Sunteți pe pagina 1din 11

International Journal of Heat and Fluid Flow 42 (2013) 94–104

Contents lists available at SciVerse ScienceDirect

International Journal of Heat and Fluid Flow


journal homepage: www.elsevier.com/locate/ijhff

A dynamic global-coefficient mixed subgrid-scale model for large-eddy


simulation of turbulent flows
Satbir Singh a, Donghyun You a,b,⇑
a
Department of Mechanical Engineering, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA 15213, United States
b
Department of Mechanical Engineering, Pohang University of Science and Technology, Pohang 790-784, South Korea

a r t i c l e i n f o a b s t r a c t

Article history: A dynamic global-coefficient mixed subgrid-scale eddy-viscosity model for large-eddy simulation of tur-
Received 12 May 2012 bulent flows in complex geometries is developed. In the present model, the subgrid-scale stress is decom-
Received in revised form 13 February 2013 posed into the modified Leonard stress, cross stress, and subgrid-scale Reynolds stress. The modified
Accepted 15 February 2013
Leonard stress is explicitly computed assuming a scale similarity, while the cross stress and the sub-
Available online 13 March 2013
grid-scale Reynolds stress are modeled using the global-coefficient eddy-viscosity model. The model
coefficient is determined by a dynamic procedure based on the global-equilibrium between the sub-
Keywords:
grid-scale dissipation and the viscous dissipation. The new model relieves some of the difficulties asso-
Subgrid-scale model
Large-eddy simulation
ciated with an eddy-viscosity closure, such as the nonalignment of the principal axes of the subgrid-scale
Turbulence simulation stress tensor and the strain rate tensor and the anisotropy of turbulent flow fields, while, like other
Global-coefficient model dynamic global-coefficient models, it does not require averaging or clipping of the model coefficient
Scale-similarity model for numerical stabilization. The combination of the global-coefficient eddy-viscosity model and a scale-
Navier–Stokes equations similarity model is demonstrated to produce improved predictions in a number of turbulent flow
simulations.
Ó 2013 Elsevier Inc. All rights reserved.

1. Introduction the dynamic model is applied to complex flow configurations in


which there are no homogeneous flow directions. A number of no-
In large-eddy simulation (LES) of turbulent flows, large scale vel approaches have been proposed (e.g., the dynamic localization
fluid motions are directly computed, while the influence of fluid model by Ghosal et al. (1994) and the Lagrangian dynamic model
motions of scales smaller than the computational cell size, is mod- by Meneveau et al. (1996)) to address this issue. However, the iter-
eled using a subgrid-scale (SGS) model. The predictive capability of ative solution procedure to solve an integral equation (Ghosal
an LES method, therefore, relies on the performance of an SGS et al., 1994) or the averaging of the model coefficient along a path-
model employed. Although many different types of SGS models line (Meneveau et al., 1996) demand non-trivial efforts in imple-
are available, the dynamic Smagorinsky model of Germano et al. mentation and computation.
(1991) has been most widely used. In the dynamic Smagorinsky The shortcoming of the dynamic Smagorinsky models, which
model, the model coefficient is dynamically determined as a func- are based on the local-equilibrium hypothesis, was overcome by
tion of space and time using the scale-similarity concept and the You and Moin (2007, 2009) and Park et al. (2006). They proposed
local-equilibrium hypothesis (i.e., an equilibrium between the dynamic procedures for determining the model coefficient of an
SGS dissipation and the viscous dissipation at the same physical eddy-viscosity model developed by Vreman (2004) utilizing a
location at a given instant). Although the dynamic model coeffi- ‘‘global equilibrium’’ hypothesis that assumes a global balance
cient vanishes where flow is laminar or fully resolved, it can cause between the SGS dissipation and the viscous dissipation. In the
numerical instability since its value often becomes negative and/or global-equilibrium approaches (You and Moin, 2007, 2009; Park
highly fluctuates in space and time. et al., 2006), the model coefficient is determined to be globally
The numerical instability has been remedied by procedures uniform in space but to vary in time and does not require any ad
such as an averaging of the model coefficient over statistically hoc numerical stabilization or clipping operations. Even with
homogeneous directions or an ad hoc clipping procedure. The non-zero uniform model coefficients, the global-coefficient models
numerical stabilization procedure becomes complicated when still guarantee vanishing eddy viscosity in the laminar and fully
resolved flow regions by the inherent advantage of Vreman’s
eddy-viscosity model in which vanishing eddy viscosity for
⇑ Corresponding author. Tel.: +1 412 268 6808; fax: +1 412 268 3348. various laminar shear flows is theoretically guaranteed. The
E-mail address: dhyou@cmu.edu (D. You).

