Analysis of The Gradient-Diffusion Hypothesis in Large-Eddy Simulations Based On Transport Equations
Analysis of The Gradient-Diffusion Hypothesis in Large-Eddy Simulations Based On Transport Equations
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-2 C. B. da Silva and J. C. F. Pereira Phys. Fluids 19, 035106 2007
Reynolds-averaged Navier-Stokes RANS/LES, or in un- pology of these terms was given recently by da Silva and
steady RANS URANS/LES, have created a renewed inter- Métais.26
est in the use of SGS models based on transport equations. The goal of the present work is to assess the modeling of
Hybrid RANS/LES models appeared to cope with the high the diffusion terms either in LES using a transport equation
demands of LES in terms of grid refinement in the near wall for the SGS kinetic energy or in hybrid RANS/LES models
region. within the “LES mode.” For this purpose classical a priori
In hybrid RANS/LES using the zonal approach, the flow tests are carried out by applying a box filter to direct numeri-
is computed with different turbulent strategies in two zones. cal simulations DNS of statistically stationary forced ho-
LES is applied in one flow region, usually away from solid mogeneous isotropic turbulence. Even if the models ana-
walls, and a RANS model is used to compute the flow in the lyzed here are to be used in much complex turbulent flows
other region, usually near the wall see Hamba,15,16 Temmer- than isotropic turbulence, they have to show good results in
man et al.17. The advantage with this approach is that the this simple flow configuration if they are to succeed in more
current LES and RANS models can be used without the need complex situations.
for any particular modification. The difficulty lies in the in- Recently, in the rather different context of RANS mod-
terface separating the RANS from the LES zones, where eling, the “gradient-diffusion” assumption was put into ques-
information regarding the turbulence statistics has to be tion by Schmitt et al.27 By analyzing LES of the flow past a
somehow exchanged or generated. long square cylinder they observed that: a the pressure dif-
fusion is not negligible when compared to the turbulent dif-
In nozonal or continuous hybrid RANS/LES, the flow is
fusion, b the gradient diffusion hypothesis is not confirmed
computed with the same model in the whole domain, which
by the LES data, particularly in complex, i.e., highly three-
reverts into “RANS mode” near the walls and into “LES
dimensional, flow regions.
mode” away from walls. The switch between one or the
Finally, in order to compare the evolution of the SGS
other “modes” is done “automatically” by comparing the lo-
kinetic energy and SGS scalar variance, as well as to explore
cal mesh size with the turbulence length scale. Examples of
the difficulties associated in the development of SGS scalar
this approach are the detached eddy simulation models of
variance transport equation models, we extend this study to
Spalart et al.,18 and more recently the model of Paterson and the modeling of the SGS scalar variance. Few authors have
Peltier.19 The great advantage from this modeling strategy addressed systematically an equation to model the evolution
consists in eliminating the difficulties associated with the of the SGS scalar variance. Exceptions are the works of
interface layer. Jiménez et al.28,29
Within this context of RANS/LES coupling, a new and This article is organized as follows. In the next section,
very promising trend consists in developing transport equa- the equations governing the exact and modeled SGS kinetic
tions that turn “naturally” from LES into URANS as the energy and SGS scalar variance are described, and the mod-
filter size increases. When the grid is fine enough, the model els currently used for the diffusion terms are reviewed. In
uses a transport equation for the SGS kinetic energy. As the Sec. III we describe the DNS of isotropic turbulence used in
grid spacing increases the SGS kinetic energy becomes the this work. Sections IV–VI focus in the modeling of the dif-
grid-scale GS kinetic energy and the LES equations based fusion term. The study begins with a detailed analysis of the
on the transport of SGS energy turn into classical URANS diffusion term and its model equations based on the gradient
equations for the evolution of the total kinetic energy. Recent diffusion hypothesis Sec. IV. The analysis identifies a ma-
examples of models using this “natural” hybrid RANS/LES jor weakness arising from the SGS pressure-velocity interac-
technique are the models of Schiestel and Dejoan20 and tions Sec. V and a new model is proposed to cope with this
Langhe et al.21,22 This approach was recently extended suc- deficiency. This model is then thoroughly tested in classical a
cessfully into a Reynolds stress tensor/LES model involving priori and a posteriori LES tests Sec. VI. Finally, in Sec.
all the components of the SGS and Reynolds stresses tensor VII, the article ends with an overview of the main results and
by Chaouat and Schiestel.14 conclusions.
Arguably, the greatest challenge when modeling the
transport equation for the SGS kinetic energy is the modeling II. EXACT AND MODELED TRANSPORT EQUATIONS
of the dissipation term, where much effort has been spent
lately see de Langhe et al.,21 Chaouat and Schiestel14. For In this section we review the exact and modeled trans-
the modeling of the diffusion terms in virtually all models, port equations for the SGS kinetic energy and SGS scalar
these terms are lumped together and are modeled using a variance, with particular emphasis placed on the modeling of
classical “gradient-diffusion” hypothesis. The model con- the diffusion terms.
stants associated with the diffusion and dissipation terms are
A. Evolution of the SGS kinetic energy
either chosen as constants for the whole flow e.g.,
Schumann,5 Schmidt and Schumann,6 Dejoan and The SGS kinetic energy ii / 2 is obtained by contracting
Schiestel10, or calculated dynamically e.g., Wong,23 Ghosal the SGS stress tensor defined by ij = uiu j − ui u j, where ui is
et al.,24 Debliquy et al.9. the ith component of the velocity field, and the overlay sym-
The importance of the diffusion terms for the evolution bol — represents a spatial filtering operation. The evolu-
of the subgrid-scale kinetic energy was first recognized by tion of ii / 2 is governed by the exact equation see, e.g,
Horiuti,25 and an exhaustive description of the role and to- Piomelli and Chasnov,30 da Silva and Métais26:
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-3 Analysis of the gradient-diffusion hypothesis Phys. Fluids 19, 035106 2007
D =
xj
t
k
+
Ksgs
xj
, 5
D =
xj
t
+
sgs
xj
, 6
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-4 C. B. da Silva and J. C. F. Pereira Phys. Fluids 19, 035106 2007
TABLE I. Constant CD from Eq. 8 used in several works. “dyn.” means III. DIRECT NUMERICAL SIMULATIONS
that either the “dynamic procedure” of Germano et al. Ref. 46 or some OF ISOTROPIC TURBULENCE
other relation based on it was used to obtain the model constant. The values
shown as “Present work” correspond to the DNS with Re = 95.6 and filter A. Numerical code and data bank description
size / x = 16.
