising in such a dimple is asymmetric and causes
a high heat flux from the exchanger surface. The heat exchange enhancement and
the increase of the ratio for the oval dimple is about six- seven per cent higher than
those for the spherical dimple with the same projected area.
4 Conclusions
The presence of oscillations with dominating frequencies in the flow over a sin-
gle dimple has been revealed using LES. At selected points inside of the dimple
these oscillations have low-frequency and high-frequency components whereas at
other points only a dominating low-frequency component can be distinguished.
Qualitatively similar effects have been observed in many measurements on dimples
(see [2], [3] and [4]). Analysis of flow structures using instantaneous streamlines
shows clearly the creation of the coherent vortex structure which is inclined to the
main flow. The vortex orientation is changed nearly periodically in time causing
long-period oscillations with the opposition-of-phase motion. These effects can be

reproduced only using LES. The flow obtained from URANS is only slightly un-
steady. The time averaged flow is asymmetric due to creation of an asymmetric
vortex monocore structure inclined to the mean flow. On the contrary, LES predicts
an unsteady flow with oscillating coherent vortex structure. The time averaged LES
flow fields are symmetric. Analysis of available experimental data shows that the
qualitative flow pattern obtained from LES is more realistic than that from URANS.
The heat transfer enhancement is caused mostly by the amplification of con-
vection due to vortices generated within the dimple whereas the effect of the wet-
ted area increase is sufficiently smaller. The most contribution to the heat transfer
enhancement is made by flat surface areas adjacent to the dimple.
Acknowledgements. The authors acknowledge gratefully the support of the German

Research Foundation (DFG) and the North German Alliance for the Advancement of
High-Performance Computing (HLRN).
References
1. Ligrani, P.M., Harrison, J.L., Mahmood, G.I., Hill, M.L.: Flow structure due to dimple
depressions on a channel surface. Physics of Fluids 13, 3442–3448 (2001)
2. Terekhov, V.I., Kalinina, S.V., Mshvidobadse, Y.M.: Heat transfer coefficient and aerody-
namic resistance on a surface with a single dimple. Enhanced Heat Transfer 4, 131–145
(1997)
3. Gromov, P.R., Zobnin, A., Rabinovich, M.I., Sushik, M.M.: Creation of solitary vortices
in a flow around shallow spherical depressions. Sov. Tech. Phys. Lett. 12, 547–548 (1986)
4. Kiknadze, G.I., Oleinikov, V.G.: Self-organization of tornado-like vortex structures in gas
and liquid flows and heat transfer intensification. Proc. Inst. Thermophysics, Preprint 227,
Novosibirsk, 1–46 (1990)
5. Isaev, S.A., Leont’ev, A.I., Baranov, P.A., Pyshnyi, I.A.: Numerical analysis of the influ-
ence of the depth of a spherical hole on a plane wall on turbulent heat transfer. J. Eng.
Physics and Thermophysics 76, 52–59 (2003)
6. Chyu, M.K., Yu, Y., Ding, H., Downs, J.P., Soechting, F.: Concavity enhanced heat transfer
in an internal cooling passage. ASME Paper, No 97-GT-437 (1997)
7. Isaev, S.A., Leontev, A.I., Mitjakov, A.V., Pyshnyi, I.A.: Intensification of tornado-like
heat transfer in asymmetric dimples. J. Engineering Physics and Thermophysics 76,
31–34 (2003)
Investigation of the Flow around a Cylinder
Plate Configuration with Respect to
Aerodynamic Noise Generation Mechanisms
Michael Kornhaas, Dörte C. Sternel, and Michael Schäfer
Abstract. In this work we present a Large Eddy Simulation of an aero-acoustic test

case consisting of a plate located in the turbulent wake of a circular cylinder. This
configuration is very attractive for the validation of low Mach number aero-acoustic
codes and coupling techniques, since a high sound pressure level is present at a very
low Mach number and also because its simple geometry. Further it seems to be an
interesting test case for future works if besides aero-acoustics also fluid induced
vibrations are of interest.
In order to identify the involved noise generation mechanisms, mean pressure
fluctuations as well as frequency spectra are investigated. The numerically obtained
frequencies are compared to experimental data.
1 Introduction
A major part of the noise production in urban environments is caused by turbulent
flows, e.g. the noise of driving cars and fans. Even though the prediction and reduc-
tion of noise is of great importance, the physical mechanisms of noise generation in
turbulent flows are still not fully understood. Computational fluid dynamics (CFD),
especially time resolved methods like the Direct Numerical Simulation (DNS) and
the Large Eddy Simulation (LES) can be a powerful tool for a better understand-
ing of flow physics and the mechanisms of aerodynamic noise generation. This is
not only because of their capability of predicting fluctuating quantities that can be
used as source terms for aero-acoustic simulations, but also because of its higher
accuracy compared to the numerically cheaper Reynolds Averaged Navier-Stokes
(RANS) simulations. However, only recently sufficient computational power has
become available to perform such computations.
Michael Kornhaas · Dörte C. Sternel · Michael Schäfer
Technische Universität Darmstadt, Chair of Numerical Methods in Mechanical Engineering
Dolivostraße 15, 64293 Darmstadt, Germany
e-mail: kornhaas@fnb.maschinenbau.tu-darmstadt.de
240 M. Kornhaas, D.C. Sternel, and M. Schäfer
We present LES results for the flow field around an aero-acoustic test case that
consists of a plate which is located in the turbulent wake of a circular cylinder. The
geometrical simplicity of the configuration in combination with an enormous in-
crease of noise at low flow speeds by adding a simple plate makes this test case very
attractive for the validation of low Mach number aero-acoustic codes and coupling
techniques. Further this test case seems to be an interesting test case for fluid struc-
ture acoustic coupling, since dependent on the plate material structural vibrations
can be obtained.
The aim of this study is to prove the capability of the applied LES solver to
capture the physical effects that account for the increase of the noise level when a
plate is present in the wake of the cylinder.
2 Test Case Description

The investigated configuration consists of a steel plate that is located in the turbulent
wake of a circular cylinder (compare figure 1). The Reynolds number based on the
cylinder diameter D and the free stream velocity is ReD = 2000. Experimental in-
vestigations of this test case are documented in [5]. Figure 2 shows the experimental
Fig. 1 Test case configuration
setup as well as the acoustic sound pressure spectrum for the test case configuration
in comparison to the one obtained for a single cylinder. The sound pressure level
increases significantly if the plate is present.
3 Numerical Setup
3.1 Numerical Method
The LES is performed using the inhouse finite volume solver FASTEST [2]. The
filtered incompressible Navier-Stokes equations are solved on a boundary fitted,
block structured hexahedral grid. Convective and diffusive fluxes are approximated
with a second-order central difference scheme. Subgrid stresses are computed us-
ing the Smagorinsky model with the dynamic approach of Germano [3]. For time
discretization the implicit second order Crank-Nicolson scheme is applied. Pressure
Investigation of the Flow around a Cylinder Plate Configuration 241
Fig. 2 Experimental setup (left) and sound pressure spectrum for the test case and a single
cylinder (right)
velocity coupling is realized with a SIMPLE type algorithm which is embedded in

a geometric multi-grid scheme with standard restriction and prolongation [1]. The
resulting linear systems of equations are solved with an ILU method.
3.2 Acoustic Source Term Modeling

Time resolved CFD simulations like LES allow the calculation of aeroacoustic
sources according to different aeroacoustic analogies e.g. Lighthill’s analogy [6]
or Hardin and Pope’s [4] acoustic/viscous splitting approach. In the latter the com-
pressible flow field at low Mach numbers is modeled as the incompressible calcu-
lated hydrodynamic part (denoted with ∗inc ) and an acoustic fluctuation (denoted
with ∗
). In the following u stands for the velocity vector, p for the pressure, ρ for
the density and κ for the adiabatic exponent. Further a hydrodynamic density cor-
rection ρ (1) is introduced. If viscous effects are neglected, a system of equations
that is equivalent to the Linearized Euler Equations (LEE) on the left hand side
and acoustic source terms - that can be computed from the incompressible flow
quantities - on the right hand side, is obtained.
∂ ρ
∂ ρ (1)
+ ρ inc ∇ · u
+ uinc · ∇ρ
= − − uinc · ∇ρ (1) (1)
∂t ∂t
∂ u

ρ inc + ρ inc uinc · ∇ u
+ ∇p
= 0 (2)
∂t
∂ p
∂ pinc
+ κ pinc∇ · u
+ uinc · ∇p
= − − uinc · ∇pinc (3)
∂t ∂t
With the hydrodynamic density correction
1 inc
ρ (1) = p − <pinc > (4)
c∞
where c∞ stands for the speed of sound and < pinc > for the mean hydrodynamic
pressure. The results of the presented LES are interpreted according to this approach
in order to identify the involved noise generation mechanisms. In doing so we focus
on the incompressible pressure fluctuations that are the dominant part in the source
term for the acoustic pressure fluctuation p
.
3.3 Simulation Setup

The computational domain is modeled periodically in spanwise directions. To obtain
a Reynolds number of ReD = 2000 the inlet velocity is set to 10m/s. Zero gradient
boundary conditions are applied at the outlet and all walls are no slip walls.
To minimize disturbances caused by boundary conditions, the size of the compu-
tational domain was chosen sufficiently big. Its size is 100 D in streamwise direction
where the inlet is located 15 D upstream the cylinder. In spanwise directions 5 D re-
spectively 50 D are modeled.
The boundary layer around the cylinder is fully resolved with corresponding y+
values of the order O(1) whereas the developing boundary layer attached to the
plate is not resolved. The resulting grid consist of approximately 2.47 × 106 control
volumes. The grid structure is displayed in figure 3.
The time step size was set to 1.6 × 10−6s which results in Courant numbers below
unity. The simulation was carried out on 8 IBM Power5 CPUs with a load balancing
efficiency of 98.5%. The computational time per time step was approximately 12
seconds. Another computation of a single cylinder at the same Reynolds number
and a comparable grid was carried out in order to compare the pressure fluctuations
to those obtained for the test case.
4 Results
4.1 Flow Features
To give an impression of the turbulent structures in the vicinity of the test case the
vorticity magnitude is shown in figure 3. It is obvious that the von Karman vortex
street can not develop if the plate is present. The interaction of the vortex shed-
ding at the cylinder with the front edge of the plate leads to high time dependent
pressure fluctuations and therefore high aeroacoustic sources (compare section 3.2).
Figure 4 shows the pressure distribution for two time steps (time between snapshots:
8 · 10−4 s). The mean pressure fluctuations < p
inc > for the test case and those ob-
tained for a single cylinder at the same Reynolds number are shown in figure 5. The
absolute values of the pressure fluctuations are approximately 5 times larger for the
test case compared to those of the single cylinder what explains the significantly
higher level of emitted noise observed in the experiments (compare figure 2).
Fig. 3 Numerical grid and vortical structures in the vicinity of the test case - snapshot of the
vorticity magnitude colored by pressure
Fig. 4 Pressure distribution for two different time steps (time between snapshots: 8 · 10−4 s)
Fig. 5 Mean pressure fluctuations for the test case (left) and a single cylinder (right)
4.2 Frequency Spectra

To get information on which frequencies the sound is emitted and to identify the
corresponding flow features, respectively to prove if the numerical method is able to
capture these frequencies, time series were recorded at different monitoring points
(MP). The locations of the monitoring points are depicted in figure 6. Since the high-
est pressure fluctuations occur around the front edge of the plate (see figure 5) and
therefore the dominant acoustic sources can be expected in this region we focus on
MP09 which is located 0.5 mm behind and 0.5 mm below the front edge of the plate.
Fig. 6 Locations of the monitoring points (MP)
Figure 7 (left) shows the frequency spectrum of the simulated incompressible

pressure in comparison with the measured sound pressure spectrum. The frequen-
cies of both spectra correlate well. Also the first and second harmonics are present.
The deviations of the main frequencies and the harmonics are below 5% (compare
table 1). In figure 7 (right) the spectrum of ∂ ∂pt as dominant part of the aeroacous-
inc
tic source term is displayed. Again the frequencies correlate well with the measured
ones. An interesting fact is that the second peak which is equivalent to the first har-
monic is much more distinct.
∂ pinc
Fig. 7 Frequency spectra of incompressible pressure (left) and ∂t (right) in comparison to
the measured sound pressure
Table 1 Numerically obtained frequencies in comparison to experimental values
f / [Hz] - Experiment f / [Hz] - Simulation Deviation %

main frequency 634 611 3.8
first harmonic 1271 1214 4.5
second harmonic 1879 1832 3.5
5 Conclusion
An LES of an aero-acoustic test case consisting of a plate located in the turbulent
wake of a circular cylinder was considered. The incompressible pressure fluctua-
tions that are the dominant part of the aeroacoustic source term are investigated.
The higher level of emitted noise compared to the noise caused by a single cylinder
can be explained by the much higher pressure fluctuations that are present at the
front edge of the plate. Further, frequency spectra are taken into account. The char-
acteristic frequencies in the measured sound pressure spectrum can also be found in
the numerical results and correlate well with experimentally obtained frequencies.
The applied numerical method seems to be able to capture the relevant physical
effects and therefore appears to be a good choice for further aero-acoustic studies.
Acknowledgement
This work was supported by the DFG within the collaborative research center “Flow and
Combustion in future Gas Turbine Combustion Chambers”, SFB 568.
References
1. Briggs, W.L., Van Emden, H., McCormick, S.F.: A Multigrid Tutorial. SIAM, Philadelphia
(2000)
2. FASTEST-Manual. Technische Universiät Darmstadt, Chair of Numerical Methods in Me-
chanical Engineering. Darmstadt (2005)
3. Germano, M., Piomelli, U., Moin, P., Cabot, W.H.: A dynamic subgrid-scale eddy viscos-
ity model. Phys. of Fluids A 3(7), 1760–1765 (1991)
4. Hardin, J.C., Pope, D.S.: An Acoustic/Visous Splitting Technique for Computational
Aeroacoustics. Theoret. Comp. Fluid Dynamics 6, 323–340 (1994)
5. Kornhaas, M., Winkler, M., Sternel, D.C., Becker, K., Schäfer, M., Kameier, F.: LES
of cylinder plate configurations and corresponding aeroacoustic effects. Submitted to
XIXème Congrès Français de Mécanique, CFM 2009 (2009)
6. Lighthill, M.J.: On Sound Generated Aerodynamically 1: General Theory. Proc. Roy. Soc.
London Series A 211, 564–587 (1952)
LES of the Flow around Ahmed Body with
Active Flow Control
Siniša Krajnović and Branislav Basara
Abstract. Three different activation strategies for active flow control around an
Ahmed body are investigated using large eddy simulations. Both the separation re-
gion on the slanted surface and the cone-like trailing vortices were influenced using
different actuation strategies. However, only one of the flow regions was influenced
at a time. The actuation of the separation slant region was done using either constant
blowing or periodic blowing and suction through the spanwise slot near the edge be-
tween the slant and the top of the body. Another actuation strategy used blowing of
constant jets into the cone-like vortices with the objective to weaken them. The latter
strategy was found to produce an increased drag in agreement with previous experi-
mental data. Only the actuation strategy using constant blowing along the spanwise
slot was found to decrease the overall drag on the body.
1 Introduction
Recent study of Krajnović and Fernandes [6] has shown that the technique of large
eddy simulation (LES) can be used for simulation of active flow control of vehicle-
like bluff bodies. An Ahmed body spanning between two lateral walls of the tunnel
was used in [6], (from now denoted 2D shape). In the present work, we go a step
forward and apply the active flow control on the original Ahmed body shape [1]
where the flow goes around all sides of the body. Furthermore, the angle of the
slanted rear surface is 25o compared to 0o used for the 2D shape [6]. One observation
from [6], was that the actuation had a larger impact on the spanwise part of the near
wake where the actuation was not applied, i.e. close to the lateral walls. This was
S. Krajnović
Chalmers University of Technology, Göteborg, Sweden
e-mail: sinisa@chalmers.se
B. Basara
AVL List GmbH, Graz, Austria
e-mail: branislav.basara@avl.com
248 S. Krajnović and B. Basara
the result of the 2D character of the body used in that work and it raised the question
about the efficiency of the actuation strategy used in [6] in the case of a 3D Ahmed
body [5, 8, 9]. This is why in the present paper we not only test the same actuation
strategy used in [6] but also two additional actuation strategies previously tested in
experiments of Brunn and Nitche [3] and Brunn et al. [4].
2 Description of the Model and Numerical Details

The geometry of the body is given in Fig. 1a. All the geometric quantities are nor-
malized with the body height, H, equal to 0.288 m. The front part is rounded with
a radius of R/H = 0.347 in the planes, y = 0 and z = 0. This body is placed in the
channel with a cross section of B × F = 6.493H × 4.861H (width × height). The
cross section of this channel is identical to the open test section of the wind tunnel
used in the experiments of Lienhart and Becker [7]. The front face of the body is
located at a distance of 7.3H from the channel inlet and the downstream length be-
tween the rear face of the body, and the channel outlet is 21H. The body is lifted
from the floor, producing a ground clearance of c/H = 0.174, as in the experiments.
The Reynolds number, based on the incoming velocity, U∞ , and the car height, H
was Re = 2 × 105 as in [5].
½ 0.43
0.42
0.41

Cd
0.4

0.39

0.38

0.37
0.36
2.5 3 3.5 4 4.5
a) b) t
Fig. 1 a) Geometry of the Ahmed body (Krajnović and Davidson [5]). b) Time history of the
drag signal from natural flow (red curve); constant blowing in corners (blue curve); constant
blowing at the spanwise edge (black curve) and periodic blowing and suction on the spanwise
edge (green curve)
2.1 Boundary Conditions and the Flow Control

A uniform velocity profile, U∞ , constant in time was used as the inlet boundary
condition in all LES. The homogeneous Neumann boundary condition for velocities
was used at the downstream boundary. The lateral surfaces and the ceiling were
treated as slip surfaces using symmetry conditions (dū/dy = dw̄/dy = v̄ = 0 for the
lateral sides and dū/dz = dv̄/dz = w̄ = 0 for the ceiling). The instantaneous wall
functions were used at all solid walls.
LES of the Flow around Ahmed Body with Active Flow Control 249
Three different actuation strategies were applied in the present work. The first
strategy that was tried was to use actuation applied to a spanwise slot by imposing a
velocity on the boundary equal to us = uA sin(ωAt)(cos(φ )i + sin(φ )j), where i and j
are the unit vectors in the x and z directions, respectively, and the actuation angle, φ ,
was 45o , in agreement with the experiments [4]. Note however that the angle of the
slanted surface in [4] used for this activation strategy was 35o and not 25o as in the
present paper. This will of course have implication on the resulting flow. However,
the aim in the present paper is not to compare the two flows at different angle of slant
but merely to investigate the applicability of the actuation strategy used in [4] for the
flow around body with 25o angle of slant. The actuation amplitude, uA , can be de-
rived from the expression for the momentum coefficient, Cμ = 2su2A /HU∞2 . The ac-
tuation frequency and the moment coefficients in the present study were StA = 0.11
and Cμ = 0.0075, respectively. These values were chosen so that the ratio uA /U∞
and the actuation frequency are the same as in [6]. This actuation is however dif-
ferent from that used in [6] where two slots were used to force symmetric vortex
shedding and thereby delay the wake instabilities. In the present case of the three-
dimensional Ahmed body it is not clear where to place the second slot or even how
many slots should be used due to shear layers coming from the lateral sides of the
body. This is why it was decided to start with investigation of influence of only one
slot in agreement with that used in [3].
The second strategy used only the constant blowing at an angle of 45o in respect
to the top of the body through a spanwise slot from the first actuation strategy.
However, the forcing intensity was changed to Cμ = 0.0014.
The last strategy was to blow two steady jets at the lateral corners of the spanwise
edge between the slanted surface and the top of the Ahmed body. The idea behind
this strategy was to investigate if the drag can be decreased by weakening the in-
tensity of the cone-like vortices that normally roll around the lateral slanted edges
for the rear angle of 25o [5]. The synthetic jet was directed normal to the mean flow
direction similar to that in the experimental part of the study presented in [4]. Simi-
lar was done in [4] and here we used the Cμ = 0.0154 where Cμ = AS u2A /ABU∞2 and
AS = 2b × l is the total area of the two quadratic regions with the width b and the
length l at the corners of the spanwise edge where two steady jets were applied.
3 Numerical Details
Large eddy simulations using the standard Smagorinsky model with the constant
CS = 0.1 were made with the commercial finite-volume code AVL FIRE v2008
(AVL AST [2]). Both convective and viscous plus sub-grid fluxes are approximated
by central differences of second-order accuracy. The second order accurate three
time level scheme was used for time integration. The computational grid used for
all simulations was the coarse grid from [5] containing 3.6 millions computational
nodes. Although this computational grid is too coarse for a well resolved LES, it
was found in [5] that it was sufficient for prediction of main flow structures and that
the prediction of the drag with this grid was acceptable.
a) b)
Fig. 2 Natural flow. a) An iso-surface of the instantaneous low static pressure. The symmetry
plane is colored with the velocity magnitude. b) An iso-surface of the second invariant of the
velocity gradient Q and static pressure on the surface of the body
Fig. 3 An iso-surface of the instantaneous low static pressure and velocity in the symmetry
plane for constant blowing in the corners (a-b) and periodic blowing and suction (c-d). Figure
(c) is from suction process and figure (d) is from blowing process
4 Results
Although the uncontrolled flow using present computational grid has already been
computed in [5], this flow was recomputed with the AVL-Fire solver in order to
eliminate the influence of different solvers on the results. However, the new results
for the uncontrolled flow are very similar to those obtained in [5]. The resulting
instantaneous flow is presented in Fig. 2. The hairpin vortices in the recirculation
region on the slanted surface and the main cone-like trailing vortices are visible in
this figure. Comparison of Figs. 2a and Fig. 3 shows that the actuation using two jets
with constant blowing has for the result only weakening of the cone-like vortices.
On the other hand, the periodic blowing and suction changes flow drastically (com-
pare Fig. 2a and Fig. 3 (c-d)). The suction phase is very efficient in preventing the
formation of the recirculation region on the slanted surface while the blowing phase
enriches this separated flow region. Another observation is that both the blowing
and the suction have small influence on the cone-like vortices. The difference in the
resulting surface pressure is visible in Figs. 2b and 4 between the natural and the
two controlled strategies. It is also interesting to observe the local influence of the
actuating jet on the surface pressure with an upstream increase and a downstream
reduction.
Although there are differences in the pressure pictures, the overall mean drag is
similar between the three cases. Unfortunately, the natural flow shows lower drag
(CDn = 0.4) compared with approximately CDc = 0.41 for both strategies illustrated
a) b) c)
d) e) f)
Fig. 4 An iso-surface of the second invariant of the velocity gradient Q and static pressure
on the surface of the body for two time steps when constant blowing in corners is applied
(a-b) and periodic suction (d) and blowing (e). Figures (c) and (f) are zoom of (b) and (e),
respectively
in Figs. 3 and 4. However, the frequency picture of the three signals is different
(see Fig. 1b). In particular, the periodic blowing and suction seems to dominate the
resulting drag signal. It should be mentioned that the inefficiency of the actuation
strategy using jets with Cμ = 0.0154 in the corners is perhaps not a surprise know-
ing that the experimental investigation of Brunn et al. [4] showed only minor drag
reduction (less than 2 per-cent) and only for Cμ ≤ 0.003. For higher values of Cμ ,
the drag coefficient in [4] was increase in agreement with the result for Cμ = 0.0154
used in the present work. In addition, it should be mentioned that the overall drag of
natural flow of CDn = 0.4 is close to CDn = 0.41 found in experimental study [4] for
two and a half higher Reynolds number. Krajnović and Davidson showed in [5] a
relative small influence of the Reynolds number to all aspect of this flow including
the pressure part of the drag. Therefore, the agreement of the overall drag between
the present simulation and the experimental value from [4] is found to be an indica-
tion of the accuracy of the prediction from the present LES.
The resulting time-averaged flows are presented in Figs. 5 and 6. The near wake
separation bubble is slightly shorter in the flow using periodic blowing and suction
than in the natural flow and the controlled flow using jets. The length of the sepa-
ration bubble in the near wake was found to be approximately 0.65H in the natural
flow and the flow actuated with two jets and 0.54H in the flow actuated with the
periodic blowing and suction. Note that the value of 0.65H in the natural flow is al-
most identical to that found in [5]. The recirculation region on the slanted surface is
slightly longer in the periodic blowing and suction flow than in the other two flows.
Finally, the constant blowing in the spanwise slot was applied as an actuation strat-
egy and this showed to decrease drag to CDcs = 0.37 (see Fig. 1b). Figure 5d shows
that the resulting time-averaged flow using this actuating strategy is very different
a) b)
c) d)
Fig. 5 Time-averaged streamlines projected on the center plane on the body for a) natural
flow, b) constant blowing in corners, c) periodic blowing and suction on spanwise edge and
d) constant blowing on spanwise edge
from all other flows discussed above. As this actuation strategy was applied only re-
cently it remains to explore the reasons for its efficiency. However, one observation
is that longer averaging time is required for this flow than for the others meaning
that the flow contains some low frequency motion.
5 Discussion
Both the cone-like trailing vortices and the separation region on the slanted surface
were influenced using actuation strategies presented in the paper. However, they
were applied on only one of these flow regions at a time and this might be the rea-
son of poor efficiency as the flow regions influence each other. Flow on and around
the slanted surface of the body is characterized by strong interaction of the cone-like
vortices and the structures in the recirculation region on the slant as found in [5].
Weakening of the cone-like structures will give more space for the recirculation re-
gion on the slant which eventualy will not reattach on the slant and produce similar
flow to that around body with slanted surface at 35o [7]. This is perhaps what hap-
pened in the case of constant blowing through the spanwise slot (see Fig. 5d). More
investigations are needed to find the reason for this very different flow in actuation
using only constant blowing through the spanwise slot. The foci of the near wake
bubble in the flow using this actuation strategy was moved further back compared
with other three flows presented in Fig. 5. Such a change in the flow was found in
[6] to result in an increase of the static pressure on the base of the body and reduc-
tion of the overall drag. The present study used only a single value of Cμ and StA for
a) b) c)
d) e) f)
Fig. 6 Time-averaged streamlines projected on the planes z/H = 0.337 (a,d); z/H = 0.675
(b,e) and z/H = 0.73 (c,f). Constant blowing in corners (a-c) and periodic blowing and suc-
tion on spanwise edge (d-f)
each of the actuation strategies. A systematic investigation of the variation of these

parameters is required in order to find the best optimal values.
Acknowledgments
Computer time at SNIC (Swedish National Infrastructure for Computing) resources at the
Center for scientific computing at Chalmers (C3SE) is gratefully acknowledged. Help of
Master thesis student Jan Östh with preparing preparing and monitoring of some of the sim-
ulations is gratefully acknowledged.
References
1. Ahmed, S.R., Ramm, G., Faltin, G.: Some salient features of the time averaged ground
vehicle wake. SAE Paper 840300 (1984)
2. AVL List GmbH, Graz, Austria. Fire Manual v2008, AVL AST (2008)
3. Brunn, A., Nitsche, W.: Active control of turbulent separated flows over slanted surfaces.
Int. J. Heat and Fluid Flow 27, 748–755 (2006)
4. Brunn, A., Wassen, E., Sperber, D., Nitsche, W., Thiele, F.: Active drag control for a
generic car model. In: Notes on Numerical Fluid Mechanics and Multidisciplinary Design,
vol. 95/2007, pp. 247–259 (2007)
5. Krajnović, S., Davidson, L.: Flow around a simplified car, part 1: Large eddy simulation.
ASME: Journal of Fluids Engineering 127, 907–918 (2005)
6. Krajnović, S., Fernandes, J.: LES of the flow around a 2-dimensional vehicle model with
active flow control. In: ERCOFTAC WORKSHOP, Direct and Large-Eddy Simulations,
Trieste - Italy, September 8-10, vol. 7 (2008)
7. Lienhart, H., Becker, S.: Flow and turbulent structure in the wake of a simplified car
model. SAE Paper 2003-01-0656 (2003)
8. Minguez, M., Pasquetti, R., Serre, E.: High-order large-eddy simulations of flow over the
ahmed body car model. Physics of Fluids 20, 095101 (2008)
9. Minguez, M., Pasquetti, R., Serre, E.: Spectral vanishing viscosity stabilized les of the
ahmed body turbulent wake. Communications in computational physics 5(2-4), 635–648
(2009)
Enhanced Bubble Migration in Turbulent
Channel Flow by an Acceleration-Dependent
Drag Coefficient
J.G.M. Kuerten, C.W.M. van der Geld, and B.J. Geurts
Abstract. DNS of turbulent bubble-laden channel flow has been used with one-way
coupling to investigate consequences of a postulated dependence of acceleration
on drag. This dependence is known to hold for converging and diverging steady
streamlines, and is here postulated to also exist for the instantaneous acceleration of
micro-bubbles that arises from turbulent velocity fluctuations. Bubble segregation is
found to be considerably increased through this dependency of drag on acceleration,
particularly for smaller bubbles.
1 Introduction
Migration of particles and bubbles in turbulent channel flow is important in the
process industry. The efficiency of air-water separators, for example, depends on
the lateral migration of air bubbles. This is influenced by several agglomeration
effects. First, gravity causes agglomeration of bubbles near the top wall. In addition,
if bubble sizes are relatively small, the drag force causes turbophoresis in turbulent
flows, which results in agglomeration of bubbles near a wall [1]. The present paper
deals with another agglomeration effect in turbulent channel flow.
Bubbles in tapwater experience a drag force that is close to that of a solid
sphere [2]. If these bubbles are accelerated in an area with converging streamlines,
as for example occurring at the entrance of a venturi, the drag force coefficient
J.G.M. Kuerten · C.W.M. van der Geld
Department of Mechanical Engineering, Eindhoven University of Technology, P.O. Box 513,
NL-5600 MB Eindhoven
e-mail: j.g.m.kuerten@tue.nl
B.J. Geurts
Department of Applied Mathematics, University of Twente, P.O. Box 217,
NL-7500 AE Enschede and Department Applied Physics,
Eindhoven University of Technology, P.O. Box 513, NL-5600 MB Eindhoven
e-mail: b.j.geurts@utwente.nl
256 J.G.M. Kuerten, C.W.M. van der Geld, and B.J. Geurts
depends not only on the relative velocity of the bubble, but also on the acceleration
of the fluid. This effect has been confirmed by both measurements [2] and numerical
computations [3].
If bubbles move in a turbulent flow, they experience local instantaneous accelera-
tions, even if the time-averaged acceleration of the fluid is equal to zero everywhere.
We hypothesize that in such a flow the drag coefficient also depends in the same
way on the local instantaneous acceleration of the fluid. There appears no physical
reason why the same mechanism would not prevail in unsteady flows when the ac-
celeration from turbulent fluctuations is nonzero. In this paper consequences of our
hypothesis for bubble migration in turbulent channel flow will be investigated by
means of direct numerical simulation (DNS). In the next section the dependence of
the drag coefficient on the fluid acceleration will be discussed. Section 3 presents
the numerical method, results are shown in Sect. 4 and conclusions are stated in
Sect. 5.
2 Drag Force in Accelerating Flow

It is known from experiments [2] and numerical computations [3] that the drag force
coefficient a bubble experiences in an accelerating fluid, CD , not only depends on
the Reynolds number,
|vrel |ρL db
Re = , (1)
μL
with vrel = vb − vL the difference of bubble and liquid velocity, ρL and μL the liquid
mass density and dynamic viscosity and db the bubble diameter, but also on the
acceleration number Ac. This acceleration number can be defined as:
vrel · Dv L
Ac = db Dt
, (2)
|vrel |3
which weighs the acceleration of a fluid particle due to motion relative to the bubble
to that in the absence of bubbles [2]. The relative velocity plays an essential role
because the pressure distribution and shear stresses at the bubble-liquid interface
are affected by acceleration only if relative velocity is nonzero [3]. For low relative
velocities obviously high acceleration numbers are obtained. For a limited range of
Ac-values measurements and numerical computations have shown the existence of
a dependence of CD on Ac of the form
CD (Re, Ac) = (1 + b1Ac)CD (Re, 0) + b2Ac, (3)
where b1 and b2 are positive constants and CD (Re, 0) is the drag coefficient in the
absence of acceleration. Expression (3) holds for Re > 10 and |Ac| < 3. In con-
verging flows the value of Ac is nonzero and drag is increased according to (3).
The physical reason for this is the increase in dissipation rate as a consequence
of increased velocity gradients near the bubble. Whether similar phenomena hap-
pen in instantaneous turbulent accelerations needs to be investigated, still. Such an
Enhanced Bubble Migration in Turbulent Channel Flow 257
investigation is beyond the scope of the present paper, as is the determination of b1

and b2 for Re < 10 and |Ac| > 1.
It is to be expected that the effect of a possible acceleration dependence of
the drag coefficient is larger for smaller and lighter particles, since they have a
smaller relative velocity and hence a larger acceleration number. Moreover, in a
wall-bounded turbulent flow the instantaneous fluid acceleration is generally larger
than in homogeneous turbulence. Therefore, in this paper a dependence of the type
of (3) will be investigated in case of bubble motion in turbulent channel flow, where
the fluid acceleration is fluctuating in both time and space. Statistically the fluid
acceleration equals zero everywhere in the flow. In the definition of the accelera-
tion number Ac (2) the fluid acceleration is equal to the convective contribution to
the undisturbed local instantaneous fluid acceleration at the position of the bubble.
The undisturbed acceleration is known in a DNS with one-way coupling between
bubbles and fluid. This is one of the reasons to compute bubble trajectories with one-
way coupling in the DNS. To prevent too large values of Ac if the relative velocity
is small, the absolute value of Ac is bounded by the value of 3. Larger values occur
in particular in case of small bubbles, since they follow the flow more closely. In
accordance with the findings of Magnaudet et al. [3] for Re > 10, in the simulations
presented here b1 = 0.25 and b2 = 0.55 have been chosen.
3 Numerical Method
The present study focuses on the migration of air microbubbles to the upper wall
in turbulent flow of water in a horizontal channel. The fluid flow is determined by
solving the Navier-Stokes equation and continuity equation by means of DNS for
incompressible flow in a channel with periodic boundary conditions in the stream-
wise and spanwise directions. The numerical method is described in detail in [4].
The two periodic directions are treated with a Fourier-Galerkin method. In the di-
rection normal to the walls a Chebyshev collocation method is applied. A total of
129 collocation points are taken in the normal direction and 128 Fourier modes in
both periodic directions. For time integration a second-order accurate combination
of a three-stage Runge-Kutta method and the Crank-Nicolson method is applied. In
every stage of the Runge-Kutta method first the nonlinear terms in the Navier-Stokes
equation are calculated by means of fast Fourier transform. In order to prevent alias-
ing the 3/2-rule is applied in the two periodic directions. Next the pressure and
wall-normal velocity component are determined with the implicit Crank-Nicolson
method and the influence matrix method [5, 6] is used to obtain a divergence-free
velocity field at the end of the Runge-Kutta stage. Finally, the other two velocity
components are calculated with the Crank-Nicolson method. Since in the implicit
parts of the algorithm all Fourier modes are decoupled, only a one-dimensional sys-
tem of equations has to be solved for each Fourier mode. A substantial amount of
computing time can be saved by calculating the LR-decomposition of the matrices
only once and storing the results in memory.
Spherical bubbles with a diameter in the range 18-96 μ m are introduced

homogeneously in the channel at a void fraction below 0.1%. The motion of the
bubbles is computed with a point force model in which all forces except the lift
force and the history force are taken into account:
dvb DvL 1
(ρb + CAM ρL )Vb = Fbuo + ρL (1 + CAM )Vb − CD ρL π db2 |vrel |vrel . (4)
dt Dt 8
Here Fbuo accounts for the body force (combination of gravity and buoyancy), Vb
is the bubble volume and the added mass coefficient CAM is taken to be 0.5. In the
following section it will be shown that the lift force does not play a role. If the value
of |Ac| experienced by a bubble exceeds 3 the value of CD is computed for Ac = 3
or -3; during the computation only 0.01 % of the biggest bubbles (diameter 96 μ m)
and 16% of the smallest bubbles (diameter 18μ m) typically had such a limited value
of CD . Equation of motion (4) is numerically solved by a second-order accurate time
integration method based on the trapezoidal rule. To this end the average of the fluid
velocity and fluid acceleration over two consecutive time instants is used. These
quantities are determined at the position of a bubble with a fourth-order accurate
interpolation method, consisting of Lagrangian interpolation in the two periodic di-
rections and Hermite interpolation in the wall-normal direction. In the following
section it will be verified that this interpolation method is sufficiently accurate.
All results have been calculated at a Reynolds number based on the friction ve-
locity and half the channel height of 150. This corresponds e.g. to a bulk water
velocity of 19 cm/s and channel height of 1.8 cm. Four different bubble sizes have
been considered with diameters ranging from 18 to 96 μ m. For each of these sizes
statistical results are obtained by averaging over 40,000 bubbles both for the case
where the drag coefficient depends on the acceleration number as specified in (3)
and for the case when CD (Re, Ac) is taken to be CD (Re, 0). Since one-way coupling
is used, the fluid is not influenced by the presence of the bubbles and all different
types of bubbles can be included in one single simulation. The simulation is started
from a fully-developed turbulent velocity field in which the bubbles are randomly,
but homogeneously distributed with velocity equal to the sum of the fluid velocity
at the bubble position and the terminal rise velocity.
4 Results
In this section, first the effect of a drag coefficient of the form (3) on particle settling
is discussed. Then the influence of the interpolation method and of the neglect of
the lift force is analyzed.
The assumed acceleration dependence of drag coefficient affects bubble migra-
tion. A typical result is presented in Fig. 1. In this figure the bubble concentration
near the top wall is plotted versus time, and for each bubble size the history is
given both with and without the dependence of CD on Ac. The differences in con-
centration histories are obvious, showing the importance of the effect of accelera-
tion on the settling of microbubbles in a channel. The figure clearly shows that the
40
35
30
25
wall
20
c
15
10
0
0 1000 2000 3000 4000 5000
+
t
Fig. 1 Bubble concentration near the top wall as a function of time. From bottom to top:
db = 18 μ m, 32 μ m, 55 μ m and 96 μ m; solid: with dependence of CD on Ac; dashed: without
dependence of CD on Ac. A value of cwall = 40 corresponds to a situation in which all bubbles
introduced homogeneously are caught at the top wall. Time t + is time divided by uτ /νL ,
where uτ is friction velocity and νL the liquid kinematic viscosity
effect of the acceleration-dependence is largest for the smallest bubbles. This is not
surprising since they experience the smallest relative velocity and hence the highest
acceleration number.
This effect can also be observed in Fig. 2, where for the bubbles with db = 18μ m
the average relative velocity component in the wall-normal direction is shown as
a function of the wall-normal coordinate. Averages are taken over the two homo-
geneous directions and time. Results are shown for both cases with and without
dependence of drag coefficient on the acceleration number. For the latter case the
particle Reynolds number is so small that the drag force is in very good approxima-
tion linear in the relative velocity. Moreover, the simulation results indicate that the
average bubble acceleration is very small. The time-averaged version of Eq. (4) then
implies that the average relative velocity is constant over the height of the channel
and proportional to the acceleration of gravity g:
gdb2
vrel = − . (5)
18νL
For the smallest bubbles considered herethis yields an average wall-normal
relative velocity of 0.0226uτ , where uτ = τw /ρL with τw the wall shear stress
is the friction velocity. This is in very good agreement with the simulation results
shown in Fig. 2. Small discrepancies are due to the second term on the right-hand
side of (4), which is largest close to the walls of the channel. For the case where the
drag coefficient does depend on the acceleration number the drag coefficient is no
longer inversely proportional to the relative velocity, which leads to a higher relative
velocity that is not constant over the height of the channel.
0.06
0.05
0.04
vrel,y
0.03
0.02
0.01
0
−1 −0.5 0 0.5 1
y/H
Fig. 2 Average relative velocity in wall units in the wall-normal direction for bubbles with
db = 18 μ m as a function of the wall-normal coordinate; solid: with dependence of CD on
Ac; dashed: without dependence of CD on Ac. Gravity points in the negative y-direction
Next the effect of the interpolation method for the fluid velocity and acceleration
at the bubble position is studied. The method used, fourth-order accurate interpo-
lation, is a compromise between accuracy and calculation time. Direct summation
of the Fourier series would be the most accurate method, but is prohibitively ex-
pensive if there are many bubbles present. In order to investigate the effect of the
interpolation method the simulation has been repeated with linear interpolation. Al-
though individual bubble tracks differ substantially after some time, average results
as shown in this paper are hardly influenced. A typical result is shown in Fig. 3,
where the bubble concentration near the top wall is shown as a function of time for
the smallest bubbles considered. Similar results are obtained for the larger bubbles.
Finally, the effect of lift force on the bubbles is studied. To this end a lift force
1
FL = cL ρ f π db3 (vL − vb ) × ω , (6)
6
is added to the right-hand side of the equation of motion (4). In this lift force ω
represents the fluid vorticity: ω = ∇× vL and the lift coefficient cL is taken to be 1/2.
The scaling of lift and drag force with bubble diameter indicates that the influence
of lift force will be largest on the largest bubbles. Indeed, lift force is proportional
to the volume of the bubble, whereas the drag force is proportional to its diameter.
We can make a crude estimate of both forces, by substituting typical values for fluid
velocity and vorticity. The result is that the ratio of lift and drag force is on the
order of 2 × 10−3 for the largest bubbles considered. In order to verify this, we also
performed simulations with lift force included. The results show that the effect of lift
force on individual bubble tracks is small but non-negligible; the effect on average
bubble properties is hardly noticeable even for the largest bubbles considered here.
20
15
wall
c
10
0
0 500 1000 1500 2000 2500 3000 3500 4000 4500
+
t
Fig. 3 Bubble concentration near the top wall as a function of time for bubbles with
db = 18μ m. solid: with dependence of CD on Ac; dashed: without dependence of CD on
Ac. The lines represent the results with fourth-order accurate interpolation of fluid velocity
and acceleration to the bubble position, the symbols represent linear interpolation
5 Conclusions
The instantaneous acceleration of micro-bubbles, by turbulent velocity fluctuations,
was hypothesized to affect the drag force coefficient. This has been found to affect
bubble segregation in channel flow. The smaller the bubble diameter, the larger the
effect has been found to be. For 18 μ m bubbles a significant influence is found with
the drag force given by Eq. 3 with b1 = 0.25 and b2 = 0.55. The lift force and
the interpolation method for the determination of the fluid velocity at the bubble
position have been found not to affect the long-time segregation trends.
References
1. Kuerten, J.G.M., Vreman, A.W.: Can turbophoresis be predicted by large-eddy simula-
tion? Phys. Fluids 17, 011701 (2005)
2. Van der Geld, C.W.M., Van Wingaarden, H., Brand, B.A.: Experiments on the effect of
acceleration on the drag of tapwater bubbles. Exp. in Fluids 31, 708–722 (2001)
3. Magnaudet, J., Rivero, M., Fabre, J.: Accelerated flows past a rigid sphere or a spehrical
bubble. Part 1. Steady straining flow. J. Fluid Mech. 284, 97–135 (1995)
4. Kuerten, J.G.M.: Subgrid modeling in particle-laden channel flow. Phys. Fluids 18,
025108 (2006)
5. Canuto, C., Hussaini, M.Y., Quarteroni, A., Zang, T.A.: Spectral methods in fluid
dynamics. Springer, Berlin (1988)
6. Kleiser, L., Schumann, U.: Treatment of incompressibility and boundary conditions in 3-D
numerical spectral simulations of plane channel flows. In: Hirschel, E.H. (ed.) Proceedings
of the Third GAMM-Conference on Numerical Methods in Fluid Mechanics, Vieweg,
Braunschweig, pp. 165–173 (1980)
Experimental and Numerical Study of
Unsteadiness in Boundary Layer / Shock Wave
Interaction
L. Larchevêque, P. Dupont, E. de Martel, E. Garnier, and J.-F. Debiève
Abstract. The unsteady flowfield induced by an interaction between an impinging

shock wave and a turbulent boundary layer is analysed by means of both Large-
Eddy simulations and experiments relying on PIV and wall pressure measurements.
A simple kinematic model is derived from these analyses and demonstrates a good
ability to reproduce the main unsteady features found in the data.
1 Introduction
From several experimental results [2, 3, 5, 9] as well as numerical simulations
[7, 10–12], it is commonly accepted that shock-induced separated flows exhibit a
very similar behaviour whatever the geometry considered: the separated region and
the shock developing upstream of it exhibit unsteadiness of low frequencies, at least
two orders of magnitude lower than the frequencies of the energetic scales found in
the upstream boundary layer.
Although the origin of these low frequencies is not yet clearly understood, the
mean and unsteady wall pressure fields exhibit distinctive features in their down-
stream evolution. The main objective of this work is to describe the links be-
tween the motion of the reflected shock wave and some of the flow features found
downstream of the separation.
The flow under examination is a turbulent boundary layer with a Mach number
M = 2.3, at a Reynolds number of Rθ = 5000, impinged by an oblique shock wave
L. Larchevêque · P. Dupont · J.-F. Debiève
IUSTI, UMR CNRS 6595, Aix-Marseille I University, 5 rue Enrico Fermi,
F-13453 Marseille, France
e-mail: lionel.larcheveque@polytech.univ-mrs.fr
E. de Martel · E. Garnier
ONERA, Applied Aerodynamics Department, 8 rue des Vertugadins,
F-92190 Meudon, France
e-mail: eric.garnier@onera.fr
264 L. Larchevêque et al.
Fig. 1 Schlieren picture of the impinging shock / boundary layer interaction
corresponding to flow deflections ranging from 8◦ to 9.5◦ . The corresponding shock

strength is large enough for the boundary layer to separate. The flow has already
been extensively documented [1, 4–6] and the global organisation of the interaction
region, based on a Schlieren picture is seen in Fig. 1.
Experiments were carried out in the low turbulence supersonic wind tunnel of
the IUSTI. Velocity fields were obtained by PIV measurements resulting in sets of
5000 uncorrelated images encompassing the whole development of the interaction.
Unsteady wall pressure measurements are carried out by means of Kulite pressure
transducers. The sampling time ranges from several hundreds to several thousands
of periods.
Complementary to the experiments, Large-Eddy Simulations have been carried
out, either at ONERA [7] or at IUSTI, using ONERA’s code FLU3M. The com-
putations, both relying on a 9.5◦ deflection, mainly differ in the inflow boundary
layer, whose thickness is about 20 % larger for the ONERA computation than for
the IUSTI one. Both simulations were performed at the full scale of the wind tunnel,
except for the spanwise extension, reduced to 1.5 δ with periodic boundary condi-
tions for cost saving reasons. In turn, such a simplification allows the computation
of 30 (ONERA) to 120 (IUSTI) periods of the shock motion, making feasible a joint
experimental-numerical analysis of the flow unsteadiness.
2 Experimental and Numerical Results

