Академический Документы
Профессиональный Документы
Культура Документы
of Turbulent Flows
ICASE/LaRC Series in Computational Science and Engineering
Edited by
Thomas B. Gatski
NASA Langley Research Center
M. Yousuff Hussaini
ICASE
John L. Lumley
Cornell University
1 3 5 7 9 8 6 4 2
Thomas B. Gatski
M. Yousuff Hussaini
John L. Lumley
CONTENTS
PREFACE v
INTRODUCTION 1
1 INTRODUCTION 5
1.1 The Energy Cascade in the Spectrum in Equilibrium
Flows 6
1.2 Kolmogorov Scales 9
1.3 Equilibrium Estimates for Dissipation 10
1.4 The Dynamics of Turbulence 11
2 EQUILIBRIUM AND NON-EQUILIBRIUM FLOWS 13
2.1 The Spectral Cascade in Non-Equilibrium Flows ... 13
2.2 Delay in Crossing the Spectrum 14
2.3 Negative Production 19
2.4 Mixing of Fluid with Different Histories 20
2.5 Deformation Work in Equilibrium and
Non-Equilibrium Situations 23
2.6 Alignment of Eigenvectors 25
2.7 Dilatational Dissipation and Irrotational Dissipation 26
2.8 Eddy Shocklets 28
3 PROPER ORTHOGONAL DECOMPOSITION AND
WAVELET REPRESENTATIONS 29
3.1 Coherent Structures 29
3.2 The Role of Coherent Structures in Turbulence
Dynamics 32
3.3 The POD as a Representation of Coherent
Structures . 33
3.4 Low-Dimensional Models Constructed Using
the POD 37
3.5 Comparison with the Wall Region 42
3.6 Generation of Eigenfunctions from Stability
Arguments 52
3.7 Wavelet Representations 67
viii Contents
1 INTRODUCTION 79
2 PROBLEM OF NUMERICAL SIMULATION 80
3 SIMULATION OF HOMOGENEOUS
INCOMPRESSIBLE TURBULENCE 85
4 WALL-BOUNDED AND INHOMOGENEOUS FLOWS 86
5 FAST, VISCOUS VORTEX METHODS 91
6 SIMULATION OF COMPRESSIBLE TURBULENCE 100
7 REFERENCES 104
1 INTRODUCTION 109
2 TURBULENCE AND ITS PREDICTION 111
2.1 The Nature of Turbulence 111
2.2 RANS Models 112
2.3 Direct Numerical Simulation (DNS) 115
3 FILTERING 116
4 SUBGRID SCALE MODELING 118
4.1 Physics of the Subgrid Scale Terms 118
4.2 Smagorinsky Model 119
4.3 A Priori Testing 123
4.4 Scale Similarity Model 125
4.5 Dynamic Procedure 127
4.6 Spectral Models 132
4.7 Effects of Other Strains 135
4.8 Other Models 137
5 WALL MODELS 138
6 NUMERICAL METHODS 141
7 ACCOMPLISHMENTS AND PROSPECTS 143
8 COHERENT STRUCTURE CAPTURING 146
8.1 The Concept 146
Contents ix
1 INTRODUCTION 155
2 PERTURBATION THEORY FOR THE
NAVIER-STOKES EQUATIONS 159
3 RENORMALIZATION GROUP METHOD FOR
RESUMMATION OF DIVERGENT SERIES 162
4 TRANSPORT MODELING 169
5 REFERENCES 182
1 INTRODUCTION 185
2 INCOMPRESSIBLE TURBULENT FLOWS 187
2.1 Reynolds Averages 187
2.2 Reynolds-Averaged Equations 189
2.3 The Closure Problem 189
2.4 Older Zero- and One-Equation Models 190
2.5 Transport Equations of Turbulence 192
2.6 Two-Equation Models 193
2.7 Full Second-Order Closures 210
3 COMPRESSIBLE TURBULENCE 220
3.1 Compressible Reynolds Averages 221
3.2 Compressible Reynolds-Averaged Equations 221
3.3 Compressible Reynolds Stress Transport Equation . . 223
3.4 Compressible Two-Equation Models 226
3.5 Illustrative Examples 227
4 CONCLUDING REMARKS 234
5 REFERENCES 236
x Contents
6 AN INTRODUCTION TO SINGLE-POINT
CLOSURE METHODOLOGY 243
Brian E. Launder
1 INTRODUCTION 243
1.1 The Reynolds Equations 243
1.2 Mean Scalar Transport 245
1.3 The Modeling Framework 246
1.4 Second-Moment Equations 247
1.5 The WET Model of Turbulence 253
2 CLOSURE AND SIMPLIFICATION OF THE
SECOND-MOMENT EQUATIONS 255
2.1 Some Basic Guidelines 255
2.2 The Dissipative Correlations 257
2.3 Non-Dispersive Pressure Interactions 258
2.4 Diffusive Transport dij, dio 273
2.5 Determining the Energy Dissipation Rate 275
2.6 Simplifications to Second-Moment Closures 278
2.7 Non-Linear Eddy Viscosity Models 281
3 LOW REYNOLDS NUMBER TURBULENCE NEAR
WALLS 284
3.1 Introduction 284
3.2 Limiting Forms of Turbulence Correlations
in the Viscous Sublayer 286
3.3 Low Reynolds Number Modelling 288
3.4 Applications 299
4 REFERENCES 302
INDEX 311
LECTURERS
The aim of this book is to provide the engineer and scientist with the
necessary understanding of the underlying physics of turbulent flows,
and to provide the user of turbulence models with the necessary back-
ground on the subject of turbulence to allow them to knowledgeably
assess the basis for many of the state-of-the-art turbulence models.
While a comprehensive review of the entire field could only be
thoroughly done in several volumes of this size, it is necessary to focus
on the key relevant issues which now face the engineer and scientist
in their utilization of the turbulent closure model technology. The
organization of this book is intended to guide the reader through the
subject starting from key observations of spectral energy transfer and
the physics of turbulence through to the development and application
of turbulence models.
Chapter 1 focuses on the fundamental aspects of turbulence phys-
ics. An insightful analysis of spectral energy transfer and scaling
parameters is presented which underlies the development of phe-
nomenological models. Distinctions between equilibrium and non-
equilibrium turbulent flows are discussed in the context of modifica-
tions to the spectral energy transfer. The non-equilibrium effects of
compressibility are presented with particular focus on the alteration
to the turbulent energy dissipation rate. The important topical issue
of coherent structures and their representation is presented in the
latter half of the chapter. Both Proper Orthogonal Decomposition
and wavelet representations are discussed.
With an understanding of the broad dynamic range covered by
both the turbulent temporal and spatial scales, as well as their modal
interactions, it is apparent that direct numerical simulation (DNS) of
turbulent flows would be highly desirable and necessary in order to
capture all the relevant dynamics of the flow. Such DNS methods,
in which all the important length scales in the energy-containing
1
2 Introduction
FUNDAMENTAL ASPECTS
OF INCOMPRESSIBLE AND
COMPRESSIBLE
TURBULENT FLOWS
John L. Lumley
Gal Berkooz, Juan Elezgaray, Philip Holmes
Andrew Poje, Cyril Volte
1 INTRODUCTION
5
6 J. L. Lumley et al.
turbulence, and the other associated with the strain rate to which the
flow is subjected. Or, a turbulence may have different length scales
in different directions. Ordinary turbulence modeling is restricted
to situations that can be approximated as having a single scale of
length and velocity. Turbulence with multiple scales is much more
complicated to predict. Some progress can be made by applying
rapid distortion theory, or one or another kind of stability theory,
to the initial turbulence, and predicting the kinds of structures that
are induced by the applied distortion. We will talk more about this
later. For now, we will restrict ourselves to a turbulence that has a
single scale of length in the energy containing range. We may take
this to be the integral length scale:
flow has a size not much larger than /. Hence, we must seek another
physical entity. Traditionally, we have talked about "eddies," but
these have never been very well denned. We want to introduce here
the wavelet. We will talk more later about wavelets, and how to use
them for a complete representation of the velocity field, and how to
construct physical models using wavelets. For now, we want to use
the simplest properties of wavelets. If we consider a clump of Fourier
modes, as the band in Fourier space becomes wider, the extent in
physical space becomes narrower. A reasonable size appears to be a
band lying between about 1.62/c and K/1.62; the numerical value can
be obtained by requiring that KO, — K/a = K, from which a = 1.62.
That is, the bandwidth is equal to the center wavenumber. This
results in a wavelet in physical space that is confined to a distance
of about a wavelength.
We can now discuss how these wavelets interact in Fourier space.
Reasoning in the crudest way, a wavelet exists in the strain rate
field of all larger wavelets. This strain rate field induces anisotropy
in the wavelet, which permits it to extract energy from the larger
wavelets. This energy extraction process is associated with vortex
stretching: when a vortex is stretched by a strain rate field, the
strain rate field does work on the vortex, increasing its energy, and
losing energy in the process. This process, of extracting energy from
the larger wavelets, and feeding it to the smaller wavelets, is known as
the energy cascade. In Tennekes and Lumley (1972), it is shown that
the cascade is not particulatly tight — a given wavelet receives half its
energy supply from the immediately adjacent larger wavelet, and the
other half from all its neighbors. Similarly, of all the energy crossing a
given wavenumber, three quarters goes to the next adjacent smaller
wavelet, and the remainder is distributed to all the even smaller
wavelets. Nevertheless, we can to a crude approximation consider
that the energy enters the spectrum at the energy containing scales,
and then is passed from wavelet to wavelet across the spectrum until
it arrives at the dissipative range, where it is converted to heat.
There is some discussion in the literature of what is called "back-
scatter." This is an unfortunate choice of words, since it suggests that
energy that started in one direction, is turned around and ends up
going in the other direction. This is not at all what is meant. Rather,
this refers to transfer of energy in the spectrum in the direction from
small to large wavelets. Now, there is no question that, taking av-
erages over long times or large regions of space, the energy transfer
8 J. L. Lnmley et al
where we are assuming that the average dissipation < £ > is deter-
mined by averaging over regions of the size of the integral scale. This
is true in most flows - i.e.- the integral scale is of the order of the size
of the flow. Using the definition of r)m, we can easily obtain
where RI — ul/v.
Fundamental Aspects 11
where < . > denotes some kind oi averaging (possibly time; one,
two or three-dimensional space; phase; or ensemble), but if we do
not indicate otherwise, we will take the average to be an ensemble
average. For the moment we will consider the flow to be incompress-
ible, Uiti = 0, and later we will consider what modifications we will
have to make for the case of compressibility. We are also designating
d ( . ) / d x j = (.),j. Then, the mean strain rate 5,-j is = ([/,-j + Ujii)/2,
and e is the dissipation of turbulent fluctuating kinetic energy. Prop-
erly speaking e = 1v < SijSjj >, where Sij is the strain rate of the
fluctuating motion; however, at high Reynolds number this can be
written as v < UijUij >. We have neglected a number of other terms
in equation (1.4.1) which can be shown to be small at high Reynolds
numbers. These are all of the form of transport terms - that is, they
can be written as divergences of something, and hence, if integrated
over a closed region, contribute nothing to the net turbulent kinetic
energy budget, but simply move kinetic energy from place to place.
These neglected terms are of importance in the neighborhood of the
wall, for example, where the local Reynolds number is low.
Recall that we are supposing that < . > is a long time average,
or an average over the full space, or an ensemble average. In an
inhomogeneous flow (where, of course, we must use either a time
average or an ensemble average), such a quantity may be a function
of position, or in an instationary flow (where we must use either a
space average or an ensemble average) it may be a function of time,
but it will be a smooth function, and will not be a random variable.
This was the kind of average envisioned originally in the early work
of Kolmogorov (1941). There are, of course, other possibilities. For
12 J. L. Lumley et al.
moments, and hence all fluxes of the form < U{U3u^ > will be zero,
and hence < (g 2 /2)u; > will vanish. In a homogeneous flow, all
transport vanishes, since the transport terms are of the form of a
divergence, and spacial derivatives are all zero in a homogeneous sit-
uation. More than this, however, by crude physical reasoning, we
expect that probably all fluxes will vanish in a homogeneous flow - if
everything is statistically the same everywhere, there is no reason for
anything to flow from one place to another. Hence, non-zero fluxes,
and thus transport, are associated with a departure from a Gaus-
sian probability density, specifically with the appearance of skewness
in the density, associated with inhomogeneity. A formal expansion
has been developed for the case of weak inhomogeneity, relating the
third moments with the gradients (Lumley, 1978), which does not
work badly in practice (Panchapakesan and Lumley, 1993).
Finally, the last term is the dissipation of turbulent kinetic energy
per unit mass discussed previously. This is, technically, the rate at
which turbulent fluctuating kinetic energy per unit mass is converted
irreversibly to heat (to entropy). We will see later, however, that in
an equilibrium situation, this is also the rate at which kinetic energy
is removed from the energy containing scales, and the rate at which
kinetic energy is passed from scale to scale across the spectrum. In
a non-equilibrium situation, of course, these three quantities are not
necessarily equal to each other.
ber, then the time required to cross the spectrum is something like
I(O,K) / U(O,K,) + l(a3K,)/u(a3K) + l(a5K)/u(a5K) + ... + l(aN K,}/u(aN K),
where K is at the peak of the energy containing range, say nl = 1.3,
and aNKr/ = 0.55, to place this at the peak of the dissipation range.
Note that these eddies, which we introduced in 1972 are, in fact,
wavelets, as pointed out by Sreenivasan (Zubair et al. 1992). Adding
this up (supposing that all terms are within the inertial subrange)
we find that the total time T = 2(//u)(l - 1.29JRJT1/2), where / and u
are scales characteristic of the energy containing range. Bear in mind
that we should not pay too much attention to the numerical values
of the coefficients in the expression for T; probably the only thing
that is significant is the general form, and the value of the exponent.
Note also that, for low Reynolds number, the time shrinks to a very
small value, since the energy is dissipated at essentially the same
wavenumber where it is produced, while for high Reynolds number
it goes to 2l/u. Now, the idea of a simple lag suggests that the cas-
cade is tight, that all the energy must pass through each wavenumber
16 J. L. Lumley et al.
Figure 2.2.2. Simplified view of spectral flux being passed from eddy
to eddy.