The numerical code used in the present simulations is a
Source CD standard pseudospectral code in which the temporal advance-
Schumann5 0.023 ment is made with an explicit third-order Runge-Kutta
Yoshizawa Ref. 7 0.066 scheme. The physical domain consists in a periodic box of
Yoshizawa Ref. 47 0.085 sides 2
and the simulations were fully dealiased using the
Yoshizawa and Horiuti Ref. 8 0.106 3 / 2 rule.
Horiuti Ref. 25 0.100 Three DNS of statistically steady forced homogeneous
Schmidt and Schumann Ref. 6 0.333 isotropic turbulence using N = 192 collocation points in each
Wong Ref. 23 dyn. direction were carried out. Table II lists the details of the
Ghosal et al. Ref. 24 dyn. simulations. Both the velocity and scalar large scales were
Menon et al. Ref. 32 0.050 forced in order to sustain the turbulence using the method
Fureby et al. Ref. 13 0.070 described by Alvelius,33 where the scalar forcing uses a simi-
Debliquy et al. Ref. 9 dyn. lar procedure as the one used to force a given velocity com-
Hamba Ref. 15 0.055 ponent. The forcing was imposed on three wave numbers
Dejoan and Schiestel Ref. 10 0.116 concentrated on k p = 3, for both the velocity and scalar fields.
Davidson and Peng Ref. 48 0.070 The same data bank was recently used by da Silva and
Krajnović and Davidson Ref. 11 dyn.
Pereira.2
Schiestel and Dejoan Ref. 20 0.090
After an initial transient that lasts about 10Tref, where
Kajishima and Nomachi Ref. 49 0.100
Tref = Vck p−1, Vc = P / k p1/3, and P is the forcing intensity,33
Present work 0.055
the flow reaches a state where all the turbulence quantities
for the velocity and scalar fields are statistically stationary.
The analysis was made using ten instantaneous fields taken
1/2
t = CmKsgs . 7 from this region, separated by about 0.5Tref.
Notice that L 4L11 in all simulations, where L is the
Cm, k, and are model constants, and is the size of the
box size and L11 is the integral scale, so that the size of the
implicit grid filter.
computational domain does not affect the larger flow
In high Reynolds number flows and away from solid
walls, the molecular viscosity is often neglected Menon et structures.34 In addition, to insure a good resolution of the
al.32 as it is much smaller than the turbulent eddy-viscosity, dissipative scales, we have kmax 1.5 and kmaxB 1.5 in
and therefore the relations 5 and 6 can be written as all simulations, where = 3 / 1/4 and B = / Sc1/2 are the
Kolmogorov and Batchelor microscales, respectively.
1/2 Ksgs Figures 1a and 1b show the kinetic energy and scalar
D = CD Ksgs 8
xj xj variance spectra for all the simulations, with the location of
and the filters used in the subsequent analysis. That the dissipa-
tive scales are indeed being well resolved is attested by the
1/2
sgs small upturns in the energy spectra at the end of the wave-
D = CD
Ksgs , 9
xj xj number range. For the case Re = 95.6 and Sc= 0.7, the ve-
respectively, where CD
and CD locity and scalar spectra have a −5 / 3 range, which shows
are model constants.
Table I lists the values of CD
used in several previous also that, at least for that simulation, the existence of an
works. In wall bounded flows, CD is about 0.10 e.g., see inertial range region. For each simulation, the values of the
Dejoan and Schiestel,10 Horiuti25, whereas in homogeneous skewness and flatness of the velocity derivative oscillate
and free shear flows, CD 0.05 e.g., see Yoshizawa,7 Menon around about −0.5 and 4.0, respectively, and increase slowly
32 6
et al. . For the constant CD Schmidt and Schumann ob- with the Reynolds number. The present values are very close
tained CD = 0.33. to the ones of Jimenez et al.34 for similar Reynolds numbers.
TABLE II. Details of the direct numerical simulations. Re = u / : Reynolds number based on the Taylor microscale and rms of the velocity fluctuations u;
: molecular viscosity; kmax: maximum resolved wave number; Sc= / : Schmidt number is the scalar diffusivity; L11: velocity integral scale; :
Kolmogorov microscale; B: Batchelor microscale; S and F are the skewness and flatness, respectively, of .
39.4 0.02 3.0 4.3 2.5 1.11 6.8 3.9 −0.46 +3.75 −0.55
95.6 0.006 0.7 1.8 2.1 1.24 2.8 3.3 −0.49 +4.63 −0.46
95.6 0.006 0.2 1.8 4.1 1.24 2.8 6.2 −0.49 +4.63 −0.50
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-5 Analysis of the gradient-diffusion hypothesis Phys. Fluids 19, 035106 2007
FIG. 1. Kinetic energy and scalar variance spectra for all DNS with the location of the several filters used in this work. a Kinetic energy spectra and b
scalar variance spectra.
Moreover, the mixed-derivative skewness, involving the ve-
locity and scalar gradients, is close to −0.5, which agrees 1, if
k
kc = ,
Ĝkc
k
=
with the numerical simulations of Kerr.35,36
0, otherwise.