The main scaling parameter for shock / boundary layer interaction with separation is
the interaction length L, defined as the distance between the location X0 where the
peak value of the wall rms pressure is found ( i.e. the median location of the foot
of the reflected shock) and the extrapolation of the incident shock down to the wall
(see Fig. 1). The interaction lengths of the experiments with 8◦ and 9.5◦ deflections
are respectively equal to 46 mm and 71 mm. The typical Strouhal number of the low
frequency modulation of the flow, based on these lengths and the external velocity,
is roughly equal to 0.03 for both cases, in concordance with values found for a wide
range of separated flows [4].
Study of Unsteadiness in Boundary Layer / Shock Wave Interaction 265
Fig. 2 Conditionally averaged mean streamwise velocity (left) and vertical velocity
fluctuation (right) for the 9.5◦ case
Conditional analysis was applied to the PIV velocity fields based on the instanta-
neous vertical extent of the recirculating region [8]. Large amplitude variation of the
bubble size were observed, as seen in the left part of Fig. 2 where streamwise ve-
locity fields, averaged over the data respectively exhibiting the lowest, highest and
average vertical extent of the bubble, are plotted. For the same subsets of realiza-
tions, the conditionally averaged location of the reflected shock have been estimated
from the standard deviation of the vertical velocity. Results for four vertical loca-
tions are plotted in the right part of Fig. 2 for the 9.5◦ case, clearly demonstrating
that the shock motion and the dynamics of the separated region are strongly linked
together.
Unfortunately, the PIV measurements are not time-resolved. Consequently, the
unsteady dynamics of the separated bubble, of the shock and their mutual influence
can only be studied from experiments by means of the wall pressure measurements.
LES data can be used to alleviate this drawback provided that the unsteady pressure
fields are in agreement with the experimental data.
However both LES computations underpredict the interaction lengths with
L = 43 mm (ONERA) and L = 57 mm (IUSTI), as already found by LES computa-
tions of the 8◦ case carried out by various partners of the European project UFAST.
The origin of this discrepancy is not clearly understood except for the influence of
the incoming boundary layer demonstrated by the differences in the two LES. How-
ever the underprediction remains, though weakened, even if the incoming bound-
ary layer is in very good agreement with the experimental measurement, as for the
IUSTI computation.
The LES datasets are nonetheless valuable when classified by the interaction
length rather than by the deflection angle. According to this criterion, the ON-
ERA computation matches the 8◦ experimental case whereas the IUSTI LES corre-
sponds to a 8.8◦ flow. LES data thus rescaled agree well with the experiments in the
interaction region, as shown by Fig. 3. The longitudinal evolution of the mean
2.8 1
2.6 0.8
Coherence
2.4 0.6
2.2 0.4
2 0.2
P*
1.8 0
-0.5 0 0.5 1 1.5
4
1.6
3
1.4
Phase
2
1.2
1
1
0
0.8
−0.5 0 0.5 1 1.5 2 -1
*
X* x
Fig. 3 Longitudinal evolution of the mean pressure (normalised by the incoming pressure,
left), of the coherence (right, top) and the phase (right, bottom) pressure spectra with refer-
ence location taken at X ∗ =(X − X0 ) /L=0. ∇: 8◦ experiments; : 9.5◦ experiments; − − −:
ONERA LES; −−−−−: IUSTI LES
pressure in the interaction, plotted on the left part of this figure, is adequately
recovered by the computations in the interaction region 0 ≤ X ∗ ≤ 1. The main
discrepancy is found for the IUSTI computation which exhibits a small plateau
classically associated with the reattachment location.
Beyond the agreement on mean values, low-frequency modulations of the flow
are found in both LES. The typical Strouhal numbers StL 0.031 are in good con-
cordance with the experiments. Moreover, the computations are able to recover the
strong linear dependence of the flowfield on the motion of the reflected shock. It
can be checked by looking at the upper right plot of Fig. 3 dedicated to the longi-
tudinal evolution of the pressure coherence for StL = 0.031 with the reference value
at X ∗ = 0. Note that the pressure recorded at the median location of the reflected
shock is used to characterise the shock motion for experimental convenience. How-
ever the LES data make possible to follow the instantaneous location of the shock
and to correlate it with the downstream wall pressure fluctuations. It yields similar
longitudinal evolutions of coherence as the ones obtained by using the pressure at
X ∗ = 0.
Experimentally, the low-pass filtered space-time correlation in the shock fre-
quency range has been used to estimate the phase between the location X ∗ = 0 and
locations in the separation region and downstream of the reattachment point[1]. The
lower right plot of Fig. 3 demonstrates that the pressure remains in phase in the first
half of the interaction region whereas a phase shift roughly equal to π is found in
the second half and beyond. The phase spectra of both LES with StL 0.031 demon-
strates a similar evolution, though the phase value beyond X ∗ = 0.5 is slightly lower
than in the experiment. This is due to the fact that LES phases are related to the
single dominant frequency whereas the phases deduced from the low-pass filtered
correlations correspond to an average over the low frequency range. Similar longi-
tudinal phase evolution but with a higher phase shift are found in the LES phase
spectra for Strouhal numbers noticeably higher than StL 0.031. The wide range
of frequencies over which a rather similar phase shift is found is a consequence
of the slow evolution over time of the frequency of the dominant fluctuation mode
encountered in the experiments as well as in the LES.
3 A Simple Model for the Wall Pressure Evolution

An interpretation of the results of Fig. 3 based on a simple scheme is proposed
hereafter. The scheme aims at reproducing the main features of the longitudinal
evolution of the instantaneous pressure in order to highlight the link between the
dynamics of the shock-boundary layer interaction and its corresponding signature
at the wall.
It may be mentioned that similar results on coherence and phase opposition
within the interaction region have already been found for compression ramps [9].
The authors built an interpretation based on a specific pressure variation at the sep-
aration and reattachment points when the bubble is contracting / dilating. However,
for the present case, no specific trends have been found for the pressure fluctuations
in the vicinity of the reattachment point.
The present scheme is built from the observations made in the previous section
and is sketched in the left part of Fig. 4. It relies on a longitudinal pressure evolution
translated in concordance with the shock motion. This hypothesis is supported by
the quasi-equivalence already mentioned between the use as reference signal of ei-
ther the wall pressure at X ∗ =0 or the instantaneous shock location when computing
the coherence of Fig 3. It is therefore assumed that the pressure downstream of the
shock explicitly depends on its instantaneous location.
Next, the pressure level beyond X ∗ 0.4 should be out of phase with respect to the
shock motion in order to reproduce the π phase shift found in the right part of Fig. 3.
It is seen from the left part of the same figure that this location roughly corresponds
to the end of the pressure rise directly induced by the shock. It implies that the
variations in the shock output pressure should be of the same sign than the variations
in the shock location. Such a feature can be obtained by adding to the shock pressure
2.6
2
Reflected Shock
2.4
P(t) / P1
1.5
2.2
1
2
0.5
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
P / P1
1.8
2
1.6
Reflected Shock
P(t) / P1
1.5
1.4
1
1.2
Upstream BL Shock Intermitency Separation bubble 1

0.5 -0.5 0 0.5 1 1.5 -0.5 0 0.5 1 1.5
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
X* x*
Fig. 4 Schematic of the wall pressure model (left part) and longitudinal pressure evolution
conditioned by the shock location (right part). For the right part, first and second plots corre-
spond respectively to the ONERA and IUSTI LES: most upstream (· · · · · ·), average (−−−−−)
and most downstream (− − −) shock locations
Fig. 5 Phase of the pressure with respect to the wall pressure at location X ∗ = 0 for
StL = 0.031. ONERA computation (left) and IUSTI computation (right). Regions where the
coherence is statistically insignificant have been blanked
jump a constant adverse pressure gradient in the region of shock excursion. Down-
stream of the shock exit, the pressure evolution is assumed to be characterised by
another constant adverse pressure gradient of lower magnitude. It allows the phase
to remain constant that region, in concordance with the experimental and LES data.
The validity of the model is explored by computing its prediction for the lon-
gitudinal evolution of the mean and rms value as well as of the phase. A random
computation assuming a Gaussian law for the shock location and using estimations
from the experiments for the length of the region of shock excursion, the pres-
sure step through the reflected shock and the pressure gradients in the shock region
and the recirculation yields good satisfactory agreements with the low-pass filtered
measurements.
Unfortunately, the key assumption of the model, namely the variation of the out-
put pressure of the shock with respect to its location, can not be checked from the
experiments since the instantaneous location of the shock can not be estimated using
the time-unresolved PIV measurements. On the contrary, LES data can be used for
that purpose. The unsteady fields issued from these computations have been sorted
according to the location of the shock extracted from both the pressure and velocity
fields. The data have then been averaged over the 10 % of the samples corresponding
respectively to the most upstream and downstream locations as well as for the 10 %
centred on the mean location. The resulting conditionally averaged wall pressures
are plotted in the right part of Fig. 4. It is obvious from these plots that the more the
shock is upstream located, the more the output pressure level of the shock is low.
Moreover the crossing of the pressure curves occurs more upstream for the IUSTI
LES than for the ONERA one, in concordance with the more upstream located phase
shift of π seen in Fig. 3.
The LES data also yields the extension of the model from the wall to the core
flow. As a matter of fact, a phase shift close to the value of π is found almost ev-
erywhere downstream of the shock, as demonstrated in Fig.5. The main exception
is a small region near the foot of the expansion wave clearly seen on the IUSTI plot
(right part). Note that the two straight regions of null phase found in ONERA’s data
are meaningless since there are due to the partial reflection of the reflected shock on
a non-conformal mesh interface used to reduce the computational cost.
4 Conclusion
The proposed model is able to reproduce the main features of the low frequency
wall pressure evolution across the interaction, found either from the experiments or
the computations. It must be emphasised that this model is not a predictive method
for computing the statistical moments, but rather a sketch to represent the links
between the various unsteady phenomena, namely the variation of the bubble size
in connection with the reflected shock motion and its pressure signature at the wall.
One may inferred from it that the phase shift of π found downstream of the shock is
kinematical in nature, as being directly induced by the motion of the shock.
Acknowledgements. Part of this work was carried out with the support of the Research Pole
ONERA/CNES Aérodynamique des Tuyères et Arrir̀e-Corps (ATAC) and of the European
STREP UFAST. Computational resources required by the IUSTI computation were provided
by the CNRS French national facility IDRIS under project n◦ 91877. Their support is grate-
fully acknowledged.
References
1. Debiève, J.F., Dupont, P.: Dependence between shock and separation bubble in a shock
wave / boundary layer interaction. In: IUTAM Symposium on Unsteady Separated Flows
and their Control, Corfu, Greece (2007)
2. Delery, J., Marvin, J.G.: Shock wave–boundary layer interactions. AGARDograph AG-
280, AGARD (1986)
3. Dolling, D.S.: Fifty years of shock-wave/boundary-layer interaction research: what next?
AIAA J. 39(8), 1517–1531 (2001)
4. Dupont, P., Haddad, C., Ardissone, J.P., Debiève, J.F.: Space and time organisation of
a shock wave/turbulent boundary layer interaction. Aerospace Science and Technol-
ogy 9(7), 561–572 (2005)
5. Dupont, P., Haddad, C., Debiève, J.F.: Space and time organization in a shock induced
boundary layer. J. Fluid Mech. 559, 255–277 (2006)
6. Dupont, P., Piponniau, S., Sidorenko, A., Debiève, J.F.: Investigation of an oblique shock
reflection with separation by PIV measurements. AIAA J. 46(6) (2008)
7. de Martel, E., Garnier, E., Sagaut, P.: Large eddy simulation of impinging shock wave
/ turbulent boundary layer interaction at M=2.3. In: IUTAM Symposium on Unsteady
Separated Flows and their Control, Corfu, Greece (2007)
8. Piponniau, S., Dussauge, J.P., Debieve, J.F., Dupont, P.: A simple model for low-
frequency unsteadiness in shock-induced separation. J. Fluid Mech. 629, 87–108 (2009)
9. Thomas, F.O., Putman, C.M., Chu, H.C.: On the mechanism of unsteady shock oscilla-
tion in shock wave/turbulent boundary layer interaction. Experiments in Fluids 18, 69–81
(1994)
10. Touber, E., Sandham, N.D.: Oblique shock impinging on a turbulent boundary layer: low-
frequency mechanisms. In: 38th AIAA Fluid Dynamics Conference, Seattle, United State
(2008)
11. Wu, M., Martin, M.P.: Direct numerical simulation of supersonic turbulent boundary
layer over a compression ramp. AIAA J. 45(4), 879–889 (2007)
12. Wu, M., Martin, M.P.: Analysis of shock motion in shockwave and turbulent boundary layer
interaction using direct numerical simulation data. J. Fluid Mech. 594, 71–83 (2008)
Measurement of Particle Accelerations with the
Laser Doppler Technique
H. Nobach, M. Kinzel, R. Zimmermann, C. Tropea, and E. Bodenschatz
Abstract. An extension of the laser Doppler technique for measuring particle

acceleration is presented. The basic principles of the technique follow closely those
introduced in [11], although numerous improvements have been implemented in
the signal processing for increasing the reliability of individual estimates of parti-
cle acceleration. The main contribution of this study is to identify and quantify the
errors due to optical fringe divergence in the detection volume of the present laser
Doppler system, to introduce an appropriate experiment involving a falling wire and
to compare the acceleration measurements of the laser Doppler system to the results
of a particle tracking system with high-speed cameras in a highly turbulent flow.
Noteworthy is the fact that all measurements were performed with a commercial
off-the-shelf laser Doppler system.
1 Introduction
The Lagrangian acceleration yields deep insights into turbulence, because the
fundamental conservation equations of fluid mechanics are cast in terms of this ac-
celeration. The measurement of particle accelerations allows both the measurement
H. Nobach · E. Bodenschatz
Max Planck institute for Dynamics and Self-Organisation, Bunsenstraße 10, 37073
Göttingen, Germany
e-mail: holger.nobach@ds.mpg.de,eberhard.bodenschatz@ds.mpg.de
M. Kinzel · C. Tropea
Chair of Fluid Mechanics and Aerodynamics, Darmstadt University of Technology,
Petersenstraße 30, 64287 Darmstadt, Germany
e-mail: m.kinzel@sla.tu-darmstadt.de,
c.tropea@sla.tu-darmstadt.de
R. Zimmermann
Laboratoire de physique ENS Lyon, 46, allee d’Italie, 69364 Lyon Cedex 07, France
e-mail: robert.zimmermann@ens-lyon.fr
springerlink.com © Springer-Verlag Berlin Heidelberg 2010
272 H. Nobach et al.
of Lagrangian acceleration of the flow with the aid of small passive tracer particles
as well as the Lagrangian acceleration of inertial particles.
The methods used to date were particle tracking velocimetry (PTV) with high-
speed cameras [15] or silicon strip detectors [8, 16, 17]. Further measurements of
Lagrangian and Eulerian accelerations have been performed with particle image ve-
locimetry (PIV) [2, 3, 4, 5, 6, 7, 12, 18]. However, the laser Doppler technique can
also be used to measure fluid acceleration, as demonstrated in [9, 10, 11], offering
the advantages of a smaller measurement volume with high spatial and temporal res-
olution. Therefore, it is of principle interest to quantify the accuracy of this method
and to compare it to other alternative methods. The big challenge is to use a com-
mercial system and replace the usual signal processing by one, which extracts the
particle velocity and acceleration instead of only the velocity.
The measurement volume of a usual laser Doppler system has a diameter of the
order of 100 m. A particle crossing the measurement volume with 10 m/s needs
10 s to pass the volume. Assuming a particle acceleration of 1000 m/s2, the change
of the particle velocity while passing the measurement volume is only 0.01 m/s
which is a relative change of 0.1 %. To resolve this small change both the optical
system and the signal processing must be highly accurate.
This paper introduces the technique, the optical alignment, the signal processing
and test measurements achieving an acceptable accuracy with an off-the-shelf com-
mercial laser Doppler system. The test measurements yield yield reliable information
about the possible accuracy of a given optical system.
2 Measurement System
One aim of this investigation was to achieve acceleration measurements with a
commercial laser Doppler system with no customized or additional components,
simultaneously demonstrating the possibility of employing existing laser Doppler
systems for such measurements, while only the signal processing must be modified.
The optical setup used here is a standard, three-velocity component laser Doppler
system from TSI, although only one component was used throughout this study.
The transmitting lens had a focal length of 363 mm. The measurement volume has
a diameter of 60 m, a length of 500 m and a fringe distance of 3.73 m. The
standard signal processor provided the frequency shift signal of 40 MHz for the
Bragg cell, but was otherwise only used for down-mixing and signal conditioning
(phase-conserving amplification and 20 MHz high-pass filter to remove the pedestal
DC-component) before digital acquisition of the signals with a high-speed digital
scope card. To extract the mean velocity and the particle acceleration from the laser
Doppler signals, a special post-processing has been developed.
3 Signal Processing
Once a laser Doppler signal is digitized, the signal processing can estimate the
(mean) velocity and the particle acceleration. The corresponding signal quantities
Measurement of Particle Accelerations with the Laser Doppler Technique 273
are the instantaneous frequency and its change over the signal length. Several signal
processing methods have been developed, investigated and discussed in the past
[11, 13, 14]. Model-based estimators minimizing the L2 norm are a common tool to
achieve highest accuracy. In that case, the signal envelope must be known a priori or
estimated by the procedure. However, in this application, the envelope may strongly
fluctuate in amplitude. Only the high-frequency modulation provides reliable infor-
mation.
Therefore, a robust signal processing as introduced in [14] is preferred. As a
signal model, this method uses a complex, analytic, harmonic signal of constant
amplitude
si = cos(πγ (ti − T )2 + 2π fD (ti − T )) + j sin(πγ (ti − T )2 + 2π fD (ti − T )) (1)
with the particle arrival time T , the sampling times ti , the mid-point frequency fD ,
the frequency gradient γ and the imaginary unit j. This complex model signal s is
correlated with the measured (real) signal ŝ.
R = ∑ si ŝi (2)
i
The two parameters of the model, the Doppler frequency fD corresponding to the
particle velocity, and γ corresponding to the acceleration are iteratively optimized
such that the absolute value of the correlation |R| is maximized. The advantage of
this procedure compared to an estimator minimizing the L2 norm is that it is al-
most independent of the signals envelope. Nonetheless, it is very efficient, reaching
almost the Cramér Rao lower bound. Furthermore, the estimator is almost indepen-
dent of the absolute phase of the signal. Therefore, it is not necessary to consider or
estimate the signal phase in the signal model.
4 Fringe Distortions
The beam waists have been aligned carefully with the beam intersection point to
avoid distortions of the fringe system as shown in Fig. 1. There are two different
types of deformations caused by possible misalignments of the two beam waists.
If the beam waists lie outside the intersection point, but both on the same side and
at the same distance relative to the intersection point (Fig. 1b), the fringe spacing
changes in the direction of the optical axis, while the spacing between fringes in
the measurement direction remains almost constant. The fringe spacing becomes
additionally inhomogeneous in the measurement direction if the beam waists are
unequally located outside of the intersection point (Fig. 1c).
With a dual scanning slit beam profiler, the position of each beam waist could
be adjusted along the beam with an accuracy of about ±0.5 mm by adjusting the
distance between the transmitting fibers and the collimation lenses of the LDV
probe. Following the expression given for the local spacing between fringes in [1]
(Eq. 7.132), this leads to a (mean) fringe spacing of Δ x = 3.73 m for the present
Fig. 1 Possible distortions of the fringe system due to misalignments of the beam waists
(strongly emphasized
optical system with a maximum fringe divergence of about 0.05 %, which can fulfill
the requirement of 0.1 %. To investigate the actual fringe divergence in the measure-
ment volume and systematic errors of the measurement system, a special validation
experiment was designed, consisting of a thin wire falling through the volume.
5 Reference Measurements
To evaluate the achievable accuracy of the measurement system, a reference
experiment has been performed with a falling wire (Fig. 2), which has the accel-
eration of gravity while falling through the measurement volume. Any deviation
from this value corresponds to a systematic error caused by optical misalignments
yielding fringe distortions.
The measurement volume is scanned with the falling wire in the measurement
direction of the laser Doppler system at different locations along the optical axis
(vertical arrows in Fig. 3). A particle crossing a measurement volume with a mis-
alignment of the beam waists as in Fig. 1c will see a change in the fringe spacing
during its passage, causing a virtual acceleration and strong systematic errors. On
the other hand, distortions as in Fig. 1b will only lead to different velocity mea-
surements at different positions along the optical axis and no additional virtual ac-
celeration. Unfortunately, the velocity depends on the falling height, which cannot
be controlled with the required accuracy. Therefore, the measurement volume is ro-
tated, so that the falling wire scans the measurement volume at two lines at an angle
of ±17 about the optical axis. In that case, also this type of distortions will lead
to changes of fringe spacing along the scan line and, therefore, significant virtual
accelerations and deviations from the acceleration of gravity expected at this angle.
Fig. 2 Sketch of the falling wire experiment
Fig. 3 LDV fringes (lines), measurement volume (ellipse) and scan lines (5 vertical and 2
tilted arrows) of the falling wire experiment
The results in Fig. 4 show that the optical system has almost no systematic errors
within the measurement volume.
However, the particle velocity influences both the change of frequency due to
fringe distortions and the signal duration. Therefore, systematic errors caused by
fringe distortions scale with the square of the velocity. A conservative (worst case)
m/s2
estimation of the error bounds from the measurements above yields about 300 (m/s) 2
(or m−1 ) derived from an error of 0.18 m/s2 at a velocity of 0.025 m/s for the
measurement at +17 .
Fig. 4 Reference measurements for scanlines at a) ±90 and b) ±17
Fig. 5 Sketch of the Lagrangian Exploration Module
6 Test Measurement
A test measurement has been done in a turbulent flow, comparing the results of the
laser Doppler system with the results obtained with a high-speed particle tracking
system. The Lagrangian Exploration Module (Fig. 5) generates nearly homogeneous
Fig. 6 Normalized probability density functions taken in a turbulent flow with two different
measurement techniques
and isotropic turbulence with small mean flow [19]. The apparatus is shaped as an
icosahedron containing 140 liters of water. The flow is driven by 12 independently
controlled propellers, each at one of the vertices of the icosahedron. In Fig. 6 the nor-
malized probability density function of the measured particle accelerations for both
measurements at a Taylor micro scale Reynolds number of Rλ ≈ 195 are shown.
The acceleration values normalized with the standard deviation are obtained from
the estimated acceleration values ai with
ai − ā
anorm,i = (3)
σa
where ā denotes the empiric mean value and σa the standard deviation. The two
measurements nicely agree verifying that the laser Doppler system is a useful alter-
native to common techniques for measuring particle accelerations.
Acknowledgements
The financial support by the German Academic Exchange Service (DAAD), the National Sci-
ence Foundation (NSF) and the German Research Foundation (DFG) are gratefully acknowl-
edged. Furthermore, the authors would like to thank K. Chang, A. Kubitzek, B. Mühlhölzer,
N. Ouellette, D. Wolf and H. Xu for their contributions to this work.
References
1. Albrecht, H.-E., Damaschke, N., Borys, M., Tropea, C.: Laser Doppler and Phase
Doppler Measurement Techniques. Springer, Berlin (2003)
2. Christensen, K.T., Adrian, R.J.: Measurement of instantaneous Eulerian acceleration
fields by particle-image velocimetry: method and accuracy. Exp. in Fluids 33, 759–769
(2002)
3. Dong, P., Hsu, T.-Y., Atsavapranee, P., Wei, T.: Digital particle image accelerometry.
Exp. in Fluids 30(6) (2001)
4. Jakobsen, M.L., Dewhirst, T.P., Greated, C.A.: Particle image velocimetry for predictions
of acceleration fields and force within fluid flows. Meas. Sci. Technol. 8, 1502–1516
(1997)
5. Jensen, A., Pedersen, G.K.: Optimization of acceleration measurements using PIV. Meas.
Sci. Technol. 15, 2275–2283 (2004)
6. Jensen, A., Pedersen, G.K., Wood, D.J.: An experimental study of wave run-up at a steep
beach. J. Fluid Mech. 486, 161–188 (2003)
7. Jensen, A., Sveen, J.K., Grue, J., Richon, J.-B., Gray, C.: Accelerations in water waves
by extended particle image velocimetry. Exp. in Fluids 30, 500–510 (2001)
8. La Porta, A., Voth, G.A., Crawford, A.M., Alexander, J., Bodenschatz, E.: Fluid particle
accelerations in fully developed turbulence. Nature 409, 1017–1019 (2001)
9. Lehmann, B., Helbig, A., Hassa, C.: LDA method to measure the acceleration of particles
and the curvature radii of particle trajectories. In: Proc. 5th Int. Symp. on Appl. of Laser
Techn. to Fluid Mechanics, Lisbon, Portugal (1990), paper 22.3
10. Lehmann, B., Helbig, J.: Local acquisition of mean and turbulent fluid acceleration in
highly turbulent flow by the means of laser-Doppler velocimetry. In: Proc. 10th Int.
Symp. on Appl. of Laser Techn. to Fluid Mechanics, Lisbon, Portugal (2000)
11. Lehmann, B., Nobach, H., Tropea, C.: Measurement of acceleration using the laser
Doppler technique. Meas. Sci. Technol. 13(9), 1367–1381 (2002)
12. Liu, X., Katz, J.: Instantaneous pressure and material acceleration measurements using a
four-exposure PIV system. Exp. in Fluids 41, 227–240 (2006)
13. Nobach, H.: Gegen Wavelets & Co. — Vorteile klassischer Signal- und Daten-
verarbeitungsverfahren in der optischen Strömungsmesstechnik. Shaker, Aachen. TU
Darmstadt, Zugl., Habil (2007) ISBN 978-3-8322-6564-9
14. Nobach, H., Kinzel, M., Tropea, C.: Measurement of Lagrangian acceleration in turbu-
lent flows using the laser Doppler technique. In: Proc. of SPIE Optical Methods of Flow
Investigation, vol. 6262, pp. 1–8 (2006)
15. Ouellette, N.T., Xu, H., Bodenschatz, E.: A quantitative study of three-dimensional
Lagrangian particle tracking algorithms. Exp. in Fluids 40 (2006)
16. Voth, G.A., La Porta, A., Crawford, A.M., Alexander, J., Bodenschatz, E.: Measure-
ment of particle accelerations in fully developed turbulence. J. of Fluid Mechanics 469,
121–160 (2002)
17. Voth, G.A., La Porta, A., Crawford, A.M., Bodenschatz, E., Ward, C., Alexander, J.: A
silicon strip detector system for high resolution particle tracking in turbulence. Rev. Sci.
Instrum. 72(12), 4348–4353 (2001)
18. Yang, T.S., Shy, S.S.: Two-way interaction between solid particles and homogeneous
air turbulence: particle settling rate and turbulence modification measurements. J. Fluid
Mech. 526, 171–216 (2005)
19. Zimmermann, R.: The Lagrangian exploration module. Master’s thesis, Georg-August-
Universität Göttingen, Fachbereich Physik (2008)
A Novel Numerical Method for Turbulent,
Two-Phase Flow
A. Pecenko and J.G.M. Kuerten
Abstract. A novel approach to the simulation of isothermal, turbulent two-phase

(liquid/vapour) flow is presented. The two-phase nature of the flow is modeled by
means of a diffuse-interface concept and of the Korteweg tensor of capillary forces
at the interface, so that a single system of compressible Navier-Stokes equations
can be written for the whole flow domain. A Van der Waals equation of state is
also included to account for the variation of pressure with density at the given value
of temperature. After explanation of a stable numerical method, results of a classic
benchmark problem are shown. Next, subgrid terms related to the nonlinear pressure
and capillary terms are studied by means of an a priori analysis based on DNS
results, and a subgrid model for these terms is proposed.
1 Introduction
The numerical simulation of multiphase flows requires a careful treatment of those
physical and mathematical issues that are absent in incompressible, single-phase
flows. Mainly, new problems arise from the presence of thin phase-transition lay-
ers called interfaces. Physically, they represent regions of the flow domain where,
within a small but finite thickness, density, temperature, pressure, viscosity and other
quantities vary rapidly but smoothly between their respective single-phase values.
Even in the absence of heat transfer, and therefore of any phase change, these
boundaries may undergo deformations and move to other locations within the flow
A. Pecenko
Eindhoven University of Technology, Dept. of Mechanical Engineering, Eindhoven,
The Netherlands
e-mail: a.pecenko@tue.nl
J.G.M. Kuerten
Eindhoven University of Technology, Dept. of Mechanical Engineering, Eindhoven,
The Netherlands
e-mail: j.g.m.kuerten@tue.nl
280 A. Pecenko and J.G.M. Kuerten
domain, in a way that depends on the particular problem. From the perspective of
numerical computation, it is evident to expect a substantial complication of the sim-
pler single-phase case. Without discussing the various methods and approaches de-
veloped for multiphase flows in the past fourty years, we remark that our formulation
of the problem does not require any special treatment for the interfacial zone, since
we make use of a diffuse interface concept [1], and of the Korteweg capillary tensor
in the Navier-Stokes equations [5]. This approach allows to overcome some typical
disadvantages of sharp interface methods, where fluid properties are discontinuous
over each phase boundary and the surface tension must be included in some way
in the system of equations. In the method we present here, instead, all variables are
continuous everywhere, including the interfacial zones, and surface tension forces
are incorporated in the model in such a way that no weak formulation of the set of
compressible Navier-Stokes equations is needed.
The contents of this paper are as follows. In Sect. 2 the governing equations are
shown. In Sect. 3 the numerical method and simulation results of a laminar problem
are briefly described. In Sect. 4 we introduce the case of turbulent two-phase flow,
and we propose a model for the subgrid terms related to the presence of the interface
stemming from filtering the equations. The model is assessed in an a priori analysis
based on the simulation shown in Sect. 3.
2 The Set of Governing Equations

From a thermodynamical point of view, the notion of phase boundary as a diffuse
interface is related to the excess of Helmholtz free energy at the interface, which can
be expressed (see [2] [3]) by means of the local density gradient. A minimization
procedure on the free energy functional for the equilibrium state of a two-phase
single-component fluid leads to the following expression for the above mentioned
Korteweg tensor [1]

1
T = −p + K ρ ∇2ρ + K|∇ρ |2 I − K∇ρ ⊗ ∇ρ , (1)
2
where p is the pressure, ρ is the density and I denotes the identity tensor. The
energy-gradient coefficient K plays the role of a capillary coefficient related to sur-
face tension, and, in general, is a function of density and temperature. In the present
work we ignore the dependence on temperature, since we only consider isothermal
situations. The full system of governing equations reads for a single-component
fluid
ρt + ∇ · (ρ u) = 0
(ρ u)t + ∇ · (ρ uu) = ∇ · (d + T) (2)
p = p(ρ ) ,
u denoting the velocity, d the usual viscous stress tensor for a Newtonian fluid, and
p(ρ ) a suitable equation of state. At a temperature slightly below its critical value,
A Novel Numerical Method for Turbulent, Two-Phase Flow 281
which we consider in this work, the Van der Waals equation of state provides a
sufficiently accurate description in both vapour and liquid phases of the fluid.
3 Numerical Method and Results

System (2) is highly nonlinear, due to the equation of state and the capillary forces.
Moreover, the third-order derivatives of density in the Korteweg tensor cause a dis-
persive effect on the solution. Furthermore, the nonmonotonic behaviour of Van der
Waals isotherms below the critical temperature brings a source of ill-posedness into
the problem.
To obtain stable numerical solutions of system (2) we generalize the transfor-
mation of variables proposed in [4] for the Riemann problem of one-dimensional,
isothermal propagation of phase boundaries in solids in a Lagrangian frame. Thus,
we write
ρ = ρ̂ ρ u = ρ.u − ν0 (ρ̂ )∇ρ̂ . (3)
where ν0 is an arbitrary parameter that can be chosen in such a way that the diagonal
term of the matrix of transformed capillary coefficients

ν0 (ρ̂ ) 4
K0 (ρ̂ ) = K(ρ̂ ) − μ (ρ̂ ) − ν0 (ρ̂ )ρ̂
ρ̂ 2 3
is identically equal to zero, so that the dispersive third-order derivatives are canceled
out of the numerical scheme.
For space discretization, we have used a central finite-volume scheme, globally
accurate at the second order. A TVD-Runge Kutta third-order accurate time integra-
tion scheme has been adopted to control oscillations in the unstable region of the
solution domain.
(a) (b)
Fig. 1 Retraction of an ellipsoidal drop surrounded by quiescent vapour, under the effect
of capillary forces, represented on a plane normal to the other Cartesian coordinate. a initial
state. b final equilibrium state. The interface is represented by means of density isolines. Time
t is expressed in arbitrary units
In Fig. 1 we show results of a benchmark test case for two-phase flow simulation.
The initial condition consists of a liquid drop of ellipsoidal shape surrounded by
vapour in absence of external forces and at rest. Due to surface tension the drop
retracts to a stationary spherical shape, in which the relation between the pressure
difference Δp and the radius R is given by the Laplace equation Δp = cγ /R where
c equals 1 in two dimensions and 2 in three dimensions, and γ is the surface tension
(constant at the assigned temperature).
In Fig. 2a the theoretical value of surface tension for this problem is compared
with the value computed numerically by using the Laplace equation for different
radii of the drop in two and three dimensions. Fig. 2b shows a study of the grid
convergence of the numerical scheme. The pressure difference Δp at equilibrium is
shown in Fig. 3.
1.1
1.05
γ/γth
2D
3D
0.95
0 0.5 1 1.5 2 2.5 3
R −7
x 10 (a) (b)
Fig. 2 a ratio of numerical and theoretical value of surface tension coefficient for different
radii of a static drop. The error bars represent the uncertainty in the drop radius. Abscissa is
expressed in meters. b pressure as a function of spatial coordinate x on a line passing through
the center of the drop, for three different grid spacings in two dimensions: (+) 4002 , (•) 8002 ,
() 16002 nodes
4 A Priori Analysis of the Interfacial Subgrid Terms

In Sect. 3, we have briefly described a method of solution for the two-phase
isothermal set of equations (2), and we have shown, as an example of numerical
application, a classic test in the laminar flow regime.
If Large Eddy Simulation is applied for turbulent two-phase flow, new subgrid
terms stem from the filtering operation of the nonlinear pressure and capillary terms.
This implies that, apart from the well-known subgrid stress tensor for the case of a
single-phase flow, a subgrid pressure term and several subgrid capillary terms ap-
pear. To devise a model for these unknown quantities, the first step is to carry out an a
priori study of the subgrid terms based on results from Direct Numerical Simulation.
Here, we present the results that we have obtained for the well-assessed case of the
drop retraction. Although this test problem is not turbulent (Re ∼ 102 at the most),
we postulate that the subgrid terms related to the pressure and capillary terms do
(a) (b)
Fig. 3 Retraction of a drop in absence of external forces. a steady solution for the pressure
as a function of the Cartesian coordinate x on a line passing through the center of the drop.
The spikes are physical and due to the Korteweg tensor across the interfacial film. b pressure
field at steady state as a function of the coordinates x, y
not strongly depend on the type of flow. The analysis has been conducted as follows.
First, the exact values of all subgrid terms have been evaluated. Particularly, we have
focused our attention on the subgrid term for the pressure psg = p(ρ ) − p(ρ̄ ) (the
overline denotes the filtered quantities), and on the subgrid terms for the capillary
terms. With the convenient choice K(ρ ) ∼ ρ −1 , the term with the highest order of
derivative in (1) becomes linear, and the capillary subgrid terms arise from filtering
the nonlinear tensor (1/2)K(ρ )|∇ρ |2I − K(ρ )(∇ρ ⊗∇ρ ). Note that the components
of this tensor are made of terms of the form K(ρ )ρxi ρx j . Hence, we adopt for the
corresponding subgrid contributions the formal notation [K(ρ )(∇ρ ⊗ ∇ρ )]sg , which
reads
[K(ρ )(∇ρ ⊗ ∇ρ )]sg = K(ρ )ρxi ρx j − K(ρ̄ )(ρ̄ )xi (ρ̄ )x j .
The pressure and capillary subgrid terms have been compared with the values given
by the similarity model (SM) and the gradient model (GM). The SM has provided
the best approximation for the capillary subgrid terms, while the GM has given the
best result for the pressure subgrid term. The difference between the exact and the
modeled term has been treated with an extra subgrid model that we call capillarity
model (CM):
sg + Cp (Ca) K(ρ̄ )|∇ρ̄ |

psg = pGM 2 2
(4)
[K(ρ )(∇ρ ⊗ ∇ρ )]sg = [K(ρ )(∇ρ ⊗ ∇ρ )]SM
sg + Cc (Ca) K(ρ̄ )(∇ρ̄
2
⊗ ∇ρ̄ ) , (5)
where Ca ∼ Δ /L is a nondimensional parameter of the simulation called the Cahn

number, representing the effect of the filter on the interface length scale, and Cp , Cc
are the constants of the model. Note that the CM contribution to the pressure subgrid
term has the same form as the components of the isotropic part of the capillary
5
x 10
2.5
4
exact x 10
2 GM 12
DCM exact
GM+DCM 10 GM
1.5 DCM
8 GM+DCM
1 6
psg
0.5 4
psg
2
0
0
−0.5 −2
−4
−1
−6
0 0.2 0.4 0.6 0.8 1 2 3 4 5 6
x −6
x 10 (a) x x 10
−7
(b)
Fig. 4 A priori analysis of the subgrid pressure term as a function of the Cartesian coordinate
x on a line passing through the center of the drop. a comparison between the exact subgrid
term and the models. b the sum of the gradient-model (GM) and of the capillarity-model with
dynamically-computed model constants (DCM) provides a much better approximation than
the two models separately
4 5
x 10 x 10
4 3
exact exact
3.5 SM 2.5 with NDCM
DCM with DCM
3 SM+DCM 2
sg
2.5 1.5
p +(K∇ρ⊗∇ρ)
x sg
2 1
(Kρ2)
1.5 0.5
sg
1 0
0.5 −0.5
0 −1
−0.5 −1.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x 10
−6
(a) x −6
x 10 (b)
Fig. 5 a comparison between the exact subgrid term of and the models, as functions K(ρ )ρx2
of x. b comparison between the exact and the modeled values of the subgrid term resulting
from the sum of the pressure subgrid term with the subgrid capillary terms appearing under
the divergence operator component ∂∂x in the x−momentum equation, as functions of x
tensor (1/2)K(ρ̄ )|∇ρ̄ |2 I. We have evaluated the constants Cp , Cc by a least-square

minimization of the quantities (CM denoting the last terms on the r.h.s. of (4) (5))
psg − [psg ]GM+CM + [K(ρ )(∇ρ ⊗ ∇ρ )]sg − {[K(ρ )(∇ρ ⊗ ∇ρ )]sg }SM+CM
that appear in each of the three scalar momentum equations (the superscripts being
a short formal notation for the sum of the modeled subgrid contributions to pressure
and capillary terms respectively). The constants can be calculated with either non-
dynamical (NDCM) or dynamical (DCM) procedure.
Results for the pressure are shown in Fig. 4, and refer to an instant of time
taken at the beginning of the simulation, when the magnitude of the velocity is
4
x 10
1.5
1 0
−1000
0.5
−2000
(Kρxρy)sg
(Kρxρy)sg
0
−3000
−0.5 −4000
exact
exact
−5000 SM
−1 SM
NDCM NDCM
SM+NDCM −6000 SM+NDCM
−1.5 DCM DCM
−7000
0 0.2 0.4 0.6 0.8 1 2.5 3 3.5 4 4.5
x −6
x 10 (a) x x 10
−7
(b)
Fig. 6 a comparison between the exact subgrid term of K(ρ )ρx ρy and the models. b the
CM model with dynamically-computed model constants largely overestimates the subgrid
contribution at the interface
maximum. The use of dynamically-computed constants appears to contribute to a

better approximation of / the exact
0 subgrid term. This trend is also found for the
capillary subgrid term K(ρ )ρx2 sg at the same instant of time (Fig. 5a), and for the
sum {psg + [K(ρ )(∇ρ ⊗ ∇ρ )]sg } (Fig. 5b). For other capillary subgrid terms, such
as [K(ρ )ρx ρy ]sg , the CM model with non-dynamical calculation of the model con-
stants seem to provide a better approximation, again at the selected instant of time
(Fig. 6). Currently, we are working on the a posteriori validation of the CM model
for the same flow problem. Next, we will test the model on other turbulent, two-
phase flow examples and, eventually, we will apply it to the LES of high Reynolds
two-phase flows.
Acknowledgements. This research is supported by the Dutch Technology Foundation STW,

applied science division of NWO and the Technology Program of the Ministry of Economic
Affairs of the Netherlands.
References
1. Anderson, D.M., McFadden, G.B., Wheeler, A.A.: Diffuse-interface methods in fluid me-
chanics. Annu. Rev. Fluid Mech. 30, 139–165 (1998)
2. Cahn, J.W., Hilliard, J.E.: Free energy of a nonuniform system. I. Interfacial free energy.
J. Chem. Phys. 28(2), 258–267 (1958)
3. Cahn, J.W.: Free energy of a nonuniform system. II. Thermodynamic basis. J. Chem.
Phys. 30(5), 1121–1124 (1959)
4. Cockburn, B., Gau, H.: A model numerical scheme for the propagation of phase transitions
in solids. SIAM J. Sci. Comput. 17(5), 1092–1121 (1996)
5. Korteweg, D.J.: Sur la forme que prennent les équations du mouvement des fluides si l’on
tient compte des forces capillaires causées par des variations de densité considérables mais
continues. Arch. Néerl. Sci. Exactes Nat. Série II, Tome VI, 1–24 (1901)
Modeling of High Reynolds Number Flows with
Solid Body Rotation or Magnetic Fields
Annick Pouquet, Julien Baerenzung, Jonathan Pietarila Graham, Pablo Mininni,