Figure 2.2.3. More realistic view of spectral liux. Now the flux
crossing the wavenumber K goes mostly to eddy &K, but a decreasing
fraction goes to a 3 K, a 5 K, etc. In its turn (at the second step), that
which had gone to a,K is redistributed to a 3 K, a 5 K, etc., at the same
time that which had gone to a3K is redistributed to &5K, etc., and
that which had gone to a 5 K is redistributed to a 7 K and a 9 K, and . ..
This is only the second step. On the third step, each packet must
again be redistributed.
where c" is another constant of order unity. Note that this is now
a non-local theory. Stan Corrsin (1975) pointed out some years ago
that the k — e model was also non-local for similar reasons.
If we apply this model to grid turbulence we may obtain some
relations among the constants. We should identify S with (propor-
tional to) the value of u/l determined by whatever mechanism. In
a grid turbulence, the initial value of u/l is determined by the grid;
thereafter, there is no further input to determine the value of u/l
(since Sjj is identically zero) and it simply relaxes, or decays. In
an equilibrium homogeneous shear flow (which may not exist), S
will take on the value [SijSij]1/2 asymptotically, which will have the
value f/'/A/2 = u/l^/2. If we identify the initial value S0 (at time
Fundamental Aspects 23
let the positive strain rate be in the x3 direction, and the negative
strain rate be in the x-i direction. This is easy to arrange in practice -
the direction that had been shrinking now begins to expand, and vice
versa. Experimentally, it is found that the production immediately
becomes negative throughout the tunnel and remains so until the
anisotropy of the turbulence can adjust itself to the new value of the
strain rate field, which takes some time. Any model for turbulence
which hopes to deal with non-equilibrium situations must take this
into account - this means, in practice, that a separate equation must
be carried for the Reynolds stress; only in equilibrium situations can
the value of the Reynolds stress be related directly to the strain rate
field.
The second term may be written as < u^jUj^ > — < O2 > (in a homo-
geneous flow), where 9 = w ti ,. Hence, we can write (in a homogeneous
situation)
In addition, we may write < TIJTIJ >=< Wj-u;; > /2, and < SijSij > =
< s'-s'- > + < O2 > /3, so that we may write
28 J. L. Lumley et ad.
the projection, which can be given by the mean square of its absolute
value. The calculus of variations reduces this problem of maximiza-
tion to a Fredholm integral equation of the first kind whose sym-
metric kernel is the autocorrelation matrix. The properties of this
integral equation are given by the Hilbert Schmidt theory. There is
a denumerable set of eigenfunctions (structures). The eigenfunctions
form a complete orthogonal set, which means that the random field
can be reconstructed. The coefficients are uncorrelated and their
mean square values are the eigenvalues themselves. The kernel can
be expanded in a uniformly and absolutely convergent series of the
eigenfunctions and the turbulent kinetic energy is the sum of the
eigenvalues. Thus, every structure makes an independent contribu-
tion to the kinetic energy and Reynolds stress. The most significant
point of the decomposition is perhaps the fact that the convergence
of the representation is optimally fast since the coefficients of the
expansion have been maximized in a mean square sense. The mean
square of the first coefficient is as large as possible, the second is
the largest in the remainder of the series once the first term has
been subtracted, etc. We have described here the simplest case, that
of a completely inhomogeneous, square-integrable, field. If the ran-
dom field is homogeneous in one or more directions, the spectrum of
the eigenvalues becomes continuous, and the eigenfunctions become
Fourier modes, so that the proper orthogonal decomposition reduces
to the harmonic orthogonal decomposition in those directions. See
Lumley (1967, 1970, 1981) for more details.
The flow of interest here is three dimensional, approximately ho-
mogeneous in the stream wise direction ( x \ ) and span wise direction
(23), approximately stationary in time (/), inhomogeneous and of
integrable energy in the normal direction (x^).
We want a three dimensional decomposition which can be sub-
stituted in the Navier-Stokes equations in order to recover the phase
information carried by the coefficients. We have to decide which
variable we want to keep. Time is a good candidate since we are
particularly interested in the temporal dynamics of the structures.
Such a decomposition is possible and we do not need a separation of
variables in the eigenfunctions of the type <f>(x,u) = A(ui)tl>(x} (as
suggested by Glauser et a/., 1985) if we do not use any decomposi-
tion in time and choose the appropriate autocorrelation tensor. The
idea is to measure the two velocities at the same time and determine
< Ui(xi, £2, £3, t~)uj(x'i, x'2, £3,2) >= Rij. Since the flow is quasista-
36 J. L. Lumley et al.
tionary, Rij does not depend on time and nor do the eigenvalues and
eigenfunctions. The information in time is carried by the coefficients
a(") which are still "stochastic," but now evolve under the constraint
of the equations of motion. Thus the decomposition becomes
and we have to solve equation (3.3.2) for each pair of wave numbers
(&i, £3). 4>ij now denotes the Fourier transform of the autocorrelation
tensor in the #1,3:3 directions.
Our second change to the decomposition is a transformation of
the Fourier integral into a Fourier series, assuming that the flow is
periodic in the x^ and x3 directions. The periods L\,Lz are de-
termined by the first non zero wave numbers chosen. Finally, each
component of the velocity field can be expanded as the triple sum
and the entire velocity field is recovered by the sum of all the struc-
tures (over n).
The candidate flow we are investigating is the wall region (which
reaches x 2 + = 40) of a pipe flow with almost pure glycerine (98%) as
the working fluid (Herzog, 1986). The Reynolds number based on the
centerline mean velocity and the diameter of the pipe is 8750. The
corresponding Reynolds number based on the shear velocity UT is 531.
From this data the autocorrelation tensor at zero time lag (t — t' = 0)
between the two velocities, RIJ(X\ — x ' l , X 2 , x ' 2 , X 3 — x'3)t_ti-0, was ob-
tained and the spatial eigenfunctions were extracted by numerical
solution of the eigenvalue problem. The results show that approxi-
mately 60% of the total kinetic energy and Reynolds stress is con-
tained in the first eigenmode and that the first three eigenmodes
capture essentially the entire flow field as far as these statistics are
Fundamental Aspects 37
and
Fundamental Aspects 39
field as tij>, while an average of <j>, say, over the unresolved modes
(the filtering process) can be designated as < d> >^. Thus:
with
and
Here < denotes the sum over all the modes (ki,k3,n) such that
k\ < kic,ks < &3C, n < nc and > denotes the sum over all the
modes (ki,kz,n) such that k\ > k\c or ^3 > k%c or n > n c , where
(k\c, ksc, n c ) is the cut-off mode. The characteristic scales of the
parameter Vf are those of the higher modes. We have introduced
an explicit dimensionless parameter ai, and will exclude adjustable
constants from i>x- By observation that the energy decreases rapidly
with increasing n and fc, we assume that these relevant scales are
given by characteristic scales of the first neglected modes. This is
probably a good approximation as far as the eigenmodes are con-
cerned since they are separated by large gaps in the spectrum and it
is a reasonable assumption for the Fourier modes since the steps of
our Fourier series are also large.
Finally, the parameter VT is taken equal to
(where u> and /> are characteristic scales of the neglected modes).
This can be expressed in terms of the eigenvalues and eigenfunctions
of the first neglected modes in the three directions (see figure 3.4.1).
We will refer to «i as a Heisenberg parameter. We will adjust a\
upward and downward to simulate greater and smaller energy loss
to the unresolved modes, corresponding to the presence of a greater
or smaller intensity of smaller scale turbulence in the neighborhood
of the wall. This might correspond, for example, to the environment
just before or just after a bursting event, which produces a large
Fundamental Aspects 41
where L and L' represent the linear terms, Q the direct quadratic
terms, Q' the quadratic pseudo-pressure term and C the cubic terms
arising from the Reynolds stress.
istic length scales is of order ten (see figure 3.5.1.). Our first approx-
imation therefore neglects streamwise variations. We need at least
three terms in the spanwise direction (see below). Thus the minimum
truncation consists of one eigenmode, one streamwise wave number
(&! = 0) and three (i.e. two active) spanwise wave numbers (0, k, 2k).
In this paper, up to six spanwise wavenumbers (0,..., 5&) will be con-
sidered. In this case k has the value 3 X 10~3 and the lengths of the
periodic box are L\ — Ly, — 333. Even the model having six spanwise
Fourier modes is still very crude, although the truncation seems to
be the one which contains an optimally large amount of the total
44 J. L. Lumley et al.
energy among those of the same dimension. The zero cut-off mode
in the streamwise direction in particular is a very rough approxima-
tion. Such a truncation causes a drop of the spanwise and normal
root mean square values of the velocity which is particularly signifi-
cant in the upper half of the layer. For this reason, we do not expect
our velocity field reconstructed without rescaling to have more than
qualitative significance.
As a preliminary approach, we study the set of equations for a
truncation limited to the first eigenfunction (n = p = q = r = 1),
the zero streamwise wavenumber fci = 0 and up to six spanwise
wavenumbers k$ = 0, k, ...5k, for a suitably chosen k.
When only the zero streamwise wavenumber is considered, the
equations become much simpler. Indeed, because of the symme-
tries of the eigenfunctions (Herzog, 1986) in the (&i, £3) wavenumber
plane, the first and second components are purely real and the third
component is imaginary on the k<3 - axis (i.e. for k\ — 0). Using these
symmetries, the equations for the complex modal coefficients e41=0 ks
can be readily derived. Letting ^3 take the values jk,j = 0,1, ...,5
and writing a^. '_0 k __-k = Xj+i j/j, a typical equation takes the form:
Figure 3.5.2. Time histories of the real (x;) and imaginary (y,-) parts
of the coefficients for a value of the Hewisenberg parameter a = 1.4
lizes (see figure 3.5.6). The mean interburst times for various values
of the pressure signal are shown in figure 3.5.7. Probably the most
significant finding of this work is the identification of the etiology of
the bursting phenomenon. That is, the bursts appear to be produced
autonomously by the wall region, but to be triggered by pressure sig-
nals from the outer layer. Whether the bursting period scales with
inner or outer variables has been a controversy in the turbulence lit-
erature for a number of years. The matter has been obscured by the
fact that the experimental evidence has been measured in boundary
layers with fairly low Reynolds numbers lying in a narrow range, so
that it is not really possible to distinguish between the two types of
scaling. The turbulent polymer drag reduction literature is particu-
larly instructive, however, since the sizes of the large eddies, and the
bursting period, all change scale with the introduction of the polymer
(see, for example, Kubo and Lumley, 1980, Lumley and Kubo, 1984).
The present work indicates clearly that the wall region is capable of
producing bursts autonomously, but the timing is determined by trig-
ger signals from the outer layer. This suggests that events during a
burst should scale unambiguously with wall variables. Time between
bursts will have a more complex scaling, since it is dependent on the
first occurrence of a large enough pressure signal long enough after
a previous burst; "long enough" is determined by wall variables, but
the pressure signal should scale with outer variables.
where D/Dt denotes the mean convective derivative, Sij the mean
rate of strain, and v the kinematic viscosity. TJJ represents the rec-
tified effects of the small scale fluctuations on the coherent field and
is defined by
The two scalar functions are then expanded in normal modes in the
streamwise and spanwise directions.
Substituting the above into equation (3.6.7) and eliminating the pres-
sure TT results in two coupled equations, forming a differential eigen-
value problem.
In order to precede we need to specify a mean velocity field and
a model for the unknown stress terms (see figure 3.6.1.).
Figure 3.6.1. Model inputs; (a) mean velocity and mean gradient;
(b) eddy viscosity and Reynolds stress.
The nature of the averaging procedure implies that the scales of the
coherent field and the background turbulence are different. Assum-
ing that the background turbulence evolves on much shorter time
and length scales then the structures, it seems plausible that a New-
tonian stress-strain relationship like that for the molecular stresses
will provide the basis for a model. We set
analogy for the complete flow: the coherent structures are fed energy
directly by the mean gradients while the small scale turbulence is in
turn fed by gradients of the coherent field. If we identify the Oth
order stresses with an eddy viscosity tensor, then the closure model
can be written as
where the tensor viscosity has the following structure in this specific
case: z/13 = 1/33 = 0, 1/33 = 1/22-
Despite the absence of mean production terms, this model is still
a major improvement over the isotropic eddy viscosity formulation.
In the simple model the effects of the mean field have been neglected
entirely. Here we have allowed for modulation of the perturbation
stresses by the mean field through the 0th order stresses appearing
in the production terms. Also we have unconstrained the model
in an important way since the tensorial form of the eddy viscosity
allows the pricipal axes of the stress tensor to be unaligned with
the axes of the rate of strain. This is more realistic considering the
three-dimensionality of the coherent field. This model leads to the
expected cross coupling of the equations through the stress terms.
Figure 3.6.3. shows eigensolutions for several values of k3. The
results compare well with the POD eigenvalues, especially for wave-
numbers at or below the peak in the POD spectrum. The improve-
ment with decreasing wavenumber is expected given the modeling
considerations. The separation of scales between the background
turbulence and the coherent structures increases as the wave number
decreases adding to the expected accuracy of the stress model. The
comparison of the two models indicates significant improvements in
the results given by the anisotropic eddy viscosity form. The energy
method procedure with the more refined closure model appears ca-
pable of extracting structures which closely approximate those given
by the POD at least at the energy containing scales of motion.
Despite the general improvement, it is still clear that more needs
to be done. From Figures 3.6.2. and 3.6.3., it is evident that the eigen
spectrum produced by solution of Equation (3.6.1),while improved
by the use of the anisotropic closure model, still predicts structures
with maximum growth rate that are a factor of 2 smaller than those
containing the most energy (as given by the POD). We next con-
sidered the effect on the spatial mean velocity field of the growing
coherent perturbation (see figure 3.6.4.).
60 J. L. Lumley et al.
interaction between the mean and the coherent field. The mean flow
is imposed and the resulting structures are calculated. The mean
profiles we have used are time averages which mask any contribution
from the coherent field. As such the stability analysis predicts that
the highest growth modes are those which can best extract energy
from the time averaged mean shear which is concentrated in the small
near-wall region. Since the structures have an aspect ratio of about
1, the narrow region of high imposed shear leads to a peak in the
eigen spectrum at a large wave number.
To allow the mean field to evolve under the influence of the coher-
ent field, we follow Liu (1988) and write time evolution equations for
the energy density of the coherent field. We allow the mean profile to
depend on the coherent velocity as it does in reality. We expect that
equilibrium solutions for the energy density as a function of cross
stream wave number will approximate the average energy content as
given by the POD spectrum.
