Articulo 5

PHYSICS OF FLUIDS 23, 011301 共2011兲
Simple models of turbulent flows*

Stephen B. Popea兲
Sibley School of Mechanical and Aerospace Engineering, Cornell University, 254 Upson Hall, Ithaca, New
York 14853, USA
共Received 25 October 2010; accepted 23 November 2010; published online 18 January 2011兲
Stochastic Lagrangian models provide a simple and direct way to model turbulent flows and the
processes that occur within them. This paper provides an introduction to this approach, aimed at the
nonspecialist, and providing some historical perspective. Basic models for the Lagrangian velocity
共i.e., the Langevin equation兲 and composition are described and applied to the simple but revealing
case of dispersion from a line source in grid turbulence. With simple extensions, these models are
applied to inhomogeneous turbulent reactive flows, where they form the core of probability density
function 共PDF兲 methods. The use of PDF methods is illustrated for the case of a lifted turbulent jet
flame. Lagrangian time series are now accessible both from experiments and from direct numerical
simulations, and this information is used to scrutinize and improve stochastic Lagrangian models. In
particular, we describe refinements to account for the observed strong Reynolds-number effects
including intermittency. It is emphasized that all models of turbulence are necessarily approximate
and incomplete, and that simple models are valuable in many applications in spite of their
limitations. © 2011 American Institute of Physics. 关doi:10.1063/1.3531744兴
I. INTRODUCTION then there would be a realistic hope for a successful statisti-

cal theory. But subsequent investigations have shown that the
For more than a century, researchers have been grappling small scales are significantly affected by the flow-dependent
with the challenges posed by turbulent flows. The research large scales. It appears now that this “ideal solution” is un-
has been motivated and sustained not only by scientific cu- realistic and unattainable. The fundamental difficulty is the
riosity and the intellectual challenges posed, but also by the nonlinear and nonlocal 共in space and scale兲 interactions of
ubiquity of turbulent flows in the natural world and in engi- the turbulent motions over the large range of scales present
neering devices, and by the importance of turbulence in en- in turbulent flows.
hancing the rates of transport and mixing, typically by many Instead of the unattainable ideal, a realistic goal, of great
orders of magnitude. practical value, is the development of tractable models and
There can be several different objectives in studies of simulation approaches for the diverse range of turbulent
turbulent flows including developing an understanding of flows encountered in engineering, oceanography, meteorol-
particular flows and the processes that occur within them, ogy, astrophysics, and elsewhere. There is not one “turbu-
developing methodologies to design engineering devices and lence problem” but a wide variety of problems of varying
to control the turbulent flows involved to achieve objectives difficulty, dependent on the geometry of the flow and the
such as reducing drag or promoting mixing, and—the focus additional processes involved 共e.g., chemical reactions and
of this paper—developing theories, models, and simulation multiple phases兲. Some of these turbulence problems have
techniques. The latter objective is of particular importance, been “solved” in the sense that current models or simulations
as the methodologies developed often can be used to achieve are adequately accurate for the application involved. And as
the other objectives. progress is made over time, a greater fraction of the turbulent
From the perspective of theory, modeling, and simula- problems will in this sense be solved.
tions, we may ask: How and when will the turbulence prob- As the title implies, this paper focuses on simple models,
lem be solved? The ideal solution would be to have a trac- and one of the morals 共due to the statistician G. E. P. Box兲 is
table quantitative theory, based soundly on the underlying that “All models are wrong, but some are useful.” Within the
physics, as described by the Navier–Stokes equations. In the turbulence research community, there is a tendency toward
middle of the last century, such a solution may have seemed more complex models. All models are incomplete and inac-
a realistic prospect. The theory proposed by Kolmogorov1 in curate to some degree, which naturally motivates the devel-
1941 postulated that the small scales of turbulence are statis- opment of increasingly more complete and more accurate
tically universal and simply parametrized by the viscosity models. But such advanced models are inevitably more com-
and the mean dissipation rate. If this picture were accurate, plicated and less tractable. On the other hand, researchers
and practitioners outside of turbulence research seek simple,
*
This paper is based on the Otto Laporte Lecture delivered by the author on tractable models for their turbulent flow, which may be but
November 22, 2009, following the award of the APS Fluid Dynamics Prize
one aspect of their larger complex problem. There is, there-
at the American Physical Society Division of Fluid Dynamics Annual
Meeting in Minneapolis. fore, an important role for simple models, in spite of their
a兲
Electronic mail: [email protected]. flaws and limitations, and an important role for the turbu-
1070-6631/2011/23共1兲/011301/20/$30.00 23, 011301-1 © 2011 American Institute of Physics
Author complimentary copy. Redistribution subject to AIP license or copyright, see http://phf.aip.org/phf/copyright.jsp
011301-2 Stephen B. Pope Phys. Fluids 23, 011301 共2011兲
lence research community is to develop and understand such M

models. z
y
The simple models considered here are called “stochas- σ ( x′)
tic Lagrangian models.” Within the turbulent flow of interest, U
we consider a fluid particle, i.e., a mathematical point mov- x′
ing with the local fluid velocity. As a function of time t, the <φ >
fluid particle has position X+共t兲, velocity U+共t兲, acceleration heated mean excess
A+共t兲, and thermochemical properties ␾+共t兲 such as tempera- wire temperature
ture and the mass fractions of chemical species. Such quan- x0 x′
tities are called Lagrangian trajectories or Lagrangian time x
series 共when sampled at discrete times in experiments or
FIG. 1. 共Color online兲 Sketch of Warhaft’s experiment on a line source in
simulations兲. The fluid-particle properties are, by definition,
grid turbulence showing the x⬘-y-z-coordinate system with origin at the
the fluid properties at the moving particle’s location. Hence, middle of the heated wire 共which is the line source of temperature兲 and the
with U共x , t兲 being the Eulerian velocity field, the fluid- 共Gaussian兲 mean excess temperature profile of width ␴共x⬘兲.
particle velocity is
U+共t兲 = U关X+共t兲,t兴. 共1兲 models and their use in PDF methods. The reader is encour-
Stochastic Lagrangian models are stochastic processes de- aged to use the on-line version in order to access the anima-
signed to mimic the behavior of Lagrangian trajectories in tions and color figures.
turbulent flows.
In the three major sections of this paper, we examine II. STOCHASTIC MODELING OF TURBULENT
different related aspects of stochastic Lagrangian models ap- DISPERSION
plied to three different flows. In Sec. II we introduce the
In this section, we consider one of the simplest and most
Langevin model for velocity, and apply it to study the dis-
basic turbulent flow problems—the dispersion of a conserved
persion of heat from a line source in grid turbulence. The
passive scalar from a line source in grid turbulence. Starting
ideas involved go back to Langevin2 and Taylor,3 and the
in the 1950s, there have been numerous experimental inves-
ability of the model to describe turbulent transport 共revealed
tigations of this flow:9–13 here we focus on that of Warhaft.11
through the mean temperature field兲 was demonstrated over
While this flow is relatively simple, it can nevertheless be
25 years ago.4 To study turbulent mixing 共revealed through
used to study two of the most fundamental turbulent pro-
the temperature variance兲 we introduce a simple stochastic
cesses: convective transport by the turbulent velocity field
Lagrangian model for a conserved passive scalar. Only re-
and turbulent mixing.
cently has a satisfactory model for the scalar been developed
and demonstrated for the line source.5,6 A. Dispersion from a line source: Experimental
The simplicity of the stochastic Lagrangian models, de- observations
veloped by reference to dispersion in homogeneous isotropic
turbulence, belies their potency. In Sec. III, we show that As sketched in Fig. 1, the experiment of Warhaft11 was
共with minor extensions兲 the same stochastic Lagrangian conducted in a wind tunnel in which a uniform flow passes
models lead to a complete turbulence closure for inhomoge- through a turbulence-generating grid of mesh spacing M. A
neous turbulent reactive flows. This closure takes the form of Cartesian coordinate system is introduced with origin at the
a modeled transport equation for the joint probability density center of the grid, with x being the direction of the mean
function 共PDF兲 of velocity and composition. This PDF flow, and y and z being normal to the mean flow. The mean
method, and its computational implementation as a particle/ velocity 具U典 is in the x-direction, while the fluctuating ve-
mesh Monte Carlo method, is demonstrated for the case of a locities in the x , y , z directions are denoted by u , v , w or al-
lifted turbulent jet flame. ternatively by u1 , u2 , u3. To an approximation, the turbulence
While the Langevin model is clearly very useful—as is isotropic, and the velocity variances decay with the down-
demonstrated by the results presented in Secs. II and III—a stream distance as power laws of the form
more detailed examination of its properties in Sec. IV shows
that it has fundamental shortcomings and inaccuracies—“All
models are wrong, but some are useful.” We describe more
具u2典
具U典 2 =A
x
M
冉冊 −m
, 共2兲
advanced models, which account for Reynolds-number where A and m = 1.32 are obtained from the measurements.
effects7 and intermittency.8 The development of such models At a distance x0 downstream of the grid, a fine heated wire is
has been made possible by the recent investigations of La- stretched across the center of the wind tunnel in the
grangian properties using direct numerical simulations z-direction 共at y = 0兲. The distance downstream of wire is
共DNS兲 and experiments with modern diagnostics. The paper denoted by x⬘ = x − x0, so that the coordinate system 共x⬘ , y , z兲
concludes with some thoughts on the current challenges and has its origin in the center of the wire. The wire is suffi-
future prospects for the modeling and simulation of turbulent ciently fine that it has a negligible effect on the turbulent
flows. velocity field. Because of the electrical heating of the wire,
This paper is intended primarily for nonspecialists, with the air passing very close to the wire is heated. Thus, to an
the aim to provide an introduction to stochastic Lagrangian approximation, the wire is a continuous line source of heat.
011301-3 Simple models of turbulent flows Phys. Fluids 23, 011301 共2011兲
To a good approximation, the 共suitably normalized兲 tempera- d␧ ␧2

ture excess—denoted by ␾共x⬘ , y , z , t兲—is a conserved pas- = − C␧2 . 共7兲
dt k
sive scalar.
Of interest in this flow is the behavior of the passive Equation 共6兲 is exact and shows that turbulent kinetic energy
scalar as it is swept downstream and affected by the turbu- k ⬅ 21 具uiui典 decays at a rate given by the viscous dissipation
lence. Away from the side walls, statistics such as the mean ␧. Equation 共7兲 is a model, which correctly yields the ob-
具␾共x⬘ , y兲典 depend solely on x⬘ and y. served power law decay of the form
The measurements show that the lateral profiles of the
mean are Gaussian, i.e.,
k共t兲
k共t0兲
=
t
冉冊 −m
, 共8兲
冉冊
t0
1 − y2
具␾共x⬘,y兲典 = exp , 共3兲 where the model constant C␧2 and the exponent m are related
␴共x⬘兲冑2␲ 2␴共x⬘兲2 by
where ␴共x⬘兲 is the characteristic width of the mean profile. 1+m
All theories and models predict the Gaussian shape: the chal- C␧2 = . 共9兲
m
lenge is to predict the downstream spreading, i.e., the depen-
dence of ␴共x⬘兲 on x⬘. Thus, the k-␧ model is successful for this simplest of turbu-
In the laboratory frame considered, the flow is statisti- lent flows in correctly describing the decay of the turbulence.
cally stationary and two-dimensional, with statistics varying According to the k-␧ model, the turbulent diffusivity is
only in the x⬘ and y directions, and spatial gradients of sta- k2
tistics are dominantly in the y-direction. It is also convenient ⌫T = C⌫ , 共10兲
to consider a frame moving with the mean velocity. Thus, we ␧
define where C⌫ is a constant, and indeed this relation is inevitable
once k and ␧ are taken to represent the turbulence.
x̂ = x⬘ − 具U典t, 共4兲
In the moving frame, the exact conservation equation for
and consider the scalar ␾共x̂ , y , z , t兲 in this frame. Now, to a the mean scalar is
good approximation, statistics depend only on y and t, and
the wire 共which is a continuous line source in the laboratory
frame兲 appears as an instantaneous plane source at y = 0, t
⳵ 具␾典 ⳵
⳵t
=
⳵y
⌫
⳵ 具␾典
⳵y
冉
− 具v␾典 . 冊共11兲
= 0. Thus, in this frame, the flow is decaying homogeneous The right-hand side represents the divergence of two fluxes:
isotropic turbulence 共with zero-mean velocity兲, and the phe- that due to molecular diffusion and that due to turbulence
nomenon under study is the evolution of ␾共x̂ , y , z , t兲 from the convection. The term 具v␾典 is called the 共turbulent兲 scalar
initial condition of a heated plane sheet 关i.e., ␾共x̂ , y , z , 0兲 flux, and it is the covariance of the scalar and the y-direction
= ␦共y兲兴. velocity v.
The turbulent diffusion model amounts to approximating
B. Turbulent diffusion model the scalar flux by
The simplest models for turbulent flows are those based ⳵ 具␾典
on the concepts of turbulent viscosity and turbulent diffusiv- 具v␾典 ⬇ − ⌫T , 共12兲
⳵y
ity. The basic notion is that the mean fields in turbulent flows
behave similarly to fields in laminar flows, but with en- so that Eq. 共11兲 becomes
冉冊
hanced “effective” viscosity and diffusivity. Specifically, the
molecular diffusivity ⌫ is enhanced by a turbulent contribu- ⳵ 具␾典 ⳵ ⳵ 具␾典
= ⌫eff . 共13兲
tion ⌫T to yield the effective diffusivity ⳵t ⳵y ⳵y
⌫eff = ⌫ + ⌫T . 共5兲 For the homogeneous turbulence considered, ⌫T is indepen-