Hélène Politano, and Yannick Ponty
Abstract. We present two models for turbulent flows with periodic boundary
conditions and with either rotation, or a magnetic field in the magnetohydrodynam-
ics (MHD) limit. One model, based on Lagrangian averaging, can be viewed as
an invariant-preserving filter, whereas the other model, based on spectral closures,
generalizes the concepts of eddy viscosity and eddy noise. These models, when
used separately or in conjunction, may lead to substantial savings for modeling high
Reynolds number flows when checked against high resolution direct numerical sim-
ulations (DNS), the examples given here being run on grids of up to 15363 points.
1 The Lagrangian Model

Turbulence modeling, in engineering as well as for geo- and astrophysics, is a
needed approach even though the power of computers is ever increasing, simply
because the number of excited modes in such flows vastly exceeds the capacity of
computers in the foreseeable future. As the Reynolds number of DNS grows, tests
can be devised which study in detail the properties of such models and thus allow
Annick Pouquet · Julien Baerenzung
NCAR, PO Box 3000, Boulder CO-80307, USA
e-mail: pouquet@ucar.edu,baeren@ucar.edu
Jonathan Pietarila Graham
MPI für Sonnensystemforschung, 37191 Katlenburg, Germany
e-mail: jpietarilagraham@mailaps.org
Pablo Mininni
Departamento de Fı́sica, Facultad de Ciencias Exactas y Naturales, Universidad de Buenos
Aires, Ciudad Universitaria, 1428 Buenos Aires, Argentina, and NCAR
e-mail: mininni@ucar.edu
Hélène Politano · Yannick Ponty
Observatoire de la Côte D’Azur, Nice, France
e-mail: politano@obs-nice.fr,ponty@obs-nice.fr
288 A. Pouquet et al.
(a)
Fig. 1 A slice of a region of space in which regions of energy transfer smaller than 1%
its mean are shown in black for (left) a LAM model for fluids on a grid of 2563 points and
(middle) a DNS of the Navier-Stokes equations on a 10243 grid at the same Reynolds number:
LAM has a substantial reduction in energy transfer, and thus of dissipation, leading to an
energy accumulation at small scales, as shown in its energy spectrum. Right: kinetic energy
spectra for a 15363 DNS of MHD (solid line), a 5123 LAMHD (dash) with filter at kα = 18
(vertical dash line), and a 5123 LAM (dots), in the latter case with no magnetic field (b ≡ 0
at all times) but otherwise identical LES run. For k ∈ [5, 40], LAMHD reproduces well the
scaling of the DNS, with no bottleneck. For k close to the filter (k ∈ [kα /2, kα ]), a k0.5 power
law (gray line) obtains for fluids using LAM, corresponding to the energy accumulation at
small scale for lack of dissipation, whereas it is not present for LAMHD; the magnetic energy
has no accumulation of energy at small scale either [8]
improvements, or else generalizations to handle more complex flows, for example

taking into account anisotropies in the presence of either rotation or magnetic fields.
The first model we have considered can be constructed as a particular filter of
small scales [1] that preserves invariants of the ideal case but in a different norm
(H1 instead of L2 ). It is called the alpha or Lagrangian averaged model (LAM) [2]-
[4] and has been tested in a variety of conditions both in two and three dimensions
[5]–[11] for Navier-Stokes and MHD. This model can be viewed as a quasi-DNS
insofar as it does not introduce by hand a model of the physical effects of the small
scales that are neglected, but rather it preserves the Hamiltonian structure of the un-
derlying equations. However, when leaving sufficient room between the filter length
α of the model and the smallest resolved scale in the computations, a peculiar feature
is observed, namely that small scales are insufficiently dissipated due to a tendency
of the model to create regions in space where the normalized energy transfer ε is !
negligible. This is shown in Fig. 1 mapping ε when below 1% its mean (respective
filling factors of regions with negligible ε are 0.26 and 0.67 for DNS and LES-LAM).
Note the larger and more numerous patches of negligible transfer in LAM (left) com-
pared to the DNS (middle), leading to an energy accumulation at small scale in the
energy spectrum (right, dotted line) before the α cut-off with a positive slope corre-
sponding to a “bottleneck” [9]. This bottleneck is absent in MHD: we observe in Fig.
1 (right) the agreement between the spectra for the MHD-DNS and LAMHD both
above and below the filter scale (α = 2π /18); this is probably due to the nonlocality
of nonlinear interactions in MHD, and this lack of accumulation of energy observed
in the energy spectrum at large wavenumbers represents a marked improvement for
Modeling of High Reynolds Number Flows with Solid Body Rotation 289
the Lagrangian model in MHD when compared to the Navier-Stokes case [11]. Thus,
LAMHD is able to reproduce a DNS on a grid of 15363 points, with savings in CPU
and memory usage by a factor of 6 in linear resolution. In fact, one can pursue the
DNS run with LAMHD up to times unreachable with reasonable resources using a
DNS (it would take ∼ 1.7 × 106 CPU hours with present day computers) [8]. In so
doing, we observe in Fig. 2 that equipartition between kinetic and magnetic energy,
imposed at t = 0, is broken in time, with the latter being enhanced by nonlinear inter-
actions; note that the DNS has a small ideal phase where energy is almost conserved
and no kinetic-magnetic exchanges take place globally, whereas both the LAMHD
and the under-resolved DNS depart from equipartition almost immediately. When
considering the total enstrophy (right), the under-resolved run overestimates it be-
cause of an accumulation of small-scale excitation not being properly dissipated,
whereas LAMHD is much closer to the DNS dat! a (with a slight under-estimation
of it near the peak). LAMHD should thus prove quite useful, since it is also known
to reproduce well the generation of magnetic fields by velocity gradients (dynamo
effect) and the inverse cascade of magnetic helicity, as well as small-scale properties
such as the variation of the cancellation exponent of the current density in two space
dimensions [12].
(a) (b)
Fig. 2 Temporal evolution of energy (left) and < J 2 + ω 2 > (right), where ω = ∇ × v and
J = ∇ × b, with v and b the velocity and magnetic field; total energy on top, kinetic EV and
magnetic EM energies below, with EM ≥ EV ∀t. The thick solid line and dots are DNS on
grids of 15363 and 2563 points respectively, and the dash line is LAMHD on 2563 points,
all with the same Reynolds numbers. Only the lower resolution computations are performed
beyond the peak of dissipation
2 Spectral Models for Rotating Flows

The second model we test in this paper is based on a two-point closure of turbulence,
the Eddy Damped Quasi Normal Markovian or EDQNM (see, e.g., [13]). In this ap-
proach, eddy viscosity and eddy noise are included, and the model allows for taking
into account non-Kolmogorovian energy spectra to assess these transport coeffi-
cients [14]–[19]; the model builds on the so-called Chollet-Lesieur formulation of
spectral eddy viscosity [20] (hereafter, CL) which is also tested separately against
Fig. 3 Left: time variation (in units of the eddy turn-over time) of the isotropy coefficient
IC = < |v1 |2 > / < |v2 |2 >, with v1, 2 the velocity projected onto e1 = k × z and
e2 = k × e1 , k being the wavenumber and z the axis of rotation; solid line: LES (643 grid
points); dash: full DNS (2563 points); triangle: full DNS data downgraded to 643 points;
Rossby number of 0.03 with a non-helical forcing at large scale and a non-helical spectral
model [18]. Note the DNS/LES agreement and the progressive return to isotropy. Right: En-
ergy spectrum of the error [Emodel − EDNS ]/EDNS when comparing the DNS on a grid of
15363 points for a rotating flow with Ro ∼ 0.03 with either an under-resolved DNS (+) on a
1603 grid, with the Chollet-Lesieur model [20] (circles) and the LES we propose here [14, 18]
(solid line), both on grids of 963 points. The error, exponential at high k for the under-resolved
run, is the lowest almost consistently for the spectral model
DNS at the same Reynolds number Rv and down to Rossby numbers of Ro ∼ 0.03.
Moreover, helical contributions to the transport coefficients, following the helical
EDQNM developed in [21], can be incorporated in the model; these contributions
depend on the helicity spectrum at small scale (where the helicity HV is defined as
usual as v · ω with ω = ∇ × v the vorticity). For example, HV (k) being the helicity
spectral density, one can write, in the temporal variation of the energy spectrum, the
small-scale contributions as:
∂t E(k) ∼ −2k2 E(k)[ν + νturb ] − 2k2 H(k)ν̃turb ; (1)

Eq. (1) uses a short-hand but hopefully self-explanatory notation to bring the struc-
ture of the model (see [14] for details), and it omits both the resolved scale contribu-
tions and the eddy-noise
contributions for simplicity. The classical EDQNM eddy
viscosity νt (k,t) ∼ > f1 (k, p, q)E(q)d pdq depends

on an integral of the energy
spectrum in the small scales (symbolized by > ) and represents the drain of energy
due to the unresolved sub-grid scales; similarly, ν̃t (k,t) ∼ > f2 (k, p, q)HV (q)d pdq
gives the contribution of small-scale helicity (with k = p+ q due to the convolution).
When modeling rotating flows (here, with no magnetic fields) [22], one intro-
duces an obvious external anisotropy in the problem. However, anisotropic models,
as those developed using for example extensions of EDQNM to such flows (see e.g.
[23]-[25] for recent works) are costly since they now depend on both the parallel
and perpendicular (referring to the rotation axis) components of the velocity. On the
other hand, one can remark that in an LES approach, one models the small scales
which can recover some degree of isotropy since the scale-dependent Rossby num-
ber Ro = v /Ω with v the velocity at scale , and Ω the rotation rate, gets larger
as → 0. In Fig. 3 (right) is given the temporal evolution of the isotropy coefficient
IC (see caption for definition) for the full DNS for a run forced with a non-helical
velocity field (the Taylor-Green flow), the DNS data being downgraded to the grid
resolution of the LES and the LES using the spectral model we propose [19]; this
coefficient, of unit value for full isotropy (rotation is introduced in the run at t ∼ 90
after the flow has settled to a turbulent state), begins to increase substantially once
the inverse cascade of energy builds up, for t ≥ 110, and then decreases under the
influence of the small-scale cascade that restores isotropy to some extent. Further-
more, Fig. 3 shows that the LES, when compared to the DNS downgraded to the LES
run, reproduces this result quite accurately. This means that, at least at the moderate
Rossby number of these computations, down to Ro ∼ 0.03 and micro-Rossby num-
ber ωrms /Ω ≈ 1 (with ωrms the rms value of the vorticity), an isotropic approach is
a workable solution for modeling such flows since the small scales are sufficiently
isotropic. Whether such an agreement will persist at lower Rossby numbers is left
for future investigations, but since Ro ≈ 0.1 in the atmosphere, this spectral model
may prove useful in this context. Fig. 3 (right) confirms this result, by plotting the
energy spectral error for three models (see caption).
Noting that it has been found recently that helicity plays an important role in
the dynamics of turbulent flows in the presence of rotation [17], a point that may
relate to a simplified dynamics of tornadoes, we test further the possibility of using
isotropic spectral models for rotating flows by performing a comparison against a
massive DNS of a rotating helical flow, on a grid of 15363 points; note that more than
700,000 CPU hours were used for this second large DNS run; the Beltrami forcing is
an ABC flow [26] set at wavenumber kF = 7, leaving room for both a direct cascade
and an inverse cascade to take place. Among the many novel features of such a flow
[27], we display here a comparison on two diagnostics, see Fig. 4: when examining
the temporal evolution of the total energy (left) and the energy spectra averaged
over a few turn-over times (right), we see that the LES model (including for this
fully helical case,! the helical contributions to eddy viscosity and eddy noise [14])
performs best, and the under-resolved DNS performs worst, in particular because
of an accumulation of energy at both small and large scales. The Chollet-Lesieur
model obtains a growth rate for the energy in the inverse cascade quite close to that
in the DNS but is somewhat more dissipative, whereas the spectral model behaves
better energetically. Similarly, for the energy spectra, the spectral model performs
best. This data thus provides an unambiguous display of the added value of a LES
when contrasted either to under-resolved DNS or to simple eddy-viscosity models
in order to approach the dynamics of complex turbulent flows, with here a huge gain
in resolution (all LES are performed on a grid of [1536/16]3 = 963 points).
Fig. 4 Helical rotating flow: comparisons between a DNS on a grid of 15363 points (solid
line), an under-resolved 1603 DNS (+), the LES-CL model (grey line), and the helical spec-
tral model (LES-PH, dash) [14]; the two LES runs use grids of 963 points. The Reynolds
number is 5600 and the Rossby number is 0.06. Left: Energy as a function of time; note the
unphysical substantial increase in the case of the under-resolved run (+). Right: Energy spec-
trum averaged from t = 20 to t = 30; again, the under-resolved run clearly underperforms
the LES models, and LES-PH is the model closest to the DNS at a substantial savings in
computational cost compared to the DNS
3 Conclusion
The increase of power in computers, with the petascale initiative and beyond, does
not mean one need not worry about modeling of turbulent flows, quite to the con-
trary. With increased capability, one will tackle more complex problems with non
trivial geometries and micro-physics, as needed in a comprehensive approach to
climate, weather, and space physics modeling for example. But because realistic pa-
rameters are still well out of range, we can foresee complementary roles for DNS
and LES, together with experiments and observations: (i) analysis of the dynamics
of complex turbulent flows with highly resolved DNS, followed by (ii) verification
and amelioration of models against such DNS runs, the models being used either
alone or in a combined fashion (see, e.g., [28] using both LAMHD and LES-PH
for the dynamo problem at low magnetic Prandtl number); then (iii) exploration
of parameter space with such models, and finally (iv) starting again the cycle with
new DNS runs! . Such a cyclic approach relies on Moore’s law of doubling of pro-
cessor speed every ≈ 18 months, leading to a doubling of resolution in a 3D run
every ≈ 6 years. In the specific cases mentioned in this paper, the savings at given
Reynolds (and magnetic Prandtl or Rossby) numbers, are already substantial since a
LES run on a grid of 963 points reproduces satisfactorily a DNS run that cost almost
104 times more. Such models thus should prove useful in exploring parametrically
dynamical regimes of geophysical and astrophysical turbulence in the presence of
rotation and/or magnetic fields in a variety of conditions such as they arise in nature.
The two large runs on 15363 grids were performed at NCAR – which is sponsored
by NSF – using special allocations for computer time.
References
1. Montgomery, D., Pouquet, A.: An alternative interpretation for the Holm “alpha” model.
Phys. Fluids 14, 3365–3366 (2002)
2. Chen, S.Y., Holm, D.D., Margolin, L.G., Zhang, R.: Direct numerical simulations of the
Navier-Stokes alpha model. Physica D 133, 66–83 (1999)
3. Foias, C., Holm, D.D., Titi, E.S.: The Navier-Stokes-alpha model of fluid turbulence.
Physica D 152-153, 505–519 (2001)
4. Holm, D.D.: Lagrangian averages, averaged Lagrangians, and the mean effects of fluc-
tuations in fluid dynamics. Chaos 12, 518–530 (2002)
5. Chen, S., et al.: A connection between the Camassa-Holm equations and turbulent ßows
in channels and pipes. Phys. Fluids 11, 2343–2353 (1999)
6. Nadiga, B., Shkoller, S.: Enhancement of the inverse-cascade of energy in the 2D
Lagrangian-averaged Navier-Stokes equations. Phys. Fluids 13, 1528–1531 (2001)
7. Mininni, P., Montgomery, D., Pouquet, A.: Numerical solutions of the three-dimensional
MHD alpha model. Phys. Fluids 17, 035112 (2005)
8. Pietarila Graham, J., et al.: Inertial Range Scaling, Kármán Theorem and Intermittency
for Forced and Decaying Lagrangian Averaged MHD in 2D. Phys. Fluids 18, 045106
(2006)
9. Pietarila Graham, J., et al.: Highly turbulent solutions of LANS-α and their LES poten-
tial. Phys. Rev. E 76, 056310 (2007)
10. Pietarila Graham, J., et al.: Three regularization models of the Navier-Stokes equations.
Phys. Fluids 20, 035107 (2008)
11. Pietarila Graham, J., Mininni, P., Pouquet, A.: The Lagrangian-averaged model for
MHD turbulence and the absence of bottleneck. Phys. Rev. E (2009) (to appear),
arxiv/0806.2054v1
12. Graham, J., Mininni, P., Pouquet, A.: Cancellation exponent and multifractal structure in
Lagrangian averaged magnetohydrodynamics. Phys. Rev. E 72, 045301(R) (2005)
13. Orszag, S., Kruskal, M.: Formulation of the theory of turbulence. Phys. Fluids 11, 43–60
(1968)
14. Baerenzung, J., et al.: Spectral Modeling of Turbulent Flows and the Role of Helicity.
Phys. Rev. E 77, 046303 (2008)
15. Baerenzung, J., et al.: Spectral Modeling of Magnetohydrodynamic Turbulent Flows.
Phys. Rev. E 78, 026310 (2008)
16. Mininni, P., Alexakis, A., Pouquet, A.: Scale interactions and scaling laws in rotating
flows at moderate Rossby numbers and large Reynolds numbers. Phys. Fluids 21, 015108
(2009)
17. Mininni, P., Pouquet, A.: Helicity cascades in rotating turbulence. Phys. Fluids (2009)
(to appear); See also arxiv:0809.0869
18. Baerenzung, J., et al.: Modeling of rotating flows at moderate Rossby numbers. Phys.
Rev. E (2009) (submitted); See also arXiv:0812.1821v1
19. Baerenzung, J., et al.: Modeling of rotating flows with helicity (2009) (in preparation)
20. Chollet, J.P., Lesieur, M.: Parametrization of small-scale three-dimensional isotropic tur-
bulence using spectral closures. J. Atmos. Sci. 38, 2747–2757 (1981)
21. André, J.C., Lesieur, M.: Influence of Helicity on the Evolution of Isotropic Turbulence
at High Reynolds Number. J. Fluid Mech. 81, 187–207 (1977)
22. Sagaut, P., Cambon, C.: Homogeneous Turbulence Dynamics. Cambridge Univ. Press,
Cambridge (2008)
23. Cambon, C., Scott, J.F.: Linear and nonlinear models of anisotropic turbulence. Ann.
Rev. Fluid Mech. 31, 1–53 (1999)
24. Cambon, C., Rubinstein, R., Godeferd, F.S.: Advances in wave turbulence: rapidly rotat-
ing flows. New J. Phys. 6(73), 1–29 (2004)
25. Cui, G.X., et al.: A new subgrid eddy-viscosity model for large-eddy simulation of
anisotropic turbulence. J. Fluid Mech. 582, 377–397 (2007)
26. Childress, S., Gilbert, A.: Stretch, Twist, Fold: The Fast Dynamo. Springer, Heidelberg
(1995)
27. Mininni, P., Pouquet, A.: Persistent cyclonic structures in self-similar turbulent flows.
Phys. Rev. Lett. arXiv:0903.2294 (2009) (submitted)
28. Ponty, Y., et al.: Critical magnetic Reynolds number for dynamo action as a function of
magnetic Prandtl number. Phys. Rev. Lett. 94, 164502 (2005)
Direct Numerical Simulation of Buoyancy
Driven Turbulence inside a Cubic Cavity
R. Puragliesi, A. Dehbi, E. Leriche, A. Soldati, and M. Deville
Direct numerical simulation (DNS) of thermally driven turbulent flows inside a fully
confined eclosure is performed using state-of-the-art Chebyshev pseudo-spectral
methods. The identification of turbulent coherent structures through λ2 method is
presented as well as the direct influence of turbulent eddies on the local Nusselt
number and the shear stress distribution at the active walls. Furthermore a first study
of turbulent kinetic energy and temperature variance together with the respective
production and dissipation terms are reported for Rayleigh number Ra = 109.
1 Introduction
Despite the fact that natural convective flows occur in many interesting fields and
applications like civil and environmental engineering, stellar dynamics, nuclear en-
gineering and safety, electronic device cooling, solar thermal collectors among the
others, experimental and numerical studies of the so called differentially heated cav-
ity are limited by different reasons [11]. It has been shown [7, 6] that sophisticated
closure models for Reynolds averaged Navier-Stokes (RANS) equations have to be
employed in order to describe properly the turbulence and avoiding relaminarization
R. Puragliesi · A. Dehbi
Paul Scherrer Institut, CH-5232 Villigen PSI Switzerland
e-mail: riccardo.puragliesi@psi.ch,abdel.dehbi@psi.ch
E. Leriche
Université Jean-Monnet, Saint-Étienne, F-42023 France
e-mail: emmanuel.leriche@univ-st-etienne.fr
A. Soldati
University of Udine, 33100 Udine, Italy
e-mail: soldati@uniud.it
M.O. Deville
École Polytechnique Fédérale de Lausanne, CH-1015 Switzerland
296 R. Puragliesi et al.
phenomena (low-Reynolds k − ε with two-layer approach and generalized gradient

diffusion hypothesis or k − ε − Θ
Θ
− εΘ for instance) increasing the global com-
plexity of the algorithm and introducing more model constants that need to be prop-
erly tuned. Concerning the turbulent regime at relatively high Rayleigh numbers in
a differentially heated cavity (DHC) the actual knowledge is not satisfactory due to
prohibitive computational costs for DNS and only few studies have dealt with three-
dimensional computations introducing spanwise homogenous direction, reducing
the overall cost of the computation but at the same time neglecting the influence of
the spanwise walls. For instance, an extensive study of 2D and 3D turbulent quanti-
ties obtained by DNS, using spectro-consistent spatial discretization and adaptative
central difference time splitting scheme, in a cavity with aspect ratio 4 and span-
wise periodic boundary conditions, for 6.4 × 108 ≤ Ra ≤ 1010 has been presented in
[16]. They concluded that the 2D and 3D mean fields are similar but the turbulent
intensity and eddy ejection process are rather stronger in the two-dimensional case
leading to turbulent activity also in the core region and higher levels of turbulent
kinetic energy and dissipation rate. To the author’s knowledge only [17] and [15]
reported accurate DNS results in a fully enclosed cubical domain at slightly sub-
critical and critical Rayleigh numbers (RaC ≈ 3.3 × 107 ), in order to address the
identification of the mechanism wich yelds the first transition to unsteadiness, and
supercritical Rayleigh number for characterising the time averaged flows and turbu-
lence intensity (Ra = 1.5 × 109). The time-averaged results reported in [15] are in
good agreement with the experimental profiles of vertical velocity component and
temperature along the horizontal line at half-height of the midplane, but on the other
hand discrepancies occur for the temperature vertical profile, as well as temperature
variance profiles probably due to the fact that in the experimental set up the lateral
walls were not perfectly adiabatic and also radiative effects between the walls might
influence the vertical thermal stratification leading also to a different temperature
variance distribution.
The present contribution will provide new informations about turbulent natural
convection inside a fully closed cavity at high Rayleigh number (Ra = 109 ) by
means of spectral accurate DNS calculations. The description of coherent struc-
tures identified by the λ2 criterion and the study of the turbulent kinetic energy and
temperature variance profiles and the associated production and dissipation terms
appearing in the corresponding transport equations are discussed.
2 Methodology
2.1 Governing Equations

The two vertical opposite walls of the cubic cavity shown in Fig.1 are kept at
different temperature, i.e. TH and TC for the hot and cold wall respectively, whereas
the horizontal and lateral walls are assumed adiabatic. No-slip condition for the
fluid velocity is applied to all solid boundaries. The fluid under examination is air
and is assumed homogenous, incompressible, Newtonian with constant kinematic
Direct Numerical Simulation of Buoyancy Driven Turbulence 297
viscosity νr and constant thermal diffusivity κr where the subscript r indexed

quantities that are measured at the reference temperature Tr = TH +TC
2 . The tem-
perature difference Δ T = TH − TC is supposed to be sufficiently small to employ
the Oberbeck-Boussinesq approximation. Accordingly to [2, 10], the appropriate
scaling quantities for velocity, length, time and pressure for convective natural
convection are
κr √ H2
U∗ = Ra, L∗ = H, t ∗ = √ , P∗ = ρrU ∗ 2 . (1)
H κr Ra
Fig. 1 Schematic rapresentation of the DHC cavity and boundary conditions
Denoting the open domain Ω = (] − 0.5, +0.5[)3 and ∂ Ω its boundary (therefore
the closed domain is Ω = Ω ∪ ∂ Ω ), the dimensionless Navier-Stokes equations,
where the local variation of density applies to the buoyancy source term only, read
in vector notation
∂u Pr g
+ u · ∇u = −∇p + √ Δ u − PrΘ in Ω , (2)
∂t Ra |g|
∇ · u = 0 in Ω , (3)
∂Θ 1
+ u · ∇Θ = √ ΔΘ in Ω , (4)
∂t Ra
where the dimensionless variables are the velocity vector u ≡ (u1 , u2 , u3 ), p the
pressure, Θ = TΔ−T T the temperature, Pr = νr /κr the Prandtl number and x =
r
(x1 , x2 , x3 ) ≡ (X1 /H, X2 /H, X3 /H) describes the position vector in the reference co-
ordinate system, H being the cube edge length. The boundary conditions for velocity
and temperature are the following
u = 0 on ∂ Ω , (5)
Θ (x1 = ±0.5, x2 , x3 ) = ∓0.5 for all x2 , x3 ∈ [−0.5, +0.5], (6)

and
∂Θ
(x1 , x2 = ±0.5, x3 ) = 0 for all x1 , x3 ∈ [−0.5, +0.5], (7)
∂ x2
∂Θ
(x1 , x2 , x3 = ±0.5) = 0 for all x2 , x3 ∈ [−0.5, +0.5] (8)
∂ x3
The initial velocity and temperature distribution are instantaneous solutions of
previous computations at lower Rayleigh (the code has been benchmarked with [17]
for the case of steady laminar regime).
2.2 Numerical Method

The spatial approximation of any field relies on the expansion in tensor product
of Chebyshev polynomials of order N along every space direction and a usual
collocation method is applied at the Chebyshev-Gauss-Lobatto points [4, 3]. The
projection-diffusion method introduced in [1] is chosen for its consistency with the
continuous space-time problem and for its optimal cost. A complete numerical anal-
ysis of this decoupling method may be found in [12, 13] where it is compared with
the high order time splitting scheme proposed by Karniadakis et al. in [9].
The time discretization is based on a second order backward Euler differentiation
formula. The diffusive terms are treated implicitly whereas the advection terms are
advanced explicitly in time by a second-order extrapolation scheme. The time dis-
cretized equations can be recast in a vectorial Helmholtz system of equations for the
velocity, a quasi-Poisson operator for the pressure, and a scalar Helmholtz equation
for temperature to solve at each time step, noting that the simple geometry allows
to use fast diagonalisation method for inverting these elliptic operators at a low
computational cost [14, 5].
3 Results
The parameters used in the present computation are summarized in Tab.1. Since the
large scale recirculation has a period based on the velocity maximum of about 20
dimensionless time units, the averaging time covers around 24 complete rotations.
Table 1 Computational parameters: polynomial degree N in each space direction, minimum
and maximum grid size Δ xi , integration time step Δ t, number of frames per unit of time and
averaging time
Ra Pr N min(Δ xi ) max(Δ xi ) Δt Sampl. freq. Ave. time

109 0.71 169 8.7 × 10−5 9.3 × 10−3 5 × 10−3 0.1 470.0

The actual local shear stress τ12 = √PrRa ∂∂ ux1 + ∂∂ ux2 and Nusselt number Nu = − ∂∂Θ
x1
2 1
on the hot wall are plotted in Fig.2 together with the cross-section views of coher-
ent structures identified by means of λ2 criterion [8] at three different spanwise
locations, namely x3 = −0.25, 0, 0.25. It can be seen that the instantaneous local
regions of low shear stress and heat flux are directly related to the presence of small
spanwise elongated eddies near the wall which are slightly bent at the extremity due
to the presence of the lateral walls, whereas high shear stress and heat flux appears
in between of two consecutive eddy structures. The latter behavior is direct conse-
quence of travelling waves and eventually detaching fluid at the vertical bound-
ary layers associated with periodic increase and decrease of the boundary layer
thickness.
0.500 0.500
0.250 0.250
x2
x2
0.000 0.000
−0.250 −0.250
−0.500 −0.500
−0.500 −0.250 0.000 0.250 0.500 0.0 0.1 0.0 0.1 0.0 0.1
x x
3 1
Fig. 2 Left: instantaneous shear stress τ12 and Nusselt number half-plane distribution at the
hot wall. Right: coherent structures on cross-planes at location x3 = −0.25, 0, 0.25
The turbulent kinetic energy, given by k = 12 u

i u
i , is shown in Fig.3-(left), while
in Fig.3-(right) the turbulent kinetic production terms due to the mean velocity gra-
dient P = −u
i u
j ∂∂ux ij and buoyancy G = PruiΘ
δi,2 , together with the dissipation
1

2
∂u ∂u
rate D = − √PrRa ∂ x ij ∂ x ij are plotted in the mid-plane (x3 = 0) at x2 = 0.3 against
y− , the distance from the wall multiplied by Ra1/4 . It is possible to note that the peak
of k is around y− ≈ 30 outside the boundary layer but in the corner eddy. The pres-
ence of the latter structure causes also the local maxima closer to the wall. On the
other hand the maximum of the production P appears around y− = 3 and for G it is
worthful to note that is always negative but a source term for 2.5 ≤ y− ≤ 7.5. Finally
the dissipation D has strong influence for y− ≤ 1. On the other hand the temperature
variance Θ
Θ
and its relative production and dissipation
1

2 curves are reported in
∂ Θ
Fig.4, being PΘ = −2uiΘ ∂ xi and DΘ = −2Pr ∂Θ

∂Θ
∂ xi ∂ xi respectively. Alike the
turbulent kinetic energy the temperature variance shows a peak at y− ≈ 30. More-
over it presents less local maxima and minima. Unlike the production term P, PΘ is
always positive and with the first maximum peak around y− ≈ 2. Finally the level
of temperature variance dissipation rate is rather low in comparison with D.
2.50 6.00
P
G
4.00 D
2.00
2.00
1.50
k [10 ]
-4
[10-4]
0.00
1.00
-2.00
0.50
-4.00
0.00 -6.00
0.01 0.1 1 10 100 0.01 0.1 1 10 100
y- y-
Fig. 3 Mean turbulent kinetic energy (left) and production/dissipation profiles (right) in the
mid-plane (x3 = 0) at x2 = 0.3
5.00 1.50
PΘ
DΘ
4.00
1.00
3.00
Θ’Θ’ [10-4]
[10 ]
-3
0.50
2.00
0.00
1.00
0.00 -0.50
0.01 0.1 1 10 100 0.01 0.1 1 10 100
y- y-
Fig. 4 Temperature variance (left) and production/dissipation profiles (right) in the mid-plane
(x3 = 0) at x2 = 0.3
4 Conclusions
Accurate pseudo-spectral DNS calculations of turbulent natural convection in a fully
closed cubic cavity are presented at Rayleigh number Ra = 109 . A first incursion
concerning the identification and the topology of turbulent coherent structures is
shown for the first time in the literature with this configuration, stressing the direct
influence on the instantaneous local distribution of the shear stress and heat flux
at the hot (or cold) wall. Finally turbulent kinetic energy profile and temperature
variance together with their respective production and dissipation terms across the
midplane are briefly discussed.
References
1. Batoul, A., Khallouf, H., Labrosse, G.: Une Méthode de Résolution Directe
(Pseudo-Spectrale) du Problème de Stokes 2D/3D Instationnaire. Application à la Cavité
Entrainée Carrée. C.R. Acad. Sci. Paris 319 Série I, 1455–1461 (1994)
2. Bejan, A.: Convection heat transfer. Wiley Interscience, Hoboken (1984)
3. Canuto, C., Hussaini, M.Y., Quarteroni, A., Zang, T.A.: Spectral Methods in Fluid Dy-
namics. Springer Series in Computational Physics. Springer, New-York (1988)
4. Gottlieb, D., Orszag, S.A.: Numerical Analysis of Spectral Methods: Theory and
Applications. SIAM-CBMS, Philadelphia (1977)
5. Haldenwang, P., Labrosse, G., Abboudi, S.A., Deville, M.: Chebyshev 3D Spectral
and 2D Pseudospectral Solvers for the Helmholtz Equation. Journal of Computational
Physics 55, 115–128 (1984)
6. Hanjalic, K., Kenjeres, S., Durst, F.: Natural convection in partitioned two-dimensional
enclosures at higher rayleigh numbers. International Journal of Heat and Mass Trans-
fer 39(7), 1407–1427 (1996)
7. Hsieh, K.J., Lien, F.S.: Numerical modeling of buoyancy-driven turbulent flows in
enclosures. International Journal of Heat and Fluid Flow 25(4), 659–670 (2004)
8. Jeong, J., Hussain, F.: On the identification of a vortex. Journal of Fluid Mechanics 285,
69–94 (1995)
9. Karniadakis, G.E.M., Israeli, M., Orszag, S.A.: High-Order Splitting Methods for the
Incompressible Navier-Stokes Equations. J. Computational Physics 97, 414–443 (1991)
10. Le Quéré, P.: Accurate solutions to the square thermally driven cavity at high Rayleigh
number. Computers and Fluids 20(1), 29–41 (1991)
11. Leong, W.H., Hollands, K.G.T., Brunger, A.P.: On a physically-realizable benchmark
problem in internal natural convection. Int. J. Heat Mass Transfer 41, 3817–3828 (1998)
12. Leriche, E., Labrosse, G.: High-Order Direct Stokes Solvers with or without Tempo-
ral Splitting: Numerical Investigations of their Comparative Properties. SIAM J. Scient.
Comput. 22(4), 1386–1410 (2000)
13. Leriche, E., Perchat, E., Labrosse, G., Deville, M.O.: Numerical evaluation of the accu-
racy and stability properties of high-order direct Stokes solvers with or without temporal
splitting. Journal of Scientific Computing 26, 25–43 (2006)
14. Lynch, R.E., Rice, J.R., Thomas, D.H.: Direct Solution of Partial Diiference Equations
by Tensor Product Methods. Numerishe Mathematik 6, 185–199 (1964)
15. Salat, J., Xin, S., Joubert, P., Sergent, A., Penot, F., Le Quéré, P.: Experimental and
numerical investigation of turbulent natural convection in a large air-filled cavity. Int. J.
Heat and Fluid Flow 23, 824–832 (2004)
16. Trias, F.X., Soria, M., Oliva, A., Pérez-Segarra, C.D.: Direct numerical simulation of
two- and three-dimensional turbulent natural convection flows in a differentially heated
cavity of aspect ratio 4. J. Fluid Mech. 586, 259–293 (2007)
17. Tric, E., Labrosse, G., Betrouni, M.: A first incursion into the 3D structure of natural con-
vection of air in a differentially heated cubic cavity, from accurate numerical solutions.
Int. J. Heat Mass Transfer 43, 4043–4056 (2000)
Numerical Simulations of a Massively Separated
Reactive Flow Using a DDES Approach for
Turbulence Modelling
Bruno Sainte-Rose, Nicolas Bertier, Sébastien Deck, and Francis Dupoirieux
Abstract. Computations of a lean premixed methane - air flame in a lean stepped

combustor are performed using a Delayed Detached Eddy Simulation approach to
model turbulence. Two conditions for the outlet section are simulated and compared
to an experimental database including mean velocity, mean temperature and instan-
taneous OH emission measurements. The main objective of this study is to assess
of the efficiency of DDES for a massively separated reactive flow.
1 Introduction
Large Eddy Simulation is now of common use for reactive flows [1]. In fact, LES
is well suited for massively separated flows such as those found in swirled burn-
ers, ramjets and is effective to reproduce combustion instabilities. Indeed, LES is
able to solve the most energetic structures of the turbulent flow while modelling the
subgrid turbulent dissipation. On the other hand, Reynolds Averaged Navier Stokes
Bruno Sainte-Rose
Onera (Fundamental and Applied Energetics Department), 29 avenue de la Division Leclerc
92322 Châtillon
e-mail: bruno.sainte-rose@onera.fr
Nicolas Bertier
Onera (Fundamental and Applied Energetics Department), 29 avenue de la Division Leclerc
92322 Châtillon
e-mail: nicolas.bertier@onera.fr
Sébastien Deck
Onera (Applied Aerodynamics Department), 8 rue des Vertugadins 92190 Meudon
Francis Dupoirieux
Onera (Fundamental and Applied Energetics Department), Chemin de la Hunière 91761
Palaiseau
e-mail: francis.dupoirieux@onera.fr
304 B. Sainte-Rose et al.
approaches are progressively limited to parametric and optimisation studies in the

design path of combustion devices. However in certain cases where an accurate
resolution of the wall flows is required, simulating thin turbulent structures encoun-
tered in the boundary layers close to the walls can become very costly with a LES,
while the low Reynolds models make such computations far more affordable with
RANS. As a consequence, hybrid RANS / LES methods such as DDES employed
in this study are proposed to take into account simultaneously attached boundary
layers by RANS approach and massively separated flows by LES approach. Such a
method has been used and is presented here for the study of a lean premixed turbu-
lent methane - air flame stabilized by a backward facing step. An analysis of both
the mixing layer and the turbulent flame is adressed in this paper. The influence of
the acoustic conditionning of the outlet boundary is also scrutinized.
2 Delayed Detached Eddy Simulation for Reactive Flow

The hybrid RANS / LES method chosen to take into account turbulence in this study
is the DDES approach originally proposed by Spalart et al. [2]. DDES is applied
to the k − ω SST model [3] for which the destruction term of the equation of the
turbulent kinetic energy k, i. e. ε , is thus modified [4],
k3/2 k3/2
ε = β ∗ω k = −→ (1)
lRANS lDDES
and where the modified length scale lDDES is calculated as
lDDES = lRANS − fDDES × max(0, lRANS − CDES Δ ) (2)

√
lRANS is the RANS length scale and is equal to k/β ∗ ω and Δ is the mesh scale
equal to (Δx Δy Δz )1/3 . In equation (2), the fDDES function allows the model to yield
the boundary layers to SST - RANS ( fDDES = 0) while switching to a Strelets DES
approach [5] far from the wall ( fDDES = 1). This modification of the original DES
was motivated by some unphysical outcomes like artificial relaminarisation, also
called Modelled Stress Depletion [6][7].
Concerning the modelling used for turbulent combustion, the source term of the
species balance equation is obtained by an Arrhenius law corresponding to a global
reaction mechanism. The interaction with turbulence is modelled by a Dynamically
Thickened Flame for LES approach introduced by Legier et al. [8].
3 Presentation of the Physical and Numerical Test Case

The test case presented in this study was used to validate the DES approach for
reactive flows in the Onera CEDRE code [4]. The main features of the flow are
presented in figure 1. The equivalence ratio of the methane air mixture is equal
to 0.8. An experimental database covering both velocity [9] and temperature [11]
Numerical Simulations of a Massively Separated Reactive Flow 305
is available along with instantaneous visualizations of the OH [10] spontaneous

emissions.
Fig. 1 Geometry and main features of the reactive flow, the background color corresponds to
modelled turbulent viscosity, height of the step h = 3.5 cm, U0 = 50 m.s−1 , Reh ≈ 40.000
The main characteristics of the computational domain are displayed on figure 2

along with a spanwise cut of the 3D grid. The number of cells for the computational
grids is 70,000 for the 2D case and 1,500,000 for the 3D block structured mesh.
Both meshes verify a grid spacing of y+ 0 < 3 for the lower wall and y0 < 12 for
+
the upper wall, periodic conditions are applied to the lateral walls. More detailed
informations about the meshing strategy and numerical methods used in the Onera
CEDRE code can be found in [4].
Fig. 2 Computational domain and measurements, spanwise cut of the 3D grid
The effect of the streamwise acoustic modes are reproduced by applying imposed
pressure conditions at the outlet of the computational domain while the case with-
out acoustic reflections at the outlet is obtained thanks to non reflecting NSCBC
boundary conditions [12]. Three computations (2D-RANS, DDES with or without
reflection at the outlet) are presented and discussed in the following sections.
4 Analysis of the Recirculating Region and of the Mixing Layer

The time - dependent field has been averaged during the computations. Our com-
putations cover 0.15 s of physical time with a time step of 1.5 μ s i. e. 100, 000
iterations. The first relevant analysis of the recirculating zone which can be made on
the results concerns the averaged reattachment length. The ratio between the mean
reattachment length and the size of the step Lr /h is compared to the experimental
value obtained by Laser Doppler Velocimetry in table 1. One can observe that all the
computations tend to overestimate this value. This is due in particular to the thermal
conditions at the lower wall. Indeed, adiabatic conditions were considered at the
wall due to the lack of measurements of wall temperature, as a consequence wall
cooling is neglected in the computations. Moreover, the comparison between the
three computed values gives rise to the following observations: RANS results are
of poor quality compared to the DDES computations, in the meantime, the imposed
pressure at the outlet section of the DDES calculation tend to shorten the length of
the bubble.
Table 1 Averaged size of the recirculating bubble
LDV DDES without reflection DDES with reflection 2D-RANS

Lr /h 2.9 < − < 3.4 3.8 3.4 4.7
Another aspect was tackled to appraise of the quality of the description of the re-
active mixing layer. Indeed, the evolution of the vorticity thickness δω in the stream-
wise direction for the three computations and for the Laser Doppler Velocimetry
measurements [9] in the range 0 < x/h < 3.8 is plotted in figure 3. The vorticity
thickness is normalised by the height of the step and by the momentum thickness θ
respectively in the left and right sides of the figure. The expressions of δω (x) and
θ (x) are reminded:
[U max[y] − U min[y] ](x)
δω (x) = (3)
max[y] ∂ U(x,y)
∂y
ymax

U(x, y) − U min[y] (x) U(x, y) − U min[y] (x)
θ (x) = 1− dy (4)
ymin U max[y] (x) − U min[y] (x) U max[y] (x) − U min[y] (x)
where U is the averaged streamwise velocity. It can be noticed in figure 3 that
6
0 .2
0.8 4
0.6 3
δω/θ
δω/h
DDES w/o reflexion 2

0.4
DDES w/ reflexion
2D - RANS
0.2 1
LDV
LDV
0
0 1 2 3 0
0 1 X/h 2 3 X/h
Fig. 3 (left) Vorticity thickness and (right) ratio between vorticity and momentum
thicknesses along the mixing layer
close to the step for 0 < x/h < 0.4, the evolution of the vorticity thickness is not
exponential as for an inert shear layer but rather linear. This can be explained
by the acceleration of the fluid located under the flame which tends to lower the
maximum velocity gradient. A good agreement between the DDES computations
without acoustic reflections and the experiments is observed. A linear growth of
about 0.26 is also evidenced. This remark is a consequence of the accuracy of the
streamwise velocities discussed in [4]. Moreover, the comparison between the two
DDES computations show that the shear layer is greatly impacted by the outlet con-
ditions. The growth of the vorticity thickness is indeed doubled for 0 < x/h < 1.5
in the case of streamwise modes which highlights a coupling between acoustic and
the size of the eddies growing from the corner of the step.
5 Analysis of the Turbulent Flame

The results obtained for the averaged temperature flowfields are discussed in this
section. In figure 4, an iso-line of T = 1500 K is displayed for the three computations
and compared to temperature measurements obtained by Coherent Anti- Stokes Ra-
man Spectroscopy (CARS) [11]. This layout allows to approximately locate the av-
eraged computed flame. The results obtained close to the step with DDES are very
satisfying since the flame angle is well reproduced whereas the RANS simulation
clearly underestimates the flame angle. The use of reflecting conditions at the outlet
results in an increase of the flame angle.
Fig. 4 1500 K temperature iso - line for DDES without reflection (solid), DDES with reflection
(dash-dot), RANS (dashed, only for T ), CARS (circles); vertical dash lines: position of the
CARS measurements
Concerning the temperature profiles in figure 5, the previous remarks on the location
of the flame are verified since the peak and levels found for the fluctuating temper-
ature are in very good agreement with the experiments except for the first profile
where the effect of the lack of
wall cooling is evidenced. For the temperature fluc-

= 2
tuations, the value of Trms T 2 − T is shown in the right part of the figure. The
difference between the two acoustic conditions discussed previsously is identified
since the important wrinkling due to the pressure waves lead to an overestimation
of the peak and of the width of the flame brush.
The acoustic modes appearing in the computations and their coupling with the
dynamic of the flame are now scrutinized. For that purpose the temperature sig-
nal obtained at a point located at the corner of the step is analyzed. In figure 6 the
temperature spectrum GT ( f ) for the two cases are plotted. The case with imposed
1
1
0 Y/h
0 Y/h
-1
-1
1000 2000 0 500