We assume that the coherent field is given by the eigenvalues of
62 J. L. Lumley et al.
where z/n = 1/22 = ^33 = ^r,^i2 = ^21 = Ar^is = "23 = 0. All that
remains is to model the mean profile. For this we adopt the quasi-
steady model used in Aubry et al. (1988). This allows the mean to
respond to growing structures providing the necessary feedback to
the evolving modes. Using the friction velocity, ur and the channel
half height, a, the scaled equation for the mean gradient is:
large scale structures, but must become important for the smaller
scales. We will consider such interactions below.
The volume averaged kinetic energy equations for the two compo-
nents of the coherent field are:
and the mean velocity model now contains contributions from both
the fundamental and the subharmonic:
New coupling terms appear in the evolution equations for the coher-
ent energy densities due to the interaction between the different size
modes:
/|2 and /I1 represent the effect on the mean shear production of
fundamental (subharmonic) coherent energy due to the presence of
the subharmonic (fundamental). /4 is a measure of the direct energy
transfer between the two modes due to the working of the sunhar-
monic stresses against the fundamental rate of strain. This quantity
appears with opposite sign in each equation.
As an example, we have calculated the interaction integrals using
the mode £3 = 20 as the fundamental. The results of integrating
equations (3.6.28) for these coefficient values are shown in figure
3.6.7. The mode-mode dynamics are dominated by terms due to the
mean velocity feedback model. The direct interaction term J4 is much
66 J. L. Lumley et al.
then, if the {&;} are the coefficients in a wavelet basis, we get for
some constant C, depending only on the process (and not on e):
for some -V(e) > N(e) slightly bigger than -/V(e) (the precise state-
ment is given in Berkooz et al. 1992).
u periodic on [0, L], where L, the length of the spatial domain, is the
only free parameter in the problem.
Although the dynamical behavior for small values of L is fairly
well understood (see Hyman et a/., 1986 for an overview), many
open questions remain concerning the limit L —» oo (Zaleski, 1989;
Pomeau et al, 1984). As can be seen from the numerical simulations,
for L —> 30, a chaotic regime involving both space and time disorder
occurs (see Figure 3.8.1, where we plot a space-time representation
of a typical solution, L = 400, 0 < t < 100).
In order to check the estimate (3.7.2), we compare the energy
resolved by a given number of modes using either a Fourier (POD)
Fundamental Aspects 69
ACKNOWLEDGMENTS
4 REFERENCES
Aubry, N., Holmes, P., Lumley, J.L., and Stone, E., 1988. "The
dynamics of coherent structures in the wall region of a turbulent
boundary layer," J. Fluid Mech. 192: pp. 1-30.
Fundamental Aspects 73
Glauser, M.N., Leib, S.J. and George W.K., 1985. "Coherent struc-
ture in the axisymmetric jet mixing layer," Proc. of the 5th
Symp. of the Turbulent Shear Flow Con/., Cornell University.
(Springer selected papers from TSF).
Herzog, S., 1986. "The large scale structure in the near-wall region
of turbulent pipe flow," Ph.D. thesis, Cornell University.
Kubo, I. and Lumley, J.L., 1980. "A study to assess the potential
for using long chain polymers dissolved in water to study tur-
bulence," Annual Report, NASA-Ames Grant No. NSG-2382,
Ithaca, NY: Cornell.
Meneveau, C., Lund, T.S. and Chasnov, J., 1992. "On the local
nature of the energy cascade," Proceedings of the Summer Pro-
gram, Center for Turbulence Research. Stanford/NASA Ames:
CTR.
76 J. L. Lumley et al.
Sarkar, S., Erlebacher, G., Hussaini, M.Y. & Kreiss, H.O., 1991.
"The analysis and modeling of dilatational terms in compress-
ible turbulence," J. Fluid Mech. 227: pp. 473-493.
Shih, T.-H., Lumley, J.L. and Janicka, J., 1987. "Second order
modeling of a variable density mixing layer," J. Fluid Mech.
180: pp. 93-116.
DIRECT NUMERICAL
SIMULATION OF
TURBULENT FLOWS
Anthony Leonard
1 INTRODUCTION
The numerical simulation of turbulent flows has a short history.
About 45 years ago von Neumann (1949) and Emmons (1949) pro-
posed an attack on the turbulence problem by numerical simulation.
But one could point to a beginning 20 years later when Deardorff
(1970) reported on a large-eddy simulation of turbulent channel flow
on a 24x20x14 mesh and a direct simulation of homogeneous, iso-
tropic turbulence was accomplished on a 323 mesh by Orszag and
Patterson (1972). Perhaps the arrival of the CDC 6600 triggered
these initial efforts. Since that time, a number of developments have
occurred along several fronts. Of course, faster computers with more
memory continue to become available and now, in 1994, 2563 sim-
ulations of homogeneous turbulence are relatively common with oc-
casional 5123 simulations being achieved on parallel supercomputers
(Chen et al., 1993) (Jimenez et a/., 1993). In addition, new algo-
rithms have been developed which extend or improve capabilities in
turbulence simulation. For example, spectral methods for the simu-
lation of arbitrary homogeneous flows and the efficient simulation of
wall-bounded flows have been available for some time for incompress-
ible flows and have recently been extended to compressible flows. In
79
80 A. Leonard
Number of
Kolmogorov length mesh points
(A = 37?)
Reynolds n+
3
Number 7?/D wall^units N = 10^(ff
5 X 103 0.0045 1.6 3.1 X 106
1 X 104 0.0028 1.8 1.3 X 107
5 X 104 0.00093 2.4 3.6 x 108
1 X 105 0.00058 2.8 1.5 x 109
5 x 105 0.00019 3.8 4.2 x 1010
For Re from 5,000 to 500,000, A + ranges from 4.8 to 11.4 (see Table
I). This spacing would be marginally sufficient resolution to repro-
duce all important wall-layer structures (such as streamwise streaks),
which have characteristic lengths of 50-100 wall units with some
structures down to 20 units in size.
The number of time-steps, N s , required to follow one realization
for a time T and obtain reasonable statistics also depends on Rey-
nolds number. The time-step At is roughly limited to
Using the above estimate for A and (2/f) 1 ' 4 = 3 we find that
Direct Numerical Simulation 85
3 SIMULATION OF HOMOGENEOUS
INCOMPRESSIBLE TURBULENCE
and I ranges over -N/2 + 1 < I < N/2 - 1. The Navier-Stokes equa-
tions become a 3(N — I) 3 system of ordinary differential equations
(ODEs),
/here k 2 = k n k n , and
86 A. Leonard
(In the above, zero mean flow, Umn = 0, has been assumed in order
to simplify the presentation.) To avoid explicit evaluation of the
convolution sums u^un requiring 0(N 6 ) operations per step, fast
Fourier transforms (FFTs) are used to return to physical space where
the required products are formed and then transformed (by FFTs)
back to Fourier space. Consequently, only 0(N 3 logN) operations
per step are required.
Suppose the tensor U^m is decomposed into a symmetric (R) and
an antisymmetric (ft)tensor; then the only constraint on R is that it
be traceless Rnn = 0, but 17, related to the vorticity, must satisfy the
evolution equation
Besides zero mean flow, four examples are: plain strain, TJ22 =
—Uss = const; axisymmetric strain, 1)22 — Uss = —(l/2)Un =
const; shear, Uia = const; and rotation, TJis = -Uai = const. See
Rogallo (1981) and Rogallo and Moin (1984) for more discussion and
application.
and Legendre polynomials are popular choices for channel flow but,
for example, in the case of pipe flow, other choices may be preferable
because of special conditions of the problem at hand. Whatever the
choice may be, this change in basis functions complicates the impo-
sition of, for example, the no-slip condition and the satisfaction of
the continuity constraint. These two conditions become global con-
straints on the expansion which are generally difficult or costly to im-
pose simultaneously. By contrast, in the simulation of homogeneous
flows using Fourier expansions in all three directions, the bound-
ary conditions are built into the expansion and the divergence-free
constraint is satisfied by a simple projection which is local in wave-
number space.
In the following, we will describe a technique for overcoming the
algorithmic difficulties described above, at least for flows in simple
geometries. The technique consists of expanding the velocity field
in terms of a set of divergence-free vector functions satisfying the
appropriate boundary conditions. First we write the Navier-Stokes
equations in rotational form:
and
If the £m form a complete set and N -> oo, this operation is equiva-
lent to applying the projection operator because
for an arbitrary scalar field (f>. The result is the following system of
ODEs:
where
If k = 0 we choose
R e = 10 4 ,k x = l,k z = 0
N Real (A)
15 .00372 . . . .
20 .0037398 . . . .
25 .003739669 . . .
30 .003739670616.
35 .0037396706227
40 .0037396706223
45 .0037396706216
50 .0037396706222
given by
such that
where t is any tangent vector at the body surface. That (41) and
(42) imply (39) may be shown as follows. Let -0, the stream function
for the imaginary fluid within the body, be given by
and
Direct Numerical Simulation 93
at every point on the surface <9B for any tangent vector t. Thus,
~y~ — 0, so that (46) reduces to
where a is the effective core radius of the particle, and £ has unit
volume integral. See Leonard (1985) and Winckelmans and Leonard
(1993) for further discussion including choices for the function £. For
example, the distribution for the gaussian particle is given by
Using (59) and the particle representation, we find that (58) can be
written
Note now that by approximating the integral on the right hand side
of (62) over the particles and using (60) we obtain the evolution
equation for the problem (58) in an infinite domain,
spatially varying flux of vorticity from the surface. This effect along
with the pressure gradient produces a wall slip (Usijp) at time t + At.
We compute this slip as an average over computational panels on
the wall directly from the Green's function integral for the velocity
potential, with a form analogous to (41), using the fast algorithm.
Thus the total flux to be emitted into the flow for the other substep
of the diffusion process is given by
If the old particle is less then a distance h from a wall the interpo-
lation is modified to maintain the same conservation properties. See
Direct Numerical Simulation 97
is the viscous stress tensor (with the bulk viscosity assumed zero),
L is the identity tensor,
102 A. Leonard
which arises when one takes the curl of the momentum equation to
obtain the vorticity transport equation.
The initial conditions for, say, isotropic, homogeneous turbulence
are, of course, more varied for compressible flows. The initial ve-
locity field may be decomposed (Helmholtz decomposition) into a
solenoidal or "incompressible" component, u1 and a dilatational or
"compressible" component, u c , i.e.,
where
where Ui and U2 are the freestream speeds on either side of the layer,
and GI and C2 are the corresponding speeds of sound. It is known from
experiment that compressible layers have growth rates significantly
104 A. Leonard
7 REFERENCES
Barnes, J. E., and Hut, P., 1986. "A hierarchical 0(N log N) force
calculations algorithm," Nature, 324, pp. 446-449.
Beale, J. T., and Majda, A., 1982. "Vortex methods II: High order
accuracy in two and three dimensions," Math. Comp. 39, pp.
29-52.
Degond, P., and Mas-Gallic, S., 1989. "The weighted particle method
for convection-diffusion equations, Part I: the case of isotropic
viscosity, Part II: the anisotropic case," Math. Comp. 53, pp.
485-526.
Lee, S., Lele, S. K., and Moin, P., 1991. "Eddy shocklets in decaying
compressible turbulence," Phys. Fluids A, 3, pp. 657-664.
Leonard, A., and Wray, A., 1982. " A new numerical method for the
simulation of three- dimensional flow in a pipe," Proceedings
of the 8th International Conference on Numerical Methods in
Fluids Dynamics, June 28-July 2, 1982, Aachen, W. Germany,
Springer-Verlag, New York, pp. 335-342.
Moser, R. D., and Rogers, M. M., 1990. "Mixed transition and the
cascade to small scales in a plane mixing layer," Phys. Fluids
A, 5, pp. 1128-1134.
Direct Numerical Simulation 107
Moser, R. D., Moin, P., and Leonard, A., 1983. "A spectral numer-
ical method for the Navier-Stokes equations with applications
to Taylor-Couette flow," J. Comput. Phys., 52, p. 524.
Patera, A. T., 1984. "A spectral element method for fluid dynamics;
laminar flow in a channel expansion," J. Comput. Phys., 54,
pp. 468-488.
Sarkar, S., Erlebacher, G., Hussaini, M. Y., and Kreiss, H. 0., 1989.
"The analysis and modelling of diltational terms in compress-
ible turbulence," J. Fluid Mech., 227, pp. 473-493.
She, Z.-S., Chen, S.-Y., Doolen, G. D., Kraichnan, R. H., and
Orszag, S. A., 1993. "Reynolds-number dependence of iso-
tropic Navier-Stokes turbulence," Phys. Rev. Letters, 70, pp.
3251-3254.
Sherwin, S. J., and Karniadakis, G. E., 1994. "A triangular spec-
tral element method: algorithms and flow simulations," Pro-
ceedings of the 14th International Conference on Numerical
Methods in Fluids Dynamics, Bangalore, 11-15 July.
Spalart, P. R., 1987. "Direct simulation of a turbulent boundary
layer up to R<? = 1410," J. Fluid Mech., 187, pp. 61-98.
Stanaway, S. K., CantweU, B. J., and Spalart, P. R., 1988. AIAA
Paper 88-0318.
von Neumann, J., 1963. "Recent theories of turbulence," 1949 re-
port to the Office of Naval Research, reprinted in John von
Neumann, Collected Works, 6, A. H. Taub, ed., Macmillan
Co., New York, pp. 437-472.
Winckelmans, G. S., Salmon, J. K., Warren, M. S., and Leonard,
A., 1995. "The fast solution of three-dimensional fluid dynam-
ical N-body problems using parallel tree codes: vortex element
method and boundary element method," Seventh SIAM Conf.
on Parallel Processing for Scientific Comp., Feb. 1995, San
Francisco.
Winckelmans, G. S., and Leonard, A., 1993. "Contributions to vor-
tex particle methods for the computation of three-dimensional
incompressible unsteady flows," J. Comput. Phys., 109, pp.
247-273.
Chapter 3
Joel H. Ferziger
1 INTRODUCTION
109
110 J. H. Ferziger
where [/,- =< ul >, the brackets denoting one of the averages listed
above, and T;J is the Reynolds stress tensor,
may be used. Eq. (2.5) allows one to use an equation for the dissi-
pation as a means of obtaining both e and L. No constant is used in
Eq (2.5) because this constant combines with others in the complete
model.