0142-727X/$ - see front matter Ó 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.ijheatfluidflow.2013.02.008
S. Singh, D. You / International Journal of Heat and Fluid Flow 42 (2013) 94–104 95

global-coefficient model of You and Moin (2007) has shown good 1


sij  dij skk ¼ 2mT Sij ; ð3Þ
predictive capability and superior stability to the conventional 3
dynamic Smagorinsky-type local-coefficient models for turbulent where mT is the eddy viscosity. The isotropic part of the stress tensor
channel flow (You and Moin, 2007), turbulent flow over a circular 1
d s is absorbed into the pressure. In the present study, an eddy-
3 ij kk
cylinder (You and Moin, 2007), and turbulent flow in a lid-driven viscosity model developed by Vreman (2004), is employed to deter-
cavity (Shetty et al., 2010). mine mT:
Like other eddy-viscosity SGS models, the global-coefficient SGS
model assumes that the principal axes of the SGS stress tensor are mT ¼ C v Pg ; ð4Þ
aligned with those of the resolved strain rate tensor, a result which where
is not supported by direct numerical simulation data (Bardina, sffiffiffiffiffiffiffiffiffiffiffiffi
1983). Furthermore, although the global equilibrium coheres with
g
Bgb
the conservation principles, it may not be fully capable of repres- P ¼ ;
a kl a kl
entating local characteristics of SGS turbulence.
The SGS stress tensor can be decomposed into two parts: a re- Bgb ¼ bg11 bg22  bg12 bg12 þ bg11 bg33  bg13 bg13 þ bg22 bg33  bg23 bg23 ;
solved part which can be explicitly computed from the resolved X
3 ð5Þ
flow field and a part which cannot be explicitly computed (Zang bgij ¼ D2m a
 mi a
 mj ;
m¼1
et al., 1993). The inconsistency associated with the assumption of
j
@u
the alignment of the SGS stress and the resolved strain rate tensors a ij ¼ ;
can be alleviated by a model that explicitly computes the resolved
@xi
part of the SGS stress tensor, such as the scale similarity model Dm is the grid-filter width in the m-direction, ( )g denotes a grid-fil-
proposed by Bardina (1983). It was found that the scale similarity ter-level quantity, and Cv is the model coefficient. A novel feature of
model alone does not dissipate SGS energy, and therefore needs to the model that makes it superior to the Smagorinsky model with a
be combined linearly with an eddy-viscosity model to dissipate constant coefficient, is that the kernel Pg becomes zero for laminar
SGS energy. Zang et al. (1993) combined the scale similarity model or fully resolved flow regions where vanishing eddy viscosity is
of Bardina (1983) with the dynamic Smagorinsky model of Ger- expected.
mano et al. (1991) to formulate the dynamic Smagorinsky mixed
model. The model was validated for three-dimensional unsteady 2.1. Dynamic global-coefficient model (DGM)
turbulent flow in a lid-driven cubical cavity (Zang et al., 1993).
The mixed model produced a better prediction compared to the You and Moin (2007) considered a transport equation for
dynamic Smagorinsky model of Germano et al. (1991). It was found Lii ð¼ T ii  s
^ii Þ, where sii ð¼ ui ui  u  i Þ; T ii ð¼ ud
i u ^  i Þ, and ðbÞ
^
i u
i ui  u
that although the magnitude of the local model coefficient be- denotes a test-filtered quantity:
comes smaller for the mixed model, the model coefficient still be- ( ! )
@Lii @ @u d i @u
iu
^ i u
^ i
comes negative and highly fluctuates in time and space. ¼ ð u diui u
j  u ^ i u ^ j Þ  2ð d
^ i u j p  u
u ^ j b
pÞ þ m   2ð sd  i  T ij u
ij u
^ i Þ
@t @xj @xj @xj
In the present new model, the dynamic global-coefficient model |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
C
of You and Moin (2007) is coupled with the scale-similarity model !
@u di @u i @ u ^ i @ u ^ i d b
proposed by Bardina (1983). The newly proposed dynamic global-  2m  þ 2ð sij Sij  T ij S ij Þ; ð6Þ
@xj @xj @xj @xj |fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl}
coefficient mixed model retains the favorable features of the dy- |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} eSGS
namic global-coefficient model with additional advantages from em