T, K X/h=1.1 X/h=4.3 X/h=6.0 X/h=9.7 X/h=13.1 X/h=20.3 T’rms (K)

DDES without reflection (solid), with
Fig. 5 Mean and fluctuating temperature T and Trms
reflection (dash-dot) and RANS (dashed only for T ), CARS measurements (diamonds)
pressure at the outlet clearly displays quarter wave frequencies. Moreover, the
analytic value of the quarter wave frequencies in a 1D wave duct of length L cor-
responds to fn = nc/4L, where c is the speed of sound in the burnt gases and n
the number of the harmonics. This formula gives f1 ≈ 174 Hz, f2 ≈ 348 Hz, f3 ≈
522 Hz etc.. These frequencies are evidenced in the right part of figure 6 where
peaks at f1 = 173 Hz, f2 = 352 Hz, f3 = 525 Hz etc. can be observed correspond-
ing to Strouhal numbers (based on an upstream velocity of 50 m.s−1 and the height
of the step) of St1 = 0.12, St2 = 0.25, St3 = 0.37 etc. . It is interesting to see that
even with non reflecting boundary conditions the first harmonic can also be seen in
the left part of the figure. However another peak at f = 433 Hz, St = 0.3 is evi-
denced, this frequency being close to the value obtained for the Kelvin - Helmholtz
instability on the inert case.
f(Hz) f(Hz)
102 103 102 103
2 2
temperature spectrum, GT(f)
temperature spectrum, GT(f)
3 3 5
6
7
1
-1 0 -1
10 Sth 10 10 Sth 100
Fig. 6 Temperature spectrum GT ( f ) for DDES without reflection (left) and with reflection
(right)
A time - dependent sequence of heat release fields is displayed in figure 7 for

DDES with acoustic reflection and enables a comparison between the current sim-
ulation with instantaneous screenshots of OH emission measurements obtained
during a campaign aiming at reproducing combustion instabilities [10]. A strong
wrinkling of the flame close to the step evidences a strong coupling between flame
and eddies in response to streamwise acoustic modes. In the current case massive
flash-back does not occur unlike the experiments.
Fig. 7 Top: time dependent heat release ω̇T during T0 = 2.8 ms in the midspan plane
corresponding to a period of the first harmonic; bottom: corresponding OH emission
screenshots, Φ is the phase in deg
6 Conclusions
In the present work, DDES computations were carried out on a BFS reactive flow.
This study constitutes an extension of the work published in [4]. It indeed gives new
insights by providing an investigation of the mixing layer and the flame behaviour
particularly to study the effect of acoustic waves inside the chamber. DDES was
demonstrated to provide a clear improvement on RANS by showing a good agree-
ment with the experimental database for velocity and temperature. The behaviour of
the flame and mixing layer with streamwise acoustic waves was evidenced. To ex-
tend the use of DES for massively separated reactive flows, the study of a separated
nozzle flow with combustion will be carried out.
Acknowledgements
The authors want to acknowledge the Centre Nationale d’Etudes Spatiales (CNES) for the
funding of B. Sainte-Rose PhD studies. A. Laverdant and V. Sabelnikov are also gratefully
acknowledged.
References
1. Poinsot, T., Veynante, D.: Theoretical and numerical combustion, R.T. Edwards Ed.
(2001)
2. Spalart, P.R., Deck, S., Shur, M.L., Squires, K.D., Strelets, M.K., Travin, A.: A new ver-
sion of detached-eddy simulation, resistant to ambiguous grid densities. Theor. Comput.
Fluid Dyn. 20, 181–195 (2006)
3. Menter, F.R.: Zonal two-equation k-omega turbulence models for aerodynamic flows,
AIAA 1993-2906 (1993)
4. Sainte-Rose, B., Bertier, N., Deck, S., Dupoirieux, F.: A DES method applied to a Back-
ward Facing Step reactive flow. C.R. Meca 337, 340–351 (2009)
5. Strelets, M.K.: Detached-Eddy Simulation of Massively Separated Flows, AIAA

2001-0879 (2001)
6. Sagaut, P., Deck, S., Terracol, M.: Multiscale and multiresolution approaches in
turbulence. Imperial College Press, London (2006)
7. Spalart, P.R.: Detached Eddy Simulation. Annu. Rev. Fluid Mech. 41, 181–202 (2009)
8. Legier, J.P., Poinsot, T., Veynante, D.: Dynamically thickened flame LES model for
premixed and non-premixed turbulent combustion. In: Proceedings of the Center for
Turbulence Research (2000)
9. Moreau, P., Labbe, J., Dupoirieux, F., Borghi, R.: Experimental and numerical study of
a turbulent recirculation zone with combustion. In: 5th Symposium on Turbulence and
Shear Flow (1985)
10. Sabelnikov, V., Grisch, F., Orain, M.: Instabilities and structure of turbulent premixed
flame in a lean stepped combustor, paper ISABE-2005-1059 (2005)
11. Magre, P., Collin, G., Bouchardy, P.: Application de la DRASC à l’opération A3C
(french). Onera Technical Report (1996)
12. Poinsot, T., Lele, S.: Boundary conditions for direct simulations of compressible reacting
flows. J. Comput. Phys. 101, 104–129 (1992)
Particle Dispersion in Large-Eddy Simulations:
Influence of Reynolds Number and of Subgrid
Velocity Deconvolution
Maria Vittoria Salvetti, Cristian Marchioli, and Alfredo Soldati
Abstract. Particle dispersion in turbulent channel flow is investigated through

Lagrangian tracking of inertial pointwise particles one-way coupled to the fluid.
First, the results obtained in direct numerical simulations at two different shear
Reynolds numbers, Relτ = 150 and Rehτ = 300, for particles having different iner-
tia are briefly summarized and the Reynolds number effects are discussed. Then,
particle dispersion in LES flow fields at Rehτ is compared to reference DNS data.
Comparison is made with and without approximate deconvolution of the LES fluid
velocity field in the particle motion equations. It is found that particle segregation
and, consequently, near wall accumulation are underestimated in LES, although ap-
proximate deconvolution improves the agreement with DNS. These findings confirm
those of analogous previous studies at Relτ .
1 Introduction
Turbulent dispersed flows in boundary layers are important in a number of indus-
trial and environmental applications where a key information is the rate at which
the particles are transported to, deposited at, and re-entrained from the wall by
turbulence. Direct numerical simulation (DNS) studies have shown that there is a
strong correlation between coherent wall structures, local particle segregation and
Maria Vittoria Salvetti
Dip. Ingegneria Aerospaziale, Università di Pisa, Via G. Caruso 8 - 56122 Pisa, (Italy)
e-mail: mv.salvetti@ing.unipi.it
Cristian Marchioli · Alfredo Soldati
Centro Interdipartimentale Fluidodinamica e Idraulica and Dip. Energetica e Macchine,
Università di Udine, Via delle scienze 208 - 33100 Udine, (Italy)
e-mail: cristian.marchioli@uniud.it,soldati@uniud.it
312 M.V. Salvetti, C. Marchioli, and A. Soldati
subsequent deposition phenomena (see, e.g., [6]). Specifically, inertial particles are
initially segregated and accumulated into specific flow regions close to the walls and
only afterwards are driven to the walls.
In the present paper we analyze from a statistical viewpoint particle segregation
and wall accumulation in directly-simulated particle-laden turbulent channel flow,
when different values of the flow Reynolds number and of the particle response
time are considered. Object of this analysis is to discuss possible Reynolds number
effects on the macroscopic phenomena characterizing particle dispersion.
Another important issue is the modeling of the segregation and accumulation
mechanisms in numerical methods coarser than DNS. Lagrangian tracking of parti-
cles in flow fields obtained by large-eddy simulation (LES) represents a potentially
useful tool for practical applications; however, the full capabilities of LES in ac-
curately predicting segregation and wall accumulation are still to be proven. More
specifically, the critical issues are: (i) how does the subgrid scale (SGS) turbulence
affect particle dispersion? (ii) How should these SGS effects be modeled in order to
obtain accurate prediction of the selective response of different-inertia particles?
Recent studies (see e.g. [2, 4]) have shown that LES tends to underestimate lo-
cal particle segregation and, in turn, deposition fluxes and near-wall accumulation.
Therefore, a way to model the effects of the filtered SGS flow velocity fluctuations
on particle motion needs to be identified. In our previous works [4, 5] we carried out
a-priori and a-posteriori LES for channel flow at shear Reynolds number Reτ = 150
(Relτ hereinafter), for different particle inertia and grid resolutions, with and without
SGS closure models in the particle motion equations. The considered closures were
based on approximate deconvolution and fractal interpolation. One of the main con-
clusions was that, even when a SGS closure model for particles is able to reintroduce
the correct amount of fluid velocity fluctuations in the particle motion equations, as
was obtained through approximate deconvolution, this may not be sufficient to war-
rant accurate prediction of particle near-wall accumulation. A similar study was
carried out in [3] for particle-laden channel flow at Reτ = 150 and 590. However,
particle-laden DNS data where not available at the higher Reτ . In the present paper
we compare the results obtained in a-posteriori LES at Reτ = 300 (Rehτ hereinafter)
with available DNS data for both fluid and particle phases, to investigate if the previ-
ous finding also holds at higher Reynolds numbers and whether an explicit Reynolds
number and/or inertia dependence is needed in SGS closure models for the particle
motion equations.
2 Physical Modeling and Numerical Methodology

The governing equations for the fluid phase in DNS are the continuity and the
Navier-Stokes equations for incompressible flows. In LES, the flow variables are de-
composed through filtering into a resolved (large-scale) part and a residual (sub-grid
scale) part. The filtered equations for the resolved scales are thus considered, closed
by a physical model of the subgrid-scale (SGS) stress tensor. The eddy-viscosity
dynamic SGS model [1] is adopted here.
Particle Dispersion in Large-Eddy Simulations 313
The reference geometry consists of two infinite flat parallel walls: the origin of
the coordinate system is located at the center of the channel and the x, y and z axes
represent streamwise, spanwise and wall-normal directions respectively. Periodic
boundary conditions are imposed on the fluid velocity field in the homogeneous
directions (x and y), no-slip boundary conditions are imposed at the walls (z = −h
and z = h). The size of the computational domain is Lx × Ly × Lz = 4π h × 2π h × 2h.
The flow is driven by an assigned mean pressure gradient. The reference Reynolds
number Reτ ≡ uτ h/ν is based on the shear (or friction) velocity, uτ , and on the
half channel
height, h, ν being the fluid kinematic viscosity. The shear velocity is
uτ ≡ τw /ρ , where τw is the mean shear stress at the wall and ρ is the fluid density.
For the numerical discretization of the fluid-phase governing equations a pseudo-
spectral approach is used, based on Fourier transforms for the periodic streamwise
and spanwise directions and on a Chebyshev representation for the wall-normal di-
rection. Time advancing is carried out through a two level explicit Adams-Bashforth
scheme for the non-linear terms and an implicit Crank-Nicolson method for the vis-
cous terms.
The motion of particles is described in the Lagrangian framework by a set of
ordinary differential equations for particle position, x p , and velocity, u p . These
equations in vector form read as:
dx p du p (u@p − u p)
= up , = 1 + 0.15 Re0.687 (1)
dt dt τp p
where u@p is the fluid velocity at the particle position, τ p ≡ ρ p d 2p /18μ is the par-
ticle relaxation time (d p and μ being the diameter of the particle and the dynamic
viscosity of the fluid, respectively) and the particle Reynolds number is defined as
Re p = |u@p − u p |d p /ν . As previously mentioned, in LES only the resolved part of
the fluid velocity field, u, is available. If u interpolated at the particle position is
plugged into Eq. (1) then no SGS model is used in the particle motion equations to
retrieve the effects of the smallest flow scales not available in LES. Another pos-
sibility is to reconstruct an approximation of the whole fluid velocity field starting
from u. This reconstruction can be done by means of different techniques, viz. fil-
ter inversion [2], approximate deconvolution of fractal interpolation [5], and can be
considered as a closure model for the particle motion equations. Based on the results
of our previous studies [5], approximate deconvolution is used herein. The filtered
velocity is defined by the convolution product of the instantaneous velocity and a
filter kernel G having a size Δ . Assuming that G has an inverse, G−1 , the inverse
operator can be expanded as an infinite series of filter operators. Filters with com-
pact transfer functions are non-invertible, but a regularized inverse operator can be
approximated by truncating the series at some N [7]. As a consequence, the approx-
imate deconvoluted velocity, u , can be obtained by: u = ∑Nα =0 (I − G)α u, where
I is the identity operator, denotes the convolution operator and N is the truncation
parameter to be fixed (following [7], we set N = 5). The transfer function of a top-
hat filter, having a width corresponding to the LES grid resolution and applied in
the streamwise and spanwise directions, is considered herein.
The Lagrangian particle tracking (LPT) is coupled to the DNS/LES equations

under the following assumptions: (i) particles are pointwise, non-rotating rigid
spheres (point-particle approach); (ii) particles are injected into the flow at concen-
tration low enough to consider dilute system conditions: the effect of particles onto
the turbulent field is neglected (one-way coupling approach) as well as inter-particle
collisions.
The equations of particle motion are time advanced by a 4th -order Runge-Kutta
scheme, starting from an initial random distribution over the computational domain,
in which particles have a velocity equal to that of the fluid at their position. At each
time step, the fluid velocity at the particle position is interpolated from the avail-
able discrete distribution using 6th -order Lagrangian polynomials (near the wall,
the interpolation scheme switches to one-sided). The timestep for particle tracking
was chosen equal to that of the fluid. Periodic boundary conditions are imposed on
particles moving outside the computational domain in the homogeneous directions.
Perfectly-elastic collisions at the smooth walls are assumed when the particle center
is at a distance lower than one particle radius from the wall. A discussion of the mo-
tivation and validity of the simplifying assumption made in the physical modeling
of the particle motion can be found in [6].
DNS and LES were carried out at two different Reynolds numbers: Relτ = 150
and Rehτ = 300. The size of the computational domain in wall units is L+ x × Ly ×
+
Lz = 1885 × 942 × 300 for the Reτ simulations and Lx × Ly × Lz = 3770 × 1885 ×
+ l + + +
600 for the Rehτ simulations. In DNS, 128 × 128 Fourier modes and 129 Chebyshev
coefficients were used for the Relτ simulations and 256 × 256 Fourier modes and
257 Chebyshev coefficients for the Rehτ simulations to maintain the same resolution
in wall units. The particle statistics and concentration obtained in these simulations,
with and without SGS closure for the particle motion equations and with two differ-
ent grids, were compared against DNS data in [4] and [5]. In Sec. 4 we present new
results for LES at Rehτ carried out on a 64 × 64 × 129 grid, with no SGS model and
with approximate deconvolution in the particle equations, respectively. Although at
Rehτ the LES grid resolution in the wall normal direction is half the DNS one, as a
first step, in the present study deconvolution is not applied in this direction, in anal-
ogy to what was done at Relτ in [5]. This choice also avoids the introduction of an
additional source of uncertainty, i.e. the definition of the filter transfer function in
a inhomogeneous direction and for a non-uniform grid spacing. Based on the anal-
ysis in [5], no result is presented for simulations with SGS model based on fractal
interpolation.
3 Reynolds Number Effects on Particle Dispersion

Reynolds number effects on particle dispersion may become significant because
the same (dimensional) value of the particle response time corresponds to different
values of the Stokes number for different Reτ . In particular, the following expression
holds:
1.6 2
a) b)
Scaled particle-to-fluid velocity RMS ratios
Scaled particle-to-fluid velocity RMS ratios

1.4 1.8
l
x-component: St =1 , Reτ=150
l
Sth=4 , Reτ=300 1.6 x-component: St =25 , Reτ=150
1.2 Sth=100 , Reτ=300
1.4
1
y-component: Stl=1 , Reτ=150 1.2 y-component: Stl=25 , Reτ=150
St =4 , Reτ=300
h
Sth=100 , Reτ=300
0.8 1
0.6 0.8
z-component: Stl=1 , Reτ=150
0.4 Sth=4 , Reτ=300 0.6
0.4 z-component: Stl=25 , Reτ=150

0.2 Sth=100 , Reτ=300
0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Normalized non-dimensional distance from the wall (z+/H+) Normalized non-dimensional distance from the wall (z+/H+)
Fig. 1 Scaled particle−to−fluid velocity rms ratios at Relτ and Rehτ . (a) St l = 1 versus St h =
4, (b) St l = 25 versus St h = 100
2 2
St h τ lf uhτ Rehτ
τ ph = τ pl → St h · τ hf = St l · τ lf → = = = , (2)
St l τ hf ulτ Relτ
where St h and St l represent the particle Stokes number in the Rehτ simulation and
in the Relτ simulation. The Stokes number is the dimensionless particle response
time, and is defined as St = τ p /τ f = τ p u2τ /ν . In this study, we considered 3 particle
classes in the Relτ -flow (corresponding to St l = 1, 5 and 25) and six particle classes
in the Rehτ -flow (corresponding to St h = 1, 4, 5, 20, 25 and 100). In the proposed scal-
ing (2) the shear velocity is assumed to be the proper fluid parameter to quantify the
Reynolds number effect on particle preferential concentration. Following Eq. (2),
for instance, the St h = 4 particles behavior in the Rehτ -flow is expected to resemble
that of the St l = 1 particles behavior in the Relτ -flow. Similarly, scaling effects are
expected to couple the St h = 100 particles to the St l = 25 particles. We checked this
scaling argument for particles tracked in the DNS flow fields. Figure 1 compares the
ratios between the rms of the particle and the fluid velocity components, urms p,i /ui ,
rms
normalized using the shear velocity as scaling parameter. The non-dimensional dis-
tance from the wall is indicated as z+ /H + where z+ is the distance in wall units and
H + is equal to either Relτ or Rehτ . In Fig. 1, thick lines are used for the Relτ -DNS
whereas open symbols refer to the Rehτ -DNS. For ease of reading, urms p,i /ui
rms pro-
files for the streamwise and wall-normal rms components are shifted by a factor of
0.4, up and down respectively. It is apparent that the profiles, though a bit ragged,
overlap quite well even in the near-wall region, where discrepancies (possibly due
to the extension of the averaging time window) are limited to very thin slabs inside
the viscous sublayer, thus supporting the validity of the adopted scaling.
The tendency of particles to segregate in a turbulent flow field can be quan-
tified in terms of the segregation parameter, D = (σ − σPoisson )/m, where m is
the mean particle number density and σ and σPoisson represent the standard devi-
ations for the measured particle number density distribution and for the Poisson
distribution respectively. D = 0 corresponds to a random distribution, D < 0 corre-
sponds to a uniform distribution, and D > 0 indicates segregation of particles. In this
3
DNS - Reτ=150
DNS - Reτ=300
2.5
Dmax
1.5
0.5
(a)
0
0.1 1 10 100 1000
St
Fig. 2 Maximum value of the segregation parameter, Dmax , in the near wall region (0 ≤ z+ ≤
5) versus particle Stokes number, St: • DNS at Relτ , ◦ DNS at Rehτ
latter case, larger values of D correspond to larger deviation from randomness and to
stronger segregation. Technical details on the computation of D, not given here for
brevity, can be found in [6]. Fig. 2 shows the maximum deviation from randomness,
Dmax in the viscous sublayer (0 < z+ < 5) obtained from the two DNS as a function
of the particle Stokes numbers, St l and St h respectively. It is clear that the degree
of segregation is indeed nearly the same for particle Stokes numbers matching the
condition of Eq. 2 and, thus, the scaling holds also for particle preferential concen-
tration. We remark that the St = 25 particles in the Relτ -flow, for which Dmax has a
peak, also exhibit the strongest tendency to sample preferentially the flow field and
to accumulate at the wall (as discussed in [4, 6]).
4 Particle Tracking in LES Flow Fields

In previous studies carried out at Relτ [5], we found that particle segregation and
near wall accumulation is generally underestimated in LES, even when the correct
amount of particle velocity fluctuations is recovered through the introduction of a
SGS model in the particle motion equations, e.g. approximate deconvolution. To
investigate further on the Reynolds number scaling described in Sec. 3, we ana-
lyze here the results obtained from particle tracking in LES flow fields at Rehτ . Fig.
3 shows the maximum value of the segregation parameter Dmax computed in the
near wall region for LES with and without approximate deconvolution of the fluid
velocity field in the particle motion equations: values are compared to reference
DNS data. It is evident that segregation is always underestimated in LES, although
approximate deconvolution improves the agreement with DNS. As previously men-
tioned, wall accumulation is deeply related to segregation of particles by the near-
wall turbulence structures (see also [6]); thus, underestimation of segregation in LES
results in underestimation of particle near-wall accumulation, as shown for instance
by the particle number density profiles of Fig. 4. Note how particles characterized by
higher values of Dmax show a stronger tendency to accumulate at the wall. Again,
2.5
DNS
LES
LES WITH DECONVOLUTION
2
Segregation parameter, Dmax

1.5
0.5
0
1 10 100
Stokes number, St
Fig. 3 Maximum value of the segregation parameter, Dmax , in the near wall region (0 ≤ z+ ≤
5) versus particle Stokes number, St: • DNS at Rehτ , LES at Rehτ , LES with SGS closure
based on approximate deconvolution at Rehτ
100 250
DNS DNS
90 LES LES
LES WITH DECONVOLUTION LES WITH DECONVOLUTION
80 200
70
60 150
C/C0
C/C0
50
40 100
30
20 50
10
0 0
0.1 1 10 100 0.1 1 10 100
z+ z+
Fig. 4 Particle concentration at Rehτ : comparison between DNS and LES. (a) St h = 5, (b)
St h = 25
approximate deconvolution improves the agreement with DNS. These results are
consistent with previous findings at Relτ [4, 5], and indicate that a scaling is possible
for small to moderate Reynolds numbers. Further simulations at higher Reynolds
numbers (Rehτ ≥ 600) and for larger particles (St h ≥ 100) are under way to explore
a broader range of parameters in the (Reτ , St) space.
References
1. Germano, M., Piomelli, U., Moin, P., Cabot, W.H.: A dynamic subgrid-scale eddy
viscosity model. Phys. Fluids 3, 1760–1765 (1991)
2. Kuerten, J.G.M., Vreman, A.W.: Can turbophoresis be predicted by large-eddy
simulation? Phys. Fluids 17, 011701 (2005)
3. Kuerten, J.G.M.: Subgrid modeling in particle-laden channel flow. Phys. Fluids 18,
025108 (2006)
4. Marchioli, C., Salvetti, M.V., Soldati, A.: Some issues concerning large-eddy simula-
tion of inertial particle dispersion in turbulent bounded flows. Phys. Fluids 20, 040603-1
(2008)
5. Marchioli, C., Salvetti, M.V., Soldati, A.: Appraisal of energy recovering sub-grid scale
models for large-eddy simulation of turbulent dispersed flows. Acta Mech. 201, 277–296
(2008)
6. Soldati, A., Marchioli, C.: Physics and modelling of turbulent particle deposition and
entrainment: review of a systematic study. Int. J. Multiphase Flow 35, 827–839 (2009)
7. Stolz, P., Adams, N.A., Kleiser, L.: An approximate deconvolution model for large-
eddy simulation with application to incompressible wall-bounded flows. Phys. Fluids 13,
997–1015 (2001)
Use of Lagrangian Statistics for the Direct
Analysis of the Turbulent Constitutive Equation
François G. Schmitt and Ivana Vinkovic
Abstract. Turbulence models often involve Reynolds averaging, with a closure

providing the Reynolds stress u
v
as function of mean velocity gradient d ū/dy,
through a turbulence constitutive equation (Eq. 1). The main limitation of this
linear closure is that it rests on an analogy with the kinetic theory. For this anal-
ogy to be valid there has to be scale separation. The aim of this work is to bet-
ter understand this hypothesis from a microscopic point of view. Therefore, fluid
elements are tracked in a turbulent channel flow. The flow is resolved by direct nu-
merical simulation (DNS). Statistics on particle trajectories are computed leading to
estimations of the turbulent mixing length scale and the Knudsen number. Compar-
ing the computed values to the values in the case of scale separation we may know
where and to what extent Eq. (1) is not verified. Finally, a new non-local formulation
for predicting the Reynolds stress is proposed.
1 Introduction
Diverse industrial applications often involve complex turbulent flows. A better
understanding of these flows is challenging since available turbulence models do
not always provide accurate predictions, [1], [2]. Turbulence models often provide
the Reynolds stress component u
v
as function of mean velocity gradient d ū/dy,
through a turbulence constitutive equation:
François G. Schmitt
Laboratoire d’Océanologie et de Géosciences, CNRS UMR LOG 8187, France
e-mail: Francois.Schmitt@univ-lille1.fr
Ivana Vinkovic
Laboratoire de Mécanique des Fluides et d’Acoustique (LMFA), CNRS UMR 5509, France
e-mail: ivana.vinkovic@ec-lyon.fr
320 F.G. Schmitt and I. Vinkovic
d ū
− u
v
= νT (1)
dy
The main limitation of this linear closure (Eq. 1) is that it rests on an analogy with
the kinetic theory. For this analogy to be valid there has to be scale separation
between the mean velocity variations and the turbulent Lagrangian free path whose
mean value is the turbulent mixing length. The aim of this work is to better
understand this hypothesis from a microscopic point of view.
Since turbulent transport process depend to a large extent upon the dynamics
of fluid particle motion, considerable benefit may be taken from analysing the
Reynolds stress from a Lagrangian point of view. This perspective has been explored
by Deardorff and Peskin [3] for channel flow computed on relatively coarse meshes.
Later, Bernard et al [4] and Bernard and Handler [5] investigated the mechanisms
for momentum transport in the wall region of channel flow by using ensembles of
computed particle paths with DNS. The authors showed that the gradient mecha-
nism overpredicts the Reynolds stress and that significant positive contributions to
Reynolds stress come from non-gradient transport processes.
In the present study, fluid elements are tracked in a turbulent channel flow. The
flow is resolved by DNS. Statistics on particle trajectories ending on a certain dis-
tance y0 from the wall are computed. This leads to estimations of the turbulent mix-
ing length scale and the Knudsen number. Comparing the computed values to the
values in the case of scale separation we may know where and to what extent Eq.
(1) is not verified. This can lead to improvements or new formulations of Eq. (1).
2 Estimates of the Reynolds Stress by Analogy with the Kinetic

Theory of Gases
In the following section we briefly describe how Eq. (1) is obtained when the mixing
length scale is small compared to other macroscopic length scales. Details about this
description may be found in [6]. Let us consider a steady channel flow. The stream-
wise, vertical and transverse directions will be given by x, y and z, respectively.
The corresponding velocity components will be denoted by u, v and w. Finally, the
total velocity will be given according to the Reynolds averaged decomposition by
u = ū + u
. The overline will therefore represent a Reynolds average.
The mean free path of molecules is denoted by y∗ . A molecule coming from
y − y∗ has a positive velocity v
. On the average, a molecule coming from y − y∗
collides with another molecule at the reference level y = y0 . As a result of this
collision the molecule coming from below adjusts its momentum in the x direction
to that of its new environment. This adjustment takes place by absorption of an
amount of momentum equal to M = m (ū(y0 ) − ū(y0 − y∗ )). The right-hand side of
this equation may be expanded in a Taylor series:
∂ ū 1 ∂ 2 ū
M = my∗ + m (y∗ )2 2 + ... (2)
∂y 2 ∂y
Use of Lagrangian Statistics for the Direct Analysis 321
∂ ū 1 ∂ 2 ū
If >> y∗ 2 the second and higher order terms in the expansion may be
∂y 2 ∂y
neglected. If we define l, a characteristic length scale as:
∂ ū
∂y
l= (3)
1 ∂ 2 ū
2 ∂ y2
then the condition expressed above writes as y∗ << l. If this condition is fulfilled, M
∂ ū
can be expressed as M = my∗ . Considering that per unit area and time there are
∂y
Nv0 collisions, with v0 the average free path velocity, we finally obtain −ρ u
v
=
∂ ū
ρ v0 y∗ . By denoting the viscosity ν = v0 y∗ we obtain the turbulence constitutive
∂y
equation (Eq. 1). Equation (1) can be applied if the condition y∗ << l is satisfied.
This condition may as well be given in terms of a Knudsen number defined by
Kn = y∗ /l.
In conclusion, Eq. (1) is valid if the Knudsen number is small, i.e. Kn << 1.
Finally, the gradient transport model used for establishing the Eq. (1) rests on the
hypothesis that fluid particles maintain their momentum over a small mixing length
y∗ before blending in with the surroundings. This assertion requires that the momen-
tum of fluid particles be preserved over y∗ and that y∗ must be less than the distance
over which a linear approximation to ū is satisfactory. In that case, Kn is small and
scale separation takes place. The extension to which these conditions are fulfilled in
channel flow can be explored with the present data.
3 Numerical Simulation
Particle paths were computed using DNS of channel flow at two different Reynolds
numbers, Reτ = 180 and Reτ = 590, based on the friction velocity uτ and the channel
half-width H, or Re = 3280 and Re = 12500 based on the velocity at the center of
the channel and H. Some parameters for both simulations may be found in table 1.
In the following the superscript + will denote quantities normalized by uτ and the
viscosity ν . For both Reynolds numbers 200, 000 fluid elements were tracked.
Table 1 Numerical simulation characteristics
Reynolds number Mesh points Friction velocity Viscosity
Reτ = 180 192 × 192 × 192 uτ = 0.05m/s ν = 3 × 10−3 m2 /s

Reτ = 590 256 × 384 × 384 uτ = 0.05m/s ν = 8 × 10−5 m2 /s
The resolution of the incompressible Navier-Stokes equations for a turbulent

channel flow, is based on a pseudo-spectral code using Fourier-Chebyshev expan-
sions [7]. The time integration is performed with a second-order Crank-Nicholson /
Adams-Bashford scheme. Fluid particle trajectories are computed by numerical in-
tegration of the equation of particle motion using an Adams-Bashford time ad-
vancement scheme. The Lagrangian velocity is extracted from the nearest grid
information using an eight-point Hermite interpolation.
4 Estimation of the Turbulent Mixing Length Scale

We consider here the trajectories of particles ending at a given distance y0 from
the wall. For each of these particles we know the time evolution of the Reynolds
stress u
v
(t). We then compute the autocorrelation function C(t), of the Lagrangian
product Z = u
v
(t) estimated along each trajectory and the “memory” t ∗ of each
Lagrangian stress by considering the first time t = t ∗ at which C(t) is equal to zero
(Fig. 1 left). In this way we obtain an estimation of the mixing timescale for each
trajectory. In other words, t ∗ can be considered as a characteristic timescale of fluid
particle momentum mixing. This mixing timescale can be computed for each trajec-
tory and for different ending positions y0 .
The momentum mixing timescale t ∗ is then used for determining the turbulent
mixing length scale y∗ , by considering the distance travelled during time t ∗ , (y∗ )2 =
(y0 − y(t0 − t ∗ ))2 + (x(t0 ) − x(t0 − t ∗ ))2 , where t0 and x0 are the time and position
in the longitudinal direction of the fluid element when it crosses the y0 section. y∗
represents the distance a fluid element has to cross before the information of its
momentum u
v
is lost. By estimating y∗ for all the fluid particles that are tracked
in the flow we obtain the probability density function (PDF) of y∗ as shown on the
right of Fig. 1. This PDF depends on y0 .
0
1 10
yo/H=1 ; Re=3280
yo/H=1 ; Re=12500 PDF Re=3280
0.8
f(y)=1.15 y^-0.55
PDF Re=12500
0.6
10-1
0.4
C(t+)
PDF
0.2
-2
10
0
-0.2
t*
-3
-0.4 10
0 100 200 300 400 0 500 1000 1500 2000 2500 3000
t+ y*
Fig. 1 Left - Example of the autocorrelation C(τ ), of u

v
for each particle ending at y0 = H
for Re = 3280 (squares) and Re = 12500 (triangles). Right - PDF of y∗ for all fluid elements
ending at y0 = H for Re = 3280 (squares) and Re = 12500 (triangles)
5 The Knudsen Number and the Scale Separation Hypothesis

In the kinetic theory the molecular free path is considered to be close to a mean free
path value which is very small compared to the macroscopic quantities, [6]. In this
theory, the Knudsen number that gives the ratio of the molecular mean free path to
a macroscopic characteristic length scale is, therefore, very small. Let us now check
the value of the Knudsen number in the “turbulent” analogy of this theory. The re-
sults developed in the previous section are used for the estimation of the Knudsen
number Kn = y∗ /l, where y∗ is the momentum mixing length scale previously de-
termined and l is the characteristic length scale of mean velocity variations. One has
l ∼ y0 in the log-law region.
Figure 2 (left) illustrates the values of Kn taken by several fluid elements for
y0 /H = 1. The Knudsen number decreases as the Reynolds number increases. How-
ever, it remains much larger than 1 in most of the cases. On the right of Fig. 2 we can
see the evolution of the mean Knudsen number for the same cases for different end-
ing positions y0 expressed in wall units. The evolution of the mean Knudsen number
has been adjusted by a hyperbolic law. Close to the wall the mean Knudsen number
is very high and much larger than 1 for both Reynolds numbers. Knowing that for
scale separation Kn must be smaller than 1, we can conclude that in the simple cases
studied here the hypothesis necessary for applying Eq. (1) is not verified.
6 A Non Local Turbulence Constitutive Equation

Having established in the previous section that the basic assumption of the turbu-
lent constitutive equation (Eq. 1) fails in turbulent channel flow whatever the wall
distance, it is of interest to now consider another way of predicting the Reynolds
stress.
3
10 200
yo/H=1 ; Re=3280
yo/H=1 ; Re=12500
2
10 mean Kn ; Re=3280
mean Kn ; Re=12500
150
f(y+)=2500 / y+
Knudsen number
101
mean Knudsen
100 100
-1
10
50
-2
10
-3
10 0
0 20 40 60 80 100 0 200 400 600 800
fluid elements ending at y=yo y+
Fig. 2 Left - Example of the Knudsen number Kn for different fluid elements ending at
y0 = H for Re = 3280 (squares) and Re = 12500 (triangles). Right - Mean Knudsen number
evolution with distance from the wall with its hyperbolic fit (line)
-1 -1
10 10
-2 -2
10 10
Q(y*) / mean ( y* )
Q(y*) / mean ( y* )
10-3 10-3
10-4 10-4
-5 -5
10 10
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000
y* y*
Fig. 3 Q(1) (y∗ )/y∗ (full squares) and Q(2) (y∗ )/y∗ (empty squares) for different wall
∗
distances y0 , for Re = 3280 with the exponential adjustment f (y∗ ) = −0.028e−0.06y (left)
∗
∗
and for Re = 12500 with f (y ) = −0.01e −0.2y (right)
Let us write the Reynolds stress at a distance y0 from the wall as:

u
v
(y0 ) = u
v
(y0 )dP(u
v
) (4)
Du
v
where P(u
v
) represents the joint probability of u
v
. Let us then separate up-
ward moving particles (v
> 0) from downward moving particles (v
< 0) by in-
troducing P+ the probability of having an upward moving particle (v
> 0) and
P− the probability of having a downward moving particle (v
< 0). Since over
the distance of a mixing length scale y∗ the information of u
v
is not lost, we
can assume that for upward moving particles u
v
(y0 ) ∼ u
v
(y0 − y∗ ) and that
for downward moving particles u
v
(y0 ) ∼ u
v
(y0 + y∗ ). If we now assume that

= P(u
)P(v
+ ) and P(u
, v
|v
< 0)) = P(u
)P(v
− ) knowing that
P(u , v
|v > 0))
∗

D
u (y0 − y )P(u )du = ū(y0 − y∗), we obtain:
u

u
v
= ū(y0 − y∗)w
(y0 − y∗ )P(1) (w
) − ū(y0 + y∗ )w
(y0 + y∗ )P(2) (w
) dw
Dw
where P(1) (w
) = P(v
+ ) and P(2) (w
) = P(v
− ) denote the PDF of |v
| = w
the ab-
solute value of the fluctuating vertical velocity component for upward and down-
ward moving particles respectively. We then suppose that w
∼ w
as well as
that P− P(1) (w
)dw
= P− Q(1) (y∗ )dy∗ and P+ P(2) (w
)dw
= P+ Q(2) (y∗ )dy∗ , where
Q(1) (y∗ ) and Q(2) (y∗ ) may respectively be regarded as the PDF of y∗ for downward
and upward moving fluid elements. Finally:
y∗
max
− u
v
(y0 ) = w
P− ū(y0 + y∗)Q(1) (y∗ ) − P+ū(y0 − y∗ )Q(2) (y∗ ) dy∗ (5)
0
Figure 3 contains the plots of Q(1) (y∗ )/y∗ and Q(2) (y∗ )/y∗ for different wall dis-
tances y0 and for Re = 3280 (left) and Re = 12500 (right). The computed values of
Q(1) (y∗ ) and Q(2) (y∗ ) are similar and adopt an exponential adjustment for small val-
ues of y∗ . In addition to this, computed values of Q(i) (y∗ )/y∗ graphically illustrate
that the non-local kernels proposed here do not depend on the wall distance y0 .
7 Conclusion and Further Studies

By analysing the trajectories of fluid elements in a turbulent channel flow we found
that the Knudsen number of individual fluid elements is often much larger than 1.
The individual Knudsen numbers as well as the mixing length scale can attain high
values indicating that there is no scale separation even for high Reynolds numbers.
The PDF of the mixing length scale y∗ is not a delta function. It may be adjusted by
a power law p(y) = Ay−B . These results suggest that the constitutive equation (Eq.
1) cannot represent a full and consistent answer to the closure problem. A non-local
approach for predicting the Reynolds stress was then proposed. This approach is
based on non-local kernels, whose values have been evaluated from the computed
results. The computed values graphically illustrate that the non-local kernels pro-
posed here do not depend on the wall distance y0 .
In subsequent studies it is planned to establish the influence of the Reynolds
number and of the flow configuration on the proposed non-local kernels. This ex-
panded database may reveal additional information about the relationship between
the Reynolds stress and the paths of fluid elements. In addition to this, it is in-
teresting to develop conditional sampling of particle paths used for obtaining the
Reynolds stress. For example, particle paths conditioned by sweeps or ejections may
play particularly important roles in the determination of turbulent mixing scales.
References
1. Wilcox, D.C.: Turbulent modeling for CFD, 2nd edn. La Canada, DCW Industries (1998)
2. Pope, S.B.: Turbulent flows. Cambridge University Press, Cambridge (2000)
3. Deardorff, J.W., Peskin, R.L.: Lagrangian statistics from numerically integrated turbulent
shear flow. Phys. Fluids 13, 584 (1970)
4. Bernard, P.S., Ashmawey, M.F., Handler, R.A.: An analysis of particle trajectories in
computer-simulated turbulent channel flow. Phys. Fluids 9, 1532 (1989)
5. Bernard, P.S., Handler, R.A.: Reynolds stress and the physics of turbulent momentum
transport. J. Fluid Mech. 220, 99 (1990)
6. Tennekes, H., Lumley, J.L.: A first course in turbulence. MIT Press, Cambridge (1972)
7. Buffat, M., Le Penven, L., Cadiou, A.: An efficient spectral projection method on
divergence-free subspaces for transition analysis in wall bounded flow. submitted to J.
Comp. Phys.
Numerical Simulation of Supersonic Jet Noise
with Overset Grid Techniques
J. Schulze and J. Sesterhenn
Abstract. Supersonic jets with a complex shock pattern appear in numerous

technical applications. Most supersonic jets, especially in modern military or civil
aircraft, are not perfectly expanded. Thereby, shocks are appearing in the jet core
and interacting with the turbulent mixing-layers and emanating shock induced noise.
Under certain conditions this upstream traveling noise can be amplified due to a
closed feedback loop. These so called screech tones can reach sound pressure levels
of up to 160 dB [11] and hence lead to immense noise pollution and even structural
fatigue.
The focus of this research project lies in the numerical simulation of supersonic
jet noise and finally the minimization of screech tones with an adjoint shape opti-
mization approach of the nozzle geometry. To this end the nozzle geometry, based
on a complex shape, has to be included in the computational domain. In the present
paper the method of overset grid techniques is presented for the simulation of super-
sonic jet noise. Direct numerical simulations with a modeled nozzle inlet showed a
good agreement of the screech frequency to a semi-empirical low found by Powell
in 1953 [6].
1 Introduction
Considering a supersonic jet, e.g. at the exit of a jet engine, in the over- or under-
expanded case, a regular pattern of compression and expansion waves will be found
J. Schulze
Institut für Mathematik und Rechneranwendung, Werner-Heisenberg-Weg 39,
85577 Neubiberg, Germany
e-mail: jan.schulze@unibw.de
J. Sesterhenn
Institut für Mathematik und Rechneranwendung, Werner-Heisenberg-Weg 39,
85577 Neubiberg, Germany
e-mail: joern.sesterhenn@unibw.de
328 J. Schulze and J. Sesterhenn
within the supersonic part of the jet flow. A compression wave incident on the sonic
line will be reflected as an expansion, and vice versa, cf. Fig. 1. At the location of
interaction between the compression wave and the turbulent mixing layer, acoustic
waves are generated. This shock-induced noise also plays an important role in what
is called jet screech. This phenomenon is caused by shock-induced acoustic waves
traveling upstream and forcing the transitional shear-layer at the nozzle exit. At
this point Kelvin-Helmholtz instabilities are growing to vortices, transported down-
stream and interacting with the shock tips which are emanating noise again and
closing a feedback loop. The presence of a nozzle-geometry seems to be important
since it has a leading effect for the sensitivity of the upstream traveling acoustics on
the instabilities in the mixing layers.
ise
no
p_2
nozzle
n
sio
ck
es
ho
pr
sc
m
M>1 p_1
co
n
io
ns
pa
ex
nozzle
vortices in the turbulent mixing layer
Fig. 1 Schematic view of the interaction of shock and turbulent mixing layer in a jet with
emanated noise. Over-expanded ⇒ p2 > p1
Experimental results indicate particularly high sound pressure levels of up to

160 dB and even beyond [11]. Prediction and reduction of shock-induced noise,
produced by modern civil aircraft with jet propulsion, traveling at high subsonic or
supersonic Mach numbers is a matter of particular interest. Besides pollution of the
environment by the radiated sound, the latter can also lead to high dynamic loads on
parts of the aircraft causing structural fatigue and even destroy them.
In this project, the method of direct numerical simulation will be used to compute
a supersonic, three-dimensional and rectangular jet that is not perfectly expanded, as
it is found at the nozzle exit of jet engines for aircraft. Numerical methods of high or-
der of accuracy are chosen for the direct solution of the compressible Navier-Stokes
equations, which gives us the possibility to compute the sound field, generated by
the supersonic jets, directly. The aim of this study lies in the minimization of super-
sonic jet-noise and in particular in the minimization of jet-screech. Since screech – a
phenomenon which is not yet understood in all details – seems to be affected by the
presence of the jet-nozzle, we will perform a shape optimization of the nozzle ge-
ometry to minimize the emanated noise. The optimization technique is based on the
Numerical Simulation of Supersonic Jet Noise with Overset Grid Techniques 329
adjoints of the compressible Navier-Stokes equations. These equations have to be

solved backwards in time to get the gradient information of the objective function.
The organization of the paper is as follows: In section two the numerical method
of overset grid techniques to include the nozzle geometry in the computational do-
main is presented. Results of direct numerical simulations with and without the
presence of a nozzle will be found in section three, followed by some concluding
remarks in section four.
2 Numerical Methods
To simulate the investigated flow case, the compressible Navier-Stokes equations
are solved, based on a characteristic-type formulation [9] on an orthogonal grid
stretched in both the stream-wise and the transverse directions. Along the span-wise
direction, periodicity and statistical homogeneity are assumed. Spatial discretization
is implemented using a finite difference compact scheme of sixth order [4] and a
spectral like method in the periodic direction. The compact schemes preserve the
dispersion relation even for higher wavenumbers which is an important issue for
aeroacoustic applications. With an additional compact high order filter, spurious
high frequent grid to grid oscillations are suppressed.
The time advancement is performed using an exponential integration based on
Krylov subspaces. This integration is a low-dimensional approximation of the ma-
trix exponential which represents the exact evolution operator for an autonomous
linear system [7]. The main advantage of this method lies in its high numerical sta-
bility compared to other standard explicit techniques like the Runge-Kutta method.
Hence, large time-steps can be used with CFL numbers of up to 50 and even be-
yond. Although, such a high CFL-number causes additional computational time,
since one needs to compute more subspaces, it is still up to twice as fast as a com-
parable Runge-Kutta method.
A sponge region with a low-pass filter (applied spatially) and a grid stretching
in the stream-wise direction is applied to avoid large structures interacting with the
outflow boundary and scattering spurious energy back in the computational domain.
Additionally, a boundary wave-acceleration is implemented to prevent acoustic re-
flections from the transverse boundaries. At the inflow two different boundary con-
ditions are implemented. As a first approach the rectangular nozzle at the inlet is
modeled with a laminar tanh-profile. The resulting shear-layers are forced slightly
downstream of the inflow boundary with a model-spectrum including the most un-
stable frequency of the shear-layer and its first sub-harmonic with a stochastical
phase-shifting (inlet 1). In a second approach the geometry of the nozzle is modeled
by two backwards facing Joukowsky profiles (inlet 2, cf. Fig. 2). This geometry is
included in the computational domain with overset grid techniques as explained in
section 2.1.
The code is parallelized using the Massage Passing Interface (MPI). For the cur-
rent setup up to 1020 CPU’s are used on a SGI-Altix 4700. Approximately 12 TB
of data are written to disk and 0.5 TB of main memory are used.
Fig. 2 Computational domain with some implemented numerical boundary conditions and
the two different inlet conditions (velocity profile and Joukowsky-like nozzle geometry)
2.1 Nozzle Geometry with Overset Grid Techniques