dent of y, and varies with t as a known power law. Hence,
The turbulent diffusivity ⌫T共x , t兲 represents the effects of the Eq. 共13兲 is readily solved analytically to show that the k-␧
turbulent motions, and hence can depend on the local char- model correctly yields a Gaussian profile for 具␾典, and that
acteristics of the turbulent velocity field. On the other hand, the predicted width ␴共t兲 is given by
since the velocity field is 共by definition兲 unaffected by pas-
sive scalars, in a consistent model, ⌫T is independent of any
passive scalar field that may be present.
␴共t兲2 = 冕0
t
2⌫eff共t⬘兲dt⬘ . 共14兲
The most widely used turbulent viscosity model is the
k-␧ model,14 and it is instructive to apply it to the line source This prediction for ␴共t兲 is compared to the experimental data
problem. In the moving frame, the flow is decaying homo- in Fig. 2 共shown in the laboratory frame兲. As may be seen,
geneous turbulence, for which case the k-␧ model reduces to for small times 共i.e., small x⬘ / x0兲 the model displays very
the pair of ordinary differential equations, significant errors and appears to be qualitatively incorrect.
This is confirmed by the theory presented in Sec. II C, which
dk also shows that the agreement between the model and the
= − ␧, 共6兲
dt data for large times 共i.e., large x⬘ / x0兲 is not fortuitous, but
For given t, X+共t兲 is a random variable 共because of the

randomness of the turbulent velocity field兲, and we can con-
sider its mean 具X+共t兲典, its variance, and its PDF. The mean is
zero, since the problem is statistically symmetric about y
= 0. As mentioned, according to this theory, the mean
具␾共y , t兲典 is proportional to the probability density of the
event X+共t兲 = y, since 具␾典 is proportional to the probability
that the fluid at y , t originates from the source. In accord with
experimental observations, all theories and models lead to a
Gaussian PDF of X+共t兲, and hence to a Gaussian profile of
具␾共y , t兲典. The characteristic width ␴共t兲 of this profile is the
standard deviation at X+共t兲,
␴共t兲2 = 具X+共t兲2典. 共18兲
The equation for the fluid-particle motion dX+ / dt = u+ can be
FIG. 2. 共Color online兲 Width ␴ of the thermal wake 共normalized by L integrated to yield
= k3/2 / ␧ at the source兲 against distance from the source x⬘ normalized by the
distance x0 of the source from the grid: symbols, experimental data 共Refs. 11
and 12兲; line, from the k-␧ model. X+共t兲 = 冕 t
0
u+共t⬘兲dt⬘ , 共19兲
and hence Eq. 共18兲 can be reexpressed as

instead the turbulent diffusion model is valid in this region.
C. Diffusion by continuous movements

␴共t兲2 = 冕冕 t
0 0
t
具u+共t⬘兲u+共t⬙兲典dt⬘dt⬙ . 共20兲
Fifty years before the k-␧ turbulence model, Taylor3 This is Taylor’s principal result, showing that the dispersion
tackled the problem of turbulent dispersion in his famous ␴共t兲2 is known in terms of the two-time Lagrangian velocity
1921 paper “Diffusion by continuous movements.” For the autocovariance.
line source problem, the equation governing the mean 具␾典 Further deductions are most simply made for the case of
关Eq. 共11兲兴 is derived from the instantaneous equation, statistically stationary turbulence 共as opposed to decaying
grid turbulence兲. In the stationary case, the autocovariance
D␾ can be reexpressed as
= ⌫ⵜ2␾ . 共15兲
Dt
具u+共t⬘兲u+共t⬙兲典 = 具u2典␳共兩t⬘ − t⬙兩兲, 共21兲
Taylor argued that at high Reynolds number, the molecular
where 具u2典 is the velocity variance and ␳共s兲 is the Lagrangian
flux in Eq. 共11兲 is negligible compared to the turbulent flux.
With the molecular term neglected, the equation for the mean velocity autocorrelation function. DNS 共Ref. 15兲 shows that
then rises from the instantaneous equation, ␳共s兲 is well approximated by the exponential
D␾
Dt
= 0, 共16兲 ␳共s兲 = exp 冉冊
− 兩s兩
TL
, 共22兲
i.e., the scalar ␾ is conserved following the fluid. This ob- where TL is the Lagrangian velocity integral time scale
servation transforms the problem into that of describing the
motion of fluid particles. For the line source problem in the
moving frame, the mean 具␾共y , t兲典 is proportional to the PDF
TL ⬅ 冕 0
⬁
␳共s兲ds. 共23兲
共at y , t兲 of the position of fluid particles originating from the 关The approximation equation 共22兲 is scrutinized in Sec. IV.兴
source 共y = 0兲 at time t = 0. Using Eq. 共21兲 and a nontrivial manipulation 共see, e.g.,
We denote by X+共t , Y兲 the position at time t of the fluid Ref. 16兲, Eq. 共20兲 can be reexpressed as
冕
particle which is at position Y at time 0. By definition, the t
fluid particle moves with its own velocity 关denoted by ␴共t兲2 = 2具u2典共t − s兲␳共s兲ds. 共24兲
u+共t , Y兲兴, which is the local Eulerian fluid velocity, 0
⳵ X+共t,Y兲 Several important deductions 共due to Taylor兲 can be made

= u+共t,Y兲 ⬅ U关X+共t,Y兲,t兴. 共17兲
⳵t from Eq. 共24兲 without invoking Eq. 共22兲:
Focusing on the line source problem, we are interested in 共1兲 For very small time 共s / TL Ⰶ 1兲, ␳共s兲 is very close to
fluid particles which originate from y = 0 at t = 0, i.e., those unity, so that Eq. 共24兲 yields
with Y 2 = 0, and, furthermore, we are interested only in their ␴共t兲 ⬇ u⬘t, for t/TL Ⰶ 1, 共25兲
motion in the y-direction. Consequently, we simplify the no-
tation to consider X+共t兲 and u+共t兲 to be the y-components of where u⬘ ⬅ 具u 典 is the rms velocity. This corresponds
2 1/2
position and velocity of such particles. to straight-line, ballistic motion.
共2兲 For very large time 共t / TL Ⰷ 1兲, Eq. 共24兲 yields
␴共t兲2 ⬇ 2具u2典tTL, for t/TL Ⰷ 1. 共26兲

共3兲 The evolution of ␴共t兲2 can be represented in terms of a
time-dependent diffusivity,
⌫ˆ 共t兲 ⬅
d 1
冉冊冕
dt 2
␴共t兲2 = 具u2典
0
t
␳共s兲ds
⬇ 具u2典t, for t/TL Ⰶ 1

⬇ 具u 典TL,
2
for t/TL Ⰷ 1. 共27兲
The last result establishes the validity of the turbulent diffu-

sion model at large times with ⌫T = 具u2典TL. But for smaller
times, ⌫ˆ 共t兲 depends on the time t since the release of the
source, contrary to the notion that ⌫T depends solely on
properties of the turbulent velocity field.
We could proceed to obtain a prediction for ␴共t兲 for the
line source in grid turbulence by substituting into Eq. 共24兲 a
model for the Lagrangian velocity autocovariance in decay-
ing turbulence. But, instead, we proceed by the more power-
ful and general route of considering a stochastic model for
u+共t兲.
D. The Langevin model FIG. 3. 共Color online兲 Samples of fluid-particle paths X+共t兲 for short times
共t / TL ⱕ 1, top兲 and for long times 共t / TL ⱕ 100, bottom兲 obtained from the
In an appendix to his paper,3 Taylor proposed a stochas- Langevin model. The dashed lines show ⫾␴共t兲, the standard deviation of
tic model for the position X+共t兲 of a fluid particle. According X+共t兲. Normalization is by the Lagrangian integral time scale TL and the rms
velocity u⬘.
to this model, over successive small time intervals ⌬t, the
position increments 关X+共t + ⌬t兲 − X+共t兲兴 and 关X+共t兲 − X+共t
− ⌬t兲兴 are highly correlated. In fact, this model is identical to
the stochastic model for velocity u+共t兲 proposed a decade Realizations 共or sample paths兲 of u+共t兲 can be generated
earlier by Langevin2 to model the velocity of a particle un- from Eq. 共28兲 and then integrated to yield corresponding
dergoing Brownian motion. sample paths of X+共t兲. Figure 3 shows such sample paths at
For the case of statistically stationary homogeneous tur- short times 共top兲 and long times 共bottom兲. The ballistic and
bulence, the Langevin model can be written as the stochastic diffusive regimes identified by Taylor are clearly evident.
differential equation 共SDE兲, In order to apply the Langevin equation to the case of
冉冊
decaying grid turbulence, we rewrite it as
dt 2u⬘2 1/2
du+共t兲 = − u+共t兲 + dW共t兲, 共28兲 du+ = − ␻uu+dt + bdW, 共30兲
TL TL
and we seek to relate the relaxation rate ␻u and the diffusion
where W共t兲 is a Wiener process. The reader unfamiliar with
coefficient b to properties of the turbulence, which is char-
SDEs can understand this equation through its finite-
acterized by the kinetic energy k and the dissipation rate ␧.
difference analog,
Two pieces of information are required to determine these
u+共t + ⌬t兲 − u+共t兲 = − u+共t兲

⌬t
TL
+
TL
冉
2u⬘2⌬t
冊 1/2
␰, 共29兲
two coefficients. The first is that the variance 具u+2典 equals the
variance of the Eulerian velocity, which 共assuming isotropy兲
equals 32 k. Thus, the variance decays as
where ␰ is a zero-mean, unit-variance Gaussian random vari-
able 共independent for each time step兲. The stochastic process
generated by the Langevin equation is called the Ornstein–
d +2
dt
具u 典 =
d 2
dt 3
冉冊 2
k = − ␧.
3
共31兲
Uhlenbeck 共OU兲 process, which is readily analyzed 共see, The second piece of information comes from Kolmogorov1
e.g., Ref. 16兲. The OU process corresponding to Eq. 共28兲 is theory. This pertains to the second-order Lagrangian velocity
fully characterized by the fact that it is a continuous statisti- structure function defined by
cally stationary Gaussian process with mean 具u+共t兲典 = 0, vari-
D共s兲 ⬅ 具关u+共t + s兲 − u+共t兲兴2典, 共32兲
ance 具u+共t兲2典 = u⬘2, and exponential autocorrelation ␳共s兲
= exp共−兩s兩 / TL兲. The mean and variance are 关by construction which is simply the variance of the velocity increment over a
of Eq. 共28兲兴 consistent with the given properties of the tur- time interval s. According to Kolmogorov, at high Reynolds
bulence, and, as mentioned above, the exponential autocor- number, and for s in the inertial range 共i.e., ␶␩ Ⰶ s Ⰶ TL,
relation is supported by DNS data. where ␶␩ is the Kolmogorov time scale兲, D共s兲 is uniquely
yields ␴ ⬇ 冑2⌫t; at intermediate times there is ballistic tur-