the incorporation of a scale-similarity model. The global-coeffi- where C is the redistribution term, and em and eSGS are the viscous
cient model does not necessitate ad hoc numerical stabilization and the SGS dissipation terms, respectively. The viscous dissipation
while guarantees vanishing eddy viscosity in laminar or fully-re- (em) and the subgrid-scale dissipation (eSGS) are balanced when they
solved flow regions. The model is expected to alleviate the assump- are averaged over the flow field since volume averaged time-deriv-
tion of the alignment of the principal axes of the SGS stress tensor ative term and redistribution terms are negligible in Eq. (6) (Park
and the resolved strain rate tensor while improving the capability et al., 2006; You and Moin, 2007). The balance between volume
for local characteristics of SGS turbulence. averaged viscous dissipation and subgrid-scale dissipation terms
The paper is organized as follows. A mathematical formulation lead to an equation for the model constant Cv:
of the global-coefficient mixed model is presented in Section 2. Re- D E
sults from simulations of decaying isotropic turbulence, turbulent m ad
ij a
^ ij a
 ij  a ^ ij
flow in a straight channel on a regular grid and a multi-resolution Cv ¼    ; ð7Þ
2
grid, and turbulent flow in a lid-driven cavity, are discussed in Sec- Pd
g b b
Sij Sij  Pt S ij S ij
tion 3, followed by concluding remarks in Section 4.
where Pg is defined in Eq. (5) and
2. Mathematical formulation 1 b
T ij  T kk dij ¼ 2C v Pt S ij ;
3sffiffiffiffiffiffiffiffiffiffiffiffi
The grid-filtered governing equations for incompressible flow
Btb
are as follows: Pt ¼ ;
a^ kl a^ kl
i @ u
@u i u
j 
@p @2ui @ sij Btb ¼ bt11 bt22  bt12 bt12 þ bt11 bt33  bt13 bt13 þ bt22 bt33  bt23 bt23 ;
þ ¼ þm  ; ð1Þ ð8Þ
@t @xj @xi @xj @xj @xj
X
3
b2 a
j
@u btij ¼ ^ ^
¼ 0; ð2Þ D m mi amj ;
@xj m¼1
^ j
@u
where u  are the filtered velocity and pressure, respectively. m is
i ; p a^ ij ¼ ;
@xi
the kinematic viscosity and sij ð¼ ui uj  u
i u
 j Þ is the SGS stress tensor.
The anisotropic part of the SGS stress tensor sij is modeled by an b is the test-filter width in the m-direction and ( )t denotes a test-
D m
eddy-viscosity model: filter-level quantity. The width of the test filter is taken as twice the
96 S. Singh, D. You / International Journal of Heat and Fluid Flow 42 (2013) 94–104

b ¼ 2D). The model coefficient is dynami-


width of the grid filter ( D
cally determined from the instantaneous flow field (You and Moin,
2007).

2.2. Dynamic global-coefficient mixed model (DGMM)

Using the decomposition method proposed by Germano (1986):


 þ u0 ;
u¼u ð9Þ

sij can be decomposed as follows:


   
sij ¼ ½ui uj  ui uj  þ ½ ui u0j þ u0i uj  ui u0j þ u0i uj 
h i
þ u0i u0j  u0i u0j ; ð10Þ
Fig. 1. Energy spectrum in forced isotropic turbulence at Rek = 100. Solid lines, the
where present DNS; dashed lines, k5/3.