A planar and rectangular jet is used for the investigation of screech tones. As it was
mentioned before the presence of a nozzle is important. For this reason a nozzle ge-
ometry is being implemented in the present setup based on overset grid techniques
[5], [8] and [10].
The nozzle geometry is modeled by two backwards facing Joukowsky profiles as
shown in Fig. 2 (inlet 2). Each profile is surrounded by a curvilinear grid which is
orthogonal due to the conformal mapping of the Joukowsky transformation. Both
curvilinear grids are embedded in a cartesian grid where most of the jet is simu-
lated. These three grids are furthermore decomposed into sub-domains which allows
a parallelization of the code. For the parallelization the Message Passing Interface
(MPI) is used to guarantee the communication of the sub-domains. The same rou-
tines are also used to enable the communication needed for the interpolation of the
curvilinear grid and the cartesian grid. This communication has to be set up carefully
to define the processes involved in the interpolation process and to avoid commu-
nication deadlock. Additional communication at the boundaries of the sub-domains
has to be implemented to receive ghost cells of the neighboring processes. The lat-
ter one are used to interpolate points close to the sub-domain boundaries where an
interpolation stencil would extend the domain of the process.
2.1.1 Interpolation
The interpolation is performed in an interpolation zone close to the boundary of the
curvilinear grids. Depending on the order of interpolation, the width of this interpo-
lation zone has to be at lest n + 1 points, where n is the order of the interpolation. It is
proposed by [3] that a fourth order Lagrange interpolation is a good compromise be-
tween accuracy and computational cost. Based on a Lagrange interpolation in three
dimensions, the values in the interpolation zone are updated within each iteration
of the time integration method. Due to the closed feedback loop in our application
4 4
2 1 2
0.8
0.6
0 → 0.4
→ 0
0.2
−2 0
0 0.5 1 1.5 2 −2
−4 −4
−2 0 2 4 −2 0 2 4
Problem: Interpolate
1. Transformation from
values from the
Joukowsky to unit-cube. Interpolated values on
colored curvilinear → →
2. Perform interpolation on cartesian grid.
grid to the Cartesian
unit-cube.
one.
Fig. 3 Inverse Joukowsky transformation of the cartesian grid to apply a Lagrangian
interpolation
(screech) the information of the curvilinear grid and the cartesian grid need to be
interpolated in both directions. Thus, a two-way coupling is implemented.
To increase the numerical stability of the interpolation method the Lagrange in-
terpolation is written in a barycentric formulation. A two-dimensional barycentric
Lagrange interpolation can be written as:
m n
∑ ∑ fi, j μi (x)μ j (y)
j=0 i=0
pnm (x, y) = m n (1)
∑ ∑ μi (x)μ j (y)
j=0 i=0
ωi
with μi (x) = , the interpolated value pnm (x, y) at the point (x, y) and the
x − xi
1
barycentric weights ωi = ; (analogous for μ j (y)), and n, m the order
∏ (xi − xk )
k=i
of the interpolation in the x and y direction, respectively.
Since the geometry in our case is time independent, the values μi (x) and μ j (y)
need to be computed only once and can be saved throughout the simulation to reduce
computational time.
Anyway, a direct interpolation of the values of the curvilinear grid on the carte-
sian grid is not possible since one needs a cartesian domain for the interpolation
stencil. To interpolate a value from the curvilinear grid on the cartesian grid the sten-
cil on the curvilinear grid needs to be in a cartesian domain (computational domain).
These cartesian values are computed using an inverse Joukowsky transformation of
the curvilinear grid. This is a different approach to [2] and [5] but fast and accurate
since no system of equations has to be solved. Anyway, this approach is restricted to
applications where an analytic solution to the inverse transformation of the domain
exists. This is the case for a standard Joukowsky transformation. An example how
the inverse transformed domain looks like is presented in Fig. 3. See also [8] for
a more detailed description and some validation of the implemented interpolation
method.
3 Results
In the present research project two different numerical inlet conditions are
implemented (cf. Fig. 2).
3.1 Nozzle Geometry (Inlet 2)

A two-dimensional supersonic simulation with included nozzle (inlet 2) is presented
in Fig. 4 with a Reynolds number based on the jet diameter of ReD = 30 000, and
a jet Mach number MJ = 1.35. To obtain oblique shocks in the jet core, a pressure
ratio of pJ /p∞ = 1.2 (under-expanded) is adjusted.
(a) (b)
Fig. 4 Two-dimensional supersonic case for the jet computation with overset grid techniques.
Mach number MJ = 1.35; Reynolds number based on the jet diameter ReD = 30 000. White
lines border the sub-domains for the MPI domain decomposition. (a) Absolute velocity,
Contour-lines: stream-wise vorticity. (b) Divergence of the velocity, Contour-lines: stream-
wise vorticity
The cartesian grid in the background is decomposed in 256 sub-domains,

bordered by weak white lines and contains in total 1024 × 1024 grid points. Each of
the two computational domains around the nozzle-profiles is as well decomposed in
each with 32 sub-domains containing 512 × 256 grid points. Hence, a total of 320
CPU-cores are used for this two-dimensional computation. The nozzle geometry is
partly shown as two backwards facing Joukowsky profiles in black color.
In Fig. 5 the interpolation zones for the current two-dimensional simulation are
visualized. Since there is a two-way coupling, two interpolation zones are needed.
The interpolation zone close to the nozzle geometry interpolates the flow vari-
ables from the body fitted grid to the cartesian one. The second interpolation zone
Fig. 5 Visualization of the interpolation zones and the cartesian grid in the background. Only
every fourth point is shown
interpolates the values the other way round. Both on the body fitted grid and on
the cartesian and only every fourth point is shown. Due to a grid stretching on all
grids one has to adjust the size an position of the interpolation zones carefully to
guarantee the minimum number of required points in the interpolation zones. The
cartesian points inside the inner interpolation zone have no physical meaning since
they belong to the simulation of the body fitted grid and can be neglected. Any-
way, to obtain a numerically stable solution one needs to treat them in a special
way. In the present study a simple but efficient method is used where the right hand
side (∂ u/∂ t) of all interior points is set to zero. The contour-lines represent the
stream-wise vorticity. One can see, that even large structures, like the recirculation
bubble close to the nozzle, which extends both interpolation zones, are well re-
solved. In Fig. 4(b) we can see, that this also holds for steady shocks extending the
interpolation zones.
In Fig. 4(a) the absolute velocity is visualized including the contour lines of the
stream-wise vorticity. The emanated noise is clearly visible in Fig. 4(b) where the
divergence of the velocity is shown again including the contour lines of the stream-
wise vorticity. Acoustics directed in the downstream direction caused by the turbu-
lence in the mixing layers and also shock induced noise directed in the upstream
direction can be identified. In addition to this a complex shock pattern is visible in
the jet core.
(a) (b)
pressure field absolute velocity (|u|); inverse color scaling.
Fig. 6 (a): x-y-plane of the pressure field, visualization of the shock-cells. (b) absolute
velocity (|u|)
3.2 Velocity Profile (Inlet 1)

The shown results in Fig. 6 are based on a three-dimensional planar jet configuration
with a Reynolds number (based on the jet height) of 30 000. To obtain a supersonic
behavior, the Mach number in the jet core is 1.35 with a uniform subsonic co-flow
of M = 0.2025. Furthermore, the jet is under-expanded to achieve shocks and con-
sequently shock induced noise. In Fig. 6(b) the absolute velocity of the jet in a plane
normal to the span-wise direction is presented. Based on the high Reynolds num-
ber, the initially laminar shear-layers undergo an early transition to turbulence about
3δω0 (shear-layer thickness) downstream of the forcing fringe. The shocks in the jet
core interact with the turbulent mixing layers and are reflected at the sonic line to
form the so called shock cell diamonds (cf. also Fig. 1). In Fig. 7 the spectrum of
the emanated acoustic signal is shown. It is measured close to the jet exit to cap-
ture mainly the upstream traveling noise, like shock induced noise. We identify a
low frequent peak at fs ≈ 8 × 10−2 f0 which corresponds to the screech tone. The
reference frequency f0 is the most unstable frequency of the mixing layer:
uc
f0 = 0.132 (2)
δω0
with the mixing layer thickness δω0 and the velocity uc = (ujet core − u∞ )/2. In
addition to this a semi-empirical analytic frequency fs for the screech tone found
by Powell [6] in 1953 is included in Fig. 7 (a) (black bullet •). This frequency can
be obtained by the equation:
Uc
fs = (3)
s(1 + Mc )
with the convective speed of the hydrodynamic disturbances Uc , its corresponding
Mach number (Uc /c) and the shock spacing s. Although this equation only predicts
the screech frequency and not its amplitude and the correct determination of the
120
100
80
SPL [dB]
60
40
20
0
10−2 10−1 100 101 102
f /f0
Fig. 7 The spectrum of the acoustic signal of the jet measured in the acoustic near-field close
to the jet exit to capture the shock induced noise. The peak at f s ≈ 8 × 10−2 f0 indicates the
screech tone. The black bullet • is the screech frequency form a semi-empirical low (Eq. (3))
found by Powell [6]
1.5
1
0.5
p /pmax
0
-0.5
-1
-1.5
21 22 23 24
t/Ts
Fig. 8 Pressure history on both sides of the jet in the upstream direction. Upper side
( ); Lower side ( )
parameters s and Uc is not clearly defined, it is still in good agreement the DNS data
of the current simulation.
The corresponding pressure history is depicted in Fig. 8 measured close to the
nozzle exit at the upper and lower side of the planar jet. The fluctuations of the two
locations seem to be 180 degrees phase shifted which is a typical phenomenon for
screech tones [1].
As mentioned above, the presence of a nozzle is important for the screech phe-
nomenon. Since there is no nozzle implemented in the present study, one needs
to force the instabilities artificially. To close the feedback loop, the acoustic signal
is measured close to the jet exit and superimposed to a forcing in a fringe region
around the mixing layers. The forcing signal consists now on three parts: a har-
monic part including f0 with a model spectrum, a random phase shifting and finally
the aforesaid screech component.
4 Conclusion
Different numerical approaches have been used to investigate the screech tones of
planar supersonic jets. A three-dimensional DNS of a jet without the presence of a
nozzle produced screech tones which are in good agreement with a semi-empirical
low found by Powell ([6]).
An improved implementation with an artificial closing of the feedback loop, by
triggering the instabilities in the mixing layers, yield to the desired screech tones.
Since the presence of the nozzle seems to be of particular importance to close the
feedback loop, a complex geometry for the nozzle based on overset grid techniques
is implemented and will de used for future screech tone investigations. A two di-
mensional case to test the interpolation method showed good results with a small
relative error.
Acknowledgements. The authors like to thank the DEISA Consortium (co-funded by the
EU, FP6 project 508830), for support within the DEISA Extreme Computing Initiative
(www.deisa.org) and the DFG/CNRS research group FOR 508 (Noise Generation in
Turbulent Flows) for support within this project.
References
1. Berland, J., Bogey, C., Bailly, C.: Numerical study of screech generation in a planar
supersonic jet. Physics of fluids 19 (2007), doi:10.1063/1.2747225
2. Desquesnes, G., et al.: On the use of a high order overlapping grid method for coupling
in CFD/CAA. J. Comp. Phys. 220, 355–382 (2006)
3. Guenanff, R., Sagaut, P., Manoha, E., Terracol, M., Lewandowski, R.: Theoretical
aspects of a multi-domain high-order method for CAA. AIAA Paper 2003-3117 (2003)
4. Lele, S.: Compact finite difference schemes with spectral-like resolution. J. Comp.
Phys. 103, 16–42 (1992)
5. Marsden, O., Bogey, C., Bailly, C.: High-Order Curvilinear Simulations of Flows Around
Non-Cartesian Bodies. J. Comp. Aeroacoustics 13(4), 731–748 (2004)
6. Powell, A.: On the mechanism of choked jet noise. Proc. Phys. Soc. London B66,
1039–1056 (1953)
7. Schulze, J., Schmid, P., Sesterhenn, J.: Exponential time integration using Krylov sub-
spaces. Int. J. Numer. Meth. Fluids (2008), doi:10.1002/fld.1902
8. Schulze, J., et al.: Numerical Simulation of Supersonic Jet Noise. Num. Sim. Turbulent
Flows & Noise Generation, NNFM 104, 29–46 (2009)
9. Sesterhenn, J.: A characteristic-type formulation of the Navier-Stokes equations for high
order upwind schemes. Computers & Fluids 30(1), 37–67 (2001)
10. Sherer, S., Scott, J.: High order compact finite-difference methods on general overset
grids. J. Comp. Phys. 210, 459–496 (2005)
11. Tam, C., Ahuja, K., Jones III, R.: Screech Tones from Free and Ducted Supersonic Jets.
AIAA J. 32 (5), 917–922 (1994)
Large Eddy Simulation of Turbulent Jet Flow
in Gas Turbine Combustors
Y. Shimada, B. Thornber, and D. Drikakis
Abstract. Implicit Large Eddy Simulation (ILES) in conjunction with high

resolution and high order computational modelling was applied to a turbulent mix-
ing jet of a fuel injector in gas turbine combustors. In the ILES calculation, the
governing equations for three dimensional, single phase, nonreactive multi-species
compressible flow were solved using a finite volume Godunov method. A fifth-order
accurate methods was used to achieve high order spatial accuracy and a second order
explicit scheme was applied for time integration. Comparison of mean and fluctu-
ating velocity components and mixture fraction with experiment and conventional
LES demonstrated that the ILES successfully captured the turbulent flow structures
without explicit subgrid scale modelling.
1 Introduction
As the performance of the primary combustor of a gas turbine engine depends
strongly on fuel injection, a variety of fuel jet flows have been numerically and
experimentally investigated. In the context of numerical research, with the aid of
recent developments in computational power, Large Eddy Simulation (LES) has
been widely applied for time dependent flows. However, conventional LES includes
some essential problems, e.g., commutation errors in the filtering operation and the
development of a universal subgrid scale model. Furthermore, generally applicable
subgrid scale models for multi-species compressible flow do not exist.
Implicit Large Eddy Simulation (ILES) techniques together with high resolution
methods were developed to overcome the disadvantages of the conventional LES
[2]. In ILES computations it is assumed the numerical discretisation itself accounts
for the unresolved scales, in other words implicitly separates scales, hence setting
Y. Shimada · B. Thornber · D. Drikakis
Department of Aerospace Sciences, School of Engineering, Cranfield University
Cranfield, Bedford, MK43 0AL, UK
e-mail: y.shimada@cranfield.ac.uk
338 Y. Shimada, B. Thornber, and D. Drikakis
explicit filters for determining the subgrid scale is unnecessary. Recently, ILES has
shown good agreement in simulations of basic flows [3, 5, 13, 14], however, only
few research [4] applied ILES to practical, engineering related jet flow.
The present work concerns numerical computations of multi-species practical jet
flows using recently developed high resolution Godunov methods which have been
incorporated in the computational fluid dynamics code CNS3D [2].
2 ILES Numerical Methods

The three dimensional compressible equations for single phase nonreactive flow
comprising flow mass, momentum and energy conservation laws, are employed.
The quasi-conservative model by [1] was applied to model multi-species mixing.
Assuming instantaneous pressure equilibrium assumption in a single cell, i.e., both
fuel and oxidiser in the mixture have the same velocity and pressure, the governing
equations are given by
∂ αf ρf
+ ∇ · (αf ρf u) = 0 , (1)
∂t
∂ αo ρo
+ ∇ · (αoρo u) = 0 , (2)
∂t
∂ ρu
+ ∇ · (ρ uu) = −∇ · P , (3)
∂t
∂ ρE
+ ∇ · (ρ E + p)u = −∇ · (P · u) − ∇ · q , (4)
∂t
∂α
+ ∇α · u = 0 , (5)
∂t
where u, α , ρ and E denote the velocity vector, volume fraction, density and total
energy per unit volume, respectively; the subscripts f and o designate the species for
fuel and oxidiser. The stress tensor P contains pressure and viscous effect,
2
P = pI + μ (∇ · u)I − μ ((∇u) + (∇u)T ) , (6)
3
where p, I and μ , stand for pressure, unit tensor and dynamic viscosity coefficient,
respectively. The heat flux q follows Fourier’s equation, q = −κ ∇T , where κ and
T are the thermal conductivity coefficient and temperature, respectively. A perfect
gas equation of state is used to close the system.
The governing equations are solved using a finite volume Godunov method. To
achieve high-order spatial accuracy, the fifth-order accurate Monotonic Upstream-
centred Scheme for Conservation Laws (MUSCL) reconstruction method [9] is
applied. For time integration, the second order explicit Runge-Kutta method is em-
ployed. In the reconstruction process, the velocity components are locally modified
Large Eddy Simulation of Turbulent Jet Flow in Gas Turbine Combustors 339
to prevent overly dissipative behaviour of the turbulent kinetic energy (TKE) which
is present in standard upwind scheme [15]. For this scheme the leading order of the
truncation error which acts as a subgrid stress in the ILES approach is given by [15],
Δ x5
ε= min(|u| , a)ux uxxxxx , (7)
60
where, Δ x, a and u denote grid spacing, speed of sound and velocity in x direction,
respectively; the subscript x denotes a derivative with respect to x direction.
3 The Turbulent Mixing Jet Flow

The jet injector geometry shown in Fig. 1 was investigated experimentally [12] and
selected as the test problem for the Turbulent Nonpremixed Flame (TNF) workshop
[6]. The fuel is injected in the axial direction through the round nozzle with diameter
3.6 mm located at the centre of a cylindrical bluff body with diameter 50.0 mm. The
oxidiser is supplied in the axial direction from an annulus inlet with width 5.0 mm
surrounding the bluff body. Constant air flow surrounding the injector was supplied
as co-flow. The mean inflow velocity conditions are 61 m/s, 20 m/s and 20 m/s for
fuel, oxidiser and co-flow, respectively. Air was selected for both fuel and oxidiser
in the study of flow structure and velocity profiles, while methane was selected as
fuel for an evaluation of multi-species mixing flow. The Reynolds number based on
the fuel jet velocity and the fuel nozzle diameter is 1.45 × 104, and the local Mach
number is 0.18 for air and 0.14 for methane flow.
The computational domain was set to have the same size as in [7]. Axial and
radial lengths are 5D and 8.8D, respectively, where D is the diameter of the bluff
body. To investigate grid convergence, the computations were carried out with three
grid resolutions. The coarse grid has 60 axial, 62 radial and 32 circumferential cells,
Fig. 1 Schematic view of the jet flow Fig. 2 Instantaneous snapshot of fuel volume
fraction in y/D = 0 plane on the medium grid
0
10
10-1
Turbulent kinetic energy
-2
10
-3
10
10-4
-5
10 2.00×106
-6
5.95×105
10 1.19×105
-7 k-5/3
10
10-8
102 103 104
Wave number k
Fig. 3 Kinetic energy spectra on three dif- Fig. 4 Vorticity magnitude in y/D = 0
ferent grid resolution with the ideal k−5/3 plane on the medium grid; values are nor-
line (dotted) malised by the fuel inflow velocity and the
smallest cell length
thus resulting in 1.19 × 105 cells, in total; the medium grid is 300 × 62 × 32 and
5.95 × 105; and the fine grid is 480 × 65 × 64 and 2.00 × 106 cells, respectively.
A fluctuating velocity component was added to the fuel and oxidiser inlet based
on previous direct numerical simulations [8].

The instantaneous fuel volume fraction profile is shown in Fig. 2, illustrating the
large scale flow structure. Note that air was applied as fuel in this computation. The
central fuel jet forms a laminar flow up to z/D = 0.4 and breaks up at z/D = 1.0,
where highly turbulent region can be seen. Recirculation zones with low volume
fraction can be observed under z/D = 0.7.
The velocity spectrum was examined to validate the representation of the small
scales, particularly unresolved scales of the flow, since the unresolved scale is not
explicitly represented in ILES computation. The one dimensional TKE spectrum
was computed using a velocity time history at a fixed spatial location; z/D=1.0 on
the centre line. Figure 3 shows the result based on the radial velocity components
on the coarse, medium and fine grid. An approximate inertial range in which the
slope of the line should be parallel to k−5/3 [10] can be seen. In the coarse grid,
the ‘cut off ’wave number which describes the deviation point from the k−5/3 line
is around 1.5 × 103, and the resolution of TKE was estimated to be 88% based on
the assumption of the k−5/3 line to the Kolmogorov scale in the real flow field. As
the grid resolution increases the inertial range extends the cut off wave number up
to an order of 104 and the resolution of TKE increases to be 92% for the medium
and 97% for the fine grid, respectively. Note that similar results were obtained with
the axial velocity component. From the vorticity magnitude illustrated in Fig. 4
80 80 80 z/D=0.8
z/D=0.06 z/D=0.4
Mean axial velocity (m/s)
ILES
60 LES 60 60
Exp.
40 40 40
20 20 20
0 0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
x/D x/D x/D
Axial rms velocity fluctuation (m/s)
20 z/D=0.06 20 z/D=0.4 20 z/D=0.8
15 15 15
10 10 10
5 5 5
0 0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
x/D x/D x/D
Fig. 5 Mean axial velocity (top) and axial rms velocity fluctuation (bottom) on the medium
grid. The solid line is ILES (CNS3D), the dot line is LES [7] and the square marks are
experiment [11]
8 8 8 z/D=0.8
z/D=0.06 z/D=0.4
Mean radial velocity (m/s)
ILES
LES
4 Exp. 4 4
0 0 0
-4 -4 -4
-8 -8
0 0.1 0.2 0.3 0.4 0.5 0.6 -80 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
x/D x/D x/D
Radial rms velocity fluctuation (m/s)
15 z/D=0.06 15 z/D=0.4 15 z/D=0.8
10 10 10
5 5 5
0 0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
x/D x/D x/D
Fig. 6 Mean radial velocity (top) and radial rms velocity fluctuation (bottom) on the medium
grid. The solid line is ILES (CNS3D), the dot line is LES [7] and the square marks are
experiment [11]
1 0.2
Rms mixture fraction fluctuation

z/D=0.4 z/D=0.4
Mean mixture fraction

0.8 ILES
Exp.
0.6
0.1
0.4
0.2
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
x/D x/D
Fig. 7 Mean (left) and rms fluctuation (right) of mixture fraction on the medium grid. The
solid line is ILES (CNS3D) and the square marks are experiment [11]
the location of the eddies containing maximum vorticity is close to the centreline,
x/D = 0. With the assumption that the eddy containing maximum vorticity produces
maximum velocity frequency, the corresponding maximum wavenumber produced
by the eddy rotation was calculated as 0.8 × 103. Since Fig. 3 shows that the cut off
wave number is above 0.8 × 103 , the ILES accounted for the energy transportation
between large and small eddies sufficiently up to this frequency.
To obtain mean profiles of the field, the instantaneous data was temporally and
spatially (circumferentially) averaged. Results of mean axial velocity and axial root-
mean-square (rms) velocity fluctuation shown in Fig. 5 match well to experimental
data [11] and are comparable to the conventional LES results [7] using Smagorin-
sky’s subgrid scale model with a second order central scheme on 5.75 × 105 grid do-
main. Figure 6 shows mean radial velocity and radial rms velocity fluctuation and,
in general, the ILES results show good agreement with the reference data. Only in
the near centreline area x/D = 0, particularly near the bluff body region z/D = 0.06,
can substantial discrepancy of the rms velocity fluctuation be observed. As previous
researchers pointed out, the flow profiles are sensitive to the fluctuation levels in the
inflow [7], hence a better setup of the velocity fluctuation component at the inflow
should improve the match to reference values in this area.
For a multi-species mixing flow analysis, the mixture fraction was computed.
Figure 7 shows temporally and spatially averaged mixture fraction and rms mix-
ture fluctuation profile, respectively. Note that the experimental data was obtained
for compressed natural gas (CNG) [11], while simulations were carried out with
methane gas. It can be implied that the ILES of multi-species mixing flow shows
good agreement with the experiment data.
5 Conclusions
An ILES technique based on high resolution and high order methods was used to
simulate a turbulent mixing jet. Analysis demonstrates that the simulation captured
a fully developed turbulent flow structure without explicit subgrid scale modelling.
The one dimensional TKE spectrum was examined and a k−5/3 decay of energy
could be seen. The resolution of TKE was estimated to be 88%, 92% and 97% with
1.19 × 105, 5.75 × 105 and 2.00 × 106 cells, respectively. Temporally and circumfer-
entially averaged velocity and rms of velocity fluctuation showed good agreement
with experimental data and conventional LES results in most parts of the flow fields
where discrepancies can be attributed to the quality of the inlet boundary conditions.
The mixture fraction profiles also matched well to experimental data.
References
1. Allaire, G., Clerc, S., Kokh, S.: A five-equation model for the simulation of interfaces
between compressible fluids. J. Comput. Phys. 181(2), 577–616 (2002)
2. Drikakis, D.: Advances in turbulent flow computations using high-resolution methods.
Prog. Aerosp. Sci. 39(6-7), 405–424 (2003)
3. Drikakis, D., Fureby, C., Grinstein, F., Youngs, D.: Simulation of transition and turbu-
lence decay in the Taylor-Green vortex. J. Turbul. 8(20) (2007)
4. Fureby, C.: ILES and LES of complex engineering turbulent flows. J. Fluids
Eng. 129(12), 1514–1523 (2007)
5. Grinstein, F.: Recent progress on monotone integrated large eddy simulation of free jets.
JSME Int. J. B-Fluids Therm. Eng. 49(4), 890–898 (2007)
6. Kempf, A.: Aspects of LES quality. In: Proceedings of the Ninth International Work-
shop on Measurement and Computation of Turbulent Nonpremixed Flames, pp. 80–111
(2008)
7. Kempf, A., Lindstedt, R.P., Janicka, J.: Large-eddy simulation of a bluff-body stabilized
nonpremixed flame. Combust. Flame 144(1-2), 170–189 (2006)
8. Kim, J., Moin, P., Moser, R.: Turbulence statistics in fully developed channel flow at low
Reynolds number. J. Fluid Mech. 177(1), 133–166 (1987)
9. Kim, K., Kim, C.: Accurate, efficient and monotonic numerical methods for multi-
dimensional compressible flows: Part II: Multi-dimensional limiting process. J. Comput.
Phys. 208(2), 570–615 (2005)
10. Kolmogorov, A.: A refinement of previous hypotheses concerning the local structure
of turbulence in a viscous incompressible fluid at high Reynolds number. J. Fluid
Mech. 13(1), 82–85 (1962)
11. Masri, A.: Bluff body and swirl flames database,
http://www.aeromech.usyd.edu.au/thermofluids/main_frame
(cited March 3, 2009)
12. Masri, A., Kelman, J., Dally, B.: The instantaneous spatial structure of the recirculation
zone in bluff-body stabilized flames. In: Symp. Int. Combust., vol. 27(1), pp. 1031–1038
(1998)
13. Thornber, B., Drikakis, D.: Implicit large-eddy simulation of a deep cavity using high-
resolution methods. AIAA J. 46(10), 2634–2645 (2008)
14. Thornber, B., Mosedale, A., Drikakis, D.: On the implicit large eddy simulations of ho-
mogeneous decaying turbulence. J. Comput. Phys. 226(2), 1902–1929 (2007)
15. Thornber, B., Mosedale, A., Drikakis, D., Youngs, D., Williams, R.: An improved
reconstruction method for compressible flows with low Mach number features. J. Com-
put. Phys. 227(10), 4873–4894 (2008)
Computations of the Flow around a Wind
Turbine: Grid Sensitivity Study and the
Influence of Inlet Conditions
R.Z. Szasz and L. Fuchs
Abstract. The flow around a complete model wind turbine is computed using LES
and the immersed boundary method. The influence of inlet velocity profile and tur-
bulence level is evaluated. The inlet velocity profile was found to have major in-
fluence in the upper regions of the flow field. The imposed turbulence level had
no major influence on the turbulent spectra but it’s effect is clearly seen on the
generated vortices.
1 Introduction
Due to the decreasing fossil fuel resources there is an intensive research to find
alternative energy sources. One promising possibility is the use of wind energy
which has a further advantage that it does not contribute at all to the chemical pol-
lution of the environment. The wind turbines used for the conversion of the kinetic
energy of the wind to electricity are getting larger and larger. This size increase is
associated with certain problems.
The major disadvantage of wind turbine parks is their negative impact on the
nearby community. Besides the visual factors, a crucial issue is the noise generated
by wind turbines. There are several studies showing that wind turbines cause low
frequency very annoying noise in their vicinity [2, 12, 10]. The low frequency noise
is generated by the motion of the blades, the tower and the generator housing.
To determine the noise, it is necessary to understand first the flow around the
wind turbines. Since detailed flow measurements are difficult and expensive even
at reduced scales, it is desired to predict the flow field numerically. The first flow
R.Z. Szasz
Lund University, Dept. Energy Sciences, LTH, PO.Box. 118, 22100, Sweden
e-mail: robert-zoltan.szasz@energy.lth.se
L. Fuchs
Linné Flow Centre, Royal Institute of Technology, Stockholm, Sweden
e-mail: lf@mech.kth.se
346 R.Z. Szasz and L. Fuchs
computations related to wind turbines focused on the prediction of the lift and drag
coefficients along a rotor blade and relatively good predictions could be obtained
even with simple methods. In some conditions however, the flow around the blade is
detached and highly three-dimensional. As it is concluded by Hansen et al. [4] even
RANS-based methods fail to accurately predict the flow in such conditions. Le Pape
and Lecanu [9] for example computed the flow around a stall-regulated wind turbine
using the k-ε model and obtained an important underestimation of the delivered
torque due to the premature stall prediction at intermediate wind speed. Johanssen
et al. [6] computed the flow around the same geometry with the help of detached
eddy simulations (DES). Even if more three-dimensional structures were captured
(compared to RANS computations), there were no particular improvements in the
prediction of global blade characteristics. As it results from the review articles of
Crespo et al. [1] and Vermeer et al. [13], just a few years ago there was a lack of LES
computations dealing with wind turbine wakes. LES begun to be applied in recent
years, but with simplifications, e.g. in Ivanell [5] the turbine tower is neglected.
Here we focus on the flow field around a complete wind turbine with a prescribed
angular speed of the rotor. Based on the grid sensitivity study, we shall discuss the
impact of grid resolution on the accuracy of the flow. A complicating issue of wind
turbine computations are the indeterminate boundary conditions. We will evaluate
the impact of inlet conditions by comparing a case with flat inlet velocity profile
with a more realistic exponential profile case. Additionally, the influence of inlet
turbulence will also be evaluated.
2 Numerical Methods
The incompressible Navier-Stokes equations are solved using finite differences. The
grid is cartesian equidistant with the possibility of having local refinements. Third
and fourth order accurate spatial discretization schemes are used for the convec-
tive and diffusive parts, respectively. This high order accuracy is achieved with the
defect correction approach. Turbulence is accounted for using LES, without any
explicit sub-grid scale (SGS) model, thus it can be considered as “under-resolved
DNS”. As it is shown e.g. in [8], for sufficiently high grid resolution, the numerical
dissipation of the third order upwind scheme is of the same order of magnitude as
the sub-grid scale terms. Multigrid is used to accelerate convergence of the iterative
process in each time-step, utilizing four global grids and a local refinement grid in
the wind turbine region. The code is parallelized based on the domain decomposi-
tion approach, using the MPI library.
A complicating issue when handling the wind-turbine is the fact that there are
moving components (the blades) and stationary parts. The near surface (blade, tower
and ground) grid resolution must be rather fine for acceptable accuracy of the LES.
We use the “Immersed Boundary Method” (IBM) c.f. [11] to account for solid ob-
jects. The approach is implemented in our in-house LES well validated code c.f.
[8, 3]. To compute the angular speed of the blades on-line one would need to in-
tegrate the forces acting on the blade every timestep. Furthermore one needs to
Computations of the Flow around a Wind Turbine 347
know the resistance of the electric generator. Here, we use a simpler approach by
prescribing a constant angular speed of the rotor. This simpler approach can be jus-
tified by the fact that even in real-life situations the angular speed of the rotor is
often limited by a control mechanism.
3 Problem Set-Up
We compute the flow field around a model three-bladed wind turbine. All lengths
are non-dimensionalized with the height of the wind turbine-tower. The set-up of the
computational domain and the coordinate system used herein are shown in Fig. 1.
The domain is divided into four blocks, these blocks being computed in parallel on
four processors. Each block has the size of 2 × 2 × 1.625 units. The wind turbine is
located in the second block, centrally in Y direction. The length of the blades is 0.7
units. At the inlet a constant wind speed of 10 m/s is imposed. The angular speed of
the rotor is set to ω = 0.39 rot/s. At the outlet flux-conserving zero-gradient condi-
tions are used. To evaluate the influence of outlet boundary conditions, an additional
case with a length of the domain extended with 25% has been computed. Compar-
isons of the average and rms axial velocity profiles one length unit downstream the
turbine (not shown) resulted in practically superimposed plots. As a conclusion, the
presently applied zero gradient boundary conditions are located far enough to not
disturb the flow field in the regions of interest. Slip conditions are imposed at the
sides.
6.5
R 0.7
1
2
1
2.1
X
Z
Fig. 1 The geometry and coordinate system Fig. 2 The geometry used to study the grid
used in the computations sensitivity
4 Results
In the followings a grid sensitivity study will be presented first. Next, the main
features of the flow for the base case will be shown. Finally, the influence of inlet
conditions will be evaluated.
4.1 Grid Sensitivity

To assess the errors associated to grid resolution, a grid sensitivity study has been
carried out as a first step. To reduce the computational resources needed for the
grid sensitivity study, only a part of the computational domain is accounted for,
shown in Fig. 2. The highest resolution being required around the turbine blades, in
the grid sensitivity study a single, non-rotating turbine blade was considered. Four
different grid resolutions have been used, having grid spacing of 0.025, 0.0125,
0.00625 and 0.003571 length units, denoted in the followings by G0–G3. The cell
size on the finest grid is not exactly half of the next finest grid resolution to keep the
computational costs on reasonable levels.
Figure 3 shows the average (top) and RMS (bottom) axial velocity profile along
a horizontal transversal line, located in the center of the domain (x = 0.5, z = 0.3).
One can observe a significant difference between cases G0 and G1. An inspection
of the flow field (not shown) revealed that the wake of the blade in the coarsest (G0)
case was extended until the outlet boundary, while in the rest of the computations
it ended further upstream. As it regards the two finest grid resolutions, the error
in the average and rms axial velocity profiles between the two grids was 1.45%
and 15.95%, respectively. These errors were considered acceptable, thus in the
computations grid resolutions corresponding to G2 have been used.
0.25
0.2
W AVG
0.15 G0
0.1 G1
G2
0.05 G3
0
0 0.1 0.2 0.3 0.4 0.5
0.007
0.006 G0
G1
0.005
W RMS
G2
0.004 G3
0.003
0.002
0.001
0
0 0.1 0.2 0.3 0.4 0.5
y
Fig. 3 Average and RMS axial velocity for four grid resolutions along a horizontal transversal
line in the center of the domain
4.2 Main Flow Features

Figure 4 shows an instantaneous snapshot of the axial velocity field in a longitudinal
plane crossing the tower of the wind turbine. For clarity, the wind turbine geometry
is also shown. One can clearly observe the wake behind the tower which has a longer
extent close to the ground. Additionally, one can follow the tip vortices and the trace
of the turbine blades as they are convected downstream.
The average axial velocity magnitude in the same cross-section is shown in
Fig. 5. Due to the tip vortices, the largest magnitudes of the axial velocity are lo-
cated along a line parallel to the ground, at an elevation corresponding to the highest
position of the blade tip. Negative values of the axial velocity indicate the presence
of recirculation zones behind the generator housing and the bottom of the tower.
The wake downstream the generator housing is considerably longer as compared to
the rest of the wake, which is due to several factors. First, the tip vortices accelerate
the flow in axial direction. Without the tower, one would expect similar values of
the axial velocity at an elevation corresponding to the lowest position of the blade
tip as the values found at the highest position. Nevertheless, due to the interaction of
the tower wake and tip vortices, even if there is a local maximum, its value is only
approximately 80% of the peak average axial velocity. The same interaction leads
also to the shortening of the wake in the central regions of the tower.
Fig. 4 Instantaneous axial velocity in the ax- Fig. 5 Average axial velocity in the axial
ial symmetry plane, base case symmetry plane, base case
4.3 Boundary Condition Effects

4.3.1 Inlet Velocity Profile
The flat velocity profile case is compared to a case where a realistic atmospheric
boundary layer profile [12] is imposed according to Eq. (1).
w(x) = 0.2561317x0.3 (1)

where the constants were chosen to have the same average axial velocity for a height
of one length unit. Figure 6 shows the average axial velocity profile along vertical
lines located at the inlet (left) and one unit downstream the turbine (right). The figure
shows that the major impact of the inlet velocity profile occurs in the upper region
of the domain.
Z=0 Z=3.1
2 2
1.5 1.5
X
1 1
0.5 0.5
Base Base
ABL ABL
0 0
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
W AVG W AVG
Fig. 6 Comparison of the axial velocity profile along lines parallel to the tower, located at
the inlet (left) and one length unit downstream the turbine (right)
4.3.2 Inlet Turbulence

To account for wind fluctuations, the base case has been compared with a case where
10% inlet fluctuations have been imposed using the digital filtering technique de-
scribed by Klein et al. [7]. Figure 7 shows the turbulent kinetic energy spectra for
the base (continuous line) and turbulent inlet (dashed line) cases at a monitoring
point located upstream the wind turbine 0.6 units upstream, at an elevation corre-
sponding to the highest position of the blade tip. The frequencies are normalized
with the blade passage frequency. The base case spectrum shows clear peaks corre-
sponding to the blade passage frequency and its harmonics. The turbulent inlet case
is characterized by considerably higher levels of fluctuations following the −5/3
line in the frequency interval 1.25–3.75. The relatively low with of the resolved part
of the inertial subrange is due to grid resolution effects. Since the turbulent inlet case
is characterized by fluctuation amplitudes approximately one order of magnitude
larger than the largest amplitudes in the base case, the influence of rotating blades is
not visible at this monitoring point. Further downstream, the influence of inlet tur-
bulence is minimal. Figure 8 plots the turbulent kinetic energy spectra for the two
cases at a point located one length unit downstream the turbine. One can observe that
the two spectra are basically superimposed, indicating that the structures generated
by the rotating blades dominate the ones imposed at the inlet. Indeed, comparing to
Fig. 7, the fluctuations in Fig. 8 are several orders of magnitude larger. Furthermore,
one can observe that at one length unit from the turbine (Fig. 8) the blade passage
frequency is not dominating any more the turbulent kinetic energy spectrum.
The vortices generated by the interaction of the wind and the wind turbine ge-
ometry are visualized by the λ2 method in Figs. 9 and 10, for the base and turbu-
lent cases, respectively. One can clearly see the trace of the tip vortices, the wake
behind the tower and the core of the swirl generated by the blades. Secondary vor-
tices around the tip vortices are also seen. Comparing the base and turbulent cases,
one can observe a more perturbed shape of the tip vortices for the later one. These
perturbations might lead to a faster break-down of the vortices further downstream.
Base P1 Base P6
Klein P1 Klein P6
-5/3 -5/3
1e-06
0.0001
E(f)
E(f)
1e-08
1e-10
0.5 1 2 4 8 1 8
f / fBPF f / fBPF
Fig. 7 Turbulent kinetic energy spectra for Fig. 8 Turbulent kinetic energy spectra for
the base and turbulent inlet cases at point P1 the base and turbulent inlet cases at point P6
(close to the inlet) (one unit downstream the turbine)
Fig. 9 Vortices visualized with the λ2 Fig. 10 Vortices visualized with the λ2
method, base case method, turbulent inlet case
5 Conclusions
The flow around a three-bladed wind turbine has been studied. Both the blades and
the tower have been considered in the computations. The grid sensitivity study re-
vealed that very fine grids are needed to evaluate the flow field accurately. There is a
strong interaction between the tower wake and the vortices generated by the turbine
blades. Imposing 10% inlet fluctuations did not change significantly the flow down-
stream the wind turbine. For future computations it is needed to consider several
turbulence levels. Also computations of wind turbine farms are needed to evaluate
the interaction of wind turbine wakes.
References
1. Crespo, A., Hernandez, J., Frandsen, S.: Survey of modelling methods for wind turbine
wakes and wind farms. Wind Energy 2, 1–24 (1999)
2. Grauthoff, M.: Utilization of wind energy in urban areas – chance or utopian dream?
Energy and Buildings 15-16, 517–523 (1990/1991)
3. Gullbrand, J., Bai, X.S., Fuchs, L.: High-order cartesian grid method for calculation of
incompressible turbulent flows. Int. J. Numerical Methods in Fluids 36, 687–709 (2001)
4. Hansen, M., Sørensen, J., Voutsinas, S., Sørensen, N., Madsen, H.: State of the art in
wind turbine aerodynamics and aeroelasticity. Progr. Aerospace Sciences 42, 285–330
(2007)
5. Ivanell, S.: Numerical computations of wind turbine wakes. Ph.D. thesis, Royal Institute
Of Technology (2009)
6. Johansen, J., Sørensen, N.N., Michelsen, J.A., Schreck, S.: Detached-eddy simulation of
flow around the NREL Phase VI blade. Wind Energy 5, 185–197 (2002)
7. Klein, M., Sadiki, A., Janicka, J.: A digital filter based generation of inflow data for
spatially developing direct numerical or large eddy simulations. J. Comput. Phys. 186,
652–665 (2003)
8. Olsson, M., Fuchs, L.: Large Eddy Simulation of the proximal region of a spatially
developing circular jet. Phys. Fluids 8(8), 2125–2137 (1996)
9. Pape, A.L., Lecanu, J.: 3D Navier-Stokes computations of a stall-regulated wind turbine.
Wind Energy 7, 309–324 (2004)
10. Pedersen, E., Waye, K.P.: Wind turbine noise, annoyance and self-reported environ-
ments health and well-being in different living environments. Occup. Environ. Med. 64,
480–486 (2007)
11. Salewski, M., Duwig, C., Milosavljevic, V., Fuchs, L.: LES of spray dispersion and mix-
ing in a swirl stabilized GT combustor. AIAA paper AIAA-2007-0924 (2007)
12. van den Berg, G.: Effects of the wind profile at night on wind turbine sound. J. Sound
and Vibration 277, 955–970 (2004)
13. Vermeer, L., Sorensen, J., Crespo, A.: Wind turbine wake aerodynamics. Progress in
Aerospace Sciences 39, 467–510 (2003)
Stochastic Synchronization of the Wall
Turbulence
Sedat Tardu
Abstract. Instantaneous amplitude and phase concept emerging from analytical

signal formulation is applied to the wavelet coefficients of streamwise velocity fluc-
tuations in the buffer layer of a near wall turbulent flow. Experiments and direct nu-
merical simulations show both the existence of long periods of inert zones wherein
the local phase is constant. These regions are separated by random phase jumps.
These behaviours are reminiscent of phase synchronization phenomena observed in
stochastic chaotic systems. The lengths of the constant phase inert (laminar) zones
reveal a type-I intermittency behaviour. The observed phenomena are related to the
footprint of coherent structures convecting in the low buffer layer that synchronizes
the wall turbulence.
1 Introduction
The discovery of coherent structures in the early 1960s has profoundly modified our
point of view of the wall turbulence structure. The incoherent turbulence occupies
only 20% of time and space in the inner layer. The coherent part is simpler to un-
derstand, since the coherent vortical structures can be identified, and tracked in time
and space, and their direct effect on the wall shear and transport of the shear stresses
and passive scalar can be clearly determined. The common consensus reached by
now points at the existence of quasi-streamwise vortices of diameters typically 10
wall units and located at 20 units from the wall. Their streamwise extent is roughly
300 units, and they generate low and high speed streaks at the wall with a spanwise
periodicity of 100 wall units. The sweep and ejection events they generate contribute
to the Reynolds shear stress by 80%. The time period of their generation is approx-
imately 100 units also and it depends on the distance from the wall. Turbulence in
general and the wall turbulence in particular can be seen as an infinite dimensional
Sedat Tardu
LEGI, B.P. 53 X, 38041 Grenoble, Cedex 9, France
e-mail: Sedat.Tardu@hmg.inpg.fr
354 S. Tardu
chaotic system. The quasi-periodicity induced by the coherent structures that are
convecting in the low buffer layer, should logically lead to the synchronization of
the turbulent quantities near the wall. Chaos synchronization is a process wherein
chaotic coupled (sub) systems subject to external forcing adjust their time scales re-
sulting in common spatial and temporal dynamics (Boccaletti et al., 2002, Pikovsky
et al., 2001). Synchronization can also be defined as the locking between the instan-
taneous phases of a state variable of the system and the phase of the external peri-
odic force. A rather weak degree of wall turbulence synchronization is expected in a
rush environment partially dominated by incoherence. The weaker synchronization
between chaotic systems, namely the phase synchronization occurs when the suit-
ably well-defined phases collapse, while the amplitudes remain highly uncorrelated.
The noisy synchronization is commonly defined as stochastic synchronization, and
the phase locking occurs for random periods of times and is interrupted by random
phase slips (Freund et al., 2000; Callenbach et al., 2002). The synchronization of
the wall turbulence driven by coherent vortices advecting in the low buffer layer, if
it occurs, should be classified in this last category the incoherent part playing the
equivalent role of noise. The occurrence of imperfect synchronization has been re-
ported through experiments in Tardu (2008). Our aim here is to give more evidence
based on direct numerical simulations.
2 Direct Numerical Simulations and Data Reduction