B. The filters The box and Gaussian filters are positive and localized
In this study, the separation between grid and subgrid in the physical space, whereas the cutoff filter is oscillatory.
scales was made using three different filters: box or top-hat, In the Fourier space, on the other hand, both the Gaussian
Gaussian, and cutoff. For each filter type, four different filter and cutoff filters are localized, while the box filter is
widths were used: m = mx, with m = 2 , 4 , 8 , 16. Their loca- oscillatory.
tion in the energy spectrum is shown in Figs. 1a and 1b, As in de Borue and Orszag1 and Liu et al.,37 we ob-
where it can be noticed that the implicit cutoff wave number served that using the top-hat and the Gaussian filters the
for the filter with / x = 16 is within the inertial range results are practically independent of the filter type. For in-
region. stance, the use of the Gaussian filter tends to increase slightly
The filtering operation is defined by the integral 5 % the level of the correlations compared to the box
filter.1 For these reasons, we will show results obtained using
¯ x =
+/2
−/2
+/2
−/2
+/2
−/2
xGx − xdx , 10
the box and cutoff filters only, but we will focus mainly in
the results obtained with the box filter and leave the discus-
sion on the differences between the box and cutoff filters for
where ¯ x represents the spatially filtered variable x, the end of Secs. IV and V.
and Gx is the filter kernel, which has to verify There are a number of important reasons to make the
+/2 +/2 +/2
−/2 −/2−/2Gxdx = 1, both in the analytical and com- present analysis with the box filter instead of using the cutoff
putational discrete sense. The kernel for the box filter is filter. The most important physical quantities under study
here are localized in the physical space. In addition, the box
1 filter corresponds to the filter implicitly associated with the
, if
x −
, discretization using centred finite difference Rogallo and
Gx − = 2
Moin38 or finite volume Schumann5 codes, which are used
0, otherwise. more often in large-eddy simulations of engineering and geo-
The Gaussian filter is given by physical flows.
There are, however, more important reasons to use the
Gx − =
2
3/2
exp
2
− 6
x −
2
,
box or Gaussian filters instead of the cutoff filter. This has
to do with the positiveness of the top-hat and Gaussian
filters. As shown by Vreman et al.39 these filters always give
where is a given constant. Finally, the cutoff filter is best rise to positive subgrid-scale kinetic energy ii / 2. On the
defined in the Fourier space by other hand, and as confirmed in the present study, using a
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-6 C. B. da Silva and J. C. F. Pereira Phys. Fluids 19, 035106 2007
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-7 Analysis of the gradient-diffusion hypothesis Phys. Fluids 19, 035106 2007
FIG. 3. Joint probability density functions between: a D and D; and b D and D . Both plots correspond to the simulation with Re = 95.6, Sc= 0.7 and
were done using the box filter with filter widths equal to / x = 2 and / x = 8.
4b, we see that no model constant could be chosen in order shown in Figs. 6a and 6b. The flatness decreases with the
to get the correct agreement for all filter sizes. filter size and thus indicates a decrease in the intermittency
Figures 5a and 5b show the values of the skewness of the diffusion for large values of / x.
for all the diffusion terms. The skewnesses are always nega- Notice that the flatness can reach values close to 103,
tive, which is not surprising from the definitions of the dif- which indicates that the diffusion terms are highly intermit-
fusion terms if one recalls that the skewness of the velocity tent, particularly for small filter sizes. These values agree
derivative is negative. Figures 5a and 5b show that there very well with the values found in the literature. For the
is a big difference between the exact and modeled diffusion subgrid-scale dissipation term P from Eq. 1 Cerutti and
skewnesses for small filter sizes, but that it decreases very Meneveau40 and da Silva and Métais26 obtained flatness fac-
fast with an increase in the filter size. For instance, for tors close to 103 when considering inertial range filter
/ x = 16, this “error” is much smaller than the one obtained sizes. On the other hand the flatness of the diffusion terms is
with / x = 2. The same observation is valid for the flatness one order of magnitude lower than the subgrid-scale dissipa-
FIG. 4. Variances for the exact and modeled diffusion terms for all the simulations and filter sizes using the box filter: a D and D and b D and D .
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-8 C. B. da Silva and J. C. F. Pereira Phys. Fluids 19, 035106 2007
FIG. 5. Skewness for the exact and modeled diffusion terms for all the simulations and filter sizes using the box filter: a D and D and b D and D .
tion, but it still is quite high. As shown by da Silva and Schmidt numbers, and for all filter sizes. Notice also that this
Métais and as confirmed in the present work, it reaches 102 situation is always much more problematic for the diffusion
when the filter is at the inertial range. However, as shown by terms from the SGS kinetic energy than from the SGS scalar
Ceruti and Meneveau,40 the intermittency level increases as variance transport equation. This fact justifies the use of a
the filter size decreases, and therefore it seems plausible that dynamic procedure in order to compute CD
, as is done, e.g.,
the intermittency of the diffusion terms is 103 for dissipa- in Ghosal et al. 24
tive range filter sizes, as shown in the present work.
Again, the fact that the skewness and flatness of the C. Probability density functions for the exact
exact and modeled diffusion terms change considerably with and modeled diffusion
the filter size suggests that it may be very difficult or even
In order to give a global picture of the exact and mod-
impossible to find constant values for CD and CD within the
“gradient diffusion” hypothesis to describe accurately the eled diffusion terms, Figs. 7a and 7b show their probabil-
statistics of the exact diffusion terms for all Reynolds and ity density functions PDFs for / x = 2 and / x = 8, re-
FIG. 6. Flatness for the exact and modeled diffusion terms for all the simulations and filter sizes using the box filter: a D and D and b D and D .
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-9 Analysis of the gradient-diffusion hypothesis Phys. Fluids 19, 035106 2007
FIG. 7. Probability density functions for the exact D and D and modeled D and D diffusion terms using the box filter for the simulation with Re
= 95.6 and Sc= 0.7, and with filter widths: a / x = 2 and b / x = 8.
spectively. Notice that all the PDFs are negatively skewed, e.g., 0.055 CD 0.678. Again, this fact justifies the use of a
and that the modeled diffusion D and D, achieve higher dynamic procedure in order to compute CD
, as done by Gho-
24
magnitudes than the real D and D, and that this difference sal et al.