bulent motion with ␴ ⬇ u⬘t; and at long times there is turbu-
lent diffusion with ␴ ⬃ t0.34—the power being less than 21
because of the decay of the turbulence.
Comparing the poor performance of the turbulent diffu-
sion model 共Fig. 2兲 with the excellent performance of simple
Langevin model 共Fig. 4兲, one is reminded of Einstein’s ex-
hortation to make things as simple as possible, but no sim-
pler. Clearly, for this simple flow, the turbulent diffusion
model is too simple.
The Langevin model, in particular its short-time behav-
ior and the specification of C0, is further examined in Sec. IV
E. Scalar variance
The success of stochastic Lagrangian models based on
FIG. 4. 共Color online兲 Width ␴ of the thermal wake 共normalized by L
= k3/2 / ␧ at the source兲 against distance from the source x⬘ normalized by the the Langevin equation to describe the mean temperature field
distance x0 of the source from the grid: symbols, experimental data 共Refs. 11 was well established 25 years ago. Only recently, however,
and 12兲; dashed line, from the Langevin model for fluid particles; solid line, has the same success been achieved for the temperature vari-
from the Langevin model for Brownian particles. ance.
For the temperature variance, which reveals the extent of
determined by s and ␧. Dimensional analysis then leads to turbulent mixing, there are two distinct modeling ap-
proaches. The first is based on pair dispersion in which one
D共s兲 = C0␧s, for ␶␩ Ⰶ s Ⰶ TL , 共33兲 models the motion of a pair of Brownian particles: see Ref.
where C0 is a universal Kolmogorov constant. 共Since this 18 for a recent review. We follow the second approach in
relation is linear in ␧, it also holds according to Kolmogor- which the stochastic Lagrangian model is extended by add-
ov’s refined similarity hypotheses17 in the presence of inter- ing composition ␾+共t兲 as a particle property. Furthermore, we
mittency.兲 consider particles uniformly distributed throughout the flow
The Langevin equation is consistent with Eq. 共33兲 pro- domain, rather than just those originally from the location of
vided that b is taken to be 共C0␧兲1/2. Then Eq. 共31兲 requires ␻u the source.
to be 共 21 + 43 C0兲␧ / k. Thus, for decaying turbulence, the Lange- An important aspect of this approach is the estimation of
vin equation can be written as mean fields from particle properties. For example, the mean
scalar field 共in a general flow兲 is obtained as
du+ = − 冉 1 3 ␧
冊
+ C0 u+dt + 共C0␧兲1/2dW.
2 4 k
共34兲具␾共x,t兲典 = 具␾+共t兲兩X+共t兲 = x典, 共36兲
i.e., the mean Eulerian field 具␾共x , t兲典 is the expectation of the
共The same equation applies to the stationary case, but with
particle composition ␾+共t兲, conditional upon the particle be-
the 21 omitted from the drift coefficient.兲
ing located at x at time t. 共In Sec. III C, we explain how such
The dispersion ␴共t兲2 = 具X+共t兲2典 given by the Langevin
means are estimated in practice, in a numerical implementa-
equation 共34兲 can be determined analytically:4,16 it depends
tion of the method.兲
solely on the decay exponent m 共known from the experi-
For the case considered of a single conserved passive
ment兲 and the constant C0. Figure 4 compares the experimen-
scalar, the simplest Lagrangian model for the scalar is the
tal data for ␴共t兲 with the Langevin-model prediction for C0
interaction by exchange with the mean 共IEM兲 model,19,20
= 2.1—the value determined by reference to these data. As
may be seen, there is excellent agreement, except for very d␾+共t兲
= − ␻m关␾+共t兲 − 具␾+共t兲兩X+共t兲典兴, 共37兲
small times. This discrepancy is due to the complete neglect dt
of molecular diffusion, which is not justified in this
moderate-Reynolds-number flow 共R␭ ⬇ 50兲. This deficiency where the composition relaxation rate ␻m is taken to be
is readily remedied by redefining the particles considered to 1 ␧
be Brownian particles, moving both with the local fluid ve- ␻m = C␾ , 共38兲
2 k
locity and by molecular diffusion. Thus, X+共t兲 evolves by the
SDE where C␾ is a model constant. Thus, according to this IEM
model, the particle composition ␾+共t兲 relaxes to the local
dX+共t兲 = u+共t兲dt + 共2⌫兲1/2dW⬘ , 共35兲
mean 具␾+共t兲兩 X共t兲典 at the rate ␻m.
where W⬘共t兲 is a Wiener process 关independent of W共t兲 in the Despite its apparent simplicity, the line source in grid
Langevin equation兴. With this modification, as may be seen turbulence contains features that are challenging to over-
from Fig. 4, the Langevin model is in excellent agreement simple models, and the IEM model is found to be deficient in
with the data. predicting the variance. Close to the heated wire, the instan-
As indicated in Fig. 4, the theory identifies three re- taneous structure of the temperature field is that of an un-
gimes: at very early times molecular diffusion dominates and steady laminar thermal wake that is flapped around by the
1.5

1
φ
φ
y= 0
0.5
0 −3 1
10 10
x /xo
FIG. 5. 共Color online兲 Axial profile of the ratio of the rms temperature
fluctuation to the mean temperature excess on the centerline downstream of FIG. 6. 共Color online兲 Radial profiles of the rms temperature fluctuation
the line source: symbols, experimental data 共Ref. 11兲; line, calculations us- normalized by its centerline value at axial locations 共from top to bottom兲
ing the IECM model 共Ref. 6兲. x⬘ / M = 0.36, 0.62, 1.0, 100: symbols, experimental data 共Ref. 11兲; line, cal-
culations using the IECM model 共Ref. 6兲.
turbulent motions. As a consequence, in this early region, the sive scalars, denoted by ␾1 and ␾2. As functions of x⬘ and y,
direct effects of molecular diffusion to broaden the thermal in addition to the means 具␾1典 and 具␾2典, and variances 具␾1⬘2典
wake are important, and the temperature field is highly cor- and 具␾2⬘2典, we can consider the covariance 具␾1⬘␾2⬘典 and the
related with the velocity 共in the y-direction兲. correlation coefficient
A refined model6 which accounts for these effects is
具␾1⬘␾2⬘典
given by ␳12 ⬅ . 共40兲
关具␾1⬘2典具␾2⬘2典兴1/2
d␾ 共t兲
+
= − ␻m关␾+共t兲 − 具␾+共t兲兩X+共t兲,u+共t兲典兴 The covariance and the correlation coefficient are important
dt
quantities, as they reveal the rate of mixing of different sca-
⳵2具␾+共t兲兩X+共t兲,u+共t兲典 lars, which is clearly relevant to reactive flows. In the experi-
+⌫ . 共39兲
⳵ y2 ment, Warhaft deduced the covariance using a superposition
principle based on three separate measurements of the tem-
The term in ␻m is similar to the IEM model, but now the
perature variance: one with the wire at y = 21 d0 heated, one
relaxation is to the mean conditioned on the particle velocity
with the wire at y = − 21 d0 heated, and one with both wires
共in addition to its position兲. This part is called the interaction
heated.
by exchange with the conditional mean 共IECM兲 model.21–23
Figure 7 shows the axial evolution of the correlation
The term in ⌫ implements the direct effects of molecular
coefficient ␳12 on the centerline 共y = 0 , z = 0兲 for different
diffusion in the evolution of the composition, instead of
wire separations d0. As may be seen, this evolution is highly
through a random walk in the position equation, Eq. 共35兲.
nontrivial and strongly dependent on the value of d0. The
Analysis shows that while the use of Brownian particles, Eq.
stochastic Lagrangian model is completely successful in rep-
共35兲, is valid for studying the mean 具␾典, it introduces a spu-
resenting these observations.
rious source of composition variance.
Figures 5 and 6 show axial and radial profiles of the
III. PROBABILITY DENSITY FUNCTION METHODS
normalized rms temperature fluctuations ␾⬘ for Warhaft’s
line source experiment. As may be seen, there is excellent With straightforward modifications and extensions, the
agreement between the model calculations and the measure- simple stochastic Lagrangian models for position, X+共t兲, ve-
ments. The calculations are obtained from the stochastic La- locity, u+共t兲, and composition, ␾+共t兲, can be used to produce
grangian model for fluid particles 共dX+ / dt = u+兲, in which the a closed modeled equation for inhomogeneous turbulent
velocity u+共t兲 evolves by the Langevin equation 关Eq. 共34兲兴, flows, including reacting flows. Specifically, the models pro-
and the composition ␾+共t兲 evolves by the IECM model 关Eq. vide a closure for the transport equation for the one-point,
共39兲兴. one-time Eulerian joint PDF of velocity and composition. In
This stochastic Lagrangian model has been applied to practice, this modeled PDF equation is solved by a particle/
the ingenious experiment by Warhaft11 on pairs of line mesh Monte Carlo method, the core of which is a large num-
sources. Instead of a single heated wire at x⬘ = 0, y = 0, there ber of particles, each evolving according to the stochastic
are two wires, separated by a distance d0, located at x⬘ = 0, Lagrangian models.
y = ⫾ 21 d0. Conceptually, and in the modeling, we can con- In this section, this PDF method is described, along with
sider the two wires to be sources of different conserved pas- its associated particle-mesh method, and its application to a
Navier–Stokes equations; but here we focus on the mean

composition equation obtained from Eq. 共41兲, which is
冉 ⳵
⳵t
冊
+ 具U典 · ⵜ 具␾典 = ⵜ · 共⌫ ⵜ 具␾典 − 具u␾典兲 + 具S共␾兲典, 共42兲
where the Reynolds decomposition has been used to express

the velocity as
U = 具U典 + u. 共43兲
In Eq. 共42兲, since 具U典 and 具␾典 are knowns, the left-hand side
and the term in ⌫ are in “closed form,” i.e., they are known
in terms of the knowns. The remaining quantities—the scalar
flux 具u␾典 and the mean source term 具S共␾兲典—are unknowns
for which models or closure approximations are required.
The simplest model for the scalar flux is a turbulent diffusion
model 共as discussed above兲 which is
具u␾典 ⬇ − ⌫T ⵜ 具␾典, 共44兲
FIG. 7. 共Color online兲 Axial profiles on the centerline of the correlation
coefficient between the scalars from a pair of line sources: symbols, experi-
mental data 共Ref. 11兲; lines, calculations using the IECM model 共Ref. 6兲.
where ⌫T is the turbulent diffusivity 共which is obtained
The separation between the sources is 共from top to bottom downstream兲 through an additional model兲. At this level of closure, the
d0 = 1.2, 8 , 14, 25, 35 mm. only information about the composition is its mean 具␾典.
Hence, for the mean source term, the only available closure
approximation is
turbulent lifted jet flame. First, we contrast the Eulerian sta- 具S共␾兲典 ⬇ S共具␾典兲. 共45兲
tistical approach and the stochastic Lagrangian approach for
obtaining turbulence closures. While this relation is exact if S共␾兲 is a linear function, for the
highly nonlinear functions encountered in combustion, for
A. Eulerian statistical and stochastic Lagrangian example, this approximation can be in error by orders of
approaches magnitude. With the models 关Eqs. 共44兲 and 共45兲兴 substituted
1. The Eulerian statistical approach into the exact mean equation 关Eq. 共42兲兴, we obtain the final
result—the closed, modeled conservation equation,
冉冊
The Eulerian statistical approach, which goes back to
Reynolds’ paper,24 consists of the following stages: ⳵
+ 具U典 · ⵜ 具␾典 = ⵜ · 共关⌫ + ⌫T兴 ⵜ 具␾典兲 + S共具␾典兲. 共46兲
⳵t
共1兲 The starting point is the set of exact, instantaneous con-
servation equations for the flow variables considered. This Eulerian statistical approach, described here for the
共2兲 We then derive the exact 共but unclosed兲 equations gov- means, can be applied to means, variances, and covariances
erning a chosen set of statistics 共e.g., means, variances兲, 共to yield a “second-moment” closure兲 or to higher statistics.
which are called “knowns.” In particular, it can be applied to the Eulerian one-point one-
共3兲 We provide models for other statistics arising in these time joint PDF of velocity and composition. This quantity,
equations 共referred to as “unknowns”兲 in terms of the denoted f共V , ␺ ; x , t兲, is the joint probability density of the
knowns. event 兵U共x , t兲 = V , ␾共x , t兲 = ␺其, where V and ␺ are the corre-
共4兲 The result is a set of closed, modeled equations for the sponding sample-space variables. In this case, the unknowns
knowns. to be modeled are conditional statistics, such as the condi-
tional diffusion
As a simple example, we consider a constant-property flow
involving a single reactive scalar ␾共x , t兲, whose conservation D共V, ␺ ;x,t兲 = 具⌫ⵜ2␾兩U = V, ␾ = ␺典. 共47兲
equation is
D␾
Dt
=
⳵
⳵t
冉冊
+ U · ⵜ ␾ = ⌫ⵜ2␾ + S共␾兲, 共41兲
2. The stochastic Lagrangian approach
where U共x , t兲 is the Eulerian velocity field, ⌫ is the molecu- In the stochastic Lagrangian approach, we select a set of
lar diffusivity, and S is the chemical source term, which is a fluid properties to be considered. For definiteness, we con-
known, highly nonlinear function of the local value of ␾. sider the velocity U and composition ␾. As in the Eulerian
In the simplest “mean flow” closure, the set of statistics statistical approach, the starting point is the instantaneous
considered consists of the mean velocity 具U典, the mean pres- conservation equations for the quantities considered. In the
sure 具p典, and the mean scalar 具␾典. Conservation equations for case considered, these are the Navier–Stokes equations for
具U典 and 具p典—the Reynolds equations—are derived from the U共x , t兲,
DU
Dt
1
= − ⵜ p + ␯ⵜ2U,
␳
ⵜ ·U=0 共48兲 dU+ + 冋 1
␳
ⵜ 具p典册 +
dt = − ␻u关U+ − 具U典+兴dt + 共C0␧兲1/2dW,
共where ␳ and ␯ are the constant density and kinematic vis- 共53兲
cosity and p共x , t兲 is the pressure兲, and Eq. 共41兲 for ␾共x , t兲. where again the relaxation rate ␻u and the mean dissipation
The end result of the approach is a closed modeled conser- rate ␧ are specified as another aspect of the modeling. The
vation equation for the one-point, one-time Eulerian joint determination of ␻n, ␻m, and ␧ is discussed in Sec. III B.
PDF of the properties considered; here f共V , ␺ ; x , t兲, the Note that for the inhomogeneous flows considered, the ap-
velocity-composition joint PDF. propriate drift term in the Langevin equation causes U+ to
We consider the joint PDF f to be “known” and hence relax to the local mean velocity. Also, if the effects of mean
any quantity that can be deduced from f is also a known. diffusion and mean viscous stresses are important, then these
Consequently, moments such as 具U典 , 具␾典 , 具uiu j典 , 具ui␾典 are terms 共which are known兲 can be included.
known, as is the mean pressure field 具p典, since it can be The final stage is to derive from the model particle equa-
obtained from a Poisson equation of known source. tions a modeled conservation equation for the joint PDF
The second step in the stochastic Lagrangian approach is f共V , ␺ ; x , t兲. By design, this equation is closed.
to write down the fluid-particle evolution equations, which
follow directly from the Eulerian conservation equations. 3. Comparison of approaches
Thus, for the case considered, the fluid-particle properties
A significant difference between the Eulerian and La-
X+共t兲 , U+共t兲 , ␾+共t兲 evolve by
grangian approaches is in the form of the quantities to be
dX+ modeled. In the Eulerian approach, it is statistics which re-
− U+共t兲 = 0, 共49兲 quire modeling, and their complexity increases with the level
dt
of closure considered. In the stochastic Lagrangian approach,
冋册冋册
+ + to be modeled are the instantaneous, time-dependent physi-
dU+ 1 1
+ ⵜ 具p典 = − ⵜ p ⬘ + ␯ ⵜ 2U , 共50兲 cal processes affecting the fluid-particle properties. Thus, the
dt ␳ ␳ modeling is very direct.
A second-moment closure and a PDF model can be com-
d␾+ pared in terms of the processes requiring modeling. To this
− S共␾+共t兲兲 = 共⌫ⵜ2␾兲+ , 共51兲
dt end, we rewrite the governing equations as
where 关兴+ indicates that the quantity is evaluated at 共x , t兲

