Large Eddy Simulation For Incompressible Flows

Scientic Computation
Editorial Board
J.-J. Chattot, Davis, CA, USA
P. Colella, Berkeley, CA, USA
Weinan E, Princeton, NJ, USA
R. Glowinski, Houston, TX, USA
M. Holt, Berkeley, CA, USA
Y. Hussaini, Tallahassee, FL, USA
P. Joly, Le Chesnay, France
H. B. Keller, Pasadena, CA, USA
D. I. Meiron, Pasadena, CA, USA
O. Pironneau, Paris, France
A. Quarteroni, Lausanne, Switzerland
J. Rappaz, Lausanne, Switzerland
R. Rosner, Chicago, IL, USA.
J. H. Seinfeld, Pasadena, CA, USA
A. Szepessy, Stockholm, Sweden
M. F. Wheeler, Austin, TX, USA
Pierre Sagaut
Large Eddy Simulation

for Incompressible Flows
An Introduction
Third Edition
With a Foreword by Massimo Germano
With 99 Figures and 15 Tables
123
Prof. Dr. Pierre Sagaut

LMM-UPMC/CNRS
Boite 162, 4 place Jussieu
75252 Paris Cedex 05, France
[email protected]
Title of the original French edition:

Introduction la simulation des grandes chelles pour les coulements de uide incompressible,
Mathmatique & Applications.
Springer Berlin Heidelberg 1998
Library of Congress Control Number: 2005930493
ISSN 1434-8322
ISBN-10 3-540-26344-6 Third Edition Springer Berlin Heidelberg New York
ISBN-13 978-3-540-26344-9 Third Edition Springer Berlin Heidelberg New York
ISBN 3-540-67841-7 Second Edition Springer-Verlag Berlin Heidelberg New York
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.
Springer is a part of Springer Science+Business Media
springeronline.com
Springer-Verlag Berlin Heidelberg 2001, 2002, 2006
Printed in Germany
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.
Typesetting: Data conversion by LE-TEX Jelonek, Schmidt & Vckler GbR, Leipzig, Germany
Cover design: design & production GmbH, Heidelberg
Printed on acid-free paper
55/3141/YL
543210
Foreword to the Third Edition
It is with a sense of great satisfaction that I write these lines introducing the
third edition of Pierre Sagauts account of the eld of Large Eddy Simulation
for Incompressible Flows. Large Eddy Simulation has evolved into a powerful
tool of central importance in the study of turbulence, and this meticulously
assembled and signicantly enlarged description of the many aspects of LES
will be a most welcome addition to the bookshelves of scientists and engineers
in uid mechanics, LES practitioners, and students of turbulence in general.
Hydrodynamic turbulence continues to be a fundamental challenge for
scientists striving to understand uid motions in elds as diverse as oceanography, acoustics, meteorology and astrophysics. The challenge also has socioeconomic attributes as engineers aim at predicting ows to control their features, and to improve thermo-uid equipment design. Drag reduction in external aerodynamics or convective heat transfer augmentation are well-known
examples. The fundamental challenges posed by turbulence to scientists and
engineers have not, in essence, changed since the appearance of the second
edition of this book, a mere two years ago. What has evolved signicantly
is the eld of Large Eddy Simulation (LES), including methods developed
to address the closure problem associated with LES (also called the problem
of subgrid-scale modeling), numerical techniques for particular applications,
and more explicit accounts of the interplay between numerical techniques and
subgrid modeling.
The original hope for LES was that simple closures would be appropriate, such as mixing length models with a single, universally applicable model
parameter. Kolmogorovs phenomenological theory of turbulence in fact supports this hope but only if the length-scale associated with the numerical
resolution of LES falls well within the ideal inertial range of turbulence, in
ows at very high Reynolds numbers. Typical applications of LES most often violate this requirement and the resolution length-scale is often close to
some externally imposed scale of physical relevance, leading to loss of universality and the need for more advanced, and often much more complex,
closure models. Fortunately, the LES modeler disposes of large amount of
raw materials from which to assemble improved models. During LES, the resolved motions present rich multi-scale elds and dynamics including highly
non-trivial nonlinear interactions which can be interrogated to learn about
VI
the local state of turbulence. This availability of dynamical information has

led to the formulation of a continuously growing number of dierent closure
models and methodologies and associated numerical approaches, including
many variations on several basic themes. In consequence, the literature on
LES has increased signicantly in recent years. Just to mention a quantitative
measure of this trend, in 2000 the ISI science citation index listed 164 papers
published including the keywords large-eddy-simulation during that year.
By 2004 this number had doubled to over 320 per year. It is clear, then, that
a signicantly enlarged version of Sagauts book, encompassing much of what
has been added to the literature since the books second edition, is a most
welcome contribution to the eld.
What are the main aspects in which this third edition has been enlarged
compared to the rst two? Sagaut has added signicantly new material in
a number of areas. To begin, the introductory chapter is enriched with an
overview of the structure of the book, including an illuminating description of
three fundamental errors one incurs when attempting to solve uid mechanics innite-dimensional, non-linear dierential equations, namely projection
error, discretization error, and in the case of turbulence and LES, the physically very important resolution error. Following the chapters describing in
signicant detail the relevant foundational aspects of ltering in LES, Sagaut
has added a new section dealing with alternative mathematical formulations
of LES. These include statistical approaches that replace spatial ltering with
conditionally averaging the unresolved motions, and alternative model equations in which the Navier-Stokes equations are replaced with mathematically
better behaved equations such as the Leray model in which the advection
velocity is regularized (i.e. ltered).
In the chapter dealing with functional modeling approaches, in which the
subgrid-scale stresses are expressed in terms of local functionals of the resolved velocity gradients, a more complete account of the various versions of
the dynamic model is given, as well as extended discussions of new structurefunction and multiscale models. The chapter on structural modeling, in which
the stress tensor is reconstructed based on its denition and various direct
hypotheses about the small-scale velocity eld is signicantly enhanced: Closures in which full prognostic transport equations are solved for the subgridscale stress tensor are reviewed in detail, and entire new subsections have been
added dealing with ltered density function models, with one-dimensional
turbulence mapping models, and variational multi-scale models, among others. The chapter focussing on numerical techniques contains an interesting
new description of the eects of pre-ltering and of the various methods to
perform grid renement. In the chapter on analysis and validation of LES,
a new detailed account is given about methods to evaluate the subgrid-scale
kinetic energy. The description of boundary and inow conditions for LES is
enhanced with new material dealing with one-dimensional-turbulence models
near walls as well as stochastic tools to generate and modulate random elds
VII
for inlet turbulence specication. Chapters dealing with coupling of multiresolution, multidomain, and adaptive grid renement techniques, as well as
LES - RANS coupling, have been extended to include recent additions to the
literature. Among others, these are areas to which Sagaut and his co-workers
have made signicant research contributions.
The most notable additions are two entirely new chapters at the end of
the book, on the prediction of scalars using LES. Both passive scalars, for
which subgrid-scale mixing is an important issue, and active scalars, of great
importance to geophysical ows, are treated. The geophysics literature on
LES of stably and unstably stratied ows is voluminous - the eld of LES
in fact traces its origins to simulating atmospheric boundary layer ows in
the early 1970s. Sagaut summarizes this vast eld using his classications of
subgrid closures introduced earlier, and the result is a conceptually elegant
and concise treatment, which will be of signicant interest to both engineering
and geophysics practitioners of LES.
The connection to geophysical ow prediction reminds us of the importance of LES and subgrid modeling from a broader viewpoint. For the eld of
large-scale numerical simulation of complex multiscale nonlinear systems is,
today, at the center of scientic discussions with important societal and political dimensions. This is most visible in the discussions surrounding the trustworthiness of global change models. Among others, these include boundarylayer parameterizations that can be studied by means of LES done at smaller
scales. And LES of turbulence is itself a prime example of large-scale computing applied to prediction of a multi-scale complex system, including issues
surrounding the verication of its predictive capabilities, the testing of the
cumulative accuracy of individual building blocks, and interesting issues on
the interplay of stochastic and deterministic aspects of the problem. Thus
the book - as well as its subject - Large Eddy Simulation of Incompressible
Flow, has much to oer to one of the most pressing issues of our times.
With this latest edition, Pierre Sagaut has fully solidied his position as
the preeminent cartographer of the complex and multifaceted world of LES.
By mapping out the eld in meticulous fashion, Sagauts work can indeed be
regarded as a detailed and evolving atlas of the world of LES. And yet, it is not
a tourist guide: as with any relatively young terrain in which the main routes
have not yet been rmly established, what is called for is unbiased, objective,
and sophisticated cartography. The cartographer describes the topography,
scenery, and landmarks as they appear, without attempting to preach to the
traveler which route is best. In return, the traveler is expected to bring along
a certain sophistication to interpret the maps and to discern which among the
many paths will most likely lead towards particular destinations of interest.
The reader of this latest edition will thus be rewarded with a most solid, insightful, and up-to-date account of an important and exciting eld of research.
Baltimore, January 2005
Charles Meneveau
Foreword to the Second Edition
It is a particular pleasure to present the second edition of the book on Large

Eddy Simulation for Incompressible Flows written by Pierre Sagaut: two editions in two years means that the interest in the topic is strong and that
a book on it was indeed required. Compared to the rst one, this second
edition is a greatly enriched version, motivated both by the increasing theoretical interest in Large Eddy Simulation (LES) and the increasing numbers
of applications and practical issues. A typical one is the need to decrease
the computational cost, and this has motivated two entirely new chapters
devoted to the coupling of LES with multiresolution multidomain techniques
and to the new hybrid approaches that relate the LES procedures to the
classical statistical methods based on the Reynolds Averaged NavierStokes
equations.
Not that literature on LES is scarce. There are many article reviews and
conference proceedings on it, but the book by Sagaut is the rst that organizes a topic that by its peculiar nature is at the crossroads of various interests
and techniques: rst of all the physics of turbulence and its dierent levels of
description, then the computational aspects, and nally the applications that
involve a lot of dierent technical elds. All that has produced, particularly
during the last decade, an enormous number of publications scattered over
scientic journals, technical notes, and symposium acta, and to select and
classify with a systematic approach all this material is a real challenge. Also,
by assuming, as the writer does, that the reader has a basic knowledge of
uid mechanics and applied mathematics, it is clear that to introduce the
procedures presently adopted in the large eddy simulation of turbulent ows
is a dicult task in itself. First of all, there is no accepted universal denition
of what LES really is. It seems that LES covers everything that lies between
RANS, the classical statistical picture of turbulence based on the Reynolds
Averaged NavierStokes equations, and DNS, the Direct Numerical Simulations resolved in all details, but till now there has not been a general unied
theory that gradually goes from one description to the other. Moreover we
should note the dierent importance that the practitioners of LES attribute
to the numerical and the modeling aspects. At one end the supporters of
the no model way of thinking argue that the numerical scheme should and
could capture by itself the resolved scales. At the other end the theoretical
modelers try to develop new universal equations for the ltered quantities.
In some cases LES is regarded as a technique imposed by the present provisional inability of the computers to solve all the details. Others think that
LES modeling is a contribution to the understanding of turbulence and the
interactions among dierent ideas are often poor.
Pierre Sagaut has elaborated on this immense material with an open mind
and in an exceptionally clear way. After three chapters devoted to the basic
problem of the scale separation and its application to the NavierStokes equations, he classies the various subgrid models presently in use as functional
and structural ones. The chapters devoted to this general review are of the
utmost interest: obviously some selection has been done, but both the student and the professional engineer will nd there a clear unbiased exposition.
After this rst part devoted to the fundamentals a second part covers many
of the interdisciplinary problems created by the practical use of LES and
its coupling with the numerical techniques. These subjects, very important
obviously from the practical point of view, are also very rich in theoretical
aspects, and one great merit of Sagaut is that he presents them always in
an attractive way without reducing the exposition to a mere set of instructions. The interpretation of the numerical solutions, the validation and the
comparison of LES databases, the general problem of the boundary conditions are mathematically, physically and numerically analyzed in great detail,
with a principal interest in the general aspects. Two entirely new chapters
are devoted to the coupling of LES with multidomain techniques, a topic in
which Pierre Sagaut and his group have made important contributions, and
to the new hybrid approaches RANS/LES, and nally in the last expanded
chapter, enriched by new examples and beautiful gures, we have a review of
the dierent applications of LES in the nuclear, aeronautical, chemical and
automotive elds.
Both for graduate students and for scientists this book is a very important reference. People involved in the large eddy simulation of turbulent ows
will nd a useful introduction to the topic and a complete and systematic
overview of the many dierent modeling procedures. At present their number
is very high and in the last chapter the author tries to draw some conclusions
concerning their eciency, but probably the person who is only interested
in the basic question What is the best model for LES? will remain a little disappointed. As remarked by the author, both the structural and the
functional models have their advantages and disadvantages that make them
seem complementary, and probably a mixed modeling procedure will be in
the future a good compromise. But for a textbook this is not the main point.
The fortunes and the misfortunes of a model are not so simple to predict,
and its success is in many cases due to many particular reasons. The results
are obviously the most important test, but they also have to be considered
in a textbook with a certain reserve, in the higher interest of a presentation
that tries as much as possible to be not only systematic but also rational.
XI
To write a textbook obliges one in some way or another to make judgements,

and to transmit ideas, sometimes hidden in procedures that for some reason or another have not till now received interest from the various groups
involved in LES and have not been explored in full detail.
Pierre Sagaut has succeeded exceptionally well in doing that. One reason
for the success is that the author is curious about every detail. The nal task
is obviously to provide a good and systematic introduction to the beginner,
as rational as a book devoted to turbulence can be, and to provide useful
information for the specialist. The research has, however, its peculiarities,
and this book is unambiguously written by a passionate researcher, disposed
to explore every problem, to search in all models and in all proposals the
germs of new potentially useful ideas. The LES procedures that mix theoretical modeling and numerical computation are often, in an inextricable way,
exceptionally rich in complex problems. What about the problem of the mesh
adaptation on unstructured grids for large eddy simulations? Or the problem of the comparison of the LES results with reference data? Practice shows
that nearly all authors make comparisons with reference data or analyze large
eddy simulation data with no processing of the data .... Pierre Sagaut has the
courage to dive deep into procedures that are sometimes very dicult to explore, with the enthusiasm of a genuine researcher interested in all aspects
and condent about every contribution. This book now in its second edition
seems really destined for a solid and durable success. Not that every aspect
of LES is covered: the rapid progress of LES in compressible and reacting
ows will shortly, we hope, motivate further additions. Other developments
will probably justify new sections. What seems, however, more important is
that the basic style of this book is exceptionally valid and open to the future
of a young, rapidly evolving discipline. This book is not an encyclopedia and
it is not simply a monograph, it provides a framework that can be used as
a text of lectures or can be used as a detailed and accurate review of modeling procedures. The references, now increased in number to nearly 500, are
given not only to extend but largely to support the material presented, and
in some cases the dialogue goes beyond the original paper. As such, the book
is recommended as a fundamental work for people interested in LES: the
graduate and postgraduate students will nd an immense number of stimulating issues, and the specialists, researchers and engineers involved in the
more and more numerous elds of application of LES will nd a reasoned and
systematic handbook of dierent procedures. Last, but not least, the applied
mathematician can nally enjoy considering the richness of challenging and
attractive problems proposed as a result of the interaction among dierent
topics.
Torino, April 2002
Massimo Germano
Foreword to the First Edition
Still today, turbulence in uids is considered as one of the most dicult

problems of modern physics. Yet we are quite far from the complexity of
microscopic molecular physics, since we only deal with Newtonian mechanics
laws applied to a continuum, in which the eect of molecular uctuations
has been smoothed out and is represented by molecular-viscosity coecients.
Such a system has a dual behaviour of determinism in the Laplacian sense,
and extreme sensitivity to initial conditions because of its very strong nonlinear character. One does not know, for instance, how to predict the critical
Reynolds number of transition to turbulence in a pipe, nor how to compute
precisely the drag of a car or an aircraft, even with todays largest computers.
We know, since the meteorologist Richardson,1 numerical schemes allowing us to solve in a deterministic manner the equations of motion, starting
with a given initial state and with prescribed boundary conditions. They
are based on momentum and energy balances. However, such a resolution
requires formidable computing power, and is only possible for low Reynolds
numbers. These Direct-Numerical Simulations may involve calculating the
interaction of several million interacting sites. Generally, industrial, natural, or experimental congurations involve Reynolds numbers that are far
too large to allow direct simulations,2 and the only possibility then is Large
Eddy Simulations, where the small-scale turbulent uctuations are themselves smoothed out and modelled via eddy-viscosity and diusivity assumptions. The history of large eddy simulations began in the 1960s with the
famous Smagorinsky model. Smagorinsky, also a meteorologist, wanted to
represent the eects upon large synoptic quasi-two-dimensional atmospheric
or oceanic motions3 of a three-dimensional subgrid turbulence cascading toward small scales according to mechanisms described by Richardson in 1926
and formalized by the famous mathematician Kolmogorov in 1941.4 It is interesting to note that Smagorinskys model was a total failure as far as the
1
2
3
4
L.F. Richardson, Weather Prediction by Numerical Process, Cambridge University Press (1922).
More than 1015 modes should be necessary for a supersonic-plane wing!
Subject to vigorous inverse-energy cascades.
L.F. Richardson, Proc. Roy. Soc. London, Ser A, 110, pp. 709737 (1926); A. Kolmogorov, Dokl. Akad. Nauk SSSR, 30, pp. 301305 (1941).
XIV
atmosphere and oceans are concerned, because it dissipates the large-scale

motions too much. It was an immense success, though, with users interested
in industrial-ow applications, which shows that the outcomes of research
are as unpredictable as turbulence itself! A little later, in the 1970s, the theoretical physicist Kraichnan5 developed the important concept of spectral
eddy viscosity, which allows us to go beyond the separation-scale assumption
inherent in the typical eddy-viscosity concept of Smagorinsky. From then
on, the history of large eddy simulations developed, rst in the wake of two
schools: StanfordTorino, where a dynamic version of Smagorinskys model
was developed; and Grenoble, which followed Kraichnans footsteps. Then
researchers, including industrial researchers, all around the world became infatuated with these techniques, being aware of the limits of classical modeling
methods based on the averaged equations of motion (Reynolds equations).
It is a complete account of this young but very rich discipline, the large
eddy simulation of turbulence, which is proposed to us by the young ONERA researcher Pierre Sagaut, in a book whose reading brings pleasure and
interest. Large-Eddy Simulation for Incompressible Flows - An Introduction
very wisely limits itself to the case of incompressible uids, which is a suitable starting point if one wants to avoid multiplying diculties. Let us point
out, however, that compressible ows quite often exhibit near-incompressible
properties in boundary layers, once the variation of the molecular viscosity
with the temperature has been taken into account, as predicted by Morkovin
in his famous hypothesis.6 Pierre Sagaut shows an impressive culture, describing exhaustively all the subgrid-modeling methods for simulating the
large scales of turbulence, without hesitating to give the mathematical details needed for a proper understanding of the subject.
After a general introduction, he presents and discusses the various lters
used, in cases of statistically homogeneous and inhomogeneous turbulence,
and their applications to NavierStokes equations. He very aptly describes
the representation of various tensors in Fourier space, Germano-type relations
obtained by double ltering, and the consequences of Galilean invariance of
the equations. He then goes into the various ways of modeling isotropic turbulence. This is done rst in Fourier space, with the essential wave-vector
triad idea, and a discussion of the transfer-localness concept. An excellent
review of spectral-viscosity models is provided, with developments going beyond the original papers. Then he goes to physical space, with a discussion of
the structure-function models and the dynamic procedures (Eulerian and Lagrangian, with energy equations and so forth). The study is then generalized
to the anisotropic case. Finally, functional approaches based on Taylor series expansions are discussed, along with non-linear models, homogenization
techniques, and simple and dynamic mixed models.
5
6
He worked as a postdoctoral student with Einstein at Princeton.

M.V. Morkovin, in Mecanique de la Turbulence, A. Favre et al. (eds.), CNRS,
pp. 367380 (1962).
XV
Pierre Sagaut also discusses the importance of numerical errors, and proposes a very interesting review of the dierent wall models in the boundary
layers. The last chapter gives a few examples of applications carried out at
ONERA and a few other French laboratories. These examples are well chosen
in order of increasing complexity: isotropic turbulence, with the non-linear
condensation of vorticity into the worms vortices discovered by Siggia;7
planar Poiseuille ow with ejection of hairpin vortices above low-speed
streaks; the round jet and its alternate pairing of vortex rings; and, nally,
the backward-facing step, the unavoidable test case of computational uid
dynamics. Also on the menu: beautiful visualizations of separation behind
a wing at high incidence, with the shedding of superb longitudinal vortices.
Completing the work are two appendices on the statistical and spectral analysis of turbulence, as well as isotropic and anisotropic EDQNM modeling.
A bold explorer, Pierre Sagaut had the daring to plunge into the jungle
of multiple modern techniques of large-scale simulation of turbulence. He
came back from his trek with an extremely complete synthesis of all the
models, giving us a very complete handbook that novices can use to start o
on this enthralling adventure, while specialists can discover models dierent
from those they use every day. Large-Eddy Simulation for Incompressible
Flows - An Introduction is a thrilling work in a somewhat austere wrapping.
I very warmly recommend it to the broad public of postgraduate students,
researchers, and engineers interested in uid mechanics and its applications
in numerous elds such as aerodynamics, combustion, energetics, and the
environment.
Grenoble, March 2000
E.D. Siggia, J. Fluid Mech., 107, pp. 375406 (1981).
Marcel Lesieur
Preface to the Third Edition
Working on the manuscript of the third edition of this book was a very
exciting task, since a lot of new developments have been published since the
second edition was printed.
The large-eddy simulation (LES) technique is now recognized as a powerful tool and real applications in several engineering elds are more and more
frequently found. This increasing demand for ecient LES tools also sustains
growing theoretical research on many aspects of LES, some of which are included in this book. Among them, it is worth noting the mathematical models of LES (the convolution lter being only one possiblity), the denition of
boundary conditions, the coupling with numerical errors, and, of course, the
problem of dening adequate subgrid models. All these issues are discussed
in more detail in this new edition. Some good news is that other monographs,
which are good complements to the present book, are now available, showing
that LES is a topic with a fastly growing audience. The reader interested in
mathematics-oriented discussions will nd many details in the monoghaphs
by Volker John (Large-Eddy Simulation of Turbulent Incompressible Flows,
Springer) and Berselli, Illiescu and Layton (Mathematics of Large-Eddy Simulation of Turbulent Flows, Springer), while people looking for a subsequent
description of numerical methods for LES and direct numerical simulation
will enjoy the book by Bernard Geurts (Elements of Direct and Large-Eddy
Simulation, Edwards). More monographs devoted to particular features of
LES (implicit LES appraoches, mathematical backgrounds, etc.) are to come
in the near future.
My purpose while writing this third edition was still to provide the reader
with an up-to-date review of existing methods, approaches and models for
LES of incompressible ows. All chapters of the previous edition have been
updated, with the hope that this nearly exhaustive review will help interested
readers avoid rediscovering old things. I would like to apologize in advance for
certainly forgetting some developments. Two entirely new chapters have been
added. The rst one deals with mathematical models for LES. Here, I believe
that the interesting point is that the ltering approach is nothing but a model
for the true LES problem, and other models have been developed that seem
to be at least as promising as this very popular one. The second new chapter
is dedicated to the scalar equation, with both passive scalar and active scalar
XVIII Preface to the Third Edition
(stable/unstable stratication eects) cases being discussed. This extension

illustrates the way the usual LES can be extended and how new physical
mechanisms can be dealt with, but also inspires new problems.
Paris, November 2004
Pierre Sagaut
Preface to the Second Edition
The astonishingly rapid development of the Large-Eddy Simulation technique

during the last two or three years, both from the theoretical and applied
points of view, have rendered the rst edition of this book lacunary in some
ways. Three to four years ago, when I was working on the manuscript of the
rst edition, coupling between LES and multiresolution/multilevel techniques
was just an emerging idea. Nowadays, several applications of this approach
have been succesfully developed and applied to several ow congurations.
Another example of interest from this exponentially growing eld is the development of hybrid RANS/LES approaches, which have been derived under
many dierent forms. Because these topics are promising and seem to be
possible ways of enhancing the applicability of LES, I felt that they should
be incorporated in a general presentation of LES.
Recent developments in LES theory also deal with older topics which have
been intensely revisited by reseachers: a unied theory for deconvolution and
scale similarity ways of modeling have now been established; the no model
approach, popularized as the MILES approach, is now based on a deeper
theoretical analysis; a lot of attention has been paid to the problem of the
denition of boundary conditions for LES; ltering has been extended to
NavierStokes equations in general coordinates and to Eulerian timedomain
ltering.
Another important fact is that LES is now used as an engineering tool
for several types of applications, mainly dealing with massively separated
ows in complex congurations. The growing need for unsteady, accurate
simulations, more and more associated with multidisciplinary applications
such as aeroacoustics, is a very powerful driver for LES, and it is certain that
this technique is of great promise.
For all these reasons, I accepted the opportunity to revise and to augment
this book when Springer oered it me. I would also like to emphasize the fruitful interactions between traditional LES researchers and mathematicians
that have very recently been developed, yielding, for example, a better understanding of the problem of boundary conditions. Mathematical foundations
for LES are under development, and will not be presented in this book, because I did not want to include specialized functional analysis discussions in
the present framework.
XX
Preface to the Second Edition
I am indebted to an increasing number of people, but I would like to

express special thanks to all my colleagues at ONERA who worked with me
on LES: Drs. E. Garnier, E. Labourasse, I. Mary, P. Quemere and M. Terracol.
All the people who provided me with material dealing with their research
are also warmly acknowledged. I also would like to thank all the readers
of the rst edition of this book who very kindly provided me with their
remarks, comments and suggestions. Mrs. J. Ryan is once again gratefully
acknowledged for her help in writing the English version.
Paris, April 2002
Pierre Sagaut
Preface to the First Edition
While giving lectures dealing with Large-Eddy Simulation (LES) to students

or senior scientists, I have found diculties indicating published references
which can serve as general and complete introductions to this technique.
I have tried therefore to write a textbook which can be used by students
or researchers showing theoretical and practical aspects of the Large Eddy
Simulation technique, with the purpose of presenting the main theoretical
problems and ways of modeling. It assumes that the reader possesses a basic
knowledge of uid mechanics and applied mathematics.
Introducing Large Eddy Simulation is not an easy task, since no unied
and universally accepted theoretical framework exists for it. It should be
remembered that the rst LES computations were carried out in the early
1960s, but the rst rigorous derivation of the LES governing equations in
general coordinates was published in 1995! Many reasons can be invoked to
explain this lack of a unied framework. Among them, the fact that LES
stands at the crossroads of physical modeling and numerical analysis is a major point, and only a few really successful interactions between physicists,
mathematicians and practitioners have been registered over the past thirty
years, each community sticking to its own language and center of interest.
Each of these three communities, though producing very interesting work,
has not yet provided a complete theoretical framework for LES by its own
means. I have tried to gather these dierent contributions in this book, in
an understandable form for readers having a basic background in applied
mathematics.
Another diculty is the very large number of existing physical models,
referred to as subgrid models. Most of them are only used by their creators,
and appear in a very small number of publications. I made the choice to
present a very large number of models, in order to give the reader a good
overview of the ways explored. The distinction between functional and structural models is made in this book, in order to provide a general classication;
this was necessary to produce an integrated presentation.
In order to provide a useful synthesis of forty years of LES development,
I had to make several choices. Firstly, the subject is restricted to incompressible ows, as the theoretical background for compressible ow is less
evolved. Secondly, it was necessary to make a unied presentation of a large
XXII
Preface to the First Edition
number of works issued from many research groups, and very often I have
had to change the original proof and to reduce it. I hope that the authors
will not feel betrayed by the present work. Thirdly, several thousand journal articles and communications dealing with LES can be found, and I had
to make a selection. I have deliberately chosen to present a large number
of theoretical approaches and physical models to give the reader the most
general view of what has been done in each eld. I think that the most important contributions are presented in this book, but I am sure that many
new physical models and results dealing with theoretical aspects will appear
in the near future.
A typical question of people who are discovering LES is what is the best
model for LES?. I have to say that I am convinced that this question cannot
be answered nowadays, because no extensive comparisons have been carried
out, and I am not even sure that the answer exists, because people do not
agree on the criterion to use to dene the best model. As a consequence,
I did not try to rank the model, but gave very generally agreed conclusions
on the model eciency.
A very important point when dealing with LES is the numerical algorithm
used to solve the governing equations. It has always been recognized that
numerical errors could aect the quality of the solution, but new emphasis
has been put on this subject during the last decade, and it seems that things
are just beginning. This point appeared as a real problem to me when writing
this book, because many conclusions are still controversial (e.g. the possibility
of using a second-order accurate numerical scheme or an articial diusion).
So I chose to mention the problems and the dierent existing points of view,
but avoided writing a part dealing entirely with numerical discretization and
time integration, discretization errors, etc. This would have required writing
a companion book on numerical methods, and that was beyond the scope of
the present work. Many good textbooks on that subject already exist, and
the reader should refer to them.
Another point is that the analysis of the coupling of LES with typical
numerical techniques, which should greatly increase the range of applications,
such as Arbitrary LagrangianEulerian methods, Adaptive Mesh-Renement
or embedded grid techniques, is still to be developed.
I am indebted to a large number of people, but I would like to express
special thanks to Dr. P. Le Quere, O. Daube, who gave me the opportunity to
write my rst manuscript on LES, and to Prof. J.M. Ghidaglia who oered me
the possibility of publishing the rst version of this book (in French). I would
also like to thank ONERA for helping me to write this new, augmented and
translated version of the book. Mrs. J. Ryan is gratefully acknowledged for
her help in writing the English version.
Paris, September 2000
Pierre Sagaut
Contents
1.
2.
3.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Computational Fluid Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Levels of Approximation: General . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Statement of the Scale Separation Problem . . . . . . . . . . . . . . . .
1.4 Usual Levels of Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Large-Eddy Simulation: from Practice to Theory.
Structure of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Formal Introduction to Scale Separation:
Band-Pass Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1 Denition and Properties of the Filter
in the Homogeneous Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1 Denition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.2 Fundamental Properties . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.3 Characterization of Dierent Approximations . . . . . . . .
2.1.4 Dierential Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.5 Three Classical Filters for Large-Eddy Simulation . . . .
2.1.6 Dierential Interpretation of the Filters . . . . . . . . . . . . .
2.2 Spatial Filtering: Extension to the Inhomogeneous Case . . . . .
2.2.1 General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.2 Non-uniform Filtering Over an Arbitrary Domain . . . .
2.2.3 Local Spectrum of Commutation Error . . . . . . . . . . . . . .
2.3 Time Filtering: a Few Properties . . . . . . . . . . . . . . . . . . . . . . . . .
Application to NavierStokes Equations . . . . . . . . . . . . . . . . . .
3.1 NavierStokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1 Formulation in Physical Space . . . . . . . . . . . . . . . . . . . . .
3.1.2 Formulation in General Coordinates . . . . . . . . . . . . . . . .
3.1.3 Formulation in Spectral Space . . . . . . . . . . . . . . . . . . . . .
3.2 Filtered NavierStokes Equations in Cartesian Coordinates
(Homogeneous Case) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1 Formulation in Physical Space . . . . . . . . . . . . . . . . . . . . .
3.2.2 Formulation in Spectral Space . . . . . . . . . . . . . . . . . . . . .
1
1
2
3
5
9
15
15
15
17
18
20
21
26
31
31
32
42
43
45
46
46
46
47
48
48
48
XXIV
Contents
3.3 Decomposition of the Non-linear Term.

Associated Equations for the Conventional Approach . . . . . . .
3.3.1 Leonards Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.2 Germano Consistent Decomposition . . . . . . . . . . . . . . . .
3.3.3 Germano Identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.4 Invariance Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.5 Realizability Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Extension to the Inhomogeneous Case
for the Conventional Approach . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.1 Second-Order Commuting Filter . . . . . . . . . . . . . . . . . . . .
3.4.2 High-Order Commuting Filters . . . . . . . . . . . . . . . . . . . . .
3.5 Filtered NavierStokes Equations in General Coordinates . . . .
3.5.1 Basic Form of the Filtered Equations . . . . . . . . . . . . . . .
3.5.2 Simplied Form of the Equations
Non-linear Terms Decomposition . . . . . . . . . . . . . . . . . . .
3.6 Closure Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6.1 Statement of the Problem . . . . . . . . . . . . . . . . . . . . . . . . .
3.6.2 Postulates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6.3 Functional and Structural Modeling . . . . . . . . . . . . . . . .
4.
5.
49
49
59
61
64
72
74
74
77
77
77
78
78
78
79
80
Other Mathematical Models for the Large-Eddy

Simulation Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1 Ensemble-Averaged Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 Yoshizawas Partial Statistical Average Model . . . . . . . .
4.1.2 McCombs Conditional Mode Elimination Procedure . .
4.2 Regularized NavierStokes Models . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Lerays Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.2 Holms NavierStokes- Model . . . . . . . . . . . . . . . . . . . . .
4.2.3 Ladyzenskajas Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
83
83
83
84
85
86
86
89
Functional Modeling (Isotropic Case) . . . . . . . . . . . . . . . . . . . . .

5.1 Phenomenology of Inter-Scale Interactions . . . . . . . . . . . . . . . . .
5.1.1 Local Isotropy Assumption: Consequences . . . . . . . . . . .
5.1.2 Interactions Between Resolved and Subgrid Scales . . . .
5.1.3 A View in Physical Space . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Basic Functional Modeling Hypothesis . . . . . . . . . . . . . . . . . . . .
5.3 Modeling of the Forward Energy Cascade Process . . . . . . . . . .
5.3.1 Spectral Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.2 Physical Space Models . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.3 Improvement of Models in the Physical Space . . . . . . .
5.3.4 Implicit Diusion: the ILES Concept . . . . . . . . . . . . . . . .
5.4 Modeling the Backward Energy Cascade Process . . . . . . . . . . .
5.4.1 Preliminary Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
91
91
92
93
102
104
104
105
105
109
133
161
171
171
Contents
XXV
5.4.2 Deterministic Statistical Models . . . . . . . . . . . . . . . . . . . . 172

5.4.3 Stochastic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
6.
7.
Functional Modeling:
Extension to Anisotropic Cases . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1 Statement of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Application of Anisotropic Filter to Isotropic Flow . . . . . . . . . .
6.2.1 Scalar Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.2 Battens Mixed Space-Time Scalar Estimator . . . . . . . .
6.2.3 Tensorial Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3 Application of an Isotropic Filter to a Shear Flow . . . . . . . . . .
6.3.1 Phenomenology of Inter-Scale Interactions . . . . . . . . . . .
6.3.2 Anisotropic Models: Scalar Subgrid Viscosities . . . . . . .
6.3.3 Anisotropic Models: Tensorial Subgrid Viscosities . . . . .
6.4 Remarks on Flows Submitted to Strong Rotation Eects . . . .
Structural Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1 Introduction and Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 Formal Series Expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2.1 Models Based on Approximate Deconvolution . . . . . . . .
7.2.2 Non-linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2.3 Homogenization-Technique-Based Models . . . . . . . . . . . .
7.3 Scale Similarity Hypotheses and Models Using Them . . . . . . . .
7.3.1 Scale Similarity Hypotheses . . . . . . . . . . . . . . . . . . . . . . . .
7.3.2 Scale Similarity Models . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3.3 A Bridge Between Scale Similarity and Approximate
Deconvolution Models. Generalized Similarity Models .
7.4 Mixed Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.4.1 Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.4.2 Examples of Mixed Models . . . . . . . . . . . . . . . . . . . . . . . .
7.5 Dierential Subgrid Stress Models . . . . . . . . . . . . . . . . . . . . . . . .
7.5.1 Deardor Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.5.2 Fureby Dierential Subgrid Stress Model . . . . . . . . . . . .
7.5.3 Velocity-Filtered-Density-Function-Based Subgrid
Stress Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.5.4 Link with the Subgrid Viscosity Models . . . . . . . . . . . . .
7.6 Stretched-Vortex Subgrid Stress Models . . . . . . . . . . . . . . . . . . .
7.6.1 General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.6.2 S3/S2 Alignment Model . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.6.3 S3/ Alignment Model . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.6.4 Kinematic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.7 Explicit Evaluation of Subgrid Scales . . . . . . . . . . . . . . . . . . . . .
7.7.1 Fractal Interpolation Procedure . . . . . . . . . . . . . . . . . . . .
7.7.2 Chaotic Map Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
187
187
187
188
191
191
193
193
198
202
208
209
209
210
210
223
228
231
231
232
236
237
237
239
243
243
244
245
248
249
249
250
250
250
251
253
254
XXVI
Contents
7.7.3 Kersteins ODT-Based Method . . . . . . . . . . . . . . . . . . . . .

7.7.4 Kinematic-Simulation-Based Reconstruction . . . . . . . . .
7.7.5 Velocity Filtered Density Function Approach . . . . . . . . .
7.7.6 Subgrid Scale Estimation Procedure . . . . . . . . . . . . . . . .
7.7.7 Multi-level Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.8 Direct Identication of Subgrid Terms . . . . . . . . . . . . . . . . . . . . .
7.8.1 Linear-Stochastic-Estimation-Based Model . . . . . . . . . .
7.8.2 Neural-Network-Based Model . . . . . . . . . . . . . . . . . . . . . .
7.9 Implicit Structural Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.9.1 Local Average Method . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.9.2 Scale Residual Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
257
259
260
261
263
272
274
275
275
276
278
Numerical Solution: Interpretation and Problems . . . . . . . . .

8.1 Dynamic Interpretation of the Large-Eddy Simulation . . . . . . .
8.1.1 Static and Dynamic Interpretations: Eective Filter . .
8.1.2 Theoretical Analysis of the Turbulence
Generated by Large-Eddy Simulation . . . . . . . . . . . . . . .
8.2 Ties Between the Filter and Computational Grid.
Pre-ltering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3 Numerical Errors and Subgrid Terms . . . . . . . . . . . . . . . . . . . . .
8.3.1 Ghosals General Analysis . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.2 Pre-ltering Eect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.4 Remarks on the Use of Articial Dissipations . . . . . . . .
8.3.5 Remarks Concerning the Time Integration Method . . .
281
281
281
Analysis and Validation of Large-Eddy Simulation Data . .

9.1.1 Type of Information Contained
in a Large-Eddy Simulation . . . . . . . . . . . . . . . . . . . . . . . .
9.1.2 Validation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.1.3 Statistical Equivalency Classes of Realizations . . . . . . .
9.1.4 Ideal LES and Optimal LES . . . . . . . . . . . . . . . . . . . . . . .
9.1.5 Mathematical Analysis of Sensitivities
and Uncertainties in Large-Eddy Simulation . . . . . . . . .
9.2 Correction Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2.1 Filtering the Reference Data . . . . . . . . . . . . . . . . . . . . . . .
9.2.2 Evaluation of Subgrid-Scale Contribution . . . . . . . . . . . .
9.2.3 Evaluation of Subgrid-Scale Kinetic Energy . . . . . . . . . .
9.3 Practical Experience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
305
305
10. Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10.1 General Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.1.1 Mathematical Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.1.2 Physical Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
323
323
323
324
8.
9.
283
288
290
290
294
297
299
303
305
306
307
310
311
313
313
314
315
318
Contents XXVII
10.2 Solid Walls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10.2.2 A Few Wall Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.2.3 Wall Models: Achievements and Problems . . . . . . . . . . .
10.3 Case of the Inow Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.3.1 Required Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.3.2 Inow Condition Generation Techniques . . . . . . . . . . . . .
11. Coupling Large-Eddy Simulation
with Multiresolution/Multidomain Techniques . . . . . . . . . . . .
11.2 Methods with Full Overlap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.2.1 One-Way Coupling Algorithm . . . . . . . . . . . . . . . . . . . . . .
11.2.2 Two-Way Coupling Algorithm . . . . . . . . . . . . . . . . . . . . .
11.2.3 FAS-like Multilevel Method . . . . . . . . . . . . . . . . . . . . . . . .
11.2.4 Kravchenko et al. Method . . . . . . . . . . . . . . . . . . . . . . . . .
11.3 Methods Without Full Overlap . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.4 Coupling Large-Eddy Simulation with Adaptive
Mesh Renement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.4.2 Error Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
326
326
332
351
354
354
354
369
369
371
372
372
373
374
376
377
377
378
12. Hybrid RANS/LES Approaches . . . . . . . . . . . . . . . . . . . . . . . . . .

12.1 Motivations and Presentation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.2 Zonal Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.2.2 Sharp Transition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.2.3 Smooth Transition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.2.4 Zonal RANS/LES Approach as Wall Model . . . . . . . . . .
12.3 Nonlinear Disturbance Equations . . . . . . . . . . . . . . . . . . . . . . . . .
12.4 Universal Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.4.1 Germanos Hybrid Model . . . . . . . . . . . . . . . . . . . . . . . . . .
12.4.2 Speziales Rescaling Method and Related Approaches .
12.4.3 Baurles Blending Strategy . . . . . . . . . . . . . . . . . . . . . . . .
12.4.4 Arunajatesans Modied Two-Equation Model . . . . . . .
12.4.5 BushMani Limiters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12.4.6 Magagnatos Two-Equation Model . . . . . . . . . . . . . . . . . .
12.5 Toward a Theoretical Status for Hybrid
RANS/LES Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
383
383
384
384
385
387
388
390
391
392
393
394
396
397
398
13. Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13.1 Filter Identication. Computing the Cuto Length . . . . . . . . .
13.2 Explicit Discrete Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13.2.1 Uniform One-Dimensional Grid Case . . . . . . . . . . . . . . . .
13.2.2 Extension to the Multi-Dimensional Case . . . . . . . . . . . .
401
401
404
404
407
399
XXVIII Contents
13.2.3 Extension to the General Case. Convolution Filters . . . 407

13.2.4 High-Order Elliptic Filters . . . . . . . . . . . . . . . . . . . . . . . . . 408
13.3 Implementation of the Structure Function Models . . . . . . . . . . 408
14. Examples of Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14.1 Homogeneous Turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14.1.1 Isotropic Homogeneous Turbulence . . . . . . . . . . . . . . . . .
14.1.2 Anisotropic Homogeneous Turbulence . . . . . . . . . . . . . . .
14.2 Flows Possessing a Direction of Inhomogeneity . . . . . . . . . . . . .
14.2.1 Time-Evolving Plane Channel . . . . . . . . . . . . . . . . . . . . . .
14.2.2 Other Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14.3 Flows Having at Most One Direction of Homogeneity . . . . . . .
14.3.1 Round Jet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14.3.2 Backward Facing Step . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14.3.3 Square-Section Cylinder . . . . . . . . . . . . . . . . . . . . . . . . . . .
14.3.4 Other Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14.4 Industrial Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14.4.1 Large-Eddy Simulation for Nuclear Power Plants . . . . .
14.4.2 Flow in a Mixed-Flow Pump . . . . . . . . . . . . . . . . . . . . . . .
14.4.3 Flow Around a Landing Gear Conguration . . . . . . . . .
14.4.4 Flow Around a Full-Scale Car . . . . . . . . . . . . . . . . . . . . . .
14.5 Lessons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14.5.1 General Lessons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14.5.2 Subgrid Model Eciency . . . . . . . . . . . . . . . . . . . . . . . . . .
14.5.3 Wall Model Eciency . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14.5.4 Mesh Generation for Building Blocks Flows . . . . . . . . . .
411
411
411
412
414
414
418
419
419
426
430
431
432
432
435
437
437
439
439
442
444
445
15. Coupling with Passive/Active Scalar . . . . . . . . . . . . . . . . . . . . . .

15.1 Scope of this Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.2 The Passive Scalar Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.2.1 Physical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.2.2 Dynamics of the Passive Scalar . . . . . . . . . . . . . . . . . . . . .
15.2.3 Extensions of Functional Models . . . . . . . . . . . . . . . . . . .
15.2.4 Extensions of Structural Models . . . . . . . . . . . . . . . . . . . .
15.2.5 Generalized Subgrid Modeling for Arbitrary Non-linear
Functions of an Advected Scalar . . . . . . . . . . . . . . . . . . . .
15.2.6 Models for Subgrid Scalar Variance and Scalar Subgrid
Mixing Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.2.7 A Few Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.3 The Active Scalar Case: Stratication and Buoyancy Eects .
15.3.1 Physical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.3.2 Some Insights into the Active Scalar Dynamics . . . . . . .
15.3.3 Extensions of Functional Models . . . . . . . . . . . . . . . . . . .
15.3.4 Extensions of Structural Models . . . . . . . . . . . . . . . . . . . .
15.3.5 Subgrid Kinetic Energy Estimates . . . . . . . . . . . . . . . . . .
449
449
450
450
453
461
466
468
469
472
472
472
474
481
487
490
Contents
XXIX
15.3.6 More Complex Physical Models . . . . . . . . . . . . . . . . . . . . 492

15.3.7 A Few Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
A. Statistical and Spectral Analysis of Turbulence . . . . . . . . . . .
A.1 Turbulence Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.2 Foundations of the Statistical Analysis of Turbulence . . . . . . .
A.2.1 Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.2.2 Statistical Average: Denition and Properties . . . . . . . .
A.2.3 Ergodicity Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.2.4 Decomposition of a Turbulent Field . . . . . . . . . . . . . . . . .
A.2.5 Isotropic Homogeneous Turbulence . . . . . . . . . . . . . . . . .
A.3 Introduction to Spectral Analysis
of the Isotropic Turbulent Fields . . . . . . . . . . . . . . . . . . . . . . . . .
A.3.1 Denitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.3.2 Modal Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.3.3 Spectral Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.4 Characteristic Scales of Turbulence . . . . . . . . . . . . . . . . . . . . . . .
A.5 Spectral Dynamics of Isotropic Homogeneous Turbulence . . . .
A.5.1 Energy Cascade and Local Isotropy . . . . . . . . . . . . . . . .
A.5.2 Equilibrium Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . .
495
495
495
495
496
496
498
499
B. EDQNM Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
B.1 Isotropic EDQNM Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
B.2 Cambons Anisotropic EDQNM Model . . . . . . . . . . . . . . . . . . . .
B.3 EDQNM Model for Isotropic Passive Scalar . . . . . . . . . . . . . . . .
507
507
509
511
499
499
501
502
504
504
504
505
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
1. Introduction
1.1 Computational Fluid Dynamics

Computational Fluid Dynamics (CFD) is the study of uids in ow by numerical simulation, and is a eld advancing by leaps and bounds. The basic idea
is to use appropriate algorithms to nd solutions to the equations describing
the uid motion.
Numerical simulations are used for two types of purposes.
The rst is to accompany research of a fundamental kind. By describing
the basic physical mechanisms governing uid dynamics better, numerical
simulation helps us understand, model, and later control these mechanisms.
This kind of study requires that the numerical simulation produce data of
very high accuracy, which implies that the physical model chosen to represent
the behavior of the uid must be pertinent and that the algorithms used,
and the way they are used by the computer system, must introduce no more
than a low level of error. The quality of the data generated by the numerical
simulation also depends on the level of resolution chosen. For the best possible
precision, the simulation has to take into account all the space-time scales
aecting the ow dynamics. When the range of scales is very large, as it is
in turbulent ows, for example, the problem becomes a sti one, in the sense
that the ratio between the largest and smallest scales becomes very large.
Numerical simulation is also used for another purpose: engineering analyses, where ow characteristics need to be predicted in equipment design
phase. Here, the goal is no longer to produce data for analyzing the ow
dynamics itself, but rather to predict certain of the ow characteristics or,
more precisely, the values of physical parameters that depend on the ow,
such as the stresses exerted on an immersed body, the production and propagation of acoustic waves, or the mixing of chemical species. The purpose
is to reduce the cost and time needed to develop a prototype. The desired
predictions may be either of the mean values of these parameters or their
extremes. If the former, the characteristics of the systems normal operating
regime are determined, such as the fuel an aircraft will consume per unit
of time in cruising ight. The question of study here is mainly the systems
performance. When extreme parameter values are desired, the question is
rather the systems characteristics in situations that have a little probability of ever existing, i.e. in the presence of rare or critical phenomena, such
1. Introduction
as rotating stall in aeronautical engines. Studies like this concern system

safety at operating points far from the cruising regime for which they were
designed.
The constraints on the quality of representation of the physical phenomena dier here from what is required in fundamental studies, because what
is wanted now is evidence that certain phenomena exist, rather than all the
physical mechanisms at play. In theory, then, the description does not have
to be as detailed as it does for fundamental studies. However, it goes without
saying that the quality of the prediction improves with the richness of the
physical model.
The various levels of approximation going into the physical model are
discussed in the following.
1.2 Levels of Approximation: General

A mathematical model for describing a physical system cannot be dened
before we have determined the level of approximation that will be needed for
obtaining the required precision on a xed set of parameters (see [307] for
a fuller discussion). This set of parameters, associated with the other variables
characterizing the evolution of the model, contain the necessary information
for describing the system completely.
The rst decision that is made concerns the scale of reality considered.
That is, physical reality can be described at several levels: in terms of particle physics, atomic physics, or micro- and macroscopic descriptions of phenomena. This latter level is the one used by classical mechanics, especially
continuum mechanics, which will serve as the framework for the explanations
given here.
A system description at a given scale can be seen as a statistical averaging
of the detailed descriptions obtained at the previous (lower) level of description. In uid mechanics, which is essentially the study of systems consisting
of a large number of interacting elements, the choice of a level of description,
and thus a level of averaging, is fundamental. A description at the molecular
level would call for a denition of a discrete system governed by Boltzmann
equations, whereas the continuum paradigm would be called for in a macroscopic description corresponding to a scale of representation larger than the
mean free path of the molecules. The system will then be governed by the
NavierStokes equations, if the uid is Newtonian.
After deciding on a level of reality, several other levels of approximation
have to be considered in order to obtain the desired information concerning
the evolution of the system:
Level of space-time resolution. This is a matter of determining the time and
space scales characteristic of the system evolution. The smallest pertinent
1.3 Statement of the Scale Separation Problem
scale is taken as the resolution reference so as to capture all the dynamic

mechanisms. The system spatial dimension (zero to three dimensions) has
to be determined in addition to this.
Level of dynamic description. Here we determine the various forces exerted on the system components, and their relative importance. In the
continuum mechanics framework, the most complete model is that of the
NavierStokes equations, complemented by empirical laws for describing
the dependency of the diusion coecients as a function of the other variables, and the state law. This can rst be simplied by considering that
the elliptic character of the ow is due only to the pressure, while the other
variables are considered to be parabolic, and we then refer to the parabolic
NavierStokes equations. Other possible simplications are, for example,
Stokes equations, which account only for the pressure and diusion eects,
and the Euler equations, which neglect the viscous mechanisms.
The dierent choices made at each of these levels make it possible to develop a mathematical model for describing the physical system. In all of the
following, we restrict ourselves to the case of a Newtonian uid of a single
species, of constant volume, isothermal, and isochoric in the absence of any
external forces. The mathematical model consists of the unsteady Navier
Stokes equations. The numerical simulation then consists in nding solutions
of these equations using algorithms for Partial Dierential Equations. Because of the way computers are structured, the numerical data thus generated
is a discrete set of degrees of freedom, and of nite dimensions. We therefore
assume that the behavior of the discrete dynamical system represented by
the numerical result will approximate that of the exact, continuous solution
of the NavierStokes equations with adequate accuracy.
1.3 Statement of the Scale Separation Problem

Solving the unsteady NavierStokes equations implies that we must take into
account all the space-time scales of the solution if we want to have a result
of maximum quality. The discretization has to be ne enough to represent
all these scales numerically. That is, the simulation is discretized in steps
x in space and t in time that must be smaller, respectively, than the
characteristic length and the characteristic time associated with the smallest
dynamically active scale of the exact solution. This is equivalent to saying
that the space-time resolution scale of the numerical result must be at least as
ne as that of the continuous problem. This solution criterion may turn out
to be extremely constrictive when the solution to the exact problem contains
scales of very dierent sizes, which is the case for turbulent ows.
This is illustrated by taking the case of the simplest turbulent ow, i.e. one
that is statistically homogeneous and isotropic (see Appendix A for a more
1. Introduction
precise denition). For this ow, the ratio between the characteristic length
of the most energetic scale, L, and that of the smallest dynamically active
scale, , is evaluated by the relation:

L
= O Re3/4
(1.1)
in which Re is the Reynolds number, which is a measure of the ratio of

the forces of inertia and the molecular viscosity eect, . We therefore need
O Re9/4 degrees of freedom in order to be able to represent all the scales in

a cubic volume of edge L. The ratio of characteristic times varies as O Re1/2 ,
but the use of explicit time-integration algorithm leads to a linear dependency
of the time step with respect to the mesh size. So in order to calculate the
evolution of the solution in a volume L3 for a duration equal to the characteristic time of the most energetic
scale, we have to solve the NavierStokes

equations numerically O Re3 times!
This type of computation for large Reynolds numbers (applications in
the aeronautical eld deal with Reynolds numbers of as much as 108 ) requires computer resources very much greater than currently available supercomputer capacities, and is therefore not practicable.
In order to be able to compute the solution, we need to reduce the number
of operations, so we no longer solve the dynamics of all the scales of the
exact solution directly. To do this, we have to introduce a new, coarser level
of description of the uid system. This comes down to picking out certain
scales that will be represented directly in the simulation while others will not
be. The non-linearity of the NavierStokes equations reects the dynamic
coupling that exists among all the scales of the solution, which implies that
these scales cannot be calculated independently of each other. So if we want
a quality representation of the scales that are resolved, their interactions with
the scales that are not have to be considered in the simulation. This is done
by introducing an additional term in the equations governing the evolution of
the resolved scales, to model these interactions. Since these terms represent
the action of a large number of other scales with those that are resolved
(without which there would be no eective gain), they reect only the global
or average action of these scales. They are therefore only statistical models: an
individual deterministic representation of the inter-scale interactions would
be equivalent to a direct numerical simulation.
Such modeling oers a gain only to the extent that it is universal, i.e.
if it can be used in cases other than the one for which it is established.
This means there exists a certain universality in the dynamic interactions
the models reect. This universality of the assumptions and models will be
discussed all through the text.
1.4 Usual Levels of Approximation

There are several common ways of reducing the number of degrees of freedom
in the numerical solution:
By calculating the statistical average of the solution directly. This is called
the Reynolds Averaged Numerical Simulation (RANS)[424], which is used
mostly for engineering calculations. The exact solution u splits into the
sum of its statistical average u and a uctuation u (see Appendix A):
u(x, t) = u(x, t) + u (x, t) .
This splitting, or decomposition, is illustrated by Fig. 1.1. The uctuation u is not represented directly by the numerical simulation, and is
included only by way of a turbulence model. The statistical averaging operation is in practice often associated with a time averaging:

1 T
u(x, t)dt .
u(x, t) u(x) = lim
T T 0
The mathematical model is then that of the steady NavierStokes equations. This averaging operation makes it possible to reduce the number of
scales in the solution considerably, and therefore the number of degrees
of freedom of the discrete system. The statistical character of the solution prevents a ne description of the physical mechanisms, so that this
approach is not usable for studies of a fundamental character, especially
so when the statistical average is combined with a time average. Nor is
it possible to isolate rare events. On the other hand, it is an appropriate
approach for analyzing performance as long as the turbulence models are
able to reect the existence of the turbulent uctuation u eectively.
Fig. 1.1. Decomposition of the energy spectrum of the solution associated with
the Reynolds Averaged Numerical Simulation (symbolic representation).
By calculating directly only certain low-frequency modes in time (of the

order of a few hundred hertz) and the average eld. This approach goes
by a number of names: Unsteady Reynolds Averaged Numerical Simula-
1. Introduction
tion (URANS), Semi-Deterministic Simulation (SDS), Very Large-Eddy

Simulation (VLES), and sometimes Coherent Structure Capturing (CSC)
[726, 44]. The eld u appears here as the sum of three contributing terms
[456, 451, 240, 726]:
u(x, t) = u(x) + u(x, t)c + u (x, t) .
The rst term is the time average of the exact solution, the second its conditional statistical average, and the third the turbulent uctuation. This
decomposition is illustrated in Fig. 1.2. The conditional average is associated with a predened class of events. When these events occur at a set
time period, this is a phase average. The u(x, t)c term is interpreted as
the contribution of the coherent modes to the ow dynamics, while the
u term, on the other hand, is supposed to represent the random part
of the turbulence. The variable described by the mathematical model is
now the sum u(x) + u(x, t)c , with the random part being represented
by a turbulence model. It should be noted that, in the case where there
exists a deterministic low-frequency forcing of the solution, the conditional
average is conventionally interpreted as a phase average of the solution, for
a frequency equal to that of the forcing term; but if this does not exist,
the interpretation of the results is still open to debate. Since this is an unsteady approach, it contains more information than the previous one; but
it still precludes a deterministic description of a particular event. It is of
use for analyzing the performance characteristics of systems in which the
unsteady character is forced by some external action (such as periodically
pulsed ows).
Fig. 1.2. Decomposition of the energy spectrum of the solution associated with the
Unsteady Reynolds Averaged Numerical Simulation approach, when a predominant
frequency exists (symbolic representation).
By projecting the solution on the ad hoc function basis and retaining only
a minimum number of modes, to get a dynamical system with fewer degrees
of freedom. The idea here is to nd an optimum decomposition basis for
representing the phenomenon, in order to minimize the number of degrees
of freedom in the discrete dynamical system. There is no averaging done
here, so the space-time and dynamics resolution of the numerical model is
still as ne as that of the continuum model, but is now optimized. Several

approaches are encountered in practice.
The rst is to use standard basis function (Fourier modes in the spectral
space or polynomials in the physical space, for example) and distribute
the degrees of freedom as best possible in space and time to minimize the
number of them, i.e. adapt the space-time resolution of the simulation to
the nature of the solution. We thus adapt the topology of the discrete
dynamical system to that of the exact solution. This approach results in
the use of self-adapting grids and time steps in the physical space. It is
not associated with an operation to reduce the complexity by switching to
a higher level of statistical description of the system. It leads to a much less
important reduction of the discrete system than those techniques based on
statistical averaging, and is limited by the complexity of the continuous
solution.
Another approach is to use optimal basis functions, a small number of
which will suce for representing the ow dynamics. The problem is then
to determine what these basis functions are. One example is the Proper
Orthogonal Decomposition (POD) mode basis, which is optimum for representing kinetic energy (see [55] for a survey). This technique allows very
high data compression, and generates a dynamical system of very small
dimensions (a few dozen degrees of freedom at most, in practice). The approach is very seldom used because it requires very complete information
concerning the solution in order to be able to determine the base functions.
The various approaches above all return complete information concerning
the solutions of the exact problem, so they are perfectly suited to studies
of a fundamental nature. They may not, on the other hand, be optimal in
terms of reducing the complexity for certain engineering analyses that do
not require such complete data.
By calculating only the low-frequency modes in space directly. This is what
is done in Large-Eddy Simulation (LES). It is this approach that is discussed in the following. It is illustrated in Fig. 1.3.
Typical results obtained by these three approaches are illustrated in
Fig. 1.4.
Fig. 1.3. Decomposition of the energy spectrum in the solution associated with
large-eddy simulation (symbolic representation).
1. Introduction
Fig. 1.4. Pressure spectrum inside a cavity. Top: experimental data (ideal directnumerical simulation) (courtesy of L. Jacquin, ONERA); Middle: large-eddy simulation (Courtesy of L. Larcheveque, ONERA); Bottom: unsteady RANS simulation
(Courtesy of V. Gleize, ONERA).
1.5 Large-Eddy Simulation: from Practice to Theory. Structure of the Book
1.5 Large-Eddy Simulation: from Practice to Theory.

Structure of the Book
As mentioned above, the Large-Eddy Simulation approach relies on the definition of large and small scales. This fuzzy and empirical concept requires
further discussion to become a tractable tool from both the theoretical and
practical points of view. Bases for the theoretical understanding and modeling
of this approach are now introduced. In practice, the Large-Eddy Simulation
technique consists in solving the set of ad hoc governing equations on a computational grid which is too coarse to represent the smallest physical scales.
Let x and be the computional mesh size (assumed to be uniform for the
sake of simplicity) and the characteristic size of the smallest physical scales.
Let u be the exact solution of the following continuous generic conservation
law (the case of the NavierStokes equations will be extensively discussed in
the core of the book)
u
+ F (u, u) = 0
(1.2)
t
where F (, ) is a non-linear ux function. The Large-Eddy Simulation problem consists in nding the best approximation of u on the computational grid
by solving the following discrete problem
ud
+ Fd (ud , ud ) = 0
t
(1.3)
where ud , /t and Fd (, ) are the discrete approximations of u, /t and

F (, ) on the computational grid, respectively. Thus, the question arise of
dening what is the best possible approximation of u, u , among all discrete
solutions ud associated with x.
Let e(u, ud ) be a measure of the dierence between u and ud , which does
not need to be explicitely dened for the present purpose. It is just emphasized here that since Large-Eddy Simulation is used to compute turbulent
ows, u exhibits a chaotic behavior and therefore e(u, ud ) should rely on statistical moments of the solutions. A consistency constraint on the denition
of the error functional is that it must vanish in the limit case of the Direct
Numerical Simulation
lim e(u, ud ) = 0
(1.4)
A careful look at the problem reveals that the error can be decomposed
as
e(u, ud ) = e (u, ud ) + ed (u, ud ) + er (u, ud )
where
(1.5)
10
1. Introduction
1. e (u, ud ) is the projection error which accounts for the fact that the exact
solution u is approximated using a nite number of degrees of freedom.
The Nyquist theorem tells us that no scale smaller than 2x can be
captured in the simulation. As a consequence, ud can never be strictly
equal to u :
(1.6)
|u ud | = 0
2. ed (u, ud ) is the discretization error which accounts for the fact that partial derivatives which appear in the continuous problem are approximated
on the computational grid using Finite Dirence, Finite Volume, Finite
Element (or other similar) schemes. Putting the emphasis on spatial
derivatives, this is expressed as
Fd (u, u) = F (u, u)
(1.7)
3. er (u, ud ) is the resolution error, which accounts for the fact that, some
scales of the exact solution being missing, the evaluation of the non-linear
ux function cannot be exact, even if the discretization error is driven to
zero:
(1.8)
F (ud , ud ) = F (u, u)
This analysis shows that the Large-Eddy Simulation problem is very complex, since it depends explicitely on the exact solution, the computational grid
and the numerical method, making each problem appearing as unique. Therefore, it is necessary to nd some mathematical models for the Large-Eddy
Simulation problem which will mimic its main features, the most important
one being the removal of the small scales of the exact solution. A simplied
heuristic view of this problem is illustrated in Fig. 1.5, where the eect of the
Nyquist lter is represented.
Several mathematical models have been proposed to handle the true
Large-Eddy Simulation problem. The most popular one (see [216, 440, 495,
619, 627]) relies on the representation of the removal of the small scales as
the result of the application of a low-pass convolution lter (in terms of
wave number) to the exact solution. The denition and the properties of
this ltering operator are presented in Chap. 2. The application of this lter
to the NavierStokes equations, described in Chap. 3, yields the corresponding constitutive mathematical model for the large-eddy simulation. Alternate
mathematical models are detailed in Chap. 4.
The second question raised by the Large-Eddy Simulation approach deals
with the search for the best approximate solution u {ud } that will minimize the error e(u, ud). The short analysis given above shows that the projection error, e (u, ud ) cannot be avoided. Therefore, the best, ideal Large-Eddy
Solution is such that
e(u, ud ) = e(u, u ) = e (u, u )
(1.9)
11
Fig. 1.5. Schematic view of the simplest scale separation operator: grid and theoretical lters are the same, yielding a sharp cuto ltering in Fourier space between
the resolved and subgrid scales. The associated cuto wave number is denoted kc ,
which is directly computed from the cuto length in the physical space. Here,
is assumed to be equal to the size of the computational mesh.
and, following relation (1.5), it is associated to

ed (u, u ) + er (u, u ) = 0
(1.10)
The best solution in sought in practice trying to enforce the sequel relation (1.10). Two basic dierent ways are identied for that purpose:
The explicit Large-Eddy Simulation approach, in which an extra forcing
term, referred to as a subgrid model, is introduced in the governing equation
to cancel the resolution error. Two modeling approaches are discussed here:
functional modeling, based on representing kinetic energy transfers (covered in Chaps. 5 and 6), and structural modeling, which aims to reproduce
the eigenvectors of the statistical correlation tensors of the subgrid modes
(presented in Chap. 7). The basic assumptions and the subgrid models
corresponding to each of these approaches are presented. In the hypothetical case where a perfect subgrid model could be found, expression (1.10)
shows that the discretization error ed (u, ud ) must also be driven to zero
to recover the ideal Large-Eddy Simulation solution. A perfect numerical
method is obviously a natural candidate for that purpose, but reminding
that the error measure is based on statistical moments, the much less stringent requirement that the numerical method must be neutral with respect
to the error denition is sucient. Chapter 8 is devoted to the theoretical
12
1. Introduction
problems related to the eects of the numerical method used in the simulation. The representation of the numerical error in the form of an additional
lter is introduced, along with the problem of the relative weight of the
various lters used in the numerical simulation.
The implicit Large-Eddy Simulation approach, in which no extra term is
introduced in the governing equations, but the numerical method is chosen
such that the numerical error and the resolution error will cancel each
other, yielding a direct fullment of relation (1.10). This approach is briey
presented in this book in Sect. 5.3.4. The interested reader can refer to [276]
for an exhaustive description.
The fact that the ideal solution u is associated to a non-vanishing projection error e (u, u ) raises the problem of the reliability of data obtained
via Large-Eddy Simulation for practical purposes. Several theoretical and
practical problems are met when addressing the issue of validating and exploiting Large-Eddy Simulation. The denition of the best solution being
intrinsically based on the denition of the error functional (which is arbitrary), a universal answer seems to be meaningless. Questions concerning the
analysis and validation of the large-eddy simulation calculations are dealt
with in Chap. 9. The concept of statistically partially equivalent simulations
is introduced, which is of major importance to interpret the nature of the
data recovered from Large-Eddy Simulation. A short survey of available results dealing with the properties of ltered NavierStokes solutions (ideally
u ) and Large-Eddy Simulation solutions (true ud elds) is presented.
The discussions presented above deal with the denition of the LargeEddy Simulation problem inside the computational domain. As all dierential
problems, it must be supplemented with ad hoc boundary conditions to yield
a well-posed problem. Thus, the new question of dening discrete boundary
conditions in a consistent way appears. The problem is similar to the previous
one: what boundary conditions should be used to reach the best solution u ?
A weaker constraint is to nd boundary conditions which do not deteriorate
the accuracy that could potentially be reached with the selected numerical
scheme and closure. The boundary conditions used for large-eddy simulation
are discussed in Chap. 10, where the main cases treated are solid walls and
turbulent inow conditions. In the solid wall case, the emphasis is put on the
problem of dening wall stress models, which are subgrid models derived for
the specic purpose of taking into account the dynamics of the inner layer
of turbulent boundary layers. The issue of dening ecient turbulent inow
condtions raises from the need to truncate the computational domain, which
leads to the requirement of nding a way to take into account upstream
turbulent uctuations in the boundary conditions.
Despite the fact that it yields very signicant complexity reduction in
terms of degrees of freedom with respect to Direct Numerical Simulation,
Large-Eddy Simulation still requires considerable computational eorts to
handle realistic applications. To obtain further complexity reduction, several
13
hybridizations of the Large-Eddy Simulation technique have been proposed.

Methods for reducing the cost of Large-Eddy Simulation by coupling it with
multiresolution and multidomain techniques are presented in Chap. 11. Hybrid RANS/LES approaches are presented in Chap. 12. The denition of such
multiresolution methods and/or hybrid RANS/LES techniques raises many
practical and theoretical issues. Among the most important ones, the emphasis is put in the dedicated chapters on the coupling strategies and the fact
that the instantaneous elds can be fully discontinuous (fully meaning here
that the velocity eld is not a priori continuous at the interfaces between
domains with dierent resolution, but also that even the number of space
dimension and the number of unknwons can be dierent).
Practical aspects concerning the implementation of subgrid models are described in Chap. 13. Lastly, the discussion is illustrated by examples of largeeddy simulation applications for dierent categories of ows, in Chap. 14.
Chapter 15 is devoted the the extension of concepts, methods and models
presented in previous chapters to the case of a more complex physical system,
in which an additional equation for a scalar is added to the NavierStokes
equations. Two cases are considered: the passive scalar case, in which there
is no feedback in the momentum equation and the new problem is restricted
to closing the ltered scalar equation, and the active scalar case, which corresponds to a two-way coupling between the scalar eld and the velocity eld.
In the latter, the denition of subgrid models for both the velocity and the
scalar is a full problem. For the sake of clarity, the discussion is limited to
stably stratied ows and buoyancy driven ows. Combustion and two-phase
ows are not treated.
2. Formal Introduction to Scale Separation:

Band-Pass Filtering
The idea of scale separation introduced in the preceding chapter will now be
formalized on the mathematical level, to show how to handle the equations
and derive the subgrid models.
This chapter is devoted to the representation of the ltering as a convolution product, which is the most common way to model the removal of
small scales in the Larg-Eddy Simulation approach. Other denitions, such
as partial statistical averaging or conditional averaging [251, 250, 465], will
be presented in Chap. 4. The ltering approach is rst presented in the ideal
case of a lter of uniform cuto length over an innite domain (Sect. 2.1).
Fundamental properties of lters and their approximation via dierential operators is presented. Extensions to the cases of a bounded domain and a lter
of variable cuto length are then discussed (Sect. 2.2). The chapter is closed
by discussing a few properties of the Eulerian time-domain lters (Sect. 2.3).
2.1 Denition and Properties of the Filter

in the Homogeneous Case
The framework is restricted here to the case of homogeneous isotropic lters,
for the sake of easier analysis, and to allow a better understanding of the
physics of the phenomena. The lter considered is isotropic. This means that
its properties are independent of the position and orientation of the frame of
reference in space, which implies that it is applied to an unbounded domain
and that the cuto scale is constant and identical in all directions of space.
This is the framework in which subgrid modeling developed historically. The
extension to anisotropic and inhomogeneous1 lters, which researchers have
only more recently begun to look into, is described in Sect. 2.2.
2.1.1 Denition
Scales are separated by applying a scale high-pass lter, i.e. low-pass in frequency, to the exact solution. This ltering is represented mathematically in
1
That is, whose characteristics, such as the mathematical form or cuto frequency,
are not invariant by translation or rotation of the frame of reference in which
they are dened.
16
2. Formal Introduction to Filtering
physical space as a convolution product. The resolved part (x, t) of a spacetime variable (x, t) is dened formally by the relation:

(x, t) =
(, t )G(x , t t )dt d3
(2.1)
in which the convolution kernel G is characteristic of the lter used, which is

associated with the cuto scales in space and time, and c , respectively.
This relation is denoted symbolically by:
= G
(2.2)
The dual denition in the Fourier space is obtained by multiplying the

) of (x, t) by the spectrum G(k,
) of the kernel G(x, t):
spectrum (k,
) = (k,
)G(k,
) ,
(k,
(2.3)
,
= G
(2.4)
or, in symbolic form:

where k and are the spatial wave number and time frequency, respectively.
is the transfer function associated with the kernel G. The
The function G
spatial cuto length is associated with the cuto wave number kc and
time c with the cuto frequency c . The unresolved part of (x, t), denoted
(x, t), is dened operationally as:
(x, t) = (x, t) (x, t)
+
= (x, t)
or:
(2.5)
+
(, t )G(x , t t )dt d3 , (2.6)
= (1 G)
.
(2.7)
The corresponding form in spectral space is:
i.e.

) (k,
) = 1 G(k,
) ,
) (k,
(k, ) = (k,
(2.8)
.
= (1 G)
(2.9)
2.1 Denition and Properties of the Filter in the Homogeneous Case
17
2.1.2 Fundamental Properties

In order to be able to manipulate the NavierStokes equations after applying
a lter, we require that the lter verify the three following properties:
1. Conservation of constants

a = a
G(, t )d3 dt = 1
(2.10)
2. Linearity
+ =+
(2.11)
This property is automatically satised, since the product of convolution

veries it independently of the characteristics of the kernel G.
3. Commutation with derivation
=
,
s
s
s = x, t
(2.12)
Introducing the commutator [f, g] of two operators f and g applied to the

dummy variable
[f, g] f g() g f () = f (g()) g(f ())
the relation (2.12) can be re-written symbolically

G
,
=0 .
s
(2.14)
The commutator dened by relation (2.13) has the

properties2 :
[f, g] = [g, f ] Skew-symmetry ,
[f g, h] = [f, h] g + f [g, h]
[f, [g, h]] + [g, [h, f ]] + [h, [f, g]] = 0
(2.13)
following
(2.15)
Leibniz identity ,
(2.16)
Jacobis identity .
(2.17)
The lters that verify these three properties are not, in the general case,
Reynolds operators (see Appendix A), i.e.
=
=
G
G
= G2
= = G

G
(1 G)
= 0 ,
(2.18)
(2.19)
In the linear case, the commutator satises all the properties of the Poissonbracket operator, as dened in classical mechanics.
18
which is equivalent to saying that G is not a projector (excluding the trivial

case of the identity application). Let us recall that an application P is dened
as being a projector if P P = P . Such an application is idempotent because
it veries the relation
... P = P, n IN+
P n P P

(2.20)
n times
When G is not a projector, the ltering can be interpreted as a change of

variable, and can be inverted, so there is no loss of information3 [243]. The
kernel of the application is reduced to the null element, i.e. ker(G) = {0}.
If the lter is a Reynolds operator, we get
G2 = 1
(2.21)
or, remembering the property of conservation of constants:

G=1
(2.22)
In the spectral space, the idempotency property implies that the transfer
function takes the following form:

0

G(k, ) =
k, .
(2.23)
1
therefore takes the form of a sum of Dirac funcThe convolution kernel G
tions and Heaviside functions associated with non-intersecting domains. The
is 1 for the modes that are constant
conservation of constants implies that G
in space and time. The application can no longer be inverted because its kernel ker(G) = { } is no longer reduced to the zero element; and consequently,
the ltering induces an irremediable loss of information.
A lter is said to be positive if:
G(x, t) > 0, x and t .
(2.24)
2.1.3 Characterization of Dierent Approximations

The various methods mentioned in the previous section for reducing the number of degrees of freedom will now be explained. We now assume that the
3
The reduction of the number of degrees of freedom comes from the fact that the
new variable, i.e. the ltered variable, is more regular than the original one in
the sense that it contains fewer high frequencies. Its characteristic scale in space
is therefore larger, which makes it possible to use a coarser solution to describe
it, and therefore fewer degrees of freedom.
The result is a direct numerical simulation of the smoothed variable. As in
all numerical simulations, a numerical cuto is imposed by the use of a nite
number of degrees of freedom. But in the case considered here the numerical
cuto is assumed to occur within the dissipative range of the spectrum, so that
no active scales are missing.
19
space-time convolution kernel G(x , t t ) in IR4 is obtained by tensorizing

mono-dimensional kernels:

G(x , t t ) = G(x )Gt (t t ) = Gt (t t )
Gi (xi i ) . (2.25)
i=1,3
The Reynolds time average over a time interval T is found by taking:

Gt (t t ) =
HT
, Gi (xi i ) = (xi i ), i = 1, 2, 3 ,
T
(2.26)
in which is a Dirac function and HT the Heaviside function corresponding

to the interval chosen. This average is extended to the ith direction of space
by letting Gi (xi i ) = HL /L, in which L is the desired integration interval.
The phase average corresponding to the frequency c is obtained by letting:
t () = ( c ), Gi (xi i ) = (xi i ), i = 1, 2, 3 .
G
(2.27)
In all of the following, the emphasis will be put on the large-eddy simulation technique based on spatial ltering, because it is the most employed
approach, with very rare exceptions [160, 161, 603, 107]. This is expressed
by:
(2.28)
Gt (t t ) = (t t ) .
Dierent forms of the kernel Gi (xi i ) in common use are described
in the following section. It should nonetheless be noted that, when a spatial
ltering is imposed, it automatically induces an implicit time ltering, since
the dynamics of the NavierStokes equations makes it possible to associate
a characteristic time with each characteristic length scale. This time scale is
evaluated as follows. Let be the cuto length associated with the lter, and
kc = / the associated wave number. Let E(k) be the energy spectrum of
the exact solution (see Appendix A for a denition). The kinetic energy associated with the wave number kc is kc E(kc ). The velocity scale vc associated
with this same wave number is estimated as:

vc = kc E(kc ) .
(2.29)
The characteristic time tc associated with the length is calculated by
dimensional arguments as follows:
tc = /vc
(2.30)
The corresponding frequency is c = 2/tc . The physical analysis shows

that, for spectrum forms E(k) considered in the large-eddy simulation framework, vc is a monotonic decreasing function of kc (resp. monotonic increasing
20
function of ), so that c is a monotonic increasing function of kc (resp.

monotonic decreasing function of ). Suppressing the spatial scales corresponding to wave numbers higher than kc induces the disappearance of the
time frequencies higher than c . We nonetheless assume that the description
with a spatial ltering alone is relevant.
Eulerian time-domain ltering for spatial large-eddy simulation is recovered taking
(2.31)
Gi (xi i ) = (xi i ) .
A reasoning similar to the one given above shows that time ltering induces an implicit spatial ltering.
2.1.4 Dierential Filters
A subset of the lters dened in the previous section is the set of dierential
lters [242, 243, 245, 248]. These lters are such that the kernel G is the
Greens function associated to an inverse linear dierential operator F :
=
=
F (G
) = F ()
+
+ l
+ lm
+ ... ,
t
xl
xl xm
(2.32)
where and l are some time and space scales, respectively. Dierential lters
can be grouped into several classes: elliptic, parabolic or hyperbolic lters. In
the framework of a generalized space-time ltering, Germano [242, 243, 245]
recommends using a parabolic or hyperbolic time lter and an elliptic space
lter, for reasons of physical consistency with the nature of the NavierStokes
equations. It is recalled that a lter is said to be elliptic (resp. parabolic or
hyperbolic) if F is an elliptic (resp. parabolic, hyperbolic) operator. Examples
are given below [248].
Time Low-Pass Filter. A rst example is the time low-pass lter. The
associated inverse dierential relation is:
=+
(2.33)
The corresponding convolution lter is:

1
=

t t
(x, t ) exp
dt
(2.34)
It is easily seen that this lter commutes with time and space derivatives. This lter is causal, because it incorporates no future information, and
therefore is applicable to real-time or post-processing of the data.
21
Helmholtz Elliptic Filter. An elliptic lter is obtained by taking:

2
x2l
(2.35)
It corresponds to a second-order elliptic operator, which depends only on

space. The convolutional integral form is:

|x |
1
(, t)
exp
=
d .
(2.36)
2
|x |
4
This lter satises the three previously mentioned basic properties.
Parabolic Filter. A parabolic lter is obtained taking
= +
t
x2l
(2.37)
yielding

(, t)
(x )2
t t
ddt .
=
exp 2
3
)3/2
3/2
(t
t
(4)
4 (t t )
(2.38)
It is easily veried that the parabolic lter satisties the three required
properties.
Convective and Lagrangian Filters. A convective lter is obtained by

adding a convective part to the parabolic lter, leading to:
=+
2
+ Vl
t
xl
x2l
(2.39)
where V is an arbitrary velocity eld. This lter is linear and constant preserving, but commutes with derivatives if and only if V is uniform. A Lagrangian lter is obtained when V is taken equal to u, the velocity eld. In
this last case, the commutation property is obviously lost.
2.1.5 Three Classical Filters for Large-Eddy Simulation
Three convolution lters are ordinarily used for performing the spatial scale
separation. For a cuto length , in the mono-dimensional case, these are
written:
Box or top-hat lter:
G(x ) =
if |x |
otherwise
(2.40)
22
sin(k/2)

G(k)
=
k/2
(2.41)
are represented in
The convolution kernel G and the transfer function G
Figs. 2.1 and 2.2, respectively.
Gaussian lter:

G(x ) =
1/2

exp
|x |2
(2.42)

G(k)
= exp
k 2
4

,
(2.43)
in which is a constant generally taken to be equal to 6. The convolution

are represented in Figs. 2.3 and 2.4,
kernel G and the transfer function G
respectively.
Spectral or sharp cuto lter:
G(x ) =
sin (kc (x ))
, with kc =
kc (x )

G(k)
=
if |k| kc
otherwise
.
0
(2.44)
(2.45)
are represented in
The convolution kernel G and the transfer function G
Figs. 2.5 and 2.6, respectively.
It is trivially veried that the rst two lters are positive while the sharp
cuto lter is not. The top-hat lter is local in the physical space (its support
is compact) and non-local in the Fourier space, inversely from the sharp cuto
lter, which is local in the spectral space and non-local in the physical space.
As for the Gaussian lter, it is non-local both in the spectral and physical
spaces. Of all the lters presented, only the sharp cuto has the property:
n

G

G G...
= G = G ,
n times
and is therefore idempotent in the spectral space. Lastly, the top-hat and
Gaussian lters are said to be smooth because there is a frequency overlap
between the quantities u and u .
Modication of the exact solution spectrum by the ltering operator is
illustrated in gure 2.7.
23
Fig. 2.1. Top-hat lter. Convolution kernel in the physical space normalized by .
Fig. 2.2. Top-hat lter. Associated transfer function.
24
Fig. 2.3. Gaussian lter. Convolution kernel in the physical space normalized
by .
Fig. 2.4. Gaussian lter. Associated transfer function.
Fig. 2.5. Sharp cuto lter. Convolution kernel in the physical space.
Fig. 2.6. Sharp cuto lter. Associated transfer function.
25
26
Fig. 2.7. Energy spectrum of the unltered and ltered solutions. Filters considered
are a projective lter (sharp cuto lter) and a smooth lter (Gaussian lter) with
the same cuto wave number kc = 500.
2.1.6 Dierential Interpretation of the Filters

General results. Convolution lters can be approximated as simple dierential operators via a Taylor series expansion, if some additional constraints
are fullled by the convolution kernel, thus yielding simplied and local ltering operators. Validation of the use of Taylor series expansions in the
representation of the ltering operator and conditions for convergence will
be discussed in the next section.
We rst consider space ltering and recall its denition using a convolution
product:

+
(x, t) =
(, t)G(x )d
(2.46)
To obtain a dierential interpretation of the lter, we perform a Taylor

series expansion of the (, t) term at (x, t):
(, t) = (x, t) + ( x)
2 (x, t)
(x, t) 1
+ ( x)2
+ ...
x
2
x2
(2.47)
Introducing this expansion into (2.46), and considering the symmetry and
conservation properties of the constants of the kernel G, we get:
27
(x, t)
1 2 (x, t) + 2
z G(z)dz + ...
2 x2

1 n (x, t)
+
z n G(z)dz + ...
n! xn
(l) l (x, t)
(x, t) +
,
l!
xl
(x, t)
(2.48)
l=1,
where (l) designates the lth-order moment of the convolution kernel:

+
(l) = (1)l
z l G(z)dz .
(2.49)
Assuming that the solution is 2-periodic, the moments of the convolution

kernel can be rewritten as follows [728, 607]:

(l) = Ml ,
(x)/
l G()d
Ml =
(2.50)
(x)/
leading to the following expression for the ltered variable :

(1)k
(x) =
k!
k=0
Mk (x)
k
(x)
xk
(2.51)
With this relation, we can interpret the ltering as the application of

a dierential operator to the primitive variable . The subgrid part can also
be rewritten using the following relation
(x)
=
=
(x) (x)
(l) l (x)
l! xl
l=1,
k+1

k=1
(1)
k!
Mk (x)
k
(x)
xk
(2.52)
The ltered variable can also be expanded using derivatives of the trans of the lter [607]. Assuming periodicity and dierentiability
fer function G
of , we can write
(x) =
+

k ekx
2 = 1 ,
(2.53)
k=
and
+

l
(x) =
(k)l k ekx
xl
k=
(2.54)
28
The ltered eld is expanded as follows:

(x) =
+

k ekx
G(k)
(2.55)
k=
The ltered eld can be expressed as a Taylor series expansion in the lter
width :
2
(x, ) = (x, 0) +
(x, 0) +
(x, 0) + ...
2 2
(2.56)
By dierentiating (2.55) with respect to , we obtain

l
l
(x, 0) = l!al
l
(x)
xl
(2.57)
with
1 lG
(0) .
l l! k l
The resulting nal expression of the ltered eld is
al =
+

(k)l l G
(0)k ekx
l! k l
(x) =
(2.58)
(2.59)
l=0 k=
Time-domain lters dened as a convolution product can be expanded in

an exactly similar way, yielding
+
(x, t) =
(x, t )G(t t )dt
(x, t) +
(l) l (x, t)
l!
tl
(2.60)
l=1,
The values of the rst moments of the box and Gaussian lters are given
in Table 2.1. It can be checked that the sharp cuto lter leads to a divergent
series, because of its non-localness.
For these two lters, we have the estimate
n
(n) = O( )
(2.61)
(n) = O(c n )
(2.62)
for space-domain ltering, and
for time-domain ltering.
29
Table 2.1. Values of the rst ve non-zero moments for the box and Gaussian
lters.
(n)
n=0
box
Gaussian
n=2
n=4
n=6
/12
2
/12
1
1
/80
4
/48
/448
6
5 /576
n=8
8
/2304
8
35 /6912
For a general spacetime lter, neglecting cross-derivatives of the kernel,

this Taylor series expansion gives [160, 161]:

(x, t)
(, t )G(x , t t )ddt
l
(l)
(l) l (x, t)
x (x, t)
t
+
, (2.63)
l!
xl
l!
tl
= (x, t) +
l=1,
with

(l)
x
and
(l)
t
l=1,
( x)l G(x , t t )ddt
(2.64)
(t t)l G(x , t t )ddt
(2.65)
Conditions for Convergence of the Taylor Series Expansions. A rst

analysis of the convergence properties of the Taylor series expansions discussed above was provided by Vasilyev et al. [728], and is given below. Assuming that the periodic one-dimensional eld does not contain wave numbers
higher than kmax , one can write the following Fourier integral:

kmax
(x) =
kx
dk
(k)e
(2.66)
kmax
where time-dependence has been omitted for the sake of simplicity. The total
energy of , E , is equal to

E =
kmax
kmax
|(k)|
dk
(2.67)
The mth derivative of can be written as

m
(x) = ()m
xm
kmax
kmax
kx
k m (k)e
dk
(2.68)
30
From this expression, we get the following bounds for the derivative:
m
kmax

2

|k|2m |(k)|
dk
xm
kmax
1/2
1/2

kmax
kmax
|k|m dk
|(k)|dk
kmax
2E kmax m
k
2m + 1 max
kmax
(2.69)
From relations (2.69) and (2.51) we obtain the following inequalities:

l

(1)l

1 l
l
l

Ml (x) l (x)
|Ml (x)| l (x)

l!
x
l!
x
l=0
l=0
l

kmax |Ml (x)|
.(2.70)
2E kmax
l! 2l + 1
l=0
From this last inequality, it can easily be seen that the series (2.51) converges for any value of if the following constraint is satised:
lim
|Ml+1 (x)|
=0 .
|Ml (x)|(l + 1)
(2.71)
For lters with compact support, the following criterion holds:

lim
(kmax )|Ml+1 (x)|

<1 .
|Ml (x)|(l + 1)
(2.72)
For symmetric lters, the analogous criterion is

(kmax )2 |M2l+2 (x)|
<1 .
l |M2l (x)|(2l + 2)(2l + 1)
lim
(2.73)
Pruett et al. [607] proved that all symmetric, non-negative4 lters satisfy
relation (2.73). This proof is now presented. If the lter is non-negative,
following an integral mean value theorem, there exists a value c, 2
x c x 2, such that

2 (x)/

|c|

2l
G()d
|M2l+2 | =

(x)/

2
|c|
|M2l | ,
(2.74)
=
It is recalled that the lter is said to be non-negative if G(x) 0, x.
2.2 Spatial Filtering: Extension to the Inhomogeneous Case
whereby
|M2l+2 |
=
|M2l |
|c|
2
31
2
.
(2.75)
The expected result is trivially deduced from this last relation.
2.2 Spatial Filtering:

Extension to the Inhomogeneous Case
2.2.1 General
In the above explanations, it was assumed that the lter is homogeneous
and isotropic. These assumptions are at time too restrictive for the resulting
conclusions to be usable. For example, the denition of bounded uid domains
forbids the use of lters that are non-local in space, since these would no
longer be dened. The problem then arises of dening lters near the domain
boundaries. At the same time, there may be some advantage in varying the
lter cuto length to adapt the structure of the solution better and thereby
ensure optimum gain in terms of reducing the number of degrees of freedom
in the system to be resolved.
From relation (2.1), we get the following general form of the commutation
error for a convolution lter G(y, (x, t)) on a domain [230, 260]:

, G
=
(G
) G
.
(2.76)
x
x
x
The rst term of the right-hand side of (2.76) can be expanded as

(G
) =
G(x , (x, t))(, t)d
(2.77)
x
x

G
+
=
G(x , (x, t))(, t)n()ds
,
(2.78)
+G
x
where n() is the outward unit normal vector to the boundary of , ,
yielding

, G
=
+
G(x , (x, t))(, t)n()ds. (2.79)

x
x
The rst term appearing in the right-hand side of relation (2.79) is due to
spatial variation of the ltering length, while a domain boundary generates
the second one. A similar development leads to:
32

G
, G
=
t
t
(2.80)
These error terms must be eliminated, or bounded, in order to be able to

dene a controlled and consistent ltering process for large-eddy simulation.
This is done by deriving new ltering operators. Several alternatives to the
classical convolution products have been proposed, which are described in
the following.
Franke and Frank [225] propose an extension of (2.79) to the case of a domain with moving boundaries and a uniform time depdendent lter length,
i.e. = (t) and = (t). Limiting the analysis to the one dimensionalcase for the sake of clarity, and taking (t) = [a(t), b(t)], one rst notices
that the constraint dealing with the preservation of the constant

b(t)
G(x , (t))d = 1 ,
(2.81)
a(t)
yields the following lter conservation law

G d(t)
db(t)
da(t)
d = G(x b(t), (t))
G(x a(t), (t))
.
dt
dt
a(t) dt
(2.82)
The commutation errors have the following forms:

, G
(x, t) = G(x a(t), (t))(a(t), t) G(x b(t), (t))(b(t), t) ,
x
(2.83)

b(t)
, G
(x, t)
t

=b(t)
d
G(x , (t))(, t)
dt =a(t)
b(t)
G d(t)
d .
+
(, t)
dt
a(t)
(2.84)
2.2.2 Non-uniform Filtering Over an Arbitrary Domain

This section presents the ndings concerning the extension of the ltering to
the case where the lter cuto length varies in space and where the domain
over which it applies is bounded or innite.
New Denition of Filters and Properties: Mono-dimensional Case.
Alternative proposals in the homogeneous case. Ghosal and Moin [262] propose to dene the ltering of a variable (), dened over the interval
33
] , +[, as
() = G
=

()d
(2.85)
in which the cuto length is constant. The convolution kernel G is made

to verify the following four properties:
1. Symmetry
G() = G()
(2.86)
We note that this property was not explicitly required before, but that
it is veried by the three lters described in Sect. 2.1.5.
2. Conservation of constants
+
a G()d, a = const.
(2.87)
3. Fast decay. G() 0 as || fast enough for all of its moments to

be nite, i.e.
+
G() n d < , n 0 .
(2.88)
4. Quasi-local in physical space. G() is localized (in a sense to be specied)

in the interval [1/2, 1/2].
Extension of the Top-Hat Filter to the Inhomogeneous Case: Properties. Considering denition (2.85), the top-hat lter (2.40) is expressed:

1 if || 12
G() =
.
(2.89)
0 otherwise
There are a number of ways of extending this lter to the inhomogeneous
case. The problem posed is strictly analogous to that of extending nite volume type schemes to the case of inhomogeneous structured grids: the control
volumes can be dened directly on the computational grid or in a reference
space carrying a uniform grid, after a change of variable. Two extensions
of the box lter are discussed in the following, each based on a dierent
approach.
Direct extension. If the cuto length varies in space, one solution is to say:
1
() =
+ () + ()
++ ()
()d
(2.90)
()
in which + () and () are positive functions and (+ () + ()) is the

cuto length at point . These dierent quantities are represented in Fig. 2.8.
34
Fig. 2.8. Direct extension of the top-hat lter. Representation of the integration
cell at point .
If the domain is nite or semi- innite, the functions + () and () must

decrease fast enough near the domain boundaries for the integration interval
[ (), ++ ()] to remain dened. The box lter is extended intuitively
here, as an average over the control cell [ (), + + ()]. This approach
is similar to the nite volume techniques based on control volumes dened
directly on the computational grid.
It is shown that this expression does not ensure the commutation property
with derivation in space. Relation (2.12) becomes (the dependency of the
functions + and as a function of is not explicitly state, to streamline
the notation):

(d/d) ( + + )
d
=
G
,
d
+ +

1
d+
d
+ ( )
( + + )
.
+ +
d
d
(2.91)
The amplitude of the error committed cannot be evaluated a priori, and
thus cannot be neglected. Also, when (2.90) is applied to the NavierStokes
equations, all the terms, including the linear ones, will introduce unknown
terms that will require a closure.
Extension by Variable Change. SOCF. To remedy this problem, a more
general alternative description than relation (2.90) is proposed by Ghosal and
Moin [262]. This new denition consists of dening lters that commute at
the second order with the derivation in space (Second Order Commuting
Filter, or SOCF). This is based on a change of variable that allows the use of
a homogeneous lter. The function is assumed to be dened over a nite
or innite interval [a, b]. Any regular monotonic function dened over this
interval can be related to a denite function over the interval [, +] by
performing the variable change:
= f (x) ,
(2.92)
in which f is a strictly monotonic dierentiable function such that:

f (a) = , f (b) = + .
(2.93)
35
The constant cuto length dened over the reference space [, +]

is associated with the variable cuto length (x) over the starting interval by
the relation:
.
(2.94)
(x) =
f (x)
In the case of a nite or semi-innite domain, the function f takes innite
values at the bounds and the convolution kernel becomes a Dirac function.
The ltering of a function (x) is dened in the inhomogeneous case in three
steps:
1. We perform the variable change x = f 1 (), which leads to the denition
of the function () = (f 1 ()).
2. The function () is then ltered by the usual homogeneous lter (2.85):

f (x)
1 +
(x) () =
G
()d .
(2.95)

3. The ltered quantity is then re-expressed in the original space:

f (x) f (y)
1 b
(x) =
G
(y)f (y)dy .
a
(2.96)
This new expression of the lter modies the commutation error with the
spatial derivation. Using (2.95) and integrating by parts, we get:
d
dx
=
+

y=b
f (x)
f (x) f (y)
G
(y)
y=a

b
f (x) f (y)
1
G
f (x) (y)dy
a
(2.97)
The fast decay property of the kernel G makes it possible to cancel the
rst term of the rigth-hand side. The commutation error is:

f (x) f (y)
d
1 b
G
G
,
=
f (y) (y)
d
a

f (x)
1
dy .
(2.98)
f (y)
In order to simplify this expression, we introduce a new variable such
that:
f (y) = f (x) + .
(2.99)
The variable y is then re-expressed as a series as a function of :
2
y = y0 () + y1 () + y2 () + ...
(2.100)
36
Then, combining relations (2.99) and (2.100), we get (the dependence of

the functions according to the variable x is not explicitly shown, to streamline
the notation):
2
f
y =x+
+ ... ,
(2.101)
f
2f 3
which allows us to re-write relation (2.98) as:

d
G
,
d

f (x)
G() (y()) 1
d
f (y())
(2.102)
= C1 + C2 + ... ,
(2.103)
in which the coecients C1 and C2 are expressed as:

C1
C2
f
f 2
G()d
(2.104)
2f f + f f 3f
2f 4
2
2 G()d
(2.105)
The symmetry property of the kernel G implies C1 = 0, which ensures

that the lter commutation error with the spatial derivation is of the second
order as a function of the cuto length . The authors call this Second-Order
Commuting Filter (SOCF).
A study of the spectral distribution of the commutation error is available
in reference [262]. Rather than detailing this analysis here, only the major
results will be explained. Considering a function of the form:
(x) = k ekx ,
2 = 1 ,
(2.106)
the two derivation operations are written:

d
= k,
dx
d
= k
dx
(2.107)
The commutation error can be measured by comparing the two wave

numbers k and k , the latter being such that k = k . The commutation
error is zero if k = k . Algebraic manipulations lead to the relation:

k
f
= 1 2
+
k
f
G() sin(k/f )d
.
G() cos(k/f )d
(2.108)
37
Using the modal decomposition (2.106), the commutation error can be

expressed in dierential form. The calculations lead to:

f 2 d2
d
+ O(k)4
(2.109)
= (2) 3
G
,
dx
dx2
f

d2
(2) 2
=
+ O(k)4 ,
(2.110)
dx2
in which (x) is the local cuto length and (2) the second-order moment of
G, i.e.
+
(2)
=
2 G()d .
(2.111)
Van der Vens Filters. Commuting lters can be dened with the spatial
derivation at an order higher than 2, at least in the case of an innite domain.
To obtain such lters, Van der Ven [725] proposes dening the ltering for
the case of a variable cuto length (x) by direct extension of the form (2.85):

+
1
xy
(x) =
G
(y)dy .
(2.112)
(x)
(x)
The function G is assumed here to be class C 1 , symmetrical, and must
conserve the constants. Also, the function (x) is also assumed to be class C 1 .
This denition is achieved by linearizing the general formula (2.96) around x,
that is by letting (y) = (x) and (x) (y) = (x)(x y) and including
relation (2.94). This linearization operation is equivalent to considering that
the function is linear in a neighbourhood of x containing the support of
the convolution kernel. By introducing the variable change y = x (x), the
corresponding commutation error is expressed:

d
G
,
(G() + G ()) (x (x))d
=
dx
(2.113)
To increase the order of the commutation error, we look for functions G

that are solutions to the equation
G + G = a G(n) ,
n>1 ,
(2.114)
in which a is a real and G(n) designates the nth derivative of the kernel G.
For such functions, the commutation error becomes:

d
G
,
dx
a (1)n
a (x)(x)n1 (n) (x)
G()
n
(x (x))d (2.115)
,
(2.116)
38
Fig. 2.9. High-order commuting lter. Graph of the associated transfer function
for dierent values of the parameter m.
and is thus formally of order n 1. Simple analysis shows that the Fourier
of the solution to problem (2.114) verifying the constant contransform G
servation property is of the form:

an n

G(k) = exp
k
.
(2.117)
n
The symmetry property of G implies that n = 2m is even, and therefore:

a(1)m 2m

G(k)
= exp
k
.
(2.118)
2m
The fast decay property is recovered for a = b(1)m , b > 0. It can be seen
that the Gaussian lter then occurs again by letting m = 1. It is important
to note that this analysis is valid only for innite domains, because when the
bounds of the uid domain are included they bring out additional error terms
with which it is no longer possible to be sure of the order of the commutation
error. The transfer function obtained for various values of the parameter m
is represented in Fig. 2.9.
High-Order Commuting Filters. Van der Vens analysis has been generalized by Vasilyev et al. [728] so as to contain previous works (SOCF and
Van der Vens lters) as special cases. As for SOCF, the ltering process is
dened thanks to the use of a reference space. We now consider that the
physical domain [a, b] is mapped into the domain [, ]. Ghosal and Moin
used = and = +. The correspondances between the two domains
are summarized in Table 2.2.
39
Table 2.2. Correspondances for Vasilyevs high-order commuting lters.

[a, b]
Domain
Coordinate
Filter length
Function
[, ]
1
x = f ()
(x) = /f (x)
(x)
= f (x)
() = (f 1 ())
Considering this new mapping, relation (2.95) is transformed as

1
() =
G
()d ,
(2.119)

and using the change of variables (2.99), we get

() =
G () ( )d
(2.120)
The next step consists in performing a Taylor expansion of ( ) in

powers of :

( ) =
(1)k k k k

()
k!
k
(2.121)
(1)k k (k) k
() k () ,
k!
(2.122)
k=0,+
Substituting (2.121) into (2.120), we get

() =
k=0,+
where the kth moment of the lter kernel is now dened as

(k)
G () k d .
() =
(2.123)
Using the relation (2.122), the space derivative of the ltered variable
expressed in the physical space can be evaluated as follows:
d
(x)
dx
=
=
d
()
(2.124)
d

(1)k k d(k)
k
k+1
f (x)
() k () + (k) () k+1 () .
k!
d
f (x)
k=0,+
(2.125)
A similar procedure is used to evaluate the second part of the commutation error. Using (2.124) and the same change of variables, we get:
40
d
1
(x) =
dx
d
()f (f 1 ())d
d
(2.126)
with
f (f 1 ()) =

l=1,+
(l 1)!

k=1,+
l1
k 1
lf
(x)
xl
(1) k k f

()
k!
k
(2.127)
and
d
() =
d

k=0,+
(1)k k k k+1

()
k!
k+1
(2.128)
Making the assumptions that all the Taylor expansion series are
convergent5, the commutation error in the physical space is equal to

G
,

d
=
dx
Ak (k) () +
k=1,+
Bk
k=0,+
d(k)
k
()
d
(2.129)
where Ak and Bk are real non-zero coecients. It is easily seen from relation
(2.129) that the commutation error is determined by the lter moments and
the mapping function. The order of the commutation error can then be governed by chosing an adequate lter kernel. Vasilyev proposes to use a function
G such that:
(0)
[, ] ,
(k)
=
<
1 k n 1, [, ]
k n, [, ] .
(k)
(2.130)
,
(2.131)
(2.132)
(2.133)
An immediate consequence is
d(k)
() = 0, 0 k n 1, [, ]
d
leading to

G
,

d
n
= O( ) .
dx
(2.134)
The commutation error can be controlled by choosing a kernel G with desired moment values. It is important noting that conditions (2.130) (2.132)
do not require that the lter kernel be symmetric. Discrete lters verifying
theses properties will be discussed in Sect. 13.2.
5
Vasilyev et al. [728] show that this is always true for practical numerical simulations.
41
Extension to the Multidimensional Case.

SOCF Filters. SOCF lters are extensible to the three-dimensional case
for nite or innite domains. Let (x1 , x2 , x3 ) be a Cartesian system, and
(X1 , X2 , X3 ) the reference axis system associated with a uniform isotropic
grid with a mesh size . The two systems are related by the relations:
X1 = H1 (x1 , x2 , x3 ),
x1 = h1 (X1 , X2 , X3 )
(2.135)
X2 = H2 (x1 , x2 , x3 ),
X3 = H3 (x1 , x2 , x3 ),
x2 = h2 (X1 , X2 , X3 )
x3 = h3 (X1 , X2 , X3 )
,
,
(2.136)
(2.137)
or, in vectorial form:

X = H(x),
x = h(X),
h = H 1
(2.138)
The ltering of a function (x) is dened analogously to the monodimensional case. We rst make a variable change to work in the reference
coordinate system, in which a homogeneous lter is applied, and then perform the inverse transformation. The three-dimensional convolution kernel is
dened by tensorizing homogeneous mono-dimensional kernels.
After making the rst change of variables, we get:

Xi Xi
1
(h(X)) = 3
G
(2.139)
(h(X ))d3 X ,
i=1,3
or, in the original space:

1
Hi (x) Xi
(h(X ))d3 X
(x) =
G
(2.140)
3
i=1,3

Hi (x) Hi (x )
1
=
G
(x )J(x )d3 x , (2.141)
3
i=1,3
where J(x) is the Jacobian of the transformation X = H(x). Analysis of
the error shows that, for lters dened this way, the commutation error with
the derivation in space is always of the second order, i.e.
= O( )
xk
xk
(2.142)
where the second term of the left-hand side is written:

1
1 Hj (x) Xj
G
=
3
xk

Hi (x) Xi
G
Hj,k (x)(h(X ))d3 X ,(2.143)
i=1,3;i=j
42
with the notation:

Hj,k (x) =
Hj (x)
xk
(2.144)
Dierential analysis of the commutation error is performed by considering

the solutions of the form:
k exp(k x) .
(x) =
(2.145)
An analogous approach to the one already made in the mono-dimensional

case leads to the relation:

2
2
G
,
+ O(k)4 ,
(2.146)
= (2) kmp
xk
xm xp
where the function kmp is dened as:
kmp = hm,jq (H(x))hp,q (H(x))Hj,k (x)
(2.147)
Van der Vens Filters. Van der Vens simplied ltering naturally extends to
the three-dimensional case in Cartesian coordinates by letting:

xi x
1
i
(x) =
G
(2.148)
(x )d3 x ,
i (x)
i (x) R3 i=1,3
i=1,3
in which i (x) is the lter cuto length in the ith direction of space at point
x. For a kernel G verifying (2.114), the commutation error is expressed:

i (x)
n (x)
G
,
i (x)n1
=a
xj
xj
xni
i=1,3
(2.149)
and is formally of order n 1.

High-Order Commuting Filters. Vasilyevs lters are generalized to the multiple dimension case in the same way as SOCF.
2.2.3 Local Spectrum of Commutation Error
A spectral tool for analyzing the wavenumber sprectrum of the commutation
error, referred to as the local spectrum analysis, was introduced by Vasilyev
and Goldstein [727]. This tool enables an accurate understanding of the impact of the commutation error on the derivatives of the ltered quantities.
Writing the convolution lter as

+
xy
1
(x) =
G
(y)dy ,
(2.150)
(x)
(x)
2.3 Time Filtering: a Few Properties
43
and introducing its local Fourier decomposition (which can be evaluated using
a windowed Fourier transform on bounded domains)

(x) =

kx
k(x) (k)e

G
dk
(2.151)
one can identify the coecients of the local Fourier transform of the ltered
quantity

x) = G

k(x) (k)
.
(2.152)
(k,
The commutation error can be written as

+
d
1 d(x)
1

, G
()(x) =
K(k(x))(k)
ekx dk
dx
2
(x) dx
(2.153)
is dened as
where the transfer function K

dG(k)

K(k)
= k
dk
(2.154)
By analogy with the previous case, the local spectrum of the commutation
error is dened as

d
1 d(x)

K(k(x))(k)
.
(2.155)
, G
()(k, x) =
dx
(x) dx
This expresion shows that the gradient of the cuto length (x) aects
the amplitude of the commutation error, while the lter shape (more precisly
the gradient of the transfer function in Fourier space) governs the spectral
repartition of the error. Analyses carried out considering convolution kernels
presented in Sect.2.1.5 reveal that the spectrum of commutation error is
global for smooth lters like the Gaussian lter (i.e. error occurs at all scales)
while it is much more local for sharp lters (i.e. the commutation error is
concentrated on a narrow wavenumber range).
2.3 Time Filtering: a Few Properties

We consider here continuous causal lters of the form [603]:

(x, t) = G
(x, t) =
(, t )G(t t )dt
(2.156)
where the kernel G satises immediately two of the three fundamental properties given in Sect. 2.1.2, namely the linearity constraint and the constantconservation constraint.
44
Due to the causality constraint, the time-domain support of these lters

is bounded, i.e.
lim G(t) = 0 .
(2.157)
t
Then, the following commutation properties hold:

=0 ,
G
,
xj

G
G
,
= (x, t)G(0)
t
t
(2.158)
(2.159)
It is observed that the use of spatially bounded domains does not introduce any commutation error terms, the lter being independent of the
position in space.
Two examples [606] are the exponential lter

t t
1 t
1 t
(, t ) exp
dt , (2.160)
G(t) = e (x, t) =
where is the characteristic cuto time, and the Heaviside lter

G(t) =
1
H(t + )
(2.161)
where H(t) is the Heaviside function, yielding

1
(x, t) =
(, t )dt
(2.162)
An interesting property of Eulerian time-domain ltering is that local

dierential expression of the lters are easily derived, whose practical implementation is easier than those of their original counterparts. Dierentiating
relation (2.162), on obtains the dierential form of the Heaviside lter:

1
(x, t)
=
(x, t) (x, t )
t
(2.163)
(2.164)
The dierential exponential lter is expressed as

1
(x, t)
=
(x, t) (x, t )
t
3. Application to NavierStokes Equations
This chapter is devoted to the derivation of the constitutive equations of the

large-eddy simulation technique, which is to say the ltered NavierStokes
equations. Our interest here is in the case of an incompressible viscous Newtonian uid of uniform density and temperature.
We rst describe the application of an isotropic spatial lter1 to the
NavierStokes equations expressed in Cartesian coordinates or in general coordinates.
The emphasis will be put on the Eulerian, velocitypressure formulation.2
An important point is that these two formulations lead to dierent commutation errors with the ltering operator, and thus yield dierent theoretical and
practical problems. The main point is that, when curvilinear grids are considered, two possibilities arise for solving numerically the ltered governing
equations:
Conventional Approach: First the lter is applied to the NavierStokes
equations written in Cartesian coordinates, and then the ltered equations
are transformed in general coordinates. Here, the lter is applied in the
physical space, and the lter kernels are developed within the usual Cartesian framework (see Chap. 2).
Alternate Approach: First the NavierStokes equations are expressed in
general coordinates, and then the lter is applied to the transformed equations. In this case, the transformed variables are ltered using uniform lter
kernels, leading to vanishing commutation errors. If physical variables are
ltered using a transformed lter kernel, some commutation errors appear
and specic lters must be employed (see Chap. 2).
The dierences originate from the fact that the transformation in general
coordinates is a nonlinear operation, yielding dierent commutation errors
between the two operations.
1
2
Refer to the denition given in Chap. 2.

A few works dealing with the velocityvorticity form of the NavierStokes equations exist [163, 485, 486]. It is important to note that the Lagrangian framework
is employed in [485, 486]. Results dealing with large-eddy simulation within the
framework of lattice-Boltzmann methods are presented in the review by Chen
and Doolen [127].
46
Most of the existing published works deal with the conventional approach.
As a consequence, this chapter will be mostly devoted to this approach.
It should be noted that this ideal framework, which implies that the uid
domain is unbounded, is the one nearly all authors use because it is only
in this framework that the theory on which the subgrid modeling is based
can be fully developed. The commutation errors between the lter and the
derivation in space are then ignored. Section 3.4 is on the application of an
inhomogeneous lter to the basic equations written in Cartesian coordinates.
We begin by deriving the ltered NavierStokes equations following the
conventional approach. The various decompositions of the nonlinear term as
a function of the ltered quantities are then discussed. We lastly introduce
the closure problem, i.e. the representation of the unknowns as a function of
the variables in the ltered problem.
3.1 NavierStokes Equations

We recall here the equations governing the evolution of an incompressible
Newtonian uid, rst in the physical space, in general coordinates, and then
in the spectral space.
3.1.1 Formulation in Physical Space
In the physical space, the velocity eld u = (u1 , u2 , u3 ) expressed in a reference Cartesian coordinate system x = (x1 , x2 , x3 ) is a solution of the system
comprising the momentum and continuity equations:

ui
ui
uj
+
(ui uj ) =
+
+
,
i = 1, 2, 3 , (3.1)
t
xj
xi
xj xj
xi
ui
=0 ,
xi
(3.2)
in which p = P/ and are, respectively, the static pressure and the assumedly constant, uniform kinematic viscosity. To obtain a well-posed problem, we have to add initial and boundary conditions to this system.
3.1.2 Formulation in General Coordinates
The incompressible NavierStokes equations written in general coordinates
in strong conservation-law form [733] read:
(J 1 ik ui ) = 0 ,
k
(3.3)
3.1 NavierStokes Equations
47
Fig. 3.1. Schematic of the coordinate transformation.
1
(J ui ) + k (U k ui ) = k (J 1 ik p) + k
t

ui
J 1 Gkl l ,
(3.4)
where k are the coordinate directions in the transformed space, ik =

k /xi , J 1 is the Jacobian of the transformation, Gkl = ik il denotes
the contravariant metric tensor, ui the Cartesian components of the velocity
eld, and U k = J 1 ik ui the contravariant velocity component (see Fig. 3.1).
3.1.3 Formulation in Spectral Space
The dual system in spectral space is obtained by applying a Fourier transform
to equations (3.1) and (3.2). By making use of the fact that the incompressibility constraint is reected geometrically by the orthogonality3 of the wave
(k) , dened as (see Appendix A for greater detail
vector k and of the mode u
on the spectral analysis of turbulence):

1
(k) =
u
(3.5)
u(x)ekx d3 x, 2 = 1 ,
(2)3
the system (3.1) - (3.2) can be reduced to a single equation:

+ k 2 u
i (k) = Ti (k) ,
t
in which the non-linear term Ti (k) is of the form:

u
j (p)
um (q)(k p q)d3 pd3 q
Ti (k) = Mijm (k)
3
(3.6)
(3.7)
This orthogonality relation is demonstrated by re-writing the incompressibility

constraint of the velocity eld in the spectral space as:
ui
(k) = 0
= 0 ki u
i (k) k u
xi
48
with:
Mijm (k) = (km Pij (k) + kj Pim (k)) ,

(3.8)
2
in which is the Kronecker symbol and Pij (k) is the projection operator on
the plane orthogonal to the vector k. This operator is written:
Pij (k) =

ki kj
ij 2
k
(3.9)
3.2 Filtered NavierStokes Equations in Cartesian

Coordinates (Homogeneous Case)
This section describes the equations of large-eddy simulation such as they
are obtained by applying a homogeneous lter verifying the properties of
linearity, conservation of constants, and commutation with derivation, to the
NavierStokes equations. These are the equations that will be resolved in the
numerical simulation.
3.2.1 Formulation in Physical Space
In light of the commutation with derivation property, the application of a lter to equations (3.1) and (3.2) is expressed:
ui
+
(ui uj ) =
+
t
xj
xi
xj
ui
=0 ,
xi
ui
uj
+
xj
xi

,
(3.10)
(3.11)
where p is the ltered pressure. The ltered momentum equation brings out
the non-linear term ui uj which, in order for this equation to be usable, will
have to be expressed as a function of u and u , which are now the only
unknowns left in the problem and where:
u = u u
(3.12)
This decomposition is not unique, and will be discussed in the following

section.
3.2.2 Formulation in Spectral Space
ui (k), the momentum equation in the
Using the equivalence u
i (k) = G(k)
spectral space obtained by multiplying equation (3.6) by the transfer function
3.3 Decomposition of the Non-linear Term. Conventional Approach

G(k)
is expressed:

ui (k) = G(k)T

+ 2k 2 G(k)
i (k) ,
t

in which the ltered non-linear term G(k)T
i (k) is written:

uj (p)

G(k)
G(k)T
i (k) = Mijm (k)
49
(3.13)
. (3.14)
The ltered non-linear term (3.14) brings out the contributions of the
(p) and u
(q). To complete the decomposition, these modes also
modes u
have to be expressed as the sum of a ltered part and a uctuation. This is
the same problem as the one encountered when writing the equations in the
physical space. This operation is described in the following section.
3.3 Decomposition of the Non-linear Term.

Associated Equations for the Conventional Approach
This section details the various existing decompositions of the non-linear term
and the associated equations written in Cartesian coordinates.
3.3.1 Leonards Decomposition
Expression in Physical Space. Leonard [436] expresses the non-linear
term in the form of a triple summation:
ui uj
(ui + ui )(uj + uj )
ui uj
ui uj +
uj ui
(3.15)
+
ui uj
(3.16)
The non-linear term is now written entirely as a function of the ltered

quantity u and the uctuation u . We then have two versions of this [762].
The rst considers that all the terms appearing in the evolution equations
of a ltered quantity must themselves be ltered quantities, because the simulation solution has to be the same for all the terms. The ltered momentum
equation is then expressed:

ui

p
uj
ui
ij
+
ui uj =
+
+
,
(3.17)
t
xj
xi
xj xj
xi
xj
in which the subgrid tensor , grouping together all the terms that are not
exclusively dependent on the large scales, is dened as:
ij = Cij + Rij = ui uj ui uj
(3.18)
50
where the cross-stress tensor, C, which represents the interactions between

large and small scales, and the Reynolds subgrid tensor, R, which reects the
interactions between subgrid scales, are expressed:
Cij
ui uj + uj ui
Rij
ui uj
(3.19)
(3.20)
In the following, this decomposition will be called double decomposition.

The other point of view consists of considering that it must be possible
to evaluate the terms directly from the ltered variables. But the ui uj term
cannot be calculated directly because it requires a second application of the
lter. To remedy this, Leonard proposes a further decomposition:

ui uj = ui uj ui uj + ui uj
= Lij + ui uj .
(3.21)
The new L term, called Leonard tensor, represents interactions among the
large scales. Using this new decomposition, the ltered momentum equation
becomes:

ui
p
uj
ij
ui
(ui uj ) =
+
+
.
(3.22)
+
t
xj
xi
xj xj
xi
xj
The subgrid tensor , which now groups all the terms that are not expressed directly from u, takes the form:
ij = Lij + Cij + Rij = ui uj ui uj
(3.23)
This decomposition will be designated hereafter the Leonard or triple

decomposition. Equation (3.22) and the subgrid term ij dened by (3.23)
can be obtained directly from the NavierStokes equations without using the
Leonard decomposition for this. It should be noted that the term ui uj is
a quadratic term and that it contains frequencies that are in theory higher
than each of the terms composing. So in order to represent it completely,
more degrees of freedom are needed than for each of the terms ui and uj 4 .
We may point out that, if the lter is a Reynolds operator, then the
tensors Cij and Lij are identically zero5 and the two decompositions are
then equivalent, since the subgrid tensor is reduced to the tensor Rij .
4
In practice, if the large-eddy simulation lter is associated with a given computational grid on which the NavierStokes equations are resolved, this means that
the grid used for composing the ui uj product has to be twice as ne (in each
direction of space) as the one used to represent the velocity eld. If the product
is composed on the same grid, then only the ui uj term can be calculated.
It is recalled that if the lter is a Reynolds operator, then we have the three
following properties (see Appendix A):
u = u, u = 0, uu = u u
51
Writing the NavierStokes equations (3.1) in the symbolic form

u
+ N S(u) = 0
t
(3.24)
the ltered NavierStokes equations are expressed

G
u
u
=
t
t
G
N S(u)
N S(u) [G
, N S](u)
(3.25)
,
(3.26)
where [., .] is the commutator operator introduced in Sect. 2.1.2. We note

that the subgrid tensor corresponds to the commutation error between the
lter and the non-linear term. Introducing the bilinear form B(, ):
B(ui , uj ) ui uj
(3.27)
we get
ij = [G
, B](ui , uj ) .
(3.28)
Double decomposition (3.18) leads to the following equation for the resolved kinetic energy qr2 = ui ui /2:
qr2
t
ui ui
ui
ui
ui uj
+ ij
xj
xj
xj xj

qr2
(ui p) +
x
x
x
i
i
i
II
IV
III

ui ui uj
(ui ij )
xj
xj

VI
(3.29)
V II
This equation shows the existence of several mechanisms exchanging kinetic energy at the resolved scales:
whence
Cij
Lij
ui uj + uj ui
=
=
ui u j + uj u i
0 ,
=
=
=
ui uj ui uj
ui uj ui uj
0 .
52
I - production
II - subgrid dissipation
III - dissipation by viscous eects
IV - diusion by pressure eect
V - diusion by viscous eects
V I - diusion by interaction among resolved scales
V II - diusion by interaction with subgrid modes.
Leonards decomposition (3.23) can be used to obtain the similar form:
qr2
t
ui
qr2 uj
ui ui
+ ij
xj
xj
xj xj

qr2
(ui p) +
x
x
x
i
i
i
V III
XI
IX
XII
ui
ui uj
(ui ij )
xj xj

XIII
(3.30)
XIV
This equation diers from the previous one only in the rst and sixth
terms of the right-hand side, and in the denition of tensor :
V III - advection
IX - idem II
X - idem III
XI - idem IV
XII - idem V
XIII - production
XIV - idem V II
The momentum equation for the small scales is obtained by subtracting the large scale equation from the unltered momentum equation (3.1),
making, for the double decomposition:

ij
p
ui
+
+
(ui + ui )(uj + uj ) ui uj =
t
xj
xi
xi

uj
ui
+
+
,
xj xj
xi
(3.31)
and, for the triple decomposition:

ij
ui

p
+
+
(ui + ui )(uj + uj ) ui uj =
t
xj
xi
xi

uj
ui
+
+
.
xj xj
xi
(3.32)
53
2
The ltered subgrid kinetic energy qsgs
= uk uk /2 equation obtained by

multiplying (3.32) by ui and ltering the relation thus derived is expressed:
2
qsgs
t
1
2
(ui ui uj uj ui ui )
(puj p uj )
qsgs uj
xj
2 xj
xj

XV
XV I
XV II

2
qsgs
+
(ij ui )
xj
xj
xj

XIX
XV III

ui ui
ui
ui ui
ij
xj xj
xj xj
xj

XX
(3.33)
XXI
XV - advection
XV I - turbulent transport
XV II - diusion by pressure eects
XV III - diusion by viscous eects
XIX - diusion by subgrid modes
XX - dissipation by viscous eects
XXI - subgrid dissipation.
For the double decomposition, equation (3.31) leads to:
2
qsgs
t

ui
ui
ui ui uj ui ui uj + ui uj
ui uj
xj
xj
xj

XXII
XXIII

2 ui
2 ui

puj p uj + ui 2 ui 2
xj
xj
xj

XXIV

ui

ij ui ij
,
xj
xj

XXV I
with:
XXII - turbulent transport

XXIII - production
XXIV - diusion by pressure eects
XXV - viscous eects
XXV I - subgrid dissipation and diusion
XXV
(3.34)
54
It is recalled that, if the lter used is not positive, the generalized subgrid
2
kinetic energy qgsgs
dened as the half-trace of the subgrid tensor,
1
u u + ui ui ,
2 i i
can admit negative values locally (see Sect. 3.3.5). If the lter is a Reynolds
operator, the subgrid tensor is then reduced to the subgrid Reynolds tensor
and the generalized subgrid kinetic energy is equal to the subgrid kinetic
energy, i.e.
1
2
2
qsgs
ui ui = qgsgs
kk /2 .
(3.35)
2
Expression in Spectral Space. Both versions of the Leonard decomposition can be transcribed in the spectral space. Using the denition of the
(k) as
uctuation u

ui (k) ,
(3.36)
u
i (k) = (1 G(k))
2
= kk /2 =
qgsgs

the ltered non-linear term G(k)T
i (k) is expressed, for the triple decomposition:

G(k)T
i (k) =

Mijm (k)

G(q)
uj (p)
G(p)

G(q)

(1 G(k))
G(p)
Mijm (k)

uj (p)

G(k)
G(p)(1
G(q))
+ G(q)(1
G(p))
Mijm (k)

uj (p)

G(k)
(1 G(q))(1
G(p))
Mijm (k)

uj (p)
(3.37)
The rst term of the right-hand side corresponds to the contribution ui uj ,

the second to the Leonard tensor L, the third to the cross stresses represented
by the tensor C, and the fourth to the subgrid Reynolds tensor R. This is
illustrated by Fig. 3.2.
The double decomposition is derived by combination of the rst two terms
of the right-hand side of (3.37):

G(k)T
i (k) =

Mijm (k)
G(q)
G(k)

G(p)

uj (p)
55
Fig. 3.2. Representation of the various Leonard decomposition terms in the spectral space, when using a sharp cuto lter with a cuto frequency kc .