An exact equation for the dissipation can be derived from the
Navier-Stokes equations and has a form similar to the energy equa-
tion. The modeling applied to the dissipation equation is so severe
that it is probably best to regard the entire equation as a model in
itself. Difficulties associated with the dissipation equation (or any
114 J. If. Ferziger
other equation used to determine the length scale) are the most dif-
ficult ones in two-equation modeling.
Some of the significant deficiencies of models based on Eq. (2.3)
are direct consequences of the eddy viscosity relationship itself. In
three dimensional flows, the eddy viscosity may no longer be a scalar;
measurements and numerical simulations show that it becomes highly
anisotropic, i.e. it is actually a tensor quantity.
The effect of an eddy viscosity can be interpreted in another
way. Enough viscosity is added to the equations to assure that the
computed flow is stable i.e., the solution of the RANS equations is
an effective laminar flow with a velocity field that is the mean (in the
Reynolds sense) of the turbulent velocity field. In two dimensions, it
is always possible to define a spatially dependent eddy viscosity that
produces the correct mean flow. In general, it is not possible to find
this eddy viscosity without knowledge of the solution but it is useful
to know that it exists in principle. In three dimensional flows, the
eddy viscosity may be a tensor of either second or fourth rank.
Anisotropic or tensor models have been proposed. If a model
is to be applicable to a wide range of flows, it should possess in-
variance properties, i.e. it must give the same results independent
of the coordinate system is used in the calculation. Many early
anisotropic models were not properly invariant. Recently, invariant
tensor and/or non-linear models have been proposed, for example,
see Speziale (1987) and Horiuti (1990). These models take the form:
where Sij is
Although these models contain more constants than the scalar eddy
viscosity model, some of them are fixed by requiring invariance. A
detailed discussion of these types of closures can be found in Chapters
5 and 6.
The most complex models in common use today are Reynolds
stress models which are based on dynamic equations for the Rey-
nolds stress tensor itself. As these are complicated, and because
their application as subgrid scale models in large eddy simulation
Large Eddy Simulation 115
has been limited, we shall not describe or discuss them here. A de-
tailed discussion of these types of closures can be found in Chapter
5 and 6.
With few exceptions, RANS models cannot be applied to flow
near a surface without modification. Special near-wall versions of the
models, especially the k — c model, have been developed and work
quite well for attached boundary layers. One of the most recent of
these models was proposed by Rodi and Mansour (1992), who include
references to other models.
3 FILTERING
The exception is the cutoff filter for which equality does hold. The
difference between the two sides of this inequality will be exploited
for modeling purposes later.
For now, we note that the difference represented by Eq (4.1)
means that the both the physics and modeling of the subgrid scale
Reynolds stresses (SGSRS) may be more complicated than for the
RANS Reynolds stresses. By using the kind of decomposition of
the velocity field used in RANS modeling i.e. writing the complete
velocity field as a combination of the filtered field and a subgrid scale
field, we can decompose the SGSRS into three sets of terms:
By far the most commonly used subgrid scale model is the one
proposed by Smagorinsky (1963). It is an eddy viscosity model that
can be thought of as an adaptation of the Boussinesq concept of Eq
(2.3) to the subgrid scale. It is:
Equating (4.5) and (4.6) shows that the eddy viscosity must take
the form:
Large Eddy Simulation 121
where y+ is the distance from the wall in viscous wall units (y+ =
yur/v, where ur is the shear velocity (r//?) 1 ' 2 and T is the shear stress
at the wall) and A+ is a constant usually taken to be approximately
25. Although this modification produces the desired results, it is
difficult to justify in the context of LES. The SGS model should
depend solely on the local properties of the flow and it is difficult
to see how the distance from the wall qualifies as an appropriate
parameter in this regard.
The purpose of the van Driest damping is to reduce the subgrid
scale eddy viscosity near the wall; it is generally believed that VT ~ y3
in this region and models ought to respect this property. It follows
that an alternative to van Driest damping is a subgrid scale model
which reduces the magnitude of the viscosity in the proper manner
when a subgrid scale Reynolds number (the obvious one is |5|A2/!/)
becomes small. Models of this kind were suggested by McMillan and
Ferziger (1980) and by Yakhot and Orszag (1986): the latter used
renormalization group theory to derive their model. It is necessary to
point out that these issues focus on the fully turbulent flow. Applica-
tion to transitional flows present further problems and modifications
which have been addressed initially by Piomelli et al.
A further problem is that, near a wall, the structure of the flow
is very anisotropic. Regions of low speed fluid (streaks) are created.
They have dimensions of approximately 1000 viscous units in the
strearriwise direction and perhaps 100 viscous units in the spanwise
Large Eddy Simulation 123
scale, A, associated with the filter. These can be extracted from the
velocity field in the following manner.
From the complete velocity field, v,j, we can compute the resolved
or large scale field ¥,• by filtering and the small or subgrid scale
field u\ — U{ — U{ by subtraction. From these we can construct a
further subdivision. The very largest resolved scales may be defined
by filtering a second time to obtain ¥; so the smallest resolved scales
are defined by Uj — u,-. The largest unresolved scales are defined by
u\, A simple calculation shows that these are identical. This leads
to the following possibility as a subgrid scale model:
When the large scales are defined by the cutoff filter, ¥; = ¥,-,
and the scale similarity model produces nothing i.e., Eq. (4.15)
evaluates to zero. This difficulty can be removed by noting that, for
the Gaussian filter, filtering twice is equivalent to a single filtering
with A replaced by \/2A. This is easily mimicked for the cutoff filter
by defining the double overline filter as a cutoff filter corresponding
to this width; however, the correlation is not as good as it is when
the Gaussian filter is used, cf. Bardina et al. (1980).
The second or test filter (the one used to determine the parame-
ter) is similar to the second filter used in the scale similarity model
but is denoted by a tilde (~) to make explicit the idea that the orig-
inal and test filters need not be identical. The subgrid scale stress
that must be modeled in the test-filter level LES is:
Now let us define the large scale component of the SGS Reynolds
stress at the test filter level. This is the portion that is directly
computable from the LES field by filtering:
and is essentially the Leonard stress associated with the test filter.
Now it follows directly from these definitions that:
where the brackets (<>) represent an average over the spatial re-
gion to which the least square method was applied. This technique
produces excellent results; it has been used to compute a variety of
homogeneous turbulent flows, fully developed channel flow, and tran-
sitional channel flow, all with excellent results. This version of the
dynamic model removes many of the difficulties described earlier:
• It was noticed that, in shear flows, the required value of the
Smagorinsky model parameter is much smaller than in isotropic tur-
bulence. The dynamic model produces this modification automati-
cally.
• Near walls, the value of the model parameter has to be reduced
even further, for example by using van Driest damping (4.13). In
channel flow, the averaging in Eq (4.24) is usually averaging over
planes parallel to the wall. When this is done, the model automati-
cally decreases the parameter near the wall.
Large Eddy Simulation 131
The Smagorinsky and scale similarity models are not the only
ones that have been used to represent subgrid scale turbulence. For
guidance as to how improved models might be constructed, one can
turn to turbulence theories. In order to deal with the distribution
of turbulent energy over a range of length scales, in most turbulence
theories the principal variables are the Fourier transforms of the ve-
locity components:
and
The models described above have been designed for flows without
'extra strains' ( e . g . , rotation, compressibility and curvature); despite
this, we have seen that the dynamic model can handle some of these
136 J. H. Ferziger
without problems.
Meteorologists and oceanographers who predict global circulation
deal with flows that are nearly two-dimensional; an eddy viscosity
is used to represent the unresolved motions. At the smallest scales,
three-dimensional equations may be used; simulations are routinely
done on several levels. A single model (with a single parameter)
that can account for phenomena at all the various scales probably
does not exist. A systematic approach is needed to build a firm
foundation for modeling in these areas. The task is difficult and
progress may come slowly. A possible approach is the following. At
the lowest level, one can simulate the small-scales e.g., the planetary
boundary layer or the ocean mixed layer and use the data produced
to construct parameterizations that to be used represent motions
that belong to the subgrid scale on the next larger scale, perhaps
the regional scale. To assure that all possibilities are included, a
range of cases containing all physically possible phenomena must be
simulated to ensure that the full range of parameters are included
in the database. By bootstrapping in this way, and allowing two-
way interaction between simulations at different scales, it may be
possible to develop methods that allow phenomena on all scales to
be predicted. It should be obvious that there are difficulties in this
scenario for which solutions are yet not available.
Extra strains can be roughly divided into two classes. Some, such
as rotation, curvature, and stratification, affect the large scales more
strongly than the small scales. In these cases, SGS models designed
for incompressible flows without extra strains can probably be used
without major modification. For example, large eddy simulations of
a stable planetary boundary layer performed with the Smagorinsky
model (Mason and Derbyshire, 1990) agree very well with both direct
simulations (Colernan et a/,, 1989) and field data.
On the other hand, for 'strains' whose action is principally in the
small scales, the situation is less clear. Compressible turbulence at
low Mach numbers can be treated with incompressible models. At
higher Mach numbers, small shock waves ('eddy shocklets') develop,
and the flow behavior can be quite different (Blaisdell et a/., 1991,
Lee et a/., 1993). If one is to do large eddy simulations of such flows,
there are two possibilities. The first approach is similar to standard
SGS modeling. In the absence of shocks, SGS models applicable to
the Favre-filtered equations can be developed in a manner analogous
to the incompressible case (Erlebacher et a/., 1992; Speziale et a/.,
Large Eddy Simulation 137
5 WALL MODELS
where y\ is the height of the first grid point, < TW > is the mean wall
shear stress, and Ui(yi) is the mean velocity at y\.
Mason and Callen (1986) assumed that the logarithmic profile
for the mean velocity in the buffer region, holds locally and instanta-
neously. This assumption is incorrect but their boundary condition
is often used by meteorologists. Piomelli et al. (1987) found it to be
inadequate for engineering applications. This is an example of how
the differing needs of two disciplines can lead to opposite conclusions
about the effectiveness of a model.
140 J. R. Ferziger
6 NUMERICAL METHODS
so
for this method. For small &, the Taylor series approximation:
for routine engineering use for a long time to come. Further, LES
can and will be used as a complement and partial substitute for
experimental testing. An occasional LES can be compared to RANS
predictions to test the adequacy of a design and/or the methods used
to develop it. It will also continue to be used to directly test the
accuracy of RANS models, a role that it has played with distinction
throughout its history. By using LES to tune RANS models, it should
be possible to obtain most of the benefits of LES at a small fraction
of the cost.
Flows that may be good candidates for LES in the near future
include the turbine blade passage and the internal combustion engine
cylinder. These are both relatively low Reynolds number flows and
of obvious technological importance. Both of these flows also contain
many extra strains that renders the development of RANS models
for them exceedingly difficult, making the possibility of using LES
directly in the design process and interesting one.
A word of caution is necessary. LES and its subgrid scale models
have been validated only for relatively simple flows at fairly low Rey-
nolds numbers. In these flows, most of the energy is in the resolved
scales and, even if the subgrid scale model is not very accurate, its ef-
fect on the results may not be too important. If one uses the success
of these simulations as a justification for applying LES to much more
complex flows, although reasonable looking results may be obtained,
placing one's trust in them may be risky. The leap is simply too great
to allow expectation of success in this kind of endeavor. Simulations
of this kind have been made but, in the author's opinion, they have
been premature and their value is questionable.
It is also important that the goal of a simulation be defined in
advance. Doing a simulation merely to show that it can be done is
of limited value. It is known in advance that it can be done and, if
enough resources are deployed, good results will be obtained. The
value is in learning about the physical nature of the flow, how it may
be modeled, or, perhaps, in making a contribution to the improve-
ment of a design.
It would be very valuable to have models that eliminate the need
to specify no-slip conditions at a wall. Boundary conditions of this
kind exist for attached flows and were discussed above. What is
not known is whether conditions of this kind can be constructed for
separated flows. Doing so for RANS models has proved exceedingly
difficult and there is no reason not to expect the task to be at least
146 J. H. Ferziger
A better choice for the near term is to perform LES and/or DNS
on 'building block' flows, i.e. flows that are structurally similar to
the ones of actual interest. From the results of such simulations,
RANS models that can be applied to the more complex flows can
be validated and improved. RANS computations can then be used
as the everyday tool. LES need be performed only when there are
significant changes in the design or as an occasional check on the
validity of the RANS results.
As noted earlier, there have been attempts at large-eddy and
direct numerical simulation of complex flows. Unfortunately, in most
of these, the subgrid scale model was uncontrolled and the results are
of uncertain value. This appears to be a case of reaching too far too
fast; we shall not present examples here.
Since answers to questions involving technologically significant
flows are required, the following questions arise. Is there a method
that will enable more complex flows to be simulated on available
machines? Are there flows of importance that are good candidates
for simulation via LES in the relatively near future?
The answer to these questions appear to be a qualified yes. Other
than the flows mentioned earlier, particularly good candidates are
flows in which there are a small number of important, energetic, and
easily identified coherent structures. In all the cases that have been
suggested, the large structures are vortices. Let us consider two such
cases.
Flows over bluff bodies usually produce strong vortices in their
wakes. The vortices produce strong fluctuating forces on the body in
both the streainwise and spanwise directions whose prediction is very
important in many applications. The latter include buildings (wind
engineering) and ocean platforms, among others. If the vortices are
sufficiently larger than the bulk of the motions that constitute the
'turbulence' it should be possible to construct a filter that allows the
vortices to be retained in the resolved field while removing all of the
smaller scale motions. We have called a method that accomplishes
this 'coherent structure capturing' or CHC (Ferziger, 1993). The au-
thor and others earlier called it very large eddy simulation or VLES,
a term we now find less descriptive. Methods of this kind were sug-
gested a long time ago (Ferziger, 1983) but deliberate simulations
of this type do not appear to have been attempted other than a few
cases which apparently gave unsteady results when a steady flow was
expected. An exception to this might be the recent work of Orszag
148 J. H. Ferziger
discussed in Chapter 4.