⳵U 1 1
u兲 · ⵜU + ⵜ 具p典 = ␯ⵜ2U − ⵜ p⬘ ,
+ 共具U典 +□ 共54兲
= 关X+共t兲 , t兴. These equations are written with known terms on ⳵t ␳ ␳
the left-hand side and unknown terms on the right. As men-
tioned, the mean pressure 具p典 is known, whereas the pressure
⳵␾
function p⬘ is unknown. u兲 · ⵜ␾ − S共␾兲 = ⌫ⵜ2␾ .
+ 共具U典 +□
In addition to statistics deduced from the joint PDF, the ⳵t 共55兲
knowns include the fluid-particle properties. Hence, impor- In both approaches, the processes represented by the terms
tantly, the source term S关␾+共t兲兴 is known directly in terms of on the right-hand side require modeling 共although their
the particle composition. means are in closed form in both approaches兲. In the PDF
In the statistical Eulerian approach, the unknowns to be approach, all the processes represented by the terms on the
modeled are statistics which, in statistically stationary flows, left-hand side are in closed form and do not require model-
are independent of time. In contrast, in the stochastic La- ing. In contrast, in a second-moment closure, the terms in
grangian approach, the unknowns 关the right-hand sides of boxes—representing convection by the turbulent velocity
Eqs. 共50兲 and 共51兲兴 are time-dependent random processes. and the chemical source term—do need to be modeled.
The right-hand side of Eq. 共50兲 is the random force due to
the fluctuating pressure gradient and viscous stresses, B. Joint PDF of velocity, composition, and turbulent
whereas the right-hand side of Eq. 共51兲 represents the rate- frequency
of-change of composition due to molecular diffusion.
The third stage in the stochastic Lagrangian approach is For application to turbulent reactive flows, the stochastic
to provide models— either stochastic or deterministic—for Lagrangian models and PDF methods described above are
the unknown random processes. Thus, using the IEM model, generalized and extended in three ways.25,26 First, the single
Eq. 共51兲 is modeled as composition variable ␾+共t兲 is replaced by a set of n␾ compo-
sition variables ␾+共t兲, which completely describe the thermo-
d␾+ chemical composition of the fluid. For low-Mach-number
− S关␾+共t兲兴 = − ␻m关␾+共t兲 − 具␾典+兴, 共52兲 flows involving ns chemical species, typically ␾+共t兲 is taken
dt
to be the ns species mass fractions and the enthalpy 共or tem-
where ␻m is the turbulent mixing rate 共specified by another perature兲: as far as the thermochemistry is concerned, the
aspect of the modeling兲 and 具␾典+ is the mean composition at pressure is adequately represented by its mean at a reference
the particle location. Similarly, using the Langevin equation location 共usually atmospheric pressure兲. For high-speed
to model the random force, Eq. 共50兲 is modeled as flows, the pressure needs to be included as an additional
variable.27 Second, the “turbulent frequency” ␻+共t兲 is added

as a 共modeled兲 property of a fluid particle.16,28 Roughly,
␻+共t兲 can be thought of as the instantaneous dissipation rate
␧+共t兲 divided by the local turbulent kinetic energy k, so that
具␻+典 corresponds to the quantity ␻ = ␧ / k in the k-␻ model,
and, similarly, k具␻+典 corresponds to ␧ in the k-␧ model. This
turbulent frequency is added to the model in order to deter-
mine the relaxation rates ␻u and ␻m 共appearing in the Lange-
vin and IEM models兲 and the mean dissipation rate ␧ 关ap-
pearing in the diffusion term of the Langevin model, Eq.
共53兲兴. A stochastic process is used to model the evolution of
␻+共t兲.16,28
Third, the assumption of constant density is removed. In
its place there is an equation of state giving the density ␳+共t兲
of the fluid particle as a function of its composition ␾+共t兲
共and of the pressure兲.
The result is a set of stochastic Lagrangian models for
the fluid-particle properties: X+共t兲, U+共t兲, ␾+共t兲, and ␻+共t兲.
From these one can deduce a closed modeled conservation
equation for the velocity-composition-frequency joint PDF,
f̃共V , ␺ , ␩ ; x , t兲, where V , ␺ and ␩ are the sample-space vari-
ables corresponding to U+ , ␾+ and ␻+.
Thus, the simple stochastic models for fluid-particle
properties lead to a general and powerful result: a single
closed modeled equation for the joint PDF f̃, describing in- FIG. 8. 共Color online兲 Sketch of a highly simplified, plane, lifted flame
homogeneous turbulent reactive flows. formed from a cold fuel jet issuing into hot quiescent air. Also shown are the
domain and the mesh used in the particle/mesh method.
C. Particle/mesh method
In order for the modeled PDF equation to be useful, we
need methods for its solution. Only for the simplest of cases X共n兲共n兲共n兲共n兲共n兲
1 共t兲, X2 共t兲; velocity U1 共t兲, U2 共t兲, U3 共t兲; frequency
共n兲共n兲共n兲
in homogeneous turbulence are analytical solutions feasible; ␻ 共t兲; and composition ␾1 共t兲 , ␾2 共t兲. In this highly simpli-
and, inevitably, for inhomogeneous flows, numerical meth- fied illustration, the two composition variables are taken to
ods are required. The PDF equations present an insurmount- be mixture fraction ␰ and the product mass fraction Y p, i.e.,
共n兲
able challenge to conventional grid-based methods because 兵␾共n兲共n兲共n兲
1 共t兲 , ␾2 共t兲其 = 兵␰ 共t兲 , Y p 共t兲其. The particle properties
of the high dimensionality of the PDF. For example, for a evolve in time according to stochastic Lagrangian models.
statistically stationary and two-dimensional flow involving The particles move with their own velocity, which evolves
just three chemical species, the joint PDF by the Langevin equation. The compositions evolve by the
f̃共V1 , V2 , V3 , ␺1 , ␺2 , ␺3 , ␺4 , ␩ ; x1 , x2兲 is ten-dimensional. IEM model, and the mass fraction of product increases due
Even with an extremely coarse grid with ten nodes in each to a chemical source term.
direction, 1010 grid nodes are required, which is challenging, Figure 9 shows the motion of the particles and 共via color
to say the least. coding兲 their mixture fraction 共left兲 and product mass frac-
The most widely used method to solve the PDF equation 共right兲. As may be seen, particles emanate from the jet
tions is a particle/mesh method,25,29 outlined below, which is exit, and particles in the coflow are entrained into the jet.
based directly on the stochastic Lagrangian models. Other New particles enter at the side boundaries of the domain, and
methods that have been developed 共especially for the com- particles exit at the downstream boundary. The 共expected兲
position PDF兲 include grid-based particle methods,30 sto- number density of the particles is uniform: this is a conse-
chastic field methods,31,32 and approximate methods based quence of the uniform number density of the initial particle
on the direct quadrature method of moments 共DQMOM兲.33 distribution and of the inflowing particles, and of the mean
We illustrate the operation of the particle/mesh method continuity equation 共ⵜ · 具U典 = 0, for this constant-density
for the highly simplified case 共sketched in Fig. 8兲 of a statis- flow兲.
tically stationary and two-dimensional lifted turbulent flame. The particles entering the domain from the jet have
There is a central plane jet of cold fuel and hot quiescent air ␰ = 1, Y 共n兲
共n兲共n兲
p = 0, while those in the coflow have ␰ = 0, Y p
共n兲
on each side. As shown in the sketch, a rectangular solution = 0. The intermediate values of ␰共n兲共i.e., 0 ⬍ ␰共n兲 ⬍ 1兲 arise
domain is selected and covered by a uniform Cartesian mesh. solely due to the IEM mixing model 共because mixture frac-
The fundamental representation of the flow is in terms of tion is a conserved scalar and so has no chemical source
a large number N of computational particles, which model term兲. The chemical source of product S共␰ , Y p兲 is appropri-
fluid particles 共on different realizations of the flow兲. At time ately specified to be zero for ␰ = 0 共pure air兲 and for ␰ = 1
t, the nth particle has the following properties: position 共pure fuel兲. Hence, mixing to intermediate values of ␰ pre-
FIG. 9. 共Color online兲 Scatter plots of particles in the x-y solution domain FIG. 10. 共Color online兲 Cell means of product mass fraction obtained from
for the lifted jet flame. The particles are grayscale-coded 共color-coded on- the particle/mesh method applied to the simple lifted flame. Left: contour
line兲 by mixture fraction 共left兲 and by product mass fraction 共right兲共en- plot of instantaneous cell means. Right: lateral profile at y ⬇ 3.2 of the in-
hanced online兲. 关URL: http://dx.doi.org/10.1063/1.3531744.1兴 stantaneous 关circles 共red兲兴 and time-averaged 关squares 共blue兲兴 cell mean 共en-
hanced online兲.关URL: http://dx.doi.org/10.1063/1.3531744.2兴
D. Application to turbulent flames