G(k)
G(p)(1
G(q))
+ G(q)(1
G(p))
Mijm (k)

uj (p)

G(k)
(1 G(q))(1
G(p))
Mijm (k)

uj (p)
(3.38)
The rst term of the right-hand side corresponds to the contribution ui uj

in the physical space, and the last two remain unchanged with respect to the
triple decomposition. Let us note that the sum of the contributions of the
cross tensor and the subgrid Reynolds tensor simplies to the form:

G(q))

Cij + Rij = Mijm (k)
(1 G(p)
G(k)

uj (p)
(3.39)
The momentum equations corresponding to these two decompositions are

found by replacing the right-hand side of equation (3.13) with the desired
terms. For the double decomposition, we get:
56
ui (k)
+ k 2 G(k)
t

= Mijm (k)
G(q)
G(k)

G(p)

uj (p)

G(q))

+ Mijm (k)
(1 G(p)
G(k)

uj (p)
,
(3.40)
and for the triple decomposition:

G(q)

ui (k) = Mijm (k)
G(p)
+ k 2 G(k)
t

uj (p)

G(q)

(1 G(k))
G(p)
Mijm (k)

uj (p)

G(k)
G(p)(1
G(q))
+ Mijm (k)

+G(q)(1
G(p))

uj (p)

G(k)
(1 G(q))(1
G(p))
+ Mijm (k)

uj (p)
(3.41)
For both decompositions, the momentum equation can be expressed in

the symbolic form:

2
ui (k) = Tr (k) + Tsgs (k) ,
+ k G(k)
(3.42)
t
in which Tr (k) designates the transfer terms calculated directly from the
resolved modes, and is therefore equivalent to the contribution of the ui uj
term in the case of the triple decomposition, and that of the ui uj term for
the double decomposition. The Tsgs (k) term designates the other non-linear
terms, and therefore corresponds to the contribution of the subsidiary term
such as dened above. Let E(k) be the energy contained on the sphere of
radius k. It is calculated as:

1 2
(k) u
(k)dS(k) ,
u
(3.43)
E(k) = k
2
where dS(k) is the surface element of the sphere, and where the asterisk designates a conjugate complex number. The kinetic energy of the resolved modes
contained on this same sphere, denoted E r (k), is dened by the relation
57
E r (k)
1 2
u(k) G(k)
u (k)dS(k)
G(k)
k
2
2 (k)E(k) .
= G
=
(3.44)
(3.45)
We return to the kinetic energy of the resolved modes, qr2 = ui ui /2, by

summation on all the wave numbers:

E r (k)dk .
(3.46)
qr2 =
0
It is important to note that E r (k) is related to the energy of the resolved

modes, which is generally not equal to the ltered part of the kinetic energy
which, for its part, is associated with the quantity denoted E(k), dened as

E(k) = G(k)E(k)
(3.47)
The identity of these two quantities is veried when the transfer func2 (k) = G(k),

tion is such that G
k, i.e. when the lter used is a projector.
The evolution equation for E r (k) is obtained by multiplying the ltered mo u (k), and then integrating the result on
mentum equation (3.13) by k 2 G(k)
the sphere of radius k. Using the double decomposition we get the following
equation:

1
2
G(q)
G
2 (k)S(k|p, q)dpdq
G(p)
+ 2k E r (k) =
t
2

1
G(q))

2 (k)S(k|p, q)dpdq ,
+
(1 G(p)
G
2
(3.48)
and the triple decomposition:

1
2
G(q)
G(k)S(k|p,

G(p)
q)dpdq
+ 2k E r (k) =
t
2

1

G(q)S(k|p,

G(k)(1
G(k))
G(p)
q)dpdq
2

1
2 (k) G(p)

+
G
2

(1 G(q))
+ G(q)(1
G(p))
S(k|p, q)dpdq

1
2 (k) (1 G(q))(1

G
+
G(p))
2
S(k|p, q)dpdq
in which
(3.49)
58
S(k|p, q) = 16 2 kpqMijm (k)

uj (p)
um (q)
ui (k)(k p q) ,
(3.50)
!!
and where the symbol
designates integration over the interval
|k p| < q < k + p.
Following the example of what was done for the momentum equations,
the kinetic energy evolution equation for the resolved modes can be expressed
in the abbreviated form

2
e
+ 2k E r (k) = Tre (k) + Tsgs
(k) .
(3.51)
t
e
(k) represent, respectively, the energy transfers
The terms Tre (k) and Tsgs
of mode k with all the other modes associated with the terms that can be
calculated directly from the resolved modes, and the subgrid terms.The kinetic energy conservation property for inviscid uids, i.e. in the case of zero
viscosity, implies:

e
(k))d3 k = 0 .
(Tre (k) + Tsgs
(3.52)
The momentum equations for the unresolved scales are obtained by algebraic manipulations strictly analogous to those used for obtaining the equations for the resolved scales, except that this time equation (3.6) is multiplied

by (1 G(k))
instead of G(k).
These equations are written:

ui (k) = (1 G(k))T
(3.53)
+ k 2 (1 G(k))
i (k) .
t

Calculations similar to those explained above lead to:

G(q)(1

G(p)
G(k))
+ k 2 u
i (k) = Mijm (k)
t

uj (p)

(1 G(k))
G(p)(1
G(q))
Mijm (k)

+G(q)(1
G(p))

uj (p)

(1 G(k))
(1 G(q))(1
G(p))
Mijm (k)

uj (p)
(3.54)
The rst term of the right-hand side represents the contribution of the
interactions between large scale modes, the second the contribution of the
cross interactions, and the last the interactions among the subgrid modes.
59
Let E sgs be the energy contained in the subgrid modes. This energy is
dened as:

1 2

k
E sgs (k) =
(1 G(k))
u(k) (1 G(k))
u (k)dS(k) (3.55)
2
=
2 (k)E(k)
(1 G)
(3.56)
and is dierent, in the general case, from the kinetic energy uctuation

though the equality of these two quantities is verE (k) = (1 G)(k)E(k),
ied when the lter is a Reynolds operator. Simple calculations give us the
following evolution equation for E sgs (k):

1
2
2
G(q)(1

G(p)
G(k))
+ 2k E sgs (k) =
S(k|p, q)dpdq
t
2

1

+
(1 G(k))
G(p)(1
G(q))
2

+G(q)(1
G(p))
S(k|p, q)dpdq

1
2

+
(1 G(k))
(1 G(q))(1
G(p))
2
S(k|p, q)dpdq ,
(3.57)
where the notation used is the same as for the kinetic energy evolution equa2
is obtained by
tion of the resolved modes. The subgrid kinetic energy qsgs
summation over the entire spectrum:

2
=
E sgs (k)dk .
(3.58)
qsgs
0
3.3.2 Germano Consistent Decomposition

This section presents the Germano consistent decomposition, which is a generalization of the Leonard decomposition.
Denition and Properties of Generalized Central Moments. For convenience, we use []G to denote the resolved part of the eld , dened as in
the rst chapter, where G is the convolution kernel, i.e.:
+
G(x )()d3 .
(3.59)
[]G (x) G
(x)
We dene the generalized central moments with the lter G, denoted G ,

as [244, 246, 247, 248]:
G (1 , 2 ) =
G (1 , 2 , 3 ) =
G (1 , 2 , 3 , 4 ) =
[1 2 ]G [1 ]G [2 ]G
(3.60)
[1 2 3 ]G [1 ]G G (2 , 3 ) [2 ]G G (1 , 3 )
(3.61)
[3 ]G G (1 , 2 ) [1 ]G [2 ]G [3 ]G ,
...
(3.62)
60
The generalized central moments thus dened verify the following properties:
G (, )
G (, a) =
G (, , a) =
G (, )/s =
G (, )
0,
(3.63)
for a = const.
(3.64)
0,
for a = const. ,
G (/s, ) + G (, /s), s = x, t
(3.65)
. (3.66)
If we perform the decomposition = 1 + 2 , = 1 + 2 , we get:

G (1 + 2 , 1 + 2 ) =
G (1 , 1 ) + G (1 , 2 )
G (2 , 1 ) + G (2 , 2 )
(3.67)
The generalized central moments also appear as the coecients of the

following formal Taylor expansion [247]:
[(a1 , ..., an )]G
= ([a1 ]G , ..., [an ]G ) +
G (al , am )
2!
l,m
ylm
G (al , am , ak )
ylmk + ... ,
3!
(3.68)
l,m,k
with
ylm =
2 ([a1 ]G , ..., [an ]G )

,
[al ]G [am ]G
ylmk =
3 ([a1 ]G , ..., [an ]G )

[al ]G [am ]G [ak ]G
and where the ai are generic turbulent quantities. The relation (3.68) establishes a link between the ltered value of the functional and its unltered
counterpart applied to the ltered variables [ai ]G .
Consistent Decomposition: Associated Equations. By applying the
property (3.67) to the decomposition = []G + , = []G + , we get:
G ([]G + , []G + ) = G ([]G , []G ) + G ( , []G )
+G ([]G , ) + G ( , )
(3.69)
This decomposition is said to be consistent because it is consistent with

the denition of the generalized central moments, ensuring that all the terms
in it are of the same form, which is not true of the Leonard decomposition.
The various terms of the right-hand side of equation (3.69) can be interpreted
as generalizations of the terms of the Leonard triple decomposition. By applying this denition to the components of the velocity elds, the subgrid
tensor (3.23) appears in a double form:
G (ui , uj )
= [ui uj ]G [ui ]G [uj ]G

= Lij + Cij + Rij
= Lij + Cij + Rij
(3.70)
61
in which the tensors L, C and R are dened as:

Lij
Cij
=
=
G ([ui ]G , [uj ]G ) ,
G ([ui ]G , uj ) + G (ui , [uj ]G ) ,
(3.71)
(3.72)
Rij
G (ui , uj )
(3.73)
and represent, respectively, the interactions between the large scales, the
cross interactions, and the interactions among subgrid scales. They therefore
represent tensors dened by Leonard, but are not the same as them in the
general case.
By bringing out the generalized central moments, the ltered momentum
equations are written in the form:

[p]G
[uj ]G
[ui ]G
[ui ]G
+
([ui ]G [uj ]G ) =
+
+
t
xj
xi
xj
xj
xi
G (ui , uj )
.
(3.74)
xj
This equation is equivalent to the one derived from the triple Leonard
decomposition. Similarly, the subgrid kinetic energy evolution equation (3.33)
is re-written as:

2
2
qsgs
qsgs
1
=
G (ui , ui , uj ) + G (p, uj )
t
xj 2
xj
G (ui /xj , ui /xj ) G (ui , uj )
[ui ]G
xj
(3.75)
It is easy to check that the structure of the ltered equations is, in terms
of generalized central moments, independent of the lter used. This is called
the ltering (or averaging) invariance property.
3.3.3 Germano Identity
Basic Germano Identity. Subgrid tensors corresponding to two dierent
ltering levels can be related by an exact relation derived by Germano [246].
A sequential application of two lters, F and G, is denoted:
[ui ]FG = [[ui ]F ]G = [[ui ]G ]F
(3.76)
or equivalently:

[ui ]FG (x) =
G(x y)d y
F (y )ui ()d3
(3.77)
62
Fig. 3.3. Illustration of the two ltering levels, F and G, involved in the Germano
identity. Associated cuto wave numbers are denoted kF and kG , respectively. Resolved velocity elds are uF and uG , and the associated subgrid tensors are F and
G , respectively.
Here, [ui ]FG corresponds to the resolved eld for the double ltering F G.
The two ltering levels are illustrated in Fig. 3.3.
The subgrid tensor associated with the level F G is dened as the following
generalized central moment:
FG (ui , uj ) = [ui uj ]FG [ui ]FG [uj ]FG
(3.78)
This expression is a trivial extension of the denition of the subgrid tensor associated with the G ltering level. By denition, the subgrid tensor
G ([ui ]F , [uj ]F ) calculated from the scales resolved for the F ltering level, is
written:
(3.79)
G ([ui ]F , [uj ]F ) = [[ui ]F [uj ]F ]G [ui ]FG [uj ]FG .
These two subgrid tensors are related by the following exact relation,
called the Germano identity:
FG (ui , uj ) = [F (ui , uj )]G + G ([ui ]F , [uj ]F ) .
(3.80)
This relation can be interpreted physically as follows. The subgrid tensor

at the F G ltering level is equal to the sum of the subgrid tensor at the F
level ltered at the G level and the subgrid tensor at the G level calculated
from the eld resolved at the F level. This relation is local in space and time
and is independent of the lter used.
It is interesting noting that re-writing the subgrid tensor as
F G (ui , uj ) = [F
G
, B](ui , uj ) ,
63
where [., .] is the commutator operator (see equation (2.13)) and B(., .) the
bilinear form dened by relation (3.27), the Germano identity (3.80) is strictly
equivalent to relation (2.16):
[F
G
, B](ui , uj ) = [F
, B] (G
)(ui , uj ) + (F
) [G
, B](ui , uj ). (3.81)
The previous Germano identity can be referred to as the multiplicative
Germano identity [251], because it is based on a sequential application of
the two lters. An additive Germano identity can also be dened considering
that the second ltering level is dened by the operator (F + G)
and not by
the operator (G
) (F
). The equivalent relation for (3.80) is
F+G (ui , uj ) = F (ui , uj ) + G (ui , uj ) ([ui ]F [uj ]G + [ui ]G [uj ]F ) . (3.82)
Multilevel Germano Identity. The Multiplicative Germano Identity can
be extended to the case of N ltering levels, Gi , i = 1, N , with associated
characteristic lengths 1 2 ... N [248, 710, 633].
n
We dene the nth level ltered variable as
n
= Gn
Gn1
...
G1
= G1n
,
(3.83)
with
n
Gm
Gn
Gn1
...
Gm , Gnn = Id,
m [1, n] .
(3.84)
Let ijn = ui uj n uni unj be the subgrid tensor associated to the nth ltering
level. The classical two-level Germano identity (3.80) reads
ijn+1 = ijn
n+1
+ Ln+1
ij ,
Ln+1
= uni unj
ij
n+1
un+1
un+1
i
j
(3.85)
Simple algebraic developments lead to the following relation between two

ltering levels n and m, with m < n:

n
n
ijn = Lnij +
Gk+1
Lkij + Gm+1
ijm ,
(3.86)
k=m+1,n1
resulting in a fully general mutilevel identity.

Generalized Germano Identity. A more general multiplicative identity is obtained by applying an arbitrary operator L to the basic identity (3.81) [629], yielding
L{[F
G
, B](ui , uj )]}
= L{[F
, B] (G
)(ui , uj )
+ (F
) [G
, B](ui , uj )} .
(3.87)
For linear operators, we get

L{[F
G
, B](ui , uj )]} =
+
L{[F
, B] (G
)(ui , uj )}
L{(F
) [G
, B](ui , uj )}
Application to the multilevel identity (3.86) is straightforward.
(3.88)
64
3.3.4 Invariance Properties

One of the basic principles of modeling in mechanics is to conserve the generic
properties of the starting equations [681, 230, 260, 572, 326, 398].
We consider in the present section the analysis of some invariance/symmetry properties of the ltered NavierStokes equations, and the resulting constraints for subgrid models. It is remembered that a dierential equation will
be said to be invariant under a transformation if it is left unchanged by this
transformation. It is important to note that these properties are not shared
by the boundary conditions. It is shown that properties of the ltered Navier
Stokes equations depend on the lter used to operate the scale separation.
The preservation of the symmetry properties of the original NavierStokes
equations will then lead to the denition of specic requirements for the lter
kernel6 G(x, ). The properties considered below are:
Galilean invariance for the spatial ltering approach (p. 64).
Galilean invariance for the time-domain ltering approach (p. 66).
General frame-invariance properties for the time-domain ltering approach
(p. 67).
Time invariance for the spatial ltering approach (p. 69).
Rotation invariance for the spatial ltering approach (p. 70).
Reection invariance for the spatial ltering approach (p. 70).
Asymptotic Material Frame Indierence for the spatial ltering approach
(p. 71).
Galilean Invariance for Spatial lter. This section is devoted to the analysis by Speziale [681] of the preservation of the Galilean invariance property
for translations of the NavierStokes equations, rst by applying a spatial
lter, then by using the dierent decompositions presented above.
Let us take the Galilean transformation (translation):
x = x + V t + b,
t = t ,
(3.89)
in which V and b are arbitrary uniform vectors in space and constant in time.
If the (x, t) frame of reference is associated with an inertial frame, then so
is (x , t ). Let u and u be the velocity vectors expressed in the base frame
of reference and the new translated one, respectively. The passage from one
system to the other is dened by the relations:
u
xi
t
6
=
=
=
u+V ,
,
xi
Vi
t
xi
(3.90)
(3.91)
.
(3.92)
We will only consider lters with constant and uniform cuto length, i.e. is
independent on both space and time. Variable length lters are anisotropic or
nonhomogeneous, and violate the following properties in the most general case.
65
The proof of the invariance of the NavierStokes equations for the transformation (3.89) is trivial and is not reproduced here. With this property in
hand, what remains to be shown in order to prove the invariance of the ltered
equations by such a transformation is that the ltering process preserves this
property.
Let there be a variable such that
= .
(3.93)
The ltering in the translated frame of reference is expressed:

= G(x x ) (x )d3 x .
(3.94)
By using the previous relations, we get:

x x = (x + V t + b) (x + V t + b) = x x

x

d3 x = i d3 x = d3 x ,
xj
and thus, by substitution, the equality:

= G(x x )(x )d3 x = ,
(3.95)
(3.96)
(3.97)
which completes the proof7 . The invariance of the NavierStokes equations

for the transformation (3.89) implies that the sum of the subgrid terms and
the convection term, calculated directly from the large scales, is also invariant,
but not that each term taken individually is invariant. In the following, we
study the properties of each term arising from the Leonard and Germano
decompositions.
The above relations imply:
u = u + V , u = u , u = u
(3.98)
which reects the fact that the velocity uctuations are invariant by Galilean
transformation, while the total velocity is not. In the spectral space, this
corresponds to the fact that only the constant mode does not remain invariant
by this type of transformation since, with the V eld being uniform, it alone
is aected by the change of coordinate system8 .
7
8
A sucient condition is that the lter kernel appears as a function of x x .

This is expressed:
V = cste = V (k) = 0 k = 0 ,
and thus
(k) = u
(k) k = 0 ,
u

(0) + V (0) .
u (0) = u
66
In the translated frame, the Leonard tensor takes the form:

Lij
= ui uj ui uj

= Lij + Vi uj + Vj ui (Vi uj + Vj ui )

.
= Lij Vi uj + Vj ui
(3.99)
(3.100)
(3.101)
So this tensor is not invariant. Similar analyses show that:

Cij
,
= Cij + Vi uj + Vj ui
Rij
Lij + Cij
=
=
Rij ,
Lij + Cij
(3.102)
(3.103)
(3.104)
The tensor C is thus not invariant in the general case, while the tensor R
and the groups L+C and L+C+R are. A dierence can be seen to appear here
between the double and triple decompositions: the double retains groups of
terms (subgrid tensor and terms computed directly) that are not individually
invariant, while the groups in the triple decomposition are.
The generalized central moments are invariant by construction. That is,
by combining relations (3.67) and (3.63), we immediately get:
(ui , uj ) = G (ui , uj ) .
G
(3.105)
This property results in all the terms in Germanos consistent decomposition being invariant by Galilean transformation, which is all the more true
for the tensors L, C and R.
Galilean Invariance and Doppler Eect for Time-Domain Filters.
Pruett [603] extended the above analysis to the case of Eulerian
time-domain ltering. Using the properties of the Eulerian time-domain
ltering and the fact that NavierStokes equations are form-invariant under Galilean transformations, one can easily prove that time-domain ltered
NavierStokes equations are also form-invariant under these transformations:

uj
ui

ui
p
+ ui uj = +
+
,
(3.106)
t
xj
xi
xj xj
xi
ui
=0 .
xi
(3.107)
It was shown in the preceding section that the spatially ltered part of
a Galilean-invariant function is itself Galilean invariant, i.e.
u = u + V = u = u + V ,
p = p = p = p
(3.108)
A fundamental dierence between spatial- and time-domain ltering is

that what applies to the former does not apply to the latter. Writing the
67
denition of the ltered velocity in the translated frame, we have:

ui
ui (t , x V t ) + Vi
t
=
ui (t , x V t )G(t t )dt + Vi

t
=

ui (t , x V t )G(t t)dt + Vi

ui (t , x V t)G(t t)dt + Vi
ui (t , x)G(t t)dt + Vi
ui (x, t) + Vi
(3.109)
It is seen from these equations that Eulerian temporally ltered quantities

experience a Doppler shift in the direction of the translational velocity V ,
and that equalities presented in (3.108) are recovered only when V = 0.
The equations remain invariant under Galilean transformations because each
term is subjected to the same shift.
This Doppler shift results in a wave number-dependent frequency shift
between the two frames [603], indicating that Eulerian time-domain ltering
may be inadequate for ows in which structures are convected at very dierent
characteristic velocities, such as boundary layers [289]. On the contrary, free
shear ows (mixing layers, jets, and wakes) seem to be better adapted.
General Investigation of Frame-Invariance Properties of TimeDomain Filtered NavierStokes Equations. Previous analysis of Galilean invariance properties of Eulerian time-domain lter was extended to the
more general case of Euclidean group of transformations by Pruett et al. [606].
In the case the observer is xed in the inertial frame (referred to as case I
below), the spatial coordinates in the noninertial frame vary with time, while
those in the inertial frame are xed. Thus, the general change of reference
frame is expressed as
(3.110)
xi (t ) = Qij [xi + Vi ] ,
where xi is refers to the coordinates of a point in a frame of reference in
arbitrary time-dependent motion (rotation and translation) relative to an
inertial frame tied to coordinates xi , and Q = Q(t) is a time-dependent
orthogonal tensor. The vector V is also time-dependent, i.e. V = V (t). The
time in the new reference frame is obtained considering a shift t0 : t = t + t0 .
The velocity in the moving frame is
"
#
.
(3.111)
ui (t , x ) = Q ij [xi + Vi ] + Qij ui + V i
The ltering cuto being frame-invariant, it will not be explicited in
the following.
68
In the opposite case (case II) where the observer is xed in the noninertial
frame, one obtains
xi (t) = Qji xj Vi
(3.112)
ui (t, x) = Q ji xj + Qji uj V i
(3.113)
The spatial coordinates in the inertial frame are now time-dependent, and
those in the noninertial frame are xed.
These two dierent cases must be treated separately when analyzing the
properties of the time-letered NavierStokes equations.
Application of the Eulerian time lter in case I yields
xi (t ) = Qij xi + Qij Vi
(3.114)
"
#
.
ui (t , x ) = Q ij [xi + Vi ] + Qij ui + V i
(3.115)
The subgrid velocity eld u = u u is then equal to

u
i = Qik uk Qik uk + Qik Qik xk + (Qik bk ) (Qik bk )
. (3.116)
In case II, the ltered and subgrid quantities are expressed as follows
xi (t) = Qji xj V i
ui (t, x) = Q ji xj + Qji uj V i
(3.117)

ui = Qki uk Qki uk + Q kj Q kj xk b i b i
(3.118)
(3.119)
A look at equations (3.115) and (3.118) show that the velocity is not frame
invariant under general Euclidean transformations. The same conclusion apply for the subgrid velocity eld. A noticeable dierence between time- and
space-ltering is that, because Q(t) is a time-dependent parameter, ltered
and unltered velocity elds do not transform in the same manner in the
time-ltering approach.
The NavierStokes equations are known to be not frame-invariant under
the Euclidean group of transformation, and can be expressed as
ui
u
P
2 u
+ uk i = + i + 2ik uk + ik xk
t
xk
xi
xk xk
(3.120)
69
where the modied pressure P and the rotation rate tensor are dened as
1
P = p + kl ln xn xk Qnk Vk xn
2
ik Q il Qkl
(3.121)
(3.122)
with Vk Qkn Vn .
The NavierStokes equations under the Euclidean transformation group
for an observer xed in the noninertial reference frame are obtained by rst
taking the material derivative of (3.111) and applying the lter, leading to
Dui
Duj
= ( ik il lk )xk + 2ik uk + Qij Vj + Qij
Dt
Dt
(3.123)
and then inserting the following expression deduced from the NavierStokes
equations written in the inertial frame:
Qij
Duj
p
2 u
= + i
Dt
xi
xk xk
(3.124)
yielding
2
ui
u i = P + ui
+
ik xk ik
+
u
+
2
u
ik
k
k
t
xk
xi
xk xk
xk
, (3.125)
where the ltered pressure P and the subgrid scale tensor are dened as
1
P = p + kl ln xn xk Vk xk
2
= ui uk u i uk
ik
(3.126)
(3.127)
A comparison of equations (3.125) and (3.120) shows that the time-ltered

NavierStokes equations do not retain the same form in the most general case,
to the contrary of the spatially ltered ones. The dierences appear in the
Coriolis terms, the centrifugal and the rotational acceleration terms.
Time Invariance (Spatial Filters). A time shift of the amount t0 yields
the following change of coordinates:
t = t + t 0 ,
x = x,
u = u
(3.128)
Since we are considering space dependent lters only, the ltered Navier
Stokes equations are automatically time-invariant, without any restriction on
70
the lter kernel. We have:

u = u,
u = u
(3.129)
and
ik
ik
Lik
Rik
=
=
Lik ,
Rik ,
(3.131)
(3.132)
Cik
Cik
(3.133)
(3.130)
All the subgrid terms are invariant.

Rotation Invariance (Spatial Filters). We now consider the following
change of reference system:
t = t,
x = Ax,
u = Au
(3.134)
where A is the rotation matrix with AT A = AAT = Id and |A| = 1. Simple

calculations similar to those shown for in the section devoted to Galilean
invariance lead to the following relations:
u = Au,
u = Au
(3.135)
if and only if the lter kernel G(x, ) satises

G(A(x )) = G(x ) = G(x, ) = G(|x |) ,
(3.136)
meaning that the lter must be spherically symmetric. The subgrid terms are
transformed as:
ik
Lik
= Aim Akn mn ,
= Aim Akn Lmn ,
(3.137)
(3.138)
Rik
Cik
= Aim Akn Rmn

= Aim Akn Cmn
(3.139)
(3.140)
,
,
and are seen to be invariant.

Reection Invariance (Spatial Filters). We now consider a reection in
the lth direction of space:
t = t; xl = xl ; xi = xi , i = l; ul = ul ; ui = ui , i = l
(3.141)
If the lter is such that G(x ) = G(x + ), i.e. is symmetric, then

ul = ul ; ui = ui , i = l; u
l = u l ; ui = u i , i = l
(3.142)
71
yielding
ik
Lik
=
=
ik ,
Lik ,
(3.143)
(3.144)
Rik
Cik
=
=
Rik
Cik
(3.145)
(3.146)
,
,
with = 1 if i = l or k = l and i = l, and = 1 otherwise. We can see

that the subgrid tensor and all the terms appearing in both the double and
triple decomposition are invariant.
Asymptotic Material Frame Indierence (Spatial Filters). The last
symmetry considered in the present section is the asymptotic material frame
indierence, which is a generalization of the preceding cases. The change of
frame is expressed as:
t = t, x = A(t)x + c(t), u = Au + d(t), d(t) = c + Ax
(3.147)
where the rotation matrix A is such that AT A = AAT = Id , |A| = 1 and

c(t) is a vector. The NavierStokes equations are not form-invariant under
this group of transformation in the general case. Form invariance is recovered
in the asymptotic limit of two-dimensional ows.
The resulting changes of the subgrid and resolved velocity eld are:
u = Au + d,
u = Au
(3.148)
yielding
with
ik
Aim mn Akn
Lik
Aim mn Lkn Bik
(3.150)
Cik
Aim mn Ckn + Bik
(3.151)
Rik
Aim Rmn Akn
Bij = ui dj + uj di
(3.149)
(3.152)
These properties are subjected to the condition G(x, ) = G(|x |). We

can see that the properties of the subgrid tensors are the same as in the case
of the Galilean invariance case.
Table 3.1 summarizes the results dealing with the symmetry properties.
72
Table 3.1. Invariance properties of spatial convolution lters and subgrid tensors.
Symmetry
G(x, )
L+C
Galilean translation
Time shift
Rotation
Reection
Asymptotic material
indierence
G(x )
G(x, )
G(|x |)
G(x ) = G( x)
G(|x |)
no
yes
yes
yes
no
no
yes
yes
yes
no
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
3.3.5 Realizability Conditions

A second-rank tensor is realizable or semi-positive denite, if the following inequalities are veried (without summation on the repeated greek indices) [746, 260]:

| |
det( )
0,
= 1, 2, 3 ,
,
, = 1, 2, 3 ,
(3.153)
(3.154)
0 .
(3.155)
These conditions can be written in several equivalent forms [260]. Some

of these are listed below.
1. The quadratic form
Q = xi ij xj
(3.156)
is positive semidenite.
2. The three principal invariants of are nonnegative:
I1
I2

2
(
)0 ,
(3.157)
(3.158)
=
I3
= det( ) 0 .
(3.159)
The positiveness of the lter as dened by relation (2.24) is a necessary

and sucient condition to ensure the realizability of the subgrid tensor .
Below, we reproduce the demonstration given by Vreman et al. [746], which
is limited to the case of a spatial lter G(x) without restricting the general
applicability of the result.
Let us rst assume that G 0. To prove that the tensor is realizable at
any position x of the uid domain , we dene the sub-domain x representing the support of the application G(x ). Let Fx be the space of real
73
functions dened on x . Since G is positive, for , Fx , the application

G(x )()()d
(, )x =
(3.160)
denes an inner product on Fx . Using the denition of the ltering, the

subgrid tensor can be re-written in the form:
ui uj (x) ui (x)uj (x)

ui uj (x) ui (x)uj (x) uj (x)ui (x) + ui (x)uj (x)

G(x )ui ()uj ()d3 ui (x)
G(x )uj ()d3
x
x

3
uj (x)
G(x )ui ()d ui (x)uj (x)
G(x )d3
x
x

G(x ) (ui () ui (x)) (uj () uj (x))
(uxi , uxj )x
ij (x) =
=
=
(3.161)
where the dierence uxi () = ui () ui (x) is dened on x . The tensor

thus appears as a Grammian 3 3 matrix of inner products, and is consequently always dened as semi-positive. This shows that the stated condition
is sucient.
Let us now assume that the condition G 0 is not veried for a piecewise
continuous kernel. There then exists a pair (x, y) , an IR+ , > 0,
and a neighbourhood V = { , | y| < }, such that G(x ) < 0,
V . For a function u1 dened on such that u1 () = 0 if V et u1 () = 0
everywhere else, then the component 11 is negative:

2
11 (x) = u21 (x) (u1 (x))
G(x ) (u1 ()) d3 < 0 .
(3.162)
The tensor is thus not semi-positive denite, which concludes the

demonstration. The properties of the three usual analytical lter presented
in Sect. 2.1.5 are summarized in Table 3.2.
Table 3.2. Positiveness property of convolution lters.
Filter
Box
Gaussian
Sharp cuto
Eq.
(2.40)
(2.42)
(2.44)
Positiveness
yes
yes
no
74
3.4 Extension to the Inhomogeneous Case

for the Conventional Approach
The results of the previous sections were obtained by applying isotropic homogeneous lters on an unbounded domain to NavierStokes equations written
in Cartesian coordinates. What is presented here are the equations obtained
by applying non-homogeneous convolution lters on bounded domains to
these equations. Using the commutator (2.13), the most general form of the
ltered NavierStokes equations is:
ui
t

ui
uj
ij
(ui uj ) +
+
=
xj
xi
xj xj
xi
xj

G
,
(ui uj )
(ui ) G
,
t
xj

2
G
,
(3.163)
(p) + G
,
(ui ) ,
xi
xk xk

ui
= G
,
(ui ) .
xi
xi
(3.164)
All the terms appearing in the right-hand side of equations (3.163)

and (3.164) are commutation errors. The rst term of the right-hand side
of the ltered momentum equation is the subgrid force, and is subject to
modeling. The other terms are artefacts due to the lter, and escape subgrid
modeling.
An interesting remark drawn from equation (3.164) is that the ltered
eld is not divergence-free if some commutation errors arise. An analysis of
the breakdown of continuity constraint in large-eddy simulation is provided
by Langford and Moser [420], which shows that for many common largeeddy simulation representations, there is no exact continuity constraint on
the ltered velocity eld. But for mean-preserving representations a bulk
continuity constraint holds.
The governing equations obtained using second-order commuting lters
(SOCF), as well as the techniques proposed by Ghosal and Moin [262] and
Iovenio and Tordella [343] to reduce the commutation error and Vasilyevs
high-order commuting lters [728], are presented in the following.
3.4.1 Second-Order Commuting Filter
Here we propose to generalize Leonards approach by applying SOCF lters.
The decomposition of the non-linear term considered here as an example is
the triple decomposition; but the double decomposition is also usable. For
convenience in writing the ltered equations, we introduce the operator Di
3.4 Extension to the Inhomogeneous Case for the Conventional Approach
75
such that:
= Di
xi
(3.165)
According to the results of Sect. 2.2.2, the operator Di is of the form:

Di =
2
4
(2) ijk 2 + O( ) ,
xi
xi
(3.166)
in which the term is dened by the relation (2.147). By applying the

lter and bringing out the subgrid tensor ij = ui uj ui uj , we get for the
momentum equation:
ui
+ Dj (ui uj ) = Di p + Dj Dj ui Dj ij
t
(3.167)
To measure the errors, we introduce the expansion as a function of :

p
u =
p(0) + p(1) + ... ,
(3.168)
u(0) + u(1) + ...
(3.169)
The terms corresponding to the odd powers of are identically zero

because of the symmetry of the convolution kernel. By substituting this decomposition in (3.167), at the rst order we get:

(0)
(0)
(0)
(0)
uj
ij
p(0)
(0) (0)
ui
ui
=
+
ui uj
+
+
,
t
xj
xi
xj
xj
xi
xj
(3.170)
(0)
(0)
in which ij is the subgrid term calculated from the eld u . The associated
continuity equation is:
(0)
ui
=0 .
(3.171)
xi
These equations are identical to those obtained in the homogeneous case,
2
but relate to a variable containing an error in O( ) with respect to the
exact solution.
2
To reduce the error, the problem of the term in has to be resolved, i.e.
solve the equations that use the variables u(1) and p(1) . Simple expansions
lead to the system:

(1)
(1)
(1)

uj
p(1)
(1) (0)
ui
ui
(0) (1)
=
+
ui uj + ui uj
+
+
t
xj
xi
xj
xj
xi
(1)
ij
(1)
+ (2) fi
xj
(3.172)
76

(1)
in which the coupling term fi
dened as:
(0) (0)
(1)
fi
(0)
2 (ui uj )
2 ij
2 p(0)
jmn
+ imn
+ jmn
xm xn
xm xn
xm xn
(0)
(0)
3 ui
kmn 2 ui
2kmn
xk xm xn
xk xm xn
(3.173)
(1)
ui
=0
xi
(3.174)
By solving this second problem, we can ensure the accuracy of the solution
4
up to the order O( ) .
Another procedure aiming at removing the commutation error was proposed by Iovenio and Tordella [343]. It relies on an approximation of the
commutation error terms up to the fourth order in terms of which is based
on the use of several ltering levels. Reminding that the commutation error
between the ltering operator and the rst-order spatial derivative can be
expressed as

G
,

d
d(x)
(x)
() =
dx
dx
(3.175)
where denotes the ltered quantity obtained applying a lter with length
on the variable , and introducing the central second order nite-dierence
approximation for the gradient of the ltered quantity with respect to the
lter width:

1
+h
h
=
(3.176)
+ O(h2 )
2h
one obtains the following explicit, two ltering level approximation for the
commutation error

2

d
d(x) 1
G
,

()
dx
dx 2
(3.177)
This evaluation is independent of the exact lter shape, and makes it possible to cancel the leading error term in each part of the ltered Navier-Stokes
equations (3.163) - (3.164). It just involves the denition of an auxiliary ltering level with a cuto length equal to 2.
3.5 Filtered NavierStokes Equations in General Coordinates
77
3.4.2 High-Order Commuting Filters

The use of Vasilyevs lters (see Sect. 2.2.2) instead of SOCF yields a set of
governing ltered equations formally equivalent to (3.167), but with:
Di =
n
+ O( ) ,
xi
(3.178)
where the order of accuracy n is xed by the number of vanishing moments

of the lter kernel. The classical ltered equations, without extra-terms accounting for the commutation errors, relate to a variable containing an error
n
scaling as O( ) with respect to the exact ltered solution.
3.5 Filtered NavierStokes Equations

in General Coordinates
3.5.1 Basic Form of the Filtered Equations
Jordan [358, 359], followed by other researchers [780, 18], proposed operating
the ltering in the transformed plane, following the alternate approach, as
dened at the beginning of this chapter. Assuming that the lter width and
local grid spacing are equal, the resolved and ltered owelds are identical.
It is recalled that the ltering operation is applied along the curvilinear lines:

k
k
( )( ) = G( k k )( k )( k )d k ,
(3.179)
where is a metric coecient or a group of metric coecients, a physical
variable (velocity component, pressure), G a homogeneous lter kernel, and
k the coordinate along the considered line. It is easily deduced from the
results presented in Sect. 2.2 that the commutation error vanishes in the
present case, thanks to the homogeneity of the kernel: G/ = 0. But it is
worth noting that the error term coming from the boundary of the domain
will not cancel in the general case.9
Application of the lter to the NavierStokes equations written in generalized coordinates (3.3) and (3.4) leads to the following set of governing
equations for large-eddy simulation:
(J 1 ik ui ) = 0 ,
k
1
(J ui ) + k (U k ui ) = k (J 1 ik p) + k
t
This point is extensively discussed in Chap. 10.
(3.180)

J 1 Gkl l (ui )
.
(3.181)
78
3.5.2 Simplied Form of the Equations Non-linear Terms

Decomposition
It is seen that many ltered nonlinear terms appear in (3.180) and (3.181)
which originate from the coordinate transformation. In order to uncouple
geometrical quantities, such as metrics and Jacobian, from quantities related
to the ow, like velocity, and to obtain a simpler problem, further assumptions
are required. The metrics being computed by a nite dierence approximation
in practice, they can be considered as ltered quantities, yielding:
U k = J 1 jk uj J 1 jk uj
(3.182)
All the terms appearing in the ltered equations can be simplied similarly. As for the conventional approach, convective nonlinear terms need to
be decomposed in order to allow us to use them for practical purpose. The
resulting equations are:
(U k ) = 0 ,
(3.183)
k
1
(J ui ) + k (U k ui ) =
t
(J 1 ik p)
(3.184)
k

+ k J 1 Gkl l (ui ) k (ik ) ,
where the contravariant counterpart of the subgrid tensor is dened as

ik = J 1 jk ui uj J 1 jk uj ui = U k ui U k ui
(3.185)
Taking into account the fact that the metrics are assumed to be smooth
ltered quantities, the contravariant subgrid tensor can be tied to the subgrid
tensor dened in Cartesian coordinates:
ik = J 1 jk ui uj J 1 jk uj ui = J 1 jk (ui uj ui uj ) = J 1 jk ij .
(3.186)
3.6 Closure Problem

3.6.1 Statement of the Problem
As was already said in the rst chapter, large-eddy simulation is a technique
for reducing the number of degrees of freedom of the solution. This is done by
separating the scales in the exact solution into two categories: resolved scales
and subgrid scales. The selection is made by the ltering technique described
above.
3.6 Closure Problem
79
The complexity of the solution is reduced by retaining only the large

scales in the numerical solution process, which entails reducing the number
of degrees of freedom in the solution in space and time. The information
concerning the small scales is consequently lost, and none of the terms that

use these scales, i.e. the terms in u in the physical space and in (1 G)
in the spectral space, can be calculated directly. They are grouped into the
subgrid tensor . This scale selection determines the level of resolution of the
mathematical model.
Nonetheless, in order for the dynamics of the resolved scales to remain
correct, the subgrid terms have to be taken into consideration, and thus have
to be modeled. The modeling consists of approximating the coupling terms
on the basis of the information contained in the resolved scales alone. The
modeling problem is twofold:
1. Since the subgrid scales are lacking in the simulation, their existence is
unknown and cannot be decided locally in space and time. The problem
thus arises of knowing if the exact solution contains, at each point in
space and time, any smaller scales than the resolution established by
the lter. In order to answer this question, additional information has
to be introduced, in either of two ways. The rst is to use additional
assumptions derived from acquired knowledge in uid mechanics to link
the existence of subgrid modes to certain properties of the resolved scales.
The second way is to enrich the simulation by introducing new unknowns
directly related to the subgrid modes, such as their kinetic energy, for
example.
2. Once the existence of the subgrid modes is determined, their interactions
with the resolved scales have to be reected. The quality of the simulation
will depend on the delity with which the subgrid model reects these
interactions.
Various modeling strategies and models that have been developed are
presented in the following.
An important remark, somewhat tautological, is that the modeling process should take into account the ltering operator [597, 171, 604]. This can
be seen by remarking that ltered and subgrid elds are dened by the ltering operator, and that a change in the lter will automatically lead to a new
denition of these quantities and modify their properties.
3.6.2 Postulates
So far, we have assumed nothing concerning the type of ow at hand, aside
from those assumptions that allowed us to demonstrate the momentum and
continuity equations. Subgrid modeling usually assumes the following hypothesis
Hypothesis 3.1 If subgrid scales exist, then the ow is locally (in space and
time) turbulent.
80
Consequently, the subgrid models will be built on the known properties

of turbulent ows.
It should be noted that theories exist that use other basic hypotheses.
We may mention as an example the description of suspensions in the form of
a uid with modied properties [423]: the solid particles are assumed to have
predened characteristics (mass, form, spatial distribution, and so forth) and
have a characteristic size very much less than the lter cuto length, i.e. at the
scale at which we want to describe the ow dynamics directly. Their actions
are taken into account globally, which means a very high saving compared
with an individual description of each particle. The dierent descriptions
obtained by homogenization techniques also enter into this framework.
3.6.3 Functional and Structural Modeling
Preliminary Remarks. Before discussing the various ways of modeling
the subgrid terms, we have to set some constraints in order to orient the
choices [627]. The subgrid modeling must be done in compliance with two
constraints:
1. Physical constraint. The model must be consistent from the viewpoint of
the phenomenon being modeled, i.e.:
Conserve the basic properties of the starting equation, such as Galilean
invariance and asymptotic behaviors;
Be zero wherever the exact solution exhibits no small scales corresponding to the subgrid scales;
Induce an eect of the same kind (dispersive or dissipative, for example) as the modeled terms;
Not destroy the dynamics of the solve scales, and thus especially not
inhibit the ow driving mechanisms.
2. Numerical constraint. A subgrid model can only be thought of as included
in a numerical simulation method, and must consequently:
Be of acceptable algorithmic cost, and especially be local in time and
space;
Not destabilize the numerical simulation;
Be insensitive to discretization, i.e. the physical eects induced theoretically by the model must not be inhibited by the discretization.
Modeling Strategies. The problem of subgrid modeling consists in taking
the interaction with the uctuating eld u , represented by the term ,
into account in the evolution equation of the ltered eld u. Two modeling
strategies exist [627]:
Structural modeling of the subgrid term, which consists in making the best
approximation of the tensor by constructing it from an evaluation of u
or a formal series expansion. The modeling assumption therefore consists
in using a relation of the form u = H(u) or = H(u).
3.6 Closure Problem
81
Functional modeling, which consists in modeling the action of the subgrid

terms on the quantity u and not the tensor itself, i.e. introducing a dissipative or dispersive term, for example, that has a similar eect but not
necessarily the same structure (not the same proper axes, for example).
The closure hypothesis can then be expressed in the form = H(u).
These two modeling approaches do not require the same foreknowledge
of the dynamics of the equations treated and theoretically do not oer the
same potential in terms of the quality of results obtained.
The structural approach requires no knowledge of the nature of the interscale interaction, but does require enough knowledge of the structure of the
small scales of the solution in order to be able to determine one of the relations
u = H(u) or = H(u) to be possible, one of the two following conditions
has to be met:
The dynamics of the equation being computed leads to a universal form of
the small scales (and therefore to their total structural independence from
the resolved motion, as all that remains to be determined is their energy
level).
The dynamics of the equation induces a suciently strong and simple interscale correlation for the structure of the subgrid scales to be deduced from
the information contained in the resolved eld.
As concerns the modeling of the inter-scale interaction by just taking
its eect into account, this requires no foreknowledge of the subgrid scale
structure, but does require knowing the nature of the interaction [184] [383].
Moreover, in order for such an approach to be practical, the eect of the small
scales on the large must be universal in character, and therefore independent
of the large scales of the ow.
4. Other Mathematical Models

for the Large-Eddy Simulation Problem
The two preceding chapters are devoted to the convolution ltering mathematical model for Large-Eddy simulation. Others approaches are now described, that can be gouped in two classes:
Mathematical models which rely on a statistical average (Sect. 4.1), recovering this way some interesting features of the Reynolds-Averaged Navier
Stokes model by precluding some drawbacks of the convolution lter approach in general domains.
Models derived from regularized versions of the NavierStokes equations
(Sect. 4.2), that were proposed to alleviate some theoretical problems dealing with the existence, the uniqueness and the regularity of the general
solution of the three-dimensional, unsteady, incompressible NavierStokes
equations. These regularized models have smooth solutions, in the sense
that their gradients remain controlled, and are re-interpreted within the
Large-Eddy Simulation framework as good candidates to account for the
removal of small scales.
4.1 Ensemble-Averaged Models

4.1.1 Yoshizawas Partial Statistical Average Model
Yoshizawa [791] proposes to combine scale decomposition and statistical average to dene an ad hoc mathematical model for Large-Eddy Simulation,
referred to as the partial statistical average procedure. Writing the generalized Fourier decomposition of a dummy variable (x, t) as

k (t)k (x)
(4.1)
(x, t) =
k=1,+
where k (t) and k (x) are the coecients of the decomposition and the basis
functions, respectively, the ltered part of (x, t) is dened as

(x, t) =
k (t)k (x) +
k (t)k (x)
(4.2)
k=1,kc
k=kc ,+
84
4. Other Mathematical Models for the Large-Eddy Simulation Problem
where denotes a statistical average operator and kc is related to the cuto

index of the decomposition. The partial statistical averaging method appears
then as the restriction of the usual ensemble average to scales which correspond to modes higher than kc . The cuto length is deduced from the
characteristic lengthscale associated to kc .
Since it relies on an ensemble average operator, this procedure does not
suer the drawbacks of the convolution ltering approach and can be applied
on curvilinear grids on bounded domains in a straightforward manner. But it
requires the computation of the coecients k (t), and therefore several realizations of the ow are necessary, rendering its practical implementation very
expensive from the computational viewpoint. In the simple case of homogeneous ows, the statistical average can be transformed into a spatial average
invoking the ergodic theorem (see Appendix A for a brief discussion).
4.1.2 McCombs Conditional Mode Elimination Procedure
Another procedure was proposed independently by McComb and coworkers [465], which is referred to as conditional mode elimination. These authors
based their approach on the local chaos hypothesis, which states that in a fully
turbulent ow the small scales are more uncertain than the large ones. This
assumption is compatible with Kolmogorovs local isotropy hypothesis (see
Sect. A.5.1 for a discussion) dealing with the universality of the small scales
and their increasing (as a function of the wavenumber) statistical decoupling
from the large ones. More precisely, McCombs interpretation says that uncertainty in the high-wavenumber modes originates in the amplication of some
degree of uncertainty in low-wavenumber modes by the non-linear chaotic
nature of turbulence. This scheme is illustrated in Fig. 4.1.
Fig. 4.1. Schematic view of the local chaos hypothesis proposed by McComb in the
Fourier space. Left: several instantaneous spectra are shown, in which increasing
uncertainty is observed. Right: ideal view, where wave numbers smaller than kc are
strictly deterministic, while higher wave number exhibit a fully chaotic behavior.
4.2 Regularized NavierStokes Models
85
The scale separation with a cuto length is achieved carying out a conditional statistical average of scales smaller than , < , based on xed realizations of scales larger than , > . The former are assumed to be uncertain
and to exhibit and innite number of dierent realizations for each realization
of the large scales.
The ltered part of (x, t) is then expressed as
(x, t) = > (x, t) + < |> (x, t)
(4.3)
where f |g denotes the conditional statistical average of f with respect to

g.
As Yoshizawas procedure, the conditional mode elimination does not suffer the drawbacks of the ltering approach. These two ensemble-average based
models for Large-Eddy Simulation are equivalent in many cases.

The mathematical models discussed in this section were not originally proposed to represent the properties of the Large-Eddy Simulation technique.
They are surrogates to the NavierStokes equations, which have better properties from the pure mathematical point of view: while the question of the
existence, uniqueness and regulatity of the solution of the three-dimensional,
unsteady and incompressible NavierStokes equations is still an open problem, these new models allow for a complete mathematical analysis. One of
the main obstacle faced in the mathematical analysis of the NavierStokes
equations is that it cannot yet be proven that its solutions remain smooth for
arbitrarily long times. More precisely, no a priori estimates has been found
which guarantees that the enstrophy remains nite everywhere in the domain
lled by the uid (but it can be proven that it is bounded in the mean). The
physical interpretation associated with this picture is that some very intermittent vorticity bursts can occur, injecting kinetic energy at scales much
smaller than the Kolmorogov scale, resulting in quasi-innite local values
of the enstrophy. Such events correspond to nite-time singularities of the
solution, and violate the axiom of continuum mechanics.
A large number of mathematical results dealing with these problems have
been published, which will not be further discussed here. The important point
is that some systems, which are very close to the NavierStokes equations,
have been proposed. A common feature is that they are well-posed from the
mathematical point of view, meaning that their solutions are proved to be
regular. As a consequence, they appear as regularized systems derived from
the original NavierStokes equations, the regularization being associated to
the disappearance of singularities. From a physical point of view, these new
systems do not allow the occurance of local innite gradient thanks to an
extra damping of the smallest scales. This smoothing property originates
their interpretation as models for Large-Eddy Simulation.
86
4.2.1 Lerays Model

The rst model was proposed by Leray in 1934, who suggested to regularize
the NavierStokes equations as follows:
ui uk ui
p
2 ui
+
=
+
t
xk
xi
xk xk
(4.4)
ui
=0
xi
(4.5)
where the regularized (i.e. ltered) velocity eld is dened as

u(x, t) =
u(x, t)
(4.6)
where the mollifying function (i.e. the lter kernel) is assumed to have
a compact support, to be C and to have an integral equal to one. It can
be proved under these assumptions that the solution of the regularized system (4.4)(4.5) is unique and C . A main drawback is that it does not share
all the frame-invariance properties of the NavierStokes equations. As quoted
by Geurts and Holm [258], the system proposed by Leray can be rewritten
in the usual Large-Eddy Simulation framework applying the lter a second
times, leading to
ijLeray
uk ui
p
2 ui
ui
+
=
+
t
xk
xi
xk xk
xj
(4.7)
ui
=0
xi
(4.8)
where the subgrid tensor ansatz is dened as

ijLeray = ui uj ui uj
(4.9)
An important dierence with the usual denition of the subgrid tensor

ij is that this new tensor is not symmetric. Lerays regularized model makes
it possible to carry out a complete mathematical analysis, but suers the
same problem when dealing with curvilinear grids on bounded domains as
the original convolution lter model described in the preceding chapters.
4.2.2 Holms NavierStokes- Model
The second regularized model presented in this chapter is the NavierStokes proposed by Holm (see [221, 282, 182, 221, 258, 281]). The regularization
is achieved by imposing an energy penalty which damps the scales smaller
than the threshold scale (to be interpreted as within the usual large-
87
eddy simulation framework)1, while still allowing for non linear sweeping of
the small scales by the largest ones. The regularization appears as a nonlinearly dispersive modication of the convection term in the NavierStokes
equations.
The system of the NavierStokes- (also referred to as the Camassa-Holm
equations) can be derived in two dierent ways, which are now presented.
Method 1: Kelvin-ltered NavierStokes equations. The rst way to
obtain the NavierStokes- model is to introduce the Kelvin-ltering. The
NavierStokes equations satisfy Kelvins circulation theorem
d
dt
$
u dx =
(u)
(2 ) dx
(4.10)
(u)
where (u) is a closed uid loop that moves with velocity u. The original
set of equations is regularized by modifying the uid loop along which the
circulation is integrated: instead of using a uid loop moving at velocity u,
an new uid loop moving at the regularized velocity u is considered. The
exact denition of u is not necessary at this point and will be given later.
The new circulation relationship is
d
dt
$
u dx =
(u)
(2 ) dx
(4.11)
(u)
and corresponds to the following modied momentum equation:

u
+ u u + T u u = p + 2 u
t
(4.12)
u = 0 .
(4.13)
with
This set of equations describes the Kelvin-ltered NavierStokes equations. The NavierStokes- equations are recovered specifying the regularized
eld u as the result of the application of the Helmholtz lter (2.35) to the
original eld u:
(4.14)
u = (1 2 2 )u .
It can be proved that the kinetic energy E dened as
1
E =
2
1

u udx =

1 2 2 2 2
|u| +
| u| dx
2
2
(4.15)
It can be shown that in the case of three-dimensional fully developed turbulence,

the solution of the NavierStokes- exhibits the usual k5/3 behavior for scales
larger than and a k3 behavior for scales smaller than .
88
is bounded, showing that the ltered eld u remains regular. The equation (4.12) can be rewritten under the usual form in Large-Eddy Simulation
as a momentum equation for the ltered velocity eld u (formally identical
to (4.4)). The corresponding denition of the subgrid tensor is

ijNS = (ui uj ui uj ) 2
uk uk
+ uj 2 ui
xi xj

.
(4.16)
Method 2: Modied Lerays Model. Guermond, Oden and Prudhomme [282] observe that the NavierStokes- system can be interpreted as
a frame-invariant modication of original Lerays regularized model. Starting
from the rotational form of the NavierStokes equations
u
+ ( u) u = + 2 u,
t
1
= p + u2
2
(4.17)
u = 0 ,
(4.18)
and regularizing it using the technique proposed by Leray, one obtains

u
+ ( u) u = + 2 u,
t
1
= p + u2
2
(4.19)
u = 0 .
(4.20)
Now using the relations

( u) u = u u (T u)u,
(u (T u)) = (T u)u + (T u)u ,

(4.21)
the following form of the regularized system is recovered
u
+ u u + (T u) u = + 2 u,
t
= u u
u = 0 .
(4.22)
(4.23)
The NavierStokes- model is obatined using the Helmholtz lter (4.14).

The corresponding equation for u is
u
+ u u = T
t
(4.24)
with
T = pId + 2(1 2 2 )S + 22 S
(4.25)
89
where S is related to the Jaumann derivative of the regularized strain rate

tensor:
S =
S
+ u S + S S,
t

1
u T u
2
(4.26)
This system is formally similar to the constitutive law of a rate-dependent

incompressible uid of second grade with slightly modied dissipation, and it
is frame-invariant. It is equivalent to the Leray model in which the term which
is responsible for the failure in the frame preservation, i.e. 2 (T u2 u), has
been removed. Therefore, the NavierStokes- equations appear as a pertur2
bation of order 2 (i.e. ) of the original Leray model.
4.2.3 Ladyzenskajas Model
Another regularized version of the NavierStokes equations was proposed
by Ladyzenskaja and Kaniel [417, 418, 377], who introduced a non-linear
modication of the stress tensor which is expected to be more relevant than
the linear relationship for Newtonian uids when velocity gradients are large.
The equation for the regularized eld u is
u
+ u u = p + 2 u T (u) ,
t
(4.27)
u = 0 ,
(4.28)
where is a strictly arbitrary constant and the non-linear stress tensor T is

dened as
T (u) = T (u2 )u
(4.29)
where the non-linear viscosity T ( ) is a positive monotonically-increasing

function of 0 that obeys the following law for large values of :
c T ( ) c ,
0 < c < c ,
1
4
(4.30)
The equivalent expression for the subgrid stress tensor is

ijLadyzenskaja = Tij (u2 ) .
(4.31)
Since T depends only on the gradient of the resolved eld u, Ladyzenskajas model is closed and does not require further modeling work.
5. Functional Modeling (Isotropic Case)
It would be illusory to try to describe the structure of the scales of motion

and the interactions in all imaginable congurations, in light of the very large
disparity of physical phenomena encountered. So we have to restrict this description to cases which by nature include scales that are too small for todays
computer facilities to solve them entirely, and which are at the same time accessible to theoretical analysis. This description will therefore be centered on
the inter-scale interactions in the case of fully developed isotropic homogeneous turbulence1 , which is moreover the only case accessible by theoretical
analysis and is consequently the only theoretical framework used today for
developing subgrid models. Attempts to extend this theory to anisotropic
and/or inhomogeneous cases are discussed in Chap. 6. The text will mainly
be oriented toward the large-eddy simulation aspects. For a detailed description of the isotropic homogeneous turbulence properties, which are reviewed
in Appendix A, the reader may refer to the works of Lesieur [439] and Batchelor [45].
5.1 Phenomenology of Inter-Scale Interactions

It is important to note here the framework of restrictions that apply to the
results we will be presenting. These results concern three-dimensional ows
and thus do not cover the physics of two-dimensional ows (in the sense
of ows with two directions2 , and not two-component3 ows), which have
a totally dierent dynamics [403, 404, 405, 438, 481]. The modeling in the
two-dimensional case leads to specic models [42, 624, 625] which will not
1
2
That is, whose statistical properties are invariant by translation, rotation, or

symmetry.
These are ows such that there exists a direction x for which we have the property:
u
0 .
x
These are ows such that there exists a framework in which the velocity eld
has an identically zero component.
92
be presented. For details on two-dimensional turbulence, the reader may also

refer to [439].
5.1.1 Local Isotropy Assumption: Consequences
In the case of fully developed turbulence, Kolmogorovs statistical description
of the small scales of the ow, based on the assumption of local isotropy, has
been the one most used for a very long time.
By introducing the idea of local isotropy, Kolmogorov assumes that the
small scales belonging to the inertial range of the energy spectrum of a fully
developed inhomogeneous turbulent ow are:
Statistically isotropic, and therefore entirely characterized by a characteristic velocity and time;
Without time memory, therefore in energy equilibrium with the large scales
of the ow by instantaneous re-adjustment.
This isotropy of the small scales implies that they are statistically independent of the large energetic scales, which are characteristic of each ow
and are therefore anisotropic. Experimental work [512] has shown that this
assumption is not valid in shear ows for all the scales belonging to the inertial range, but only for those whose size is of the order of the Kolmogorov
scale. Numerical experiments [32] show that turbulent stresses are nearly
isotropic for wave numbers k such that kL > 50, where L is the integral
dissipation length4 . These experiments have also shown that the existence
of an inertial region does not depend on the local isotropy hypothesis. The
causes of this persistence of the anisotropy in the inertial range due to interactions existing between the various scales of the ow will be mentioned
in Chap. 6. Works based on direct numerical simulations have also shown
that the assumption of equilibrium between the resolved and subgrid scales
may be faulted, at least temporarily, when the ow is subject to unsteady
forcing [594, 570, 454, 504]. This is due to the fact that the relaxation times
of these two scale ranges are dierent. In the case of impulsively accelerated
ows (plane channel, boundary layer, axisymmetric straining) the subgrid
scales react more quickly than the resolved ones, and then also relax more
quickly toward an equilibrium solution.
The existence of a zone of the spectrum, corresponding to the higher frequencies, where the scales of motion are statistically isotropic, justies the
study of the inter-modal interactions in the ideal case of isotropic homogeneous turbulence. Strictly speaking, the results can be used for determining
subgrid models only if the cuto associated with the lter is in this region,
4
The integral dissipation length is dened as

L =
ui ui 3/2
where is the energy dissipation rate.
93
because the dynamics of the unresolved scales then corresponds well to that
of the isotropic homogeneous turbulence. It should be noted that this last
condition implies that the representation of the dynamics, while incomplete,
is nonetheless very ne, which theoretically limits the gain in complexity that
can be expected from large-eddy simulation technique.
Another point is that the local isotropy hypothesis is formulated for fully
developed turbulent ows at very high Reynolds numbers. As it arms the
universal character of the small scales behavior for these ows, it ensures
the possibility using the large-eddy simulation technique strictly, if the lter
cuto frequency is set suciently high. There is no theoretical justication,
though, for applying the results of this analysis to other ows, such as transitional ows.
5.1.2 Interactions Between Resolved and Subgrid Scales
In order to study the interactions between the resolved and subgrid scales,
we adopt an isotropic lter by a cuto wave number kc . The subgrid scales
are those represented by the k modes such that k kc .
In the case of fully developed isotropic homogeneous turbulence, the statistical description of the inter-scale interactions is reduced to that of the
kinetic energy transfers. Consequently, only the information associated with
the amplitude of the uctuations is conserved, and none concerning the phase
is taken into account.
These transfers are analyzed using several tools:
Analytical theories of turbulence, also called two-point closures, which describe triadic interactions on the basis of certain assumptions. They will
therefore express the non-linear term S(k|p, q), dened by relation (3.50)
completely. For a description of these theories, the reader may refer to
Lesieurs book [439], and we also mention Walees analysis [748, 749],
certain conclusions of which are presented in the following.
Direct numerical simulations, which provide a complete description of the
dynamics.
Renormalization Group Theory [622, 328, 812, 809, 464, 775, 804, 805, 813,
802, 803, 810], with several variants.
Typology of the Triadic Interactions. It appears from the developments
(k) mode interacts only with
of Sect. 3.1.3 (also see Appendix A) that the u
those modes whose wave vectors p and q form a closed triangle with k. The
wave vector triads (k, p, q) thus dened are classied in several groups [805]
which are represented in Fig. 5.1:
Local triads for which
%p q&
1
max
,
a,
a
k k
a = O(1)
94
Fig. 5.1. Dierent types of triads.
which correspond to interactions among wave vectors of neighboring modules, and therefore to interactions among scales of slightly dierent sizes;
Non-local triads, which are all those interactions that do not fall within the
rst category, i.e. interactions among scales of widely diering sizes. Here,
we adopt the terminology proposed in [74], which distinguishes between
two sub-classes of non-local triads, one being distant triads of interactions
in which k p q or k q p. It should be noted that these terms
are not unequivocal, as certain authors [439, 442] refer to these distant
triads as being just non-local.
By extension, a phenomenon will be called local if it involves wave vectors
k and p such that 1/a p/k a, and otherwise non-local or distant.
Canonical Analysis. This section presents the results from analysis of the
simplest theoretical case, which we call here canonical analysis. This consists
of assuming the following two hypotheses:
1. Hypothesis concerning the ow. The energy spectrum E(k) of the exact
solution is a Kolmogorov spectrum, i.e.
E(k) = K0 2/3 k 5/3 ,
k [0, ] ,
(5.1)
where K0 is the Komogorov constant and the kinetic energy dissipation rate. We point out that this spectrum is not integrable since its
corresponds to an innite kinetic energy.
2. Hypothesis concerning the lter. The lter is a sharp cuto type. The
subgrid tensor is thus reduced to the subgrid Reynolds tensor.
e
In analyzing the energy transfers Tsgs
(k) (see relation (3.51)) between the
modes to either side of a cuto wave number kc located in the inertial range of
the spectrum, Kraichnan [405] uses the Test Field Model (TFM) to bring out
the existence of two spectral bands (see Fig. 5.2) for which the interactions
with the small scales (p and/or q kc ) are of dierent kinds.
95
Fig. 5.2. Interaction regions between resolved and subgrid scales.
1. In the rst region (1 in Fig. 5.2), which corresponds to the modes such
that k kc , the dominant dynamic mechanism is a random displacement of the momentum associated with k by disturbances associated
with p and q. This phenomenon, analogous to the eects of the molecular viscosity, entails a kinetic energy decay associated with k and, since
the total kinetic energy is conserved, a resulting increase of it associated
with p and q. So here it is a matter of a non-local transfer of energy associated with non-local triadic interactions. These transfers, which induce
a damping of the uctuations, are associated with what Walee [748, 749]
classies as type F triads (represented in Fig. 5.3).
Fig. 5.3. Non-local triad (k, p, q) of the F type according to Walees classication, and the associated non-local energy transfers. The kinetic energy of the mode
corresponding to the smallest wave vector k is distributed to the other two modes
p and q, creating a forward energy cascade in the region where k kc .
96
Subsequent analyses using the Direct Interaction Approximation (DIA)

and the Eddy Damped Quasi-Normal Markovian (EDQNM) models [120,
136, 442, 443, 647] or Walees analyses [748, 749] have rened this representation by showing the existence of two competitive mechanisms in
the region where k kc . The rst region is where the energy of the large
scales is drained by the small ones, as already shown by Kraichnan. The
second mechanism, of much lesser intensity, is a return of energy from
the small scales p and q to the large scale k. This mechanism also corresponds to a non-local energy transfer associated with non-local triadic
interactions that Walee classies as type R (see Fig. 5.4). It represents
a backward stochastic energy cascade associated with an energy spectrum
in k 4 for very small wave numbers. This phenomenon has been predicted
analytically [442] and veried by numerical experimentation [441, 120].
The analytical studies and numerical simulations show that this backward cascade process is dominant for very small wave numbers. On the
average, these modes receive more energy from the subgrid modes than
they give to them.
2. In the second region (region 2 in Fig. 5.2), which corresponds to the k
modes such that (kc k) kc , the mechanisms already present in region
1 persist. The energy transfer to the small scales is at the origin of the
forward kinetic energy cascade.
Moreover, another mechanism appears involving triads such that p or
q kc , which is that the interactions between the scales of this region
and the subgrid scales are much more intense than in the rst. Let us take
q kc . This mechanism is a coherent straining of the small scales k and
p by the shear associated with q, resulting in a wave number diusion
process between k and p through the cuto, with one of the structures being stretched (vortex stretching phenomenon) and the other unstretched.
What we are observing here is a local energy transfer between k and p
associated with non-local triadic interactions due to the type R triads
(see Fig. 5.4). Walee renes the analysis of this phenomenon: a very
large part of the energy is transferred locally from the intermediate wave
number located just ahead of the cuto toward the larger wave number
just after it, and the remaining fraction of energy is transferred to the
smaller wave number. These ndings have been corroborated by numerical data [120, 185, 189] and other theoretical analyses [136, 443].
e
(k) (see relation (3.51)) between mode k and
The energy transfers Tsgs
the subgrid modes can be represented in a form analogous to molecular dissipation. To do this, by following Heisenberg (see [688] for a description of
Heisenbergs theory), we dene an eective viscosity e (k|kc ), which represents the energy transfers between the k mode and the modes located beyond
the kc cuto such that:
e
Tsgs
(k) = 2e(k|kc )k 2 E(k) .
(5.2)
97
Fig. 5.4. Non-local (k, p, q) triad of the R type according to Walees classication,
and the associated energy transfers in the case q kc . The kinetic energy of the
mode corresponding to the intermediate wave vector k is distributed locally to the
largest wave vector p and non-locally to the smallest wave vector, q. The former
transfer originates the intensication of the coupling in the (kc k) kc spectral
band, while the latter originates the backward kinetic energy cascade.
It should be pointed out that this viscosity is real, i.e. e (k|kc ) IR, and
that if any information related to the phase were included, it would lead
the denition of a complex term having an a priori non-zero imaginary part,
which may seem to be more natural for representing a dispersive type of
coupling. Such a term is obtained not by starting with the kinetic energy
equation, but with the momentum equation5 .
The two energy cascades, forward and backward, can be introduced separately by introducing distinct eective viscosities, constructed in such a way
as to ensure energy transfers equivalent to those of these cascades. We get
the following two forms:
e+ (k|kc , t) =
+
(k|kc , t)
Tsgs
2
2k E(k, t)
(5.3)
e (k|kc , t) =
Tsgs
(k|kc , t)
2
2k E(k, t)
(5.4)
in which Tsgs
(k|kc , t) (resp. Tsgs
(k|kc , t)) is the energy transfer term from the
k mode to the subgrid modes (resp. from the subgrid modes to the k mode).
This leads to the decomposition:
e
(k) =
Tsgs
Tsgs
(k|kc , t) + Tsgs
(k|kc , t)

2
+
2k E(k, t) e (k|kc , t) + e (k|kc , t)
(5.5)
.
(5.6)
These two viscosities depend explicitly on the wave number k and the
cuto wave vector kc , as well as the shape of the spectrum. The result of
5
This possibility is only mentioned here, because no works have been published
on it to date.
98
these dependencies on the ow is that the viscosities are not, because they
characterize the ow and not the uid. They are of opposite sign: e+ (k|kc , t)
ensures a loss of energy of the resolved scales and is consequently positive,
like the molecular viscosity, whereas e (k|kc , t), which represents an energy
gain in the resolved scales, is negative.
The conclusions of the theoretical analyses [405, 443] and numerical studies [120] are in agreement on the form of these two viscosities. Their behavior
is presented in Fig. 5.5 in the canonical case.
Fig. 5.5. Representation of eective viscosities in the canonical case. Short dashes:
e+ (k|kc , t); long dashes: e (k|kc , t); solid: e+ (k|kc , t) + e (k|kc , t).
We may note that these two viscosities become very high for wave numbers
close to the cuto. These two eective viscosities diverge as (kc k)2/3 as
k tends toward kc . However, their sum e (k|kc , t) remains nite and Leslie et
al. [443] proposes the estimation:
e (kc |kc , t) = 5.24e+ (0|kc , t) .
(5.7)
The interactions with the subgrid scales is therefore especially important

in the dynamics of the smallest resolved scales. More precisely, Kraichnans
theoretical analysis leads to the conclusion that about 75% of the energy
transfers of a k mode occur with the modes located in the [k/2, 2k] spectral
99
band6 . No transfers outside this spectral band have been observed in direct
numerical simulations at low Reynolds numbers [185, 804, 805]. The dierence with the theoretical analysis stems from the fact that this analysis is
performed in the limit of the innite Reynolds numbers.
In the limit of the very small wave numbers, we have the asymptotic
behaviors:
2k 2 E(k, t)e+ (k|kc , t)
2k 2 E(k, t)e (k|kc , t)
k 1/3 ,
k4 .
(5.8)
(5.9)
The eective viscosity associated with the energy cascade takes the constant asymptotic value:
e+ (0|kc , t) = 0.2921/3kc4/3
(5.10)
We put the emphasis on the fact that the eective viscosity discussed
here is dened considering the kinetic enery transfer between unresolved and
resolved modes. As quoted by McComb et al. [465], it is possible to dene
dierent eective viscosities by considering other balance equations, such
as the enstrophy transfer. An important consequence is that kinetic-energybased eective viscosities are ecient surrogates of true transfer terms in
the kinetic energy equation, but may be very bad representations of subgrid
eects for other physical mechanisms.
Dependency According to the Filter. Leslie and Quarini [443] extended
the above analysis to the case of the Gaussian lter. The spectrum considered
is always of the Kolmogorov type. The Leonard term is now non-zero. The
results of the analysis show very pronounced dierences from the canonical
analysis. Two regions of the spectrum are still distinguishable, though, with
regard to the variation of the eective viscosities e+ and e , which are shown
in Fig. 5.6:
In the rst region, where k kc , the transfer terms still observe a constant
asymptotic behavior, independent of the wave number considered, as in the
canonical case. The backward cascade term is negligible compared with the
forward cascade term.
In the second region, on the other hand, when approaching cuto, the two
transfer terms do not have divergent behavior, contrary to what is observed
in the canonical case. The forward cascade term decreases monotonically
and cancels out after the cuto for wave numbers more than a decade
beyond it. The backward cascade term increases up to cuto and exhibits
a decreasing behavior analogous to that of the forward cascade term. The
maximum intensity of the backward cascade is encountered for modes just
after the cuto.
6
The same local character of kinetic energy transfer is observed in nonhomogeneous ow, such has the plane channel ow [186].
100
Fig. 5.6. Eective viscosities in the application of a Gaussian lter to a Kolmogorov

spectrum. Long dots e+ (k|kc , t); dots: e (k|kc , t); solid: e+ (k|kc , t) + e (k|kc , t).
Fig. 5.7. Eective viscosity corresponding to the Leonard term in the case of the
application of a Gaussian lter to a Kolmogorov spectrum.
101
In contrast to the sharp cuto lter used for the canonical analysis, the
Gaussian lter makes it possible to dene Leonard terms and non-identically
zero cross terms. The eective viscosity associated with these terms is shown
in Fig. 5.7, where it can be seen that it is negligible for all the modes more
than a decade away from the cuto. In the same way as for the backward
cascade term, the maximum amplitude is observed for modes located just
after the cuto. This term remains smaller than the forward and backward
cascade terms for all the wave numbers.
Dependency According to Spectrum Shape. The results of the canonical analysis are also dependent on the shape of the spectrum considered. The
analysis is repeated for the case of the application of the sharp cuto lter
to a production spectrum of the form:
E(k) = As (k/kp )K0 2/3 k 5/3
(5.11)
with
As (x) =
xs+5/3
1 + xs+5/3
(5.12)
and where kp is the wave number that corresponds to the maximum of the
energy spectrum [443]. The shape of the spectrum thus dened is illustrated
in Fig. 5.8 for several values of the s parameter.
The variation of the total eective viscosity e for dierent values of the
quotient kc /kp is diagrammed in Fig. 5.9. For low values of this quotient, i.e.
when the cuto is located at the beginning of the inertial range, we observe
Fig. 5.8. Production spectrum for dierent values of the shape parameter s.
102
Fig. 5.9. Total eective viscosity e (k|kc ) in the case of the application of a sharp
cuto lter to a production spectrum for dierent values of the quotient kc /kp ,
normalized by its value at the origin.
that the viscosity may decrease at the approach to the cuto, while it is
strictly increasing in the canonical case. This dierence is due to the fact
that the asymptotic reasoning that was applicable in the canonical case is
no longer valid, because the non-localness of the triadic interactions involved
relay the dierence in spectrum shape to the whole of it. For higher values of
this quotient, i.e. when the cuto is located suciently far into the inertial
range (for large values of the ratio kc /kp ), a behavior that is qualitatively
similar to that observed in the canonical case is once again found7 .
For kc = kp , no increase is observed in the energy transfers as k tends toward kc . The behavior approximates that observed for the canonical analysis
as the ratio kp /kc decreases.
5.1.3 A View in Physical Space
Analyses described in the preceding section were all performed in the Fourier
space, and do not give any information about the location of the subgrid
transfer in the physical space and its correlation with the resolved scale features8 . Complementary informations on the subgrid transfer in the physical
space have been found by several authors using direct numerical simulation.
7
8
In practice, kc /kp =8 seems appropriate.

This is a prerequisite for designing a functional subgrid model in physical space.
103
Kerr et al. [383] propose to use the rotational form of the non-linear term of
the momentum equation:
N (x) = u(x) (x) ph (x)
(5.13)
where = u and ph the pressure term. By splitting the velocity and

vorticity eld into a resolved and a subgrid contribution, we get:
u u = u + u + u
II
III
(5.14)
IV
The four terms represent dierent coupling mechanisms between the resolved motion and the subgrid scales:
I - exact subgrid term,

II - interaction between resolved velocity and subgrid vorticity,
III - interaction between subgrid velocity and resolved vorticity,
IV - interaction between subgrid velocity and subgrid vorticity.
The corresponding complete non-linear terms N I , ..., N IV are built by

adding the specic pressure term. The associated subgrid kinetic energy
transfer terms are computed as l = u N l . The authors made three signicant observations for isotropic turbulence:
Subgrid kinetic energy transfer is strongly correlated with the boundaries of
regions of large vorticity production (stretching), i.e. regions where i S ij j
is large;
Term II, u , has a correlation with subgrid non-linear term I up to 0.9.
This term dominates the backward energy cascade;
Up to 90% of the subgrid kinetic energy transfer comes from term III, i.e.
from the interaction of subgrid velocity with resolved vorticity. This term
mostly contributes to the forward energy cascade.
Additional results of Borue and Orszag [72] show that the subgrid transfer
takes place in regions where the vorticity stretching term is positive or in
3
regions with negative skewness of the resolved strain rate tensor, Tr(S ).
These authors also found that there is only a very poor local correlation
between the subgrid transfer ij S ij and the local strain S ij S ij , where S ij is
the resolved strain rate tensor.
Horiuti [325, 327] decomposed the subgrid tensor into several contributions,9 and used direct numerical simulation data of isotropic turbulence to
analyze their contributions. A rst remark is that the eigenvectors of the total subgrid tensor have a preferred orientation of 42 relative to those of S.
Eigenvectors of (S ik S kj ik kj ) are highly aligned with those of S, while
9
This decomposition is discussed in the section devoted to nonlinear models,

p. 223.
104
those of (S ik kj S ik kj ) exhibit a 42 angle, from which stems the global

observed dierence. The rst term is associated with the forward energy
cascade. The second one makes no contribution to the total production of
subgrid kinetic energy, but is relevant to the vortex stretching and the backward energy cascade process.10 Similar results were obtained by Meneveau
and coworkers [704, 703].
The role of coherent structures in interscale transfer is of major importance in shear ows. Da Silva and Metais [158] carried out an exhaustive
study in the plane jet case: the most intense forward cascade events occur
near these coherent structures and not randomly in space. The local equilibrium assumption is observed to hold globally but not locally as most viscous
dissipation of subgrid kinetic energy takes place within coherent structure
cores, while forward and backward cascade occur at dierent locations.
5.1.4 Summary
The dierent analyses performed in the framework of fully developed isotropic
turbulence show that:
1. Interactions between the small and large scales is reected by two main
mechanisms:
A drainage of energy from the resolved scales by the subgrid scales
(forward energy cascade phenomenon);
A weak feedback of energy, proportional to k 4 to the resolved scales
(backward energy cascade phenomenon).
2. The interactions between the subgrid scales and the smallest of the resolved scales depend on the lter used and on the shape of the spectrum.
In certain cases, the coupling with the subgrid scales is strengthened
for wave numbers close to the cuto and the energy toward the subgrid
modes is intensied.
3. These cascade mechanisms are associated to specic features of the velocity and vorticity eld in physical space.
5.2 Basic Functional Modeling Hypothesis

All the subgrid models entering into this category make more or less implicit
use of the following hypothesis:
Hypothesis 5.1 The action of the subgrid scales on the resolved scales is
essentially an energetic action, so that the balance of the energy transfers
alone between the two scale ranges is sucient to describe the action of the
subgrid scales.
10
This is an indication that the backward energy cascade is not associated with
negative subgrid viscosity from the theoretical point of view.
5.3 Modeling of the Forward Energy Cascade Process
105
Using this hypothesis as a basis for modeling, then, we neglect a part of the
information contained in the small scales, such as the structural information
related to the anisotropy. As was seen above, the energy transfers between
subgrid scales and resolved scales mainly exhibit two mechanisms: a forward
energy transfer toward the subgrid scales and a backward transfer to the
resolved scales which, it seems, is much weaker in intensity. All the approaches
existing today for numerical simulation at high Reynolds numbers consider
the energy lost by the resolved scales, while only a few rare attempts have
been made to consider the backward energy cascade.
Once hypothesis 5.1 is assumed, the modeling consists in modifying the
dierent evolution equations of the system in such a way as to integrate the
desired dissipation or energy production eects into them. To do this, two
dierent approaches can be found in todays works:
Explicit modeling of the desired eects, i.e. including them by adding additional terms to the equations: the actual subgrid models;
Implicit inclusion by the numerical scheme used, by arranging it so the
truncation error induces the desired eects.
Let us note that while the explicit approach is what would have to be
called the classical modeling approach, the implicit one appears generally
only as an a posteriori interpretation of dissipative properties for certain
numerical methods used.

This section describes the main functional models of the energy cascade mechanism. Those derived in the Fourier space, conceived for simulations based on
spectral numerical methods, and models derived in the physical space, suited
to the other numerical methods, are presented separately.
5.3.1 Spectral Models
The models belonging to this category are all eective viscosity models drawing upon the analyses of Kraichnan for the canonical case presented above.
The following models are described:
1. The CholletLesieur model (p. 106) which, based on the results of the
canonical analysis (inertial range of the spectrum with a slope of -5/3,
sharp cuto lter, no eects associated with a production type spectrum)
yields an analytical expression for the eective viscosity as a function of
the wave number considered and the cuto wave number. It will reect
the local eects at the cuto, i.e. the increase in the energy transfer toward the subgrid scales. This model explicitly brings out a dependency
106
2.
3.
4.
5.
of the eective viscosity as a function of the kinetic energy at the cuto.

This guarantees that, when all the modes of the exact solution are resolved, the subgrid model automatically cancels out. The fact that this
information is local in frequency allows the model to consider (at least
partially) the spectral disequilibrium phenomena that occur at the level
of the resolved scales11 , though without relaxing the hypotheses underlying the canonical analysis. Only the amplitude of the transfers is variable,
and not their pre-supposed shape.
The eective viscosity model (p. 107), which is a simplication of the
previous one and is based on the same assumptions. The eective viscosity is then independent of the wave number and is calculated so as to
ensure the same average value as the CholletLesieur model. It is simpler
to compute, but does not reect the local eects at the cuto.
The dynamic spectral model (p. 107), which is an extension of the
CholletLesieur model for spectra having a slope dierent from that of
the canonical case (i.e. - 5/3). Richer information is considered here:
while the CholletLesieur model is based only on the energy level at the
cuto, the dynamic spectral model also incorporates the spectrum slope
at the cuto. With this improvement, we can cancel the subgrid model in
certain cases for which the kinetic energy at cuto is non-zero but where
the kinetic energy transfer to the subgrid modes is zero12 . This model
also reects the local eects at the cuto. The other basic assumptions
underlying the CholletLesieur model are maintained.
The LesieurRogallo model (p. 108), which computes the intensity of the
transfers by a dynamic procedure. This is an extension of the Chollet
Lesieur model for ows in spectral disequilibrium, as modications in
the nature of the transfers to the subgrid scales can be considered. The
dynamic procedure consists in including in the model information relative
to the energy transfers at play with the highest resolved frequencies. The
assumptions concerning the lter are not relaxed, though.
Models based on the analytical theories of turbulence (p. 108), which
compute the eective viscosity without assuming anything about the
spectrum shape of the resolved scales, are thus very general. On the other
hand, the spectrum shape of the subgrid scales is assumed to be that of
a canonical inertial range. These models, which are capable of including
very complex physical phenomena, require very much more implementation and computation eort than the previous models. The assumptions
concerning the lter are the same as for the previous models.
CholletLesieur Model. Subsequent to Kraichnans investigations, Chollet and Lesieur [136] proposed an eective viscosity model using the results
of the EDQNM closure on the canonical case. The full subgrid transfer term
11
12
This is by their action on the transfers between resolved scales and the variations
induced on the energy level at cuto.
As is the case, for example, for two-dimensional ows.
107
including the backward cascade is written:

e
(k|kc ) = 2k 2 E(k)e (k|kc ) ,
Tsgs
(5.15)
in which the eective viscosity e (k|kc ) is dened as the product

e (k|kc ) = e+ (k|kc )e
(5.16)
The constant term e , independent of k, corresponds the asymptotic

value of the eective viscosity for wave numbers that are small compared
with the cuto wave number kc . This value is evaluated using the cuto
energy E(kc ):
'
3/2
e = 0.441K0
E(kc )
kc
(5.17)
The function e (k|kc ) reects the variations of the eective viscosity in

the proximity of the cuto. The authors propose the following form, which
is obtained by approximating the exact solution with a law of exponential
form:
(5.18)
e+ (k|kc ) = 1 + 34.59 exp(3.03kc /k) .
This form makes it possible to obtain an eective viscosity that is nearly
independent of k for wave numbers that are small compared with kc , with
a nite increase near the cuto. There is a limited inclusion of the backward
cascade with this model: the eective viscosity remains strictly positive for all
wave numbers, while the backward cascade is dominant for very small wave
numbers, which would correspond to negative values of the eective viscosity.
Constant Eective Viscosity Model. A simplied form of the eective
viscosity of (5.16) can be derived independently of the wave number k [440].
By averaging the eective viscosity along k and assuming that the subgrid
modes are in a state of energy balance, we get:
'
2 3/2 E(kc )
e (k|kc ) = e = K0
.
(5.19)
3
kc
Dynamic Spectral Model. The asymptotic value of the eective viscosity
(5.17) has been extended to the case of spectra of slope m by Metais
and Lesieur [514] using the EDQNM closure. For a spectrum proportional to
k m , m 3, we get:
'
E(kc )
5
m
3/2
e (m) = 0.31
3 mK0
.
(5.20)
m+1
kc
For m > 3, the energy transfer cancels out, inducing zero eective viscosity. Here, we nd a behavior similar to that of two-dimensional turbulence.
Extension of this idea in physical space has been derived by Lamballais and
his coworkers [422, 675].
108
LesieurRogallo Model. By introducing a new ltering level corresponding to the wave number km < kc , Lesieur and Rogallo [441] propose a dynamic
algorithm for adapting the CholletLesieur model. The contribution to the
transfer T (k), k < kc , corresponding to the (k, p, q) triads such that p and/or
q are in the interval [km , kc ], can be computed explicitly by Fourier transforms. This contribution is denoted Tsub (k|km , kc ) and is associated with the
eective viscosity:
e (k|km , kc ) =
Tsub (k|km , kc )
2k 2 E(k)
(5.21)
The eective viscosity corresponding to the interactions with wave numbers located beyond km is the sum:
e (k|km ) = e (k|km , kc ) + e (k|kc ) .
(5.22)
This relation corresponds exactly to Germanos identity and was previously derived by the authors. The two terms e (k|km ) and e (k|kc ) are then
modeled by the CholletLesieur model. We adopt the hypothesis that when
k < km , then k kc , which leads to e+ (k|kc ) = e+ (0). Relation (5.22) then
leads to the equation:
'

4/3
km
km
+
e (k|km ) = e (k|km , kc )
+ e+ (0)
.
(5.23)
E(km )
kc
The factor e+ (0) is evaluated by considering that we have the relations
e+ (k|km ) e+ (0),
e (k|km , kc ) e (0|km , kc ) ,
(5.24)
for k km , which leads to:

'
e+ (0)
= e (0|km , kc )
(

4/3 )1
km
km
1
E(km )
kc
(5.25)
Models Based on Analytical Theories of Turbulence. The eective

viscosity models presented above are all based on an approximation of the
eective viscosity prole obtained in the canonical case, and are therefore
intrinsically linked to the underlying hypotheses, especially those concerning
the shape of the energy spectrum. One way of relaxing this constraint is to
compute the eective viscosity directly from the computed spectrum using
analytical theories of turbulence. This approach has been used by Aupoix [24],
Chollet [132, 133], and Bertoglio [56, 57, 58].
More recently, following the recommendations of Leslie and Quarini,
which are to model the forward and backward cascade mechanisms separately,
109
Chasnov [120] in 1991 proposed an eective viscosity model considering only

the energy draining eects, with the backward cascade being modeled separately (see Sect. 5.4). Starting with an EDQNM analysis, Chasnov proposes
computing the eective viscosity e (k|kc ) as:

p2
q2
3
3
(xy + z )E(q) + (xz + y )E(p) ,
dp
dqkpq
q
p
kc
pk
(5.26)
in which x, y and z are geometric factors associated with the (k, p, q) triads
and kpq a relaxation time. These terms are explained in Appendix B. To
compute this integral, the shape of the energy spectrum beyond the cuto kc
must be known. As it is not known a priori, it must be specied elsewhere.
In practice, Chasnov uses a Kolmogorov spectrum extending from the cuto to innity. To simplify the computations, the relation (5.26) is not used
outside the interval [kc p 3kc ]. For wave numbers p > 3kc , the following
simplied asymptotic form already proposed by Kraichnan is used:
1
e (k|kc ) = 2
2k
1
e (k|kc ) =
15
kc

E(p)
dpkpq 5E(p) + p
p
(5.27)
5.3.2 Physical Space Models

Subgrid Viscosity Concept. The forward energy cascade mechanism to
the subgrid scales is modeled explicitly using the following hypothesis:
Hypothesis 5.2 The energy transfer mechanism from the resolved to the
subgrid scales is analogous to the molecular mechanisms represented by the
diusion term, in which the viscosity appears.
This hypothesis is equivalent to assuming that the behavior of the subgrid
scales is analogous to the Brownian motion superimposed on the motion of
the resolved scales. In gaskinetics theory, molecular agitation draws energy
from the ow by way of molecular viscosity. So the energy cascade mechanism
will be modeled by a term having a mathematical structure similar to that
of molecular diusion, but in which the molecular viscosity will be replaced
by a subgrid viscosity denoted sgs . As Boussinesq proposed, this choice of
mathematical form of the subgrid model is written:

d = sgs (u + T u)
(5.28)
in which d is the deviator of , i.e.:

1
ijd ij kk ij
3
(5.29)
110
The complementary spherical tensor 13 kk ij is added to the ltered static

pressure term and consequently requires no modeling. This decomposition is
necessary since the tensor (u + T u) has a zero trace, and we can only
model a tensor that also has a zero trace. This leads to the denition of the
modied pressure :
1
(5.30)
= p + kk .
3
It is important to note that the modied pressure and ltered pressure
p may take very dierent values when the generalized subgrid kinetic energy becomes large [374]. The closure thus now consists in determining the
relation:
sgs = N (u) .
(5.31)
The use of hypothesis (5.2) and of a model structured as above calls for
a few comments.
Obtaining a scalar subgrid viscosity requires the adoption of the following
hypothesis:
Hypothesis 5.3 A characteristic length l0 and a characteristic time t0 are
sucient for describing the subgrid scales.
Then, by dimensional reasoning similar to Prandtls, we arrive at:
sgs
l02
t0
(5.32)
Models of the form (5.28) are local in space and time, which is a necessity
if they are to be used in practice. This local character, similar to that of the
molecular diusion terms, implies [26, 405, 813]:
Hypothesis 5.4 (Scale Separation Hypothesis) There exists a total
separation between the subgrid and resolved scales.
A spectrum verifying this hypothesis is presented in Fig. 5.10.
Using L0 and T0 to denote the characteristic scales, respectively, of the
resolved eld in space and time, this hypothesis can be reformulated as:
l0
1,
L0
t0
1 .
T0
(5.33)
This hypothesis is veried in the case of molecular viscosity. The ratio

between the size of the smallest dynamically active scale, K , and the mean
free path fp of the molecules of a gas is evaluated as:
fp
Ma

K
Re1/4
(5.34)
111
Fig. 5.10. Energy spectrum corresponding to a total scale separation for cuto
wave number kc .
where Ma is the Mach number, dened as the ratio of the uid velocity
to the speed of sound, and Re the Reynolds number [708]. In most of the
cases encountered, this ratio is less than 103 , which ensures the pertinence
of using a continuum model. For applications involving rareed gases, this
ratio can take on much higher values of the order of unity, and the Navier
Stokes equations are then no longer an adequate model for describing the
uid dynamics.
Filtering associated to large-eddy simulation does not introduce such
a separation between resolved and subgrid scales because the turbulent energy spectrum is continuous. The characteristic scales of the smallest resolved
scales are consequently very close to those of the largest subgrid scales13 . This
continuity originates the existence of the spectrum region located near the
cuto, in which the eective viscosity varies rapidly as a function of the wave
number. The result of this dierence in nature with the molecular viscosity is
that the subgrid viscosity is not a characteristic of the uid but of the ow.
Let us not that Yoshizawa [786, 788], using a re-normalization technique, has
shown that the subgrid viscosity is dened as a fourth-order non-local tensor
in space and time, in the most general case. The use of the scale separation hypothesis therefore turns out to be indispensable for constructing local
models, although it is contrary to the scale similar hypothesis of Bardina et
al. [40], which is discussed in Chap. 7.
It is worth noting that subgrid-viscosity based models for the forward
energy cascace induce a spurious alignment of the eigenvectors for resolved
strain rate tensor and subgrid-scale tensor, because they are expressed as
13
This is all the more true for smooth lters such as the Gaussian and box lters,
which allow a frequency overlapping between the resolved and subgrid scales.
112
d (u + T u).14 Tao et al. [703, 704] and Horiuti [325] have shown that
this alignment is unphysical: the eigenvectors for the subgrid tensor have
a strongly preferred relative orientation of 35 to 45 degrees with the resolved
strain rate eigenvectors.
The modeling problem consists in determining the characteristic scales l0
and t0 .
Model Types. The subgrid viscosity models can be classied in three categories according to the quantities they bring into play [26]:
1. Models based on the resolved scales (p. 113): the subgrid viscosity is
evaluated using global quantities related to the resolved scales. The existence of subgrid scales at a given point in space and time will therefore
be deduced from the global characteristics of the resolved scales, which
requires the introduction of assumptions.
2. Models based on the energy at the cuto (p. 116): the subgrid viscosity
is calculated from the energy of the highest resolved frequency. Here, it
is a matter of information contained in the resolved eld, but localized in
frequency and therefore theoretically more pertinent for describing the
phenomena at cuto than the quantities that are global and thus not
localized in frequency, which enter into the models of the previous class.
The existence of subgrid scales is associated with a non-zero value of the
energy at cuto15 .
3. Models based on the subgrid scales (p. 116), which use information directly related to the subgrid scales. The existence of the subgrid scales is
no longer determined on the basis of assumptions concerning the characteristics of the resolved scales as it is in the previous cases, but rather
directly from this additional information. These models, because they are
richer, also theoretically allow a better description of these scales than
the previous models.
These model classes are presented in the following. All the developments
are based on the analysis of the energy transfers in the canonical case. In
order to be able to apply the models formulated from these analyses to more
realistic ows, such as the homogeneous isotropic ows associated with a production type spectrum, we adopt the assumption that the lter cuto fre14
15
But it must be also remembered that the purpose of these functional models
is not to predict the subgrid tensor, but just to enforce the correct resolved
kinetic energy balance. This reconstruction of the subgrid tensor is nothing but
an a posteriori interpretation. This fact is used by Germano to derive new subgrid
viscosity models [252].
This hypothesis is based on the fact that the energy spectrum E(k) of an isotropic
turbulent ow in spectral equilibrium corresponding to a Kolmogorov spectrum
is a monotonic continuous decreasing function of the wave number k. If there
exists a wave number k such that E(k ) = 0, then E(k) = 0, k > k . Also, if
the energy is non-zero at the cuto, then subgrid modes exist, i.e. if E(kc ) = 0,
then there exists a neighbourhood kc = [kc , kc + c ], c > 0 such that E(kc )
E(k) 0 k kc .
113
quency is located suciently far into the inertial range for these analyses to
remain valid (refer to Sect. 5.1.2). The use of these subgrid models for arbitrary developed turbulent ows (anisotropic, inhomogeneous) is justied by
the local isotropy hypothesis: we assume then that the cuto occurs in the
scale range that veries this hypothesis.
The case corresponding to an isotropic homogeneous ow associated with
a production spectrum is represented in Fig. 5.11. Three energy uxes are
dened: the injection rate of turbulent kinetic energy into the ow by the
driving mechanisms (forcing, instabilities), denoted I ; the kinetic energy
transfer rate through the cuto, denoted *; and the kinetic energy dissipation
rate by the viscous eects, denoted .
Models Based on the Resolved Scales. These models are of the generic form:

sgs = sgs ,
,
(5.35)
in which is the characteristic cuto length of the lter and the instantaneous energy ux through the cuto. We implicitly adopt the assumption
here, then, that the subgrid modes exist, i.e. that the exact solution is not
entirely represented by the ltered eld when this ux is non-zero.
First Method. Simple dimensional analysis shows that:
4/3
sgs 1/3
(5.36)
Fig. 5.11. Dynamics of the kinetic energy in the spectral space. The energy is
injected at the rate I . The transfer rate through the cuto, located wave number
kc , is denoted . The dissipation rate due to viscous eects is denoted . The local
equilibrium hypothesis is expressed by the equality I = = .
114
Reasoning as in the framework of Kolmogorovs hypotheses for isotropic

homogeneous turbulence, for the case of an innite inertial spectrum of the
form
(5.37)
E(k) = K0 2/3 k 5/3 , K0 1.4 ,
in which is the kinetic energy dissipation rate, we get the equation:
sgs =
A
4/3

1/3
K0 4/3
(5.38)
in which the constant A is evaluated as A = 0.438 with the TFM model

and as A = 0.441 by the EDQNM theory [26]. The angle brackets operator
, designates a statistical average. This statistical averaging operation is
intrinsically associated with a spatial mean by the fact of the ows spatial
homogeneity and isotropy hypotheses. This notation is used in the following
to symbolize the fact that the reasoning followed in the framework of isotropic
homogeneous turbulence applies only to the statistical averages and not to
the local values in the physical space. The problem is then to evaluate the
average ux
. In the isotropic homogeneous case, we have:

2|S|2 = 2S ij S ij =
kc
2k 2 E(k)dk, kc =
(5.39)
If the cuto kc is located far enough into the inertial range, the above
relation can be expressed solely as a function of this regions characteristic
quantities. Using a spectrum of the shape (5.37), we get:
3
4/3
2|S|2 = 4/3 K0 2/3
2
(5.40)
Using the hypothesis16 [447]:

|S|3/2 |S|3/2
(5.41)
we get the equality:

=
1
2
3K0
2
3/2
2|S|2 3/2
(5.42)
In order to evaluate the dissipation rate from the information contained in the resolved scales, we assume the following:
Hypothesis 5.5 (Local Equilibrium Hypothesis) The ow is in constant spectral equilibrium, so there is no accumulation of energy at any frequency and the shape of the energy spectrum remains invariant with time.
16
The error margin measured in direct numerical simulations of isotropic homogeneous turbulence is of the order of 20% [507].
115
This implies an instantaneous adjustment of all the scales of the solution

to the turbulent kinetic energy production mechanism, and therefore equality
between the production, dissipation, and energy ux through the cuto:
I =
= .
(5.43)
Using this equality and relations (5.38) and (5.42), we get the closure
relation:

2
sgs = C 2|S|2 1/2 ,
(5.44)
where the constant C is evaluated as:

1/4
A
3K0
C=
0.148 .
2
K0
(5.45)
Second Method. The local equilibrium hypothesis allows:

=
S ij ij = sgs 2S ij S ij .
(5.46)
The idea is then to assume that:

sgs 2S ij S ij = sgs 2S ij S ij .
(5.47)
By stating at the outset that the subgrid viscosity is of the form (5.44)
and using relation (5.40), a new value is found for the constant C:
C=
3K0
2
3/4
0.18 .
(5.48)
We note that the value of this constant is independent of the cuto wave
number kc , but because of the way it is calculated, we can expect a dependency as a function of the spectrum shape.
Alternate Form. This modeling induces a dependency as a function of the
cuto length and the strain rate tensor S of the resolved velocity eld. In
the isotropic homogeneous case, we have the equality:
2|S|2 = , = u .
(5.49)
By substitution, we get the equivalent form [487]:

2
sgs = C 1/2
(5.50)
These two versions bring in the gradients of the resolved velocity eld.
This poses a problem of physical consistency since the subgrid viscosity is
non-zero as soon as the velocity eld exhibits spatial variations, even if it is
laminar and all the scales are resolved. The hypothesis that links the existence
of the subgrid modes to that of the mean eld gradients therefore prevents
116
us from considering the large scale intermittency and thereby requires us

to develop models which by nature can only be eective for dealing with
ows that are completely turbulent and under-resolved everywhere17. Poor
behavior can therefore be expected when treating intermittent or weakly
developed turbulent ows (i.e. in which the inertial range does not appear in
the spectrum) due to too strong an action by the model.
Models Based on the Energy at Cuto. The models of this category are based
on the intrinsic hypothesis that if the energy at the cuto is non-zero, then
subgrid modes exist.
First Method. Using relation (5.38) and supposing that the cuto occurs
within an inertial region, i.e.:
E(kc ) = K0 2/3 kc5/3
(5.51)
by substitution, we get:
'
A
sgs =
K0
E(kc )
, kc = / .
kc
(5.52)
This model raises the problem of determining the energy at the cuto
in the physical space, but on the other hand ensures that the subgrid viscosity will be null if the ow is well resolved, i.e. if the highest-frequency
mode captured by the grid is zero. This type of model thus ensures a better
physical consistency than those models based on the large scales. It should
be noted that it is equivalent to the spectral model of constant eective viscosity.
Second Method. As in the case of models based on the large scales, there
is a second way of determining the model constant. By combining relations (5.46) and (5.51), we get:
'
E(kc )
2
sgs =
.
(5.53)
3/2
kc
3K
0
Models Based on Subgrid Scales. Here we considers models of the form:

2
sgs = sgs , qsgs
,
,
(5.54)
2
in which qsgs
is the kinetic energy of the subgrid scales and the kinetic
energy dissipation rate18 . These models contain more information about the
17
18
In the sense that the subgrid modes exist at each point in space and at each time
step.
Other models are of course possible using other subgrid scale quantities like
a length or time scale, but we limit ourselves here to classes of models for which
practical results exist.
117
subgrid modes than those belonging to the two categories described above,
and thereby make it possible to do without the local equilibrium hypothesis (5.5) by introducing characteristic scales specic to the subgrid modes by
2
and . This capacity to handle the energy disequilibrium is
way of qsgs
expressed by the relation:
*
ij S ij = ,
(5.55)
which should be compared with (5.43). In the case of an inertial range extending to innity beyond the cuto, we have the relation:
2

qsgs
1
ui ui =
2
E(k)dk =
kc
3
K0 2/3 kc2/3
2
(5.56)
from which we deduce:

=
kc
q 2 3/2
(3K0 /2)3/2 sgs
(5.57)
By introducing this last equation into relation (5.38), we come to the

general form:
1+/3
2 (1)/2
sgs = C /3 qsgs
in which
A
C =
K0 4/3
3K0
2
(5.58)
(5.59)
(1)/2
(1)/3
and in which is a real weighting parameter. Interesting forms of sgs have

been found for certain values:
For = 1, we get
sgs =
A
4/3
1/3
K0 4/3
(5.60)
This form uses only the dissipation and is analogous to that of the models
based on the resolved scales. If the local equilibrium hypothesis is used,
these two types of models are formally equivalent.
For = 0, we get

2 1/2
2 A
sgs =
qsgs

.
(5.61)
3/2
3 K
0
This model uses only the kinetic energy of the subgrid scales. As such,
it is formally analogous to the denition of the diusion coecient of an
ideal gas in the framework of gaskinetics theory. In the case of an inertial
118
spectrum extending to innity beyond the cuto, this model is strictly

equivalent to the model based on the energy at cuto, since in this precise
case we have the relation:
3
2
kc E(kc ) = qsgs

2
(5.62)
For = 3, we have:
2 2

4A qsgs
sgs =
3
9K0
(5.63)
This model is formally analogous to the k statistical model of turbulence

for the Reynolds Averaged NavierStokes equations, and does not bring in
the lter cuto length explicitly.
2
The closure problem consists in determining the quantities and qsgs
.
To do this, we can introduce one or more equations for the evolution of
these quantities or we can deduce them from the information contained in
the resolved eld. As these quantities represent the subgrid scales, we are
justied in thinking that, if they are correctly evaluated, the subgrid viscosity
will be negligible when the ow is well resolved numerically. However, it
should be noted that these models in principle require more computation than
those based on the resolved scales, because they produce more information
concerning the subgrid scales.
Extension to Other Spectrum Shapes. The above developments are based on

a Kolmogorov spectrum, which reects only the existence of a region of similarity of the real spectra. This approach can be extended to other more
realistic spectrum shapes, mainly including the viscous eects. Several extensions of the models based on the large scales were proposed by Voke [737]
for this. The total dissipation can be decomposed into the sum of the
dissipation associated with the large scales, denoted r , and the dissipation
associated with the subgrid scales, denoted sgs , (see Fig. 5.12):
= r + sgs .
(5.64)
These three quantities can be evaluated as:

=
r =
2(sgs + )|S|2 ,
kc
2
2|S| = 2
k 2 E(k)dk
0
sgs =
2sgs |S|2 = Cs
(5.65)
,

3/2
2|S|2
(5.66)
,
(5.67)
from which we get:

1
r
=
,

1 +
sgs
(5.68)
119
Fig. 5.12. Kinetic energy dynamics in the spectral space. The energy is injected
at the rate I . The transfer rate through the cuto located at the wave number
kc is denoted . The dissipation rate in the form of heat by the viscous eects
associated with the scales located before and after the cuto kc are denoted r and
sgs , respectively.
This ratio is evaluated by calculating the r term analytically from the

chosen spectrum shapes, which provides a way of then computing the subgrid
viscosity sgs .
We dene the three following parameters:
=
k
=k
kd
3
1/4
,
Re =
c =
kc
kd
(5.69)
+
2|S|2
(5.70)
in which kd is the wave number associated with the Kolmogorov scale (see
Appendix A), and Re is the mesh-Reynolds number. Algebraic substitutions
lead to:
1/2
(5.71)
= Re (1 + )1/4 .
The spectra studied here are of the generic form:
E(k) = K0 2/3 k 5/3 f () ,
(5.72)
in which f is a damping function for large wave numbers. The following are
some commonly used forms of this function:
120
HeisenbergChandrasekhar spectrum:
(
f () = 1 +
3K0
2
)4/3
3
(5.73)
Kovasznay spectrum:
f () =

2
K0 4/3
1
(5.74)
Note that this function cancels out for = (2/K0 )3/4 , which requires that
the spectrum be forced to zero for wave numbers beyond this limit.
Pao spectrum:

3K0 4/3
f () = exp
.
(5.75)
2
These three spectrum shapes are graphed in Fig. 5.13. An analytical integration leads to:
For the HeisenbergChandrasekhar spectrum:
r
= 4/3
c
(
2
3K0
)1/3
3
+
4c
(5.76)
Fig. 5.13. Graph of HeisenbergChandrasekhar, Kovasznay, and Pao spectra, for

kd = 1000.
or:
sgs =
4/3
c
(
2
3K0
)1/3
3
+ 4c
121
(5.77)
For the Kovazsnay spectrum:

3
K0 4/3
r
=1 1
c

2
or:
(5.78)

3 )1
K0 4/3
c
sgs =
1 1
1
(5.79)
For the Pao spectrum:

r
= 1 exp

or:
sgs =

1 exp
3K0 4/3
2 c
3K0 4/3
2 c
3
,
3 )1
1
(5.80)
(5.81)
These new estimates of the subgrid viscosity sgs make it possible to

take the viscous eects into account, but requires that the spectrum shape
be set a priori, as well as the value of the ratio c between the cuto wave
number kc and the wave number kd associated with the Kolmogorov scale.
Inclusion of the Local Eects at Cuto. The subgrid viscosity models in the
physical space, such as they have been developed, do not reect the increase in
the coupling intensity with the subgrid modes when we consider modes near
the cuto. These models are therefore analogous to that of constant eective
viscosity. To bring out these eects in the proximity of the cuto, Chollet [134], Ferziger [217], Lesieur and Metais [440], Deschamps [176], Borue
and Orszag [68, 69, 70, 71], Winckelmans and co-workers [762, 163, 759] and
Layton [427] propose introducing high-order dissipation terms that will have
a strong eect on the high frequencies of the resolved eld without aecting
the low frequencies.
e
can
Chollet [134], when pointing out that the energy transfer term Tsgs
be written in the general form
e
(k|kc ) = 2e(n) (k|kc )k 2n E(k) ,
Tsgs
(5.82)
122

(n)
in which e (k|kc ) is a hyper-viscosity, proposes modeling the subgrid term

in the physical space as the sum of an ordinary subgrid viscosity model and
a sixth-order dissipation. This new model is expressed:

(5.83)
= sgs C1 2 + C2 6 u ,
in which C1 and C2 are constants. Ferziger proposes introducing a fourthorder dissipation by adding to the subgrid tensor the tensor (4) , dened
as:

2
2
(4)
(4) ui
(4) uj
ij =
sgs
+
sgs
,
(5.84)
xj
xk xk
xi
xk xk
or as
(4)
ij =
2
xk xk

ui
uj
(4)
+
sgs
xj
xi
(5.85)
(4)
in which the hyper-viscosity sgs is dened by dimensional arguments as

4
(4)
sgs
= Cm |S| .
(5.86)
The full subgrid term that appears in the momentum equations is then
written:
(2)
(4)
,
(5.87)
ij = ij + ij
(2)
in which ij is a subgrid viscosity model described above. A similar form is

proposed by Lesieur and Metais: after dening the velocity eld u as
u = 2p u
(5.88)
the two authors propose the composite form:

Sij
ij = sgs S ij + (1)p+1 sgs
(5.89)

hyper-viscosity obtained by applying a subgrid viscosity model
in which sgs

to the u eld, and S the strain rate tensor computed from this same eld.
The constant of the subgrid model used should be modied to verify the local
equilibrium relation, which is
ij S ij = .
This composite form of the subgrid dissipation has been validated experimentaly by Cerutti et al. [115], who computed the spectral distribution of
dissipation and the corresponding spectral viscosity from experimental data.
It is worth noting that subgrid dissipations dened thusly, as the sum of
second- and fourth-order dissipations, are similar in form to certain numerical schemes designed for capturing strong gradients, like that of Jameson et
al. [346].
123
Borue and Orszag [68, 69, 70, 71] propose to eliminate the molecular and
the subgrid viscosities by replacing them by a higher power of the Laplacian
operator. Numerical tests show that three-dimensional inertial-range dynamics is relatively independent of the form of the hyperviscosity. It was also
shown that for a given numerical resolution, hyperviscous dissipations increase the extent of the inertial range of three-dimensional turbulence by an
order of magnitude. It is worth noting that this type of iterated Laplacian
is commonly used for two-dimensional simulations. Borue and Orszag used
a height-time iterated Laplacian to get these conclusions. Such operators are
easily dened when using spectral methods, but are of poor interest when
dealing with nite dierence of nite volume techniques.
Subgrid-Viscosity Models. Various subgrid viscosity models belonging
to the three categories dened above will now be described. These are the
following:
1. The Smagorinsky model (p. 124), which is based on the resolved scales.
This model, though very simple to implant, suers from the defects already mentioned for the models based on the large scales.
2. The second-order Structure Function model developed by Metais and
Lesieur (p. 124), which is an extension into physical space of the models based on the energy at cuto. Theoretically based on local frequency
information, this model should be capable of treating large-scale intermittency better than the Smagorinsky model. However, the impossibility
of localizing the information in both space and frequency (see discussion
further on) reduces its eciency.
3. The third-order Structure Function models developed by Shao (p. 126),
which can be interpreted as an extension of the previous model based
on the second-order structure function. The use of the Kolmogorov
Meneveau equation [510] for the ltered third-order structure function
enables the denition of several models which do not contain arbitrary
constants and have improved potentiality for non-equilibrium ows.
4. A model based on the kinetic energy of the subgrid modes (p. 128). This
energy is considered as an additional variable of the problem, and is
evaluated by solving an evolution equation. Since it contains information
relative to the subgrid scales, it is theoretically more capable of handling
large-scale intermittency than the previous model. Moreover, the local
equilibrium hypothesis can be relaxed, so that the spectral nonequilibrium can be integrated better. The model requires additional hypotheses,
though (modeling, boundary conditions).
5. The Yoshizawa model (p. 129), which includes an additional evolution
equation for a quantity linked to a characteristic subgrid scale, by which
it can be classed among models based on the subgrid scales. It has the
same advantages and disadvantages as the previous model.
6. The Mixed Scale Model (p. 130), which uses information related both to
the subgrid modes and to the resolved scales, though without incorpo-
124
rating additional evolution equations. The subgrid scale data is deduced

from that contained in the resolved scales by extrapolation in the frequency domain. This model is of intermediate level of complexity (and
quality) between those based on the large scales and those that use additional variables.
Smagorinsky Model. The Smagorinsky model [676] is based on the large
scales. It is generally used in a local form the physical space, i.e. variable
in space, in order to be more adaptable to the ow being calculated. It is
obtained by space and time localization of the statistical relations given in
the previous section. There is no particular justication for this local use
of relations that are on average true for the whole, since they only ensure
that the energy transfers through the cuto are expressed correctly on the
average, and not locally.
This model is expressed:

2
sgs (x, t) = Cs (2|S(x, t)|2 )1/2
(5.90)
The constant theoretical value Cs is evaluated by the relations (5.45)

or (5.48). It should nonetheless be noted that the value of this constant is,
in practice, adjusted to improve the results. Clark et al. [143] use Cs = 0.2
for a case of isotropic homogeneous turbulence, while Deardor [172] uses
Cs = 0.1 for a plane channel ow. Studies of shear ows using experimental
data yield similar evaluations (Cs 0.1 0.12) [503, 570, 724]. This decrease
in the value of the constant with respect to its theoretical value is due to
the fact that the eld gradient is now non-zero and that it contributes to
the |S(x, t)| term. To enforce the local equilibrium relation, the value of the
constant has to be reduced. It should be noted that this new value ensures
only that the right quantity of resolved kinetic energy will be dissipated on
the whole throughout the eld, but that the quality of the level of local
dissipation is uncontrolled. 19
Second-Order Structure Function Model. This model is a transposition
of Metais and Lesieurs constant eective viscosity model into the physical
space, and can consequently be interpreted as a model based on the energy
at cuto, expressed in physical space. The authors [514] propose evaluating
the energy at cuto E(kc ) by means of the second-order velocity structure
19
Canuto and Cheng [99] derived a more general expression for the constant Cs ,
which appears as an explicit function of the subgrid kinetic energy and the local
shear:

1/2
2
qsgs
Cs
,
2
|S|
which is eectively a decreasing function of the local shear |S|. That demonstrates
the limited theoretical range of application of the usual Smagorinsky model.
function. This is dened as:

DLL (x, r, t) =
[u(x, t) u(x + x , t)] d3 x

2
|x |=r
125
(5.91)
In the case of isotropic homogeneous turbulence, we have the relation:

sin(kr)
E(k, t) 1
DLL (r, t) = DLL (x, r, t)d3 x = 4
dk . (5.92)
kr
0
Using a Kolmogorov spectrum, the calculation of (5.92) leads to:
DLL (r, t) =
9
(1/3)K02/3 r2/3
5
(5.93)
or, expressing the dissipation , as a function of DLL (r, t) in the expression

for the Kolmogorov spectrum:
E(k) =
5
DLL (r, t)r2/3 k 5/3
9 (1/3)
(5.94)
To derive a subgrid model, we now have to evaluate the second-order

structure function from the resolved scales alone. To do this, we decompose
by:
(5.95)
DLL (r, t) = DLL (r, t) + C0 (r, t) ,
in which DLL (r, t) is computed from the resolved scales and C0 (r, t) corresponds to the contribution of the subgrid scales:

C0 (r, t) = 4
kc

sin(kr)
E(k, t) 1
dk
kr
(5.96)
By replacing the quantity E(k, t) in equation (5.96) by its value (5.94),

we get:
2/3
r
Hsf (r/) ,
(5.97)
C0 (r, t) = DLL (r, t)
in which Hsf is the function

20
3
2/3
Hsf (x) =
+
x
Im
{exp(5/6)
(5/3,
x)}
9 (1/3) 2 2/3
. (5.98)
Once it is substituted in (5.95), this equation makes it possible to evaluate

the energy at the cuto. The second-order Structure Function model takes
the form:
+
sgs (r) = A(r/) DLL (r, t) ,
(5.99)
126
in which
A(x) =
3/2

1/2
2K
0
x4/3 1 x2/3 Hsf (x)
3 4/3 (9/5) (1/3)
(5.100)
In the same way as for the Smagorinsky model, a local model in space
can be had by using relation (5.94) locally in order to include the local intermittency of the turbulence. The model is then written:
+
sgs (x, r) = A(r/) DLL (x, r, t) .
(5.101)
In the case where r = , the model takes the simplied form:
+
sgs (x, , t) = 0.105 DLL (x, , t) .
(5.102)
A link can be established with the models based on the resolved scale
gradients by noting that:
u(x, t) u(x + x , t) = x u(x, t) + O(|x |2 )
(5.103)
This last relation shows that the function F 2 is homogeneous to a norm of

the resolved velocity eld gradient. If this function is evaluated in the simulation in a way similar to how the resolved strain rate tensor is computed for the
Smagorinsky model, we can in theory expect the Structure Function model to
suer some of the same weaknesses: the information contained in the model
will be local in space, therefore non-local in frequency, which induces a poor
estimation of the kinetic energy at cuto and a loss of precision of the model
in the treatment of large-scale intermittency and spectral nonequilibrium.
Third-Order Structure Functions Models. Shao et al. [669, 153] dened
a more general class of structure function-based subgrid viscosities considering the KolmogorovMeneveau equation for the third-order velocity structure
function in isotropic turbulence [510]
4
r = DLLL 6GLLL
5
(5.104)
where DLLL is the third-order longitudinal velocity correlation of the ltered

eld
DLLL (r) = [u(x + r) u(x)]3 ,
(5.105)
where denotes the statistical average operator (which is equivalent to the
integral sequence introduced in the presentation of the second-order structure
function model). The two other quantities are the longitudinal velocity-stress
correlation tensor
GLLL (r) = u1 (x)11 (x + r) ,
(5.106)
127
and the average subgrid dissipation

= ij S ij
(5.107)
Shaos procedure consists in using relation (5.104) to compute the subgrid

viscosity. To this end, he assumes that the velocity-stress correlation obeys
the following scale-similarity hypothesis
GLLL (r) rp
(5.108)
where p = 1/3 corresponds to the Kolmogorov local isotropy hypotheses.

Using that assumption, one obtains the following relationship for two space
increments r1 and r2 :
0.8r1 + DLLL (r1 )
=
0.8r2 + DLLL (r2 )
r1
r2
1/3
.
(5.109)
Now introducing a subgrid viscosity model with subgrid viscosity sgs ,

several possibilities exist for the evaluation of the subgrid viscosity. The three
following models have been proposed by Shao and his coworkers:
The one-scale, constant subgrid viscosity model. Assuming that the subgrid
viscosity is constant, one obtains
ij = 2sgs S ij ,
= sgs |S|2 ,
GLLL = sgs DLL,r
(5.110)
where comma separated indices denote derivatives. Inserting these relations

into (5.104), an expression for the eddy viscosity is recovered:
+
Sk
DLL ,
(5.111)
r
sgs =
2
0.8 D|S| 4
LL
where the skewness of the longitudinal ltered velocity increment is dened

as
DLLL
Sk = 3/2 .
(5.112)
DLL
The Metais-Lesieur model is recovered taking r = and using properties
of isotropic turbulence to evaluate the dierent terms appearing in (5.111).
The one-scale, variable subgrid viscosity model. Relaxing the previous contraint dealing with the constant appearing in the structure-function model,
one obtains the following asymptotic expression
+
Sk
r DLL .
(5.113)
sgs =
8
128
The constant is observed to be self-adpative, depending on the computed

value of the skewness parameter Sk .
The multiscale structure function model. The last model derived by Shao
is more general and is based on the two-scale relation (5.109). Using the
relation = sgs |S|2 and considering two dirent separation distances r1
and r2 , one obtain the following expression
sgs
1/3
DLLL (r1 ) rr12
D LLL (r2 )

=
4/3
0.4|S|2 1 rr12
r1
(5.114)
These expressions are made local in the physical space by using local
values of the dierent parameters involving the resolved velocity eld u.
Model Based on the Subgrid Kinetic Energy. One model, of the
form (5.61), based on the subgrid scales, was developed independently by
a number of authors [318, 653, 791, 792, 525, 690, 391]. The subgrid viscosity
2
:
is computed from the kinetic energy of the subgrid modes qsgs
sgs (x, t) = Cm
+
2 (x, t) ,
qsgs
(5.115)
where, for reference:

2
qsgs
(x, t) =
1
(ui (x, t) ui (x, t))2
2
(5.116)
The constant Cm is evaluated by the relation (5.61). This energy constitutes another variable of the problem and is evaluated by solving an evolution
equation. This equation is obtained from the exact evolution equation (3.33),
whose unknown terms are modeled according to Lillys proposals [447], or by
a re-normalization method. The various terms are modeled as follows (refer
to the work of McComb [464], for example):
The diusion term is modeled by a gradient hypothesis, by stating that
2
gradient
the non-linear term is proportional to the kinetic energy qsgs
(Kolmogorov-Prandtl relation):

2
+
qsgs
1

2
u u u + uj p = C2
qsgs
.
(5.117)
xj 2 i i j
xj
xj
The dissipation term is modeled using dimensional reasoning, by:
2
(qsgs
)
ui ui
=
= C1
2 xj xj
3/2
(5.118)
129
The resulting evolution equation is:

2
2
qsgs
uj qsgs
+
=
t
xj

3/2
2
qsgs
ij S ij C1

II
C2
xj

III

2
2
+
2 qsgs
q
sgs
2
qsgs
+
, (5.119)
xj
xj xj

IV
in which C1 and C2 are two positive constants and the various terms represent:
I - advection by the resolved modes,

II - production by the resolved modes,
III - turbulent dissipation,
IV - turbulent diusion,
V - viscous dissipation.
Using an analytical theory of turbulence, Yoshizawa [791, 792] and Horiuti [318] propose C1 = 1 and C2 = 0.1.
Yoshizawa Model. The lter cuto length, , is the only length scale
used in deriving models based on the large scales, as this derivation has
been explained above. The characteristic length associated with the subgrid
scales, denoted f , is assumed to be proportional to this length, and the
developments of Sect. 5.3.2 show that:
f = Cs .
(5.120)
The variations in the structure of the subgrid modes cannot be included

by setting a constant value for the coecient Cs , as is done in the case of
the Smagorinsky model, for example. To remedy this, Yoshizawa [787, 790]
proposes dierentiating these two characteristic scales and introducing an
additional evolution equation to evaluate f . This length can be evaluated
2
from the dissipation and the subgrid kinetic energy qsgs
by the relation:
f = C1
2 3/2
2 3/2
2
2 5/2
(qsgs
)
(qsgs
)
Dqsgs
(qsgs
)
D
+ C2
C
3
2
Dt
3
Dt
(5.121)
in which D/Dt is the material derivative associated with the resolved velocity
eld. The values of the constants appearing in equation (5.121) can be determined by an analysis conducted with the TSDIA technique [790]: C1 = 1.84,
C2 = 4.95 et C3 = 2.91. We now express the proportionality relation between
the two lengths as:
f = (1 + r(x, t)) .
(5.122)
130
By evaluating the subgrid kinetic energy as:

2/3

2
qsgs
= /C1
(5.123)
relations (5.121) and (5.122) lead to:

2/3 4/3 D
r = C4
Dt
(5.124)
with C4 = 0.04. Using the local equilibrium hypothesis, we get:

= ij S ij C5 2f |S|3
(5.125)
in which C5 = 6.52.103. This denition completes the calculation of the

factor r and the length f . This variation of the characteristic length f can
be re-interpreted as a variation of the constant in the Smagorinsky model:

2
2 D|S|
2
2 |S|
+ Cb |S|
Cs = Cs0 1 Ca |S|
|S|
Dt
xj
xj
. (5.126)
The constants Cs0 , Ca and Cb are evaluated at 0.16, 1.8, and 0.047, respectively, by Yoshizawa [790] and Murakami [554]. In practice, Cb is taken
to be equal to zero and the constant Cs is bounded in order to ensure the
stability of the simulation: 0.1 Cs 0.27. Morinishi and Kobayashi [541]
recommend using the values Ca = 32 and Cs0 = 0.1.
Mixed Scale Model. Ta Phuoc Loc and Sagaut [627, 626] dened models
having a triple dependency on the large and small structures of the resolved
eld as a function of the cuto length. These models, which make up the
one-parameter Mixed Scale Model family, are derived by taking a weighted
geometric average of the models based on the large scales and those based
on the energy at cuto:
sgs ()(x, t) = Cm |F(u(x, t))| (qc2 )
1
2
(x, t)
1+
(5.127)
with
F (u(x, t)) = S(x, t) or u(x, t) .
(5.128)
It should be noted that localized versions of the models are used here, so
that any ows that do not verify the spatial homogeneity property can be
processed better. The kinetic energy qc2 can be evaluated using any method
presented in Sect. 9.2.3. In the original formulation of the model, it is evaluated in the physical space by the formula:
qc2 (x, t) =
1
(ui (x, t)) (ui (x, t))
2
(5.129)
131
* is the resolved
Fig. 5.14. Spectral subdivisions for double sharp-cuto ltering. u
eld in the sense of the test lter, (u) the test eld, and u the unresolved scales
in the sense of the initial lter.
The test eld (u) represents the high-frequency part of the resolved velocity eld, dened using a second lter, referred to as the test lter, des* >
ignated by the tilde symbol and associated with the cuto length
(see Fig. 5.14):
* .
(u) = u u
(5.130)
The resulting model can be interpreted in two ways:
As a model based on the kinetic energy of the subgrid scales, i.e. the second
form of the models based on the subgrid scales in Sect. 5.3.2, if we use
Bardinas hypothesis of scale similarity (described in Chap. 7), which allows
us to set:
2
,
(5.131)
qc2 qsgs
2
is the kinetic energy of the subgrid scales. This assumption
in which qsgs
can be rened in the framework of the canonical analysis. Assuming that
the two cutos occur in the inertial range of the spectrum, we get:

qc2 =
kc
E(k)dk =
kc

3
2/3
K0 2/3 kc
kc 2/3
2
(5.132)
* respectively.
in which kc and kc are wave numbers associated with and ,
132
We then dene the relation:

(
2
, =
qc2 = qsgs
kc
kc
2/3
1
(5.133)
It can be seen that the approximation is exact if = 1, i.e. if:

1
kc = kc
8
(5.134)
This approximation is also used by Bardina et al. [40] and Yoshizawa et

al. [794] to derive models based on the subgrid kinetic energy without using
any additional transport equation.
As a model based on the energy at cuto, and therefore as a generalization
of the spectral model of constant eective viscosity into the physical space.
That is, using the same assumptions as before, we get:
3
kc E(kc ) .
2
Here, the approximation is exact if kc = kc / 8.

qc2 =
(5.135)
It is important to note that the Mixed Scale Model makes no use of any
commutation property between the test lter and the derivation operators.
Also, we note that for [0, 1] the subgrid viscosity sgs () is always dened,
whereas the model appears in the form of a quotient for other values of
can then raise problems of numerical stability once it is discretized, because
the denominator may cancel out.
The model constant can be evaluated theoretically by analytical theories
of turbulence in the canonical case. Resuming the results of Sect. 5.3.2, we
get:
Cm = Cq1 Cs2
(5.136)
in which Cs 0.18 or Cs 0.148 and Cq 0.20.