increases with the filter size compare the PDFs for / x However, the constants exhibit an asymptotic behavior
= 2 and / x = 8. These observations agree with the results as the filter size increases, which allows us to at least use a
described before. The same occurs for / x = 4 and / x single constant value for engineering applications of high
= 16 not shown. Finally, note that the difference between Reynolds number flows in coarse meshes and away from
the exact and modeled diffusion is always bigger for D and solid walls. For Re = 95.6 and Sc= 0.7 at / x = 16, we get
D, than for D and D , which reflects the fact that the CD
= 0.055 and CD
= 0.040, respectively, which is close
gradient-diffusion hypothesis works better for the diffusion to the values used in previous work, as shown in Table I. In
terms in the SGS scalar variance D, than for the SGS ki- particular, CD
= 0.055 is very close to the value of CD
d =
xj
1/2 Ksgs
Ksgs
xj
, 11
which leads to D = CD d , through Eq. 8. Since the mean
value of any diffusion term is zero, we have d2
= d2, and therefore the constant CD can be obtained
through
CD = D2
d2
. 12
Figure 8 shows CD obtained with Eq. 12 and CD
obtained in a similar way. Notice that the constants display FIG. 8. Constants CD and CD for all DNS and filter sizes using the box
big variations with the Reynolds number and the filter size; filter.
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-10 C. B. da Silva and J. C. F. Pereira Phys. Fluids 19, 035106 2007
FIG. 9. Joint probability density functions between a D and D and b D and D . Both plots correspond to the simulation with Re = 95.6, Sc= 0.7 and were
done using the cutoff filter with filter widths equal to / x = 2 and / x = 8.
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-11 Analysis of the gradient-diffusion hypothesis Phys. Fluids 19, 035106 2007
FIG. 10. Variances for the exact and modeled diffusion terms for all the simulations and filter sizes using the cutoff filter: a D and D and b D and D .
more insight into the weaknesses of the classical gradient- served between Dturb and Dgs/sgs, and between Dturb and
diffusion assumption. The results discussed here were ob- Dgs/sgs, decreasing only slightly with an increase of the filter
tained using the box filter and are equivalent to the ones width. This means that there is at least a partial cancellation
obtained with the Gaussian filter. Results for the cutoff filter between the diffusion by SGS turbulent transport and GS/
are discussed in Sec. V D. SGS interactions, for both the SGS kinetic energy and SGS
scalar variance, for all Reynolds and Schmidt numbers, and
A. Cancellation between SGS turbulent transport for all filter sizes considered in this work.
and GS/SGS interactions
When analyzing each separate diffusion term, we ob- B. Spectra of D and D
served that terms Dturb and Dgs/sgs from Eq. 1, and terms We now turn into the spatial three-dimensional 3D
Dturb and Dgs/sgs in Eq. 2 are highly anti-correlated. For spectrum of each separate diffusion term appearing in Eqs.
terms Dturb and Dgs/sgs a similar cancellation was observed by 1 and 2, computed on a given instantaneous field. These
Balaras and Piomelli41 in a turbulent channel flow and by da are shown in Figs. 12a and 12b, for the simulation with
Silva and Métais26 in a turbulent plane jet. Here we carried Re = 95.6 and Sc= 0.7, and for the filter size / x = 2. The
out a similar analysis into the SGS scalar variance transport smaller filter size was chosen to shown spectra of the diffu-
Eq. 2 and tested several filter sizes in order to see whether sion terms because since the a priori tests are done in the
this had any impact in the correlations. For the simulation LES grid, there are more wave numbers to describe each
with Re = 95.6 and Sc= 0.7, the computed correlation coef- spectra using this filter size. It must be stressed, however,
ficients CorrDturb , Dgs/sgs and CorrDturb , Dgs/sgs were that results qualitatively similar to the ones described here
seen to decrease slightly when the filter size increases from were obtained when considering the other larger filter sizes.
/ x = 2 to / x = 16, but stayed always above For the SGS pressure-velocity diffusion term, we define
CorrDturb , Dgs/sgs −0.90 and CorrDturb , Dgs/sgs −0.93, its spatial 3D spectrum D̂pressK by
respectively. The joint PDFs for these terms for / x = 8 are
shown in Figs. 11a and 11b. The figures show that the D̂pressK
anti-correlation comes equally from positive as well as from d̂pressKd̂pressK = K + K, 13
2
K2
negative values of Dturb and Dgs/sgs, and Dturb and Dgs/sgs.
In conclusion, strong anti-correlations are indeed ob- where d̂pressK is the Fourier transform of Dpress,
d̂pressk1,k2,k3,t =
1
2
3
−
−
−
Dpressx,y,z,te−ixk1+yk2+zk3dk1dk2dk3 , 14
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-12 C. B. da Silva and J. C. F. Pereira Phys. Fluids 19, 035106 2007
FIG. 11. Joint probability density functions between a Dturb and Dgs/sgs and b Dturb and Dgs/sgs for the simulation with Re = 95.6, Sc= 0.7 and the filter
width equal to / x = 8.
and K = k1 , k2 , k3 is the wave-number vector, of norm K almost equal both to the exact diffusion D̂K, and to the
=
K
. The spectra D̂pressK relates to the variance of the SGS molecular diffusion D̂molecK, for all the range of wave
pressure-velocity diffusion through numbers. This is easy to explain by looking into the indi-
1
2
Dpress2 =
0
D̂pressKdK. 15
vidual terms D̂turb, D̂molec and D̂gs/ssg. As we saw above,
the terms Dturb and Dgs/sgs are strongly anti-correlated and
cancel out almost exactly. The strong similarity between
Similar definitions apply to the other diffusion terms. Dturb and Dgs/sgs can be appreciated also in Fig. 12b,
Starting with Fig. 12b for the SGS scalar variance diffusion where we see that D̂turbK D̂gs/sgsK for all K. This
terms, we see that the modeled total diffusion D̂K, is leaves us with D Dmolec. Since the SGS molecular diffu-
FIG. 12. Spectra for all the exact diffusion terms from Eqs. 1 and 2 and their models defined by Eqs. 8 and 9 using the box filter. The graphs correspond
to the simulation with Re = 95.6 and Sc= 0.7, and the filter width is / x = 2: a diffusion terms for Eq. 1 and b diffusion terms for Eq. 2.