cedes reaction, and consequently significant values of Y 共n兲
p As described above, the particle/mesh method, based on
are observed only downstream of y ⬇ 1, corresponding ap- stochastic Lagrangian models, amounts to a numerical solu-
proximately to the base of the lifted flame. tion to the modeled joint PDF transport equation. Based on
Several important observations can be made about the this approach, there have been numerous PDF-method stud-
ensemble of particles within a given mesh cell, for example, ies of turbulent flames, which have recently been reviewed
a cell around x = 0, y = 4, in the center of the flame. First, it is by Haworth.26 Among the demonstrated successes of the
possible for two particles to have the same position but sig- approach35,36 is its ability to account quantitatively for the
nificantly different velocity and composition. This is because local extinction and reignition observed in the Sandia series
the computational particles model fluid particles on different of piloted jet diffusion flames.37 Here, we illustrate the per-
realizations of the flow. For the same reason, it is not pos- formance of the PDF method for the single case of a lifted jet
sible to extract two-point turbulence statistics from the par- flame in a heated coflow—a laboratory flame, qualitatively
ticles. Second, within a cell there is a distribution of particle similar to that considered in Sec. III C. The flame
properties, reflecting the distribution of fluid properties in the considered38–40 consists of a central fuel jet 共of diameter D兲
turbulent flow: in the particle/mesh method, the PDF is not of an H2 / N2 mixture at 300 K surrounded by a hot coflow
represented directly, but the joint PDF of particle properties consisting of the products of lean H2 / air combustion at tem-
within a cell models the same joint PDF in the turbulent flow. perature Tc around 1000 K. A striking observation is that the
lift-off height H of the flame increases rapidly as the coflow
Third, ensemble averages of properties 共e.g., U+ , ␾+ , ␻+兲 of
temperature is decreased in the narrow range from 1010 to
particles within a cell provide estimates of the corresponding
1050 K.
means 共i.e., 具U典 , 具␾典 , 具␻典兲 at cell centers.
PDF calculations of this flame have been performed by
Figure 10 illustrates that instantaneous cell means con-
Cao et al.41 and Wang and Pope42 using the method de-
tain significant statistical fluctuations, which are here larger scribed above, based on the particle/mesh method and sto-
than normal because of the relatively small number of par- chastic models for velocity, composition, and frequency. The
ticles used in this simple simulation. However, as illustrated 11 composition variables are the enthalpy and ten chemical
in the right-hand frame, for this statistically stationary flow, species, which react according to a detailed chemical
the statistical fluctuations can be reduced at will by time mechanism.43 Figure 11 compares the measured and calcu-
averaging to yield stable estimates of means. lated normalized lift-off height H / D as a function of coflow
The stochastic Lagrangian models governing the par- temperature. As may be seen, there is excellent agreement
ticles’ evolution contain both particle properties 共especially considering the ⫾25 K experimental uncertainty
关X+共t兲 , U+共t兲 , ␾+共t兲 , ␻+共t兲兴 and also means evaluated at par- in the absolute temperature兲. The calculated profiles of mean
ticle locations 共e.g., 具U典+ , 具␾典+兲. In particle/mesh methods, and rms temperature and species41 are also in excellent
the mesh is used in mean estimation 共most simply by en- agreement with the experimental data.
semble averaging over cells兲 and also in the interpolation of Insight into the flame stabilization mechanism can be
means onto the particles 共most simply by multilinear inter- gained by examining particle properties in the PDF calcula-
polation兲. In practice, more sophisticated techniques may be tions. Specifically, we examine the evolution of the particles’
used for mean estimation and for interpolation.34 mixture fraction ␰+共t兲 and temperature T+共t兲 as they move up
50
40
30
H/D
20
10
0
1000 1020 1040 1060 1080 1100
Tc (K)
FIG. 11. Lift-off height H 共normalized by the jet diameter D兲 against coflow FIG. 13. 共Color online兲 Scatter plot of mixture fraction and temperature of
temperature Tc: symbols, experimental data 共Ref. 39兲; line with symbols, particles emanating from the fuel jet in PDF calculations 共Ref. 42兲
PDF calculations 共Ref. 41兲. of the Cabra flame. Print version, particles at x / D = 13; on-line version,
animation with time corresponding to x / D 共enhanced online兲. 关URL:
http://dx.doi.org/10.1063/1.3531744.3兴
in the flame. Figure 12 shows the ␰-T composition space.

Pure fuel and pure coflow correspond to the points 共␰ , T兲
are close to 共␰ , T兲 = 共1 , 300 K兲. As time evolves, the particles
= 共1 , 300 K兲 and 共␰ , T兲 = 共0 , 1045 K兲, respectively. If there
move upward 关increasing X+1 共t兲兴, and the animation shows
were inert mixing between these two streams, then all par-
ticles would lie on the line between these two points 共labeled the particle properties evolving as their axial position in-
“inert mixing”兲. On the right-hand part of this line, the tem- creases. That is, “time” in the animation corresponds to
peratures are low, and the mixture is essentially inert. In downstream distance x / D. As may be seen in the animation,
contrast, on the left-hand part of the line, the temperatures early on 共x / D ⬍ 9兲 there is essentially inert mixing. At x / D
are sufficiently high for the mixture to autoignite. The most ⬇ 9, a few particles approach the most reactive composition
reactive mixture—that with the shortest ignition-delay and then move upward, away from the inert-mixing line, by
time—occurs at a quite lean mixture, ␰ ⬇ 0.07. If the particles virtue of chemical reactions. As more particles react, mixing
reacted to the point of reaching chemical equilibrium, then increases the temperature of other particles until they are
they would lie on the indicated upper curve. sufficiently hot 共T ⱖ 1000 K兲 to react at a significant rate.
Figure 13, from the PDF calculations of Wang and Eventually, nearly all of the particles are close to the equi-
Pope,42 shows the evolution of the properties of particles that librium line, and they move along it due to mixing.
emanate from the fuel jet. Hence, initially, all of the particles The picture that emerges from these PDF calculations is
that the route to flame stabilization is inert mixing of the fuel
and coflow leading to mixtures that autoignite and then
1800 propagate the combustion. In contrast to most other flames,
this stabilization mechanism does not require propagation of
1600 heat or products against the flow. Subsequent studies44,45
m have confirmed and amplified this view.
ib riu
uil
Eq
1400
Eq
uil
E. Large-eddy simulation/PDF methods

Temperature/(K)
ibr
1200
ium
Pure co−flow In the PDF method described above, the velocity-

1000 frequency joint PDF provides a statistical description of the
flow and turbulent motions of all scales. Over the past two
800 decades, large-eddy simulation 共LES兲 has emerged as a
Iner Pure fuel popular alternative to purely statistical approaches. In LES,
t mi
600 Most Reactive xing the large-scale turbulent motions are explicitly represented,
mixture while statistical models account for the influence of the un-
400
resolved, smaller scales. The statistical models for the small
scales are usually simple algebraic models—for example, the
0 0.2 0.4 0.6 0.8 1
Mixture Fraction Smagorinsky model—but PDF methods are also used.
In the LES context, there are nontrivial issues concerned
FIG. 12. 共Color online兲 The composition space 共projected onto the mixture
with the appropriate definition of the PDF 共or related quan-
fraction/temperature plane兲 for the Cabra lifted flame, showing compositions
corresponding to pure fuel, pure coflow, inert mixing, and chemical tity兲. Based on the filtering approach to LES, Pope46 intro-
equilibrium. duced the filtered density function 共FDF兲. The corresponding
governing equations have been derived,47 modeled, and molecular mixing is modeled to be determined by the large-
implemented.48 However, the FDF pertains to a single real- scale turbulent motions, independent of the molecular prop-
ization and does not account for the distribution of subfilter- erties. Given that the Reynolds numbers of the turbulent
scale fields. In the author’s view, a better definition is that of flames mentioned are not large, it is perhaps surprising how
a PDF conditional on the resolved fields, as proposed by successful this scaling argument appears to be.
Fox33 and developed by Pope.49 At this stage of develop- In spite of the successes mentioned above, it is certainly
ment, the difference between LES/FDF and LES/PDF is the case that the modeling of molecular mixing is not satis-
purely conceptual, with the modeled equations used being factory in all respects, and many attempts have been made to
the same. However, there are real differences in their exact construct improved models, e.g., Refs. 57–59. Perhaps the
evolution equations, especially in the molecular transport greatest challenge is posed by premixed turbulent combus-
terms as the LES is refined to approach DNS, and we can tion in the flamelet regime. In that case, the steepest scalar
expect modeled LES/PDF equations to be developed to con- gradients result from a reaction-diffusion balance in the
form to this known limiting behavior. flamelets, rather than from the straining-diffusion balance ex-
There are several different PDF approaches, dependent perienced by nonreactive scalars. In several studies,60,61 PDF
on the set of variables considered, primarily combinations of methods have been applied to premixed combustion in the
composition, velocity, and frequency. Most of these ap- flamelet regime.
proaches have been implemented in the LES context by Givi There are two interesting observations concerning mo-
and co-workers.50 However, the most widely used LES/PDF lecular diffusion in LES/PDF. First, in contrast to PDF meth-
approach is for reactive flows, and it uses conventional LES ods, the direct effect of molecular diffusion can be signifi-
modeling for the velocity field, and the PDF approach for the cant, and even dominant. In LES simulations of the Sandia
compositions. In this case, the position X+共t兲 of the compu- Flame D, it is found that 共on reasonable grids兲 the molecular
tational particles evolves by the SDE diffusivity is larger than the turbulent diffusivity in the near
field and at all but the lowest temperatures.62 共This observa-
冋册冑冉冊
tion is relevant to all LES approaches, not just to LES/PDF.兲
+ +
1 2⌫T To make the second observation, we consider a high-
dX+共t兲 = Ũ + ⵜ ⌫T dt + dW, 共56兲
具␳典具␳典 Reynolds-number flow involving the mixing of a conserved
scalar ␾, the mixture fraction, which is zero in one stream
and unity in the other. The integral scale of the turbulence is
where Ũ共x , t兲 is the resolved LES velocity field, 具␳典 is the L, and the LES resolution parameter 共e.g., the filter width兲 ⌬
resolved density, and ⌫T is the turbulent 共subgrid-scale兲 dif- is in the inertial range 共L Ⰷ ⌬ Ⰷ ␩兲.
fusivity. With ␾¯ denoting the LES field, and with 具典 denoting the
In recent years, there has been an increasing use of LES/ mean, the scalar can be decomposed as
PDF methods for turbulent combustion, both in academic
research and in industry.51 Reviews of this work are provided
by Pitsch,52 Haworth,26 and Haworth and Pope,53 and there ␾ = 具␾
¯ 典 + 共␾
¯ − 具␾
¯ 典兲 + 共␾ − ␾
¯ 兲. 共57兲
are recent examples using a particle/mesh method,54 using
the stochastic fields method55 and using DQMOM.56 While The three terms on the right-hand side correspond to the
LES/PDF is computationally more expensive than both LES mean, the resolved fluctuation, and the residual fluctuation.
and PDF methods, it combines the merits of both in provid- To a good approximation, the composition variance can be
ing an accurate description of the turbulent velocity field and expressed as the sum of resolved and residual contributions,
of the turbulence-chemistry interactions, which typically oc-
cur on the smallest, unresolved scales. 具共␾ − 具␾典兲2典 ⬇ 具共␾
¯ − 具␾
¯ 典兲2典 + 具共␾ − ␾
¯ 兲2典. 共58兲
F. Discussion on the modeling of mixing in PDF Standard scaling arguments16 give the residual variance de-
and LES/PDF methods
creasing as 共⌬ / L兲2/3 as ⌬ decreases, and hence this modeled
Models are sometimes criticized for lacking a represen- contribution becomes small compared to the known resolved
tation of a physical process deemed essential to the phenom- contribution. This is the normal picture of LES—the mod-
enon at hand. Clearly, in reactive flows, the processes of eled unresolved contributions become progressively less im-
molecular diffusion and thermal conductivity are essential, portant as ⌬ decreases.
since only by molecular diffusion can fuel and oxidant mix Consider now the composition ␾+共t兲 of a computational
to form a reactive mixture, and only by conduction can hot particle in a LES/PDF calculation of this flow. For the case
and cold fluids produce warm fluid. It is these processes that considered 共⌬ Ⰷ ␩兲, the direct effects of molecular diffusion
in PDF methods are modeled by the IEM and similar models. are negligible 共since ⌫ Ⰶ ⌫T兲, whereas molecular mixing is
It is interesting to observe, therefore, that PDF methods are modeled, for example, by the IEM model. We observe, there-
successful in accounting for challenging phenomena such as fore, that ␾+共t兲 evolves solely due to the modeled term: if the
local extinction and reignition, and stabilization of lifted mixing model were omitted, then there would be no mecha-
flames in hot coflows, and yet these models do not involve nism for ␾+共t兲 to depart from its initial value 共0 or 1兲, and
the molecular diffusivity or conductivity of the fluid. Instead, hence the predicted composition PDF would be a double-
in accord with the cascade picture of turbulence, the rate of delta function everywhere. Furthermore, scaling arguments
show that the 共normalized兲 mixing rate ␻m increases as 8