Some other particular cases of the Mixed Scale Model can be found. Wong
and Lilly [766], Carati [103] and Tsubokura [717] proposed using = 1,
yielding a model independent of the cuto length . Yoshizawa et al. [793]
used = 0, but introduced an exponential damping term in order to enforce
a satisfactory asymptotic near-wall behavior (see p. 159):
(
sgs =
0.03(qc2 )1/2
1 exp 21
qc2
2
|S|2
)
.
(5.137)
Mathematical analysis in the case = 0 was provided by Iliescu and

Layton [342] and Layton and Lewandowski [430].
133
5.3.3 Improvement of Models in the Physical Space

Statement of the Problem. Experience shows that the various models
yield good results when they are applied to homogeneous turbulent ows
and that the cuto is placed suciently far into the inertial range of the
spectrum, i.e. when a large part of the total kinetic energy is contained in
the resolved scales20 .
In other cases, as in transitional ows, highly anisotropic ows, highly
under-resolved ows, or those in high energetic disequilibrium, the subgrid
models behave much less satisfactorily. Aside from the problem stemming
from numerical errors, there are mainly two reasons for this:
1. The characteristics of these ows does not correspond to the hypotheses
on which the models are derived, which means that the models are at
fault. We then have two possibilities: deriving models from new physical
hypotheses or adjusting existing ones, more or less empirically. The rst
choice is theoretically more exact, but there is a lack of descriptions of
turbulence for frameworks other than that of isotropic homogeneous turbulence. Still, a few attempts have been made to consider the anisotropy
appearing in this category. These are discussed in Chap. 6. The other
solution, if the physics of the models is put to fault, consists in reducing
their importance, i.e. increasing the cuto frequency to capture a larger
part of the ow physics directly. This means increasing the number of degrees of freedom and striking a compromise between the grid enrichment
techniques and subgrid modeling eorts.
2. Deriving models based on the energy at cuto or the subgrid scales
(with no additional evolution equation) for simulations in the physical
space runs up against Gabor-Heisenbergs generalized principle of uncertainty [204, 627], which stipulates that the precision of the information
cannot be improved in space and in frequency at the same time. This
is illustrated by Fig. 5.15. Very good frequency localization implies high
non-localization in space, which reduces the possibilities of taking the
intermittency21 into account and precludes the treatment of highly inhomogeneous ows. Inversely, very good localization of the information in
space prevents any good spectral resolution, which leads to high errors,
e.g. in computing the energy at the cuto. Yet this frequency localization
is very important, since it alone can be used to detect the presence of
the subgrid scales. It is important to recall here that large-eddy simulation is based on a selection in frequency of modes making up the exact
20
21
Certain authors estimate this share to be between 80% and 90% [119]. Another
criterion sometimes mentioned is that the cuto scale should be of the order
of Taylors microscale. Bagget et al. [32] propose to dene the cuto length in
such a way that the subgrid scales will be quasi-isotropic, leading to L /10,
where L is the integral dissipation length.
Direct numerical simulations and experimental data show that the true subgrid
dissipation and its surrogates do not have the same scaling laws [114, 510].
134
Fig. 5.15. Representation of the resolution in the space-frequency plane. The spat
ial resolution is associated with frequency resolution k . Gabor-Heisenbergs
uncertainty principle stipulates that the product k remains constant, i.e.
that the area of the gray domain keeps the same value (from [204], courtesy of
F. Ducros).
solution. Problems arise here, induced by the localization of statistical

average relations that are exact, as this localization may correspond to
a statistical average. Two solutions may be considered: developing an
acceptable compromise between the precision in space and frequency, or
enriching the information contained in the simulation, which is done either by adding variables to it as in the case of models based on the kinetic
energy of the subgrid modes, or by assuming further hypotheses when
deriving models.
In the following, we present techniques developed to improve the simulation results, though without modifying the structure of the subgrid models
deeply. The purpose of all these modications is to adapt the subgrid model
better to the local state of the ow and remedy the lack of frequency localization of the information.
135
We will be describing:
1. Dynamic procedures for computing subgrid model constants (p. 137).
These constants are computed in such a way as to reduce an a priori
estimate of the error committed with the model considered, locally in
space and time, in the least squares sense. This estimation is made using the Germano identity, and requires the use of an analytical lter. It
should be noted that the dynamic procedures do not change the model
in the sense that its form (e.g. subgrid viscosity) remains the same. All
that is done here is to minimize a norm of the error associated with the
form of the model considered. The errors committed intrinsically22 by
adopting an a priori form of the subgrid tensor are not modied. These
procedures, while theoretically very attractive, do pose problems of numerical stability and can induce non-negligible extra computational costs.
This variation of the constant at each point and each time step makes it
possible to minimize the error locally for each degree of freedom, while
determining a constant value oers only the less ecient possibility of an
overall minimization. This is illustrated by the above discussion of the
constant in the Smagorinsky model.
2. Dynamic procedures that are not directly based on the Germano identity (p. 152): the multilevel dynamic procedure by Terracol and Sagaut
and the multiscale structure function method by Shao. These procedures
have basically the same capability to monitor the constant in the subgrid
model as the previous dynamic procedures. They also involve the denition of a test lter level, and are both based on considerations dealing
with the dissipation scaling as a function of the resolution. Like the procedures based on the Germano identity, they are based on the implicit
assumption that some degree of self-similarity exists in the computed
ow.
3. Structural sensors (p. 154), which condition the existence of the subgrid
scales to the verication of certain constraints by the highest frequencies
of the resolved scales. More precisely, we consider here that the subgrid
scales exist if the highest resolved frequencies verify topological properties that are expected in the case of isotropic homogeneous turbulence.
When these criteria are veried, we adopt the hypothesis that the highest
resolved frequencies have a dynamics close to that of the scales contained
in the inertial range. On the basis of energy spectrum continuity (see the
note of page p. 112), we then deduce that unresolved scales exist, and
the subgrid model is then used, but is otherwise canceled.
4. The accentuation technique (p. 156), which consists in articially increasing the contribution of the highest resolved frequencies when evaluating
22
For example, the subgrid viscosity models described above all induce a linear
dependency between the subgrid tensor and the resolved-scale tensor:
d
= sgs S ij
ij
136
the subgrid viscosity. This technique allows a better frequency localization of the information included in the model, and therefore a better
treatment of the intermittence phenomena, as the model is sensitive only
to the higher resolved frequencies. This result is obtained by applying
a frequency high-pass lter to the resolved eld.
5. The damping functions for the near-wall region (p. 159), by which certain
modications in the turbulence dynamics and characteristic scales of the
subgrid modes in the boundary layers can be taken into account. These
functions are established in such a way as to cancel the subgrid viscosity
models in the near-wall region so that they will not inhibit the driving
mechanisms occurring in this area. These models are of limited generality
as they presuppose a particular form of the ow dynamics in the region
considered. They also require that the relative position of each point
with respect to the solid wall be known, which can raise problems in
practice such as when using multidomain techniques or when several
surfaces exist. And lastly, they constitute only an amplitude correction
of the subgrid viscosity models for the forward energy cascade: they are
not able to include any changes in the form of this mechanism, or the
emergence of new mechanisms.
The three generalist techniques (dynamic procedure, structural sensor,
accentuation) for adapting the subgrid viscosity models are all based on extracting a test eld from the resolved scales by applying a test lter to these
scales. This eld corresponds to the highest frequencies catured by the simulation, so we can see that all these techniques are based on a frequency
localization of the information contained in the subgrid models. The loss of
localness in space is reected by the fact that the number of neighbors involved in the subgrid model computation is increased by using the test lter.
Dynamic Procedures for Computing the Constants.
Dynamic Models. Many dynamic procedures have been proposed to evaluate
the parameters in the subgrid models. The following methods are presented
1. The original method proposed by Germano, and its modication proposed by Lilly to improve its robustness (p. 137). Its recent improvements
for complex kinetic energy spectrum shapes are also discussed.
2. The Lagrangian dynamic procedure (p. 144), which is well suited for fully
non-homogeneous ows.
3. The constrained localized dynamic procedure (p. 146), which relax some
strong assumptions used in the GermanoLilly approach. q
4. The approximate localized dynamic procedure (p. 148), which is a simplication of the constrained localized dynamic procedure that do nor
requires to solve an integral problem to compute the dynamic constant.
5. The generalized dynamic procedures (p. 149), which aim at optimizing
the approximation of the subgrid acceleration and make it possible to
account for discretization errors.
137
6. The dynamic inverse procedure (p. 150), which is designed to improve

the dynamic procedure when the cuto is located at the very begining of
the inertial range of the kinetic energy spectrum.
7. The Taylor series expansion based dynamic procedure (p. 151), which
results in a dierential expression for the dynamic constant, the test
lter being replaced by its dierential approximation.
8. The dynamic procedure based on dimensional parameters (p. 151), which
yields a very simple expression.
GermanoLilly Dynamic Procedure. In order to adapt the models better to
the local structure of the ow, Germano et al. [253] proposed an algorithm for
adapting the Smagorinsky model by automatically adjusting the constant at
each point in space and at each time step. This procedure, described below, is
applicable to any model that makes explicit use of an arbitrary constant Cd ,
such that the constant now becomes time- and space-dependent: Cd becomes
Cd (x, t).
The dynamic procedure is based on the multiplicative Germano identity (3.80) , now written in the form:
Lij = Tij ij
(5.138)
in which
ij
Tij
Lij + Cij + Rij = ui uj ui uj

*i u*j ,
u/
i uj u
Lij
*i u*j
u/
i uj u
(5.139)
(5.140)
(5.141)
in which the tilde symbol tilde designates the test lter. The tensors and
T are the subgrid tensors corresponding, respectively, to the rst and second
ltering levels. The latter ltering level is associated with the characteristic
* with
* > . Numerical tests show that an optimal value is
* = 2.
length ,
The tensor L can be computed directly from the resolved eld.
We then assume that the two subgrid tensors and T can be modeled by
the same constant Cd for both ltering levels. Formally, this is expressed:
1
ij kk ij
3
1
Tij Tkk ij
3
= Cd ij
(5.142)
= Cd ij
(5.143)
in which the tensors and designate the deviators of the subgrid tensors
obtained using the subgrid model deprived of its constant. It is important
noting that the use of the same subgrid model with the same constant is
equivalent to a scale-invariance assumption on both the subgrid uxes and
the lter, to be discussed in the following.
138
Table 5.1. Examples of subgrid model kernels for the dynamic procedure.
Model
(5.90)
(5.102)
(5.127)
ij
ij
* 2 |S|
*S
*
2
+ ij
* ()
* F
* S
*
2
2 |S|S ij
+
2 F ()S ij
2
1+
|F(u)| (qc2 )
1
2
S ij
*
2
1+
ij
* (
|F(u)|
qc2 )
1
2
*
S
ij
Some examples of subgrid model kernels for ij and ij are given in

Table 5.1.
Introducing the above two formulas in the relation (5.138), we get23 :
1
Lij Lkk ij Ldij = Cd ij C/
d ij
3
(5.144)
We cannot use this equation directly for determining the constant Cd because the second term uses the constant only through a ltered product [621].
In order to continue modeling, we need to make the approximation:
*
C/
d ij = Cd ij
(5.145)
which is equivalent to considering that Cd is constant over an interval at least

equal to the test lter cuto length. The parameter Cd will thus be computed
in such a way as to minimize the error committed24 , which is evaluated using
the residual Eij :
1
Eij = Lij Lkk ij Cd ij + Cd *ij
3
(5.146)
This denition consists of six independent relations, which in theory

makes it possible to determine six values of the constant25 . In order to conserve a single relation and thereby determine a single value of the constant,
Germano et al. propose contracting the relation (5.146) with the resolved
strain rate tensor. The value sought for the constant is a solution of the
problem:
Eij S ij
=0 .
(5.147)
Cd
23
24
25
It is important to note that, for the present case, the tensor Lij is replaced by
its deviatoric part Ldij , because we are dealing with a zero-trace subgrid viscosity
modeling.
Meneveau and Katz [505] propose to use the dynamic procedure to rank the subgrid models, the best one being associated with the lowest value of the residual.
Which would lead to the denition of a tensorial subgrid viscosity model.
139
This method can be ecient, but does raise the problem of indetermination when the tensor S ij cancels out. To remedy this problem, Lilly [448]
proposes calculating the constant Cd by a least-squares method, by which
the constant Cd now becomes a solution of the problem:
Eij Eij
=0 ,
Cd
or
Cd =
in which
mij Ldij
mkl mkl
mij = ij *ij
(5.148)
(5.149)
(5.150)
The constant Cd thus computed has the following properties:

It can take negative values, so the model can have an anti-dissipative eect
locally. This is a characteristic that is often interpreted as a modeling of the
backward energy cascade mechanism. This point is detailed in Sect. 5.4.
It is not bounded, since it appears in the form of a fraction whose denominator can cancel out26 .
These two properties have important practical consequences on the numerical solution because they are both potentially destructive of the stability
of the simulation. Numerical tests have shown that the constant can remain
negative over long time intervals, causing an exponential growth in the high
frequency uctuations of the resolved eld. The constant therefore needs an
ad hoc process to ensure the models good numerical properties. There are
a number of dierent ways of performing this process on the constant: statistical average in the directions of statistical homogeneity [253, 779], in time
or local in space [799]; limitation using arbitrary bounds [799] (clipping); or
by a combination of these methods [779, 799]. Let us note that the averaging
procedures can be dened in two non-equivalent ways [801]: by averaging the
denominator and numerator separately, which is denoted symbolically:
Cd =
mij Ldij
mkl mkl
or by averaging the quotient, i.e. on the constant itself:

0
1
mij Ldij
Cd = Cd =
.
mkl mkl
26
(5.151)
(5.152)
This problem is linked to the implementation of the model in the simulation. In

the continuous case, if the denominator tends toward zero, then the numerator
cancels out too. These are calculation errors that lead to a problem of division
by zero.
140
The ensemble average can be performed over homogeneous directions of

the simulation (if they exist) or over dierent realizations, i.e. over several
statistically equivalent simulations carried out in parallel [102, 108].
The time average process is expressed:
Cd (x, (n + 1)t) = a1 Cd (x, (n + 1)t) + (1 a1 )Cd (x, nt) ,
(5.153)
in which t is the time step used for the simulation and a1 1 a constant.
Lastly, the constant clipping process is intended to ensure that the following
two conditions are veried:
+ sgs 0
(5.154)
Cd Cmax
(5.155)
The rst condition ensures that the total resolved dissipation = S ij S ij

ij S ij remains positive or zero. The second establishes an upper bound. In
practice, Cmax is of the order of the theoretical value of the Smagorinsky
constant, i.e. Cmax (0.2)2 .
The models in which the constant is computed by this procedure are called
dynamic because they automatically adapt to the local state of the ow.
When the Smagorinsky model is coupled with this procedure, it is habitually
called the dynamic model, because this combination was the rst to be put
to the test and is still the one most extensively used among the dynamic
models.
The dynamic character of the constant Cd is illustrated in Fig. 5.16, which
displays the time history of the square root of the dynamic constant in freely
decaying isotropic turbulence. It is observed that during the rst stage of
the computation the constant is smaller than the theoretical value of the
Smagorinsky constant Cd 0.18 given by equation (5.48), because the spectrum is not fully developed. In the second stage, when a self-similar state is
reached, the theoretical value is automatically recovered.
The use of the same value of the constant for the subgrid model at the
two ltering levels appearing in equation (5.138) implicitely relies on the two
following self-similarity assumptions:
The two cuto wave numbers are located in the inertial range of the kinetic
energy spectrum;
The lter kernels associated to the two ltering levels are themselves selfsimilar.
These two constraints are not automatically satised, and the validity of
the dynamic procedure for computing the constant requires a careful analysis.
Meneveau and Lund [507] propose an extension of the dynamic procedure
for a cuto located in the viscous range of the spectrum. Writing the constant
141
Fig. 5.16. Time history of the square root of the dynamic constant in large-eddy
simulation of freely decaying isotropic turbulence (dynamic Smagorinsky model).
Courtesy of E. Garnier, ONERA.
of the subgrid-scale model C as an explicit function of the lter characteristic

length, the GermanoLilly procedure leads to
* =C
C() = C()
d
(5.156)
Let be the Kolmogorov length scale. It was said in the introduction

that the ow is fully resolved if = . Therefore, the dynamic procedure is
consistent if, and only if
lim Cd = C() = 0
(5.157)
Numerical experiments carried out by the two authors show that the
GermanoLilly procedure is not consistent, because it returns the value of
the constant associated to the test lter level
*
Cd = C()
,
(5.158)
*
lim Cd = C(r) = 0, r = /
.
(5.159)
yielding
Numerical tests also showed that taking the limit r 1 or computing the
two values C() and C(r) using least-square-error minimization without
142
assuming them to be equal yield inconsistent or ill-behaved solutions. A solution is to use prior knowledge to compute the dynamic constant. A robust
algorithm is obtained by rewriting equation (5.146) as follows:

Eij = Ldij C() f (, r)ij *ij
,
(5.160)
where f (, r) = C(r)/C() is evaluated by calculations similar to those of
Voke (see page 118). A simple analytical tting is obtained in the case r = 2:
f (, 2) max(100, 10x), x = 3.23(Re0.92
Re0.92
)
2
(5.161)
where the mesh-Reynolds numbers are evaluated as (see equation (5.70)):

Re =
2
2 *
|S|
4 |S|
, Re2 =
Other cases can be considered where the similarity hypothesis between

the subgrid stresses at dierent resolution levels may be violated, leading to
dierent values of the constant [601]. Among them:
The case of a very coarse resolution, with a cuto located at the very
beginning of the inertial range or in the production range.
The case of a turbulence undergoing rapid strains, where a transition length
T S 3/2 1/2 appears. Here, S and are the strain magnitude and the
dissipation rate, respectively. Dimensional arguments show that, roughly
speaking, scales larger than T are rapidly distorted but have no time
to adjust dynamically, while scales smaller than T can relax faster via
nonlinear interactions.
For each of these cases, scale dependence of the model near the critical
length scale is expected, which leads to a possible loss of eciency of the
classical GermanoLilly dynamic procedure.
A more general dynamic procedure, which does not rely on the assumption
of scale similarity or location of the cuto in the dissipation range, was proposed by Porte-Agel et al. [601]. This new scale-dependent dynamic procedure
is obtained by considering a third ltering level (i.e. a second test-ltering
* Filtered variables at
> .
level) with a characteristic cuto length scale
this new level are denoted by a caret.
leads to
Writing the Germano identity between level and level
1

j = C()
ui u
Qij Qkk ij ui uj
ij C()ij
3
(5.162)
and
where ij and ij denote the expression of the subgrid model at levels
, respectively. By taking

)
,
(5.163)
nij = (,
ij
ij
with
) =
(,

C()
C()
143
(5.164)
we obtain the following value for the constant at level :

C() =
nij Qij
nij nij
(5.165)
By now considering relation (5.149), which expresses the Germano identity between the rst two ltering levels, where mij is now written as

* ) 2
,
(5.166)
mij = (,
ij
ij
and by equating the values of C() obtained using the two test-ltering
levels, we obtain the following relation:
(Lij mij )(nij nij ) (Qij nij )(mij mij ) = 0 ,
(5.167)
* ) and (,
). In order to obtain a closed
which has two unknowns, (,
system, some additional assumptions are needed. It is proposed in [601] to
assume a power-law scaling of the dynamic constant, C(x) xp , leading to
C(a) = C()ar
(5.168)
For this power-law behavior, the function (., .) does not depend on the
scales but only on the ratio of the scales, i.e. (x, y) = (x/y)r . Using this simplication, (5.167) appears as a fth-order polynomial in C(). The dynamic
constant is taken equal to the largest root.
We now consider the problem of the lter self-similarity. Let G1 and G2
be the lter kernels associated with the rst and second ltering level. For
* = . We assume
the sake of simplicity, we use the notations = 1 and
2
that the lter kernels are rewritten in a form such that:

|x |
u()d ,
(5.169)
u(x) = G1
u(x) = G1
1

|x |
*
u(x)
= G2
u(x) = G2
u()d .
(5.170)
2
We also introduce the test lter Gt , which is dened such that
* = G2
u = Gt
u = Gt
G1
u .
u
The lters G1 and G2 are self-similar if and only if
y
1
G1 (y) = d G2
, r = 2 /1 .
r
r
(5.171)
(5.172)
144
Hence, the two lters must have identical shapes and may only dier by
their associated characteristic length. The problem is that in practice only
Gt is known, and the self-similarity property might not be a priori veried.
Carati and Vanden Eijnden [104] show that the interpretation of the resolved
eld is fully determined by the choice of the test lter Gt , and that the use
of the same model for the two levels of ltering is fully justied. This is
demonstrated by re-interpreting previous lters in the following way. Let us
consider an innite set of self-similar lters {Fn F (ln )} dened as

x
1
(5.173)
Fn (x) = n F
, ln = r n l0 ,
r
ln
where F , r > 1 and l0 are the lter kernel, an arbitrary parameter and
a reference length, respectively. Let us introduce a second set {Fn F (ln )}
dened by
(5.174)
Fn Fn
Fn1
...
F .
For positive kernel F , we get the following properties:
The length ln obeys the same geometrical law as ln :
ln = rln1
,
r
and ln =
ln
r2 1
(5.175)
{Fn } constitute a set of self-similar lters.

Using these two set of lters, the classical lters involved in the dynamic
procedure can be dened as self-similar lters:
Gt (t ) =
G1 (1 ) =
Fn (ln ) ,
Fn1
(ln1
) ,
(5.176)
(5.177)
G2 (2 ) =
Fn (ln ) .
(5.178)
For any test-lter Gt and any value of r, the rst lter operator can be
constructed explicitly:
G1 = Gt (t /r)
Gt (t /r2 )
...
Gt (t /r ) .
(5.179)
This relation shows that for any test lter of the form (5.176), the two
ltering operators can be rewritten as self-similar ones, justifying the use of
the same model at all the ltering levels.
Lagrangian Dynamic Procedure. The constant regularization procedures
based on averages in the homogeneous directions have the drawback of not
being usable in complex congurations, which are totally inhomogeneous.
One technique for remedying this problem is to take this average along the
uid particle trajectories. This new procedure [508], called the dynamic Lagrangian procedure, has the advantage of being applicable in all congurations.
145
The trajectory of a uid particle located at position x at time t is, for

times t previous to t, denoted as:

z(t ) = x
u[z(t ), t ]dt
(5.180)
t
The residual (5.146) is written in the following Lagrangian form:

Eij (z, t ) = Lij (z, t ) Cd (x, t)mij (z, t ) .
(5.181)
We see that the value of the constant is xed at point x at time t, which
is equivalent to the same linearization operation as for the GermanoLilly
procedure. The value of the constant that should be used for computing the
subgrid model at x at time t is determined by minimizing the error along
the uid particle trajectories. Here too, we reduce to a well-posed problem
by dening a scalar residual Elag , which is dened as the weighted integral
along the trajectories of the residual proposed by Lilly:

Elag =
Eij (z(t ), t )Eij (z(t ), t )W (t t )dt
(5.182)
in which the weighting function W (tt ) is introduced to control the memory

eect. The constant is a solution of the problem:
Elag
=
Cd
2Eij (z(t ), t )
Eij (z(t ), t )
W (t t )dt = 0 ,
Cd
or:
Cd (x, t) =
JLM
JMM
(5.183)
(5.184)
in which

JLM (x, t) =

JMM (x, t) =
Lij mij (z(t ), t )W (t t )dt
(5.185)
mij mij (z(t ), t )W (t t )dt
(5.186)
These expressions are non-local in time, which makes them unusable for
the simulation, because they require that the entire history of the simulation
be kept in memory, which exceeds the storage capacities of todays supercomputers. To remedy this, we choose a fast-decay memory function W :
W (t t ) =
1
Tlag

t t
exp
Tlag
(5.187)
146
in which Tlag is the Lagrangian correlation time. With the memory function
in this form, we can get the following equations:
JLM
JLM
1
DJLM
+ ui
=
(Lij mij JLM )
Dt
t
xi
Tlag
JMM
JMM
DJMM
1
+ ui
=
(mij mij JMM )
Dt
t
xi
Tlag
(5.188)
(5.189)
the solution of which can be used to compute the subgrid model constant at
each point and at each time step. The correlation time Tlag is estimated by
tests in isotropic homogeneous turbulence at:
Tlag (x, t) = 1.5 (JMM JLM )
1/8
(5.190)
which comes down to considering that the correlation time is reduced in

the high-shear regions where JMM is large, and in those regions where the
non-linear transfers are high, i.e. where JLM is large.
This procedure does not guarantee that the constant will be positive,
and must therefore be coupled with a regularization procedure. Meneveau et
al. [508] recommend a clipping procedure.
Solving equations (5.188) and (5.189) yields a large amount of additional
numerical work, resulting in a very expensive subgrid model. To alleviate this
problem, the solution to these two equations may be approximated using the
following Lagrangian tracking technique [596]:
JLM (x, nt) =
+
a Lij (x, nt)mij (x, nt)

(1 a)JLM (x tu(x, nt), (n 1)t) , (5.191)
JMM (x, nt) =

+
a mij (x, nt)mij (x, nt)

(1 a)JMM (x tu(x, nt), (n 1)t) , (5.192)
where
a=
t/Tlag
1 + t/Tlag
(5.193)
This new procedure requires only the storage of the two parameters JLM
and JMM at the previous time step and the use of an interpolation procedure.
The authors indicate that a linear interpolation is acceptable.
Constrained Localized Dynamic Procedure. Another generalization of the
GermanoLilly dynamic procedure was proposed for inhomogeneous cases
by Ghosal et al. [261]. This new procedure is based on the idea of minimizing an integral problem rather than a local one in space, as is done in the
GermanoLilly procedure, which avoids the need to linearize the constant
147
when applying the test lter. We now look for the constant Cd that will
minimize the function F [Cd ], with

F [Cd ] = Eij (x)Eij (x)d3 x ,
(5.194)
in which Eij is dened from relation (5.144) and not by (5.146) as was the
case for the previously explained versions of the dynamic procedure. The
constant sought is such that the variation of F [Cd ] is zero:

F [Cd ] = 2 Eij (x)Eij (x)d3 x = 0 ,
(5.195)
or, by replacing Eij with its value:

Cd d3 x = 0 .
ij Eij Cd + Eij ij/
(5.196)
Expressing the convolution product associated with the test lter, we get:

3
ij Eij + ij Eij (y)G(x y)d y Cd (x)d3 x = 0 ,
(5.197)
from which we deduce the following Euler-Lagrange equation:

ij Eij + ij Eij (y)G(x y)d3 y = 0 .
(5.198)
This equation can be re-written in the form of a Fredholms integral equation of the second kind for the constant Cd :

f (x) = Cd (x) K(x, y)Cd (y)d3 y ,
(5.199)
where
f (x) =
1
kl (x)kl (x)

ij (x)Lij (x) ij (x) Lij (y)G(x y)d3 y
(5.200)
K(x, y) =
KA (x, y) + KA (y, x) + KS (x, y)

kl (x)kl (x)
(5.201)
and
KA (x, y) = ij (x)ij (y)G(x y) ,

KS (x, y) =ij (x)ij (y) G(z x)G(z y)d3 z
(5.202)
.
(5.203)
148
This new formulation raises no problems concerning the linearization of

the constant, but does not solve the instability problems stemming from the
negative values it may take. This procedure is called the localized dynamic
procedure.
To remedy the instability problem, the authors propose constraining the
constant to remain positive. The constant Cd (x) is then expressed as the
square of a new real variable (x). Replacing the constant with its decomposition as a function of , the Euler-Lagrange equation (5.198) becomes:

3
(5.204)
ij Eij + ij Eij (y)G(x y)d y (x) = 0 .
This equality is true if either of the factors is zero, i.e. if (x) = 0 or if the
relation (5.198) is veried, which is denoted symbolically Cd (x) = G[Cd (x)].
In the rst case, the constant is also zero. To make sure it remains positive,
the constant is computed by an iterative procedure:
G[Cd(n) (x)] if G[Cd(n) (x)] 0

(n+1)
Cd
(x) =
,
(5.205)
0
otherwise

in which
G[Cd (x)] = f (x)
K(x, y)Cd (y)d3 y
(5.206)
This completes the description of the constrained localized dynamic procedure. It is applicable to all congurations and ensures that the subgrid
model constant remains positive. This solution is denoted symbolically:

Cd (x) = f (x) + K(x, y)Cd (y)d3 y
,
(5.207)
+
in which + designates the positive part.

Approximate Localized Dynamic Procedure. The localized dynamic procedure
decribed in the preceding paragraph makes it possible to regularize the dynamic procedure in fully non-homogeneous ows, and removes the mathematical inconsistency of the GermanoLilly procedure. But it requires to solve
an integral equation, and thus induces a signicant overhead.
To alleviate this problem, Piomelli and Liu [596] propose an Approximate
Localized Dynamic Procedure, which is not based on a variational approach
but on a time extrapolation process. Equation (5.144) is recast in the form
Ldij = Cd ij C/
d ij
(5.208)
where Cd is an estimate of the dynamic constant Cd , which is assumed to

be known. Writing the new formulation of the residual Eij , the dynamic
149
constant is now evaluated as

Cd =
ij (Ldij + C/
d ij )
ij ij
(5.209)
The authors propose to evaluate the estimate Cd by a time extrapolation:

Cd
(n1)
Cd
(n1)
Cd
+ t
+ ... ,
t
(5.210)
where the superscript (n 1) is related to the value of the variable at the

(n 1)th time step, and t is the value of the time step. In practice, Piomelli
and Liu consider rst- and second-order extrapolation schemes. The resulting
dynamic procedure is fully local, and does not induce large extra computational eort as the original localized procedure does. Numerical experiments
carried out by these authors demonstrate that it still requires clipping to
yield a well-behaved algorithm.
Generalized Dynamic Procedure. It is also possible to derive a dynamic procedure using the generalized Germano identity (3.87) [629]. We assume that
the operator L appearing in equation (3.88) is linear, and that there exists
a linear operator L such that
L(a N ) = aL(N ) + L (a, N ) ,
(5.211)
where a is a scalar real function and N an arbitrary second rank tensor. The
computation of the dynamic constant Cd is now based on the minimization
of the residual Eij
Eij = L(Ldij ) Cd L(mij ) ,
(5.212)
where Ldij and mij are dened by equations (5.144) and (5.150). A leastsquare minimizations yields:
Cd =
L(Ldij )L(mij )
L(mij )L(mij )
(5.213)
The reduction of the residual obtained using this new relation with respect
to the classical one is analyzed by evaluating the dierence:

L(Eij ) ,
Eij = Eij
(5.214)

is given by relation (5.212) and Eij by (5.146). Inserting the
where Eij
two dynamic constants Cd and Cd , dened respectively by relations (5.213)
and (5.149), we get:
Eij = (Cd Cd )L(mij ) + L (Cd , mij )
(5.215)
150
An obvious example for the linear operator L is the divergence operator.

The associated L is the gradient operator.
An alternative consisting in minimizing a dierent form of the residual
has been proposed by Morinishi and Vasilyev [542, 544] and Mossi [550]:
Eij
L(Ldij ) L(Cd mij )
L(Ldij )
(5.216)
Cd L(mij ) L (Cd , mij )
(5.217)
The use of this new form of the residual generally requires solving a differential equation, and then yields a more complex procedure than the
form (5.212).
These two procedures theoretically more accurate results than the classical one, because they provide reduce the error committed on the subgrid force
term itself, rather than on the subgrid tensor. They also take into account
for the numerical error associated to the discrete form of L.
Dynamic Inverse Procedure. We have already seen that the use of the dynamic procedure may induce some problems if the cuto is not located in
the inertial range of the spectra, but in the viscous one. A similar problem
arises if the cuto wave number associated to the test lter occurs at the very
beginning of the inertial range, or in the production range of the spectrum.
In order to compensate inaccuracies arising from the use of a large lter
length associated with the test lter, Kuerten et al. [415] developed a new
approach, referred to as the Dynamic Inverse Procedure. It relies on the idea
that if a dynamic procedure is developed involving only length scales comparable to the basic lter length, self-similarity properties will be preserved and
consistent modeling may result. Such a procedure is obtained by dening the
rst ltering operator G and the second one F by
G = H 1 L, F = H
(5.218)
where L is the classical lter level and H an explicit test lter, whose inverse
H 1 is assumed to be known explicitly. Inserting these denitions into the
Germano identity (3.80), we get a direct evaluation of the subgrid tensor :
[F
G
, B](ui , uj ) = [L
, B](ui , uj )
= ui uj ui uj
ij
(5.219)
(5.220)
(5.221)
= [H
, B] (H 1
L
)(ui , uj )
+(H
) [H 1
L
, B](ui , uj ) .
(5.222)
This new identity can be recast in a form similar to the original Germano
identity
(5.223)
Lij = ij H
Tij ,
151
with
Lij = H
((H 1
ui )(H 1
uj )) ui uj
Tij = H 1
ui uj (H 1
ui )(H 1
uj ) .
The term Lij is explicitly known in practice, and does not require
any modeling. Using the same notation as in the section dedicated to the
GermanoLilly procedure, we get, for the Smagorinsky model:
ij
Tij
Cd ij , ij = 2 |S|S ij
(5.224)
(5.225)
|S|
S

Cd ij , ij = 2
ij
and S
are the characteristic length and the strain rate tensor assowhere
ciated to the H 1 L ltering level, respectively. Building the residual Eij
as
Eij = Lij Cd (ij H
ij ) = Lij Cd mij
(5.226)
the least-square-error minimization procedure yields:

Cd =
Lij mij
mij mij
(5.227)
Since the
In this new procedure, the two lengths involved are and .
, ensurlatter is associated to an inverse ltering operator, we get
ing that the dynamic procedure will not bring in lengths associated to the
production range of the spectrum. In practice, this procedure is observed to
suer the same stability problems than the GermanoLilly procedure, and
needs to be used together with a stabilization procedure (averaging, clipping,
etc.).
Taylor Series Expansion Based Dynamic Models. The dynamic procedures
presented above rely on the use of a discrete test lter. Chester et al. [129]
proposed a new formulation for the dynamic procedure based on the dierential approximation of the test lter. All quantities apprearing at the test lter
level can therefore be rewritten as sums and products of partial derivatives of
the resolved velocity eld, leading to a new expression of dynamic constants
which involves only higher-order derivatives of the velocity eld.
Dynamic Procedure with Dimensional Constants. The dynamic procedures
described in the preceding paragraphs are designed to nd the best values of
non-dimensional constants in subgrid scale models. Wong and Lilly [766], followed by Carati and his co-workers [103] propose to extend this procedure to
evaluate dimensional parameters which appear in some models. They applied
152
this idea to the so-called Kolmogov formulation for the subgrid viscosity:
sgs = C
4/3 1/3
4/3
= C
(5.228)
where the parameter C = C1/3 has the dimension of the cubic root of
the subgrid dissipation rate . Introducing this closure at both grid and test
ltering levels, one obtains (the tilde symbol is related to the test lter level):
ij
4/3
2C
4/3
Tij
*
2C
S ij
(5.229)
*
S
ij
(5.230)
leading to the following expression of the residual

4/3
* 4/3 S
* + C /
Eij = Ldij C
S ij
ij
(5.231)
This expression can be used to generate integral expressions for C . A very

4/3
simple local denition is recovered further assuming that C
is almost
* Using the additional property that
constant over distance of the order of .
the test lter perfectly commutes with spatial derivatives, relation (5.231)
simplies as

4/3
* 4/3 S
*
Eij = Ldij 2C
.
(5.232)
ij
The least-square optimization method therefore yields the following formula for the dynamic C :
C =
*
Ldij S
1
ij

4/3
*
*
4/3
*
S ij S ij
2
(5.233)
As the original Germano-Lilly procedure for non-dimensional parameters,

this procedure suers some numerical instability problems and must therefore
be regularized using clipping and/or averaging.
Dynamic Procedures Without the Germano Identity.
Multilevel Procedure by Terracol and Sagaut. This method proposed by Terracol and Sagaut [709] relies on the hypothesis that the computed resolved
kinetic energy spectrum obeys a power-law like
E(k) = E0 k
(5.234)
where is the scaling parameter. It is worth noting that Barenblatt [41]

suggests that both E0 and might be Reynolds-number dependent. A more
153
accurate expression for the kinetic energy spectrum is

E(k) = K0 2/3 k 5/3 (k)
(5.235)
where K0 = 1.4 is the Kolmogorov constant, a length scale and an

intermittency factor. Under this assumption, the mean subgrid dissipation
rate across a cuto wave number kc , (kc ), scales like
(kc ) = 0 kc ,
3
3 + 5
=
2
2
(5.236)
where 0 is a kc -independent parameter. It is observed that in the Kolmogorov

case ( = 5/3), one obtains = 0, leading to a constant dissipation rate.
Let us now introduce a set of cuto wave numbers kn , with k1 > k2 > ...
The following recursive law is straithgforwardly derived from (5.236)
(kn )
= Rn,n+1
,
(kn+1 )
Rn,n+1 =
kn
kn+1
(5.237)
leading to the following two-level evaluation of the parameter :

=
log((kn )/(kn+1 ))
log(Rn,n+1 )
(5.238)
Now introducing a generic subgrid model for the nth cuto level
n
ijn = Cfij (un , ) ,
(5.239)
where C is the constant of the model to be dynamically computed, un the

n
resolved eld at the considered cuto level and /kn the current cuto
length, the dissipation rate can also be expressed as
n
(kn ) = ijn S ij = Cfij (un , )S ij
(5.240)
where S is the resolved strain rate at level n. Equation (5.237) shows that
the ratio (kn )/(kn+1 ) is independent of the model constant C. Using this
property, Terracol and Sagaut propose to introduce two test lter levels k2
and k3 (k1 being the grid lter level where the equations must be closed, i.e.
1
= ). The intermittency factor is then computed using relation (5.238),
and one obtains the following evaluation for the subgrid dissipation rate at
the grid lter level:
(k2 ) ,
(5.241)
(k1 ) = R1,2
where (k2 ) is evaluated using a reliable approximation of the subgrid tensor to close the sequence (in practice, a scale-similarity model is used in
Ref. [709]). The corresponding value of C for the model at the grid level is
then deduced from (5.240):
C = R1,2
(k2 )
fij (u, )S ij
(5.242)
154
Multiscale Method Based on the Kolmogorov-Meneveau Equation. Another

procedure was developed by Shao [669, 153] starting from the KolmogorovMeneveau equation for ltered third-order velocity structure function:
4
r = DLLL 6GLLL
5
(5.243)
where DLLL is the third-order longitudinal velocity correlation of the ltered eld, GLLL (r) the longitudinal velocity-stress correlation tensor and
= ij S ij the average subgrid dissipation (see p. 126 for additional details).
Now assuming that the following self-similarity law is valid
GLLL (r) rp
(5.244)
where p = 1/3 corresponds to the Kolmogorov local isotropy hypotheses,

one obtains the following relationship for two space increments r1 and r2 :
0.8r1 + DLLL (r1 )
=
0.8r2 + DLLL (r2 )
r1
r2
1/3
.
(5.245)
Now introducing the same generic subgrid closure as for the Terracol
Sagaut procedure
ij = Cfij (u, )
(5.246)
and inserting it into (5.245) to evaluate = ij S ij , taking r1 = and

r2 > r1 , the dynamic value of the constant C is
C=
1/3
DLLL (r2 ) DLLL ()

1/3
0.8fij (u, )S ij r2
r2
r2
(5.247)
The only xed parameter in the Shao procedure is the scaling parameter p
in (5.244). This parameter can be computed dynamically introducing a third
space increment r3 , leading to the denition of a dynamic procedure with the
same properties as the one proposed by Terracol and Sagaut. The proposal of
Shao can also be extended to subgrid models with several adjustable constant
by introducing an additional space increment for each new constant and
solving a linear algebra problem.
Structural Sensors. Selective Models. In order to improve the prediction of intermittent phenomena, we introduce a sensor based on structural information. This is done by incorporating a selection function in the
model, based on the local angular uctuations of the vorticity, developed by
David [166, 440].
155
The idea here is to modulate the subgrid model in such a way as to apply
it only when the assumptions underlying the modeling are veried, i.e. when
all the scales of the exact solution are not resolved and the ow is of the
fully developed turbulence type. The problem therefore consists in determining if these two hypotheses are veried at each point and each time step.
Davids structural sensor tests the second hypothesis. To do this, we assume
that, if the ow is turbulent and developed, the highest resolved frequencies
have certain characteristics specic to isotropic homogeneous turbulence, and
particularly structural properties.
So the properties specic to isotropic homogeneous turbulence need to
be identied. David, taking direct numerical simulations as a base, observed
that the probability density function of the local angular uctuation of the
vorticity vector exhibit a peak around the value of 20o . Consequently, he
proposes identifying the ow as being locally under-resolved and turbulent
at those points for which the local angular uctuations of the vorticity vector
corresponding to the highest resolved frequencies are greater than or equal
to a threshold value 0 .
The selection criterion will therefore be based on an estimation of the
angle between the instantaneous vorticity vector and the local average
(see Fig. 5.17), which is computed by applying a test lter
vortcity vector
to the vorticity vector.
The angle is given by the following relation:

(x)
(x)
(x) = arcsin
.
(5.248)
(x).(x)
We dene a selection function to damp the subgrid model when the angle
is less than a threshold angle 0 .
In the original version developed by David, the selection function f0 is
a Boolean operator:

1 if 0
f0 () =
.
(5.249)
0 otherwise
Fig. 5.17. Local angular uctuation of the vorticity vector.
156
This function is discontinuous, which may pose problems in the numerical

solution. One variant of it that exhibits no discontinuity for the threshold
value is dened as follows [636]:

1
if 0
f0 () =
,
(5.250)
r()n otherwise
in which 0 is the chosen threshold value and r the function:
r() =
tan2 (/2)
tan2 (0 /2)
(5.251)
where the exponent n is positive. In practice, it is taken to be equal to 2.

Considering the fact that we can express the angle as a function of the
and the norm
norms of the vorticity vector , the average vorticity vector ,
by the relation:
of the uctuating vorticity vector dened as = ,
2 + 2 2
cos
=
2
and the trigonometric relation:

tan2 (/2) =
1 cos
1 + cos
the quantity tan2 (/2) is estimated using the relation:

2

2 2 +
2
+
2 + 2 2
2
tan2 (/2) =
(5.252)
The selection function is used as a multiplicative factor of the subgrid

viscosity, leading to the denition of selective models:
sgs = sgs (x, t)f0 ((x)) ,
(5.253)
in which sgs is calculated by an arbitrary subgrid viscosity model. It should

be noted that, in order to keep the same average subgrid viscosity value over
the entire uid domain, the constant that appears in the subgrid model has
to be multiplied by a factor of 1.65. This factor is evaluated on the basis of
isotropic homogeneous turbulence simulations.
Accentuation Technique. Filtered Models.
Accentuation Technique. Since large-eddy simulation is based on a frequency
selection, improving the subgrid models in the physical space requires a better
diagnostic concerning the spectral distribution of the energy in the calculated
solution. More precisely, what we want to do here is to determine if the exact solution is entirely resolved, in which case the subgrid model should be
reduced to zero, or if there exist subgrid scales that have to be taken into
account by means of a model. When models expressed in the physical space
157
do not use additional variables, they suer from imprecision due to GaborHeisenbergs principle of uncertainty already mentioned above, because the
contribution of the low frequencies precludes any precise determination of
the energy at the cuto. Let us recall that, if this energy is zero, the exact
solution is completely represented and, if otherwise, then subgrid modes exist. In order to be able to detect the existence of the subgrid modes better,
Ducros [204, 205] proposes an accentuation technique which consists in applying the subgrid models to a modied velocity eld obtained by applying
a frequency high-pass lter to the resolved velocity eld. This lter, denoted
HPn , is dened recursively as:
HP1 (u)
n
HP (u)
2 u
= HP(HP
,
n1
(5.254)
(u))
(5.255)
We note that the application of this lter in the discrete case results in
a loss of localness in the physical space, which is in conformity with GaborHeisenbergs principle of uncertainty. We use EHPn (k) to denote the energy
spectrum of the eld thus obtained. This spectrum is related to the initial
spectrum E(k) of the resolved scales by:
E HPn (k) = THPn (k)E(k) ,
(5.256)
in which THPn (k) is a transfer function which Ducros evaluates in the form:
n
k
.
(5.257)
THPn (k) = bn
kc
Here, b and are positive constants that depend on the discrete lter used
in the numerical simulation27 . The shape of the spectrum obtained by the
transfer function to a Kolmogorov spectrum is graphed in Fig. 5.18 for several
values of the parameter n. This type of lter modies the spectrum of the
initial solution by emphasizing the contribution of the highest frequencies.
The resulting eld therefore represents mainly the high frequencies of the
initial eld and serves to compute the subgrid model. To remain consistent,
the subgrid model has to be modied. Such models are called ltered models.
The case of the Structure Function model is given as an example. Filtered
versions of the Smagorinsky and Mixed Scale models have been developped
by Sagaut, Comte and Ducros [628].
Filtered Second-Order Structure Function Model. We dene the second-order
structure function of the ltered eld:

HPn
2
DLL (x, r, t) =
[HPn (u)(x, t) HPn (u)(x + x , t)] d3 x , (5.258)
|x |=r
27
For a Laplacian type lter discretized by second-order accurate nite dierence

scheme iterated three times (n = 3), Ducros nds b3 = 64, 000 and 3 = 9.16.
158
Fig. 5.18. Energy spectrum of the accentuation solution for dierent values of the
parameter n (b = = 1, kc = 1000).
for which the statistical average over the entire uid domain, denoted
HPn
D LL (r, t), is related to the kinetic energy spectrum by the relation:

HPn
kc
DLL (r, t) = 4
0

sin(k)
E HPn (k) 1
dk
k
(5.259)
According to the theorem of averages, there exists a wave number k

[0, kc ] such that:
HPn
HPn
D
(r, t)
LL
(k ) =
4(/kc )
(1 sin()/) d
(5.260)
Using a Kolmogorov spectrum, we can state the equality:

E(kc )
5/3
kc
E HPn (k )
k
(5.261)
Considering this last relation, along with (5.256) and (5.257), the subgrid
viscosity models based on the energy at cuto are expressed:
3/2
2 K0
sgs =
3 kc1/2
'

k
kc
5/3n
1
E HPn (k ) ,
bn
(5.262)
in which:

k
kc
5/3+n
=
5/3+n
5/3+n (1 sin()/) d

(1 sin()/) d
159
(5.263)
By localizing these relations in the physical space, we deduce the ltered

structure function model:
HPn
1/2
1/2

3/2
DLL (x, , t)
n
2 K0
sgs (x, , t) =
1/2

3 4/3 2
bn
5/3+n
(1 sin()/) d
0
+
HPn
= C (n) DLL (x, , t) .
(5.264)
The values of the constant C (n) are given in the following Table:
In practice, Ducros recommends using n = 3.
Table 5.2. Values of the Structure Function model constant for dierent iterations
of the high-pass lter.
n
C
(n)
0.0637
0.020
0.0043
0.000841
1.57 104
Damping Functions for the Near-wall Region. The presence of a solid

wall modies the turbulence dynamics in several ways, which are discussed
in Chap. 10. The only fact concerning us here is that the presence of a wall
inhibits the growth of the small scales. This phenomenon implies that the
characteristic mixing length of the subgrid modes f has to be reduced in
the near-surface region, which corresponds to a reduction in the intensity
of the subgrid viscosity. To represent the dynamics in the near-wall region
correctly, it is important to make sure that the subgrid models verify the
good properties in this region. In the case of a canonical boundary layer (see
Chap. 10), the statistical asymptotic behavior of the velocity components and
subgrid tensions can be determined analytically. Let u be the main velocity
component in the x direction, v the transverse component in the y direction,
and w the velocity component normal to the wall, in the z direction. Using
the incompressibility constraint, a Taylor series expansion of the velocity
component in the region very near the wall yields:
u z, v z, w z 2 ,
11 z 2 , 22 z 2 ,
13 z 3 , 12 z 2 ,
33 z 4 ,
23 z 3 .
(5.265)
(5.266)
160
Experience shows that it is important to reproduce the behavior of the

component 13 in order to ensure the quality of the simulation results. It is
generally assumed that the most important stress in the near-wall region is
13 , because it is directly linked to the mean turbulence production term, P ,
which is evaluated as
du
P 13
.
(5.267)
dz
Thus, it is expected that subgrid-viscosity models will be such that
sgs
du
13 z 3
dz
(5.268)
We deduce the following law from relations (5.265) and (5.268):

sgs z 3
(5.269)
We verify that the subgrid-viscosity models based on the large scales alone
do not verify this asymptotic behavior. This is understood by looking at the
wall value of the subgrid viscosity associated with the mean velocity eld
u. A second-order Taylor series expansion of some zero-equation subgrid
viscosity models presented in the preceding section yields:

2 u
(Smagorinsky) ,
sgs |w |w
z w

u

(2nd order Structure Function) ,
sgs |w |w z1
z w

u 1/2 2 u 1/2
3/2

sgs |w |w z1
(Mixed Scale) ,
z w z 2 w
(5.270)
where the w subscript denotes values taken at the wall, and z1 is the distance to the wall at which the model is evaluated. Because u/z is not
zero at the wall,28 , we have the following asymptotic scalings of the modeled
subgrid viscosity at solid walls:
sgs |w
sgs |w
= O(|2w ) (Smagorinsky) ,
= O(|w z1 ) (2nd order Structure Function) ,
sgs |w
= 0
(Mixed Scale) .
(5.271)
In practice, the Mixed Scale model can predict a zero subgrid viscosity
at the wall if the computational grid is ne enough to make it possible to
28
Or, equivalently, the skin friction is not zero.
161
evaluate correctly the second-order wallnormal velocity derivative, i.e. if at

least three grid points are located within the region where the mean velocity
prole obeys a linear law. Consequently, the rst two models must be modied
in the near-wall region in order to enforce a correct asymptotic behavior of
the subgrid terms in that region.
This is done by introducing damping functions. The usual relation:
f = C
(5.272)
is replaced by:
f = Cfw (z) ,
(5.273)
in which fw (z) is the damping function and z the distance to the wall. From
Van Driests results, we dene:
fw (z) = 1 exp (zu /25)
(5.274)
in which the friction velocity u is dened in Sect. 10.2.1. Piomelli et al. [600]
propose the alternate form:

1/2
fw (z) = 1 exp (zu /25)3
(5.275)
From this last form we can get a correct asymptotic behavior of the sub3
grid viscosity, i.e. a decrease in z + in the near-wall region, contrary to the
Van Driest function. Experience shows that we can avoid recourse to these
functions by using a dynamic procedure, a ltered model, a selective model, or
the Yoshizawa model. It is worth noting that subgrid viscosity models can be
designed, which automatically follow the correct behavior in the near-wall region. An example is the WALE model, developed by Nicoud and Ducros [567].
5.3.4 Implicit Diusion: the ILES Concept
Large-eddy simulation approaches using a numerical viscosity with no explicit
modeling are all based implicitly on the hypothesis:
Hypothesis 5.6 The action of subgrid scales on the resolved scales is equivalent to a strictly dissipative action.
This approach is referred to as Implicit Large-Eddy Simulation (ILES). Simulations belonging to this category use dissipation terms introduced either
in the framework of upwind schemes for the convection or explicit articial
dissipation term, or by the use of implicit [716] or explicit [210] frequency lowpass lters. The approach most used is doubtless the use of upwind schemes
for the convective term. The diusive term introduced then varies both in degree and order, depending on the scheme used (QUICK [437], Godunov [776],
PPM [145], TVD [150], FCT [66], MPDATA [489, 488], among others) and
162
the dissipation induced can in certain cases be very close29 to that introduced
by a physical model [275]. Let us note that most of the schemes introduce
dissipations of the second and/or fourth order and, in so doing, are very close
to subgrid models. This point is discussed more precisely in Chap. 8. This approach is widely used in cases where the other modeling approaches become
dicult for one of the two following reasons:
The dynamic mechanisms escape the physical modeling because they are
unknown or too complex to be modeled exactly and explicitly, which is true
when complex thermodynamic mechanisms, for example, interact strongly
with the hydrodynamic mechanisms (e.g. in cases of combustion [135] or
shock/turbulence interaction [435]).
Explicit modeling oers no a priori guarantee of certain realizability constraints related to the quantities studied (such as the temperature [125]
or molar concentrations of pollutants [474]). This point is illustrated in
Fig. 5.19, which displays the probability density function of a passive scalar
computed by Large-Eddy Simulation with dierent numerical schemes for
the convection term.
In cases belonging to one of these two classes, the error committed by
using an implicit viscosity may in theory have no more harmful consequence
on the quality of the result obtained than that which would be introduced
by using an explicit model based on inadequate physical considerations. This
approach is used essentially for dealing with very complex congurations or
those harboring numerical diculties, because it allows the use of robust
numerical methods. Nonetheless, high-resolution simulations of ows are beginning to make their appearance [756, 602, 776, 274].
A large number of stabilized numerical methods have been used for largeeddy simulation, but only a few of them have been designed for this specic
purpose or more simply have been analyzed in that sense. A few general
approaches for designing stabilized methods which mimic functional subgrid
modeling are discussed below:
1. The MILES (Monotone Integrated Large Eddy Simulation) approach
within the framework of ux-limiting nite volume methods, as discussed
by Grinstein and Fureby (p. 163).
2. The adaptive ux reconstruction technique within the framework of nonlimited nite volume methods, proposed by (p. 165).
3. Finite element schemes with embedded subgrid stabilization (p. 166).
4. The use of Spectral Vanishing Viscosities (p. 169) which are well suited
for numerical methods with spectral-like accuracy.
5. The high-order ltering technique (p. 170), originally developed within
the nite-dierence framework, and wich is equivalent to some approximate deconvolution based structural models.
29
In the sense where these dissipations are localized at the same points and are of
the same order of magnitude.
163
Fig. 5.19. Probability of the density function of the temperature (modeled as

a passive scalar) in a channel ow obtained via Large-Eddy Simulation. Vertical
lines denote physical bounds. It is observed that the simulation carried out with
centered fourth-order accurate scheme admits non-physical values, while the use
of the stabilized schemes to solve the passive scalar equation cures this problem.
Courtesy of F. Chatelain (CEA).
MILES Approach. A theoretical analysis of the MILES approach within

the framework of ux-limiting nite volume discretizations has been carried
out by Fureby and Grinstein [273, 228, 231, 274, 229], which puts the emphasis on the existing relationship between leading numerical error terms and
tensorial subgrid viscosities.
Dening a control cell of face-normal unit vector n, the convective
uxes are usually discretized using Greens theorem

(u u)d =
(u n)udS ,
(5.276)
where denotes the tensorial product and is the boundary of . The

associated discrete relation is

(u n)udS
FfC (u) ,
(5.277)
where f are the faces of and the discrete ux function is expressed as

FfC (u) = ((u dA)u)f
(5.278)
164
where dA is the face-area vector of face f of , and ()f is the integrated value
on face f. For ux-limiting methods, the numerical ux is decomposed as the
weighted sum of a high-order ux function FfH that works well in smooth
regions and a low-order ux function FfL :
3
4
,
(5.279)
FfC (u) = FfH (u) + (1 (u)) FfH (u) FfL (u)
where (u) is the ux limiter.30 Fureby and Grinstein analyzed the leading
error term using the following assumptions: (i) time integration is performed
using a three-point backward scheme, (ii) the high-order ux functions use
rst-order functional reconstruction, and (iii) the low-order ux functions
use upwind dierencing. Retaining the leading dissipative error term, the
continuous equivalent formulation for the discretized uxes is
(u u) (u r + r u + r r) ,

exact
(5.280)
dissipative error

1
ud
r = (u)d, = (1 (u))sgn
,
(5.281)
2
|d|
where d is the topology vector connecting neighboring control volumes. Comparison of the error term in (5.280) and the usual subgrid term appearing in
ltered NavierStokes equations (3.17) yields the following identication for
the MILES subgrid tensor:
with
MILES
= (u r + r u + r r)
3
4
= (u d)T u + u(u d)T

+ 2 (u)d (u)d
(5.282)
II
Term I appears as a general subgrid-viscosity model with a tensorial

diusivity (u d), while term II mimics the Leonard tensor, leading to
the denition of an implicitmixed model (see Sect. 7.4).31 A scalar-valued
measure of the viscosity is 2/8|u|MILES
, where the characteristic length
associated with the grid is MILES = tr[(T d)(d d)(d)].
The authors remarked that these error terms are invariant under the
Galilean group of transformations, but are not frame indierent. Realizability
and non-negative dissipation of subgrid kinetic energy may be enforced for
some choice of the limiter.
30
31
Many ux limiters can be found in the literature: minmod, superbee, FCT limiter, ... The reader is referred to specialized reference books [307] for a detailed
discussion of these functions.
MILES can also be interpreted as an implicit deconvolution model, using the
analogy discussed in Sect. 7.3.3.
165
Adaptive Flux Reconstruction. Adams [2] established a theoretical bridge

between high-order adaptive ux reconstruction used in certain nite-volume
schemes and the use of a subgrid viscosity. In the simplied case of the following one-dimensional conservation law
u F (u)
+
=0
t
x
(5.283)
the nite volume technique leads to the computation of the cell-averaged

variable:
xj+1/2
uj =
u()d ,
(5.284)
xj1/2
with x = xj+1/2 xj1/2 the cell spacing of the jth cell. High-order nite
volume methods rely on the reconstruction of the unltered value uj+1/2 on
both sides of the cell face xj+1/2 on each cell j. This is achieved by deltering
the variable u and dening a high-order polynomial interpolant.
The deltering step is similar to the deconvolution approach, whose related results are presented in Sect. 7.2.1 and will not be repeated here. Sticking to Adams demonstration, a second-order deconvolution is employed:
uj = uj
x2 2 uj
24 x2
(5.285)
Following the WENO (Weightest Essentially Non-Oscillatory) concept [350], a hierarchical family of left-hand-side interpolants of increasing
order is:
P +,(0) (x)j
+,(1)
(x)
Pj
+,(2)
Pj
(x)
= uj
(5.286)
+,(0)
Pj
(x)
+,(1)
Pj
+
1,1 (x
(1)
xj )j
(5.287)

(2)
(2)
+
, (5.288)
+(xxj )(xxj+1 ) +
1,2 j1 +2,2 j
... = ...
+,(k)
(p)
where Pj
(x) is the kth-order interpolant, j the divided dierence of
degree p of the variable, and +
m,n some weigthing parameters. Right-handsides are dened in the same way, except that P ,(0) (x)j = uj+1 and weights
are noted
m,n . A kth-order interpolation is obtained under the following
constraints:

m
m,n = 1,
m,n > 0,
n = 1, .., k 1
(5.289)
166
Applying this procedure to (5.283), the numerical convection term can be

expressed as
4
F (u)
1 3
fj+1/2 (xj+1/2 ) fj1/2 (xj1/2 )

x
x
(5.290)
where fj1/2 is a numerical ux function. Adams analysis is based on the

local LaxFriedrichs ux:
3
4
fj+1/2 (x) = f (Pj+ (x)) + f (Pj+1

(x)) j+1/2 (Pj+1
(x)Pj+ (x))
, (5.291)
with
j+1/2 = max |f (u)| .
uj ,uj+1
(5.292)
In the simplied case of the Burgers equation, i.e. F (u) = u2 /2, the
leading error term is:

1 1
3 uj
uj 2 uj
2
+
=
x uj
8 16
x3
x x2
1
3 uj
+ (j+1/2 1 j1/2 2 )x2
8
x3
2
uj
+(j+1/2 j1/2 )2 x
,
x2
(5.293)
where the coecients are dened as

+
+
+
1 = (+
1,2 + 2,2 + 1,2 + 2,2 ) , 2 = (1,2 + 2,2 1,2 2,2 ) , (5.294)
+
1 = (
2,2 2,2 ),
+
2 = (
1,2 1,2 ) .
(5.295)
Subgrid-viscosity models can be recovered by chosing adequately the values of the constants appearing in (5.293). The Smagorinsky model with length
scale = x and constant CS is obtained by taking:
1 = 2, 2 = CS , 1 = 2 = 0, j1/2 = |uj+1/2 uj1/2 | .
(5.296)
Variational Schemes with Embedded Subgrid Stabilization. We now

present nite element methods with some built-in subgrid stabilization [280,
79, 144, 299, 623, 331, 332, 336]. The presentation will be limited to the main
ideas for a simple linear advectiondiusion equation. The reader is referred
to original articles for detailed mathematical results and extension to Navier
Stokes equations. These methods are all based on the variational formulation
167
of the problem. For a passive scalar , we have:

dV +
t
u dV
x
=
=

2
dV +
f dV
2
x

dV +
f dV
x x
(5.297)
,
where is the uid domain, a weighting function, u the advection velocity

and f a source term. Boundary terms are assumed to vanish for the sake of
simplicity. Let L be the time-dependent advectiondiusion operator:
L=
+u
2
t
x
x
(5.298)
Using this operator, (5.297) can be recast under the symbolic compact
form
(, L) = (L , ) = a(, ) = (, f ) ,
(5.299)
where (., .) is a scalar product, a(., .) the bilinear form deduced from the
preceding equations and L the adjoint operator:
L =
u
2
t
x
x
(5.300)
We now split the trial and weighting functions as the sum of a resolved
and a subgrid function, i.e. = + and = + . Inserting these
decompositions into relation (5.299), we obtain
a(, ) = a( + , + ) = ( + , f )
(5.301)
and, assuming that and are linearly independent, we get the two following subproblems:
a(, ) + a(, ) = (, f )
(5.302)
and
a( , ) + a( , ) = ( , f )
(5.303)
a(, ) + (L , ) = (, f )
(5.304)
or, equivalently,
and
( , L) + ( , L ) = ( , f )
(5.305)
A rst solution consists of discretizing (5.304) and (5.305) using standard

shape functions for the resolved scales and oscillatory bubble functions for the
168
Fig. 5.20. Schematic of the embedded subgrid stabilization approach: linear nite
element shape functions in one dimension plus typical bubbles
subgrid scales (see Fig. 5.20). The resulting method is a two-scale method,
with embedded subgrid stabilization. It is important to note that degrees of
freedom associated with bubble functions are eliminated by static condensation, i.e. are expressed as functions of the resolved scales, and do not require
the solution of additional evolution equations.
Hughes and Stewart [336] proposed regularizing (5.304), which governs
the motion of resolved scales, yielding
a(, ) + (L , M (L f )) = (, f )
(5.306)
where (L f ) is the residual of the resolved scales and M an operator originating from an elliptic regularization, which can be evaluated using bubble
functions.
Usual stabilized methods also rely on the regularization of (5.304) without considering the subgrid scale equation. A general form of the stabilized
problem is
(5.307)
a(, ) + (IL, stab (L f )) = (, f ) ,
where, typically, IL is a dierential operator and stab is an algebraic operator
which approximates the integral operator (M ) of (5.306).
Classical examples are:
Standard Galerkin method, which does not introduce any stabilizing term:
IL = 0
(5.308)
169
Streamwise upwind PetrovGalerkin method, which introduces a stabilizing term based on the advection operator:
IL = u
(5.309)
Galerkin/least-squares method, which extends the preceding method by

including the whole dierential operator:
IL = L .
(5.310)
Subgrid Stabilization (Bubbles), which recovers the method of Hughes:

IL = L
(5.311)
The amount of numerical dissipation is governed by the parameter stab ,

which can assume either tensorial or scalar expression. Many denitions can
be found, most of them yielding |stab | x2 , which is the right scale for
a subgrid dissipation. Results dealing with the tuning of this parameter for
turbulent ow simulations are almost nonexistent.
Spectral Vanishing Viscosities. Karniadakis et al. [379, 395] propose to
adapt the Tadmor spectral viscosity [700] for large-eddy simulation purpose.
This approach will be presented using a simplied non-linear conservation
law for the sake of clarity. Considering the model Burgers equation
u
+
t
x
u2
2

=0
(5.312)
which can develop singularities, Tadmor proposes to regularize it for numerical purpose as

u2
u
u
+
=
Q
,
(5.313)
t
x 2
x
x
where , Q and u are the articial viscosity parameter, the articial viscosity
kernel and the regularized solution (interpreted as the resolved eld in largeeddy simulation), respectively. The original formulation of the regularization
proposed by Tadmor is expressed in the spectral space as

x

u
Q
=
x
u
k eikx
k 2 Q(k)
(5.314)
M|k|N
where k is the wave number, N the number of Fourier modes, and M the wave
number above which the articial viscosity is activated. Several forms for the
viscosity kernel have been suggested, among which the continuous kernel of
170
Maday [475] for pseudo-spectral methods based on Legendre polynomials:

(k N )2

Q(k) = exp
,
(k M )2
k>M
(5.315)
with M 5 N and 1/N . A major dierence with spectral functional

model in Fourier space is that Tadmor-type regularizations vanish at low
wave numbers, while functional subgrid viscosities dont. The extension of
the method to multidimensional curvilinear grids has been extensivly studied
by Pasquetti and Xu [773, 582, 581].
Karniadakis et al. [395] further modify this model by dening a dynamic
version of the articial viscosity in which the parameter is tuned regarding
the local state of the ow. To this end, the regularization term is redened
as
2 Qu
,
(5.316)
c(x, t)Q
x2
where the self-adaptive amplitude parameter c(x, t) can be computed considering either the gradient of the solution
c(x, t) =
|u|
N u
(5.317)
where is an adjustable arbitrary parameter, or the strain tensor

c(x, t) =
|S|
S
(5.318)
To prevent a too high dissipation near solid walls, Kirby and Karniadakis
multiply Q by a damping function
2
g(y ) = tan1
2ky +

+ 2
y
1 exp
C
(5.319)
where y is he distance to the wall, the superscript + refers to quantities

expressed in wall units and C is a parameter.
High-Order Filtered Methods. Visbal and Rizetta dene [734] another
procedure to perform large-eddy simulation based on numerical stabilization
without explicit physical subgrid model. Their procedure is based on the
application, at the end of each time, of an high-order low-pass lter to the
solution. This stabilizing procedure and the associated corresponding method
originating in the works by Visbal and Gaitonde is discussed in references
given therein. In practice, they use a symmetric compact nite dierence
lter with the followng properties:
5.4 Modeling the Backward Energy Cascade Process
1.
2.
3.
4.
it
it
it
it
171
is non-dispersive, i.e. it is strictly dissipative,

does not amplify any waves,
preserves constant functions,
completely eliminates the odd-even mode.
A tenth-order compact lter is observed to yield satisfactory results in

simple cases (decaying isotropic turbulence). It is important to notice that
this procedure is formally equivalent to the ltering-form of the full deconvolution procedure proposed by Mathew et al. [499] (see p. 220 for details).
Therefore, this implicit procedure can be completely rewritten within the
structural modeling framework. Other authors [63] develop similar strategies, based on the use of very-high order accurate nite dierence schemes.

5.4.1 Preliminary Remarks
The above models reect only the forward cascade process, i.e. the dominant
average eect of the subgrid scales. The second energy transfer mechanism,
the backward energy cascade, is much less often taken into account in simulations. We may mention two reasons for this. Firstly, the intensity of this
return is very weak compared with that of the forward cascade toward the
small scales (at least on the average in the isotropic homogeneous case) and
its role in the ow dynamics is still very poorly understood. Secondly, modeling it requires the addition of an energy source term to the equations being
computed, which is potentially a generator of numerical problems.
Two methods are used for modeling the backward energy cascade:
Adding a stochastic forcing term constructed from random variables and
the information contained in the resolved eld. This approach makes it
possible to include a random character of the subgrid scales, and each simulation can be considered a particular realization. The space-time correlations characteristic of the scales originating the backward cascade cannot
be represented by this approach, though, which limits its physical representativeness.
Modifying the viscosity associated with the forward cascade mechanism dened in the previous section, so as to take the energy injected at the large
scales into account. The backward cascade is then represented by a negative viscosity, which is added to that of the cascade model. This approach
is statistical and deterministic, and also subject to caution because it is
not based on a physical description of the backward cascade phenomenon
and, in particular, possesses no spectral distribution in k 4 predicted by the
analytical theories like EDQNM (see also footnote p. 104). Its advantage
resides mainly in the fact that it allows a reduction of the total dissipation
172
of the simulation, which is generally too high. Certain dynamic procedures

for automatically computing the constants can generate negative values of
them, inducing an energy injection in the resolved eld. This property is
sometimes interpreted as the capacity of the dynamic procedure to reect
the backward cascade process. This approach can therefore be classed in
the category of statistical deterministic backward cascade models.
Representing the backward cascade by way of a negative viscosity is controversial because the theoretical analyses, such as by the EDQNM model,
distinguish very clearly between the cascade and backward cascade terms,
both in their intensity and in their mathematical form [443, 442]. This representation is therefore to be linked to other statistical deterministic descriptions of the backward cascade, which take into account only an average reduction of the eective viscosity, such as the CholletLesieur eective viscosity
spectral model.
The main backward cascade models belonging to these two categories are
described in the following.
5.4.2 Deterministic Statistical Models
This section describes the deterministic models for the backward cascade.
These models, which are based on a modication of the subgrid viscosity
associated with the forward cascade process, are:
1. The spectral model based on the theories of turbulence proposed by Chasnov (p. 172). A negative subgrid viscosity is computed directly from the
EDQNM theory. No hypothesis is adopted concerning the spectrum shape
of the resolved scales, so that the spectral disequilibrium mechanisms can
be taken into account at the level of these scales, but the spectrum shape
of the subgrid scales is set arbitrarily. Also, the lter is assumed to be of
the sharp cuto type.
2. The dynamic model with an equation for the subgrid kinetic energy
(p. 173), to make sure this energy remains positive. This ensures that
the backward cascade process is represented physically, in the sense that
a limited quantity of energy can be restored to the resolved scales by the
subgrid modes. However, this approach does not allow a correct representation of the spectral distribution of the backward cascade. Only the
quantity of restored energy is controlled.
Chasnovs Spectral Model. Chasnov [120] adds a model for the backward cascade, also based on an EDQNM analysis, to the forward cascade
model already described (see Sect. 5.3.1). The backward cascade process is
represented deterministically by a negative eective viscosity term e (k|kc ),
which is of the form:
e (k|kc , t) =
F (k|kc , t)
2k 2 E(k, t)
(5.320)
173
The stochastic forcing term is computed as:

p
k3
F (k|kc , t) =
dp
dqkpq (1 2x2 z 2 xyz)E(q, t)E(p, t), (5.321)
pq
kc
pk
in which x, y, and z are geometric factors associated with the triad (k, p, q),
and kpq is a relaxation time described in Appendix B. As is done when
computing the draining term (see Chasnovs eective viscosity model in
Sect. 5.3.1), we assume that the spectrum takes the Kolmogorov form beyond the cuto kc . To simplify the computations, formula (5.321) is not used
for wave numbers kc p 3kc . For the other wave numbers, we use the
asymptotic form

14 4
E 2 (p, t)
k
F (k|kc , t) =
dpkpp (t)
.
(5.322)
15
p2
kc
This expression complete Chasnovs spectral subgrid model which, though
quite close to the Kraichnan type eective viscosity models, makes it possible
to take into account the backward cascade eects that are dominant for very
small wave numbers.
Localized Dynamic Model with Energy Equation. The GermanoLilly
dynamic procedure and the localized dynamic procedure lead to the denition of subgrid models that raise numerical stability problems because the
model constant can take negative values over long time intervals, leading to
exponential growth of the disturbances.
This excessive duration of the dynamic constant in the negative state corresponds to too large a return of kinetic energy toward the large scales [101].
This phenomenon can be interpreted as a violation of the spectrum realizability constraint: when the backward cascade is over-estimated, a negative
kinetic energy is implicitly dened in the subgrid scales. A simple idea for
limiting the backward cascade consists in guaranteeing spectrum realizability32 . The subgrid scales cannot then restore more energy than they contain.
To verify this constraint, local information is needed on the subgrid kinetic
energy, which naturally means dening this as an additional variable in the
simulation.
A localized dynamic model including an energy equation is proposed by
Ghosal et al. [261]. Similar models have been proposed independently by
Ronchi et al. [621, 511], Wong [765] and Kim and Menon [391, 392, 583].
The subgrid model used is based on the kinetic energy of the subgrid modes.
Using the same notation as in Sect. (5.3.3), we get:
+
* Q2 S
*
,
(5.323)
ij = 2
sgs ij
+
2 S
,
(5.324)
ij = 2 qsgs
ij
32
The spectrum E(k) is said to be realizable if E(k) 0, k.
174
2
in which the energies Q2sgs and qsgs
are dened as:
Q2sgs =
1 / * * 1
ui ui ui ui = Tii
2
2
1
1
(ui ui ui ui ) = ii
2
2
Germanos identity (5.138) is written:
2
qsgs
=
1
2
Q2sgs = q/
sgs + Lii
2
,
.
(5.325)
(5.326)
(5.327)
2
The model is completed by calculating qsgs
by means of an additional
evolution equation. We use the equation already used by Schumann, Horiuti,
and Yoshizawa, among others (see Sect. 5.3.1):
2
2
uj qsgs
qsgs
+
t
xj
3/2
2
(qsgs
)
ij S ij C1

2
2
+
2 qsgs
qsgs
2
qsgs
, (5.328)
+
+C2
xj
xj
xj xj
in which the constants C1 and C2 are computed by a constrained localized

dynamic procedure described above. The dynamic constant Cd is computed
by a localized dynamic procedure.
2
This model ensures that the kinetic energy qsgs
will remain positive, i.e.
that the subgrid scale spectrum will be realizable. This property ensures that
the dynamic constant cannot remain negative too long and thereby destabilize
the simulation. However, ner analysis shows that the realizability conditions
concerning the subgrid tensor (see Sect. 3.3.5) are veried only on the
condition:
+
+
2
2
qsgs
qsgs
Cd
,
(5.329)
3|s |
3s
where s and s are, respectively, the largest and smallest eigenvalues of
the strain rate tensor S. The model proposed therefore does not ensure the
realizability of the subgrid tensor.
The two constants C1 and C2 are computed using an extension of the
constrained localized dynamic procedure. To do this, we express the kinetic
energy Q2sgs evolution equation as:
*
uj Q2sgs
Q2sgs
+
t
xj
3/2
* C (Qsgs )
= Tij S
ij
1
*

2
+
Q
2 Q2sgs
sgs
Q2sgs
. (5.330)
+
+ C2
xj
xj
xj xj
175
One variant of the Germanos relation relates the subgrid kinetic energy
ux fj to its analog at the level of the test lter Fj :
2 /
2/
*j (p + qsgs
+ ui ui /2) uj (p + qsgs
+ ui ui /2) ,
Fj f*j = Zj u
(5.331)
in which p is the resolved pressure.

To determine the constant C2 , we substitute in this relation the modeled
uxes:
2
+
qsgs
2
fj = C2 qsgs
,
(5.332)
xj
2
+
* Q2 Qsgs
Fj = C2
sgs
xj
(5.333)
which leads to:

Zj = Xj C2 Y/
j C2
in which
2
+
* Q2 Qsgs
Xj =
sgs
xj
2
+
qsgs
2
Yj = qsgs
xj
(5.334)
(5.335)
(5.336)
Using the same method as was explained for the localized dynamic procedure, the constant C2 is evaluated by minimizing the quantity:

Zj Xj C2 + Y/
j C2

Zj Xj C2 + Y/
j C2
(5.337)
By analogy with the preceding developments, the solution is obtained in

the form:

3
C2 (x) = fC2 (x) + KC2 (x, y)C2 (y)d y
,
(5.338)
+
in which:
fC2 (x) =
1
Xj (x)Xj (x)

Xj (x)Zj (x) Yj (x) Zj (y)G(x y)d3 y
(5.339)
KC2 (x, y) =
C2
KA
(x, y)
C2
KA
(y, x)
Xj (x)Xj (x)
KSC2 (x, y)
(5.340)
176
in which
C2
KA
(x, y) = Xj (x)Yj (y)G(x y) ,
(5.341)

KSC2 (x, y)
G(z x)G(z y)d3 z
= Yj (x)Yj (y)
(5.342)
This completes the computation of constant C2 . To determine the constant C1 , we substitute (5.327) in (5.330) and get:
2
2
2
*
q/
uj q/
2 q/
Fj
sgs
sgs
sgs
+
= E
+
t
xj
xj
xj xj
(5.343)
in which E is dened as:

3/2
2
* + C1 (Qsgs )
E = Tij S
ij
*
1
1 2 Lii
+
2 xj xj
2
*j Lii
Lii
u
+
t
xj

. (5.344)
Applying the test lter to relation (5.328), we get:

2
2
q/
*
uj q/
sgs
sgs
/
+
= ij
S ij
t
xj

C1

/
2 )3/2
(qsgs
2
2 q/
f*j
sgs
+
xj
xj xj
. (5.345)
2
By eliminating the term q/
sgs /t between relations (5.343) and (5.345),
then replacing the quantity Fj f*j by its expression (5.331) and the quantity
Tij by its value as provided by the Germano identity, we get:
/1
= C1 C
(5.346)
in which
j
1
1 2 Lii
/
* L S
*
= ij
S ij *ij S
Dt Lii +
ij
ij ij +
xj
2
2 xj xj
= (Q2sgs )
(5.347)
* ,
/
(5.348)
/ ,
(5.349)
3/2
3/2
2
= (qsgs
)
and
*j (p + /
ui ui /2) uj (p +/
ui ui /2) .
j = u
(5.350)
177
*j /xj . The
The symbol Dt designates the material derivative /t + u
constant C1 is computed by minimizing the quantity

/1
C1 + C

/1
C1 + C
(5.351)
by a constrained localized dynamic procedure, which is written:

C1 (x) = fC1 (x) + KC1 (x, y)C1 (y)d3 y
(5.352)
in which
fC1 (x) =
1
(x)(x)
KC1 (x, y) =

(x)(x) (x)
(y)G(x y)d3 y
C1
C1
KA
(x, y) + KA
(y, x) KSC1 (x, y)
(x)(x)
(5.353)
(5.354)
in which
C1
KA
(x, y) = (x)(y)G(x y) ,
(5.355)

KSC1 (x, y) = (x)(y)
G(z x)G(z y)d3 z
(5.356)
which completes the computation of the constant C1 .

The version by Menon et al. [391, 392, 583], also extensively used by
Davidson and his coworkers [576, 577] is much simpler as far as the practical
implementation is addressed. This simplied formulation is dened as follows:
C1 =
test
(Q2sgs )3/2
and
C2 = Cd =
*
1 Ldij S ij
* S
*
2S
ij
(5.357)
(5.358)
ij
where Q2sgs is computed using relation (5.327) and the dissipation at the test
lter level is evaluated using a scale-similarity hypothesis, yielding
(
test = ( + sgs )
u/
i ui
xj xj
*
ui *
ui
xj xj
)
.
(5.359)
178
Since it is strictly local in the sense that no integral problem is involved,

this new formulation is much less demanding than the previous one in terms
of computational eort.
Another simplied local one-equation dynamic model was proposed by
Fureby [231], which is dened by the following relations:
C1 =
where
m
mm
(5.360)
/
2 3/2
3/2
2
qsgs
Qsgs
m=
*
*
= ij/
S ij Tij S
ij
t
1
Lkk
2
xj
1
*j
Lkk u
2
(5.361)

.
(5.362)
The remaining parameter is computed as follows:

Ldij Mij
Mij Mij
(5.363)

1
ij 2
ij
2
(5.364)
C2 = Cd =
where
Mij =
A more complex model is proposed by Krajnovic and Davidson [406], who

use a linear-combination model (see Sect. 7.4) to close both the momentum
equations and the prognostic equation for the subgrid kinetic energy.
5.4.3 Stochastic Models
Models belonging to this category are based on introducing a random forcing
term into the momentum equations. It should be noted that this random
character does not reect the space-time correlation scales of the subgrid
uctuations, which limits the physical validity of this approach and can raise
numerical stability problems. It does, however, obtain forcing term formulations at low algorithmic cost. The models described here are:
1. Bertoglios model in the spectral space (p. 179). The forcing term is constructed using a stochastic process, which is designed in order to induce
the desired backward energy ux and to possess a nite correlation time
scale. This is the only random model for the backward cascade derived
in the spectral space.
179
2. Leiths model (p. 180). The forcing term is represented by an acceleration

vector deriving from a vector potential, whose amplitude is evaluated by
simple dimensional arguments. The backward cascade is completely decoupled from forward cascade here: there is no control on the realizability
of the subgrid scales.
3. MasonThomson model (p. 182), which can be considered as an improvement of the preceding model. The evaluations of the vector potential amplitude and subgrid viscosity modeling the forward cascade are coupled,
so as to ensure that the local equilibrium hypothesis is veried. This
ensures that the subgrid kinetic energy remains positive.
4. Schumann model (p. 183), in which the backward cascade is represented
not as a force deriving from a vector potential but rather as the divergence
of a tensor constructed from a random solenoidal velocity eld whose
kinetic energy is equal to the subgrid kinetic energy.
5. Stochastic dynamic model (p. 184), which makes it possible to calculate
the subgrid viscosity and a random forcing term simultaneously and dynamically. This coupling guarantees that the subgrid scales are realizable,
but at the cost of a considerable increase in the algorithmic complexity
of the model.
Bertoglio Model. Bertoglio and Mathieu [57, 58] propose a spectral stochastic subgrid model based on the EDQNM analysis. This model appears as
a new source term fi (k, t) in the ltered mometum equations, and is evaluated as a stochastic process. The following constraints are enforced:
f must not modify the velocity eld incompressibility, i.e. ki fi (k, t) = 0;

f will have a Gaussian probability density function;
The correlation time of f , noted tf , is nite;
f must induce the desired eect on the statistical second-order moments
of the resolved velocity eld:
fi (k, t)

uj (k, t) + fj (k, t)
ui (k, t) = Tij (k, t)
2
L
3
,
(5.365)
where Tij (k, t) is the exact backward transfer term appearing in the vari
ui (k, t)
uj (k, t) and L the size of the computational
ation equation for
domain in physical space.
Assuming that the response function of the simulated eld is isotropic
and independent of f , and that the time correlations exhibit an exponential
decay, we get the following velocity-independent relation:
fi (k, t)fj (k, t) + fi (k, t)fj (k, t) = Tij (k, t)
2
L
3
1
1
+
(k, t) tf

,
(5.366)
180
where (k, t) is a relaxation time evaluated from the resolved scales. We now
have to compute the stochastic variable fi . The authors propose the following
algorithm, which is based on three random variables a, b and c:
'

+
t
t
(n+1)
(n)
(n) (n+1)
=
1
exp(2a(n+1) )
f1 + h11 11
f1
tf
tf
'
+
t
(n) (n+1)
+ h22 12
exp(2c(n+1) ) ,
(5.367)
tf
'

+
t
t
(n+1)
(n)
(n) (n+1)
=
1
exp(2b(n+1) )
f2
f2 + h22 22
tf
tf
'
+
t
(n) (n+1)
exp(2c(n+1) ) ,
(5.368)
+ h11 21
tf
where the superscript (n) denotes the value at the nth time step, t is the
value of the time step, and hij (k, t) = fi (k, t)fj (k, t). Moreover, we get the
complementary set of equations, which close the system:

1
(n+1) 2
(n+1)
(n) tf
(n) (n+1) 2
h22 (12 )
h11 )
(h11
(11 ) =
(n)
t
h
11
(n+1) 2
(n+1) (n+1)
12
(n+1) 2
(22
12
(21
t
+2
,
(5.369)
tf

1
(n+1)
(n) tf
(n) (n+1) 2
h
)
(
)
(h
22
22
11
21
(n)
t
h22
t
+2
,
(5.370)
tf

1
(n+1)
(n) tf
+
h12 )
(h12
t
(n) (n)
h11 h11

t
(n)
+h12 2
,
(5.371)
tf
(n+1) 2
(12
(5.372)
which completes the description of the model. The resulting random force
satises all the cited constraints, but it requires the foreknowledge of the hij
tensor. This tensor is evaluated using the EDQNM theory, which requires
the spectrum of the subgrid scales to be known. To alleviate this problem,
arbitrary form of the spectrum can be employed.
Leith Model. A stochastic backward cascade model expressed in the physical space was derived by Leith in 1990 [434]. This model takes the form of
a random forcing term that is added to the momentum equations. This term
181
is computed at each point in space and each time step with the introduction
of a vector potential b for the acceleration, in the form of a white isotropic
noise in space and time. The random forcing term with null divergence f b is
deduced from this vector potential.
We rst assume that the space and time auto-correlation scales of the
subgrid modes are small compared with the cuto lengths in space and in
time t associated with the lter33 . This way, the subgrid modes appear to
be de-correlated in space and time. The correlation at two points and two
times of the vector potential b is then expressed:
bi (x, t)bk (x , t ) = (x, t)(x x )(t t )ik
(5.373)
in which is the variance. This is computed as:

(x, t) =
1
3
dt
d3 x bk (x, t)bk (x , t )
(5.374)
Simple dimensional reasoning shows that:

7
(x, t) |S|3
(5.375)
Also, as the vector potential appears as a white noise in space and time
at the xed resolution level, the integral (5.374) is written:
(x, t) =
1 b
3
(x, t)bk (x, t) t .
3 k
(5.376)
Considering relations (5.375) and (5.376), we get:

4
bk (x, t)bk (x, t) |S|3
1
t
(5.377)
The shape proposed for the kth component of the vector potential is:
2
bk = Cb |S|3/2 t1/2 g
(5.378)
in which Cb is a constant of the order of unity, t the simulation time cuto

length (i.e. the time step), and g the random Gaussian variable of zero average
and variance equal to unity. The vector f b is then computed by taking the
rotational of the vector potential, which guarantees that it is solenoidal.
In practice, Leith sets the value of the constant Cb at 0.4 and applies
a spatial lter with a cuto length of 2, so as to ensure better algorithm
stability.
33
We again nd here a total scale separation hypothesis that is not veried in

reality.
182
MasonThomson Model. A similar model is proposed by Mason and

Thomson [498]. The dierence from the Leith model resides in the scaling of
the vector potential. By calling f and the characteristic lengths of the subgrid scales and spatial lter, respectively, the variants of the resolved stresses
due to the subgrid uctuations is, if f , of the order of (f /)3 u4e ,
in which ue is the characteristic subgrid velocity. The amplitude a of the
uctuations in the gradients of the stresses is:
3/2
5/2
u2e
(5.379)
which is also the amplitude of the associated acceleration. The corresponding

kinetic energy variation rate of the resolved scales, qr2 , is estimated as:
qr2
3
a2 te f5 u4e te
t
(5.380)
in which te is the characteristic time of the subgrid scales. As te f /ue and

the dissipation rate is evaluated by dimensional arguments as u3e /f , we
can say:
5
qr2
= Cb f5 .
(5.381)
t
The ratio f / is evaluated as the ratio of the subgrid scale mixing length
to the lter cuto length, and is thus equal to the constant of the subgrid
viscosity models discussed in Sect. 5.3.2. Previous developments have shown
that this constant is not unequivocally determinate, but that it is close to
0.2. The constant Cb is evaluated at 1.4 by an EDQNM analysis.
The dissipation rate that appears in equation (5.381) is evaluated in light
of the backward cascade. The local subgrid scale equilibrium hypothesis is
expressed by:
5
(5.382)
ij S ij = + Cb f5 ,
in which ij is the subgrid tensor. The term on the left represents the
subgrid kinetic energy production, the rst term in the right-hand side the
dissipation, and the last term the energy loss to the resolved scales by the
backward cascade. The dissipation rate is evaluated using this last relation:
=
ij S ij
1 + (f /)5
(5.383)
which completes the computation of the right-hand side of equation (5.381),

with the tensor ij being evaluated using a subgrid viscosity model.
183
This equation can be re-written as:

qr2
= a2 t ,
t
(5.384)
in which a2 is the sum of the variances of the acceleration component amplitudes. From the equality of the two relations (5.381) and (5.384), we can
say:
5
.
(5.385)
a2 = Cb f5
t
The vector potential scaling factor a and a2 are related by:

t
a = a2
.
te
(5.386)
To complete the model, we now have to evaluate the ratio of the subgrid
scale characteristic time to the time resolution scale. This is done simply by
evaluating the characteristic time te from the subgrid viscosity sgs computed
by the model used, to reect the cascade:
te =
2f
sgs
(5.387)
which completes the description of the model, since the rest of the procedure
is the same as what Leith dened.
Schumann Model. Schumann proposed a stochastic model for subgrid tensor uctuations that originate the backward cascade of kinetic energy [654].
The subgrid tensor is represented as the sum of a turbulent viscosity model
and a stochastic part Rst :
2 2
st
ij = sgs S ij + qsgs
ij + Rij
3
(5.388)
st
are zero:
The average random stresses Rij
st
=0 .
Rij
(5.389)
They are dened as:

st
Rij

2 2
= m vi vj qsgs ij
3
(5.390)
in which m is a parameter and vi a random velocity. From dimensional

arguments, we can dene this as:

2
2qsgs
gi ,
vi =
(5.391)
3
184
in which gi is a white random number in space and has a characteristic

correlation time v :
gi = 0 ,
gi (x, t)gj (x , t ) = ij (x x ) exp(|t t |/v ) .
(5.392)
(5.393)
The vi eld is made solenoidal by applying a projection step. We note

that the time scale v is such that:
+
2 / 1 .
(5.394)
v qsgs
The parameter m determines the portion of random stresses that generate the backward cascade. Assuming that only the scales belonging to the
interval [kc , nkc ] are active, for a spectrum of slope of m we get:

2
m
=
nkc
k 2m dk
kc
k
2m
= 1 n12m
(5.395)
dk
kc
For n = 2 and m = 5/3, we get m = 0.90. The subgrid kinetic energy

2
qsgs
is evaluated from the subgrid viscosity model.
Stochastic Localized Dynamic Model. A localized dynamic procedure
including a stochastic forcing term was proposed by Carati et al. [101]. The
contribution of the subgrid terms in the momentum equation appears here
as the sum of a subgrid viscosity model, denoted Cd ij using the notation of
Sect. 5.3.3, which models the energy cascade, and a forcing term denoted f :
ij
Cd ij
=
+ fi
xj
xj
(5.396)
The ij term can be computed using any subgrid viscosity model. The
force f is chosen in the form of a white noise in time with null divergence in
space. The correlation of this term at two points in space and two times is
therefore expressed:
fi (x, t)fj (x , t ) = A2 (x, t)Hij (x x )(t t ) .
(5.397)
The statistical average here is an average over all the realizations of f

conditioned by a given velocity eld u(x, t). The factor A2 is such that
Hii (0) = 1. Since a stochastic term has been introduced into the subgrid
model, the residual Eij on which the dynamic procedure for computing the
constant Cd is founded also possesses a stochastic nature. This property will
therefore be shared by the dynamically computed constant, which is not acceptable. To nd the original properties of the dynamic constant, we take
a statistical average of the residual, denoted Eij , which gets rid of the
185
random terms. The constant of the subgrid viscosity model is computed by

a localized dynamic procedure based on the statistical average of the residual,
which is written:
.
(5.398)
Eij = Lij + C/
d ij Cd ij
The amplitude of the random forcing term can also be computed dynamically. To bring out the non-zero contribution of the stochastic term in the
statistical average, we base this new procedure on the resolved kinetic energy
*i /2. The evolution equation of
*i u
balance at the level of the test lter Q2r = u
this quantity is obtained in two dierent forms (only the pertinent terms are
detailed, the others are symbolized):
Q2r
*i (Cd ij + P ij ) + EF
= ... u
t
xj
(5.399)

Q2r
*i C/
= ... u
*ij + Ef* .
d ij + Lij + p
t
xj
(5.400)
The pressure terms P and p are in equilibrium with the velocity elds
* and u, respectively. The quantities EF and E * are the backward cascade
u
f
energy injections associated, respectively, with the forcing term F computed
* computed
directly at the level of the test lter, and with the forcing term f
at the rst level and then ltered. The dierence between equations (5.399)
and (5.400) leads to:
(5.401)
Z EF Ef* g = 0 ,
in which the fully known term g is of the form:
*i
g=u

Cd ij + P ij C/
*ij
d ij Lij p
xj
(5.402)
The quantity Z plays a role for the kinetic energy that is analogous to
the residual Eij for the momentum. Minimizing the quantity

Z=
Z2
(5.403)
can thus serve as a basis for dening a dynamic procedure for evaluating the
stochastic forcing.
To go any further, the shape of the f term has to be specied. To simplify
the use, we assume that the correlation length of f is small compared with
the cuto length . The function f thus appears as de-correlated in space,
which is reected by:
1
(5.404)
Ef = A2 (x, t) .
2
186
In order to be able to calculate Ef dynamically, we assume that the backward cascade is of equal intensity at the two ltering levels considered, i.e.
Ef = EF .
(5.405)
* scale, we assume:
Also, since f is de-correlated at the
Ef* Ef = EF ,
(5.406)
which makes it possible to change relation (5.401) to become

Z = EF g
(5.407)
We now choose f in the form:

fi = Pij (Aej ) ,
(5.408)
in which ej is a random isotropic Gaussian function, A a dimensioned constant that will play the same role as the subgrid viscosity model constant,
and Pij the projection operator on a space of zero divergence. We have the
relations:
ei (x, t) = 0 ,
(5.409)
1
ij (t t )(x x ) .
3
Considering (5.408), (5.410) and (5.404), we get:
ei (x, t)ei (x , t ) =
Ef =
1 2
1
A = A2
2
3
(5.410)
(5.411)
The computation of the model is completed by evaluating the constant

A by a constrained localized dynamic procedure based on minimizing the
functional (5.403), which can be re-written in the form:

Z[A] =
A2
g
3
2
.
(5.412)
6. Functional Modeling:
Extension to Anisotropic Cases
6.1 Statement of the Problem

The developments of the previous chapters are all conducted in the isotropic
framework, which implies that both the lter used and the ow are isotropic.
They can be extended to anisotropic or inhomogeneous cases only by localizing the statistical relations in space and time and introducing heuristic procedures for adjusting the models. But when large-eddy simulation is applied
to inhomogeneous ows, we very often have to use anisotropic grids, which
correspond to using a anisotropic lter. So there are two factors contributing
to the violation of the hypotheses underlying the models presented so far: lter anisotropy (respectively inhomogeneity) and ow anisotropy (respectively
inhomogeneity).
This chapter is devoted to extensions of the modeling to anisotropic cases.
Two situations are considered: application of a anisotropic homogeneous lter
to an isotropic homogeneous turbulent ow (Sect. 6.2), and application of an
isotropic lter to an anisotropic ow (Sect. 6.3).
6.2 Application of Anisotropic Filter to Isotropic Flow

The lters considered in the following are anisotropic in the sense that the
lter cuto length is dierent in each direction of space. The dierent types
of anisotropy possible for Cartesian ltering cells are represented in Fig. 6.1.
In order to use an anisotropic lter to describe an isotropic ow, we are
rst required to modify the subgrid models, because theoretical work and
numerical experiments have shown that the resolved elds and the subgrid
thus dened are anisotropic [368]. For example, for a mesh cell with an aspect
ratio 2 /1 = 8, 3 /1 = 4, the subgrid stresses will dier from their values
obtained with an isotropic lter by about ten percent. It is very important to
note, though, that this anisotropy is an artifact due to the lter but that the
dynamic of the subgrid scales still corresponds that of isotropic homogeneous
turbulence.
On the functional modeling level, the problem is in determining the characteristic length that has to be used to compute the model.
188
6. Functional Modeling: Extension to Anisotropic Cases
Fig. 6.1. Dierent types of ltering cells. Isotropic cell (on the left): 1 = 2 = 3 ;
pancake-type anisotropic cell (center): 1 2 3 ; cigar-type anisotropic cell
(right): 1 2 3 .
Two approaches are available:

The rst consists in dening a single length scale for representing the lter.
This lets us keep models analogous to those dened in the isotropic case,
using for example scalar subgrid viscosities for representing the forward
cascade process. This involves only a minor modication of the subgrid
models since only the computation of the characteristic cuto scale is modied. But it should be noted that such an approach can in theory be valid
only for cases of low anisotropy, for which the dierent cuto lengths are
of the same order of magnitude.
The second approach is based on the introduction of several characteristic
length scales in the model. This sometimes yields major modications in
the isotropic models, such as the denition of tensorial subgrid viscosities
to represent the forward cascade process. In theory, this approach takes the
lter anisotropy better into account, but complicates the modeling stage.
6.2.1 Scalar Models
These models are all of the generic form = (1 , 2 , 3 ). We present
here:
1. Deardors original model and its variants (p. 189). These forms are empirical and have no theoretical basis. All we do is simply to show that they
are consistent with the isotropic case, i.e. = 1 when 1 = 2 = 3 .
2. The model of Scotti et al. (p. 189), which is based on a theoretical analysis considering a Kolmogorov spectrum with an anisotropic homogeneous
lter. This model makes a complex evaluation possible of the lter cuto
length, but is limited to the case of Cartesian ltering cells.
189
Deardor s Proposal. The method most widely used today is without

doubt the one proposed by Deardor [172], which consists in evaluating the
lter cuto length as the cube root of the volume V of the ltering cell .
Or, in the Cartesian case:

1/3
(x) = 1 (x)2 (x)3 (x)
(6.1)
in which i (x) is the lter cuto length in the ith direction of space at
position x.
Extensions of Deardor s Proposal. Simple extensions of denition (6.1)
are often used, but are limited to the case of Cartesian ltering cells:
(x) =
+
2
2
2
(1 (x) + 2 (x) + 3 (x))/3 ,

(x) = max 1 (x), 2 (x), 3 (x)
(6.2)
(6.3)
Another way to compute the characteristic length on a strongly nonuniform mesh, which prevents the occurrence of large values of subgrid viscosity, was proposed by Arad [16]. It relies on the use of the harmonic mean
of the usual length scales:

1/3
(x) = 1 (x)2 (x)3 (x)
with
1/

i (x) = (i (x)) + (M )
i = 1, 2, 3 ,
(6.4)
(6.5)
where i is a prescribed bound for i (x), and > 0.

Proposal of Scotti et al. More recently, Scotti, Meneveau, and Lilly [664]
proposed a new denition of based on an improved estimate of the dissipation rate in the anisotropic case. The lter is assumed to be anisotropic
but homogeneous, i.e. the cuto length is constant in each direction of space.
We dene max = max(1 , 2 , 3 ). Aspect ratios of less than unity,
constructed from the other two cuto lengths with respect to max , are
denoted a1 and a2 1 . The form physically sought for the anisotropy correction
is:
= iso f (a1 , a2 )
(6.6)
in which iso is Deardors isotropic evaluation computed by relation (6.1).

1
For example, by taking max = 1 , we get a1 = 2 /1 and a2 = 3 /1 .
190
Using the approximation:

2
= 2S ij S ij 3/2
(6.7)
and the following equality, which is valid for a Kolmogorov spectrum,

K0
2 5/3 3

|G(k)|
k
d k ,
(6.8)
S ij S ij = 2/3
2

where G(k)
is the kernel of the anisotropic lter considered, after calculation
we get:

3/4

K0
2 5/3 3

|G(k)| k
=
d k
.
(6.9)
2
Considering a sharp cuto lter, we get the following approximate relation
by integrating equation (6.9):

4
[(ln a1 )2 ln a1 ln a2 + (ln a2 )2 ] .
f (a1 , a2 ) = cosh
(6.10)
27
It is interesting to note that the dynamic procedure (see Sect. 5.3.3) for the
computation of the Smagorinsky constant can be interpreted as an implicit
way to compute the f (a1 , a2 ) function [663]. Introducing the subgrid mixing
length f , the Smagorinsky model reads
sgs
=
=
=
2f |S|
(6.11)
2
Cd iso |S|
(CS iso f (a1 , a2 )) |S| ,

2
(6.12)
(6.13)
where Cd is the value of the constant computed using a dynamic procedure,

and CS the theoretical value of the Smagorinsky constant evaluated through
the canonical analysis. A trivial identication leads to:

f (a1 , a2 ) = Cd /CS .
(6.14)
This interpretation is meaningful for positive values of the dynamic constant. A variant can be derived by using the anisotropy measure f (a1 , a2 ) instead of the isotropic one inside the dynamic procedure (see equation (6.12)),
yielding new denitions of the tensors ij and ij appearing in the dynamic
procedure (see Table 5.1). Taking the Smagorinsky model as an example, we
get:
* f (*
*S
*
a2 ))2 |S|
ij = 2(iso f (a1 , a2 ))2 |S|S ij , ij = 2(
iso a1 , *
ij
, (6.15)
where f (a1 , a2 ) and f (*

a1 , *
a2 ) are the anisotropy measures associated to the
rst and second ltering levels, respectively. The corresponding formulation
of the f function is now:

f (a1 , a2 ) = Cd /(CS iso ) .
(6.16)
191
6.2.2 Battens Mixed Space-Time Scalar Estimator

It was shown in Sect. 2.1.3 that spatial ltering induces a time ltering. In
a reciprocal manner, enforcing a time-frequency cuto leads to the denition of an intrinsic spatial cuto length. To account for that phenomenon,
Batten [49] denes the cuto length as
+

2 t
(x) = 2 max 1 (x), 2 (x), 3 (x), qsgs
(6.17)
2
where t and qsgs
are the time step of the simulation and the subgrid
kinetic energy, respectively. The subgrid kinetic energy can be computed
solving an prognostic transport equation (p. 128) or one of the methods
discussed in Sect. 9.2.3.
6.2.3 Tensorial Models

The tensorial models presented in the following are constructed empirically,
with no physical basis. They are justied only by intuition and only for
highly anisotropic ltering cells of the cigar type, for example (see Fig. 6.1).
Representing the lter by a single and unique characteristic length is no longer
relevant. The lters characteristic scales and their inclusion in the subgrid
viscosity model are determined intuitively. Two such models are described:
1. The model of Bardina et al. (p. 191), which describes the geometry of the
ltering cell by means of six characteristic lengths calculated from the
inertia tensor of the ltering cell. This approach is completely general and
is a applicable to all possible types of ltering cells (Cartesian, curvilinear,
and other), but entrains a high complexication in the subgrid models.
2. The model of Zahrai et al. (p. 192), which is applicable only to Cartesian
cells and is simple to include in the subgrid viscosity models.
Proposal of Bardina et al.
Denition of a Characteristic Tensor. These authors [39] propose replacing
the isotropic scalar evaluation of the cuto length associated with the grid
by an anisotropic tensorial evaluation linked directly to the ltering cell geometry: V (x) = (1 (x)2 (x)3 (x)). To do this, we introduce the moments
of the inertia tensor I associated at each point x:
Iij (x) =
1
V (x)

xi xj dV
(6.18)
Since the components of the inertia tensor are homogeneous at the square
of a length, the tensor of characteristic lengths is obtained by taking the
square root of them. In the case of a pancake ltering cell aligned with the
192
axes of the Cartesian coordinate system, we get the diagonal matrix:
2 1
Iij = 0
3
0
0
2
2
0
0
2
3
(6.19)
Application to the Smagorinsky Model. As we model only the anisotropic part

of the subgrid tensor, the tensor I is decomposed into the sum of a spherical
term I i and an anisotropic term I d :
d
Iij = I i ij + (Iij I i ij ) = I i ij + Iij
(6.20)
with
Ii =
1
1 2
2
2
Ikk = (1 + 2 + 3 ) .
3
3
(6.21)
Modifying the usual Smagorinsky model, the authors nally propose the
following anisotropic tensorial model for deviator of the subgrid tensor :
1
ij kk ij
3
=
+
+
C1 I i |S|S ij

1
C2 |S| Iik S kj + Ijk S ki Ilk S kl ij
3

1
|S|
,
C3 i Iik Ijl S kl Imk Iml S kl ij
I
3
(6.22)
in which C1 , C2 and C3 are constants to be evaluated.

Proposal of Zahrai et al.
Principle. Zahrai et al. [795] proposed conserving the isotropic evaluation of
the dissipation rate determined by Deardor and further considering that
this quantity is constant over each mesh cell:
2/3

2S ij S ij 3/2
= 1 (x)2 (x)3 (x)
(6.23)
On the other hand, when deriving the subgrid model, we consider that
the lters characteristic length in each direction is equal to the cuto length
in that direction. This procedure calls for the denition of a tensorial model
for the subgrid viscosity.
Application to the Smagorinsky model. In the case of the Smagorinsky model,
we get for component k:
(sgs )k = C1 (1 2 3 )2/9 (k )4/3 2S ij S ij 3/2
where C1 is a constant.
(6.24)
6.3 Application of an Isotropic Filter to a Shear Flow
193

We will now be examining the inclusion of subgrid scale anisotropy in the
functional models.
The rst part of this section presents theoretical results concerning subgrid scale anisotropy and the interaction mechanisms between the large and
small scales in this case. These results are obtained either by the EDQNM
theory or by asymptotic analysis of the triadic interactions.
The second part of the section describes the modications that have been
proposed for functional type subgrid models. Only models for the forward
energy cascade will be presented, because no model for the backward cascade
has yet been proposed in the anisotropic case.
6.3.1 Phenomenology of Inter-Scale Interactions
Anisotropic EDQNM Analysis. Aupoix [24] proposes a basic analysis of
the eects of anisotropy in the homogeneous case using Cambons anisotropic
EDQNM model. The essential details of this model are given in Appendix B.
The velocity eld u is decomposed as usual into average part u and
a uctuating part u :
u = u + u
(6.25)
To study anisotropic homogeneous ows, we dene the spectral tensor

ij (k) =
u
uj (k)
i (k)
(6.26)
which is related to the double correlations in the physical space by the relation:

ui uj (x) =
ij (k)d3 k .
(6.27)
Starting with the NavierStokes equations, we obtain the evolution equation (see Appendices A and B):

2
+ 2k ij (k) +
t
ui
uj
jl (k) +
il (k)
xl
xl
ul
(ki jm (k) + kj mi (k))
xm
ul
(kl ij (k))
xm km
Pil (k)Tlj (k) + Pjl (k)Tli (k) ,
(6.28)
194
where

ui (k)ul (p)uj (k p)d3 p
Tij (k) = kl
and
Pij (k) =

ki kj
ij 2
k
(6.29)
(6.30)
and where the * designates the complex conjugate number. We then simplify
the equations by integrating the tensor on spheres of radius k=cste:

ij (k) =
ij (k)dA(k) ,
(6.31)
and obtain the evolution equations:

2
+ 2k ij (k) =
t
+
ui
uj
jl (k)
il (k)
xk
xl
l
nl
Pijl (k) + Sij
(k) + Pijnl (k) + Sij
(k)
,
(6.32)
where the terms P l , S l , P nl and S nl are the linear pressure, linear transfer,
non-linear pressure, and non-linear transfer contributions, respectively. The
linear terms are associated with the action of the average velocity gradient,
and the non-linear terms with the action of the turbulence on itself.
The expression of these terms and their closure by the anisotropic
EDQNM approximation are given in Appendix B. Using these relations,
Aupoix derives an expression for the interaction between the modes corresponding to wave numbers greater than a given cuto wave number kc (i.e.
the small or subgrid scales) and those associated with small wave numbers
such that k kc (i.e. the large or resolved scales). To obtain a simple expression for the coupling among the dierent scales by the non-linear terms
P nl and S nl , we adopt the hypothesis that there exists a total separation of
scales (in the sense dened in Sect. 5.3.2) between the subgrid and resolved
modes, so that we can obtain the following two asymptotic forms:
Pijnl (k) =
+
E 2 (p)Hij (p)
0pp [10 + a(p)]
dp
p2
kc

16 2
k E(k)
0pp (a(p) + 3)p (E(p)Hij (p))
105
p
kc

a(p)
+E(p)Hij (p) 5 {a(p) + 3} + p
dp ,
(6.33)
p
32 4
k
175

nl
Sij
(k) =
195

E 2 (p) 14 1
8
ij + 2Hij (p) + a(p)Hij (p) dp
2k 4
0pp 2
p
15 3
25
kc

E(p)
1
2k 2 ij (k)
0pp 5E(p) + p
dp
15 kc
p
7

2
2
0pp
2k E(k)
5E(p)Hij (p) + p (E(p)Hij (p))
15
p
kc
7

8
8
(a(p) + 3) + a(p) dp ,
(6.34)
+E(p)Hij (p)
15
25

where E(k) is the energy spectrum, dened as:

E(k) =
1
ll (k)
2
(6.35)
and Hij (k) the anisotropy spectrum:

Hij (k) =
ij (k) 1
ij
2E(k) 3
(6.36)
It is easily veried that, in the isotropic case, Hij cancels out by construction. The function a(k) is a structural parameter that represents the
anisotropic distribution on the sphere of radius k, and kpq the characteristic relaxation time evaluated by the EDQNM hypotheses. The expression of
this term is given in Appendix B.
These equations can be simplied by using the asymptotic value of the
structural parameter a(k). By taking a(k) = 4.5, we get:

E 2 (p) 28
368
nl
nl
4
ij
Hij (p) dp
0pp 2
Pij (k) + Sij (k) = k
p
45
175
kc

1
E(p)
2
2k ij (k)
0pp 5E(p) + p
dp
15 kc
p

1052
E(p)Hij (p)
0pp
+ k 2 E(k)
525
kc

52
(E(p)Hij (p)) dp .
(6.37)
105 p
From this equation, it can be seen that the anisotropy of the small scales
takes on a certain importance. In a case where the anisotropic spectrum has
the same (resp. opposite) sign for the small scales as it does for the large, the
term in k 4 constitutes a return of energy that has the eect of a return toward
isotropy (resp. departure from isotropy), and the term in k 2 E(k) represents
an backward energy cascade associated with an increasing anisotropy (resp.
a return to isotropy). Lastly, the term in k 2 ij (k) is a term of isotropic
drainage of energy to the large scales by the small, and represents here the
energy cascade phenomenon modeled by the isotropic subgrid models.
196
Asymptotic Analysis of Triadic Interactions. Another analysis of interscale interactions in the isotropic case is the asymptotic analysis of triadic
interactions [74, 785].
(k) is written in the symbolic
The evolution equation of the Fourier mode u
form:
(k)
u
= u(k)
= [u(k)]
,
(6.38)
nl + [u(k)]
vis
t
where [u(k)]
nl and [u(k)]
vis represent, respectively, the non-linear terms associated with the convection and pressure, and the linear term associated
with the viscous eects, dened as:
[u(k)]
nl = i
(p)k (k u
(k p))
u
(6.39)
with
u
i (p)k

ki kj
= ij 2 u
j (p) ,
k
(6.40)
(k) .
[u(k)]
vis = k u
(6.41)
(k) u
(k), is of the
The evolution equation of the modal energy, e(k) = u
form:
e(k)
(k) u (k) + cc = [e(k)]
=u
,
(6.42)
nl + [e(k)]
vis
t
with
[e(k)]
nl = i
3
4
(k) u
(p) [k u
(k p)] + cc
u
(6.43)
p
2
[e(k)]
vis = 2k e(k) ,
(6.44)
where the symbol cc designates the complex conjugate number of the term
that precedes it. The non-linear energy transfer term brings in three wave
vectors (k, p, q = kp) and is consequently a linear sum of non-linear triadic
interactions. We recall (see Sect. 5.1.2) that the interactions can be classied
into various categories ranging from local interactions, for which the norms
of the three wave vectors are similar (i.e. k p q), to distant interactions
for which the norm of one of the wave vectors is very small compared with
the other two (for example k p q). The local interactions therefore
correspond to the inter-scale interactions of the same size and the distant
interactions to the interactions between a large scale and two small scales.
197
Also, any interaction that introduces a (k, p, q) triad that does not verify the
relation k p q is called a non-local interaction.
In the following, we will be analyzing an isolated distant triadic interaction
associated with three modes: k, p and q. We adopt the conguration k
p q and assume that k is large scale located in the energetic portion of the
spectrum. An asymptotic analysis shows that:
[u(k)]
nl = O()
(6.45)

3
4
(k) + O()
(q) p u
[u(p)]
nl = i u
(6.46)

3
4
(p) p u
(k) + O()
[u(q)]
nl = i u
(6.47)
where is the small parameter dened as

=
k
1
p
The corresponding energy transfer analysis leads to the following relations:

[e(k)]
nl = O()
(q) [p u
(k)] + cc} + O()
u(p) u
[e(p)]
nl = [e(q)]
nl = i {
(6.48)
.
(6.49)
Several remarks can be made:

The interaction between large and small scales persists in the limit of the
innite Reynolds numbers. Consistently with the Kolmogorov hypotheses,
these interactions occur with no energy transfer between the large and
small scales. Numerical simulations have shown that the energy transfers
are negligible between two modes separated by more than two decades.
(p) and u
(q) is directly pro The variation rate of the high frequencies u
(k). This implies
portional to the amplitude of the low-frequency mode u
that the strength of the coupling with the low-frequency modes increases
with the energy of the modes.
Moreover, complementary analysis shows that, for modes whose wavelength is of the order of the Taylor micro-scale dened as (see Appendix
A):
8
9
9 u 2
(6.50)
=9
9 2 ,
: u
x
198
the ratio between the energy transfers due to the distant interactions and
those due to the local interactions vary as:
[e(k
)]distant
11/6
Re
[e(k
)]local
(6.51)
where Re is the Reynolds number referenced to the Taylor micro-scale and

the velocity uctuation u . This relation shows that the coupling increases
with the Reynolds number, with the result that an anisotropic distribution
of the energy at the low frequencies creates an anisotropic forcing of the high
frequencies, leading to a deviation from isotropy of these high frequencies.
A competitive mechanism exists that has an isotropy reduction eect
at the small scales. This is the energy cascade associated with non-local
triadic interactions that do not enter into the asymptotic limit of the distant
interactions.
For a wave vector of norm k, the ratio of the characteristic times
(k)cascade and (k)distant , associated respectively with the energy transfer
of the cascade mechanism and that due to the distant interactions, is evaluated as:
(k)cascade
constant (k/kinjection)11/6 ,
(6.52)
(k)distant
where kinjection is the mode in which the energy injection occurs in the
spectrum. So we see that the distant interactions are much faster than the
energy cascade. Also, the rst eect of a sudden imposition of large scale
anisotropy will be to anisotropize the small scales, followed by competition
between the two mechanisms. The dominance of one of the two depends on
a number of factors, such as the separation between the k and kinjection scales,
or the intensity and coherence of the anisotropy at the large scale.
Numerical simulations [785] performed in the framework of homogeneous
turbulence have shown a persistence of anisotropy at the small scales. However, it should be noted that this anisotropy is detected only on statistical
moments of the velocity eld of order three or more, with rst- and secondorder moments being isotropic.
6.3.2 Anisotropic Models: Scalar Subgrid Viscosities
The subgrid viscosity models presented in this section have been designed to
alleviate the problem observed with basic subgrid viscosities, i.e. to prevent
the occurance of too high dissipation levels in shear ows, which are known
to have disastrous eects in near-wall regions.2 The models presented below
are:
2
Problems encountered in free shear ows are usually less important, since large
scales are often driven by inviscid instabilities, while the existence of a critical
Reynolds number may lead to relaminarization if the subgrid viscosity is too
high.
199
1. WALE model by Nicoud and Ducros (p. 199), which is built to recover
the expected asymptotic behavior in the near-wall region in equilibrium
turbulent boundary layers on ne grids, without any additional damping
function.
2. CasalinoJacob Weighted Gradient Model (p. 199), which is based on
a modication of the of the Smagorinsky constant to make it sensitive to
the mean shear stresses, rendering it more local in terms of wave number.
3. Models based on the idea of separating the eld into an isotropic part
and inhomogeneous part (p. 200), in order to be able to isolate the contribution of the mean eld in the computation of the subgrid viscosity, for
models based on the large scales, and thereby better localize the information contained in these models by frequency. This technique, however,
is applicable only to ows whose mean velocity prole is known or can
be computed on the y.
Wall-Adapting Local Eddy-Viscosity Model. It has been seen before
(p. 159) that most subgrid viscosity models do not exhibit the correct behavior in the vicinity of solid walls in equilibrium boundary layers on ne grids,
resulting in a too high damping of uctuations in that region and to a wrong
prediction of the skin friction. The common way to alleviate this problem
is to add a damping function, which requires the distance to the wall and
the skin friction as input parameters, leading to complex implementation issues. Another possibility is to use self-adpative models, which involve a larger
algorithmic complexity.
An elegant solution to solve the near-wall region problem on ne grids is
proposed by Nicoud and Ducros [567], who found a combination of resolved
velocity spatial derivatives that exhibits the expected asymptotic behavior
3
sgs z + , where z + is the distance to the wall expressed in wall units.
The subgrid viscosity is dened as
sgs
d d 3/2
Sij Sij
= (Cw )
5/2 d d 5/4
S ij S ij
+ Sij Sij
2
(6.53)
with Cw = 0.55 0.60 and

d
= S ik S kj + ik kj
Sij

1
S mn S mn mn mn ij
3
(6.54)
This model also possesses the interesting property that the subgrid viscosity vanishes when the ow is two-dimensional, in agreement with the physical
analysis.
Weighted Gradient Subgrid Viscosity Model. Casalino, Boudet and
Jacob [112] introduced a modication in the evaluation of the Smagorinsky
constant to render it sensitive to resolved gradients, with the purpose of
recovering a better accuracy in shear ows.
200
The weighted gradient subgrid viscosity model is written as

sgs = (C(x, t))2 |S ij | ,
(6.55)
where the self-adaptive constant is equal to

C(x, t) = CS

Sij
Sij
m/4
,
S ij S ij
(6.56)
where CS = 0.18 is the conventional Smagorinsky constant, and m are

free parameters and the weighted strain tensor S is dened as (without
summation over repeated greek indices)
S
= W S
(6.57)
The weighting matrix coecients are inversely proportional to the third

moment of corresponding strain rate tensor coecients:
;
<
3
S
;

W =
(6.58)
3 <1 .
=
S

,=1,3

The bracket operator is related to a local average. Numerical tests have
shown that (m, ) = (1, 3) and (3, 10) yield satisfactory results in free shear
ows, the last values yielding the recovery of the theoretical behavior of the
subgrid viscosity in the vicinity of solid walls on ne meshes.
Models Based on a Splitting Technique. Subgrid viscosity models are
mostly developed in the framework of the hypotheses of the canonical analysis, i.e. for homogeneous turbulent ows. Experience shows that the performance of these models declines when they are used in an inhomogeneous
framework, which corresponds to a non-uniform average ow. One simple
idea initially proposed by Schumann [653] is to split the velocity eld into
inhomogeneous and isotropic parts and to compute a specic subgrid term
for each of these parts.
In practice, Schumann proposes an anisotropic subgrid viscosity model
for dealing with ows whose average gradient is non-zero, and in particular
any ow regions close to solid walls. The model is obtained by splitting the
deviator part of the subgrid tensor d into one locally isotropic part and one
inhomogeneous:

a
ijd = 2sgs S ij S ij 2sgs
S ij ,
(6.59)
where the angle brackets . designate an statistical average, which in practice
is a spatial average in the directions of homogeneity in the solution. The coa
ecients sgs and sgs
are the scalar subgrid viscosities representing a locally
201
isotropic turbulence and an inhomogeneous turbulence, respectively. Moin

and Kim [537] and Horiuti [319] give the following denitions:

2 +

sgs = C1
2 S ij S ij S ij S ij ,
(6.60)
+

2
a
sgs
= C2 z
2S ij S ij ,
(6.61)
where C1 and C2 are two constants. Horiuti recommends C1 = 0.1 and C2
= 0.254, while Moin and Kim use C1 = C2 = 0.254. The isotropic part is
a function of the uctuation of the viscosity gradients, so as to make sure that
the extra-diagonal components thus predicted for the subgrid tensor cancel
out on the average over time. This is consistent with the isotropic hypothesis.
The two characteristic lengths and z represent the cuto lengths for
the two types of structures, and are evaluated as:
(z) = (1 2 3 )1/3 (1 exp(zu /A))
z (z) = 3 (1 exp([zu /A]2 ))
(6.62)
(6.63)
where z is the distance to the solid wall, 3 the cuto length in the direction normal to the surface, and u the friction velocity at the surface (see
Sect. 10.2.1). The constant A is taken to be equal to 25.
This model was initially designed for the case of a plane channel ow. It
requires being able to compute the statistical average of the velocity eld, and
thus can be extended only to sheared ows exhibiting at least one direction of
homogeneity, or requires the use of several statistically equivalent simulations
to perform the ensemble average [102, 108].
Sullivan et al. [698] propose a variant of it that incorporates an anisotropy
factor (so that the model constant can be varied to represent the eld
anisotropy better):
a
S ij
ijd = 2sgs S ij 2sgs
(6.64)
a
The authors propose computing the viscosity sgs
as before. The sgs term,
on the other hand, is now calculated by a model with one evolution equation
for the subgrid kinetic energy (see equation (5.119) in Chap. 5). Only the
subgrid kinetic energy production by the isotropic part is included, which is
equivalent to replacing the II term in equation (5.119) with

.
(6.65)
2sgs S ij S ij S ij S ij
The authors evaluate the anisotropy factor from the shearing rates of the
large and small scales. The average per plane of uctuation homogeneity of
the resolved strain rate tensor, calculated by
+

S = 2 S ij S ij S ij S ij ,
(6.66)
202
is used for evaluating the shear of the small scales. The shear of the large
scales is estimated as
+
S = 2S ij S ij .
(6.67)
The isotropy factor is evaluated as:
=
S
S
+ S
(6.68)
6.3.3 Anisotropic Models: Tensorial Subgrid Viscosities

Here we describe the main models proposed in the anisotropic framework.
Except for Aupoixs spectral model, none of these take explicit account of
the backward cascade mechanism. They are:
1. Aupoixs spectral model (p. 203), which is based on the anisotropic
EDQNM analysis. The interaction terms are evaluated by adopting a preset shape of the energy spectra and subgrid mode anisotropy. This model,
which requires a great deal of computation, has the advantage of including all the coupling mechanisms between large and small scales.
2. Horiutis model (p. 204), which is based on an evaluation of the anisotropy
tensor of the subgrid modes from the equivalent tensor constructed from
the highest frequencies in the resolved eld. This tensor is then used to
modulate the subgrid viscosity empirically in each direction of space. This
is equivalent to considering several characteristic velocity scales for representing the subgrid modes. This model can only modulate the subgrid
dissipation dierently for each velocity component and each direction of
space, but does not include the more complex anisotropic transfer mechanisms through the cuto.
3. The model of Carati and Cabot (p. 205), who propose a general form of
the subgrid viscosity in the form of a fourth-rank tensor. The components
of this tensor are determined on the basis of symmetry relations. However,
this model is a applicable only when the ow statistically exhibits an axial
symmetry, which restricts its eld of validity.
4. The model of Abba et al. (p. 207) which, as in the previous example,
considers the subgrid viscosity in the form of a fourth-rank tensor. The
model is based on the choice of a local adapted reference system for
representing the subgrid modes, and which is chosen empirically when
the ow possesses no obvious symmetries.
5. The model proposed by Carati (p. 207), which is based on the evaluation of the statistical anisotropy tensor. The requirement (mean velocity
prole) is the same as in the Schumann splitted model.
203
Aupoix Spectral Model. In order to take the anisotropy of the subgrid

scales into account, Aupoix [24] proposes adopting preset shapes of the energy
spectra and anisotropy so that the relations stemming from the previously
described EDQNM analysis of anisotropy can be used. Aupoix proposes the
following model for the energy spectrum:
E(k) = K0 2/3 k 5/3 exp {f (k/kd )}
(6.69)
where
"

%
&#

f (x) = exp 3.5x2 1 exp 6x + 1.2 196x2 33.6x + 1.4532
(6.70)
This spectrum is illustrated in Fig. 6.2. The anisotropy spectrum is modeled by:

k E(k)
Hij (k) = bij 5 +
E(k) k
>
?)
(

2/3

k
k
1 H (|F(u)|)
1
,
1+H
kmax
kmax
(6.71)
where F (u) = u, kmax is the wave number corresponding to the energy
spectrum maximum, and H the Heaviside function dened by:

0
if x 0
H(x) =
,
1 otherwise
Fig. 6.2. Aupoix spectrum (kd = 1000).
204
and where bij is the anisotropy tensor dened as:

bij =
ui uj
1
ij
2
qsgs
3
(6.72)
Horiutis Model. Horiuti [322] proposes extending the Smagorinsky model

to the isotropic case by choosing a dierent velocity scale for characterizing
each component of the subgrid tensor.
Starting with an ordinary dimensional analysis, the subgrid viscosity sgs
2
is expressed as a function of the subgrid kinetic energy qsgs
and the dissipation
rate :
2 2
(qsgs
)
sgs = C1
.
(6.73)
To make a better adjustment of the dissipation induced by the subgrid

model to the local state of the ow, Horiuti proposes replacing equation (6.73)
by:
2 2
(qsgs
)
sgs = C1
,
(6.74)
in which is a dimensionless parameter whose function is to regulate the

dissipation rate as a function of the anisotropy of the resolved eld. The
proposed form for is:
3E s
= 2
,
(6.75)
2qsgs
where E s is the square of a characteristic velocity scale of the subgrid modes.
For example, near solid walls, Horiuti proposes using the uctuation of the
velocity component normal to the wall, which makes it possible for the model
to cancel out automatically. To generalize this approach, we associate a chars
with each subgrid stress ij .
acteristic velocity Eij
In practice, the author proposes evaluating these characteristic velocities
by the scale similarity hypothesis by means of a test lter indicated by a tilde:
s
i )(uj u
j ) ,
= (ui u
Eij
(6.76)
which makes it possible to dene a tensorial parameter ij as:

i )(uj u
j )
3(ui u
ij = =
2
l=1,3 (ul ul )
(6.77)
This tensorial parameter characterizes the anisotropy of the test eld

and can be considered as an approximation of the anisotropy tensor
(u u)
205
associated with this velocity eld (to within a coecient of 1/3 ij ). Using a model based on the large scales, Horiuti derives the tensorial subgrid
viscosity eij :

2
eij = C1 |F(u)| ij ,
(6.78)
with
F (u) = u,
or

u + T u
where the constant C1 is evaluated as it is for the scalar models. He proposes

a model of the general form for the subgrid tensor :

ij = ij
where
K=
2
2
K+ P
3
3

eil
1
l )2 ,
(ul u
2
uj
ui
ejl
xl
xl
P = elm
l=1,3
um
xl
(6.79)
It is important to note that this is a model for the entire subgrid tensor
and not for its deviatoric part alone, as is the case for the isotropic models.
Carati and Cabot Model. Carati and Cabot [100] propose a tensorial
anisotropic extension of the subgrid viscosity models. Generally, the deviator d of the subgrid tensor is modeled as:
(1)
(2)
ijd = ijkl S kl + ijkl kl
(6.80)
where the tensors S and are dened as:

S=

1
u + T u ,
2

1
u T u
2
The two viscosities (1) and (2) are fourth-rank tensors theoretically dened by 81 independent coecients. However, the properties of the tensors d ,
S and make it possible to reduce the number of these parameters.
The tensors d and S are symmetrical and have zero trace, which entails:
ijkl
(1)
= jikl
(1)
(1)
ijkl
(1)
jilk
iikl
(1)
= 0
(1)
ijkk
= 0
206
The tensor (1) therefore contains 25 independent coecients. By a similar

analysis, we can say:
(2)
= jikl
ijkl
(2)
= jilk
(2)
iikl
= 0
ijkl
(2)
(2)
.
(6.81)
The tensor (2) therefore contains 15 independent coecients, which raises

the number of coecients to be determined to 40.
Further reductions can be made using the symmetry properties of the
ow. For the case of symmetry about the axis dened by the vector n =
(n1 , n2 , n3 ), the authors show that the model takes a reduced form that now
uses only four coecients, C1 , ..., C4 :
ijd

2
2C1 S ij 2C2 ni sj + si nj sk nk ij
3

1
C3 ni nj n2 ij sk nk 2C4 (r i nj + ni r j )
3
(6.82)
where si = S ik nk and r i = ik nk .
Adopting the additional hypothesis that the tensors (1) and (2) verify
the Onsager symmetry relations for the covariant vector n and the contravariant vector p:
(1)
klij (n) ,
(2)
klij (n) ,
ijkl (p) =
(1)
klij (p)
(2)
ijkl (p)
(2)
klij (p)
ijkl (n) =
ijkl (n) =
(1)
(2)
(1)
(6.83)
we get the following reduced form:
ijd = 21 S ij 22 n2 S ij
(6.84)
where 1 and 2 are two scalar viscosities and
S ij =
1
1
(ni sj + si nj ) 2 sk nk ij ,
n2
3n
S ij = S ij S ij
Carati then proposes determining the two parameters 1 and 2 by an

ordinary dynamic procedure.
207
Model of Abba et al. Another tensor formulation was proposed by Abba

et al. [1]. These authors propose dening the subgrid viscosity in the form of
the fourth-rank tensor denoted ijkl . This tensor is dened as the product of
a scalar isotropic subgrid viscosity iso and an fourth-rank tensor denoted C,
whose components are dimensionless constants which will play the role of
the scalar constants ordinarily used. The tensor subgrid viscosity ijkl thus
dened is expressed:

ijkl = Cijkl iso =
C ai aj ak al iso ,
(6.85)
,
where ai designates the ith component of the unit vector a ( = 1, 2, 3),

C is a symmetrical 3 3 matrix that replaces the scalar Smagorinsky constant. The three vectors a are arbitrary and have to be dened as a function
of some foreknowledge of the ow topology and its symmetries. When this
information is not known, the authors propose using the local framework
dened by the following three vectors:
a1 =
u
,
u
a3 =
(|u|2 ) u
,
|(|u|2 ) u|
a2 = a3 a1
(6.86)
The authors apply this modication to the Smagorinsky model. The scalar
viscosity is thus evaluated by the formula:
2
iso = |S| .
(6.87)
The subgrid tensor deviator is then modeled as:

2
2
2
ijd = 2
Cijkl |S|S kl + ij Cmmkl |S|S kl
3
(6.88)
k,l
The model constants are then evaluated by means of a dynamic procedure.