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-13 Analysis of the gradient-diffusion hypothesis Phys. Fluids 19, 035106 2007
FIG. 13. Variances for the exact diffusion terms from Eqs. 1 and 2, using the box filter for both Reynolds numbers and filter sizes considered in this work.
a Equation 1 and b Eq. 2.
sion is exactly governed by a gradient diffusion process pressure-velocity interactions do not suffice to explain the
compare Dmolec in Eq. 2 with D defined in Eq. 9, this differences between D and D, as will be shown below.
explains the strong resemblance between D and D, and First, consider the correlation between D and the sum
thus, their relatively high correlation. of all the exact diffusion terms without the SGS pressure-
Considering now the diffusion terms for the SGS kinetic velocity diffusion; i.e., CorrD , Dturb + Dvisc + Dgs/sgs. We
energy equation shown in Fig. 12a, we have an additional obtain correlation coefficients ranging from 0.91 to 0.18 with
new term: the SGS pressure-velocity diffusion Dpress. Again, increasing filter sizes, instead of the 0.68 to 0.11 obtained
SGS diffusion caused by SGS motions and GS/SGS interac- before, when considering all terms see Sec. IV A. This
tions Dturb and Dgs/sgs cancels out almost exactly, and we means that the sole addition of the SGS pressure-velocity
can see that the two terms are almost alike for all wave- diffusion term Dpress, is responsible for a decrease in the final
number ranges. As before, we may be tempted to assume that correlation between D and D of about 7% to 25% only. The
D Dpress + Dvisc, and since D is quite similar to Dvisc com- 7% figure corresponds to the filter size / x = 16, but for
pare Dvisc in Eq. 1 with D defined by Eq. 8, this could / x = 8, we have already an 11% decrease in the correla-
indicate that the SGS pressure-velocity diffusion term Dpress, tion CorrD , D caused by Dpress alone. Thus, the relatively
explains the smaller correlation between D and D, as com- small correlation obtained between D and D for / x = 16
pared to that between D and D. Indeed, the spectral shape cannot be explained by Dpress only, although it does play a
of D̂pressK is quite different than the other terms for K major role for small to intermediate filter sizes.
20, and the slope of D̂pressK for low K is much smaller To elucidate this point further, the Figs. 13a and 13b
than for the other terms. show the variances of each individual diffusion term in Eqs.
1 and 2, respectively. Theses graphs allow us to compare
C. Correlations and variances for individual the local “intensity” of each term compared to the others.
diffusion terms Shown is also the variance of the term resulting from the
In order to see whether the SGS pressure-velocity diffu- sum Dturb + Dgs/sgs, and from Dturb + Dgs/sgs, for which there
sion term is indeed responsible for the lower correlation is a strong partial cancellation, as we saw above. Note that
found between D and D as compared to D and D, we the variances are shown in logarithmic coordinates so that
started to compute the correlation coefficients between D the differences between the terms are quite high in most
and Dpress for the simulation with Re = 95.6, and for various cases.
filter sizes. The highest value was obtained for / x = 16, Again, it is better to start with the SGS scalar variance
and its value is quite small: CorrD , Dpress = 0.056. Thus, it diffusion terms in Fig. 13b. When considered separately,
seems that the weak correlation found in the modeling of the the terms Dturb and Dgs/sgs largely dominate the diffusion
SGS kinetic energy diffusion terms, compared to the SGS process. However, since they are anti-correlated, the variance
scalar variance, could be explained by a deficient modeling of their sum is much smaller than the variance of each of
of the SGS pressure-velocity diffusion, Dpress. them taken separately. For / x = 2, the sum Dturb
However, it turns out that the deficiencies of the + Dgs/sgs accounts for a variance that is more than two orders
gradient-diffusion assumption in the modeling the SGS of magnitude smaller than the variance obtained for each
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-14 C. B. da Silva and J. C. F. Pereira Phys. Fluids 19, 035106 2007
FIG. 14. Variances for the exact diffusion terms from Eqs. 1 and 2, using the cutoff filter for both Reynolds numbers and filter sizes considered in this
work. a Equation 1 and b Eq. 2.
separate term. Nevertheless, as the filter size increases, the between the diffusion terms due to SGS turbulent transport
variance of Dturb + Dgs/sgs becomes more important. and GS/SGS interactions is absent when using a cutoff filter,
Whereas for small filter sizes the SGS molecular diffusion since the highest value observed for the correlation coeffi-
dominates the total diffusion term D, for large filter sizes it cients CorrDturb , Dgs/sgs and CorrDturb , Dgs/sgs was
is the sum of terms Dturb + Dgs/sgs that commands the total smaller than 2%. Therefore, no partial cancellation exists
diffusion. This happens because for some reason, between these terms when using the cutoff filter. Conse-
Dgs/sgs2 is always slightly greater than Dturb2, al- quently, the relative local importance of the several diffusion
though their anti- correlation is very high, as seen before. terms as estimated by their variances is also quite different
This can be noticed particularly well for the filter sizes than the one observed before see Figs. 14a and 14b. The
/ x = 8 and / x = 16. We therefore arrive at the important first important observation is that the diffusion terms due to
conclusion that even if Dturb and Dgs/sgs are strongly anti- GS/SGS interactions Dgs/sgs and Dgs/sgs are negligible com-
correlated, they do not cancel out completely, such that their pared to the other terms. In addition, as expected, the diffu-
sum is still the dominant term for the total diffusion D. sion due to either viscous or molecular effects Dvisc and
Looking now into the SGS kinetic energy diffusion D in Dmolec is also much less important that the other terms.