The cylinders of internal combustion engines provide another ex-
ample. This flow is inherently unsteady, so there is no possibility of
modeling it as a steady flow. Several interesting issues arise which
lead to the following questions. What does RANS mean in such a
flow and how should RANS results be compared with experimental
results? Since the flow is unsteady, LES can only produce a single re-
alization; can such a simulation provide sufficient information about
the flow? A partial answer to the first question is that the RANS
mean velocity should probably be defined as an average over many
cycles and the turbulence as the deviation from the multi-cycle aver-
age. LES should simulate a single cycle. To see what the differences
are consider the following. After the intake stroke, the flow contains
a strong vortex whose location, size, and strength varies from cycle
to cycle; this vortex is important to engine behavior. In a RANS
calculation, the result should contain an average vortex, one that is
relatively large and of average circulation. In LES, the vortex should
be smaller, of similar circulation, but its location should vary from
realization to realization, it should be possible to construct a filter
that can separate the vortex from the rest of the turbulence field.
ACKNOWLEDGMENTS
The author has been active in this field for over twenty years and
the list of people who have helped him is too long to be covered in a
short acknowledgment. I will, therefore, limit myself to mentioning a
few people who have influenced my thinking in this area in the past
few years. These include my colleagues: Profs. Peter Bradshaw, Jef-
frey Koseff, Parviz Moin, Stephen Monismith and William Reynolds,
and my students Matthew Bohnert and Kishan Shah. The support
received from a number of agencies over the year has also been very
150 J. H. Ferziger
10 REFERENCES
Akselvoll, K. and Moin, P., 1993."Large eddy simulation of a back-
ward facing step flow," ASME Fluids Engr. Conf., Washington,
DC, June.
Bagwell, T.G., Adrian, R.J., Moser, R.D. and Kim, J., 1993. "
Improved approximation of wall shear stress boundary condi-
tion for large eddy simulation," in Near Wall Turbulent Flows
(R.M.C. So, C.G. Speziale and B.E. Launder eds.), Elsevier.
Berger, M.J. and Oliger, J., 1984. "Adaptive mesh refinement for
hyperbolic partial differential equations," J. Comp. Phys,, 53,
p. 484.
Bohnert, M.J. and Ferziger, J.H., 1993. "The dynamic subgrid scale
model in large eddy simulation of the turbulent Ekman layer,"
in Engineering Turbulence Modeling and Experiments 2, W.
Rodi and F. Martelli eds., Elsevier.
Large Eddy Simulation 151
Cain, A.B., Reynolds, W.C., and Ferziger, J.H., 1981. "A three
dimensional simulation of transition and early turbulence in a
time developing mixing layer," Report TF-14, Dept. of Mech.
Engr., Stanford Univ.
Caruso, S.C., Ferziger, J.H., and Oliger, J., 1985. "An adaptive grid
method for incompressible flows," Report TF-23, Dept. Mech.
Engr., Stanford Univ.
Chapman, D.R. and Kuhn, G.D., 1986. "The limiting behavior of
turbulence near a wall," J. Fluid Mech., 70, pp. 265-92.
Clark, R.A., Ferziger, J.H., and Reynolds, W.C., 1979. "Evalua-
tion of subgrid scale turbulence models using a fully simulated
turbulent flow," J. Fluid Mech., 91, p. 92.
Coleman, G.N., Ferziger, J.H. and Spalart, P.R., 1990. "A nu-
merical study of the stratified turbulent Ekman layer," Rept.
TF-48, Thermosciences Div., Dept. of Mech. Engr., Stanford
Univ.
Deardorff, J.W., 1970. "A numerical study of three-dimensional
turbulent channel flow at large Reynolds number," J. Fluid
Mech., 41, p. 452.
Deardorff, J.W., 1974. "Three dimensional numerical modeling of
the planetary boundary layer," Boundary Layer Meteorology,
1, p. 191.
Erlebacher, G., Hussaini, M.Y., Speziale, C.G., and Zang, T.A.,
1992. "Toward the large-eddy simulation of compressible tur-
bulent flows," J. Fluid Mech., 238, p. 155.
Ferziger, J.H., 1983. "Higher level simulations of turbulent flow," in
Computational Methods for Turbulent, Transonic, and Viscous
Flows, J.-A. Essers, ed., Hemisphere.
Ferziger, J.H., 1993. "Simulation of complex turbulent flows: recent
advances and prospects in wind engineering," in Computational
Wind Engineering 1, S. Murakami ed., Elsevier.
Ferziger, J.H. and Peric, M., 1994. "Computational methods for
incompressible flow," in Computational Fluid Dynamics, (M.
Lesieur and J. Zinn-Justin eds.), Elsevier.
152 J. E. Ferziger
Gao, F. and O'Brien, 1993. "A large eddy simulation scheme for
turbulent reacting flows," Phys. Fluids A 5, p. 1282.
Germane, M., Piomelli, U., Moin, P., and Cabot, W.H., 1990. "A
dynamic subgrid scale eddy viscosity model," Proc. Summer
Workshop, Center for Turbulence Research, Stanford CA.
Ghosal, S., Lund, T.S., Moin, P., and Akselvoll, K., 1994. "A dy-
namic localization model for large eddy simulation of turbulent
flows," submitted to /. Fluid Mech..
Horiuti, K., 1990. "Higher order terms in anisotropic representation
of the Reynolds stress," Phys. Fluids A 2, p. 1708.
Hussaini, M. Y., Speziale, C. G., and Zang, T. A., 1990. "The
potential and limitations of direct and large eddy simulations,"
Whither Turbulence? Turbulence at the Crossroads, Springer-
Verlag, Berlin, pp. 354-368.
Kawamura, T., and Kuwahara, K., 1985. "Direct simulation of
a turbulent inner flow by a finite difference method," AIAA
paper 85-0376.
Le, H., 1993. "Direct numerical solution of turbulent flow over a
backward facing step," Dissertation, Dept. of Mech. Engr.,
Stanford Univ.
Lee, S., Lele, S.K., and Moin, P., 1993. "Simulation of spatially
evolving turbulence and the applicability of Taylor's hypothesis
to compressible flow," Phys. Fluids A 5, pp. 1521-1530.
Leonard, A., 1974. "Energy cascade in large eddy simulations of
turbulent fluid flows," Adv. in Geophys., ISA, p. 237.
Lesieur, M., 1991. "Turbulence in fluids," Second ed., Kluwer, Dor-
drecht.
Leslie, D.C., 1973. "Theories of turbulence," Oxford U. Press.
Lilly, D.K., 1965. "On the computational stability of numerical so-
lutions of time-dependent, nonlinear, geophysical fluid dynamic
problems," Mon. Wea. Rev., 93, p. 11.
Lilly, D.K., 1992. "A proposed modification of the Germano subgrid
scale closure method," Phys. Fluids A 4, p. 633.
Large Eddy Simulation 153
INTRODUCTION TO
RENORMALIZATION
GROUP MODELING OF
TURBULENCE
Steven A. Orszag
I. Staroselsky, W. S. Flannery, Y. Zhang
1 INTRODUCTION
155
156 S. A. Orszag et al.
where
By applying these rules for averaging over the gaussian random noise
/ to the velocity field and the nonlinear Green's function G, we ob-
tain the full perturbation series for U and. G. The terms in these
Renormaliz&tion Group 161
dynamic variables w > (^) in the band A — 6A < |k| < A by using the
Neumann series (9) to obtain new expressions for the remaining vari-
ables u < (A;), 0 < |k| < A — £A, expressing the dynamic variables for
A — <*>A < |k| < A in terms of the random force in this narrow band,
and then averaging over the corresponding subensemble of random
forces. At each stage of this process, the use of the Neumann series
(9) is justified because <5A is small. This perturbation technique in <5A
generates so-called recursion relations for the terms in the reduced
dynamical system satisfied by n < (A;). The next step of the RNG pro-
cedure is to solve these relations to obtain new dynamic equations
in the case where many narrow bands of modes are eliminated. Of
course, the errors in the reduced dynamical system that result from
solving these relations need no longer be small because we are accu-
mulating many small contributions. However, under the assumption
that our Reff is small we proceed to use the dynamical system that
results from the recursion relations. The justification for this last
step has not been given for turbulence, so it is only possible to judge
the validity of the RNG equations by the accuracy of the results to
which they lead. The advantage of the present explanation of the
RNG technique is that the e—expansion (Re/f <C 1) is required only
to eliminate high-order nonlinearities in the dynamical equations for
tt < (A;) and not to evaluate constants appearing in the lowest order
dynamical equations for u < (A;).
When the velocity field is decomposed as
Upon taking the functional derivative < <5/<5/ < (6) > in (18) and
gaussian averaging over the subensemble in which the Fourier modes
of / are held fixed for |k| < A — <5A, one obtains the correction to the
Green's function which is second-order in the coupling constant:
where the symbol / indicates integration over the band A — <5A <
q < A, and we have included the factor Pa;g(k) on the left side of
(20) due to the incompressibility of u(k). Notice that this procedure
of unconditional shell-averaging is asymmetric with respect to the
arguments of the Green's function 6*0 and the correlation function
UQ] the wavenumber of UQ is restricted to the shell A — 6A < k < A
while there is no similar constraint upon the wavenumber of GQ.
To compute the integral in (20) we proceed as follows. First,
using (6) and performing the integral over the internal frequency %,
noting that
we obtain
Renormalization Group 165
where
In the limit k/q —> 0, this latter integral has the asymtotic form
we find
where r = log (fcj/A). Using analogous methods one finds that there
is no "bare force renormalization", i.e. no contribution to the forcing
D0k~y. However, a "thermal" noise proportional to k2 is generated,
namely,
When r —>• oo, the RNG transformation given by (25) has a simple
fixed point in which the coupling constant is indeed small:
The coefficients of (29) and (30) agree well with the experimentally
observed parameters.
Renormalization Group 169
4 TRANSPORT MODELING
The methods described above have also been used to derive trans-
port models in which all the scales up to the integral scale of tur-
bulent flow are averaged over. The RNG transport model describes
the dynamics of the mean flow /7;, turbulence kinetic energy li', and
energy dissipation rate £ and defines the eddy transport coefficients
in terms of the latter. For example, the eddy viscosity in the RNG
K — £ model is given by
where for the heat transfer problem a0 refers to the molecular inverse
Prandtl number.
There are two interesting properties of the RNG transport model
described above which lead to reduced eddy viscosity. First, the re-
duced value of (7f2 compared with the standard S -equation coefficient
(^standard ^ ^<^ j^g ^g interesting consequence of decreasing both
the rate of production of K and the rate of dissipation of S, leading
to smaller eddy viscosities. In regions of small 77, the .R-term tends
to increase eddy viscosity somewhat, but it is still typically smaller
than its value in the standard model. In regions of very large 77,
where strong anisotropy exists, R can become negative and reduces
the eddy viscosity even more. This feature of the RNG model is
perhaps responsible of the marked improvement in anisotropic large-
scale eddies. The /^-dependency of C'£2 makes it possible to have a
spatially as well as a temporally varying balance between production
and dissipation terms in the ^-equation. This feature of dynamic
balance, based on the RNG theory, is unique. Secondly, in regions
of low turbulence Reynolds numbers the RNG eddy viscosity has a
cut-off below which the eddy viscosity is zero.
The RNG K — £ model has been successfully applied to a number
of difficult turbulent flow problems. Below, we present some results
Renormalization Group 171
Some recent results from the computation of the flow over a com-
pressor trailing edge indicate that steady-state computations prove
inadequate in flows with large-scale unsteady structure. It will be
seen that when these calculations have a steady-state solution it does
not correspond to the time-averaged flow field. On the other hand,
the RNG transport model, which has a higher effective Reynolds
Renormalization Group 175
number than the standard K—£ model, may be used to perform time-
dependent, so-called Very Large Eddy Simulations (VLES). VLES
refers to the capability of the model to correctly predict the behav-
ior of anisotropic eddies with a size of the order of the characteristic
size of the problem, in this case the thickness of the trailing edge.
The Reynolds number for the flow under consideration is 56,400
based on the diameter of the half-cylinder trailing edge. The inlet,
taken at 10.6 diameters upstream, is a symmetrical zero pressure
gradient turbulent boundary layer set to match experimental con-
ditions. A number of runs were computed including steady-state
and time-dependent cases, and a steady-state case with a splitter
176 S. A. Orszag et al.
respectively.
Finally, a time-aver aged calculation using the standard K — £
model was performed for comparison. A comparison of the time-
averaged results for Cp with experiment is shown in Fig. 8. As with
the steady-state RNG computations, the base pressure in the wake is
significantly overpredicted by the standard model. Excessive turbu-
lent production leads to large eddy viscosities and hence mean wake
vortices which are too weak. It appears that the VLES RNG model,
by producing less turbulence as a result of lower eddy viscosities, can
predict the time-dependent behavior of the (very) largest eddies.
Flapping-Hydrofoil Simulations
The flow over a two dimensional hydrofoil subject to vertical gusts
at high reduced frequency has been simulated, and a comparison
with the experiments performed in the MIT Variable Pressure Water
Tunnel was done. A modified NAG A16 hydrofoil with chord length
178 S. A. Orszag et al.
gions between upper/bottom tunnel wall and the flapper wake where
experimental data is not available, uniform velocity profiles were as-
signed.
Two sets of runs were performed with the same grid and bound-
ary conditions, but using different turbulence models. One used the
standard K — C model and the other used RNG K — £ model. The
skin friction coefficients are plotted in Fig 11. On the suction side
with x/C > 0.4 where experiment data were available, the computa-
tion with the RNG K — £ model produced excellent agreement while
the computation results using the standard K — E model severely
overpredicted the skin friction. The results from the RNG K — £
model predicted the separation point (shown by the location where
skin friction first turns negative) at x/C = 0.968, while the exper-
iment data showed the separation point between x/C — 0.972 and
x/C — 0.990. The standard K — £ model, on the other hand, in-
180 5. A. Orszag et al.
ACKNOWLEDGMENTS
This work has been supported by the ONR, NASA, and United
Technologies Research Center. Discussions with Victor Yakhot have
been a constant source of stimulation. We would also like to ac-
knowledge discussions with Torn Barber and Dochul Choi of UTRC.
Finally, we should like to acknowledge our collaborations with Fluent,
Inc., especially Ferit Boysan, Nelson Carter, Dipankar Choudbury,
182 S. A. Orszag et al.
5 REFERENCES
Chen, S., Doolen, G., Herring, J. R., Kraichnan, R. H., Orszag,
S. A. and She, Z.-S., 1993. "Far dissipation range of turbu-
lence," Phys. Rev. Letts. 70, p. 3051.
Dannevik, W., Yakhot, V., and Orszag, S. A., 1987. "Analytical
theories of turbulence and the e—expansion," Phys. Fluids A
30, p. 2021.