Models Based on a Splitting Technique. An anisotropic tensorial subgrid viscosity model is proposed by Carati et al. [103]. The resulting form of
the subgrid stress tensor is
ij = scalar ik jl S kl
(6.89)
where scalar plays the role of a scalar subgrid viscosity to be computed using
an arbitrary functional model and
ij = 3
ui uj
uk uk
(6.90)
where u (x, t) = u(x, t) u(x, t) is the local instantaneous uctuation
of the resolved eld around its statistical mean value. In practice, the mean
ow values can be obtained performing statistical averages over homogeneous
directions of the ow, over several realizations computed in parallel, or by
performing a steady RANS computation.
208
6.4 Remarks on Flows Submitted to Strong Rotation

Eects
All developements presented above deal with shear ows. Rotation is known
to also lead to isotropy breakdown and to very complex changes in the interscale dynamics [97]:
The turbulent kinetic energy dissipation rate is observed to diminish, due to
a scrambling in the non-linear triadic interactions. These phenomena have
been extensively described using the weak wave turbulence framework and
advanced EDQNM closures.
The Kolmogorov spectrum is no longer valid when strong rotational eects
are present.
Anisotropy is seen to rise, even starting from initially isotropic state. This
trend is observed on all statistical moments, including Reynolds stresses, integral length scales, ... Spectral analysis reveals that the induced anisotropy
escapes the classical description in the physical space, and that special representations in the Fourier space must be used to describe it acuurately.
Very complex coupling of rotation with strain is observed, leading to a very
large class of dynamical regimes.
The question thus arises of the validity of the subgrid models in the
presence of dominant rotational eects, since they do not account for this
change in the interscale transfers. In fact, none of the subgrid model presented above is able to account for rotation eects. But numerical experiments [183, 702, 557, 577, 152, 398, 422, 576, 596, 777, 215] show that good
results in ows driven by strong rotation are obtained using models that are
local in terms of wave numbers on ne grids: dynamic models (and other selfadaptive models), approximate deconvolution models, ... The explanation for
this success is that rotation eects are already taken into account by resolved
scales, and that local subgrid models will be built on scales that are already
modied by the rotation.
7. Structural Modeling
7.1 Introduction and Motivations

This chapter describes some of the family of structural models. As has already
been said, these are established with no prior knowledge of the nature of the
interactions between the subgrid scales and those that make up the resolved
eld.
These models can be grouped into several categories:
Those derived by formal series expansions (Sect. 7.2). These models make
no use of any foreknowledge of the physics of the ows, and are based
only on series expansions of the various terms that appear in the ltered
NavierStokes equations. This group of model encompasses models based
on deconvolution procedures, nonlinear models and those based on the
homogenization technique.
Those that use the physical hypothesis of scale similarity (Sect. 7.3). These
models are based on the scale similarity hypothesis, which establishes a correspondence between the statistical structure of the ow at dierent ltering levels. Despite the fact that these models are formally equivalent to
deconvolution-type models, I chose to present them in a separate section,
because they were originally derived on the grounds of physical assumptions rather than on mathematical considerations. The link between the
two classes of model is explicitly discussed in Sect. 7.3.3.
The mixed models, which are based on linear combinations of the functional and structural types, are presented in Sect. 7.4. These models have
historically been developed within the framework of the scale-similarity hypothesis, but recent developments dealing with the deconvolution approach
show that they are a natural part of deconvolution-based subgrid models.
Here again, I chose to t these to the classical presentation, in order to
allow the reader to establish the link with published references more easily.
The theoretical equivalence with full deconvolution models is discussed at
the end of this section.
Those based on transport equations for the subgrid tensor components
(Sect. 7.5). These models, though they require no information concerning the way the subgrid modes act on the resolved scales, require a very
210
complex level of modeling since all the unknown terms in the transport
equations for the subgrid tensor components have to be evaluated.
Those constructed from deterministic models for the subgrid structures
(Sect. 7.6). They assume that preferential directions of alignments are
known for the subgrid structures.
Those based on an explicit reconstruction of the subgrid velocity uctuations on an auxiliary grid (Sect. 7.7). These models are the only ones which
aim at reconstructing the subgrid motion directly. They can be interpreted
as solutions for the full deconvolution problem, as dened in Sect. 7.2.
The main dierence with deconvolution-like models is that they require
the denition of a ner auxiliary grid, on which the solution of the hard
deconvolution problem is explicitly reconstructed.
Those based on a direct identication of subgrid terms using advanced
mathematical tools, such as linear stochastic estimation or neural networks
(Sect. 7.8).
Those based on specic numerical algorithms, whose errors are designed
to mimic the subgrid forces (Sect. 7.9).
7.2 Formal Series Expansions

The structural models presented in this section belong to one of the three
following families:
1. Models based on approximate deconvolution (Sect. 7.2.1). They rely on
an attempt to recover, at least partially, the original unltered velocity
eld by inverting the ltering operator. The full deconvolution being
impossible to compute in practice, only approximate deconvolution is
used.
2. Nonlinear models (Sect. 7.2.2), which rely on the formal derivation of
a surrogate of the subgrid stress tensor as a function of the gradient of
the resolved velocity eld.
3. Models based on the homogenization technique (Sect. 7.2.3).
7.2.1 Models Based on Approximate Deconvolution
General Statement of the Deconvolution Problem. The deconvolution approach, also sometimes referred to as the deltering approach, aims
at reconstructing the unltered eld from the ltered one. The subgrid modes
are no longer modeled, but reconstructed using an ad hoc mathematical procedure [181].
We recall that, writing the NavierStokes equations symbolically as
u
+ N S(u) = 0
t
(7.1)
211
we get the following for the ltered eld evolution equation (see Chap. 3)
u
+ N S(u) = [N S, G
](u)
t
(7.2)
where G is the lter kernel, and [, ] is the commutator operator. The exact
subgrid term, which corresponds to the right-hand side of relation (7.2), appears as a function of the exact nonltered eld u. This eld being unknown
during the computation, the idea here is to approximate it using a deconvolution procedure:
1
u u G1
l
u = Gl
G
u ,
(7.3)
is an lth-order approximate inverse of the lter G

where G1
l
l
G1
l
G = Id + O( ) .
The subgrid term is then approximated as
[N S, G
](u) [N S, G
](u ) = [N S, G
](G1
l
u) ,
(7.4)
achieving the description of the procedure. Combining the right-hand side

and the left-hand side of the resulting equation, we get:
u
+ G
N S(G1
l
u) = 0
t
(7.5)
The nonlinear term appears as (G

) (N S) (G1
l ), i.e. as the sequential
application of: (i) the approximate deconvolution operator, (ii) the Navier
Stokes operator, and (iii) a regularization operator, referred to as primary
regularization in the parlance of Adams [691, 4, 692, 693, 694]. It is worth
noting that the eciency of the present strategy will be conditioned by our
capability of nding the approximate inverse operator, the two others being
a priori known.
The deconvolution procedure, in the general presentation given above,
calls for several important remarks:
1. It is ecient for invertible lters only, i.e. non-projective lters. Projective lters induce an irreversible loss of information, which cannot be
recovered (see Sect. 2.1.2). For smooth lters, the unltered eld can
theoretically be reconstructed [106, 807].
2. In practice, the grid lter is always present, because of the nite number
of degrees of freedom used to compute the solution. As a consequence of
the Nyquist theorem, a projective lter with space and time cutos equal
to 2x and 2t, respectively, is always present. Here, x is the mesh
size of the computational grid used for the large-eddy simulation, and t
the time step employed for the numerical time integration. As a consequence, the lter to be considered in a practical simulation, referred to as
212
the eective lter, is a combination of the convolution lter with cuto

length scale and the projective grid lter. The latter can be modeled
as a sharp cuto lter with cuto wave number kc = /x. A complete
discussion about the eective lter is given in Sect. 8.2. This implies
that the deconvolution procedure cannot reconstruct structures smaller
than 2x, and, following the terminology of Adams, two problems can
be identied in the deconvolution approach:
a) The soft deconvolution problem, which corresponds to the reconstruction of the unltered eld for wave numbers k [0, /x]. The corresponding reconstruction procedures described below are:
Procedures relying on an iterative reconstruction of the inverse of
the ltering operator (p. 212);
Procedures based on a truncated Taylor series expansion of the
ltering operator (p. 213).
b) The hard deconvolution problem: solving the soft deconvolution problem does not suce for closing the reconstruction problem, because
interactions with scales smaller than 2x are not taken into account.
In order to alleviate this problem, the soft deconvolution procedure
must be supplemented with a secondary procedure, sometimes referred to as secondary regularization. These scales being denitively
lost, they must be modeled (and not reconstructed) using a functional model. All the models presented in Chap. 5 can be used. The
specic penalization procedure developed by Stolz and Adams and
other techniques are discussed in a dedicated section (p. 218).
These two processes are illustrated in Fig. 7.1.
Solving the Soft Deconvolution Problem: Iterative Deconvolution.
Adams et al. [691, 692, 4, 694, 693, 743, 744, 742] developed an iterative deconvolution procedure based on the Van Cittert method. If the lter kernel G
has an inverse G1 , the latter can be obtained using the following expansion:
G1
= (Id (Id G))1 ,

=
(Id G)p ,
(7.6)
(7.7)
p=0,
yielding the following reconstruction for the deltered variable :

= + ( ) + ( 2 + ) + ... ,
(7.8)
= ( ) + ( 2 + ) + ...
(7.9)
or equivalently
The series are known to be convergent if Id G < 1. A practical
model is obtained by truncating the expansion at a given power. Stolz and
Adams [691] recommend using a fth-order (p = 5) expansion.
213
Fig. 7.1. Schematic of the full deconvolution problem. Top: soft deconvolution
problem; Bottom: ideal solution of the full (soft+hard) deconvolution problem.
Other possible reconstruction techniques, such as the Tikhonov regularization, the singular value decomposition or the conjugate gradient method,
are discussed in reference [4].
Solving the Soft Deconvolution Problem: Truncated Taylor SeriesExpansion-Based Deconvolution Procedures. The deconvolution procedures presented in this section are all based on the representation of the
ltering operator in the form of a Taylor series expansion (see Sect. 2.1.6).
We recall the general form of such an expansion for the dummy variable :
(x) =

(1)k
k=0
k!
Mk (x)
k
(x)
xk
where Mk is the kth-order moment of the lter kernel G.
(7.10)
214
The common idea shared by all these models is to approximate the ltering operator by truncating the Taylor series expansion, dening a low-order
dierential operator:
(x)
N

(1)k
k=0
k!
Mk (x)
k
(x)
xk
(7.11)
with N 4 in practice. The accuracy of the reconstruction obviously depends on the convergence rate of the Taylor series expansion for the ltering
operator. Pruett et al. [607] proved that it is fastly converging for some typical lters, such as Gaussian and top-hat lters. The unltered eld can
formally be expressed as the solution of the following inverse problem:

(x)
N

(1)k
k=0
k
Mk (x) k
k!
x
k
1
(x)
(7.12)
The last step consists of inverting this operator, being considered as

a small parameter. Two possibilities arise at this stage of the soft deconvolution procedure.
The rst one consists of using an implicit method to solve (7.12). This
procedure is rarely used, because of its algorithmic cost.
The second one, considered by a large number of authors, relies on the use
of an explicit approximation of the solution of (7.12), which is obtained using
once again a Taylor series expansion. Recalling that for a small parameter
we have
(7.13)
(1 + )1 = 1 + 2 3 + 4 ... ,
and assuming that can play the role of a small parameter, we get the
following approximate inverse relation, which is valid for symmetric lters:

2
1 2
(7.14)
(x) Id M2 (x) 2 + ... (x) .
2
x
This last form can be computed immediately from the resolved eld. Limiting the expansion to the second order, the subgrid part is expressed as:
(x)
2 (x)
1 2
4
M2
+ O( )
2
2
x

1 2
2
2
=
M2 2 + O( )
2
x

2
= (x) + O( ) .
=
(7.15)
This can be used to express all the contributions as a function of the

resolved eld, with second-order accuracy. The various terms of the Leonard
215
decomposition are approximated to the second order as:

2
1 2
4
M2 2 (ui uj ) + O( ) ,
(7.16)
2
x

2
2
1 2
4

Cij ui uj + uj ui = M2 ui 2 uj + uj 2 ui + O( ) . (7.17)
2
x
x
Lij ui uj ui uj =
The combination of these two terms leads to:

2
Lij + Cij = M2
ui uj
4
+ O( ) .
x x
(7.18)
As for the subgrid Reynolds tensor, it appears only as a fourth-order term:

Rij ui uj =
1 2 2 2 ui 2 uj
6
M2
+ O( ) ,
4
x2 x2
(7.19)
so that it disappears in a second-order expansion of the full subgrid tensor. The resulting model (7.18) is referred to as the gradient model, tensordiusivity model or Clarks model. It can also be rewritten using undivided dierences, leading to the denition of the increment model proposed by Brun and Friedrich [82]. In practice, this approach is used only
to derive models for the tensors L and C, which escape functional modeling [98, 143, 160, 161, 459]. Certain authors also use these evaluations to
neglect these tensors when the numerical scheme produces errors of the same
order, which is the case for second-order accurate schemes.
Finer analysis allows a better evaluation of the order of magnitude of the
subgrid tensor. By using a subgrid viscosity model, i.e.
ij = 2sgs S ij
(7.20)
and using the local equilibrium hypothesis:

= ij S ij = sgs |S|2
(7.21)
the amplitude of the subgrid tensor can be evaluated as:

|ij | sgs |S|
sgs
(7.22)
By basing the computation of the subgrid viscosity on the subgrid kinetic

energy:
+
sgs
2
qsgs
(7.23)
and computing this energy from a Kolmogorov spectrum:

+
2
qsgs

1/2
E(k)dk
kc
216

5/3
1/2
dk
kc
1/3
(kc )
1/3
we get for the subgrid viscosity:

1/3
sgs
4/3
(7.24)
The order of magnitude of the corresponding subgrid tensor is:

|ij |
2/3
sgs
(7.25)
This estimation is clearly dierent from those given previously, and shows
that the subgrid tensor is theoretically dominant compared with the terms
2
in . This last evaluation is usually interpreted as being that of the subgrid
Reynolds tensor Rij , while the estimations of the tensors Cij and Lij given
above are generally considered to be correct.
A generalized expansion for the whole subgrid tensor is proposed by Carati
et al. [105]. These authors have proved that the dierential expansion
=
Clm (G)
l,m=0
l m
xl xm
(7.26)
where (x) and (x) are two C real functions, and Clm (G) are some real
coecients which depend explicitly on the lter, is valid for all the lter
kernels G such that:
G(i( + ))
IR,
G(i)G(i)
i2 = 1 .
(7.27)
This is particularly true of all symmetric kernels. The resulting general

form of the gradient model deduced from (7.26) is:
ij = ui uj ui uj =

l,m=0,;(l,m)=(0,0)
Clm (G)
l ui m uj
xl xm
(7.28)
The use of the Gaussian lter yields the following simplied form:
2 m

m ui m uj
ij =
.
(7.29)
16
xm xm
m=0,
By retaining only the rst term in (7.29), one recovers the gradient
model (7.18).
217
Expression (7.28) can be recast in a general tensor-diusivity form, as

demonstrated by Adams and Leonard [3]. These authors have shown that,
for discontinuous, but otherwise smooth, functions and the series (7.26)
can be summed up, leading to
= +
(/4)2

R( )R( )
2
x x
(7.30)
where a non-dimensional eddy-diusivity R() is introduced. It is dened by

the following relation:
1
R () = 2
2
2

G1 ()
2
G( )d
(7.31)
A consistency constraint is R() 1 for 0. This constraint is satised for the top-hat lter, but is violated for the Gaussian lter. Numerical
experiments show that this form is ill-conditionned and leads to unstable results, and must be supplemented by a secondary regularization. To this end,
most of the authors explicitly add another dissipative term to the momentum
equation (see p. 218).
Layton et al. [341, 206, 234] proposed regularizing the gradient model by
applying a smoothing operator. By choosing a convolution lter with a kernel
G2 , the resulting model, referred to as the rational approximation model, is
written as

2
ui uj
.
(7.32)
ij = G2
16 x x
Iliescu et al. [341] discuss both explicit and implicit implementation
of (7.32), which was applied to channel ows [219, 340, 339].
The need for a secondary regularization to get a stable computation using the gradient-like models may be understood by analysing their dissipative
properties. The link with functional models of subgrid viscosity type is established by looking at the corresponding subgrid force term appearing in
the momentum equations [761]. In three dimensions, we have:
2
M2
2
xj
ui uj
xk xk
= M2 S jk

ui
xj xk
(7.33)
showing that M2 S jk plays the role of a tensorial subgrid viscosity. Since

the tensor S jk is not positive-denite, antidissipation occurs along stretching
directions, which are associated with negative eigenvalues.
Full three-dimensional expressions up to the fourth-order terms of the
subgrid tensor have been derived by several authors [81, 607], but these are
very cumbersome and will not be presented here.
218
Taylor series expansion has also been used by several authors [81, 717] to
derive an equivalent dierential expression for the test lter and some tensorial quantities appearing in Germanos dynamic procedure for evaluation of
the constants (see p. 137).
A worthy remark can be made dealing with the evaluation of the constants appearing in the models derived from truncated Taylor series expansions, which are considered as truncated polynomials in . These parameters
are commonly taken equal to the one appearing in the full (untruncated) expansions, and do not correspond to the coecients of the best truncated
polynomial approximation of the full expansions. Considering the best approximation, dierent coecients are usually found, which may lead to different properties with respect to the preservation of symmetries of the ltered
NavierStokes equations discussed in Sect. 3.3.4.
Connection between Taylor Expansion-Based Deconvolution and
Iterative Deconvolution. Stolz, Adams and Kleiser [693] have demonstrated the practical equivalence between the gradient-type models (or tensor
diusivity models) and the use of the Van Cittert iterative technique for the
soft deconvolution problem.
In the particular case of the Gaussian lter, an exact fourth-order deltered variable is
2
2
4
6
=+
+
+ O( ) .
(7.34)
24 x2
1152 x4
The use of this relation yields the following expression for the approximated subgrid tensor:
G1
l
ui uj ui uj
1
(G1
l
ui )(Gl
uj ) ui uj
2
ui uj
2 ui 2 uj
6
+
=
+ O( ) . (7.35)
24 x x
288 x2 x2
It is observed that each iterative deconvolution procedure yields the denition of a particular tensor-diusivity model. Practical dierences may arise
from the discretization of the continuous derivate operators.
Solving the Hard Deconvolution Problem: Secondary Regularization. The secondary regularization is needed to obtain stable numerical simulation. This can be seen by two dierent ways:
1. The soft deconvolution procedure is restricted to the reconstruction of
scales which are resolvable on the considered computational grid. Interactions with unresolved scales need to be taken into account from
a theoretical point of view.
2. It has been demonstrated (at least in the particular case of the tensordiusivity model) that negative dissipation can occur, yielding possible
numerical troubles. The net drain of resolved kinetic energy by unresolved scales needs to be taken into account (see Chap. 5 for a detailed
discussion).
219
In practice, this secondary regularization is achieved by adding a dissipative term to the deltered equations. Starting from relation (7.5), we arrive
at the following formal evolution equation:
u
+ G
N S(G1
l
u) = S(u, ) ,
t
(7.36)
where the dissipative source term S(u, ) can a priori be considered as a function of u and .
The StolzAdams Penalty Term. In order to account for kinetic energy transfer with scales which are not recovered by the deconvolution procedure, Stolz
and Adams [691, 692, 694, 4] introduced a relaxation term, yielding
S(u, ) = (Id G1
l
G)
u ,
(7.37)
where is an empirical relaxation time. Since (Id G1
G) is positive
l
semidenite, this relaxation term is purely dissipative. The use of this term
can be interpreted as applying a second ltering operator (G1
l
) (G
) to
u every 1/t time steps, t being the time step selected to perform the
numerical time integration. It can also be interpreted as a penalization of the
ltered solution. An important point is that this relaxation regularization
is not equivalent to the use of subgrid viscosity type dissipation, since the
associated spectral distributions of the dissipation are very dierent.
The relaxation time can be empiricaly chosen or dynamically evaluated.
Numerical simulations show that the results may not be very sensitive to the
value of this parameter. Stolz et al. [692, 693] reported no important changes
for the channel ow case in the range 12.5u /h 100u /h, where u
is the skin friction and h the channel height. This relaxation time can also
be evaluated dynamically in order to enforce a constant kinetic energy of the
resolved sublter modes, i.e. of the modes which are ltered out of the exact
solution by the convolution lter but which are resolved on the computational
grid. These modes correspond to the spectral band [/, /x]. A measure
of this energy is
2
1
(Id G1
.
(7.38)
EHF =
l
G)
u
2
The equilibrium hypothesis can be expressed as
EHF
=0 .
t
(7.39)
By combining relations (7.36) and (7.37), we get

43
4
EHF 3
1
1
= (Id G1
l
G)
u G
N S(Gl
u) (Id Gl
G)
u ,
t
(7.40)
220
leading to
4 3
4
3
1
(Id G1
l
G)
u G
N S(Gl
u)
4 3
4
= 3
1
(Id G1
l
G)
u (Id Gl
G)
u
(7.41)
Other Possibilities. The secondary regularization is achieved by many authors using subgrid-viscosity models, which provide the desired drain of
resolved kinetic energy. All the subgrid-viscosity models can be used. The
most employed ones are the Smagorinsky model and the dynamic Smagorinsky model. But it is worth noting that the secondary regularization can
be achieved by means of the Implicit Large-Eddy Simulation approach
(Sect. 5.3.4). As an example, Pasquetti and Xu used the Spectral Vanishing Viscosity approach in Refs. [582, 581].
Full Deconvolution Model Examples. As seen at the beginning of this
section, the deconvolution approach necessitates the use of two models. An
a priori innite number of combinations between models for the soft and the
hard deconvolution problems can be dened. Some examples are listed in
Table 7.1. These two-part models for the full deconvolution problem can also
be recast within the framework of mixed modeling. This point is discussed in
Sect. 7.4.
Table 7.1. Examples of solutions to the Full Deconvolution Problem.
Ref.
Soft Deconvolution
Hard Deconvolution
[4, 694, 693, 692, 691]

[759]
[143, 759]
[763, 764, 12]
[341]
Van Cittert (7.8)

Van Cittert (7.8)
tensor-diusivity (7.18)
tensor-diusivity (7.18)
rational model (7.32)
relaxation (7.37)
Smagorinsky (5.90)
Smagorinsky (5.90)
dynamic Smagorinsky
Smagorinsky (5.90)
Single-Step Filtering Implementation of the Full Approximate Deconvolution Model. The practical implementation of the full approximate
deconvolution model has been shown by Mathew et al. [499] to simplify as
the application of a single ltering operator to the solution at the end of each
time step of the numerical computation. Observing that relation (7.5) can be
recast as

u
+ N S(u ) = 0 ,
(7.42)
G
t
under the assumption that u is sucient close to u, one obtains the following
relation:
u
u
= G
.
(7.43)
t
t
221
Therefore, the basic implementation requires three steps to obtain the

solution at time (n+ 1)t (noticed u(n+ 1) below) starting from the solution
at the previous time step, u(n)
1. Evaluation of the approximate unltered eld at time nt :
u (n) = G1
l
u(n)
2. Time advancement of the approximate unltered solution :

u (n + 1) = u (n) + t
u
+ O(t2 ) .
t
3. Restriction of the new approximate unltered solution :

u(n + 1) = G
u (n + 1) .
Looking at this sequence, Mathew observes that the rst and third steps
can be combined in a unique one, leading to the following two-step algorithm
1. Time advancement of the approximate unltered solution :
u = u (n) + t
u
+ O(t2 ) .
t
u (n + 1) = (G1
l
G)
u = Ql
u
It is seen that the sole lter involved in practice is the lter Ql , which is an
lth order perturbation of the identity whose eect is mostly concentrated on
high resolved wavenumbers. This two-step method accounts for the soft deconvolution model only. The hard deconvolution problem, or equivalently the
secondary regularization, is reintroduced within the same framework using
the fact that it is equivalent (at least in the case of the penalty term proposed
by Stolz and Adams) to the application of the lter (G1
l
)(G
) = Ql to the
solution every 1/t time steps. The full deconvolution problem can thus be
implemented in a simple two-step procedure:
1. Time advancement of the approximate unltered solution:
u = u (n) + t
u
+ O(t2 ) .
t

u (n + 1) = G
u
where the basic value G = Ql is changed into G = Ql

Ql = Q2l every
1/t time steps.
222
Toward Higher-Order Deconvolution Models. The approximate deconvolution approach presented above can be seen as the lowest-order member of a general class of deconvolution approaches. This fact is emphasized
by Mathew and his coworkers [499], who carried out the development at the
next order. Restricting for the sake of simplicity the discussion to the formal,
one-dimensional scalar conservation law :
u f (u)
+
=0 ,
t
x
(7.44)
the deconvolution approach can be written as

u f (u)
f (u)
f (u)
+
=
G
=R .
t
x
x
x
(7.45)
The low-order approximate deconvolution procedure presented above corresponds to the following closure
R R1 =
f (u )
f (u)
G
x
x
(7.46)
The higher-order method is derived writing the remainder in this closure

relation as

f (u ) f (u)
R = R1 + R2 , R2 = G
,
(7.47)
x
x
and nding a computable approximation for R2 . Such an expansion is found
introducing the Taylor series expansion

f
2
R2 = G
(u u) + O(u u)
.
(7.48)
x u u
An estimation of the leading term of this expansion is

f
1
(G
G Id)
u
R2 = G
x u u l
which can be approximated as

f
(G
Id)
u
R2 = G
x u u=u l
(7.49)
(7.50)
The resulting higher-order formulation of the approximate deconvolution

approach is

f (u )
u
f
1
+G
=G
(G
Id)
u
. (7.51)
t
x
x u u=u l
It is worth noticing that this rened modeling makes a term appearing in
the right hand side of equation (7.51) that is similar to the empirical penalty
term for secondary regularization (7.37) introduced by Stolz and Adams.
223
7.2.2 Non-linear Models

There are a number of ways of deriving nonlinear models: Horiuti [321],
Speziale [683], Yoshizawa [788], and Wong [765] start with an expansion in
a small parameter, while Lund and Novikov [462] use the mathematical properties of the tensors considered. It is this last approach that will be described
rst, because it is the one that best reveals the dierence with the functional
models. Kosovics simplied model [401] and Wongs dynamic model [765]
are then described.
Generic Model of Lund and Novikov. We assume that the deviator of
the subgrid tensor can be expressed as a function of the resolved velocity eld
gradients (and not the velocity eld itself, to ensure the Galilean invariance
property), the unit tensor, and the square of the cuto length :
1
2
ij kk ij ijd = F (S ij , ij , ij , )
3
(7.52)
The isotropic part of is not taken into account, and is integrated in the
pressure term because S and have zero traces. To simplify the expansions
in the following, we use the reduced notation:
S = S ik kj ,
tr(S ) = S ij jk ki
The most general form for relation (7.52) is a polynomial of innite dea1 a2 a3 a4
gree of tensors whose terms are of the form S S ..., where the ai are
positive integers. Each terms in the series is multiplied by a coecient, which
is itself a function of the invariants of S and . This series can be reduced
to a nite number of linearly independent terms by the Cayley- Hamilton
theorem. Since the tensor d is symmetrical, we retain only the symmetrical terms here. The computations lead to the denition of eleven tensors,
m1 , ..., m11 , with which I1 , ..., I6 are associated:
m1
m3
m5
m7
m9
m11
= S,
m2
2
= ,
m4
2
2
= S S , m6
2
= S + S,
m8
2
2
= S S S S,
m10
2 2
2 2
= S S ,
I1
I3
I5
=
=
=
tr(S ),
3
tr(S ),
2 2
tr(S ),
I2
I4
I6
where Id designates the identity tensor.
=
=
=
=
=
S ,
S S,
Id,
2
S S ,
2 2
2 2
S + S ,
(7.53)
= tr( ),
2
= tr(S ),
2 2
= tr(S S ),
(7.54)
224
These tensors are independent in the sense that none can be decomposed
into a linear sum of the ten others, if the coecients are constrained to
appear as polynomials of the six invariants dened above. If we relax this last
constraint by considering the polynomial quotients of the invariants too, then
only six of the eleven tensors are linearly independent. The tensors dened
above are no longer linearly independent in two cases: when the tensor S
has a double eigenvalue and when two components of the vorticity disappear
when expressed in the specic reference of S. The rst case corresponds to
an axisymmetrical shear and the second to a situation where the rotation is
about a single axis aligned with one of the eigenvectors of S. Assuming that
neither of these conditions is veried, six of the terms of (7.53) are sucient
for representing the tensor , and ve for representing its deviator part, which
is consistent with the fact that a second-order symmetrical tensor with zero
trace has only ve degrees of freedom in the third dimension. We then obtain
the generic polynomial form:
d
= C1 |S|S + C2 (S )d + C3 ( )d
2
+ C4 (S S) + C5
1
2
2
(S S ) ,
|S|
(7.55)
where the Ci , i = 1, 5 are constants to be determined. This type of model is

analogous in form to the non-linear statistical turbulence models [682, 683].
Numerical experiments performed by the authors on cases of isotropic homogeneous turbulence have shown that this modeling, while yielding good
results, is very costly. Also, computing the dierent constants raises problems because their dependence as a function of the tensor invariants involved
is complex. Meneveau et al. [509] attempted to compute these components
by statistical techniques, but achieved no signicant improvement over the
linear model in the prediction of the subgrid tensor eigenvectors. A priori
tests carried out by Horiuti [325] have shown that the (S S) term is
responsible for a signicant improvement of the correlation coecient with
the true subgrid tensor.
We note that the rst term of the expansion corresponds to subgrid viscosity models for the forward energy cascade based on large scales, which makes
it possible to interpret this type of expansion as a sequence of departures
from symmetry: the isotropic part of the tensor is represented by a spherical
tensor, and the rst term represents a rst departure from symmetry but
prevents the inclusion of the inequality of the normal subgrid stresses1 . The
anisotropy of the normal stresses is included by the following terms, which
therefore represent a new departure from symmetry.
1
This is true for all modeling of the form = (V V ) in which V is an arbitrary

vector. It is trivially veried that the tensor (V V ) admits only a single nonzero eigenvalue = (V12 + V22 + V32 ), while the subgrid tensor in the most general
case has three distinct eigenvalues.
225
Kosovics Simplied Non-Linear Model. In order to reduce the algorithmic cost of the subgrid model, Kosovic [401] proposes neglecting certain
terms in the generic model presented above. After neglecting the high-order
terms on the basis of an analysis of their orders of magnitude, the author
proposes the following model:

1
2
2 1/2
ij = (Cs ) 2(2|S| ) S ij + C1 S ik S kj S mn S mn ij
3

4
,
(7.56)
+ C2 S ik kj ik S kj
where Cs is the constant of the subgrid viscosity model based on the large
scales (see Sect. 5.3.2) and C1 and C2 two constants to be determined. After
computation, the local equilibrium hypothesis is expressed:
=
=
ij S ij
#
"
(Cs )2 2 (2|S|2 )1/2 S ij S ij + C1 S ik S kj S ji .
(7.57)
In the framework of the canonical case (isotropic turbulence, innite inertial range, sharp cuto lter), we get (see [45]):
S ij S ij =
=

2
30 u1

4
x1
3
K0 2/3 kc4/3
4
(7.58)

3
105 u1
S ik S kj S ji =

8
x1
3/2

105
1
=
kc2
S(kc )
K0
8
10
(7.59)
where coecient S(kc ) is dened as:

3
2
u1
u1
/
3/2
S(kc ) =
x1
x1
(7.60)
Substituting these expressions in relation (7.57) yields:

= (Cs )

3/2
3
7
K0
C1 S(kc )
kc2
1
2
960
(7.61)
This relation provides a way of relating the constants Cs and C1 and

thereby computing C1 once Cs is determined by reasoning similar to that explained in the chapter on functional models. The asymptotic value of S(kc )
is evaluated by theory and experimental observation at between 0.4 and 0.8,
226
as kc . The constant C2 cannot be determined this way, since the contribution of the anti-symmetrical of the velocity gradient to the energy transfer
is null2 .
On the basis of simple examples of anisotropic homogeneous turbulence,
Kosovic proposes:
(7.62)
C2 C1 ,
which completes the description of the model.
Dynamic Non-Linear Model. Kosovics approach uses some hypotheses
intrinsic to the subgrid modes, for example the existence of a theoretical
the spectrum shape and the local equilibrium hypothesis. To relax these
constraints, Wong [765] proposes computing the constants of the non-linear
models by means of a dynamic procedure.
To do this, the author proposes a model of the form (we use the same
notation here as in the description of the dynamic model with one equation
for the kinetic energy, in Sect. 5.4.2):
ij =
+
2 2
2 S C N
qsgs ij 2C1 qsgs
ij
2 ij
3
(7.63)
2
where C1 and C2 are constants and qsgs
the subgrid kinetic energy, and
1
1
N ij = S ik S kj S mn S mn ij + S ij S mm ij
3
3
(7.64)
where S ij is the Oldroyd3 derivative of S ij :

ui
DS ij
uj
S ij =
S kj
S ki
Dt
xk
xk
(7.65)
where D/Dt is the material derivative associated with the velocity eld u.
The isotropic part of this model is based on the kinetic energy of the subgrid
modes (see Sect. 5.3.2). Usually, we introduce a test lter symbolized by
* Using the same model, the
a tilde, the cuto length of which is denoted .
subgrid tensor corresponding to the test lter is expressed:
+
2
* Q2 S
*
*
,
Tij = Q2sgs ij 2C1
sgs ij C2 H ij
3
This is because we have the relation

ij S ij 0
(7.66)
since the tensors and S are anti-symmetrical and symmetrical, respectively.

This derivative responds to the principle of objectivity, i.e. it is invariant if the
frame of reference in which the motion is observed is changed.
227
where Q2sgs is the subgrid kinetic energy corresponding to the test lter,
* the tensor analogous to N , constructed from the velocity eld u.
*
and H
ij
ij
Using the two expressions (7.63) and (7.66), the Germano identity (5.138) is
expressed:
Lij
=

Tij *ij
2 2
2
2
(Q q/
sgs )ij + 2C1 Aij + C2 Bij
3 sgs
in which
Aij = S ij
+
+
*
*
2
qsgs
S
Q2sgs
ij
*
Bij = N
ij
2
*
H
ij
(7.67)
(7.68)
(7.69)
We then dene the residual Eij :

2
2
2
Eij = Lij (Q2sgs q/
sgs )ij + 2C1 Aij + C2 Bij
3
(7.70)
The two constants C1 and C2 are then computed in such a way as to

minimize the scalar residual Eij Eij , i.e.
Eij Eij
Eij Eij
=
=0
C1
C2
(7.71)
A simultaneous evaluation of these two parameters leads to:

2C1
C2
Lmn (Amn Bpq Bpq Bmn Apq Bpq )

Akl Akl Bij Bij (Aij Bij )2
(7.72)
Lmn (Bmn Apq Apq Amn Apq Bpq )

Akl Akl Bij Bij (Aij Bij )2
(7.73)
2
and Q2sgs are obtained by solving the corresponding
The quantities qsgs
evolution equations, which are described in the chapter on functional models.
This completes computation of the subgrid model.
One variant that does not require the use of additional evolution equations
is derived using a model based on the gradient of the resolved scales instead
of one based on the subgrid kinetic energy, to describe the isotropic term.
The subgrid tensor deviator is now modeled as:
1
2
ij kk ij = 2C1 |S|S ij C2 N ij
3
(7.74)
228
2
The two parameters computed by the dynamic procedure are now C1

2
and C2 . The expressions obtained are identical in form to relations (7.72)
and (7.73), where the tensor Aij is dened as:
*S
*
Aij = |S|S ij |S|
ij
2
.
(7.75)
7.2.3 Homogenization-Technique-Based Models

General Description. The theory of homogenization is a two-scale expansion technique originally developed in structural mechanics to model inhomogeneous materials with a periodic microstructure. If the slow (i.e. large) scale
and the rapid (i.e. small) scale are very dierent (i.e. if the microstructure
is very ne compared with the large scale variations of the material), the
composite material can be represented by an homogeneous material whose
characteristics can be computed theoretically through the two-scale expansion and an averaging step. Using the scale separation assumption between
the resolved scales of motion and the subgrid scales, the homogenization was
introduced by Perrier and Pironneau [588] to obtain an theoretical evaluation
for the subgrid viscosity. This approach was resurrected twenty years later
by Persson, Fureby and Svanstedt [589] who derived a new homogenizationbased tensorial subgrid viscosity model and performed the rst simulations
with this class of models.
The homogenization approach, which consists in solving the evolution
equations of the ltered eld separately from those of the subgrid modes, is
based on the assumption that the cuto is located within the inertial range
at each point. The resolved eld u and the subgrid eld u are computed on
two dierent grids by a coupling algorithm. In all of the following, we adopt
the hypothesis that u = 0. The subgrid modes u are then represented by
a random process v , which depends on the dissipation , and the viscosity
, and which is transported by the resolved eld u. This modeling is denoted
symbolically:

x ut t
u = v ,
,
(7.76)
, 2
in which 1 is the largest wave number in the inertial range and 2 the
highest frequency considered. As the inertial range is assumed to extend to
the high wave numbers, is taken as small parameter. Let u be the solution
to the problem:
ui (ui + vi )(uj + vj )
2 vi
2 ui
p vi
+
+
t
xj
xk xk
xi
t
xk xk
. (7.77)
229
If v is close to u , then u is close to u. More precisely, we have:

u = u + u1 + 2 u2 + ...
(7.78)
The lowest-order terms in the ltered expansion yields a homogenized

problem for u which involves a subgrid stress tensor v u1 + u1 v . The
homogenization technique relies on the computation of u1 and the denition
of stochastic model for v .
A modeling of this kind, while satisfactory on the theoretical level, is not
so in practice because the function v oscillates very quickly in space and
time, and the number of degrees of freedom needed in the discrete system to
describe its variations remains very high. To reduce the size of the discrete
system signicantly, other hypotheses are needed, leading to the denition
of simplied models which are described in the following. The description
of the models is limited to the minimum amount of details for the sake of
simplicity. The reader is referred to Ref. [589] for a detailed presentation of
the two-scale expansion of the NavierStokes equations.
PerrierPironneau Models. The rst simplication introduced by Perrier
and Pironneau [588] consists in choosing the random process in the form:
v (x, t) =
1
v (x, t, x , t )
(7.79)
in the space and time scales x and t , respectively, of the subgrid modes are
dened as:
x ut
t
, t = 2 .
(7.80)
x =
The new variable v (x, t, x , t ) oscillates slowly and can thus be represented with fewer degrees of freedom. Assuming that v is periodical depending on the variables x and t on a domain v = Z]0, T [, and that the
average of v is null over this domain4 , it is demonstrated that the subgrid
tensor is expressed in the form:
= Bu
(7.81)
where the term Bu is computed by taking the average on the cell of periodicity v of the term (v u1 + u1 v), where u1 is the a solution on this
cell of the problem:
u1
2x u1 + v x u1 + u1 x v = q v u u v
t
x u1 = 0 ,
4
, (7.82)
(7.83)
This is equivalent to considering that v (x, t, x , t ) is statistically homogeneous

and isotropic, which is theoretically justiable by the physical hypothesis of local
isotropy.
230
where x designates the gradient with respect to the x variables and q the
Lagrange multiplier that enforces the constraint (7.83). This model, though
simpler, is still dicult to use because the variable (x ut) is dicult to
manipulate. So other simplications are needed.
To arrive at a usable model, the authors propose neglecting the transport
of the random variable by the ltered eld in the elds evolution equation.
This way, the random variable can be chosen in the form:
v (x, t) =
1
v (x, t, x , t )
(7.84)
with
x =
(7.85)
and where the time t is dened as before. Assuming that v is periodic along
x and t on the domain v and has an average of zero over this interval, the
subgrid term takes the form:
= Au
(7.86)
where A is a denite positive tensor such that the term Au is equal to the
average of the term (v u1 ) over v , in which u1 is a solution on v of the
problem:
u1
2x u1 + v x u1 = q + v u ,
t
x u1 = 0
(7.87)
(7.88)
Persson Tensorial Subgrid Viscosity. Persson, Fureby and Svanstedt

[589, 227] followed a similar procedure to derive a fourth-rank tensorial subgrid viscosity. They use the following equation for u1 as a starting point (the
rapid scale system coordinates is the same as in the second model of Perrier
and Pironneau):
u1
2x u1 + x (u1 v + v u1 ) x q = (v u)
t
x u1 = 0
, (7.89)
(7.90)
Based on this system, the authors derived the following expression for the
subgrid tensor:
uk
ij = Aijkl
,
(7.91)
xl
7.3 Scale Similarity Hypotheses and Models Using Them
231
where the anisotropic fourth-rank tensorial subgrid viscosity is dened as

Ahjkl =
2
K0 C1 qsgs
*hjkl
A
11/3
(2)
(7.92)
where the constant parameters are K0 = 1.4 (Kolmogorov constant) and

2
C1 = 1.05. The subgrid kinetic energy qsgs
is evaluated solving an additional
evolution equation deduced from (5.119) that includes the new denition of
the subgrid viscosity as a tensorial quantity instead of a scalar one. The nondimensional parameter is given by the following spectral summation over the
wave vectors m
(

*k )2 + (ml /
*l )2
(mk /
17/3
*hjkl =
A
1
Rm
2
*m
R
mZ
Z

*k )2 (ml /
*l )2 (2 kl )
(mk /
(hl jk + jl hk )
+
*4
R
m
)
*h )2 (mk /
*k )2 (1 jl )
(mh /
+2
,
(7.93)
*4
R
m
where
*m
R
8
2
9
9
mn
=:
,
*n
n=1,3
Rm =
*
Rm
(7.94)
where the scalar cuto length on a Cartesian grid is dened as =

(1 2 3 )1/3 , and is the expansion parameter. The cell aspect ratio are
*l = l /.
dened as
7.3 Scale Similarity Hypotheses

and Models Using Them
7.3.1 Scale Similarity Hypotheses
Basic Hypothesis. The scale similarity hypothesis such as proposed by
Bardina et al. [39, 40] consists in assuming that the statistical structure of
the tensors constructed on the basis of the subgrid scales is similar to that of
their equivalents evaluated on the basis of the smallest resolved scales. The
spectrum of the solution based on this hypothesis is therefore broken down
into three bands: the largest resolved scales, the smallest resolved scales (i.e.
the test eld), and the unresolved scales (see Fig. 5.14).
This statistical consistency can be interpreted in two complementary
ways. The rst uses the energy cascade idea. That is, the unresolved scales
232
and the smallest resolved scales have a common history due to their interactions with the largest resolved scales. The classical representation of the
cascade has it that the eect of the largest resolved scales is exerted on the
smallest resolved scales, which in turn inuence the subgrid scales, which
are therefore indirectly forced by the largest resolved scales, but similarly
to the smallest. The second interpretation is based on the idea of coherent
structures. These structures have a non-local frequency signature5 , i.e. they
have a contribution on the three spectral bands considered. Scale similarity
is therefore associated with the fact that certain structures appear in each of
the three bands, inducing a strong correlation of the eld among the various
levels of decomposition.
Extended Hypothesis. This hypothesis was generalized by Liu et al. [455]
(see [506] for a more complete discussion) to a spectrum split into an arbitrary
number of bands, as illustrated in Fig. 7.2. The scale similarity hypothesis is
then re-formulated for two consecutive spectrum bands, with the consistent
forcing being associated with the low frequency band closest to those considered. Thus the specic elements of the tensors constructed from the velocity
eld un and their analogous elements constructed from un+1 are assumed to
be the same. This hypothesis has been successfully veried in experiments
in the case of a jet turbulence [455] and plane wake turbulence [570]. Liu et
al. have also demonstrated that scale similarity persists during rapid straining [454].
7.3.2 Scale Similarity Models
This section presents the structural models constructed on the basis of the
scale similarity hypothesis. All of them make use of a frequency extrapolation technique: the subgrid tensor is a approximated by an analogous tensor
computed from the highest resolved frequencies. The following are described:
1. Bardinas model (p. 233) in which the subgrid tensor is computed by
applying the analytical lter a second time and thereby evaluating the
uctuation of the resolved scales. This model is therefore inoperative
when the lter is idempotent, because this uctuation is then null.
2. Filtered Bardina model (p. 234), which is an improvement on the previous
one. By construction, the subgrid tensor is a ltered quantity, which
results in the application of a convolution product and is therefore nonlocal in the sense that it incorporates all the information contained in
5
This is due to the fact that the variations of the velocity components associated
with a vortex cannot be represented by a monochromatic wave. For example, the
Lamb-Oseen vortex tangential velocity radial distribution is:

2
q
1 er
,
U =
r
where r is the distance to the center and q the maximum velocity.
233
Fig. 7.2. Spectral decomposition based on the extended scale similarity hypothesis.
the support of the lter convolution kernel. It is proposed in this model

to recover this non-local character by applying the lter to the modeled
subgrid tensor.
3. LiuMeneveauKatz model (p. 234), which generalizes the Bardina model
to the use of two consecutive lters of dierent shapes and cuto frequencies, for computing the small scale uctuations. This model can therefore
be used for any type of lter.
4. The dynamic similarity model (p. 236), which can be used to compute
the intensity of the modeled subgrid stresses by a dynamic procedure,
whereas in the previous cases this intensity is prescribed by hypotheses
on the form of the energy spectrum.
Bardina Model. Starting with the hypothesis, Bardina, Ferziger, and
Reynolds [40] proposed modeling the C and R terms of the Leonard decomposition by a second application of the lter that was used to separate
the scales. We furthermore have the approximation:

(7.95)
which allows us to say:

Rij = (ui ui )(uj uj ) ,
(7.96)
Cij = (ui ui )uj + (uj uj )ui
(7.97)
or
Rij + Cij = (ui uj ui uj )
(7.98)
234
Adding Leonards term, which is computed directly from the resolved

scales, we get:
ij = Lij + Rij + Cij = (ui uj ui uj )
(7.99)
This can be re-written in another using the generalized central moments

proposed by Germano [244]:
ij = G ([ui ]G , [uj ]G ) Lij
(7.100)
The subgrid tensor is therefore approximated by the generalized central

moment of the ltered eld dened like the tensor Lij in the Germano decomposition (see Sect. 3.3.2). Experience shows that this model is not eective
when the lter is a Reynolds operator, because the contribution thus computed then cancels out. Contrary to the subgrid viscosity models, this one
does not induce an alignment of the proper axis system of the subgrid tensor
on those of the strain rate tensor. Tests performed on databases generated by
direct numerical simulation have shown that this model leads to a very good
level of correlation with the true subgrid tensor, including when the ow is
anisotropic [320].
Despite its very good level of correlation6, experience shows that this
model is only slightly dissipative and that it underestimates the energy cascade. It does, on the other hand, include the backward cascade mechanism.
Filtered Bardina Model. The Bardina model (7.99) is local in space in
the sense that it appears as a product of local values. This local character
is in contradiction with the non-local nature of the subgrid tensor, so that
each component appears in the form of a convolution product. To remedy
this problem, Horiuti [323] and Layton and his colleagues [429, 428] propose
the ltered Bardina model:
ij = (ui uj ui uj ) = Lij
(7.101)
With this additional ltering operation, we recover the non-local character

of the subgrid tensor.
LiuMeneveauKatz Model. The Bardina model uses a second application of the same lter, and therefore a single cuto scale. This model is
generalized to the case of two cuto levels as [455]:
*i *
ij = Cl (u/
uj ) = Cl Lm
i uj u
ij
(7.102)
where the tensor Lm

ij is now dened by two dierent levels of ltering. The test
lter cuto length designated by the tilde is larger than that of the rst level.
6
The correlation coecient at the scalar level is generally higher than 0.8.
235
The constant Cl can be evaluated theoretically to ensure that the average

value of the modeled generalized subgrid kinetic energy is equal to its exact
counterpart [148]. This leads to the relation:
Cl =
uk uk uk uk
*k
*k u
u/
k uk u
(7.103)

Let F(k) and G(k)
be transfer functions associated with the grid lter
and test lter, respectively, and let E(k) be the energy spectrum of the exact
solution. Relation (7.103) can then be re-written as:
!
(1 F 2 (k))E(k)dk
.
(7.104)
Cl = ! 0
2 (k))F2 (k)E(k)dk
(1 G
0
Evaluations made using experimental data come to Cl 1 [570, 455]7 .

Shah and Ferziger [668] propose extending this model to the case of nonsymmetric lters.
To control the amplitude of the backward cascade induced by the model,
especially near solid walls, Liu et al. [455] propose the modied form:
ij = Cl f (ILS )Lm
ij
(7.105)
where the dimensionless invariant ILS , dened as

Lm
lk S lk

ILS =
m
Llk Lm
S lk S lk
lk
(7.106)
measures the alignment of the proper axes of the tensors Lm and S. As the
kinetic energy dissipated by the subgrid model is expressed
= ij S ij
(7.107)
we get, using model (7.105):

= Cl f (ILS )ILS
(7.108)
The backward energy cascade is modulated by controlling the sign and

amplitude of the product f (ILS )ILS . The authors considered a number of
choices. The rst is:

1
if
ILS 0
f (ILS ) =
.
(7.109)
0 otherwise
This solution makes it possible to cancel out the representation of the
backward cascade completely by forcing the model to be strictly dissipative.
7
The initial value of 0.45 0.15 given in [455] does not take the backward cascade
into account.
236
One drawback to this is that the function f is discontinuous, which can

generate numerical problems. A second solution that is continuous consists
in taking:

ILS
if
ILS 0
f (ILS ) =
.
(7.110)
0
otherwise
One last positive, continuous, upper-bounded solution is of the form:

f (ILS ) =
2
(1 exp(ILS
))
0
if
ILS 0
otherwise
(7.111)
in which = 10.
Dynamic Similarity Model. A dynamic version of the LiuMeneveau
Katz model (7.102) was also proposed [455] for which the constant Cl will no
longer be set arbitrarily. To compute this model, we introduce a third level
of ltering identied by .. The Q analogous to tensor Lm for this new level
of ltering is expressed:

*j u
*
*i u
*i
uj ) .
Qij = (u
(7.112)
The GermanoLilly dynamic procedure, based here on the dierence:

m
Mij = f (IQS )Qij f (I/
LS )Lij
(7.113)
where
IQS =
*
Qmn S
mn
*
|Q| |S|
Cl =
Lm
lk Mlk
Mpq Mpq
(7.114)
yields:
.
(7.115)
7.3.3 A Bridge Between Scale Similarity and Approximate

Deconvolution Models. Generalized Similarity Models
The Bardina model can be interpreted as a particular case of the approximate
soft deconvolution based models described in Sect. 7.2.
Using the second order dierential approximation
=+
(2) 2
2 x2
(7.116)
the Bardina model (7.99) is strictly equivalent to the second order gradient
model given by relations (7.18) and (7.19).
7.4 Mixed Modeling
237
It can also be derived using the Van Cittert deconvolution procedure:

a zeroth-order truncation in (7.7) is used to recover relation (7.95), while
a rst-order expansion is employed to derive (7.96).
The Bardina model then appears as a low-order formal expansion model
for the subgrid tensor. Generalized scale similarity models can then be dened using higher-order truncations for the formal expansion [254]. They are
formulated as
1
1
uj ) (G1
ij = (G1
l
ui )(Gl
l
u)i (Gl
u)j
(7.117)
where G1
l
designates the approximate soft deconvolution operator, dened
in Sect. 7.2.1.
The same key idea can be extended to close the NavierStokes on a truncated wavelet basis: this was achieved by Homan for several simplied systems [315, 311, 309, 308].
7.4 Mixed Modeling

7.4.1 Motivations
The structural models based on the scale similarity idea and the soft deconvolution models/techniques on the one hand, and the functional models
on the other hand, each have their advantages and disadvantages that make
them seem complementary:
The functional models, generally, correctly take into account the level of
the energy transfers between the resolved scales and the subgrid modes.
However, their prediction of the subgrid tensor structure, i.e. its eigenvectors, is very poor.
The models based on the scale-similarity hypothesis or an approximate deconvolution procedure generally predict well the structure of the subgrid
tensor better (and then are able to capture anisotropic eects and disequilibrium), but are less ecient for dealing with the level of the energy
transfers. It is also observed that these models yield a poor prediction of
the subgrid vorticity production [491], a fact coherent with their underdissipative character.
Dubois et al [202] also observed that these two parts yield dierent correlations with reference data. Tests have shown that mixed models are able to
capture disequilibrium and anisotropy eects [8, 454, 504, 506, 594, 12].
Shao et al. [670] propose a splitting of the kinetic energy transfer across
the cuto that enlights the role of each one of these two model classes. These
authors combine the classical large-eddy simulation convolution lter to the
238
ensemble average, yielding the following decompositions:

u =
=
=
u + ue
(7.118)
u + u
u + u + ue + u
(7.119)
(7.120)
Using this hybrid decomposition, the subgrid tensor splits into

ij = ijrapid + ijslow
(7.121)
with
e
e e
ijslow = ue
i uj u i u j
ijrapid
(7.122)
e
ui uj ui uj + ue
i uj u i uj
e
+ue
j ui u j ui .
(7.123)
These two parts can be analysed as follows:

The rapid part explicitly depends on the mean ow. This contribution
arises only if the convolution lter is applied in directions where the mean
ow gradients are non-zero. It is referred to as rapid because the time scale
of its response to variations of the mean ow is small. Numerical experiments show that this part plays an important role when the turbulence
is in a desiquilibrium state when: (i) production of kinetic energy is much
larger than dissipation or (ii) the lter length is of the same order as the
integral scale of turbulence. Subgrid stresses anisotropy is observed to be
due to the interaction of this rapid part and the mean shear. Numerical
simulations have shown that the rapid part escapes the functional modeling, but scale-similarity models and soft deconvolution models succeed
in representing anisotropic energy transfer (both forward and backward
cascades) associated to the rapid part.
The slow part is always present in large-eddy simulation, because it does
not depend on the mean ow gradients. It corresponds to the subgrid tensor
analyzed through the previously described canonical analysis. It is referred
to as slow because its relaxation time is long with respect to rapid part.
Numerical tests show that subgrid viscosity model correctly capture the
associated kinetic energy transfer.
One simple idea for generating subgrid models possessing good qualities
on both the structural and energy levels is to combine a functional with
a structural model, making what is called mixed models. This is generally
done by combining a subgrid viscosity model for representing the energy
cascade mechanism with a scale similarity. The stochastic backward cascade
models are usually not included because the structural models are capable of
including this phenomenon.
7.4 Mixed Modeling
239
The resulting form is

1
1
ij kk ij = 2sgs S ij + (Lij Lkk ij ) ,
3
3
(7.124)
where sgs is the subgrid viscosity (evaluated using one of the previously
described model), and Lij the evaluation obtained using one of the structural
model8 .
Another argument for using a mixed model originates from the splitting
of the full deconvolution model as the sum of the soft deconvolution model
and the hard deconvolution model (see Sect. 7.2.1): the scale-similarity models are formally equivalent to truncated Taylor-series-expansion-based soft
deconvolution models, and thus are not able to account for interactions between resolved scales and scales smaller than the mesh size on which the
equations are solved. Consequently, they must be supplemented by another
model, which will play the role of the secondary regularization within the
framework of the deconvolution approach.
Examples of such models are described in the following.
7.4.2 Examples of Mixed Models
We present several examples of mixed models here:
1. The SmagorinskyBardina model (p. 240), for which the respective
weights of each of the contributions are preset. This model is limited
by the hypotheses underlying each of the two parts constituting it: the
subgrid viscosity is still based on arguments of the innite inertial range
type. Experience shows, though, that combining the two models reduces
the importance of the constraints associated with these underlying hypotheses, which improves the results.
2. A one-parameter mixed model whose subgrid viscosity is computed by
a dynamic procedure of the GermanoLilly type (p. 240). With this procedure, the respective weights of the structural and functional parts of
the model can be modied, so that the subgrid viscosity model is now
computed as a complement to the scale similarity model, which allows
a better control of the dissipation induced. It can be said, though, that
this procedure innately prefers the structural part.
3. The general form of N -parameter dynamic mixed model, as derived by
Sagaut et al. (p. 241). This procedure is an extension of the previous one:
the weights of the dierent parts of the model are dynamically computed,
resulting in a possibly better approximation of the true subgrid stresses.
The case of two-parameter dynamic mixed model is emphasized.
8
Only scale-similarity models or approximate deconvolution models are used in

practice to derive mixed models, because they are very easy to implement.
240
Other mixed models have already been presented within the framework
of the deconvolution approach (see Sect. 7.2.1, p. 220), which will not be
repeated in the present section. Application of the results dealing with the
dynamic evaluation of the constants presented below to the Taylor series
expansion based deconvolution models is straightforward.
Mixed SmagorinskyBardina Model. The rst example is proposed by
Bardina et al. [40] in the form of a linear combination of the Smagorinsky
model (5.90) and the scale similarity model (7.99). The subgrid tensor deviator is then written:
1
1
ij kk ij =
3
2
in which

1
2sgs S ij + Lij Lkk ij
3

Lij = ui uj ui uj
(7.125)
(7.126)
and
2
sgs = Cs |S| .
(7.127)
Variants are obtained either by changing the subgrid viscosity model

used or by replacing the tensor L with the tensor Lm (7.105) or the tensor L (7.101).
One-Parameter Mixed Dynamic Model. A mixed dynamic model was
proposed by Zang, Street, and Kose [799]. This is based initially on the
Bardina model coupled with the Smagorinsky model, but the latter can be
replaced by any other subgrid viscosity model. The subgrid viscosity model
constant is computed by a dynamic procedure. The subgrid tensors corresponding to the two ltering levels are modeled by a mixed model:
1
1 m
ij kk ij = 2sgs S ij + Lm
ij Lkk ij
3
3
(7.128)
1
* + Q 1Q
Tij Tkk ij = 2sgs S
ij
ij
kk ij
3
3
(7.129)
in which
**
Qij = u/
i uj ui uj
(7.130)
and
sgs = Cd |S|
(7.131)
7.4 Mixed Modeling
The residual Eij is now of the form:

2
2C
m
+
P
Eij = Lm
ij
d
ij
ij
kk
ij
in which
** ,
Hij = u/
i uj ui uj
m
*i u
*j ,
u/
L
i uj u
ij =
* 2
*
*
/
mij =
|S|S ij |S|S
ij
241
(7.132)
(7.133)
(7.134)
,
(7.135)
and where Pkk represents the trace of the subgrid tensor. The GermanoLilly
dynamic procedure leads to:
Cd =
(Lm
ij Hij )mij
mij mij
(7.136)
In simulations performed with this model, the authors observed a reduction in the value of the dynamic constant with respect to that predicted by
the usual dynamic model (i.e. based on the Smagorinsky model alone). This
can be explained by the fact that the dierence between the Lm and H terms
appears in the numerator of the fraction (7.136) and that this dierence is
small because these terms are very similar. This shows that the subgrid viscosity model serves only to model a residual part of the full subgrid tensor
and not its entirety, as in the usual dynamic model.
Vreman et al. [747] propose a variant of this model. For the sake of mathematical consistency, by making the model for the tensor Tij dependent only
*
on the velocity eld that corresponds to the same level of ltering, i.e. u,
these authors propose the following alternate form for the tensor Qij :
**
*/
*j u
*
*j
Qij = u
iu
iu
(7.137)
N-Parameter Dynamic Mixed Model.

General Formulation and Formal Resolution. A general form of multiparameter dynamic model was derived by Sagaut et al. [630]. Considering a formal
N -part parametrization of the subgrid tensor, each term being associated to
a real constant Cl , k = 1, .., N

Cl fijl (u, ) ,
(7.138)
ij =
l=1,N
where the functions fijl are the kernels of the dierent parts of the complete
model. The equivalent formulation obtained at the test lter level is

*
* )
Cl fijl (u,
.
(7.139)
Tij =
l=1,N
242
Inserting (7.138) and (7.139) into the Germano identity (5.138), we get
the following denition of the residual Eij :
Eij = Lij
* f2l (u, ) .
* )
Cl mlij , mlij = fijl (u,
ij
(7.140)
l=1,N
In order to obtain N linearly independent relations to compute the constants Cl , a rst solution is to operate the contraction of the residual (7.140)
with N independent tensors Alij . The constants will then appear as the solutions of the following linear algebraic problem of rank N :

Cl mlij Akji = Lij Akji , k = 1, N .
(7.141)
l=1,N
It is worth noting that the N constants are coupled, resulting in a global

self-adaption of each constant. The particular case of the least-square minimization is recovered by taking Akij = mkij , k = 1, N .
In the case where some constants are not computed dynamically but are
arbitrarily set, the linear system (7.141) corresponds to a ill-posed problem
containing more constraints than degrees of freedom. Assuming that the N
rst constants are arbitrarily chosen, we recover a well-posed problem of rank
N N by replacing Lij with Lij , where
Lij = Lij
Cl mlij
(7.142)
l=1,N
Two-Parameter Dynamic Models. Mixed models have also been proposed by

Salvetti [640, 643] and Horiuti [324] with two dynamic constants (one for
the subgrid-viscosity part and one for the scale-similarity), corresponding to
the N = 2 case in the previous section. These models have the advantage of
avoiding any a priori preference for the contribution of one or the other model
component. An extensive study of dynamic mixed model has been carried out
by Sarghini et al. [646] for the plane channel case.
Numerical simulations show that two-paramater mixed models may yield
disappointing results, because of a too low dissipation level. This is due to the
fact that the coupled dynamic procedure described in the previous section
gives a heavy weigth to the scale-similarity part of the model, because its correlation coecient with the exact subgrid tensor is much higher than the one
of the subgrid-viscosity model. This conclusion was conrmed by Anderson
and Meneveau [12] in isotropic turbulence. These authors also observed a serious lack of robustness in regard to the dynamic Smagorinsky model when
the test lter cuto exceeds the integral scale of turbulence: negative values
of the dynamic constant are returned in this case. To relieve this problem,
Morinishi [542, 543] proposes to uncouple the computation of the dynamic
constants. The modied algorithm for the dynamic procedure is:
7.5 Dierential Subgrid Stress Models
243
1. Compute the constant associated to the subgrid-viscosity part of the

model using a classical dynamic procedure, without taking the scale similarity part into account. This corresponds to the N = 1 case in the
previous section. The resulting constant will ensure a correct level of
dissipation.
2. Compute the constant associated to the scale-similarity part using a twoparameter dynamic procedure, but considering that the constant of the
subgrid-viscosity part is xed. This corresponds to N = 2 and N = 1 in
the previous section.

A natural way to nd an expression for the subgrid stress tensor is to solve
a prognostic equation for each component. This approach leads to the denition of six additional equations, and thus a very signicant increase in both
the model complexity and the computational cost. But the expected advantage is that such a model would a priori be able to account for a large class
of physical mechanisms, yielding the denition of a very robust model. Three
proposal have been published by dierent research groups:
1. The pioneering model of Deardor (p. 243).
2. The model by Fureby et al. (p. 244), which is an improved version of the
Deardor model with better theoretical properties, such as realizability.
3. The models based of the use of transport equations for the velocity ltered
probability density function (p. 245), which do not rely on an explicit
model for the triple correlations.
7.5.1 Deardor Model
Another approach for obtaining a model for the subgrid tensor consists in
solving an evolution equation for each of its components. This approach proposed by Deardor [173] is analogous in form to two-point statistical modeling. Here, we adopt the case where the lter is a Reynolds operator. The
subgrid tensor ij is thus reduced to the subgrid Reynolds tensor Rij . We
deduce the evolution equation of the subgrid tensor components from that of
the subgrid modes (3.31)9 :
ij
t
9
uj
ui
(uk ij ) ik
jk
xk
xk
xk
This is done by applying the lter to the relation obtained by multiplying (3.31)
by uj and taking the half-sum with the relation obtained by inverting the subscripts i and j.
244

uj

ui
u u u +p
+
xk i j k
xj
xi

u uj

ui p
uj p 2 i
xj
xi
xk xk
(7.143)
The various terms in this equation have to be modeled. The models Deardor proposes are:
For the pressurestrain correlation term:
+

2

qsgs
u
u
2 2
2 2
j
i
+
ij + qsgs
S ij
p
= Cm
ij qsgs
xj
xi
3
5
(7.144)
2
where Cm is a constant, qsgs
the subgrid kinetic energy, and S ij the strain
rate tensor of the resolved eld.
For the dissipation term:
2 3/2
(qsgs
)
ui uj
= ij Ce
xk xk
(7.145)
where Ce is a constant.
For the triple correlations:

+
2
ui uj uk = C3m qsgs
jk +
ik +
ij
xi
xj
xk
(7.146)
The pressurevelocity correlation terms p ui are neglected. The values of

the constants are determined in the case of isotropic homogeneous turbulence:
Cm = 4.13, Ce = 0.70, C3m = 0.2
(7.147)
Lastly, the subgrid kinetic energy is determined using evolution equation (5.119).
7.5.2 Fureby Dierential Subgrid Stress Model
An alternate form of the dierential stress model is proposed by Fureby et
al. [232], which has better symmetry preservation properties than the original
Deardor model. The triple correlation term is approaximated as
+
ij
2
ui uj uk = cq qsgs

xk
(7.148)
sgs
with cq = 0.07. Other terms are kept unchanged. Each term in the closed
subgrid stress equations is now frame indierent. The set of closed equations
245
is realizable, and has the same transformation properties as the exact Navier
Stokes equations under a change of frame.
An alternate form of the subgrid viscosity in (7.148) which allows the
representation of backscatter is also dened by Fureby:
sgs =
ij S ij
S ij S ij
(7.149)
7.5.3 Velocity-Filtered-Density-Function-Based Subgrid

Stress Models
A methodology referred to as the Velocity Filtered Density Function to close
the ltered NavierStokes equations is proposed by Gicquel et al. [264] on the
grounds of the previous works by Pope and Givi. The subgrid scale stresses
are reconstructed by considering the joint probability function of all of the
components of the velocity vector. To this end, an exact evolution equation is
derived for the velocity ltered density function in which unclosed terms are
modeled. Two implementations have been proposed: a rst one consists in
discretizing transport equations for the subgrid stresses associated to the velocity ltered density function (this is the one emphasized below); the second
one consists in solving it via a Lagrangian Monte Carlo scheme, which leads
to the denition of an equivalent stochastic system. This last form should
be classied as a structural model based on a stochastic reconstruction of
subgrid scales, and will be mentioned in Sect. 7.7.
Denitions. The rst step consists in dening the velocity ltered density
function PL
+
PL (v; x, t) =
[v, u(x , t)]G(x x)dx ,
(7.150)
where G is the convolution lter kernel and [v, u(x , t)] is the ne-grained
density
[v, u(x , t)] = (v u(x, t)) .
(7.151)
It is observed that PL has all the properties of the probability density

function when the lter kernel is positive. The conditional ltered value of
a dummy variable (x, t) is therefore dened as
! +
(x, t)|u(x, t) = vL = |vL
(x , t)[v, u(x , t)]G(x x)dx

PL (v; x, t)
(7.152)
where |L denotes the ltered value of conditioned on .
246
First Model. The starting point of the procedure is to write an expression

for the time-derivative of PL . Such an expression is found by applying the
time-derivative operator to Eq. (7.150), yielding
PL (v; x, t)
t
ui (x , t) [v, u(x , t)]

G(x x)dx
t
vi
A

@
ui
v
=
P
(v;
x,
t)
(7.153)
L
vi
t
L
+
The next step consists in eliminating the velocity time derivative using
the momentum equation, leading to
PL
t
PL
p PL
=
[(vk uk ) PL ] +
xk
xk
xi vi
A
@

p
S ik PL
p
v
PL
2
+
xk vi
vi
xi
xi
L
A
@

Sik
S ik
v
2
PL
,
vi
xi
xi
uk
(7.154)
where the relation

G
=
|vL PL (v; x, t)dv
(7.155)
was used. The last two terms in the right hand side of (7.154) are unclosed
terms which require the denition of ad hoc subgrid models. The rst term
in the right hand side is related to the subgrid advection and is closed. The
sum of the two unknown terms is modeled as follows
vi
@
A
A

@

p
Sik
p
S ik
2
PL
v
v
L
xi
xi
vi
xi
xi
L
L

1
2 PL
[Gij (vj ui )PL ] C0

(7.156)
,
vi
2
vi vi
with

1
+
2
with C0 = 2.1 and where the subgrid
mixing frequency are evaluated as
Gij =
2 3/2
qsgs

3
C0 ij ,
(7.157)
4
dissipation rate and the subgrid
2
qsgs
(7.158)
247
2
The subgrid kinetic energy qsgs
is taken equal to half the trace of the
subgrid stress tensor. This closure is equivalent the the one proposed by
Rotta for the Reynolds-averaged NavierStokes equations.
The nal stage consists in writing the corresponding equations for the
subgrid stresses, which are dened here using the generalized central moment framework of Germano (Sect. 3.3.2), i.e. ij = G (ui , uj ). Evolution
equations consistant with the previous closed form of the velocity ltered
density function are
ij
t
ijk
(uk ij ) =
+ Gik jk + Gjk ik
xk
xk
uj
ui
ik
jk
+ C0 ij .
xk
xk
(7.159)
The last unknown term is the third-order generalized central moment

ijk = G (ui , uj , uk ). Its value are taken from the Lagrangian Monte Carlo
solver used to generate a prognostic stochastic velocity eld whose pdf PL
satises equation (7.154). Details of the numerical implementation can be
found in Ref. [264] and will not be reproduced here.
Second Model. An alternative form for relation (7.154) is
PL
t
p PL
PL
=
[(vk uk ) PL ] +
xk
xk
xi vi
A
@

p
p
v
PL
vi
xi
xi
L
A
@

2
2 PL
ui uj
v
PL
xk xk
vi vi
xk xk
L
uk
+
+
+
(7.160)
Using the same procedure as in the previous case, the following closed
form is obtained
PL
t
PL
p PL
=
[(vk uk ) PL ] +
xk
xk
xi vi
ui uj 2 PL
ui 2 PL
+
+ 2
xk xk vi vj
xk vi vk
1
2 PL
(Gij (vj uj )PL ) + C0

.
vi
2
vi vj
uk
(7.161)
The consistent subgrid stress equations are now

ij
t
ijk
(uk ij ) =
+ Gik jk + Gjk ik
xk
xk
2 ij
uj
ui
+
ik
jk
+ C0 ij .
xk xk
xk
xk
(7.162)
248
The dierence with the previous model is that viscous diusion is taken
into account in the subgrid stress transport equations, leading to a better
accuracy in ows where viscous eects are inuencial. The numerical implementation relies on the same Monte Carlo approach as the previous model.
7.5.4 Link with the Subgrid Viscosity Models
We reach the functional subgrid viscosity models again starting with a model
with transport equations for the subgrid stresses, at the cost of additional
assumptions. For example, Yoshizawa et al. [794] proposed neglecting all the
terms of equation (7.143), except those of production. The evolution equation
thus reduced comes to:
ij
uj
ui
= ik
jk
t
xk
xk
(7.163)
Assuming that the subgrid modes are isotropic or quasi-isotropic, i.e. that
the extra-diagonal elements of the subgrid tensor are very small compared
with the diagonal elements, and that the latter are almost mutually equal, the
right-hand side of the reduced equation (7.163) comes down to the simplied
form:
2
S ij ,
(7.164)
qsgs
2
in which qsgs
= uk uk /2 is the subgrid kinetic energy. Let t0 be the characteristic time of the subgrid modes. Considering the relations (7.163) and (7.164),
and assuming that the relaxation time of the subgrid modes is much shorter
than that of the resolved scales10 , we get
1
2
S ij
ij kk ij t0 qsgs
3
(7.165)
The time t0 can be evaluated by dimensional argument using the cuto

length and the subgrid kinetic energy:
t0 +
2
qsgs
(7.166)
By entering this estimate into equation (7.165), we get an expression

analogous to the one used in the functional modeling framework:
+
1
2 S
.
(7.167)
ij kk ij qsgs
ij
3
10
We again nd here the total scale-separation hypothesis 5.4.
7.6 Stretched-Vortex Subgrid Stress Models
249
7.6 Stretched-Vortex Subgrid Stress Models

7.6.1 General
Misra and Pullin [517, 608, 735], following on the works of Pullin and
Saman [609], proposed subgrid models using the assumption that the subgrid modes can be represented by stretched vortices whose orientation is
governed by the resolved scales.
Supposing that the subgrid modes can be linked to a random superimposition of elds generated by axisymmetrical vortices, the subgrid tensor can
be written in the form:

E(k)dkEpi Zpq Eqj ,
(7.168)
ij = 2
kc
in which E(k) is the energy spectrum, Elm the rotation matrix used to switch
from the vortex coordinate system to the reference system, Zij the diagonal
tensor whose main elements are (1/2, 1/2, 0) and Epi Zpq Eqj , the moment
of the probability density function P (, ) of the Euler angles and giving
the orientation of the vortex axis with respect to the frame of reference. The
statistical average performed on the Euler angles of a function f is dened
as:
2
1
f (Eij ) =
f (Eij )P (, ) sin()dd .
(7.169)
4 0 0
Two pieces of information are therefore needed to compute the subgrid
term: the shape of the energy spectrum for the subgrid modes and the subgrid
structure orientation distribution function. As the use of an evolution equation for the probability density function yielded no satisfactory results, Misra
and Pullin propose modeling this function as a product of Dirac functions or
a linear combination of such products. These are of the general form:
P (, ) =
4
( )( )
sin()
(7.170)
where (x, t) and (x, t) determine the specic orientation considered. Dening the two unit vectors e and ev :
e1 = sin() cos(), e2 = sin() sin(), e3 = cos(),
(7.171)
ev1 = sin() cos(), ev2 = sin() sin(), ev3 = cos(),
(7.172)
the subgrid tensor can be re-written in the form:

2

E(k)dk = ij evi evj qsgs
ij = ij evi evj
kc
(7.173)
250
The various models must thus specify the specic orientation directions of
the subgrid structures. Three models are presented in the following. The sub2
grid kinetic energy qsgs
can be computed in dierent ways (see Sect. 9.2.3),
for example by solving an additional evolution equation, or by using a double ltering technique. A local evaluation procedure in the physical space
based on the second-order velocity structure function is proposed by Voekl
et al. [735].
7.6.2 S3/S2 Alignment Model
A rst hypothesis is to assume that the subgrid structures are oriented along
the eigenvectors of the resolved strain rate tensor S ij that corresponds to
its two largest eigenvalues. This is equivalent to assuming that they respond
instantaneously to the forcing of the large scales. Using es2 and es3 to denote
these two vectors, and 2 and 3 2 the associated eigenvalues, we get
3

4
2
s3
s2
ij es3
+ (1 ) ij es2
,
(7.174)
ij = qsgs
i ej
i ej
where the weighting coecient is taken proportional to the norms of the
eigenvalues:
3
.
(7.175)
=
3 + |2 |
7.6.3 S3/ Alignment Model
The second model is derived on the assumption that the subgrid structures
are oriented along the third eigenvector of the tensor S ij , denoted es3 as
before, and the vorticity vector of the resolved eld. The unit vector it carries
is denoted e and is computed as:
e =
u
| u|
The subgrid tensor is evaluated as:

3

4
2
s3
ij = qsgs
ij es3
+ (1 ) ij e
i ej
i ej
(7.176)
(7.177)
The weighting parameter is chosen arbitrarily. The authors performed

tests considering the three values 0, 0.5, and 1.
7.6.4 Kinematic Model
Starting with the kinematics of a vortex lament entrained by a xed velocity
eld, Misra and Pullin propose a third model, for which the vector ev is obtained by solving an evolution equation. The equation for the ith component
of this vector is:
ui
uk
evi
= evj
evi evk evj
.
(7.178)
t
xj
xj
7.7 Explicit Evaluation of Subgrid Scales
251
The subgrid tensor is then evaluated by inserting the vector ev thus computed into the expression (7.173).

The models described in the present section are all based on an explicit
evaluation of the subgrid scales u (Id G)
u. Because the subgrid
modes correspond to scales of motion that can not be represented at the
considered ltering level (i.e. in practice on the computational grid), a new
higher-resolution ltering level is introduced. Numerically, this is done by
introducing an auxiliary computational grid (or a set of embedded auxiliary
grids), whose mesh size is smaller than the original one. The subgrid eld
u is evaluated on that grid using one of the model presented below, and
then the non-linear G
((u + u ) (u + u )) is computed. The corresponding
general algorithmic frame is
1. u is known from a previous calculation, on the computational grid, i.e.
at the G ltering level, whose characteristic length is .
2. Dene an auxiliary grid, associated to a new ltering level F with char* < , and interpolate u on the auxiliary grid.
acteristic length
3. Compute the approximate subgrid eld ua = (F G)
u using a model
on the auxiliary grid.
4. Compute the approximate non-ltered non-linear term at the F level on
the auxiliary grid:
(u + ua ) (u + ua ) .
5. Compute the approximate ltered non-linear term at the G level on the
computational grid:
G
((u + ua ) (u + ua ))
and use it to compute the evolution of u.

It is worth noting that this class of models can be interpreted as a generalization of deconvolution-based models (see Sect. 7.2). Classical deconvolution
models require the use of a second regularization to take into account the interactions with modes that cannot be reconstructed on the mesh. In the
present case, these unresolved scales are explicitly reconstructed on a ner
grid, rendering the approach more general. Looking at the algorithmic framework presented above, we can see that (i) the second step (interpolation) is
equivalent to the soft deconvolution, the third step is an extension of the
hard deconvolution and, (iii) the fth step is associated to the primary regularization.
Several ways to compute the subgrid motion on the auxiliary grid have
been proposed by dierent authors. They are classied by increasing order of
complexity (computational cost):
252
1. Fractal Interpolation Procedure of the uctuations, as proposed by Scotti

and Meneveau (p. 253). The subgrid uctuations are reconstructed in
a deterministic way on the ne grid using an iterative fractal interpolation technique (several similar fractal reconstruction techniques can be
found in [364]). This model is based on geometrical considerations only,
and does not take into account any information dealing with the ow dynamics such as disequilibrium, anisotropy, ... But it provides an estimate
of the subgrid motion at a very low cost.
2. Chaotic Map Model of McDonough et al. (p. 254). The subgrid uctuations are approximated in a deterministic way using a very simple chaotic
dynamical system, which is chosen in order to mimic some properties of
the real turbulent uctuations (amplitude, autocorrelation, distribution
of velocity uctuations, ...).
This model is the easiest to implement, and induces a very small overhead. A problem is that it requires the denition of a realistic dynamical
system, and then a complete knowledge of the turbulent motion characteristics at each point of the numerical simulation.
3. One-Dimensional Turbulence model for the uctuations, as proposed by
Kerstein and his co-workers (p. 257). This approach relies on the denition of a simplied model for the three subgrid velocity component along
lines located inside the computational cells. This model can be seen as
an improvement of the Chaotic Map Model, since the non linear cascade
eect are taken into account via the use of chaotic map, but diusion
eects and resolved pressure coupling are incorporated by solving a differential equation.
4. Reconstruction of the subgrid velocity eld using kinematic simulations
(p. 259). This approach, proposed by Flohr and Vassilicos, provides an
incompressible, random, statistically steady, isotropic turbulent velocity
eld with prescribed energy spectrum. This model does not require that
we solve any dierential equation, and thus has a very low algorithmic
cost.
5. The Velocity Filtered Density Function proposed by Gicquel et al. (p.
260) is another model belonging to this family, based on the use of
a stochastic model for the subgrid uctuations. Its equivalent dierential
formulation being presented in Sect. 7.5.3, the emphasis is put here on
the equivalent stochastic system. In this form, it can be interepreted as
the most advanced stochastic reconstruction technique for the subgrid
scales, since it retains all the complexity of the NavierStokes dynamics.
6. Subgrid Scale Estimation Procedure proposed by Domaradzki and his
coworkers (p. 261). The subgrid uctuation are now deduced from a simplied advection equation, deduced from the ltered NavierStokes operator. An evaluation of the subgrid motion production term is derived,
and integrated over a time interval associated to characteristic relaxation
time of the subgrid scale. This model makes it possible to evaluate the
253
subgrid motion at a very low computational cost, but requires the computation of an approximate inverse lter.
7. Multilevel Simulations (p. 263), which are based on the use of the exact
NavierStokes equations on a set of embedded computational grids. The
reduction of the computational eort with respect to the Direct Numerical Simulation is obtained by freezing (quasi-static approximation) the
high-frequencies represented on ne grids for some time interval, leading
to the denition of a cyclic strategy. These methods can be interpreted
as a time-consistent extension of the classical multigrid procedures for
steady computations. They correspond to the maximal computational
eort, but also to the most realistic approach.
7.7.1 Fractal Interpolation Procedure
Scotti and Meneveau [661, 662] propose to reconstruct the subgrid velocity
eld using two informations: (i) the resolved velocity eld, which is known on
the coarsest grid, and (ii) the fractality of the velocity eld. The uctuations
are evaluated by interpolating the resolved coarse-grid velocity eld on the
ne grid using a fractal interpolation technique.
We rst describe this interpolation technique in the monodimensional
case. It is based on an iterative mapping procedure. The uctuating eld
ua is reconstructed within each interval of the coarse grid by introducing
a local coordinate [0, 1]. Let us consider the interval [xi1 , xi+1 ], where
i 1 and i + 1 are related to the grid index on the coarse grid. We have =
(x xi1 )/2. The proposed map kernel W for a function to interpolated
on the considered interval is:

W []() =
di,1 (2) + qi,1 (2)

di,2 (2) + qi,2 (2)
if [0, 1/2]
if ]1/2, 1]
(7.179)
where qi,j are polynomials and di,j are stretching parameters. The authors
propose to use the following linear polynomials:
qi,1 () =
((xi ) (xi1 ) di,1 ((xi+1 ) (xi1 ))

+(xi1 )(1 di,1 )
qi,2 () =
(7.180)
((xi+1 ) (xi ) di,2 ((xi+1 ) (xi1 ))

(xi1 )di,2
(7.181)
The uctuation is dened as

ua = lim W n [u] = W W ... W [u]
n

n times
(7.182)
254
The stretching parameters are such that the Hausdor dimension D of

the synthetic signal is equal to
>
log(|di,1 |+|di,2 |)
if 1 < |di,1 | + |di,2 | < 2
1+
log(2)
.
(7.183)
D=
1
if |di,1 | + |di,2 | 1
In order to conserve then mean value of the signal over the considered
interval, we have di,1 = di,2 = d. For three-dimensional isotropic turbulence,
we have D = 5/3, yielding d = 21/3 .
This procedure theoretically requires an innite number of iterations to
build the uctuating eld. In practice, a nite number of iterations is used.
The statistical convergence rate of process being exponential, it still remains
a good approximation. A limited number of iterations can also be seen as
a way to account for viscous eects.
The extension to the multidimensional case is straightforward, each direction of space being treated sequentially.
This procedure also makes it possible to compute analytically the subgrid tensor. The resulting model will not be presented here (see [662] for
a complete description).
7.7.2 Chaotic Map Model
McDonough and his coworkers [552, 338, 469] propose an estimation procedure based on the denition of a chaotic dynamical system. The resulting
model generates a contravariant subgrid-scale velocity eld, represented at
discrete time intervals on the computational grid:
ua = Au V
(7.184)
where A is an amplitude coecient evaluated from canonical analysis, an

anisotropy correction vector consisting mainly of rst-order structure function of high-pass ltered resolved scales, and V is a vector of chaotic algebraic
maps. It is important noting that the two vectors are multiplied using a vector Hadamard product, dened for two vectors and a unit vector i according
to:
( V ) i ( i)(V i) .
(7.185)
The amplitude factor is chosen such that the kinetic energy of the synthetic subgrid motion is equal to the energy contained in all the scales not
resolved by the simulation. It is given by the expression:
1/6
Au = Cu u Re
with
(7.186)
1/2
u = (|u|)
, Re =
|u|
255
where is the molecular viscosity. The scalar coecient Cu is evaluated from

classical inertial range arguments. The suggested value is Cu = 0.62.
The anisotropy vector is computed making the assumption that the ow
anisotropy is smoothly varying in wave-number. In a way similar to the one
proposed by Horiuti (see Sect. 6.3.3), the rst step consists in evaluating the
anisotropy vector from the highest resolved frequency. In order to account
for the anisotropy of the lter, the resolved contravariant velocity eld uc is
considered. The resulting expression for is:
=
s
|J 1
s|
(7.187)
where J 1 is the inverse of the coordinate transformation matrix associated

to the computational grid (and to the lter). The vector s is dened according
to

|(2
uc i)|
,
(7.188)
si = 3

|2
uc |
2c is related to the test eld computed thanks to the use of the

where the u
* > .
test lter of characteristic length
We now describe the estimation procedure for the stochastic vector V .
In order to recover the desired cross-correlation between the subgrid velocity
component, the vector V is dened as:
V = AM
(7.189)
where A is a tensor such that R = A AT , where R is the correlation tensor

of the subgrid scale velocity. In practice, McDonough proposes to use the
evaluation:
(*
ui )j
.
(7.190)
Aij =
|*
ui |
Each component Mi , i = 1, 2, 3 of the vector M is of the form:

al
Mlm
,
Mi =
l=0,N
(7.191)
m=1,Nl
where Nl is the binomial coecient

N
Nl
l
and = 1.67 is the standard deviation for the variable, and the weights al
are given by
1/2

al = 3 pl (1 p)(N l)
, p = 0.7 .
(7.192)
256

The maps Mlm
are all independent instances of one of the three following
normalized maps:
The tent map:
(n+1)
R(2 3m(n) )
=
R(3m(n) )
R(2 3m(n) )
if m(n) < 1/3

if 1/3 m(n) 1/3
if m(n) > 1/3
(7.193)
where m(n) is the nth instance of the discrete dynamical system, and R
[1, 1].
The logistic map:
m(n+1) = RAR m(n) (1 |m(n) |Am )
with
AR = 2 + 2 2,
Am =

1
3
1+
AR
2
(7.194)
The sawtooth map:
(n+1)
R(2 + 3m(n) )
=
R(3m(n) )
R(2 + 3m(n) )
if m(n) < 1/3

if 1/3 m(n) 1/3
if m(n) > 1/3
(7.195)
The map parameter R is related to some physical ow parameter, since

the bifurcation and autocorrelation behaviors of the map are governed by
R. An ad hoc choice for R will make it possible to model some of the local
history eects in a turbulent ow in a way that is quantitatively and qualitatively correct. It is chosen here to set the bifurcation parameter R on the
basis of local ow values, rather than on global values such as the Reynolds
number. That choice allows us to account for large-scale intermittency eects.
Selecting the ratio of the Taylor and Kolmogorov scales, a possible choice
is:
7
r

(/)
1
tanh (Rc )
,
(7.196)
R = tanh
(/)c
where r is a scaling exponent empirically assumed to lie in the range [4, 6],
and (/)c is a critical value of the microscale ratio that is mapped onto Rc ,
the critical value of R. Suggested values are given in Table 7.2.
The last point is related to the time scale of the subgrid scales. Let te
be the characteristic relaxation time of the subgrid scales, to be evaluated
using inertial range considerations. If this time scale is smaller than the time
step t of the simulation (the characteristic lter time), then the stochastic
257
Table 7.2. Parameters of the Chaotic Map Model.

Map
Rc
(/)c
26
Logistic
(2 + 2 2)1/2
Tent
-1/3
28.6
Sawtooth
-1/3
28.6
variables Mi must be updated nu times per time step, with

t|u|
t
1/3
nu
=
,
Re
te
fM
(7.197)
where fM is a fundamental frequency associated with the chaotic maps used

to generate the variables. It is dened as:
fM =
(7.198)
where C is some positive constant and the integral iteration scale

=

m(n) m(n+l)
1
(0) +
(l), (l) =
2
m(n) m(n)
l=1,
(7.199)
which completes the description of the model. This model is Galilean- and
frame-invariant, and automatically generates realizable Reynolds stresses. It
reproduces the desired root-mean-square amplitude of subgrid uctuations,
along with the probability density function for this amplitude. Finally, the
proper temporal auto-correlation function can be enforced.
7.7.3 Kersteins ODT-Based Method
A more complex chaotic map model based on Kersteins One-Dimensional
Turbulence (ODT) approach11 was also proposed [650, 387, 389, 390, 388,
305, 306, 208, 198, 386, 771]. This is a method for simulating turbulent uctuations along one-dimensional lines of sight through a three-dimensional
turbulent ows. The velocity uctuations evolve by two mechanisms, namely
the molecular diusion and turbulent stirring. The latter mechanisms is taken
into account by a sequence of fractal transformations denoted eddy events.
An eddy event may be interpreted as a model of an individual eddy, whose
location, length scale and frequency are determined using a non-linear probabilistic model.
11
It is worth noting that ODT originates in the Linear Eddy Model [384, 385, 117,
118, 177, 696, 414, 413, 473].
258
The diusive step consists in solving the following one-dimensional advection-diusion equation for each subgrid velocity component along the line
ui
p
2 ui
+
(Vj ui ) =
+
t
xj
xi
x2
(7.200)
where is the molecular viscosity and V is the local advective eld such that

ui =
Vi ()d ,
(7.201)
where is the volume based on the cuto length .

The second step, which accounts for non-linear eects, is more complex
and consists in two mathematical operations. The rst one is a measurepreserving map representing the turbulent stirring, while the second one is
a modication of the velocity proles in order to implement energy transfers
among velocity components. These two steps can be expressed as
ui (x) ui (f (x)) + ci K(x) ,
where the stirring-related mapping
3(x x0 )
2l 3(x x0 )
f (x) = x0 +
3(x x0 ) 2l
x x0
(7.202)
f (x) is dened as
if x0 x x0 + l/3
if x0 + l/3 x x0 + 2l/3
if x0 + 2l/3 x x0 + l
otherwise
(7.203)
where l is the length of the segment aected by the eddy event. The second
term in the right hand side of (7.202) is implemented to capture pressureinduced energy redistribution between velocity components and therefore
makes it possible to account for the return to isotropy of subgrid uctuations. The kernel K is dened as
K(x) = x f (x)
(7.204)
The amplitude coecients ci are determined for each eddy to enforce

= the
two
following
constraints:
(i)
the
total
subgrid
kinetic
energy
E
=
i Ei =
= 1!

u
(x)u
(x)dx
remains
constant,
and
(ii)
the
subgrid
scale
spectrum
i
i
i 2
must be realizable, i.e. the energy extracted from a velocity component cannot
exceed the available energy in this component. The resulting denition of the
coecients is
'

27
wi + sign(wi ) (1 )wi2 +
(7.205)
wj2 ,
ci =
4l
2
j=i
259
where
1
wi = 2
l
ui (f (x))K(x)dx
4
= 2
9l
x0 +l
ui (x)(l 2(x x0 ))dx
. (7.206)
x0
The degree of energy redistribution is governed by the parameter , which

is taken equal to 2/3 in Ref. [650] (corresponding to equipartition of the
available energy among the velocity components).
The last element of the method is the eddy selection step, which give
access to the time sequence of eddy events. All events are implemented instantaneously, but occur with frequencies comparable to turnover frequencies
of associated turbulent structures. At each time step, the event-rate distribution is obtained by rst associating a time scale (x0 , l) with every eddy
event. Using l/ and l3 / 2 as an eddy velocity scale and a measure of the
energy of the eddy motion, respetively, the time scale is computed using
the following relation
2
l
2
(1 )w12 + (w22 + w32 ) Z 2 ,
(7.207)
2
l
where Z is the amplitude of the viscous penalty term that governs the size
of the smallest eddies for given local strain conditions. A probabilistic model
can be derived dening an event-rate distribution
(x0 , l, t) =
C
l2 (x0 , l, t)
(7.208)
where C is an arbitrary parameter which determines the relative strength of

turbulent stirring.
7.7.4 Kinematic-Simulation-Based Reconstruction
Following Flohr and Vassilicos [220], the incompressible, turbulent-like subgrid velocity eld is generated by summing dierent Fourier modes

u (x, t) =
(an cos(kn x + n t) + bn sin(kn x + n t)) , (7.209)
n=1,N
where N is the number of Fourier modes, an and bn are the amplitudes

corresponding to wave vector kn , and n is a time frequency. The wave
vectors are randomly distributed in spherical shells:
kn = kn (sin cos , sin sin , cos ) ,
(7.210)
where and are uniformly distributed random angles within [0, 2[ and
[0, ], respectively. The random uncorrelated amplitude vectors an and bn
260
are chosen such that

an kn = bn kn = 0
(7.211)
to ensure incompressibility, and

|an |2 = |bn |n = 2E(kn )kn
(7.212)
where E(k) is the prescribed energy spectrum, and kn is the wave number increment between the shells. Recommended shell distributions in the
spectral space are:
Linear distribution
kn = k1 +
kN k1
(n 1) ;
N 1
(7.213)
Geometric distribution

kn = k1
kN
k1
(n1)/(N 1)
;
(7.214)
Algebraic distribution
kn = k1 nlog(kN /k1 )/ log N
(7.215)
The time frequency n is arbitrary. Possible choices are n

= Uc kn if all
the modes are advected with a constant velocity Uc , and n = kn3 E(kn ) if
it is proportional to the eddy-turnover time of mode n.
In practice, Flohr and Vassilicos use this model to evaluate the dynamics
of a passive tracer, but do not couple it with the momentum equations.
Nevertheless, it could be used to close the momentum equation too.
7.7.5 Velocity Filtered Density Function Approach
The reconstruction of the subgrid motion via a stochastic system which obeys
the required probability density function is proposed by Gicquel et al. [264].
This method is also equivalent (up to the second order) to solving the dierential equations for the subgrid stresses presented in Sect. 7.5.3. The bases
of the method are presented in this section, and will not be repeated here.
The key of the present method is the denition of a Lagrangian Monte
Carlo method, which is used to evaluate both the position Xi (in space) and
the value of a surrogate of the subgrid velocity, Ui , associated to a set of
virtual particules. The value of the subgrid velocity in each cell of the LargeEddy Simulation grid is dened as the statistical average over all the virtual
particules that cross the cell during a xed time interval.
261
The stochastic dierential equations equivalent to the rst model presented in Sect. 7.5.3 is
dXi (t) = Ui (t)dt
dUi (t) =

p
sik
+ 2
+ Gij (Uj (t) uj (t)) dt
xi
xk

+ C0 dWiv (t) ,
(7.216)
(7.217)
where Gij , C0 and are dened in Sect. 7.5.3, and Wiv denotes and independent WienerLevy process.
The second model proposed by Gicquel accounts for viscous diusion and
is expressed as
dXi (t) = Ui (t)dt + 2 dWix (t) ,

(7.218)
dUi (t) =

p
sik
+ 2
+ Gij (Uj (t) uj (t)) dt
xi
xk

ui
dWjx (t) ,
+ C0 dWiv (t) + 2
xj
(7.219)
where is the molecular viscosity and Wix is another independent Wiener

Levy process. In practice, convergence of the statistical average over the particules within each cell must be carefully checked to recover reliable results.
7.7.6 Subgrid Scale Estimation Procedure
A two-step subgrid scale estimation procedure in the physical space12 is proposed by Domaradzki and his coworkers [458, 187, 394, 188]. In the rst
(kinematic) step, an approximate inversion of the ltering operator is performed, providing the value of the deltered velocity eld on the auxiliary
grid. In the second (non-linear dynamic) step, scales smaller than the lter
length associated to the primary grid are generated, resulting in an approximation of the full solution.
Let u be the ltered eld obtained on the primary computational grid,
and u the deltered eld on the secondary grid. That secondary grid is
chosen such that the associated mesh size is twice as ne as the mesh size
of the primary grid. We introduce the discrete ltering operator Gd , dened
such that
Gd u = u .
12
(7.220)
A corresponding procedure in the spectral space is described in reference [190].
262
It is important to note that in this two-grid implementation, the righthand side of equation (7.220) must rst be interpolated on the auxiliary grid
to recover a well-posed linear algebra problem. To avoid this interpolation
step, Domaradzki proposes to solve directly the ltered NavierStokes equations on the nest grid, and to dene formally the G ltering level by taking
* The deltered eld u is obtained by solving the inverse problem
= 2.
u = (Gd )1 u .
(7.221)
This is done in practice by solving the corresponding linear system. In

practice, the authors use an three-point discrete approximation of the box
lter for Gd (see Sect. 13.2 for a description of discrete test lters). This
step corresponds to an implicit deconvolution procedure (the previous ones
were explicit procedures, based on the construction of the inverse operator
via Taylor expansions or iterative procedures), and can be interpreted as an
interpolation step of the ltered eld on the auxiliary grid.
The ua subgrid velocity eld is then evaluated using an approximation of
its associated non-linear production term:
ua = a N
(7.222)
where a and N are a characteristic time scale and N the production rate.
These terms are evaluated as follows. The full convection term on the auxiliary grid is
u
uj i , j = 1, 2, 3 .
(7.223)
xj
This term accounts for the production of all the frequencies resolved on
the auxiliary grid. Since we are interested in the production of the small
scales only, we must remove the advection by the large scales, and restrict
the resulting term tho the desired frequency range. The resulting term Ni is

u
Ni = (Id G)
(uj uj ) i
xj
(7.224)
In practice, the convolution lter G is replaced by the discrete operator

Gd . The production time a is evaluated making the assumption that the
subgrid kinetic energy is equal to the kinetic energy contained in the smallest
resolved scales:
|ua |2 = a2 |N |2 = 2 |u u|2 = a =
2|u u|
|N |
(7.225)
where is a proportionality constant, nearly equal to 0.5 for the box lter.
This completes the description of the model.
263
7.7.7 Multi-level Simulations

This class of simulation relies on the resolution of an evolution equation
for ua on the auxiliary grid. These simulations can be analyzed within the
framework of the multiresolution representation of the data [293, 295, 17,
294], or similar theories such as the Additive Turbulent Decomposition [471,
338, 80].
Let us consider N lters G1 , ..., GN , with associated cuto lengths 1
... N . We dene the two following sets of velocity elds:
un = Gn
...
G1
u = G1n
u ,
(7.226)
v n = un un+1 = (G1n G1n+1 )

u = Fn
u .
(7.227)
The elds u and v are, respectively, the resolved eld at the nth level
of ltering and the nth level details. We have the decomposition

un = unk +
v nl ,
(7.228)
l=1,k
yielding the following multiresolution representation of the data:

u {uN , v 1 , ..., v N 1 }
(7.229)
The multilevel simulations are based on the use of embedded computational grids or a hierarchical polynomial basis to solve the evolution equations
associated with each ltering level/details level. The evolution equations are
expressed as
un
+ N S(un ) = n = [G1n
, N S](u),
t
n [1, N ] ,
(7.230)
where N S is the symbolic NavierStokes operator and [., .] the commutator

operator. The equations for the details are
vn
+ N S(v n ) = n = [Fn
, N S](u),
t
n [1, N 1] ,
(7.231)
or, equivalently,
v n
+ N S(un ) N S(un+1 ) = n + n+1 ,
t
n [1, N 1] .
(7.232)
There are three possibilities for reducing the complexity of the simulation
with respect to Direct Numerical Simulation:
The use of a cycling strategy between the dierent grid levels. Freezing the
high-frequency details over some time while integrating the equations for
the low-frequency part of the solution results in a reduction of the simulation complexity. This is referred to as the quasistatic approximation for the
264
high frequencies. The main problem associated with the cycling strategy is
the determination of the time over which the high frequencies can be frozen
without destroying the quality of the solution. Some examples of such a cycling strategy can be found in the Multimesh method of Voke [736], the
Non-Linear Galerkin Method [174, 579, 705, 222, 223, 224, 91], the Incremental Unknowns technique [73, 201, 128, 200], Tzipermans MTS algorithm [722], Lius multigrid method [452, 453] and the Multilevel algorithm
proposed by Terracol et al. [711, 710, 633].
The use of simplied evolution equations for the details instead of (7.231).
A linear model equation is often used, which can be solved more easily than
the full nonlinear mathematical model. Some examples among others are
the Non-Linear Galerkin method, early versions of the Variational Multiscale approach proposed by Hughes et al. [332, 331, 336, 335, 333, 334],
and the dynamic model of Dubrulle et al. [203, 426]. Another possibility
is to assume that the nth-level details are periodic within the ltering cell
associated with the (n 1)th ltering level. Each cell can then be treated
separately from the others. An example is the Local Galerkin method of
McDonough [468, 466, 467]. It is interesting to note that this last assumption is shared by the Homogenization approach developed by Perrier and
Pironneau (see Sect. 7.2.3). Menon and Kemenov [382] use a simplied
set of one-dimensional equations along lines. A reminiscent approach was
developed by Kerstein on the grounds of the stochastic One-Dimensional
Turbulence (ODT) model (presented in Sect. 7.7.3).
The use of a limited number of ltering levels. In this case, even at the nest
description level, subgrid scales exist and have to be parametrized. The gain
is eective because it is assumed that simple subgrid models can be used
at the nest ltering level, the associated subgrid motion being closer to
isotropy and containing much less energy than at the coarser ltering levels.
Examples, among others, are the Multilevel algorithm of Terracol [710,
633], the Modied Estimation Procedure of Domaradzki [193, 183], and
the Resolvable Sublter Scales (RSFS) model [806].
Some strategies combining these three possibilities can of course be dened. The eciency of the method can be further improved by using a local
grid renement [699, 62, 378]. Non-overlapping multidomain techniques can
also be used to get a local enrichment of the solution [611, 633]. These methods are presented in Chap. 11.
We now present a few multilevel models for large-eddy simulation. The
emphasis is put here on methods based on two grid levels, and which can be
interpreted as models, in the sense that they rely on some simplications and
cannot be considered just as multilevel algorithms applied to classical largeeddy simulations. In these methods, the secondary grid level is introduced to
compute the uctuations u , and not for the purpose of reducing the cost of
the primary grid computation. This latter class of methods, which escapes
the simple closure problem, is presented in Chap. 11. It is worth noting that
265
Fig. 7.3. Schematic of the three-level models. Left: spectral decomposition; Middle:
computational grid; Right: time cycling.
these two-grid methods correspond to a three-level decomposition of the exact

solution: two lters are applied in order to split the exact solution into three
spectral bands (see Fig. 7.3).
Using Hartens representation (7.229), this decomposition is expressed as
u = {u2 , v 1 , v 0 }
(7.233)
where u2 is the resolved ltered eld at the coarsest level, u1 = u2 + v 1 is

the resolved ltered eld at the nest level, and v 0 is the unresolved eld at
the nest level, i.e. the true subgrid velocity eld. The detail v 1 corresponds
to the part of the unresolved eld at the coarsest level which is resolved at
the nest ltering level.
The coupling between these three spectral bands and the associated closure problem can be understood by looking at the nonlinear term. For the
sake of simplicity, but without restricting the generality of the results, we will
assume here that the ltering operators perfectly commute with dierential
operators, and that the domain is unbounded. The remaining coupling comes
from the nonlinear convective term. For the exact solution, it is expressed as
B(u, u) = B(u2 + v 1 + v 0 , u2 + v 1 + v 0 ) ,
(7.234)
where B is the bilinear form dened by relation (3.27).

At the coarsest resolution level, the nonlinear term can be split as follows:
2
B(u, u)
= B(u2 , u2 )

I
266
2
+ B(u2 , v 1 ) + B(v1 , u2 ) + B(v 1 , v 1 )

II
2
+ B(u2 + v 1 , v 0 ) + B(v0 , u2 + v 1 + v 0 )

(7.235)
III
Term I can be computed directly at the coarsest grid level. Term II

represents the direct coupling between the two levels of resolution, and can
be computed exactly during the simulation. Term III represents the direct
coupling with the true subgrid modes. It is worth noting that, at least theoretically, the non-local interaction between u2 and v 0 is not zero and requires
the use of an ad hoc subgrid model.
At the nest resolution level, the analogous decomposition yields
1
B(u, u)
B(u2 , u2 )

IV
1
+ B(u2 , v 1 ) + B(v1 , u2 )

V
1
+ B(v 1 , v 1 )

(7.236)
VI
1
+ B(u2 + v 1 , v 0 ) + B(v 0 , u2 + v 1 ) + B(v0 , v 0 ) .

V II
Terms IV and V represent the coupling with the coarsest resolution level,
and can be computed explicitly. Term V I is the nonlinear self-interaction of
the detail v 1 , while term V II is associated with the interaction with subgrid
scales v 0 and must be modeled.
By looking at relations (7.235) and (7.236), we can see that specic subgrid models are required at each level of resolution. Several questions arise
dealing with this closure problem:
1. Is a subgrid model necessary in practice for terms III and V II?
2. What kind of model should be used?
3. Is it possible to use the same model for both terms?
Many researchers have worked on these problems, leading to the denition
of dierent three-level strategies. A few of them are presented below:
1. The Variational Multiscale Method (VMS) proposed by Hughes et al.
(p. 267).
2. The Resolvable Sublter-scale Model (RSFR) of Zhou et al. (p. 269).
3. The Dynamic Sublter-scale Model (DSF) developed by Dubrulle et al.
(p. 270).
267
4. The Local Galerkin Approximation (LGA), as dened by McDonough et

al. (p. 270).
5. The Two-Level-Simulation (TLS) method, proposed by Menon et al.
(p. 271).
6. The Modied Subgrid-scale Estimation Procedure (MSEP) of Domaradzki et al. (p. 269).
7. Terracols multilevel algorithm (TMA) with explicit modeling of term III
(p. 271).
The underlying coupling strategies are summarized in Table 7.3.
Table 7.3. Characteristics of multilevel subgrid models. + means that the term is
taken into account, and that it is neglected.
Model
II
III
IV
VI
VII
VMS
RSFR
MSEP
DSF
LGA
TLS
TMA
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
Variational Multiscale Method. Hughes et al. [332, 331, 336, 335, 333,
334] rst introduced the Variational Multiscale Method within the framework
of nite element methods, and then generalized it considering a fully general
framework.
The coupling term III is neglected (see Table 7.3), at least in the original
formulation of the method. The need for a full coupling including term III
was advocated by Scott Collis [660].
2
In practical applications, the coarse resolution cuto length scale is
1
taken equal to twice the ne resolution cuto length scale , and the solution
is integrated with the same time step at the two levels.
The subgrid term V II is parametrized using a Smagorinsky-like functional model

vj1
vi1
+
,
(7.237)
(V II)ij = 2sgs
xi
xj
with two possible variants:
The smallsmall model13 :
13
This form of the dissipation is very close to the variational embedded stabilization previously proposed by Hughes (see Sect. 5.3.4). Similar expressions for the
dissipation term have been proposed by Layton [429, 428] and Guermond [280].
268
sgs

= (CS 1 ) 2|S 1 |,
2

1
Sij
vj1
v 1
+ i
xi
xj

.
(7.238)
(7.239)
The largesmall model:

sgs
+
1
= (CS 1 ) 2|S |,
2

1
S ij
u1j
u1
+ i
xi
xj
The recommended value of the constant CS is 0.1 for isotropic turbulence

and plane channel ow. This value is arbitrary, and numerical experiments
show that it could be optimized.
Numerical results show that both variants lead to satisfactory results on
academic test cases, including non-equilibrium ow. This can be explained by
2
the fact that the subgrid tensor is evaluated using the S 1 tensor, instead of S
in classical subgrid-viscosity methods. Thus, the subgrid model dependency
is more local in Fourier space, the emphasis being put on the highest resolved
frequency, yielding more accurate results (see Sect. 5.3.3).
This increased localness in terms of wavenumber with respect to the usual
Smagorinsky model is emphasized rewriting the VMS method as a special
class of hyperviscosity models. The link between these two approaches is
enlightened assuming that the following dierential approximation for secondary lter utilized to operate the splitting u1 = u2 + v 1 holds14 :
2
u2 = G2
u1 (Id + 2 2 )u1
(7.240)
where is a lter-dependent parameter, yielding

2
v 1 = 2 2 u1
(7.241)
Inserting that denition for v 1 into previous expressions for both the
Small-Small (7.238) and the Large-Small (7.239) models shows that these
models are equivalent to fourth-order hyperviscosity models (see p. 121).
Higher-order hyperviscosities are recovered using higher-order elliptic lters
to operate the splitting.
The splitting of the resolved eld into two parts can be interpreted as
the denition of a more complex accentuation technique (see Sect. 5.3.3, p.
156). In the parlance of Hughes, the ltered Smagorinsky model corresponds
the the Small-Large model (while the classical Smagorinsky model is the
Large-Large model).
The accuracy of the results is observed to depend on the spectral properties of the lter used to extract v 1 from u1 . It is observed that kinetic energy
pile-up can occur if a spectral sharp cuto is utilized. The reason why is that
in this case the subgrid viscosity acts only on scales encompassed within v 1
14
It is known from results of Sect. 2.1.6 that this approximation is valid for smooth
symmetric lters and for most discrete lters.
269
and non-local energy transfers between largest scales and v 0 are neglected.
The use of a smooth lter which allows a frequency overlap between u1 and
v 1 alleviates this problem.
The use of the classical Smagorinsky model may also lead to an excessive
damping of scales contained in v 1 . To cure this problem, Holmen et al. [316]
proposed to use the Germano-Lilly procedure to evaluate the constant of the
Large-Small model.
Another formulation for the Variational Multiscale Method is proposed
by Vreman [745] who build some analogous models on the grounds of the test
lter G2 . Keeping in mind that one essential feature of the VMS approach is
that the action of the subgrid scale is restricted to the small scale eld v 1 ,
the following possibilities arise
Model 1: restriction of the usual Smagorinsky model

+
1
1 1
(V II)ij = (Id G2 )
2(CS )2 2|S |S ij
(7.242)
Model 2 : Smagorinsky model based on the small scales

1
(V II)ij = 2(CS )2

1
2|S 1 |Sij
(7.243)
Model 3 : restriction of the Smagorinsky model based on the small scales

1
1
(V II)ij = (Id G2 )
2(CS )2 2|S 1 |Sij
(7.244)
Models 2 and 3 are equivalent if the lter G2 is a sharp cuto lter, but
are dierent in the general case. If a dierential second-order elliptic lter is
used, model 3 will be equivalent to a sixth-order hyperviscosity, while model
2 is associated to a fourth-order hyperviscosity.
Resolvable Sublter-scale Model. Zhou, Brasseur and Juneja [806] developed independently a three-level model which is almost theoretically equivalent to the Variational Multiscale Method of Hughes. The terms taken into
account at the two resolution levels are the same (term III is ignored in
both cases), and term V II is modeled using the Smagorinsky model. The
model used is the smallsmall model (7.238) following Hughes terminology.
As in the VMS implementation described above, simulations of homogeneous
2
1
anisotropic turbulence are carried out with = 2 .
Modied Subgrid-scale Estimation Procedure. Domaradzki et al. [782,
193, 183] proposed a modication of the original subgrid-scale estimation
procedure (see Sect. 7.7.6) in order to improve its robustness. The key point
270
of this method is to account directly for the production of small scales v 1 by

the forward energy cascade rather than using the production estimate (7.222).
The governing equations of this method are the same as those of VMS and
RSFR, with the exception that the subgrid term V II in the detail equation is
neglected. The main dierence with VMS and RSFR is the time integration
procedure: in the two previous approaches the coarsely and nely resolved
elds are advanced at each time step, while Domaradzki and Yee proposed
advancing the ne grid solution u1 over an evolution time T between 1% and
3% of the large-eddy turnover time.
Dynamic Sublter-scale Model. The dynamic sublter-scale model of
Dubrulle et al. [203, 426] appears as a linearized version of the three-level
approach: the nonlinear term V I in the evolution equation of the details is
neglected, and the subgrid term V II is not taken into account. This linearization process renders the detail equation similar to those of the Rapid
Distortion Theory. Only a priori tests have been carried out on parallel wallbounded ows.
Local Galerkin Approximation. The Local Galerkin Approach proposed
by McDonough et al. [468, 466, 467, 470, 471] can be seen as a simplication
of a typical three-level model. All subgrid terms are neglected, and, due to the
fact that the original presentation of the method is not based on a ltering
operator, term IV is not taken into account.
The key idea of the method is to make the assumption that the uctuating
eld v 1 is periodic in space within each cell associated with the coarse level
resolution (see Fig. 7.4). Consequently, a spectral simulation is performed
2
within each cell of size , but the eld v 1 is not continuous at the interface of
1
each cell. The number of Fourier modes determine the cuto length scale .
Fig. 7.4. Schematic of the Local Galerkin Approach.
271
This method can be seen as a dynamic extension of the Kinematic Simulation approach (see Sect. 7.7.4) and is very close, from a practical point of
view, to the PerrierPironneau homogenization technique (see Sect. 7.2.3).
Menons Two-Level-Simulation Method. Menon and Kemenov [382]
developed a two-level method based on a simplied model for the subgrid
scales. Instead of dening a three-dimensional grid to compute the subgrid
modes inside each cell of the large-eddy simulation grid, the authors chose to
solve simplied one-dimensional equations along lines (one in each direction
in practice) inside each cell, leading to a large cost reduction. This feature
makes it reminiscent of the Kerstein subgrid closure based on the stochastic
ODT model (see Sect. 7.7.3). In Menons approach, terms III and V II are
neglected.
In each cell, the three-dimensional subgrid velocity eld is modeled as
a family of one-dimensional velocity vector elds dened on the underlying
family of lines {l1 , l2 , l3 } (which are in practice aligned with the axes of the
reference Cartesian frame). Assuming that the derivatives of the modeled
subgrid velocity eld are such that
v 1
v 1
vi1
i i,
l1
l2
l3
i = 1, 2, 3 ,
(7.245)
and that the incompressibility constraint can be expressed as

1
1
1
+ v3,j
v + v2,j
=0
lj 1,j
(7.246)
1
refers to the jth component of the subgrid eld computed along
where vk,j
the line of index k, the following mometum-like equations are found:
1
1
vi,j
2 vi,j
p1
1
+ N L(vi,j
, u, lj ) =
+ 3
2
t
lj
lj
(7.247)
1
, u, lj ) contains the surrogates for terms
where the non-linear term N L(vi,j
IV , V and V I.
Terracols Multilevel Algorithm. The last three-level model presented in

this section is the multilevel closure proposed by Terracol et al. [710, 633, 711].
It is the only one which considers the full closure problem by taking into
account the non-local interaction term III, and can then be considered as
the most general one. The original method presented in [711] is able to handle
an arbitrary number of ltering levels, but the present presentation will be
restricted to the three-level case.
272
The key points of the method are:

The use of a specic closure at each level u2 and u1 . The proposed model at
the coarse level is an extension of the one-parameter dynamic mixed model
(see Sect. 7.4.2, p. 240), where the scale-similarity part is replaced by the
explicitly computable term II. Term III is modeled using the Smagorinsky part of the mixed dynamic model, with a dynamically computed constant. At the ne resolution level, term V II is parametrized using a oneparameter dynamic mixed model. Numerical results demonstrate that the
use of a specic model for term III is mandatory, since the use of a classical
dynamic Smagorinsky model yields poor results.
The denition of a cycling strategy between the dierent resolution levels,
in order to decrease the computational cost while maintaining the accuracy
of the results. The idea is here to freeze the details v 1 and to carry out the
computation at the coarse level only during a time T .15 The problem is to
nd the optimal T in order to maximize the cost reduction while limiting
the loss of coherence between u2 and v 1 . A simple solution is to advance
the solution for one time step at each level, alternatively, with the same
value of the Courant number at each level. Numerical experiments show
that this solution leads to good results with a gain of about a factor two.
Results obtained using this method on a plane mixing layer conguration
are illustrated in Fig. 7.5.
7.8 Direct Identication of Subgrid Terms

Introduction. This section is dedicated to the presentation of approaches
which aim at reconstructing the subgrid terms using direct identication
mathematical tools.
Like all other subgrid models, either of functional or structural types, they
answer the following question: given a ltered velocity eld u, what is the
subgrid acceleration? Subgrid models described previously were all based on
some a priori knowledge of the nature of the interactions between resolved
and subgrid scales, on a description of the lter, or on a structure of the
subgrid scales hypothesized a priori. The models presented in this section
do not require any of this information, and do not rely on any assumptions
about the internal structure of the subgrid modes.
They are based on mathematical tools which are commonly used within
the framework of pattern recognition, and do not really correspond to what
is usually called a model. Using Mosers words, they represent a radical
approach to large-eddy simulation.
15
This part is close to the quasistatic approximation for small scales introduced
within the context of the nonlinear Galerkin method [174].
7.8 Direct Identication of Subgrid Terms
273
Fig. 7.5. Terracols three-level method. Plane mixing layer. Streamwise energy
spectrum during the self-similar phase. Crosses: large-eddy simulation on the coarse
grid only. Other symbols and lines: direct numerical simulation and large-eddy
simulation using the multilevel closure. The vertical dashed lines denote the cuto
wave numbers of the two grids. Courtesy of M. Terracol, ONERA.
Let us consider a scalar-, vector- or tensor-valued variable , which is to

be estimated, and a set of solutions (u, )n , 1 n N . What we are looking
for is an estimation of the value of for any new arbitrary velocity eld. This
problem is equivalent to estimating the following functional
M (u, K(u))
(7.248)
where K(u) can be any arbitrary function based on the solution (gradient,
correlations, ...). The two classes of approaches presented below are:
1. The approach developed by Moser et al., which relies on linear stochastic
estimation (Sect. 7.8.1).
2. The proposal of Sarghini et al. dealing with the use of neural networks
(Sect. 7.8.2).
Both approaches share the same very dicult practical problem: they require the existence of a set of realizations to achieve the identication process
(computation of correlation tensors in the rst case, and training phase of
the neural network in the second case). In other approaches, this systematic
identication process is replaced by the subgrid modeling phase, in which
the modeler plays the role of the identication algorithm. As a consequence,
the potential success of these identication methods depends on trade-o between the increase of computing power and the capability of researchers to
improve typical subgrid models.
274
7.8.1 Linear-Stochastic-Estimation-Based Model

Moser et al. [419, 549, 548, 165, 296, 741, 797] proposed an identication
procedure based on Linear Stochastic Estimation. This approach can be interpreted in several ways. An important point is that it is closely tied to the
denition of large-eddy simulation as an optimal control problem, where the
subgrid model plays the role of a controller. This interpretation is discussed
in Sect. 9.1.4 and will not be repeated here. The linear stochastic estimation
approach can also be seen as the best linear approximation for the subgrid
term in the least-squares sense. Starting from the general formulation (7.248),
the linear estimation is written as
M (E) = + L E T , with
E = (u, K(u))
(7.249)
where the tensorial dimension of L depends on those of , u and K(u). The

linear stochastic estimation procedure leads to the best values of the coefcient of L in the least-squares sense. Considering vectorial unknowns, we
obtain the following spatially non-local estimation

i (x) = i + Lij (x, x )Ej (x )dx ;
(7.250)
the best coecients Lij are computed by solving the following linear problem:

(7.251)
Ei (x )j (x) = Ljk (x, x )Ej (x)Ek (x )dx .
Local estimates can also be dened using one-point correlations instead
of two-point correlations to dene L.
Practial applications of this approach have been carried out in homogeneous turbulence and plane channel ow [549, 548]. Numerical experiments
have shown that both the subgrid stresses and the subgrid energy transfer
must be taken into account to obtain stable and accurate numerical simulations. This means that both ui ij and S ij ij must be recovered. This is
achieved by estimating the subgrid acceleration
i =
ij
xj
(7.252)
as a function of the velocity eld and its gradients

E = (u, u) .
(7.253)
The mean value is computed from the original data set and stored.
In practice, some simplications can be assumed in denitions (7.252) and
(7.253) in parallel shear ows.
7.9 Implicit Structural Models
275
7.8.2 Neural-Network-Based Model

Sarghini et al. [645] proposed estimating the subgrid terms using a multilayer,
feed-forward neural network. Rather than estimating directly the subgrid
acceleration in the momentum equation, the authors decided to decrease the
complexity of the problem by identifying the value of a subgrid-viscosity
coecient, yielding a new dynamic Smagorinsky model. To this end, a threelayer network is employed. The number of neurons in each layer is 15, 12
and 6, respectively, with a single output (the subgrid viscosity). The input
vector is
(7.254)
E = (u, u u ) IR15 .
The output is the subgrid model constant: = CS IR. The training
of the neural network is achieved using data elds originating from a classical large-eddy simulation of the same plane channel ow conguration. The
learning rule used to adjust the weights and the biases of the network is
chosen so as to minimize the summed-squared-error between the output of
the network and the original set of data. Six thousand samples were found
necessary for the training and validation steps. The training was performed
through a backpropagation with weight decay technique in less than 500 iterations.
A priori and a posteriori tests show that the resulting model leads to
stable numerical simulations, whose results are very close to those obtained
using typical subgrid viscosity models. An interesting feature of the model is
that the predicted subgrid viscosity exhibits the correct asymptotic behavior
in the near-wall region (see p. 159).

The last class of structural subgrid models discussed in this chapter is the implicit structural model family. These models are structural ones, i.e. they do
not rely on any foreknowledge about the nature of the interactions between
the resolved scales and the subgrid scales. They can be classied as implicit,
because they can be interpreted as improvements of basic numerical methods
for solving the ltered NavierStokes equations, leading to the denition of
higher-order accurate numerical uxes. We note that, because the modication of the numerical method can be isolated as a new source term in the
momentum equation, these models could also be classied as exotic formal
expansion models. A major specicity of these models is that they all aim
at reproducing directly the subgrid force appearing in the momentum equation, and not the subgrid tensor . They dier from the stabilized numerical
methods presented in Sect. 5.3.4 within the MILES framework because they
are not designed to induce numerical dissipation.
276
The two models presented in the following are:

1. The Local Average Method of Denaro (p. 276), which consists in a particular reconstruction of the discretized non-linear uxes associated to
the convection term. This approach incorporates a strategy to lter the
subgrid-scale by means of an integration over a control volume and to recover the contribution of the subgrid scales with an integral formulation.
It can be interpreted as a high-order space-time reconstruction procedure
for the convective numerical uxes based on a deltering process.
2. The Scale Residual Model of Maurer and Fey (p. 278). As for the Approximate Deconvolution Procedure, the purpose is to evaluate the commutation error which denes the subgrid term. This evaluation is carried
out using the residual between the time evolution of the solutions of the
NavierStokes equations on two dierent grids (i.e. at two dierent ltering levels) and assuming some self-similarity properties of this residual.
This model can be considered as: (i) a generalization of the previous one,
which does not involve the deconvolution process anymore, but requires
the use of the second computational grid and (ii) a generalization of the
scale-similarity models, the use of a test lter for dening the test eld
being replaced by the explicit computation (by solving the NavierStokes
equations) of the eld at the test lter level.
Other implicit approaches for large-eddy simulation exist, which make it
possible to obtain reliable results without subgrid scale model (in the common sense given to that term), and without explicit addition of numerical
diusion16 . An example is the Spectro-Consistent Discretization proposed by
Verstappen and Veldman [730, 729]. Because these approaches rely on numerical considerations only, they escape the modeling concept and will not
be presented here.
7.9.1 Local Average Method
An other approach to the traditional large-eddy simulation technique was
proposed by Denaro and his co-workers in a serie of papers [175, 169, 170].
It is based on a space-time high-order accurate reconstruction/deconvolution
of the convective uxes, which account for the subgrid-scale contribution.
As a consequence, it can be seen as a particular numerical scheme based on
a dierential approximation of the ltering process. For sake of simplicity,
we will present the method in the case of a dummy variable advected by
a velocity eld u, whose evolution equation is (only convective terms are
retained):
= (u) = A(u, ) .
(7.255)
t
16
Dissipative numerical methods should be classied as Implicit Functional Modeling.
277
The local average of in a ltering cell is dened as the mean value of

in this cell17 :

1
(x, t)
(, t)d = (t), x ,
(7.256)
V
where V is the measure of . We now consider an arbitrary ltering cell.
Applying this operator to equation (7.255), and integrating the resulting
evolution equation over the time interval [t, t + t], we get:
t+t
n u(, t )ddt ,
(7.257)
((t + t) (t))V =
t
where is the boundary of , and n the vector normal to it. The righthand side of this equation, which appears as the application of a time-box
lter to the boundary uxes, can be approximated by means of a dierential
operator, exactly in the same way as for the space-box lter (see Sect. 7.2.1),
yielding:
t+t

tl1 l
(, t)d.
n u(, t )ddt t
n u Id +
l! tl
t
l=1,
(7.258)
The time expansion is then writen as a space dierential operator using
the balance equation (7.255):
tl1
tl1 l
(, t) = Id +
Id +
Al1 (u, ) (, t) ,
l! tl
l!
l=1,
l=1,
(7.259)
with
Al (u, ) A(u, ) A(u, ) ... A(u, )

l times
The second step of this method consists in the reconstruction step. At

each point x located inside the ltering cell , we have
(x, t) =
(x, t + t) =
17
(t) + (x, t) ,
(t + t) + (x, t + t)
(t) + ((t + t) (t)) + (x, t + t)

+( (x, t + t) (x, t))
(x, t) + ((t + t) (t))

+( (x, t + t) (x, t)) .
(7.260)
(7.261)
This ltering operator corresponds to a modication of the box lter dened in

Sect. 2.1.5: the original box lter is dened as a IR IR operator, while the local
average is a IR IN operator. It is worth noting that the local average operator
is a projector.
278
The rst term in the left-hand side of relation (7.261) is known. The
second one, which corresponds to the contribution of the low frequency part of
the solution (i.e. the local averaged part), is computed using equation (7.257).
The third term remains to be evaluated. This is done using the dierential
operator (2.52), leading to the nal expression:
(x, t + t) = (x, t) + (Id Pd ) ((t + t) (t))
with
l
11

Pd =
(i xci )
d
l! V
xi
(7.262)
l=1,
(7.263)
i=1,d
where d is the dimension of space and xci the ith coordinate of the center of
the ltering cell. In practice, the serie expansions are truncated to a nite
order. The repeated use of equation (7.262) makes it possible to compute the
value of the new pointwise value at each time step.
7.9.2 Scale Residual Model
Maurer and Fey [500] propose to evaluate the full subgrid term, still dened
as the commutation error between the NavierStokes operator and the lter
(see Chap. 3 or equation (7.2)), by means of a two-grid level procedure. A deconvolution procedure is no longer needed, but some self-invariance properties
of the subgrid term have to be assumed. First we note that a subgrid model,
referred to as m(u), is dened in order to minimize the residual E, with
E = [N S, G
](u) m(G
u) .
(7.264)
Assuming that the lter G has the two following properties:

G is a projector,
G commutes with the NavierStokes operator in the sense that
N S (G
)u = N S (G
) (G
)u = (G
) N S (G
)u
the residual can be rewritten as

E = (G
) (N S Id N S (G
))u m (G
)u
(7.265)
We now introduce a set of lter Gk , k = 0, N , whose characteristic lengths

k are such that 0 = N < N 1 < .... < 0 . The residual Ek obtained for
the kth level of ltering is easily deduced from relation (7.265):
Ek
= (Gk
) (N S (GN
) N S (Gk
))u
m (Gk
)u
(7.266)
= (Gk
)
279
(N S (Gj
) N S (Gj+1
))u
j=0,k1
m (Gk
)u
(7.267)
To construct the model m, we now make the two following assumptions:

The interactions between spectral bands are local, in that sense that the
inuence of each spectral band gets smaller with decreasing values of j < k.
The residuals between two ltering levels have the following self-invariance
property:
(N S (Gj+1
)N S (Gj
))u = (N S (Gj
)N S (Gj1
))u, (7.268)
where < 1 is a constant parameter. It is important noting that this can
only be true if the cutos occur in the inertial range of the spectrum (see
the discussion about the validity of the dynamic procedure in Sect. 5.3.3).
Using these hypotheses, the following model is derived:
m (Gk
)u = Gk
(
j )(N S (Gk
) N S (Gk+1
))u , (7.269)
j=1,k
where the operator (Gk

) N S (Gk+1
) corresponds to a local reconstruction of the evolution of the coarse solution Gk+1
u according to
the uctuations of the ne solution Gk
u. The implementation of the
model is carried out as follows: a short history of both the coarse and the
ne solutions are computed on two dierent computational grids, and the
model (7.269) is computed and added as a source term into the momentum equations solved on the ne grid. This algorithm can be written in
the following symbolic form:

= N S 2k,t + (N S 2k,t N S 1k+1,2t ) un+1
un+1
k
k
(7.270)
where un+1
designates the solution on the ne grid (kth ltering level) at
k
the (n + 1)th time step, (N S nk,t refers to n applications of the discretized
NavierStokes operator on the grid associated to the ltering level k with
a time step t (i.e. the computation of n time steps on that grid without
any subgrid model), and is a parameter deduced from relation (7.269).
The weight is evaluated analytically through some inertial range consideration, and is assumed to be equal to the ratio of the kinetic energy
contained in the two spectral bands (see equation (5.132)). An additional
correction factor (lower than 1) can also be introduced to account for the
numerical errors.
8. Numerical Solution:
Interpretation and Problems
This chapter is devoted to analyzing certain practical aspects of large-eddy

simulation.
The rst point concerns the dierences between the ltering such as it is
dened by a convolution product and such as it is imposed on the solution
during the computation by the subgrid model. We distinguish here between
static and dynamic interpretations of the ltering process. The analysis is
developed only for subgrid viscosity models because their mathematical form
makes this possible. However, the general ideas resulting from this analysis
can in theory be extended to other types of models. The second point has to
do with the link between the lter cuto length and the mesh cell size used
in the numerical solution. It is important to note that all of the previous
developments proceed in a continuous, non-discrete framework and make no
mention of the spatial discretization used for solving the equations of the
problem numerically. The third point addressed is the comparative analysis
of the numerical error and the subgrid terms. We propose here to compare
the amplitude of the subgrid terms and numerical discretization errors to try
to establish criteria for the required numerical scheme accuracy so that the
errors committed will not overly mar the computed solution.
8.1 Dynamic Interpretation

of the Large-Eddy Simulation
8.1.1 Static and Dynamic Interpretations: Eective Filter
The approach that has been followed so far in explaining large-eddy simulation consists in ltering the momentum equations explicitly, decomposing the
non-linear terms that appear, and then modeling the unknown terms. If the
subgrid model is well designed (in a sense dened in the following chapter),
then the energy spectrum of the computed solution, for an exact solution
verifying the Kolmogorov spectrum, is of the form
2 (k)
E(k) = K0 2/3 k 5/3 G
(8.1)
282
8. Numerical Solution: Interpretation and Problems

where G(k)
is the transfer function associated with the lter. This is the
classical approach corresponding to a static and explicit view of the ltering
process.
An alternate approach is proposed by Mason et al. [495, 496, 497], who
rst point out that the subgrid viscosity models use an intrinsic length scale
denoted f , which can be interpreted as the mixing length associated with
the subgrid scales. A subgrid viscosity model based on the large scales is
written thus (see Sect. 5.3.2):
sgs = 2f |S| .
(8.2)
The ratio between this mixing length and the lter cuto length is:
f
= Cs
(8.3)
Referring to the results explained in the section on subgrid viscosity models, Cs can be recognized as the subgrid model constant. Varying this constant
is therefore equivalent to modifying the ratio between the lter cuto length
and the length scale included in the model. These two scales can consequently
be considered as independent. Also, during the simulation, the subgrid scales
are represented only by the subgrid models which, by their eects, impose the
lter on the computed solution1 . But since the subgrid models are not perfect, going from the exact solution to the computed one does not correspond
to the application of the desired theoretical lter. This switch is ensured by
applying an implicit lter, which is intrinsically contained in each subgrid
model. Here we have a dynamic, implicit concept of the ltering process that
takes the modeling errors into account. The question then arises of the qualication of the lters associated with the dierent subgrid models, both for
their form and for their cuto length.
The discrete dynamical system represented by the numerical simulation
is therefore subjected to two ltering operations:
The rst is imposed by the choice of a level of representation of the physical
system and is represented by application of a lter using the NavierStokes
equations in the form of a convolution product.
The second is induced by the existence of an intrinsic cuto length in the
subgrid model to be used.
In order to represent the sum of these two ltering processes, we dene
the eective lter, which is the lter actually seen by the dynamical system.
To qualify this lter, we therefore raise the problem of knowing what is the
share of each of the two ltering operations mentioned above.
1
They do so by a dissipation of the resolved kinetic energy ij S ij equal to the

ux * through the cuto located at the desired wave number.
8.1 Dynamic Interpretation of the Large-Eddy Simulation
283
8.1.2 Theoretical Analysis of the Turbulence

Generated by Large-Eddy Simulation
We rst go into the analysis of the lter associated with a subgrid viscosity
model.
This section resumes Muschinskys [551] analysis of the properties of a homogeneous turbulence simulated by a Smagorinsky model. The analysis proceeds by establishing an analogy between the large-eddy simulation equations
incorporating a subgrid viscosity model and those that describe the motions
of a non-Newtonian uid. The properties of the latter are studied in the
framework of isotropic homogeneous turbulence, so as to bring out the role
of the dierent subgrid model parameters.
Analogy with Generalized Newtonian Fluids. Smagorinsky Fluid.
The constitutive equations of large-eddy simulation for a Newtonian uid, at
least in the case where a subgrid viscosity model is used, can be interpreted
dierently as being those that describe the dynamics of a non-Newtonian
uid of the generalized Newtonian type, in the framework of direct numerical
simulation, for which the constitutive equation is expressed
ij = pij + sgs Sij
(8.4)
where ij is the stress tensor, S the strain rate tensor dened as above, and
sgs will be a function of the invariants of S. Eects stemming from the
molecular viscosity are ignored because this is a canonical analysis using the
idea of an inertial range. It should be noted that the ltering bar symbol no
longer appears because we now interpret the simulation as a direct one of
a uid having a non-linear constitutive equation. If the Smagorinsky model
is used, i.e.

2
(8.5)
sgs = 2f |S| = Cs |S| ,
such a uid will be called a Smagorinsky uid.
Laws of Similarity of the Smagorinsky Fluid. The rst step consists in
extending the Kolmogorov similarity hypotheses (recalled in Appendix A):
1. First similarity hypothesis. E(k) depends only on , f and .
2. Second similarity hypothesis. E(k) depends only on and for wave
numbers k much greater than 1/f .
3. Third similarity hypothesis. E(k) depends only on and f if f .
The spectrum can then be put in the form:
E(k) = 2/3 k 5/3 Gs (1 , 2 ) ,
(8.6)
where Gs is a dimensionless function whose two arguments are dened as:

1 = kf ,
2 =
1
=
f
Cs
(8.7)
284
By analogy, the limit in the inertial range of Gs is a quantity equivalent

to the Kolmogorov constant for large-eddy simulation, denoted Kles (Cs ):
Kles (Cs ) = Gs (0, 2 ) .
(8.8)
By introducing the shape function

fles (kf , Cs ) =
Gs (1 , 2 )
Gs (0, 2 )
(8.9)
the spectrum is expressed:

E(k) = Kles (Cs )2/3 k 5/3 fles (kf , Cs ) .
(8.10)
By analogy with Kolmogorovs work, we dene the dissipation scale of

the non-Newtonian uid les as:
1/4

3
sgs
les =
.
(8.11)
For the Smagorinsky model, by replacing with its value, we get:

les = f = Cs .
(8.12)
Using this denition and postulating that Kolmogorovs similarity theory

for the usual turbulence remains valid, the third similarity hypothesis stated
implies, for large values of the constant Cs :

2/3 5/3
E(k) =
lim Kles (Cs ) k
,
(8.13)
lim fles (kles , Cs )
Cs
Cs
which allows us to presume that the two following relations are valid:
lim Kles (Cs ) = K0
Cs
lim fles (x, Cs ) = f (x)
Cs
(8.14)
,
(8.15)
where f (x) is the damping function including the small scale viscous eects,
for which the HeisenbergChandresekhar, Kovazsnay, and Pao models have
already been discussed in Sect. 5.3.2.
The corresponding normalized spectrum of the dissipation2 is of the form:
gles (x, Cs ) = x1/3 fles (x, Cs ) ,
(8.16)
where x is the reduced variable x = kles .