We use a DNS data basis of a fully developed homogeneous channel flow (Reτ =
180, see Tardu and Doche, 2007 for the details), multiple scale edge detection (Mal-
lat and Zhong, 1992) and directional Hilbert transforms (Granlund and Knutsson,
1995) together with 2D Hardy wavelets to detect two-dimensional singularities. We
consequently extract information related to the local amplitudes and phases versus
the scale. Let 2 wavelets defined by:
∂θ ∂θ
ψ1 (x, z) =
, ψ2 (x, z) = (1)
∂x ∂z
where θ is the smoothing function taken as a Gaussian here, while x and z stands
respectively for the streamwise and spanwise directions. The wavelet coefficients at
scale s are given by:
∂
s
W1 u(x, z) = u ⊗s ψ1 (x, z) = s (u ⊗s θ )(x, z)
∂x
∂
s
W2 u(x, z) = u ⊗s ψ2 (x, z) = s (u ⊗s θ )(x, z) (2)
∂z
The local maxima of sW1 u(x, z) correspond to the sharp smoothed gradients of the
velocity field, or to the inflection points of the convolution u ⊗s θ . The computations
are performed in Fourier domain in practice, and a fast algorithm is available but the
details are omitted here. The scale is not a continuous parameter and varies along
the dyadic sequence 2 jj∈N . The largest scale here is m = 8 . The wavenumber k+ at a
Stochastic Synchronization of the Wall Turbulence 355
2π
given scale j = m is defined as k+ = where Δ x+ = 13 and Δ z+ = 6 are
2m (Δ x+ Δ z+ )1/2
the mesh sizes in the streamwise and spanwise directions. Let us introduce now, sW ˘1
and sW˘ 2 the directional Hilbert transform of the wavelet coefficients. Let sW ˆ 1 (ωx , ωz )
denotes the Fourier transform of W1 . The directional Hilbert transform following x
√
is defined as W̆ˆ1 = jŴ sign(ωx ) where sign(ωx ) is the sign of ωx and j = −1. The
computations are performed in Fourier domain in practice, and a fast algorithm is
available but the details are omitted here. The wavelet ψ1 is then defined in Fourier
domain by ψ1 (2ωx , 2ωz ) = G(ωx )φ0 (ωx )φ0 (ωz ) with G(ω ) = 4 je jω /2 sinω /2 and
φ0 (ω ) = [ sin(ωω/2/2) ]3 . Similar relations hold for ψ2 and the details can be found in
Mala and Zhong (1992). For the band-pass wavelet transforms at a given scale s,
one can define a local amplitude and wavenumber through along the prescribed
direction i(i = 1 and i = 2) corresponds respectively to the streamwise and spanwise
directions), by using analytical signal concept:
s
s
Wi =s Ai cos( ki dxi ) (3)
where, the amplitude is:

s sW 2 +s W̆ 2
Ai (x, z) = i i (4)
The representation (3) is not unique and different characterizations are possible.
The relation (4) comes from the Rice canonical representation that is optimum in
the sense of minimizing the average rate. The local wavenumber that minimizes the
space variations of the envelope in the Rice representation is given by:
sW ∂ sW̆i −s W̆i ∂∂ W
s
i
s i ∂ xi xi
ki = s A2
(5)
i
The carrier wavenumber is defined by:
s E[s A2i ki ]
kci = (6)
E[s A2i ]
with E standing for spatial averaging. With the introduction of the carrier frequency,
one can define a local phase through:
s
Wi =s Ai cos[s kic xi +s Φi ] (7)
3 Results
We set the direction i = 1 hereafter and replace i by the corresponding streamwise
direction x. In the streamwise constant phase zones, the local wavenumber obeys to
s k (x, z) =s k at a given scale s . Fig. 1 shows the contours of the events defined as
x cx
the ensemble of points where: 0.9s kcx ≤s kx ≤ 1.1s kcx and obtained at y+ = 20.
The emerging contours were more regular by using these thresholds, instead of
356 S. Tardu
taking s kx (x, z) =s kcx directly, but the results did not change significantly. The
spurious zones where the phase is only punctually constant are eliminated. Elon-
gated constant streaky zones exist at each scale, and occupy a large amount of the
homogeneous plane as seen in Fig.1. It is clearly seen that at k+ = 0.044 (Fig. 1b)
the constant phase zones are of larger scales than at k+ = 0.0178 (Fig. 1a) although
their relative occupations of the whole plane are relatively comparable with respec-
tively 0.23 and 0.17. Their length scales are several times larger than 2π /s kcx , i.e.
one wavelength. The physical interpretation of the occurrence of constant phase
zones is related to the phase synchronization between the wall turbulent quantities
and the forcing imposed by the coherent vortices generated and convecting near the
wall. For phase synchronization of coupled chaotic oscillators, a very large constant
phase locked zone is followed by a very short turbulent stage (the turbulent stage
here refers to periods wherein the signal is chaotic according to the terminology
used in chaos synchronization). The difference here, is that, not only the duration
of phase locked zones is random, but also that the phase smoothly fluctuates in the
so called phase locked zones before sharp increases or phase jumps. These are the
main characteristics of noisy (or stochastic) synchronization as discussed and de-
fined before. In order to quantify the space occupied by constant phase events, we
define the occupancy Θ by dividing the number of constant phase events to the to-
tal number of points. In perfect chaos synchronization, the occupancy of the phase
locking regions separated by successive phase slips scales as:
Θ ∝ |C − C ps |−γ (8)
where C is either the coupling strength or the frequency (wavenumber) of the driving
signal. The exponent γ > 0 is γ = 1 in the case of on-off intermittency, and γ = 0.5
for the type I intermittency. The parameter C ps is the critical value of the phase syn-
chronization. It is impossible to observe perfect synchronization in stochastic sys-
tems with incoherent turbulence (IT). The phase locked regions are interrupted by IT
induced phase slips, and the system does not exhibit infinite laminar lengths. Fig. 2
shows lnΘ versus with ln|k+ − k∗+ | at which the maximum occupancy occurs. This
value differs only slightly from k∗+ = 0.25 corresponding to experimental u
data
at y+ = 10 (Tardu, 2008). Type I intermittency clearly occurs for |k+ − k∗+ | ≥ 0.1.
There is also reasonable correspondence between the experimental and DNS data,
taking into account the differences in the methodology (Tardu 2008) ,the mother
wavelets and the detection points.
The local amplitude Ax (x, z) defined through the equation (4) varies slowly in
the phase locked zones. We show in Fig. 3 (left) the flatness factor of Ax (x, z) de-
fined in the usual way as to characterize the associated intermittency. The flatness
FA
x is computed globally but also separately in the phase varying and phase locked
regions, first taking the mean values individually in these zones, and then comput-
ing the normalized fourth order moment of the fluctuating amplitudes A
x (x, z) =
Ax (x, z) − E(Ax ). It is seen in Fig. 3 that the intermittency of the local amplitude
increases with the wavenumber as expected, both globally and in the varying phase
zones, the increase being more pronounced in the latter case. In the phase varying
Fig. 1 Phase locked zones at y+ = 20 for the wavenumbers k+ = 0.178 and k+ = 0.044
Fig. 2 Type I intermittency emerging from dyadic wavelet decomposition of the DNS data at
y+ = 20. Comparison with experimental data of fluctuating longitudinal velocity at y+ = 10.
The experiments have been conducted in a turbulent boundary layer at the Reynolds number
based on the local momentum thickness Re = 913. The experimental procedure is detailed in
Tardu (2008)
zones for instance the flatness reaches values as high as 9 at the highest wavenumber.
The amplitude in the phase locked zones, however is significantly less intermittent,
FA
x is practically uniform across the whole wavenumber range, it is smaller than the
Gaussian value and varies from 1 to 2. This shows that the amplitude in the constant
phase zones approach a uniform distribution in the high wavenumber range, and the
uniformity of FA
x at large k+ indicates clearly that the wavelet coefficients oscillate
smoothly in the constant phase zones. Coherent structures near the wall are situated
in the low buffer layer concentrated roughly at δ + = 20 . Let us define an external
forcing at the scale 0 by 0 Ω = 2π /δ + . That represents the direct effect of coherent
vortices convecting near the wall and inducing phase locking of the wall turbulence.
Suppose that the forcing acting at the dyadic scale s = 2i is self-similar and directly
of the form s Ω =s Ω /2i . The frequency locking between the fluctuating near wall
velocity at a given scale and the external forcing occurs when the local wavenum-
ber s kx (x, z) frequency is s kx = m/ns Ω where m and n are integers (Rosenblum
et al., 1996). That reduces to s kcx = m/ns Ω in terms of carrier wavenumber during
358 S. Tardu
Fig. 3 Flatness of the amplitudes (left) and carrier wavenumber (right) . See the text for
details
the constant phase zones. Fig. 3 (right) shows the ratio s kcx /s Ω versus k+ . There
is quite a reasonable agreement between this simplified theory and the DNS data
leading to m/n = 1/4 .
4 Conclusion
The analytic signal concept is applied to the wavelet coefficients of fluctuating
streamwise velocity in the low buffer layer. Automatic separation of different time
scales through both the Hilbert transform that freezes the slow variables and the
wavelet analysis allows to sort out the hidden phase synchronization. Long quies-
cent periods of about 100 wall units wherein the phase oscillates around constant
values are noticed near the critical scale parameter. The constant phase zones are
interrupted by rapid phase jumps. A parallelism is constructed between these behav-
iors and the stochastic synchronization of chaotic systems that are under the effect
of noise, or incoherence as is the case of the near wall turbulence. Both experimen-
tal and direct numerical simulation results converge to the more or less similar re-
sults. They both reveal the existence of type-I intermittency connected to the saddle-
node bifurcation of the locking periods. The occurrence of long quiescent periods of
constant phases of the wavelet coefficients is interesting and may be used in some
control strategies. The phase jumps in particular are unambiguously well defined
near the critical wavelet scale parameter. The long time periods between the jumps
annunciate the arrival of active structures. A gain in effectiveness may presumably
be achieved if the decision and action stages of active control schemes coincide
with these periods. Strategies similar to the chaos control can also be developed on
the basis of the results presented here. It is expected that local excitation the wall
turbulence near the critical scale parameter improves the efficiency of active control
strategies. Tardu and Doche (2007) have shown for instance that a local blowing of
frequency near the bursting frequency doubles the efficiency of suboptimal control,
and synchronizes the fluctuating wall shear stress over significantly long streamwise
distances downstream of local forcing.
References
1. Aulin, T., Sundberg, C.E.-W.: Continuous phase modulation. IEEE Trans.
Commun. COM-29 (1981)
2. Boccaletti, S., Kurths, J., Osipov, G., Valladeres, D.L., Zhou, C.S.: The synchronization
of chaotic systems. Phys. Rep. 366, 1 (2002)
3. Freund, J., Neiman, A., Schiemansky-Geier, L.: Analytic description of noise induced
synchronization. Europhysics Letters 50, 8 (2000)
4. Granlund, G.H., Knutsson, H.: Signal processing for computer vision. Kluwer Academic
Publishers, Dordrecht (1995)
5. Mallat, S., Zhong, S.: IEEE Transactions on Pattern Analysis and Machine Intelligence
14, 710 (1992)
6. Pikovsky, A.S., Rosenblum, M.G., Kurths, J.: Synchronization. A Universal Concept in
Nonlinear Sciences. Cambridge University Press, Cambridge (2001)
7. Rosenblum, M., Pikovsky, A.S., Kurths, J.: Phase synchronization of chaotic oscillators.
Phys. Rev. Lett. 76, 1804 (1996)
8. Tardu, S., Doche, O.: Active control effectiveness and synchronization of wall turbulence
under localized imposed unsteadiness. Phys. Fluids 19, 108103–108107 (2007)
9. Tardu, S.: Stochastic synchronization of the near wall turbulence. Phys. Fluids 20, 045105
(2008)
Large-Eddy Simulations of an Oblique Shock
Impinging on a Turbulent Boundary Layer:
Effect of the Spanwise Confinement on the
Low-Frequency Oscillations
Emile Touber and Neil D. Sandham
Abstract. The low-frequency motions found in two large-eddy simulations of the

same oblique-shock/turbulent-boundary-layer interaction with significantly differ-
ent domain widths are investigated. The narrow domain artificially confines the
shock-induced separation bubble, which is seen to grow significantly. In addition,
the low-frequency/large-amplitude shock oscillations are found to be enhanced,
therefore suggesting that they originate from an intrinsic two-dimensional mech-
anism. By reducing the spanwise confinement, large coherent structures as wide
as one separation-bubble length are found to develop inside the interaction. Those
structures can move sideways and survive for extended periods of time. Their proper
resolution is therefore challenging in terms of computational cost and their meander-
ing motions can significantly bias the interpretation of a spectral analysis performed
at a fixed point.
1 Introduction and Previous Work

Shock-wave/turbulent-boundary-layer interactions (SBLI) are of practical interest in
a number of external and internal problems in aerospace, where the observed low-
frequency unsteadiness of the reflected shock can jeopardise the stability and/or the
structural integrity of flight vehicles. Such low-frequency oscillations are commonly
Emile Touber
University of Southampton, Southampton SO17 1BJ, U.K.
e-mail: etouber@soton.ac.uk
Neil D. Sandham
University of Southampton, Southampton SO17 1BJ, U.K.
e-mail: n.sandham@soton.ac.uk
362 E. Touber and N.D. Sandham
found at frequencies one or two orders of magnitude smaller than the frequency
associated with the upstream free-stream velocity U∞ and the 99%-boundary-layer
thickness δ99 . The driving mechanism of the reflected-shock low-frequency motions
is currently not understood and only partial theories have been proposed [5, 3, 6, 2, 7].
With the increasing computational power available, detailed numerical investiga-
tions covering long-enough time series and at the same Reynolds number as in ex-
periments are within reach of Large-Eddy Simulation (LES). Garnier et al. [4] first
reported the capability of LES to reproduce the time-averaged experimental flow-
field of an oblique shock impinging on a turbulent boundary layer. Very recently,
the ability of LES to correctly capture the low-frequency dynamics was shown in
Touber and Sandham [7]. Yet, due to the broadband nature of the low-frequency os-
cillations, it is necessary to acquire time series about hundred times longer than the
timescale associated with the most energetic low frequency. Given the two orders
of magnitude of difference with the characteristic timescale associated with the in-
coming turbulent boundary layer, the study of low-frequency oscillations requires
much longer integration times than commonly used in LES. To this end, the LES in
[7] was performed on a relatively narrow domain. Despite the encouraging results
on the possibility of LES to capture the low-frequency motions, the authors found
significant differences in the mean fields obtained between the narrow-span simula-
tion and shorter-time integrations of simulations with exactly the same settings but
using wider computational domains.
In this paper, the effect of the computational-domain width on the low-frequency
oscillations is investigated. The earlier narrow-span LES results [7] are compared
with a relatively long-run LES of the same case with a domain nearly ten times
wider. The wall-pressure spectra near separation will be compared and the principal
difference explained. This numerical study will give insight into the relevance of
the global instability reported in Touber and Sandham [7] and the development of
important flow structures inside the interaction region which can affect the interpre-
tation of wall-pressure spectra. It will also offer some guidelines on the necessary
domain width to correctly simulate such interactions.
2 Brief Description of the Simulations

Two LES of an oblique shock generated by an 8-degree-wedge angle impinging on
a Mach 2.3 turbulent boundary layer at Reynolds number of about 2 × 104, based on
the boundary-layer displacement thickness, were performed. The first one, which is
labelled “narrow-span LES” was aimed at minimising the overall number of grid
points in order to cover the longest time series possible. In particular, the spanwise
extent was reduced to the minimum required in order to properly resolve the in-
coming turbulent boundary layer. The second, labelled “large-span LES” is essen-
tially the same as the narrow-span LES but the spanwise extent is 8.75 times longer.
Details about those two simulations can be found in table 1.
Large-Eddy Simulations of an Oblique Shock Impinging 363
Table 1 Narrow- and large-span simulation details
Flow case large span narrow span

Flow conditions 8◦ -wedge angle, M = 2.3, Reδa1 ≈ 2 × 104
Domain sizeb,c
Lx × Ly × Lz in δ1vd units 450 × 90 × 105 450 × 70 × 12
Lx × Ly × Lz in Lsep units 6.9 × 1.4 × 1.6 4.3 × 0.7 × 0.1
Grid size, Nx × Ny × Nz 451 × 151 × 281 451 × 81 × 37
Stretchingd, βy 5.0 5.5
Grid resolutione, Δ x+ × Δ y+
min × Δ z
+ 33 × 1.3 × 12 41 × 1.6 × 14
Runtimef 3 × 101 6 × 101
a based on the upstream freestream velocity U∞ , the boundary-layer edge dy-
namic viscosity and
the boundary-layer displacement thickness (upstream of
interaction), δ1 = 0∞ (1 − ρ u/ρ∞ u∞ )dy
b δ1vd is the inlet displacement thickness computed on the van Driest velocity

profile ( 0∞ (1 − uvd /uvd
∞ )dy) and Lsep is the separation-bubble length
c x, y, z are the streamwise, wall-normal and spanwise directions, respectively
d where y = Ly sinh (βy ( j − 1)/(Ny − 1))/ sinh (βy ), j ∈ {1, . . . , Ny }
e using the classical wall-unit definition upstream of interaction
f based on the number of periods covered by a sine wave beating a frequency
f = 0.035U∞ /Lsep , excluding the relatively long start-up transients
The filtered 3D compressible Navier-Stokes equations are solved using a 4th -

order central spatial differencing scheme for the spatial derivatives and a 3rd -order
explicit Runge–Kutta scheme to integrate in time. Periodic boundary conditions are
used in the spanwise direction, while the no-slip condition is enforced at the wall,
which is set to be isothermal. The top (freestream) and outflow boundaries make
use of an integrated characteristic scheme in order to minimise unwanted reflections
from the computational-box boundaries. The oblique shock is introduced at the top
boundary using the Rankine–Hugoniot relationships. All details about this numer-
ical approach can be found in [7] and the references therein, together with details
about the specification of realistic three-dimensional and time-varying inflow con-
ditions, where great care is taken to ensure that no artificial low-frequency forcing is
introduced in the simulation domain, potentially interfering with the reflected-shock
dynamics.
3 Results
The flow conditions considered correspond to the experiments reported in [1]. Com-
paring the experimental and LES results is not the focus of the present paper. Never-
theless, figures 1 and 2 show the velocity and Reynolds stress profiles (in Morkovin’s
scaling) obtained in both LES and in the experiment at a station corresponding to
y+
0 1 2 3
10 10 10 10
12 9
large-span LES large-span LES
25 8
narrow-span LES
narrow-span LES 10 ρ u u
7
ρw u2τ
experiment (PIV)
experiment (PIV)
20
Reynolds stresses
6
5 .1 8
1 +
0 .4 5
y+ )/
y (mm)
+
15 (y
lo g
Uvd
+
6 4
3
10
4
2
ρ vv
ρw u2τ
1
5 2
0
(ρ/ρw ) u v /u2τ
0 0 −1
0 100 200 300 400 500 0 0.2 0.4 0.6 0.8 1 1.2
U (m s– 1 ) y/δ99
Fig. 1 Velocity profiles at x = 260 mm Fig. 2 Reynolds-stress profiles at x = 260 mm
x = 260 mm in [1]. The incoming-boundary-layer match is found to be excellent

while the Reynolds stresses are not as close to the PIV results. The LES profiles de-
crease more rapidly with distance from the wall compared with the PIV. In the case
of the narrow-span LES, the main reason is that the prescribed Reynolds-stress pro-
files in the digital-filter settings were too small in the outer region of the boundary
layer. This was improved in the large-span LES and the reduced level of fluctuations
in the outer region is mainly due to the choice of reference station, which is still in
the inflow transient produced by the digital-filter approach [7].
Figure 3 gives the wall-pressure and skin-friction distribution along the interac-
tion. A striking difference in the separation lengths can be seen, where the small-
span separation bubble is about 60% longer and 300% taller (not shown here) than
the one found in the large-span case. Despite some differences in the numerical set-
tings (see table 1), the main parameter responsible for such differences is the change
of spanwise extent [7]. Before focusing on the reason for such an influence of the
spanwise-domain size, we investigate whether similar differences can be found in
the flow unsteadiness.
Figure 4 provides the wall-pressure power-spectral-density (PSD) distribution at
separation. It was computed using Welch’s method with segments spanning
50Lsep /U∞ with a 50% overlap. The difference between the two cases is clearly seen,
with a drop of nearly one decade in the PSD levels at f = 0.03U∞/Lsep . Therefore,
the low-frequency motions in the large-span case appear to be significantly reduced.
In order to explain the significant damping of the low-frequency motions, it is
interesting to look at the one-dimensional weighted spectra of the spanwise auto-
correlation function (computed on the streamwise-velocity component) as a function
of the streamwise position, as shown in figure 5. While all the relevant scales of the in-
coming turbulent boundary layer are correctly resolved in both simulations, it is clear
−1
2.5 2 10
Cf large-span LES
small-span LES large-span LES
1.5 small-span LES
p1
p w/
wall-pressure PSD
−2
2 10
Cf × 103
1
pw /p1
0.5
1.5 −3
10
1 [p3 /p1 ]theo. = 2.5 −4

10
260 280 300 320 340 360 −2 −1 0 1
10 10 10 10
x (mm) f Lsep /U∞
Fig. 3 Wall-pressure and skin friction evolution Fig. 4 Wall-pressure PSD at separation
that the narrow-span configuration does not capture all the energetically-significant
spanwise wavelengths within the interaction. This artificially forces the separation
region to be mostly of a two-dimensional nature.
1 1
10 10
kz × δ99
0 0
10 10
large-span LES narrow-span LES
−15 −10 −5 0 5 −15 −10 −5 0 5

(x – Ximp )/δ99 (x – Ximp )/δ99
Fig. 5 One-dimensional weighted spectra maps of the spanwise auto-correlations of the

streamwise-velocity component at y+ ≈ 12. Contours show kz · |Euu |, where Euu (kz ) =
L /2
(1/Lz ) −Lz /2 Ruu (z) exp (−2iπ kz z)dz, with Ruu (z) = u
(z0 ,t)u
(z0 +z,t)/u
(z0 ,t)u
(z0 ,t).
z
The white dash lines indicate the location of separation and reattachment
The narrow-span time-averaged flow properties were analysed in [7] and the flow
topology was found to be globally unstable. Moreover, the global mode was shown
to act on timescales two orders of magnitude longer than the characteristic timescale
of the upstream boundary layer. By confining the separation bubble so that it re-
mains mostly two-dimensional, one forces the flow to be close to the one stud-
ied in the stability analysis. Therefore, the narrow-span LES greatly enhanced the
1.5
1
z/Lsep
0.5
0
60 70 80 90 100 110 120 130 140 150 160
t × U∞ /Lsep
1.5
1
z/Lsep
0.5
0
440 450 460 470 480 490 500 510 520 530 540
t × U∞ /Lsep
Fig. 6 Wake signal from the reversed-mass flow-rate (per unit width) at (x−Xsep )/Lsep ≈ 0.5
possibility for the global mode to exist [7]. Since the most unstable global-mode
structure was found to be two dimensional with a tendency to enhance or reduce
the separation, it could have made the narrow-span bubble follow more significant
expansion/compression motions, leading to more pronounced low-frequency oscil-
lations and enhanced separation. Although the large-span time-averaged flow field is
found to be globally unstable, the growth rate is about an order of magnitude smaller
than that found in the narrow-span case. This is consistent with the aforementioned
explanation and the observed drop in the PSD levels at low frequencies. Finally,
the spanwise structure of the flow in the large-span case is investigated.
Figure 6 was obtained by monitoring in time the reversed mass flowrate per unit
width at a fixed streamwise station located half the way down the separation bubble.
One can see that there are large pockets of reversed flow (in white) which can grow
in the spanwise direction by as much as one separation-bubble length. Furthermore,
they are seen to meander and be able to survive for about 30Lsep /U∞ . Sometimes,
two of these structures merge. Obviously, such structures could not fit in a confined
geometry as in the narrow-span case, causing significant differences in the spectra at
low frequency. Note that these structures are generated inside the interaction and do
not correspond to the long structures mentioned in [3]. It should be noted that due to
their meandering, interpreting the spectrum obtained from a single point is problem-
atic and these pockets of reversed flow should be tracked instead, to avoid any bias
related to the observation from a fixed position. Finally, although not shown here, it is
worth mentioning that the significant low-frequency shock motions in the large-span
LES are found to correlate well with the presence of the reverse-flow pockets.
4 Conclusion
The narrow- and large-span LES of the interaction of an oblique shock wave with a
turbulent supersonic boundary layer have revealed that, for the range of
computational-domain widths tested, the resulting flow field significantly differed.
First, the time-averaged separation bubble greatly varied in size (up to 60% in
length and 300% in height). Second, the low-frequency oscillations were observed
to be greatly affected. In the narrow-span case, the flow was constrained to be two-
dimensional, making the global mode described in [7] more relevant and potentially
responsible for the observed enhanced low-frequency oscillations. By relaxing the
spanwise confinement, the flow inside the interaction was found to develop wide
structures (of the size of the separation length) which could survive extended pe-
riods of time (O(30Lsep /U∞ )). Such structures were found to meander, making the
spectral analysis from a fixed position less relevant. The results obtained by arti-
ficially confining the bubble suggest that the low-frequency oscillations have their
origin in a two dimensional mechanism, but are weaker in unconfined flows that
contain meandering pockets of reverse flow.
Acknowledgements. The authors would like to acknowledge the UK Turbulence Consor-

tium EP/D044073/1 for the computational time provided on the HPCx and HECToR facili-
ties, the UK’s national high-performance computing service, which is provided by EPCC at
the University of Edinburgh and by CCLRC Daresbury Laboratory, and funded by the Office
of Science and Technology through EPSRC’s High End Computing Program. We would also
like to acknowledge the financial support of the European Union through the Sixth Frame-
work Program with the UFAST project (www.ufast.gda.pl). Finally, we are grateful
to J.-P. Dussauge and his co-workers at IUSTI for kindly making the data in figures 1 and 2
available.
References
1. Dupont, P., Piponniau, S., Sidorenko, A., Debiève, J.F.: Investigation by Particle Im-
age Velocimetry Measurements of Oblique Shock Reflection with Separation. AIAA
Journal 46(6) (2008)
2. Dussauge, J.P., Piponniau, S.: Shock/boundary-layer interactions: Possible sources of
unsteadiness. Journal of Fluids and Structures 24, 1166–1175 (2008)
3. Ganapathisubramani, B., Clemens, N.T., Dolling, D.S.: Effects of Upstream Boundary
Layer on the Unsteadiness of Shock-Induced Separation. J. Fluid Mech. 585, 369–394
(2007)
4. Garnier, E., Sagaut, P., Deville, M.: Large Eddy Simulation of Shock/Boundary-Layer
Interaction. AIAA Journal 40(10) (2002)
5. Pirozzoli, S., Grasso, F.: Direct Numerical Simulation of Impinging Shock
Wave/Turbulent Boundary Layer Interaction at M = 2.25. Physics of Fluids 18(6)
(2006)
6. Robinet, J.C.: Bifurcations in Shock-Wave/Laminar-Boundary-Layer Interaction:
Global Instability Approach. J. Fluid Mech. 579, 85–112 (2007)
7. Touber, E., Sandham, N.D.: Large-eddy simulation of low-frequency unsteadiness in a
turbulent shock-induced separation bubble (2009)
Parameter-Free Symmetry-Preserving
Regularization Modelling of Turbulent Natural
Convection Flows
F.X. Trias, R.W.C.P. Verstappen, M. Soria, and A. Oliva
Abstract. Since direct numerical simulations of natural convection flows cannot

be performed at high Ra-numbers, a dynamically less complex mathematical for-
mulation is sought. In the quest for such a formulation, we consider regularizations
(smooth approximations) of the nonlinearity. The regularization method basically
alters the convective terms to reduce the production of small scales of motion by
means of vortex stretching. In doing so, we propose to preserve the symmetry and
conservation properties of the convective terms exactly. This requirement yields a
novel class of regularizations that restrain the convective production of smaller and
smaller scales of motion by means of vortex stretching in an unconditional stable
manner, meaning that the velocity cannot blow up in the energy-norm (in 2D also:
enstrophy-norm). The numerical algorithm used to solve the governing equations
preserves the symmetry and conservation properties too. The regularization model
is successfully tested for a 3D natural convection flow in air-filled (Pr = 0.71) dif-
ferentially heated cavity of height aspect ratio 4 at Ra = 1010 and 1011 . Moreover,
a method to dynamically determine the regularization parameter (local filter length)
is also proposed and tested.
1 Introduction
Turbulent natural convection in enclosed cavities has been the subject of numer-
ous studies over the past decades. They model many engineering applications such
F.X. Trias · M. Soria · A. Oliva
Centre Tecnològic de Transferència de Calor, ETSEIAT, Technical University of Catalonia,
C/Colom 11, 08222 Terrassa, Spain
e-mail: cttc@cttc.upc.edu
F.X. Trias · R.W.C.P. Verstappen
Institute of Mathematics and Computing Science, University of Groningen, P.O. Box 407,
9700 AK Groningen, The Netherlands
e-mail: R.W.C.P.Verstappen@rug.nl
370 F.X. Trias et al.
as ventilation of rooms, cooling of electronics devices or air flow in buildings.

Simultaneously, this configuration has served as prototype for the development of
numerical algorithms. Therefore, an accurate prediction of the flow structure and
the heat transfer in such configuration is of great interest. However, due to the com-
plex behaviour exhibit (see figure 1), and despite the great effort devoted, an ac-
curate turbulence modelling of this configuration remains as a great challenge. On
the other hand, direct numerical simulation (DNS) has become an essential tool to
give new insights into the physics of turbulence and provide indispensable data for
the development of better turbulence models. However, at high Ra-numbers, DNS
simulations are not feasible.
In this context, a dynamically less complex mathematical formulation is sought.
In the quest for such a formulation, we consider regularizations (smooth approx-
imations) of the nonlinear convective term. The first outstanding approach in this
direction goes back to Leray [1]; the Navier-Stokes-α model also forms an example
of regularization modelling (see [2], for instance). The regularization methods basi-
cally alter the convective terms to reduce the production of small scales of motion.
In doing so, we proposed to preserve the symmetry and conservation properties of
the convective terms exactly [3]. This requirement yielded a family of symmetry-
preserving regularization models [3]: a novel class of regularizations that restrain
the convective production of smaller and smaller scales of motion in an uncondi-
tional stable manner, meaning that the velocity cannot blow up in the energy-norm
(in 2D also: enstrophy-norm). In this work, a method to dynamically determine the
regularization parameter (local filter length) with the requirement that the vortex
stretching must be stopped at the scale set by the grid is also proposed and tested.
The numerical algorithm used to solve the governing equations preserves the conser-
vation properties too [4] and is therefore well-suited to test the proposed simulation
model.
Fig. 1 Left: several instantaneous temperature fields at Ra = 1011 and Pr = 0.71 (air). Right:
zoom around the point where the vertical boundary layer becomes totally disrupted and large
eddies are ejected
Parameter-Free Symmetry-Preserving Regularization Modelling 371
2 C4 -Regularization Modelling
We restrict ourselves to the C4 approximation (see [3], for details): the
convective term in the Navier-Stokes equations is then replaced by the following
O(ε 4 )-accurate smooth approximation C4 (u, v) given by
C4 (u, v) = C (u, v) + C (u, v ) + C (u , v) (1)
where C (u, v) = (u · ∇)v represents the convective operator. Note that here a prime
indicates the residual of the filter, e.g. u = u − u, which can be explicitly evalu-
ated, and (·) represents a normalised self-adjoint linear filter with filter length ε .
Therefore, the governing equations result to
1
∂t u + C4(u, u) = Δ u − ∇p + f ; ∇·u = 0 (2)
Re
Note that the C4 approximation is also a skew-symmetric operator like the original
convective operator. Hence, the same inviscid invariants -kinetic energy, enstrophy
in 2D and helicity- that the original Navier-Stokes equations are kept for the new set
of partial differential equations to be solved (2).
Since now the performance of the C4 approximation has been successfully tested
for a turbulent channel flow [3] and an air-filled (Pr = 0.71) differentially heated
cavity (DHC) of height aspect ratio A = 4 at Ra-numbers (based on the cavity
height) 1010 [5, 6] and 1011 [7] by means of direct comparison with the DNS re-
sults [8, 9, 10]. In the latter cases, the filter width ε was treated as a parameter.
Then, the value of ε was prescribed in advanced. Figure 2 displays illustrative re-
sults for the aforesaid DHC problem at Ra = 1010 as a function of ε /h ratio. These
results show that the numerical solution is weakly dependent for sufficiently large
values of ε . Similar behaviour has also been observed for the other aforementioned
configurations.
130 0.24
DNS
C4 ( 8x17x40)
125 C4 (16x34x80) 0.22
Parameter-free C4 ( 8x17x40)
Parameter-free C4 (16x34x80)
120 0.2
Nusselt
(u3)max
115 0.18
110 0.16
DNS
C4 ( 8x17x40)
105 0.14 C4 (16x34x80)
Parameter-free C4 ( 8x17x40)
100 0.12
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
ε/h ε/h
Fig. 2 The overall averaged Nusselt number (left) and the maximum of the averaged vertical
velocity at the horizontal mid-height plane (right) as a function of the ratio of the filter length
ε to the average grid width h
3 On the Dynamic Determination of the Filter Length

In the present work, we propose to determine ε dynamically with the requirement
that the vortex stretching must be stopped at the scale set by the grid. The idea
behind is to modify convective operator sufficiently to guarantee that the following
inequality is hold
λk (ε ) − ν k2 ≤ 0 (3)
where λk (ε ) = ωk · C4 (ω , u)/(ωk · ωk ) is the Rayleigh quotient of the vortex-
stretching at the grid scale, k = π /h. In practice, the value of λk has been bounded
by the largest (positive) eigenvalue of the straintensor S,
λk (ε ) ≤ f4,k (ε ) λmax (S) (4)

For the C4 -approximation, the damping function, 0 < f4,k ≤ 1, at the highest fre-
quency is given by 3ĝ2k (ε ) − 2ĝ3k (ε ) (see [3], for details), where 0 < ĝk (ε ) ≤ 1
is the transfer function of the linear filter. Therefore, it suffices that the following
inequality be locally hold
ν k2
3ĝ2k (ε ) − 2ĝ3k (ε ) ≤ −→ ĝk (ε ) −→ ε (5)
λmax (S)
to guarantee that the vortex-stretching mechanism be stopped at the smallest scale
set by the mesh.
Table 1 Description of meshes: the spanwise (Nx ), the wall-normal horizontal (Ny ), and the
vertical (Nz ) resolutions for tested cases. Further details about the meshing can be found in [8]
DNS RM1 RM2
Ra = 1010 128 × 190 × 462 16 × 34 × 80 8 × 17 × 40
Ra = 1011 128 × 682 × 1278 12 × 45 × 85 8 × 30 × 56
4 Results for a Turbulent Differentially Heated Cavity

The performance of the proposed method to dynamically determine the regulariza-
tion parameter ε of the C4 approximation has also been tested for the aforesaid DHC
problem. Again two very coarse meshes (see table 1) have been solved. In this work,
a fourth-order accurate Gaussian filter [11] has been chosen. In 1D it becomes
ε 4 − 4ε 2 16ε 2 − ε 4 ε 4 − 20ε 2 + 192

ui = (ui+2 + ui−2) + (ui+1 + ui−1) + ui (6)
1152 288 192
where the filter length, ε , need not be restricted to integers multiples of the grid
width. The boundary conditions that supplement the Navier-Stokes equations are
also applied to (6).
First results displayed in figure 2 exhibit the great potential of this method. Note
that in the parameter-free C4 cases, the results do not depend on ε /h. At least as
good results as the optimal ε /h ratio determined by trial-and-error procedure are
being obtained.
4.1 Grid Dependence Analysis

A reliable modelling of turbulence at (very) coarse grids is a great challenge. The
coarser the grid, more convincing model quality is perceived. However, it might
happen that solution be strongly dependant on meshing parameters and thus some
particular combinations could ’accidentally’ provide good results. An example of
this behaviour can be found in [12] for a turbulent channel flow. In order to elucidate
this point, the same DHC problem has been solved on a series of 50 randomly gen-
erated meshes: with (Nx , Ny , Nz )-values limited by those given by meshes RM1 and
RM2 (see table 1), i.e. 8 ≤ Nx ≤ 16, 17 ≤ Ny ≤ 34, and 40 ≤ Nz ≤ 80. Results for
the overall Nusselt and the centerline stratification values are displayed in figure 3.
At first sight, we can observe that the parameter-free C4 modelling predicts good
results irrespective of the meshing whereas very poor and dispersed results are ob-
tained when model is switched off. Specially significant is the fairly good prediction
of the stratification (note the big dispersion obtained without model).
140
120
100
avg(Nu)
80
60
40
DNS
20 No model
Parameter-free C4
0
0 0.2 0.4 0.6 0.8 1 1.2
Strat
Fig. 3 The overall Nusselt number and the centerline stratification for randomly generated
grids
4.2 Performance at Higher Ra-Numbers

The performance of the parameter-free C4 -regularization model has also been tested
at a higher Ra-number (Ra = 1011 ). Results for two coarse meshes (see table 1) are
displayed in figure 4 and compared with the DNS reference results [9, 10]. Al-
though in this case the Ra-number is one order of magnitude higher and the mesh
280 0.75
DNS
270 C4 ( 8x30x56) 0.7
C4 (12x45x85)
260 Parameter-free C4 ( 8x30x56)
Parameter-free C4 (12x45x85) 0.65
250
0.6
240
Nusselt
(u3)max
0.55
230
0.5
220
0.45
210
DNS
200 0.4 C4 ( 8x30x56)
C4 (12x45x85)
190 0.35 Parameter-free C4 ( 8x30x56)
180 0.3
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
ε/h ε/h
Fig. 4 The overall averaged Nusselt number (left) and the maximum of the averaged vertical
velocity at the horizontal mid-height plane (right) as a function of the ratio of the filter length
ε to the average grid width h at Ra = 1011
resolutions are kept almost the same respect to the case at Ra = 1010 , the proposed
method is able to predict reasonable well the mean flow values.
5 Conclusions
The performance of the proposed method to dynamically determine the regulariza-
tion parameter ε of the C4 approximation has been tested for an air-filled DHC
problem of aspect ratio 4 and Rayleigh numbers 1010 and 1011 . Results presented
here exhibit the great potential of the proposed method as new turbulence model: the
parameter-free C4 approach is able to predict at least as good results as the optimal
ε /h ratio determined by trial-and-error procedure.
In conclusion, the proposed method constitutes a parameter-free turbulence
model. Moreover, since no ad hoc phenomenological arguments that cannot be for-
mally derived for the Navier-Stokes equations are used it suggest that this method
should be valid for other configurations. Nevertheless, more simulations for a wide
variety of cases and meshes will be necessary to confirm these conclusions.
Acknowledgements
This work has been financially supported by the Ministerio de Educación y
Ciencia, Spain, (Project: “Estudio y parametrización de fenómenos de transferéncia de calor
y dinámica de fluidos involucrados en diseños de arquitectura bioclimática”. Contract/grant
number ENE2006-14247) and a postdoctoral fellowship Beatriu de Pinós (2006 BP-A 10075)
by the Generalitat de Catalunya.
References
1. Leray, J.: Sur le movement d’un liquide visqueaux emplissant l’espace. Acta
Mathematica 63, 193–248 (1934)
2. Geurts, B.J., Holm, D.D.: Regularization modeling for large-eddy simulation. Physics of
Fluids 16, L13–L16 (2003)
3. Verstappen, R.: On restraining the production of small scales of motion in a turbulent
channel flow. Computers and Fluids 37, 887–897 (2008)
4. Verstappen, R.W.C.P., Veldman, A.E.P.: Symmetry-Preserving Discretization of Turbu-
lent Flow. Journal of Computational Physics 187, 343–368 (2003)
5. Trias, F.X., Soria, M., Oliva, A., Verstappen, R.W.C.P.: Regularization models for the
simulation of turbulence in a differentially heated cavity. In: Proceedings of the European
Computational Fluid Dynamics Conference (ECCOMAS CFD 2006), Egmond aan Zee,
The Netherlands (September 2006)
6. Trias, F.X., Verstappen, R., Soria, M., Oliva, A.: Symmetry-preserving regularization
modelling of a turbulent differentially heated cavity. In: LES in Science and Technology,
Poznan, Poland (April 2008)
7. Trias, F.X., Verstappen, R.W.C.P., Soria, M., Gorobets, A., Oliva, A.: Regularization
modelling of a turbulent differentially heated cavity at Ra = 1011 . In: 5th European
Thermal-Sciences Conference, EUROTHERM 2008, Eindhoven, The Netherlands (May
2008)
8. Trias, F.X., Soria, M., Oliva, A., Pérez-Segarra, C.D.: Direct numerical simulations of
two- and three-dimensional turbulent natural convection flows in a differentially heated
cavity of aspect ratio 4. Journal of Fluid Mechanics 586, 259–293 (2007)
9. Trias, F.X., Gorobets, A., Soria, M., Oliva, A.: Direct numerical simulation of a differ-
entially heated cavity of aspect ratio 4 with Ra-number up to 1011 - Part I: Numerical
methods and time-averaged flow. International Journal of Heat and Mass Transfer 53,
674–683 (2010)
10. Trias, F.X., Gorobets, A., Soria, M., Oliva, A.: Direct numerical simulation of a differen-
tially heated cavity of aspect ratio 4 with Ra-number up to 1011 - Part II: Heat transfer and
flow dynamics. International Journal of Heat and Mass Transfer 31, 1195–1220 (1999)
11. Sagaut, P., Grohens, R.: Discrete Filters for Large Eddy Simulations. International Jour-
nal for Numerical Methods in Fluids 31, 1195–1220 (1999)
12. Meyers, J., Sagaut, P.: Is plane-channel flow a friendly case for the testing of large-eddy
simulation subgrid-scale models? Physics of Fluids 19, 048105 (2007)
An a Priori Study for the Modeling of Subgrid
Terms in Multiphase Flows
P. Trontin, S. Vincent, J.L. Estivalezes, and J.P. Caltagirone
Abstract. In the framework of turbulent interfacial multiphase flows, an a priori

study is performed in the case of the turbulence/interface interaction. Density and
viscosity ratios are set to 1 to perform a parametric study on surface tension forces.
When using the ghost-fluid method, subgrid-scale terms deriving from jump con-
ditions appear in the pressure gradient and viscous terms. A model is tested for
advective SGS terms deriving from both the momentum equation and the advection
of the interface.
1 Introduction
If nowadays Large Eddy Simulation (LES) of single-phase flows is already widely
used even for industrial applications, LES of two-phase interfacial flows, i.e. two-
phase flows where an interface separates liquid and gas phases, still remains a
P. Trontin
ONERA/DMAE, 2, avn Edouard Belin, 31055 Toulouse, France
e-mail: pierre.trontin@onera.fr
S. Vincent
Université Bordeaux I, TREFLE-ENSCPB, UMR 8508, 16, avn Pey-Berland, 33607 Pessac
Cedex, France
e-mail: vincent@enscpb.fr
J.L. Estivalezes
ONERA/DMAE, 2, avn Edouard Belin, 31055 Toulouse, France
e-mail: jean-luc.estivalezes@onera.fr
J.P. Caltagirone
Université Bordeaux I, TREFLE-ENSCPB, UMR 8508, 16, avn Pey-Berland, 33607 Pessac
Cedex, France
e-mail: calta@enscpb.fr
378 P. Trontin et al.
challenging task. The main issue is the development of subgrid-scale (SGS) models
well suited for two-phase interfacial flows. Freely decaying homogeneous isotropic
turbulence flow in a periodic cubic box has been chosen to perform a Direct Nu-
merical Simulation (DNS) database to be filtered. Two planar interfaces are initially
inserted in the middle of the computational box separated by 5% of the length box.
In such a configuration the sheet shows distorsions comparable to those that can be
observed in the atomization process of a sheet. The aim of this work is to carry out
a a priori analysis of this DNS database. Explicit filtering of 3D Direct Numerical
Simulations has been employed to evaluate the order of magnitude of subgrid contri-
butions. Surface tension forces are taken into account with the ghost-fluid method,
and so, the subgrid scale terms deriving from capillary forces do not appear ob-
viously after the filtering operation of the Navier-Stokes equations. In this work,
subgrid terms of jump conditions are shown to derive from the filtering of both the
pressure gradient term and the viscous term. After a short introduction of the nu-
merical methods, Navier-Stokes equations are filtered. Then, the numerical test case
is presented, and different SGS terms are compared. Finally, a model and its limits
is presented for the SGS terms deriving from the convective terms in the momentum
and φ advection equations.
2 Governing Equations
As the simulation of liquid-gas flow at moderate velocities is performed, the incom-
pressible Navier-Stokes equations are used. Interfacial multiphase flows involving
immiscible phases are considered. They are modelled as follows:
∂u