Fig. 13a, a similar phenomenon occurs; i.e., although Dturb Indeed, Figs. 14a and 14b show that it is the diffusion due
and Dgs/sgs partially cancel out, the result from their sum is to SGS turbulent motions Dturb and Dturb that dominates
still very important, because the GS/SGS diffusion is slightly the diffusion processes, for both the SGS kinetic energy and
more important than the SGS turbulent diffusion. For small the SGS scalar variance transport equations. Moreover, it can
filter sizes, the SGS viscous diffusion Dvisc overshadows the be seen that the SGS pressure-velocity diffusion Dpress is of
sum Dturb + Dgs/sgs, but as the filter size increases Dvisc loses comparable local magnitude to the SGS turbulent diffusion
its relevance. A new feature in this equation is the existence Dturb.
of a SGS pressure-velocity diffusion Dpress. As shown in Fig. Finally, the correlation coefficients between the total dif-
13a, when / x = 16, this term is almost as important as fusion given by the gradient-diffusion model D and the
the sum Dturb + Dgs/sgs. Therefore, we conclude that improv- SGS pressure-velocity diffusion Dpress, was also computed
ing the modeling of term Dpress alone will not solve the prob- using the cutoff filter, for the simulation with Re = 95.6. The
lem of low correlation between the exact and modeled SGS values obtained are also quite small, as with the box filter.
diffusion D and D. The highest value is CorrD , Dpress = 0.053, obtained for
/ x = 16.
D. Influence of the filter type: Gaussian and cutoff
filters E. The need for an improved SGS pressure-velocity
diffusion model
The analysis described above was carried out using the
Gaussian and cutoff filters also. As before, the results ob- The results discussed above show that the classical
tained with the Gaussian and box filters are virtually equiva- gradient-diffusion assumption currently used to model the
lent. However, fundamentally different results are obtained diffusion terms in the SGS kinetic energy and SGS scalar
when using the cutoff filter. The strong anti- correlation variance transport equations has some serious limitations.
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-15 Analysis of the gradient-diffusion hypothesis Phys. Fluids 19, 035106 2007
Two reasons concur to this fact. One is due to a local loss of VI. NEW MODEL FOR THE SGS PRESSURE-
balance between the SGS turbulent transport and the VELOCITY DIFFUSION: Dpress
GS/SGS interactions. A second one, and this only for the
This section presents and discusses a new model for the
SGS kinetic energy transport equation, has to do with the
SGS pressure-velocity diffusion: Dpress . We start by present-
deficient modeling of the SGS pressure-velocity diffusion.
ing the model equations and then analyze in a priori tests the
This explains the greater inadequacy of the gradient-
performance of the new model using correlations, one-point
diffusion hypothesis when applied to the SGS kinetic energy statistics, and spectra. The results show that the new model
than to the SGS scalar variance equations. represents a substantial improvement over the classical
In order to improve this situation a possible strategy to gradient-diffusion hypothesis defined by Eq. 8. Finally,
improve the modeling of the diffusion term would be the LES of decaying isotropic turbulence are carried out and
following: i remove the gradient-diffusion model altogether confirm the improvement obtained with the new model.
and keep the exact viscous and molecular diffusion terms
Dvisc and Dmolec since they can be computed exactly with A. Model formulation
the available information; ii model the SGS turbulent trans-
Inspection of the term Dpress shows that the variable
port and GS/SGS interactions terms together, i.e., find a new
within the divergence operator resembles the subgrid-stresses
model for terms Dturb + Dgs/sgs and Dturb + Dgs/sgs; iii in the
tensor, except that instead of the velocity product uiu j, we
SGS kinetic energy equation, find a new model for the SGS find the product of one velocity component with the pressure
pressure-velocity diffusion term Dpress. The remainder of this field pu j. Defining the first-order tensor
article will focus on the last task outlined above; i.e., the
modeling of Dpress. M pj = p̄u j − pu j , 16
One question that one might raise at this moment is the exact SGS pressure-velocity diffusion term is given by
whether it is worthy to improve the modeling of the SGS Dpress = / x jM pj . Due to the close resemblance between the
pressure-velocity diffusion Dpress alone, bearing in mind that ij and M pj tensors, one may be tempted to try to apply the
at least in isotropic turbulence, it will only improve the local gradient or Clark’s relation to model M pj .
correlation for the total diffusion by about 25% at most. Indeed, many studies have shown that the gradient
To answer this, recall that in one-point closures based model leads to very high correlations between the exact and
upon transport equations for the Reynolds stresses, the role modeled subgrid-scale stresses, usually higher than 80%
of the pressure strain terms pui / x j is to redistribute e.g., see Liu et al.,37 de Borue and Orszag,1 Meneveau and
kinetic energy among all the components of the Reynolds Katz43. Even if the pressure field has a very different physi-
stresses tensor see, e.g., Tennekes and Lumley42. It is easy cal nature than the velocity field, particularly due its intrinsic
to see that the role of the SGS pressure-velocity interactions nonlocal nature Davidson44, it is reasonable to assume that
is similarly to redistribute energy among all the elements of locally it will not behave very differently than any other flow
the SGS stresses tensor ij, and therefore this term should be variable, and thus provides a good candidate for modeling
more important in anisotropic turbulent shear flows than in with the Clark’s model.
the isotropic turbulent flow used here. Moreover, as shown Applying the gradient or Clark’s model to M pj , we get
by da Silva and Métais,26 the term Dpress is particularly im- p̄ ū j
portant near very large flow vortices present in many turbu- M p
j = C pm
2
, 17
xk xk
lent shear flows that are absent from forced isotropic turbu-
lence where only much smaller structures “worms” exist. where Cpm is a model constant and is the filter size. The
Thus, one may assume that in flows dominated by these large modeled SGS pressure-velocity diffusion can now be ob-
vortices, such as in wakes, jets, or in separated shear flows, tained through
the term Dpress will be much more important to the dynamics
of the SGS kinetic energy than in the present isotropic tur-
bulence configuration.