2
The dissipation spectrum, denoted D(k), associated with the energy spectrum
E(k) is dened by the relation:
D(k) = k2 E(k)
285
By comparing the dissipation computed by integrating this spectrum with

the one evaluated from the energy spectrum (8.10), the dependency of the
Kolmogorov constant as a function of the Smagorinsky constant is formulated
as:
1
1
.
(8.17)
! Cs
Kles (Cs ) = !
2 0 gles (x, Cs )
2 0 gles (x, Cs )
When this expression is computed using the formulas of Heisenberg
Chandrasekhar and Pao, it shows that the function Kles does tend asymptotically to the value K0 = 1.5 for large values values of Cs , as the error is negligible beyond Cs = 0.5. The variation of the parameter Kles as
a function of Cs for the spectra of HeisenbergChandrasekhar and Pao is
presented in Fig. 8.1. When Cs is less than 0.5, the Kolmogorov constant is
over-evaluated, as has actually been observed in the course of numerical experiments [478, 479]. These numerical simulations, carried out by Magnient
et al. [479], have shown that:
The damping function depends on Cs . A clear bifurcation is observed in
the behavior of the models. The theoretical value of Cs , referred to as
Cs0 , obtained by the canonical analysis corresponds to the case where the
resolved kinetic energy transfer is equal to the energy transfer across the
wave number /. In this case, we obtain fles = 1 for all subgrid viscosity
models. For larger values, the resulting damping function is not equal to
a Heaviside function, and depends on the subgrid model. An interesting
feature is that scales larger than are progressively damped. This damping
Fig. 8.1. Variation of the Kolmogorov constant as a function of the Smagorinsky

constant for the HeisenbergChandrasekhar spectrum and the Pao spectrum.
286
originates from two dierent phenomena: (i) the energy drain induced by
the subgrid-viscosity model, and (ii) the forward energy cascade, which is
responsible for a net drain of kinetic energy by the modes located within
the spectral band [(Cs /Cs0 )/, /].
The damping function is subgrid-model dependent: each subgrid model
leads to a dierent equilibrium between the two sources of resolved energy
drain, yielding dierent spectra and associated damping functions.
Interpretation of Simulation Parameters.
Eective Filter. The above results allow us to rene the analysis concerning
the eective lter. For large values of the Smagorinsky constant (Cs 0.5),
the characteristic cuto length is the mixing length produced by the model.
The model then dissipates more energy than if it were actually located at the
scale because it ensures the energy ux balance through the cuto associated with a longer characteristic length. The eective lter is therefore fully
determined by the subgrid model. This solution criterion should be compared
with the one dened for hot-wire measurements, which recommends that the
wire length be less than twice the Kolmogorov scale in developed turbulent
ows.
For small values of the constant, it is the cuto length that plays the
role of characteristic length and the eective lter corresponds to the usual
analytical lter. It should be noted in this case that the energy drainage
induced by the model is less than the transfer of kinetic energy through the
cuto, so the energy balance is no longer maintained. This is reected in
an accumulation of energy in the resolved scales, and the pertinence of the
simulation results should be taken with caution.
For intermediate values of the constant, i.e. values close to the theoretical
one predicted in Sect. 5.3.2 (i.e. Cs 0.2), the eective lter is a combination
of the analytical lter and models implicit lter, which makes it dicult to
interpret the dynamics of the smallest resolved scales. The dissipation induced
by the model in this case correctly insures the equilibrium of the energy uxes
through the cuto.
Microstructure Knudsen Number. It has already been seen (relation (8.12))
that the mixing length can be interpreted as playing a role analogous to
that of the Kolmogorov scale for the direct numerical simulation. The cuto
length , for its part, can be linked to the mean free path for Newtonian
uids. We can use the ratio of these two quantities to dene an equivalent of
the microstucture Knudsen number Knm for the large-eddy simulation:
Knm =
1
=
f
Cs
(8.18)
Eective Reynolds Number. Let us also note that the eective Reynolds number of the simulation, denoted Reles , which measures the ratio of the inertia
287
eects to the dissipation eects, is taken in ratio to the Reynolds number Re

corresponding to the exact solution by the relation:

Reles =
les
4/3
Re ,
(8.19)
where is the dissipative scale of the full solution. This decrease in the
eective Reynolds number in the simulation may pose some problems, if the
physical mechanism determining the dynamics of the resolved scales depends
explicitly on it. This will, for example, be the case for all ows where critical
Reynolds numbers can be dened for which bifurcations in the solution are
associated3 .
Sublter Scale Concept. By analysis of the decoupling between the cuto
length of the analytical lter and the mixing length f , we can dene
three families of scales [495, 551] instead of the usual two families of resolved
and subgrid scales. These three categories, illustrated in Fig. 8.2, are the:
1. Subgrid scales, which are those that are excluded from the solution by
the analytical lter.
2. Sublter scales, which are those of a size less than the eective lter cuto
length, denoted e , which are scales resolved in the usual sense but
whose dynamics is strongly aected by the subgrid model. Such scales
exist only if the eective lter is determined by the subgrid viscosity
model. There is still the problem of evaluating e , and depends both
on the presumed shape of the spectrum and on the point beyond which we
consider to be strongly aected. For example, by using Paos spectrum
and dening the non-physically resolved modes as those for which the
energy level is reduced by a factor e = 2.7181..., we get:
e =
Cs
,
Ctheo
(8.20)
where Ctheo is the theoretical value of the constant that corresponds to

the cuto length .
3. Physically resolved scales, which are those of a size greater than the
eective lter cuto length, whose dynamics is perfectly captured by the
simulation, as in the case of direct numerical simulations.
Characterization of the Filter Associated with the Subgrid Model.
The above discussion is based on a similarity hypothesis between the properties of isotropic homogeneous turbulence and those of the ow simulated
3
Numerical experiments show that too strong a dissipation induced by the subgrid
model in such ows may inhibit the ow driving mechanisms and consequently
lead to unreliabable simulations. One known example is the use of a Smagorinsky model to simulate a plane channel ow: the dissipation is strong enough to
prevent the transition to turbulence.
288
Fig. 8.2. Representation of dierent scale families in the cases of e < (Right)
and e > (Left).
using a subgrid viscosity model. This is mainly true of the dissipative eects,
which are described using the Pao spectrum or that of HeisenbergKovazsnay.
So here, we adopt the hypothesis that the subgrid dissipation acts like an ordinary dissipation (which was already partly assumed by using a subgrid
viscosity model). The spectrum E(k) of the solution from the simulation can
therefore be interpreted as the product of the spectrum of the exact solution
Etot (k) by the square of the transfer function associated with the eective
e (k):
lter G
2e (k) .
(8.21)
E(k) = Etot (k)G
Considering that the exact solution corresponds to the Kolmogorov spectrum, and using the form (8.10), we get:
'
e (k) = Kles (Cs ) fles (kf , Cs ) .
G
(8.22)
K0
The lter associated with the Smagorinsky model is therefore a smooth
lter in the spectral space, which corresponds to a gradual damping, very
dierent from the sharp cuto lter.
8.2 Ties Between the Filter and Computational Grid.

Pre-ltering
The above developments completely ignore the computational grid used for
solving the constitutive equations of the large-eddy simulation numerically.
If we consider this new element, it introduces another space scale: the spatial
discretization step x for simulations in the physical space, and the maximum
wave number kmax for simulations based on spectral methods.
8.2 Ties Between the Filter and Computational Grid. Pre-ltering
289
The discretization step has to be small enough to be able to correctly

integrate the convolution product that denes the analytical ltering. For
lters with fastly-decaying kernel, we have the relation:
x .
(8.23)
The case where x = is the optimal case as concerns the number of degrees of freedom needed in the discrete system for performing the simulation.
This case is illustrated in Fig. 8.3.
Fig. 8.3. Representation of spectral decompositions associated with pre-ltering

(Left) and in the optimal case (Right).
The numerical errors stemming from the resolution of the discretized system still have to be evaluated. To ensure the quality of the results, the numerical error committed on the physically resolved modes has to be negligible,
and therefore committed only on the sublter scales. The theoretical analysis
of this error by Ghosal in the simple case of isotropic homogeneous turbulence
is presented in the following.
As the numerical schemes used are consistent, the discretization error cancels out as the space and time steps tend toward zero. One way of minimizing
the eect of numerical error is to reduce the grid spacing while maintaining
the lter cuto length, which comes down to increasing the ratio /x (see
Fig. 8.3). This technique, based on the decoupling of the two space scales, is
called pre-ltering [14], and aims to ensure the convergence of the solution
regardless of the grid4 . It minimizes the numerical error but induces more
computations because it increases the number of degree of freedoms in the
numerical solution without increasing the number of degrees of freedom in
the physically resolved solution, and requires that the analytical ltering be
4
A simplied analysis shows that, for an nth-order accurate numerical method,

the weight of the numerical error theoretically decreases as (/x)n . A ner
estimate is given in the remainder of this chapter.
290
performed explicitly [14] [59]. Because of its cost5 , this solution is rarely used
in practice.
Another approach is to link the analytical lter to the computed grid. The
analytical cuto length is associated with the space step using the optimal
ratio of these quantities and the form of the convolution kernel is associated
with the numerical method. Let us point out a problem here that is analogous
to that of the eective lter already mentioned: the eective numerical lter
and therefore the eective numerical cuto length, are generally unknown.
This method has the advantage of reducing the size of the system as best
possible and not requiring the use of an analytical lter, but it allows no
explicit control of the eective numerical lter, which makes it dicult to
calibrate the subgrid models. This method, because of its simplicity, is used
by nearly all the authors.
8.3 Numerical Errors and Subgrid Terms

8.3.1 Ghosals General Analysis
Ghosal [259] proposes a non-linear analysis of the numerical error in the solution of the NavierStokes equations for an isotropic homogeneous turbulent
ow whose energy spectrum is approximated by the Von Karman model.
Classication of Dierent Sources of Error. In order to analyze and
estimate the discretization error, we rst need a precise denition of it. In all
of the following, we consider a uniform Cartesian grid of N 3 points, which
are the degrees of freedom of the numerical solution. Periodicity conditions
are used on the domain boundaries.
A rst source of error stems from the approximation we make of a continuous solution u by a making a discrete solution ud with a set of N 3 values.
This is evaluated as:
|ud P(u)|
(8.24)
where P is a denite projection operator of the space of continuous solutions to that of the discrete solutions. This error is minimum (in the L2
sense) if P is associated with the decomposition of the continuous solution
on a nite base of trigonometric polynomials, with the components of ud
being the associated Fourier coecients. This error is intrinsic and cannot be
canceled. Consequently, it will not enter into the denition of the numerical
5
For a xed value of , increasing the ratio /x by a factor n leads to an

increase in the number of points of the simulation by a factor of n3 and increases
the number of time steps by a factor n in order to maintain the same ratio
between the time and space steps. In all, this makes an overall increase in the
cost of the simulation by a factor n4 .
291
error discussed in this present section. The best possible discrete solution is
uopt P(u).
The equations of the continuous problem are written in the symbolic form:
u
= N S(u) ,
t
(8.25)
where N S is the NavierStokes operator. The optimal discrete solution uopt

is a solution of the problem:
Pu
= P N S(u) ,
t
(8.26)
where P N S is the optimal discrete NavierStokes operator which, in the

xed framework, corresponds to the discrete operators obtained by a spectral
method.
Also, we note the discrete problem associated with a xed discrete scheme
as:
ud
= N S d (ud ) .
(8.27)
t
By taking the dierence between (8.26) and (8.27), it appears that the
best possible numerical method, denoted N S opt , is the one that veries the
relation:
(8.28)
N S opt P = P N S .
The numerical error Enum associated with the N S d scheme, and which is
analyzed in the following, is dened as:
Enum (P N S N S d P) (u) .
(8.29)
This represents the discrepancy between the numerical solution and the
optimal discrete one. To simplify the analysis, we consider in the following
that the subgrid models are perfect, i.e. that they induce no error with respect
to the exact solution of the ltered problem. By assuming this, we can clearly
separate the numerical errors from the modeling errors.
The numerical error Enum (k) associated with the wave number k is decomposed as the sum of two terms of distinct origins:
The dierentiation error Edf (k), which measures the error the discrete operators make in evaluating the derivatives of the wave associated with k.
Let us note that this error is null for a spectral method if the cuto frequency of the numerical approximation is high enough.
The spectrum aliasing error Ers (k), which is due to the fact that we are
computing non-linear terms in the physical space in a discrete space of nite
dimension. For example, a quadratic term will bring in higher frequencies
than those of each of the arguments in the product. While some of these
frequencies are too high to be represented directly on the discrete base,
292
they do combine with the low frequencies and introduce an error in the
representation of them6 .
Estimations of the Error Terms. For a solution whose spectrum is of the
form proposed by Von Karman:
E(k) =
a k4
17/6
(b + k 2 )
(8.30)
with a = 2.682 and b = 0.417, and using a quasi-normality hypothesis for

evaluating certain non-linear terms, Ghosal proposes a quantitative evaluation of the dierent error terms, the subgrid terms, and the convection term,
for various Finite Dierence schemes as well as for a spectral scheme. The
convection term is written in conservative form and all the schemes in space
are centered. The time integration is assumed to be exact.
The exact forms of these terms, available in the original reference work,
are not reproduced here. For a cuto wave number kc and a sharp cuto lter,
simplied approximate estimates of the average amplitude can be derived for
some of these terms.
The amplitude of the subgrid term sgs (kc ), dened as
(
kc
sgs (kc ) =
)1/2
| (k)|dk
(8.31)
where (k) is the subgrid term for the wave number k, is bounded by:

0.36 kc0.39 upper limit
sgs (kc ) =
,
(8.32)
0.62 kc0.48 lower limit
that of the sum of the convection term and subgrid term by:
tot (kc ) = 1.04 kc0.97
6
(8.33)
Let us take the Fourier expansions of two discrete functions u and v represented
by N degrees of freedom. At the point of subscript j, the expansions are expressed:

N/21
uj =
N/21
u
n e(i(2/N)jn) , vj =
n=N/2
vm e(i(2/N)jm) j = 1, N
m=N/2
The Fourier coecient of the product wj = uj vj (without summing on j) splits

into the form:

wk =
u
n vm +
u
n vm .
n+m=k
n+m=kN
The last term in the right-hand side represents the spectrum aliasing error. These
are terms of frequencies higher than the Nyquist frequency, associated the sampling, which will generate spurious resolved frequencies.
293
in which
sgs (kc )
kc0.5 .
tot (kc )
The amplitude of the dierentiation error df (kc ), dened by:
(
)1/2
(8.34)
kc
df (kc ) =
Edf (k)dk
(8.35)
is evaluated as:
1.03 (second order)
0.82 (fourth order)

0.70 (sixth order)
df (kc ) = kc0.75
0.5 (heigth order)
0
(spectral)
(8.36)
and the spectrum aliasing error rs (kc ), which is equal to:

(
)1/2
kc
rs (kc ) =
Ers (k)dk
(8.37)
is estimated as:
0.90
2.20
rs =
0.46
1.29
kc0.46
kc0.66
kc0.41
kc0.65
(minimum estimation, spectral, no de-aliasing)

(maximum estimation, spectral, no de-aliasing)
.
(minimum estimation, second order)
(maximum estimation, second order)
(8.38)
The spectrum aliasing error for the spectral method can be reduced to
zero by using the 2/3 rule, which consists of not considering the last third
of the wave numbers represented by the discrete solution. It should be noted
that, in this case, only the rst two-thirds of the modes of the solution are
correctly represented numerically. The error of the nite dierence schemes
of higher order is intermediate between that of the second-order accurate
scheme and that of the spectral scheme.
From these estimations, we can see that the discretization error dominates the subgrid terms for all the nite dierence schemes considered. The
same is true for the spectrum aliasing error, including for the nite dierence schemes. Finer analysis on the basis of the spectra of the various terms
shows that the discretization error is dominant for all wave numbers for the
second-order accurate scheme, whereas the subgrid terms are dominant at
the low frequencies for the heigth-order accurate scheme. In the former case,
the eective numerical lter is dominant and governs the solution dynamics.
Its cuto length can be considered as being of the order of the size of the
computational domain. In the latter, its cuto length, dened as the wavelength of the mode beyond which it becomes dominant with respect to the
subgrid terms, is smaller and there exist numerically well-resolved scales.
294
8.3.2 Pre-ltering Eect

The pre-ltering eect is clearly visible from relations (8.32) to (8.38). By
decoupling the analytical from the numerical lter, two dierent cuto scales
are introduced and thereby two dierent wave numbers for evaluating the
numerical error terms and the subgrid terms: while the cuto scale associated with the lter remains constant, the scale associated with the numerical
error (i.e. x) is now variable.
By designating the ratio of the two cuto lengths by Crap = x/ < 1,
we see that the dierentiation error df (kc ) of the nite dierence scheme is
3/4
reduced by a factor Crap with respect to the previous case, since it varies
3/4
as kc . This reduction is much greater than the one obtained by increasing
the order of accuracy of the the schemes.
Thus, more detailed analysis shows that, for the second-order accurate
scheme, the dominance of the subgrid term on the whole of the solution
spectrum is ensured for Crap = 1/8. For a ratio of 1/2, this dominance is
once again found for schemes of order of accuracy of 4 or more.
These theoretical evaluations do not take into account the nonlinear feedback of the computed solution on the numerical error. Numerical experiments
were conducted by Chow and Moin [138] in isotropic turbulence to assess
Ghosals results. Their results show that to ensure that the subgrid terms
will dominate the numerical error, a lter-grid ratio /x of at least four is
desired for a second-order centered nite dierence scheme. This minimum
is a decreasing function of the scheme order of accuracy, and a ratio of two
is found to be sucient for a sixth-order centered Pade scheme.
The eciency of the preltering technique was exhaustively checked by
Geurts and Fr
ohlich [256, 257] on the plane mixing layer conguration. The
main conclusion of this study is that the best solution for improving the results of a large-eddy simulation in practice is to rene the computational grid,
i.e. to lower x, while keeping a low value of the ratio /x. The numerical
tests conducted by these authors show that /x = 1 is optimal for highresolution large-eddy simulations, i.e. for simulations which are close to direct
numerical simulation. For coarser grid simulation, the optimum was found to
be /x 2 3, combined with a fourth-order accurate non-dissipative
numerical method. Nevertheless, the recommended strategy to improve the
results at a xed computational cost is to rene the grid using /x = 12
rather than augmenting the ratio /x for a given value of . This conclusion is conrmed by a large set of results published by Gullbrand and
Chow [283], who carried out several simulations of turbulent plane channel
ow with second-order and fourth-order accurate nite dierence methods.
In these simulations, the use of a prelter with size 2x didnt lead to a clear
improvement of the results.
The nest study of the eect of preltering was conducted by Meyers,
Geurts and Baelmans [515] in large-eddy simulation of isotropic turbulence
using the Smagorinsky model. Dening the error etotal (, x) for a dummy
295
variable as the dierence between the ltered exact solution (obtained

by Direct Numerical Simulation) and the solution found using large-eddy
simulation with cuto length on a mesh of size x:
etotal (, x) = DNS LES (, x)
(8.39)
the authors investigate the relative inuence of dierent sources of error.

Subgrid modeling and discretization errors being respectively dened as
emodel = DNS LES (, 0) ,
ediscr = LES (, 0) LES (, x)
(8.40)
,
(8.41)
exhaustive analyses carried out using a direct numerical simulation database

yields the following conclusions:
The global error behavior is parametrized by the subgrid activity parameter s dened like
s=
,
(8.42)
+
where = ij S ij is the subgrid dissipation and = Sij Sij is the
molecular dissipation. Direct numerical simulation corresponds to s = 0,
while s = 1 characterizes large-eddy simulation in the limit of innite
Reynolds number. Tests prove there there exists a threshold value sc 7 : for
s sc the total error is dominated by discretization error eects, while for
s > sc the modeling error is the most important source of uncertainty.
For s sc , the relative error err :
! 2
dt
e
,
(8.43)
err = ! total2
DNS dt
can be either a decreasing or increasing function of s, depending on s and
x when it is based on the kinetic energy (see Fig. 8.4). No general trend
is observed in this regime, in which strong interactions can occur between
discretization errors and modeling errors. Partial cancellation sometimes
occurs leading to a signicant reduction of the total error. For s > sc , the
relative error is a monotone exponentially increasing function of s.
The relative error based on the Taylor length scale has a dierent behavior
in the s sc regime (see Fig. 8.5): strong interactions between modeling and discretization errors are observed, but the relative error exhibits
a monotonic increasing behavior as a function of sc . Partial error cancellation is less intense than in the case of the turbulent kinetic energy.
Because of the strong non-linear interactions between modeling and discretization errors, which leads to a non-monotone behavior of the total
7
Typical values are sc = 0.4 for isotropic turbulence at Re = 50, sc = 0.8 for
isotropic turbulence at Re = 100 and sc = 0.5 for a plane mixing layer.
296
Fig. 8.4. Evolution of the relative error on kinetic energy as a function of the
subgrid activity, in isotropic turbulence (Courtesy of J. Meyers and B. Geurts,
Univ. Twente). Top: Re = 50, Bottom: Re = 100.
297
Fig. 8.5. Evolution of the relative error on the Taylor scale as a function of the
subgrid activity, in isotropic turbulence (Courtesy of J. Meyers and B. Geurts, Univ.
Twente). Left: Re = 50, Right: Re = 100.
error. As a consequence, grid renement (decrease of x at constant )

must be coupled to a change in the Smagorinsky constant CS to obtain
an optimal error minimization. The ratio CS /x which yields the minimum error follows non-trivial trajectories in both the (x, CS ) and the
(x, s/sc ) planes. This behavior is illustrated in Fig. 8.6.
8.3.3 Conclusions
This analysis can be used only for reference, because it is based on very
restrictive hypotheses. It nonetheless indicates that the numerical error is
not negligible and that it can even be dominant in certain cases over the
subgrid terms. The eective numerical lter is then dominant over the scale
separation lter.
This error can be reduced either by increasing the order of accuracy of
the numerical scheme or by using a pre-ltering technique that decouples
the cuto length of the analytical lter of the discretization step. Ghosals
ndings seem to indicate that a combination of these two techniques would
be the most eective solution.
These theoretical ndings are conrmed by the numerical experiments
of Najjar and Tafti [563] and Kravenchko and Moin [409], who observed
that the eect of the subgrid models is completely or partially masked by
the numerical error when second-order accurate methods are employed. It
should be noted here that practical experience leads us to less pessimistic
conclusions than the theoretical analyses: large-eddy simulations performed
with a scheme accurate to the second order show a dependency with respect to
the subgrid model used. The eects of these models are not entirely masked,
which justies using them. However, no precise qualication exists today of
the information loss due to the use of a given scheme. These observations are
made empirically, case by case [76, 480, 94].
298
Fig. 8.6. Top: Map of the relative error in isotropic turbulence (Courtesy of J.
Meyers and B. Geurts, Univ. Twente). Left: Re = 50, Right: Re = 100. The
dashed-dotted line is related to the optimal trajectory corresponding to the minimal
error. The dashed line corresponds to the trajectory associated to a xed value of
the subgrid activity parameter. Bottom: Optimal renement strategy for dierent
Reynolds numbers and error denitions, shown in dierent planes. The shadded
areas are related to minimal value of the error.
Another eect was emphasized by Geurts et al. [255, 515], who found that
partial cancellation of modeling and numerical errors may occur, leading to
a signicant improvement of the accuracy of the simulation. This cancellation was observed in both plane mixing layer conguration and isotropic
turbulence for fourth-order and spectral schemes. An important and counterintuitive conclusion is that using higher-order accurate schemes or improved
subgrid models may lead to worse results with regards to a simulation in
which this cancellation occurs. This behavior is observed to be strongly dependent on the value of the subgrid activity parameter s and the considered
denition of the error.
An important nding dealing with the preltering technique is that rening the the grid at constant lter cuto (i.e. /x at xed )
must be coupled to a change in subgrid viscosity model constant to obtain
the optimal error reduction.
299
8.3.4 Remarks on the Use of Articial Dissipations

Many comments have been made over recent decades on the sensitivity of
large-eddy simulation results, for example concerning the formulation of the
convection term [319, 409], the discrete form of the test lter [597, 563, 81,
641, 605], and the formulation of the subgrid term [634], but there are far too
many, too dispersed, and too far from general to be detailed here. Moreover,
countless analyses have been made of the numerical error associated with
various schemes, especially as concerns the treatment of the non-linear terms,
which will not be resumed here, but we will still take more special note of
the ndings of Fabignon et al. [211] concerning the characterization of the
eective numerical lter of several schemes.
Special attention should still be paid to the discretization of the convective terms. To capture strong gradients without having the numerical solution
polluted with spurious high-frequency wiggles, the scheme is very often stabilized by introducing articial dissipation. This dissipation is added explicitly
or implicitly using an upwind scheme for the convection term. Introducing
an additional dissipation term for the large-eddy simulation is still controversial [563, 521, 239, 228, 665, 501, 94] because the eective lter is then very
similar in form to that which would be imposed by subgrid viscosity model,
making for two competing resolved kinetic energy spectrum mechanisms. The
similarity between the numerical dissipation and that associated with the energy cascade model is still being investigated, but a few conclusions have
already been drawn.
It seems that the total numerical dissipation induced by most upwind
schemes is still greater than that of the subgrid viscosity models, if no preltering method is used. This is true even for seventh-order accurate upwind
schemes [52].
Garnier et al. [239] developed the generalized Smagorinsky constant as
a tool to compare numerical and physical subgrid dissipations. The generalized Smagorinsky constant is the value that should take the constant of the
Smagorinsky model to obtain a total dissipation equal to the numerical dissipation. Numerical tests carried out on decaying isotropic turbulence have
shown that all the numerical upwind schemes, up to the fth order of accuracy, are more dissipative than the usual Smagorinsky model. Typical results
are displayed in Fig. 8.7.
It has also been demonstrated [501] that the use of a stabilized sixth-order
accurate scheme may lead to the same quality of results as a second-order
accurate scheme, because of the very high dissipation applied to the highest
resolved frequency, which is responsible for the largest part of the interactions
with subgrid modes. These two dissipations are correlated in space (especially
in the case of the Smagorinsky model), but have dierent spectral distributions: a subgrid viscosity model corresponds to a second-order dissipation
associated with a spectrum of the form (k/kc )2 E(k), while an nth-order numerical dissipation is associated with a spectrum of the form (k/kc )n E(k).
300
Fig. 8.7. Time history of the generalized Smagorinsky constant for various dissipative schemes (second- to fth-order of accuracy) in freely decaying isotropic
turbulence. Courtesy of E. Garnier, ONERA.
For n > 2 (resp. n < 2), the numerical (resp. subgrid) dissipation may be
dominant for the highest resolved frequencies and the subgrid (resp. numerical) dissipation will govern the dynamics of the low frequencies. This point
is illustrated in Figs. 8.8 and 8.9.
Fig. 8.8. Numerical and subgrid dissipations for a Von Karman spectrum. The
peak of the Von Karman spectrum is at kc /5. The dissipation spectra have been
normalized so that the total dissipation is the same in all cases. It is worth noting
that typical numerical schemes lead to a total dissipation higher than the subgrid
dissipation.
301
Fig. 8.9. Turbulent kinetic energy spectra (isotropic turbulence) computed by

large-eddy simulation. Solid line, fth-order accurate dissipative scheme without
subgrid model. Dashed lines, second-order accurate non-dissipative scheme with (i)
Smagorinsky model and (ii) dynamic Smagorinsky model. Courtesy of E. Garnier,
ONERA.
The studies that have been made show a sensitivity of the results to
the subgrid model used, which proves that the eects of the model are not
entirely masked by the numerical dissipation. The theoretical analysis presented above should therefore be taken relative to this. But consistent with it,
Beaudan et al. [52] have observed a reduction in the numerical cuto length
as the order of accuracy of the scheme increases. This type of nding should
nonetheless be treated with caution, because the conclusions may be reversed
if we bring in an additional parameter, which is the grid renement. Certain
studies have shown that, for coarse grids, i.e. high values of the numerical
cuto length, increasing the order of accuracy of the upwind scheme can lead
to a degradation of the results [701]. But some specic numerical stabilization
procedures can be dened, which tune the numerical dissipation in such a way
that the results remain sensitive to subgrid modeling [94, 15, 95]. These numerical methods allow the use of relatively coarse grids, but no general theory
for them exist at present time.
This relative similarity between articial dissipation and the direct energy
cascade model has induced certain authors to perform no-model large-eddy
simulations, with the ltering based entirely on the numerical method (leading to the Implicit Large-Eddy Simulation technique, see Sect. 5.3.4 for a detailed discussion). Thus many ow simulations have been seen in complex
geometries, based on the use of an third-order accurate upwind scheme proposed by Kawamura and Kuwahara [381], yielding interesting results. In the
compressible case, this approach has been called the Monotone Integrated
302
Large-Eddy Simulation (MILES) method (see Sect. 5.3.4 for a description of

ILES numerical methods). The experimental analysis of some particular
numerical methods reveals that there exists some numerical dissipation procedures which mimic very well the theoretical subgrid viscosity [192]. This
is illustrated in Fig. 8.10, which compares the computes spectral viscosity
associated to the MPDATA scheme with the theoretical spectral viscosity
prole. The observed agreement prove that the Implict Large-Eddy Simulation approach can yield very good results if the numerical scheme is carefully
chosen.
The use of articial dissipation therefore raises many questions, but is
very common in simulations that are physically very strongly under-resolved
in complex congurations, because experience shows that adding subgrid
models does not ensure a monotonic solution. To ensure that certain variables
remain positive, such as concentrations of pollutants or the temperature, it
seems to be necessary to resort to such numerical methods. Alternatives based
on local renement of the solution, i.e. decreasing the eective cuto length
by enriching or adapting the grid, have been studied by certain authors but
no nal conclusion has been drawn.
A few studies dealing with some particular stabilization techniques such
as the Galerkin Least-Square method, reveal that the numerical dissipation
may happen to be be too low to prevent a pile-up in the resolved kinetic
energy, leading to the growth of bounded wiggles. In such a case, the use of
a subgrid model seems to be required to recoved physical results. The main
problem arising in this case is that the use of a basic subgrid model will lead
to a global overdamping of the computed solution. The use of self-adaptive
subgrid models able to tune their induced dissipation so that the sum of
Fig. 8.10. Analysis of the numerical viscosity associated to the MPDATA scheme.
Dashed line: theoretical spectral subgrid viscosity. Solid line: MPDATA equivalent spectralk viscosity. Solid line with symbol: spectral viscosity recomputed from
kinetic energy transfer of the unltered velocity eld. From [192].
303
the numerical and subgrid dissipation will reach the required level to obtain
physical results. The identication of subgrid models and/or procedures to
dene such subgrid models is still a open issue. An a priori requirement is
that the subgrid model must be local in terms of wave number, i.e. it must
emphasized the highest resolved frequencies. Numerical experiments show
that self-adaptive models and some multilevel approaches have this property.
While the NavierStokes equations contain energy information, they also
contain information concerning the signal phase. Using centered schemes for
the convection term therefore raises problems too, because of the dispersive
errors they induce in the highest resolved frequencies.
Generally, estimates of the wave number beyond which the modes are
considered to be well resolved numerically vary from 2x to 20x, depending
on the schemes and authors [604].
8.3.5 Remarks Concerning the Time Integration Method
Large-eddy simulation is ordinarily addressed using a spatial ltering, but
without explicitly stating the associated time ltering. This is due to the fact
that most computations are made for moderate time steps (CFL ut/
x < 1) and it is felt that the time ltering eects are masked by those of
the space ltering. Choi and Moin [131], however, have shown by direct simulations of a plane channel ow that the time ltering eects can be very large,
even for CFLs of the order of 0.5, since the turbulence cannot be maintained
numerically if the time step is greater than the characteristic time associated
with the Kolmogorov scale. Most authors use second-order accurate integration methods, but no complete study has been published to date to determine
what timescales are well resolved numerically and physically. We should also
note the results of Beaudan and Moin [52] and Mittal and Moin [521], who
showed that the use of articial viscosity aects the solution of a very large
share of the simulated time frequencies (about 75% for the particular case
studied).
9. Analysis and Validation

of Large-Eddy Simulation Data

9.1.1 Type of Information Contained in a Large-Eddy Simulation
The solution to the equations that dene the large-eddy simulation furnishes
explicit information only on the scales that are resolved, i.e. those that are
left after reduction of the number of degrees of freedom in the exact solution.
We are therefore dealing with information that is truncated in space and
time. The time ltering is induced implicitly by the spatial ltering because,
as the ltering eliminates certain space scales, it eliminates the corresponding
time scales with them (see p. 19).
The information of use for analysis or validation is what is contained in
those scales that are both physically and numerically well-resolved. It should
be remembered that, since the eective numerical and physical lters are
unknown, the usable scales are most often identied empirically.
Adopting the assumption that all the scales represented by the simulation are physically and numerically well-resolved, the statistical average of
the usable resolved eld is expressed u. The statistical uctuation of the
resolved eld, denoted u , is dened by:
ui = ui ui
(9.1)
The dierence between the statistical average of the resolved scales and
that of the exact solution is dened as:
ui ui = ui ,
(9.2)
which corresponds to the statistical average of the unresolved scales. The

Reynolds stresses computed from the resolved scales are equal to ui uj .
e
The dierence from the exact stresses ue
i uj , where the exact uctuation is
e
dened as u = u u, is:
e
ue
i uj = (ui ui )(uj uj )
= ui uj ui uj
= ui uj + ij ui + ui uj + uj
306
= ui uj + ij ui uj ui uj ui uj ui uj
= ui uj + ij ui uj ui uj ui uj
where ij = ui uj + uj ui + ui uj is the deltered subgrid tensor.
Since the subgrid scales are not known, the terms containing the contribution u , cannot be computed from the simulation. When the statistical
average of the subgrid modes is very small compared with the other terms,
we get:
e

(9.3)
ue
i uj ui uj + ij .
The two terms on the right-hand side can be evaluated from the numerical
simulation, but the quality of the models representation of the subgrid tensor
partly conditions that of the result. We can easily see that subgrid-viscosity
models, which only account for the deviatoric part of the subgrid-stress tensor, make it possible only to recover the deviatoric part of the Reynolds
stresses [760, 397], at least theoretically.
9.1.2 Validation Methods
The subgrid models and their various underlying hypotheses can be validated
in two ways [218]:
A priori validation. The exact solution, which is known in this case, is
ltered analytically, leading to the denition of a fully determined resolved eld and subgrid eld. The various hypotheses or models can then
be tested. The exact solutions are usually generated by direct numerical
simulations at moderate or low Reynolds numbers, which limits the eld of
investigation. A priori tests like this have also been performed using experimental data, making it possible to reach higher Reynolds numbers. This
type of validation raises a fundamental problem, though. By comparing
the exact subgrid stresses with those predicted by a subgrid model evaluated on the basis of the ltered exact solution, the eects of the modeling
errors are neglected and the implicit lter associated with the model is not
considered1 . This means that the results of a priori validations are only
relative in value.
A posteriori validation. Here, we perform a large-eddy simulation computation and validate by comparing its results with a reference solution. This
is a dynamic validation that takes all the simulation factors into consideration, while the previous method is static. Experience shows that models yielding poor a priori results can be satisfactory a posteriori, and vice
1
This eld could not have been obtained by a large-eddy simulation since it is
a solution of the ltered momentum equations in which the exact subgrid tensor
appears. In the course of a simulation, the subgrid model is applied to a velocity
eld that is a solution of the momentum equation where the modeled subgrid
tensor appears. These two elds are therefore dierent in theory. Consequently,
in order to be fully representative, an a priori test has to be performed on the
basis of a velocity eld that can be obtained from the subgrid model studied.
307
versa [597]. It is more advantageous to validate models a posteriori because

it corresponds to their use in the simulation; but it is sometimes dicult to
draw any conclusions on a precise point because of the multitude of often
imperfectly controlled factors at play in a numerical simulation.
9.1.3 Statistical Equivalency Classes of Realizations
The subgrid models are statistical models and it seems pointless to expect
them to produce deterministic simulations in which the resolved scales coincide exactly with those of other realizations, for example of the experimental
sort. On the other hand, large-eddy simulation should correctly reproduce
the statistical behavior of the scales making up the resolved eld. Equivalency classes can thus be dened among the realizations [503] by considering
that one of the classes consists of realizations that lead to the same values of
certain statistical quantities computed from the resolved scales.
Belonging to the same class of equivalency as a reference solution is a validation criterion for the other realizations. If we set aside the numerical errors,
we can dene the necessary conditions on the subgrid models such that two
realizations will be equivalent, by verifying these validity criteria. These conditions will be discussed in the following sections. A subgrid model can thus
be considered validated if it can generate realizations that are equivalent to
a reference solution, in a sense dened below.
Theoretically, while we overlook the eect of the discretization on the
modeling, it can be justiably thought that a model reproducing the interscale interactions exactly will produce good results, whereas the opposite
proposition is not true. That is, the idea of sucient complexity of a model
has to be introduced in order to obtain a type of result on a given conguration with a tolerated margin of error in order to say what a good model
is. The idea of a universal or best model might not be rejected outright,
but should be taken relatively. The question is thus raised of knowing what
statistical properties two subgrid models should share in order for the two
resulting solutions to have common properties.
Let u and u be the ltered exact solution and the solution computed
with a subgrid model, respectively, for the same lter. The exact (unmodeled)
subgrid tensor corresponding to u is denoted ij , and the modeled subgrid
tensor computed from the u eld is denoted ij (u ). The two velocity elds
are solutions of the following momentum equations:
u
+ (u u) = p + 2 u
t
(9.4)
u
+ (u u ) = p + 2 u (u ) .
(9.5)
t
A simple analysis shows that, if all the statistical moments (at all points
of space and time) of ij conditioned by the u eld are equal to those of
308
ij (u ) conditioned by u , then all the statistical moments of u and u

will be equal. This is a full statistical equivalency, which implies that the
subgrid models fulll an innity of conditions. To relax this constraint, we
dene less restrictive equivalency classes of solutions which are described
in the following sections. They are dened in such a way as to bring out
the necessary conditions applying to the subgrid models, in order to qualify
them [503]. We try to dene conditions such that the statistical moments of
moderate order2 (1 and 2) of the eld resulting from the large-eddy simulation
u are equal to those of a reference solution u.
Equivalency of First-Order Moments. The equivalency relation is built
on the equality of the rst-order statistical moments of the realizations. A velocity and a pressure eld are associated with each realization. Let (u, p) and
(u , p ) be the doublets associated with the rst and second realizations, respectively. The evolution equations of the rst-order statistical moments of
the velocity eld of these two realizations are expressed:
u
+ (u u)
t
p + 2 u
(u u u u) ,

u
+ (u u ) =
t
(9.6)
p + 2 u (u )
(u u u u ) , (9.7)
where designates an ensemble average performed using independent realizations. The two realizations will be called equivalent if their rst- and
second-order moments are equivalent, i.e.
ui =
p =
ui uj =
ui ,
(9.8)
p ,
ui uj .
(9.9)
(9.10)
Analysis of evolution equations (9.6) and (9.7) shows that one necessary
condition is that the resolved and subgrid stresses be statistically equivalent.
The last condition is expressed:
ij = ij + Cij
(9.11)
where Cij is a null-divergence tensor. This condition is not sucient because

a model that leads to a good prediction of the mean stresses can generate
an error on the mean eld if the mean resolved stresses are not correct. To
obtain a sucient condition, the equivalency of the stresses ui uj and ui uj
must be ensured by another relation.
2
Because these are the quantities sought in practice.
309
Equivalency of Second-Order Moments. We now base the equivalency

relation on the equality of the second-order moments of the resolved scales.
Two realizations will be called equivalent if the following conditions are satised:
ui =
ui uj =
ui uj uk =
pui =
pS ij =
; u u <
i
i
=
xk xk
ui
ui uj ,
ui uj uk ,
p ui ,
p S ij ,
; u u <
i
i
xk xk
(9.12)
(9.13)
(9.14)
(9.15)
(9.16)
.
(9.17)
Analysis of the equation for the second-order moments ui uj shows that,
in order for two realizations to be equivalent, the following necessary condition must be satised:
(ui jk + uj ik ) =
xk

S kj + jk
S ki

.
ik
ui jk + uj ik
xk
ik S kj + jk S ki
This condition is not sucient. To obtain such an realization, the equality

of the third-order moments also has to be ensured. It is noted that the nonlinear coupling prohibits the denition of sucient conditions on the subgrid
model to ensure the equality of the nth-order moments of the resolved eld
without adding necessary conditions on the equality of the (n + 1)th-order
moments.
Equivalency of the Probability Density Functions. We now base the
denition of the equivalency classes on the probability density function
fprob (V , x, t) of the resolved scales. The eld V is the test velocity eld
from which the conditional average is taken. The function fprob is dened as
the statistical average of the one-point probabilities:
fprob (V , x, t) (u(x, t) V )
and is a solution of the following transport equation:

; p
<7
fprob
ij
fprob
2
+Vj
=
+
uj |u = V
fprob
t
xj
Vj
xj
xj
(9.18)
. (9.19)
Two realizations can be called equivalent if:

fprob (V , x, t) =
ui (y)|u = V =
ui (y)uj (y)|u = V =
fprob
(V , x, t) ,
ui (y)|u = V ,
(9.20)
(9.21)
ui uj (y)|u = V .
(9.22)
310
Once the pressure gradient is expressed as a function of the velocity (by

an integral formulation using a Green function) and the conditional average
of the strain rate tensor is expressed using gradients of the two-point conditional averages, equation (9.19) can be used to obtain the following necessary
condition:

2
xj yj 3
1
d y + lim
ik |u(x) = V
ij |u(x) = V
yx yi
4
yi yk
|x y|

1
xj yj 3
2
d y
=
ik
|u (x) = V
4
yi yk
|x y|
ij |u (x) = V + Cj
yx yi
+ lim
in which the divergence of vector Cj is null. It is noted that the condition

dened from the one-point probability density uses two-point probabilities.
We again nd here the problem of non-localness already encountered when
the equivalency class is based on statistical moments. A more restrictive
condition is:
|u (x) = V .
ik |u(x) = V = ik
(9.23)
9.1.4 Ideal LES and Optimal LES

An abstract subgrid model can be dened, which is in all senses ideal [419].
An LES using this model will exactly reproduce all single-time, multipoint
statistics, and at the same time will exhibit minimum possible error in instantaneous dynamics. Such a LES will be referred to as ideal LES. Using
the same notations as in Sect. 9.1.3, ideal LES is governed by the conditional
average

@
A
du
du
=
u
=
u
,
(9.24)
dt
dt
where u and u are the solution of the exact LES equation and the LES
equation with a subgrid model, respectively. It can be shown that such ideal
LES is associated to the minimum mean-square error between the evolution
of the LES eld u (t) and the exact solution u(t), dened as an instantaneous
pointwise measurement on u/t:
ei (x) =
ui
ui
t
t
(9.25)
Equivalently, this error can be evaluated using the exact and the modeled
subgrid forces, referred to as M = and m = :
e(x) = M (x) m(x) .
(9.26)
311
The ideal subgrid model is then such that

= m = M |u = u
(9.27)
This model is written as an average over the real turbulent elds whose
resolved scales match the current LES eld, making it impossible to compute
in practical applications. In order to approximate in an optimal sense this
ideal model, several authors [419, 54, 6, 137] propose to formally approximate
the conditional average by a stochastic estimation. These new models can be
referred to as optimal or nearly-optimal models, leading to optimal LES. The
estimation of the subgrid force is based on the convolution of an estimation
kernel Kij with velocity event data at N points (1 , ..., N ):

(9.28)
mi (x) = Kij (x, 1 , ..., N )Ej (u ; 1 , ..., N )d1 ...dN ,
where Ej is an event vector. Chosing
E(1 , ..., N ) = (1, ui (1 ), uj (1 )uk (2 ), ...)
(9.29)
we recover the expansion

mi (x) = Ai (x) + Bij (x, 1 )uj (1 )d1

+ Cijk (x, 1 , 2 )uj (1 )uk (2 )d1 d2
(9.30)
The random mean square error between Mi and mi is minimal when

ei (x)Ek (1 , ..., N ) = 0
(9.31)
yielding the following denition of the optimal kernel Kij

Mi (x)Ek (1 , ..., N ) =
Kij (x, 1 , ..., N )
.
Ej (1 , ..., N )Ek (1 , ..., N ) d1 ...dN (9.32)
The resulting optimal subgrid models have the property that the correlation of the parametrized subgrid force with any event data is the same as
the correlation of the exact subgrid force with the same event data:
mi (x)Ej (1 , ..., N ) = Mi (x)Ej (1 , ..., N )
(9.33)
9.1.5 Mathematical Analysis of Sensitivities and Uncertainties

in Large-Eddy Simulation
The developments presented above deal with the nature of the information
retrieved from large eddy simulation and some statistical equivalency constraints. But a key problem in practical cases is to evaluate the sensitivity of
312
the computed results with respect to the subgrid model (or its inputs), i.e. to
obtain an estimate for the uncertainties associated to the fact that all scales
are not resolved and that small ones are parameterized. The mathematical
framework for such an analysis was proposed by Anitescu and Layton [13].
Starting from the governing equations for large eddy simulation in the case
where a subgrid viscosity model is utilized
u
+ (u u) = p + 2 u + (sgs (, u)S(u))
t
u = 0 ,
(9.34)
(9.35)
where the notation S(u) was used instead of the usual S for the sake of
convenience. Let us now introduce the sensitivities of the velocity and the
pressure dened as
p
u
, q
.
(9.36)
v
These sensitivities are solutions of the following equations, which are obtained dierentiating (9.34) and (9.35) with respect to
v
t
(v u + u v) = q + 2 v

sgs (, u) v S(u)
sgs (, u) +
+
u

+ sgs (, u)S(v)
,
v =0
(9.37)
(9.38)
Thus, once the large-eddy simulation elds u and p are computed, their
sensitivities can be evaluated solving the linear problem given above. Let
us illustrate this approach considering the Smagorinsky model (5.90), which
corresponds to
(9.39)
sgs (, u) = (CS )2 |S(u)| .
Applying the dierentiation rules, one obtains the following Jacobian of
the subgrid viscosity

S(u)
2
2
sgs v = 2CS |S(u)| + (CS )
sgs +
: S(v) . (9.40)
u
|S(u)|
That expression is also linear with respect to the sensitivity vector v and
can easily be computed.
Once the sensitivies of the solution are known, one can derive an estimate
for the uncertainties on the quantities calculated from it. Let J (, u) be
any smooth functional estimated using the large-eddy simulation results (e.g.
drag and lift of an immersed body). The best value one can sought is the one
associated to the exact, unltered velocity eld u, i.e. J (0, u).
9.2 Correction Techniques
313
The error commited on J can be expressed as a function of the cuto

length writing the following rst-order Taylor series expansion:
J (0, u) = J (, u) J (, u) v
(9.41)
where J is the Jacobian of J . The error is given by the second term in the
right hand side. An interesting point is that this relation can also be used to
estimate the exact value of the functional by correcting the value computed
from the large-eddy simulation data.

As relations (9.2) and (9.3) show, the statistical moments computed from
the resolved eld cannot be equal to those computed from the exact solution.
In order to be able to compare these moments for validation purposes, or
analyze the large-eddy simulation data, the error term has to be evaluated
or eliminated. Several possible techniques are described in the following for
doing this.
9.2.1 Filtering the Reference Data
The rst solution is to apply the same ltering as was used for the scale separation to the reference solution [533, 7]. Strict comparisons can be made with
this technique, but it does not provide access to theoretically usable values,
which makes it dicult to use the data generated by large-eddy simulation
for predicting physical phenomena, because only ltered data are available.
In order for physical analyses to be fully satisfactory, they should be made on
complete data. However, analysis is possible when the quantities considered
are independent or weakly dependent on the subgrid scales3 .
Moreover, this approach is dicult to apply when the eective lter is not
known analytically, because the reference data cannot be ltered consistently.
It may also be dicult to apply an analytical to experimental data, because
in order to do so, access is needed to the data spectra that are to serve
for validation or analysis. We see another source of problems cropping up
here [562]: experimentally measured spectra are time spectra in the vast
majority of cases, while the large-eddy simulation is based on space ltering.
This may introduce essential dierences, especially when the ow is highly
anisotropic in space, as it is in the regions near a solid wall. Similar remarks
can be made concerning the spatial ltering of data from a direct numerical
simulation for a priori test purposes: applying a one- or two-dimensional lter
can produce observations that are dierent from those that would be obtained
with a three-dimensional lter.
3
As is generally the case for the mean velocity eld. See the examples given in
Chap. 14.
314
9.2.2 Evaluation of Subgrid-Scale Contribution

A second solution is to evaluate the error term and reconstruct from the
ltered solution moments that are equal to those obtained from the full eld.
Use of a De-ltering Technique. One way is to try to reconstruct the full
eld from the resolved one, and compute the statistical moments from the
reconstructed eld. In theory, this makes it possible to obtain exact results if
the reconstruction itself is exact. This reconstruction operation can be interpreted as de-ltering, i.e. as an inversion of the scale separation operation. As
was seen in Chap. 2, this operation is possible if the lter is an analytical one
not belonging to the class of Reynolds operators. In other cases, i.e. when the
eective lter is unknown or possesses projector properties, this technique is
not strictly applicable and we have to do with an approximate recontruction.
We then use a technique based on the dierential interpretation of the lter
analogous to the one described in Sect. 7.2.1. With this interpretation, we
can express the ltered eld u as:

2n

2n
u = Id +
Cn
u .
(9.42)
x2n
n=1
This relation can be formally inverted writing:

u=
Id +

n=1
2n
Cn
2n
x2n
1
u
(9.43)
and, by interpreting the dierential operator as an expansion function of the

small parameter , we get:

2n

2n

u = Id +
Cn
u .
(9.44)
x2n
n=1
By truncating the series at some arbitrary order, we thus get a recontruction method that is local in space and easy to use. The diculty resides in
the choice of the coecients Cn , which describe the eective lter and can
only be determined empirically.
Use of a Subgrid Model. Another means that is easier to use is to compute the contribution of the subgrid terms by means of the subgrid stresses
representation generated by the model used in the simulation. This technique
cannot evaluate all the error terms present in (9.2) and (9.3) and can only
reduce the error committed in computing the second-order moments.
It does, however, oer the advantage of not requiring additional computations as in the recontruction technique.
It should be noted here that this technique seems to be appropriate when
the models used are structural, representing the subgrid tensor, but that it
315
is no longer justied when functional models are used because these ensure
only an energy balance.
It is worth noting that dierent subgrid models can be used for dierent
purposes: a model can be used during the computation to close the ltered
momentum equation, while another model can be used to recover a better reconstruction of subgrid contributions. This is more particulary true when the
data extraction is related to multiphysics purposes. As an example, specic
subgrid models have been derived for predicition of subgrid acoustic sources
by Seror et al. [666, 667], while an usual functional model was used in the
momentum equations.
9.2.3 Evaluation of Subgrid-Scale Kinetic Energy
2
= ui ui /2
The specic case of the evaluation of the subgrid kinetic energy qsgs
(not to be assimilated to the generalized subgrid kinetic energy, see p. 54
received a lot of attention, since it is required in many applications. Several
proposals have been made, which are presented below:
1. Yoshizawas method (p. 315), which relies on a simple dimensional analysis and requires the same inputs as the Smagorinsky functional model.
Its dynamic variants are also discussed.
2. Knaepens dynamic model (p. 316), which is based on scale similarity
concepts.
3. Models based on the integration of a spectrum shape (p. 317). These
methods require the foreknowledge of the analytical spectrum shape,
whose parameters (dissipation rate, Kolmogorov scale) are computed using simple subgrid models.
This section presents the original version of the models. But it is observed
that many elements of each model can be transposed in the other ones, leading
to the denition of new models.
Yoshizawas Model. A simple dimensional analysis yields
2
2
= 2CI |S|2
qsgs
(9.45)
The value of the constant is taken equal to CI = 1/ 2 in [231] and to 0.01

in [538]. A dynamic evaluation of the parameter CI based on the Germano
identity was proposed by Moin et al. [538]. Denoting the test lter level with
a tilde symbol, one obtains
CI =
*k
uk u
1 u/
k uk *
2
/
2
2* *2
|S| |S|2
(9.46)
This dynamic model may requires some numerical stabilization (like

averaging over homogeneous directions) in practical applications. Another
316
dynamic evaluation procedure is proposed by Wong and Lilly [766], which is

based on the so-called Kolmogorov scaling:
4/3
2
= CI
qsgs
|S| ,
(9.47)
where the dimensional constant is evaluated as

CI =
*k
1
u/
uk u
k uk *
4/3

2 *
4/3
*

|S|
(9.48)
2
This new formulation ensures that the realizability constraint qsgs
is fullled.
Knaepens Model. Rewriting the multiplicative Germano identity (3.80)

and taking the trace of it, one obtains (the test lter level being noted with
the tilde symbol)
*i *
*i = ui ui u
Lii u/
ui u
ui
i ui *
(9.49)
showing that the trace of the Leonard tensor L is equal to twice the dierence
of the resolved kinetic energy at the two considered ltering level. Assuming
that the two lters are sharp cuto lters with cuto wavenumbers kc and
kc , the following expression is recovered
1
Lii =
2
kc
E(k)dk
(9.50)
kc
If the case the spectrum shape is of the form E(k) = A k 5/3 , (9.50)
yields
Lii
1
.
(9.51)
A=
3 kc2/3 kc 2/3
Using this last relationship and assuming that the inertial range spectrum
shape is valid at all subgrid wave numbers, one obtains the following estimate
for the subgrid kinetic energy

Lii
1
2
qsgs
=
E(k)dk = 2/3
.
(9.52)
*
2
kc
This model can be modied to account for the existence of the dissipative
range of the spectrum, leading to

Lii
1
2
qsgs =
E(k)dk = 2/3
(9.53)
(1 2/3 ) ,
*
2
kc
317
where = K / is the ratio of the lter cuto length and the Kolmogorov
scale. The Kolmogorov scale can also be evaluated in the same way. Assuming
that the local equilibrium hypothesis applies and that most of the viscous
dissipation occurs in the inertial range, one obtains
k
2k 2 E(k) = ij S ij ,
(9.54)
kc
where k = 2/K and is the subgrid tensor (to be evaluated with any
subgrid model). As a consequence, one obtains
k =

3/4
2ij S ij
kc4/3
3A
(9.55)
leading to a closed model.

Dissipation-based models. Another way to compute the subgrid kinetic
energy is to assume a spectrum shape and to integrate it for all wave numbers
larger than the cuto kc [517, 735, 608].
A simple spectrum shape is obtained assuming that the Kolmogorov spectrum shape is valid from kc to Jk (with k the Kolmogorov wave number
and J a cuto parameter) and that no scale exist at higher wave number,
leading to:

K0 2/3 k 5/3 kc k Jk
.
(9.56)
E(k) =
0
otherwise
Integrating this expression, one obtains

2/3

3K0 2/3
kc
1
kc Jk
2
2/3
Jk
2kc
qsgs
=
0
otherwise
(9.57)
The subgrid dissipation rate is not known a priori and must be evaluated.
Assuming that the subgrid scale are in statistical equilibrium, the relation
= ij S ij is valid an can be used to compute , the subgrid stresses ij being
parameterized using an arbitrary subgrid model. A simple subgrid viscosity
model is usually sucient to this end, but more complex models can be used
to account for anisotropy, as proposed by Pullin [608] who uses a stretchedvortex structural model. Equation (9.57) is explicit if J = . For nite
values of J, the factor Jk remains to be evaluated. A simple method to
do that [517, 735] is to assume a local balance between the total dissipation
on one side and the sum of the resolved-scale dissipation and the subgrid
dissipation on the other side
= |S|2 ij S ij
(9.58)
Expressing and/or ij as a function of the kinetic energy spectrum, one

obtains a non-linear equation for Jk which can be solved thanks to Newton
algorithm.
318
9.3 Practical Experience

Practice shows that nearly all authors make comparisons with reference data
or analyze large-eddy simulation data with no processing of the data. The
agreements observed with the reference data can then be explained by the fact
that the quantities being compared are essentially related to scale ranges contained in the resolved eld. This is generally true of the rst-order moments
(i.e. the mean velocity eld) and, in certain cases, of the second-order moments (the Reynolds stresses). This lack of processing prior to data analysis
seems to be due mainly to the uncertainties in the techniques for evaluating the contributions of the subgrid scales and to the diculty of ad hoc
ltering of the reference data. Large-eddy simulation also allows a satisfactory prediction of the time frequency associated with large-scale periodic or
quasi-periodic phenomena (such as vortex shedding) and the rst harmonics
of this frequency for ne mesh.
One has to be careful when trying to recover high-order statistical data
from large-eddy simulation. The two main reasons are:
The properties of the exact ltered solution, i.e. the solution of an ideal
large-eddy simulation without numerical and modeling errors referred to as
u in Chap. 1, may be dierent from those of the exact unltered solution.
Numerical and modeling errors can corrupt the eld, yielding new errors.
Fig. 9.1. Comparison of time correlation versus a normalized time lag in isotropic
turbulence. Upper curves are from Large-Eddy Simulation and lower curves from
Direct Numerical Simulation. Courtesy of S. Rubinstein, NASA.
319
Fig. 9.2. Time microscale versus wave number in isotropic turbulence. Solid line:
Large-Eddy Simulation; Dashed line: Direct Numerical Simulation; Dotted line:
theoretical 1 slope in the inertial range. Courtesy of S. Rubinstein, NASA.
Fig. 9.3. Ratio of LES time microscale with the DNS value (Symbols). Solid line
shows the best linear t. Courtesy of S. Rubinstein, NASA.
320
Fig. 9.4. Probability density function of ltered velocity increments (experimental

data), for dierent lter size. Tails of the PDF is observed to be damped with
increasing lter size. Courtesy of C. Meneveau, Johns Hopkins Univ.
Fig. 9.5. Comparison of probability density functions of ltered velocity increments. Dashed line: ltered experimental data. Others: LES computations based on
dierent subgrid models (Smagorinsky, dynamic Smagorinsky, dynamic functionalstrucrtural models). Courtesy of C. Meneveau, Johns Hopkins Univ.
321
Fig. 9.6. Longitudinal velocity structure of order 6 and 10 plotted versus displacement. Squares: unltered data; Circles: ltered data. From [116].
322
Very little information is available on these two sources of error. He et

al. [300, 301] have shown that spectral viscosity models generate elds which
are more correlated than the corresponding unltered or perfect large-eddy
simulation solutions (see Fig. 9.1). This increased correlation is shown to
lead to a modication of Eulerian time correlations. The subgrid closure
is observed to have a signicant impact on the time microscale at all wave
numbers (see Figs. 9.2 9.3): the value found in Large-Eddy Simulation is 1.8
times greater than in Direct Numerical Simulation, while theoretical analyses
show that the value in ltered Direct Numerical Simulation (or ideal LargeEddy Simulation) should lie in the range 11.15.
From experimental grid and wake turbulence data, Cerutti and Meneveau [116] and Kang et al. [375] examined fundamental dierences between
ltered and unltered velocity elds through probability density functions
and the scaling behavior of high-order structure functions (see Figs. 9.4
9.6). This comparative study dealing with probability density functions of
velocity increments yields the conclusion that the tails of the distributions
are aected by the ltering even at scales much larger than the lter scale.
Large discrepancies are also observed with respect to the scaling of structure
functions. But it is worth noting that using simple shell models of turbulence,
Benzi et al. [53] have shown that the use of subgrid viscosity-like models does
not preclude internal intermittency. Inertial intermittent exponents were observed to be fairly independent of the way energy is dissipated at small scales
for this very simplied dynamical model of turbulence. Other eects, such
as the limited size of the computational domain, can also generate some discrepancies with experimental data [7]. Nevertheles, Large-Eddy Simulation
can be used to recover some useful informations if the numerical error is
controlled and the grid ne enough: Alvelius and Johansson [11] observed
a good prediction of two-point pressure-velocity correlations in homogeneous
turbulence.
10. Boundary Conditions
Like all the other approaches mentioned in the introduction, large-eddy simulation requires the setting of boundary conditions in order to fully determine
the system and obtain a mathematically well-posed problem. This chapter is
devoted to questions of determining suitable boundary conditions for largeeddy simulation computations. The rst section is a discussion of general
order, the second is devoted to the representation of solid walls, and the
third discusses methods used for representing an unsteady upstream ow.
10.1 General Problem

10.1.1 Mathematical Aspects
Many theoretical problems dealing with boundary conditions for large-eddy
simulation can be identied. Two of the most important ones are the possible
interaction between the lter and the exact boundary conditions, and the
specication of boundary conditions leading to the denition of a well-posed
problem.
From the results of Sects. 2.2 and 3.4, we deduce that the general form of
the ltered NavierStokes equations on a bounded domain is

ui
ui
uj
+
(ui uj ) +
+
=
t
xj
xi
xj xj
xi

ui () uj ()
G(x) ui ()uj ()+p()

+
nj ()d , (10.1)
xj
xi

ui
=
G(x )ui ()ni ()d ,
(10.2)
xi
where n is the outward unit normal vector to the boundary of , . It

must be noted that only terms induced by the interaction of the lter and the
boundaries have been retained, i.e. other commutation errors are neglected.
Right-hand side terms in (10.1) and (10.2) are additional source terms, that
must be modeled because they involve the non-ltered velocity and pressure
elds.
324
It is seen from these equations that two possibilities arise when ltering
the NavierStokes equations on bounded domains:
The rst one, which is the most commonly adopted (classical approach),
consists of considering that the lter width decreases when approaching
the boundaries such that the interaction term cancels out (see Fig. 10.1).
Thus, the source term can be neglected and the basic ltered equations
are left unchanged. The remaining problem is to dene classical boundary
conditions for the ltered eld.
The second solution (embedded boundary conditions), rst advocated by
Layton et al. [206, 234, 429, 357, 207], consists of ltering through the
boundary (see Fig. 10.1). As a consequence, there exists a layer along the
boundary, whose width is of the order of the lter cuto lengthscale, in
which the source term cannot be neglected. This source term must then be
explicitly computed, i.e. modeled.
The discussions so far clearly show that the constitutive equations of
large-eddy simulation can be of a degree dierent from that of the original
NavierStokes equations. This is trivially veried by considering the dierential interpretation of the lters: the resolved equations are obtained by applying a dierential operator of arbitrarily high order to the basic equations.
Moreover, it has been seen that certain subgrid models generate high-order
derivatives of the velocity eld.
This change of degree in the resolved equations raises the problem of
determining the associated boundary conditions, because those associated
with the equations governing the evolution of the exact solution can no longer
be used in theory for obtaining a mathematically well-posed problem [728,
234]. This problem is generally not considered, arguing the fact that the
p
higher-order terms appear only in the form of O( ), p 1 perturbations
of the NavierStokes equations and the same boundary conditions are used
both for the large-eddy simulation and for direct numerical simulation of the
NavierStokes equations. Moreover, when the eective lter is unknown, it is
no longer possible to derive suitable boundary conditions strictly, which also
leads to the use of the boundary conditions of the basic problem.
10.1.2 Physical Aspects
The boundary conditions, along with the similarity parameters of the equations, specify the ow, i.e. determine the solution. These conditions represent
the whole uid domain beyond the computational domain. To specify the solution completely, these conditions must apply to all of its scales, i.e. to all
the space-time modes it comprises.
So in order to characterize a particular ow, the amount of information
in the boundary conditions is a function of the number of degrees of freedom
of the boundary condition system. This poses the problem of representing
a particular solution, in order to be able to reproduce it numerically. We
10.1 General Problem
325
Fig. 10.1. Schematic of the classical approach (Top) and the embedded boundary
condition approach (Bottom), in a solid wall conguration. In the classical approach
the cuto length is reduced in the vicinity of the wall so that the interaction
term cancels out. In the second approach, the cuto length is constant. Courtesy
of E. Garnier, ONERA.
326
have a new modeling problem here, which is that of modeling the physical
test conguration.
This diculty is increased for the large-eddy simulation and direct numerical simulation, because these simulations contain a large number of degrees
of freedom and require a precise space-time deterministic representation of
the solution at the computational domain boundaries.
Two special cases will be discussed in the following sections: that of representing solid walls and that of representing a turbulent inow. The problem
of the outow conditions, which is not specic to the large-eddy simulation
technique, will not be addressed1 .
10.2 Solid Walls

Specic Features of the Near-Wall Region. The structure of the boundary layer ow has certain characteristics that call for special treatment in the
framework of large-eddy simulation. In this section, we describe the elements
characteristic of the boundary layer dynamics and kinematics, which shows
up the dierence with an isotropic homogeneous turbulence. For a detailed
description, the reader may refer to [648, 150].
Denitions. Here we adopt the ideal framework of a at-plate, turbulent
boundary layer, without pressure gradient. The external ow is in the (Ox)
direction and the (Oz) direction is normal to the wall. The external velocity is
denoted Ue . In the following, the Cartesian coordinate system will be denoted
either (x, y, z) or (x1 , x2 , x3 ), for convenience. Similarly, the velocity vector
is denoted (u, v, w) or (u1 , u2 , u3 ).
We rst recall a few denitions. The boundary layer thickness is dened
as the distance from the plate beyond which the uid becomes irrotational,
and thus where the uid velocity is equal to the external velocity.
The wall shear stress p is dened as:
+
2
2
+ p,23
,
(10.3)
p = p,13
in which p,ij = S ij (x, y, 0). The friction velocity u is dened as:
u = p .
In the case of the canonical boundary layer, we get:

u1
u =
(x, y, 0) .
z
1
(10.4)
(10.5)
See [156] for a specic study of exit boundary conditions for the plane channel
ow case.
10.2 Solid Walls
327
We dene the Reynolds number Re by:

Re =
(10.6)
The reduced velocity u+ , expressed in wall units, is dened as:

u+ = u/u
(10.7)
The wall coordinates (x+ , y + , z + ) are obtained by the transformation:

(x+ , y + , z + ) = (x/l , y/l , z/l )
(10.8)
where the viscous length l is dened as l = /u .

Statistical Description of the Canonical Boundary Layer. The boundary layer
is divided into two parts: the inner region (0 z < 0.2) and the outer region
(0.2 z). This decomposition is illustrated in Fig. 10.2. In the inner region,
the dynamics is dominated by the viscous eects. In the outer region, it is
controlled by the turbulence. Each of these regions is split into several layers,
corresponding to dierent types of dynamics.
In the case of the canonical boundary layer, we have three layers in the
inner region in which the mean longitudinal velocity prole follows special
laws. The positions of these layers are referenced in the reduced coordinate
system, because the dynamics of the inner region is dominated by the wall
eects and l is the pertinent length scale for describing the dynamics. The
characteristic velocity scale is the friction velocity. These three layers are the:
Viscous sublayer: z + 5, in which
+
+
u+
1 (z ) = z
(10.9)
Fig. 10.2. Mean streamwise velocity prole for the canonical turbulent boundary
layer, and its decomposition into inner and outer regions. Left: mean velocity prole
(external units). Right: mean velocity prole (wall units).
328
Buer layer: 5 < z + 30, where

+
+
u+
1 (z ) 5 ln z 3.05 .
(10.10)
Prandtl or logarithmic inertial layer: 30 < z + ; z/ 1, for which

+
u+
1 (z )
1
ln z + + 5, 5 0, 1, = 0, 4 .
(10.11)
The outer region includes the end of the logarithmic inertial region and
the wake region. In this zone, the characteristic length is no longer l but
rather the thickness . The characteristic velocity scale remains unchanged,
though. The average velocity proles are described by:
For the logarithmic inertial region:
zu
u1 (z)
+B
= A ln
u
(10.12)
where A and B are constants;

For the wake region:
z
zu
u1 (z)
+B+ W
= A ln
u
(10.13)
where A, B and are constants and W the wake function dened by

Clauser as:
(10.14)
W (x) = 2 sin2 (x/2) .
Concerning the Dynamics of the Canonical Boundary Layer. Experimental
and numerical studies have identied dynamic processes within the boundary layer. We will summarize here the main elements of the boundary layer
dynamics that originate the turbulence in the near-wall region.
Observations show that the ow is highly agitated very close to the wall,
consisting of pockets of fast and slow uid that organize in ribbons parallel
to the outer velocity (streaks, see Fig. 10.3). The low-velocity pockets migrate slowly outward in the boundary layer (ejection) and are subject to an
instability that makes them explode near the outer edge of the inner region.
This burst is followed by an arrival of fast uid toward the wall, sweeping
the near-wall region almost parallel to it. These highly intermittent events
in time and space induce strong variation in the unsteady Reynolds stresses
and originate a very large part of the production and dissipation of the turbulent kinetic energy. These variations produce uctuations in the subgrid
dissipation that can reach 300% of the average value and can make it change
sign. Analyses of direct numerical simulations [291, 598, 449, 363] indicate
that a very intense small scale dissipation in the buer region is correlated
with the presence of sheared layers that form the interfaces between the uid
pockets of dierent velocities. These mechanisms are highly anisotropic. Their
10.2 Solid Walls
329
Fig. 10.3. Visualization of the streaks in a plane channel ow (Re = 590). Dark
and light gray denote opposite sign of streamwise vorticity. Courtesy of M. Terracol,
ONERA.
characteristic scales in the longitudinal and transverse directions x and y ,

+
respectively, are such that +
x 200 1000 and y 100. The maximum
+
turbulent energy production is observed at z 15. This energy production
at small scales gives rise to a high backward energy cascade and associated
with the sweeping type events. The forward cascade, for its part, is associated
with the ejections.
In the outer regions of the boundary layer where the viscous eects no
longer dominate the dynamics, the energy cascade mechanism is predominant. Both cascade mechanisms are associated preferentially with the ejections.
Numerical and theoretical results [353, 354, 355, 750] show that wallbounded turbulence below z + 80 is a relatively autonomous system, which
is responsible for the generation of a signicant part of the turbulent kinetic
energy dissipated in the outer part of the boundary layer. This turbulent cycle
is often referred to as an autonomous cycle because it is observed to remain
active in the absence of the outer ow. This cycle involves the formation of
velocity streaks from the advection of the mean velocity prole by streamwise
vortices, and the generation of the vortices from the instability of the streaks.
330
The presence of the wall seems to be only necessary to maintain the mean
shear. The way that this inner turbulent cycle and the outer ow interact are
still under investigation.
H
artel and his coworkers [291, 290, 289] give a more precise analysis of
the subgrid transfer in the boundary layer by introducing a new splitting2 of
the subgrid dissipation
= ij S ij = MS + FS
(10.15)
with
MS = ij S ij
FS
= (ij ij )(S ij S ij )
(10.16)
.
(10.17)
The MS is related to the mean strain, and accounts for an enhancement

of subgrid kinetic energy in the presence of mean-ow gradients. The second
term, which is linked to the strain uctuations, represents the redistribution
of energy without aecting the mean ow directly.
A priori tests [291, 290, 289] perfomed using plane channel ow and circular pipe data reveal that the net eect of the coupling is a forward energy
transfer, and:
The mean strain part is always associated to a net forward kinetic energy
cascade.
The uctuating strain part results in a net backward kinetic cascade in
a zone located in the buer layer, with a maximun near z + = 15. This net
backward cascade is correlated to the presence of coherent events associated
to turbulence production.
Typical distributions of MS and FS are shown in Fig. 10.4.
Kinematics of the Turbulent Boundary Layer. The processes described above
are associated with existence of coherent structures [617].
The buer layer is dominated by isolated quasi-longitudinal structures
that form an average angle with the wall of 5 at z + = 15 and 15 at z + = 30.
Their mean diameter increases with their distance from the wall3 .
The logarithmic inertial region belongs both to the inner and outer regions, and thus contains characteristic space scales, which is compatible with
the existence of two dierent types of structures. The dynamics is governed
by quasi-longitudinal and arch structures. The quasi-longitudinal structures
can be connected to transverse structures and form an angle with the surface that varies from 15 to 30 . The span of the arch structures is of the
order of the width of the slow-uid pockets at the bottom of the layer, and
2
3
It diers from the splitting proposed by Shao (see Sect. 7.4.1).

It should be noted that contradictory observations can be found. Lamballais [421]
observes that the most probable angle of the vorticity (projected on a plane
perpendicular to the wall) is close to 90 for 5 < z + < 25, which goes against
the model of longitudinal vortices at the wall.
10.2 Solid Walls
331
Fig. 10.4. Distribution of mean strain (MS ) and uctuating strain (FS ) dissipations in a plane channel ow. Fluctuating strain is observed to yield dominant
backscatter near z + 15, but the total dissipation remains positive.
increases linearly with the distance from the wall. The relative number of
quasi-longitudinal structures decreases with the distance from the wall, until
it cancels out at the beginning of the wake region.
The wake region is populated with arch structures forming an angle of
45 with the wall. Their x and y spacing is of the order of .
Resolving or Modeling. The description we have just made of the boundary layer ow structure clearly shows the problem of applying the large-eddy
simulation technique in this case. Firstly, the mechanisms originating the
turbulence, i.e. the ow driving mechanisms, are associated with xed characteristic length scales on the average. Also, this turbulence production is
associated with a backward energy cascade, which is largely dominant over
the cascade mechanism in certain regions of the boundary layer. These two
factors make it so that the subgrid models presented in the previous chapters become inoperative because they no longer permit a reduction of the
number of degrees of freedom while ensuring at the same time a ne representation of the ow driving mechanisms. There are then two possible approaches [536]:
Resolving the near-wall dynamics directly. Since the production mechanisms escape the usual subgrid modeling, if we want to take them into
account, we have to use a suciently ne resolution to capture them. The
solid wall is then represented by a no-slip condition: the uid velocity is
332
set equal to that of the solid wall. This equality implicitly relies on the hypothesis that the mean free path of the molecules is small compared with
the characteristic scales of the motion, and that these scales are large compared with the distance of the rst grid point from the wall. In practice,
this is done by placing the rst point in the zone (0 z + 1). To represent the turbulence production mechanisms completely, Schumann [655]
+
+
+
recommends a spatial resolution such that 1 < 10, 2 < 5 and 3 < 2.
Also, Zang [798] indicates that the minimum resolution for capturing the
+
+
existence of these mechanisms is 1 < 80, 2 < 30 and that three grid
+
points should be located in the z 10 zone. Zahrai et al. [795] indicate
+
+
that 1 100, 2 = 12 should be used as an upper limit if a secondorder accurate numerical method is used. These values are given here only
for reference, since larger values can also be found in the literature. For
+
example, Piomelli [591] uses 1 = 244 for a plane channel ow. Chapman [119] estimates that representing the dynamics of the inner region,
which contributes about one percent to the thickness of the full boundary
layer, requires O(Re1.8 ) degrees of freedom, while only O(Re0.4 ) are needed
+
+
to represent the outer zone. This corresponds to 1 100, 2 20 and
+
3 < 2. Considering that non-isotropic modes must be directly resolved,
Bagget et al. [32] show that the number of degrees of freedom of the solution (in space) scales as Re2 .
Modeling the near-wall dynamics. To reduce the number degrees of freedom
and especially avoid having to represent the inner region, we use a model
for representing the dynamics of that zone. This is a special subgrid model
called the wall model. Since the distance from the rst grid point to the
wall is greater than the characteristic scales of the modes existing in the
modeled region, the no-slip condition can no longer be used. The boundary
condition will apply to the values of the velocity components and/or their
gradients, which will be provided by the wall model. This approach makes
it possible to place the rst point in the logarithmic layer (in practice,
20 z + 200). The main advantage of this approach is that the number
of degrees of freedom in the simulation can be reduced greatly; but since
a part of the dynamics is modeled, it constitutes an additional source of
error.
10.2.2 A Few Wall Models
In the following, we present the most popular wall models for large-eddy
simulation. These models all represent an impermeable wall, and most of
them have been implemented using a staggered grid (see Fig. 10.5). The discussion will be restricted to wall models dened on Cartesian body-tted
computational grids. Details dealing with implementation on curvilinear
body-tted grids or on Cartesian grids using the immersed boundary technique [35, 713, 271, 270] will not be discussed, since they would require an
10.2 Solid Walls
333
Fig. 10.5. Illustration of a staggered grid system in the streamwise/wall-normal

plane.
extensive discussion about numerical methods which is far beyond the scope
of this book.
Existing strategies for the denition of wall models for large-eddy simulation following the classical approach can be grouped in several classes:
Higher-order boundary conditions: velocity gradients at the wall are controlled by enforcing boundary conditions on second-order derivatives. This
class is represented by Deardors model (p. 337).
Wall stress models: the rst mesh at the wall is chosen to be very large, so
that it is not able to respresent correctly the dynamics of the inner layer.
Typical dimensions are 100 wall units in the wall-normal and spanwise
directions and 500 wall units in the streamwise direction. This is illustrated
in Fig. 10.7, where it is clearly observed that the near-wall events are
averaged over the rst grid cell. The wall model should provide the value
of wall stresses, which cannot be accurately directly computed on the grid
because of its coarseness (see Fig. 10.6), and the value of the wall-normal
velocity component.
The basic form of these stress models is
i,3 = F (u1 , z2 ) ,
i = 1, 2 ,
(10.18)
where z2 is the height of the rst cell. If the ow is bidimensional and

laminar, a trivial relation is
1,3 =
u1
z2 /2
(10.19)
A rst group of wall-stress models is based on the extrapolation of this

linear, laminar law, the molecular viscosity being replaced by an eective
turbulent viscosity, e . Recalling that the total mean shear stress is almost
334
Fig. 10.6. Computed value of the mean wall shear stress (with reference to its
exact value), expressed as a function of the distance of the rst o-wall grid point.
constant across the inner part of the boundary layer, we have:

13 (z) = tot (z)
u1 (z)
p,13 ,
z
(10.20)
where tot (z) is the sum of the molecular and the turbulent viscosity. By
integrating (10.20) in the wall-normal direction, we obtain:
z
z
13 (z)
u1 (z)
dz =
dz ,
(10.21)
z
0 tot (z)
0
leading to

p,13
0
1
dz = u1 (z) ,
tot (z)
(10.22)
and, nally,
p,13 =
1
z
1
1
u1 (z)
dz
z 0 tot (z)
z

(10.23)
The wall-stress models discussed below are:

1. The Schumann model (p. 339), which relies on a linear relation between
the wall stress and the velocity component at the rst o-wall grid point.
The skin friction is an entry parameter for the model.
10.2 Solid Walls
335
Fig. 10.7. Representation of the wall-stress approach. The contours of a typical

cell are superposed onto a boundary-layer instantaneous ow obtained with a wallresolving mesh (instantaneous velocity vectors and isolevels of streamwise velocity
uctuation). Length scales are expressed in wall units. The dimensions of the cell
are 500 wall units in the streamwise direction and 100 wall units in the spanwise
and wall-normal directions. Top: view in an (x, z) plane. Bottom: view in an (y, z)
plane. Courtesy of E. Tromeur and E. Garnier, ONERA.
336
2. The Gr
otzbach model (p. 339), which is an extension of the Schumann
model. The skin friction is now computed assuming that the ow corresponds to the canonical at-plate boundary layer. The skin friction
is computed by inverting the logarithmic law prole for the streamwise
velocity.
3. The shifted correlations model (p. 340), which extends the Gr
otzbach
model by taking into account explicitly the fact that uctuations are
governed by coherent structures with given time- and lengthscales.
4. The ejection model (p. 341), which is another extension of the model of
Gr
otzbach. It takes into account the eects of sweep and ejection events
on the wall shear stress.
5. The optimized ejection model (p. 341), which is based on experimental
correlation data and yields better correlation coecient in a priori tests.
6. The model of Werner and Wengle (p. 344), which can be seen as a variant
of the Gr
otzbach model based on power-law proles for the streamwise
velocity instead of the logarithmic law. The main advantage is that the
power law can be inverted explicitly.
7. The modied WernerWengle model (p. 345), which accounts for the
existence of ejection, as the model proposed by Piomelli.
8. The model of Murakami et al. (p. 343), which can be interpreted as
a simplied version of the preceding model.
9. The model of Mason and Callen for rough walls (p. 340).
10. The suboptimal-control-based wall models (p. 345). Numerical experiments show that the previous wall stress models are not robust, i.e.
they lead to disappointing results on very coarse mesh for high Reynolds
numbers. This is mainly due to the fact that they can not account for
large numerical and physical errors occuring on such coarse grids. These
new models, developed within the framework of suboptimal control, aim
at producing the best possible results.
The second group of wall-stress models relies on the use of an internal layer
near the wall, leading to the denition of two-layer simulations. Another
set of governing equations is solved on an auxiliary grid located inside the
rst cell of the large-eddy simulation grid. An eective gain is obtained
if the auxiliary simulation can be run on a much coarser grid than largeeddy simulation while guaranteeing the accuracy of the results, or if the
auxiliary equations are much simpler than the NavierStokes equations.
Models belonging to this group are:
1. The thin-boundary-layer equation model developed by Balaras et al.
(p. 342). The governing equations solved in the inner layer are the boundary layer equations derived from the NavierStokes equations. The gain
comes from the fact that the pressure is assumed to be constant across
the boundary layer, and the pressure is not computed.
2. The wall model based on Kersteins ODT approach (p. 350). This
model relies on the reconstruction of the solution in the inner layer via
10.2 Solid Walls
337
a simplied one-dimensional stochastic system. It can be interpreted as

a surrogate of the model based on the full boundary layer equations.
3. Hybrid RANS/LES approaches, in which a RANS-like simulation is performed in the near-wall region, while the core of the ow is treated by
large-eddy simulation. All these models and techniques are discussed in
Chap. 12, with emphasis in Sect. 12.2.
O-wall boundary conditions: the mesh is built so that the rst point is
located in the uid region, and not on the wall. The rst grid line parallel
to the wall must be located in the inner region of the boundary layer. This
approach is illustrated in Fig. 10.8. An important point is that the grid
must still be able to represent details of the ow, and thus the mesh should
be the same as those used for wall-resolving simulation away from the wall.
Several types of boundary conditions can be used following this approach,
dealing either with the velocity components or their derivatives [30, 568,
356]. Poor results have been obtained using that approach, which requires
very accurate structural information on the uctuations to yield accurate
results. Jimenez and Vasco [356] showed that the ow is very sensitive
to the prescribed wall-normal velocity component (transpiration velocity),
which must satisfy the continuity constraint.4 The need for an instantaneous non-zero transpiration velocity can easily be understood by looking
at the velocity eld displayed in Fig. 10.8.
The usual failure of these models leads to the appearance of a strong,
spurious boundary layer above the articial boundary.
Deterministic minimal boundary-layer unit simulation. Pascarelli et al.
[580] proposed performing a wall-resolved temporal large-eddy simulation
on the smallest domain allowing the existence of the near-wall autonomous
cycle, and duplicating it. This approach leads to the denition of a crystal
of elementary chaotic dynamical systems. The associated numerical technique corresponds to a multiblock approach. Nonlinear interactions are
expected to scramble the data in the outer part of the boundary layer and
to break possible periodicity.
The wall-model developed by Das and Moser within the framework of
embedded boundary conditions is presented on p. 349.
Deardor s Model. In the framework of a plane channel simulation with
innite Reynolds number, Deardor [172] proposes using the following conditions for representing the solid walls:
1
2 u1
2 u1
=
+
2
2
z
(z2 /2)
y 2
4
(10.24)
This conclusion must be considered together with the fact that the best results
obtained with wall-stress models are with non-zero transpiration velocity (see
suboptimal models, p. 345). A general conclusion is that even for impermeable
walls a non-zero wall-normal velocity must be prescribed to accurately describe
the near-wall dynamics.
338
Fig. 10.8. Representation of the o-wall approach. The grid lines are superposed
to a boundary-layer instantaneous ow obtained with a wall-resolving mesh (instantaneous velocity vectors and isolevels of streamwise velocity uctuation). Length
scales are expressed in wall units. The rst grid point is located at 100 wall units.
Top: view in an (x, z) plane. Bottom: view in an (y, z) plane.
10.2 Solid Walls
u3 = 0
2
339
(10.25)
u2
u2
=
2
z
x2
(10.26)
where z2 is the distance from the rst point to the wall and = 0.4 the
Von Karman constant. The rst condition assumes that the average velocity
prole veries the logarithmic law and that the second derivatives of the uctuation u = u u in the y and z directions are equal. The impermeability
condition (10.25) implies that the resolved stresses u1 u3 , u3 u3 and u2 u3 are
zero at the wall. This model suers from a number of defects. Namely, it
shows no dependency as a function of the Reynolds number, and assumes
that the shear-stress near the wall is entirely due to the subgrid scales.
Schumann Model. Schumann [653] has developed a wall model for performing a plane channel ow simulation at a nite Reynolds number. It is
based on the extended turbulent relation (10.23). Using dimensional analysis,
the eective viscosity can be evaluated using
e =
p
z2
2 u1 (x, y, z2 )
The resulting boundary conditions are:

u1 (x, y, z2 )
p,13 (x, y) =
p
u1 (x, y, z2 )
u3 = 0 ,

u3 (x, y, z2 )
2
p,23 (x, y) =
Re
z2
(10.27)
(10.28)
(10.29)
(10.30)
where designates a statistical average (associated here with a time average), and z2 the distance of the rst point to the wall. The condition (10.28)
is equivalent to adopting the hypothesis that the longitudinal velocity component at position z2 is in phase with the instantaneous wall shear stress. The
mean velocity prole can be obtained by the logarithmic law, and the mean
wall shear stress p is, for a plane channel ow, equal to the driving pressure gradient. This wall model therefore implies that the mean velocity eld
veries the logarithmic law and can be applied only to plane channel ows
for which the value of the driving pressure gradient is known a priori. The
second condition is the impermeability condition, and the third corresponds
to a no-slip condition for the transverse velocity component u2 .
Gr
otzbach Model. Gr
otzbach [278] proposes extending the Schumann
model to avoid having to know the mean wall shear stress a priori. To do
this, the statistical average is now associated with a mean on the plane
parallel to the solid wall located at z = z2 . Knowing u1 (z2 ), the mean wall
340
shear stress p is computed from the logarithmic law. The friction velocity
is computed from (10.11), i.e.:
u+
1 (z2 ) = u1 (z2 )/u =
1
log(z2 u /) + 5.5 0.1 ,
(10.31)
then p , by relation (10.4). This model is more general than Schumanns,

but it still requires that the mean velocity prole verify the logarithmic law.
Another advantage of Gr
otzbachs modication is that it allows variations of
the total mass ux through the channel.
Shifted correlations Model. Another modication of Schumanns model
can be made on the basis of the experimental works of Rajagopalan and
Antonia [614]. These two authors observed that the correlation between the
wall shear stress and the velocity increases when we consider a relaxation time
between these two evaluations. This phenomenon can be explained by the
existence of coherent inclined structures that are responsible for the velocity
uctuations and the wall shear stress. The modied model is expressed [595]:

u1 (x + s , y, z2 )
p,13 (x, y) =
(10.32)
p ,
u1 (x, y, z2 )

p,23 (x, y) =
u3 = 0
u2 (x + s , y, z2 )
u1 (x, y, z2 )
(10.33)

p ,
(10.34)
where the value of the length s is given by the approximate relation:
for 30 z2+ 5060

(1 z2 ) cot(8 )
s =
.
(10.35)
(1 z2 ) cot(13 )
for z2+ 60
Rough Wall Model. Mason and Callen [497] propose a wall model including the roughness eects. The three velocity components are specied at the
rst computation point by the relations:

u (x, y)
u1 (x, y, z2 ) = cos
(10.36)
ln(1 + z2 /z0 ) ,

u (x, y)
u2 (x, y, z2 ) = sin
(10.37)
ln(1 + z2 /z0 ) ,
u3 (x, y, z2 ) = 0 ,
(10.38)
where z0 is the roughness thickness of the wall and angle is given by the
relation = arctan(u2 (z2 )/u1 (z2 )). These equations can be used to compute
the friction velocity u as a function of the instantaneous velocity components
10.2 Solid Walls
341
u1 and u2 . The instantaneous surface friction vector u2 is then evaluated as:

u2 =
1
|u
|u
(10.39)
where u
is the vector (u1 (x, y, z2 ), u2 (x, y, z2 ), 0) and
1
1
= 2 ln2 (1 + z2 /z0 ) .
M
The instantaneous wall shear stresses in the x and y directions are then
evaluated respectively as |u2 | cos and |u2 | sin . This model is based on the
hypothesis that the logarithmic distribution is veried locally and instantaneously by the velocity eld. This becomes even truer as the grid is coarsened,
and the large scale velocity approaches the mean velocity.
Ejection Model. Another wall model is proposed by Piomelli, Ferziger,
Moin, and Kim [595] in consideration of the fact that the fast uid motions
toward or away from the wall greatly modify the wall shear stress. The impact
of fast uid pockets on the wall causes the longitudinal and lateral vortex lines
to stretch out, increasing the velocity uctuations near the wall. The ejection
of fast uid masses induces the inverse eect, i.e. reduces the wall shear stress.
To represent the correlation between the wall shear stress and the velocity
uctuations, the authors propose the following conditions:
p,13 (x, y) = p Cu u3 (x + s , y, z2 ) ,

p
u2 (x + s , y, z2 ) ,
p,23 (x, y) =
u1 (z2 )
u3 (x, y) = 0
(10.40)
(10.41)
(10.42)
where C is a constant of the order of unity, p is computed from the logarithmic law as it is for the Gr
otzbach model, and s is computed by the
relation (10.35).
Marusics Optimized Ejection Model. Piomellis ejection wall model
was futher improved by Marusic et al. [492] on the grounds of very accurate
wind tunnel experiments. The proposed generalization for relation (10.40)
based on the experimental correlations is
p,13 (x, y) = p u (u1 (x + s , y, z2 ) u1 (x, y, z2 ))
(10.43)
where is a parameter taken equal to 0.1 for zero pressure gradient ows.
This new ejection model, which is based on the streamwise velocity component instead of the wall-normal velocity in Piomellis original model, is found
to yield better results on priori tests carried out using experimental data: the
computed peak correlation coecient is in the range 0.340.53 for the new
model while it is between 0.19 and 0.24 for the original model.
342
Thin Boundary Layer Models. Balaras et al. [37] and Cabot [86, 87]
propose more sophisticated models based on a system of simplied equations
derived from the boundary layer equations. A secondary grid is embedded
within the rst cell at the wall (see Fig. 10.9), on which the following system
is resolved:

ui
ui
+
(u1 ui ) +
(u3 ui ) =
+
( + sgs )
, i = 1, 2 ,
t
x
z
xi
z
z
(10.44)
where z is the direction normal to the wall. Equation (10.44) can be recast
as an equation for the shear stresses i3 = ui /z, i = 1, 2:
ui
p
(( + sgs )
+
(u1 ui ) +
(u3 ui ) +
i3 ) =
i = 1, 2 . (10.45)
z
t
x
z
xi
Simplied models can be derived by neglecting some source terms in the
right-hand side of (10.45) or by approximating them using values from the
outer ow [90].
This approach is equivalent to assuming that the inner zone of the boundary layer behaves like a Stokes layer forced by the outer ow. Balaras et al.
propose computing the viscosity sgs by the simplied mixing length model:
sgs = (z)2 Db (z)|S| ,
(10.46)
where z is the distance to the wall, the Von Karman constant, and Db (z)
the damping function:

,
(10.47)
Db (z) = 1 exp((z + /A+ )3 )
Fig. 10.9. Representation of the primary and secondary grids.
10.2 Solid Walls
343
with A+ = 25. Cabot proposes the alternate denition:

2
sgs = us zDC
(z) ,
(10.48)
in which
DC (z) = (1 exp(zud/A))
(10.49)
where us and ud are velocity scales to be determined, and A = 19. The

simplest choice is us = ud = u .
Cabot and Moin [90] observed that the constant of the mixing length
model must be lowered in regard to its usual value in RANS computations in
order to account for resolved stresses. A dynamic evaluation of this constant
was achieved by Wang [751] to deal with ows with strong favorable/adverse
pressure gradient and incipient separation. The dynamic adjustment is performed by imposing that the mixing-length viscosity and the subgrid viscosity
are equal at the interface of the two simulations.
When this system is solved, it generates longitudinal and transverse velocity component distributions at each time step, so that the value of the wall
shear stress can be calculated for solving the ltered NavierStokes equations
on the main grid. The pressure gradient appears as a source term, because
this is obtained using the relation p/xn = 0.
The vertical velocity component is obtained from the continuity equation:

z2
u1
u2
(x, y, ) +
(x, y, ) d .
(10.50)
u3 (x, y, z2 ) =
x
y
0
The boundary conditions applied to the secondary system are:
On the solid wall: no-slip condition;
On the upper boundary: Dirichlet condition obtained from the value of the
velocity eld computed on the rst cell of the main grid.
Model of Murakami et al. Murakami, Mochida, and Hibi [556] developed
a wall model for dealing with the case of the separated ow around a cube
mounted on a at plate. This model is based on power-law solutions for the
mean longitudinal velocity prole of the form:
u1 (z) z n

Ue
(10.51)
The authors recommend using n = 1/4 on the at plate and n = 1/2 on

the cube surface. When the rst grid point is located close enough to the
wall, the following boundary conditions are used:

n
z2
ui (x, y) =
ui (x, y, z2 + z), i = 1, 2 ,
(10.52)
z2 + z
u3 (x, y) = 0
(10.53)
344
where z is the size of the rst cell. The rst equation is obtained by assuming
that the instantaneous prole also veries the law (10.51). When the distance
of the rst point from the wall is too large for the convection eects to be
neglected, the relation (10.53) is replaced by:
u3
=0
z
(10.54)
WernerWengle Model. In order to be able to compute the same ow as

Murakami et al., Werner and Wengle [757] propose a wall model based on
the following hypotheses:
The instantaneous tangential velocity components at the wall u2 (x, y, z2 )
and u3 (x, y, z2 ) are in phase with the associated instantaneous wall shear
stresses.
The instantaneous velocity prole follows the law:

z+
if z + 11.81
u+ (z) =
,
(10.55)
+ B
otherwise
A(z )
in which A = 8.3 and B = 1/7.
The values of the tangential velocity components can be related to the
corresponding values of the wall shear stress components by integrating the
velocity prole (10.55) over the distance separating the rst cell from the wall.
This allows a direct analytical evaluation of the wall shear stress components
from the velocity eld:
If |ui (x, y, z2 )|
2/(1B)
,
2zm A
then:
p,i3 (x, y) =
2ui (x, y, z2 )
z2
(10.56)
and otherwise:
p,i3 (x, y) =
(
1+B
1+B
ui (x, y, z2 ) 1 B 1B
A
|ui (x, y, z2 )|
2
z2
2
) 1+B
B
1+B
+
|ui (x, y, z2 )|
,
A
z2
(10.57)
where zm is the distance to the wall that corresponds to z + = 11.81. This

model has the advantage of not using average statistical values of the velocity
and/or wall shear stresses, which makes it easier to use for inhomogeneous
congurations. An impermeability condition is used to specify the value of
the velocity component normal to the wall:
u3 = 0
(10.58)
10.2 Solid Walls
345
WernerWengle-Type Ejection Model. A version of the WernerWengle

model which accounts for the shift that exists in the correlation between
the wall friction and the instantaneous velocity is proposed by Hassan and
Barsamian [298]. The authors recommend to account for this shift as in
Piomellis shifted model, leading to the following modications for relation (10.57)
(
1+B
1+B
ui (x + s , y, z2 ) 1 B 1B
A
p,i3 (x, y) =
|ui (x + s , y, z2 )|
2
z2
2
) 1+B
B
1+B
+
|ui (x + s , y, z2 )|
A
z2
Cu u3 (x + s , y, z2 ) ,
(10.59)
where C is a constant of the order of the unity and s is computed using

the relation (10.35). Equation (10.56) is kept unchanged, since the shift is
assumed to be negligible in the viscous sublayer.
Suboptimal-Control-Based Wall Models. The goal of this approach
is to provide numerical boundary conditions so that the overall error (to
be dened) is minimum in a given norm. The boundary conditions (wall
stresses and wall-normal velocity component) are used as a control to minimize a cost function at each time step. Many variants of this approach can
be dened [566, 33], considering dierent degrees of freedom at the boundary (i.e. dierent controllers), dierent cost functions and dierent ways to
evaluate the gradient of the cost function with respect to the controller.
Nicoud et al. [566] and Bagget et al. [33] considered a control vector on
the boundary whose components are the usual output of a wall-stress model:
= (p,13 , u3 , p,23 ) .
(10.60)
The control can then be exerted by modifying both the stress at the wall
and the wall-normal transpiration velocity.
The general form of the cost function J (u; ) is dened as
J (u; ) =
3

i=1
Jmean,i (u; ) +
3

i=1
Jrms,i (u; ) +
3

Jpenalty ()
, (10.61)
i=1
where the terms appearing on the right-hand side are, from left to right: the
part of the cost based on the mean ow, the part of the cost based on the rms
velocity uctuations, and a penalty term representing the cost of the control.
The mean-ow part of the cost function is typically a measure of the
dierence between the computed mean ow and a target mean ow uref . For
the plane channel ow, it can be expressed as
h
1
eu (z)2 dz ,
(10.62)
Jmean,i (u; ) = i
2h h i
346

1
(10.63)
eui (z) =
(ui (x, y, z) Uref,i (z))dxdy ,
A
where A is the surface of computational planes parallel to the walls and i
is an arbitrary weighting factor. The target mean ow can be prescribed
using experimental data, RANS simulations or theory. In a similar way, the
rms-based cost function is dened as
h
1
eu (z)2 dz ,
(10.64)
Jrms,i (u; ) = i
2h h i
with
with
1
eui (z) =
A

(ui (x, y, z) ui (z))2 u2
ref,i (z) dxdy
(10.65)
where ui is the average of the computed ow over homogeneous directions,

i is an arbitrary parameter, and u2
ref,i (z) are prescribed rms velocity proles.
Bagget et al. proposed the following form of the penalty term:

i
2ui dxdy +
i3 4u3 dxdy ,
(10.66)
Jpenalty () =
A z=h
A z=h
where i and are arbitrary constants. The last term of the penalty term
prevents the transpiration velocity from becoming too high.
Several tests have been carried out, dealing with dierent weights of the
three parts of the cost function and the possibility of having a non-zero
transpiration velocity. The main results are:
Suboptimal-control-based models lead to better results than usual wallstress models (see Fig. 10.10). An interesting feature of these models is
that they are able to break the spurious linear dependence of the predicted
wall stresses with respect to the instantaneous velocity at the rst o-wall
grid point. This is observed by looking at Figs. 10.11 and 10.12.
The use of a non-zero transpiration velocity makes it possible to improve
signicantly the mean ow prole compared to usual wall-stress models, if
the rms part of the cost function is not considered (i = 0). But, in that
case, rms velocity proles are not improved and prediction can even be
worse.
Rms velocity proles can be improved if both Jmean,i and Jrms,i are taken
into account, but the improvement of the prediction of the mean velocity
prole is less important than in the previous case. The use of the non-zero
transpiration velocity is also observed to be benecial.
This class of wall model based on the suboptimal control theory can be
considered as the best achievable wall-stress model, and it seems dicult
to get much better results controlling the same parameters (wall stresses and
transpiration velocity). Then, an interesting conclusion is that getting the
10.2 Solid Walls
347
Fig. 10.10. Large-eddy simulations of plane channel ow (Re = 4000) on a uniform 32 32 32 grid. Mean velocity prole: Symbols: 2.41 log(z + ) + 5.2; Solid line:
suboptimal wall stress model, without transpiration velocity; Upper dashed line:
suboptimal estimation of 13 , with shifted wall-stress model for 23 ; Lower dashed
line: shifted wall-stress model. Courtesy of F. Nicoud, University of Montpellier.
correct mean velocity prole and the correct rms velocities may be competing
objectives.
The preceding models are based on suboptimal control theory, and necessitate computing the gradient of the cost function. This involve a large
computational eort, whatever solution is adopted to compute the gradient
(nite dierences or solving the adjoint problem).5 The use of the incomplete gradient approach of Mohammadi and Pironneau [535] was observed to
yield poor results for this problem by Templeton et al. [707]. Following the
pioneering works of Bagwell et al. [34], Nicoud et al. [566] proposed a more
practical wall model based on linear stochastic estimation. This model is the
best possible least-square estimate of the suboptimal wall stresses as explicit
functions of the local velocity eld. It can be expressed as the conditional
average of the wall stress given the local velocity eld:
i,3 |E ,
(10.67)
where E is a vector of events containing the local instantaneous velocity.

This formal expression does not lead to a tractable wall model, and it is
approximated via a polynomial expansion. Restricting this expansion to the
5
The reported cost of a large-eddy simulation based on these models is 20 times

greater than that of a simulation on the same grid with explicit wall-stress models.
348
Fig. 10.11. Large-eddy simulations of plane channel ow (Re = 4000) on a uniform 32 32 32 grid. Shifted correlation model. Instantaneous isocontours of
predicted stresses and velocity components at the rst o-wall grid point. Courtesy
of F. Nicoud, University of Montpellier.
rst term, one obtains the following linear stochastic estimate for the wall
stresses:
(10.68)
i,3 |E Lij Ej ,
where the estimation coecients Lij are governed by
i,3 Ek = Lij Ej Ek .
(10.69)
In practice, these coecients are computed using reference data (direct

numerical simulation data or suboptimal-control-based prediction of the wall
stresses). The resulting model is an explicit model, whose cost is of the same
order as those of the other explicit wall-stress models, which are able to reproduce accurately the results of the suboptimally controlled simulations.
Unfortunately, numerical experiments have shown this model to be very sensitive to numerical and modeling errors, indicating that it might be impossible
to nd an accurate and robust6 linear wall stress model.
6
A model whose accuracy will be the same whatever numerical method and subgrid model are employed.
10.2 Solid Walls
349
Fig. 10.12. Large-eddy simulations of plane channel ow (Re = 4000) on a uniform 32 32 32 grid. Suboptimal prediction of the wall stresses, without transpiration velocity. Instantaneous isocontours of predicted stresses and velocity components at the rst o-wall grid point. Courtesy of F. Nicoud, University of Montpellier.
DasMoser Embedded Wall Model. In the case of an impermeable wall,

the general equations (10.1) and (10.2) simplify to

ui
p
uj
ui
=
(ui uj ) +
+
+
t
xj
xi
xj xj
xi

ui
uj
G(x ) p()
() +
() nj ()d , (10.70)
xj
xi
ui
=0 .
xi
(10.71)
The source term in the continuity equation cancels out, while the momentum source term is associated with the total force exerted on the wall. When
ltering through the wall, it is necessary to prolong the velocity eld inside
the solid wall. This is done by assuming a zero velocity eld inside the wall
in a buer zone .
350
In the present formulation, the problem is to compute the source term

in (10.70). Because the unltered velocity is set to zero in the wall, the
wall stress is the surface forcing required to ensure that momentum is not
transfered to a buer domain located inside the wall, i.e. that the velocity will
remain zero. Using this remark, Das and Moser [164] suggested computing
the source term at each time step in order to minimize the transport of
momentum from the uid domain toward the buer zone. The embedded wall
boundary condition is then a control minimizing the following cost function:
2

u
dx ,
J =
|u|2 +
(10.72)
t
where the rst term forces the energy in the buer zone to be small, and
the second one ensures that the momentum transfer of energy across the
wall surface is small. The constant scales like t2 in plane channel ow
computation, with t being the time step.
ODT-Based Wall Model. Structural models presented in Sect. 7.7 can be
used to reconstruct subgrid uctuations inside the rst cell near the solid
walls. This strategy was adopted by Schmidt et al. [650], who used the ODTbased model (see p. 257) to predict the near wall dynamics, the outer part
of the boundary layer being computed using a classical large-eddy simulation method. It is worth noting that this approach is reminiscent of the one
proposed by Balaras based on thin boundary layer equations within the rst
grid cell.
The proposed coupling strategy, based on the existence of an overlap
region between ODT and large-eddy simulation (see p. 257 for a detailed
description of the ODT-based model), is the following. This zonal decomposition is illustrated in Fig. 10.13. An ODT-type simulation is performed
within the rst grid cell of the large-eddy simulation grid. The large-eddy
simulation provides the ODT with boundary conditions at the top of the
ODT domain, and also appears in the evaluation of the advection speed in
the ODT equation. The ODT simulation provides the large-eddy simulation
with subgrid uxes in two dierent regions: (i) in the inner region (i.e. the
rst grid cell o the wall), the full uxes appearing in the large-eddy simulation equations is computed using the ODT uctuations and (ii) in the outer
region, which has a thickness L (L being the maximum size for eddy events
allowed in the ODT simulation within the inner region). In the outer region,
ODT equations are not solved, but the inuence of large events belonging
to the ODT solution computed in the inner region is taken into account.
This is achieved by adding the contribution of ODT eddy events of sucient
size to reach the considered position in the large-eddy simulation grid to the
conventional uxes in that region.
10.2 Solid Walls
351
Fig. 10.13. Schematic of the ODT-based wall model.
10.2.3 Wall Models: Achievements and Problems

Most of the wall models presented above have been developed on the grounds
of the dynamics of the zero-pressure gradient, equilibrium ate plate boundary layer. In practical cases, most of them exhibit the same behavior:
In attached, equilibrium ows, satisfactory results (skin friction predicted
within a 20% error) on the mean ow are recovered on medium grids, i.e.
on grids such that the rst cell o the wall has the following dimensions:
x+ 500,
y + 200 300,
z + = 50 150 .
On such a grid, spurious bumps are observed in the turbulence intensities

just near the rst grid point (see Fig. 10.14). These unphysical overshoots
are associated to the existence of large spurious streaky structures, whose
size can be governed by either the mesh size or the numerical and subgrid
model dissipation. The growth of these structures might be explained by
the same mechanisms as the physical streaky vortices: they might be parts
of an autonomous self-sustaining cycle feeded by the mean shear involving
low-/high-speed streaks and streamwise vortices. Another possible cause is
that they could arise from the splatting of turbulent eddies coming from
the outer part of the boundary layer on the boundary condition (almost
all models account for the impermeability constraint enforcing a vanishing
wall normal velocity).
No general cure for this problem is known. These bumps can be damped
by scrambling the spurious streamwise vortices by adding a random noise
in the region where they are detected [592]. Partial error cancellation when
using a very dissipative subgrid model has also been observed: the spurious
352
Fig. 10.14. Resolved Reynolds stresses in a plane channel ow with Gr

otzbach
wall stress model. Courtesy of Y. Benarafa and F. Ducros, CEA.
streaks are then damped. It was also shown by Nicoud that some control
on the turbulence intensities can be achieved playing of the wall stresses
and the wall normal velocity, but with a higher error level on the mean ow
prole. It is worth noticing that the problem of the reduction of this error
is close to the one of the active control of the boundary layer dynamics.
On very coarse grids, the models are no longer ables to yield an accurate
prediction of the mean ow prole: the correct logarithmic slope is not
recovered in the logarithmic layer, the skin friction is poorly predicted (see
Fig. 10.15), ... It must be observed that on very coarse grids the problem
is much more complicated: (i) the near-wall layer dynamics is not resolved,
(ii) large subgrid modeling errors occur in the core of the ow since a large
part of the turbulent kinetic energy is contained in subgrid scales (and most
subgrid scale models are not good at taking into account a large part of the
full turbulent kinetic energy), and (iii) since the mesh is coarse, numerical
errors may become dominant.
Most models are not very ecient at predicting separation, since they
are based on very stringent assumptions. This was observed, among other
studies, by Temmerman et al. [706] in a wavy channel conguration where
the location of the separation point (and therefore of the reattachment
point) is seen to be sensitive to the wall model. A noticeable exception is
the Thin Boundary Layer Model, which has been proved to yield a satisfac-
10.2 Solid Walls
353
Fig. 10.15. Error on the computed skin friction versus the size of the mesh in
the streamwise direction (in wall units) in plane channel ow computation, with
dierent wall models. Steady RANS solution is shown for comparison. Courtesy of
Y. Benarafa and F. Ducros, CEA.
tory prediction of an incipient separated region on smooth airfoil trailing

edge [754] and on a circular cylinder at high Reynolds number [113]. In
congurations where the separation point is imposed like in the backward
facing step case investigated by Cabot [87], main features of the ow are
relatively insensitive to the way the solid wall is treated inside the separated region. But skin friction in the seprated region is of course subject to
consequent errors. First trials for the denition of wall models devoted to
separated regions have been done [706, 370], but satisfactory results have
not yet been obtained. A main diculty is the lack of universal scaling law
for the mean velocity prole in separated ows.
Some wall-stress models require values related to the mean ow as inputs.
This need lead to a severe limitation of the models, since the mean ow
must be known, or some homogeneous directions must exists to enable
the evaluation of some statistiscal moments of the instantaneous elds at
each time step. Possible solutions to overcome this problem are: (i) to
run several statistically equivalent simulations in parallel and to perform
true statistical average or (ii) to use a local spatial averaging instead of
a statistical average. The latter solution is used by Hassan and Barsamian
[298] to obtain a localized version of the Gr
otzbach model.
354
10.3 Case of the Inow Conditions

10.3.1 Required Conditions
Representing the ow upstream of the computational domain also raises difculties when this ow is not fully known deterministically, because the lack
of information introduces sources of error. This situation is encountered for
transitional or turbulent ows that generally contain a very large number of
spacetime modes [631]. Several boundary condition generation techniques
are used for furnishing information about all the modes contained to the
large-eddy simulation computation.
Apart the purely mathematical problem of dening well-posed inow
boundary conditions, the Large-Eddy Simulation and the Direct Numerical Simulation techniques raise the problem of reconstructing the turbulent
uctuations at the inlet plane in an accurate way. The exact denition of
accurate turbulent inow conditions is still an open question, but the accumlulated experience proves that both kinetic energy and coherence of the inlet
uctuations must be taken into account to minimize the size of the buer
region that exists downstream the inlet plane, in which turbulent uctuations consistent with the NavierStokes dynamics are reconstructed by the
non-linear eects. This need is illustrated in Fig. 10.16, which displays results
obtained in Direct Numerical Simulation of a two-dimensional mixing layer.
10.3.2 Inow Condition Generation Techniques
Stochastic Recontruction from a Statistical One-Point Description.
When the freestream ow is described statistically (usually the mean velocity
eld and the one-point second-order moments), the deterministic information
is denitively lost. The solution is then to generate instantaneous realizations
that are statistically equivalent to the freestream ow, i.e. that have the same
statistical moments.
In practice, this is done by superimposing random noises having the same
statistical moments as the velocity uctuations, on the mean statistical prole. This is expressed as
u(x0 , t) = U (x0 ) + u (x0 , t) ,
(10.73)
where the mean eld U is given by experiment, theory or steady computations, and where the uctuation u is generated from random numbers. This
technique makes it possible to remain in keeping with the energy level of the
uctuations as well as the one-point correlations (Reynolds stresses) in the
directions of statistical homogeneity of the solution, but does not reproduce
the two-point (and two-time) spacetime correlations [432, 540, 493]. The
information concerning the phase is lost, which can have very harmful consequences when the consistency of the uctuations is important, as is the case
355
Fig. 10.16. Illustration of the inuence of the turbulent inlet boundary condition
(DNS of a 2D mixing layer). Iso-contours of instantaneous vorticity are shown.
Top: reference 2D simulation. Below: Truncated simulation using as inow conditions: a) exact instantaneous velocity eld stored at the x0 section; b) random
velocity uctuations spatially and temporally uncorrelated (white noise) having the
same Reynolds stress tensor components as in case (a); c) instantaneous velocity
eld preserving temporal two point correlation tensor of case (a); d) instantaneous
velocity eld preserving spatial two point correlation tensor of case (a); e) reconstructed velocity eld with the aid of Linear Stochastic Estimation procedure from
the knowledge of exact instantaneous velocity eld at 3 reference locations (center
of the mixing layer and /2 where is the local vorticity thickness). Courtesy
of Ph. Druault and J.P. Bonnet, LEA.
356
for shear ows (mixing layer, jet, boundary layer, and so forth). That is, the
computations performed show the existence of a region in the computational
domain in which the solution regenerates the spacetime consistency specic
to the NavierStokes equations [139]. The solution is not usable in this region, which can cover a large part7 of the computational domain, and this
entails an excess cost for the simulation. Also, it appears that this technique
prevents the precise control of the dynamics of the solution, in the sense that
it is very dicult to reproduce a particular solution for a given geometry.
A few ways to generate the random part of the inlet ow are now presented:
1. The LeeLeleMoin procedure (p. 356).
2. The SmirnovShiCelik procedure and Battens simplied version
(p. 356).
3. The LiWang procedure (p. 358).
4. The Weighted Amplitude Wave Superposition procedure (p. 358).
5. The digital lter based method (p. 359).
6. The Arad procedure (p. 360).
7. The YaoSandham model (p. 361).
The LLM Procedure. The rst method was proposed by Lee, Lele and
Moin [432] for a ow evolving in the direction x and homogeneous in the
two other directions, and statistically stationary in time. Assuming that the
energy spectrum of a ow variable , E , is prescribed in terms of frequency
and two transverse wave numbers, the Fourier coecients of the uctuating
part of are prescribed as
+
y , kz , , t) = E (ky , kz , ) exp [r (ky , kz , , t)] ,
(10.74)
(k
where r is the phase factor and = 1. The dependence of this phase

factor on time and transverse wave numbers is necessary so that the signal
generated is not periodic. The authors propose changing r only once in
a given time interval Tr at a random instance by a random bounded amount
r . The resulting signal is not continuous, and the frequency spectrum of
the generated turbulence is not equal to E .
The authors get satisfactory results for decaying isotropic turbulence by
applying this procedure to each uctuating velocity component ui , but its
application to advanced transitional ows or turbulent ows yields the occurrence of large non-physical transition regions.
The SSC Procedure. Another procedure was proposed by Smirnov, Shi and
Celik [677] with application to wall-bounded ows. It involves scaling and
orthogonal transformation operations applied to a continuous eld generated
as a superposition of harmonic functions.
7
Numerical experiments show that this region can cover more than 50% of the
total number of simulation points.
357
Let Rij = ui uj be the (anisotropic) velocity correlation tensor at the
inlet plane (see Appendix A for a precise denition). The rst step of the
SSC procedure consists of nding an orthogonal transformation tensor Aij
that would diagonalize Rij (without summation over Greek indices)
Ai Aj Rij = 2
Aik Akj = ij
(10.75)
(10.76)
where the coecients 1 , 2 and 3 play the role of turbulent uctuating

velocities u1 , u2 and u3 in the new coordinate system. It is worth noting that
both the transformation matrix and new coecients are functions of space.
The second step consists of generating a transient ow-eld V in a threedimensional domain and rescaling it. This eld is computed using the modied Kraichnans method:

N
#
2 " n
ai cos(kjn x
j + n t) + bni sin(kjn x
j + n t)
, (10.77)
Vi (x, t) =
N n=1
with
x
j =
xj
t
L0
c
, t =
, c=
, kjn = kjn
L0
T0
T0
j
n
n
ani = ijm jn km
, bni = ijk jn km
,
jn , jn , n
N (0, 1),
kin
N (0, 1/2) ,
(10.78)
(10.79)
(10.80)
where L0 and T0 are the length- and timescales of turbulence, ijm is the
permutation tensor, and N (M, ) is a normal distribution with mean M
and standard deviation . Quantities n and kjn represent a sample of n
frequencies and wave number vectors of the turbulent spectrum. In practice,
the authors use the following model spectrum:

2 4
k exp(2k 2 ) .
(10.81)
E(k) = 16
The last step consists of applying scaling and orthogonal transformations

to V to recover the synthetic uctuating eld in physical space:
W = V ,
ui = Aik Wk
(10.82)
The resulting uctuating eld is nearly divergence-free, and has correlation scales L0 and T0 with the correlation tensor Rij . This method was
successfully applied to boundary-layer ows.
A simplied formulation is proposed by Batten et al. [49], which does
not require the use of the orthogonal transformation. This simplication is
achieved redening the parameters j as
'
j
3Rlm klj km
j =
,
(10.83)
j j
2kn kn
and using the relation (10.77) to prescribe directly the velocity uctuations
u instead of V .
358
The LiWang Procedure. Another random generation technique for uctuations in a boundary-layer ow was proposed by Li and Wang [446]. Fluctuations are reconstructed using the following equation:
u (x, y, z, 0) =
N1
N2
N3 +

2
Eu (xl , y m , z n )x y z
l=1 m=1 n=1
cos(x l x + y m y + z n z + lmn ) ,
(10.84)
where Eu is the target spectrum, lmn a random phase with uniform distribution, and x l = (l 1)x the angular frequency in the x direction. The
periodicity is eliminated by dening
x l = x l + x
(10.85)
where x is a small random frequency. The time-evolving uctuating eld

at the inlet plane is then reconstructed using Taylors frozen turbulence hypothesis:
u (0, y, z, t) = u (x , y, z, 0),
x = Uc t ,
(10.86)
where Uc is an advecting velocity.

Weighted Amplitude Wave Superposition (WAWS) spectral representation
method. Another procedure relying on modied random time series to generate velocity uctuations was proposed by Glaze and Frankel [265]. This
method, referred to as the WAWS method, is capable of simulating both
spatial and temporal correlation. It is based on the regeneration of the uctuating signal from its cross-spectral density at the inlet plane. As a result,
both spatial correlation across the inlet plane and the power spectrum of each
velocity component can be enforced.
Let M and N be the number of grid points at the inlet plane and the
number of frequencies to be prescribed, respectively. Each velocity component
is synthetized at the ith grid point using the following relation
i
N

|Him (n )| 2 cos(n t + im (n ) + mn ) , (10.87)
ui (t) = 2
m=1 n=1
with
n =

1
n
,
2
n = n + n
(10.88)
where = u /N , u and n [ /20, + /20] are the frequency

resolution, the maximum frequency and a small random perturbation, respectively. The parameters mn are random phases perturbations uniformly
distributed between 0 and 2. The key parameters are the components of
the transfer function matrix, Him (n ), which are linked to the cross spectral
359
density matrix components Sij (n ) through the relation
Sij (n ) = Hik (n )Hkj

(n )
(10.89)
where H is the Hermitian transpose of H. Since the H matrix is not unique

for a given S, a method should be chosen to compute it. Glaze and Frankel
used a Cholesky decomposition to obtain a lower-triangular matrix, leading
to simple calculations. The im (n ) parameter in (10.87) are dened writing
the transfer function matrix components in polar form:
Him (n ) = |Him (n )|eim (n )
(10.90)
Since the foreknowledge of the cross-spectral density matrix S for each

velocity component is not a realistic requirement in practice, the next step
consists in modeling it from available data. Noticing that it can be expressed
as a function of one-point spectral density and the complex coherence function
im :

Sim (n ) = im Sii (n )Smm (n ) ,
(10.91)
the problem is equivalent to prescribing the power spectrum at each grid point
and the coherence function. This is achieved using informations available on
each ow.
The case of the near-eld of a turbulent jet studied in [265] is given below
as an illustration. The coherence function is estimated like follows

2
A(rim + Brim
)
, A = 1, B = 4 ,
(10.92)
im = exp
Uim
where rim is the distance between points i and m and Uim is the average
mean velocity between these points. The power spectrum is obtained using
the von Karman model:
Sii (f ) =
4f(u )2
f (1 + 70.8f2)5/6
(10.93)
where u is the rms turbulence intensity, f the frequency (in hertz), f =

Lu f /U the associated Strouhal number, with U the mean velocity and Lu
the turbulent integral scale.
Digital Filter Based Method. Klein, Sadiki and Janicka [396] introduced
a new approach based on signal modeling through the use of linear nonrecursive lters. The general form of the discrete time series for the u velocity
component at any grid point of the inow plane is
um =
N

n=N
bn rm+n
(10.94)
360
where rm is a series of random data with zero mean and such that rm rn =
mn and bn are the digital lter coecients. The autocorrelation of the synthetized signal is
um um+k
=
um um
=N
j=N +k bj bj+k
=N
2
j=N bj
(10.95)
The problem consists in inverting this relation to compute the coecients

of the digital lter associated to a given autocorrelation tensor. The authors
propose to use a model autocorrelation function to obtain a simple form of
the lter coecients. They use the following form, which is valid for fully
developed homogeneous turbulence (autocorrelation of the u component in
the direction associated to it):
2
r

Ru u (r, 0, 0) = exp
,
(10.96)
4L2
where L is a prescribed integral length scale. Combining this model autocorrelation function with relation (10.95) and setting L = nx, one obtains
=N
j=N +k bj bj+k
=N
2
j=N bj

(kx)2
k 2
= exp
=
exp
4(nx)2
4n2
whose accurate approximate explicit solution is

ak
k 2
bk = +=
, ak = exp 2
N
4n
2
a
j=N k
(10.97)
(10.98)
This method yields the generation of a set of values with the targeted
autocorrelation. Cross-correlations between velocity component can be enforced using the same change of variable as in the SSC procedure. Extension
to the three-dimensional case straightforward, applying the procedure in sequentially in the three directions. This method was shown to give satifactory
results in plane jet simulations.
Arads Procedure. Arad [16] proposed a reconstruction technique for an initial
condition based on physical observations related to the turbulence production
process in a boundary layer. Assuming a general form of the perturbation
corresponding to linearly unstable modes of the mean prole,
i (z) exp((x + y t)),
ui (x, y, z, t) = u
u
i (z) = Ai exp(z 2 ) ,
(10.99)
where Ai is the amplitude of the mode and a random number. This idea
of Arad is to design the uctuations so that their growth rate will be maximized, resulting in a short unphysical transient region near the inlet plane.
361
Remarking that in a boundary layer the turbulence production term has the
following form
u w dU +
P =
,
(10.100)
u2 dz +
and taking into account the fact that quadrant Q2 (ejection: u < 0, w > 0)
and Q4 (sweep: u > 0, w < 0) events have a dominant contribution to the
shear stress u w , Arad proposed introducing a phase shift between u and
w :
u (x, y, z, t) =
w (x, y, z, t) =
u
(z) cos(x + y t) ,
u
(z) cos(x + y t + )
(10.101)
(10.102)
The spanwise component uctuation, v , is assumed to be in phase with u

in Arads work. The resulting uctuations correspond to sweep and ejection
events.
The YaoSandham Procedure. A more sophisticated procedure was proposed
by Yao and Sandham [781, 644], which relies on the observations that uctuations in the inner and outer parts of the boundary layer have dierent
characteristic scales. As a consequence, specic disturbances are introduced
in each part of the boundary layer. The inner-part uctuations, uinner are
+
:
designed to represent lifted streaks with an energy maximum at zp,j
+
(y, z, t) = cij exp(z + /zp,j
) sin(j t) cos(ky,j y + j )
uinner
i
(10.103)
The outer-part uctuation is assumed to be of the following form (with

a peak at zp,j )
(y, z, t) = cij
uouter
i
z
zp,j
exp(z/zp,j ) sin(j t) cos(ky,j y + j )
(10.104)
where subscripts i = 1, 2 and j are related to the velocity component and

to the mode indices, respectively, and cij are constants. The + superscript
refers to inner coordinates (wall units). The j are phase shifts, j are forcing
frequencies, and ky,j are spanwise wave numbers. These parameters are to be
adjusted using information on the boundary-layer dynamics. The spanwise
velocity component is deduced from the continuity constraint.
In the inner region of the boundary layer, it is assumed that the disturbances travel downstream for a distance of 1000 wall units at a convective
velocity Uc 10 u , where u is the friction velocity within a time period.
The wave numbers ky,j are chosen such that there will be four streaks with
a typical characteristic length of 100 wall units.
In the outer region, the downstream travelling distance is taken equal to
16 and the convection velocity is Uc 0.75 U, where U is the external velocity. The spanwise wave number is chosen to be of the order of the spanwise
extent of the computational domain.
362
Yao and Sandham applied this procedure to a turbulent boundary layer,

taking one mode in the inner region and three in the outer region. Corresponding parameters are given in Table 10.1. They also add a random noise
with a maximum amplitude of 4% of the external velocity to prevent possible
spurious symmetries.
Table 10.1. Coecients of the four-mode YaoSandham model of uctuations for
boundary-layers.
inner region
outer region
outer region
outer region
c1j
c2j
ky,j
+
zp,j
zp,j
0
1
2
3
0.1
0.3
0.3
0.3
0.0016
0.06
0.06
0.06
0.1
0.25
0.125
0.0625
0.75
0.5
0.25
0.
0.
0.1
0.15
12
1.
1.5
2.0
Deterministic Computation.
Precursor Simulation. One way of minimizing the errors is to perform a simulation of the upstream ow [740, 226, 649], called a precursor simulation,
with a degree of resolution equivalent to that desired for the nal simulation
(see Fig. 10.17). This technique almost completely eliminates the errors encountered before, and oers very good results. On the other hand, it is hardly
practical in the general case because it requires reproducing the entire history
Fig. 10.17. Schematic of the precursor simulation technique. A precursor simulation of an attached boundary layer ow is performed. An extraction plane is
dened, whose data are used as an inlet boundary condition for a simulation of the
ow past a trailing edge.
363
of the ow which, for complex congurations, implies very high computation

costs. Another problem stemming from this approach is that of causality:
since the precursor is computed separately, no feedback of information from
the second simulation is possible. This is a one-way coupling between two
simulations that can become problematic when a signal (acoustic wave, for
example) is emitted by the second.
Li et al. [445] reduced the cost of the precursor technique by storing the
results of the precursor simulation over a (relatively) short time, and cycling
in time over these data to generate the ow at the inlet plane of the main
computation. In practice, the precursor results are stored over a time of the
order of the integral timescale of the ow, and windowed to get a periodic
signal. This technique has been applied to a plane mixing layer ow. Numerical results show that non-linear interactions quickly eliminate the spurious
periodicity imposed at the inlet plane (about 25% of the total computational
domain is contaminated). The eciency of the method for ows with lower
scrambling eects, such as wall-bounded ows, remains to be investigated.
Lunds Extraction/Rescaling Technique. Lund et al. [463] developed a variant
of the precursor approach for boundary layers, in which the information at
the inlet plane is produced from that contained in the computation. There
is no longer any need for a precursor. The heart of the method is a means of
estimating the velocity at the inlet plane, based on the velocity eld extracted
from the simulation on a plane downstream, as illustrated in Fig. 10.18.
The main diculty arises from the fact that the mean ow is not parallel,
i.e. the boundary layer thickness increases, and the ow at the extraction
plane must rst be rescaled before being used at the inlet plane.
The rst step consists of decomposing the extracted ow, ue (x, t), as the
sum of a mean and a uctuating part:
e
uei (x, y, z, t) = ue
i (x, y, z, t) + Ui (y, z) .
(10.105)
The second step consists of rescaling the mean ow part using classical
scalings related to the mean velocity prole of the turbulent boundary layer
(see Sect. 10.2.1). In practice, the rescaling is carried out according to the
law of the wall in the inner region and the defect law in the outer region
of the boundary layer, leading to the following relations for the streamwise
component:
U inner
U U outer
=
=
u (x)f1 (z + ) ,
u (x)f2 () ,
(10.106)
(10.107)
where x is assumed to be the streamwise direction, z the wall-normal direction, U the external velocity, u the friction velocity, = z/ the outer coordinate, the boundary-layer thickness, and f1 and f2 two universal functions
to be determined. These two scaling laws dictate that the extracted mean
364
Fig. 10.18. Schematic of Lunds extraction/rescaling technique. Instantaneous

isolevels of streamwise velocity in a boundary layer are shown. Courtesy of
E. Tromeur and E. Garnier, ONERA.
velocity U e and the rescaled mean velocity at the inow, U r , are related in
the inner and outer regions via
U r,inner
U e (z +,r ) ,
U r,outer
U e ( r ) + (1 )U
(10.108)
,
(10.109)
with
=
ur
ue
(10.110)
and where ur and ue are skin friction at the inlet plane and extraction plane,
respectively, z +,r is the inner coordinate computed at the inlet plane, and r
is the external coordinate computed at the inlet plane. A linear interpolation
is used between the grid points of the planes.
A similar technique is used to rescale the mean wall-normal component,
yielding:
W r,inner
W
r,outer
=
=
W e (z +,r ) ,
e
W ( ) .
(10.111)
(10.112)
365
The third step consists of rescaling the uctuating part of the instantaneous eld:
ui,r,inner
= u,e (y, z +,r , t) ,
(10.113)
ui,r,outer
= u,e (y, r , t) .
(10.114)
The last step consists of writing a composite prole for the full instantaneous velocity at the inlet plane, ur , that is approximately valid over the
entire boundary layer. It is dened as a weighted average of the inner and
outer proles
#
"
uri = Uir,inner + ui,r,inner (1 ( r ))
3
4
+ Uir,outer + u,r,outer
( r ) ,
i
(10.115)
with
() =

1
( b)
1 + tanh
/tanh()
2
(1 2b) + b
(10.116)
where = 4 and = 0.2.