∂ t + (u · ∇)u = − ρ ∇p + ρ ∇ · μ (∇u + ∇u)
1 1 T
(1)
∇·u = 0
The boundary conditions at the interface between two immiscible fluids are the
continuity of the velocity components:
[u]Γ = 0 (2)
and the dynamic boundary conditions for the normal and tangential stresses:
[p]Γ − n · [μ (∇u + ∇uT )]Γ · n = σ κ

(3)
t · [μ (∇u + ∇uT )]Γ · n = 0
where n is the normal to the interface, t the tangent. The brackets [u]Γ for example
stand for [u]Γ = u+ − u− , with u+ the velocity in the phase 1 and u− the velocity
in the phase 2. As a level-set method [5, 7] is used to capture the interface, it is
implicitly given by the zero of the smooth function φ (x, y, z,t). φ is imposed to be
the signed distance function to the interface. This function obeys to the following
equation:
∂φ
+ (u · ∇)φ = 0 (4)
∂t
An a Priori Study for the Modeling of Subgrid Terms in Multiphase Flows 379
3 Numerical Methods
3.1 Navier-Stokes Solver
Classical projection methods are performed to enforce the incompressibility
constraint [1]. The convection terms in the momentum equations are approximated
in a conservative way with 5th order accurate WENO schemes [6]. Time integration
is performed with 3rd order accurate TVD Runge-Kutta scheme. The Poisson equa-
tion for the pressure is solved by a fast multigrid preconditioned conjugate gradient
method [8].
3.2 The Ghost Fluid Method (GFM) for the Jump Conditions at
the Interface
The jump conditions for pressure and pressure gradient in the Poisson pressure equa-
tion as well as jump condition for the viscous terms are taken into account by the
Ghost Fluid Method (GFM) [3]. The full mathematical details of this method can be
found in [2]. An artificial fluid (the Ghost Fluid) is created which implicitely gives
the proper conditions at the interface. The GFM can be used to model the Navier-
Stokes equations directly, without the addition of source terms to model the effects
of surface tension.
4 Filtering Operation
To conduct a priori testing, a filtering operation is performed on the equations (1)
and (4). The classical convolution operator is used. Let G be the filter function and
f a fluctuating field. f is then defined by:

f = (G f )(x) = f (x )G(x − x ; Δ ) dx (5)
Ω
The selected filter is a top hat filter. The filter size is symbolised by Δ . The same
filter size along each direction was used : Δ = Δ x = Δ y = Δ z . As a first example,
the filter size used in this paper is 5Δx where Δx is the spatial step.
The filtering operation is applied on the motion equations of the two-phase flow
and the LES motion equations are obtained.
∇ · u = τmass
∂u 1 1
+ (u.∇)u = − ∇P + ∇. μ (∇u +t ∇u) + τmom,trans + τconv − τ press − τdi f f
∂t ρ ρ
∂φ
+ (u · ∇)φ = τ phi,trans + τinter f
∂t
(6)
with the following subgrid scale (SGS) terms:
∂u
τmom,trans = ∂t − ∂∂ut
∂φ
τ phi,trans = ∂t − ∂∂φt
τmass = ∇.u − ∇.u
τconv = (u.∇)u − (u.∇)u (7)
τinter f = (u · ∇)φ − (u · ∇)φ
τ press = − ρ1 ∇P + ρ1 ∇P
τdi f f = ρ ∇.(μ (∇u +
1 t ∇u)) − ρ1 ∇. ( μ (∇u +t ∇u))
The filtering of jump conditions does not derive from an explicit term in the
Navier-Stokes equations. As explained previously, the GFM allows to capture im-
plicitly the proper conditions at the interface when discretizing equations. So the
filtering of jump conditions at the interface is taken into consideration in the calcu-
lation of the terms τ press and τdi f f . That is why the commutation between the spatial
filter · and the derivatives cannot be assumed for the pressure term. Transient terms
τmom,trans and τ phi,trans are zero.
5 Numerical Simulation and a Priori Filtering

A freely decaying homogeneous isotropic turbulence (HIT) case in a 2π periodic
cubic box has been chosen to peform a DNS database to be filtered. The spatial
resolution is 512 × 512 × 512. All the dimensionless parameters of the simulation
are summarized in the following table.
ν u ε Lf λ ηk Te RL f Rλ
0.003 1.0 0.911 0.278 0.222 1.31210−2 0.278 93 74
An initially plane sheet is added in the HIT field. This sheet divides the square
box in two equal parts. Its thickness δ is 5% of the box size. Initially plane, the sheet
follows the movements of the surrounding turbulent flow (Fig. 1). A parametric
study based on σ , the surface tension coefficient, is carried out. The work is focused
on the role played by surface tension forces in the interface/turbulence coupling,
and density and viscosity are both the same in the two fluids: ρρ12 = 1 and μμ12 = 1.
Before analyzing the different SGS terms, the anisotropy due to the presence of
the interface is studied. In Fig. 2, normal and tangential Reynolds stresses are repre-
sented for different We at t/Te = 13. Near the interface, both normal and tangential
Reynolds stresses decrease except for We = 2 and 0.2 where they both increase. For
We = 2, the interface is composed of large drops oscillating under surface tension
forces and an energy transfer happens from surface energy to the surrounding turbu-
lence, which explains the increase of both normal and tangential Reynolds stresses.
For We = 0.2, the same behaviour happens but with less intensity. Therefore, for ev-
ery We number, the interface leads to local modifications of turbulence in its vicin-
ity. This will have consequences on SGS terms. The magnitude of τmass is negligible
compared to the one of ∇.u and no further details are given about τmass in this paper.
The magnitude of τdi f f appears to be negligible and is of the order of the machine
accuracy. This is shown in the following table at t/Te = 13.
We 110 63 19 2 0.2
||τmass ||L1 /||∇ · u||L∞ 5.4 10−5 1.2 10−5 4.2 10−6 4.4 10−7 5.6 10−6
||τdi f f ||L1 /|| ρ1 ∇.(μ (∇u +t ∇u)) ||L∞ 1.9 10−15 1.8 10−15 7.1 10−16 6.7 10−16 1.6 10−15
Therefore, only the terms τconv , τ press and τinter f are significant in our study.
τ press (φ ) and τconv (φ ) are represented in Fig. 1 where they have to be understood
as follows: 3 4
1
N∑
τ (φ ) = τ (xi ) , φ (xi ) ∈ [φ , φ + Δ x] (8)
i
where N is the number of points for which φ satisfies φ (xi ) ∈ [φ , φ + Δ x], Δ x being
the spatial step. Only the z component of τconv and τ press is represented. Conclusions
for x and y components are similar. The use of φ as a parameter allows to study the
magnitude of SGS terms in function of the distance to the interface. τconv and τ press
are non-dimensioned by ||(u.∇)u||L∞ . In Fig. 1, two regions appear. For φ /δ ≥
0.15, τ press is zero and a modeling is necessary only for τconv . For φ /δ ≤ 0.15,
τconv and τ press increase near the interface. For the two highest We (110 and 63),
τ press /τconv < 1 at the interface. Therefore, for small surface tension coefficients,
the main SGS contribution is τconv even in the vicinity of the interface. For lower We
(except 0.2), τ press > τconv at the interface for φ /δ < 0.1. But τ press vanishes quickly,
and the magnitude of τ press is negligible compared to τconv very quickly. We = 0.2 is
a particular case. Indeed, the interface remains flat, only disturbed by surface waves.
Curvatures are low and then pressure jumps at the interface are moderate. Even if
σWe =0.2 > σWe =2 , mean curvatures for We = 0.2 are less than for We = 2 inducing
weaker surface tension forces for We = 0.2, and therefore smaller τ press .
0.0008 0.0008
0.0007 0.0007
0.0006 0.0006
τpress (dimensionless)
τconv (dimensionless)
0.0005 0.0005
We=110 We=110
We=63 We=63
0.0004 0.0004
We=19 We=19
We=2 We=2
0.0003 We=0.2 We=0.2 0.0003
0.0002 0.0002
0.0001 0.0001
0 0
0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3
φ/δ φ/δ
Fig. 1 Left: sheet in a surrounding HIT flow. 5123 grid, We = 110, t/Te = 5. Right: τ press (φ )
and τconv (φ ) for different We numbers at t/Te = 13 (z component)
2.2 0.02
We=110 We=110 We=110
tangential Reynolds stresses

1.4
normal Reynolds stresses 2 We=63 We=63 We=63
1.8 We=19 We=19 We=19
We=2 We=2 0.015
We=2
error (L norm)
1.2 We=0.2
1.6 We=0.2 We=0.2
1
1.4 0.01
1.2 1
1 0.005
0.8
0.8
0.6 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 0.2 0.4 0.6 0.8 1
φ/δ φ/δ φ/δ
Fig. 2 Left and middle: normal and tangential Reynolds stresses vs. φ /δ . t/Te = 13. Right:
absolute L1 error between τconv and τconv
mod vs. φ /δ for different W . t/T = 13
e e
1 1 1
0.8 y=x 0.8 y=x 0.8 y=x
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
mod
mod
mod
τinterf
τinterf
τinterf
0 0 0
-0.2 -0.2 -0.2
-0.4 -0.4 -0.4
-0.6 -0.6 -0.6
-0.8 -0.8 -0.8
-1 -1 -1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
τinterf τinterf τinterf
We = 110 We = 19 We = 0.2
Fig. 3 τinter
mod vs. τ
f inter f . y = x is represented. t/Te = 13
1 1 1
0.8 y=0.8*x 0.8 y=0.8*x 0.8 y=0.8*x
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
mod
mod
mod
0 0 0
τconv
τconv
τconv
-0.2 -0.2 -0.2

-0.4 -0.4 -0.4
-0.6 -0.6 -0.6
-0.8 -0.8 -0.8
-1 -1 -1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
τconv τconv τconv
We = 110 We = 19 We = 0.2
Fig. 4 τconv
mod vs. τ
conv . y = 0.8 x is represented. t/Te = 13
In the filtering of the momentum equation, the only significant SGS term is τconv .
In the vicinity of the interface, τ press vanishes very quickly. Now, modeling will be
studied for τconv and τinter f .
6 Modeling of τconv and τinter f

For τconv and τinter f , the Leonard decomposition improved by Germano [4] is used.
τinter
mod
f = u · ∇φ − u · ∇φ
(9)
τconv
mod
= (u · ∇)u − (u · ∇)u
where τinter
mod and τ mod are the modeled terms for τ
f conv inter f and τconv . In Fig. 3, τinter f is
mod
represented vs. τinter f for three different We numbers at t/Te = 13. The correlation
between the SGS term and the model proposed in Eq. (9) is strong for the interfacial
term. Indeed, points in Fig. 3 spread along the line y = x. For τconv , the same kind
of correlation is plotted. Points spread on a larger band around the line y = 0.8 x,
and the model is not as accurate for high We numbers. At t/Te = 13 for high We
numbers, both phases are present in a large part of the computational box, with a
multiphase flow behavior. As the single-phase model proposed in Eq. (9) does not
take mutiphase interactions into account, this explains a less accurate model for high
We . In Fig. 2 (right), the L1 error between τconv and τconv
mod is represented vs. φ /δ . That
shows that the model for τconv proposed in Eq. (9) is not accurate is the vicinity of
the interface. The model has to be adapted near the interface to deal with anisotropy
due to the presence of the interface. The model (9)is accurate for φ /δ > 0.3.
7 Conclusion
An a priori study has been performed, and SGS terms deriving from the convective
and pressure gradient terms appear when ρ1 = ρ2 and μ1 = μ2 in the momentum
equation. As the GFM is used, the filtering of jump conditions is taken into consid-
eration in the computation of the pressure gradient (to compute the gradient across
the surface in an accurate way). Therefore, the commutation between the filtering
operator and the pressure gradient is not valid anymore, and τ press is not zero in the
close vicinity of the interface. The magnitude of τ press is higher than τconv for low
We numbers, except We = 0.2. For the lowest We number, the interface is slightely
disturbed with moderate curvatures, and therefore the jump at the interface is moder-
ate, inducing low magnitudes for τ press . However, the magnitude of τ press decreases
quickly far from the interface and can be neglected compared to τconv . For the inter-
face advection, the SGS term deriving from the non-linear terms have to be taken
into consideration. A model has been tested for τconv and τinter f . τinter
mod is accurate,
f
whereas τconv has difficulties to take into account the anisotropy near the interface.
mod
References
1. Chorin, A.J.: Numerical solution of the Navier-Stokes equations. Math. Comp. 22, 745–
762 (1968)
2. Couderc, F.: Développement d’un code de calcul pour la simulation d’écoulements de flu-
ides non miscibles. Application à la désintégration assistée d’un jet liquide par un courant
gazeux. PhD thesis (2007)
3. Fedkiw, R., Aslam, T., Merriman, B., Osher, S.: A non-oscillatory Eulerian approach to
interfaces in multimaterial flows (the ghost fluid method). J. Comput. Phys. 152, 457–492
(1999)
4. Germano, M.: A proposal for a redefinition of the turbulent stresses in the filtered Navier-
Stokes equations. Phys. Fluids. 29, 2323–2324 (1986)
5. Osher, S., Sethian, J.: Fronts Propagating with Curvature Dependant Speed: Algoritms
based on Hamilton-Jacobi Formulations. J. Comput. Phys. 79, 12–49 (1988)
6. Shu, C.-W.: Essentially non-oscillatory and weighted essentially non-oscillatory schemes

for hyperbolic conservation laws. ICASE - NASA Report (1997)
7. Sussman, M., Smereka, P., Osher, S.: A Level Set Approach for Computing Solutions to
Incompressible Two-Phase Flow. J. Comput. Phys. 114, 146–159 (1994)
8. Trottenberg, U., Schuller, A.: Multigrid. Academic Press, Inc., London (2001)
Computation of Flow in a 3D Diffuser Using a
Two-Velocity Field Hybrid RANS/LES
J.C. Uribe, A. Revell, and C. Moulinec
Abstract. The flow inside a three-dimensional diffuser is computed with a two-

velocity hybrid RANS/LES model that ensues the separation of dissipative effects
of the mean and fluctuating fields to be treated individually; by extracting a running
average velocity field from instantaneous quantities. The averaged field is then used
to calculate the contribution of the mean shear which is larger than that from the
fluctuating one over the near wall region. Results with the hybrid model are pre-
sented and compared to those from a DES on the same grid. RANS results obtained
with models upon which both these approaches are based are also reported. The
RANS results are unable to capture essential characteristics of the flow whereas the
hybrid method produces results which generally agree well with the experiment.
1 Introduction
Turbulence modelling of separated flows remains a challenging subject. Industrial
codes rely heavily on the Reynolds Averaged Navier-Stokes (RANS) approach, in
which the effect of turbulence is modelled throughout the entire domain. This ap-
proach, although relatively low resource-demanding, has been proven to lack accu-
racy in regions of flow separation and the subsequent recirculation zone also present
physical characteristics that are not within the usual assumptions of standard RANS
models. In cases where the adverse pressure gradient is strong, rapid changes across
the boundary layer region remains a challenging feature to simulate.
In order to be able to compute such flows, many research groups have recently
been developing so-called Hybrid RANS/LES approaches, by attempting to com-
bine the virtues of RANS models with the accuracy provided by Large Eddy
J.C. Uribe · A. Revell
School of MACE, University of Manchester, M60 1QD, UK
e-mail: juan.uribe@manchester.ac.uk
C. Moulinec
STFC Daresbury Lab, Daresbury, Warrington WA4 4AD, UK
386 J.C. Uribe, A. Revell, and C. Moulinec
Simulation (LES) - where only the small scales are modelled (See [3] or [10] for
example). LES has recently enjoyed increasing popularity due to the rapid increase
of computational power, but its use on industrial applications remains limited. One
major reason for this lies in the modelling of the small scales; they are assumed to be
isotropic and as such it is necessary to refine the mesh as the anisotropy increases.
In the present paper, a zonal method that combines RANS and LES approaches is
employed to compute the flow on a three dimensional diverging diffuser [2]. The re-
sults of the model are compared with the standard hybrid Detached Eddy Simulation
(DES) method and with the base-line RANS models.
2 Modelling
In many hybrid approaches the main challenge resides in coupling the two velocity
fields coming from RANS (statistically averaged) and LES (filtered). This is often
done by applying a matching criteria, i.e. assuming the same turbulent viscosity at
the interface, the same kinetic energy or dissipation etc. [3], [10]. This poses a prob-
lem since these values represent intrinsically different properties in the two velocity
fields. Instead of a single velocity field to couple both RANS and LES, the model
presented here allows an overlap of both fields, with the RANS model driving the
near wall RANS velocity field, without damping in any way the dissipation of re-
solved fluctuations. Many sub-grid models assume the flow to contain an inertial
subrange and hence the sub-grid motions can be assumed to be isotropic. This is
true only if the grid is small enough for the anisotropy introduced by the mean shear
to be neglected. At high Reynolds numbers, the refinement of the grid becomes too
costly, therefore restricts standard LES methods to lower Reynolds numbers flows.
As the solid boundary is approached, the mean shear becomes sufficiently high to
introduce anisotropy across a range of diminishing scales. It is then necessary for
model to represent both the subgrid-scale contributions to the mean shear stress
and isotropic dissipation effects. The two-velocity hybrid model is based on defin-
ing separate contributions from the fluctuating field and from the averaged one, by
modelling each of them with a separate eddy viscosity, in a similar way as proposed
by Schumann [7].
The instantaneous velocity can be decomposed as U = U + u , where U is
time averaged and u is fluctuating. The residual stress is modelled as:
2
τirj − τkk δi j = −2νr fb (Si j − Si j ) −2(1 − fb)νa Si j , (1)
3 $ %& '$ %& '
locally isotropic inhomogeneous

Here fb = tanh 1.5 LΔt is a blending function devised to avoid double counting of
the stresses. The model requires the computation of a running averaged velocity
field to evaluate the inhomogeneous contribution, but this is performed at no ex-
tra CPU cost since only one set of momentum equations are solved. The isotropic
contribution is modelled using the standard Smagorinsky [8] model whereas the in-
homogeneous contribution is modelled using the elliptic relaxation eddy viscosity
Computation of Flow in a 3D Diffuser 387
v2
model of Laurence et al. [5] that solves a transport equation for the ratio ϕ = k in
order to take into account the near wall effects.

νr = (Cs Δ )2 2 Si j − Si j Si j − Si j (2)

νa = Cμ ϕ kT with T = max εk ,CT νε (3)
More details of this scheme can be found in [11].

The DES [9] approach used here is based on the SST model [6]. The destruction
term in the ω equation is multiplied by a function FDES that switches between the
turbulent length scale predicted by the RANS model (Lt = k3/2 /ε ) and a LES-type
length scale (CDES Δ ) in order to reduce the viscosity in regions of flow separation.:

Lt
FDES = max (1 − F1), 1 (4)
CDES Δ
where F1 is the blending function of the SST model that combines k − ε and k −
ω formulations based on the distance from the wall. The F1 blending function of
the model is used to “shield” the boundary layer from grid induced separation [6];
prevents erroneous switching from RANS to LES inside the boundary layer in the
presence of local mesh refinement.
3 3D Diffuser
The flow inside a three-dimensional diffuser has been studied experimentally by [2]
using Magnetic Resonance Velocimetry. Figure 1 shows the diffuser geometry, in
which two of the four sides of the duct are diverging, whilst the other two remain
orthogonal. The Reynolds number based on the inlet height, H and bulk velocity Ub
is Re = 10000. The inlet conditions are defined by a fully developed flow through
a rectangular duct of aspect ratio 1 : 3.33. The flow is characterised by a separation
zone that starts in the diverging (upper) corner and continues to grow along the top
inclined wall.
The case has been computed with the DES and the two-velocity model and with
the corresponding two RANS base-line models, the SST and the ϕ − f .
The mesh used in these computations has Nx = 212, Ny = 60, Nz = 180 cells,
with a nondimensionall wall-distance, y+ max < 1.2 in all the domain. The maximum
values of Δ x+ and Δ z+ on all the wall surfaces can be seen in figure 2 (calculated
from the DES results).
A precursor 2D simulation of a rectangular duct was carried out in order to obtain
inlet conditions with both RANS models. The resulting profiles were used for the
hybrid simulations, where in the case of the the two-velocity hybrid model, turbulent
fluctuations were added using the Synthetic Eddy Method (SEM) of [4].
Fig. 1 Diffuser Geometry
200 200
+ +
Δx lower wall Δx side wall
+ +
Δz lower wall Δz side wall
+ +
150 Δx top wall (diverging) 150 Δx side wall (diverging)
+ +
Δz top wall (diverging) Δz side wall (diverging)
+
+
Δx Δz
Δx Δz
100 100
+
50 50
0 0
0 5 10 15 20 0 5 10 15 20
x/H x/H
Fig. 2 Maximum Δ x+ and Δ z+ at the walls
4 Results
The case has been computed using the open-source CFD software Code Saturne
[1]. The code is based on the finite volume method, and a second order scheme for
both space and time has been used.
Both RANS models tested failed to reproduce correctly the detached flow fea-
tures. Figure 3 displays a qualitatively comparison of results from all models by
extracting YZ planes at several locations where shaded areas represent regions of
flow reversal. The early onset of separation in the diverging corner, as predicted by
the SST and the ϕ − f models causes the zone of recirculating flow to remain in
the vicinity of the side wall instead of to the top wall. Both RANS models seem
to exhibit a similar behaviour despite the inherent capability the ϕ − f model to
better resolve the near wall layer. Since both models are based upon the Boussi-
nesq approximation, they are not capable of resolving secondary vortices produced
in a rectangular duct. This in turn means that the boundary layer interaction on the
diverging wall are not accurately modelled.
On the other hand both hybrid models correctly show separation to appear on the
top wall, although the DES predicts a much larger recirculation region. The flow
initially separates at the diverging corner but almost immediately it creates a small
separation bubble on the opposite top corner. This results in a massive separated
region that accelerates the rest of the flow. Eventually the separated region reaches
the diverging side. The two-velocity hybrid model predicts a smaller detached flow
region than the experimental but although the location of the separation point is
close to observed in the the experiment, the recirculation length fails to span the
a) b) c)
d) e)
Fig. 3 Separated zone (dark). a) Experiment, b) SST, c) ϕ − f , d) DES, e) Hybrid
Exp.
3 DES
Hybrid
y/H
0 5 10 15 20
x/H, 2 U/Ubulk
Fig. 4 Stream wise velocity profiles at z/H = 0.5
entire top wall as expected. The streamwise velocity profiles at the z/H = 0.5 plane
are shown in figure 4 where the size of the recirculation length predicted by the
models are compared.
Figure 5 shows the urms contours, where DES levels are low at the inlet since
the model is not designed to reproduce turbulence in attached flows. The turbulence
levels grow from the edge of the recirculation region towards the core of the at-
tached flow which in the DES is strongly accelerated in the lower left corner. On the
contrary, the two-velocity hybrid model is able to sustain these fluctuations from the
inlet but with level that are too high near the top wall.
The DES is influenced to a greater extent by the RANS prediction than the two-
velocity model, since a larger proportion of the near wall layer is computed by
the SST. In figure 6 the zone treated only in RANS mode (dark) is shown at the
z/H = 0.5 plane. It can be seen that the LES mode is never activated in the near wall
region thus avoiding under-resolved LES-like results. In contrast, the two-velocity
model allows more contend by blending the contributions of RANS and LES.
a)
b)
c)
Fig. 5 urms contours at x/h = 2, 5, 8, 12, 15. a) Experiment, b) DES, c) Hybrid
a) b)
Fig. 6 RANS-LES Zones. a) DES, b) Hybrid
Because the hybrid method allows both regions to co-exist, it is less sensitive to the
choice of RANS model. The blending function used in equation 1 does not damp
the turbulent fluctuations since the fluctuating stress is combined with the averaged
via two different velocity fields.
5 Conclusions
The flow inside a three dimensional diffuser has been computed. The geometry
of the diffuser has two diverging wall at different angles which creates a complex
separation pattern. The eddy viscosity models used here fail to predict the correct
separation point and the region of recirculating flow is observed to remain close to
the side wall with the smaller diverging angle as oppose to the findings of the exper-
iments. The two-velocity model and the DES are able to predict the separated region
on the top wall but not to the correct extend. DES predicts an early separation point,
similar to the base SST model, which leads to a large recirculation region extending
across almost the entire streamwise length of the domain. The two-velocity hybrid
model does a better job at predicting the separation but it is unable to sustain the
separated layer leading to an early reattachment.
Acknowledgements. This work was partly supported by the UK Engineering and Physical
Sciences Research Council (EPSRC) under Grant No. EP/D053994/1. Additional support
was provided by EPSRC under the auspices of Collaborative Computational Project 12.
References
1. Archambeau, F., Mechitoua, N., Sakiz, M.: A finite volume method for the computation
of turbulent incompressible flows - industrial applications. International Journal on Finite
Volumes 1(1), 1–62 (2004)
2. Cherrye, E.M., Elkins, C.J., Eaton, J.K.: Geometric sensitivity of three-dimensional
separated ows. International Journal of Heat and Fluid Flow 29, 803–811 (2008)
3. Davidson, L., Peng, S.H.: Hybrid RANS-LES: A one equation SGS model combined
with a k − ω model for predicting recirculating flows. International Journal of Numerical
Methods in Fluids 43, 1003–1018 (2003)
4. Jarrin, N., Benhamadouche, S., Laurence, D., Prosser, R.: A synthetic-eddy method for
generating inflow conditions for large-eddy simulations. International Journal of Heat
and Fluid Flow 27, 585–593 (2006)
5. Laurence, D., Uribe, J., Utyuzhnikov, S.: A robust formulation of the v2-f model. Flow,
Turbulence and Combustion 73, 169–185 (2004)
6. Menter, F., Kuntz, M., Langtry, R.: Ten years of industrial experience with the SST
model. In: Hanjalić, K., Nagano, Y., Tummers, M. (eds.) Turbulence, Heat and Mass
Transfer 4, vol. 4, pp. 625–632 (2003)
7. Schumann, U.: Subgrid scale model for finite difference simulations of turbulent flows
in plane channels and annuli. Journal of Computational Physics 18, 376–404 (1975)
8. Smagorinsky, J.: General circulation experiments with the primitive equations: I the basic
equations. Monthly Weather Review 91, 99–164 (1963)
9. Spalart, P., Jou, W., Strelets, M., Allmaras, S.: Comments of feasibility of les for wings,
and on a hybrid RANS/LES approach. In: International Conference on DNS/LES, Rus-
ton, Louisiana, August 4-8 (1997)
10. Temmerman, L., Hadziabdić, M., Leschziner, M.: A hybrid two layer URANS-LES ap-
proach for large eddy simulation at high Reynolds numbers. International Journal of Heat
and Fluid Flow 26, 173–190 (2005)
11. Uribe, J., Jarrin, N., Prosser, R., Laurence, D.: Hybrid v2f rans les model and syn-
thetic inlet turbulence applied to a trailing edge flow. In: Turbulence and Shear Flow
Phenomena, pp. 701–706 (2007)
On the Dynamics of High Reynolds Number
Turbulent Axisymmetric and Plane
Separating/Reattaching Flows
Pierre-Élie Weiss, Sébastien Deck, Jean-Christophe Robinet, and Pierre Sagaut
Abstract. The present work focuses on the intrinsic properties of an axisymmetric

separating/reattaching flow. A numerical simulation of a compressible flow over
a cylinder extended by another cylinder of smaller diameter is performed at a
Reynolds number based on the diameter of the larger cylinder of 1.2 × 106. Sta-
tistical and fluctuating properties are compared with the available experimental data
and those of two additional configurations. First the plane counterpart of the axisym-
metric case allows us to assess the influence of three-dimensionality. Then a double
backward facing step designed from the half upper part of the plane case permits
us to survey the flow interactions. Finally a linear stability analysis is coupled with
two-point correlations unveiling the importance of the highest coherent modes in
the flow behaviour.
Pierre-Élie Weiss
ONERA, Applied Aerodynamics Department, Missiles Hypersonic Launchers Unit, 8,
rue des Vertugadins, 92190 Meudon, France
e-mail: peweiss@onera.fr
Sébastien Deck
ONERA, Applied Aerodynamics Department, Missiles Hypersonic Launchers Unit, 8,
rue des Vertugadins, 92190 Meudon, France
Jean-Christophe Robinet
SINUMEF - Arts et Métiers - ParisTech, 151 boulevard de l’Hôpital, 75013 Paris, France
e-mail: Jean-Christophe.Robinet@paris.ensam.fr
Pierre Sagaut
Institut Jean Le Rond d’Alembert, Université Pierre et Marie Curie, case 162, 4, Place Jussieu,
75252 Paris Cedex 05, France
e-mail: sagaut@lmm.jussieu.fr
394 P.-É. Weiss et al.
1 Introduction
Unsteady flow mechanisms such as wall-bounded turbulent shear phenomena with
separation and reattachment have been studied for decades. Among them, axisym-
metric and plane step flows present similarities regarding the shear layer instability
process as emphasized by Deck and Thorigny [5]. Nevertheless, the amount of un-
steady analysis and experimental data about the disparities between plane and ax-
isymmetric step flows is very limited in the literature. Roshko [10] has stressed the
lack of a theory analogous to the one of Von Kàrmàn for the plane vortex street.
Recently, Deprés et al. [6] have highlighted the features of several axisymmetric
base flows, depending on the occurrence of a downstream solid reattachment of the
free shear layer or not. Concerning the flow interactions Ho and Huerre [7] have re-
minded the importance of the global feedback mechanism in the shear layer process.
Finally, instability mechanisms have been investigated numerically by Sandberg
and Fasel [11] in supersonic axisymmetric wakes behind a bluff body for several
Reynolds numbers and related unstable modes to coherent structures.
The purpose of this work is to couple the unsteady analysis with the linear
stability theory after a preliminary introduction of the configurations and the general
instantaneous behaviour of the flows. Finally, footprints are proposed to explain the
underlying mechanisms related to the dynamics of the step flows.
2 Methods and Configuration Set Up

The axisymmetric afterbody dimensions (see figure 1) were chosen to dovetail with
the experiments of the ONERA’s S3Ch wind tunnel described by Deprés et al. [6]
and later by Meliga and Reijasse [8]. The configuration is placed into a flow with
a free stream Mach number of 0.702 yielding a Reynolds number based on the
forebody diameter ReD ≈ 1.2 × 106.
Fig. 1 Sketch of the axisymmetric geometry exhibiting the mesh, velocity vectors upstream
from the separated area, characteristic dimensions and the mean organisation of the flow
On the Dynamics of High Reynolds Number Turbulent Axisymmetric 395
Fig. 2 Schematic view of the plane case showing the mesh, velocity vectors upstream from
the separated area, characteristic dimensions and the mean organisation of the flow
Fig. 3 Cutaway through the double backward facing step geometry depicting the mesh,
velocity vectors upstream from the separated area, characteristic dimensions and the mean
organisation of the flow
As represented in figure 2, the plane case consists in an extrusion of a longitudinal

plane of symmetry of the axisymmetric configuration over 1.2D and the double
backward facing step (DBFS) is the upper part of this plane case providing two step
heights equal to 0.3D and 0.2D (see figure 3). The approach used to model the flow
is Zonal Detached Eddy Simulation (ZDES) proposed by Deck [4] and the mesh1
includes 12 million hexaedric cells, with 240 cells in the azimuthal direction. The
upstream part is computed using a URANS model, which ensures a boundary layer
with the expected integral properties. In other words, the coupling of URANS and
LES ignores possible interaction of coherent structures from the oncoming boundary
layer over the cylinder and the separation bubble. However, this possible interaction
is assumed to be negligible which is broadly accepted for high Reynolds number
massively separated flows.
1 Meshes represented in figures 1,2 and 3 are the part of interest of the whole meshes which
are wider in all directions.

The validation of the axisymmetric simulation has been exhibited in Weiss
et al. [12]. The behaviour of the instantaneous flow field for the three cases is illus-
trated in figure 4, 5 and 6. The turbulent structures are evidenced plotting values of
the second invariant of the velocity gradient tensor Q defining the footprints of vor-
tex tubes. In all cases, a roll-up of toroidal eddies apparently occurs which grow by
pairing and are rapidly replaced by large three-dimensional structures. They develop
as the shear layer approaches reattachment and are destabilized for the axisymmetric
case by azimuthal instability modes according to Deck and Thorigny [5]. The plane
case exhibits a vigorous shedding whereas the DBFS seems to be led by the de-
velopment of the mixing layer. Some similarities with the axisymmetric bluff body
discussed by Sandberg and Fasel [11] such as oblique or helical structures, longitu-
dinal structures and several vortex loops can be noticed. However, their scales differ
because the convective Mach number and the Reynolds number are quite different.
Fig. 4 Map of non dimensional Q (i.e. Q.U∞2 /D2 ) evidencing the coherent structures down-
stream the axisymmetric step flow (black stands for negative values and grey for positive
values)
Fig. 5 Dimensionless Q criterion evidencing the coherent structures downstream the plane
case (black stands for negative values and grey for positive values)
Fig. 6 Visualization of non dimensional Q criterion planes evidencing the coherent structures
downstream the double backward facing step flow (black stands for negative values and grey
for positive values)
Fig. 7 Spectral sheet of the azimuthal pressure mode Cr,1 for every streamwise location along
the emergence in function of nondimensional frequencies, −− : spatial limits of the absolute
area for m = 1
To get deeper insight into the spatial organization of the flow at these frequencies,
the azimuthal coherence of two pressure sensors p1 (r, x, φ1 ) and p2 (r, x, φ2 ) located
in a plane normal to the inflow at a constant position x and a constant radius r can
be considered. Assuming the hypothesis of an homogeneous flow, i.e. without any
preferred angle of reference φ1 , the complex coherence function may be expressed
as:
C ( f , r, x, Δ φ ) = (Cr + jCi ) ( f , r, x, Δ φ )
S12 ( f , r, Δ φ , x)
= (1)
S1 ( f , r, φ1 , x) S2 ( f , r, φ2 , x)
√
where j = −1, Cr and Ci are the real and imaginary part of the cross-spectral
density function S12 and Δ φ = φ1 − φ2 , respectively. Assuming the disturbances do
not exhibit any particular direction of propagation one has S12 (Δ φ ) = S12 (−Δ φ ).
Besides, the hypothesis of isotropy yields Ci = 0. Consequently, the Cr function is
2π periodic with respect to Δ φ and can be expressed thanks to a Fourier transform
in azimuthal modes:
∞
Cr ( f , Δ φ ) = ∑ Cr,m ( f ) cos (mΔ φ ) (2)
m=0
Cr,m represents the percentage of the fluctuating energy at frequency f relative

to the azimuthal mode m since ∑m Cr,m = 1. Let us be reminded that m = 0 and
m = 1 modes are respectively characterized by an inphase and anti-phase relation-
ship of signals recorded simultaneously at two diametrically opposed locations. In
the present case, the Cr,m spectrum has been plotted for every streamwise location
providing a spectral layer at the wall (figure 7). It exhibits the highest percentages
of the fluctuating energy for Strouhal numbers StD based on the forebody diameter
close to 0.2 and which are localized between x/D ≈ 0.35 and x/D ≈ 0.75. Let us
now consider a perturbation approach of this flow.
In the context of spatio-temporal stability analysis performed for a locally paral-
lel flow assumption, the singularities of the resulting dispersion relation are deter-
mined and the ones satisfying the Briggs-Bers pinching criterion (see Briggs [2] and
Bers [1]) have been selected. Results summarized in figure 8 show that the mean flow
is linearly absolutely stable for an axisymmetric disturbance (m = 0) for all stream-
wise locations and frequencies. Concerning a helical disturbance (m = 1), the mean
flow is linearly absolutely unstable for a broad range of abscissa x/D ∈ [0.35; 0.75]
Fig. 8 Streamwise evolution of the growth rate Im(ω ). ◦: m = 0 mode for the axisymmetric
case, : m = 1 mode for the axisymmetric case, 2: plane case, : double backward facing
step. Solid line represents the threshold of marginal absolute instability for Im(ω ) = 0
Fig. 9 Streamwise evolution of the Strouhal number StD based on the diameter of the larger
cylinder. ◦: m = 0 mode for the axisymmetric case, : m = 1 mode for the axisymmetric case,
2: plane case, : double backward facing step
close to StD ≈ 0.2 (see figure 9). The co-existence of absolute helicoidal (m = 1)
unstable global modes within the recirculation region and convectively unstable
shear-layer modes corroborates the results of Sandberg and Fasel [11]. Indeed, the
linear criterion proposed by Chomaz et al. [3] points out that the dominant global
mode is close to StD = 0.227 which is in excellent agreement with ZDES results.
Indeed, the spectrum of the azimuthal antisymmetric pressure mode, depicted in fig-
ure 7, finds a maximum of fluctuating energy (centered around StD 0.2) precisely
in the absolutely unstable area of the mean flow. Concerning the plane and DBFS
cases, instabilities remain of convective nature. The explanation for the DBFS seems
to lie in the presence of the walls strongly stabilizing the flow. Regarding the instan-
taneous behaviour of the plane case flow which acts as an oscillator, an absolute
instability could govern the dynamics of the flow downstream the extension as in
two-dimensional bluff body flows [9].
4 Conclusions
Axisymmetric turbulent separating/reattaching flow has been investigated at ReD ≈
1.2 × 106. It is shown that a large scale coherent motion at frequency f .D/U∞ ≈ 0.2
dominates an ordered structure in this separating/reattaching flow. The stability anal-
ysis has allowed us to show that axisymmetric turbulent afterbody dynamics is
clearly dominated by a helical absolute instability while other modes (m = 0 and
m ≥ 2) remain of convective nature. These results are in excellent agreement with
the azimuthal antisymmetric pressure mode spectra. This last point is of major im-
portance since the analysis of the flow dynamics and the perturbation technique
converge on the same spatial information. Concerning the double backward fac-
ing step, the flow remains convectively unstable likely due to the presence of the
walls stabilizing the flow. Finally the plane case flow, contrary to its axisymmetric
counterpart which exhibits an absolute instability above the extension, could be gov-
erned by an absolute instability downstream the extension as for two-dimensional
oscillator configurations. This should be demonstrated in future works.
Acknowledgements. The authors thank the Centre National d’Etudes Spatiales (CNES) for
financial support within the framework of the research and technology program ATAC (aero-
dynamics of nozzles and after-bodies). The Ph.D. work of P.E. Weiss is funded by CNES and
ONERA.
References
1. Bers, A.: Linear waves and instabilities. In: Peyraud, C.D.J. (ed.) Physique des Plasmas,
p. 117, Gordon and Breach (1975)
2. Briggs, R.J.: Electron-Stream Interaction with Plasmas. MIT Press, Cambridge (1964)
3. Chomaz, J.M., Huerre, P., Redekopp, L.M.: A frequency selection criterion in spatially
developing flows. Studies in Applied Mathematics 84, 119–144 (1991)
4. Deck, S.: Zonal-detached-eddy simulation of the flow around a high-lift configuration.
AIAA Journal 43(11), 2372–2384 (2005)
5. Deck, S., Thorigny, P.: Unsteadiness of an axisymmetric separating-reattaching flow.
Phys. Fluids 19, 065103 (2007)
6. Deprés, D., Reijasse, P., Dussauge, J.P.: Analysis of unsteadiness in afterbody transonic
flows. AIAA Journal 42(12), 2541–2550 (2004)
7. Ho, C., Huerre, P.: Perturbated free shear layer. Ann. Rev. Fluid Mech. 16, 365–424
(1984)
8. Meliga, P., Reijasse, P.: Unsteady transonic flow behind an axisymmetric afterbody with
two boosters. In: Proceedings of the 25th AIAA Applied Aerodynamics Conference,
Miami, June 2007, vol. 2007-4564 (2007)
9. Monkewitz, P.A., Nguyen, L.N.: Absolute instability in the near-wake of two-
dimensional bluff bodies. J. Fluids Struct. 1, 165–184 (1987)
10. Roshko, A.: On the wake and drag of bluff bodies. J. Aero. Sci. 22, 124–132 (1955)
11. Sandberg, R.D., Fasel, H.F.: Numerical investigation of transitional supersonic axisym-
metric wakes. J. Fluid Mech. 563, 1–41 (2006)
12. Weiss, P.E., Deck, S., Sagaut, P.: Zonal-Detached-Eddy-Simulation of a Two-
Dimensional and Axisymmetric Separating/Reattaching Flow. In: AIAA paper
2008-4377 (2008)
Numerical Simulation and Statistical Modeling
of Inertial Droplet Coalescence in Homogeneous
Isotropic Turbulence
Dirk Wunsch, Pascal Fede, Olivier Simonin, and Philippe Villedieu
Abstract. A comparative parameter study is performed in order to analyze the

influence of turbulence on the rate of droplet coalescence. Therefore, Direct Nu-
merical Simulations (DNS) of the fluid turbulence are coupled with a Lagrangian
tracking of the particle phase (DPS) accounting for collisions leading to coales-
cence and to a broad droplet size distribution. In addition the accuracy of stochastic
collision models is evaluated by comparison of Monte-Carlo predictions with the
obtained results from the DNS/DPS simulations and statistical collision models are
evaluated.
1 Introduction
The study of collisions between particles in turbulent dispersed multiphase flows is
of interest in many engineering applications. In particular, flows in which droplet
coalescence is likely to play an important role include many topics such as solid-
fuel rocket propulsion, internal combustion engines and electric generation by a
liquid fuel turbine. The droplet size generally has a major influence on the global
performance of the system and must be accurately taken into account in numerical
Dirk Wunsch · Pascal Fede · Olivier Simonin