Lm  
ij ¼ ui uj  ui uj ; ð11Þ
Cm
ij
 i u0j þ u0i u
¼ ðu  i u0j þ u0i u
 j Þ  ðu  j Þ; ð12Þ
Rm 0 0 0 0
ij ¼ ui uj  ui uj : ð13Þ
m
Lm m
ij ; C ij , and Rij are referred to as the modified Leonard stress,
the modified cross stress, and the modified SGS Reynolds stress,
respectively. All three stress terms are invariant under Galilean
transformation (Germano, 1986). The modified Leonard stress
can be calculated using the resolved flow field, while the other
two terms involve unresolved terms and require a model for
closure.
In the mixed model proposed by Zang et al. (1993), the modified
cross and Reynolds stresses were modeled using the Smagorinsky
model. In the present new mixed model, the modified cross and
Reynolds stresses are modeled using the eddy-viscosity model of
Vreman with the dynamic global-coefficient closure procedure of
You and Moin (2007). (a)
SGS stress tensors at the grid and test filter levels are expressed
as follows:

1 dij m
sij  dij skk ¼ Lmij  L  2C v Pg Sij ; ð14Þ
3 3 kk

1 dij b
T ij  dij T kk ¼ LTij  LT  2C v Pt S ij ; ð15Þ
3 3
where the modified Leonard stress at the test-filter level is given as
follows:

^d
LTij ¼ u iu
^^ ^^
^ j  u i uj : ð16Þ

Substituting Eqs. (14) and (15) into Eq. (6) and taking the vol-
ume average of Eq. (6) assuming global equilibrium results in the
following form of the dynamic model coefficient: (b)
       
m ad
ij a
 ij  a ^ ij  Lm d
^ ij a dij m
L S ij  L T

dij T b
L S ij Fig. 2. (a) Resolved kinetic energy as a function of time and (b) energy spectrum as
ij 3 kk ij 3
1 a function of wavenumber at three different times for decaying isotropic turbulence
Cv ¼     : at Rek = 100. Solid lines, DGMM; dashed lines, DGM; symbols, filtered DNS.
2
Pd
g b b
Sij Sij  Pt S ij S ij

ð17Þ

Table 1
3. Results and discussion
Grid parameters for large-eddy simulations of turbulent channel flow. Lx, Ly, and Lz are
the streamwise, wall-normal, and spanwise domain sizes, respectively. d is the
The predictive capability of the new dynamic global-coefficient channel half height. Nx(y, z) and Dx(y, z)+ are the number of grid points and resolution
mixed model (DGMM) compared to that of the dynamic global- in wall units, respectively.
coefficient model (DGM), is evaluated in large-eddy simulations ReT Lx Ly Lz Nx Ny Nz Dx + Dy+ Dz+
of decaying isotropic homogeneous turbulence, turbulent flow in
180 4pd 2d 4
pd 48 48 48 47 0.32–18 16
a planar channel on a regular grid and a multi-resolution grid, 3
395 2pd 2d pd 64 64 64 39 0.52–60 19
and turbulent flow in a lid-driven cubical cavity.
S. Singh, D. You / International Journal of Heat and Fluid Flow 42 (2013) 94–104 97

(a)
Fig. 5. Temporal history of dynamic model coefficients (Cv) for channel flow at
Res = 395 on a regular grid obtained using a fourth-order numerical scheme. Solid
line, DGMM; dashed line, DGM.

(b)
Fig. 3. Profiles of (a) the mean streamwise velocity and (b) rms velocity fluctuations
for channel flow at Res = 180 on a regular grid. Solid lines, DGMM; dashed lines,
DGM; symbols, non-filtered DNS (Moser et al., 1999).

(a)

(a)

(b)
Fig. 6. Profiles of the mean eddy viscosity (mt) for channel flow at (a) Res = 180 and
(b) Res = 395 on a regular grid obtained using a fourth-order numerical scheme.
Solid lines, DGMM; dashed lines, DGM.