This extraction/rescaling technique is observed to be ecient in practice,
but must be used with care. The rst point is that the extraction plane must
be located far enough from the inlet to prevent spurious couplings in the
computed solution. This constraint is satised by taking a distance between
the two planes larger than the correlation length of the uctuations in the
streamwise direction. The second point is that it is valid for fully turbulent
self-similar boundary layers only, and that the scaling laws must hold to
obtain a relevant procedure.
Spille-KohoKaltenbach Method. The extraction/rescaling technique presented above suers some lack of generality, because it relies on self-similarity
assumptions and can introduce some spurious couplings inside the computational domain. A more general method was proposed by Spille-Koho and
Kaltenbach [686], with application to a boundary-layer.
The core of the method is the denition of a buer region, referred to
as the control region, near the inlet plane, where a body force is adjusted
in order to recover targeted proles of turbulent uctuations at a position
located downstream of this buer region. The body force is dened within
the closed-loop control theory, and makes it possible, at least theoretically,
to control both rms proles and integral properties of the boundary layer.
This procedure is illustrated in Fig. 10.19.
A random uctuation is rst specied at the inlet plane, together with
a mean velocity prole. The body force is applied to the wall-normal velocity
component only. The rationale for this is the observation that w w dU/dz
366
Fig. 10.19. Schematic of the Spille-KohoKaltenbach method. Courtesy of

H. Kaltenbach, University of Berlin.
is the dominant production term in the balance equation for the shear stress
u w . The amplitude of the body force at a streamwise position x0 is
adjusted via a PI controller in order to achieve a prescribed shearstress
prole u w target . This target can be dened using experimental data or
RANS simulations. Another way to dene the targeted prole is to extract
and rescale it from a position xR downstream of the control region.8
The instantaneous body force in the plane x = x0 is computed as follows
f (x0 , y, z, t) = A(z, t) [u(x0 , y, z, t) uy,t (x0 , z)]
8
(10.117)
It is important to remark that only mean proles are extracted, and not instantaneous elds as in Lunds approach. This prevents the occurrence of spurious
feedback.
367
where the amplitude A(z, t) is given by

A(z, t) = e(z, t) +
e(z, t )dt
(10.118)
where and are two arbitrary parameters. The averaging operator y,t
is associated with the average in the spanwise (homogeneous) direction and
in time over a sliding window of width equal to O(10)/U . The error term
e(z, t) is a measure of the dierence between the computed and the prescribed
shear stress:
e(z, t) = u w target (x0 , z) u w y,t (x0 , z, t) .
(10.119)
In order to prevent unphysically large values of the shear stress, the body
force is applied only at grid points satisfying the following four instantaneous
constraints:
|u | < 0.6 U ,
|w | < 0.4 U ,
u w < 0,
2
|u w | > 0.0015 U
.
(10.120)
If the extraction technique is used to specify target values inside the control region, another closed-loop controller is used to control the boundarylayer thickness at the inlet plane until the target value is obtained at the
extraction plane. Similarly, the wall-normal velocity component at the top of
the computational domain is controlled in order to obtain the desired streamwise pressure gradient. In this case, the authors observed that the adjustment
time for reaching statistically steady values is of the order of 100 /U .
Semideterministic Recontruction. Bonnet et al. [64, 199] propose an

intermediate approach between the two previous ones, to recover the twopoint correlations of the inow with no preliminary computations. The signal
at the inow plane is decomposed in the form
u(x0 , t) = U (x0 ) + Uc (x0 , t) + u (x0 , t) ,
(10.121)
where U (x0 ) is the mean eld, Uc (x0 , t) the coherent part of turbulent uctuations, and u (x0 , t) the random part of these uctuations. In practice, this
last part is generated by means of random variables and the coherent part
is provided by a dynamical system with a low number of degrees of freedom
(like the POD, as seen in the Introduction), or by linear stochastic estimation,
which gives access to the two-point correlations.
11. Coupling Large-Eddy Simulation

with Multiresolution/Multidomain Techniques

This chapter is devoted to the presentation of the coupling of large-eddy
simulation with multiresolution and/or multidomain approaches. The main
purpose of these couplings is to decrease the computational cost of the largeeddy simulations by clustering the degrees of freedom in regions of interest.
The key idea is to adapt locally the cuto length scale of the simulation, i.e.
to rene the computational grid.1 This grid renement is associated with the
denition of dierent subdomains with varying resolution.2
The methods proposed by various research groups can be classied as
follows:
Methods relying on fully overlapping subdomains (Sect. 11.2), as shown
in Fig. 11.1. The term fully overlapping means here that the subdomain
with the nest resolution is totally embedded within the coarsest resolution
subdomain.
Fig. 11.1. Multiresolution decomposition with full overlap.
This corresponds to h-adaptivity within the framework of hp methods. We will

focus on h-adaptivity only, because these are the most employed methods, even
within the nite-element framework.
The problem of mesh adaptation on unstructured grids for large-eddy simulation
will not be discussed here, because it has not yet been treated.
370
11. Multiresolution/Multidomain LES Techniques
The full overlap feature makes it possible to dene two dierent strategies:
Cycling between the dierent grid resolutions, which will be considered
as a time-consistent extension of the multigrid acceleration technique
for steady simulations. The emphasis is put here on the fact that the
cycling strategies discussed below are not associated with convergence
acceleration for implicit methods for unsteady simulations but are based
on consistent time-integration at each grid level.
Global resolution methods, in which time integration on all the subdomains is carried out at each time step.
Methods with partial or no overlap between the subdomains (Sect. 11.3),
as illustrated in Fig. 11.2.
Fig. 11.2. Partial-overlap conguration.
From mathematical and physical points of view, the underlying problem

can be interpreted as coupling two solutions obtained with dierent lter
kernels and dierent cuto length scales. An idealized problem with two
domains is discussed below, which shows the full complexity of the problem.
Let us note G1 and G2 , the lter kernels associated with the ne and
coarse resolution levels, respectively. The associated cuto length scales are
1 and 2 . The domains with ne and coarse resolutions are referred to as
1 and 2 , respectively.
The solution is decomposed as
u = u1 + u1 in 1 ,
u = u2 + u2 in 2
(11.1)
The key problem is the transfer of information between 1 and 2 . Let

v 12 be the complementary eld dened as
v 12 = u1 u2 = G1
u G2
u = (G1 G2 )
u = G12
u
(11.2)
11.2 Methods with Full Overlap
371
where G12 is the restriction operator governing the transfer from the ne to
the coarse resolution level.
The problem of interfacing the two domains is thus the following:
Transfer from 1 to 2 : restrict u1 on 2 , or equivalently substract v 12
from u1 . This corresponds to the denition of the new ltering operator
G21 such that
G2
u = G21
G1
u .
(11.3)
Transfer from 2 to 1 : delter u2 at the 1 level, or equivalently add
v 12 to u2 .
It is worth noting that most methods presented in this chapter can be
dened as static methods, because the number of levels of resolution is arbitrarily xed before the computation. Results dealing with dynamic methods,
in which the number of levels is not xed but automatically adjusted, are
very rare. Most advanced results dealing with the coupling of large-eddy simulation with Adaptive Mesh Renement (AMR) are given in Sect. 11.4.

According to the notation used in the previous section, the full overlap corresponds to 1 2 = 1 .
The methods described in this section are:
Methods with separate time-integration at each level. These methods are
true multidomain methods, in the sense that the solution is integrated
separately on each subdomain. The solutions are coupled via information
transfer from time to time. The methods presented below are:
1. The one-way coupling procedure of Khanna and Brasseur (Sect. 11.2.1).
This method is the simplest one, and represents the minimal degree
of coupling between the two domains, the coarse level solution being
independent of the ne level solution.
2. The two-way coupling procedure of Sullivan et al. (Sect. 11.2.2). This
method can be seen as an extension of the previous one, because the
information transfer is now taken into account in both ways: from ne
to coarse level, and from coarse to ne level.
3. The multilevel of Terracol et al. (Sect. 11.2.3), which is the most complete
one. It relies on a two-way coupling, and also incorporates a dynamic
cycling strategy. It is also the only one to incorporate a specic subgrid
modeling for multilevel computations.3
Methods with a single time-integration step. These methods are not true
multidomain techniques, but rather multiblock methods. Time-integration
3
This multilevel method is an extension of the multilevel closure presented in

Sect. 7.7.7 (p. 271).
372
is carried out at the same time at each level, without any distinction
between the dierent resolution levels. The example given below is the
method proposed by Kravchenko et al. (Sect. 11.2.4), which is based on
a Galerkin method with overlapping trial functions.
11.2.1 One-Way Coupling Algorithm
Khanna and Brasseur [365] developed a one-way coupling embedded grid
technique for large-eddy simulation of atmospheric boundary layers. The ne
resolution (i.e. ne grid) domain 1 is located in the near-wall region, in
order to permit a more accurate capture of details of the ow in that zone.
The coarse resolution domain is noted 2 .
The solution is integrated in each domain independently for an arbitrary
time. The coupling is enforced by imposing boundary conditions on 1 using
data coming from the coarse solution 2 .
The proposed boundary conditions are the following:
Dirichlet conditions on the velocity components:

u1 = u2 + v 12 ,
1
(11.4)
where the coarse resolution eld u2 is obtained on the boundary of 1 , 1

by a simple linear interpolation procedure. The complementary eld v 12
dened by relation (11.2) is here approximated by adding random noise perturbation following a k 5/3 law within the spectral band k [/2 , /1 ].
Rescaling of the subgrid viscosity. The second boundary condition is applied to the subgrid viscosity, which is rescaled in order to take into account
the grid renement. Assuming that both cutos are within the inertial subrange, (5.36) leads to the following rescaling law for the subgrid viscosity
at the boundary
1 4/3

1
2
sgs
,
(11.5)
sgs =
2
1
1
1
2
where sgs
and sgs
are the values of the subgrid viscosity in 1 and 2 ,
respectively.
In practice, the coupling is operated at the end of each time step.

11.2.2 Two-Way Coupling Algorithm
A two-way coupling procedure was proposed by Sullivan, McWilliams and
Moeng [698] (also used in [62]). These authors consider a set of nested grids
with increasing resolution for simulating the planetary boundary layer. The
presentation of the method is restricted to a two-grid case for the sake of
simplicity. The extension to an arbitrary number of grids is straightforward.
373
The coupling is achieved at each time step in the following way:

From the coarse resolution level to the ne resolution level: boundary conditions on 1 are obtained by interpolating the low-resolution eld u2
(see (11.4)). The dierence between this and the one-way coupling algorithm presented above is that the complementary eld v 12 is now neglected.
From the ne resolution level to the coarse resolution level: the numerical
uxes at the coarse resolution level are computed by ltering the uxes
computed at the ne resolution level on the overlap region. This can be
written as follows:
u2
+ G21
N S(u1 ) = 0, in 1 2
t
(11.6)
where N S denotes the NavierStokes operator. Another possibility, as proposed by Manhart and Friedrich [483, 482] for direct numerical simulation,
is to replace the coarse velocity eld with the restricted nely resolved
velocity eld:
u2 = G21
u1 in 1 2 .
(11.7)
The data transfer is done at each time step.
11.2.3 FAS-like Multilevel Method
We now discuss the most general approaches, which can be seen as generalizations of the Full Approximation Scheme (FAS) classical multigrid acceleration
technique. The method proposed by Terracol et al. [633, 711, 710, 712] appears as the most general one. It deals with the use of an arbitrary number of
nested resolution levels, and the gain in computational time is optimized by
considering a self-adaptive time-cycling strategy between the dierent levels.
Other FAS-type methods have been proposed by Voke [736], Tziperman
et al. [722] and Liu et al. [452, 453].
Using the multilevel framework developed in Sect. 7.7.7 (p. 271), the
governing equations at each resolution level can be expressed as
un
+ N S(un ) = n = [Gn
, N S](u),
t
n = 1, 2 .
(11.8)
Equation (11.8) shows that the coupling between the dierent levels is
theoretically achieved through the generalized subgrid term which arises on
the right-hand side. This is seen by decomposing the commutation error at
the coarse resolution level as
2
[G2
, N S](u)
=
=
G21
G1
N S(u) N S(G21
G1
u)
G21
(N S(G1
u) + [G1
, N S](u)) N S(G21
u1 )
G21
N S(u1 ) + G21
1 N S(u2 ) .
(11.9)
374
The rst and last terms on the right-hand side of (11.9) correspond to
the direct coupling term between the two levels of resolution. It is worth
noting that it is equivalent to the so-called forcing function in FAS multigrid
methods. The remaining term represents the coupling between unresolved
scales (true subgrid scales) and the coarsest level of resolution. It is the only
term which requires a physical modeling, all the other terms being directly
computable.
On the grounds of the preceding developments, Terracol proposed a multilevel algorithm whose main elements are:
The subgrid term at level 2 is computed according to relation (11.9).
The subgrid term at the nest resolution level can be computed with any
subgrid model. In Lius multigrid approach, the subgrid term is replaced by
a numerical stabilization term, which is essentialy equivalent to low-pass
ltering.
Time cycling is optimized by freezing the velocity eld at the nest level
during a time T . During this period, time integration is carried out at the
coarse resolution level only. The coupling from the ne to the coarse level
is carried out when evaluating 2 . The coupling from the coarse to the ne
level is performed by refreshing the low-frequency part of u1 at the end of
the period T using the new value of u2 :
u1 (t + T ) = u2 (t + T ) + v 12 (t) .
(11.10)
The time scale T is evaluated in order to satisfy the following constraint:

|v 12 |2

max u1 u1
T
,
(11.11)

2
t 2
where max is a prescribed error tolerance. Typical values for this parameter
range from 104 to 103 . The resulting self-adaptive cycling obtained in
a plane mixing layer conguration is illustrated in Fig. 11.3. These bounds
are experimentally observed to allow a decrease of the cost by a factor N
for an N -grid simulation for both wall-bounded and free shear ows. Vokes
and Lius methods are based on static cycling strategies, with empirically
determined integration times at each level. Typical results are illustrated
in Fig. 11.4, which displays the computed time evolution of the momentum
thickness of the mixing layer obtained with dierent subgrid closures at the
nest resolution level.
11.2.4 Kravchenko et al. Method

A zonal embedded grid technique for wall bounded ows was proposed by
Kravchenko et al. [411]. Contrary to the other methods with full overlap
presented in this section, it does not rely on a separated time integration
375
Fig. 11.3. Terracols self-adaptive time-cycling algorithm. Two-grid simulation of

a plane mixing layer ow. Time history of the number of consecutive time-steps on
the coarse grid. Courtesy of M. Terracol, ONERA.
Fig. 11.4. Large-eddy simulation of a plane mixing layer using Terracols selfadaptive time-cycling algorithm with two resolution levels. Time evolution of the
momentum thickness for dierent subgrid closures at the nest resolution level.
Crosses correspond to classical large-eddy simulation on the coarse level. Courtesy
of M. Terracol, ONERA.
376
on each grid. The main element of the method is the use of a Galerkintype numerical method with wide-stencil trial and weighting functions (fully
non-local, such as Fourier basis, or B-splines). The transfer of information
is intrinsically achieved thanks to the fact that the basis functions are not
orthogonal.
In fact, this method can be interpreted as a particular case of the implementation of Galerkin methods on unstructured grids.
11.3 Methods Without Full Overlap

We now describe the methods without full overlap, i.e. methods designed for
the case where the ne-resolution domain 1 is not totally embedded into the
coarse-resolution domain 2 : 1 2 = 1 . The partial overlap precludes the
use of multigrid-type algorithms like those presented in the previous section,
and time integration must be carried synchronously in each domain. The data
transfer is done at each time step (or substep for multistep schemes) in order
to provide refreshed boundary conditions to the subdomains.
Quemere et al. [612] carried out an extensive analysis of the interfacing
problem between subdomains with dierent resolutions within the large-eddy
simulation framework. The main results of their analysis are the following:
From ne to coarse resolution subdomain: boundary conditions are obtained by ltering the data (point values or uxes):

u2 = G21 u1 , and/or G2
N S(u)| = G21
G1
N S(u)| ,
(11.12)
with
(11.13)
= 1 (1 2 ) .
From coarse to ne resolution subdomain: the denition of boundary conditions for u1 requires two successive steps:
1. Interpolation of u2 at the ne resolution level 1 . This step is easy to
implement, and does not induce specic theoretical problems.
2. Reconstruction of the complementary eld v 12 . If corresponds to an
exit boundary for the subdomain 1 , numerical experiments show that
the enrichment step is not mandatory and v 12 = 0 can be used. On the
contrary, if is an inow boundary condition for 1 , the reconstruction
of v 12 is necessary, and the enrichment step is theoretically equivalent
to the denition of turbulent inow conditions (see Sect. 10.3). All the
methods proposed for generating inow turbulence can be used to predict v 12 . A simple extrapolation procedure from the interior of 1 can
also be used, if the elementary convection length Uc t (with Uc the
characteristic advection velocity across and t the time step) is much
smaller than the local correlation length scale of turbulent uctuations.
11.4 Coupling Large-Eddy Simulation with Adaptive Mesh Renement
377
Quemere et al. emphasized the fact that a weighted extrapolation can

be used to account for the spatial variation of the mean Reynolds stress
proles.
3. Rescaling of the subgrid models. The coupling can also be completed
by enforcing some compatibility conditions on the subgrid terms at the
interface. Inertial range arguments make it possible to derive proportionality factors. The following technique for rescaling subgrid viscosity
models was proposed in [612]. Starting with the classical relations

1
2
sgs
1 u1 u1 , sgs
2 u2 u2 ,
(11.14)
and assuming the orthogonality property
u2 u2 = (u1 + v 12 ) (u1 + v 12 ) u1 u1 + v 12 v 12
we obtain the following relation

2
2
sgs
1 u1 u1 + v 12 v 12
1
(11.15)
(11.16)
From this last relation, we deduce a scaling law between the two viscosities
2

2

2
sgs
v 12 v 12
2
1+
,
(11.17)
1
sgs
u1 u1
1
which can be used at the interface.
11.4 Coupling Large-Eddy Simulation with Adaptive

Mesh Renement
The theoretical advantage of coupling with Adaptive Mesh Renement
(AMR) is twofold:
Decrease in the mesh size x results in a reduction of the discretization
error, yielding an improved numerical accuracy.
If the cuto length is tied to x, mesh renement will also induce
a decrease in and make the subgrid model less inuential on the results
(since a larger part of the exact solution is directly captured).
The philosophy of the AMR approach is that the grid renement must be
performed automatically during the computation, leading to the denition of
several open problems:
378
Finding an error estimate (the global error or each error source individually) and dening a criterion (usually a threshold value) that will trigger
the renement.
Finding a bound for the algorithm: since projection error, discretization
error and modeling error are present, an accurate error sensor should detect all of them, and a consistent unbounded AMR algorithm will converge
toward a Direct Numerical Simulation, leading to very high computational
cost. The method must then be bounded in the sense that the nal solution should correspond to a large-eddy simulation with potentially large
projection error. This bound can be imposed in several ways: by xing
a maximum number of degrees of freedom, by giving a minimum mesh size
or playing on the threshold level in the evaluation of the error.
11.4.2 Error Estimation
The criterion used to rene the grid is based on the denition of an error
estimate. Therefore, it relies on an a priori choice dealing with the quantities
whose accuracy is the most important for the considered application. This is
an arbitrary, user-dependent decision.
Since it should account for both the numerical and modeling errors, the
error estimate is closely tied to the numerical method and the subgrid model
used in the simulation. The optimal choice is therefore case dependent, and
reveals a certain degree of empiricism.
The most striking achievements have been obtained by Homan [314,
310, 312, 313] using an a posteriori error estimate within a nite element
framework. Less developed methods [149, 518, 268] will not be described
here. The denition of an error estimate brings in many questions dealing
with mathematical analysis that will not be mentioned here. The interested
reader can refer to the original publications and the references given therein
for more details. The presentation given below will be restricted to the main
features of the method.
The purpose is to improve the accuracy of the simulation by error control
of the quantity

(u ud ) dtdx ,
(11.18)
err(u ud ) =
Q
where is a vectorial test function and the integration domain Q is dened

as Q = I, where is the volume of the computational domain and I
a time interval. The two velocity elds u and ud are related to the ideal
solution in which only the projection error is present and the computed solution including modeling and discretization errors, respectively. The aim of
an AMR procedure based on (11.18) is therefore to eliminate modeling and
discretization errors. This denition of the error is the most general one, and
allows the denition of optimal control methods.
379
Homan introduces the following linearized dual problem to dene an

optimal error control method:
u
u u + ud u + p 2 u =
t
u = 0 ,
(11.19)
(11.20)
where u and p are the dual variables. The dual system is supplemented with
adequate boundary and initial conditions which account for the denition
of the physical quantity of interest (drag/lift of a body, vortex shedding
frequency, ...). The exact projected eld u being unknown, it is replaced in
practice by ud . Using these new variables, the error can be expanded as

err(u ud ) =
( (ud ) ud ) : u dxdt
Q

discretization error

ud
+
+ ud ud u dxdt
t
Q


+ ( ud )p + pd ( u )dxdt
Q


+ (( (u)) (ud )) : u dxdt , (11.21)
Q

modeling error
where (ud ) denotes the modeled subgrid tensor and ( (u)) the projection
of the exact subgrid tensor onto the considered basis of degrees of freedom.
This expression enlights the denition of both the modeling and the numerical errors. The discretization error is a closed quantity which can be used in
a straightforward way, to the contrary of the modeling error, which requires
the knowledge of the projection of the exact subgrid tensor and therefore
appears as an unclosed quantity. To close this equation, Homan proposes
to use a structural subgrid model which exhibits a high degree of correlation
with the exact subgrid tensor, like the Bardina model or approximate deconvolution models. More accurate formula including interpolation errors on the
dual variables can be derived, which are not presented here for the sake of
simplicity.
It is worth noting that the error estimate is based on a volume integral,
even in the case where the physical quantities can be expressed as surface
integrals (e.g. drag, lift). This volumic formulation makes it possible to rene
the grid in regions where the error originates.
380
Fig. 11.5. LES-AMR simulation of the ow around a surface-mounted cube. View

of the instantaneous ow Top: (xy) plane: Bottom: (xz) plane. Courtesy of J. Homan, Courant Institute.
Fig. 11.6. LES-AMR simulation of the ow around a surface-mounted cube. View

of the grid after 9 renement steps. Courtesy of J. Homan, Courant Institute .
381
Fig. 11.7. LES-AMR simulation of the ow around a surface-mounted cube. Computed cube drag versus the number of grid points (in arbitrary units). Courtesy of
J. Homan, Courant Institute .
Fig. 11.8. LES-AMR simulation of the ow around a surface-mounted cube. Discretization (Circle) and modelling (Cross) errors (as dened in 11.21) versus the
number of grid points (in arbitrary units). Courtesy of J. Homan, Courant Institute.
382
The resulting algorithm is

1. Compute the solution (ud , pd ) solving the large-eddy simulation governing equation over a time interval I.
2. Solve the dual problem (11.19) (11.20) to obtain (u , p ) over I and
compute the discretization and modeling error according to (11.21).
3. Rene the grid in zones where one or both error sources are higher than
a xed threshold, taking into account possible bounds.
4. Loop until the targeted level of accuracy is reached, i.e. until err(u ud )
is lower than a required value.
Good results were obtained dealing with the prediction of aerodynamic
forces on blu bodies, using a nite-element method on unstructured meshes.
Typical results obtained dealing with the prediction of the drag of a surfacemounted cube are displayed in Figs. 11.5 11.8.
12. Hybrid RANS/LES Approaches
12.1 Motivations and Presentation

This chapter is devoted to the presentation of hybrid RANS/LES methods.
The main motivation for hybridizing the two methods is to decrease the cost
of the traditional large-eddy simulation method, which is large because of:
the requirement to directly capture all the scales of motion responsible for
turbulence production;
the observed inability of most subgrid models to correctly account for
anisotropy and disequilibrium.
These two weaknesses lead to the use of very-ne-resolution meshes, which
can be a real problem if the characteristic lengthscale of turbulence-production events is a decreasing function of the Reynolds number. A famous
problematic example is the inner region of boundary layers, whose intrinsic
scale is the wall unit (see Sect. 10.2).
In order to alleviate this problem, a possible solution is to blend large-eddy
simulation with another technique which must be able to provide relevant
lower-frequency solutions at a much lower cost. A natural proposition is to
use the RANS approach, which relies on a statistical average of the exact
solution and leads to a very large reduction of the number of degrees of
freedom in comparison with large-eddy simulation.
The presently existing techniques can be classied as follows:
Zonal decomposition (Sect. 12.2): the global computational domain is
divided into subdomains, some of them being treated with the RANS
method, the other ones being computed using large-eddy simulation. The
gain comes from the fact that the grid resolution can be coarser in the
RANS subdomains. Further cost reduction can be obtained when one or
two spatial directions can be suppressed in the RANS subdomains thanks
to statistical homogeneity of the ow in these directions.
Nonlinear Disturbance Equations (Sect. 12.3): the idea here is to compute
the mean oweld or the low-frequency part of the solution using a RANS
or unsteady RANS simulation, and to reconstruct the missing part of the
uctuating eld using a large-eddy-type simulation. This approach can be
interpreted as a zonal multidomain approach in the frequency domain.
384
Universal modeling (Sect. 12.4): the subgrid model is replaced by a new

model, which appears as a generalized turbulence model dened as a combination of a RANS model and a typical subgrid model. The hope here is
to obtain robust subgrid models, which are able to deal with very coarse
grids similar to those used in unsteady RANS computation. The merging
with a RANS model is expected to introduce more physics into the subgrid
model, rendering it ecient if the cuto is located in the low-frequency part
of the spectrum, outside or at the very beginning of the inertial range.
12.2 Zonal Decomposition

The zonal decomposition approach can be recast as a generalized multidomain/multiresolution problem. The main dierence with the multidomain
methods presented in Chap. 11 is that the coupling is now performed between subdomains where the scale separation is performed using operators of
a dierent nature: ltering in large-eddy simulation and statistical averaging
in RANS.
Equations presented in Sect. 11.1 can be reused in the present framework,
keeping in mind that now G2 is related to statistical averaging:
1
2
G2 () =
i = ,
(12.1)
N
i=1,N
where N is the number of samples chosen to compute the statistical average,

while G1 is still dened as a convolution lter.
All the theoretical analyses presented in Sect. 11.1 can be directly extended to the RANS/LES coupling method. The main practical dierence is
that now the problem of the reconstruction of the complementary eld v 12 is
strictly equivalent to that of the denition of turbulent inow conditions for
large-eddy simulation described in Sect. 10.3. The mean ow is now predicted
using the RANS computation.
Another pratical dierence is the denition of the restriction operator G21 ,
which is dened so that:
G21 (G1
u) = G2 (u)
(12.2)
or, equivalently, and using the usual notation:

G21 (u1 ) = u
(12.3)
G21
From a purely theoretical point of view, we see that

must be dened
as the sequential application of a deconvolution operator and statistical averaging. It can be simplied as a simple statistical average when the mean
velocity proles of the exact and ltered solutions are equal.1
1
This condition is satised if the lter is applied in homogeneous directions only.
385
Two main approaches are identied:

Sharp transition (Sect. 12.2.2): the reconstruction of the complementary
eld is carried out explicitly at the interface between RANS and LES subdomains, yielding a sharp transition over one grid mesh from one solution
to the other.
Smooth transition (Sect. 12.2.3): the reconstruction of v 12 is not performed
at the interface, and the high-frequency part of the uctuating eld must
be regenerated by instabilities of the mean ow and the forward energy
cascade. This approach leads to the existence of a transition region near
the interface, whose width is case-dependent.
12.2.2 Sharp Transition
A method based on the sharp transition approach was proposed by Quemere
et al. [610], and is an extension of the multidomain method proposed by the
same authors (see Sect. 11.3). It relies on a strict analogy between the reconstruction of the complementary eld v 12 and the denition of the turbulent
inow condition.
The G21 restriction operator is simply computed as a statistical average,
without incorporating a deltering operator. Several techniques for reconstructing the uctuations at the interface have been implemented (listed in
decreasing order of eciency): use of a predictor simulation, Lunds extraction/rescaling technique, and random perturbation. Numerical experiments
show that the smaller the transition region near the interface the more realistic the uctuating eld option becomes. When a predictor simulation is
implemented, it is reduced to one grid cell.
The method was applied with varying success to plane channel ow and
ow around a blunt trailing edge. The RANS model was the JonesLaunder
k model. The boundary conditions for the turbulent quantities at the interface were derived from a simple extrapolation procedure. Typical results
for a plane channel ow conguration are presented in Figs. 12.1 and 12.2.
A simplied version of Quemeres approach was implemented by Georgiadis et al. [241], who operated the switch from RANS to LES while neglecting the reconstruction of the complementary eld. The purpose was the
simulation of a mixing layer ow, in which the primary instabilities have
a very large growth rate, resulting in a small inuence of the exact denition
of the complementary eld.
Applications of this approach to the denition of sophisticated two-layer
wall models for large-eddy simulation have also been proposed by Davidson
et al. [167] and Diurno et al. [180]. In both cases, the complementary eld
is not reconstructed. The resolved eld is assumed to be continuous at the
interface in Davidsons method, which is based on a k RANS model, while
the coupling is achieved by transmiting wall stresses at the interface in Diurnos simulations relying on the SpalartAllmaras model. Additional boundary conditions must be provided for turbulent quantities at the RANS/LES
386
Fig. 12.1. Large-eddy simulation of plane channel ow with zonal RANS/LES

coupling. RANS subdomains are used in the near-wall region, while the core of the
channel is simulated using LES. The computed mean velocity prole is compared
with reference data, for two position of the interface . Solid and dashed lines:
hybrid RANS/LES computations. Squares: classical RANS simulation. Crosses:
classical LES simulation. Courtesy of P. Quemere, ONERA.
Fig. 12.2. Large-eddy simulation of plane channel ow with zonal RANS/LES

coupling. RANS subdomains are used in the near-wall region, while the core of the
channel is simulated using LES. The computed rms velocity proles are compared
with reference data obtained via typical LES simulation, for two positions of the
interface . Solid and dashed lines: hybrid RANS/LES computations. Symbols:
classical LES simulation. Courtesy of P. Quemere, ONERA.
387
interface. Davidson assumed no coupling at the interface for variables related

to models and imposed arbitrary conditions, while explicit coupling is enforced in Diurnos method. The latter assumes that the total stress is equal
on both sides of the interface.
12.2.3 Smooth Transition
Hybrid RANS/LES strategies presenting a smooth transition have all been
derived to alleviate the classical grid resolution requirements of classical
large-eddy simulation in the near-wall region. Consequently, they can be
interpreted as hybrid wall-models for large-eddy simulation. Combinations
between the two basic techniques are derived in two dierent ways:
By performing an explicit blending of the two basic models. Bagget [31]
classied the existing approaches in two groups. The rst one corresponds
to the linear combination of the predicted turbulent/subgrid stresses:
4
3
1
ij kk ij = sgs S ij (1 (z))S ij (z)rans S ij , (12.4)
3
where sgs and rans are the subgrid and turbulent viscosities, respectively.
The arbitrary weight (z) is expressed as a function of the distance to the
wall, and should be prescribed. This form of hybrid RANS/LES model is an
extension of the splitting technique proposed by Schumann (see Sect. 6.3.3,
p. 200) to account for anisotropy in the near-wall region. Methods belonging to the second group are based on the blending of the modeled viscosities:
1
ij kk ij = [(1 (z))sgs + (z)rans ] S ij
3
(12.5)
In both cases = 0 corresponds to classical large-eddy simulations and

= 1 to classical RANS simulations.
Replacing the characteristic scales which explicitly appear in the RANS
model equations by new scales corresponding to the LES lter. The important point is that the turbulent viscosity or the turbulent stresses are not
directly rescaled, and consequently these quantities exhibit a continuous
behaviour, yielding smooth transition. The most famous example belonging to this category is the Detached Eddy Simulation proposed by Spalart
et al. [680, 569]. This model is based on a modication of the Spalart
Allmaras RANS model, in which the distance to the wall d is replaced
by d:
(12.6)
d = min(d, Cdes ) ,
in the destruction term of the transport equation. The constant Cdes is
calibrated in order to recover a Smagorinsky-like behaviour in isotropic
turbulence at equilibrium, yielding Cdes = 0.65. The Detached Eddy Simulation approach has been extented by Strelets [16] to a large family of
388
RANS two-equation
models. For a k model, the characteristic length
lk k/ is replaced in the dissipative term of the k-transport equation by

l = min(lk , Cdes ), Cdes = 0.78 .
(12.7)
A similar approach was developed by Tucker and Davidson [720] which relies on two-equation kl models. In the RANS domain, a usual kl model
is used, while the Yoshizawa kl subgrid model is used in the LES subdomain. The equations of these two models are formally equivalent, but with
dierent values of the constants appearing in them. The switch is achieved
by changing the length scale in the transport equations: the usual RANS
length scale is used in the RANS subdomain, while the subgrid lengthscale
is considered in the LES subdomain. This switch results in a discontinuous value of the length scale at the interface. Tucker and Davidson
obtain a regular distribution at the interface by applying a 3-point test
lter across the interface to the lengthscale.
Another smooth transition method was proposed by Hamba [284]. In this
method, the RANS and LES subdomains are overlapping, and the length
scale in the subgrid model is assumed to vary linearly from the RANS
length scale to the usual subgrid length scale tied to the computational
mesh. Hamba applied this technique to a channel ow, hybridizing a k
model and a Smagorinsky model. This method yields some error in the
mean velocity prole, and was further improved in [285]. In the improved
method, the dissipation is also linearly interpolated between the RANS
value and the one predicted using the subgrid model. But the most important improvement factor is that the convective uxes on each side at the
interface are estimated using an ad hoc interpolated velocity eld which
is more compatible with the discrete mass preservation constraint. This
blending of the two solutions at the interface can also be interpreted as
a surrogate for the complementary eld v 12 . The improved method is observed to yield much satisfactory results on a plane channel conguration.
12.2.4 Zonal RANS/LES Approach as Wall Model
Both sharp- and smooth-transition hybrid methods have been used to derive
new robust wall models for large-eddy simulation. Results are generally disappointing for methods which neglect the reconstruction of the complementary
eld at the interface. An articial turbulent boundary layer develops, which
is composed of overly large streamwise streaks and vortices. This turbulent
process is observed to be self-sustaining. This can be explained by theoretical
and numerical results dealing with the existence of a near-wall autonomous
cycle [353, 354, 355, 750], which is created by the mean shear.
For sharp-transition methods, the problem is identical to that of wall
models for large-eddy simulation, and it seems that zonal modeling cannot
389
perform better than the suboptimal-control-theory-based wall stress models

(see Sect. 10.2).
For smooth-transition methods, the interface is replaced by a transition region. An explanation [31] for the existence of a spurious near-wall cycle is that
the no-slip boundary condition produces a near-wall viscous region in which
the mean pressure gradient is balanced only by the mean viscous stress. The
ow is fully turbulent outside the transition region, so that the mean shear
scales as 1/z, z being the wall-normal coordinate. In the transition region, the
mean shear has to be reduced by wall-normal streamwise momentum transport, i.e. mean Reynolds stress, to couple the core ow to the wall. If the subgrid model does not carry a large enough amount of Reynolds stress, the resolved motion will adjust in order to enforce the necessary balance. Some authors tried to cure this problem by articially enhancing the dissipation in the
transition region, but this method did not prove to be ecient and general.
For high-Reynolds-number boundary layers, the use of too coarse a grid
also yields signicant errors in the mean velocity proles [569]. This demonstrates that, even if the RANS region is able to provide the large-eddy simulation with correct boundary conditions, the grid resolution in the LES region
must be the same as for classical wall-resolving large-eddy simulations.
Another point is that Quemeres results seem to indicate that accounting
for the complementary eld v 12 at the interface improves the results and can
prevent the appearance of the spurious cycle.2
Applications to massively separated ows have also been perfomed [180,
167, 146, 680, 695] which demonstrate that these hybrid approaches can yield
much more satisfactory results than for attached ows. The main reason is
that massively separated ows are mostly governed by inviscid, large-scale
instabilities of the mean ow, which occur in the LES region. The quality of
the results is then equivalent to that of a typical large-eddy simulation, but
with an eective cost reduction. Strelets [695] remarked that the numerical
scheme must be adjusted to the selected approach in each region to obtain
reliable results: a low-order accurate dissipative method can be used in the
RANS region, but not in the LES region. As a consequence, the zonal approach for physical modeling can also imply the denition of a zonal approach
for the numerical method.
The concept of hybridizing RANS and large-eddy simulation in the nearwall region is fully general and does not depends, from a purely formal standpoint, on the modeling approaches used in both parts. The RANS simulation
is almost always carried out using models relying on zero to three additional
equations, while subgrid models involving zero to two equations are very often
met. A noticeable exception is provided by Tucker [719], who couples a oneequation RANS model with a MILES approach in the large-eddy simulation
domain.
2
This is consistent with the nding that the use of a non-zero transpiration velocity for wall stress models yields improved results.
390
12.3 Nonlinear Disturbance Equations

Another hybrid RANS/LES method was proposed by Morris et al. [545,
140, 288, 547], which is referred to as the Nonlinear Disturbance Equations
(NLDE) method. The underlying idea is to split the eld into a low-frequency
or steady part on the one hand, and a high-frequency uctuating part. The
former can be computed using steady or unsteady RANS simulations, or theoretical laws, while the latter is computed via large-eddy simulation. The
resulting set of governing equations appears as a generalized form of the
NavierStokes equations in perturbation form:
u2
+ N S(u2 ) = 2
t
(12.8)
v 12
+ N S(u1 ) N S(u2 ) = 1 2 .
(12.9)
t
An interesting dierence with other methods presented above is that the
equation for the detail v 12 is solved, rather than the equation for the ltered
eld u1 . If the carrying eld u2 is steady, the time derivative in (12.8) cancels
out, leading to
u1
v 12
=
.
(12.10)
t
t
If it is laminar, then 2 = 0.
Using the bilinear form (3.27), the nonlinear convective term appearing
in (12.9) can be recast as
B(u1 , u1 ) B(u2 , u2 ) = B(u2 , v 12 ) + B(v12 , u2 ) + B(v12 , v 12 )

. (12.11)
II
Terms I and II are related to the coupling between the two levels of resolution and to the nonlinear interactions between the uctuations, respectively.
Another coupling is achieved through the source term in the right-hand side
of (12.9), which is dened as the dierence between the subgrid force and the
Reynolds forcing term.
This approach can be interpreted as a generalized multilevel simulation (see Sects. 7.7.7 and 11.2), in which the coarsest level is dened using
a statistical-average operator and not a convolution lter.
This technique was rst developed by Morris et al. to evaluate acoustic
sources from a steady jet computation, on the basis of a nonlinear inviscid model equation for the uctuations. Their second main approximation is
that mean-ow source terms in the nonlinear disturbance equations can be
neglected. More recently, these authors added a subgrid model to the disturbance equations, in order to take into account the eect of unresolved modes,
leading to a true hybrid RANS/LES approach. They also reintroduced a part
of the mean ow source term into the nonlinear disturbances equation. The
12.4 Universal Modeling
391
NLDE approach was used by Hansen et al. [288] to simulate the unsteady,
two-dimensional laminar ow around a circular cylinder, corresponding to the
rst coupling between NLDE and an unsteady mean ow. The rst attempt
to use it for wall-bounded ow is due to Chyczewski et al. [140], with fairly
good results due to the use of very coarse grids.
A fully general expression of the governing equations for nonlinear disturbance equations for compressible ows was derived by Labourasse and
Sagaut [416] who also proposed using this method to locally reconstruct the
unsteady turbulent motion in subdomains embedded within the full ow conguration. This corresponds to a coupling between the nonlinear disturbance
equations approach and the multidomain method presented above. The expected gain with respect to the classical large-eddy simulation approach is
twofold:
Firstly, the mean eld being prescribed, the errors commited on turbulent
uctuations will not pollute it, and we can expect this method to be more
robust than classical large-eddy simulation.
Secondly, if the turbulence-producing events are localized in a small region,
it will be possible to restrict the LES-type computation to a small subdomain included in the global domain, while classical large-eddy simulation
would require us to consider the full domain.
Numerical experiments show that this method is more robust than the
usual large-eddy simulation if the mean ow is correctly prescribed, i.e.
coarser grids can be employed without loss of accuracy. Errors on the mean
ow eld u2 can be corrected using the hybrid approach on a ne grid: the
uctuating eld will adjust such that (u2 + G21
v 12 ) will be correct.

The last class of hybrid RANS/LES methods presented in this chapter is that
of the universal models. The underlying idea here is to design new models
which can asymptotically recover typical RANS or typical LES capabilities.
The existing models are based on the rescaling of the RANS models thanks
to inertial range arguments. They all aim at decreasing the resolved kinetic
energy dissipation induced by the model for the unresolved scales. This decrease in the induced dissipation leads to a weaker damping of high frequencies/high wavenumbers which can be sustained on the selected computational
grid, yielding the recovery more irregular LES-like owelds. Despite this
strategy is clear and does not suer any aw, it is worth noting that the
dierent methods used to achieve this reduction in the small scale damping3
are mostly empirical. The most common strategies found in the literature
3
The most common method is to reduce the amplitude of the turbulent viscosity
of the RANS model.
392
deal with the modication of eddy-viscosity-type RANS models, and can be

grouped in two classes:
1. The methods based on a rescaling of the eddy viscosity provided by the
RANS model.
2. The methods based on the modication of some terms in the transport
equations of the original RANS model, so as to reduce the amplitude of
the eddy viscosity. In the case of most RANS models which include an
evolution equation for the turbulent kinetic energy, k, this is achieved by
increasing the destruction of k and/or modifying the second variable (if
any).
A few hybrid models are presented below:
1.
2.
3.
4.
Germanos mixed modeling (Sect. 12.4.1).

Speziales general rescaling method for turbulent stresses (Sect. 12.4.2).
Arunajatesans modied two-equation model (Sect 12.4.4).
BushMani limiters (Sect. 12.4.5).
12.4.1 Germanos Hybrid Model

A general approach for the denition of hybrid RANS/LES models was proposed by Germano [249]. This approach relies on two basic hypotheses:
The statistical value of a ltered quantity is equal to the statistical value
of the unltered quantity:
(12.12)
= .
The ltered solution is characterized by the associated mean production of
subgrid kinetic energy:
les,ij S ij = Cf rans,ij Sij
(12.13)
where les (u) and rans (u) are the subgrid and Reynolds stress tensor,
respectively, and Cf is a constant.
The rst hypothesis makes it possible to decompose the modeled Reynolds
stress tensor as follows:
rans (u) = les (u) + rans (u) ,
(12.14)
while the second hypothesis leads to

les,ij Sij =
Cf
rans,ij (u)Sij .
1 Cf
(12.15)
This equation establishes a bridge between the modeled tensors at both

ltered and statistically averaged levels. For subgrid viscosity models, we
393
obtain the following denition:

les =
Cf |S| rans,ij (u)Sij

2(1 Cf ) |S|S ij Sij
(12.16)
The parameter Cf [0, 1] determines the respective weights of the ltered and the Reynolds-averaged parts. The RANS solution is recovered as
Cf 1. The convergence of the method depends on the capability of the
RANS model for rans,ij , once applied to the ltered eld u, to alleviate the
singular behavior of (12.16).
12.4.2 Speziales Rescaling Method and Related Approaches
Models belonging to this category can be expressed as
ij = FR ijrans
(12.17)
where FR is the rescaling function to be dened, and ijrans the modeled

turbulent stress tensor predicted using any RANS model.
Speziale [685, 684] proposed a general rescaling function
3
4n
,
(12.18)
FR = 1 exp(/K )
where is a constant, n an arbitrary power, and K the Kolmogorov length
scale computed from the variables of the RANS model. The direct numerical simulation regime is recovered as the limit of a very ne resolution,
/K 0. As the resolution becomes very coarse, /K , the
RANS model is recovered. As noted by Magnient [477], a consistency problem of the Speziales transition function is that large-eddy simulation is not
recovered as a limit for very high Reynolds number.4 In practice, Speziale
recommended applying this scaling law to complex RANS models able to
deal with anisotropy and disequilibrium, such as Reynolds Stress Models or
Explicit Algebraic Reynolds Stress Models. Proposed values in [685, 684] are
= 0.001 and n = 1, but the model is expected to be very sensitive to these
parameters.
Fasel [212] later used = 0.004, = (1 2 3 )1/3 and K = Re3/4 1/4
together with an Explicit Algebraic Resynolds Stress Model to compute a at
plate boundary layer ( being evaluated from the outputs of the RANS
model). The same authors used a slightly modied version to compute a wall
jet ow. Parameters for this conguration are
+
3
4
1
2 2 2
1 2 3 ,
FR = 1 exp( max(0, 5 10K )/N K ) , =
3
(12.19)
4
A well known example is isotropic turbulence in the limit of an innite Reynolds

number.
394
where N is an adjustable parameter that governs the cuto, which taken in

the range 10002000 in the wall-jet application.
A much simpler rescaling for computing subgrid viscosity was proposed
by Peltier et al. [796, 584]. Using inertial range argument and numerical tests,
these authors proposed computing the subgrid viscosity sgs from the RANS
eddy viscosity rans as
2
sgs =
rans ,
(12.20)
l
where l is the turbulent length computed using the RANS model outputs.
Magnient [477] proposed a similar rescaling law for the RANS viscosity with
the power 4/3 instead of 2.
The rescaling approach was further developed by Batten and his coworkers [46, 47, 48, 49] as the Limited-Numerical-Scales method, who introduced the following rescaling factor:
FR =
min [(Lsgs Usgs ), (Lrans Urans )]

(Lrans Urans )
(12.21)
where Lsgs , Usgs , Lrans , Urans are the characteristic subgrid length scale, subgrid velocity scale, turbulent length scale and turbulent velocity scale, respectively. In the simple case where eddy viscosity type models are used in
both cases, the products LU are equal to the eddy viscosities. The original implementation by Batten for several ows rely on the nonlinear k
model of Goldberg and the Smagorinsky model. It is important noticing that
the subgrid closure is used here only to evaluate the rescaling factor FR
through (12.21). This method, coined as the Limited Numerical Scale approach, was mainly used through a zonal implementation. To improve the
results, Batten strongly recommend to add explicit uctuations at the interface, i.e. to synthetize explicitly the eld v 12 .
12.4.3 Baurles Blending Strategy
A more complex method, which can be interpreted as an evolution of Battens proposal using the same strategy as the one underlying Menters hybrid
RANS model, is presented by Baurle and his co-workers [50]. The underlying
idea is still to hybridize a classical RANS model with a standard subgrid scale
model. A requirement is that both RANS and LES models include an equation on the unresolved turbulent kinetic energy (TKE). The hybrid model is
therefore dened by a linear combination of each model equation set:
Hybrid RANS/LES
TKE equation =
F [RANS TKE equation]
+(1 F) [LES TKE equation] , (12.22)
395
and
hybrid = F rans + (1 F)sgs
(12.23)
where F is a weighting function to be discussed below. As remarked by

Baurle, an algebraic subgrid viscosity model or even an implicit model can
also be considered in this approach by constructing a production/destruction
balance equation for the subgrid kinetic energy that recovers the desired subgrid viscosity. The key parameter in this approach is the weighting function
F . In the case F depends only on geometric factors such as the distance to
solid walls and on grid topology, the blending strategy results in a purely
zonal approach reminiscent to those described in Sect. 12.2. The blending
function proposed in Ref. [50] is more general and includes some information
tied to the level of resolution of the ow. It is dened as
F = max(tanh( 4 ), F1 ) ,

whith
= max
Lrans
Cd Lrans
, 500
d
krans d2
(12.24)

,
(12.25)
where Lrans , d and krans are the turbulent length scale computed from the
RANS model outputs, the distance to the nearest wall and the turbulent
kinetic energy provided by the RANS model, respectively. The constant Cd
is taken equal to 0.01. The function F1 is dened on the grounds of Battens
blending function (12.21):

1 if sgs < rans
F1 =
.
(12.26)
0 otherwise
The resulting weigthing function is such that the RANS treatment will be
invoked near solid surfaces, while the LES treatment will govern the simulation in separated and free shear regions. The introduction of the F1 parameter
allows to maintain the RANS treatment when sudden grid renements are
encountered and that no resolved uctuations are present.
A simplied version of the weighting function was developed in [772],
which is expressed as
2

sgs
.
(12.27)
F = tanh
rans
This approach is now illustrated considering the case of the two-equations
k model ( being the turbulent enstrophy) treated in [772]. The RANS
equation for the turbulent kinetic energy k is:

rans k
k uj k
ui
+
+
=
+ ij
t
xj
xj
3
k
xj
xj
k
1 rans p
C1
, (12.28)
2
Ck xi xi
p
396
where
rans =
k2
rans
= C
(12.29)
with C = 0.09, Ck , C1 and k model-dependent parameters. The subgrid

viscosity being evaluated as
sgs = Cs k, Cs = 0.01 ,
(12.30)
the hybrid viscosity is dened according to (12.23) while the original turbulent kinetic energy equation (12.28) is transformed into

hybrid k
k uj k
ui
+
+
=
+ ij
t
xj
xj
3
k
xj
xj

k
1 hybrid p
(1 F)
+ C1
+
Ck 2 xi xi
p
FCd
k 3/2
(12.31)
where Cd = 0.01 is a constant. The RANS turbulent length scale appearing

in the computation of the blending function is
Lrans =
k 3/2
(12.32)
12.4.4 Arunajatesans Modied Two-Equation Model

A modied two-equation k model was proposed by Arunajatesan et al. [20,
21, 22] as a basis for hybrid RANS/LES simulations. The key idea is similar to that of the Detached Eddy Simulation, i.e. some characteristic scales
appearing in the classical RANS model are replaced by new ones associated
with the ltering operator. Two transport equations are solved along with the
NavierStokes equations: one for the subgrid kinetic energy qsgs and one for
the total dissipation rate . Based on the values of these quantities and the
lter length , a RANS and a subgrid viscosity referred to as rans and sgs ,
respectively, are computed. The subgrid viscosity is used in the equations for
momentum and subgrid kinetic energy, while the RANS viscosity is used in
the equation for the turbulent dissipation rate.
The proposed transport equations for the subgrid kinetic energy and the
dissipation rate are

qsgs
qsgs
+
ui qsgs ( + sgs /q )
(12.33)
= Pq ,
t
xi
xi

+
ui ( + rans / )
(12.34)
= P D ,
t
xi
xi
397
where k and are modeling constants equal to their usual RANS counterparts. Production and dissipation terms appearing on the right-hand sides
of (12.33) and (12.34) are identical to those of the original RANS model and
will not be detailed here (see [424]).
Once these two quantities are known, it is assumed that the local turbulent
energy spectrum can be represented everywhere using the following hybrid
Von Karman/Pao form:

= C1 k5/3
k)
E(
k
k
4
1 +
k
k
2 17/6

exp
9 4/3
k
4

,
(12.35)
where k = kK is the normalized wave number, K the Kolmogorov length

= E/( 5 )1/4 the normalized
scale, k the energy-containing wave number, E
spectrum, and C1 a constant to be determined.
The Kolmogorov scale is computed as K = ( 3 /)1/4 . The constant C1
is chosen so that the integral of the dissipation range spectrum is equal to
the local turbulence dissipation rate.
The energy-containing wave number k is computed in order to enforce
the following relation:

qsgs =
E(k)dk
(12.36)
Once the local turbulence spectrum is completely determined, the authors

propose computing the value of the subgrid viscosity for any cuto wave
number using the constant spectral subgrid viscosity model (5.19), yielding
'
E(/)
sgs = 0.28
,
(12.37)
'
E(k )
rans = 0.28
.
(12.38)
k
12.4.5 BushMani Limiters
Another possibility proposed by Bush and Mani [792] for deriving a hybrid
model is to generalize the basic idea underlying Spalarts Detached Eddy
Simulation by applying it to all the turbulent variables. Considering the twoequation RANS model of the general form k , the proposed solution is
summarized in Table 12.1. In order to account for both space- and timelterings, the authors proposed redening the lter length in the dissipation
term as
* = max(, |u|t, kt) ,
(12.39)
398
Table 12.1. BushMani limiters for hybrid RANS/LES models. Ci are constants,
* is the cuto length.
and
limiter
*
min(l, C1 )
*
max(, C2 k3/2 /)
1/2 *
max(, C3 k /)
where is the usual cuto length used in large-eddy simulation, and t the
time step of the computation.
The same approach was followed by Allen and Mendonca [10], who tested
extended versions of Spalarts Detached Eddy Simulation based on several
two-equations models. The values of the constants appearing in Table 12.1
given by these authors are C2 = 0.73 and C3 = 0.61. But numerical experiments reported by several authors indicate that the optimal value is casedependent.
12.4.6 Magagnatos Two-Equation Model
A modied two-equation model approach is proposed by Magnagato and
Gabi [476], which includes an explicit random backscatter term. The original
formulation of the model is based on either a non-linear k or a non-linear
k RANS model, in which the characteristic length scale L is evaluated as

L = max , |u|t
.
(12.40)
Here, quantities computed solving the equations of the RANS model with
the modied length scale are assumed to be related to the unresolved turbulent uctuations. The dissipative part of the model is then written as
2
ij = C k S ij + kij
3
(12.41)
(12.42)
for the k model, and

ij = C
k2
2
S ij + kij
for the k model. In both cases, the value C = 0.09 is retained. The
authors thus add a non-dissipative term to the expressions given above for
the subgrid tensor to account for the backscatter phenomenon. The generic
form of the backscatter model is
2
ij = vi vj kij
3
(12.43)
12.5 Toward a Theoretical Status for Hybrid RANS/LES Approaches
399
where the random velocities vi are calculated at each time step using
a Langevin-type equation. Introducing i an independent random vector
ranging from 1 to 1, the value of vi at the nth time step is dened as

'

t
t
t 2 n
n1
n
k
,
(12.44)
vi = vi
1
+
2
3 i
where the subgrid time scale is directly provided by the k model and
is evaluated as
L
,
(12.45)
=
k
for the k model, where the length scale L is given by relation (12.40) and
k is the sum of the resolved and the unresolved kinetic energy. The full
model formulation is obtained by adding the backscatter term to the selected
dissipative part.
12.5 Toward a Theoretical Status for Hybrid

RANS/LES Approaches
Since they are based on the hybridization of the classical Large-Eddy Simulation and Reynolds-Averaged Numerical Simulation methods, most of the
hybrid approaches presented above escape the usual theoretical framework
developed to present them. As a matter of fact, the resulting ow eld can
neither be interpreted in terms of statistical average of the exact solution nor
as the result of a conventional ltering operation. Therefore, the question
arises of nding a relevant paradigm to analyze and understand these hybrid
approaches.
A powerful framework to understand the properties of these methods
is the unresolved-scale-model induced eective lter paradigm developed by
Muchinsky (see Sect. 8.1.1). This analysis is based on the observation that,
during the computation, the only term which carries some information about
the unresolved scales is the turbulence/subgrid model5 . Therefore, it is the
model for the unresolved scales which determines the amount of damping (the
numerical errors are assumed to be negligible here for the sake of clarity) of
the resolved scales and governs the eective ltering of the exact solution.
This is illustrated writing the resolved discrete problem as
ud
+ Fd (ud , ud ) = Sd
t
(12.46)
where ud , /t and Fd (, ) are the discrete approximations of the exact terms

u, /t and F (, ) on the computational grid, respectively. Here, F (, ) is
5
In the case of the Implicit Large-Eddy Simulation approach, this information is

contained in the numerical errors.
400
related to the uxes in the exact NavierStokes equations. The source term
Sd stands for the model for the unresolved scales.
The discrete equation for the resolved kinetic energy is
1 u2d
+ ud Fd (ud , ud ) = ud Sd
2 t
(12.47)
The term in the right hand side of this new equation shows how the
model for unresolved scales governs the dissipation process. A decrease in
this dissipation will results in a larger amount of resolved kinetic energy, and
therefore to the existence of smaller scales (the limit being xed by the grid
resolution). All hybrid RANS/LES methods based on the modication or the
rescaling of a RANS model aim at decreasing this term.
Following Muchinskys analysis, which relies on the idea that the RANS
or LES computations can be seen as direct numerical simulations of a nonnewtonian uid, the observed eects of the modication of the original RANS
models can be easily interpreted. The steady RANS solution corresponds to
a steady laminar non-newtonian ow. Reduction in the dissipation results in
a higher Reynolds number. Unsteady RANS solutions are therefore similar
to direct numerical simulations of unsteady non-newtonian ow. The bifurcations sometimes observed in unsteady RANS simulations when the total
dissipation is further reduced can be interpreted as analogous to bifurcations of a laminar newtonian ow at low-Reynolds number (transtition to
three-dimensional modes, growth of small scales, ...). The further dissipation reduction achieved by the denition of hybrid models leads to higher
Reynolds numbers, and to the possiblity for smaller scales to be sustained.
13. Implementation
This chapter is devoted to the practical details of implementing the largeeddy simulation technique. The following are described:
Cuto length computation procedures for an arbitrary grid;
Discrete test lters used for computing the subgrid models or in a preltering technique;
Computing the Structure Function model on an arbitrary grid.
This part of the implementation of large-eddy simulation is more and
more recognized as one of the keys of the success of a simulation. Most of
the theoretical developments rely on an abstract lter, which is characterized
by its cuto frequency. But practical experience show that computational
results can be very sensitive to the eective properties (transfer function,
cuto length) of the tering operators used during the simulation.
This is especially true of subgrid models relying on the use of a test
lter, such as dynamic models and scale-similarity models. Thus, the problem
of the consistency of the lter with the subgrid model must be taken into
account [597, 563, 81, 641, 605].
13.1 Filter Identication. Computing the Cuto Length

The theoretical developments of the previous chapters have identied several
lters of dierent origins:
1. Analytical lter, represented by a convolution product. This is the lter
used for expressing the ltered NavierStokes equations.
2. Filter associated with a given computational grid. No frequency higher
than the Nyquist frequency associated with this grid can be represented
in the simulation.
3. Filter induced by the numerical scheme. The error committed by approximating the partial derivative operators by discrete operators modies the
computed solution mainly the high-frequency modes.
4. Filter associated with the subgrid model, which acts like a control process
on the computed solution.
402
13. Implementation
The computed solution is the result of these four ltering processes constituting the simulation eective lter. When performing a computation, then,
the question arises as to what the eective lter is, that governs the dynamics of the numerical solution, in order to determine the characteristic cuto
length. This length is needed for several reasons.
In order to be able to determine the physically and numerically wellresolved scale beyond which we will be able to start using the results for
analysis.
In order to be able to use the subgrid models like the subgrid viscosity
models that use this cuto length explicitly.
While the lters mentioned above are denable theoretically, they are
almost never quantiable in practice. This is particularly true of the lter
associated with the numerical schemes used. In face of this uncertainty, practitioners have one of two positions they can adopt:
1. Arrange it so that one of the four lters becomes predominant over the
others and is controllable. The eective lter is then known. This is done
in practice by using a pre-ltering technique.
Normally, this is done by ensuring the dominance of the analytical lter,
which allows us strict control of the form of the lter and of its cuto length, so that we can get the most out of the theoretical analyses
and thereby minimize the relative uncertainty concerning the nature of
the computed solution. In the numerical solution, an analytical lter is
then applied here to each computed term. In order for this lter to be
dominant, its cuto length must be large compared with the other three.
Theoretically, this analytical lter should be a convolution lter which, to
keep the computation cost within acceptable limits, can only be applied
for simulations performed in the spectral space1 . For the simulations performed in the physical space, discrete lters are used, based on weighted
averages with compact support. These operators enter into the category
of explicit discrete lters, which are discussed in the following section.
We may point out here that the methods based on implicit diusion
with no physical subgrid model can be re-interpreted as a pre-ltering
method, in which case it is the numerical lter that is dominant. We can
see the major problem of this approach looming here: the lter associated
with a numerical method is often unknown and is highly dependent on
the simulation parameters (grid, boundary conditions, regularity of the
solution, and so forth). This approach is therefore an empirical one that
oers little in the way of an a priori guarantee of the quality of the results.
It does, however, have the advantage of minimizing the computation
costs because we are then limited to solving the NavierStokes equations
without implanting any subgrid model or explicit discrete lter.
1
The convolution product is then reduced to a simple product of two arrays.
13.1 Filter Identication. Computing the Cuto Length
403
2. Considering that the eective lter is associated with the computational

grid. This position, which can be qualied as minimalist on the theoretical
level, is based on the intuitive idea that the frequency cuto associated
with a xed computational grid is unavoidable and that this lter is
therefore always present. The problem then consists in determining the
cuto length associated with the grid at each point, in order to be able
to use the subgrid models.
In the case of a Cartesian grid, we take the ltering cell itself as Cartesian.
The cuto length is evaluated locally as follows:
For uniform grid, the characteristic ltering length in each direction is
taken equal to the mesh size in this same direction:
i = xi
(13.1)
The cuto length is then evaluated by means of one of the formulas

presented in Chap. 6.
For a variable mesh size grid, the cuto length in the ith direction of
the grid point of index l is computed as:
i |l = (xi |l+1 xi |l1 )/2 .
(13.2)
The cuto length is then computed locally according to the results of

Chap. 6.
In the case of a curvilinear structured grid, two options are possible
depending on the way the partial derivative operators are constructed:
If the method is of the nite volume type in the sense of Vinokur [733],
i.e. if the control volumes are dened directly on the grid in the physical space and their topologies are described by the volume of the
control cells, by the area and the unit normal vector to each of their
facets, the lter cuto length can be computed at each point either by
taking it equal to the cube root of the control volume to which the
point considered belongs, or by using what Bardina et al. propose (see
Sect. 6.2.3).
If the method is of the nite dierences type in the Vinokur sense [733],
i.e. if the partial derivative operators are computed on a uniform Cartesian grid after a change of variables whose Jacobian is denoted J, then
the cuto length can be evaluated at the point of index (l, m, n) either
by Bardinas method or by the relation:
l,m,n = (Jl,m,n )
1/3
(13.3)
where , and are the grid steps in the reference space.

In the case of an unstructured grid, we use the same evaluations as for
a structured curvilinear grid with a nite volume type method, in the
sense given above.
404
13. Implementation
13.2 Explicit Discrete Filters

Several techniques and subgrid models described in the previous chapters use
a test lter. For reference, these are the:
Pre-ltering technique;
Soft deconvolution models and scale similarity models;
Mixed Scale Model;
Dynamic constant adjustment procedures;
Models incorporating a structural sensor;
Accentuation procedure.
The corresponding theoretical developments all assume that we are able

to apply an analytical lter in the simulation. This operation comes down
to a product of two arrays in the spectral space, which is a simple operation
of little cost, and all the analytical lters whose transfer function is known
explicitly can be used. The problem is very dierent, though, when we consider the simulations performed in the physical space on bounded domains:
applying a convolution lter becomes very costly and non-local lters cannot
be employed. In order to be able to use the models and techniques mentioned
above, we have to use discrete lters with compact support in the physical
space. These are described in the rest of this section. These discrete lters
are dened as linear combinations of the values at points neighboring the one
where the ltered quantity is computed [632, 728, 563, 461].
The weighting coecients of these linear combinations can be computed
in several ways, which are described in the following. We rst present the
one-dimensional case and then that of the Cartesian grids of more than one
dimension, and lastly extend this to arbitrary grids.
The discrete approximation of the convolution lters is then discussed.
13.2.1 Uniform One-Dimensional Grid Case
We restrict ourselves here to the case of a uniform one-dimensional grid of
mesh size x. The abscissa of the grid point of index i is denoted xi , such
that we can say xi+1 xi = x. The ltered value of the variable at the
grid point of index i is dened by the relation:
i
N

al i+l
(13.4)
l=N
where N is the radius of the discrete lter stencil. The lter is said to be symmetrical if al = al l and anti-symmetrical if a0 = 0 and al = al l = 0.
The constant preservation property is represented by the following relation:
N

l=N
al = 1 .
(13.5)
405
A discrete lter dened by the relation (13.4) is associated with the continuous convolution kernel:
N

G(x y) =
al (x y + lx)
(13.6)
l=N
where is a Dirac function. Simple computations show that the associated

transfer function G(k)
is of the form:
N

G(k)
=
al eklx
(13.7)
l=N
The real and imaginary parts of this transfer function are:

(G(k))
=
a0 +
N

(al + al ) cos(klx)
l=1

(G(k))
=
N

(al al ) sin(klx)
l=1
The continuous dierential operator can be associated with the discrete

lter (13.4). To do this, we introduce the Taylor expansion of the variable
about the point i:
in

(nx)l l
=
l!
xl i
(13.8)
(13.9)
l=0
By substituting in relation (13.4), we get:

l

l
i = 1 +
al x
i
xl
l=1
in which
al =
N
1
an n l
l!
n=N
We note that these lters belong to the class of elliptic lters as dened in
Sect. 2.1.3. In practice, the lters most used are the two following three-point
symmetrical lters:
a0 =
1
1
, a1 = a1 =
2
4
(13.10)
a0 =
2
1
, a1 = a1 =
3
6
(13.11)
406
13. Implementation
Table 13.1. Coecients of discrete nonsymmetrical lters. N is the number of

vanishing moments
N
a2
1
2
2
3
3
3
a1
a0
a1
a2
a3
1/4
1/2
7/8
5/8
15/16
3/4
5/8
1/4
3/8
3/8
1/4
3/8
1/4
3/8
1/8
3/8
1/4
1/16
1/8
1/8
1/16
1/16
1/4
1/4
1/16
a4
1/16
Vasilyev et al. [728] have dened nonsymmetric lters, which have a large
number of vanishing moments2 . These lters are presented in Table 13.1.
Linearly constrained lters can also be dened, which satisfy additional
constraints.
Optimized lters, whose coecients are computed to minimize the functional
/x
/x

t (k)})2 dk +

t (k)})2 dk , (13.12)
({G(k)
G
({G(k)
G
0
t (k) is the targeted transfer function, have been proposed [632, 728].
where G
These lters ensure a better spectral response of the lter, resulting in a better
localization of the information in spectral space.
For certain uses, such as in the Germano-Lilly dynamic procedure, the
characteristic length of the discrete lter, denoted d , has to be known. For
a denite positive lter, one measure of this length is obtained by computing
the standard deviation of the associated convolution lter [563, 461]:
'
+
d =
2 G()d
12
(13.13)
The characteristic lengths of the two three-point

lters mentioned above
are 2x for the (1/6, 2/3, 1/6) lter and 6x for the (1/4, 1/2, 1/4) lter.
This method of evaluating the characteristic lengths of the discrete lters
is inecient for lters whose second-order moment is zero. One alternative
is work directly with the associated transfer function and dene the wave
number associated with the discrete lter, as for the one for which the transfer
function takes the value 1/2. Let kd be this wave number. The discrete lter
cuto length is now evaluated as:
d =
2
kd
(13.14)
These lters are necessary to obtain high-order commuting discrete lters (see
Sect. 2.2.2).
407
Implementation of test lters for the dynamic procedure within the spectral element framework is discussed by Blackburn and Schmidt [60]. The
general unstructured case is discussed in [297]. The case of the nite element
method is addressed by Kollman et al. [400].
13.2.2 Extension to the Multi-Dimensional Case
For Cartesian grids, we extend to the multidimensional case by applying
a one-dimensional lter in each direction of space. This application can be
performed simultaneously or sequentially. When simultaneously, the multidimensional lter is written symbolically as a summation:
1
Gi
n i=1
n
Gn =
(13.15)
where n is the dimension of the space and Gi the one-dimensional lter in

the ith direction of space. If applied sequentially, the resulting lter takes the
form of a product:
n

n
G =
Gi .
(13.16)
i=1
The multidimensional lters constructed by these two techniques from the

same one-dimensional lter are not the same in the sense that their transfer
functions and equivalent dierential operators are not the same. In practice,
it is the product construction that is most often used, for two reasons:
This approach makes it possible to call the easily implemented onedimensional ltering routines sequentially.
Such lters are more sensitive to the cross modes than are the lters
constructed by summation, and allow a better analysis of the threedimensional aspect of the eld.
13.2.3 Extension to the General Case. Convolution Filters
For structured curvilinear grids (or Cartesian grids with variable mesh size),
one method is to employ the lters dened in the uniform Cartesian grid
and take no account of the variations of the metric coecients. This method,
which is equivalent to applying the lter in a reference space, is very easy
to implement but allows no control of the discrete lter transfer function or
its equivalent dierential operator. So it should be used only for grids whose
metric coecients vary slowly.
Another method that is completely general and applicable to unstructured
grids consists in dening the discrete lter by discretizing a chosen dierential
operator. The weighting coecients of the neighboring nodes are then the
coecients of the discrete scheme associated with this dierential operator. In
408
13. Implementation
practice, this method is most often used by discretizing second-order elliptic

operators:
2
= (Id + 2 ) ,
(13.17)
where is a positive constant and the desired cuto length. Limiting the
operator to the second order yields lters with compact support using only
the immediate neighbors of each node. This has the advantages of:
Making it possible to dene operators that cost little to implement;
Making a multiblock and/or multidomain technique easier to use, and the
boundary conditions easier to process.
The fast-decay convolution lters (box or Gaussian) can thus be approximated by discretizing the dierential operators associated with them. These
operators are described in Sect. 7.2.1. The sharp cuto lter, which is not of
compact support, is used only when fast Fourier transforms are usable, which
implies that the grid step is constant and the data periodic.
Another possibility for deriving discrete lters on general meshes is to
compute the weight of neighbouring points by solving a linear system based
on Taylor series expansions [632, 490].
13.2.4 High-Order Elliptic Filters
Convolution lters are non-local, and may sometimes be dicult to use together with complex numerical algorithms (multidomain topology, unstructured grid, ...). An alternative, that can be implemented with all numerical
methods, consists in high-order elliptic lters [553].
The ltered variable is computed as being the solution of the general
elliptic equation:
[(2 )m + Id] = ,
m1 .
(13.18)
High values of m make it possible to obtain very sharp lters in the

spectral space. Mullen and Fischer show that the solution of equation (13.18)
can be approximated through numerical solution of a much simpler problem,
namely the Poisson equation
2 = .
(13.19)
13.3 Implementation of the Structure Function Models

In order to use the subgrid viscosity model based on the second-order structure function or the third-order structure function of the velocity (see p. 124
and p. 126), we have to establish a discrete approximation of the operator:

2
[u(x, t) u(x + x , t)] d3 x .
(13.20)
DLL (x, r, t) =
|x |=r
13.3 Implementation of the Structure Function Models
409
In practice, this integration is approximated as a sum of the contributions

of the neighboring points. In the case of uniform Cartesian grid with x = r,
the structure function is evaluated at the index point (i, j, k) by the relation:
DLL (x, t)i,j,k
1
|ui,j,k ui+1,j,k |2 + |ui,j,k ui1,j,k |2
6
|ui,j,k ui,j+1,k |2 + |ui,j,k ui,j1,k |2

|ui,j,k ui,j,k+1 |2 + |ui,j,k ui,j,k1 |2 . (13.21)
=
+
+
When the grid is non-uniform or when x = r, an interpolation technique

has to be used to compute the integral. Rather than use a linear interpolation, it is recommended that the interpolation method be based on physical
knowledge. So in the isotropic homogeneous turbulence case, when we see
that we have:
DLL (x, r, t) = 4.82K0(r)2/3
2/3
DLL (x, r , t) = 4.82K0(r )
we deduce the proportionality relation:

DLL (x, r, t) = DLL (x, r , t)
r 2/3
r
(13.22)
Relation (13.21) is thus generalized to the form:

1
|u(x) u(x + i )|2
n i=1
n
DLL (x, r, t) =
r
i
2/3
,
(13.23)
where n is the number of neighboring points retained for computing the

structure function and i the distance of the ith point to the point where
this function is evaluated.
It has already been said that the second-order Structure Function model
in its original form exhibits defects similar to those of the Smagorinsky model
because of the uncertainty relation that prevents any good frequency localization of the information. One way of at least partly remedying this problem
is to look for the structure function evaluation information only in the directions of statistical homogeneity of the solution. This is done by evaluating the
structure function only from points located in the directions of periodicity of
the solution. This way, the mean gradient of the solution is not taken into
account in the evaluation of the subgrid viscosity. We again nd here an idea
similar to the one on which the splitting technique is based, in Sect. 6.3.3.
14. Examples of Applications
This chapter gives a few examples of large-eddy simulation applications that

are representative of their accomplishments in the sense that they correspond
either to ows that are very frequently treated or to congurations that
stretch the technique of today to its limits.
14.1 Homogeneous Turbulence

14.1.1 Isotropic Homogeneous Turbulence
Problem Description. Isotropic homogeneous turbulence is the simplest
turbulent ow on which subgrid models can be validated. The physical description of this ow is precisely the one on which the very great majority
of these models are constructed. Moreover, the ows statistical homogeneity
makes it possible to use periodicity conditions for the computation, and highaccuracy numerical methods: pseudo-spectral methods can be used, optimally
reducing the eect of the numerical error on the solution.
Because of the great simplicity of this ow, most subgrid models yield very
satisfactory results in terms of the statistical moments of the velocity eld
and the integral scales, which reduces the discriminatory range of this test
case. It is nonetheless widely used for fundamental type studies of turbulence
and modeling.
Two types of such ow are considered:
Freely decaying isotropic homogeneous turbulence in which the energy is
initially distributed in a narrow spectral band and then, as the energy
cascade sets in, is directed toward the small scales and nally dissipated at
the cuto by the subgrid model. During the time the cascade is setting in,
the kinetic energy remains constant, and later declines. The computation
can be validated by comparison with decay laws developed by analytical
theories (see [439]) or by comparison with experimental data.
Sustained isotropic homogeneous turbulence, in which total dissipation of
the kinetic energy is prevented by injecting energy at each time step, for example by maintaining a constant energy level in the wave vectors of a given
norm. After a transitory phase, an equilibrium solution is established including an inertial range. The computation is validated by comparison
412
with theoretical or experimental data concerning the inertial region, and

quantities associated with the large scales.
A few Realizations. The rst large-eddy simulations of the free-decaying
type were performed at the end of the seventies and early eighties [136] with
resolutions of the order of 163 and 323 . Self-similar solutions could not be
obtained with these resolutions because the integral scale becomes larger
than the computational domain. However, the comparison with ltered experimental data turns out to be satisfactory [40]. More recent simulations
(for example [441, 514]) performed with dierent subgrid models on grids of
1283 points have yielded data in agreement analytical theories for the kinetic
energy decay. Higher-resolution simulations have been performed.
In the sustained case, Chasnov [120] is an example of achieving self-similar
solutions in agreement with theory for resolutions of 643 and 1283, though
with an over-evaluation of the Kolmogorov constant. More recently, Fureby
et al. [231] have tested six subgrid models and a case of implicit numerical
diusion on a 323 grid. The conclusions of this work are that the dierent
realizations, including the one based on articial dissipation, are nearly indiscernable in terms of the quantities linked to the resolved eld, and are in
good agreement with data yielded by a direct numerical simulation.
Though isotropic homogeneous turbulence is statistically the simplest case
of turbulent ow, it possesses a complex dynamics resulting from the interactions of very many elongated vortex structures called worms. These structures are illustrated in Fig. 14.1, which comes from a large-eddy simulation
of freely decaying isotropic homogeneous turbulence on a 1283 grid. Obtaining good results therefore implies that the simulation is capable of reecting
the dynamics of these structures correctly. We clearly see here the dierence
with the RANS approach (see Chap. 1), for which isotropic homogeneous
turbulence is a zero-dimension problem: for the large-eddy simulation, this
problem is fully three-dimensional and reveals all the aspects of this technique
(modeling errors, lter competition, and so forth).
14.1.2 Anisotropic Homogeneous Turbulence
Anisotropic homogeneous turbulence allows a better analysis of the subgrid
models because the dynamics is more complex, while optimal numerical methods are retained. So it can be expected that this type of ow oers more
discriminatory test cases for the subgrid models than do isotropic ows.
Bardina et al. [40] performed a set of simulations corresponding to the
following three cases in the early eighties:
Homogeneous turbulence subjected to a solid-body rotation. Good agreement is measured with experimental data using a de-ltering technique, on
a 323 grid with a Smagorinsky model (5.90). The eects of rotation on the
turbulence are conrmed, i.e. a reduction in the dissipation of the kinetic
energy.
14.1 Homogeneous Turbulence
413
Fig. 14.1. Isotropic homogeneous turbulence. Instantaneous view of vortices (illustrated by an iso-value surface of the vorticity). Courtesy of E. Garnier, ONERA.
Homogeneous turbulence subjected to pure strain: still a 323 grid, with

results in good agreement with experimental data concerning the turbulent
intensity, using the Smagorinsky model and mixed SmagorinskyBardina
model (7.125). The best results are obtained with the latter.
Homogeneous turbulence subjected to a deformation and rotation: simulations are performed on a 323 grid with the two previously mentioned
models. No validation is presented, for lack of reference data.
Simulations of homogeneous turbulence subjected to sequential shearing
have also been performed by Dang [162] on a 163 grid with several eective viscosity models, yielding good results concerning the prediction of the
414
anisotropy of the resolved scales. Similar computations have also been performed by Aupoix [23].
14.2 Flows Possessing a Direction of Inhomogeneity

These ows represent the next level of complexity. The presence of a direction
of inhomogeneity prompts the use of lower-order numerical methods, at least
for this inhomogeneity, and boundary conditions. Also, more complex physical mechanisms are at play that can exceed the possibilities of the subgrid
models.
14.2.1 Time-Evolving Plane Channel
Problem Description. The time-evolving plane channel ow is a ow between two innite parallel at plates having the same velocity. The time character is due to the fact that we consider the velocity eld as being periodic
in both directions parallel to the plates. Since the pressure is not periodic,
a forcing term corresponding to the mean pressure gradient is added in the
form of a source term in the momentum equations. The ow is characterized
by the uid viscosity, the distance between the plates, and the uid velocity. This academic conguration is used for investigating the properties of
a turbulent ow in the presence of solid walls, and is a widely used test case.
Turbulence is generated within the boundary layers that develop along each
solid wall (see Sect. 10.2.1). It is the driving mechanism here, which must
imperatively be simulated with accuracy to obtain reliable results. To do so,
the grid has to be rened near the surfaces, which raises numerical problems
with respect to the homogeneous turbulence. Moreover, the subgrid models
must be able to preserve these driving mechanisms.
The ow topology is illustrated in the iso-value surface plot of the streamwise velocity in Fig. 14.2.
A Few Realizations. There are dozens of numerical realizations of plane
channel ows. The rst are from Deardor [172] in 1970. The rst landmark
results obtained by solving the dynamics of the near-wall region are due
to Moin and Kim [537] in 1982. The characteristics of the computations
presented in the four reference works [537, 591, 653, 411] are reported in
Table 14.1. These computations are representative of the various techniques
employed by most authors. The Table summarizes the following information:
The Reynolds number Rec referenced to the channel mid-height and mean
velocity at the center of the channel.
The dimensions of the computational domain expressed as a function of the
channel mid-height. The domain dimensions must be greater than those of
the driving mechanisms in the near-wall region (see Sect. 10.2.1).
415
Fig. 14.2. Plane channel ow. Iso-surface of instantaneous streamwise velocity

uctuations. Courtesy of E. Montreuil, ONERA.
The number of grid points. Simulations generally include few points because the solution is bi-periodical. The computations at high Reynolds
number without wall model presented [411] use a hierarchic grid technique
with nine grid levels (symbol +H).
The subgrid model used (Sc is the Schumann subgrid viscosity model
(6.59) and Dyn the dynamic Smagorinsky model (5.149)). Only two
Table 14.1. Characteristics of time-evolving plane channel ow computations.
Ref.
Rec
Lx Ly Lz
Nx Ny Nz
SGS Model
Wall
O(x )
O(t )
[537]
13800
2 2
64 64 128
Sc
S/2
2
[591]
47100
5/2 /2 2
64 81 80
Dyn
S/T
3
[653]
1, 5.105
421
64 32 32
Sc
MSc
2
2
[411]
1, 09.105
2 /2 2
2.106 +H
Dyn
S/Gsp
3
416
models are used in the computations presented, but most existing models have been applied to this conguration.
The treatment of the solid walls ( is the no-slip condition, MSc the
Schumann wall model (10.28) to (10.30)). A single computation based on
a wall model is presented, knowing that nearly all the models mentioned
in Chap. 10 have been used for dealing with this ow.
The accuracy of the space discretization schemes. Since the directions of
statistical homogeneity are linked to directions of periodicity in the solution, pseudo-spectral methods are often used for processing them. This is
true of all the computations presented, identied by an S, except for reference [653], which presents a second-order accurate nite volume method.
In the normal direction, three cases are presented here: use of second-order
accurate schemes (identied by a 2), of a Chebyshev method (T), and
a Galerkin method based on B-splines (Gsp). The eect of the numerical
error on the solution can be reduced by using higher-order methods, which
are consequently recommended by many authors.
The accuracy of the time integration. The convection term is usually
treated explicitly (Runge-Kutta or Adams-Bashforth scheme) and the diffusion terms implicitly (Crank-Nicolson or second-order backward Euler
scheme). Nearly all the computations are performed with second- or thirdorder accuracy.
The results obtained on this conguration are usually in good agreement
with experimental data, and especially as concerns the rst-order (mean eld)
and second-order (Reynolds stresses) statistical moments. Examples of data
for these quantities are shown in Figs. 14.3 and 14.4. The mean longitudinal
Fig. 14.3. Plane channel ow. Mean longitudinal velocity prole referenced to
the friction velocity, compared with a theoretical turbulent boundary layer prole. Small circle symbols: LES computation. Lines: theoretical prole. Courtesy of
E. Montreuil, ONERA.
417
Fig. 14.4. Plane channel ow. Proles of solved Reynolds stresses with respect
to the friction velocity, compared with data from a direct numerical simulation
computation. Dot symbols: direct numerical simulation. Lines: LES computation.
Courtesy of E. Montreuil, ONERA.
velocity prole is compared here with a theoretical turbulent boundary layer

solution, and very good agreement with it is observed. It should be noted
that the logarithmic region is relatively small, which is due to the fact that
the Reynolds number for the computation is low (Re = 180). The proles of
the three main Reynolds stresses are compared with those obtained by direct
numerical simulation on a grid including about twenty times more degrees
of freedom. Although these stresses are calculated only from the resolved
eld, such that the contribution of the subgrid scales is not included, we
observe that the agreement with the reference solution is very satisfactory.
This illustrates the fact that data obtained by large-eddy simulation can be
used directly in practice without recourse to a de-ltering technique. In the
present case, the very good quality of the results can be explained by the fact
that a large part of the kinetic energy of the exact solution is contained in
the resolved scales.
The quality of the results is essentially due to the resolution of the dynamics in the near-wall region (z + < 100). This implies that, if a wall
model is not used, the computational grid is ne enough to represent the
dynamics of the vortex structures present, and that the subgrid models
employed do not alter this dynamics. Because of the necessary volumes
of the grids, this resolution constraint limits the possible Reynolds number. The largest friction Reynolds number achieved to date, using a hierarchic grid generation method, is Re = 4000 [411]. More results dealing
with high Reynolds number simulations can be found in [36]. The standard
subgrid viscosity models (Smagorinsky, Structure Function, and so forth)
are generally too dissipative and have to be used with caution (modi-
418
cation of the value of the model constant, wall damping function, and so
forth) [635]. Results concerning the transition to turbulence in this conguration are available in [600, 599]. Lastly, the results obtained for this ow
have been found to be very sensitive to numerical errors induced either by
the discrete numerical scheme or by the continuous form of the convection
term [409, 563].
14.2.2 Other Flows
Other examples of shear ows treated in the framework of the time-evolution
approximation can be found for:
plane mixing layer, in [674, 38];

rotating boundary layer, in [152];
free-surface ow, in [672];
boundary layers, in [401, 496, 498, 533];
round jet, in [214];
plane wake, in [263];
rotating plane channel, in [523, 524, 702, 422, 596, 421];
plane jet, in [431].
As in the case of the plane channel ows described above, periodicity

conditions are used in the directions of statistical homogeneity. The numerical
methods are generally dedicated to the particular conguration being treated
(with spectral methods used in certain directions) and are therefore optimal.
A forcing term is added in the momentum equations to take the driving
pressure gradient into account or avoid diusion of the base prole.
Transitional ows are more sensitive to the subgrid model and to the numerical errors, as an inhibition of the transition or re-laminarization of the
ow are possible. This is more especially true of ows (for example boundary layers) for which there exists a critical Reynolds number: the eective
Reynolds number of the simulation must remain above the threshold within
which the ow is laminar.
It should be noted that the boundary conditions in the inhomogeneous
direction raises little diculty for the ow congurations mentioned above.
These are either solid walls that are easily included numerically (except for
the procedure of including the dynamics), or outow conditions in regions
where the ow is potential. In the latter case, the computation domain boundary is generally pushed back as far as possible from the region being studied,
which reduces any spurious eects.
The types of results obtained, and their quality, are comparable to what
has already been presented for the plane channel ow.
14.3 Flows Having at Most One Direction of Homogeneity
419
14.3 Flows Having at Most One Direction

of Homogeneity
This type of ow introduces several additional diculties compared with the
previous cases. The limited number or total absence of directions of homogeneity makes it necessary in practice to use numerical methods of moderate order of accuracy (generally two, rarely more than four), and highly
anisotropic grids. The eect of the numerical error will therefore be high.
Moreover, most of these ows are in spatial expansion and the problems
related to the denition of the inow and outow conditions then appear.
Lastly, the ow dynamics becomes very complex, which accentuates the modeling problems.
14.3.1 Round Jet
Problem Description. The example of the round jet ow in spatial expansion is representative of the category of free shear ows in spatial expansion.
The case is restricted here to an isothermal, isochoric round jet ow piped
into a uniform, steady outer ow in a direction parallel to that of the jet.
Two main regions can be identied:
First, we nd a region at the pipe exit where the ow consists of a laminar
core called a potential cone, which is surrounded by an annular mixing
layer. The mixing layer is created by the inectional instability associated
Fig. 14.5. Round jet. Iso-surface of instantaneous vorticity (LES computation).

Exit plane in black. Courtesy of P. Comte, LEGI.
420
Fig. 14.6. Round jet. Iso-surface of instantaneous vorticity (LES computation).

Courtesy of P. Comte, LEGI.
with the decit velocity prole of the boundary layer on the wall of the
circular pipe. As the mixing layer thickens while moving away from the
pipe exit section, it reduces the diameter of the potential cone and also
induces an increase in the jet diameter.
After the potential cone disappears, we have a pure jet region where the
ow gradually reaches a regime corresponding to a similarity solution.
The rst region can be decomposed into two: a transition region where
the mixing layer has not yet reached its self-similar regime, and the similarity
region where it has. This organization is illustrated in Figs. 14.5 and 14.6,
representing respectively the iso-surfaces of vorticity and pressure obtained
from large-eddy simulation results. The vorticity eld very clearly shows the
transition of the annular mixing layer. The topology of the pressure eld
shows the existence of coherent structures.
Experimental and numerical analyses have shown that this ow is strongly
dependent on many parameters, which makes it highly discriminatory.
A Few Realizations. There are far fewer round jet simulations in the literature than there are plane channel ows. This is mainly due to the increased
diculty. Four of these realizations are described in the following, with their
characteristics listed in Table 14.2, which gives:
the Reynolds number ReD referenced to the initial jet diameter D and the
maximum of the mean initial velocity prole;
the computational domain dimensions referenced to the length D;
421
Table 14.2. Characteristics of the round jet computations.

Ref.
ReD
Lx Ly Lz
Nx Ny Nz
SGS Model
Inow
O(x )
O(t )
[634]
21000
10 11 11
101 121 121
MSM
U+b
2+up3
2
[573]
4
50.10
12 8 8
270000+H
Dyn
U+b
3+up3
2
[83]
21000
10 11 11
101 288 288
FSF
U+b
S/6
3
the number of grid points. All the grids used by the authors mentioned are
Cartesian. The symbol H designates the use of embedded grids (four grid
levels for [573]).
the subgrid model (MSM standing for the Mixed Scale Model (5.127);
Dyn the dynamic model (5.149); FSF the ltered Structure Function
model (5.264)). It should be noted that, for all the known realizations of
this ow, only the subgrid viscosity models have been used.
the freestream condition generation mode. The symbol U + b indicates
that the non-steady inow condition was generated by superimposing an
average steady prole and a random noise, as indicated in Sect. 10.3.2.
the overall order of accuracy of accuracy in space of the numerical method.
The symbol +up3 indicates that a third-order accurate upwind scheme is
used for the convection term to ensure computation stability. The computations presented in [83] rely on spectral schemes in the directions normal
to that of the jet.
the time accuracy of the method employed.
Examples of results obtained on this conguration are compared with experimental data in Figs. 14.7 to 14.11. The axial evolution of the location
of the point where the mean velocity is equal to half the maximum velocity is represented in Fig. 14.7. This quantity, which gives some indication
concerning the development of the annular mixing layer, remains constant
during the rst phases of evolution of the jet, which conrms the existence
of a potential cone. After the cone disappears, this quantity increases, which
indicates the beginning of the pure jet region. It is observed that the length
of the potential cone predicted by the computation is less than is observed
experimentally. Similar conclusions can be drawn from the axial evolution of
the average longitudinal velocity, which is presented in Fig. 14.8. The toorapid expansion of the pure jet region is accompanied by a strong decay of
the mean velocity1 . These symptoms are observed on all known large-eddy
simulation computations on this conguration and still have no precise explanation. Several hypotheses have been formulated concerning the dependency
1
This results from the conservation of the mass.
422
Fig. 14.7. Round jet. Axial evolution of the radial position of the point where
the mean velocity is half the maximum velocity. Dots: experimental data. Dotdashed lines: extrapolation of this data. Solid line: LES computation. Courtesy of
P. Comte, LEGI.
Fig. 14.8. Round jet. Axial evolution of the mean longitudinal velocity. Dots:
experimental data. Line: LES computation. Courtesy of P. Comte, LEGI.
423
on the initial perturbation, on the boundary conditions, or on the computational grid. The axial proles of two main Reynolds stresses are presented in
Figs. 14.9 and 14.10. These results are qualitatively correct. The Reynolds
stresses increase along the axis and exhibit a maximum in a region close to
the tip of the potential cone, which is in agreement with the experimental
observations. It is noted that the level of the longitudinal stress predicted by
the computation is higher than the experimental level in the pure jet region.
The peak observed on the downstream boundary of the computational domain is a spurious eect that is no doubt related to the outow condition.
Generally, it is noticed that the quality of the results is not as good as in
the case of the plane channel ow, which illustrates the fact that this ow is
a more complicated case for large-eddy simulation.
Lastly, the velocity spectra generated from the computation are presented
in Fig. 14.11. Over a decade, the computations recover a slope close to the
5/3 predicted by theory, and which is the foundation of the theoretical
analyses presented in the previous chapters. This indicates that the resolved
turbulent scales have physical behavior.
More generally, the conclusions given by the various authors are the following.
The dynamics of the numerical solution is consistent, i.e. the values produced are located within the bounds xed by the collected set of experimental measurements and the topology of the simulated ow exhibits the expected characteristics (potential cone, annular mixing layer, and so forth).
While the dynamics is consistent, it is nonetheless very dicult to reproduce a particular realization (for example with xed potential cone length
and maximum turbulent intensity).
The numerical solution exhibits a strong dependency on many parameters,
among which we nd:
the subgrid model, which allows a more or less rapid transition of the
annular mixing layer are consequently inuences the length of the potential cone and the turbulent intensity by modifying the eective Reynolds
number in the simulation. More dissipative models delay the development of the mixing layer, inducing the existence of a very long potential
cone.
the inow condition: the mixing layer transition is also strongly dependent on the amplitude and shape of the perturbations.
the numerical error, which can aect the turbulent of the annular mixing
layer and of the developed jet, especially during the transition phases.
Here it is a matter of an error controlled by the computational grid
and the numerical method. A dispersive error will have a tendency to
accelerate the transition and thereby shorten the potential cone. A dissipative error will have the inverse eect. With too coarse a grid, the
annular mixing layer cannot be represented correctly, which can induce
424
Fig. 14.9. Round jet. Axial evolution of the normalized longitudinal turbulent
intensity. Dots: experimental data. Line: LES computation. Courtesy of P. Comte,
LEGI.
Fig. 14.10. Round jet. Axial evolution of the normal turbulent intensity. Dots:
experimental data. Line: LES computation. Courtesy of P. Comte, LEGI.
425
Fig. 14.11. Round jet. Time (Left) and space (Right) spectra of solved turbulent
kinetic energy at dierent positions along the axis. Courtesy of P. Comte, LEGI.
it to thicken too quickly and thereby decrease the length of the potential
cone.
the size of the computational domain. The computation is sensitive to
the size of the computational domain, which modulates the eect of the
boundary conditions, especially the outow condition.
426
All the computations predict the dominant time frequency of the jet correctly, which is therefore not a pertinent parameter for analyzing the models nely.
The quality of the simulation is not a global character. Certain parameters
can be correctly predicted while others are not. This diversity in the robustness of the results with respect to the simulation parameters sometimes
makes it dicult to dene discriminatory parameters.
Other Examples of Free Shear Flows. Other examples of free shear ows
in spatial expansion have been simulated:
plane mixing layer (see [674, 38]);

planar jet (see [159]);
rectangular jet (see [272]);
swirling round jet (see [472]);
controlled round jet (see [157]);
plane wake (see [269, 740, 237]).
The conclusions drawn from the analysis of these dierent cases corroborate those explained previously for the round jet as concerns the quality of
the results and their dependency as a function of the computation parameters (subgrid model, grid, inow condition, computational domain, and so
forth). These conclusions are therefore valid for all free shear ows in spatial
expansion.
14.3.2 Backward Facing Step
Problem description. The ow over a backward facing step of innite span
is a generic example for understanding separated internal ows. It involves
most of the physical mechanisms encountered in this type of ow and is doubtless the best documented, both experimentally and numerically, of the ows
in this category. Its dynamics can be decomposed as follows. The boundary
layer that develops upstream of the step separates at the step corner, becoming a free shear layer. This layer expands in the recirculation region, thereby
entraining turbulent uid volumes. This entrainment phenomenon may inuence the development of the shear layer, which curves inward toward the
wall in the reattachment region and impacts with it. After the reattachment,
the boundary layer re-develops while relaxing toward a prole in equilibrium.
The topology of this ow is illustrated in Fig. 14.12, which is developed from
large-eddy simulation data. We observe rst the transition of the separated
shear layer, the formation of vortex structures in the impact area, and then
of hairpin structures in the boundary layer after the reattachment.
This ow brings out diculties in addition to those of the round jet,
because it adds the dynamics both of the free shear layers and of the near
wall region.
427
Fig. 14.12. Backward Facing Step. Iso-surface of instantaneous vorticity. Courtesy

of F. Delcayre, LEGI.
A Few Realizations. The methods used for simulating this ow are illustrated by four computations presented in Table 14.3. The parameters indicated are.
Reynolds number ReH , referenced to the step height H and the inow
velocity prole;
the dimensions of the computational domain, referenced to the length H;
the number of grid points used;
subgrid model used (Sc means the Schumann model (6.59); MSM the
Mixed Scale Model (5.127); SF the Structure Function model (5.102);
DynLoc the constrained localized dynamic model (5.207)). As before
for the round jet, only subgrid viscosity models have been used in the
conguration.
inow condition generation mode: U + b means the same thing as before,
while P designates the use of a precursor, which in this case is a large-eddy
simulation of a plane channel ow in [226]; Ca indicates the use of an inow
channel to allow the development of a realistic turbulence upstream of
the separation. Depending on the author, the length of this channel is
between four and ten H.
treatment of the solid walls: is the no-slip condition; MSc the Schumann
wall model, (10.28) to (10.30); MGz the Gr
otzbach wall model (10.31). It
428
Table 14.3. Backward facing step computation characteristics.

Ref.
ReH
Lx Ly Lz
Nx Ny Nz
SGS Model
Inow
Wall
O(x )
O(t )
[226]
[637]
5
1, 65.10
16 4 2
128 32 32
Sc
P
MSc
2
2
11200
20 4 2, 5
201 31 51
MSM
U+b
2+up3
2
[673]
38000
30 5 2, 5
200 30 30
FS
U+b
MLog
2+up3
2
[261]
28 000
30 3 4
244 96 96
DynLoc
U+b,Ca
2
3
can be seen that the use of wall models reduces the number of points
considerably and makes it possible to simulate ows with high Reynolds
numbers.
spatial accuracy of the numerical method;
time accuracy of the numerical method.
The results the various authors have obtained are generally in good qualitative agreement with experimental data: the ow topology is recovered and
the realizations show the existence of coherent structures similar to those
observed in the laboratory. On the other hand, there is much more diculty
obtaining quantitative agreement whenever this is possible at all (only reference [261] produces results in satisfactory agreement on the mean velocity
eld and turbulent intensity). This is due to the very high sensitivity of the
result to the computation parameters. For example, variations of the order
of 100% of the average length of the recirculation region have been recorded
when the inow boundary condition or subgrid model are manipulated. This
sensitivity stems from the fact that the ow dynamics is governed by that
of the separated shear layer, so the problems mentioned before concerning
free shear ows crop up here. We also note a tendency to under-estimate the
value of the mean velocity in the recirculation area. However, as in the case
of the round jet, the simulated physics does correspond to that of a backward
facing step ow. This is illustrated by the mean velocity proles and resolved
Reynolds stresses in Fig. 14.13, and the pressure spectra in Fig. 14.14. The
good agreement with experimental data in the prediction of the mean eld
and Reynolds stresses proves the theoretical consistency of the computation. This agreement is even clearer if we analyze the spectra. Near the step,
the mixing layer dynamics is dominated by frequencies associated with the
Kelvin-Helmholtz instability. The predicted value of the dominant frequency
is in very good agreement with experimental observations. The double peak
at the second measurement point shows that the simulation is capable of reecting the low-frequency apping mechanism of the separated region, and
still remain in good agreement with experimental observations.
429
Fig. 14.13. Backward Facing Step. Mean velocity and Reynolds stresses proles
at the reattachment point. Triangles and solid line: LES computations. Squares:
experimental data. Courtesy of F. Delcayre, LEGI.
Fig. 14.14. Backward Facing Step. Pressure spectra. Squares: in the free shear
layer near the step corner. Triangles: in the separated region near the reattachment
point. Courtesy of F. Delcayre, LEGI.
Also, it seems that the use of wall models does not aect the dynamics
of this shear layer greatly. It becomes possible to deal with higher Reynolds
numbers, but at the price of losing some of the quality of the results as concerns the wall surface terms (friction, pressure coecient) in the recirculation
zone [87]. Solutions for this conguration turn out to be highly dependent on
the subgrid model: a too dissipative model will delay the development of the
separated shear layer, pushing away the position of the reattachment point.
430
The results are also found to be dependent on the size of the domain and
the neness of the mesh in the spanwise direction, because these parameters
aect the development of the mixing layer emanating from the step corner.
A spanwise domain width of 4 to 6 H is considered to be a minimum in order
to be able to capture the three-dimensional mechanisms at low frequencies.
Lastly, the time frequencies associated with the separated zone dynamics are
robust parameters in the sense that they are often predicted with precision.
14.3.3 Square-Section Cylinder
Problem description. The square-section innite-span cylinder is a good
example of separated external ows around blu bodies. This type of conguration involves phenomena as diversied as the impact of the ow on a body,
its separation (and possible reattachment) on the body surface, the formation
of a near-wake region, and alternating vortex street, and the development of
the wake up to a self-similar solution. Each of these phenomena poses its own
particular numerical and modeling problems.
Realizations. This ow was chosen as an international test case for largeeddy simulation, and has consequently served as a basis for many computations, which are mostly summarized in [618] (see also [555, 729, 730, 678,
233, 94, 652] for a discussion of this test case). The test case parameters are:
a Reynolds number ReD , referenced to the length D of the cylinder edge
and the freestream velocity, equal to 22,000, and a computational domain
of 20D 4D 14D. The span is assumed to be innite and a periodicity
condition is used in this direction.
None of the sixteen computations collected in [618] produces an overall
good agreement with experimental data, i.e. is capable of predicting all of the
following parameters with an error of less than 30%: average lift and drag;
drag and lift variances; main vortex shedding frequency; and average length
of the separated region behind the cylinder. Average lift and drag, as well
as the vortex shedding frequency, are very often predicted very satisfactorily.
This is due to the fact that these quantities do not depend on the small
scale turbulence and are governed by Von Karman structures, which are very
large in size. The length of the recirculation region behind the cylinder is
very often under-estimated, as is the amplitude of the mean velocity in this
region. Also, the mean velocity in the wake is only very rarely in agreement
the experimental data.
The numerical methods used are of moderate order of accuracy (at
most second-order in space and third-order in time), and third-order upwind schemes are often used for the convection term. Only subgrid viscosity
models have been used (Smagorinsky model, various dynamic models, Mixed
Scale Model). Certain authors use wall models at the cylinder surface.
The lack of agreement with experimental data is explained by the very
high sensitivity of the dierent driving mechanisms to the numerical errors
431
and to the diusion introduced by the models. So we again nd here the problems mentioned above for the case of the backward facing step, but now they
are amplied by the fact that, in order to master the impact phenomenon
numerically, the numerical diusion introduced is much stronger than in the
former case. Also, as most of the grids used are Cartesian and monodomain,
the resolution near the cylinder is too weak to allow a satisfactory representation of the boundary layers.
14.3.4 Other Examples
Many other ows have been examined by large-eddy simulation.
Among wall-bounded ows without separation, we may mention: at plate
boundary layer [237, 205, 463, 457, 585]; boundary layer on a curved surface in the presence of Gortler vortices [460]; ow in a circular-section toric
pipe [653]; three-dimensional boundary layer [767, 366]; juncture ow [687];
ow in a rotating pipe [777, 215]; ow in a rotating square duct [576]; ow
in a square/rectangular annular duct [774, 65].
Examples of recirculating ows are: conned coaxial round jets [9]; ow
around a wing section of innite span at incidence [347, 371, 348, 349, 425,
78, 502, 494] (see Fig. 14.15); ow in a planar asymmetric diuser [367,
213, 372, 369]; ow around a cube mounted on a at plate [618, 556, 757,
493, 652, 407]; ow around a circular-section cylinder [52, 519, 520, 362,
75, 410, 715, 77, 360, 76]; ow in tube bundles [620]; ow in a lid-driven
cavity [799, 179]; ow in a ribbed channel/duct [779, 142, 778, 151, 361]; jet
Fig. 14.15. Flow around a wing at high incidence: isosurface of instantaneous

vorticity. Courtesy of R. Lardat and L. Ta Phuoc, LIMSI.
432
impacting a at plate [739, 616, 574, 51, 718]; boundary layer on a wavy
surface or a bump [412, 197, 770, 302, 768, 769, 642, 522]; ow over a swept
wedge [330]; ow past a blunt trailing edge [484, 754, 564, 571]; separated
boundary layer [88, 758]; aircraft wake vortices [286]; axisymmetric pistoncylinder ow [732]; ow around an oscillating cylinder [721]; ow around
a road vehicle [731, 408]; ow around a 3D wing [329]; ow around a square
cylinder [399, 393]; ow around a forward-backward facing step [5]; ow in
reversing systems [61]; ow in a blade passage [147, 516]; ow pas a swept
fence [178]; ow over a sphere [146].
14.4 Industrial Applications

14.4.1 Large-Eddy Simulation for Nuclear Power Plants
The increasing power of supercomputers and of numerical description of unsteady ows allows the simulation of complex ows relevant to industrial
congurations in accident scenarios obviously not available through experiments. Simulation is used to capture large features of the ow and to focus on
major events, at least concerning qualitative behavior. The present example
deals with induced rupture of the primary circuit in a pressurized water reactor. The simulation domain includes the reactor vessel, the steam generator
and their connection through the hot leg (see Fig. 14.16): the dimensions of
the real domain cover several meters. For the calculation, the steam generator tubes have been grouped into nine, thus decreasing the number of control
volumes (a model has been developed to take into account this change in the
geometry). Furthermore, the tubes length has been reduced in order to limit
the computer time consumption.
In the case of the total loss of the heat sink (very hypothetical severe accident scenario) the ow circulation driven by the primary pump is stopped,
the coolant is pure vapour (T > 1000 C) and natural convection develops in
the circuit. The question here is to investigate the temperature distribution
in the wall of the hot leg and in the tubes of the steam generator, in order
to analyze their mechanical constraints. For that it is crucial to get the best
discretization of the related zones. This is achieved by concentrating the grid
points in this area. Figure 14.16 shows the mesh of each tube of a simplied
steam generator composed of 400 tubes (3300 in reality). Figure 14.17 (top)
provides a global view of the ow in the hot leg, showing a clear stratication
and a return of ow into the core. Figure 14.17 (bottom) displays a cut in the
hot leg at dierent locations, showing an increase of turbulence and mixing
as the section is closer to the steam generator: this mixing leads to homogenization of the temperature and reduced mechanical constraints. The present
simulation was performed using the TRIO-U/PRICELES code developed at
CEA Grenoble (in the DTP/SMTH/LDTA laboratory): it required around
106 grid points using a tetrahedral discretization. The underlying numerical
433
Fig. 14.16. Top: global view of the unstructured mesh. A coarse discretization is
chosen for the vessel, taking into account here a realistic boundary condition. The
resolution is ner in a region of interest, say the steam generator region (tubes and
plenum). Bottom: local view of the mesh of the steam generator: 400 individual
tubes are independently meshed. Courtesy of IRSN.
434
Fig. 14.17. Top: a cut in the instantaneous temperature eld in the whole system,
showing a stratication in the hot leg that propagates in the direction of the
steam generator tubes; a back ow develops in the vessel. Bottom: cuts of the same
eld near the steam generator, plus velocity vectors. Stratication is perturbated
by the turbulent ow. Courtesy of IRSN.
method is a centered second-order accurate Finite Element based discretization, used together with a standard Smagorinsky model. All simulations were
performed by U. Bieder (CEA) and H. Mutelle (IRSN) for the French Nuclear
Safety Institute (IRSN).
435
For other situations where experimental data are available, qualitative as

well as quantitative features are looked at, as in more traditional disciplines.
14.4.2 Flow in a Mixed-Flow Pump
This example of the use of large-eddy simulation for turbomachinery ows
is due to C. Kato et al. [380]. These authors computed the internal ow in
a mixed-ow pump stage with a high designed specic speed.
A view of the computational domain is shown in Fig. 14.18. The computational domain includes: the upstream inlet pipe, the two regulation plates,
a four-blade impeller and the diuser, leading to a very complex conguration, both from the geometrical and physical points of view.
In order to take the rotation into account, the authors make use of
a Chimera-type technique: a moving boundary interface in the ow eld is
treated with overset grids from multiple frames of reference. A computational
grid that rotates along with the impeller is used to compute the ow within
the impeller, and stationary grids are used for stationary parts of the pump.
Each grid includes appropriate margins of overlap with its neighboring grids
in order to allow appropriate interpolations, and coordinate transformation
to take into account the dierent frames of reference.
Two mesh resolutions have been investigated: a coarse grid with a total
of 1.7 106 points and a ne grid with 5 106 points. A partial view of the
surface mesh on the impeller is shown in Fig. 14.19. The boundary conditions
are the following: no-slip boundary conditions are used on solid walls, and
the turbulent pipe ow at the inlet is obtained by performing an auxiliary
large-eddy simulation of an innite pipe ow.
Fig. 14.18. Flow in a mixed-ow pump. View of the global computational domain.
Courtesy of C. Kato, University of Tokyo.
436
Fig. 14.19. Partial view of the surface mesh. Courtesy of C. Kato, University of
Tokyo.
A classical Smagorinsky model is employed, with a Van Driest damping function enforcing a vanishing subgrid viscosity in the near-wall region.
A streamwise-upwind nite-element method is used to discretize the Navier
Stokes equations. Time integration is based on the explicit Euler scheme,
but shifts the spatial residuals of the governing equations in the upstream
direction of the local ow. The magnitude of this shift is one half of the time
increment multiplied by the magnitude of the local ow velocity, yielding
an exact cancellation of the rst-order error terms. The resulting method is
second-order accurate in space and time.
Several operating conditions of the pump have been investigated using
large-eddy simulations. All of them correspond to partial load conditions, i.e.
o-design operating conditions.
Comparison of computed and measured head-ow characteristics is shown
in Fig. 14.20, showing a very good agreement between ne-grid simulations
and experiments. Coarse grid simulation is seen to yield larger discrepancies
when design operating conditions are approached, because the mesh is not
ne enough to capture accurately the attached boundary-layer dynamics. The
normalized head-ow characteristic is dened here as the relation between the
ow-rate of the pump and its total-pressure rise. After stall onset (Q/Qd
0.55, where Q is the mass owrate), both grids yield very similar descriptions
of the ow. As in the previous case, Qd is the mass ow-rate related to the
design point.
437
Fig. 14.20. Comparison of measured and predicted head-ow characteristics as

a function of the mass ow rate Q with reference to the design ow rate Qd .
Courtesy of C. Kato, University of Tokyo.
An example of comparisons of computed and experimental data is shown

in Fig. 14.21.
14.4.3 Flow Around a Landing Gear Conguration
The ow around a realistic landing gear conguration was studied by Souliez
et al. [679]. The main purpose of this work was the prediction of the far-eld
noise generated by the turbulent uctuations.
The most complex geometry includes four wheels and two lateral struts.
An unstructured mesh with 135 000 triangles on the surface of the landing
gear and 1.2 million tetrahedral cells was used.
The simulation is run without an explicit subgrid model, and then belongs
to the MILES-like group of large-eddy simulations. The PUMA code used
for this simulation is based on the nite volume approach. Explicit timeintegration is carried out using a RungeKutta scheme.
Typical results dealing with the mean surface pressure and the topology
of the instantaneous ow are shown in Figs. 14.22 and 14.23.
14.4.4 Flow Around a Full-Scale Car
Large-eddy simulations is also used in the automotive industry during the
design of new cars. The main purposes are the prediction of traditional aerodynamic parameters (drag, lift, etc.) and aeroacoustics.
438
Fig. 14.21. Predicted and measured phase-averaged velocity proles at the impellers inlet cross-section for Q/Qd = 0.43 (post-stall operating condition). Courtesy of C. Kato, University of Tokyo.
The example presented here deals with the large-eddy simulations of the
ow around a full-scale car model. The wheels and the oor are xed, as in
the wind tunnel experiments. The Reynolds number is 2 684 563 per meter.
The numerical simulation is carried out using the Powerow code, which
is based on the Lattice Boltzmann approach (see [127] for an introduction).
A Cartesian grid is used, which is composed of 20 106 cells. The size of the
smallest mesh is 7 mm. Subgrid scales are taken into account using a twoequation k subgrid model. A wall model is used at solid boundaries.
Fig. 14.22. Predicted mean surface pressure on the landing gear. Courtesy of
F. Souliez and L. Long, Pennsylvania State University.
14.5 Lessons
439
Fig. 14.23. Instantaneous vorticity laments. Courtesy of F. Souliez and L. Long,

Pennsylvania State University.
Typical results are displayed in Figs. 14.2414.27. The very complex topology of the ow is recovered, exhibiting very intense vortical structures and
pressure variations. The drag is predicted to be within 1% error in comparison
with experimental data.
14.5 Lessons
14.5.1 General Lessons
We can draw the following lessons concerning the large-eddy simulation technique from the computations mentioned above:
When the technique is used for dealing with the ideal case in which it was
derived (homogeneous turbulence, optimal numerical method), it yields
very good results. The vast majority of subgrid models produce results that
are indiscernable from reality, which removes any discriminatory character
from this type of test case, which in fact can only be used to assess the
consistency of the method.
Extending the technique to inhomogeneous cases brings up many other
problems, concerning both the physical modeling (subgrid models) and
the numerical method. The latter point becomes crucial because the use of
numerical methods of moderate order of accuracy (generally two) greatly
increases the eect of the numerical error. This is accentuated by the use of
articial dissipations for stabilizing the simulation in sti cases (strong
under-resolution, strong gradients). This error seems to be reducable by
rening the computational grid, which is done more and more by using
adaptive grids (local adaptation or enrichment).
440
Fig. 14.24. Isocontours of total pressure. Courtesy of Renault-Aerodynamic Department.
Fig. 14.25. Streamlines. Courtesy of Renault-Aerodynamic Department.
14.5 Lessons
441
Shear ows show themselves to be very strongly dependent on the inow

condition when this is unsteady. Generating these conditions is still an
open problem.
The quality of the results is variable but, for each conguration, robust,
correctly predicted parameters exist. The physics simulated is often consistent in that it exhibits the generic features that are observed experimentally
but does not necessarily correspond to a desired target realization. This is
due to the dependency on the many simulation parameters.
The quality of the results is subordinate to the correct representation of
the ow driving mechanisms (transition, near-wall dynamics, and so forth).
Low numerical error and consistent modeling are therefore mandatory in
those regions where these mechanisms occur. The other regions of the ow
where the energy cascade is the dominant mechanism are of lesser importance.
When the ow dynamics becomes complex, subgrid viscosity models are
often used. This is because they provide a clear kinetic energy dissipation
and therefore stabilize the simulation. This stabilizing character seems to
become predominant compared with the physical quality of the modeling
insofar as the numerical diculties increase (with the presence of strong
shear and highly heterogeneous grids, and so forth).
There is a consensus today that the numerical method used must be accurate to at least the second order in space and time. First-order accurate numerical dissipations are totally proscribed. Third-order accurate methods
in time are rarely used. As concerns the spatial accuracy, satisfactory results are obtained by certain authors with second-order accurate methods,
but higher-order accurate schemes are often used. Numerical stabilization
methods (upwind scheme, articial dissipation, smoothing, and so forth)
should be used only when absolutely necessary.
Large-eddy simulation is presently a powerful tool for investigating massively separated industrial ows, where the large scales and the turbulence
production is not driven by ne details of the dynamics of the boundary
442
layers. For other ows, such as fully attached ows, large-eddy simulation
is still too expensive to be used on full-scale congurations for realistic
Reynolds numbers because of the huge number of grid points required to
obtain an accurate resolution of the boundary layers. Existing wall models
have not been validated in realistic cases up to now, and the question of
their validity for such ows remains an open question.
14.5.2 Subgrid Model Eciency
Here we will try to draw some conclusions concerning the eciency of the
subgrid models for processing a few generic ows. These conclusions should be
taken with caution. As we have seen all through this book, very many factors
(numerical method, grid, subgrid model, and others) are involved and are
almost indissociable, so it is very dicult to try to isolate the eect of a model
in a simulation. The conclusions presented are statistical in the sense that
they are the fruit of an analysis of simulations performed on similar (at least
geometrically) ow congurations with dierent methods. A deterministic
analysis could lead to contradictory conclusions, depending on the other.
Also, there is no question of ranking the models, as the available information
is too lacunary to draw up a reliable list. Lastly, very many factors like the
discretization of the subgrid models still remain to be studied.
We may, however, sketch out the following conclusions.
1. To simulate a homogeneous ow:
a) All subgrid models including a subgrid viscosity yield similar results.
The eciency of functional models for the forward energy cascade
in isotropic turbulence, despite their lack of accuracy in representing
the subgrid tensor eigenvectors, was analyzed by Jimenez [351]. It is
explained by the existence of a feedback between the resolved scales
and the net energy drain provided by the subgrid model. Errors in
subgrid models are localized at the highest resolved frequency, and
do not contaminate the low frequency which is responsible for the
turbulence production. The error accumulation at high frequencies
leads to an adjustment of the subgrid model, which is expressed as
a function of the resolved scales. This is easily seen by writing the
induced subgrid dissipation: = ij S ij = 2sgs |S|2 . A classical
example is the Smagorinsky model, which leads to = CS |S|3 . A local
underestimation of CS will result in an energy accumulation at the
cuto, leading to an increase of the resolved shear |S| and an increase
of the net drain of kinetic energy. This drain mostly aects the highest
resolved frequencies, leading to a decrease of |S|. The global eect
on the ow is small, because the highest resolved frequencies contain
a small part of the total resolved kinetic energy. The eciency of
this adjustment depends of course on the way the subgrid viscosity
is computed.
14.5 Lessons
443
For anisotropic ows the loss of eciency of the basic subgrid viscosity models is a well recognized fact. This is explained by the fact
that the shear magnitude |S| is not governed by the highest resolved
frequency any more, but by very large scales. Thus, the dynamic feedback loop described above is not sucient any more to yield good results. Self-adaptive models (dynamic, ltered, selective, etc.) must be
used to ensure the quality of the results. The available experimental
data suggest that the correlation obtained using the subgrid-viscosity
model is of the order of 20%. This implies that it is impossible to get
both the right energy spectrum and the right stresses from a subgridviscosity model if its eect on the resolved eld is not negligible.
Grid renement is known to yield improved results for two reasons:
the error is commited on a smaller fraction of the total kinetic energy, and the Kolmogorov theory predicts that the small scales are
more isotropic than the large ones, rendering the structural error
less important. For a weak shear S, the Kolmogorov theory leads to
the following scaling laws for the normal-stresses spectrum E11 and
o-diagonal stress cospectrum E12 :
E11 2/3 k 5/3 ,
and
12 (LS /)4/3 ,
E12 S1/3 k 7/3
12 /11 (LS /)2/3 ,

where the shear length LS = /S 3 is proportional to the integral
scale L in equilibrium ows. These scaling laws show that rening
the resolution, i.e. decreasing , makes it possible to reduce quickly
the anisotropy of the subgrid scales. Jimenez estimates that 1% error
is obtained for L / = 1020.
b) Scale similarity or soft-deconvolution-type models do not yield good
results if used alone. This is also true for all the other types of ows.
A possible explanation for the improvement observed when using
mixed models or explicit random models for the backward energy
cascade is that these models weaken the spurious correlation between
the resolved strain rate tensor and the modeled subgrid tensor. This
improvement is also expected from theoretical results dealing with
the full deconvolution problem or the rapid/slow decomposition introduced by Shao et al. It is important to note that on very coarse
grids mixed models may not result in improved results.
c) These results hold locally for all other ows.
2. To simulate a free shear ow (mixing layer, jet, wake):
a) Subgrid viscosity models based on large scales can delay the transition. This problem can be remedied by using a dynamic procedure,
a selection function, or an accentuation technique. The other subgrid
viscosity models seem to allow the transition without any harmful effects.
444
3.
4.
5.
6.
b) Using a mixed structural/functional model improves the results obtained with a subgrid viscosity model based on the large scales.
To simulate a boundary layer or plane channel ow:
a) Subgrid viscosity models based on the resolved scales may inhibit
the driving mechanisms and relaminarize the ow. As before, this
problem is resolved by using a dynamic procedure, selection function,
or accentuation technique. The other subgrid viscosity models do not
seem to exhibit this defect.
b) Using a mixed functional/structural model can improve the results
by better taking the driving mechanisms into account.
c) Using a model for the backward cascade can also improve the results.
For separated ows (backward facing step, for example), use a model that
can yield good data on a free shear ow (to capture the dynamics of the
recirculating area) and on a boundary layer (to represent the dynamics
after the reattachment point).
For transitional ows:
a) Subgrid viscosity models based on the gradients of the resolved scales
generally yield poor results because they are too dissipative and damp
the phenomena. This problem can be remedied by using a dynamic
procedure, a selection function, or the accentuation technique.
b) Anisotropic tensorial models can inhibit the growth of certain threedimensional modes and lead to unexpected scenarios of transition to
turbulence.
For fully developed turbulent ows, the problems with subgrid viscosity
models based on the large scales are less pronounced than in the previous cases. Because these ows have a marked dissipative character, they
produce results that are sometimes better than the other models because
they ensure numerical stability properties in the simulation.
14.5.3 Wall Model Eciency

Numerical experiments show that wall stress models based on a linear relationship between wall stresses and the instantaneous velocity component
yield satisfactory results for well-resolved large-eddy simulation of attached
ows.
For separated ows, this class of wall models is inadequate, leading to
a poor prediction of the skin friction inside the separation zone.
For very coarse grids, i.e. when the cuto is not located inside the inertial
range of the spectrum, large errors are observed and the mean velocity prole
is not recovered [363, 806, 89]. Several reasons for this can be identied:
On coarse grids, the numerical error is large and pollutes the solution,
yielding erroneous input for the wall model.
On coarse grids, most subgrid models induce large errors on the resolved
scales. This is especially true for all subgrid viscosity models, which are
14.5 Lessons
445
not able to account for the strong ow anisotropy in the near-wall region.
It has been shown [363, 806] that this error is mainly due to the fact that
the subgrid acceleration term in the ltered momentum equations is not
properly predicted on very coarse grids.2 This bad prediction can lead to
the occurrence of spurious coupling with linear wall models, yielding very
bad results. The use of dynamic models may result in worse results if the
test lter is applied outside the inertial range [89].
A very important conclusion of studies dealing with suboptimal-based
wall models is that predicting the mean velocity prole and recovering the
best possible rms velocity prole seem to be competing objectives.
Wall models based on auxiliary simulations performed on secondary embedded grids (RANS, thin boundary layer equations) may be an alternative,
which remains to be assessed in critical situations. The use of more complex subgrid models, at least in the resolved near-wall region, which are able
to predict the subgrid acceleration is also expected to improve the response
of wall stress models. Multilevel subgrid models have demonstrated a clear
superiority in academic test cases.
14.5.4 Mesh Generation for Building Blocks Flows
We discuss here the basic rules for mesh generation for two classical building
blocks of complex ows, namely the attached equilibrium boundary layer and
the plane mixing layer.
The two basic rules are:
The key idea is that the driving mechanisms, i.e. the events responsible
for turbulence production and mean prole instabilities, must be correctly
captured by the simulation to recover reliable results.
The size of the computational domain must be larger than the correlation
length of the uctuations, in each direction of space. Too small a domain
size will yield spurious coupling between the dynamics of the ow and the
boundary conditions.
Exact values of mesh size, number of grid points and their repartition,
and domain size for plug and play simulations are not available: these parameters depend on many parameters, including numerical methods. What
are presented below are the commonly accepted ideas underlying the design
of a large number of published works.
Equilibrium Boundary Layer. As discussed in Sect. 10.2, the boundary
layer exhibits two dierent scalings.
In the inner layer the viscous length l is relevant to describe the dynamics. The typical correlation lenghts are 1000 wall units in the streamwise
2
It is recalled that subgrid viscosity models are designed to yield the correct
amount of dissipation, not the proper subgrid acceleration.
446
Fig. 14.28. Streaks in the near-wall region of a plane channel ow (Re = 180).
Left: well-resolved large-eddy simulation. Right: Coarse-grid large-eddy simulation.
Courtesy of E. Montreuil, ONERA.
(x) direction, and 100 wall units in the spanwise direction (y). Based on
a study of the anisotropy of the ow, Bagget et al. [32] derived the following
general criterion: /L 0.1, where the local inertial length L = k 3/2 / is
associated with the peak of the turbulent energy spectrum (with k the turbulent kinetic energy and the dissipation rate). Typical criteria dealing
with the mesh size for a wall-resolving simulation are:
x+ 50,
y + 15 .
(14.1)
The rst grid point must be located at one wall unit from the wall, with
three grid points in the viscous region 1 z + 10. Thirty to fty grid
points across the boundary layer are generally enough to get acceptable
results. Grids which are too coarse to allow a good resolution of the nearwall events usually yield to the occurrence of an overshoot in the streamwise
rms velocity prole. This peak is associated with fat streaky structures (see
Fig. 14.28). In order to prevent spurious correlations the minimum size of
the computational domain in the streamwise direction should be larger
than 2000 wall units, with a spanwise extent of 400 wall units.
In the outer layer, the visual boundary layer thickness 99 is the relevant
scale. The large scales have a correlation length which scales with 99 , and
are advected in the streamwise direction at a speed roughly equal to 0.8
U . Consequently, a minimum domain size of 3 99 to 5 99 is required in
the streamwise direction, with a spanwise extent of 2 99 to 3 99 .
Plane Mixing Layer. The choice of the extent of the computational domain
and the mesh size for the plane mixing layer is often guided by results from
the linear instability theory. For a hyperbolic tangent type mean velocity
prole, it is known that the two-dimensional linearly most unstable mode
has a streamwise wavelength x nearly equal to 7 , where the vorticity
14.5 Lessons
447
thickness is dened as = U/max(dU /dz), with U the velocity jump

across the shear layer. The associated spanwise length scale is y 2x /3.
Classical analysis of numerical errors shows that at least 5 to 20 grid
points per wavelength are required to get a reliable description of physical
phenomena. Thus, in the streamwise direction, large scales associated with
KelvinHelmholtz instability should be captured with x /2. Square-like
meshes in the (x, y) plane are recommended, yielding x = y. Numerical
experiments show that a minimum of 20 grid points is required across the
shear layer, yielding z /20.
Fig. 14.29. Large-eddy simulation of a time-developing plane mixing layer. Instantaneous view of the coherent structures at three dierent times. Left: early
transition stage. Right: advanced transition stage. Bottom: fully developed ow.
Courtesy of M. Terracol, ONERA.
448
The size of the domain is dictated by the purpose of the simulation.