Université de Toulouse, INPT, UPS, IMFT, 31400 Toulouse, France,
CNRS, Institut de Mécanique des Fluides de Toulouse, 31400 Toulouse, France
e-mail: wunsch@imft.fr,fede@imft.fr,simonin@imft.fr
Philippe Villedieu
ONERA; DMAE, 2 avenue Edouard Belin, 31055 Toulouse, France
e-mail: Philippe.Villedieu@onera.fr
402 D. Wunsch et al.
simulations. As the carrier flow is often turbulent, droplets located in the vicinity
of the same point may have different velocities, collide and perhaps coalesce lead-
ing to a strong modification of the initial droplet size distribution. Several numer-
ical approaches can be used to investigate the effect of the fluid turbulence on the
coalescence rate in a cloud of inertial droplets (with a particle relaxation time ap-
preciably larger than the Kolmogorov time scale). One possibility is the stochastic
approach based on the Monte-Carlo method in order to solve the kinetic equation
governing the probability density function of the droplet cloud. Villedieu and Si-
monin [15] proposed a kinetic approach for the calculation of the collision probabil-
ity that takes into account the correlations between neighboring particles due to the
interaction with the fluid turbulence [13]. Model evaluations for simplified test cases
have shown that the correlation effect may modify appreciably the droplet growth
rate prediction in comparison with results obtained assuming standard molecular
chaos assumption. In order to analyze the influence of the colliding particle velocity
correlations induced by the fluid turbulence on the rate of droplet coalescence, re-
veal coalescence mechanisms and to validate this new statistical modeling approach,
deterministic Lagrangian simulations of inertial particles (DPS) coupled with Direct
Numerical Simulations (DNS) of the fluid turbulence have been carried out.
2 DNS/DPS Approach
Direct Numerical Simulations (DNS) coupled with a Lagrangian tracking of the
particle phase (DPS) are performed here and have been extensively used to investi-
gate gas-particle flows [6], [9]. The flow configuration is a Homogeneous Isotropic
Turbulence (HIT) forced by a scheme initially proposed by Eswaran and Pope [5],
which assures a statistical steadiness. The particles are considered as rigid spheres
with diameters smaller than the Kolmogorov turbulence length scale ηK . The tur-
bulence modulation by the dispersed phase (two-way coupling) is not considered,
as the particle mass fraction is small. Assuming that the particle fluid density ratio
(ρ p ρ f ) is large, the forces acting on the particle are reduced to the drag force
only. Thus, the governing equations of the N p particle system in interaction with the
surrounding flow field and undergoing particle-particle collisions is written as
dx p
= vp
dt
/ 0 Np
dv p v p − u f @p
mp = mp + ∑ F p,i j (1)
dt τp i=1; j=i
where x p , v p are the position and velocity vectors of the particle p and m p is the
particle mass. u f @p is the fluid velocity at the position of the particle and F p,i j rep-
resents the impulsive force resulting from particle-particle collisions. As two-way
coupling is neglected, u f @p is computed with an accurate interpolation scheme [3].
The particle response time τ p is given using the relation of Schiller and Naumann
[11] by
Numerical Simulation and Statistical Modeling 403
4 ρp dp 1
τp =
3 ρ f CD v p − u f @p

24 v p − u f @p d p
CD = 1 + 0.15Re0.687 Re p = (2)
Re p p
νf
with d p the particle diameter and ν f the kinematic viscosity of the fluid. Coalescence
is modeled assuming that each collision leads to permanent coalescence. Other col-
lision outcomes as identified by several authors [1],[4],[10] are not regarded for the
sake distinctness. The mass and momentum conservation equations for two particles
undergoing coalescence are written as
m∗ = m p + mq
m∗ v ∗ = m p v p + mq v q (3)
with m p and mq the mass of the particles before coalescence and m∗ after. Analo-
gous for the particle velocities v p , vq and v∗ . The corresponding particle diameter
is directly deductible from the mass conservation equations as the particle density
is constant and the particles are modeled as rigid spheres, as mentioned above. The
position of the new particle that arises from coalescence is given as
d 3p x p + dq3 xq
x∗ = (4)
d∗
3
with x∗ the position of the new particle and d p , dq and d ∗ the particle diameters.
Coalescence is detected using a recently developed algorithm allowing detecting
collisions in a poly-dispersed particle mixture [16].
3 Monte-Carlo Approach
The Monte-Carlo method represents a stochastic approach to solve the kinetic equa-
tion governing the probability density function of the droplet cloud. The particle
PDF transport equation is written by Simonin in [12] introducing the fluid-particle
joint PDF. Villedieu and Simonin [15] proposed a new kinetic approach in order to
account for correlations between neighboring particles in the collision probability
due to their interaction with the fluid. The here applied method is of the Babovsky
type [2] and described shortly in the following. The particles are sorted by their fluid
velocity and grouped into pairs, the collision probability for each pair is then calcu-
lated. A rejection method is used to decide whether the collision takes place or not.
This procedure is applied for all pairs of particles, then the system advances to the
next time step. In order to account for the correlation of the particle velocities by the
fluid, the particles fluid velocity is discretized into several classes. The Babovsky
algorithm is applied within each of these classes. Therefore, only particles collide
that show a similar fluid velocity.
4 Fluid-Particle Flow Fields

The fluid flow field is a Homogeneous Isotropic Turbulence. Three turbulent flow
fields are compared in this study, in order to investigate the influence of the turbu-
lent Reynolds number on particle statistics. Once the fluid flow field reaches a steady
state the particulate phase is initiated and reaches an equilibrium defined by Tchen
and Hinze [14] with the fluid turbulence. From this equilibrium state on the coa-
lescence effect is started and collisions between particles lead now to coalescence.
The particle Stokes numbers are chosen to be identical for different Reynolds num-
bers. The initial particle diameter d p is the same in all simulations, only the particle
density ρ p is modified for different particle inertia.
5 Coalescence in DNS/DPS and Model Comparison
5.1 Influence of Particle Inertia

Figure 1 (left) shows the reduction of particle number due to coalescence in a
turbulent fluid flow at Reynolds number ReL = 61 for different particle inertia.
d 3p
Figure 1 (right) shows the corresponding Sauter diameter ds = d 2p
. The lowest coa-
lescence rate is found for particles with a small Stokes number and increases with an
increasing Stokes number. This means that higher particle inertia lead to more col-
lisions, which is consequential. The highest coalescence rate, however, is not found
for St = 2.4, which represents the highest particle inertia investigated here, but for
intermediate Stokes numbers. This seems to be surprising but may be explained by
preferential particle concentration. Simonin et al. [13] showed in turbulent gas-solid
1.05 1.2
St = 0.03 St = 0.03
1 St = 0.23 St = 0.23
St = 0.80 1.16 St = 0.80
St = 1.65 St = 1.65
0.95 St = 2.40 St = 2.40
1.12
0.9
Np
Np
0.85 1.08
0.8
1.04
0.75
1
0.7
0.65 0.96
0 1 2 3 4 0 1 2 3 4
time [s] time [s]
Fig. 1 Number of particles (left panel) and Sauter diameter (right panel) in coalescence with
respect to Stokes number. Numerical simulation with ReL = 61
flows that particles with a particle response time in the order of the Lagrangian time
scale tend to accumulate in regions of low vorticity due to the interaction with fluid
turbulence. This effect increases the particle volume fraction locally and may lead
to a higher coalescence rate and consequently explain an elevated coalescence rate
for these particle inertia. The behavior of the Sauter diameter, in figure 1 (right),
corresponds to the coalescence rates and the effect of the Stokes number accords.
The initial particle size distribution is a monodisperse Dirac function. Figures 2
show the time development of the initial particle distribution for the smallest and
largest Stokes number investigated here.
0 0
10 10
t=0s t=0s
-1 -1
10 t = 1.9 s 10 t = 1.6 s
t = 4.0 s t = 4.0 s
-2 -2
10 10
Np
Np
-3 -3
10 10
-4 -4
10 10
-5 -5
10 10
0.00015 0.0002 0.00025 0.0003 0.00035 0.00015 0.0002 0.00025 0.0003 0.00035
diameter [m] diameter [m]
Fig. 2 Time-evolution of the particle diameter distribution measured from DPS / DNS for
St = 0.03 (left) and St = 2.4 (right)
5.2 Comparison with Monte-Carlo Predictions

The coalescence rates in figure 3 (left panels) and the particle kinetic energy in
figure 3 (right panels) obtained from the DNS/DPS simulations are compared to
predictions from Monte-Carlo simulations. First, applying the standard uncorrelated
collision model based the ’molecular chaos’ assumption (dashed lines) and second,
applying a correlated model (solid lines), which takes fluid-particle correlations into
consideration. This is done by applying the same collision algorithm as in the uncor-
related case to subsections of the particle vector, which is sorted by means of fluid
velocity. This has a consequence that particles will collide with particles, which see
a similar fluid velocity. In these simulations 10 sections are used.
It can be seen that the correlated collision model predicts a considerably better
coalescence rate than the uncorrelated model for either small and large particle in-
ertia, as seen in figure 3 (left). Also the particle kinetic energy is better predicted
using the correlated model (figure 3 (right)).
1.1 0.01
St = 0.03 MC correlated
St = 0.03 MC uncorrelated St = 0.03 MC correlated
1
DNS/DPS St = 0.03 MC uncorrelated
DNS/DPS
0.008
0.9
0.8
2
qp
Np
0.006
0.7
0.6
0.004
0.5
0.4 0.002
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
time [s] time [s]
1.1 0.004
St = 0.80 MC correlated
1
DNS/DPS
0.9
0.8
2
qp
Np
0.003
0.7
0.6
0.5
0.4 0.002
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
time [s] time [s]
1.1
St = 2.4 MC correlated 0.0024

1
DNS/DPS
0.9
0.002
0.8
2
qp
Np
0.7
0.0016
0.6
0.5 0.0012
0.4
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
time [s] time [s]
Fig. 3 Comparison of particle number (left panel) and of particle kinetic energy (right panel)
in the coalescence regime for DNS/DPS and Monte-Carlo simulations with respect to Stokes
number. White symbols represent the uncorrelated collision model and black symbols a cor-
related collision model. For St = 0.03 (top), St = 0.8 (middle) and Stokes St = 2.4 (bottom)
Acknowledgments
Numerical simulations were performed on NEC-SX8 supercomputer. The CPU time has been
provided by the Institut de Développement et des Ressources en Informatique Scientifique
(IDRIS) in the frame of the Project 91066. This research project has been supported by a
Marie Curie Early Stage Training Fellowship of the European Community’s Sixth Framework
Programme under contract number ’MEST-CT-2005-020429’.
References
1. Ashgriz, N., Poo, J.Y.: Coalescence and separation in binary collisions of liquid drops. J.
Fluid. Mech. 221, 183–204 (1990)
2. Babovsky, H.: On a simulation scheme for the Boltzmann equation. Mth Meth. in Appl.
Sc. 8, 223–233 (1986)
3. Balachandar, S., Maxey, M.R.: Method for evaluating fluid velocities in spectral simula-
tions of turbulence. J. of Comput Physics 83, 96–125 (1988)
4. Estrade, J.P.: Etude expérimentale et numérique de la collision de gouttelettes. Ecole
national supérieure de l’aéronautique et de l’espace: Thèse (1998)
5. Eswaran, V., Pope, S.B.: An examination of forcing in direct numerical simulations of
turbulence. Computers and Fluids 16(3), 257–278 (1988)
6. Fede, P., Simonin, O.: Modelling of kinetic energy transfer by collision in a non-settling
binary mixture of particles suspended in a turbulent homogeneous isotropic flow. In:
Proceedings of the ASME FEDSM2003 FEDSM2003–45735 (2003)
7. Fede, P., Simonin, O., Villedieu, P.: Monte-Carlo simulation of colliding particles in gas-
solid turbulent flows from a joint fluid-particle PDF equation. In: Proceedings of the
ASME FEDSM2002 FEDSM2002–31226 (2002)
8. Hylkema, J.J., Villedieu, P.: A random particle method to simulate coalescence phenom-
ena in dense liquid sprays. In: Proc. of 16th Int. Conf. on Numerical Methods in Fluid
Dynamics. Lecture notes in physics, vol. 515. Springer, Heidelberg (1999)
9. Laviéville, J., Deutsch, E., Simonin, O.: Large Eddy Simulation of interaction between
colliding particles and a homogeneous isotropic turbulence field. In: 6th Int Symp on
Gas-Solid Flows, FEDSM 2005, vol. 228, pp. 347–357 (2005)
10. Qian, J., Law, C.K.: Regimes of coalescence and separation in droplet collision. J. Fluid
Mech. 331, 59–80 (1997)
11. Schiller, L., Naumann, A.: A drag coefficient correlation. VDI Zeitung 77, 318–320
(1935)
12. Simonin, O.: Combustion and turbulence in two-phase-flows. Lecture Series 1996-02.
von Karman Institute for Fluid Dynamics (1996)
13. Simonin, O., Février, P., Laviéville, J.: On the spatial distribution of heavy-particle veloc-
ities in turbulent flow: from continuous field to particulate chaos. Journal of Turbulence
(2002), doi:10.1088/1468-5248/3/1/040
14. Tchen, C.M.: Mean value and correlation problems connected with the motion of small
particles suspended in a turbulent fluid. Thesis: Delft, Martinus Nijhoff, The Hague
(1947)
15. Villedieu, P., Simonin, O.: Modeling of coalescence in turbulent gas-droplet flows.
Comm. Math. Sci. Suppl. Issue 1, 13–33 (2004)
16. Wunsch, D., Fede, P., Simonin, O.: Development and validation of a binary collision
detection algorithm for a poly-dispersed particle mixture. In: Proceedings of ASME
FEDSM2008 FEDSM2008–55159 (2008)
Gas-Phase Mixing in Droplet Arrays
M.R.G. Zoby, S. Navarro-Martinez, A. Kronenburg, and A.J. Marquis
Abstract. Droplet evaporation is usually modelled as a subgrid process and induces

local inhomogeneities in the mixture fraction probability density function (PDF) and
its scalar dissipation. These inhomogeneities are usually neglected, however, they
can be significant and determine the combustion regime. In the present work, Di-
rect Numerical Simulations (DNS) of fully resolved evaporating methanol droplets
are analysed, assessing fuel vapour mixing in laminar and turbulent flows. The re-
sults show that scalar probability distributions and scalar dissipation vary greatly
depending on the position relative to the droplet position, on droplet loading and on
flow conditions. The β -PDF seems to capture the global behaviour for laminar flows
around droplet arrays with low droplet density, however, mixing characteristics for
higher droplet densities in stagnant and turbulent flows cannot be approximated by a
β -PDF, and modelling approaches based on cell mean values will lead to erroneous
results.
1 Introduction
Modelling sprays and spray combustion is a challenging task due to the wide range
of the associated length and time scales and the complex interaction of turbu-
lence and the physico-chemical processes, such as heat and mass transfer across
Maria Regina Gomes Zoby
Imperial College, SW7 2AZ, London, UK
e-mail: m.zoby06@imperial.ac.uk
Salvador Navarro-Martinez
e-mail: s.navarro@imperial.ac.uk
Andreas Kronenburg
University of Stuttgart, 70569 Stuttgart, Germany
e-mail: kronenburg@itv.uni-stuttgart.de
Andrew J. Marquis
e-mail: a.marquis@imperial.ac.uk
410 M.R.G. Zoby et al.
interfaces. Most computational studies of evaporating sprays do not resolve the

liquid phase nor the near field and the droplets are treated as point sources of
mass, momentum, energy and species. Even if evaporation -and to some degree
combustion- are directly dependent on local conditions, these studies neglect the
conditions in the immediate neighbourhood of the individual droplet. Large differ-
ences will be found between local and cell-averaged fuel concentrations in the inter-
droplet region [2]. These unresolved inhomogeneities can have profound effects on
the accuracy of mixture fraction based combustion models that rely on strong de-
pendencies of scalar dissipation and reactive species on mixture fraction within one
CFD cell. Schroll et al. [4] pointed out that satisfactory closures may not be ob-
tained using the source point approximation due to lack of resolution in the near
liquid field.
In the present work, 2D DNS of methanol droplets are analysed. The droplets are
organised in infinite layers of droplets with different thicknesses. Several droplet
densities in stagnant and convective environments are simulated and the sensitivity
of the local fuel mixing field to the droplet density and to the inflow conditions is
quantified, i.e. the evolution of mixture fraction PDF and its dissipation are analysed.
Droplets and interdroplet spaces are fully resolved. The relative velocity between
droplets and surroundings varies for the cases with convection from 5 m/s to 35
m/s. We recognize that 2D turbulence does not realistically represent the evolution
of real 3D turbulence, however, the focus of this work is the variation of the mixing
field due to small scale turbulent structures of the order of the droplet spacing and
an accurate description of the turbulence seems to be of secondary importance.
The paper is organised as follows: The second section presents the computational
approach and the model validation. The results of the gas-phase mixing are then
presented and discussed in the following section.
2 Computational Approach and Model Validation

The model used in the present work combines the one-fluid and two-fluid formula-
tions for multiphase flows where energy is solved based on a one-fluid formulation
while species, velocities and pressure are solved with a two-fluid formulation. This
method avoids the introduction of interface thickness and the smoothing of fluid
properties. The physical properties (density, diffusivity, specific heat and conductiv-
ity) are defined in each cell according to a phase indicator, local composition and
temperature. Density ratios up to 700 were tested and the results show that the code
is stable and the jumps in properties, scalars and velocities are captured.
In order to track the interfaces between liquid and gas, the Level Set method [5]
has been combined with the Ghost Fluid method [1] to account for pressure jumps
in a mass conserving approach. The Level Set method is based on the definition
of a smooth function, such as the minimum distance of a point to the interface, in
the entire computational domain. Negative values correspond to one phase while
positive values to the other and the interface is represented by the zero level of the
Gas-Phase Mixing in Droplet Arrays 411
function. Advection of the Level Set function, φ , is given by the following equation
in standard notation,
∂φ ∂φ
+ uiI =0 (1)
∂t ∂ xi
where uiI is the local interface velocity in the ith direction.
The model is implemented in an in-house low-Mach number CFD code
(BOFFIN). The method and its validation are described in detail elsewhere [6].
Potential sources of errors in the model are surface tension modelling and spu-
rious currents of mass, interface area calculation and momentum transfer between
phases. For the model used in this work, the maximum velocities generated due to
spurious currents vary linearly with the ratio of surface tension and viscosity, σμ , and
are lower than 0.1 m/s for 0.1 m/s< σμ < 100 m/s. In addition, the errors in the cal-
culation of the total surface are lower than 7% for droplets with 4 cells accross the
diameter and lower than 1% for droplets with more than 10 cells when compared
with the surface area of perfectly spherical droplets. The coupling of momentum
between the phases is also satisfactory and computed drag coefficients are within
15% of empirical values given in the literature [6].
3 Results
A 2D space that extends infinitely in x-direction and 3 mm in y-direction is popu-
lated with one, two, three or four infinite rows of equidistant droplets. Inflow/outflow
boundary conditions perpendicular to the droplet layers are used with periodic
boundary conditions in the remaining direction. Scalar dissipation and the mixture
fraction field are investigated for arrays of droplets of diameter of 100 μ m. The grid
mesh is Δ x=Δ y=5 μ m and the simulations run until 90% of the droplet mass has
evaporated. Different regions need to be defined to allow a zone dependent analysis
of the local conditions of the combustible mixture. Three regions are defined: the
global region comprises all of the computational domain; the inner and outer re-
gions refer to inter-droplet spaces and regions outside the droplet cloud. In addition,
in the four-droplet case, different interdroplet regions are also identified (see figure
1) and analysed. Liquid is assumed at 338 K with ρl = 750 kg/m3, σ = 1.85 10−2
N/m, μ = 3.5 10−4 Pa.s and h f g = 1.097 MJ/kg.
In the laminar flow test case, hot air (2530 K) is injected at 5 m/s and in the
turbulent flow test the mean velocity is 10 m/s with maximum fluctuations of 25
m/s. Furthermore, in the turbulence test, the flow field is initialised with velocities
scaled from a 2D turbulent field and the inflow is given by 2D turbulent DNS data.
In order to sustain the turbulence through the domain, the gas phase viscosity was
fixed at a lower value (μ = 10−6 kg/m.s) and pressure was increased from 1 atm to
5 atm. The Kolmogorov length scale is of the order of the droplet radius. A typical
vorticity field is presented in figure 2, and it is evident from the figure that the flow
is strongly affected by the droplets and develops differently for each of the cases.
Fig. 1 Geometry of the 4-droplet layer test case domain and defnition of the regions (left
figure). The right figure illustrates the droplet configurations for the 1-, 2- and 3-droplet layer
test cases
Fig. 2 Vorticity field after 5 ms for the 4- (left), 2- (centre) and 1- (right) droplet layer test
cases
A first quantitative analysis of the spatial evolution of the mixture fraction field
involves a comparison of the computational results with an analytic description of
the near droplet field for inertial droplets. Following [3], mixture fraction, f , along
a transversal line perpendicular to the mean flow field in the near droplet region is
given by Equation 2.
−r2
Jm ( fd − f∞ ) 4D u l
f = f∞ + e mean , (2)
4πρ D umean
l
f∞ is the value of mixture fraction far from the droplets, fd is the value of mixture
fraction inside the droplet initially centered at (xd , yd ), r is the transverse distance
from the droplet centre, r = |x − xd |, l is the distance from the line to the droplet,
l = y − yd , umean is the flow mean velocity, Jm is the mass evaporation rate of a
single droplet, ρ and D are the density and diffusivity in the interdroplet space,
respectively. Equation (2) gives good agreement with the simulations for the laminar
case as depicted in figure 3, where results from the one- and two-droplet layer cases
are shown. In the two-droplet layer case, one line per droplet is assessed.
0.3
Asymptotic Solutions
1 droplet
2 droplets - lower droplet
0.25 2 droplets - higher droplet
0.2
Mixture Fraction
0.15
0.1
0.05
0
0 0.1 0.2 0.3 0.4
Radial distance / Interdroplet distance
Fig. 3 Mixture fraction along a transversal line (l = 500μ m) in the laminar flow test
With respect to the probability distribution, it has been shown in [6] that in
stagnant environments, where expansion of the gas phase occurs only due to the
Stefan flow, the presumed β -PDF does not capture two dominant peaks that de-
velop at two different mixture fraction values in the three-droplet layer case. As
shown in figure 4, in the stagnant environment, the double peak does not occur
when the distance between the droplets is large (one-droplet case). However, once
the interdroplet distance decreases (two- and three-droplet cases) the bimodal nature
of the PDF appears. For the laminar flows, the β -PDF seems to capture better the
behaviour. In the turbulent environments, the β -PDF represents reasonably well the
probability for the lower droplets density cases but it does not capture the profile
for the higher density case (figure 4). It is clear that structural information must be
taken into account when modelling the mixture fraction PDF in two-phase flows,
however, defining the appropriate parameters is not evident.
1000 1000 1000

1 droplet - Turbulent 2 droplets - Turbulent 3 droplets - Turbulent
Laminar Laminar Laminar
Stagnant Stagnant Stagnant
100 Turbulent-β-pdf 100 Turbulent-β-pdf 100 Turbulent-β-pdf
Laminar-β-pdf Laminar-β-pdf Laminar-β-pdf
Stagnant-β-pdf Stagnant-β-pdf Stagnant-β-pdf
10 10 10
PDF
1 1 1
0.1 0.1 0.1
0.01 0.01 0.01
0.001 0.001 0.001

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Mixture Fraction Mixture Fraction Mixture Fraction
Fig. 4 Comparison of DNS data and a presumed β -shape for the mixture fraction PDF for
various flow fields and various droplet loadings
Figure 5 compares scalar dissipation for the different droplet arrays. It can be seen
that the higher the number of droplets (reduced interdroplet space) the lower is the
scalar dissipation in the laminar cases. However, for turbulent flows, this trend is not
repeated, and it cannot be assumed that a higher droplet concentration necessarily
decreases scalar dissipation in all the mixture fraction range.
10000
Laminar - 1 droplet 10000 Turbulent - 1 droplet
2 droplets 2 droplets
1000 3 droplets 4 droplets
1000
Scalar Dissipation [1/s]

100
100
10
10
1 1
0.1 0.1
0.01 0.01
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0 0.1 0.2 0.3 0.4 0.5 0.6
Mixture Fraction Mixture Fraction
Fig. 5 Global conditionally averaged scalar dissipation for laminar (left) and turbulent (right)
convective environments
Different regimes are identified when the global computational domain is split
into different areas depending on their position relative to the droplets (as defined
in figure 1). In figure 6, the PDF and the scalar dissipation of the four-droplet case
with turbulent flow are presented. Unsurprisingly, we can observe a clear shift of
the PDF peak towards higher mixture fraction values with downstream distance due
to accumulation of the fuel vapour as the flow passes a succession of droplet layers.
Less obvious are the results for scalar dissipation. Scalar dissipation in region 1 is
higher than in the interdroplet regions (regions 2, 3 and 4) for most of the mixture
fraction range.
16
Region 1 100 Region 1
Region 2 Region 2
14 Region 3 Region 3
Region 4 Region 4
12
10 10
PDF
6
1
4
0 0.1
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4
Mixture Fraction Mixture Fraction
Fig. 6 Mixture fraction and Scalar dissipation in the 4-droplet turbulent case
Figure 7 shows that the stagnant, laminar and turbulent flows have completely
different scalar dissipation in the two-droplet cases. While in the stagnant case the
inner and outer region’s branches are well defined, in the laminar and specially in
the turbulent tests, it is possible to have the same value of scalar dissipation over
a wide range of radial locations. It is clear that the flow affects significantly the
vapour distribution and in terms of scalar dissipation, a clear distinction between
inner and outer region is no longer possible once turbulent mixing occurs and the
radial distance does not seem to give sufficient structural information needed for
accurate modelling.
10000 1000 10
Stagnant - Inner Laminar - Inner Turbulent - Inner
Outer Outer Outer
1000 100
1
100 10
10 1 0.1
1 0.1
0.01
0.1 0.01
0.01 0.001 0.001

0 2 4 6 8 10 12 0 2 4 6 8 10 12 0 2 4 6 8 10 12
Radial distance / Diameter Radial distance / Diameter Radial distance / Diameter
Fig. 7 Scatter plot of global Scalar dissipation in the 2-droplet cases
4 Conclusions
A novel computational method is used to simulate evaporating droplet arrays in lam-
inar and turbulent environments. Mixture fraction PDFs and scalar dissipation are
assessed in order to investigate their dependence on the position relative to dense
droplet layers. As in stagnant environments, different dependencies of PDF and
scalar dissipation can be identified and large inhomogeneities exist for all test cases
within the computational domain. The β -PDF seems to capture the global behaviour
for laminar environment test cases well, but not for higher droplet densities under
stagnant and turbulent conditions. Radial distance to the droplets and droplet density
do not provide sufficient structural information for accurate modelling. The analysis
of a suitable parameterization for the modelling of the PDF and scalar dissipation is
subject to ongoing work at Imperial and Stuttgart.
Acknowledgements. The authors would like to thank the CNPQ/Ministery of Science and
Technology of Brazil for the financial support of this research (Process 200277/2006-5).
References
1. Fedkiw, R., Aslam, T., Merriman, B., Osher, S.: A non-oscillatory eulerian approach
to interfaces in multimaterial flows (the ghost fluid method). J. Comput. Phys. 152,
457–492 (1999)
2. Imaoka, R., Sirignano, W.: Vaporization and combustion in three-dimensional droplet
arrays. Proc. Combustion Institute 30, 1981–1989 (2005)
3. Klimenko, A., Bilger, R.: Conditional moment closure for turbulent combustion.
Progress in Energy and Combustion Science 25, 595–688 (1999)
4. Schroll, P., Wandel, A., Cant, R., Mastorakos, E.: Direct numerical simulations of au-
toignition in turbulent two-phase flows. Proc. Combust. Inst. 32(2), 2275–2282 (2009)
5. Sussman, M., Smereka, P., Osher, S.: A level set approach for computing solutions to
incompressible two-phase flow. J. Comput. Phys. 114, 146–159 (1994)
6. Zoby, M., Navarro-Martinez, S., Kronenburg, A.: Mixture fraction pdf and dissipation in
dense sprays. In: Proc. of the 4th European Combustion Meeting (2009)
Author Index
Accary, G. 143 Emilion, R. 79

Agelinchaab, M. 71 Estivalezes, Jean-Luc 103, 377
Baerenzung, Julien 287 Falconi, C.J. 111

Balaras, E. 25 Farge, Marie 207
Basara, Branislav 63, 247 Fede, Pascal 119, 401
Bejatovic, S. 71 Fröhlich, J. 111
Bernard, D. 79 Frohnapfel, Bettina 127, 191
Bertier, Nicolas 303 Fuchs, L. 1, 345
Bessonov, O. 143
Bockhorn, H. 111 Garnier, E. 135, 263
Bodenschatz, E. 271 Gavrilov, K. 143
Bouffanais, Roland 87, 159 Geurts, B.J. 255
Girimaji, Sharath 63
Glockner, Stéphane 103
Caltagirone, Jean-Paul 95, 377 Godeferd, Fabien S. 151
Cambon, Claude 151
Caruyer, Céline 95 Habisreutinger, Marc A. 159
Chassaing, P. 175 Hasegawa, Yosuke 127
Corre, Cédric 103 Hassel, Egon 231
Courbebaisse, Guy 87
Crespo, M. 175 Iervolino, Michele 167
Isaev, Sergei 231
Debiève, J.-F. 263 Iyer, S.K. 79
Deck, Sébastien 135, 303, 393
Dehbi, A. 295 Jamme, S. 175
Delache, Alexandre 151 Jayaraju, S.T. 183
Delgado, Antonio 191 Jiménez, Javier 3
de Martel, E. 263 Jovanović, Jovan 191
Denev, J.A. 111
Deville, Michel O. 87, 159, 295 Kaczorowski, M. 199
Drikakis, D. 337 Kadoch, Benjamin 207
Dupoirieux, Francis 303 Kasagi, Nobuhide 127
Dupont, P. 263 Kholmyansky, Michael 215
418 Author Index
Kinzel, M. 271 Schäfer, Michael 239

Knacke, Thilo 223 Schmitt, François G. 319
Kornev, Nikolai 231 Schneider, Kai 207
Kornhaas, Michael 239 Schulze, J. 327
Krajnović, Siniša 63, 247 Scotti, A. 25
Kronenburg, A. 409 Sesterhenn, J. 327
Kuerten, J.G.M. 255, 279 Shen, Lian 37
Shimada, Y. 337
Lacor, C. 183 Simonin, Olivier 103, 119, 401
Larchevêque, L. 263 Smits, Alexander J. 51
Leriche, Emmanuel 207, 295 Soldati, Alfredo 295, 311
Lyubimov, D. 143 Soria, M. 369
Sternel, Dörte C. 239
Manna, Marcello 167 Szasz, R.Z. 345
Marchioli, Cristian 311
Marquis, A.J. 409 Tachie, M.F. 71
Meillot, Erick 95 Tardu, Sedat 353
Méradji, S. 143 Thiele, Frank 223
Mininni, Pablo 287 Thornber, B. 337
Moeng, Chin-Hoh 15 Tossa, A. 79
Morvan, D. 143 Touber, Emile 361
Moulinec, C. 385 Trias, F.X. 369
Trontin, P. 377
Navarro, Laurent 87
Tropea, C. 271
Navarro-Martinez, S. 409
Tsinober, Arkady 215
Nobach, H. 271
Turnow, Johann 231
Oliva, A. 369
Uribe, J.C. 385
Paul, S.S. 71
Pecenko, A. 279 Vacca, Andrea 167
Pietarila Graham, Jonathan 287 van der Geld, C.W.M. 255
Politano, Hélène 287 Verbanck, S. 183
Ponty, Yannick 287 Verstappen, R.W.C.P. 369
Pouquet, Annick 287 Villedieu, Philippe 401
Puragliesi, R. 295 Vincent, Stéphane 95, 103, 377
Vinkovic, Ivana 319
Revell, A. 385
Robinet, Jean-Christophe 393 Wagner, C. 199
Rubinat, M. Gasser i 25 Weil, Jeffrey 15
Weiss, Pierre-Élie 393
Sabelnikov, Vladimir 215 Wunsch, Dirk 401
Sagaut, Pierre 393 Wurm, Frank-Hendrik 231
Sainte-Rose, Bruno 303
Salvetti, Maria Vittoria 311 Zimmermann, R. 271
Sandham, Neil D. 361 Zoby, M.R.G. 409
Notes on Numerical Fluid Mechanics and Multidisciplinary Design
Available Volumes
Volume 110: Michel Deville, Thien-Hiep Lê, and Pierre Sagaut (eds.): Turbulence and Interactions –
Proceedings the TI 2009 Conference. ISBN 978-3-642-14138-6
Volume 109: Wolfgang Schröder (ed.): Summary of Flow Modulation and Fluid-Structure Interaction
Findings – Results of the Collaborative Research Center SFB 401 at the RWTH Aachen University,
Aachen, Germany, 1997-2008. ISBN 978-3-642-04087-0
Volume 108: Rudibert King (ed.): Active Flow Control II – Papers Contributed to the Conference “Active
Flow Control II 2010”, Berlin, Germany, May 26–28, 2010. ISBN 978-3-642-11734-3
Volume 107: Norbert Kroll, Dieter Schwamborn, Klaus Becker, Herbert Rieger, Frank Thiele (eds.):
MEGADESIGN and MegaOpt – German Initiatives for Aerodynamic Simulation and Optimization in
Aircraft Design. ISBN 978-3-642-04092-4
Volume 106: Wolfgang Nitsche, Christoph Dobriloff (eds.): Imaging Measurement Methods for Flow
Analysis - Results of the DFG Priority Programme 1147 “Imaging Measurement Methods for Flow
Analysis” 2003–2009. ISBN 978-3-642-01105-4
Volume 105: Michel Deville, Thien-Hiep Lê, Pierre Sagaut (eds.): Turbulence and Interactions - Keynote
Lectures of the TI 2006 Conference. ISBN 978-3-642-00261-8
Volume 104: Christophe Brun, Daniel Juvé, Michael Manhart, Claus-Dieter Munz: Numerical Simula-
tion of Turbulent Flows and Noise Generation - Results of the DFG/CNRS Research Groups FOR 507
and FOR 508. ISBN 978-3-540-89955-6
Volume 103: Werner Haase, Marianna Braza, Alistair Revell (eds.): DESider – A European Effort on
Hybrid RANS-LES Modelling - Results of the European-Union Funded Project, 2004–2007. ISBN 978-
3-540-92772-3
Volume 102: Rolf Radespiel, Cord-Christian Rossow, Benjamin Winfried Brinkmann (eds.):
Hermann Schlichting – 100 Years - Scientific Colloquium Celebrating the Anniversary of His Birthday,
Braunschweig, Germany 2007. ISBN 978-3-540-95997-7
Volume 101: Egon Krause, Yurii I. Shokin, Michael Resch, Nina Shokina (eds.): Computational Sci-
ence and High Performance Computing III - The 3rd Russian-German Advanced Research Workshop,
Novosibirsk, Russia, 23–27 July 2007. ISBN 978-3-540-69008-5
Volume 100: Ernst Heinrich Hirschel, Egon Krause (eds.): 100 Volumes of ’Notes on Numerical Fluid
Mechanics’ - 40 Years of Numerical Fluid Mechanics and Aerodynamics in Retrospect. ISBN 978-3-
540-70804-9
Volume 99: Burkhard Schulte-Werning, David Thompson, Pierre-Etienne Gautier, Carl Hanson, Brian
Hemsworth, James Nelson, Tatsuo Maeda, Paul de Vos (eds.): Noise and Vibration Mitigation for Rail
Transportation Systems - Proceedings of the 9th International Workshop on Railway Noise, Munich,
Germany, 4–8 September 2007. ISBN 978-3-540-74892-2
Volume 98: Ali Gülhan (ed.): RESPACE – Key Technologies for Reusable Space Systems - Results
of a Virtual Institute Programme of the German Helmholtz-Association, 2003–2007. ISBN 978-3-540-
77818-9
Volume 97: Shia-Hui Peng, Werner Haase (eds.): Advances in Hybrid RANS-LES Modelling - Papers
contributed to the 2007 Symposium of Hybrid RANS-LES Methods, Corfu, Greece, 17–18 June 2007.
ISBN 978-3-540-77813-4
Volume 96: C. Tropea, S. Jakirlic, H.-J. Heinemann, R. Henke, H. Hönlinger (eds.): New Results in
Numerical and Experimental Fluid Mechanics VI - Contributions to the 15th STAB/DGLR Symposium
Darmstadt, Germany, 2006. ISBN 978-3-540-74458-0
Volume 95: R. King (ed.): Active Flow Control - Papers contributed to the Conference “Active Flow
Control 2006”, Berlin, Germany, September 27 to 29, 2006. ISBN 978-3-540-71438-5
Volume 94: W. Haase, B. Aupoix, U. Bunge, D. Schwamborn (eds.): FLOMANIA - A European Ini-
tiative on Flow Physics Modelling - Results of the European-Union funded project 2002 - 2004. ISBN
978-3-540-28786-5
Volume 93: Yu. Shokin, M. Resch, N. Danaev, M. Orunkhanov, N. Shokina (eds.): Advances in High
Performance Computing and Computational Sciences - The Ith Khazakh-German Advanced Research
Workshop, Almaty, Kazakhstan, September 25 to October 1, 2005. ISBN 978-3-540-33864-2
Volume 92: H.J. Rath, C. Holze, H.-J. Heinemann, R. Henke, H. Hönlinger (eds.): New Results in Nu-
merical and Experimental Fluid Mechanics V - Contributions to the 14th STAB/DGLR Symposium
Bremen, Germany 2004. ISBN 978-3-540-33286-2
Volume 91: E. Krause, Yu. Shokin, M. Resch, N. Shokina (eds.): Computational Science and High Per-
formance Computing II - The 2nd Russian-German Advanced Research Workshop, Stuttgart, Germany,
March 14 to 16, 2005. ISBN 978-3-540-31767-8
Volume 87: Ch. Breitsamter, B. Laschka, H.-J. Heinemann, R. Hilbig (eds.): New Results in Numerical
and Experimental Fluid Mechanics IV. ISBN 978-3-540-20258-5
Volume 86: S. Wagner, M. Kloker, U. Rist (eds.): Recent Results in Laminar-Turbulent Transition -
Selected numerical and experimental contributions from the DFG priority programme ’Transition’ in
Germany. ISBN 978-3-540-40490-3
Volume 85: N.G. Barton, J. Periaux (eds.): Coupling of Fluids, Structures and Waves in Aeronautics -
Proceedings of a French-Australian Workshop in Melbourne, Australia 3-6 December 2001. ISBN 978-
3-540-40222-0
Volume 83: L. Davidson, D. Cokljat, J. Fröhlich, M.A. Leschziner, C. Mellen, W. Rodi (eds.): LESFOIL:
Large Eddy Simulation of Flow around a High Lift Airfoil - Results of the Project LESFOIL supported
by the European Union 1998 - 2001. ISBN 978-3-540-00533-9
Volume 82: E.H. Hirschel (ed.): Numerical Flow Simulation III - CNRS-DFG Collaborative Research
Programme, Results 2000-2002. ISBN 978-3-540-44130-4
Volume 81: W. Haase, V. Selmin, B. Winzell (eds.): Progress in Computational Flow Structure Inter-
action - Results of the Project UNSI, supported by the European Union 1998-2000. ISBN 978-3-540-
43902-8
Volume 80: E. Stanewsky, J. Delery, J. Fulker, P. de Matteis (eds.): Drag Reduction by Shock and
Boundary Layer Control - Results of the Project EUROSHOCK II, supported by the European Union
1996-1999. ISBN 978-3-540-43317-0
Volume 79: B. Schulte-Werning, R. Gregoire, A. Malfatti, G. Matschke (eds.): TRANSAERO - A Euro-

pean Initiative on Transient Aerodynamics for Railway System Optimisation. ISBN 978-3-540-43316-3
Volume 78: M. Hafez, K. Morinishi, J. Periaux (eds.): Computational Fluid Dynamics for the 21st Cen-
tury. Proceedings of a Symposium Honoring Prof. Satofuka on the Occasion of his 60th Birthday, Kyoto,
Japan, 15-17 July 2000. ISBN 978-3-540-42053-8
Volume 77: S. Wagner, U. Rist, H.-J. Heinemann, R. Hilbig (eds.): New Results in Numerical and Ex-
perimental Fluid Mechanics III. Contributions to the 12th STAB/DGLR Symposium, Stuttgart, Germany
2000. ISBN 978-3-540-42696-7

Turbulence Interactions

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Turbulence Interactions

Caricato da

Copyright:

Formati disponibili

110 Notes on Numerical Fluid Mechanics and Multidisciplinary Design (NNFM)

ISBN 978-3-642-14138-6 e-ISBN 978-3-642-14139-3

Notes on Numerical Fluid Mechanics

Library of Congress Control Number: 2010929478

Printed on acid-free paper

Prof. Dr. Wolfgang Schröder Prof. Dr. Sergio Pirozzoli

Prof. Dr. Michael A. Leschziner

Lausanne, Paris M. O. Deville

A Model for Dissipation: Cascade SDE with Markov

Eﬀect of Particle-Particle Collisions on the Spatial

Pulsating Flow through Porous Media . . . . . . . . . . . . . . . . . . . . . . . 167

Experimental and Numerical Study of Unsteadiness in

Large-Eddy Simulations of an Oblique Shock Impinging

Numerical Simulation and Statistical Modeling of

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417

2 The General Organisation of Wall-Bounded Turbulence

2.1 The Buffer Layer

to sinuous perturbations associated with inflection points of the distorted velocity

2.2 The Logarithmic Layer

3 The Direction of Causality

Chin-Hoh Moeng and Jeffrey Weil

Abstract. This article reviews the planetary-boundary-layer (PBL) turbulence and

bear more resemblance to two-dimensional turbulence than three-dimensional (3D)

2 PBL Turbulence and Its Interactions

2.1 Interaction with Land Processes and Diurnal Cycle

consumption, and turbulence is re-generated. Turbulence in the very stable PBL is

2.2 Interaction with Ocean Waves

2.3 Interaction with Radiation and Cloud Microphysics

3 Turbulence in Complex, Real-World Environments

Acknowledgements. The National Center for Atmospheric Research is sponsored by the

A. Scotti, M. Gasser i Rubinat, and E. Balaras

Abstract. Flows driven by a pressure gradient that oscillates periodically around a

streamlines[3]. However, no formal stability calculation has been performed under

viscosity over horizontal planes to remove numerically unstable negative values.

dimensional analysis argument, well supported by the numerical experiments [18],

3.2 Wavy Wall

8 (a) (b) (c)

10−1 101 103 10−1 101 103 10−1 101 103

Orbital-ripple cases. While suborbital flows are characterized by an enhanced

3.3 Temporal Evolution of the Near Wall Momentum Flux

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

a/λ=0.27 a/λ=0.55 a/λ=0.82 a/λ=1.66 a/λ=3.31 a/λ=6.62

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 5 Temporal evolution of the Reynolds Stress u

Acknowledgements. This work was supported by NSF, grant OCE-0351667.

Abstract. We develop numerical capabilities of direct numerical simulation and

with nonlinear wave simulation. Mechanistic study is performed in order to obtain

∂ (uv) Hy − ζηy ∂ (uv) 1 ∂ (uw)

∂ (vv) Hy − ζηy ∂ (vv) 1 ∂ (vw)

∂ (wv) Hy − ζHy ∂ (wv) 1 ∂ (ww)

ηt = w − uηx − vηy , (10)

Here P = p − ρ gη, where g is the gravitational acceleration. The subscripts

approach based on moving grid fitted to the wave boundary is complemented by

3 Mechanistic Study of Turbulence over Plane Progressive

Fig. 2 3D flow field illustration of turbulence over a monochromatic surface wave.

0.5 0.5 0.5

Fig. 4 Contours of normalized Reynolds stress −u

dash-dot-dot line. In the wave-following frame, the critical layer is surrounded by

Fig. 5 Snapshot of near-surface coherent vortical structures in instantaneous turbulence field

Fig. 7 Conditionally-averaged near-surface coherent vortical structures, represented by

vortex (2) is a quasi-streamwise vortex with negative ωx . As shown in the figure,

4 Interaction of Nonlinear Wave Evolution with Turbulence

and an eigenfunction expansion of each Φ(m) with N modes,

where ψi et D i are respectively the concentration and the diffusion coefficient of