共⌬ / L兲−2/3, indicating that in some sense the “strength” of the
model increases as ⌬ decreases. Nevertheless, since the re-
sidual variance decreases as ⌬ decreases, the sensitivity of 6
the LES/PDF calculations to the model also decreases.
D(s)/(εs)
IV. ADVANCED STOCHASTIC LAGRANGIAN MODELS 4
In this section, we examine more closely the Lagrangian
velocity in homogeneous isotropic turbulence, and the extent
to which it is described by the Langevin model. Referring to 2
Box’s statement, “All models are wrong; but some are use-
ful,” we have already seen the usefulness of the Langevin
model not only for isotropic turbulence, but as an important 0 −2 0 2
component of PDF methods applied to inhomogeneous flows 10 10 10
such as turbulent flames. Nevertheless, when examined in s/TL
detail, the model is found to be wrong—or, more charitably,
it provides an incomplete description of the phenomena. In FIG. 14. 共Color online兲 Compensated second-order Lagrangian structure
Secs. IV A–IV D that follow, we examine the Lagrangian functions: solid lines, DNS 共Ref. 70兲 at Reynolds numbers 共from bottom to
top兲 R␭ = 43, 86, 140, 235, 393, 595, 1000; dashed line, Langevin model
velocity on small time scales, observe deficiencies of the with C0 = 2.1.
Langevin model, and show how they can be remedied.
Prior to 1989, we had little detailed knowledge of La-
grangian statistics in turbulence due to the obvious experi-
mental difficulties. However, over the past 20 years, starting
s
with the work of Yeung and Pope,15 we have obtained exten- D共s兲 ⬇ C0␧s, for Ⰶ 1, 共64兲
sive, detailed information from DNS;63–65 and in the past TL
decade, remarkable progress has been made in experimental consistent with the Kolmogorov hypotheses.
techniques enabling the measurement of Lagrangian velocity Figure 14 shows compensated structure functions, i.e.,
and acceleration.66–69 It is this information from DNS and D共s兲 / 共␧s兲, obtained from DNS,70 for Taylor-scale Reynolds
experiments which has enabled the development of more ad- numbers R␭ from 43 to 1000, compared to the Langevin-
vanced stochastic Lagrangian models. model result, Eq. 共62兲. The following clear observations can
be made from the DNS data:
A. Lagrangian velocity increments
共1兲 There is a strong Reynolds-number dependence. Indeed,
We consider statistically stationary, homogeneous, iso- it is generally found that compared to Eulerian statistics,
tropic turbulence with turbulence intensity u⬘ and mean dis- Lagrangian statistics exhibit a stronger Reynolds-
sipation rate ␧. A component of the Lagrangian velocity 共of a number dependence, and do so up to higher Reynolds
fluid particle兲 is denoted by u+共t兲. This has mean zero and numbers.
variance u⬘2. The most basic multitime statistics are the au- 共2兲 Even at the highest Reynolds number, D共s兲 / 共␧s兲 does
tocorrelation function not convincingly display the plateau predicted by Kol-
␳共s兲 ⬅ 具u+共s兲u+共0兲典, 共59兲 mogorov theory 关Eq. 共64兲兴.
共3兲 At larger times 关for s to the right of the peak of
and the second-order Lagrangian structure function D共s兲 / 共␧s兲兴 there is a collapse of the compensated struc-
D共s兲 ⬅ 具关u+共s兲 − u+共0兲兴2典, 共60兲 ture functions at different Reynolds number 共when plot-
ted against s / TL兲.
which is just the variance of the increment over the time 共4兲 At very small times, D共s兲 / 共␧s兲 increases linearly with s,
interval s ⱖ 0. These two functions are related to each other since a Taylor series for 关u+共s兲 − u+共0兲兴 yields
冓冉冊冔
by 2
du+ s
D共s兲 = 2u⬘2共1 − ␳共s兲兲. 共61兲 D共s兲 ⬇ s2 , for Ⰶ 1. 共65兲
dt ␶␩
According to the Langevin model, the structure function
is In comparison, the Langevin model has no Reynolds-number
dependence, it does yield a plateau, and this extends to s
D共s兲 = C0␧TL共1 − e−s/TL兲, 共62兲 = 0. 关Note that u+共t兲 given by the Langevin equation is con-
where the Lagrangian velocity integral time scale TL is re- tinuous but not differentiable, and so Eq. 共65兲 does not apply
lated to the Eulerian turbulence properties by to it.兴
2u⬘2 4k B. Reynolds number
TL = = . 共63兲
C0␧ 3C0␧
It is straightforward to incorporate Reynolds-number de-
Note that for small times 共s / TL Ⰶ 1兲, Eq. 共62兲 yields pendence in the Langevin model simply by making the
FIG. 15. 共Color online兲 The Langevin-model constant C0 against Reynolds FIG. 16. 共Color online兲 Compensated second-order Lagrangian structure
number: symbols, from DNS 共Ref. 70兲 and Eq. 共67兲; line, empirical fit, functions: solid lines, DNS 共Ref. 70兲 at Reynolds numbers 共from bottom to
Eq. 共68兲. top兲 R␭ = 43, 86, 140, 235, 393, 595, 1000; dashed lines, Langevin model
with C0 obtained from Eq. 共67兲.
model coefficient C0 depend on R␭, i.e., C0共R␭兲. Consistency

with the Kolmogorov hypotheses requires only In 1991, Sawford7 introduced a linear stochastic model
lim C0共R␭兲 = C0 , 共66兲 for acceleration,
R␭→⬁
where now we distinguish between the Kolmogorov constant

da+ = − 1 +冉冊 ␶ a+
T⬁ ␶
dt −
u+
T ⬁␶
dt
C0 and the model coefficient C0共R␭兲. Furthermore, C0共R␭兲
can be determined directly from DNS data via Eq. 共63兲, i.e.,
4k
再冉冊冎
+ 2a⬘2
1 1
+
␶ T⬁
1/2
dW. 共69兲
C0 = . 共67兲 The two specified time scales T⬁ and ␶ are related to the
3␧TL
integral scale and the Kolmogorov scale, respectively, and
Figure 15 shows values of C0 thus obtained from DNS,70 their ratio increases with Reynolds number. Based in these
compared to the empirical fit time scales and the rms velocity u⬘, the acceleration scale is
6.5 defined by
C0共R␭兲 = , 共68兲
共1 + 140R␭−4/3兲3/4 u ⬘2
a ⬘2 = . 共70兲
T ⬁␶
which is based on a suggestion by Sawford et al.71 As may
be seen, the fit represents the data well, and is consistent with Analytic expressions for the autocorrelation functions
the Kolmogorov hypotheses with C0 = 6.5. and structure functions can be deduced from Sawford’s
Figure 16 compares the compensated structure functions model. Figure 17 compares the predicted compensated struc-
from DNS with those from the Langevin model with C0共R␭兲 ture functions with those from DNS. As may be seen, there is
specified by Eq. 共67兲. As may be seen, the Langevin model good agreement, and in particular the small time scales are
now provides an accurate representation of the structure well represented.
function for s / ␶␩ ⬎ 10, say. Even though C0 is specified
based on the data, this agreement for the structure function is D. Intermittency
not inevitable. Instead, it indicates that, except in the dissi-
As is to be expected, the statistics of acceleration and
pation range 共s ⬍ 10␶␩兲, D共s兲 is characterized by the single
velocity increments over small time intervals are found to be
time scale TL.
highly non-Gaussian—a manifestation of interval intermit-
tency. For example, in an experiment at R␭ = 680, Mordant et
C. Stochastic model for acceleration
al.72 measured the kurtosis of acceleration to be greater than
As observed from Fig. 16, the behavior of the Langevin 100, compared to the Gaussian value of 3. From the DNS of
model is qualitatively incorrect at small times. This is inevi- Yeung et al.,73 Fig. 18 shows the PDF of acceleration. As
table given that u+共t兲 is modeled as a diffusion process. The may be seen, this is much broader than the Gaussian distri-
problem can be removed, however, by moving the stochastic bution 共of the same standard deviation兲.
modeling to the next level; that is, by constructing a stochas- As originally suggested by Oboukhov74 and
tic model for the Lagrangian acceleration a+共t兲, and then Kolmogorov17 in 1962, the standard way to approach inter-
obtaining the velocity from du+ / dt = a+. nal intermittency is to condition statistics based on the local
FIG. 19. 共Color online兲 Kurtosis of ␹, the logarithm of pseudodissipation,

FIG. 17. 共Color online兲 Compensated second-order Lagrangian structure obtained from DNS 共Ref. 73兲 against Reynolds number compared to the
functions: solid lines, DNS 共Ref. 70兲 at Reynolds numbers 共from bottom to Gaussian value of 3.
top兲 R␭ = 43, 86, 140, 235, 393, 595, 1000; dashed lines, from Sawford’s
model 共Ref. 7兲, Eq. 共69兲.
to log-normal, or, equivalently, the quantity

dissipation rate. An early lesson from DNS 共Ref. 15兲 is that,
for this purpose, the pseudodissipation ␸ ⬅ ␯ ⳵ ui / ⳵x j ⳵ ui / ⳵x j
is superior to the dissipation ␧ ⬅ 21 ␯共⳵ui / ⳵x j + ⳵u j / ⳵xi兲
␹+共t兲 ⬅ ln 冉冊
␸+共t兲
具␸典
共71兲
⫻共⳵ui / ⳵x j + ⳵u j / ⳵xi兲. is close to Gaussian. Figure 19 shows that the kurtosis of ␹+