For unforced mixing layers, the turbulent self-similar state is reached after
the second pairing. For temporal simulations, this indicates that the size of
the domain in the streamwise direction must be greater than or equal to
8x = 56 . The need for a third pairing originates from the fact that,
after the last pairing permitted by the size of the computational domain,
uctuations are correlated over the domain, leading to corrupted results. The
self-similar state can be observed between the second and the last pairing.
Typical results for a time-developing plane mixing layer are displayed in
Fig. 14.29.
15. Coupling with Passive/Active Scalar
15.1 Scope of this Chapter

This chapter is devoted to the extension of the previous results dealing with
Large-Eddy Simulation of incompressible ows to the case where a scalar
quantity is added to the velocity and the pressure to describe the physical system. Depending on the application, the scalar can be related to the
temperature (or temperature increment with respect to a given value), the
density, the concentration of a pollutant, ...
The chapter details two dierent cases:
1. The case of the passive scalar (Sect. 15.2), where there is no feedback of
the scalar equation in the momentum equation. In this regime, which represents a one-way coupling between the velocity eld and the scalar eld,
the scalar dynamics is enslaved to the velocity eld, while the dynamics of
the later is not aected. The new closure problem is therefore restricted
to the scalar equation, the treatment of the momentum equation being
exactly the same as in the previous chapters.
2. The case of the active scalar (Sect. 15.3), which corresponds to physical
systems in which a two-way coupling exist between the velocity and the
scalar. The emphasis will be put on simple (destabilizing) buoyancy and
stabilizing stratication eects. In the rst case the turbulent production
is enhanced by the coupling with the scalar dynamics, while in the second case a new turbulence damping mechanism is involved. These two
simple cases are used to illustrate the new problems dealing with subgrid
modeling induced by the active scalar model: the ltered scalar equation
need to be closed, but the subgrid closure in the momentum equation
must a priori be changed to account for the fact that the dynamics is
now more complex than the simple kinetic energy cascade process.
For the sake of simplicity, the chapter is not made exhaustive: in many
cases, strategies developed to close the scalar equation are nothing but direct
extrapolations of models/methods developed to close the mometum equations. In these cases, the models for the scalar equations will not be extensively described, and the way the original model is extended is indicated.
A detailed description will be given only when the closure strategy or the
model is not a straightforward extension of a former proposal. It is worth
450
noticing here that many models developed for the momentum equation have
not yet been extended to the scalar problem, despite it may be done very
easily.
This chapter is restricted to presentation of the classical subgrid closure
problem, and the issues dealing with the development of wall models and
turbulent inlet conditions for the scalar (and the velocity eld in the active
scalar case) will not be discussed.
15.2 The Passive Scalar Case

15.2.1 Physical Model
Denitions and Filtered Equations in Physical Space. The following
passive scalar equation is used as a starting point:
+ (u) = 2
t
(15.1)
where is the scalar quantity (temperature, pollutant concentration, ...) and

the associated molecular diusivity. Since there is no feedback in the momentum equation, the later will not be considered here. Applying a lter1
to (15.1), one obtains
+ (u) = 2
t
(15.2)
Splitting the ltered non-linear term u into a resolved and a subgrid

part, one obtains
+ (u ) = 2
t
(15.3)
where the subgrid scalar ux is dened as

(u u ) .
(15.4)
The subgrid scalar ux can be further decomposed in exactly the same

way as the subgrid tensor in the momentum equation (see Sect. 3.3), yielding
= u
u + u

+ u +

u

L
(15.5)
As for the momentum equation, the mathematical model used here to represent
the true Large-Eddy Simulation problem is the convolution lter paradigm. But
other mathematical models presented in Chap. 4 can also be used to this end.
451
The subgrid uxes L , C and R are analogous to the Leonard stress

tensor, the Cross stress tensor and the subgrid Reynolds stress tensor, respectively. The subgrid scalar ux can also be decomposed using Germanos
consistent decomposition approach (Sect. 3.3.3):
= G (u, ) = G (u, ) + G (u, ) + G (u , ) + G (u , ) ,
(15.6)
where G (, ) is the generalized central moment of and

associated to the lter kernel G.
Two quantities of interest to characterize the scalar dynamics are the
subgrid scalar ux (which is equal to u if the lter belongs to the class
2
of the Reynolds operators) and the scalar subgrid variance sgs
. The
former is related to the mixing/stirring transport process of the scalar eld
at small scales, while the latter is tied to the existence of uctuations of the
scalar eld at the subgrid level and therefore is a measure of its unmixedness.
It is worth noting that these two quantities are generalized within the LargeEddy Simulation framework when arbitrary convolution lters are used in
the same way that the subgrid kinetic energy ui ui /2 is extended considering
the trace of the subgrid tensor. In the scalar case, u and the subgrid scalar
variance are extended as G(u, ) and G (, ), respectively.
Transport equations for these two quantities are easily derived using the
scalar equation and the momentum equation. The most general expressions
based on the generalized central moments are
G (ui , )
G (ui , )
+ uk
t
xk
ui
= G (uk , ui )
G (uk , )
xk
xk

I
II

ui
+
, + G
, ui
G
xk
xk
xk

III

ui
p
( + )G
,
G ,
xk xk
xi

IV
G (uk , ui , )
,
xk

VI
for the generalized subgrid scalar ux and

G (, )
G (, )
+ uk
= 2G (uk , )
t
xk
xk

V II
(15.7)
452
2 G (, )
+
2G
xk xk

V III

,
xk xk

IX
G (uk , , )
,
xk

(15.8)
for the generalized subgrid scalar variance, where it is recalled that

G (a, b, c) = abc aG (b, c) bG (a, c) cG (a, b) a b c
The physical meaning of the terms appearing in the preceding equations

are
I: Production by interaction between the subgrid stresses and the resolved
scalar gradient
II: Production by interaction between the subgrid scalar uxes and the
resolved velocity gradient
III: Viscous diusion
IV : Viscous dissipation
V : Scalar-pressure subgrid ux
V I: Diusion by subgrid motion
V II: Subgrid scalar variance production by interaction with the resolved
scalar gradient
V III: Viscous diusion
IX: Subgrid scalar variance diusion, referred to as
X: Diusion by subgrid motion.
Denitions and Filtered Equations in Spectral Space. Physical quantities and corresponding equations can be rewritten in the spectral space per
forming a Fourier transform. Denoting (k)
the Fourier transform of (x),
one has the following relation dealing with the two-point correlation
6
;
<

1

)(k)
=
e(k x+kk ) (x)(x ) dxdx
(k
2
E (k)
=
(k + k ) ,
(15.9)
2k 2
where E (k) is the scalar spectrum (dened here as an average over shell k =
cste.). The scalar variance is equal to
+
1
(x)(x) =
E (k)dk .
(15.10)
2
0
The spectral analogue of (15.1) is

2
+ k (k) = T (k)
t
(15.11)
in which the non-linear term T (k) is equal to

3

T (k) = kj
p
u
j (k p)(p)d
453
(15.12)
The corresponding evolution equation for the non-stationary spectrum of

scalar variance is

2
+ 2k E (k) = T (k) ,
(15.13)
t
where the scalar spectrum transfer is expressed as

;
<
p)

um (p)(k)
d3 p
T (k) = 8k 2 km (k
(15.14)
The conservation property for the scalar variance takes the following form:

T (k)d3 k = 0 .
(15.15)

(k),
The ltered equation associated to the ltered variable (k)
G(k)
where G(k) denotes the transfer function of the selected lter, is

+ k 2 (k)
(k)
(15.16)
= G(k)T
t
=
Tr (k) + Tsgs
(k)
(15.17)
(k) are the resolved and subgrid spectral scalar uxes,

where Tr (k) and Tsgs
only, while the
respectively. The former involves resolved quantities and u
latter contains all terms involving at least one subgrid component among

(1 G(k))
(k)
and (1 G(k))
u(k). The closure problem in the spectral
space consists in nding an expression for Tsgs

(k) which involves only known
quantities.
15.2.2 Dynamics of the Passive Scalar

This section is devoted to a brief survey of the dynamics of the passive scalar
in isotropic turbulence. The main purpose here is to enlight the fact that
the modeling task is far from being a trivial one, even within this simplied
framework. The very reason for this is that several physical regimes exist,
which are associated to dierent values of the molecular Prandtl number (or
Schmidt number, or Peclet number depending on the physical signicance of
the scalar eld) dened as
.
(15.18)
Pr
454
The spectral properties of the three ideal regimes P r 1, P r 1 and

P r 1 are surveyed in Sect. 15.2.2. The interested reader can refer to specialized books [714, 439, 464, 708] for a detailed discussion. Results obtained via
Direct Numerical Simulation and EDQNM analysis dealing with the spectral
dynamics of the passive scalar are displayed rst (see p. 456). The concept of
subgrid Prandtl number is introduced and discussed in a second step (p. 459).
Dierent Regimes and Associated Dynamics. Three regimes for the
passive scalar in isotropic turbulence are identied, each one being associated with a range of values for the Prandtl number. The existence of these
three archetypal2 regimes originates in the dierence between the viscous
cuto scales for the scalar and the velocity eld. Each regime is associated to
a specic scalar spectrum shape (the kinetic energy spectrum does not vary,
since the scalar is strictly passive).
To analyze the spectral characteristic features of these dynamical regimes,
we introduce the scalar diusion cuto length, referred to as the Obukhov
Corrsin scale, which is the analogue of the Kolmogorov scale K for the
velocity:
1/4 1 3/4
=
K .
(15.19)
=
3
Pr
The ratio of the wave numbers associated to these cutos is therefore
estimated as
k
= P r3/4 .
(15.20)
k
The denition of ObukhovCorrsin cuto scale is not valid if P r 1,
since it is based on Kolmogorov-type hypotheses on the nature of the uctuations which are no longer adequate. In this case, Batchelor derived the
following diusive cuto wave number (referred to as the Batchelor wave
number):
1/4
.
(15.21)
kB =
2
These three regimes are
1. P r 1 : the molecular diusivity is much larger than the molecular
viscosity. Looking at relations (15.19) and (15.20), it is seen that k is
much smaller than k , meaning that the the scalar diusive cuto occurs
within the inertial range of the Kolmogorov spectrum for the kinetic
energy. One can therefore infer that two inertial ranges will be observed
in the scalar spectrum (see Fig. 15.1):
a) The inertial-convective range, which corresponds to wave numbers
k k k . These scales are not subject to viscous and diusive
eects accordingly to Kolmogorovs picture. Scalar uctuations are
2
It is important noticing that the results presented here deal with asymptotic
cases, and that real-life ows are more complex.
455
Fig. 15.1. Schematic of the kinetic energy spectrum and the scalar variance spectrum in the very-low Prandtl number case.
driven by the stirring induced by velocity uctuations. The corresponding form of the scalar spectrum is
E (k) = 1/3 k 5/3
(15.22)
where ObukhovCorrsin constant is in the range 0.68 0.83.

b) The inertial-diusive range, which is observed for scales which belong
to the Kolmogorov intertial range, meaning that the velocity uctuations are not directly sensitive to the molecular viscosity, but at
which scalar uctuations experience a strong action of the molecular
diusivity. The corresponding wave number band is k k k .
The dynamics of the scalar at these scales is governed by a balance
between the turbulent advection and the molecular diusion. The
scalar spectrum shape is
E (k) =
K0
2/3 3 k 17/3
3
(15.23)
Smaller scales (k kc ) are governed by molecular viscosity and diusivity

eects, and exhibit an exponentially decaying behavior.
2. P r 1. The two cuto scales are very close, leading to the existence of
a unique inertial-convective inertial range. The scalar spectrum shape is
the same as in the previous case (15.22).
3. P r 1. This last case corresponds to congurations in which the diffusive cuto scale is much smaller than the viscous cuto scale for the
velocity uctuations. Two inertial ranges exist (see Fig. 15.2):
456
Fig. 15.2. Schematic of the kinetic energy spectrum and the scalar variance spectrum in the very-high Prandtl number case.
a) An inertial-convective range, for wave numbers such that k k

kB , which is similar to those already described above.
b) A viscous-convective range, which is associated to scales (k k
kB ) where the velocity uctuations are severely damped by viscous
eects but scalar uctuations are not aected by molecular diusion.
The associated spectrum is

1
E (k) =
k 1 .
(15.24)
2
This brief survey obviously shows that the spectral scalar transfers across
the large-eddy simulation cuto may strongly depend on the type of inertial
range it is located in. As a consequence, the development of a closure based
on the functional modeling approach may be much more dicult than in the
previous case of the momentum equations, since the latter exhibits only one
type of inertial range dynamics.
Interscale Scalar Transfers. We now present some results dealing with
the analyis of passive scalar uxes in the spectral space. These results are
the corner stone of the functional modeling approach. The emphasis is put
on theoretical results coming from the EDQNM analysis of the passive scalar
dynamics in isotropic turbulence (see B.3 for a short presentation).
The main results of the EDQNM analysis deal with the non-local interactions in the scalar variance equation. Considering a non-local triad (k, p, q)
satisfying the constraint k + p + q = 0, two cases are identied
The considered scale, associated with k, is interacting with one much larger
scale : p k q. The EDQNM form of the associated non-local scalar
457
variance ux is
T+ (k)
=
+

ak
2
2
2 E
(k)
kkq q E(q)dq 2kE (k) k
15 0
k

1 ak 3
kkq q E (q)dqE(k)
4 0

1 ak 5 E(k)E (k)
kkq q dq
.
(15.25)
4 0
k2
The parameter a denes the local/non-local triads (see p. 93). The rst
term in the right hand side corresponds to interactions which are responsible for the existence of the viscous-convective range.
The scale k is involved in a non-local triadic interaction with much smaller
scales: k p q. The asymptotic expression for the scalar variance ux
is
+

4 k 2
T (k) =
q E (q)dq
qpp
E(p)dp
3 0
sup(k,q/a)
+

4 k 4
E(p)E (p)
+
q dq
qpp
dp . (15.26)
3 0
p2
sup(k,q/a)
A simplied expression, which is valid at low wave numbers, is
+
4 2
T (k) = k E (k)
0pp
E(p)dp
3
ak

4 4 + E(p)E (p)
k
+
0pp
dp .
3
p2
ak
(15.27)
The rst term in the right hand side is associated with a forward scalar variance cascade, whose intensity is governed by both the kinetic energy and
scalar variance spectral distribution. The second term represents a backward scalar cascade.
The analysis of Direct Numerical Simulation results [783] reveals that,
in the inertial-convective range, the overall spectral transfer is forward cascading. It is dominated by the energy-containing scales in the velocity eld.
More precisely, the dominant physical process at high wave numbers belonging to this inertial range is a local transfer associated with a non-local
triadic interaction that links two high wave number scalar modes and one low
wave number velocity mode. The other non-local interactions (one high wave
number scalar mode, one low wave number scalar mode and one high wave
number velocity mode) and local interactions are much weaker. Therefore,
the suggested dominant physical process is the breaking of blobs of scalar
uctuations into smaller-scale fragments by large energetic eddies.
458
The large-eddy simulation closure problem is now directly addressed by

writing the equation associated to the resolved scalar variance density, de 2 (k)E (k):
noted E (k) = G

+ 2k E (k) = Tr (k) + Tsgs

(k) ,
(15.28)
t
where Tr (k) and Tsgs

(k) are the resolved and the subgrid spectral scalar
variance transfer, respectively. In a way similar to the one used to characterize
subgrid kinetic energy transfers in Sect. 5.1.2, the intensity of these transfers
across a cuto wave number kc can be parameterized dening an eective
spectral diusivity e (k|kc ) such that
Tsgs
(k) = 2e (k|kc )E (k) .
(15.29)
In the case the cuto is located within the inertial-conductive range, and
assuming that all the basic hypotheses of the canonical analysis (inertial
ranges extending to innity, sharp cuto lter, see Sect. 5.1.2), both the
EDQNM analysis conducted by Chollet and Zhous RNG theory [811] show
that the spectral eective diusivity shape is strictly similar to the one of the
eective spectral viscosity e (k|kc ):
for k kc , it is independent of the wave number. The value of the plateau
deduced from the non-local forward cascade term in (15.27) is

2 +
0pp E(p)dp
(15.30)
e (k|kc ) = e
3 kc
In practice, this plateau is observed for wave numbers in the range 0
k kc /3. Theoretical analysis carried out by Chollet and Lesieur using
the EDQNM analysis in the case of a very wide inertial-convective range
also shows that the value of the plateau depends on both the kinetic energy
at the cuto and the slope of the spectrum. A general expression is [514]:
'
4 3 m E(kc )
m<3
3a m + 1
kc
,
(15.31)
=
1 1
1 E(kc )
m>3
3 m 1 Dr kc
where m is the slope of the kinetic energy spectrum, a = 0.218K0 a struc!k
tural EDQNM parameter and Dr = 0 c k 2 E(k)dk a norm of the resolved
velocity gradient. Considering a Kolmogorov spectrum (m = 5/3, K0 = 1.4)
one obtains
e = 0.3.
Near the cuto, i.e. at wave numbers kc /3 k kc , the eective diusivity
exhibits a cusp, showing that the spectral transfers are more intense. In
this region, local interactions are not negligible. It is important noting
that this cusp is as sensitive as the one in the eective viscosity: it is
459
not observed with smooth lters, and can also disappear if the cuto is
located at the very beginning of the inertial range. In the canonical case
of an innite inertial-convective range, the maximum value of the subgrid
diusivity found using EDQNM analysis is e (kc |kc ) = 0.6.
Detailed investigations of the spectral transfers across a cuto located
within the viscous-convective or the inertial-diusive range are missing. This
lack in the theory may be not very important for most practical applications,
since putting the cuto within one of these inertial range requires the use of
very ne computational grids.
Numerical experiments carried out by several authors [717] also show that
the eective diusivity spectral shape is very sensitive to both the scalar variance spectrum shape and the kinetic energy spectrum: it has been observed
to be either an increasing or a decreasing function of the wave number. Eective viscosities computed from simulated isotropic turbulence by Metais and
Lesieur are displayed in Figs. 15.3 and 15.4.
The Subgrid Prandtl Number Paradigm. The molecular Prandtl number (15.18) is a useful tool to compare turbulent scales which characterize
dissipative/diusive cuto scales of the velocity and scalar elds. This is why
the idea of introducing a subgrid Prandtl number, P rsgs , is attractive. But
preceding remarks dealing with the sensitivity of the eective spectral diu-
Fig. 15.3. Eective

subgrid viscosity (Solid line) and subgrid diusivity (Dashed
line) normalized by E(kc )/kc in the LesieurMetais simulation (Case 1).
460
Fig. 15.4. Eective

subgrid viscosity (Solid line) and subgrid diusivity (Dashed
line) normalized by E(kc )/kc in the LesieurMetais simulation (Case 2).
sivity indicate that a universal distribution for the spectral eective subgrid
Prandtl number
e (k|kc )
e
,
(15.32)
(k|kc ) =
P rsgs
e (k|kc )
cannot exist. As a consequence, subgrid closure strategies based on the concept of subgrid Prandtl number necessarily introduce some errors, since they
do not account for subtle discrepancies that exist between the velocity and
the scalar dynamics.
In the very simple case of a quasi-innite inertial-convective range, the
value found for wave numbers located within the plateau (k 0.3kc ) via the
EDQNM analysis is [514]
e
P rsgs
(k|kc )
5m
20
(15.33)
But such a simple expression should not mask the fact that the subgrid
Prandtl number is fully case dependent, since it characterizes the dierences
that may exist between kinetic energy and scalar spectral transfers. As an
example, in the very simple case of passive scalar in isotropic turbulence,
Lesieur and Rogallo [441] observed two very dierent spectral subgrid Prandtl
distributions (see Fig 15.5). In the rst case, the scalar eld exhibits a k 1
461
Fig. 15.5. Eective Prandtl number versus the wavenumber in LesieurRogallo

simulations of isotropic turbulence.
inertial range (this anomalous exponent is supposed to be due to shearing by

large scales), and the subgrid Prandtl number approximately follows the law
e
P rsgs
(k|kc ) = 0.35 + 0.2 log(k/kc ) + 9e3.099kc /k
(15.34)
while in the second set of simulations the following distribution was observed
e
P rsgs
(k|kc ) = 0.45 + 0.25 log(k/kc ) .
(15.35)
15.2.3 Extensions of Functional Models

We now discuss the extension of functional models in the physical space. The
construction of functional models in the Fourier space is a straightforward
utilization of spectral subgrid difusivity laws given in the preceding section,
and will not be further detailed.
All explicit functional models for the passive scalar equations are based
on the subgrid diusivity paradigm, yielding the following general closure
relation:
(15.36)
= sgs .
As a consequence, the emphasis is put on the dominant mechanism observed in isotropic turbulence, namely the forward scalar cascade. This approach is strictly equivalent to the use of a subgrid viscosity to close the
462
ltered momentum equations. Almost all functional subgrid models and/or

closure strategies proposed for the momentum equations and presented in
Chaps. 5 and 6 can be very easily extended to the scalar equation. An exhaustive description of all possibilities is not of interest, and it is worth noting that many straightforward extensions have not been published yet3 . The
implicit Large-Eddy Simulation approach can also be used, if an adequate
numerical scheme for the scalar equation is utilized.
General Expression. Scalar Subgrid Diusivity. As in the case of subgrid viscosity, a fully general expression of the subgrid diusivity sgs as
a function of a set of selected basic quantities can be used to obtain a unied
view of most published models. A possible expression is
2
2
sgs = sgs (P r, , |S|, qsgs
, , ||, sgs
, ) ,
(15.37)
2
2
where quantities directly tied to subgrid quantities (i.e. qsgs
, , sgs
, ) can
be evaluated solving corresponding prognostic equations, or using some local
equilibrium and/or scale similarity assumptions to recover expressions involving resolved scales only. The Prandtl number is a priori included as an input
of this general expression, since it can be used to distinguish the dierent inertial range regimes. Weighting coecients in the constitutive equation (15.37)
are found performing a simple dimensional analysis.
An example is the model proposed by Schmidt and Schumann [651]:
+
2
,
(15.38)
sgs = C qsgs
where interial-conductive range considerations lead to

C =
2 1
3K0 3
(15.39)
Yoshizawa [789] proposes a more complex model:

sgs = C
2 2
)
(sgs
,
2
C = 0.446 .
(15.40)
A subgrid diusivity model which accounts for molecular Prandtl number

eects was developed by Gr
otzbach [277, 279] to describe the temperature
eld in liquid metals, which have very small Prandtl number (P r = 0.007
for liquid sodium under nuclear reactor conditions). We reproduce here the
method developed by Gr
otzbach to account for small Prandtl number eects
rather than its model, since it can be applied to any subgrid viscosity model
with one arbitrary constant. Let us consider a generic subgrid diusivity
3
The interested reader will quickly nd material for a few dozens of papers dealing
with new improved models!
463
model under the form

sgs = CK
(15.41)
where C is the constant to be adjusted and K a constant-free parameter

which has the dimension of a diusivity. The method proposed by Gr
otzbach
relies on the local equilibrium assumption (i.e. production = dissipation) for
the scalar variance:
= ,
(15.42)
where is the full turbulent scalar variance dissipation. Now introducing the
model in the left hand side of this relation, one obtains the following relation
for the constant C:

C=
.
(15.43)
K
The last parameter to be evaluated is . Assuming that the cuto occurs
in a very wide inertial-convective range, an analytic expression for the scalar
variance dissipation as a function of the kinetic energy dissipation can be
found, yielding

1
4/3
1 1/3
,
(15.44)
C
K0
where the subgrid kinetic energy dissipation rate is retrieved from the work
done to close the momentum equation, leading to a fully determined denition
of the subgrid diusivity.
The Subgrid Prandtl Number Approach. A very common closure approach is to use a subgrid Prandtl number, leading to
sgs =
sgs
P rsgs
(15.45)
where the subgrid viscosity sgs can be evaluated using any model described in
the preceding chapters devoted to the momentum equation closure. Despite it
is awed from a purely theoretical point of view, this approach is very often
used in simulations in the physical space and the subgrid Prandtl number
appears as an adjustable parameter which can be tuned in an ad hoc way to
obtain the best t with the reference data. Values found in the literature range
from 0.1 to 1, the most common one being 0.6. It is important noting that
this approach is not valid in cases where the velocity eld if fully resolved,
yielding sgs = 0, while some subgrid scalar uctuations exist. A typical
example would be to put the resolution cuto within the viscous-convective
range.
The subgrid Prandtl number can be made more accurate, but still not
adequate to represent the cases mentioned above, rendering it self-adaptive,
meaning that the value of the Prandtl number will be made space and time
dependent. Such a modication is expected to make the subgrid Prandtl
number based models adequate to treat the case where all scalar uctuations are resolved while subgrid kinetic energy transfers exist. A common
464
example is to use the Germano identity to dene a dynamic procedure, as

proposed by Moin et al. [538]. The Germano identity for the scalar equation
is

2 u
2 u
**
**
u
= u
u/
u ,
(15.46)

L
where the tilde symbol refers to the test lter level. Vectors T and are the
subgrid scalar uxes at the grid and test lter levels, respectively. As in the
analogous relation for the momentum equation, the left hand side of (15.46)
can be directly computed, while replacing subgrid uxes in the right hand
side by the corresponding subgrid models makes it possible the nd the best
value of the model constant in the least-square sense.
* the subgrid viscosity values computed at the
Denoting sgs () and sgs ()
grid and test lter levels, respectively, the residual associated to the Germano
identity (15.46) is equal to
E = L +
*
()
sgs ()
sgs/
*
P rsgs
P rsgs
(15.47)
Assuming that the subgrid Prandtl number is the same at the two levels
* the
and that it does not vay signicantly over distances of the order of ,
least-square minimization of E leads to
m m
P rsgs =
,
(15.48)
m L
where
* *
m sgs ()
sgs/
()
(15.49)
Other expressions can be derived. As an example, Moin et al. [538] do not

use the least-square minimization but chose to nd the value of P rsgs which
corresponds to a zero of E *
:
P rsgs =
m *
*
L
(15.50)
Wong and Lilly [766] dene a dynamic procedure based on dimensional

parameters, yielding
P rsgs =
* (L *
)
(L : S)
*
*
2
|S|
*
(15.51)
All these dynamic subgrid Prandtl number denitions suer some numerical stability problems. The methods used to cure this problem are the same
as for the momentum equation (see Sect. 5.3.3 ): clipping, time/space averaging, ... All variants of the dynamic procedure described in preceding chapters
can be applied to compute the dynamic subgrid Prandtl number.
465
Anisotropic Subgrid Diusivity. All models mentioned above are based

on inertial range considerations, which are characteristic features of homogeneous isotropic turbulence. Practical applications involve much more complex
congurations, in which the scalar eld can be non homogeneous while the
velocity eld remains isotropic, or even more complex ows were both the
velocity and the scalar elds are not homogeneous. Therefore, the question
arises of accounting for the velocity eld anisotropy in the scalar diusion.
This task is expected to be much more complex than the equivalent one for
the velocity eld, since the scalar variance spectrum exhibits much less universal features than the kinetic energy spectrum. Experiments conducted by
Kang et al. [376] also show that the trend of decreasing anisotropy at small
scales observed on velocity uctuations is not recovered on scalar uctuations.
It is also known that, in the presence of a mean scalar gradient, structure
functions and the derivative skewness of the scalar eld do not follow predictions from isotropy at inertial and even dissipative scales. This local isotropy
breakdown is tied to the direct action of large velocity structures on small
scalar scales (this is consistent with the EDQNM analyis results dealing with
dominant triadic interactions).
Taking anisotropy into account can be an important issue when the cuto
is such that the subgrid scalar uctuations are governed by velocity uctuations with high local Reynolds number (the localness is understood here in
terms of wave number). In such cases, anisotropy in the velocity eld will have
an eect on the scalar diusion process, and the most simple idea consists in
dening a tensorial subgrid diusivity.
Yoshizawa [789] proposes the following anisotropic diusivity, which is
derived using a two-scale expansion:

(sgs )ij = C
2
sgs
3
S ij
(15.52)
where C = 0.366. The subgrid scalar dissipation rate and the subgrid scalar
variance remain to be evaluated. Models for these quantities are presented in
Sect. 15.2.6. The use of the resolved strain tensor S is coherent with the simple
physical picture that scalar blobs are stretched by the velocity uctuations,
and will therefore be elongated in the principal shear direction, leading to
the development of smaller scalar scales in transverse directions. A similar
model is proposed by Peng and Davidson [587]. Rotational eects can also
be included by taking into account the skewsymmetric part of the resolved
velocity gradient, ij :

(sgs )ij = 2
with C1 = 0.29 and C2 = 0.08.
2
sgs
3
3
4
C1 S ij + C2 ij
(15.53)
466
Another tensorial subgrid diusivity model was proposed by Pullin [608],

who extended the stretched-vortex approach (see Sect. 7.6) to the passive
scalar case. In the extended model, it is assumed that the subgrid mixing
is restricted to the plane normal to the vortex axis. A secondary property
of this model is that the subgrid scalar gradient induced by the transport
is orthogonal to the subgrid vorticity, the latter being represented by the
stretched vortex. This is consistent with the observations that the probability
density of the alignement between the vorticity and the scalar gradient is
maximum when these vectors are orthogonal. The corresponding formulation
is

+ 2
,
(15.54)
qsgs ij evi evj
(sgs )ij =
2kc
where the vector ev is the same as in Sect. 7.6. Assuming that the cuto
occurs within an innite inertial-convective range, the structural parameter
is given by

2
2 1
.
(15.55)
=
3K0
15.2.4 Extensions of Structural Models
Many structural models have been applied to the subgrid scalar ux vector.
The most illustrative ones are presented below.
Approximate Deconvolution and Scale Similarity Models. Scale similarity, or, in an equivalent way, soft approximate deconvolution models are
dened writing the following approximation:
u u u u
(15.56)
where the approximate deltered elds are expressed as

u = G1
l
u,
= G1
l
(15.57)
an approximate inverse of the lter (see Sect. 7.2.1 for a detailed

with G1
l
presentation). A large number of structural models can be generated using
this simple form, as for the momentum equation: models based on iterative
deconvolution, models based on Taylor series expansions and tensor diusivity models. Because they are all approximations of the exact soft deconvolution solution, these models do not take into account scalar transfer towards
subgrid scales and they must be supplemented with a dissipative term, which
can be either a numerical regularization or a functional subgrid model. That
leads to the denition of linear combination models for the subgrid scalar
uxes.
The Bardina-type model obtained using the lowest-order deconvolution is
= u u
(15.58)
467
The associated gradient-type model is (for a Gaussian or Box lter)

2
2
(u) .
24
(15.59)
Dierential Scalar Fluxes Model. Another solution consists in solving

a dierential equation for each component of the subgrid scalar ux vector.
This approach requires to close (15.7), in which terms V (pressure-scalar
correlations) and V I (diusion by subgrid uctuations) are not directly computable.
Deardor [173] proposes to close term V as
+

2
qsgs
p
(15.60)
= C1
G (ui , ) ,
G ,
xi
and to express the third-order term as

+
2
G (uk , ui , ) = C2 qsgs
G (uk , ) +
G (ui , )
xi
xk
(15.61)
with C1 = 4.8 and C2 = 0.2.

Other closures are derived by Sheikhi et al. [671], who use the velocityscalar ltered density function approach. This method is strictly equivalent
to the one developed by Gicquel et al. for the momentum equation (see
Sect. 7.5.3): a stochastic eld whose probability density function is the solution of the evolution equation of the ltered density probability function
is computed via a Lagrangian-Monte-Carlo method. In the present case, the
method is utilized to reconstruct both u and . These reconstructed stochastic variables can be used to close the ltered momentum and scalar equations
directly, or to close the equations for the subgrid scalar uxes and the subgrid scalar variance. The latter approach is considered here, since the former
belongs to the family of the methods based on a direct evaluation of the subgrid scales (to be discussed below). This approach yields the following closure
relation for the sum of the pressure-scalar correlations and the dissipation:

ui
3 3
p
+ C0 G (ui , ) , (15.62)
,
2G
+ G ,
=
xk xk
xi
2 4
where C0 = 2.1 and the time scale is equal to
+
2
qsgs
=
.
(15.63)
The triple correlation term is directly computed using the stochastic subgrid eld.
468
Explicit Evaluation of Subgrid Scales, Multilevel Simulations and

Others. The technique consisting in regenerating the subgrid scales of the
scalar eld on a ner mesh using a low-cost stochastic model, a simplied
determinitic model or multilevel simulation has also been extended to the
passive scalar case. These extensions being very reminiscent of their counterparts for the momentum equation, they will not be detailed here.
15.2.5 Generalized Subgrid Modeling for Arbitrary Non-linear
Functions of an Advected Scalar
The preceding sections addressed the problem of nding subgrid closures
for the convection term, the other ones being assumed to be linear. But in
many ows, additional physical processes associated to scalar source/sink are
present, which appear as non-linear functions of the scalar eld . Common
examples are chemical reactions, uids with temperature-dependent molecular viscosity/diusivity, or coupling of the scalar eld (density, temperature,
concentration) with micro-physics such as radiative transfer. Let us consider
a general form for this additional source term:
+ (u) = 2 +
t
f ()

(15.64)
source term
The corresponding source term in the ltered scalar equation is f (),

which is decomposed as

f () = f () + f () f ()
.
(15.65)

computable
f :subgrid term
The new closure problem consists in nding a surrogate for f . Since we

are addressing a fully general problem which can cover a very wide range
a physical mechanisms, the denition of a general functional model is hopeless. Specic functional models can be found for individual process, as it is
done for the convection term.
The most general approach being the structural modeling, Pantano and
Sarkar [578] propose to use an approximate deconvolution approach to model
general subgrid source terms (in practice, they applied this strategy to
a chemical reaction term) . The key idea developed by these authors is to approximate by a synthetic eld which is optimized so that it will minimize
an error functional, leading to
f () f () = f ( ) f ( ) .
(15.66)
Since physical processes may be very sensitive to small errors and are associated to high-order non-linearities in terms of (e.g. the radiative transfer
469
ux behaves as 4 , where is the temperature), it is important to enforce

some accuracy requirements when dening .
The simple Taylor series expansion
f () f () =
1
2
f ()(2 ) + ... ,
2
(15.67)
illustrates the fact that a good control of the error committed on the subgrid
scalar variance is necessary to ensure a satisfactory representation of the
subgrid source/sink term. As a consequence, Pantano and Sarkar propose to
modify the usual soft deconvolution procedure, leading to
= + C0 ( ) + C1 ( 2 + ) + ... ,
(15.68)
were the coecients Ci are allowed to vary, instead of being xed to be

unity as in the original approximate deconvolution procedure based on the
Van Cittert iterative method (7.8). These coecients are chosen in order to
make the statistical mean ltered moments that appear in the Taylor series expansion of the modeled eld equal to their counterparts dened using
the exact eld. In practice, it is aimed to enforce the following global constraints

k
k
k dx =
k dx, k = 2, ..., N ,
(15.69)
where N is an arbitrary parameter and is the uid domain under consideration. Practical subgrid models are derived truncating expansion (15.68) at
an arbitrary order and inserting it in (15.69). To get a closed set of non-linear
polynomial equations for the coecients Ci , it is also required to select an
analytic expression of the scalar spectrum and the lter transfer function.
15.2.6 Models for Subgrid Scalar Variance and Scalar Subgrid
Mixing Rate
This section is devoted to models aiming at evaluating the two following subgrid quantities, which characterize the dynamics of the subgrid uctuations
of the scalar eld :
2
The subgrid scalar variance, sgs
, also referred to as the scalar unmixedness,
since it measures the degree of local non homogeneity of within the
volume of characteristic diameter : a uniform distribution is associated
with a zero subgrid variance.
The scalar variance dissipation rate, , also referred to as the scalar variance destruction rate, which is related to the stirring and mixing process
at subgrid scales. A high destruction rate means that diusion is quickly
homogenezing the scalar eld.
470
A large number of subgrid models for these two quantities have been
proposed, since they are important inputs of subgrid models for combustion
related terms. Since the specic problem of reactive ows is beyond the scope
of this chapter, this section illustrates the dierent modeling ways and does
not provide the reader with an exhaustive description.
Models for the Subgrid Scalar Variance. It is worth noting that the
2
is reminiscent of the one dealing with the evaluation
issue of predicting sgs
2
. Therefore, all methods and models dened
of the subgrid kinetic energy, qsgs
to compute the latter (see Sect. 9.2.3) can be modied to evaluate the former.
This work is straightforward and is left to the interested reader. Only general
approaches will be discussed in this section.
The subgrid scalar variance being dened as
2
2
sgs
(15.70)
a natural way to model it is to use a deconvolution-type approach (or its

optimized version proposed by Pantano and Sarkar for arbitrary non-linear
functions of , see Sect. 15.2.5). The usual deconvolution procedure yields
2
2
sgs
(15.71)
where = G1
l
is the approximate deltered eld. Using the zeroth-order
expansion of the deconvolution operator, one obtains a scale-similarity type
model:
2
2
2
C( ) ,
(15.72)
sgs
where the constant C is equal to one if the Van Cittert method is used.
Other values can be found using a dynamic procedure based on a Germanotype identity. A gradient type model is recovered replacing the convolution
lters by their equivalent dierential expansion, and then truncating these
expansion at an arbitrary order. For box and Gaussian lters, the rst order
term is
2
2
C ||2 ,
(15.73)
sgs
where the parameter C is lter dependent and can be adjusted using a dynamic procedure.
The subgrid scalar variance can also be easily retrieved if a structural
model based on an explicit reconstruction of the subgrid scales is used to
close the ltered scalar equation. In this case, it is directly computed from
the synthetic subgrid scalar eld. It is also straightforwardly extracted if
a dierential model for the scalar uxes is used.
2
, in
Another possiblity consists in solving a prognostic equation for sgs
a way similar to what is done for the subgrid kinetic energy. Using (15.8)
as a starting point, a closed equation is obtained. The resulting model is
471
presented in (15.135) for the sake of brevity. Prognostic equations with dynamic coecients can also be derived using techniques presented in Sect. 5.4
(p. 173) for the transport equation of the subgrid kinetic energy.
A simple expression is obtained in the simplied case where the inertialconvective range spectrum (15.22) is assumed to be valid at all subgrid wave
numbers:
+
3
2/3
2
=
E (k)dk =
1/3
.
(15.74)
sgs
2 2/3
/
Models for the Subgrid Scalar Dissipation Rate. A rst simple evaluation of is obtained assuming that the cuto is located within an innite
inertial-convective range:
+
2
2
sgs
qsgs
= C
,
(15.75)
where the constant C can by either computed analytically or adjusted using

a dynamic procedure. Considering an innite inertial-convective range, one
obtains

2
2
.
(15.76)
C=
3 3K0
The other way to evaluate the subgrid scalar variance dissipation rate is
to assume that the local equilibrium hypothesis holds: in this case, it is equal
to the local subgrid scalar variance production rate, leading to
=
(15.77)
where the subgrid scalar ux can be evaluated using any adequate subgrid
model. If a subgrid diusity model is utilized, one obtains
= 2sgs ||2
(15.78)
This expression can be modied to account for the dissipation at the

resolved scales, yielding
= 2( + sgs )||2
(15.79)
In order to obtain a more general expression which does not rely on the
local equilibrium assumption, Jimenez et al. [352] assumed that the characteristic subgrid mixing time is proportional to the subgrid turbulent time:
2
2
sgs
qsgs
(15.80)
(15.81)
yielding the following model

= C
sgs
2
qsgs
472
where C is a parameter. Tests show that C = 1/P r leads to satisfactory

results. All quantities which appear in the right hand side of (15.81) can be
evaluated using ad hoc models.
15.2.7 A Few Applications
Heat transfer at free surface [93]

Heat transfer in plane channel ow [92, 755, 196, 753]
Isotropic turbulence [121, 441, 538, 122, 220, 608, 344, 345, 752]
Complex cavities and ducts [279, 590, 639]
Homogeneous turbulence [344, 345]
Time-developing shear layer [344, 345]
Flow past a backward facing step [27]
Jet impinging on a plate [738, 155, 126, 236]
Flow in a corrugated passage [141]
Mixing Layer [168]
Flow in rotating/steady duct with/without rib [557, 558, 559, 560]
Flow in S-shaped duct [561]
Wall-mounted cube matrix [565]
15.3 The Active Scalar Case:

Stratication and Buoyancy Eects
We now turn to the case of the coupling with an active scalar, i.e. with
a eld which has a feedback eect on the velocity eld, leading to a two-way
coupling between the NavierStokes and the scalar equations. Since there are
a very huge number of possible interactions, it is chosen to put the emphasis
on buoyancy and stable stratication eects. Other physical models, such as
EulerianEulerian models for two-phase ows, will not be considered.
15.3.1 Physical Model
Buoyancy and stable stratication eects originate in the force to which
a blob of uid is submitted when it is immersed in a uid with dierent
density. Therefore, a natural physical parameter to describe these eects is
the density. When very weak compressiblity eects are taken into account,
the density of usual uids is assumed to decrease when the temperature is
increased. Thus, the temperature can also be used to represent the dynamics
of these ows. Both solutions are found in the literature, depending on the
authors and specic purposes of the studies. In the following, the scalar eld
will be related to the temperature eld.
Using the Boussinesq approximation (see reference books for a detailed
discussion of the range of validity of this model, e.g. [714, 708, 439]), the
15.3 The Active Scalar Case: Stratication and Buoyancy Eects
473
basic set of unltered equations consists in the NavierStokes equations

with a gravitational source term supplemented by a scalar equation identical to (15.1):
u
+ (u u) = p + 2 u +
t
g
0

(15.82)
feedback term
u = 0 ,
(15.83)
+ (u) = 2 ,
(15.84)
t
where is potential temperature, dened as the thermodynamic temperature
increased by the normalized gravitational potential, g = (0, 0, g) the gravity
vector and is the potential temperature eld associated to the hydrostatic
equilibrium. The vector g is related to the gravitational acceleration. 0 is
related to a reference value, which is assumed to be unique for the whole ow
domain. The molecular viscosity and the molecular diusivity are assumed
to be constant and uniform.
The corresponding set of letered equations is very similar to those used
in previous chapters, since the original set of equations diers by only one
source term:
u
+ (u u + ) = p + 2 u +
g
t
0
u = 0 ,
(15.85)
(15.86)
+ (u + ) = 2 ,
(15.87)
t
where it has been assumed that the scale separation operator perfectly commutes with all partial derivatives operator and that the mean eld is
varying slowly enough in space to have = . The subgrid uxes and
have exactly the same expressions as in the passive scalar case.
A deeper insight into the two-way coupling is gained rewriting the evolution equations for the subgrid quantities. Assuming that the vertical direction
is associated to x3 , the equations for the subgrid mometum uxes are:
G (ui , uj )
G (ui , uj )
+ uk
t
xk
uj
ui
G (uk , uj )
xk
xk

ui uj
2 G (ui , uj )
+
2G
,
xk xk
xk xk

p
G (uk , ui , uj )
G uj ,
xi
xk
g
+
(i3 G (uj , ) + j3 G (ui , )) ,
= G (uk , ui )
coupling term
(15.88)
474
and the corresponding equations for the generalized subgrid kinetic energy
is:

1
G (ui , ui )
G (ui , ui )
=
G (ui , ui , uj ) + G (p, uj )
t
xj 2
xj

ui ui
ui
G
,
G (ui , uj )
xj xj
xj
g
+ i3
G (, ui ) .
(15.89)
0

coupling term
The subgrid scalar uxes are solutions of

G (ui , )
G (ui , )
+ uk
t
xk
ui
= G (uk , ui )
G (uk , )
xk
xk

ui
+
, + G
, ui
G
xk
xk
xk

ui
p
( + )G
,
G ,
xk xk
xi
g
G (uk , ui , )
+ i3
G (, ) , (15.90)
xk
0

coupling term
while the generalized subgrid scalar variance equation is kept unchanged,

since there is no gravitational source term in the scalar equation. Therefore,
the subgrid variance is not directly aected by the feedback eect on the
velocity eld, but it is sensitive to it via changes in the latter.
15.3.2 Some Insights into the Active Scalar Dynamics
Flows governed by (15.82)(15.84) exhibit a very wide and complex range
of physical mechanisms, which originate in the coupling that may exists between convection, diusion, stratication and other features such as mean
shear, rotation and boundary conditions. It is meaningless to try to give an
exhaustive description of all these possibilities. The present section will focus
on very simple cases and the emphasis will be put on results dealing with
interscale transfers, which are of primary interest for subgrid modeling.
The inuence of the stratication eects on the subgrid kinetic energy is
seen looking at the last term of (15.89), which can lead to an increase associated to buoyancy eect or a decrease due to stable stratication damping,
depending on its sign. Its relative importance with respect to the production
term is measured by the ux Richardson number:
Rif =
g
0 G (u3 , )
ui
G (ui , uk ) x
k
(15.91)
475
High values of Rif are associated to ows in which stratication eects are
dominant, while they can be neglected when Rif 1. Neglecting energy
transport across the ow, one obtains
Rif 1
ui
G (ui , uk ) x
k
(15.92)
showing that destabilizing eects are present if Rif > 1 (the denominator of
the last term being assumed to be positive).
Two basic cases are discussed in the present section:
The case of stable stratication (p. 475), in which stratication eects tend
to damp the subgrid kinetic energy. An important feature of ows with
stable stratication is the existence of dispersive internal gravity waves.
These waves are important in many applications dealing with meteorology
and oceanology. They induce irrotational large-scale horizontal motions,
whose interactions with small scale turbulence will be one of the most
important issue discussed below.
The case of unstable stratication (p. 480), where buoyancy eects generate
turbulence, a well known example being thermal plumes.
Another relevant parameter is the group
N =
g
0 z
(15.93)
Under the assumption that the subgrid heat ux G (, u) obeys a Fickianlike law, i.e. G (, u) C, where C is a positive constant, it is seen
from (15.88), (15.89) and (15.90) that the local stablizing/destabilizing eect
of stratication depends on the sign of N , i.e. on the local sign of the vertical
resolved gradient /z:
1. If /z > 0, the eect is stabilizing, since it appears as a sink term in the
subgrid kinetic energy equation. In this regime, gravity waves are stable
and their frequency is characterized by the BruntVais
al
a frequency, N
N.
2. If /z < 0, gravity waves are unstable and break up into turbulence:
the buoyancy eects are seen to increase the production ofsubgrid kinetic
energy. The buoyancy time scale is estimated as Tb 1/ N .
Energy Transfers in Stably Stratied Flows. Energy transfers in stably stratied ows have been addressed by many authors, and are observed
to be case dependent. Therefore, no general description of the transfer across
a cuto similar to what exists for the momentum equation is available, the
main explanation for that being that such a general description is nearly
impossible. The dicult points which preclude such an analysis are the following
476
These ows are strongly anisotropic, and it has already been seen that
kinetic energy transfers within homogeneous anisotropic ows cannot be
described in a simple and general way, since they are case-dependent.
The description in terms of interscale transfers is not sucient to obtain
an accurate picture of the governing physical processes. It appears that
the concept of modes must be introduced to achieve a meaningful analysis:
the energy must be split into the vortex kinetic energy and the total wave
energy, the latter being further decomposed into the potential energy and
the wave kinetic energy [266]. The analysis of triadic interactions using this
scheme leads to a global description which is too complex to yield results
that can be straightforwardly utilized for subgrid parameterization. This
modal decomposition results from a local decomposition of the velocity and
temperature eld in the Fourier space:
(k) = u
1 (k) + u
2 (k) ,
u
(15.94)
1 (k) and u
2 (k) are related to the vortex and internal gravity wave
where u
components, respectively, with
1 (k) = 1 (k)e1 (k),
u
where
e1 (k) =
2 (k) = 2 (k)e2 (k) ,

u
kg
,
|k g|
e2 (k) =
k e1 (k)
|k e1 (k)|
(15.95)
(15.96)
The two vectors (e1 (k), e2 (k)) generate an orthonormal basis for the plane
perpendicular to k. The vortex kinetic energy spectrum, 1 (k), the wave
kinetic energy spectrum, 2 (k), and the potential energy spectrum P (k)
are dened as follows:
1 (k) =
<
1 ;
1 (k)1 (k) ,
2
1
P (k)(k) =
2N 2
<
1 ;
2 (k)2 (k)
2
;
<
(k)

(k)
,
2 (k) =
where N 2 is equal to the square of the Brunt-Vais

al
a frequency built on
the mean temperature gradient:
N2 =
g d
0 dz
(15.97)
The total kinetic energy spectrum is recovered summing the vortex and
wave contributions: E(k) = 1 (k) + 2 (k).
As a consequence, this section will be devoted to the description of the
main features of very simple cases which have been very accurately analyzed
using both theoretical tools and direct numerical simulation.
477
The most simple case is initially isotropic turbulence submitted the

strong stabilizing eects. The main observed phenomenon is the breakdown of isotropy, associated to a drastic restriction of motion along the
mean stratication direction. After some times, the ow is organized in thin
layers with a strong variability along the stratication direction (pancakeshaped large-scale vortices), leading to the occurance of a velocity eld which
is almost two-component4 Detailed investigations of energy transfers using
Direct Numerical Simulation and EDQNM models have been carried out
[235, 304, 513, 266, 689, 267]. The main results are the following
Several phases are observed, the number of which being dependent on the
inital condition and the value of the Richardson number. The rst phase (if
initial potential energy is large enough) corresponds to a ow whose large
scales are governed by stratication eects and inertial gravity waves and
exhibit a N 2 k 3 kinetic energy spectrum, while small scales are not aected
by stratication eects and are governed by the usual isotropic turbulent
dynamics. In the last phase, all scales are anisotropic and stratication
eects drive the whole ow.
During the rst phase (pre-collapse phase), the ow dynamics is controlled
by the potential energy being transferred from large to small scales at
a higher rate than kinetic energy, leading to a higher total energy decay as
in unstratied isotropic turbulence. The kinetic energy along the stratication direction is converted into potential energy, feeding the irreversible
forward cascade of potential energy. A consequence of the potential energy
cascade is the existence of a persisting counter-gradient scalar ux. Such
a counter-gradient scalar ux should be associated to a negative subgrid
diusivity at small scales, making all previous models irrelevant. But this
counter-gradient ux seems to play no major dynamical role on the smallest
scales during the rst phase if the molecular diusivity is high enough.
The nal phase (post-collapse phase) is associated to a scalar ux collapse
at large scales. The ratio of the kinetic energy along the the stratication
direction and the potential energy reach a constant value, while large-scale
energetic motion in directions perpendicular to the mean stratication direction arise from turbulence, forming the vortex part of the ow. The
small scale dynamics is dominated by stratication eects. An explanation is that, because of the collapse of large scales which reduces the rate
of the forward energy cascade, the smallest scales are made sensitive to
stratication.
A detailed analysis of the transfers reveals that the pure vortex-vortex
interactions (i.e. triadic interactions involving vortex modes only) are initially important among all triadic interactions, the resonant interactions
involving at least one wavy mode being less important. The former are
4
But not two-dimensional, since its dynamics is very dierent from the one of
two-dimensional turbulence. In particular, no strong backward energy cascade is
observed.
478
responsible for the isotropy breakdown and the blocking of a possible

backward energy cascade. A very important feature, which limits the accuracy of the modeling of triadic transfers in the physical space, is that
the energy transfers are strongly dependent on both the modulus and the
direction of the three wave numbers. Oscillatory exchanges between wave
kinetic energy and potential energy are also observed (their sum exhibiting
a non-oscillatory behavior), associated to irreversible anisotropy creation,
leading to angular variations (in the spectral space) of the vortex kinetic
energy and the total wave energy.
The counter-gradient scalar ux at small scales can inhibit mixing if the
molecular diusivity is too low, leading to a complex behavior of the scalar
variance dissipation rate.
These results show that the triadic transfer patterns are much more complicated than in unstratied ows. The physical picture presented above is to
be further complexied to account for additional physical mechanisms, such
as turbulence generation by a forcing term, interaction with a mean shear,
coupling with dynamics, etc.
The main information retrieved from the previous analysis is that the
forward kinetic energy cascade is decreased by stratication eects in the
post-collapse phase. The analysis of the transfers across a cuto wave number during the pre- and post-collapse phase was carried out by Metais and
Lesieur [514] using numerical simulations. Using the orthogonality of the local reference frame in the Fourier space, the authors extend the analysis
presented in Sect. 5.1.2 by decomposing the total kinetic energy transfer into
a vortex and wave part, and introduce a spectral subgrid viscosity representation for each part:
e,i
Tsgs
(k|kc )
,
(15.98)
ei (k|kc ) = 2
2k i (k)
where i = 1, 2 refers to the subgrid transfer associated with the corresponding
mode in the modal decomposition. As in the unstratied case, these subgrid
eective viscosties can be normalized using the kinetic energy at the cuto,
yielding
'
ei (k|kc ) = ei+ (k|kc )
E(kc )
kc
(15.99)
The total eective subgrid viscosity, e (k|kc ), is obtained using the following relationship
e (k|kc )E(k) = e1 (k|kc )1 (k) + e2 (k|kc )2 (k) .
(15.100)
This last expression illustrates the high diculty which arises in the modeling of the subgrid transfers via a single subgrid viscosity, since its value
depends on both the energy cascade rate of the vortex and wave components
and the repartition of the kinetic energy among these two modes. This double dependency shows that the total eective viscosity must be sensitive to
479
the initial conditions, which seems to preclude any general accurate denition. The subgrid potential temperature transfer is formally dened as in the
passive scalar case, since the evolution equations are the same.
It is observed in the numerical simulations carried out by Metais and
Lesieur that
In the pre-collapse phase (see Fig. 15.6), e1+ (k|kc ) and e2+ (k|kc ) are almost identical and do not dier from the eective subgrid viscosity observed in unstratied isotropic turbulence: the same plateau (at low wave
numbers) and cusp (near the cuto) behaviors are observed. The eective
subgrid diusivity exhibits a plateau at low wave numbers but do not show
any cusp at high wave numbers.
In the post-collapse phase (see Fig. 15.7), the cusp in the two subgrid
viscosities is higher than in the pre-collapse phase because of the shifting
of the spectrum maxima towards low wave numbers. The vortex-related
viscosity, e1+ (k|kc ), exhibits a plateau with constant value 0.09 at low
wave numbers, while the wave mode related viscosity exhibits the same
plateau value at intermediate scales but increases with decreasing wave
number at low wave numbers. The cusp is also increased in the subgrid
diusivity. Its is important noting that despite this increase in the relative
intensity of the transfer at small scales, the absolute level of the eective
Fig. 15.6. Eective subgrid viscosities e1+ (k|kc ) (Solid line) and e2+ (k|kc )
(Dashed line) and subgrid diusivity (Dotted line) in the LesieurMetais stablystratied case (pre-collapse phase).
480
Fig. 15.7. Eective subgrid viscosities e1+ (k|kc ) (Solid line) and e2+ (k|kc )
(Dashed line) and subgrid diusivity (Dotted line) and subgrid diusivity (Dotted line) in the LesieurMetais stably-stratied case (post-collapse phase).

subgrid viscosities, which scales like E(kc ), is decreased in the postcollapse phase, since stable stratication inhibits the forward kinetic energy
cascade, leaving less energy in the small scales.
These results show that, depending on the considered ow regime and
initial conditions [287], both the intensity and the spectral shape of the global
transfer operator vary, rendering its accurate modeling in the physical space
nearly impossible.
Energy Transfers in Buoyancy-Driven Flows. We now turn to the case
were the buoyancy force acts as a turbulence generator, as in natural convection ows. In these ows, buoyancy-generated instabilities create unsteady
motion which will lead to turbulence. Similarly to the case of stable stratication eects, buoyancy-driven ows exhibit a wide range of physical mechanisms which prevent an exhaustive detailed analysis.
As an example, let us consider the analysis of energy transfers in thermal
plumes carried out by Baastians et al. [29]. Using an experimental database,
these authors observe a time mean backscatter in the lower part of the plume
and instantaneous backscatter in the total up-owing area. The energy transfer rate across the cuto exhibits large uctuations, instantaneous values being about ten times larger than the mean value. Forward and backward energy
cascades are equally important in the mean in the plume region. These results
481
show that a reliable subgrid model should be able to account for backscatter, and to distinguish between dierent regions of the ow, rendering the
modeling task even more dicult as in the stably stratied ow case.
But it is important to remark that, in the basic philosophy which underlies most of the developments in the eld of Large-Eddy Simulation, it
is assumed that driving mechanisms must be directly captured. In the case
of buoyancy-driven ows, buoyancy eects should therefore be directly simulated if they are the only source of turbulence production (free convection).
In the case of mixed convection or forced convection, the requirement might
be less stringent since other physical mechanisms are at play. This is why
some authors use unmodied passive scalar models (and usual closures for
the momentume equations) to treat RayleighBenard convection at medium
Prandtl number.
15.3.3 Extensions of Functional Models
We now discuss the most illustrative functional models in the physical space
for the active scalar case. Since a two-way coupling exists, subgrid models for
both the momentum and the scalar equations must be revisited to account for
stratication/buoyancy eects. Modeling strategies based on the denition of
scalar subgrid viscosity/diusivity parameters are rst considered (p. 481).
Dynamic models based on the Germano identity are presented in the second
part of the section (p. 486).
Scalar Subgrid Viscosity/Diusivity Models. The brief review of results dealing with interscale energy transfers in ows with active scalar given
in the previous section obviously shows that they are strongly aected in both
stable and unstable stratication cases. This is why most scalar models are derived considering a simplied kinetic energy balance equation which includes
buoyancy eects. The most popular simplied energy balance expression is
obtained by neglecting all diusive and convective eects in (15.89), yielding
an extended local equilibrium (production = dissipation) assumption:
= : S b ,
1
g
0
(15.101)
where it is recalled that , and are the subgrid kinetic energy dissipation
rate, the momentum subgrid tensor and the potential temperature subgrid
ux vector, respectively. Now specializing Eq. (15.101) by inserting scalar
functional models of the form
= 2sgs S,
= sgs
(15.102)
one obtains
sgs |S|2 sgs Nc2 =
(15.103)
482
where the local parameter Nc2 is dened as

Nc2 b =
g
0 z
(15.104)
and can be either positive or negative. Negative (resp. positive) values are
associated with a decrease (resp. increase) of the subgrid kinetic energy dissipation rate by the buoyancy eects.
Equation (15.103) shows that the two-way coupling observed in the active
scalar case precludes any decoupled denitions for sgs and sgs .
Now assuming that the cuto is located within an inertial-convective-like
range, dimensional analysis yields
4/3
sgs = C ||1/3
4/3
sgs = C ||1/3
(15.105)
where C and C are positive parameters. The combination of (15.103)

and (15.105) leads to the following expression for the dissipation rate:
4/3
= ||1/3

C |S|2 C Nc2 .
(15.106)
This expression leads to the following denition for the subgrid viscosity
and the subgrid diusivity:

1/2

2
Nc2
2
2 1/2
2

= C C |S| C Nc
= C |S|
,
P rsgs

(15.107)

1/2
1/2
Nc2
2
2
sgs = C C |S|2 C Nc2
= C |S|2
,
P rsgs

(15.108)
sgs
where the subgrid Prandtl number is dened as P rsgs C /C = C /C .

The main modication with respect to the passive scalar lies in the evaluation
of the characteristic time scale T 1 : it now combines the velocity gradient
and the buoyancy eects. Model (15.107) can be seen as an extension of the
classical Smagorinsky subgrid viscosity model (5.90). It is seen that stable
stratication, which is associated with positive values of Nc2 , corresponds
to a decrease of the subgrid viscosity and diusivity. This is in agreement
with results dealing with the evolution of subgrid transfers discussed above.
Unstable stratication corresponds to an increase of subgrid viscosity and
diusivity, in agreement with the physical picture that the forward kinetic
energy cascade is enhanced since the subgrid kinetic energy is increased.
The subgrid time scale in (15.107) and (15.108) is dened as an absolute
value and leads to the denition of positive subgrid parameters. The reason
is that the rough estimates used to obtain the expressions given above do not
483
ensure that realizability is enforced, i.e. that the energy destroyed by stabilizing stratication eects is smaller than the available subgrid kinetic energy.
But when the backscatter is dominant, the denition is no longer relevant
to account for subgrid transfers if the subgrid Prandtl number is kept positive. In this case, Eidson [209] proposes to clip negative values of Nc2 so that
T remains well-posed. A simple solution is to consider Nc2+ = max(0, Nc2 ).
The resulting subgrid viscosity model can be rewritten as a function of the
original Smagorinsky model (5.90):
'
1 Nc2+
2
sgs = C |S|
1
.
(15.109)

P rsgs |S|2
Smagorinsky
This modied expression can be recast in a form which emphasizes the

role of the Richardson number [690]:

Ri
2
sgs = C |S|
1
,
(15.110)

Riref
Smagorinsky
where the local subgrid Richardson number Ri is dened as Ri = Nc2+ /|S|2

and Riref P rsgs is a reference value. This new expression also illustrates the
physical meaning of the subgrid Prandtl number in stably stratied ows.
A modied evaluation of the time scale is proposed by Peng and Davidson [587], which is always well-posed from a mathematical standpoint:

Nc2
T = |S|
.
(15.111)
|S|P rsgs
A more general scalar model is proposed by Schumann, which contains
an anisotropic residual heat ux associated to buoyancy eects. It is derived
considering simplied evolution equation for the subgrid stresses and subgrid
heat uxes and assuming that the local equilibrium applies, yielding
2
= 2C T S,
= C T (b ) ,
(15.112)
where the time scale T 1 and the tensor b are dened as

C2
b ,
(15.113)
T + C2 Nc2

2
1 2
1 2
2
2
4
|S| (C1 + C2 )Nc +
|S| (C1 C2 )Nc2 + C1 C2 Nm
,
T =
2
4
(15.114)
where Nc2 is given by (15.104) and
+
2
Nm = (b b)( ) .
(15.115)
b = Id +
484
Table 15.1. Values of the constants in buoyancy-aected subgrid models.

Model
C1
C2
Passive scalar
Buoyant
Eidson
Schumann
0
1/P rsgs
2.5
2.5
0
0
0.
3.
As noted by Cabot [85], this general model encompasses the previous

ones. Corresponding values of the constants C1 and C2 are summarized in
Table 15.1.
The Schumann model appears to be the most complex one, and includes
more information related to anisotropy. It is therefore expected that it should
yield better results in buoyancy-driven ows with high Prandtl numbers.
More complex models relying on the denition of tensorial subgrid viscosity/diusivity to account for anisotropy can also be dened. Such models,
very similar to the one proposed by Yoshizawa for the passive scalar case
(Sect. 15.2.3, p. 465) were tested by Peng and Davidson [587] for buoyancydriven ows. These authors use the following expression for the subgrid heat
ux
,
(15.116)
G (ui , ) = CTsgs G (ui , uk )
xk
where C is a constant, Tsgs and appropriate time scale and the subgrid momentum ux is modeled by an ad hoc model. In the case a subgrid-viscosity
type model is considered for G (u, u) and the cascade time scale is evaluated
2
as Tsgs = /sgs , the resulting model is
2
G (ui , ) = C S ik
xk
(15.117)
The value of the subgrid viscosity can be modied accounting for this
expression in the following way. Starting from the local equilibrium hypoth4
3
/ ,
esis (15.101) and approximating the subgrid dissipation rate as sgs
one obtains
'
1 gi
2
G (ui , ) .
(15.118)
sgs = C |S|2
sgs 0
Now using (15.116), one obtains the following evaluation of the subgrid
viscosity
'
gi
2
sgs = C |S|2
S ik
,
(15.119)
xk
|S| 0
where C and are adjustable parameters, which can be computed using
inertial-range considerations or a dynamic procedure.
485
All functional models presented above are built on the same bases, and
are based on the resolved strain. They can be interpreted as extensions of the
Smagorinsky model. Another class of models, which is very popular among
researchers working in the eld of meteorology, is the one of models based on
the subgrid kinetic energy:
+
+
2 ,
2
sgs = C l qsgs
,
(15.120)
sgs = C l qsgs
where C and C are constants and l a characteristic length scale. The subgrid
kinetic energy must be evaluated using a prognostic equation or a simplied
approximation (see Sect. 15.3.5). Since the time scale T does not appear
explicitely in these expressions, the stratication eect must be taken into
account in the denition of the length scale l. The subgrid Prandtl number
associated with this corrected length scale is
P rsgs =
+ 2l
(15.121)
In the case of unstable stratication, the length scale is taken equal to its
usual value tied to the lter cuto, i.e. l = . In the case of stably stratied
ows, a rst simple model proposed by Schumann for meteorology-related
purpose is
'
2

qsgs
.
(15.122)
l = min , LN , LN = 0.76
Nc2
A more general corrected expression, which accounts for both stable stratication and shear eects is proposed by Canuto and Cheng [99]. It is written
as follows:
(15.123)
l = f (|S|, Nc ) ,
with

3/2
[1 X log(1 + aQ )] dQ
2
f (|S|, Nc ) =
(15.124)

3 3/2 P rsgs Sh2
X =
K0
1
,
16
F ri2
Q = (k/)2/3
a=
2
F r2 f (|S|, Nc )2/3 ,
2 i
(15.125)
where the non-dimensional shear number and the inverse Froude number are
dened as
|S|
Sh = +
,
2
qsgs
Nc
F ri = +
2
qsgs
(15.126)
486
Dynamic Models. A weakness shared by all the models presented above is

the existence of preset constants. They can be evaluated using inertial range
considerations, or they can be automatically adjusted using a dynamic procedure. The use of a dynamic procedure based on the Germano identity yields
much more dicult problems than in other cases, because of the intrinsic
coupling between the momentum and the temperature subgrid uxes. This
problem is clearly seen considering the following generic closures
ij = C fij (, u, , P rsgs ),
i =
C
hi (, u, , P rsgs ) .
P rsgs
(15.127)
The two unknown parameters are C and P rsgs . Inserting these expressions in the Germano identities for the momentum and temperature uxes
and using the classical assumptions, one denes the two following residuals:

* u,
* *
, P rsgs ) f2
,
(15.128)
Eij = Lij C fij (,
ij (, u, , P rsgs )
E i = hi

C * * *
hi (, u, , P rsgs ) h*i (, u, , P rsgs )
P rsgs
(15.129)
The two unknown parameters are chosen so as to minimize the global

error in the least square sense, leading to the denition of four constraints:
E : E
E
= 2E :
= 0,
C
C
E : E
E
= 2E :
= 0,
C
C
E : E
E
= 2E :
=0
P rsgs
P rsgs
E : E
E
= 2E :
=0 .
P rsgs
P rsgs
(15.130)
(15.131)
The complexity of this problem is enlightened looking at the denition of

the subgrid Prandtl number: since both fij and hi are non linear functions of
P rsgs , the use of either (15.130) or (15.131) leads to the denition of a system
of non-linear equations for it, which cannot be solved explicitely as in previous
cases. Thus, the dynamic evaluation of the constants now require to solve
a non linear system, leading to a subsequent increase in the computational
complexity. It also raises the problem of nding the best among the multiple
possible roots of the system.
This is why many authors use the same dynamic models as in the passive
scalar case (i.e. with basic models which are linear with respect to the subgrid
Prandtl number), with the idea that, if the grid is ne enough, most of the
stratication eects will be captured. This is especially true of the stable
stratication eects, where the decrease in the subgrid energy transfer may
be partially captured using a dynamic procedure, since resolved scales are
already aected. The case of buoyancy driven ows is much more dicult,
since turbulence production exist at small scales which cannot be inferred
from the resolved scale dynamics. Therefore, the use of purposely modied
487
models for buoyancy-driven ows is more important than in the case of stable
stratication.
15.3.4 Extensions of Structural Models
We now consider structural models. As in the passive scalar case, models
relying on the explicit evaluation of subgrid scales (soft deconvolution models,
scale similarity models, linear combination models [19], explicit stochastic or
deterministic models for subgrid uctuations, multilevel simulations) can be
used in a straightforward way, since they do not rely on any assumptions
dealing with the nature of subgrid transfers. Consequently, these models will
not be explicitly detailed in this section, since they dont need to be modied
to account for the active scalar dynamics. Another comment is that many
structural models have not yet been tested in stably/unstably stratied ows.
In this section, we will put the emphasis on models which contain explicit
modications and whose accuracy has already been assessed in real LargeEddy Simulations:
Deardors dierential stress model (p. 487), which requires to solve ten
additional transport equations.
Schumanns algebraic stress model (p. 488), wich can be seen as a simplication of the previous model and necessitates to solve only one additional
prognostic equation to compute the subgrid kinetic energy.
Deardor Dierential Model. The dierential model proposed by Deardor [173] relies on solving closed expressions of (15.88) and (15.90).
In the equations for the subgrid momentum uxes (15.88), the only new
terms with respect to the case treated in Sect. 7.5.1 is the coupling term.
Since the closures proposed by Deardor are kept unmodied in the active
scalar case, all the dynamic coupling eects are contained in the coupling
term. Since this term is directly proportional to the subgrid heat uxes which
are explicitely computed, no further closure assumptions is needed for these
equations.
The equations for the subgrid scalar uxes and the associated closures
have already been discussed in Sect. 15.2.4. Among already closed terms in
the passive scalar case, only the pressure-temperature term is modied to
account for stratication eects: Equation (15.60) is changed into
+

2
qsgs
1
p
G (ui , ) bi G (, ) ,
(15.132)
= C1
G ,
xi
3
which necessitates an evaluation of the generalized subgrid scalar variance,

G (, ). The new term appearing in (15.90) is the coupling term, which is
also proportional to the generalized subgrid scalar variance.
Therefore, an additional prognostic equation for G (, ) deduced from
(15.8) is solved. Term V II is directly computed, since the subgrid scalar
488
uxes are explicitly computed. Molecular diusion IX, is neglected, while

the molecular dissipation IX is evaluated as
+

2
qsgs

2G
G (, ), CIX = 0.42 ,
,
(15.133)
= CIX
xk xk
while the subgrid diusion cubic term X is closed as

+
2
G (uk , , ) = CX qsgs
G (, ), CX = 0.2
xk
(15.134)
The resulting solvable equation for the subgrid temperature variance is

+
2
qsgs
G (, )
G (, )
+ uk
G (, )
= 2G (uk , )
CIX
t
xk
x

+k
2
+CX
qsgs
G (, ) (15.135)
,
xk
xk
achieving the description of the model.
Schumann Algebraic Stress Model. A much simpler model, which does
not require to solve a large amount of additional transport equations, was
developed by Schmidt and Schumann [651] to analyze the convective boundary layer dynamics. This model can be interpreted as a simplication of the
Deardor model, in which several contributions will be neglected.
Starting from (15.88) and making the following assumptions:
1. The subgrid uxes respond instantaneously to large-scale forcing, i.e. the
time derivative term is negligible.
2. Diusive terms are small, and can be neglected.
3. Importance of the deviatoric part of the subgrid stress tensor in the
production terms is small, with respect to the one of the isotropic part.
4. The pressure-velocity term can be modeled as
+

2
qsgs
g
ui
2
G
(ui , uj ) + C3
= C1 qsgs S ik C2
G p,
xk
0

2
G (ui , )k3 + G (uk , )i3 G (u3 , )ij
,
3
(15.136)
where C1 , C2 and C3 are model parameters. The star superscript denotes
the deviatoric part of the tensor.
5. The subgrid dissipation term is approximated as

uj ui
2 2
,
S ij .
(15.137)
= qsgs
2G
xk xk
3
489
one obtains the following linear relationship for the subgrid stresses:
+

2
qsgs
g
G (ui , )k3 + G (uk , )i3

G (ui , uj ) = (1 CBm )
CRm
0

2
2 2
G (u3 , )ij (1 CGm ) qsgs
S ij ,
3
3
(15.138)
where CRm = 3.5, CBm = 0.55 and CGm = 0.55. The subgrid heat uxes are
known since they are also outputs of the model, and the subgrid kinetic energy
is obtained solving a prognostic equation derived from (15.89). Since the exact
equation diers from the one associated to the unstratied case only because
of the coupling term, and that this coupling term is directly proportional
to the subgrid heat ux which is an output of the model, a simple solution
consists in using the same closed equation as in the unstratied case (see
Sect. 5.4.2, p. 173) with the additional coupling term. More complex variants
of the prognostic equation for the subgrid kinetic energy are presented in
a dedicated section (Sect. 15.3.5).
The subgrid heat ux equation (15.90) is simplied using similar assumptions:
1.
2.
3.
4.
The time derivative is negligible.

Convective and diusive uxes are negligible.
The dissipation term is negligible.
Contribution of the deviatoric part of the subgrid stress tensor in the
production term is very small compared to the one of its isotropic part.
5. The pressure-temperature subgrid ux can be modeled as
+

2
qsgs
2
a2 2
G (, ui )
= 2a1 qsgs
G p,
xi
xi
g
G (, )i3 ,
(15.139)
+ 2a3
0
with ai are adjustable parameters.
The resulting algebraic equation is
+
2
qsgs
u
i
2
G (, uk ) = 2 (1 CG )qsgs
CR
ik +
x
3
xi
k
+ (1 + CB )
g
G (, )i3
0
. (15.140)
To obtain a fully explicit model for the subgrid heat ux, it is necessary
to eliminate the term related to the resolved velocity gradient in the left hand
490
side, i.e. to neglect production eects associated with the interaction of the
subgrid heat ux with the resolved velocity gradient, yielding
+
2
qsgs
2
g
2
G (, ui ) = (1 CG )qsgs
+ 2(1 CB ) G (, )i3 .
3
xi
0
(15.141)
The constants are set equal to: CR = 1.63, CG = 0.5, CB = 0.5. The
subgrid scalar variance equation is simplied using a local equilibrium hypothesis, leading to the following explicit expression
+
2
qsgs
G (, ) = 2G (, ui )
,
(15.142)
C
xi
CR
with C = 2.02. The explicit model is obtained by solving (15.138), (15.141)

and (15.142) supplemented with a prognostic equation for the subgrid kinetic
energy. Therefore, it appears much more simple than a dierential stress
model. Another interesting feature of this model is that it do not rely on
the eddy diusivity/viscosity paradigm and is fully anisotropic. Nevertheless,
since it is based on strong assumptions, its domain of validity is expected to
be narrower than the one of the full dierential model.
In the absence of buoyancy, the model simplies as a simple subgrid viscosity/subgrid diusivity model:
+
+
2 ,
2
sgs = C qsgs
,
(15.143)
sgs = C qsgs
with
C =
2 1 CGm
,
3 CRm
C =
2 1 CG
3 CR
(15.144)
The corresponding value of the subgrid Prandtl number is

P rsgs =
2K0
(15.145)
15.3.5 Subgrid Kinetic Energy Estimates

2
The modeled prognostic equation for the subgrid kinetic energy qsgs
(u
,
u
)
can
be
modied
to
account
for
buoyancy/stratication
eects
i
2 G i
[402]. Rewritting the transport equation for this quantity as
2
2
qsgs
qsgs
g
+ uj
=P D+
G (u3 , )
t
xj
0
(15.146)
491
where the production P and the diusion D are dened as

2
D = (uqsgs
+ pId)
P = : S,
(15.147)
and where is the dissipation rate. The model for the production term and
the diusion term are not modied to account for active scalar dynamics
+
2 |S|2
,
(15.148)
P = Ce qsgs
D = 2Ce
xk
2
+
qsgs
2
qsgs
xk

,
(15.149)
where Ce is a constant. The subgrid kinetic energy dissipation rate is modied

to account for stabilizing stratication eects by writing it as
= Ce
2 3/2
qsgs
L
(15.150)
where the characteristic dissipative length scale L is dened as

L =
1
2
1
1
+ 2
L2N
LS
1/2
.
(15.151)
The subgrid buoyancy lenghtscale, LN , and the subgrid shear length scale,
LS , are dened as
'
'
2
2
qsgs
qsgs
LN = 0.76
,
L
=
2.76
,
(15.152)
S
Nc2
|S|2
where Nc2 is dened as in (15.104). This expression was derived to account
for the dynamics of a shear-driven stable atmospheric boundary layer. The
predicted value of the dissipation length scale is larger or equal to , yielding a decrease of the predicted dissipation, in agreement with the observed
diminution in the forward kinetic energy cascade rate.
An algebraic expression which takes into account the correction in the
denition of the dissipation length scale is obtained by assuming that the
dissipation is locally balanced by conversion of the mean ow energy into
subgrid kinetic energy (production = dissipation):
=P D
(15.153)
Using (15.150) and (15.103), one obtains

2
=
qsgs

2/3
L
sgs |S|2 sgs Nc2
Ce
(15.154)
492
This equation is fully non-linear if models based on the subgrid kinetic

energy are used for the subgrid viscosity and the subgrid diusivity, and/or if
the corrected model for L is utilized. In the absence of stability correction for
the dissipation length scale, and using the following subgrid kinetic energybased models:
+
+
2 ,
2
sgs = C qsgs
,
(15.155)
sgs = C qsgs
one obtains the following explicit algebraic expression [690]:
2
2
qsgs

=
C |S|2 C Nc2
Ce
(15.156)
More expressions for the subgrid kinetic energy can be found using other
models for the subgrid viscosity and diusivity.
15.3.6 More Complex Physical Models
The active scalar model discussed above can be further complexied to account for more complex physics. This is currently done in studies dealing
with the atmospheric boundary layer, in which detailed microphysical models (icing, moisture) and infrared radiative cooling are taken into account.
The related subgrid models are made more complex to account for new physical mechanisms. They are nor presented here, since they are very specic
to the eld of application. The interested reader can refer to the original
publications.
15.3.7 A Few Applications
Stably stratied ows:
Stably stratied channel ow [19, 195, 238]

Forced homogeneous stably stratied turbulence [110]
Decaying homogeneous stably stratied ows [514, 373]
Turbulent penetrative convection [124]
Wake in a weakly stably stratied uid [194]
Buoyancy-driven ows:
Thermal plume [29, 28]

Turbulent penetrative convection [124]
Buoyancy-generated homogeneous turbulence [123]
RayleighBenard convection [209, 394, 766, 85, 575]
Rotating RayleighBenard ow [152]
Forced and mixed convection in rotating and non-rotating square duct [577]
Natural convection in a cavity [586, 800]
493
Turbulent convection driven by free-surface cooling [814]

Buoyant jet [808]
Buoyant wake [539]
Buoyant pipe ow [433]
Meteorology-related applications: [601, 488, 525, 444, 698, 651, 130, 656,
173, 402, 277, 690, 337, 615, 109, 154, 111, 527, 534, 528, 317, 657, 529, 658,
530, 450, 699, 697, 690, 638, 526, 532, 531]
A. Statistical and Spectral Analysis

of Turbulence
A.1 Turbulence Properties

Flows qualied as turbulent can be found in most elds that make use of
uid mechanics. These ows posses a very complex dynamics whose intimate
mechanisms and repercussions on some of their characteristics of interest to
the engineer should be understood in order to be able to control them. The
criteria for dening a turbulent ow are varied and nebulous because there
is no true denition of turbulence. Among the criteria most often retained,
we may mention [150]:
the random character of the spatial and time uctuations of the velocities, which reect the existence of nite characteristic scales of statistical
correlation (in space and time);
the velocity eld is three-dimensional and rotational;
the various modes are strongly coupled, which is reected in the nonlinearity of the mathematical model retained (NavierStokes equations);
the large mixing capacity due to the agitation induced by the various scales;
the chaotic character of the solution, which exhibits a very strong dependency on the initial condition and boundary conditions.
A.2 Foundations of the Statistical Analysis

of Turbulence
A.2.1 Motivations
The very great dynamical complexity of turbulent ows makes for a very
lengthy deterministic description of them. To analyze and model them, we
usually refer to a statistical representation of the uctuations. This reduces
the description to that of the various statistical moments in the solution,
which sharply reduces the volume of information. Moreover, the random character of the uctuations make this approach natural.
496
A. Statistical and Spectral Analysis of Turbulence
A.2.2 Statistical Average: Denition and Properties

We use to denote the stochastic mean (or statistical average, or mathematical expectation, or ensemble average) of a random variable calculated
from n independent realizations of the same phenomenon {l }:
1
l
n n
n
= lim
(A.1)
l=1
The turbulent uctuation l associated with the realization l is dened

as its deviation from the mathematical expectation:
l = l
(A.2)
By construction, we have the property:

0
(A.3)
On the other hand, uctuation moments of second or higher order are not
necessarily zero. The standard deviation can be dened as:
2 =
2
(A.4)
We dene the turbulence intensity as /.

The correlation at two points in space and two times, (x, x ) and (t, t ) of
the two random variables and , denoted R (x, x , t, t ) is:
R (x, x , t, t ) = (x, t)(x , t )
(A.5)
A.2.3 Ergodicity Principle

When is a random steady function in time (i.e. its probability density function is independent of time), we can apply the ergodicity principle according
to which it is equivalent, statistically speaking, to consider indenitely repeated experiments with a single drawing or a single experiment with an
innite number of drawings. We will therefore admit that a single experiment of innite duration can be considered as representative of all possible
scenarios.
The theorem of ergodicity says that the quadratic mean of the random
function T (t) dened by:
T (t) =
1
T

t
t+T
(t )dt
(A.6)
A.2 Foundations of the Statistical Analysis of Turbulence
497
converges to a non-random limit equal to the stochastic mean as T

only on the condition that:

1 T
lim
R (t)dt = 0 ,
(A.7)
T T 0
where R (t) is the time autocorrelation (or covariance) of the uctuations
of over time interval t:
R (t) = ((t ) )((t + t) )
(A.8)
For turbulent uctuations, the random character reects the fact that
R (t) 0 as t . So if we dene the mean in time as the limit of T
as T , i.e.:

1 T
= lim
(t)dt ,
(A.9)
T T 0
we get the equality:
=
(A.10)
We establish that the standard error varies as 1/ T for suciently

large T . Another way of estimating is to construct the experimental
average n dened as the arithmetic mean from experiments:
1
i (t)
n i=1
n
n (t) =
(A.11)
where the time t is arbitrary since the ow is assumed

to be statistically
steady. We show that the standard error decreases as 1/ n if the experiments
l are independent.
Let and be two random variables. The operator thus dened veries
the following properties, sometimes called Reynolds rules:
+ = + ,
a = a
a = const.
(A.12)
,
= ,
; <
=
s = x, t ,
s
s

;
<
(x, t)d3 xdt = (x, t)d3 xdt .
(A.13)
(A.14)
(A.15)
(A.16)
Any operator that veries these properties is called a Reynolds operator.

We deduce from these relations the properties:
=
=
,
0 .
(A.17)
(A.18)
498
A.2.4 Decomposition of a Turbulent Field

Decomposition Principle. One technique very commonly used for describing a turbulent eld is statistical representation. The velocity eld at time t
and position x splits into:
u(x, t) = u(x, t) + u (x, t) .
(A.19)
Using this decomposition and the stochastic mean, we dene an evolution

equation for the quantity u(x, t). To recover all the information contained
in the u(x, t) eld, we have to handle an innite set of equations for the
statistical moments of it. The quadratic non-linearity of the NavierStokes
equations induces an intrinsic coupling among the various moments of the
solution: the evolution equation of the moment of order n in the solution
uses the moment of order (n + 1). To recover all the information in the
exact solution, it is thus necessary to solve an innite hierarchy of coupled
equations. As this is impossible in practice, this hierarchy is truncated at
an arbitrarily chosen level so as to obtain a nite number of equations. This
truncation brings out an unknown term that will be modeled using closure
hypotheses. If the degree of precision of the information obtained theoretically
increases with the number of equations retained, the consequences of the
truncation and of the hypotheses used are dicult to predict.
Equations of the Stochastic Moments. The evolution equations of the
mean eld are obtained by applying the averaging operator to the Navier
Stokes equations. By applying the rules of commutation with the derivation
in the case of an incompressible Newtonian uid and with no external forces,
we get
p
2 u
ui
+
ui uj =
+
,
(A.20)
t
xj
xi
xj xj
ui
=0
xi
(A.21)
where is the kinematic viscosity. The non-linear term ui uj is unknown

and has to be decomposed as a function of u and u . By introducing relation (A.19) and considering the properties (A.12) to (A.18), we get:
ui uj = ui uj + ui uj .
(A.22)
The last term of the right-hand side, called the Reynolds tensor, is unknown and has to be evaluated. It represents the coupling between the uctuations and the mean eld. This evaluation can be made by solving the
corresponding evolution equation, either by employing a model, called closure or turbulence model.
A.3 Introduction to Spectral Analysis of the Isotropic Turbulent Fields
499
A.2.5 Isotropic Homogeneous Turbulence

Denitions. A eld is said to be statistically homogeneous along the parameter x, or imprecisely just homogeneous, if its statistical moments are
independent of the value of x where the measurements are made. This is
expressed:
1 ....n = 0 .
(A.23)
x
A homogeneous eld is said to be statistically isotropic (in the Taylor
sense), or more simply isotropic, if all statistical moments relative to a set
of points (x1 , ..., xn ) at times (t1 , ..., tn ) remains invariant when the set of n
points and the coordinate axis are rotated, and if there is statistical invariance
for symmetry about an arbitrary plane.
We may note that there exists an idea of quasi-isotropy introduced by
Moat, which does not require the invariance by symmetry.
A Few Properties. A turbulent eld is said to be homogeneous (resp.
homogeneous isotropic) if its velocity uctuation u is homogeneous (resp.
homogeneous isotropic). One necessary condition for achieving homogeneity
is that the mean velocity gradient be constant in space:
ui
= const.
xj
(A.24)
Isotropy requires that the mean eld u be zero. When the turbulence
is isotropic, only the diagonal elements of the Reynolds tensor are non-zero.
Moreover, these are mutually equal:
2
Kij
3
where K is the turbulent kinetic energy.
ui uj =
(A.25)
A.3 Introduction to Spectral Analysis

of the Isotropic Turbulent Fields
A.3.1 Denitions
The tensor of correlations at two points R (r) of a statistically homogeneous
vector eld u dened as:
R (r) = u (x + r)u (x)
(A.26)
can be related to a spectral tensor (k) by the following two relations:

R (r) =
(k)eikj rj d3 k ,
(A.27)

1
(k) =
(A.28)
R (r)eikj rj d3 k ,
(2)3
500
where i2 = 1. The tensor at the origin, R (0), is the Reynolds tensor.

In the case of an isotropic eld, the general form of the correlation tensor
becomes:

r r
,
(A.29)
R (r) = K [f (r) g(r)] 2 + g(r)
r
where f (r) and g(r) are two real scalar functions. When the velocity eld is
solenoidal, these two functions are related by:
g(r) = f (r) +
r f (r)
2 r
(A.30)
The incompressibility constraint also allows us to establish the following

relation for the tensor (k):

E(k)
k k
(k) =
,
(A.31)
4k 2
k2
where the scalar function E(k) is called a three-dimensional spectrum. It
represents the contribution of the wave vectors of k to the turbulent kinetic
energy, i.e. wave vectors whose tips are included in the region located between
two spheres of radius k and k + dk. The spectral energy density, denoted
A(k), is therefore equal to E(k)/4k 2 . The three-dimensional spectrum is
computed from the spectral tensor by integration over the sphere of radius k:

1
E(k) =
(A.32)
ii (k)dS(k) ,
2
where dS(k) is the integration element on the sphere of radius k. This quantity can also be related to the function f (r) by the relation:

K
E(k) =
kr (sin(kr) kr cos(kr)) f (r)dr .
(A.33)
0
The turbulent kinetic energy, K, is found by summation over the entire
spectrum:

ui ui
K
=
E(k)d3 k .
(A.34)
2
0
By construction, the spectral tensor has the property:
ij (k) = ij (k) ,
(A.35)
where the asterisk indicates the complex conjugate number. The homogeneity
property of the turbulent eld implies:
ij (k) = ji (k)
(A.36)
501
The spectral tensor can also be related to the velocity uctuation u and
dened as:
to its Fourier transform u

1
i (k) =
u
(A.37)
u (x)ekj xj d3 x .
(2)3
Simple expansions lead to the equality:

ui (k )
uj (k) = (k + k )ij (k) .
(A.38)
So we see that the two modes are correlated statistically only if k+k = 0.
An equivalent denition of the spectral tensor is:

ij (k) =
u
uj (k )d3 k .
(A.39)
i (k)
A.3.2 Modal Interactions
The nature of the interactions among the various modes can be brought
out by analyzing the non-linear term that appears in the evolution equation
associated with them. This equation, for the mode associated with the wave
vector k (the dependency on k is not expressed, for the sake of simplicity) is:

ui
+ kj
aij = ki p k 2 u
i .
(A.40)
t
The two quantities
aij and p are related to ui uj and the pressure p by
the relations:

aij (k)ekl xl d3 k ,
(A.41)
ui (x)uj (x) =

1
p(x) =
p(k)ekl xl d3 k .
(A.42)
By introducing the spectral decompositions:

ui (x) =
u
i (k )ekl xl d3 k ,

uj (x) =
u
j (k )ekl xl d3 k ,
the non-linear term becomes:

u
i (k )
uj (k k )d3 k ekl xl d3 k
ui (x)uj (x) =

(A.43)
(A.44)
(A.45)
aij (k)

where we have performed the variable change k = k + k . The pressure term
is computed by the Poisson equation:
2 ui uj
1 2p
=
xi xi
xi xj
(A.46)
502
or, in the spectral space:

k 2 p = kl km
alm
The momentum equation therefore takes the form:

2
+ k u
m (k )
uj (k k )d3 k
i (k) = Mijm (k) u
t
(A.47)
(A.48)
in which
(A.49)
Mijm (k) = (km Pij (k) + kj Pim (k)) ,
2
where Pij (k) is the projection operator on the plane orthogonal to the vector k. This operator is expressed:

ki kj
.
(A.50)
Pij (k) = ij 2
k
The linear terms are grouped into the left-hand side and the non-linear
terms in the right. The rst linear term represents the time dependency and
the second the viscous eects. The non-linear term represents the eect of
convection and pressure. We can see that the mode k interacts with the modes
p = k and q = (k k ) such that k + p = q. This triadic nature of the
non-linear interactions is intrinsically related to the mathematical structure
of the NavierStokes equations.
A.3.3 Spectral Equations
The equations for the spectral tensor components ij are obtained by applying an inverse Fourier transform to the transport equations of the two-point
double correlations. After computation, we get:
ij
ij
lm kl
t
km
where:
ilj
=
=
ij
+ il lj + jl il + 2k 2 ij =

+ ki j + kj j
kl ilj + kl jli
ui (x)ul (x)uj (x + r)ekn rn d3 r

(2)3

1
p (x)uj (x + r)ekn rn d3 r
(2)3
kl
kl km
2lm 2 mj 2 mlj ,
k
k
ui
.
xj
, (A.51)
(A.52)
(A.53)
(A.54)
(A.55)
503
By expanding the terms (A.52) and (A.54), equation (A.51) takes the
form:

ui
uj
+ 2k 2 ij (k) +
jl (k) +
il (k)
t
xl
xl
ul
2
(ki jm (k) + kj mi (k))
xm
ul
(kl ij (k))
xm km
= Pil (k)Tlj (k) + Pjl (k)Tli (k) ,
(A.56)

where
ui (k)ul (p)uj (k p)d3 p
Tij (k) = kl
(A.57)
The evolution equation for the energy spectrum E(k), derived from (A.51)
by integration over the sphere of radius k, is:
E(k)
= P (k) + T (k) + D(k) ,
(A.58)
t
where the kinetic energy production term P (k) by interaction with the mean
eld, the transfer term T (k) and the dissipation term D(k) are given by:
P (k) =
T (k) =
D(k) =
ij ij (k) ,

(kl ii )
1
) + lm
dS(k)
kl (ili + ili
2
km
(A.59)
,
2k 2 E(k) ,
(A.60)
(A.61)
where the tensor ij (k) is dened as the integral of ij (k) over the sphere of
radius k:

ij (k) = ij (k)dS(k) .
(A.62)
The kinetic energy conservation property for ideal uid is expressed by:

T (k)dk = 0 .
(A.63)
0
We come up with the kinetic energy evolution equation in the physical

space by integrating (A.58) over the entire spectrum:

K
E(k)
=
dk =
P (k)dk +
T (k)dk +
D(k)dk . (A.64)
t
t
0
0
0
0
In the isotropic homogeneous case, production is zero and we get:
K
= ,
t
where the kinetic energy dissipation rate is given by:

2k 2 E(k)dk .
=
0
(A.65)
(A.66)
504
A.4 Characteristic Scales of Turbulence

Several characteristic scales of turbulence can be dened. We dene the integral scale Llij as:
+
Rij (r)dr .
(A.67)
Llij =
This scale is representative of the length over which the turbulent uctuations are mutually correlated, so it is directly related to the size of the
structures that form the turbulent eld. Another scale, called the Taylor microscale, is denoted and is dened as:
0
2 1
2
u
u
=
.
(A.68)
2
x
While the rst scale is associated with all the turbulent structures, this
latter scale is related directly to the small scales of the turbulence. Considering that the dissipation , can be written as:
0
2 1
u
,
(A.69)
= 2
x
we get the relation:
u
2
2
= 2
(A.70)
The Taylor microscale thus appears as characteristic of the dissipative

phenomena.
A.5 Spectral Dynamics

of Isotropic Homogeneous Turbulence
A.5.1 Energy Cascade and Local Isotropy
Analyses on the basis of (A.58) show that the dynamical mechanism associated with the term T (k) is a kinetic energy transfer from the small wave
numbers to the large. This process is called the energy cascade. It is relatively local in frequency: the transfers are negligible among wave numbers
separated by more than two decades. It repeats itself until such time as the
structures are so small that the viscous mechanisms, represented by the D(k)
term, become preponderant.
The local isotropy hypothesis formalized by Kolmogorov assumes statistical homogeneity and isotropy in a small spacetime domain and not throughout the ow. This is equivalent to the hypothesis that the ow scales, while
A.5 Spectral Dynamics of Isotropic Homogeneous Turbulence
505
suciently small, are governed by a dynamics similar to that of isotropic

homogeneous turbulence. They are thus independent of the large scales and
their statistical structure acquires a universal character.
Kolmogorovs rst hypothesis is that the statistical moments of the scales
located in such a domain depend only on the separation distance r, the total
dissipation by viscosity per unit mass , and the viscosity .
The second hypothesis is that the statistical moments for large separation
distances become independent of the viscosity and are no longer a function
of r and .
A.5.2 Equilibrium Spectrum
With the hypothesis of local isotropy, we nd three distinct regions in the
energy spectrum E(k):
the large scale region where the turbulence associated with the P (k) term is
produced. These scales are coupled with the mean eld and are aected by
the boundary conditions, so they possess no universal character. However,
following arguments related to the nite character of the energy spectral
density A(k), we can say that:
E(k) k 4 or E(k) k 2 for k 1
(A.71)
the inertial range, associated with the intermediate scales, in which the
energy is transferred by non-linear interaction with no action by viscosity or
production. The energy spectrum depends only on k and . Since the energy
is transferred without loss, remains constant. Assuming that there exists
a self-similar form of the power-law spectrum, by dimensional arguments
we get:
(A.72)
E(k) = K0 2/3 k 5/3 ,
where the constant K0 , called the Kolmogorov constant, is close to 1.5.
The dissipation region, which comprises the smallest scales where the kinetic energy is dissipated by the viscous eects. In this area, the relaxation
time d associated with the viscous eects is at least equal to that of the
non-linear transfers, denoted c . For a length scale l, these two times are
evaluated as:
d l2 /, c 2 / .
(A.73)
The dissipation region is characterized by the relation:

d c = l 3 / .
(A.74)
We call the Kolmogorov scale, which is denoted K , the scale for which
these two times are equal and which marks the beginning of the dissipation
region:

(A.75)
K = 3 / .
506
The characteristic velocity associated with this scale is:

1/4
vK = ()
(A.76)
The energy spectrum depends explicitly only on k, and , or equivalently

on k, the Kolmogorov scales K , and vK . The dimensional arguments do not
lead to a unique form E(k), and several solutions have been proposed. Arguments concerning the regularity of the velocity eld and these gradients
suggest an exponential decay of E(k) in this region.
B. EDQNM Modeling
The EDQNM model is briey described here in its isotropic and anisotropic
versions. For more details concerning the isotropic version, the reader may
refer to the work of Lesieur [439].
B.1 Isotropic EDQNM Model

Starting with the NavierStokes equations written in symbolic form:

2
+ k u = uu ,
t
we derive an innite hierarchical set of evolution
statistical moments of the velocity u:

+ k 2 uu =
t

+ (k 2 + p2 + q 2 ) uuu =
t
... =
(B.1)
equations as usual for the

uuu ,
(B.2)
uuuu ,
(B.3)
...
where the symbol designates a statistical average. We then adopt the

following hypothesis.
Hypothesis B.1 (Quasi-normality Hypothesis) The velocity distribution law is close to the Gaussian bell curve and its fourth-order cumulant,
denoted uuuuc is zero.
The evolution equation of the triple correlations then become:

2
2
2
+ (k + p + q ) uuu =
uuuu .
t
(B.4)
The quasi-normal approximation does not provide the realizability condition, i.e. the spectrum E(k) can take negative values. To recover this property,
Orszag proposes introducing a triple-correlation damping term and then gets:

+ (k 2 +p2 +q 2 ) uuu =
uuuu (k +p +q )uuu . (B.5)
t
508
B. EDQNM Modeling
The solution to this equation is:

t
uuu(t) =
uuuue(k +p +q )t dt ,
(B.6)
with
k = k + k 2
(B.7)
To get a solution that is easier to calculate, we adopt the following hypothesis:

Hypothesis B.2 The relaxation time of the triple correlations is small compared with the relaxation time of the double correlations (which is also that
of the energy spectrum).
With this hypothesis, we can Markovianize equation (B.6), which leads
to:
uuu(t) =
=

uuuu
e(k +p +q )t dt
uuuu
1 e(k +p +q )t
k + p + q
(B.8)
This relation closes the derivation equation of the second-order moments.

This closure is equivalent to replacing the solution of the NavierStokes equations with that of the following Langevin-type stochastic model:

2
+ ( + (k, t))k u = f (k, t) ,
(B.9)
t
in which
(k, t) =
1
2

kpq (t)
p
bkpq E(q, t)dpdq
kq
(B.10)
The forcing term f is such that:

t
F (k, t) 4k 2
f (k, t)f (k, s)|k|=cste ds
0

k3
=
kpq (t) akpq E(p, t)E(q, t)dpdq
pq
(B.11)
where akpq and bkpq are coecients linked to the geometry of the triad
(k, p, q), dened as:
akpq =
1
p
(1 xyz 2y 2 z 2 ), bkpq = (xy + z 3 )
2
k
(B.12)
B.2 Cambons Anisotropic EDQNM Model
509
where x, y and z are the cosines of the angles of the triangle formed by the
wave vectors (k, p, q), opposed respectively to k, p and q. The relaxation
time kpq (t) is evaluated as:
kpq (t) =
1 e(k +p +q )t
k + p + q
where the damping factor k is chosen as follows:

'
k
2
k = k + 0.36
p2 E(p)dp
(B.13)
(B.14)
B.2 Cambons Anisotropic EDQNM Model

To study anisotropic homogeneous ows, we dene the following two spectral
tensors [96] (see Appendix A):
ij (k)
u
uj (k) ,
i (k)

ij (k) = ij (k)dA(k) .
(B.15)
(B.16)
The evolution equation for the quantities ij (k) in the presence of mean
velocity gradients is:

ui
uj
2
+ 2k ij (k) =
jl (k)
il (k)
t
xk
xl
+
l
Pijl (k) + Sij
(k)
nl
Pijnl (k) + Sij
(k)
(B.17)
where
ul
xm
kl
[ki mj (k) + kj mi (k)] dA(k) ,
k2

ul
l
Sij
(k) =
(kl ij (k)) dA(k) ,
xm
km

kl
[ki Tlj (k) + kj Tli (k)] dA(k) ,
Pijnl (k) =
k2

3
4
nl
Sij
(k) =
Tij (k) + Tij (k) dA(k) .
Pijl (k) = 2
(B.18)
(B.19)
(B.20)
(B.21)
510
B. EDQNM Modeling
Equation (B.17) is closed by replacing (k) with a modeled form as a function of (k), where the direction of k is controlled in the integrals (B.18)
to (B.21):

2
Pijl (k) = 2E(k) Sij
5

2
3D Slj Hli (k) + Sli Hlj (k) ij Slm Hlm (k)
3

14
4
,
(B.22)
+
D+
(il Hlj (k) + jl Hli (k))
3
7
l
Sij
(k) =
2
Sij
(kE(k)) + 2Sil
(kDE(k)Hjl (k))
15
k
k
(kDE(k)Hil (k))
2Sjl
k
1
ij Slm
([2 + 11D] kE(k)Hlm (k)) ,
(B.23)
3
k

Pijnl (k)
=
+
"

2
(x + yz)Hij (q) k 1 pE(p)E(q) y(z 2 y 2 )(a(q) + 3)
pq
#
a(q)
(y + xz)
p3 E(k)E(q)y(z 2 x2 )(a(q) + 3) dpdq , (B.24)
5
kpq

1
2
3
2
ij + Hij (p)
kpq
(xy + z ) k pE(p)E(q)
pq
3
7

7
1
3
+Hij (q) p E(k)E(q)
ij + Hij (k) + Hij (q)
3

2

3
Hij (q) k pE(p)E(q)ckpq p E(k)E(q)cpkq dpdq , (B.25)

nl
Sij
(k)
+
where
S =

1
u + T u ,
2
=

1
u T u
2
and x, y and z are the cosines of the interior angles opposite the wave vectors k, p and q, respectively, in the triangle formed by these vectors. The
anisotropy parameter a(k) is optimized by the Rapid Distortion Theory. The
factor D is dened as:

2
4
.
(B.26)
D=
1+ a
7
5
B.3 EDQNM Model for Isotropic Passive Scalar
511
The energy and anisotropy spectra, denoted respectively E(k) and Hij (k),
are given by the relations:
E(k) =
Hij (k) =
1
ll (k)
2
ij (k) 1
ij
2E(k) 3
(B.27)
.
(B.28)
The geometric factor ckpq is dened as:

1
2
ckpq = (xy + z) (y 2 z 2 )(a(q) + 3) + a(q)(1 + z 2 )
2
5
(B.29)
The relaxation time kpq (t) is evaluated as:

kpq (t) =
1 e(k +p +q )t
k + p + q
(B.30)
where the damping term k is:

1/2
p E(p)dp + ij ij
k = k + 0, 36
(B.31)
It should be noted that fully anisotropic versions, which call for no angular
parameter setting, have been proposed and compared with simulations for the
case of pure rotation and stable stratication [97].
B.3 EDQNM Model for Isotropic Passive Scalar

The transport equation for the passive scalar (15.1) can be rewritten in the
Fourier space under the following symbolic form:

+ k 2 = u ,
(B.32)
t
leading to a hierarchy of symbolic equtions for the statistical moments:

+ (k 2 + p2 ) u =
t

+ [(k 2 + p2 ) + q 2 ] u =
t
... =

u
uu
...
(B.33)
,
(B.34)
(B.35)
512
B. EDQNM Modeling
The EDQNM procedure can be applied to this set of equations to obtain

a simple, closed equation for the temperature spectrum E (k). Details will
not be given here, since the process that yields the closed EDQNM model
involves the same steps as in the isotropic model for the velocity moments
discussed above. The reader can refer to Lesieurs book [439] for a detailed
presentation. The model is expressed as

k
+ 2k 2 E (k) =
(1 y 2 )E(q)[k 2 E (p) p2 E (k)]
kpq
t
pq
(k p q)dpdq
(B.36)
where the triad geometrical parameter is the same as in above sections, and
the characteristic time is dened as
2
kpq
1 e[(k +p )+q +1 (k)+1 (p)+2 (q)]t

=
(k 2 + p2 ) + q 2 + 1 (k) + 1 (p) + 2 (q)
(B.37)
where
'

p2 E(p)dp,
i (k) = ai
i = 1, 2 .
(B.38)
The two constants a1 and a2 are not uniquely dened, and several choices
are possible. Setting a1 = 0 and a2 = 1.3 enables the solve analytically the
system and yields a satisfactory prediction from a physical point of view.
3/2
Another possibility is to take a1 = a2 = 0.218K0 .
Anisotropic passive scalar models [303] and models for stably stratied
ows [266, 689, 267] are very complex and will not be discussed here. This is
also the case of EDQNM models for mixed scalars [723].
References
1. Abba, A., Bucci, R., Cercignani, C., Valdettaro, L. (1996): A new approach
to the dynamic subgrid scale model. Unpublished
2. Adams, N. (2001): The role of deconvolution and numerical discretization
in subgrid-scale modeling. (Direct and large-eddy simulation IV, Geurts,
Friedrich and Metais eds.), Kluwer, 311320
3. Adams, N.A., Leonard, A. (1999): Deconvolution of subgrid-scales for the
simulation of shock-turbulence interaction. (Direct and Large Eddy Simulation
III, Voke, Sandham and Kleiser eds.) Kluwer, 201212
4. Adams, N.A., Stolz, S. (2001): Deconvolution methods for subgrid-scale approximation in LES. (Modern simulation strategies for turbulent ow, Geurts
ed.), Edwards, 2144
5. Addad, Y., Laurence, D., Talotte, C., Jacob, M.C. (2003): Large eddy simulation of a forward facing step for acoustic source identication. Int. J. Heat
Fluid Flow 24, 562571
6. Adrian, J. (1990): Stochastic estimation of subgrid-scale motions. Appl. Mech.
Rev. 43(5), 214218
7. Agee, E., Gluhovsky, A. (1999): LES model sensitivities to domains, grids,
and large-eddy timescales. J. Atmos. Sci. 56, 599604
8. Akhavan, R., Ansari, A., Kang, S., Mangiavacchi, N. (2000): Subgrid-scale
interactions in a numerically simulated planar turbulent jet an implications
for modelling. J. Fluid Mech. 408, 83120
9. Akselvoll, K., Moin, P. (1996): Large-eddy simulation of turbulent conned
coannular jets. J. Fluid Mech. 315, 387411
10. Allen, R., Mendonca, F. (2004): DES validations of cavity acoustics over the
subsonic to supersonic range. AIAA Paper 2004-2862
11. Alvelius, K., Johansson, A. (2000): LES computations and comparison with
Kolmogorov theory for two-point pressure-velocity correlations and structure
functions for globally anisotropic turbulence. J. Fluid Mech. 403, 2336
12. Anderson, R., Meneveau, C. (1999): Eects of the similarity model in nitedierence LES of isotropic turbulence using a Lagrangian dynamic mixed
model. Flow, Turbulence and Combustion 62, 201225
13. Anitescu, M., Layton, W. (2002): Uncertainties in large-eddy simulation and
improved estimates of turbulent ow functionals. Unpublished
14. Antonopoulos-Domis, M. (1981): Aspects of large-eddy simulation of homogeneous isotropic turbulence. Int. J. Numer. Meth. Fluids 1, 273290
15. Aprovitola, A., Denaro, F. (2004): On the application of congruent upwind
discretizations for large eddy simulations. J. Comput. Phys. 194, 329343
16. Arad, E. (2001): Analysis of boundary layer separation over a bump using
large-eddy simulation. AIAA Paper 2001-2558
514
References
17. Arandiga, F., Donat, R., Harten, A. (1999): Multiresolution based on weighted
averages of the hat function II: nonlinear reconstruction techniques. SIAM J.
Sci. Comput. 20(3), 10531093
18. Armenio, V., Piomelli, U. (2000): A Lagrangian mixed subgrid-scale model in
generalized coordinates. Flow, Turbulence and Combustion 65, 5181
19. Armenio, V., Sarkar, S. (2002): An investigation of stably stratied turbulent
channel ow using large-eddy simulation. J. Fluid Mech. 459, 142
20. Arunajatesan, S., Sinha, N. (2000): Towards hybrid LES-RANS computations
of cavity owleds. AIAA Paper 2000-0401
21. Arunajatesan, S., Sinha, N. (2001): Unied unsteady RANS-LES simulations
of cavity owelds. AIAA Paper 2001-0516
22. Arunajatesan, S., Sinha, N., Ukeiley, L. (2001): On the application of hybrid
RANS-LES and Proper Orthogonal Decomposition techniques to control of
cavity ows. (DNS/LES Progress and Challenges, Liu, Sakell and Beutner
eds.), Greyden Press, 673688
23. Aupoix, B. (1984): Eddy viscosity subgrid models for homogeneous turbulence. (Proceedings of Macroscopic modelling of turbulent ows, Lecture
notes in physics, Vol. 230), Springer-Verlag, 4564
24. Aupoix, B. (1985): Subgrid scale models for homogeneous anisotropic turbulence. (Proceedings of the Euromech Colloquium 199, Direct and large
eddy simulation of turbulence, Notes on numerical uid mechanics, Vol. 15)
Vieweg, 3666
25. Aupoix, B. (1989): Application of two-point closures to subgrid scale modelling
for homogeneous 3D turbulence. (Turbulent shear ows, von Karman Institute
for Fluid Dynamics, Lecture Series 1989-03)
26. Aupoix, B., Cousteix, J. (1982): Modèles simples de tensions de sous-maille
en turbulence homogène isotrope (in french). Rech. Aero. 4, 273283
27. Avancha, R., Pletcher, R. (2002): Large eddy simulation of the ow past
a backward facing step with heat transfer and property variations. Int. J.
Heat Fluid Flow 23, 601614
28. Baastians, R., Rindt, C., Nieuwstadt, F., van Steenhoven, A. (2000): Direct
and large-eddy simulation of the transition of two- and three-dimensional
plane plumes in a conned enclosure. Int. J. Heat Mass Transfer 43, 23752393
29. Baastians, R., Rindt, C., van Steenhoven, A. (1998): Experimental analysis of
a conned transitional plume with respect to subgrid-scale modeling. Int. J.
Heat Mass Transfer 41, 39894007
30. Bagget, J.S. (1997): Some modeling requirements for wall models in largeeddy simulation. Annual Research Briefs Center for Turbulence Research,
123134
31. Bagget, J.S. (1998): On the feasability of merging LES with RANS for the
near-wall region of attached turbulent ows. Annual Research Briefs Center
for Turbulence Research, 267277
32. Bagget, J.S., Jimenez, J., Kravchenko, A.G. (1997): Resolution requirements
in large-eddy simulations of shear ows. Annual Research Briefs Center for
Turbulence Research, 5166
33. Bagget, J.S., Nicoud, F., Mohammadi, B., Bewley, T., Gullbrand, J., Botella, O. (2000): Sub-optimal control based wall models for LES including
transpiration velocity. Proceedings of the Summer Program Center for Turbulence Research, 331342
34. Bagwell, T.G., Adrian, R.J., Moser, R.D., Kim, J. (1993): Improved approximation of wall shear stress boundary conditions for large-eddy simulation.
(Near-wall turbulent ows, So, Speziale and Launder eds.), Elsevier
References
515
35. Balaras, E. (2003): Modeling complex boundaries using an external force eld
on xed Cartesian grids in large-eddy simulations. Computers and Fluids 33,
375404
36. Balaras, E., Benocci, C., Piomelli, U. (1995): Finite-dierence computations of
high Reynolds number ows using the dynamic subgrid-scale model. Theoret.
Comput. Fluid Dynamics 7, 207216
37. Balaras, E., Benocci, C., Piomelli, U. (1996): Two-layer approximate boundary
conditions for large-eddy simulations. AIAA Journal 34(6), 11111119
38. Balaras, E., Piomelli, U., Wallace, J.M. (2001): Self-similar states in turbulent
mixing layers. J. Fluid Mech. 446, 124
39. Bardina, J., Ferziger, J.H., Reynolds, W.C. (1980): Improved subgrid scale
models for large eddy simulation. AIAA Paper 80-1357
40. Bardina, J., Ferziger, J.H., Reynolds, W.C. (1983): Improved turbulence models based on large eddy simulation of homogeneous, incompressible, turbulent
ows. Report TF-19, Thermosciences Division, Dept. Mechanical Engineering,
Stanford University
41. Barenblatt, G.I. (1996): Scaling, self-similarity, and intermediate asymptotics.
Cambridge University Press, UK
42. Basdevant, C., Lesieur, M., Sadourny, R. (1978): Subgrid-scale modeling of enstrophy transfer in two-dimensional turbulence. J. Atmos. Sci. 35, 10281042
43. Basdevant, C., Sadourny, R. (1983): Modelisation des echelles virtuelles dans
la modelisation numerique des ecoulements turbulents bidimensionnels (in
french). J. Mech. Theor. Appl., numero special, 243269
44. Bastin, F., Lafon, P., Candel, S. (1997): Computation of jet mixing noise due
to coherent structures: the plane jet case. J. Fluid Mech. 335, 261304
45. Batchelor, G.K. (1953): The theory of homogeneous turbulence. Cambridge
University Press
46. Batten, P., Goldberg, U., Chakravarthy, S. (2000): Sub-grid turbulence moeling for unsteady ow with acoustic resonance. AIAA Paper 2000-0473
47. Batten, P., Goldberg, U., Chakravarthy, S. (2002): Reconstructed subgrid
methods for acoustics predictions at all Reynolds numbers. AIAA Paper 20022511
48. Batten, P., Goldberg, U., Chakravarthy, S. (2003): Using synthetic turbulence
to interface RANS and LES. AIAA Paper 2003-0081
49. Batten, P., Goldberg, U., Chakravarthy, S. (2004): Interfacing statistical turbulence closures with large-eddy simulation. AIAA Journal 42(3), 485492
50. Baurle, R.A., Tam, C.J., Edwards, J.R., Hassan, H.A. (2003): Hybrid simulation approach for cavity ows: blending, algorithm, and boundary treatment
issues. AIAA Journal 41(8), 14631480
51. Beaubert, F., Viazzo, S. (2003): Large eddy simulations of plane turbulent
impinging jets at moderate Reynolds numbers. Int. J. Heat Fluid Flow 24,
512519
52. Beaudan, P., Moin, P. (1994): Numerical experiments on the ow past a circular cylinder at subcritical Reynolds number. Report TF-62, Thermosciences
Division, Dept. Mechanical Engineering, Stanford University
53. Benzi, R., Biferale, L., Succi, S., Toschi, F. (1999): Intermittency and eddy
viscosities in dynamical models of turbulence. Phys. Fluids 11(5), 12211228
54. Berkooz, G. (1994): An observation on probability density equations, or, when
do simulations reproduce statistics? Nonlinearity 7, 313328
55. Berkooz, G., Holmes, P., Lumley, J.L. (1993): The proper othogonal decomposition in the analysis of turbulent ows. Ann. Rev. Fluid Mech. 25, 539575
516
References
56. Bertoglio, J.P. (1984): A stochastic subgrid model for sheared turbulence.
(Proceedings of Macroscopic modelling of turbulent ows, Lecture notes in
physics, Vol. 230) Springer-Verlag, 100119
57. Bertoglio, J.P., Mathieu, J. (1984): A stochastic subgrid model for large-eddy
simulation: general formulation. C. R. Acad. Sc. Paris 299, serie II, (12),
751754
58. Bertoglio, J.P., Mathieu, J. (1984): A stochastic subgrid model for large-eddy
simulation: generation of a stochastic process. C. R. Acad. Sc. Paris 299, serie
II, (13), 835838
59. Biringen, S., Reynolds, W.C. (1981): Large-eddy simulation of the shear-free
turbulent boundary-layer. J. Fluid Mech. 103, 5363
60. Blackburn, H., Schmidt, S. (2003): Spectral element ltering techniques for
large eddy simulation with dynamic estimation. J. Comput. Phys. 186, 610
629
61. Blin, L., Hadjadj, A., Vervisch, L. (2003): Large eddy simulation of turbulent
ows in reversing systems. JOT, 119
62. Boersma, B.J., Kooper, M.N., Nieuwstadt, F.T.M., Wesseling, P. (1997): Local
grid renement in large-eddy simulation. J. Engng. Math.32, 161175
63. Bogey, C., Bailly, C. (2004): A family of low dispersive and low dissipative
explicit schemes for ow and noise computations. J. Comput. Phys. 194, 194
214
64. Bonnet, J.P., Delville, J., Druault, P., Sagaut, P., Grohens, R. (1997): Linear stochastic estimation of LES inow conditions. (Advances in DNS/LES,
C. Liu, Z. Liu (eds.)) Greyden Press, 341348
65. Booij, R. (2003): Measurements and large eddy simulations of the ows in
some curved umes. JOT 4, 117
66. Boris, J.P., Book, D.L. (1973): Flux-Corrected Transport. I. SHASTA, a uid
transport algorithm that works. J. Comput. Phys. 11, 3869
67. Boris, J.P., Grinstein, F.F., Oran, E.S., Kolbe, R.L. (1992): New insights into
large-eddy simulation. Fluid Dyn. Res. 10, 199228
68. Borue, V., Orszag, S.A. (1995): Forced three-dimensional homogeneous turbulence with hyperviscosity. Europhys. Lett. 29, 687692
69. Borue, V., Orszag, S.A. (1995): Self-similar decay of three-dimensional turbulence with hyperviscosity. Phys. Rev. E. 52, R856859
70. Borue, V., Orszag, S.A. (1996): Numerical study of Kolmogorov ow at high
Reynolds numbers. J. Fluid Mech. 306, 293323
71. Borue, V., Orszag, S.A. (1996): Kolmogorovs rened similarity hypothesis for
hyperviscous turbulence. Phys. Rev. E. 366, R2124
72. Borue, V., Orszag, S.A. (1998): Local energy ux and subgrid-scale statistics
in three-dimensional turbulence. J. Fluid Mech. 366, 131
73. Bouchon, F., Dubois, T. (1999): Incremental Unknowns: a tool for large-eddy
simulations ?. (Direct and Large Eddy Simulation III, Voke, Sandham and
Kleiser eds) Kluwer, 275286
74. Brasseur, J.G., Wei, C.H. (1994): Interscale dynamics and local isotropy in
high Reynolds number turbulence within triadic interactions. Phys. Fluids A
6(2), 842870
75. Breuer, M. (1998): Large eddy simulation of the subcritical ow past a circular
cylinder: numerical and modeling aspects. Int. J. Numer. Meth. Fluids 28,
12811302
76. Breuer, M. (1998): Numerical and modeling inuences on large eddy simulations for the ow past a circular cylinder. Int. J. Heat Fluid Flow 19, 512521
77. Breuer, M. (2000): A challenging test case for large eddy simulation: high
Reynolds number circular cylinder ow. Int. J. Heat Fluid Flow 21, 648654
References
517
78. Breuer, M., Jovicic, N., Mazaev, K. (2003): Comparison of DES, RANS and
LES for the separated ow around a at plate at high incidence. Int. J. Numer.
Meth. Fluids 41, 357388
79. Brezzi, F., Houston, P., Marini, D., S
uli, E. (2000): Modeling subgrid viscosity
for advection-diusion problems. Comput. Methods Appl. Mech. Engrg. 190,
16011610
80. Brown, R.M., Perry, P., Shen, Z. (1998): The additive turbulent decomposition
for the two-dimensional incompressible NavierStokes equations: convergence
theorems and error estimates. SIAM J. Appl. Math. 59(1), 139155
81. Brun, C., Friedrich, R. (2001): Modeling the test SGS tensor Tij : an issue in
the dynamic approach. Phys. Fluids 13(8), 23732385
82. Brun, C., Friedrich, R. (2001): The spatial velocity increment as a tool for
SGS modeling. (Modern simulation strategies for turbulent ow, Geurts (ed.)),
Edwards, 5784
83. Brun, C., Kessler, P., Comte, P., Lesieur, M. (1997): Simulation des grandes
echelles de jets ronds (in french). Rapport de synthèse, Contrat DGA/DRET
95-2557 A
84. Bush, R.H., Mani, M. (2001): A two-equation large eddy stress model for high
subgrid shear. AIAA Paper 2001-2561
85. Cabot, W. (1992): Large eddy simulations of time-dependent and buoyancydriven channel ows. Annual Research Briefs Center for Turbulence Research, 4560
86. Cabot, W. (1995): Large-eddy simulations with wall models. Annual Research
Briefs Center for Turbulence Research, 4149
87. Cabot, W. (1996): Near-wall models in large-eddy simulations of ow behind
a backward-facing step. Annual Research Briefs Center for Turbulence Research, 199210
88. Cabot, W. (1998): Large-eddy simulation of a separated boundary layer. Annual Research Briefs Center for Turbulence Research, 279288
89. Cabot, W., Jimenez, J., Bagget, J.S. (1999): On wakes and near-wall behavior in coarse large-eddy simulation of channel ow with wall models and
second-order nite-dierence methods. Annual Research Briefs Center for
90. Cabot, W., Moin, P. (1999): Approximate wall boundary conditions in the
large-eddy simulation of high Reynolds number ow. Flow, Turbulence and
Combustion 63, 269291
91. Calgaro, C., Debussche, A., Laminie, J. (1998): On a multilevel approach
for the two-dimensional NavierStokes equations with nite elements. Int. J.
Numer. Meth. Fluids 27, 241258
92. Calmet, I., Magnaudet, J. (1997): Large-eddy simulation of high-Schmidt
number mass transfer in a turbulent channel ow. Phys. Fluids 9(2), 438455
93. Calmet, I., Magnaudet, J. (1998): High-Schmidt number mass transfer through
turbulent gas-liquid interfaces. Int. J. Heat Fluid Flow 19, 522532
94. Camarri, S., Salvetti, M., Koobus, B., Dervieux, A. (2002): Large-eddy simulation of a blu-body ow on unstructurd grids. Int. J. Numer. Meth. Fluids
40, 14311460
95. Camarri, S., Salvetti, M., Koobus, B., Dervieux, A. (2004): A low-diusion
MUSCL scheme for LES on unstructured grids. Computers and Fluids 33,
11011129
96. Cambon, C., Jeandel, D., Mathieu, J. (1981): Spectral modelling of homogeneous non-isotropic turbulence. J. Fluid Mech. 104, 247262
97. Cambon, C., Mansour, N.N., Godeferd, F.S. (1997): Energy transfer in rotating turbulence. J. Fluid Mech. 337, 303332
518
References
98. Cantekin, M.E., Westerink, J.J., Luettich, R.A. (1994): Low and moderate
Reynolds number transient ow simulations using space ltered NavierStokes
equations. Numer. Meth. Partial Di. Eq. 10, 491524
99. Canuto, V.M., Cheng, Y. (1997): Determination of the Smagorinsky-Lilly constant CS . Phys. Fluids 9(5), 13681378
100. Carati, D., Cabot, W. (1996): Anisotropic eddy viscosity models. Proceedings
of the Summer Program Center for Turbulence Research, 249259
101. Carati, D., Ghosal, S., Moin, P. (1995): On the representation of backscatter
in dynamic localization models. Phys. Fluids 7(3), 606616
102. Carati, D., Rogers, M.M. (1998): Ensemble-averaged LES of time-evolving
plane wake. Proceedings of the Summer Program Center for Turbulence
Research, 325336
103. Carati, D., Rogers, M., Wray, A. (2002): Statistical ensemble of large-eddy
simulations. J. Fluid Mech. 455, 195212
104. Carati, D., Vanden Eijnden, E. (1997): On the self-similarity assumption in
dynamic models for large-eddy simulations. Phys. Fluids 9 (7), 21652167
105. Carati, D., Winckelmans, G.S., Jeanmart, H. (1999): Exact expansions for
ltered-scales modelling with a wide class of LES lters. (Direct and Large
Eddy Simulation III, Voke, Sandham and Kleiser eds) Kluwer, 213224
106. Carati, D., Winckelmans, G., Jeanmart, H. (2001): On the modelling of the
subgrid-scale and ltered-scale stress tensors in large-eddy simulation. J. Fluid
Mech. 441, 119138
107. Carati, D., Wray, A. (2000): Time ltering in large-eddy simulations. Proceedings of the Summer Program Center for Turbulence Research, 263270
108. Carati, D., Wray, A., Cabot, W. (1996): Ensemble-averaged dynamic modeling. Proceedings of the Summer Program Center for Turbulence Research,
237248
109. Carlotti, P. (2002): Two-point properties of atmospheric turbulence very close
to the ground: comparison of a high resolution LES with theoretical models.
Boundary-Layer Meteorol. 104, 381410
110. Carnevale, G., Briscolini, M., Orlandi, P. (2001): Buoyancy- to inertial-range
transition in forced stratied turbulence. J. Fluid Mech. 427, 205239
111. Carruthers, D.J., Moeng, C.H. (1987): Waves in the overlying inversion of the
convective boundary layer. J. Atmos. Sci. 44(4), 18011808
112. Casalino, D., Boudet, J., Jacob, M.: A shear ow subgrid scale model for large
eddy simulations. Submitted
113. Catalano, P., Wang, M., Iaccarino, G., Moin, P. (2003): Numerical simulation
of the ow around a circular cylinder at high Reynolds numbers. Int. J. Heat
114. Cerutti, S., Meneveau, C. (1998): Intermittency and relative scaling of subgridscale energy dissipation in isotropic turbulence. Phys. Fluids 10 (4), 928937
115. Cerutti, S., Meneveau, C., Knio, O.M. (2000): Spectral and hyper eddy viscosity in high-Reynolds-number turbulence. J. Fluid Mech. 421, 307338
116. Cerutti, S., Meneveau, C. (2000): Statistics of ltered velocity in grid and
wake turbulence. Phys. Fluids 12(5), 11431165
117. Chakravarthy, V.K., Menon, S. (2001): Large-eddy simulation of turbulent
premixed ames in the amelet regime. Combust. Sci. and Tech. 162, 175
222
118. Chakravarthy, V.K., Menon, S. (2001): Subgrid modeling of turbulent premixed ames in the amelet regime. Flow, Turbulence and Combustion 65,
133161
119. Chapman, D.R. (1979): Computational aerodynamics development and outlook. AIAA Journal 17(12), 12931313
References
519
120. Chasnov, J.R. (1990): Simulation of the Kolmogorov inertial subrange using
an improved subgrid model. Phys. Fluids A 3(1), 188200
121. Chasnov, J. (1991): Simulation of the inertial-conductive subrange. Phys. Fluids A 3(5), 11641168
122. Chasnov, J. (1994): Similarity states of passive scalar transport in isotropic
turbulence. Phys. Fluids 6(2), 10361051
123. Chasnov, J. (1995): Similarity states of passive scalar transport in buoyancygenerated turbulence. Phys. Fluids 7(6), 14991506
124. Chasnov, J., Tse, K.L. (2001): Turbulent penetrative convection with an internal heat source. Fluid Dyn. Res. 28, 397421
125. Chatelain, A., Ducros, F., Metais, O. (2004): LES of turbulent heat transfer: proper convection numerical schemes for temperature transport. Int. J.
126. Chattopadhyay, H., Saha, S. (2003): Turbulent ow and heat transfer from
a slot jet impinging on a moving plate. Int. J. Heat Fluid Flow 24, 685697
127. Chen, S., Doolen, G.D. (1998): Lattice Boltzmann method for uid ows. Ann.
Rev. Fluid Mech. 30, 329364
128. Chen, M., Temam, R. (1991): Incremental unknowns for solving partial differential equations. Numer. Math. 59, 255271
129. Chester, S., Charlette, F., Meneveau, C. (2001): Dynamic model for LES without test ltering: quantifying the accuracy of Taylor series approximations.
Theoret. Comput. Fluid Dynamics 15, 165181
130. Chlond, A. (1998): Large-eddy simulation of contrails. J. Atmos. Sci. 55, 796
819
131. Choi, H., Moin, P. (1994): Eects of the computational time step on numerical
solutions of turbulent ow. J. Comput. Phys. 113, 14
132. Chollet, J.P. (1983): Two-point closures as a subgrid modelling for large-eddy
simulations. (Fourth Symposium on Turbulent Shear Flows, Karlsruhe, Allemagne)
133. Chollet, J.P. (1984): Spectral closures to derive a subgrid scale modelling for
large eddy simulation. (Proceedings of Macroscopic modelling of turbulent
ows, Lecture notes in physics, Vol. 230), Springer-Verlag, 161176
134. Chollet, J.P. (1984): Turbulence tridimensionnelle isotrope: modelisation
statistique des petites echelles et simulation numerique des grandes echelles
(in french). Thèse de Doctorat es-sciences, Grenoble, France
135. Chollet, J.P. (1992): LES and subgrid models for reactive ows and combustion. ERCOFTAC Summer School/Workshop on modelling turbulent ows,
Lyon, France
136. Chollet, J.P., Lesieur, M. (1981): Parametrization of small scales of threedimensional isotropic turbulence utilizing spectral closures. J. Atmos. Sci. 38,
27472757
137. Chorin, A., Kast, A., Kupferman, R. (1998): Optimal prediction of underresolved dynamics. Proc. Natl. Acad. Sci. USA, Applied Mathematics 95,
40944098
138. Chow, F.K., Moin, P. (2003): A further study of numerical errors in large-eddy
simulations. J. Comput. Phys. 184, 366380
139. Chung, Y.M., Sung, H.J. (1997): Comparative study of inow conditions for
spatially evolving simulation. AIAA Journal 35(2), 269274
140. Chyczewski, T., Morris, P., Long, L. (2000): Large-eddy simulation of
wall bounded ow using the nonlinear disturbance equations. AIAA Paper
2000-2007
141. Ciofalo, M. (1996): Investigation of ow and heat transfer in corrugated passagesII. Numerical simulations. Int. J. Heat Mass Transfer 39(1), 165192
520
References
142. Ciofalo, M., Collins, M.W. (1992): Large-eddy simulation of a turbulent ow

and heat transfer in plane and rib-roughened channels. Int. J. Numer. Meth.
Fluids 15, 453489
143. Clark, R.A., Ferziger, J.H., Reynolds, W.C. (1979): Evaluation of subgrid-scale
models using an accurately simulated turbulent ow. J. Fluid Mech. 91(1),
116
144. Codina, R. (2000): Stabilization of incompressibility and convection through
orthogonal sub-scales in nite elements methods. Comput. Methods Appl.
Mech. Engrg. 190, 15791599
145. Colella, P., Woodward, P.R. (1984): The Piecewise Parabolic Method (PPM)
for gas-dynamical simulations. J. Comput. Phys. 54, 174201
146. Constantinescu, G.S., Squires, K.D. (2000): LES and DES investigations of
turbulent ow over a sphere. AIAA Paper 2000-0540
147. Conway, S., Caraeni, D., Fuchs, L. (2000): Large eddy simulation of the ow
through the blade of a swirl generator. Int. J. Heat Fluid Flow 21, 664673
148. Cook, A.W. (1997): Determination of the constant coecient in the scale similarity models of turbulence. Phys. Fluids 9(5), 14851487
149. Cook, A. (2001): A consistent approach to large-eddy simulation using adaptive mesh renement. J. Comput. Phys. 154, 117133
150. Cousteix, J. (1989): Turbulence et couche limite (in french). CEPADUES
Editions, France
151. Cui, J., Patel, V., Lin, C.L. (2003): Large eddy simulation of turbulent ow
in a channel with rib roughness. Int. J. Heat Fluid Flow 24, 372388
152. Cui, A., Street, R.L. (2001): Large-eddy simulation of turbulent rotating convective ow development. J. Fluid Mech. 447, 5384
153. Cui, G., Zhou, H., Zhang, Z., Shao, L. (2004): A new dynamic subgrid eddy
viscosity model with application to turbulent channel ow. Phys. Fluids 16(8),
28352842
154. Cuxart, J., Bougeault, P., Redelsperger, J.L. (2000): A turbulence scheme
allowing for mesoscale and large-eddy simulations. Q.J.R. Meteorol. Soc. 126,
130
155. Cziesla, T., Biswas, G., Chattopadhyay, H., Mitra, N. (2001): Large eddy
simulation of ow and heat transfer in an impinging slot jet. Int. J. Heat
Fluid Flow 22, 500-508
156. Cziesla, T., Biswas, G., Mitra, N.K. (1999): Large eddy simulation in a turbulent channel ow using exit boundary conditions. Int. J. Numer. Meth. Fluids
30, 763773
157. da Silva, C., Metais, O. (2002): Vortex control of bifurcating jets: a numerical
study. Phys. Fluids 14(11), 37983818
158. da Silva, C., Metais, O. (2002): On the inuence of coherent structures upon
interscale interactions in turbulent plane jets. J. Fluid Mech. 473, 103145
159. Dai, Y., Kobayashi, T., Taniguchi, N. (1994): Large eddy simulation of plane
turbulent jet ow using a new outow velocity boundary condition. JSME
International Journal Series B 37(2), 242253
160. Dakhoul, Y.M., Bedford, K.W. (1986): Improved averaging method for turbulent ow simulation. Part 1: theoretical development and application to
Burgers transport equation. Int. J. Numer. Meth. Fluids 6, 4964
161. Dakhoul, Y.M., Bedford, K.W. (1986): Improved averaging method for turbulent ow simulation. Part 2: calculations and verication. Int. J. Numer.
162. Dang, K.T. (1985): Evaluation of simple subgrid-scale models for the numerical simulation of homogeneous turbulence. AIAA Journal 23(2), 221227
References
521
163. Dantinne, G., Jeanmart, H., Winckelmans, G.S., Legat, V. (1998): Hyperviscosity and vorticity-based models for subgrid scale modeling. Applied Scientic Research 59, 409420
164. Das, A., Moser, R.D. (2001): Filtering boundary conditions for LES and embedded boundary simulations. (DNS/LES Progress and Challenges, Liu, Sakell
and Beutner eds.), Greyden Press, 389396
165. Das, A., Moser, R. (2002): Optimal large-eddy simulation of forced burgers
equation. Phys. Fluids 14(12), 43444351
166. David, E. (1993): Modelisation des ecoulements compressibles et hypersoniques (in french). Thèse de Doctorat de lINPG, Grenoble, France
167. Davidson, L. (2001): Hybrid LES-RANS: a combination of a one-equation
SGS model and a K model for predicting recirculating ows. (Proceedings
of ECCOMAS CFD conference)
168. de Bruyn Kops, S., Riley, J. (2000): Re-examining the thermal mixing layer
with numerical simulations. Phys. Fluids. 12(1), 185192
169. De Stefano, G., Denaro, F.M., Riccardi, G. (1998): Analysis of 3D backwardfacing step incompressible ows via a local averaged-based numerical procedure. Int. J. Numer. Meth. Fluids 28, 10731091
170. De Stefano, G., Denaro, F.M., Riccardi, G. (2001): High-order ltering for
control volumes ow simulation. Int. J. Numer. Meth. Fluids 37(7), 797835
171. De Stefano, G., Vasilyev, O.V. (2002): Sharp cuto versus smooth ltering in
large eddy simulation. Phys. Fluids 14(1), 362369
172. Deardor, J.W. (1970): A numerical study of three-dimensional turbulent
channel ow at large Reynolds numbers. J. Fluid Mech. 41, 453465
173. Deardor, J.W. (1973): The use of subgrid transport equations in a threedimensional model of atmospheric turbulence. ASME J. Fluids Engng.,
429438
174. Debussche, A., Dubois, T., Temam, R. (1995): The nonlinear Galerkin
method: a multiscale method applied to simulation of homogeneous turbulent ows. Theoret. Comp. Fluid Dynamics 7, 279315
175. Denaro, F.M. (1996): Towards a new model-free simulation of high-Reynoldsnumber ows: local average direct numerical simulation. Int. J. Numer. Methods Fluids 23, 125142
176. Deschamps, V. (1988): Simulation numerique de la turbulence homogène incompressible dans un ecoulement de canal plan (in french). ONERA, Note
technique 1988-5
177. Desjardin, P.E., Frankel, S.H. (1996): Assessment of turbulent combustion
submodels using the linear eddy model. Combust. Flame 104, 343357
178. di Mare, L., Jones, W.P. (2003): LES of turbulent ow past a swept fence.
Int. J. Heat Fluid Flow 24, 606615
179. Ding, X., Tsang, T.T.H. (2001): Large eddy simulation of turbulent ows by
a least-squares nite element method. Int. J. Numer. Meth. Fluids 37, 297319
180. Diurno, G.V., Balaras, E., Piomelli, U. (2001): Wall-layer models for LES of
separated ows. (Modern simulation strategies for turbulent ow, Geurts ed.),
Edwards, 207222
181. Domaradzki, J.A., Adams, N.A. (2002): Direct modelling of subgrid scales of
turbulence in large eddy simulations. JOT 3(024), 119
182. Domaradzki, J.A., Holm, D.D. (2001): The NavierStokes-alpha model: LES
equations with nonlinear dispersion. (Modern simulation strategies for turbulent ow, Geurts ed.), Edwards, 107122
183. Domaradzki, J.A., Horiuti, K. (2001): Similarity modeling on an expanded
mesh applied to rotating turbulence. Phys. Fluids 13(11), 35103512
522
References
184. Domaradzki, J.A., Liu, W. (1995): Approximation of subgrid-scale energy