Dpress = Cpm2
p̄ ū j
x j xk xk
. 18
The justification to test a new model for Dpress in the The model constant Cpm will be determined below using
present forced homogeneous isotropic turbulence data set the DNS data bank, but before that it is important to assess
comes from the fact that it is generally agreed that isotropic the performance of the new model.
turbulence represents the “simplest” possible turbulent flow
configuration, and therefore it is highly unlikely that any new B. Correlation with exact SGS pressure-velocity
model will ever succeed in more complicated flows if it fails diffusion
in this simplest of all canonical flow configurations. The correspondence between the exact and modeled
For these reasons, it was decided to develop and test a SGS pressure-velocity diffusion can be assessed in Fig. 15.
new model for the SGS pressure-velocity interactions using The figure shows the correlation coefficients between Dpress
the present isotropic turbulence simulations. The model was
and Dpress for the two Reynolds numbers and all the four
tested in classical a priori tests in isotropic turbulence as filter sizes considered in this study. As can be seen, the cor-
well in a posteriori tests LES of decaying isotropic turbu- relations are quite high. The correlations tend to decrease
lence. This is the subject of the following section. slightly as the filter size increases, particularly for the higher
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-16 C. B. da Silva and J. C. F. Pereira Phys. Fluids 19, 035106 2007
FIG. 15. Correlation coefficient and joint probability density functions be-
tween the exact and the modeled SGS pressure-velocity diffusion terms,
Dpress and Dpress .
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-17 Analysis of the gradient-diffusion hypothesis Phys. Fluids 19, 035106 2007
FIG. 18. Constant Cpm from Eq. 18 for the two Reynolds numbers and all
FIG. 17. Spectra for the exact and modeled SGS pressure-velocity diffusion filter sizes considered in this work.
terms Dpress and Dpress , for the simulation with Re = 95.6 and filter width
/ x = 2.
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-18 C. B. da Silva and J. C. F. Pereira Phys. Fluids 19, 035106 2007
VII. CONCLUSIONS
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-19 Analysis of the gradient-diffusion hypothesis Phys. Fluids 19, 035106 2007
GS/SGS interactions terms, i.e., find a new model for large-eddy simulation,” Int. J. Heat Fluid Flow 23, 413 2002.
12
J. Deardorf, “A numerical study of three-dimensional turbulent channel
Dturb + Dgs/sgs and Dturb + Dgs/sgs. flow at large Reynolds numbers,” J. Fluids Eng. 95, 429 1973.
iii Model the SGS pressure-velocity diffusion term 13
C. Fureby, G. Tabor, H. G. Weller, and A. D. Gosman, “Differential sub-
Dpress, in the SGS kinetic energy equation. grid stress models in large eddy simulations,” Phys. Fluids 9, 3578
1997.
14
Related to task iii outlined above, a new model was B. Chaouat and R. Schiestel, “A new partially integrated transport model
for subgrid stresses and dissipation rate for turbulent developing flows,”
proposed to compute the SGS pressure-velocity diffusion Phys. Fluids 17, 065106 2005.
Dpress. The model constant arising from the new model 15
F. Hamba, “An attempt to combine large eddy simulation with the k −
proved to be quite insensitive to changes in Reynolds num- model in a channel-flow calculation,” Theor. Comput. Fluid Dyn. 14, 323
ber and filter size, which shows that is dispenses with more 2001.
16
F. Hamba, “A hybrid RANS/LES simulation of turbulent channel flow,”
complicated dynamical procedures for its computation. The Theor. Comput. Fluid Dyn. 16, 387 2003.
new model was assessed both in a priori and in a posteriori 17
L. Temmerman, M. Hadziabdic, M. Leschiner, and K. Hanjaliac, “A hy-
tests giving better results than classical approach in isotropic brid two-layer URANS-LES approach for large eddy simulation at high
turbulence. Reynolds numbers,” Int. J. Heat Fluid Flow 26, 173 2005.
18
P. R. Spalart, W.-H. Strelets, and S. R. Allmaras, “Comments on the fea-
Although the present results seem promising, it is clear sibility of LES for wings, and on a hybrid RANS/LES approach,” in
that new tests in inhomogeneous turbulent shear flows where Advances in DNS/LES Greyden, Columbus, OH, 1997.
19
the mean diffusion terms are not zero, should be carried out E. G. Paterson and L. J. Peltier, “Detached-eddy simulation of high-
Reynolds-number beveled-trailing-edge boundary layers and wakes,” J.
with the new model. It is expected that the new model will
Fluids Eng. 127, 897 2005.
lead to improved results in computations in these flows, 20
R. Schiestel and A. Dejoan, “Towards a new partially integrated transport
since the Dpress diffusion term is known to play a much more model for coarse grid and unsteady turbulent flow simulations,” Theor.
important role in these flows than in the present isotropic Comput. Fluid Dyn. 18, 443 2005.
21
C. de Langhe, B. Merci, and E. Dick, “Hybrid RANS/LES modeling with
turbulence configuration. an approximate renormalization group. I: Model development,” J. Turbul.
613, 1 2005.
22
C. de Langhe, B. Merci, K. Lodefier, and E. Dick, “Hybrid RANS/LES
ACKNOWLEDGMENTS modeling with an approximate renormalization group. II: Applications,” J.
Turbul. 614, 1 2005.
23
V. C. Wong, “A proposed statistical-dynamic closure method for the linear
The authors would like to acknowledge Professor Julian or nonlinear subgrid-scale stresses,” Phys. Fluids A 4, 1080 1992.
24
Andrzej Domaradzki for comments and suggestions made S. Ghosal, T. Lund, P. Moin, and K. Akselvol, “A dynamic localisation
during the revision of the manuscript during his stay in model for large-eddy simulation of turbulent flows,” J. Fluid Mech. 286,
229 1995.
Lisbon in November 2006, and an anonymous referee who 25
K. Horiuti, “Large-eddy simulation of turbulent channel flow by one-
suggested a way to compute the diffusion constants de- equation modeling,” J. Phys. Soc. Jpn. 54, 2855 1985.
26
scribed in Sec. IV D. This work was supported by the Min- C. B. da Silva and O. Métais, “On the influence of coherent structures
istério da Ciência e Tecnologia under Contract No. upon interscale interactions in turbulent plane jets,” J. Fluid Mech. 473,
103 2002.
SFRH/BPD/26756/2006. 27
F. Schmitt, B. Merci, E. Dick, and C. Hirsh, “Direct investigation of the
k-transport equation for a complex turbulent flow,” J. Turbul. 421, 100
1
V. de Borue and S. Orszag, “Local energy flux and subgrid-scale statistics 2003.
in three dimensional turbulence,” J. Fluid Mech. 366, 1 1998. 28
C. Jiménez, L. Valino, and C. Dopazo, “A priori and a posteriori tests of
2
C. B. da Silva and J. C. F. Pereira, “On the local equilibrium of the subgrid models for scalar transport,” Phys. Fluids 13, 2433 2001.