Also shown in Fig. 18 are the PDFs of acceleration con- is remarkably close to the Gaussian value of 3 over the large
ditional on ␸. Specifically, there are five conditional PDFs range of Reynolds numbers investigated. Furthermore, the
corresponding to the five quintiles of ␸. As may be seen, the autocorrelation function of ␹+共t兲, shown in Fig. 20, is well
conditional PDFs are much narrower that the unconditional approximated by an exponential. It is thus natural to model
PDF, and, remarkably, they are essentially independent of the ␹+共t兲 as an Ornstein–Uhlenbeck process, governed by a lin-
conditioning variable, i.e., the five curves collapse. ear SDE similar to the Langevin equation.
Several stochastic models have been constructed which Let ␴a共␹兲 denote the standard deviation of the accelera-
account for intermittency.8,75,76 The starting point is a sto- tion a+共t兲 conditioned on ␹+共t兲 = ␹. The DNS data73 show, as
chastic model for pseudodissipation, ␸+共t兲. A striking obser- expected, that ␴a共␹兲 increases steeply with ␹. The condition-
vation from DNS is that the one-time PDF of ␸+ is very close ally standardized acceleration â共t兲 is defined by
FIG. 18. 共Color online兲 Standardized PDFs of acceleration from DNS 共Ref.
73兲: outer solid line 共red兲, unconditional PDF; inner solid lines 共six indis-
tinguishable lines兲, PDFs conditional on pseudodissipation quintiles 共ma- FIG. 20. 共Color online兲 Autocorrelation function of ␹, the logarithm of
genta, online兲, and the cubic Gaussian equation 关Eq. 共73兲兴共green, online兲; pseudodissipation, obtained from DNS 共Ref. 77兲 compared to the exponen-
inner dashed line, Gaussian 共blue兲. tial 共dashed line兲.
a+共t兲 than modeling the velocity and related quantities. Some rea-
â共t兲 ⬅ . 共72兲 sons for these challenges 共not all independent兲 are the fol-
␴a关␹+共t兲兴
lowing:
The fact that the PDFs of conditional acceleration collapse in 共1兲 In the conservation equation for compositions, there is
Fig. 18 indicates that the one-time PDF of â共t兲 is indepen- no term analogous to the pressure gradient that appears
dent of ␹+共t兲 and is that observed in the figure. An empirical in the velocity equation, and which has a randomizing
finding is that this PDF is well approximated by a “cubic effect.
Gaussian.” That is, the PDF of â共t兲 is well approximated by 共2兲 Mixing occurs predominantly at the smallest scales,
that of whereas the particle composition 共or, equivalently, the
one-point joint PDF of composition兲 contains no scale
information, but is dominated by the large scales.
ã ⬅ 共1 − p兲ā + pā3 , 共73兲共3兲 Different compositions can have different diffusivities,
leading to differential diffusion effects, including
where ā is a Gaussian random variable and p = 0.1. thermal-diffusive instabilities.
Based on the above observations, Lamorgese et al.8 pro- 共4兲 Individual compositions 共e.g., species mass fractions兲
posed a stochastic model for velocity u+共t兲, acceleration are bounded 共i.e., between 0 and 1兲, and sets of compo-
a+共t兲, and pseudodissipation ␸+共t兲, which, by construction, sitions with equal diffusivities satisfy joint boundedness
yields the correct one-time joint PDF of these quantities, thus conditions.
appropriately accounting for internal intermittency. 共5兲 Sets of compositions with equal diffusivities satisfy lin-
earity and independence conditions,92 to which models
E. Further stochastic Lagrangian models should adhere.
共6兲 In reacting flows 共especially premixed turbulent com-
Without attempting a comprehensive review, we mention bustion in the flamelet regime兲 the steepest composition
here some further stochastic Lagrangian models that have gradients can be caused by reaction fronts, rather than
been developed. For velocity and acceleration, the models by turbulent straining.
described above focus on the small time scales. For inhomo-
geneous flows, of more importance is the effect of mean Over the past 30 years, several models have been devel-
velocity gradients. The generalized Langevin model 共GLM兲 oped based solely on particle composition. These include the
共Ref. 78兲 accounts for these effects by making the drift term modified Curl model,93,94 the binomial Langevin model,95
in the Langevin equation for the velocity vector depend lin- the mapping closure,96,97 the Euclidean minimum spanning
early on the mean velocity gradient tensor. There is a corre- tree 共EMST兲 model,57 and multiple mapping conditioning.59
spondence between the tensor drift coefficient and models It is natural to attempt to improve the description of the
for the pressure-rate-of-strain tensor in Reynolds-stress physics of mixing by incorporating scale information. It has
models.21 In a similar manner, Sawford’s acceleration model proved difficult to do so in a way that leads to a tractable
can be generalized to incorporate the effects of mean veloc- model for inhomogeneous reactive flows. Some attempts
ity gradients.79 have been based on composition gradients,98 and on spectral
The GLM models 共in part兲 the “rapid” pressure, and it is representations,58 including combining PDF methods with
consistent with rapid distortion theory to the limited extent the eddy-damped quasi-normal Markovian 共EDQNM兲
that it correctly represents the initial response of isotropic approach.99
turbulence. The stochastic wave-vector model80 introduces a
unit wavenumber vector e+共t兲 as an additional particle prop-
V. CONCLUSIONS AND FUTURE CHALLENGES
erty. Coupled equations for u+共t兲 and e+共t兲 are then con-
structed using a mathematical analogy to the Navier–Stokes In this paper, we have illustrated the potency and broad
equations in wavenumber space such that the Reynolds- applicability of stochastic Lagrangian models of turbulence.
stress evolution is correct for all rapid distortions of homo- Dispersion from a line source in grid turbulence is a funda-
geneous turbulence. It is interesting to note that this exact mental flow in the study of turbulent transport and mixing.
treatment of rapid distortion is possible through stochastic Turbulent diffusion models 共e.g., based on the k-␧ model兲 are
modeling, but it cannot be achieved in moment closures. qualitatively incorrect, except far from the source. In con-
Other quantities and processes related to turbulent veloc- trast, the simple Langevin model for the fluid-particle veloc-
ity fields for which stochastic Lagrangian models have been ity and the IECM model for composition yield accurate pre-
constructed include the velocity gradient tensor,81–83 two- dictions of the mean, variances, and covariances from single
particle dispersion,84–87 and inertial particles.88–90 Some re- and multiple line sources.
cent reviews are provided by Meneveau91 and Sawford and In Sec. III, it is shown that with straightforward exten-
Pinton.18 sions, these simple stochastic Lagrangian models for velocity
Above we have described the IEM and IECM models and composition can be used to effect a turbulence closure,
which model the evolution of the composition ␾+共t兲 of a in terms of the joint PDF of velocity and composition, and
fluid particle due to molecular diffusion. Such “mixing mod- the stochastic models form the basis for a natural particle/
els” are central to PDF methods for turbulent reactive flows. mesh numerical method to solve the modeled PDF equation.
Arguably, this general problem is much more challenging These PDF methods have been applied to several turbulent
4
flames, and they have proved most successful in accounting M. S. Anand and S. B. Pope, in Turbulent Shear Flows 4, edited by L. J.
S. Bradbury, F. Durst, B. E. Launder, F. W. Schmidt, and J. H. Whitelaw
for important turbulence-chemistry interactions.
共Springer-Verlag, Berlin, 1985兲, pp. 46–61.
With the advent of DNS and modern diagnostics, it has 5
B. L. Sawford, “Micro-mixing modelling of scalar fluctuations in plumes
been possible to examine in detail Lagrangian time series in in homogeneous turbulence,” Flow, Turbul. Combust. 72, 133 共2004兲.
6
turbulent flows. This examination has shown that the simple S. Viswanathan and S. B. Pope, “Turbulent dispersion behind line sources
in grid turbulence,” Phys. Fluids 20, 101514 共2008兲.
Langevin model has several limitations and deficiencies. All 7
B. L. Sawford, “Reynolds number effects in Lagrangian stochastic models
of these can be remedied, but at the cost of more complexity. of turbulent dispersion,” Phys. Fluids A 3, 1577 共1991兲.
8
For example, a model for velocity, acceleration, and pseudo- A. G. Lamorgese, S. B. Pope, P. K. Yeung, and B. L. Sawford, “A condi-
dissipation is able to describe accurately the Lagrangian ve- tionally cubic-Gaussian stochastic Lagrangian model for acceleration in
isotropic turbulence,” J. Fluid Mech. 582, 423 共2007兲.
locity on all time scales, including Reynolds-number effects 9
M. S. Uberoi and S. Corrsin, “Diffusion of heat from a line source in
and internal intermittency. isotropic turbulence,” National Aeronautics and Space Administration Re-
Looking to the future, there are of course enumerable port No. 1142, 1953.
10
opportunities and challenges for the further development of A. A. Townsend, “The diffusion behind a line source in homogeneous
turbulence,” Proc. R. Soc. London, Ser. A 224, 487 共1954兲.
modeling and simulation methodologies for turbulent flows. 11
Z. Warhaft, “The interference of thermal fields from line sources in grid
We mention here just two of these, in the context of stochas- turbulence,” J. Fluid Mech. 144, 363 共1984兲.
12
tic Lagrangian modeling in conjunction with LES. H. Stapountzis, B. L. Sawford, J. C. R. Hunt, and R. E. Britter, “Structure
First, in a LES of a high-Reynolds-number turbulent of the temperature field downwind of a line source in grid turbulence,” J.
Fluid Mech. 165, 401 共1986兲.
flow, there is a significant separation of scale between the 13
B. L. Sawford and C. M. Tivendale, “Measurements of concentration sta-
smallest scales in the flow and the smallest scales resolved in tistics downstream of a line source in grid turbulence,” Proceedings of the
the LES. As discussed in Sec. IV E, it remains a challenge to 11th Australasian Fluid Mechanics Conference 共University of Tasmania,
model small-scale processes such as molecular mixing, espe- Hobart, 1992兲, pp. 945–948.
14
W. P. Jones and B. E. Launder, “The prediction of laminarization with a
cially when processes on these small scales are rate limiting,
two-equation model of turbulence,” Int. J. Heat Mass Transfer 15, 301
or when they create 共rather than dissipate兲 fluctuations. High- 共1972兲.
15
Reynolds-number DNS now provides the information P. K. Yeung and S. B. Pope, “Lagrangian statistics from direct numerical
needed to develop and test such models. simulations of isotropic turbulence,” J. Fluid Mech. 207, 531 共1989兲.
16
S. B. Pope, Turbulent Flows 共Cambridge University Press, Cambridge,
Second, while LES provides significant modeling advan-
2000兲.
tages over Reynolds-averaged Navier–Stokes 共RANS兲 ap- 17
A. N. Kolmogorov, “A refinement of previous hypotheses concerning the
proaches, it has several problematic aspects.100 In a RANS local structure of turbulence in a viscous incompressible fluid at high
calculation, numerical errors can be quantified and reduced Reynolds number,” J. Fluid Mech. 13, 82 共1962兲.
18
B. L. Sawford and J.-F. Pinton, in The Nature of Turbulence, edited by P.
below acceptable levels. On the other hand, in LES as it is
A. Davidson, Y. Kaneda, and K. R. Sreenivasan 共Cambridge University
generally practiced, the calculated statistics depend both on Press, Cambridge, 2011兲.
19
the numerical method and on the grid used. A worthwhile J. Villermaux and J. C. Devillon, “Représentation de la coalescence et de
challenge for future research is the development of a LES la redispersion des domaines de ségrégation dans un fluide par un modèle
d’interaction phénoménologique,” Proceedings of the Second Interna-
methodology which yields calculations with controllably
tional Symposium on Chemical Reaction Engineering 共Elsevier, New
small numerical effects. Such a methodology most likely re- York, 1972兲, pp. 1–13.
20
quires adaptive mesh refinement and calculations on multiple C. Dopazo and E. E. O’Brien, “An approach to the autoignition of a
grids. turbulent mixture,” Acta Astronaut. 1, 1239 共1974兲.
21
S. B. Pope, “On the relationship between stochastic Lagrangian models of
turbulence and second-moment closures,” Phys. Fluids 6, 973 共1994兲.
ACKNOWLEDGMENTS 22
R. O. Fox, “On velocity-conditioned scalar mixing in homogeneous tur-
bulence,” Phys. Fluids 8, 2678 共1996兲.
This paper is based on the Otto Laporte Lecture deliv- 23
S. B. Pope, “The vanishing effect of molecular diffusivity on turbulent
ered at the 62nd Annual Meeting of the American Physical dispersion: Implications for turbulent mixing and the scalar flux,” J. Fluid
Society Division of Fluid Dynamics on 22 November 2009, Mech. 359, 299 共1998兲.
24
O. Reynolds, “On the dynamical theory of incompressible viscous flows
following the award of the APS Fluid Dynamics Prize. I
and the determination of the criterion,” Philos. Trans. R. Soc. London, Ser.
would like to thank all my current and former students, post- A 186, 123 共1894兲.
25
docs, and collaborators who have contributed to the body of S. B. Pope, “PDF methods for turbulent reactive flows,” Prog. Energy
research described here. I am particularly grateful to Brian Combust. Sci. 11, 119 共1985兲.
26
D. C. Haworth, “Progress in probability density function methods for tur-
Sawford, Haifeng Wang, Sharadha Viswanathan, Zellman
bulent reacting flows,” Prog. Energy Combust. Sci. 36, 168 共2010兲.
Warhaft, and P. K. Yeung who have directly contributed to 27
B. J. Delarue and S. B. Pope, “Application of PDF methods to compress-
the paper. This work is supported in part by Air Force Office ible turbulent flows,” Phys. Fluids 9, 2704 共1997兲.
28
of Scientific Research under Grant No. FA-9550-09-1-0047 P. R. Van Slooten, Jayesh, and S. B. Pope, “Advances in PDF modeling for
inhomogeneous turbulent flows,” Phys. Fluids 10, 246 共1998兲.
and in part by the Department of Energy under Grant No. 29
M. Muradoglu, P. Jenny, S. B. Pope, and D. A. Caughey, “A consistent
DE-FG02-90ER14128. hybrid finite-volume/particle method for the PDF equations of turbulent
reactive flows,” J. Comput. Phys. 154, 342 共1999兲.
1 30
A. N. Kolmogorov, “The local structure of turbulence in incompressible S. B. Pope, “A Monte Carlo method for the PDF equations of turbulent
viscous fluid for very large Reynolds numbers,” Dokl. Akad. Nauk SSSR reactive flow,” Combust. Sci. Technol. 25, 159 共1981兲.
31
30, 299 共1941兲. L. Valiño, “A field Monte Carlo formulation for calculating the probability
2
P. Langevin, “Sur la théorie du mouvement Brownien,” Acad. Sci., Paris, density function of a single scalar in a turbulent flow,” Flow, Turbul.
C. R. 146, 530 共1908兲. Combust. 60, 157 共1998兲.
3 32
G. I. Taylor, “Diffusion by continuous movements,” Proc. London Math. V. A. Sabel’nikov and O. Soulard, “Rapidly decorrelating velocity-field
Soc. 20, 196 共1921兲. model as a tool for solving one-point Fokker-Planck equations for prob-
ability density functions of turbulent reactive scalars,” Phys. Rev. E 72, pation in homogeneous turbulence,” Phys. Fluids 9, 2364 共1997兲.
016301 共2005兲. 59
A. Y. Klimenko and S. B. Pope, “A model for turbulent reactive flows
33
R. O. Fox, Computational Models for Turbulent Reactive Flows 共Cam- based on multiple mapping conditioning,” Phys. Fluids 15, 1907 共2003兲.
60
bridge University Press, New York, 2003兲. S. B. Pope and M. S. Anand, “Flamelet and distributed combustion in
34
R. McDermott and S. B. Pope, “The parabolic edge reconstruction method premixed turbulent flames,” Proc. Combust. Inst. 20, 403 共1985兲.
共PERM兲 for Lagrangian particle advection,” J. Comput. Phys. 227, 5447 61
R. P. Lindstedt and E. M. Vaos, “Transported PDF modeling of high-
共2008兲. Reynolds-number premixed turbulent flames,” Combust. Flame 145, 495
35
J. Xu and S. B. Pope, “PDF calculations of turbulent nonpremixed flames 共2006兲.
with local extinction,” Combust. Flame 123, 281 共2000兲. 62
K. A. Kemenov and S. B. Pope, “Molecular diffusion effects in LES of a
36
R. Cao and S. B. Pope, “The influence of chemical mechanisms on PDF piloted methane-air flame,” Combust. Flame 157, 240 共2011兲.
63
calculations of nonpremixed piloted jet flames,” Combust. Flame 143, Y. Kaneda and T. Gotoh, “Lagrangian velocity autocorrelation in isotropic
450 共2005兲. turbulence,” Phys. Fluids A 3, 1924 共1991兲.
37 64
R. S. Barlow and J. H. Frank, “Effects of turbulence on species mass Y. Kimura and J. R. Herring, “Diffusion in stably stratified turbulence,” J.
fraction in methane/air jet flames,” Proc. Combust. Inst. 27, 1087 共1998兲. Fluid Mech. 328, 253 共1996兲.
38 65
R. Cabra, T. Myhrvold, J. Y. Chen, R. W. Dibble, A. N. Karpetis, and R. L. Biferale, G. Boffetta, A. Celani, A. Lanotte, and F. Toschi, “Particle
S. Barlow, “Simultaneous laser Raman-Rayleigh-LIF measurements and trapping in three-dimensional fully developed turbulence,” Phys. Fluids
numerical modeling results of a lifted turbulent H2 / N2 jet flame in a viti- 17, 021701 共2005兲.
ated coflow,” Proc. Combust. Inst. 29, 1881 共2002兲. 66
G. A. Voth, K. Satyanarayan, and E. Bodenschatz, “Lagrangian accelera-
39
Z. Wu, S. H. Starner, and R. W. Bilger, “Lift-off heights of turbulent tion measurements at large Reynolds number,” Phys. Fluids 10, 2268
H2 / N2 jet flames in a vitiated co-flow,” in Proceedings of the 2003 Aus- 共1998兲.
67
tralian Symposium on Combustion and the Eighth Australian Flame Days, S. Ott and J. Mann, “An experimental investigation of the relative diffu-
edited by D. Honnery 共Monash University, Victoria, 2003兲. sion of particle pairs in three-dimensional turbulent flow,” J. Fluid Mech.
40
R. L. Gordon, S. H. Starner, A. R. Masri, and R. W. Bilger, “Further 422, 207 共2000兲.
68
characterisation of lifted hydrogen and methane flames issuing into a vi- G. A. Voth, A. LaPorta, A. M. Crawford, J. Alexander, and E. Boden-
tiated coflow,” Proceedings of the Fifth Asia-Pacific Conference on Com- schatz, “Measurement of particle accelerations in fully developed turbu-
bustion, 2005. lence,” J. Fluid Mech. 469, 121 共2002兲.
41 69
R. Cao, S. B. Pope, and A. R. Masri, “Turbulent lifted flames in a vitiated N. Mordant, P. Metz, O. Michel, and J.-F. Pinton, “Measurement of La-
coflow investigated using joint PDF calculations,” Combust. Flame 142, grangian velocity in fully developed turbulence,” Phys. Rev. Lett. 87,
438 共2005兲. 214501 共2001兲.
42 70
H. Wang and S. B. Pope, “Lagrangian investigation of local extinction, P. K. Yeung, S. B. Pope, and B. L. Sawford, “Reynolds number depen-
re-ignition and auto-ignition in turbulent flames,” Combust. Theory Mod- dence of Lagrangian statistics in large numerical simulations of isotropic
ell. 12, 857 共2008兲. turbulence,” J. Turbul. 7, 58 共2006兲.
43 71
J. Li, Z. Zhao, A. Kazakov, and F. L. Dryer, “An updated comprehensive B. L. Sawford, P. K. Yeung, and J. F. Hackl, “Reynolds number depen-
kinetic model for H2 combustion,” Technical Report, Fall Technical Meet- dence of relative dispersion statistics in isotropic turbulence,” Phys. Fluids
ing of the Eastern States Section of the Combustion Institute, Penn State 20, 065111 共2008兲.
72
University, University Park, PA, 2003. N. Mordant, A. M. Crawford, and E. Bodenschatz, “Experimental La-
44
R. L. Gordon, A. R. Masri, S. B. Pope, and G. M. Goldin, “A numerical grangian acceleration probability density function measurement,” Physica
study of auto-ignition in turbulent lifted flames issuing into a vitiated D 193, 245 共2004兲.
co-flow,” Combust. Theory Modell. 11, 351 共2007兲. 73
P. K. Yeung, S. B. Pope, A. G. Lamorgese, and D. A. Donzis, “Accelera-
45
C. S. Yoo, R. Sankaran, and J. H. Chen, “Three-dimensional direct nu- tion and dissipation statistics in numerical simulations of isotropic turbu-
merical simulation of a turbulent lifted hydrogen jet flame in heated cof- lence,” Phys. Fluids 18, 065103 共2006兲.
low: Flame stabilization and structure,” J. Fluid Mech. 640, 453 共2009兲. 74
A. M. Oboukhov, “Some specific features of atmospheric turbulence,” J.
46
S. B. Pope, “Computations of turbulent combustion: Progress and chal- Fluid Mech. 13, 77 共1962兲.
lenges,” Proc. Combust. Inst. 23, 591 共1990兲. 75
S. B. Pope and Y. L. Chen, “The velocity-dissipation probability density
47
F. Gao and E. E. O’Brien, “A large-eddy simulation scheme for turbulent function model for turbulent flows,” Phys. Fluids A 2, 1437 共1990兲.
reacting flows,” Phys. Fluids A 5, 1282 共1993兲. 76
A. M. Reynolds, “Superstatistical mechanics of tracer-particle motions in
48
P. J. Colucci, F. A. Jaberi, P. Givi, and S. B. Pope, “Filtered density turbulence,” Phys. Rev. Lett. 91, 084503 共2003兲.
77
function for large eddy simulation of turbulent reacting flows,” Phys. Flu- P. K. Yeung, S. B. Pope, E. A. Kurth, and A. G. Lamorgese, “Lagrangian
ids 10, 499 共1998兲. conditional statistics, acceleration and local relative motion in numerically
49
S. B. Pope, “Self-conditioned fields for large-eddy simulations of turbulent simulated isotropic turbulence,” J. Fluid Mech. 582, 399 共2007兲.
flows,” J. Fluid Mech. 652, 139 共2010兲. 78
D. C. Haworth and S. B. Pope, “A generalized Langevin model for turbu-
50
P. Givi, “Filtered density function for subgrid scale modeling of turbulent lent flows,” Phys. Fluids 29, 387 共1986兲.
combustion,” AIAA J. 44, 16 共2006兲. 79
S. B. Pope, “A stochastic Lagrangian model for acceleration in turbulent
51
S. James, J. Zhu, and M. S. Anand, “Large eddy simulations of turbulent flows,” Phys. Fluids 14, 2360 共2002兲.
80
flames using the filtered density function model,” Proc. Combust. Inst. 31, P. R. Van Slooten and S. B. Pope, “PDF modeling of inhomogeneous
1737 共2007兲. turbulence with exact representation of rapid distortions,” Phys. Fluids 9,
52
H. Pitsch, “Large-eddy simulation of turbulent combustion,” Annu. Rev. 1085 共1997兲.
Fluid Mech. 38, 453 共2006兲. 81
S. S. Girimaji and S. B. Pope, “A stochastic model for velocity gradients
53
D. C. Haworth and S. B. Pope, “Transported probability density function in turbulence,” Phys. Fluids A 2, 242 共1990兲.
82
methods for Reynolds-averaged and large-eddy simulations,” Turbulent J. Martín, C. Dopazo, and L. Valiño, “Dynamics of velocity gradient in-
Combustion 共Springer, Dordrecht/Hedielberg/London/New York, 2011兲. variants in turbulence: Restricted Euler and linear diffusion models,” Phys.
54
H. Wang and S. B. Pope, “Large eddy simulation/probability density func- Fluids 10, 2012 共1998兲.
tion modeling of a turbulent CH4 / H2 / N2 jet flame,” Proc. Combust. Inst. 83
L. Chevillard and C. Meneveau, “Intermittency and universality in a La-
33, 1319 共2011兲. grangian model of velocity gradients in three-dimensional turbulence,” C.
55
W. P. Jones and V. N. Prasad, “Large eddy simulation of the Sandia flame R. Mec. 335, 187 共2007兲.
series 共D-F兲 using the Eulerian stochastic field method,” Combust. Flame 84
P. A. Durbin, “A stochastic model of two-particle dispersion and concen-
157, 1621 共2010兲. tration fluctuations in homogeneous turbulence,” J. Fluid Mech. 100, 279
56
H. Koo, P. Donde, and V. Raman, “A quadrature-based LES/transported 共1980兲.
85
probability density function approach for modeling supersonic combus- D. J. Thomson, “A stochastic model for the motion of particle pairs in
tion,” Proc. Combust. Inst. 33, 2203 共2011兲. isotropic high-Reynolds-number turbulence, and its application to the
57
S. Subramaniam and S. B. Pope, “A mixing model for turbulent reactive problem of concentration variance,” J. Fluid Mech. 210, 113 共1990兲.
86
flows based on Euclidean minimum spanning trees,” Combust. Flame M. S. Borgas and B. L. Sawford, “A family of stochastic models for
115, 487 共1998兲. two-particle dispersion in isotropic, homogeneous and stationary turbu-
58
R. O. Fox, “The Lagrangian spectral relaxation model of the scalar dissi- lence,” J. Fluid Mech. 279, 69 共1994兲.
87
M. S. Borgas and P. K. Yeung, “Relative dispersion in isotropic turbu- for the probability density function of turbulent scalar fields,” J. Non-
lence. Part 2. A new stochastic model with Reynolds-number depen- Equilib. Thermodyn. 4, 47 共1977兲.
dence,” J. Fluid Mech. 503, 125 共2004兲. 95
L. Valiño and C. Dopazo, “A binomial Langevin model for turbulent mix-
88
B. L. Sawford and F. M. Guest, “Lagrangian statistical simulation of the ing,” Phys. Fluids A 3, 3034 共1991兲.
96
turbulent motion of heavy-particles,” Boundary-Layer Meteorol. 54, 147 H. Chen, S. Chen, and R. H. Kraichnan, “Probability distribution of a
共1991兲. stochastically advected scalar field,” Phys. Rev. Lett. 63, 2657 共1989兲.
89 97
M. Guingo and J. P. Minier, “A stochastic model of coherent structures for S. B. Pope, “Mapping closures for turbulent mixing and reaction,” Theor.
particle deposition in turbulent flows,” Phys. Fluids 20, 053303 共2008兲. Comput. Fluid Dyn. 2, 255 共1991兲.
90 98
I. Fouxon and P. Horvai, “Separation of heavy particles in turbulence,” J. Martín, C. Dopazo, and L. Valiño, “Joint statistics of the scalar gradient
Phys. Rev. Lett. 100, 040601 共2008兲. and the velocity gradient in turbulence using linear diffusion models,”
91
C. Meneveau, “Lagrangian dynamics and models of the velocity gradient Phys. Fluids 17, 028101 共2005兲.
tensor in turbulent flows,” Annu. Rev. Fluid Mech. 43, 219 共2011兲. 99
Y. Xia, Y. Liu, T. Vaithianathan, and L. R. Collins, “Eddy damped quasi
92
S. B. Pope, “Consistent modeling of scalars in turbulent flows,” Phys. normal Markovian theory for chemically reactive scalars in isotropic tur-
Fluids 26, 404 共1983兲. bulence,” Phys. Fluids 22, 045103 共2010兲.
93
R. L. Curl, “Dispersed phase mixing. I,” AIChE J. 9, 175 共1963兲. 100
S. B. Pope, “Ten questions concerning the large-eddy simulation of tur-
94
J. Janicka, W. Kolbe, and W. Kollmann, “Closure of the transport equation bulent flows,” New J. Phys. 6, 35 共2004兲.