transfer based on the dynamics of the resolved scales of turbulence. Phys.
Fluids 7(8), 20252035
185. Domaradzki, J.A., Liu, W., Brachet, M.E. (1993): An analysis of subgrid-scale
interactions in numerically simulated isotropic turbulence. Phys. Fluids 5(7),
17471759
186. Domaradzki, J.A., Liu, W., H
artel, C., Kleiser, L. (1994): Energy transfer in
numerically simulated wall-bounded turbulent ows. Phys. Fluids 6(4), 1583
1599
187. Domaradzki, J.A., Loh, K.C. (1999): The subgrid-scale estimation model in
the physical space representation. Phys. Fluids 11 (8), 23302342
188. Domaradzki, J.A., Loh, K.C. (2002): Large eddy simulations using the subgridscale estimation model and truncated NavierStokes dynamics. Theoret. Comput. Fluid Dynamics 15, 421450
189. Domaradzki, J.A., Metcalfe, R.W., Rogallo, R.S., Riley, J.J. (1987): Analysis
of subgrid-scale eddy viscosity with the use of results from direct numerical
simulations. Phys. Rev. Letter 58(6), 546550
190. Domaradzki, J.A., Saiki, E.M. (1997): A subgrid-scale model based on the
estimation of unresolved scales of turbulence. Phys. Fluids 9 (7), 21482164
191. Domaradzki, J.A., Saiki, E. (1997): Backscatter models for large-eddy simulations. Theoret. Comput. Fluid Dynamics 9, 7583
192. Domaradzki, J.A., Xiao, Z., Smolarkiewicz, P. (2003): Eective eddy viscosities in implicit large eddy simulations of turbulent ows. Phys. Fluids 15(12),
38903893
193. Domaradzki, J.A., Yee, P.P. (2000): The subgrid-scale estimation model for
high Reynolds number turbulence. Phys. Fluids 12(1), 193196
194. Dommermuth, D., Rottman, J., Innis, G., Novikov, E. (2002): Numerical simulation of the wake of a towed sphere in a weakly stratied uid. J. Fluid
Mech. 473, 83101
195. Dong, Y.H., Lu, X.Y. (2004): Large eddy simulation of a thermally stratied
turbulent channel ow with temperature oscillation on the wall. Int. J. Heat
Mass Transfer 47, 21092122
196. Dong, Y.H., Lu, X.Y., Zhuang, L.X. (2003): Large eddy simulation of turbulent
channel ow with mass transfer at high Schmidt numbers. Int. J. Heat Mass
Transfer 46, 15291539
197. Dornbrack, A., Schumann, U. (1993): Numerical simulation of turbulent convective ow over a wavy terrain. Boundary-Layer Meteorol. 65, 323355
198. Dreeben, T., Kerstein, A. (2000): Simulation of vertical slot convection using
one-dimensional turbulence. Int. J. Heat Mass TRansfer 43, 38233834
199. Druault, P., Lardeau, S., Bonnet, J.P., Coiet, F., Delville, J., Lamballais, E.,
Largeau, J.F., Perret, L. (2004): Generation of three-dimensional turbulent
inlet conditions for large-eddy simulation. AIAA Journal 42(3), 447456
200. Dubois, T., Bouchon, F. (1998): Subgrid-scale models based on incremental
unknowns for large eddy simulations.. Annual Research Briefs Center for
201. Dubois, T., Jauberteau, F., Temam, R. (1999): Dynamic multilevel methods
and the numerical simulation of turbulence. Cambridge University Press
202. Dubois, T., Jauberteau, F., Zhou, Y. (1997): Inuences of subgrid scale dynamics on resolvable statistics in large eddy simulations. Physica D 100, 390
406
203. Dubrulle, B., Laval, J.P., Nazarenko, S., Kevlahan, N.K.R. (2001): A dynamic
sublter-scale model for plane parallel ows. Phys. Fluids 13(7), 20452064
References
523
204. Ducros, F. (1995): Simulations numeriques directes et des grandes echelles

de couches limites compressibles (in french). Thèse de Doctorat de lINPG,
Grenoble, France
205. Ducros, F., Comte, P., Lesieur, M. (1996): Large-eddy simulation of transition
to turbulence in a boundary layer developing spatially over a at plate. J. Fluid
Mech. 326, 136
206. Dunca, A., John, V., Layton, W.J. (2002): The commutation error of the
spaced averaged NavierStokes equations on a bounded domain. Unpublished
207. Dunca, A., John, V., Layton, W.J., Sahin, N. (2001): Numerical analysis of
large-eddy simulation. (DNS/LES Progress and Challenges, Liu, Sakell and
Beutner eds.), Greyden Press, 359364
208. Echekki, T., Kerstein, A., Dreeben, T., Chen, J.Y. (2001): One dimensional
turbulence simulation of turbulent jet diusion ames: model formulation and
illustrative applications. Combust. Flame 125, 10831105
209. Eidson, T. (1985): Numerical simulation of the turbulent RayleighBenard
problem using subgrid model. J. Fluid Mech. 158, 245268
210. Engquist, B., L
otstedt, P., Sj
ogreen, B. (1989): Nonlinear lters for ecient
shock computation. Math. Comput. 52(186), 509537
211. Fabignon, Y., Beddini, R.A., Lee, Y. (1995): Analytic evaluation of nite difference methods for compressible direct and large eddy simulations. ONERA,
TP 1995-128
212. Fasel, H.F., Seidel, J., Wernz, S. (2002): A methodology for simulations of
complex turbulent ows. J. Fluid Engng. 124, 933942
213. Fatica, M., Mittal, R. (1996): Progress in the large-eddy simulation of ow
through an asymmetric plane diuser. Annual Research Briefs Center for
214. Fatica, M., Orlandi, P., Verzicco, R. (1994): Direct and large-eddy simulations
of round jets. (Direct and Large Eddy Simulation I, Voke, C-hollet & Kleiser
eds.) Kluwer, 4961
215. Feiz, A.A., Ould-Rouis, M., Lauriat, G. (2003): Large eddy simulation of turbulent ow in a rotating pipe. Int. J. Heat Fluid Flow 24, 412420
216. Ferziger, J.H. (1977): Large eddy simulations of turbulent ows. AIAA Journal
15(9), 12611267
217. Ferziger, J.H. (1997): Large eddy simulation: an introduction and perspective. (New tools in turbulence modelling, O. Metais and J. Ferziger eds.) Les
editions de physique, Springer, 2948
218. Ferziger, J.H., Leslie, D.C. (1979): Large eddy simulation: A predictive approach to turbulent ow computation. (AIAA Comput. Fluid Conf., Williamsburg, USA)
219. Fischer, P., Iliescu, T. (2001): A 3D channel ow simulation at Re = 180
using a rational LES model. (DNS/LES Progress and Challenges, Liu, Sakell
and Beutner eds.), Greyden Press, 283290
220. Flohr, P., Vassilicos, J.C. (2000): A scalar subgrid model with ow structure
for large-eddy simulations of scalar variances. J. Fluid Mech. 407, 315349
221. Foias, C., Holm, D., Titi, E. (2001): The NavierStokes-alpha model of uid
turbulence. Physica D 152-153, 505519
222. Foias, C., Manley, O., Temam, R. (1987): Sur linteraction des petits et grands
tourbillons dans des ecoulements turbulents (in french). C.R. Acad. Sci. Paris
t. 305, serie I, 497500
223. Foias, C., Manley, O., Temam, R. (1988): Modelling of the interaction of small
and large eddies in two dimensional turbulent ows. MMAN 22(1), 93114
224. Foias, C., Manley, O., Temam, R. (1991): Approximate inertial manifolds and
eective viscosity in turbulent ows. Phys. Fluids A 3(5), 898911
524
References
225. Franke, J., Frank, W. (2001): Temporal commuation errors in large-eddy simulation. Z. Angew. Math. Mech. 81 S3, S467S468
226. Friedrich, R., Arnal, M. (1990): Analysing turbulent backward-facing step ow
with the lowpass-ltered NavierStokes equations. J. Wind Eng. Ind. Aerodyn.
35, 101128
227. Fureby, C., Alin, N., Wikstr
om, N., Menon, S., Svanstedt, N., Persson, L.
(2004): Large-eddy simulation of high-Reynolds-number wall-bounded ows.
AIAA Journal 42(3), 457468
228. Fureby, C., Grinstein, F.F. (1999): Monotonically Integrated large eddy simulation of free shear ows. AIAA Journal 37(5), 544556
229. Fureby, C., Grinstein, F.F. (2002): Large eddy simulation of high-Reynoldsnumber free and wall-bounded ows. J. Comput. Phys. 181, 6897
230. Fureby, C., Tabor, G. (1997): Mathematical and physical constraints on largeeddy simulations. Theoret. Comput. Fluid Dynamics 9, 85102
231. Fureby, C., Tabor, G., Weller, H.G., Gosman, A.D. (1997): A comparative
study of subgrid scale models in homogeneous isotropic turbulence. Phys.
Fluids 9(5), 14161429
232. Fureby, C., Tabor, G., Weller, H.G., Gosman, A.D. (1997): Dierential subgrid
stress models in large eddy simulations. Phys. Fluids 9 (11), 35783580
233. Fureby, C., Tabor, G., Weller, H.G., Gosman, A.D. (2000): Large-eddy simulations of the ow around a square prism. AIAA Journal 38(3), 442452
234. Galdi, G.P., Layton, W.J. (2000): Approximating the larger eddies in uid
motion II: a model for space-ltered ow. Math. Models and Meth. in Appl.
Sciences 10(3), 343350
235. Galmiche, M., Thual, O., Bonneton, P. (2002): Direct numerical simulation of
turbulence-mean eld interactions in a stably stratied uid. J. Fluid Mech.
455, 213242
236. Gao, S., Voke, P. (1995): Large eddy simulation of turbulent heat transport
in enclosed impinging jets. Int. J. heat Fluid Flow 16, 349356
237. Gao S., Voke, P., Gough, T. (1997): Turbulent simulation of at plate boundary layer and near wake. (Direct and Large Eddy Simulation II, Chollet, Voke
and Kleiser eds) Kluwer, 115124
238. Garg, R., Ferziger, J., Monismith, S., Kose, J. (2000): Stably stratied turbulent channel ows. I. Stratication regimes and turbulence suppression mechanism. Phys. Fluids 12(10), 25692593
239. Garnier, E., Mossi, M., Sagaut, P., Deville, M., Comte, P. (1999): On the use
of shock-capturing schemes for large-eddy simulation. J. Comput. Phys. 153,
273311
240. Gatski, T.B., Liu, J.T.C. (1980): On the interactions between large-scale structure and ne-grained turbulence in a free shear ow. III. A numerical solution.
Phil. Trans. Royal Soc. Lond. A 293, 473509
241. Georgiadis, N.J., Iwan, J., Alexander, D., Reshotko, E. (2001): Development
of a hybrid RANS/LES method for compressible mixing layer simulations.
AIAA Paper 2001-0289
242. Germano, M. (1986): Dierential lters for the large eddy numerical simulation
of turbulent ows. Phys. Fluids 29(6), 17551757
243. Germano, M. (1986): Dierential lters of elliptic type. Phys. Fluids 29(6),
17571758
244. Germano, M. (1986): A proposal for a redenition of the turbulent stresses in
the ltered NavierStokes equations. Phys. Fluids 29(7), 23232324
245. Germano, M. (1987): On the non-Reynolds averages in turbulence. AIAA
Paper 87-1297
References
525
246. Germano, M. (1992): Turbulence: The ltering approach. J. Fluid Mech. 238,
325336
247. Germano, M. (1996): A statistical formulation of the dynamic model. Phys.
Fluids 8 (2), 565570
248. Germano, M. (1998): Fundamentals of large-eddy simulation. (Advanced Turbulent Flows Computations, Peyret and Krause eds.), CISM Courses and Lectures 395, Springer, 81130
249. Germano, M. (1999): From RANS to DNS: towards a bridging model. (Direct
and Large Eddy Simulation III, Voke, Sandham and Kleiser eds.) Kluwer,
225236
250. Germano, M. (2001): LES overview. (DNS/LES Progress and Challenges, Liu,
Sakell and Beutner eds.), Greyden Press, 112
251. Germano, M. (2001): Ten years of the dynamic model. (Modern simulation
strategies for turbulent ow, Geurts ed.), Edwards, 173190
252. Germano, M. (2002): On a possible direct eect of the eddy viscosity gradient
in turbulence modeling. Phys. Fluids 14(10), 37453747
253. Germano, M., Piomelli, U., Moin, P., Cabot, W.H. (1991): A dynamic subgridscale eddy viscosity model. Phys. Fluids A 3(7), 17601765
254. Geurts, B. (1997): Inverse modeling for large-eddy simulation. Phys. Fluids 9
(12), 35853587
255. Geurts, B.J. (1999): Balancing errors in LES. (Direct and Large Eddy Simulation III, Voke, Sandham and Kleiser eds) Kluwer, 112
256. Geurts, B.J., Froehlich, J. (2001): Numerical eects contaminating LES;
a mixed story. (Modern simulation strategies for turbulent ow, Geurts ed.),
Edwards, 309327
257. Geurts, B.J., Froehlich, J. (2002): A framework for predicting accuracy limitations in large-eddy simulation. Phys. Fluids 14(6), L41L44
258. Geurts, B.J., Holm, D. (2003): Regularization modeling for large-eddy simulation. Phys. Fluids 15(1), L13L16
259. Ghosal, S. (1996): An analysis of numerical errors in large-eddy simulations
of turbulence. J. Comput. Phys. 125, 187206
260. Ghosal, S. (1999): Mathematical and physical constraints on large-eddy simulation of turbulence. AIAA J. 37 (4), 425433
261. Ghosal, S., Lund, T.S., Moin, P., Akselvoll, K. (1995): A dynamic localization model for large-eddy simulation of turbulent ows. J. Fluid Mech. 286,
229255
262. Ghosal, S., Moin, P. (1995): The basic equations for the large-eddy simulation
of turbulent ows in complex geometry. J. Comput. Phys. 118, 2437
263. Ghosal, S., Rogers, M. (1997): A numerical study of self-similarity in a turbulent plane wake using large-eddy simulation. Phys. Fluids 9(6), 17291739
264. Gicquel, L., Givi, P., Jaberi, F., Pope, S. (2002): Velocity ltered density
function for large eddy simulation of turbulent ows. Phys. Fluids 14(3), 1196
1213
265. Glaze, D., Frankel, S. (2003): Stochastic inlet conditions for large-eddy simulation of a fully turbulent jet. AIAA Journal 41(6), 10641073
266. Godeferd, F., Cambon, C. (1994): Detailed investigation of energy transfers
in homogeneous stratied turbulence. Phys. Fluids 6(6), 20842100
267. Godeferd, F., Staquet, C. (2003): Statistical modelling and direct numerical
simulations of decaying stably stratied turbulence. Part 2. Large-scale and
small-scale anisotropy. J. Fluid Mech. 486, 115159
268. Goldstein, D., Vasilyev, O. (2004): Stochastic coherent adaptive large eddy
simulation method. Phys Fluids 16(7), 24972513
526
References
269. Gonze, M.A. (1993): Simulation numerique des sillages en transition `

a la turbulence (in french). Thèse de Doctorat de lINPG, Grenoble, France
270. Grigoriadis, D., Bartzis, J., Goulas, A. (2003): LES of the ow past a rectangular cylinder using the immersed boundary concept. Int. J. Numer. Meth.
Fluids 41, 615632
271. Grigoriadis, D., Bartzis, J., Goulas, A. (2004): Ecient treatment of complex
geometries for large eddy simulations of turbulent ows. Computers and Fluids
33, 201222
272. Grinstein, F.F. (2001): Vortex dynamics and entrainement in rectangular free
jets. J. Fluid Mech. 437, 69101
273. Grinstein, F., Fureby, C. (2002): Recent progress on MILES for high Reynoldsnumber ows. AIAA Paper 2002-0134
274. Grinstein, F., Fureby, C. (2002): Recent progress on MILES for high Reynolds
Number ows. J. Fluid Engng. 124, 848861
275. Grinstein, F.F., Guirguis, R.H. (1992): Eective viscosity in the simulation of
spatially evolving shear ows with monotonic FCT models. J. Comput. Phys.
101, 165175
276. Grinstein, F., Margolin, L., and Rider, W. (eds.): Implicit Large Eddy Simulation: Computing Turbulent Fluid Dynamics. Cambridge University Press
277. Gr
otzbach, G. (1983): Spatial resolution requirements for direct numerical
simulation of the RayleighBenard convection. J. Comput. Phys. 49, 241264
278. Gr
otzbach, G. (1987): in Encyclopedia of Fluid Mechanics, N.P. Chereminiso
editor (Gulf, West Orange, NJ), Vol. 6
279. Gr
otzbach, G., W
orner, M. (1999): Direct and large eddy simulations in nuclear applications. Int. J. Heat Fluid Flow 20, 222240
280. Guermond, J.L. (1999): Stabilization of Galerkin approximations of transport
equations by subgrid modeling. M2AN 33(6), 12931316
281. Guermond, J.L., Oden, T.J., Prudhomme, S. (2003): An interpretation of
the NavierStokes-alpha model as a frame-indierent Leray regularization.
Physica D 177, 2330
282. Guermond, J.L., Oden, J.T., Prudhomme, S. (2004): Some mathematical issues concerning large-eddy simulation models for turbulent ows. J. Meth.
Fluid Mech. 6(2), 194248
283. Gullbrand, J., Chow, F.K. (2003): The eect of numerical errors and turbulence models in large-eddy simulations of channel ow, with and without
explicit ltering. J. Fluid Mech. 495, 323341
284. Hamba, F. (2001): An attempt to combine large eddy simulation with the
k-epsilon model in a channel ow calculation. Theoret. Comput. Fluid Dynamics 14, 323336
285. Hamba, F. (2003): A hybrid RANS/LES simulation of turbulent channel ow.
Theoret. Comput. Fluid Dynamics 16(5), 387403
286. Han, J., Lin, Y.L., Schowalter, D.G., Pal Arya, S., Proctor, F.H. (2000):
Large-eddy simulation of aircraft wake vortices within homogeneous turbulence: Crow instability. AIAA Journal 38(2), 292300
287. Hanazaki, H. (2003): Eects of initial conditions on the passive and active
scalar uxes in unsteady stably stratied turbulence. Phys. Fluids 15(4), 841
848
288. Hansen, R.P., Long, L.N., Morris, P.J. (2000) Unsteady, laminar ow simulations using the nonlinear disturbance equations. AIAA Paper 2000-1981
289. H
artel, C., Kleiser, L. (1997): Galilean invariance and ltering dependence of
near-wall grid-scale/subgrid-scale interactions in large-eddy simulation. Phys.
Fluids 9(2), 473475
References
527
290. H
artel, C., Kleiser, L. (1998): Analysis and modelling of subgrid-scale motions
in near-wall turbulence. J. Fluid Mech. 356, 327352
291. H
artel, C., Kleiser, L., Unger, F., Friedrich, R. (1994): Subgrid-scale energy transfer in the near-wall region of turbulent ows. Phys. Fluids 6(9),
31303143
292. Harten, A. (1984): On a class of high resolution total-variation-stable nitedierence schemes. SIAM J. Numer. Anal. 21, 123
293. Harten, A. (1993): Discrete multi-resolution analysis and generalized wavelets.
Applied Numerical Mathematics 12, 153192
294. Harten, A. (1995): Multiresolution algorithms for the numerical solution of
hyperbolic conservation laws. Communication on Pure and Applied Mathematics XLVIII, 13051342
295. Harten, A. (1996): Multiresolution representation of data: a general framework. SIAM J. Numer. Anal. 33(3), 12051256
296. Haselbacher, A., Moser, R., Constantinescu, G., Mahesh, K. (2002): Toward
optimal LES on unstructured meshes. Proceedings of the Summer Program,
Center for Turbulence Research, 129140
297. Haselbacher, A., Vasilyev, O. (2003): Commutative discrete ltering on unstructured grids based on least-squares techniques. J. Comput. Phys. 187,
197211
298. Hassan, Y., Barsamian, H. (2001): New-wall modeling for complex ows using
the large eddy simulation technique in curvilinear coordinates. Int. J. Heat
Mass Trans. 44, 40094026
299. Hauke, G., Garcia-Olivares, A. (2001): Variational subgrid scale formulations
for the advection-diusion-reaction equation. Comput. Methods Appl. Mech.
Engrg. 190, 68476865
300. He, G.W., Rubinstein, R., Wang, L.P. (2001): Eects of subgrid-scale modeling
on time correlations in large-eddy simulation. ICASE Report No. 2001-10
301. He, G.W., Rubinstein, R., Wang, L.P. (2002): Eects of subgrid-scale modeling
on time correlations in large-eddy simulation. Phys. Fluids 14(7), 21862193
302. Henn, D.S., Sykes, R.I. (1999): Large-eddy simulation of ow over wavy surfaces. J. Fluid Mech. 383, 75112
303. Herr, S., Wang, L.P., Collins, L. (1996): EDQNM model of a passive scalar
with uniform mean gradient. Phys. Fluids 8(6), 15881608
304. Herring, J., Metais, O. (1989): Numerical experiments in forced stably stratied turbulence. J. Fluid Mech. 202, 97115
305. Hewson, J.C., Kerstein, A. (2002): Local extinction and reignition in nonpremixed turbulent CH/H2/N2 jet ames. Combust. Sci. and Tech. 174, 35
66
306. Hewson, J.C., Kerstein, A. (2001): Stochastic simulation of transport and
chemical kinetics in turbulent CH/H2/N2 ames. Combust. Theory Modelling
5, 669697
307. Hirsch, C. (1987): Numerical computation of internal and external ows. John
Wiley & Son
308. Homan, J. (2001): Dynamic subgrid modeling for scalar convection-diusionreaction equations with fractal coecients. Multiscale and multiresolution
methods: theory and applications, Lecture Notes in Computational Science
and Engineering, vol. 20. Springer, Heidelberg
309. Homan, J. (2002): Dynamic subgrid modelling for time dependent
convection-diusion-reaction equations with fractal solutions. Int. J. Numer.
310. Homan, J. (2003): Adaptive DNS/LES: a new agenda in CFD. Chalmers
Finite Element Center preprint 2003-23. Chalmers University of Technology
528
References
311. Homan, J. (2003): Subgrid modeling for convection-diusion-reaction in 2

space dimensions using a Haar multiresolution analysis. Math. Models &
Meth. Appl. Sci. 13(10), 15151536
312. Homan, J. (2004): Duality based a posteriori error estimation in various
norms and linear functionals for LES. SIAM J. Sci. Comput. 26(1), 178195
313. Homan, J.: Adaptive nite element methods for LES: computation of the
mean drag coecient in a turbulent ow around a surface mounted cube
using adpative mesh renement. Submitted
314. Homan, J., Johnson, C. (2004): Computability and adaptivity in CFD. Encyclopedia of Computational Mechanics, Stein, E., de Horz, R. and Hughes,
T.J.R. eds, John Wiley & Sons
315. Homan, J., Johnson, C., Bertoluzza, S. (2005): Subgrid modeling for
convection-diusion-reaction in 1 space dimension using a Haar multiresolution analysis. Comupt. Meth. Appl. Mech. Engng. 194(1), 1944
316. Holmen, J., Hughes, T.J.R., Oberai, A., Wells, G. (2004): Sensitivity of the
scale partition for variational multiscale large-eddy simulation of channel ow.
Phys. Fluids 16(3), 824827
317. Holtslag, A., Moeng, C.H. (1991): Eddy diusivity and countergradient transport in the convective atmospheric boundary layer. J. Atmos. Sci. 48(14),
16901698
318. Horiuti, K. (1985): Large eddy simulation of turbulent channel ow by oneequation modeling. J. Phys. Soc. Japan 54(8), 28552865
319. Horiuti, K. (1987): Comparison of conservative and rotational forms in large
eddy simulation of turbulent channel ow. J. Comput. Phys. 71, 343370
320. Horiuti, K. (1989): The role of the Bardina model in large eddy simulation of
turbulent channel ow. Phys. Fluids A 1(2), 426428
321. Horiuti, K. (1990): Higher-order terms in the anisotropic representation of
Reynolds stresses. Phys. Fluids A 2(10), 17081710
322. Horiuti, K. (1993): A proper velocity scale for modeling subgrid-scale eddy
viscosities in large eddy simulation. Phys. Fluids A 5(1), 146157
323. Horiuti, K. (1997): Backward scatter of subgrid-scale energy in wall-bounded
turbulence and free shear ow. J. Phys. Soc. Japan 66(1), 91107
324. Horiuti, K. (1997): A new dynamic two-parameter mixed model for large-eddy
simulation. Phys. Fluids 9(11), 34433464
325. Horiuti, K. (2001): Alignment of eigenvectors for strain rate and subgrid-scale
stress tensors. (Direct and large-eddy simulation IV, Geurts, Friedrich and
Metais eds.), Kluwer, 6772
326. Horiuti, K. (2001): Rotational transformation and geometrical correlation of
SGS models. (Modern simulation strategies for turbulent ow, Geurts ed.),
Edwards, 123140
327. Horiuti, K. (2003): Roles of non-aligned eigenvectors of strain-rate and
subgrid-scale stress tensors in turbulence generation. J. Fluid Mech. 491, 65
100
328. Hossain, M. (1991): Non-diusive subgrid-scale modelling of turbulence. Phys.
Letters A 161, 277282
329. Hsiao, C.T., Pauley, L.L. (1999): Direct numerical simulation of unsteady
nite-span hydrofoil ow. AIAA Journal 37(5), 529536
330. Huai, X., Joslin, R.D., Piomelli, U. (1999): Large-eddy simulation of boundarylayer on a swept wedge. J. Fluid Mech. 381, 357380
331. Hughes, T.J.R. (1995): Multiscale phenomena: Greens function, the Dirichletto-Neumann formulation, subgrid scale models, bubbles and the origin of stabilized methods. Comput. Methods Appl. Mech. Engrg. 127, 387401
References
529
332. Hughes, T.J.R., Feijoo, G.R., Mazzei, L., Quincy, J.B. (1998): The variational
multiscale method a paradigm for computational mechanics. Comput. Methods Appl. Mech. Engrg. 166, 224
333. Hughes, T.J.R., Mazzei, L., Jansen, K.E. (2000): Large eddy simulation and
the variational multiscale method. Comput. Visual Sci. 3, 4759
334. Hughes, T.J.R., Mazzel, L., Oberai, A.A., Wray, A.A. (2001): The multiscale
formulation of large-eddy simulation: decay of homogeneous isotropic turbulence. Phys. Fluids 13(2), 505512
335. Hughes, T.J.R., Oberai, A.A., Mazzei, L. (2001): Large-eddy simulation of
turbulent channel ows by the variational multiscale method. Phys. Fluids
13(6), 17841799
336. Hughes, T.J.R., Stewart, J. (1996): A space-time formulation for multiscale
phenomena. J. Comput. Appl. Math. 74, 217229
337. Hunt, J.C.R., Carlotti, P. (2001): Statistical structure at the wall of the high
Reynolds number turbulent boundary layer. Flow, Turbulence and Combustion 66, 453475
338. Hylin, E.C., McDonough, J.M.(1996): Theoretical development of a stochastic
model for small-scale turbulence in an additive decomposition of the NavierStokes equations. ME Report No CFD-02-96, Dept. of Mechanical Engineering, University of Kentucky
339. Iliescu, T., Fischer, P. (2003): Large eddy simulation of turbulent channel ows
by the rational large eddy simulation model. Phys. Fluids 15(10), 30363047
340. Iliescu, T., Fischer, P. (2004): Backscatter in the rational LES model. Computers and Fluids 33, 783790
341. Iliescu, T., John, V., Layton, W.J., Matthies, G., Tobiska, L. (2000): An assessment of models in large eddy simulation. Unpublished
342. Iliescu, T., Layton, W.J. (1998): Approximating the larger eddies in uid
motion III: the Boussinesq model for turbulent uctuations. An. St. Univ.
Al. I. Cuza 44, 245261
343. Iovenio, M., Tordella, D. (2003): Variable scale ltered NavierStokes equations: a new procedure to deal with the associated commutation error. Phys.
Fluids 15(7), 19261936
344. Jaberi, F., Colucci, P. (2003): Large eddy simulation of heat and mass transport in turbulent ows. Part 1: velocity eld. Int. J. Heat Mass Transfer 46,
18111825
345. Jaberi, F., Colucci, P. (2003): Large eddy simulation of heat and mass transport in turbulent ows. Part 2: scalar eld. Int. J. Heat Mass Transfer 46,
18271840
346. Jameson, A., Schmidt, W., Turkel, E. (1981): Numerical solutions of the Euler
equations by nite volume methods using Runge-Kutta time stepping schemes.
AIAA Paper 81-1259
347. Jansen, K. (1994): Unstructured-grid large-eddy simulation of ow over an
airfoil. Annual Research Briefs Center for Turbulence Research, 161175
348. Jansen, K. (1995): Preliminary large-eddy simulations of ow around a NACA
4412 airfoil using unstructured meshes. Annual Research Briefs Center for
349. Jansen, K. (1996): Large-eddy simulations of ow around a NACA 4412 airfoil
using unstructured grids. Annual Research Briefs Center for Turbulence
Research, 225233
350. Jiang, G.S., Shu, C.W. (1996): Ecient implementation of weighted ENO
schemes. J. Comput. Phys. 126, 202228
351. Jimenez, J. (1999): On eddy-viscosity sub-grid models. (Direct and Large
Eddy Simulation III, Voke, Sandham and Kleiser eds.) Kluwer, 7586
530
References
352. Jimenez, C., Ducros, F., Cuenot, B., Bedat, B. (2001): Subgrid scale variance
and dissipation of a scalar eld in large eddy simulations. Phys. Fluids 13(6),
17481754
353. Jimenez, J., Moin, P. (1991) The minimal ow unit in near-wall turbulence.
J. Fluid Mech. 225, 213240
354. Jimenez, J., Pinelli, A. (1999) The autonomous cycle of near-wall turbulence.
J. Fluid Mech. 389, 335359
355. Jimenez, J., Simens, M.P. (2001) Low-dimensional dynamics of a turbulent
wall ow. J. Fluid Mech. 435, 8191
356. Jimenez, J., Vasco, C. (1998): Approximate lateral boundary conditions for
turbulent simulations. Proceedings of the Summer Program Center for Turbulence Research, 399412
357. John, V., Layton, W.J. (2001): Analysis of numerical errors in large eddy
simulation. Unpublished
358. Jordan, S.A. (1999): A large-eddy simulation methodology in generalized
curvilinear coordinates. J. Comput. Phys. 148, 322340
359. Jordan, S.A. (2001): Dynamic subgrid-scale modeling for large-eddy simulations in complex topologies. J. Fluids Engng. 123, 619627
360. Jordan, S.A. (2002): Investigation of the cylinder separated shear-layer physics
by large-eddy simulation. Int. J. Heat Fluid Flow 23, 112
361. Jordan, S.A. (2003): The turbulent character and pressure loss produced by
periodic symmetric ribs in a circular duct. Int. J. Heat Fluid Flow 24, 795806
362. Jordan, S.A., Ragab, S.A. (1998): A large eddy simulation of the near wake
of a circular cylinder. Journal of Fluids Engineering 120, 243252
363. Juneja, A., Brasseur, J.G. (1999): Characterictics of subgrid-resolved-scale
dynamics in anisotropic turbulence, with application to rough-wall boundary
layers. Phys Fluids 11 (10), 30543068
364. Juneja, A., Lathrop, D.P., Sreenivasan, K.R., Stolovitzky, G. (1994): Synthetic
turbulence. Phys. Rev. E. 49(6), 51795194
365. Khanna, S., Brasseur, J.G. (1997): Analysis of Monin-Obukhov similarity from
large-eddy simulation. J. Fluid Mech. 435, 251286
366. Kannepalli, C, Piomelli, U. (2000): Large-eddy simulation of a threedimensional shear-driven turbulent boundary layer. J. Fluid Mech. 423,
175203
367. Kaltenbach, H.J. (1994): Large-eddy simulation of ow through a plane, asymmetric diuser. Annual Research Briefs Center for Turbulence Research,
175185
368. Kaltenbach, H.J. (1997): Cell aspect ratio dependence of anisotropy measures
for resolved and subgrid scale stresses. J. Comput. Phys. 136, 399410
369. Kaltenbach, H.J. (1998): Towards a near-wall model for LES of a separated diffuser ow. Annual Research Briefs Center for Turbulence Research, 255265
370. Kaltenbach, H.J. (2003): A priori testing of wall models for separated ows.
Phys. Fluids 15(10), 30483064
371. Kaltenbach, H.J., Choi, H. (1995): Large-eddy simulation of ow around an
airfoil on structured mesh. Annual Research Briefs Center for Turbulence
Research, 5161
372. Kaltenbach, H.J., Fatica, M., Mittal, R., Lund, T.S., Moin, P. (1999): Study
of ow in a planar asymmetric diuser using large-eddy simulation. J. Fluid
Mech. 390, 151185
373. Kaltenbach, H., Gerz, T., Schumann, U. (1994): Large-eddy simulation of
homogeneous turbulence and diusion in stably stratied shear ows. J. Fluid
Mech. 280, 140
References
531
374. Kaneda, Y., Leslie, D.C. (1983): Tests of subgrid models in the near-wall
region using represented velocity elds. J. Fluid Mech. 132, 349373
375. Kang, H.S., Chester, S., Meneveau, C. (2003): Decaying turbulence in
an active-grid-generated ow and comparisons with large-eddy simulation.
J. Fluid Mech. 480, 129160
376. Kang, H.S., Meneveau, C. (2001): Passive scalar anisotropy in a heated turbulent wake: new observations and implications for large-eddy simulations.
J. Fluid Mech. 442, 161170
377. Kaniel, S. (1970): On the initial value problem for an incompressible uid with
nonlinear viscosity. J. Math. Mech. 19(8), 681706
378. Kanna, S., Brasseur, J.G. (1997): Analysis of Monin-Obukhov similarity from
large-eddy simulation. J. Fluid Mech. 345, 251286
379. Karamanos, G.S., Karniadakis, G.E. (2000): A spectral vanishing viscosity
method for large-eddy simulation. J. Comput. Phys. 163, 2250
380. Kato, C., Kaiho, M., Manabe, A. (2001): Industrial applications of LES in
mechanical engineering. (DNS/LES Progress and Challenges, Liu, Sakell and
381. Kawamura, T., Kuwahara, K. (1984): Computation of high Reynolds number
ow around a circular cylinder with surface roughness. AIAA Paper 84-0340
382. Kemenov, K., Menon, S. (2003): Two level simulation of high-Reynolds number non-homogeneous turbulent ows. AIAA Paper 2003-0084
383. Kerr, R.M., Domaradzki, J.A., Barbier, G. (1996): Small-scale properties of
nonlinear interactions and subgrid-scale energy transfer in isotropic turbulence. Phys. Fluids 8(1), 197208
384. Kerstein, A. (1991): Linear-eddy modelling of turbulent transport. Part 6.
Microstructure of diusive scalar mixing elds. J. Fluid Mech. 231, 361394
385. Kerstein, A. (1992): Linear-eddy modelling of turbulent transport. Part 7.
Finite-rate chemistry and multi-stream mixing. J. Fluid Mech. 240, 289313
386. Kerstein, A. (1999): One-dimensional turbulence Part 2. Staircases in doublediusive convection. Dynamics of Atmospheres and Oceans 30, 2546
387. Kerstein, A.R. (1999): One-dimensional turbulence: model formulation and
application to homogeneous turbulence, shear ows, and buoyant stratied
ows. J. Fluid Mech. 392, 277334
388. Kerstein, A. (2002): One dimensional turbulence: a new approach to highdelity subgrid closure of turbulent ow simulations. Comput. Phys. Comm.
148, 116
389. Kerstein, A.R., Ashurst, W.T., Wunsch, S., Nilsen, V. (2001): One-dimensional turbulence: vector formulation and application to free shear ows.
J. Fluid Mech. 447, 853109
390. Kerstein, A.R., Dreeben, T.D. (2000): Prediction of turbulent free shear ow
statistics using a simple stochastic model. Phys. Fluids 12(2), 418424
391. Kim, W.W., Menon, S. (1999): An unsteady incompressible NavierStokes
solver for large-eddy simulation of turbulent ows. Int. J. Numer. Meth. Fluids
31, 9831017
392. Kim, W.W., Menon, S., Mongia, H.C. (1999): Large eddy simulation of a gas
turbine combustor ow. Combust. Sci. Tech. 143, 2562
393. Kim, D.H., Yang, K.S., Send, M. (2004): Large eddy simulation of turbulent
ow past a square cylinder conned in a channel. Computers and Fluids 33,
8196
394. Kimmel, S.J., Domaradzki, A.J. (2000): Large-eddy simulations of Rayleigh
Benard convection using subgrid scale estimation model. Phys. Fluids 12(1),
169183
532
References
395. Kirby, R., Karniadakis, G. (2002): Coarse resolution turbulence simulations

with spectral vanishing viscositylarge eddy simulations. J. Fluid Engng. 124,
886891
396. Klein, M., Sadiki, A., Janicka, J. (2003): A digital lter based generation of
inow data for spatially developing direct numerical or large eddy simulation.
J. Comput. Phys. 186, 652665
397. Knaepen, B., Debliquy, O., Carati, D. (2002): Subgrid-scale energy and pseudo
pressure in large-eddy simulation. Phys. Fluids 14(12), 42354241
398. Kobayashi, H., Shimomura, Y. (2001): The performance of dynamic subgridscale models in the large-eddy simulation of rotating homogeneous turbulence.
Phys. Fluids 13(8), 23502360
399. Kogaki, T., Kobayashi, T., Taniguchi, N. (1997): Large eddy simulation of
ow around a rectangular cylinder. Fluid Dyn. res. 20, 1124
400. Kollman, W., McCallen, R.C., Leone Jr., J.M. (2002): An examination of LES
ltering within the nite element method. Commun. Numer. Meth. Engng. 18,
513528
401. Kosovic, B. (1997): Subgrid-scale modelling for the large-eddy simulation of
high-Reynolds number boundary layers. J. Fluid Mech. 336, 151182
402. Kosovic, B., Curry, J. (2000): A large eddy simulation study of a quasi-steady,
stably stratied atmospheric boundary layer. J. Atmos. Sci. 57, 10521068
403. Kraichnan, R.H. (1967): Inertial ranges in two-dimensional turbulence.
J. Fluid Mech. 10(7), 14171423
404. Kraichnan, R.H. (1971): Inertial-range transfer in two- and three-dimensional
turbulence. J. Fluid Mech. 47(3), 525535
405. Kraichnan, R.H. (1976): Eddy viscosity in two and three dimensions. J. Atmos.
Sci. 33, 15211536
406. Krajnovic, S., Davidson, L. (2002): A mixed one-equation subgrid model for
large-eddy simulation. Int. J. Heat Fluid Flow 23, 413425
407. Krajnovic, S., Davidson, L. (2002): Large-eddy simulation of the ow around
a blu body. AIAA Journal 40(5), 927936
408. Krajnovic, S., Davidson, L. (2003): Numerical study of the ow around a busshaped body. J. Fluids Engng. 125, 500509
409. Kravchenko, A.G., Moin, P. (1997): On the eect of numerical errors in largeeddy simulations of turbulent ows. J. Comput. Phys. 131, 310322
410. Kravchenko, A.G., Moin, P. (2000): Numerical studies of ow over a circular
cylinder at ReD = 3900. Phys. Fluids 12(2), 403417
411. Kravchenko, A.G., Moin, P., Moser, R. (1996): Zonal embedded grids for numerical simulations of wall-bounded turbulent ows. J. Comput. Phys. 127,
412423
412. Krettenauer, K., Schumann, U. (1992): Numerical simulation of turbulent convection over a wavy terrain. J. Fluid Mech. 237, 261299
413. Krueger, S. (1993): Linear eddy modeling of entrainment and mixing in stratus
clouds. J. Atmos. Sci. 50(18), 30783090
414. Krueger, S., Su, C.W., McMurty, P. (1997): Modeling entrainment and
nescale mixing in cumulus clouds. J. Atmos. Sci. 54(23), 26972712
415. Kuerten, J.M.G., Geurts, B.J., Vreman, A.W., Germano, M. (1999): Dynamic
inverse modeling and its testing in large-eddy simulations of the mixing layer.
Phys. Fluids 11 (12), 37783785
416. Labourasse, E., Sagaut, P. (2002) Reconstruction of turbulent uctuations
using a hybrid RANS/LES approach. J. Comput. Phys. 182, 301336
417. Ladyzenskaja, O.A. (1970): Modication of the NavierStokes equations for
large velocity gradients. (Seminar in Mathematics V.A. Stheklov Mathematical Institute, Vol. 7, Boundary value problems of mathematical physics and
References
418.
419.
420.
421.
422.
423.
424.
425.
426.
427.
428.
429.
430.
431.
432.
433.
434.
435.
436.
437.
438.
533
related problems of function theory, Part II, New York, London 1970, Consultant Bureau)
Ladyzenskaja, O.A. (1970): New equations for description of motion of viscous
incompressible uids and solvability in the large of boundary value problems
for them. (Seminar in Mathematics V.A. Stheklov Mathematical Institute,
Vol. 102, Boundary value problems of mathematical physics, Part V, Providence, Rhode Island, AMS)
Langford, J., Moser, R. (2000): Optimal LES formulations for isotropic turbulence. J. Fluid Mech. 398, 321346
Langford, J.A., Moser, R.D. (2001): Breakdown of continuity in large-eddy
simulation. Phys. Fluids 13(5), 15241527
Lamballais, E. (1996): Simulations numeriques de la turbulence dans un canal
plan tournant (in french). Thèse de Doctorat de lINPG, Grenoble, France
Lamballais, E., Metais, O., Lesieur, M. (1998): Spectral-dynamic model for
large-eddy simulations of turbulent rotating channel ow. Theoret. Comput.
Fluid Dynamics 12, 149177
Landau, L., Lifchitz, E. (1967): Fluid Mechanics. MIR, Moscou
Launder, B.E., Spalding, D.B. (1972): Mathematical models of turbulence.
Academic Press, Londres
Lardat, R. (1997): Simulation numeriques decoulements externes instationnaires decolles autour dune aile avec des modèles de sous-maille (in french).
Notes et documents du LIMSI 97-12
Laval, J.P., Dubrulle, B., McWilliams, J.C. (2003): Langevin models of turbulence: renormalization group, distant interaction algorithms or rapide distortion theory ? Phys. Fluids 15(5), 13271339
Layton, W.J. (1996): A nonlinear, subgridscale model for incompressible viscous ow problems. SIAM J. Sci. Comput. 17 (2), 347357
Layton, W.J. (2000): Approximating the larger eddies in uid motion V: kinetic energy balance of scale similarity models. Mathematical and Computer
Modelling 31, 17
Layton, W.J. (2000): Analysis of a scale-similarity model of the motion of
large eddies in turbulent ows. Unpublished
Layton, W., Lewandowski, R. (2002): Analysis of an eddy viscosity model for
large eddy simulation of turbulent ows. J. Math. Fluid Mech. 4, 374399
Le Ribault, C., Sarkar, S., Stanley, S.A. (1999): Large eddy simulation of
a plane jet. Phys. Fluids 11(10), 30693083
Lee, S., Lele, S.K., Moin, P. (1992): Simulation of spatially evolving turbulence
and the application of Taylors hypothesis in compressible ow. Phys. Fluids
A 4(7), 15211530
Lee, J.S., Xu, X., Pletcher, R. (2004): Large eddy simulation of heated vertical
annular pipe ow in fully developed turbulent mixed convection. Int. J. Heat
Leith, C.E. (1990): Stochastic backscatter in a subgrid-scale model: Plane
shear mixing layer. Phys. Fluids A 2(3), 297299
Lele, S.K. (1994): Compressibility eects on turbulence. Ann. Rev. Fluid.
Mech. 26, 211254
Leonard, A. (1974): Energy cascade in large-eddy simulations of turbulent
uid ows. Adv. in Geophys. A 18, 237248
Leonard, B.P. (1979): A stable and accurate convective modelling procedure
based on quadratic upstream interpolation. Comp. Meth. Appl. Mech. Eng.
19, 5998
Lesieur, M. (1983): Introduction `
a la turbulence bidimensionnelle (in french).
J. Mec. Theor. Appl., numero special, 520
534
References
439. Lesieur, M. (1997): Turbulence in uids, 3rd edition. Kluwer

440. Lesieur, M., Metais, O. (1996): New trends in large-eddy simulations of turbulence. Ann. Rev. Fluid Mech. 28, 4582
441. Lesieur, M., Rogallo, R.S. (1989): Large-eddy simulation of passive scalar diffusion in isotropic turbulence. Phys. Fluids A 1(4), 718722
442. Lesieur, M., Schertzer, D. (1978): Amortissement autosimilaire dune turbulence `
a grand nombre de Reynolds (in french). Journal de Mecanique 17(4),
609646
443. Leslie, D.C., Quarini, G.L. (1979): The application of turbulence theory to
the formulation of subgrid modelling procedures. J. Fluid Mech. 91(1), 6591
444. Lewellen, D., Lewellen, W., Yoh, S. (1996): Inuence of Bowen ratio on
boundary-layer cloud structure. J. Atmos. Sci 53(1), 175187
445. Li, N., Balaras, E., Piomelli, U. (2000): Inow conditions for large-eddy simulations of mixing layers. Phys. Fluids 12(4), 935938
446. Li, C.W., Wang, J.H. (2000): Large-eddy simulation of free surface shallowwater ow. Int. J. Numer. Meth. Fluids 34, 3146
447. Lilly, D.K. (1967): The representation of small-scale turbulence in numerical
simulation experiments. (Proceedings of the IBM Scientic Computing Symposium on Environmental Sciences, Yorktown Heights, USA)
448. Lilly, D.K. (1992): A proposed modication of the Germano subgrid-scale
closure method. Phys. Fluids A 4(3), 633635
449. Lin, C.C. (1999): Near-grid-scale energy transfer and coherent structures in
the convective planetary boundary layer. Phys. Fluids 11 (11), 34823494
450. Lin, C.L., McWilliams, J.C., Moeng, C.H., Sullivan, P. (1996): Coherent structures and dynamics in a neutrally stratied planetary boundary layer ow.
Phys. Fluids 8(10), 26262639
451. Liu, J.T.C. (1988): Contributions to the understanding of large-scale coherent
structures in a developing free turbulent shear ows. Advances in Applied
Mechanics 20, 183307
452. Liu, C., Liu, Z. (1994): Fourth order nite dierence and multigrid methods
for modeling instabilities in at plate boundary layers - 2D and 3D approaches.
Computers Fluids 23(7), 955982
453. Liu, C., Liu, Z. (1995): Multigrid mapping in box relaxation for simulation of
the whole process of ow transition in 3D boundary layers. J. Comput. Phys.
119, 325341
454. Liu, S., Katz, J., Meneveau, C. (1999): Evolution and modelling of subgrid
scales during rapid straining of turbulence. J. Fluid Mech. 387, 281320
455. Liu, S., Meneveau, C., Katz, J. (1994): On the properties of similarity subgridscale models as deduced from measurements in a turbulent jet. J. Fluid Mech.
275, 83119
456. Liu, J.T.C., Merkine, L. (1976): On the interactions betzeen large-scale structure and ne-grained turbulence in a free shear ow. I. The development of
temporal interactions in the mean. Proc. R. Soc. Lond. A 352, 213247
457. Lo, S.H., Voke, P.R., Rockli, N.J. (2000): Eddy structures in a simulated low
Reynolds number turbulent boundary layer. Flow, Turbulence and Combustion 64, 128
458. Loh, K.C., Domaradzki, J.A. (1999): The subgrid-scale estimation model on
nonuniform grids. Phys. Fluids 11 (12), 37863792
459. Love, M.D. (1980): Subgrid modelling studies with Burgers equation. J. Fluid
Mech. 100(1), 87110
460. Lund, T.S. (1994): Large-eddy simulation of a boundary layer with concave
streamwise curvature. Annual Research Briefs Center for Turbulence Research, 185197
References
535
461. Lund, T.S. (1997): On the use of discrete lters for large eddy simulation.
Annual Research Briefs Center for Turbulence Research, 8395
462. Lund, T.S., Novikov, E.A. (1992): Parametrization of subgrid-scale stress by
the velocity gradient tensor. Annual Research Briefs Center for Turbulence
Research, 2743
463. Lund, T.S., Wu, X., Squires, K.D. (1998): On the generation of turbulent
inow conditions for boundary-layer simulations. J. Comput. Phys. 140,
233258
464. McComb, W.D. (1990): The physics of uid turbulence. Clarendon Press,
Oxford
465. McComb, W.D., Hunter, A., Johnston, C. (2001): Conditional mode-elimination and the subgrid-modeling problem for isotropic turbulence. Phys. Fluids
13(7), 20302044
466. McDonough, J.M., Bywater, R.J. (1985): Eects of local large-scale parameters on the small-scale chaotic solutions to Burgers equation. AIAA Paper
85-1653
467. McDonough, J.M., Bywater, R.J. (1986): Large-scale eects on local smallscale chaotic solutions to Burgers equation. AIAA J. 24 (12), 19241930
468. McDonough, J.M., Bywater, R.J., Buell, J.C. (1984): An investigation of
strange attractor theory and small-scale turbulence. AIAA Paper 84-1674
469. McDonough, J.M., Mukerji, S., Chung, S. (1998): A data-tting procedure for
chaotic time series. Appl. Math. and Comput. 95, 219243
470. McDonough, J.M., Saito, K. (1994): Local, small-scale interaction of turbulence with chemical reactions in H2-O2 combustion. Fire Science and Technology 14 (1), 118
471. McDonough, J.M., Wang, D. (1995): Additive turbulent decomposition:
a highly parallelizable turbulence simulation technique. (New algorithms ans
applications, Satofuka, Periaux, Ecer eds.) Elsevier, 129136
472. Mcllwain, S., Pollard, A. (2002): Large eddy simulation of the eects of mild
swirl on the near eld of a round free jet. Phys. Fluids 14(2), 653661
473. McMurty, P., Menon, S., Kerstein, A. (1993): Linear eddy modeling of turbulent combustion. Energy and Fuels 7, 817826
474. McRae, G.J., Goodin, W.R., Seinfeld, J. (1982): Numerical solution of the
atmospheric diusion equation for chemically reacting ows. J. Comput. Phys.
45, 142
475. Maday, Y., Ould Kaber, M., Tadmor, E. (1993): Legendre pseudospectral viscosity method for nonlinear conservation laws. SIAM J. Numer. Anal. 30(2),
321342
476. Magnagato, F., Gabi, M. (2002): A new adaptive turbulence model for unstedy
ow elds in rotating machinery. Int. J. Rotating Machinery 8(3), 175183
477. Magnient, J.C. (2001) Simulation des grandes echelles decoulements de uide quasi-incompressibles (in french). Thèse de Doctorat, Universite Paris 11,
Orsay, France
478. Magnient, J.C., Sagaut, P., Deville, M. (1999): Analysis of mesh-independent
sublter-scale models for turbulent ows. (Direct and Large Eddy Simulation
III, Voke, Sandham and Kleiser eds.) Kluwer, 263274
479. Magnient, J.C., Sagaut, P., Deville, M. (2001): A study of built-in lter for
some eddy viscosity models in large-eddy simulation. Phys. Fluids 13(5),
14401449
480. Majander, P., Siikonen, T. (2002): Evaluation of Smagorinsky-based subgrid
scale models in a nite-volume computation. Int. J. Numer. Meth. Fluids 40,
735774
536
References
481. Maltrud, M.E., Vallis, G.K. (1993): Energy and enstrophy transfer in numerical simulations of two-dimensional turbulence. Phys. Fluids A 5, 17601775
482. Manhart, M. (2004): A zonal grid algorithm for DNS of turbulent boundary
layers. Computers and Fluids 33, 435461
483. Manhart, M., Friedrich, R. (1999): Towards DNS of separated turbulent
boundary layers. (Direct and Large Eddy Simulation III, Voke, Sandham and
Kleiser eds.) Kluwer, 429440
484. Manoha, E., Tro, B., Sagaut, P. (2000): Trailing edge noise prediction using
large-eddy simulation and acoustic analogy. AIAA Journal 38(8), 13401350
485. Manseld, J.R., Knio, O.M., Meneveau, C. (1998): A dynamic LES scheme for
the vorticity transport equation: formulation and a priori tests. J. Comput.
Phys. 145, 693730
486. Manseld, J.R., Knio, O.M., Meneveau, C. (1999): Dynamic LES of colliding
vortex rings using a 3D vortex method. J. Comput. Phys. 152, 305345
487. Mansour, N.N., Moin, P., Reynolds, W.C., Ferziger, J.H. (1977): Improved
methods for large-eddy simulations of turbulence. (Symposium on Turbulent
Shear Flow, Penn State, USA)
488. Margolin, L., Smolarkiewicz, P., Sorbjan, Z. (1999): Large-eddy simulations
of convective boundary layers using nonoscillatory schemes. Physica D 133,
390397
489. Margolin, L., Smolarkiewicz, P., Wyszogrodzki, A. (2002): Implicit turbulence
modeling for high Reynolds number ows. J. Fluid Engng. 124, 862867
490. Marsden, A.L., Vasilyev, O.V., Moin, P. (2000): Construction of commutative
lters for LES on unstructured meshes. Annual Research Briefs Center for
491. Marshall, J.S., Beninati, M.L. (2003): Analysis of subgrid scale torque for
large-eddy simulation of turbulence. AIAA Journal 41(10), 18751881
492. Marusic, I., Kunkel, G., Porte-Agel, F. (2001): Experimental study of wall
boundary conditions fro large-eddy simulation. J. Fluid Mech. 446, 309320
493. Maruyama, T., Rodi, W., Maruyama, Y., Hiroaka, H. (1999): Large eddy
simulation of the turbulent boundary layer behind roughness elements using
an articially generated inow. J. Wind Eng. Ind. Aerodyn. 83, 381392
494. Mary, I., Sagaut, P. (2003): Large-eddy simulation of the ow around an airfoil
near stall. AIAA Journal 40(6), 11391145
495. Mason, P.J. (1994): Large-eddy simulation: A critical review of the technique.
Q. J. R. Meteorol. Soc. 120, 126
496. Mason, P.J., Brown, A.R. (1994): The sensitivity of large-eddy simulations of
turbulent shear ow to subgrid models. Boundary Layer Meteorol. 70, 133150
497. Mason, P.J., Callen, N.S. (1986): On the magnitude of the subgrid-scale eddy
coecient in large-eddy simulations of turbulent channel ow. J. Fluid Mech.
162, 439462
498. Mason, P.J., Thomson, D.J. (1992): Stochastic backscatter in large-eddy simulations of boundary layers. J. Fluid Mech. 242, 5178
499. Mathew, J., Lechner, R., Foysi, H., Sesterhenn, J., Friedrich, R. (2003): An
explicit ltering method for large-eddy simulation of compressible ows. Phys.
Fluids 15(8), 22792288
500. Maurer, J., Fey, M. (1999): A scale-residual model for large-eddy simulation.
(Direct and Large Eddy Simulation III, Voke, Sandham and Kleiser eds.)
Kluwer, 237248
501. Meinke, M., Schr
oder, W., Krause, E., Rister, Th. (2002): A comparison of
second- and sixth-order methods for large-eddy simulations. Computers and
Fluids, in press
References
537
502. Mellen, C., Fr

ohlich, J., Rodi, W. (2003): Lessons from LESFOIL Project on
large-eddy simulation of ow around an airfoil. AIAA Journal 41(4), 573581
503. Meneveau, C. (1994): Statistics of turbulence subgrid-scale stresses: Necessary
conditions and experimental tests. Phys. Fluids 6(2), 815833
504. Meneveau, C., Katz, J. (1999): Conditional subgrid force and dissipation in locally isotropic and rapidly strained turbulence. Phys Fluids 11 (8), 23172329
505. Meneveau, C., Katz, J. (1999): Dynamic testing of subgrid models in largeeddy simulation based on the Germano identity. Phys. Fluids 11 (2), 245247
506. Meneveau, C., Katz, J. (2000): Scale-invariance and turbulence models for
large-eddy simulation. Ann. Rev. Fluid Mech. 32, 132
507. Meneveau, C., Lund, T.S. (1997): The dynamic Smagorinsky model and scaledependent coecients in the viscous range of turbulence. Phys. Fluids 9(12),
39323934
508. Meneveau, C., Lund, T.S., Cabot, W.H. (1996): A Lagrangian dynamic
subgrid-scale model of turbulence. J. Fluid Mech. 319, 353385
509. Meneveau, C., Lund, T.S., Moin, P. (1992): Search for subgrid scale
parametrization by projection pursuit regression. Proceedings of the Summer
Program Center for Turbulence Research, 6181
510. Meneveau, C., ONeil, J. (1994): Scaling laws of the dissipation rate of turbulent subgrid-scale kinetic energy. Phys. Rev. E. 49(4), 28662874
511. Menon, S., Yeung, P.K., Kim, W.W. (1996): Eect of subgrid models on the
computed interscale energy transfer in isotropic turbulence. Computer and
Fluids 25(2), 165180
512. Mestayer, P. (1982): Local isotropy and anisotropy in high-Reynolds-number
turbulent boundary layer. J. Fluid Mech. 125, 475503
513. Metais, O., Herring, J. (1989): Numerical simulations of freely evolving turbulence in stably stratied uids. J. Fluid Mech. 202, 117148
514. Metais, O., Lesieur, M. (1992): Spectral large-eddy simulation of isotropic and
stably stratied turbulence. J. Fluid Mech. 256, 157194
515. Meyers, J., Geurts, B., Baelmans, M. (2003): Database analysis of errors in
large-eddy simulation. Phys. Fluids 15(9), 27402755
516. Michelassi, V., Wissink, J.G., Frohlich, J., Rodi, W. (2003): Large eddy simulation of ow around low-pressure turbine blade with incoming wakes. AIAA
Journal 41(11), 21432156
517. Misra, A., Pullin, D.I. (1997): A vortex-based subgrid stress model for largeeddy simulation. Phys. Fluids 9(8), 24432454
518. Mitran, S. (2001): A comparison of adaptive mesh renement approaches for
large-eddy simulation. (DNS/LES Progress and Challenges, Liu, Sakell and
519. Mittal, R. (1995): Large-eddy simulation of ow past a circular cylinder. Annual Research Briefs Center for Turbulence Research, 107117
520. Mittal, R. (1996): Progress on LES of ow past a circular cylinder. Annual
Research Briefs Center for Turbulence Research, 233243
521. Mittal, R., Moin, P. (1997): Suitability of upwind-biased nite-dierence
schemes for large-eddy simulation of turbulent ows. AIAA Journal 35(8),
14151417
522. Mittal, R., Simmons, S.P., Najjar, F. (2003): Numerical study of pulsatile ow
in a constricted channel. J. Fluid Mech. 485, 337378
523. Miyake, Y., Kajishima, T. (1986): Numerical simulation of the eects of
Coriolis force on the stucture of turbulence. Global eects. Bull. JSME 29,
33413346
538
References
524. Miyake, Y., Kajishima, T. (1986): Numerical simulation of the eects of Coriolis force on the stucture of turbulence. Structure of turbulence. Bull. JSME
29, 33473351
525. Moeng, C.H. (1984): A large-eddy simulation model for the study of planetary
boundary-layer turbulence. J. Atmos. Sci. 41(13), 20522062
526. Moeng, C.H. (1987): Large-eddy simulation of a stratus-topped boundary
layer. Part I: structure and budgets. J. Atmos. Sci. 43(23), 28862900
527. Moeng, C.H. (1987): Large-eddy simulation of a stratus-topped boundary
layer. Part II: implications for a mixed-layer modeling. J. Atmos. Sci. 44(12),
16051614
528. Moeng, C.H., Rotunno, R. (1990): Vertical-velocity skewness in the buoyancydriven boundary layer. J. Atmos. Sci. 47(9), 11491162
529. Moeng, C.H., Schumann, U. (1991): Composite structure of plumes in stratustopped boundary layers. J. Atmos. Sci. 48(20), 22802291
530. Moeng, C.H., Shen, S., Randall, D. (1992): Physical processes within the nocturnal stratus-topped boundary-layer. J. Atmos. Sci. 49(24), 23842401
531. Moeng, C.H., Wyngaard, J.C. (1984): Statistics of conservative scalars in the
convective boundary layer. J. Atmos. Sci. 41(21), 31613169
532. Moeng, C.H., Wyngaard, J.C. (1986): An analysis of closures for pressurescalar covariances in the convective boundary layer. J. Atmos. Sci. 43(21),
24992513
533. Moeng, C.H., Wyngaard, J.C. (1988): Spectral analysis of large-eddy simulations of the convective boundary layer. J. Atmos. Sci. 45(23), 35733587
534. Moeng, C.H., Wyngaard, J.C. (1989): Evaluation of turbulent transport and
dissipation closures in second-order modeling. J. Atmos. Sci. 46(14), 2311
2330
535. Mohammadi, B., Pironneau, O. (2000): Applied shape optimization for uids.
Oxford University Press, UK
536. Moin, P., Jimenez, J. (1993): Large eddy simulation of complex ows. (24th
AIAA Fluid Dynamics Conference, Orlando, USA)
537. Moin, P., Kim, J. (1982): Numerical investigation of turbulent channel ow.
J. Fluid Mech. 118, 341377
538. Moin, P., Squires, K., Cabot, W., Lee, S. (1991): A dynamic subgrid-scale
model for compressible turbulence and scalar transport. Phys. Fluids A 3(11),
27462757
539. Morales, R., Balparda, A., Silveira-Neto, A. (1999): Large-eddy simulation of
the combined convection around a heated rotating cylinder. Int. J. Heat Mass
Transfer 42, 941949
540. Morikawa, H., Maruyama, T. (1999): Theory of conditional random eld and
its application to wind engineering. Kyoto University Press
541. Morinishi, Y., Kobayashi, T. (1990): in (Engineering turbulence modelling and
experiments, Rodi and Ganic eds.) Elsevier, New York, 279
542. Morinishi, Y., Vasilyev, O. (1998): Subgrid scale modeling taking the numerical error into consideration. Annual Research Briefs Center for Turbulence
Research, 237253
543. Morinishi, Y., Vasilyev, O.V. (2001): A recommended modication to the
dynamic two-parameter mixed subgrid scale model for large eddy simulation
of turbulent ow. Phys. Fluids, submitted
544. Morinishi, Y., Vasilyev, O. (2002): Vector level identity for dynamic subgrid
scale modeling in large eddy simulation. Phys. Fluids 14(10), 36163623
545. Morris, P.J., Long, L.N., Bangalore, A., Wang, Q. (1997): A parallel threedimensional computational aeroacoustics method using nonlinear disturbance
equations. J. Comput. Phys., 133, 5674
References
539
546. Morris, P.J., Long, L.N., Wang, Q., Lockard, D.P. (1997): Numerical prediction of high-speed jet noise. AIAA Paper 97-1598
547. Morris, P.J., Long, L.N., Wang, Q., Pilon, A.R. (1998): High-speed jet noise
simulations. AIAA Paper 98-2290
548. Moser, R.D., Langford, J.A., Volker, S (2001): A radical approach to large
eddy simulation. AIAA Paper 2001-2835
549. Moser, R.D., Volker, S., Venugopal, P. (2001):Characterization of optimal LES
in turbulent channel ow. (DNS/LES Progress and Challenges, Liu, Sakell and
550. Mossi, M. (1999): Simulation of benchmark and industrial unsteady compressible turbulent uid ows. PhD Thesis No 1958, EPFL, Lausanne, Switzerland
551. Muchinsky, A. (1996): A similarity theory of locally homogeneous anisotropic
turbulence generated by a Smagorinsky-type LES. J. Fluid Mech. 325,
239260
552. Mukerji, S., McDonough, J.M., Menguc, M.P., Manickavasagam, S., Chung, S.
(1998): Chaotic map models of soot uctuations in turbulent diusion ames.
Int. J. Heat Mass Transfer 41, 40954112
553. Mullen, J.S., Fischer, P.F. (1999): Filtering technique for complex geometries
uid ows. Commun. Numer. Meth. Engng. 15, 918
554. Murakami, S. (1993): Comparison of various turbulence models applied to
a blu body. J. Wind Eng. Ind. Aerodyn. 46 & 47, 2136
555. Murakami, S., Iizuka, S., Ooka, R. (1999): CFD analysis of turbulent ow
past square cylinder using dynamic LES. Journal of Fluids and Structures 13,
10971112
556. Murakami, S., Mochida, A., Hibi, K. (1987): Three-dimensional numerical
simulation of air ow around a cubic model by means of large-eddy simulation.
J. Wind Eng. Ind. Aerodyn. 25, 291305
557. Murata, A., Mochizuki, S. (1999): Eect of cross-sectional aspect ratio on
turbulent heat transfer in an orthogonally rotating rectangular smooth duct.
558. Murata, A., Mochizuki, S. (2000): Large eddy simulation with a dynamic
subgrid-scale model of turbulent heat transfer in an orthogonally rotating
rectangular duct with transverse rib turbulators. Int. J. Heat Mass Transfer
43, 12431259
559. Murata, A., Mochizuki, S. (2001): Eect of centrifugal buoyancy on turbulent
heat transfer in an orthogonally rotating square duct with transverse or angled
rib turbulators. Int. J. Heat Mass Transfer 44, 27392750
560. Murata, A., Mochizuki, S. (2003): Eect of cross-sectional aspect ratio on turbulent heat transfer in an orthogonally rotating rectangular duct with angled
rib turbulators. Int. J. Heat Mass Transfer 46, 31193133
561. Murata, A., Mochizuki, S. (2004): Large eddy simulation of turbulent heat
transfer in a rotating two-pass smooth square channel with sharp 180o turns.
562. Murray, J.A., Piomelli, U., Wallace, J.M. (1996): Spatial and temporal ltering
of experimental data for a priori studies of subgrid-scale stresses. Phys. Fluids
8(7), 19781980
563. Najjar, F.M., Tafti, D.K. (1996): Study of discrete test lters and nite dierence approximations for the dynamic subgrid-scale stress model. Phys. Fluids
8(4), 10761088
564. Nakayama, A., Noda, H. (2000): Large-eddy simulation of ow around a blu
body tted with splitter plate. J. Wind Eng. Ind. Aerodyn. 85, 8596
540
References
565. Niceno, B., Dronkers, A., Hanjalic, K. (2002): Turbulent heat transfer from
a multi-layered wall-mounted cube matrix: large eddy simulation. Int. J. Heat
566. Nicoud, F., Bagget, J.S., Moin, P., Cabot, W. (2001): Large-eddy simulation
wall-modeling based on sub-optimal control theory and linear stochastic estimation. Phys. Fluids 13(10), 29682984
567. Nicoud, F., Ducros, F. (1999): Subgrid stress modeling based on the square of
the velocity gradient tensor. Flow, Turbulence and Combustion 62(3), 183200
568. Nicoud, F., Winckelmans, G., Carati, D., Bagget, J., Cabot, W. (1998):
Boundary conditions for LES away from the wall. Proceedings of the Summer
Program Center for Turbulence Research, 413422
569. Nikitin, N.V., Nicoud, F., Washito, B., Squires, K.D., Spalart, P.R. (2000):
An approach to wall modeling in large-eddy simulations. Phys. Fluids 12(7),
16291632
570. ONeil, J., Meneveau, C. (1997): Subgrid-scale stresses and their modelling in
a turbulent plane wake. J. Fluid Mech. 349, 253293
571. Oberai, A., Roknaldin, F., Hughes, T.J.R. (2002): Computation of trailingedge noise due to turbulent ow over an airfoil. AIAA Journal 40(11), 2206
2216
572. Oberlack, M. (1997): Invariant modeling in large-eddy simulation of turbulence. Annual Research Briefs Center for Turbulence Research, 322
573. Olsson, M., Fuchs, L. (1996): Large eddy simulation of the proximal region of
a spatially developing circular jet. Phys. Fluids 8(8), 21252137
574. Olsson, M., Fuchs, L. (1998): Large eddy simulations of a forced semiconned
circular impinging jet. Phys. Fluids 10(2), 476486
575. Pallares, J., Cuesta, I., Grau, F. (2002): Laminar and turbulent Rayleigh
Benard convection in a perfectly conducting cubical cavity. Int. J. Heat Fluid
Flow 23, 346358
576. Pallares, J., Davidson, L. (2000): Large-eddy simulation of turbulent ow in
a rotating square duct. Phys. Fluids 12(11), 28782894
577. Pallares, J., Davidson, L. (2002): Large-eddy simulations of turbulent heat
transfer in stationary and rotating square ducts. Phys. Fluids 14(8), 2804
2816
578. Pantano, C., Sarkar, S. (2001): A subgrid model for nonlinear functions of
a scalar. Phys. Fluids 13(12), 38033819
579. Pascal, F., Basdevant, C. (1992):. Nonlinear Galerkin method and subgridscale model for two-dimensional turbulent ows. Theoret. Comp. Fluid Dynamics 3, 267284
580. Pascarelli, A., Piomelli, U., Candler, G.V. (2000): Multi-block large-eddy simulations of turbulent boundary layers. J. Comput. Phys. 157(7), 256279
581. Pasquetti, R.: High-order LES modeling of turbulent incompressible ow, submitted
582. Pasquetti, R., Xu, C. (2002): High-order algorithms for large-eddy simulation
of incompressible ows. J. Sci. Comput. 17(1-4), 273284
583. Patel, N., Stone, C., Menon, S. (2003): Large eddy simulation of turbulent
ow over an axisymmetric hill. AIAA Paper 2003-0967
584. Peltier, L.J., Zajaczkowski, F.J. (2001): Maintenace of the near-wall cycle of
turbulence for hybrid RANS/LES of fully developed channel ow. (DNS/LES
Progress and Challenges, Liu, Sakell and Beutner eds.), Greyden Press,
829834
585. Peneau, F., Boisson, H., Djilali, N. (2000): Large eddy simulation of the inuence of high free-stream turbulence on a spatially evolving boundary layer.
References
541
586. Peng, S.H., Davidson, L. (2001): Large eddy simulation for turbulent buoyant
ow in a conned cavity. Int. J. Heat Fluid Flow 22, 323331
587. Peng, S.H., Davidson, L. (2002): On a subgrid-scale heat ux model for large
eddy simulation of turbulent thermal ow. Int. J. Heat Mass Transfer 45,
13931405
588. Perrier, P., Pironneau, O. (1981): Subgrid turbulence modelling by homogeneization. Mathematical Modelling 2, 295317
589. Persson, L., Fureby, C., Svanstedt, N. (2002): On homogenization-based methods for large-eddy simulation. J. Fluid Engng. 124, 892903
590. Pierce, C., Moin, P. (1998): A dynamic model for subgrid-scale variance and
dissipation rate of a conserved scalar. Phys. Fluids 10(12), 30413044
591. Piomelli, U. (1993): High Reynolds number calculations using the dynamic
subgrid-scale stress model. Phys. Fluids A 5(6), 14841490
592. Piomelli, U., Balaras, E., Pasinato, H., Squires, K., Spalart, P. (2003): The
inner-outer layer interface in large-eddy simulations with wall-layer models.
Int. J. heat Fluid Flow 24, 538550
593. Piomelli, U., Cabot, W.H., Moin, P., Lee, S. (1990): Subgrid-scale backscatter
in transitional and turbulent ows. Proceedings of the Summer Program
Center for Turbulence Research, 1930
594. Piomelli, U., Coleman, G.N., Kim, J. (1997): On the eects of nonequilibrium
on the subgrid-scale stresses. Phys. Fluids 9(9), 27402748
595. Piomelli, U., Ferziger, J.H., Moin, P., Kim, J. (1989): New approximate boundary conditions for large eddy simulations of wall-bounded ows. Phys. Fluids
A 1(6), 10611068
596. Piomelli, U., Liu, J. (1995): Large-eddy simulation of rotating channel ows
using a localized dynamic model. Phys. Fluids 7 (4), 839848
597. Piomelli, U., Moin, P., Ferziger, J.H. (1988): Model consistency in large eddy
simulation of turbulent channel ows. Phys. Fluids 31(7), 18841891
598. Piomelli, U., Yunfang, X., Adrian, R.J. (1996): Subgrid-scale energy transfer
and near-wall turbulence structure. Phys. Fluids 8(1), 215224
599. Piomelli, U., Zang, T.A. (1991): Large-eddy simulation of transitional channel
ow. Computer Physics Communications 65, 224230
600. Piomelli, U., Zang, T.A., Speziale, C.G., Hussaini, M.Y. (1990): On the largeeddy simulation of transitional wall-bounded ows. Phys. Fluids A 2(2),
257265
601. Porte-Agel, F., Meneveau, C., Parlange, M.B. (2000): A scale-dependent dynamic model for large-eddy simulation: application to a neutral atmospheric
boundary layer. J. Fluid Mech. 415, 261284
602. Porter, D.H., Pouquet, A., Woodward, P.R. (1994): Kolmogorov-like spectra
in decaying three-dimensional supersonic ows. Phys. Fluids 6(6), 21332142
603. Pruett, C.D. (2000): Eulerian time-domain ltering for spatial large-eddy simulation. AIAA Journal 38(9), 16341642
604. Pruett, C.D. (2001): Toward the de-mystication of LES. (DNS/LES Progress
and Challenges, Liu, Sakell and Beutner eds.), Greyden Press, 231238
605. Pruett, C.D., Adams, N.A. (2000): A priori analyses of three subgrid-scale
models for one-parameter families of lters. Phys. Fluids 12(5), 11331142
606. Pruett, C.D., Gatski, T.B., Grosch, C.E., Thacker, W.D. (2003): The temporally ltered NavierStokes equations: properties of the residual stress. Phys.
Fluids 15(8), 21272140
607. Pruett, C.D., Sochacki, J.S., Adams, N.A. (2001): On Taylor-series expansions
of residual stresses. Phys. Fluids 13(9), 25782589
608. Pullin, D.I. (2000): A vortex-based model for the subgrid ux of a passive
scalar. Phys. Fluids 12(9), 23112319
542
References
609. Pullin, D.I., Saman, P.G. (1994): Reynolds stresses and one-dimensional
spectra for a vortex model of homogeneous anisotropic turbulence. Phys. Fluids 6(5), 17871796
610. Quemere, P., Sagaut, P.: Zonal multidomain RANS/LES simulations of turbulent ows. Int. J. Numer. Methods Fluids, submitted
611. Quemere, P., Sagaut, P., Couaillier, V. (2000): Une methode multidomaine/multiresolution avec application `
a la simulation des grandes echelles.
C. R. Acad. Sci. Paris, Serie II.b 328, 8790
612. Quemere, P., Sagaut, P., Couaillier, V. (2001): A new multidomain/
multiresolution technique for large-eddy simulation. Int. J. Numer. Methods
Fluids 36, 391416
613. Quirk, J.J. (1991): An adaptative grid algorithm for computational shock
hydrodynamics. PhD Thesis, College of Aeronautics
614. Rajagopalan, S., Antonia, R.A. (1979): Some properties of the large structure
in a fully developed turbulent duct ow. Phys. Fluids 22(4), 614622
615. Redelsperger, J.L., Mahe, F., Carlotti, P. (2001): A simple and general subgrid
model suitable both for surface layer and free-stream turbulence. BoundaryLayer Meteorol. 101, 375408
616. Rizk, M.H., Menon, S. (1988): Large-eddy simulations of axisymmetric excitation eects on a row of impinging jets. Phys. Fluids 31(7), 18921903
617. Robinson, S.K. (1991): The kinematics of turbulent boundary layer structure.
NASA, Tech. Memo. TM 103859
618. Rodi, W., Ferziger, J.H., Breuer, M., Pourquie, M. (1997): Status of large-eddy
simulation: results of a Workshop. ASME J. Fluid Engng. 119(2), 248262
619. Rogallo, R.S., Moin, P. (1984): Numerical simulation of turbulent ows. Ann.
Rev. Fluid Mech. 16, 99137
620. Rollet-Miet, P., Laurence, D., Ferziger, J. (1999): LES and RANS of turbulent
ow in tube bundles. Int. J. Heat Fluid Flow 20, 241254
621. Ronchi, C., Ypma, M., Canuto, V.M. (1992): On the application of the Germano identity to subgrid-scale modeling. Phys. Fluids A 4(12), 29272929
622. Rose, H.A. (1977): Eddy diusivity. eddy noise and subgrid-scale modelling.
J. Fluid Mech. 81(4), 719734
623. Roy, S., Baker, A.J. (1997): Nonlinear, subgrid embedded nite-element basis
for accurate, monotone, steady CFD solutions. Numerical Heat Transfer, Part
B 31, 135175
624. Sadourny, R., Basdevant, C. (1981): Une classe doperateurs adaptes `
a la
modelisation de la diusion turbulente en dimension deux (in french). C. R.
Acad. Sc. Paris 292, 10611064
625. Sadourny, R., Basdevant, C. (1985): Parametrization of subgrid-scale
barotropic and baroclinic eddies in quasi-geostrophic models: anticipated potential vorticity method. J. Atmos. Sci. 42, 13531363
626. Sagaut, P. (1996): Numerical simulations of separated ows with subgrid models. Rech. Aero. 1, 5163
627. Sagaut, P. (1998): Introduction a
` la simulation des grandes echelles pour les
ecoulements de uide incompressible (in french). Springer-Verlag, Berlin
628. Sagaut, P., Comte, P., Ducros, F. (2000): Filtered subgrid-scale models. Phys.
Fluids 12(1), 233236
629. Sagaut, P., Garnier, E., Seror, C. (1999): Generalisation de lidentite de Germano et application `
a la modelisation sous-maille (in french). C. R. Acad. Sci.
Paris, Serie II.b, t. 327, 463466
630. Sagaut, P., Garnier, E., Terracol, M. (2000): A general algebraic formulation
for multi-parameter dynamic subgrid-scale modeling. International Journal of
Computational Fluid Dynamics 13, 251257
References
543
631. Sagaut, P., Garnier, E., Tromeur, E., Larcheveque, L., Labourasse, E. (2004):
Turbulent inow conditions for large-eddy simulation of compressible wallbounded ows. AIAA Journal 42(3), 469477
632. Sagaut, P., Grohens, R. (1999): Discrete lters for large-eddy simulation. Int.
J. Numer. Methods Fluids 31, 11951220
633. Sagaut, P., Labourasse, E., Quemere, P., Terracol, M. (2000): Multiscale approaches for unsteady simulation of turbulent ows. Int. J. Nonlinear Sciences
and Numerical Simulation 1(4)
634. Sagaut, P., Le, T.H. (1997): Some investigations on the sensitivity of largeeddy simulation. (Direct and Large Eddy Simulation II, Chollet, Voke and
Kleiser eds.) Kluwer, 8192
635. Sagaut, P., Montreuil, E., Labbe, O. (1999): Assessment of some self-adaptive
SGS models for wall bounded ows. Aerospace Science & Technology 3(6),
335344
636. Sagaut, P., Tro, B. (1997): Subgrid-scale improvements for non-homogeneous
ows. (Advances in DNS/LES, C. Liu, Z. Liu eds.) Greyden Press
637. Sagaut, P., Tro, B., Le, T.H., Ta, P.L. (1996): Large eddy simulation of
turbulent ow past a backward facing step with a new mixed scale SGS model.
(Computation of three-dimensional complex ows, Notes on Numerical Fluid
Mechanics 53, Deville, Gavrilakis and Rhyming eds.) Vieweg, 271278
638. Saiki, E., Moeng, C.H., Sullivan, P. (2000): Large-eddy simulation of the stably
stratied planetary boundary layer. Boundary-Layer Meteorol. 95, 130
639. Salinas-Vasquez, M., Metais, O. (2002): Large-eddy simulation of the turbulent ow through a heated square duct. J. Fluid Mech. 453, 201238
640. Salvetti, M.V., Banerjee, S. (1994): A priori tests of a new dynamic subgridscale model for nite-dierence large-eddy simulations. Phys. Fluids 7(11),
28312847
641. Salvetti, M.V., Beux, F. (1998): The eect of the numerical scheme on the
subgrid scale term in large-eddy simulation. Phys. Fluids 10 (11), 30203022
642. Salvetti, M.V., Damiani, R., Beux, F. (2001): Three-dimensional coarse largeeddy simulations of the ow above two-dimensional sinusoidal waves. Int. J.
643. Salvetti, M.V., Zang, Y., Street, R.L., Banerjee, S. (1997): Large-eddy simulation of free-surface decaying turbulence with dynamic subgrid-scale models.
Phys. Fluids 9(8), 24052419
644. Sandham, N., Yao, Y., Lawal, A. (2003): Large-eddy simulation sof transonic
turbulent ow over a bump. Int. J. Heat Fluid Flow 24, 584595
645. Sarghini, F., De Felice, G., Santini, S. (2003): Neural networks based subgrid
scale modeling in large-eddy simulations. Computers and Fluids 32(1), 97108
646. Sarghini, F., Piomelli, U., Balaras, E. (1999): Scale-similar models for largeeddy simulations. Phys. Fluids 11 (6), 15961607
647. Schilling, O., Zhou, Y. (2002): Analysis of spectral eddy viscosity and
backscatter in incompressible, isotropic turbulence using statistical closure
theory. Phys. Fluids 14(3), 12441258
648. Schlichting, H., Gersten, K. (2000): Boundary layer theory, 8th revised and
enlarged edition. Springer, Berlin
649. Schl
uter, J.U., Pitsch, H., Moin, P. (2004): Large eddy simulation inow conditions for coupling with Reynolds-averaged ow solvers. AIAA Journal 2(3),
478484
650. Schmidt, R., Kerstein, A., Wunsch, S., Nilsen, V. (2003): Near-wall LES closure based on one-dimensional turbulence modeling. J. Comput. Phys. 186,
317355
544
References
651. Schmidt, H., Schumann, U. (1989): Coherent structure of the convective

boundary layer derived from large-eddy simulations. J. Fluid Mech. 200, 511
562
652. Schmidt, S., Thiele, F. (2002): Comparison of numerical methods applied to
the ow over wall-mounted cubes. Int. J. Heat Fluid Flow 23, 330339
653. Schumann, U. (1975): Subgrid scale model for nite dierence simulations of
turbulent ows in plane channels and annuli. J. Comput. Phys. 18, 376404
654. Schumann, U. (1995): Stochastic backscatter of turbulence energy and scalar
variance by random subgrid-scale uxes. Proc. R. Soc. Lond. A 451, 293318
655. Schumann, U. (1995): Boundary conditions at walls - The unsolved problem.
Unpublished
656. Schumann, U., D
ornbrack, A., D
urbeck, T., Gerz, T. (1997): Large eddy simulation of turbulence in the free atmosphere and behind aircraft. Fluid Dyn.
Res. 20, 110
657. Schumann, U., Moeng, C.H. (1991): Plume uxes in clear and cloudy convective boundary layers. J. Atmos. Sci. 48(15), 17461757
658. Schumann, U., Moeng, C.H. (1991): Plume budgets in clear and cloudy convective boundary layers. J. Atmos. Sci. 48(15), 17581770
659. Schwarz, K.W. (1990): Evidence for organized small-scale structure in fully
developed turbulence. Phys. Rev. Letters 64(4), 415418
660. Scott Collis, S. (2001): Monitoring unresolved scales in multiscale turbulence
modeling. Phys. Fluids 13(6), 18001806
661. Scotti, A., Meneveau, C. (1997): Fractal model for coarse-grained nonlinear
partial dierential equations. Phys. Rev. Lett. 78(5), 867870
662. Scotti, A., Meneveau, C. (1999): A fractal model for large-eddy simulation of
turbulent ow. Physica D 127, 198232
663. Scotti, A., Meneveau, C., Fatica, M. (1997): Dynamic Smagorinsky model on
anisotropic grids. Phys. Fluids 9 (6), 18561858
664. Scotti, A., Meneveau, C., Lilly, D.K. (1993): Generalized Smagorinsky model
for anisotropic grids. Phys. Fluids A 5(9), 23062308
665. Sengupta, T.K., Nair, M.T. (1999): Upwind schemes and large-eddy simulation. Int. J. Numer. Meth. Fluids 31, 879889
666. Seror, C., Sagaut, P., Bailly, C., Juve, D. (2000): Subgrid-scale contribution
to noise production in decaying isotropic turbulence. AIAA Journal 38(10),
17951803
667. Seror, C., Sagaut, P., Bailly, C., Juve, D. (2001): On the radiated noise computed by large-eddy simulation. Phys. Fluids 13(2), 476487
668. Shah, K.B., Ferziger, J.H. (1995): A new non-eddy viscosity subgrid-scale
model and its application to channel ow. Annual Research Briefs Center
for Turbulence Research, 7391
669. Shao, L., Bertoglio, J.P., Cui, G.X., Zhou, H.B., Zhang, Z.S. (2003): Kolmogorov equation for large eddy simulation and its use for subgrid modeling.
Proceedings of the 5th ERCOFTAC Workshop on Direct and Large Eddy
Simulation, Munich, Germany
670. Shao, L., Sarkar, S., Pantano, C. (1999): On the relationship between the
mean ow and subgrid stresses in large eddy simulation of turbulent shear
ows. Phys. Fluids 11(5), 12291248
671. Sheiki, M., Drozda, T., Givi, P., Pope, S. (2003): Velocity-scalar ltered density function for large-eddy simulation of turbulent ows. Phys. Fluids 15(8),
23212337
672. Shen, L., Yue, D.K. (2001): Large-eddy simulation of free-surface turbulence.
J. Fluid Mech. 440, 75116
References
545
673. Silveira Neto, A., Grand, D., Metais, O., Lesieur, M. (1993): A numerical
investigation of the coherent vortices in turbulence behind a backward facing
step. J. Fluid Mech. 256, 125
674. Silvestrini, J. (1996): Simulation des grandes echelles des zones de melange Application `
a la propulsion solide des lanceurs spatiaux (in french). Thèse de
Doctorat de lINPG, Grenoble, France
675. Silvestrini, J.H., Lamballais, E., Lesieur, M. (1998): Spectral-dynamic model
for LES of free and wall shear ows. Int. J. Heat and Fluid Flow 19, 492504
676. Smagorinsky, J. (1963): General circulation experiments with the primitive
equations. I: The basic experiment. Month. Weath. Rev. 91(3), 99165
677. Smirnov, A., Shi, S., Celik, I. (2001): Random ow generation technique for
large-eddy simulations and particle-dynamics modeling. J. Fluids Engng. 123,
359371
678. Sohankar, A., Davidson, L., Norberg, C. (2000): Large eddy simulation of ow
past a square cylinder: comparison of dierent subgrid scale models. J. Fluids
Engng. 122, 3947
679. Souliez, F.J., Long, L.N., Morris, P.J., Sharma, A. (2002): Landing gear aerodynamic noise prediction using unstructured grids. AIAA Paper 2002-0799
680. Spalart, P.R., Jou, W.H., Strelets, M., Allmaras, S.R. (1997): Comments on
the feasability of LES for wings, and on a hybrid RANS/LES approach. (Advances in DNS/LES, C. Liu, Z. Liu eds.) Greyden Press, 137147
681. Speziale, C.G. (1985): Galilean invariance of subgrid-scale stress models in the
large-eddy simulation of turbulence. J. Fluid Mech. 156, 5562
682. Speziale, C.G. (1987): On nonlinear K-l and K- models of turbulence. J. Fluid
Mech. 178, 459475
683. Speziale, C.G. (1991): Analytical methods for the development of Reynoldsstress closures in turbulence. Ann. Rev. Fluid Mech. 23, 107157
684. Speziale, C.G. (1998): A combined large-eddy simulation and time-dependent
RANS capability for high-speed compressible ows. Journal of Scientic Computing 13(3), 253274
685. Speziale, C.G. (1998): Turbulence modeling for time-dependent RANS and
VLES: a review. AIAA Journal 36(2), 173184
686. Spille-Koho, A., Kaltenbach, H.J. (2001): Generation of turbulent inow data
with a prescribed shear stress prole. (DNS/LES Progress and Challenges,
Liu, Sakell and Beutner eds.), Greyden Press, 319326
687. Sreedhar, M., Stern, F. (1998): Large eddy simulation of temporally developing
juncture ows. Int. J. Numer. Meth. Fluids 28, 4772
688. Stanisic, M.M. (1985): The mathematical theory of turbulence. SpringerVerlag, Berlin
689. Staquet, C., Godeferd, F. (1998): Statistical modelling and direct numerical
simulations of decaying stably stratied turbulence. Part 1. Flow energetics.
J. Fluid Mech. 360, 295340
690. Stevens, B., Moeng, C.H., Sullivan, P. (1999): Large-eddy simulations of radiatively driven convection: sensitivities to the representation of small scales.
J. Atmos. Sci. 56, 39633984
691. Stolz, S., Adams, N.A. (1999): An approximate deconvolution procedure for
large-eddy simulation. Phys. Fluids 11 (7), 16991701
692. Stolz, S., Adams, N.A., Kleiser, L. (1999): The approximate deconvolution
model applied to LES of turbulent channel ow. (Direct and Large Eddy
Simulation III, Voke, Sandham and Kleiser eds.) Kluwer, 163174
693. Stolz, S., Adams, N., Kleiser, L. (2001): An approximate deconvolution model
for large-eddy simulations with application to incompressible wall-bounded
ows. Phys. Fluids 13(4), 9971015
546
References
694. Stolz, S., Adams, N., Kleiser, L. (2001): The approximate deconvolution model
for large-eddy simulations of compressible ows and its application to shockturbulent-boundary-layer interaction. Phys. Fluids 13(10), 29853001
695. Strelets, M. (2001): Detached eddy simulation of massively separated ows.
AIAA Paper 2001-0879
696. Su, C.W., Krueger, S., McMurty, P., Austin, P. (1998): Linear eddy modeling of droplet spectral evolution during entrainment and mixing in cumulus
clouds. Atmospheric Research 4748, 4158
697. Sullivan, P., Moeng, C.H., Stevens, B., Lenschow, D., Mayor, S. (1998): Structure of the entrainment zone capping the convective atmospheric boundary
layer. J. Atmos. Sci. 55, 30423064
698. Sullivan, P.P., McWilliams, J.C., Moeng, C.H. (1994): A subgrid-scale model
for large-eddy simulation of planetary boundary-layer ows. Boundary-Layer
Meteorol. 71, 247276
699. Sullivan, P., McWilliams, J.C., Moeng, C.H. (1996): A grid nesting method
for large-eddy simulation of planetary boundary-layer ows. Boundary-Layer
Meteorology 80, 167202
700. Tadmor, E. (1989): Convergence of spectral methods for nonlinear conservation laws. SIAM J. Numer. Anal. 26(1), 3044
701. Tafti, D. (1996): Comparison of some upwind-biased high-order formulations
with a second-order central-dierence scheme for time integration of the incompressible NavierStokes equations. Computers & Fluids 25(7), 647665
702. Tafti, D., Vanka, S.P. (1991): A numerical study of the eects of spanwise
rotation on turbulent channel ow. Phys. Fluids A 3, 642656
703. Tao, B., Katz, J., Meneveau, C. (2000): Geometry and scale relationships in
high Reynolds number turbulence determined from three-dimensional holographic velocimetry. Phys. Fluids 12(5), 941944
704. Tao, B., Katz, J., Meneveau, C. (2002): Statistical geometry of subgrid-scale
stresses determined from holographic PIV measurements. J. Fluid Mech. 457,
3578
705. Temam, R. (1991): Approximation of attractors, large eddy simulations and
multiscale methods. Proc. R. Soc. Lond. A 434, 2339
706. Temmerman, L., Leschziner, M., Melle, C., Froehlich, J. (2003): Investigation
of wall-function approximations and subgrid-scale models in large-eddy simulation of separated ow in a channel with streamwise periodic constrictions.
707. Templeton, J., Wang, M., Moin, P. (2002): Towards LES wall models using
optimization techniques. Annual Research briefs - Center for Turbulence Research, 189200
708. Tennekes, H., Lumley, J.L. (1972): A rst course in turbulence. MIT Press
709. Terracol, M., Sagaut, P. (2003): A multilevel-based dynamic approach for
subgrid-scale modeling in large-eddy simulation. Phys. Fluids 15(12), 3671
3682a
710. Terracol, M., Sagaut, P., Basdevant, C. (2000): Une methode multiniveau pour
la simulation des grandes echelles des ecoulements turbulents compressibles
(in french). C. R. Acad. Sci. Paris, Serie II.b 328, 8186
711. Terracol, M., Sagaut, P., Basdevant, C. (2001): A multilevel algorithm for
large-eddy simulation of turbulent compressible ows. J. Comput. Phys. 167,
439474
712. Terracol, M., Sagaut, P., Basdevant, C. (2003): A time self-adaptive multilevel algorithm for large-eddy simulation with application to the compressible
mixing layer. J. Comput. Phys. 184, 339365
References
547
713. Tessinici, F., Iaccarino, G., Fatica, M., Wang, M., Verzicco, R. (2002): Wall
modeling for large-eddy simulation using an immersed boundary method. Annual Research briefs - Center for Turbulence Research, 181188
714. Townsend, A. (1976): The structure of turbulent shear ows. Cambridge University Press, Cambridge
715. Tremblay, F., Manhart, M., Friedrich, R. (2001): LES of ow around a circular
cylinder at high subcritical Reynolds number with Cartesian grids. (Direct and
large-eddy simulation IV, Geurts, Friedrich and Metais eds.), Kluwer, 329336
716. Tro, B., Le, T.H., Loc, T.P. (1991): A numerical method for the threedimensional unsteady incompressible NavierStokes equations. J. Comput.
Appl. Math. 35, 311318
717. Tsubokura, M. (2001): Proper representation of the subgrid-scale eddy viscosity for the dynamic procedure in large eddy simulation using nite dierence
method. Phys. Fluids 13(2), 500504
718. Tsubokura, M., Kobayashi, T., Taniguchi, N., Jones, W.P. (2003): A numerical
study on the structures on impinging jets excited at the inlet. Int. J. Heat Fluid
Flow 24, 500511
719. Tucker, P. (2004): Novel MILES computations for jet ows and noise. Int. J.
Heat Fluid Flow 25(4), 625635
720. Tucker, P., Davidson, L. (2004): Zonal k-l based large eddy simulations. Computers and Fluids 33, 267287
721. Tutar, M., Holdo, A.E. (2000): Large eddy simulation of a smooth circular
cylinder oscillating normal to a uniform ow. J. Fluids Engng. 122, 694702
722. Tziperman, E., Yavneh, I., Taasan, S. (1993): Multilevel turbulence simulations. Europhys. Lett. 24(4), 239244
723. Ulitsky, M., Collins, L. (2000): On constructing relizable, conservative mixed
scalar equations using the EDQNM theory. J. Fluid Mech. 412, 303329
724. Uzun, A., Blaisdell, G., Lyrintzis, A. (2003): Sensitivity to the Smagorinsky
constant in turbulent jet simulations. AIAA Journal 41(10), 20772079
725. van der Ven, H. (1995): A family of large eddy simulation (LES) lters with
nonuniform lter widths. Phys. Fluids 7(5), 11711172
726. Vandromme, D., Haminh, H. (1991): The compressible mixing layer. (Turbulence and coherent structures, O. Metais, M. Lesieur eds.) Kluwer Academic
Press, 508523
727. Vasilyev, O., Goldstein, D. (2004): Local spectrum of commutation error in
large eddy simulations. Phys. Fluids 16(2), 471473
728. Vasilyev, O., Lund, T.S., Moin, P. (1998): A general class of commutative
lters for LES in complex geometries. J. Comput. Phys. 146, 82104
729. Verstappen, R., Veldman, E. (1997): Direct numerical simulation of turbulence
at lower costs. J. Engng. Math. 32, 143159
730. Verstappen, R., Veldman, E. (1998): Spectro-consistent discretization of
NavierStokes: a challenge to RANS and LES. J. Engng. Math. 34, 163179
731. Verzicco, R., Fatica, M., Iaccarino, G., Moin, P., Khalighi, B. (2002): Large
eddy simulation of a road vehicle with drag-reduction devices. AIAA Journal
40(12), 24472455
732. Verzicco, R., Mohd-Yusof, J., Orlandi, P., Haworth, D. (2000): Large-eddy
simulation in complex geometric congurations using boundary body forces.
AIAA Journal 38(3), 427433
733. Vinokur, M. (1989): An analysis of nite-dierence and nite-volume formulations of conservation laws. J. Comput. Phys 81, 152
734. Visbal, M., Gaitonde, D. (2002): Large-eddy simulation on curvilinear grids
using compact dierencing and ltering schemes. J. Fluid Engng. 124, 836
847
548
References
735. Voelkl, T., Pullin, D.I., Chan, D.C. (2000): A physical-space version of the
stretched-vortex subgrid-stress model for large-eddy simulation. Phys. Fluids
12(7), 18101825
736. Voke, P.R. (1990): Multiple Mesh simulation of turbulent ow. Technical report QMW EP-1082, Queen Mary and Westeld College, University of London, London, U.K.
737. Voke, P.R. (1996): Subgrid-scale modelling at low mesh Reynolds number.
Theoret. Comput. Fluid Dynamics 8, 131143
738. Voke, P., Gao, S. (1998): Numerical study of heat transfer from an impinging
jet. Int. J. Heat Mass Transfer 41(45), 671680
739. Voke, P.R., Gao, S., Leslie, D. (1995): Large-eddy simulations of plane impinging jets. Int. J. Numer. Meth. Fluids 38, 489507
740. Voke, P.R., Potamitis, S.G. (1994): Numerical simulation of a low-Reynoldsnumber turbulent wake behind a ate plate. Int. J. Numer. Meth. Fluids 19,
377393
741. V
olker, S., Moser, R., Venugopal, P. (2002): Optimal large eddy simulation of
turbulent channel ow based on direct numerical simulation statistical data.
Phys. Fluids 14(10), 36753691
742. von Kaenel, R., Adams, N.A., Kleiser, L., Vos, J.B. (2002): The approximate
deconvolution model for large-eddy simulation of compressible ows with nite
volume schemes. J. Fluid Engng. 124, 829835
743. von Kaenel, R., Adams, N.A., Kleiser, L., Vos, J.B. (2003): The approximate
deconvolution model for large-eddy simulation of compresible ows with nite
volume schemes. J. Fluids Engng. 125, 375381
744. von Kaenel, R., Adams, N.A., Kleiser, L., Vos, J.B. (2003): Eect of articial
dissipation on large-eddy simulation with deconvolution modeling. AIAA J.
41(8), 16061609
745. Vreman, A. (2003): The ltering analog of the variational multiscale method
in large-eddy simulation. Phys. Fluids 15(8), L61L64
746. Vreman, B., Geurts, B., Kuerten, H. (1994): Realizability conditions for the
turbulent stress tensor in large-eddy simulation. J. Fluid Mech. 278, 351362
747. Vreman, B., Geurts, B., Kuerten, H. (1994): On the formulation of the dynamic mixed subgrid-scale model. Phys. Fluids 6(12), 40574059
748. Walee, F. (1992): The nature of triad interactions in homogeneous turbulence. Phys. Fluids A 4(2), 350363
749. Walee, F. (1993): Inertial transfer in the helical decomposition. Phys. Fluids
A 5(3), 677685
750. Walee, F. (2001): Exact coherent structures in channel ow. J. Fluid Mech.
435, 93102
751. Wang, M. (2000): Dynamic wall modeling for LES of complex turbulent
ows. Proceedings of the Summer Program Center for Turbulence Research,
241250
752. Wang, L.P., Chen, S., Brasseur, J. (1999): Examination of hypotheses in the
Kolmogorov rened turbulence theory through high-resolution simulations.
Part 2. Passive scalar eld. J. Fluid Mech. 400, 163197
753. Wang, L., Lu, X.Y. (2004): An investigation of turbulent oscillatory heat transfer in channel ows by large eddy simulation. Int. J. Heat Mass Transfer 47,
21612172
754. Wang, M., Moin, P. (2000): Computation of trailing-edge ow and noise using
large-eddy simulation. AIAA Journal 38(12), 22012209
755. Wang, W.P., Pletcher, R. (1997): On the large eddy simulation of a turbulent
channel ow with signicant heat transfer. Phys. Fluids 8(12), 33543366
References
549
756. Weinan, E., Shu, C.W. (1994): A numerical resolution study of high order
essentially non-oscillatory schemes applied to incompressible ow. J. Comput.
Phys. 110, 3946
757. Werner, H., Wengle, H. (1991): Large-eddy simulation of turbulent ow over
and around a cube in a plate channel. (8th Symposium on Turbulent Shear
Flows, Munich, Germany)
758. Wilson, P.G., Pauley, L.L. (1998): Two- and three-dimensional large-eddy simulations of a transitional separation bubble. Phys. Fluids 10(11), 29322940
759. Winckelmans, G.S., Jeanmart, H. (2001): Assessment of some models for LES
without/with explicit ltering. (Direct and large-eddy simulation IV, Geurts,
Friedrich and Metais eds.), Kluwer, 5566
760. Winckelmans, G.S., Jeanmart, H., Carati, D. (2002): On the comparison between Reynolds stresses from large-eddy simulation with those from direct
numerical simulation. Phys. Fluids 14(5), 18091811
761. Winckelmans, G.S., Jeanmart, H., Wray, A., Carati, D., Geurts, B.J. (2001):
Tensor-diusivity mixed model: balancing reconstruction and truncation.
(Modern simulation strategies for turbulent ow, Geurts ed.), Edwards,
85106
762. Winckelmans, G.S., Lund, T.S., Carati, D., Wray, A. (1996): A priori testing of
subgrid-scale models for the velocity-pressure and vorticity-velocity formulations. Proceedings of the Summer Program Center for Turbulence Research,
Stanford, 309329
763. Winckelmans, G., Wray, A.A., Vasilyev, O.V. (1998): Testing of a new mixed
model for LES: the Leonard model supplemented by a dynamic Smagorinsky
term. Proceedings of the Summer Program Center for Turbulence Research,
367388
764. Winckelmans, G.S., Wray, A.A., Vasilyev, O.V., Jeanmart, H. (2001): Explicitltering large-eddy simulation using the tensor-diusivity model supplemented
by a dynamic Smagorinsky term. Phys. Fluids 13(5), 13851403
765. Wong, V.C. (1992): A proposed statistical-dynamic closure method for the
linear or nonlinear subgrid-scale stresses. Phys. Fluids A 4(5), 10801082
766. Wong, V., Lilly, D.K. (1994): A comparison of two subgrid closure methods
for turbulent thermal convection. Phys. Fluids 6(2), 10171023
767. Wu, X., Squires, K.D. (1997): Large eddy simulation of an equilibrium threedimensional turbulent boundary layer. AIAA Journal 35(1), 6774
768. Wu, X., Squires, K.D. (1998): Prediction of the three-dimensional turbulent
boundary layer over a swept bump. AIAA Journal 36(4), 505514
769. Wu, X., Squires, K.D. (1998): Prediction of the high-Reynolds-number ow
over a two-dimensional bump. AIAA Journal 36(5), 799808
770. Wu, X., Squires, K.D. (1998): Numerical investigation of the turbulent boundary layer over a bump. J. Fluid Mech. 362, 229271
771. Wunsch, S., Kerstein, A. (2001): A model for layer formation in stably stratied turbulence. Phys. Fluids 13(3), 702712
772. Xiao, X., Edwards, J., Hassan, H., Baurle, R. (2003): Inow boundary conditions for hybrid large eddy/Reynolds averaged NavierStokes simulations.
AIAA Journal 41(8), 14811489
773. Xu, C., Pasquetti, R. (2004): Stabilized spectral element computations of high
Reynolds number incompressible ows. J. Comput. Phys. 196(2), 680704
774. Xu, H., Pollard, A. (2001): Large eddy simulation of turbulent ow in a square
annular duct. Phys. Fluids 13(11), 33213337
775. Yakhot, V., Orszag, S.A. (1986): Renormalization group analysis of turbulence. I. Basic Theory. J. Sci. Comput. 1, 351
550
References
776. Yan, H., Knight, D. (2002): Large-eddy simulation of supersonic at-plate

boundary layers using the MILES technique. J. Fluid Engng. 124, 868875
777. Yang, Z. (2000): Large eddy simulation of fully developed turbulent ow in
a rotating pipe. Int. J. Numer. Meth. Fluids 33, 681694
778. Yang, K.S. (2000): Large-eddy simulation of turbulent ows in periodically
grooved channel. J. Wind Eng. Ind. Aerodyn. 84, 4764
779. Yang, K.S., Ferziger, J.H. (1993): Large-eddy simulation of turbulent obstacle
ow using a dynamic subgrid-scale model. AIAA Journal 31(8), 14061413
780. Yang, Z., Voke, P.R. (2000): Large-eddy simulation of separated leading-edge
ow in general coordinates. Int. J. Numer. Mech. Engng. 49, 681696
781. Yao, F.W., Sandham, N.D. (2002): DNS of turbulent ow over a bump with
shock/boundary-layer interactions. (Proceedings of the Fifth Engineering Turbulence Modelling and Measurments, Rodi and Fueyo eds.)
782. Yee, P.P., Domaradzki, J.A. (2001): The subgrid-scale estimation model for decaying isotropic turbulence. (Modern simulation strategies for turbulent ow,
Geurts ed.), Edwards, 4556
783. Yeung, P.K. (1994): Spectral transfer of self-similar passive scalar elds in
isotropic turbulence. Phys. Fluids 6(7), 22452247
784. Yeung, P.K., Brasseur, J.G. (1991): The response of isotropic turbulence to
isotropic and anisotropic forcing at large scales. Phys. Fluids A 3(5), 884897
785. Yeung, P.K., Brasseur, J.G., Wang, Q. (1995): Dynamics of direct largesmall scale couplings in coherently forced turbulence: concurrent physicaland Fourier-space views. J. Fluid Mech. 283, 4395
786. Yoshizawa, A. (1979): A statistical investigation upon the eddy viscosity in
incompressible turbulence. J. Phys. Soc. Japan 47(5), 16651669
787. Yoshizawa, A. (1991): Eddy-viscosity-type subgrid-scale model with a variable
Smagorinsky coecient and its relationship with the one-equation model in
large eddy simulation. Phys. Fluids A 3(8), 20072009
788. Yoshizawa, A. (1984): Statistical analysis of the deviation of the Reynolds
stress from its eddy-viscosity representation. Phys. Fluids 27(6), 13771387
789. Yoshizawa, A. (1988): Statistical modeling of passive scalar diusion in turbulent shear ow. J. Fluid Mech. 195, 544555
790. Yoshizawa, A. (1989): Subgrid-scale modeling with a variable length scale
Phys. Fluids A 1(7), 12931295
791. Yoshizawa, A. (1991): A statistically-derived subgrid model for the large-eddy
simulation of turbulence. Phys. Fluids A 3(8), 20072009
792. Yoshizawa, A., Horiuti, K. (1985): A statistically-derived subgrid-scale kinetic
energy model for the large-eddy simulation of turbulent ows. J. Phys. Soc.
Japan 54(8), 28342839
793. Yoshizawa, A., Kobayashi, K., Kobayashi, T., Taniguchi, N. (2000): A nonequilibrium xed-parameter subgrid-scale model obeying the near-wall asymptotic constraint. Phys. Fluids 12(9), 23382344
794. Yoshizawa, A., Tsubokura, M., Kobayashi, T., Taniguchi, N. (1996): Modeling
of the dynamic subgrid-scale viscosity in large-eddy simulation. Phys. Fluids
8(8), 22542256
795. Zahrai, S., Bark, F.H., Karlsson, R.I. (1995): On anisotropic subgrid modeling.
Eur. J. Mech. B/Fluids 14(4), 459486
796. Zajaczkowski, F.J., Peltier, L.J. (2001): Energy-containing-range modeling of
fully-developed channel ow using a hybrid RANS/LES technique. (DNS/LES
Progress and Challenges, Liu, Sakell and Beutner eds.), Greyden Press,
823828
797. Zandonade, P., Langford, J., Moser, R. (2004): Finite-volume optimal largeeddy simulation of isotropic turbulence. Phys. Fluids 16(7), 22552271
References
551
798. Zang, T.A. (1991): Numerical simulation of the dynamics of turbulent boundary layers: Perspectives of a transition simulator. Philos. Trans. R. Soc. Lond.
Ser. A 336, 95102
799. Zang, Y., Street, R.L., Kose, J.R. (1993): A dynamic mixed subgrid-scale
model and its application to turbulent recirulating ows. Phys. Fluids A 5(12),
31863196
800. Zhang, W., Chen, Q. (2000): Large eddy simulation of indoor airow with
a ltered dynamic subgrid scale model. Int. J. Heat Mass Transfer 43, 3219
3231
801. Zhao, H., Voke, P.R. (1996): A dynamic subgrid-scale model for low-Reynoldsnumber channel ow. Int. J. Numer. Meth. Fluids 23, 1927
802. Zhou, Y. (1990): Eect of helicity on renomalized eddy viscosity and subgrid
scale closure for hydrodynamic turbulence. Phys. Rev. A 41(10), 56835686
803. Zhou, Y. (1991): Eddy damping, backscatter, and subgrid stresses in subgrid
modeling of turbulence. Phys. Rev. A 43(12), 70497052
804. Zhou, Y. (1993): Degrees of locality of energy transfer in the inertial range.
Phys. Fluids A 5(5), 10921094
805. Zhou, Y. (1993): Interacting scales and energy transfer in isotropic turbulence.
Phys. Fluids A 5(10), 25112524
806. Zhou, Y., Brasseur, J.G., Juneja, A. (2001): A resolvable sublter-scale model
specic to large-eddy simulation of under-resolved turbulence. Phys. Fluids
13(9), 26022610
807. Zhou, Y., Hossain, M., Vahala, G. (1989): A critical look at the use of lters
in large eddy simulation. Phys. Lett. A. 139(7), 330332
808. Zhou, X., Luo, K., Williams, J. (2001): Study of density eects in turbulent
buoyant jets using large-eddy simulation. Theoret. Comput. Fluid Dynamics
15, 95120
809. Zhou, Y., Vahala, G. (1990): Hydrodynamic turbulence and subgrid scale closure. Phys. Letters A 147, 4348
810. Zhou, Y., Vahala, G. (1992): Loca intercations in renormalization methods for
NavierStokes turbulence. Phys. Rev. A 46(2), 11361139
811. Zhou, Y., Vahala, G. (1993): Renormalization-group estimates of transport
coecients in the advection of a passive scalar by incompressible turbulence.
Phys. Rev. E 48(6), 43874398
812. Zhou, Y., Vahala, G., Hossain, M. (1988): Renormalization-group theory for
the eddy viscosity in subgrid modeling. Phys. Rev. A 37(7), 25902598
813. Zhou, Y., Vahala, G., Hossain, M. (1989): Renormalized eddy viscosity and
Kolmogorovs constant in forced NavierStokes turbulence. Phys. Rev. A
40(10), 58655874
814. Zikanov, O., Slinn, D., Dhanak, M. (2002): Turbulent convection driven by
surface cooling in shallow water. J. Fluid Mech. 464, 81111
Index
Accentuation Technique
156
Bilinear Form 51
Boundary Conditions
Classical approach 324
Embedded boundary conditions
324
General 323
Solid walls
Deterministic minimal boundarylayer unit simulation 337
General 326
O-wall boundary conditions 337
RANS/LES approaches 388
Wall stress models 333
Inow conditions
Arad procedure 360
Deterministic reconstruction,
general 362
Digital Filter 359
General 354
LiWang procedure 358
LLM procedure 356
Lunds extraction/rescaling
technique 363
Precursor simulation 362
Semi-deterministic reconstruction
367
Spille-KohoKaltenbach method
365
Stochastic reconstruction, general
354
SSC procedure 356
WAWS procedure 358
YaoSandham procedure 361
Canonical Analysis 94
Cascade (Anisotropy) 195
Cascade (Kinetic Energy)
Backward 96, 104, 171, 195, 328
Forward 96, 104, 104, 109, 195, 198,

328, 501
Commutator Operator 17, 31, 51
Continuity Breakdown 74
Decomposition
Germano (consistent decomposition)
59, 57, 234
Double 50, 54, 66
Triple, Leonard 50, 54, 60, 66
Deltering 210, 314
Dynamic Procedure (Germano identity
based)
Generalized 149
GermanoLilly 137
Inverse 150
Lagrangian 144
Localized, approximate 148
Localized, constrained 146
One-equation model 173
Taylor series expansion
based 151
With dimensional constant 151
Dynamic Procedures (without Germano
identity) 152
Equivalency Class
General 307
Ideal and Optimal LES 310
Error of Commutation
Commutation with derivatives
74
Boundary terms 31, 323
31,
Filter
Box 21
Convective 21
Convolution product 16, 33
Dierential approximation, denition
26
554
Index
Dierential approximation,
convergence 29
Eective 282, 286
Eective, numerical 290
Elliptic 21
Elliptic, high-order 408
Eulerian time-domain 28, 43, 67
Fundamental properties 17, 33
Gaussian 22, 99
High-Order Commuting (Vasilyev)
38, 42, 77
Implicit 282
Inhomogeneous, anisotropic
31,
187
Invariance properties 64
Isotropic 15
Lagrangian 21
Moments 27
Parabolic 21
Positive 18, 22
Projector 18
Second-Order Commuting Filter
34, 41, 74
Self-similarity 143
Sharp cut-o 22
Smooth 22
Spacetime 29, 43, 66
Time Low-Pass 20, 43
Transfer function 16
Van der Ven 37, 42
Filter, test 131, 137, 204
Filtering the NavierStokes Equations
Conventional approach 45, 48, 74
Alternate approach 45, 77
Generalized Central Moment
59, 234
Germano Identity, additive form 63
Germano Identity, mutiplicative form
Classical 61, 108, 137
Multilevel 63
Generalized 63, 149
Kinetic Energy
Resolved 51, 57
Subgrid 53, 59, 128, 131
Subgrid, generalized 54
Estimates 315
Large-Eddy Simulation
Denition 9, 83
Sensitivity 311
Solution 318
Level of Approximation
Denition 2
Dynamical 3
Spacetime 2
Level of Approximation, usual
Large-Eddy Simulation 7
Optimal base 6
Reynolds Averaged Numerical
Simulation (RANS) 5
Unsteady Reynolds Averaged
Numerical Simulation (RANS)
Local Isotropy 92, 501
MILES/ILES Approach 161, 275, 302

Adaptive ux reconstruction 165
Denition 161
FurebyGrinstein analysis 163
High-order Filtered Methods 170
Spectral Vanishing Viscosity 169
Variational embedded subgrid
stabilization 166, 267
Modeling
Constraints 80
Error 9, 378
Functional 81, 104, 237
Mixed 237
Postulates 79
Statement of the problem 78
Structural 80, 209, 237
Multidomain/Multiresolution Approach
AMR 377
Full overlap
FAS-like 373
General 371
Kravchenko method 374
One-way coupling 372
Two-way coupling 372
General 369
Partial overlap 376
Multilevel Simulations
Dynamic sublter scale model 270
General 263
Local Galerkin Approximation 270
Modied subgrid-scale estimation
procedure 269
Resolvable sublter-scale model
269
Terracol multilevel algorithm 271,
373
Variational multiscale method 267,
166
Multiresolution Decomposition of Data
(Harten) 263
Index
Numerical Error
294
Preltering
9, 161, 290, 161, 378,
289, 294
RANS/LES Coupling
General 383
Nonlinear disturbance equations
390
Universal modeling
Arunajatesan two-equation model
396
BushMani limiters 397
Detached eddy simulation 387
General 391
Germano hybrid model 392
Speziale rescaling 393
Zonal decomposition 384
Link with wall models 388
Sharp transition 385
Smooth transition 387
Scalar eld
Active 472
Passive 450
Scale
Sublter 287
Subgrid 287
Physically resolved 287
Scale Similarity Hypothesis 231
Spectrum, Kinetic Energy
Aupoix 203
Equilibrium 505
HeisenbergChandrasekhar 120
Kolmogorov 94
Kovasznay 120
Pao 120
Production 101
Von Karman 292
Structural Sensor 154
Subgrid Model
Functional model, backward cascade
Bertoglio, stochastic, spectral
179
Chasnov, deterministic, spectral
172
Dynamic, localized, stochastic
184
Dynamic, one-equation, deterministic 173
Leith 180
MasonThomson 182
Schumann 183
555
Functional model, forward cascade

Abba, anisotropic, tensorial 207
Based on kinetic energy at the
cut-o 128
CaratiCabot, anisotropic, tensorial
205
Damping function 159
Dynamic 140
Dynamic, one-equation 173
Filtered 156
Filtered, structure function 157
Horiuti, anisotropic, tensorial 204
ILES approach 161
Local interactions at the cut-o
121
Mixed scale 130
Schumann, anisotropic (splitting)
200
Selective 154
Shao 126
Smagorinsky 124
Smagorinsky, anisotropic, tensorial
192
Spectral, anisotropic, from EDQNM
(Aupoix) 203
Spectral, CholletLesieur 106
Spectral, constant eective viscosity
107
Spectral, dynamic 107
Spectral, LesieurRogallo 108
Spectral, isotropic, from EDQNM
108
Structure function 124
Subgrid viscosity (types) 112
Sullivan, anisotropic (splitting)
201
Viscous eects 118
WALE 199
Weighted Gradient 199
Yoshizawa 129
Structural model
Approximate deconvolution,
general 210
Approximate deconvolution, full
220, 239
Approximate deconvolution, hard
218
Approximate deconvolution, soft
212, 236
Bardina, scale similarity 233
Bardina, scale similarity, ltered
234
Chaotic map 254
556
Index
Clark, dierential approximation

26
Deardor, dierential stress model
243
Deconvolution, dierential
approximation 26
Deconvolution, iterative approximation 212
Deterministic subgrid structure,
kinematic 250
S3/S2 250
S3/ 250
Direct identication 272
Fractal interpolation 253
Fureby, dierential stress model
244
Homogenization based 228
Kerstein, ODT based 257
Kinematic simulation 259
Kosovic, nonlinear 225
Linear stochastic estimation 274,
310
LiuMeneveauKatz, scale
similarity 234
Local average approach 276
LundNovikov, nonlinear 223
Multilevel simulations 263
Neural network 275
Nonlinear, dynamic 226
Scale residual 278
Scale-similarity, dynamic 236
Scale-similarity, generalized 236
Subgrid estimation procedure
261
VFDF, dierential stress model
245
VFDF, Lagrangian Stochastic
model 260
Mixed structural/functional models
SmagorinskyBardina 240
SmagorinskyBardina, dynamic
240
N-parameter, dynamic 241
Subgrid Tensor
Cross stresses 50, 54, 215
Denition
Classical 50
As a commutation error 51
Estimates 215
Invariance properties 64
Leonard stresses 50, 54, 99, 215
Realizability conditions 72
Reynolds stresses 50, 54, 215
Splitting
Mean strain/uctuating strain
330
Rapid part/slow part 237
Test
A posteriori 314
A priori 306
Test Field 131, 204, 231
Triad of Wave Vector 93, 196
Triadic Interaction 93, 196, 501
Viscosity, eective 96, 107
Viscosity, subgrid
Classical 109
Drawbacks 133
Near-wall asymptotic behavior 159
Hyperviscosity 121
Tensorial 192
Wall Model
DasMoser embedded wall model
349
Deardor 337
Ejection 341
Ejection, optimized 341
Ejection, WernerWengle 345
Gr
otzbach 339
Murakami 343
ODT based 350
RANS/LES hybrid approaches 388
Roughness 340
Schumann 339
Shifted correlations 340
Suboptimal-control-based models
345
Thin boundary layer
equations 342
WernerWengle 344
A Computational Method in Plasma Physics
F. Bauer, O. Betancourt, P. Garabechan
Implementation of Finite Element Methods
for Navier-Stokes Equations
F. Thomasset
Finite-Different Techniques
for Vectorized Fluid Dynamics Calculations
Edited by D. Book
Unsteady Viscous Flows
D. P. Telionis
Computational Methods for Fluid Flow
R. Peyret, T. D. Taylor
Computational Methods in Bifurcation
Theory and Dissipative Structures
M. Kubicek, M. Marek
Optimal Shape Design for Elliptic Systems
O. Pironneau
The Method of Differential Approximation
Yu. I. Shokin
Computational Techniques for Fluid

Dynamics 2Specific Techniques for Different
Flow Categories Second Edition
C. A. J. Fletcher
Methods for the Localization of Singularities
in Numerical Solutions
of Gas Dynamics Problems
E. V. Vorozhtsov, N. N. Yanenko
Classical Orthogonal Polynomials
of a Discrete Variable
A. F. Nikiforov, S. K. Suslov, V. B. Uvarov
Flux Coordinates and Magnetic Filed
Structure: A Guide to a Fundamental Tool
of Plasma Theory
W. D. Dhaeseleer, W. N. G. Hitchon,
J. D. Callen, J. L. Shohet
Monte Carlo Methods
in Boundary Value Problems
K. K. Sabelfeld
Computational Galerkin Methods

C. A. J. Fletcher
The Least-Squares Finite Element Method

Theory and Applications in Computational
Fluid Dynamics and Electromagnetics
Bo-nan Jiang
Numerical Methods
for Nonlinear Variational Problems
R. Glowinski
Computer Simulation
of Dynamic Phenomena
M. L. Wilkins
Numerical Methods in Fluid Dynamics

Second Edition M. Holt
Grid Generation Methods

V. D. Liseikin
Computer Studies of Phase Transitions

and Critical Phenomena O. G. Mouritsen
Radiation in Enclosures
A. Mbiock, R. Weber
Finite Element Methods

in Linear Ideal Magnetohydrodynamics
R. Gruber, J. Rappaz
Large Eddy Simulation for Incompressible

Flows An Introduction Second Edition
P. Sagaut
Numerical Simulation of Plasmas

Y. N. Dnestrovskii, D. P. Kostomarov
Higher-Order Numerical Methods

for Transient Wave Equations
G. C. Cohen
Computational Methods for Kinetic Models

of Magnetically Confined Plasmas
J. Killeen, G. D. Kerbel, M. C. McCoy,
A. A. Mirin
Spectral Methods in Fluid Dynamics
Second Edition C. Canuto, M. Y. Hussaini,
A. Quarteroni, T. A. Zang
Computational Techniques for Fluid
Dynamics 1 Fundamental and General
Techniques Second Edition
C. A. J. Fletcher
springeronline.com
Fundamentals of Computational
Fluid Dynamics
H. Lomax, T. H. Pulliam, D. W. Zingg
The Hybrid Multiscale Simulation
Technology An Introduction with
Application to Astrophysical and Laboratory
Plasmas A. S. Lipatov
Computational Aerodynamics and Fluid
Dynamics An Introduction J.-J. Chattot
Nonclassical Thermoelastic Problems
in Nonlinear Dynamics of Shells
Applications of the BubnovGalerkin
and Finite Difference Numerical Methods
J. Awrejcewicz, V. A. Krysko
A Computational Differential Geometry
Approach to Grid Generation V. D. Liseikin
Stochastic Numerics for Mathematical Physics
G. N. Milstein, M. V. Tretyakov
Conjugate Gradient Algorithms
and Finite Element Methods
M. Kr
zek, P. Neittaanm
aki, R. Glowinski,
S. Korotov (Eds.)
Flux-Corrected Transport
D. Kuzmin, R. Lhner, S. Turek (Eds.)
Finite Element Methods
and Their Applications Z. Chen
Mathematics of Large Eddy Simulation
of Turbulent Flows
L. B. Berselli, T. Iliescu, W. J. Layton
Large Eddy Simulation for Incompressible
Flows P. Sagaut
springeronline.com

Large Eddy Simulation For Incompressible Flows

Uploaded by

Copyright:

Available Formats

Large Eddy Simulation For Incompressible Flows

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Large Eddy Simulation For Incompressible Flows

Uploaded by

Copyright:

Available Formats

Scientic Computation

Large Eddy Simulation

With 99 Figures and 15 Tables

Prof. Dr. Pierre Sagaut

Title of the original French edition:

Library of Congress Control Number: 2005930493

Foreword to the Third Edition

Foreword to the Third Edition

the local state of turbulence. This availability of dynamical information has

Foreword to the Third Edition

Foreword to the Second Edition

It is a particular pleasure to present the second edition of the book on Large

Foreword to the Second Edition

Foreword to the Second Edition

To write a textbook obliges one in some way or another to make judgements,

Foreword to the First Edition

Still today, turbulence in uids is considered as one of the most dicult

Foreword to the First Edition

atmosphere and oceans are concerned, because it dissipates the large-scale

He worked as a postdoctoral student with Einstein at Princeton.

Foreword to the First Edition

E.D. Siggia, J. Fluid Mech., 107, pp. 375406 (1981).

Preface to the Third Edition

XVIII Preface to the Third Edition

(stable/unstable stratication eects) cases being discussed. This extension

Preface to the Second Edition

The astonishingly rapid development of the Large-Eddy Simulation technique

Preface to the Second Edition

I am indebted to an increasing number of people, but I would like to

Preface to the First Edition

While giving lectures dealing with Large-Eddy Simulation (LES) to students

Preface to the First Edition

3.3 Decomposition of the Non-linear Term.

Other Mathematical Models for the Large-Eddy

Functional Modeling (Isotropic Case) . . . . . . . . . . . . . . . . . . . . .

5.4.2 Deterministic Statistical Models . . . . . . . . . . . . . . . . . . . . 172

7.7.3 Kersteins ODT-Based Method . . . . . . . . . . . . . . . . . . . . .

Numerical Solution: Interpretation and Problems . . . . . . . . .

Analysis and Validation of Large-Eddy Simulation Data . .

10. Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10.2 Solid Walls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

12. Hybrid RANS/LES Approaches . . . . . . . . . . . . . . . . . . . . . . . . . .

13.2.3 Extension to the General Case. Convolution Filters . . . 407

15. Coupling with Passive/Active Scalar . . . . . . . . . . . . . . . . . . . . . .

15.3.6 More Complex Physical Models . . . . . . . . . . . . . . . . . . . . 492

1.1 Computational Fluid Dynamics

as rotating stall in aeronautical engines. Studies like this concern system

1.2 Levels of Approximation: General

1.3 Statement of the Scale Separation Problem

scale is taken as the resolution reference so as to capture all the dynamic

1.3 Statement of the Scale Separation Problem

in which Re is the Reynolds number, which is a measure of the ratio of

1.4 Usual Levels of Approximation

1.4 Usual Levels of Approximation

By calculating directly only certain low-frequency modes in time (of the

tion (URANS), Semi-Deterministic Simulation (SDS), Very Large-Eddy

1.4 Usual Levels of Approximation

still as ne as that of the continuum model, but is now optimized. Several

1.5 Large-Eddy Simulation: from Practice to Theory. Structure of the Book