29
subgrid-scales: The velocity and scalar fields,” Phys. Fluids 17, 108103 C. Jiménez, F. Ducros, B. Cuenot, and B. Bédat, “Subgrid scale variance
2005. and dissipation of a scalar field in large eddy simulations,” Phys. Fluids
3
O. Métais, “Direct and large-eddy simulations of coherent vortices in 13, 1748 2001.
30
three-dimensional turbulence: Geophysical and industrial applications,” in U. Piomelli and J. R. Chasnov, “Large eddy simulations: Theory and
Identification of Coherent Structures, edited by J. P. Bonnet Springer- applications,” in Turbulence and Transition Modeling, edited by M. Hall-
Verlag, Berlin, 1996. back, D. S. Henningson, A. V. Jhoannson, and P. H. Alfredssoon Kluwer,
4
E. Garnier, O. Métais, and M. Lesieur, “Synoptic and frontal-cyclone scale Dordrecht, 1996.
instabilities in baroclinic jet flows,” J. Atmos. Sci. 55, 1316 1998. 31
H. Kang and C. Meneveau, “Universality of large eddy simulation param-
5
U. Schumann, “Subgrid-scale model for finite difference simulations of eters across a turbulent wake behind a cylinder heated cylinder,” J. Turbul.
turbulent flows on plane channels and annuli,” J. Comput. Phys. 18, 376 332, 1 2002.
1975. 32
S. Menon, P. K. Yeung, and W. W. Kim, “Effect of subgrid models on the
6
H. Schmidt and U. Schumann, “Coherent structure of the convective computed interscale energy transfer in isotropic turbulence,” Comput.
boundary layer derived from large-eddy simulations,” J. Fluid Mech. 200, Fluids 25, 165 1996.
511 1989. 33
K. Alvelius, “Random forcing of three-dimensional homogeneous turbu-
7
A. Yoshizawa, “A statistically-derived subgrid model for the large-eddy lence,” Phys. Fluids 11, 1880 1999.
simulation of turbulence,” Phys. Fluids 25, 1532 1982. 34
J. Jiménez and A. Wray, “On the characteristics of vortex filaments in
8
A. Yoshizawa and K. Horiuti, “A statistically-derived subgrid-scale kinetic isotropic turbulence,” J. Fluid Mech. 373, 255 1998.
35
energy model for the large-eddy simulation of turbulent flows,” J. Phys. R. M. Kerr, “High-order derivative correlations and the alignment of
Soc. Jpn. 54, 2834 1985. small-scale structures isotropic numerical turbulence,” J. Fluid Mech.
9
O. Debliquy, B. Knapen, and D. Carati, “A dynamic subgrid-scale model 153, 31 1985.
36
based on the turbulent kinetic energy,” in Direct and Large-Eddy Simula- R. M. Kerr, “Velocity, scalar and transfer spectra in numerical turbulence,”
tions IV, edited by B. J. Geurts, R. Friedrich, and O. Métais Kluwer J. Fluid Mech. 211, 309 1990.
37
Academic, Dordrecht, 2001, pp. 89–96. S. Liu, C. Meneveau, and J. Katz, “On the properties of similarity subgrid-
10
A. Dejoan and R. Schiestel, “LES of unsteady turbulence via a one- scale models as deduced from measurements in a turbulent jet,” J. Fluid
equation subgrid-scale transport model,” Int. J. Heat Fluid Flow 23, 398 Mech. 275, 83 1994.
38
2002. R. Rogallo and P. Moin, “Numerical simulation of turbulent flows,” Annu.
11
S. Krajnović and L. Davidson, “A mixed one equation subgrid model for Rev. Fluid Mech. 16, 99 1984.
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp
035106-20 C. B. da Silva and J. C. F. Pereira Phys. Fluids 19, 035106 2007
39 45
B. Vreman, B. Geurts, and H. Kuerten, “Realizability conditions for the C. B. da Silva and J. C. F. Pereira, “The effect of subgrid-scale models on
turbulent stress tensor in large-eddy simulation,” J. Fluid Mech. 278, 351 the vortices obtained from large-eddy simulations,” Phys. Fluids 16, 4506
1994. 2004.
40 46
S. Cerutti and C. Meneveau, “Intermittency and relative scaling of M. Germano, U. Piomelli, P. Moin, and W. Cabot, “A dynamic subgrid-
subgrid-scale energy dissipation in isotropic turbulence,” Phys. Fluids 10, scale eddy viscosity model,” Phys. Fluids A 3, 1760 1991.
47
928 1998. A. Yoshizawa, “Eddy-viscosity-type subgrid-scale model with a variable
41
E. Balaras and U. Piomelli, “Subgrid-scale energy budgets in wall layer,” Smagorinsky coefficient and its relationship with the one-equation model
Bull. Am. Phys. Soc. 39, 1969 1994. in large-eddy simulation,” Phys. Fluids A 3, 2007 1991.
42 48
H. Tennekes and J. L. Lumley, A First Course in Turbulence MIT Press, L. Davidson and S. Peng, “Hybrid LES-RANS modeling: a one-equation
Cambridge, 1972. SGS model combined with a k − w model for predicting recirculating
43
C. Meneveau and J. Katz, “Scale inveriance and turbulence models for flows,” Int. J. Numer. Methods Fluids 43, 1003 2003.
49
large-eddy simulation,” Annu. Rev. Fluid Mech. 32, 1 2000. T. Kajishima and T. Nomachi, “One-equation subgrid scale model using
44
P. A. Davidson, Turbulence, An Introduction for Scientists and Engineers dynamic procedure for the energy dissipation,” Trans. ASME, J. Appl.
Oxford University Press, New York, 2004. Mech. 73, 368 2006.
Downloaded 19 Mar 2007 to 193.136.128.14. Redistribution subject to AIP license or copyright, see http://pof.aip.org/pof/copyright.jsp