Articulo 5

Uploaded by

Copyright:

Available Formats

Articulo 5

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Articulo 5

Uploaded by

Copyright:

Available Formats

PHYSICS OF FLUIDS 23, 011301 共2011兲

Simple models of turbulent flows*

I. INTRODUCTION then there would be a realistic hope for a successful statisti-

1070-6631/2011/23共1兲/011301/20/$30.00 23, 011301-1 © 2011 American Institute of Physics

lence research community is to develop and understand such M

To a good approximation, the 共suitably normalized兲 tempera- d␧ ␧2

⌫eff = ⌫ + ⌫T . 共5兲 For the homogeneous turbulence considered, ⌫T is indepen-

For given t, X+共t兲 is a random variable 共because of the

and hence Eq. 共18兲 can be reexpressed as

C. Diffusion by continuous movements

⳵ X+共t,Y兲 Several important deductions 共due to Taylor兲 can be made

position and velocity of such particles. to straight-line, ballistic motion.

共2兲 For very large time 共t / TL Ⰷ 1兲, Eq. 共24兲 yields

␴共t兲2 ⬇ 2具u2典tTL, for t/TL Ⰷ 1. 共26兲

⬇ 具u2典t, for t/TL Ⰶ 1

The last result establishes the validity of the turbulent diffu-

u+共t + ⌬t兲 − u+共t兲 = − u+共t兲

yields ␴ ⬇ 冑2⌫t; at intermediate times there is ballistic tur-

Navier–Stokes equations; but here we focus on the mean

where the Reynolds decomposition has been used to express

where 关 兴+ indicates that the quantity is evaluated at 共x , t兲

variable.27 Second, the “turbulent frequency” ␻+共t兲 is added

D. Application to turbulent flames

in the flame. Figure 12 shows the ␰-T composition space.

E. Large-eddy simulation/PDF methods

Pure co−flow In the PDF method described above, the velocity-

show that the 共normalized兲 mixing rate ␻m increases as 8

model coefficient C0 depend on R␭, i.e., C0共R␭兲. Consistency

where now we distinguish between the Kolmogorov constant

FIG. 19. 共Color online兲 Kurtosis of ␹, the logarithm of pseudodissipation,

to log-normal, or, equivalently, the quantity

⫻共⳵ui / ⳵x j + ⳵u j / ⳵xi兲. is close to Gaussian. Figure 19 shows that the kurtosis of ␹+

You might also like

where 关兴+ indicates that the quantity is evaluated at 共x , t兲