Graziani, "R. A., Blair, M. F., Taylor, J. R., and Mayle, R. E.,
1979. "An Experimental Study of Endwall and Airfoil Surface
Heat Transfer in a Large Scale Blade Cascade," ASME Paper
79-GT-99.
Lurie, E. H., 1993. "Unsteady Response of a Two-Dimensional Hy-
drofoil Subject to a High Reduced Frequency Gust Loading,"
M.S. Thesis, Massachusetts Institute of Technology.
Kraichnan, R.H., 1961. "Dynamics of nonlinear stochastic sys-
tems," J. Math. Phys. 2, pp. 124-148. Erratum: 3, p. 205,
1962.
Orszag, S.A., 1977. "Statistical theory of turbulence," In: Fluid
Dynamics 1973, Les Houches Summer School in Physics. Eds.
R. Balian and J.-L. Peabe, Gordon and Breach, pp. 237—374.
Orszag, S.A., Staroselsky, I., and Yakhot, V., 1993a. "Some basic
challenges for large eddy simulation research," In: Large Eddy
Simulation of Complex Engineering and Geophysical Flows. (B.
Galperin and S.A. Orszag, eds). Cambridge University Press.
Orszag, S. A., Yakhot, V., Flannery, W. S., Boysan, F., Choudbury,
D., Maruzewski, J., and Patel, B., 1993b. "Renormalization
group modeling and turbulence simulations," In: Near-Wall
Turbulent Flows. Eds. So, R. M. C., Speziale, C. G. and
Launder, B. E., Elsevier Science Publishers, pp. 1031-1046.
Paterson, R.W., 1984. "Experimental Investigation of a Simulated
Compressor Airfoil Trailing Edge Flowneld," AIAA Paper 84-
0101.
Renormalization Group 183
MODELING OF
TURBULENT TRANSPORT
EQUATIONS
Charles G. Speziale
1 INTRODUCTION
185
186 C. G. Spezide
Continuity Eq.
where
The velocity and pressure are decomposed into mean and fluctu-
ating parts, respectively, as follows (cf. Hinze 1975):
General Turbulence
then
Turbulent Transport Equations 189
Continuity Eq
where
The oldest Reynolds stress closures are based on the Boussinesq eddy
viscosity hypothesis:
where
where
where u> = Vxv is the mean vorticity vector. The former model has
been primarily used as a subgrid scale model for large-eddy simula-
tions whereas the latter model has been used for Reynolds-averaged
Navier-Stokes computations in aerodynamics (see Wilcox 1993).
Turbulent Transport Equations 191
One-Equation Models
The eddy viscosity in these models is typically given by
where
where
Turbulent Transport Equations 193
or
it follows that
Here,
where
Equilibrium
Values LRR Model SSG Model Experiments
(&ll)oo 0.158 0.204 0.201
(^22)00 -0.123 -0.148 -0.147
(£"12)00 -0.187 -0.156 -0.150
(StfA)oo 5.32 5.98 6.08
in equilibrium. Since,
it then follows from (24) that the Reynolds stress transport equation
has the equilibrium form
By the use of integrity bases from linear algebra, Pope (1975) first
showed that the general solution to the implicit algebraic stress equa-
tion (26) is of the form:
where
Turbulent Transport Equations 197
and
where
198 C. G. Speziale
Here, 0:1,0:2 and 0:3 are not constants but rather are related to the
coefficients C\ — C$ and g. In mathematical terms, they are "projec-
tions" of the fixed points of Aij and Mijkt onto the fixed points of
bij - quantities that can vary from one flow to the next. However,
for two-dimensional mean turbulent flows, it appears that C\ — €4
can be approximated by constants due to the linear dependence on
b^ which allows us to use the principle of superposition. (For three-
dimensional turbulent flows, the general representation for Mijkf is
nonlinear in bij which leads to an inconsistency that will be discussed
later).
It should be noted that these models obtained via the algebraic
stress approximation need to be regularized before they are applied
to complex turbulent flows. This can be accomplished via a Fade
type approximation whereby we take (Gatski and Speziale 1993):
which forms the basis for the standard K — e model of Launder and
Spalding (1974). However, in practical turbulent flows we do not
have a separation of scales; T] and £ are of O(l). Nonetheless, for
two-dimensional turbulent shear flows in equilibrium, the new model
of Gatski and Speziale (1993) yields
with
where
where
Hence,
where
In physical terms,
where
where
where
Applications
For weakly inhomogeneous turbulent flows that are near equilibrium,
we can extend the K and s transport equations by the addition of
gradient transport terms that are obtained by a formal expansion
technique:
Turbulent Transport Equations 205
where Uk and cre are constants that typically assume the values of
1.0 and 1.3, respectively. These forms are assumed to be valid ap-
proximations at high Reynolds numbers.
We will now consider several non-trivial applications of the two-
equation model derived herein which can be referred to as an explicit
algebraic stress model (ASM) based on the SSG second-order closure.
The first case that will be considered is homogeneous shear flow in a
rotating frame (see Figure 1). In this flow, an initially isotropic tur-
bulence (with turbulent kinetic energy KQ and turbulent dissipation
rate £Q) is suddenly subjected to a uniform shear (with constant shear
rate 5) in a reference frame rotating steadily with angular velocity
ft. In Figures 2(a)-2(c), the time evolution of the turbulent kinetic
energy predicted by this new two-equation model is compared with
the large-eddy simulations (LES) of Bardina, Ferziger and Reynolds
(1983), as well as with the predictions of the standard K — e model
and the full SSG second-order closure. From these results, it is clear
that the new two-equation model yields the correct growth rate for
pure shear flow (ft/5 = 0) and properly responds to the stabilizing
effect of the rotations ft/5 = 0.5 and ft/5 = —0.5. These results
are remarkably close to those obtained from the full SSG second-
order closure as shown in Figure 2. In contrast to these results,
the standard K — £ model overpredicts the growth rate of the turbu-
lent kinetic energy in pure shear flow (ft/5 = 0) and fails to predict
the stabilizing effect of the rotations illustrated in Figures 2(b)-2(c).
Since the standard K — £ model makes use of the Boussinesq eddy
viscosity hypothesis, it is oblivious to the application of a system
rotation (i.e., it yields the same solution for all values of ft/5). The
new two-equation model predicts unstable flow for the intermediate
band of rotation rates —0.09 < ft/5 < 0.53; this result is generally
consistent with linear stability theory that predicts unstable flow for
0 < ft/5 < 0.5.
In Figure 3, the prediction of the new two-equation model for
the mean velocity profile in rotating channel flow is compared with
the experimental data of Johnston, Halleen and Lezius (1972) for a
rotation number Ro — 0.068. It is clear from these results that the
model correctly predicts that the mean velocity profile is asymmetric
in line with the experimental results - an effect that arises from
Coriolis forces. In contrast to these results, the standard K — e
model incorrectly predicts a symmetric mean velocity profile identical
to that obtained in an inertial frame (the standard K — £ model is
206 C. G. Speziale
ston (1980) and Eaton and Johnston (1981) for turbulent flow past
a backward facing step. It is clear that these results are excellent:
reattachment is predicted at x/H sa 7.0 in close agreement with the
experimental data. In contrast to these results, the standard K — e
model predicts reattachment at x/H K 6.25 - an 11% underpredic-
tion (see Thangam and Speziale 1992). This predominantly results
from the inaccurate prediction of normal Reynolds stress differences
in the recirculation zone as discussed by Speziale and Ngo (1988).
These more complex closures are based on the full Reynolds stress
transport equation with turbulent diffusion:
where
(then, £,-j = §£<% and a modeled transport equation for the scalar
dissipation rate £ is solved that is of the same general form as that
discussed in Section 2.6). However, this assumption is debatable as
discussed by Durbin and Speziale (1991). More generally, a repre-
sentation of the form
can be used where the algebraic model of Speziale and Gatski (1992)
discussed in Section 2.6 is implemented.
The only additional model that is needed for closure in high-
Reynolds-number inhomogeneous turbulent flows is a model for the
third-order diffusion correlation Cijk- This is typically modeled using
a gradient transport hypothesis:
a homogeneous turbulence are shut off; the flow then gradually re-
turns to isotropy (i.e., 6,-j —>• 0 as t —> oo). In Figure 6, results for the
Reynolds stress anisotropy tensor obtained from the Speziale, Sarkar
and Gatski (SSG) and Launder, Reece and Rodi (LRR) models are
compared with the experimental data of Choi and Lumley (1984) for
the return-to-isotropy from plane strain (here, T = Sot/Ko). It is
clear from these results that the models predict a gradual return to
isotropy in line with the experimental data. In contrast to these re-
sults, all two-equation models — including the more sophisticated one
based on an anisotropic eddy viscosity derived herein - erroneously
predict that at T = 0, fyj abruptly goes to zero.
It is worth noting at this point that while the SSG model was
derived and calibrated based on near equilibrium two-dimensional
mean turbulent flows, it performs remarkably well on certain three-
dimensional, homogeneously strained turbulent flows. The predic-
tions of the SSG and LRR models for the normal Reynolds stress
anisotropies are compared in Figure 7 with the direct simulations
of Lee and Reynolds (1985) for the axisymmetric expansion (here,
t* = Tt where F is the strain rate).
are compared with experimental data (Laufer 1951) for the log-layer
of turbulent channel flow. Most of the models yield errors rang-
ing from 30% to 100%. These models are then typically forced into
agreement with the experimental data by the addition of ad hoc wall
reflection terms that depend inversely on the distance from the wall
— an alteration that compromises the ability to apply the model in
complex geometries where the wall distance is not always uniquely
defined. Only the SSG model yields acceptable results for the log-
Turbulent Transport Equations 215
layer without a wall reflection term. This results from two factors:
(a) a careful and accurate calibration of homogeneous shear flow (see
Table 3) and (b) the use of a Rotta coefficient that is not too far re-
moved from one (see Abid and Speziale 1993). The significance of
these results is demonstrated in Figure 8 where full Reynolds stress
computations of turbulent channel flow are compared with the ex-
perimental data of Laufer (1951). It is clear that the same favorable
216 C. G. Speziale
CHANNEL FLOW
Equilibrium LRR SL FIT SSG Experimental
Values Model Model Model Model Data
611 0.129 0.079 0.141 0.201 0.22
&12 -0.178 -0.116 -0.162 -0.160 -0.16
&22 -0.101 -0.082 -0.099 -0.127 -0.15
&33 -0.028 0.003 -0.042 -0.074 -0.07
SK/e 2.80 4.30 3.09 3.12 3.1
where
it follows that models for Mijke that are nonlinear in 6;j are also
nonlinear in E^j. This is a fundamental inconsistency that dooms
these models to failure. It is clear that it is impossible to describe a
range of RDTflows- which are linear - with nonlinear models (the
principle of superposition is violated). Furthermore, Shih and Lum-
ley (1985, 1993) unnecessarily introduce higher degree nonlinearities
and non-analyticity to satisfy realizability. In the process of doing
220 C. G. Speziale
3 COMPRESSIBLE TURBULENCE
Continuity
Momentum
Energy
Turbulent Transport Equations 221
where
p = mass density
Ui = velocity vector
p = thermodynamic pressure
[JL = dynamic viscosity
<Tj-j = viscous stress tensor
T = absolute temperature
K = thermal conductivity
R = ideal gas constant
Cv = specific heat at constant volume
$ = viscous dissipation
d
(\ \A*• -= dxi\( \)
where
Continuity
222 C. G. Speziale
Momentum
Energy
where
Transport Terms
Speziale and Sarkar (1991) proposed the following models for the
turbulent transport terms:
where
given that w,' is the fluctuating vorticity. Then, the Sarkar et al.
(1991) model can be used for ec yielding
where
where
Pressure-Dilatation Model
For turbulent shear flows, Sarkar (1992) proposed the following model
for the pressure-dilatation correlation:
where
where
where C*, C*D and C*E are variable density extensions of the incom-
pressible coefficients derived earlier, in Eq. (32), that depend on
both 77 and £.
This anisotropic eddy viscosity model can be solved in conjunc-
tion with modeled transport equations for K and es that take the
form
where
and the compressible terms e c , p'w''» and u" are modeled as before.
(see Reynolds 1987). According to (95), the integral length scale will
decrease under an expansion (F > 0) and increase under a compres-
sion (F < 0) - results that are clearly in error as first pointed out by
Reynolds (1987).
Finally, in regard to homogeneous turbulence, we will consider the
problem of compressible homogeneous shear flow. Here, an initially
isotropic turbulence is subjected to a uniform shear rate S with the
corresponding mean velocity gradient tensor
230 C. G. Speziale
The time evolution of the turbulent kinetic energy and turbulent dis-
sipation rate in compressible homogeneous shear flow obtained from
the SSG model with the dilatational models of Sarkar are compared
with the DNS results of BlaisdeU et al. (1991), for Mto = 0 and
Mt0 = 0.307, in Figure 11. It is clear from these results that the
model does an excellent job in reproducing the dramatic reduction
in the growth rate that arises from compressible effects (without
explicit dilatational terms, all existing second-order closures drasti-
cally overpredict the growth rate of the turbulent kinetic energy). In
Table 4, the equilibrium values obtained using a variety of pressure-
strain models - combined with the dilatational models of Sarkar -
are compared with the DNS results of BlaisdeU et al. (1991). Here,
the results are not so favorable: the normal Reynolds stress anisot-
ropies are drastically underpredicted and the Reynolds shear stress
anisotropy is overpredicted by more than 25%.
4 CONCLUDING REMARKS
ACKNOWLEDGMENTS
5 REFERENCES
Abid, R. and Speziale, C. G., 1993. "Predicting equilibrium states
with Reynolds stress closures in channel flow and homogeneous
shear flow," Phys. Fluids A 5, pp. 1776-1782.
Durbin, P. A., 1993. "A Reynolds stress model for near-wall turbu-
lence," J. Fluid Mech. 249, pp. 465-498.
Prandtl, L., 1945. "Uber ein neues formelsystem fur die ausge-
bildete turbulenz," Nachr. Akad. Wiss. Gottingen Math.
Phys. Kl 1945, pp. 6-19.
Raj, R., 1975. "Form of the turbulence dissipation equation as
applied to curved and rotating turbulent flows," Phys. Fluids
18, pp. 1241-1244.
Rodi, W., 1976. "A new algebraic relation for calculating the Rey-
nolds stresses," ZAMM 56, pp. T219-221.
Rogers, M. M., Moin, P. and Reynolds, W. C., 1986. "The structure
and modeling of the hydrodynamic and passive scalar fields in
homogeneous turbulent shear flow," Stanford University Tech-
nical Report No. TF-25.
Rubinstein, R. and Barton, J. M., 1990. "Nonlinear Reynolds stress
models and the renorrnalization group," Phys. Fluids A 2, pp.
1472-1476.
240 C. G. Speziale
Speziale, C. G., Abid, R. and Durbin, P. A., 1994. "On the read-
ability of Reynolds stress turbulence closures," J. Sci. Comp.
9, pp. 369-403.
AN INTRODUCTION TO
SINGLE-POINT CLOSURE
METHODOLOGY
Brian E. Launder
1 INTRODUCTION
1.1 The Reynolds Equations
The material presented in Section 1 will replicate that covered
in earlier chapters. It may nevertheless be helpful to include it to
familiarize the reader both with the nomenclature adopted and the
assumptions being made.
The instantaneous velocity field in a turbulent flow is described
by the continuity and the Navier-Stokes equations which in conser-
vative form may be written:
Here /5,P,/i and F{ denote the density, pressure, viscosity and body
force per unit volume. The tildes appearing above all quantities serve
as a reminder that each, potentially, will display fluctuations due to
turbulence. Here, however, density and viscosity variations will be
243
244 B. E. Launder
Since the first term within the square brackets is the viscous stress,
the second, -puiuj, has, naturally, been re-interpreted as a turbulent
stress or, more usually, the Reynolds stress tensor. As the tensor
is symmetric there are six independent Reynolds stress components.
They are unknown elements in the averaged equations of motion and
the theme of these notes is that of obtaining a satisfactory approxi-
mation to their magnitude.
The quantity pvivj is known as the mass-weighted Reynolds stress. Notice that,
in contrast with conventional "volume" averaging, the number of turbulent cor-
relations arising from averaging the convective terms is just one - as in a uniform
density flow. This is the feature that has made the use of mass weighted quantities
in variable-density flow so popular.
246 B. E. Launder
The left side of the equation expresses the total rate of increase
of the correlation ujuj for a small identified packet of fluid. The rate
of change arises from an imbalance of the terms on the right. Here
the terms have been grouped, following well-established practice, so
as best to allow a physical interpretation of the processes. One line is
given to each process and, beneath each term, appears a shorthand
symbol for the process in question which we shall use to simplify later
equations. The first two processes represent rates of creation of UiUj,
in one case by the effects of mean strain, P,-j, and in the other by
body forces,G;j. The first of these, comprising products of Reynolds
stresses and mean velocity gradients, can clearly be treated exactly
in a second-moment closure. If the body force is linear, as when one
examines the flow in a rotating coordinate frame, that too can be
handled without further approximation.
The correlation between fluctuating pressure and fluctuating
strain, faj, is a very important one. We note that its trace is zero,
since by continuity
Due to primary
^
9Z2 = A
Due to curvature
QUi _ g
9si
dU!/dx2 -^"g-
Duetoashear 0 ~ulj^ °
dUt/dxa
Thus:
So, the model may be organized in a form with just one undetermined
coefficient. After some algebra eq (2.6) may be expressed
where
It equals unity when the stress field is isotropic but its most useful
property (Lumley, 1978) is that it vanishes in two-component tur-
bulence. It is thus tempting to refer to A as the "flatness factor",
262 B. E. Launder
where
The first two lines of this equation are identical to the quadratic
model while the coefficients c-2 and c'2 of the cubic terms are freely
assignable. The quantities that these coefficients multiply are evi-
dently of very different length, that associated with c'2 being particu-
larly long. It has been found, however, that the behaviour of simple
free shear flows can be very satisfactorily mimicked by setting c% to
0.6 and c'2 to zero, Fu et al. (1987) (see also Launder (1989)), thus
reducing the algebra to more manageable proportions.
Notice that
With the coefficient c2# set to about 0.5 this formula is distinctly more
successful than the linear quasi-isotropic form (i.e. the equivalent
of eq (2.7a) for 0jj) which in t his case is obtained with no free
coefficient:
where c^g from a QI analysis equals 1/3 for buoyant flows while,
again, most applications have tended to chose a value of about 0.5.
The accuracy with which these simple models can hope to capture
the real processes hardly justifies a separate optimization for each
constant. A choice in the range 0.3 to 0.5 would nearly always be
appropriate.
In the case of Coriolis forces, there are grounds for taking the co-
efficient €3 to be half that adopted for c%, a topic discussed extensively
by Launder et al. (1987). To illustrate this point, an axisymmetric
swirling jet can be analysed perfectly well in a stationary reference
frame but it can equally well be examined with the coordinate frame
rotating about the jet's axis. If, however, one takes any ratio of 03 : c2
other than one half, the predicted results are dependent on the rate of
rotation of the coordinate system. Obviously, the predictions should
be independent of the observer's frame of reference.
In fact, there are various ways of ensuring independence of the
results from the observer's motion as discussed by Fu et al. (1987).
One is the practice noted above; another is to include the convec-
tive transport tensor DUiUj/Dt - which is hereafter denoted dj -
along with the production tensor and Coriolis tensors in applying
the isotropization of production approach:
To model the parts of 4>ij and fag arising from p^ we seek forms
containing only turbulent quantities. Experiments indicate that grid
turbulence made strongly non-isotropic by passing it through a duct
of rapidly changing cross-sectional shape will revert towards isotropy
once the mean strain is removed. In the absence of any alternative
266 B. E. Laundei
process, we must conclude that <$>(j\ (which, like the other parts o:
(f>ij, is traceless) is the agency promoting this reversion. In mos1
computations of complex flows Rotta's (1951) linear return mode'
has been adopted:
Equations (2.8) and (2.14) have been the basis of many differ-
ent proposals and have been incorporated into several commercial
software packages. For this reason the pair of equations (including,
where appropriate, wall-reflection terms as discussed below) is often
referred to as the basic model.
The question arises, however, as to what values should be as-
signed to the coefficients c\ and c 2 . Values proposed for c\ range
from 1 to 5 while recommendations for c2 cover the range from zero
to 0.8, Fig 2.1. The range of different choices suggests, at first glance,
that since such disparate values have been put forward the whole
approach is worthless. Looked at with an engineer's eye, however,
one might be tempted to fit a straight line through the 'data points'.
Now, in the case of a simple shear flow in local equilibrium (i.e. where
turbulence generation and dissipation processes are in balance) it is
readily shown that with this basic model the resultant stress ten-
sor depends not on the individual values of c\ and c2 but rather on
the single parameter (1 — c 2 )/ci. The line in Fig 2.1 is simply that
corresponding to (1 — c 2 )/ci = 0.23 which evidently does rather a
good job of fitting the various proposals. What we conclude is that,
for equilibrium simple shear flows, the very different pairs of ci and
c2 will lead to nearly the same results. In order to pick the 'best'
pairing one needs to look at non-equilibrium cases. We have already
268 B. E. Launder
Wall Effects on ^
where
In the above / = fc3/2/€ and y denotes the wall distance. Figure 2.2,
from Craft and Launder (1992) shows the application of the basic
Single-Point Closures 271
model using the wall correction of eq (2.18) to compute the flat plate
boundary layer and the impinging turbulent jet. The latter flow is a
critical turbulence modelling test case: numerous models developed
by reference to flows parallel to walls give seriously incorrect pre-
diction of the radial wall jet development as the flow develops away
from the vicinity of the stagnation point (Craft et a/., 1993a). The
figure indicates that, for the impinging jet, the form recommended
for 4>™j (solid line) achieves much better agreement with experiment
than does the earlier (and more widely used) proposal of Gibson and
Launder (1978) shown by a broken line.
It would, however, seriously misrepresent the situation to leave
the impression that modelling of near-wall influences was in a satis-
factory state. Equation (2.18) and similar approaches may be ade-
quate if one is dealing with a single plane or mildly curved surface.
In most engineering applications, however, one needs to predict flows
within an enclosure or around bodies with several distinct faces. In
these cases the approach indicated by eq (2.18) is, at best, a scheme
that requires ad hoc, case-specific interpretation while in others it is
simply unworkable. In flow through a square duct, for example, there
are four vector directions normal to a wall and, correspondingly, four
wall-normal distances. Analogous problems arise in handling flow
through tube banks, within internal combustion engines or in turbo-
machinery flows. It is this geometrical complexity that has spurred
efforts to eliminate from the closure the very parameters on which
traditional 'wall-proximity' corrections depend, namely wall distance
and unit vectors.
What has come in their place? Firstly, as noted earlier, a benefit
of using the non-linear models for <$>{j\ and <^,-j2 is that there is a
much diminished wall-proximity correction. Secondly, there is the
recognition that, within the immediate wall vicinity, the turbulence
is varying so rapidly in the direction normal to the wall that some
explicit correction for inhomogeneity should be made to the pressure-
strain process, Bradshaw et al. (1987). For example in replacing the
exact expression for faji in eq (2.4a), there is the assumption that
dU'k/dx'j (that is, the mean velocity gradient evaluated at a distance r
from the point where (f>ij is to be determined) can be replaced by that
at the point itself. Launder and Tselepidakis (1991) first proposed
improving this assumption by adopting, instead, an effective mean
velocity gradient
272 B. E. Launder
One evident weakness of these forms is that while the indices i and k
on the left side of the equation can be interchanged without altering
the resultant product, such a rearrangement on the right intrinsi-
cally alters the form. No ambiguity arises because in the u^uj and
Ufd transport equations a d/dx^ operation is applied to the triple
moments; nevertheless, the difference in character between the exact
and the modelled form would appear to be a significant shortcoming.
In fact, forms very similar to (2.21) and (2.22) can be obtained by
making sweeping closure simplifications to the transport equations
for the triple moments (Hanjalic and Launder, 1972; Launder, 1976).
In that case, the model for u^UjUk consists of three terms identical to
the right side of (2.21) but with a permutation of the indices i, j and
k; likewise, that for u^u^Q also consists of three terms in which w/t, it,-,
0 successively occupy the position of u^ in (2.22). These somewhat
more elaborate and superficially more correct models do not, in prac-
tice, seem to bring better agreement when used in numerical solvers.
This could be due, at least partly, to the fact that those workers who
have adopted (2.21) and (2.22) for the triple moments have generally
not included any model for the pressure diffusion terms:
and
where Tt denotes the net transport of c and EST stands for "extra
276 B. E. Launder
strain terms".
Originally, both coefficients cei and c £ 2 were taken as constant
with "standard" values of about 1.44 and 1.92 respectively. Lum-
ley (1975), however, argued that ct\ should be taken as zero and
c e 2 should become a function of the second invariant A^. The subse-
quent studies on buoyant diffusion undertaken by Zeman and Lumley
(1979), however, recommend the retention of both types of process.
Specifically they proposed
so finally we obtain
and finally,
The first such 1-equation EVM was proposed by Prandtl (1945) and
subsequently reinvented by Emmons (1954). It is not just in eddy
viscosity treatments that a prescribed length scale has been adopted.
Much of Donaldson's group's work with second-moment closures ( e . g .
Donaldson et a/., 1972) has used a prescribed profile of /, as has the
ASM study of turbulence driven secondary flows by Launder and
Single-Point Closures 281
or
or
This more elaborate form of the mixing length hypothesis has been
used with a surprising degree of success in the computation of flows
near spinning discs and cones where the velocity vector is strongly
skewed.
early 1970 's but only relatively recently, with the prospect of applying
CFD to very complex flows, has an impetus developed to devise mod-
els with a width of applicability approaching that of second-moment
closure for a computational cost similar to a linear EVM. Such non-
linear EVMs have many similarities with ASM's but, for complex
flows, they require much less computational effort (typically, one
half to one quarter) due to their improved stability characteristics.
Recent contributors to models of this type include Speziale (1987),
Nisizima and Yoshizawa (1987), Rubinstein and Barton (1990), My-
ong and Kasagi (1990), Shin et al. (1993) and Taulbee et al. (1993).
Our experience at UMIST is that most of the weaknesses of the lin-
ear EVM cannot be rectified by introducing just quadratic terms to
the stress-strain relation. Only by proceeding to cubic level does one
find sufficient variety in the stress-strain couplings to achieve the
desired effects. The UMIST work thus adopts the following stress
strain relation:
where
Ci C2 C3 C4 Cs CQ C'j
Figures 2.4 and 2.5 show two recent applications of the above
model (Craft et a/., 1993b) to flows that defeat conventional lin-
ear models, namely fully developed swirling flow in a pipe and the
impinging jet. In the former the roughly parabolic variation of the
circumferential velocity with radius is correctly reproduced while any
linear eddy viscosity model gives rise to solid body rotation. Equally,
the dynamic field of the impinging jet is captured just as well with the
non-linear model as with any of the tested second-moment closures.
3.1 Introduction
The closure proposals made so far imply negligible effect of vis-
cosity on the energy containing motions and no influence of the
mean-strain field on the dissipative ones. While we may expect these
assumptions to be reasonably valid throughout most of a turbulent
flow, the no-slip condition at a rigid boundary ensures that over some
region of a turbulent wall layer, however thin, viscous effects on the
transport processes must be large. The present chapter provides a
brief account of turbulent transport in this viscosity-modified sub-
layer. A more extensive treatment, albeit less up to date, has been
provided by the writer elsewhere, Launder (1986).
In round terms, molecular influences may be expected to be in-
Single-Point Closures 285
fluential over a region extending from the surface to where the local
"eddy" Reynolds number, based on a typical eddy dimension normal
to the wall and the intensity of velocity fluctuation in that direction,
is of order 102. Turbulence models for this region may reasonably
contain elements that introduce appropriately these viscous influ-
ences. The region in question is thin (even in low-speed laboratory
studies rarely extending over more than 2mm), the processes are
highly complex and the acquisition of accurate experimental data is
greatly complicated not just by the thinness of this sublayer but by
wall-proximity influences of various kinds on the instruments them-
selves. It is thus no wonder that , despite the important contributions
being made by direct numerical simulations, our knowledge of this
important region of flow is still incomplete.
Turbulence models applied in this low-Reynolds-number region
have usually been developed from some high-Reynolds-number clo-
sure by introducing various viscosity-dependent terms and/or by
making the coefficients of existing terms functions of a turbulence
Reynolds number. These adaptations, while largely empirical, are
chosen to ensure that certain general kinematic constraints are satis-
fied. Models of this type, which we refer to as "low-Reynolds-number-
treatments", are discussed in Section 3.3
Although the extreme thinness of the viscosity-affected sublayer,
in some respects, complicates the task of model development, in an-
other it simplifies the problem; for, streamwise convective transport
within this sublayer is often sufficiently small compared with diffu-
sive or (in the case of properties of the turbulent field) source or
sink processes that it may be neglected. In cases where surface
transpiration is absent and where the influence of force fields (in-
cluding pressure gradients) on the region is negligible, the flow-field
properties, suitably normalized, are then functions of only a normal-
distance Reynolds number. For the case of the mean velocity, the
resultant distribution is known as the 'Law of the Wall'. This connec-
tion between velocity, friction velocity, \/TW/p , kinematic viscosity
and normal distance can be used in place of the no-slip boundary
condition to avoid the need to extend what may be a costly two-
or three-dimensional numerical solution to the wall itself. This ap-
proach is especially advantageous if the matching is applied outside
of the viscosity dependent region (though close enough to the wall
for streamwise transport to be negligible); for then the turbulence
model used for the numerical computation does not have to include
286 B. E. Launder
viscous effects and, moreover, one escapes the need for the especially
fine mesh that is inevitably required to resolve the viscous region
because there the curvature in the profiles of both the mean velocity
and turbulence properties is so high.
However, as CFD gains in maturity and computational power
per dollar continues to double every 18 months or so, an increas-
ingly large proportion of problems being tackled are ones where it is
not safe to treat the near-wall sublayer as though it is in its quasi-
equilibrium, 'universal' state. That is why, in the present article, a
detailed modelling of the sublayer is the only route considered.
where the a's, 6's and c's are functions of time whose mean value must
be zero, since ul = 0. The linear variation of u\ over a significant
sublayer region is well confirmed by experiment and, more recently,
by direct numerical simulation.
It follows from (3.1) that
This cubic variation of turbulent stress helps explain why the viscous
layer exhibits a distinct, if very thin, region where turbulent stress
is negligible compared with viscous stress which gives way, rather
rapidly, for small increase in X2, to a zone where molecular transport
is of only minor importance in comparison with that due to turbulent
exchange processes.
Likewise, the turbulent kinetic energy variation is given by:
where
Preamble
where C M normally takes the constant value 0.09 and /^ is the "vis-
cous" damping function employed in models of that type. On com-
paring these two equations it is evident that they would be identical if
290 B. E. Launder
u\lk were equal to 0.342 /M. Figure 3.1, from Launder (1986), shows
the variation of the quantity across the sublayer region as deduced by
Patel, Rodi and Scheuerer, from admittedly rather imprecise experi-
mental data (the uncertainty band attaching to /^ is probably larger
than that shown for u\/k}. The figure does show quite convincingly,
however, that the two parameters vary in essentially the same way
across the sublayer. The more recent direct simulation data lead
to essentially the same distribution of f ^ . Thus, provided, as the
wall is approached, the proper diminution of u\ can be modelled,
there would be little need for other "viscous" damping. The pro-
cesses that are most influential in determining the level of u\ are the
pressure-strain and dissipative correlations. There are good reasons
for supposing that the former is dependent on viscosity only within
the near wall sublayer while u^/k increases with distance from the
wall over a more extensive region. It implies, therefore, that the
requisite damping of u\/k should, to a large extent, be provided by
non-viscous parameters.
The "store of new information about near-wall turbulence" noted
above refers to the results of full simulations of turbulent shear flows
that have become available over the last five years. The first of these,
the flow between plane and slightly curved parallel surfaces by Kim
et al. (1987) and Moser and Moin (1984), present detailed budgets of
each Reynolds-stress component right up to the wall itself, including
the contributions made by the pressure-strain correlation. Thus, for
the first time, the modeller can make direct appeal to data of the
processes to be approximated — data generated not by experiment
but by computer simulation.
Component dissipation
so that, by the start of the fully turbulent region (Rt « 150), the
dissipation is very nearly isotropic.
Equation (3.7), while attractively simple, is not an exact limiting
form. It may readily be deduced (Launder and Reynolds, 1983) from
the polynomial expansions for the velocity components that, while
it is correct if neither i nor j takes the value 2, if one of the indices
refers to the direction normal to the wall:
where 'a' takes the value 1 or 3. In the immediate vicinity of the wall u\ is
negligible compared with u\ or v% giving the desired limiting behaviour.
Single-Point Closures 293
early model. The more recent results to emerge from the UMIST
group (to be shown below) have been obtained with the following
choice of the coefficient of <^,-ji:
(2.19), is the need to correct eq (2.6) for the rapid spatial variation of
the mean velocity gradient. In a channel flow the correction, given by
eq (2.19) or similar forms, is of substantial importance in the buffer
region but becomes negligible further from the wall. This behaviour
is consistent with the conclusions reached by Bradshaw et al. (1987)
from a processing of the DNS data.
Diffusion
This formula embodies the' same underlying idea but does not employ
the unit vector. However, while the imbalance in the u\ equation is
entirely eliminated with this form, that in the ~u\u^ equation is only
reduced by one half. Nevertheless, since Tselepidakis (1992) found
his formulation to have only a secondary effect on the stress profiles,
the above replacement — which removes about 75% of the (small)
problem - is probably satisfactory.
Having regard for the fact that, in a thin shear flow, the mean
velocity is significant only in directions parallel to the wall (so index
'i' denotes 1 or 3), the second term is evidently small compared with
the first since, within the sublayer, rates of change of the instanta-
neous turbulence field will be much larger in direction x% than in x\
or £3. The first term may, in fact, be expressed in terms of eq (3.12)
or (3.13). If direction x\ is aligned with the near-wall mean velocity
both lead to the result:
As noted in Section 3.2, e varies as x\ near the wall and thus the
term c^f^l/k tends to a constant value as the wall is approached.
Virtually all proposals for extending the e equation to the wall adopt
this form.
Finally, there remains the matter of viscous effects on the diffu-
sion of e to be accounted for. _The writer's group has retained the
high-Reynolds-number form, c^fc/e, as the appropriate diffusion co-
efficient in the low-Reynolds-number region while Prud'homme and
Elghobashi (1983) multiply the coefficient ce by a viscous damping
function, /^. All proposals have included the exact viscous trans-
port vdt/dxj as an addition to the high-Reynolds-number diffusion
model. There is also the question of pressure diffusion to be consid-
ered. In high-Reynolds-number turbulence no explicit accounting of
this process was attempted. An order-of-magnitude estimate of this
term in the exact e equation suggests, however, that as the wall is
approached the process becomes significant (Launder, 1986). This
suggests that a separate approximation ought to be incorporated. A
possible model for the additional term that is significant only in the
viscous region has been proposed in that article:
3.4 Applications
sublayer with the DNS results of Kim et al. (1987). The adopted
model of <pij2 was the complete cubic form (c2 = 0.55; c'2 = 0.6 in eq
(2.7(d)) with an inhomogeneity correction (eq (2.19)) but no "wall-
reflection" term. The agreement of the closure computations with
the DNS results is reasonably satisfactory.
The second application, from Launder and Tselepidakis (1994),
also relates to channel flow but here the channel rotates about the
#3 axis, i.e. in orthogonal mode rotation. The resultant asymme-
try in the turbulence intensity profiles arises from the direct effect
of the Coriolis forces on the turbulent stresses in the x\ — x^ plane,
augmenting fluctuations normal to the wall on the pressure (left)
side of the channel which, in turn, increases the shear stress u\u-i
and thus of streamwise fluctuations too. The model in this case was
similar to that adopted in the previous example except that, because
the computations were carried out rather earlier, the simplified cubic
model of </>,-j2 was adopted (c2 = 0.6 ; c'2 = 0) and, in consequence a
small 'wall-reflection' contribution was added. Generally the closure
Single-Point Closures 301
ACKNOWLEDGEMENTS
4 REFERENCES
Andre, J.C., de Moor, G., Lacarrere, P., Therry, G. and du Vachat,
R., 1979. "The clipping approximation and inhomogeneous
turbulence simulations," Turbulent Shear Flows - 1 (ed. F.J.
Durst et a/.), 307, Springer-Verlag, Heidelberg.
Andre, J.C., de Moor, G., Lacarrere, P. and du Vachat, R., 1976.
"Turbulence approximation for inhomogeneous flows: Part I:
The clipping approximation," J. Atmos. Sciences, 33, p.476.
Bradshaw, P., Ferriss, D.H. and Atwell, N.P., 1967. "Calculation of
boundary-layer development using the turbulent energy equa-
tion," J. Fluid Mech. 28, p.593.
Bradshaw, P., Mansour, N.N. and Piomelli, U., 1987. "On local
approximations of the pressure strain-term in turbulence mod-
elling," Proc. Summer Program, Center for Turbulence Re-
search, Stanford University, p.159.
Cheah, S.C., Cheng, L., Cooper, D. and Launder, B.E. 1993. "On
the structure of flow in spirally fluted tubes," 5th IAHR Con-
ference on Refined Flow Modelling and Turbulence Measure-
ments, pp.293-300, Presse Ponts et Chaussees, Paris.
Single-Point Closures 303
Chou, P.Y., 1945. "On velocity correlations and the solution of the
equations of turbulent fluctuation," Quart. J. Appl. Math. 3,
pp.38-54.
Cooper, D., Jackson, D.C., Launder, B.E. and Liao, G.X., 1993.
"Impinging jet studies for turbulence model assessment: Part
I: Flow-field experiments," Int. J. Heat Mass Transfer 36,
p.2675.
Craft, T.J. and Launder, B.E., 1989. "A new model for the pres-
sure/scalar gradient correlation and its application to homoge-
neous and inhomogeneous flows," Proc. 7th Symposium Tur-
bulent Shear Flows, Stanford University, Paper 17-1.
Craft, T.J., Launder, B.E. and Suga, K., 1993b. "Extending the
applicability of eddy viscosity models through the use of defor-
mation invariants and non-linear elements," 5th IAHR Confer-
ence on Refined Flow Modelling and Turbulence Measurements,
Presse Fonts et Chaussees, Paris, pp. 125-132.
Fu, S., Huang, P.G., Launder, B.E. and Leschziner, M.A., 1986. "A
comparison of algebraic and differential second-moment clo-
sures for axisymmetric turbulent shear flows with and without
swirl," Mech. Eng. Dept., Rep. TFD/86/9, UMIST.
Kebede, W., Launder, B.E. and Younis, B.A., 1985. "Large am-
plitude periodic flow: A second-moment closure study," Proc.
5th Turbulent Shear Flows Symp., Cornell University, Ithaca.
306 B. E. Launder
Koo Sin Lin, M., Launder, B.E. and Sharma, B.I., 1974. "Prediction
of momentum, heat and mass transfer in swirling turbulent
boundary layers," ASME J. Heat Transfer 94C, p.204.
Launder, B.E. and Li, S-P, 1993. "On the prediction of flow over
riblets via 2nd-moment closure," Proc. Conf. Near-Wall Tur-
bulence (ed. R.M.C. So et a/.) Elsevier, New York.
Launder, B.E. and Li, S-P., 1994. "On the elimination of wall
topography parameters from second-moment closure," Phys.
Fluids 6 (2), p.999.
Launder, B.E. and Shima, N., 1989. "A second-moment closure for
the near-wall sublayer: development and application," AIAA
J. 27, p.1319.
Launder, B.E., Tselepidakis, D.P. and Younis, B.A., 1987. "A
second-moment closure study of rotating channel flow," J. Fluid
Mech. 183, p.63.
Launder, B.E. and Ying, W.M., 1973. "Prediction of flow and heat
transfer in ducts of square cross section," Proc. Inst. Mech.
Engrs. 187, p.455.
Morse, A.P., 1980. "Axisymmetric free shear flows with and with-
out swirl," PhD Thesis, Faculty of Engineering, University of
London.
Newman, G., Launder, B.E. and Lumley, J.L., 1981. "Modelling the
behavior of homogeneous, scalar turbulence," J. Fluid Mech.
lll,p.217.
Nisizima, S. and Yoshizawa, A., 1987. "Turbulent channel and Cou-
ette flows using an anisotropic k-e model," AIAA J. 25, p.414.
Orszag, S., 1993. "Renormalization group modelling and turbu-
lence simulations," Near-Wall Turbulent Flows (Ed R.M.C. So
et a/.), Elsevier, Amsterdam.
Patel, V.C., Rodi, W. and Scheuercr, G., 1985. "Turbulence models
for near-wall and low Reynolds number flows: a review," AIAA
J. 23, p.1320.
Single-Point Closures 309
Prandtl, L., 1945. "Uber ein neues Formelsystem fur die ausge-
bildete Turbulenz," Nachrichten von der Akad der Wissenschaft,
Gottingen.
Rodi, W., 1976. "A new algebraic relation for calculating the Rey-
nolds stresses," ZAMM 56, p.219.
Speziale, C.G., 1987. "On non-linear k-L and k-e models of turbu-
lence," J. Fluid Mech. 178, p.459.
Taulbee, D.B., Sonnenmeier, J.R. and Wall, K.M., 1993. "Appli-
cation of a new non-linear stress-strain model to axisymmetric
swirling flow," Proc. 2nd Int. Symp. on Engrg. Turbulence
Modelling and Measurement (ed. W. Rodi and F. Martelli),
Elsevier, Amsterdam.
Tselepidakis, D.P., 1992. "Development and application of a new
second-moment closure for turbulent near- wall flows," PhD
Thesis, Faculty of Technology, University of Manchester.
Wyngaard, J.C. and Sundararajan, A., 1979. "The temperature
skewness budget in the lower atmosphere and its implications
for turbulence modelling," Turbulent Shear Flows- 1, Springer-
Verlag, Heidelberg, p.319.
Younis, B.A., 1984. "On modelling the effects of streamline curva-
ture on turbulent shear flows," PhD Thesis, Faculty of Engi-
neering, University of London.
Zeman, O. and Lumley, J.L., 1979. "Buoyancy effects on entrain-
ing turbulent boundary layers: a second order closure study,"
Turbulent Shear Flows - 1 (ed. F.J. Durst et a/.), Springer-
Verlag, Heidelberg, pp.295-306.
Index
311
312 Index