3.1. Decaying isotropic homogeneous turbulence

Decaying isotropic homogeneous turbulence is simulated using


direct numerical simulation (DNS) and large-eddy simulation (LES)
with DGMM and DGM. Initial conditions for decaying isotropic tur-
bulence are generated by performing DNS of forced isotropic tur-
(b) bulence. The incompressible Navier–Stokes equations are solved
Fig. 4. Profiles of (a) the mean streamwise velocity and (b) rms velocity fluctuations
in a periodic box with sides of length 2p. A pseudospectral code,
for channel flow at Res = 395 on a regular grid. Solid lines, DGMM; dashed lines, which is dealiased by a combination of spherical truncation and
DGM; symbols, non-filtered DNS (Moser et al., 1999). phase shifting, is employed. A forcing term (Machiels, 1997)
98 S. Singh, D. You / International Journal of Heat and Fluid Flow 42 (2013) 94–104

(a) (a)

(b) (b)
Fig. 7. Profiles of (a) the mean modeled shear stress s12 and (b) the mean total Fig. 8. Profiles of (a) the mean modeled shear stress s12 and (b) the mean total
shear stress (sum of resolved and modeled shear stresses) for channel flow at shear stress (sum of resolved and modeled shear stresses) for channel flow at
Res = 180 on a regular grid. Solid lines, DGMM; dashed lines, DGM; symbols, non- Res = 395 on a regular grid. Solid lines, DGMM; dashed lines, DGM; symbols, non-
filtered DNS (Moser et al., 1999). filtered DNS (Moser et al., 1999).

^f ðkÞ ¼ P u
^ i ðkÞ
ð18Þ Res is based on the wall friction velocity us and the half-height of
i 2
^
k62 jui ðkÞj the channel d. Periodic boundary conditions are used in the
streamwise and spanwise directions. A fixed pressure gradient is
is imposed to retain Rek = u0 k/m = 100, where u0 is root-mean-square imposed to drive the flow in the streamwise direction. Computa-
velocity fluctuations, k is the Taylor microscale, and f(=0.5) is a pre- tional domain sizes and grid parameters for simulations at both
scribed dissipation rate that determines the energy injection rate. A Reynolds numbers are shown in Table 1. Uniform grid spacings
grid of 128 computational cells in each direction is used for DNS. are used in the streamwise and spanwise directions, while a hyper-
Fig. 1 shows an energy spectrum obtained from the DNS flow field bolic tangent stretching function is used in the wall-normal
after 20 eddy turnover times. This flow field is used as the initial direction:
condition for simulations of decaying turbulence.
Large-eddy simulations are performed on a 323 computational
grid with DGMM and DGM. Initial conditions for LES are generated tanhðcð1  2j=N 2 Þ
yðjÞ ¼  ; j ¼ 0; 1; 2 . . . . . . ; N2 ; ð19Þ
by filtering the DNS velocity field using the standard Gaussian filter tanhðcÞ
function in the wavenumber space. For both DNS and LES, the
time-step size is chosen to satisfy the Courant–Friedrichs–Lewy where N2 is the number of grid points in the wall-normal direction
(CFL) condition of 1.0. Fig. 2a shows the evolution of the normal- and c is the grid stretching parameter and is fixed at 2.4. Simula-
ized resolved kinetic energy (j(t)/j0) as a function of normalized tions are performed using the CFL number of 1.0.
time (t0/j0), while Fig. 2b shows the comparison of energy spectra Simulations are performed using two different numerical
obtained using LES and DNS at t0/j0 = 0, 0.27, and 0.54. The en- schemes, a fourth-order finite-difference scheme of Morinishi
ergy decay rates obtained using DGMM and DGM are predicted et al. (1998) and a second-order accurate structured-grid finite-
to be similar to each other and to agree well with DNS data volume scheme of Singh and You (2011). Different filter functions
(Fig. 2a). As shown in Fig. 2b, compared to DGM, DGMM predicts were employed for the two numerical schemes. Details of the filter
slightly enhanced dissipation at higher wave numbers. Overall, functions are provided in Appendix A. The simulations are initiated
both models predict an energy spectrum close to DNS data. with a log law velocity profile which is randomly perturbed with a
magnitude of 5% of the streamwise velocity and allowed to evolve
3.2. Turbulent channel flow on a regular grid for 30d/us. Thereafter, each case is integrated for another 60d/us to
produce a statistically converged solution. Flow statistics are ob-
Large-eddy simulations of turbulent flow through a straight tained by averaging the flow field in time and over horizontal
channel are performed at Res = 180 and 395. The Reynolds number planes.
S. Singh, D. You / International Journal of Heat and Fluid Flow 42 (2013) 94–104 99

(a)
(a)

(b)
Fig. 10. Profiles of (a) the mean streamwise velocity and (b) rms velocity
(b) fluctuations in channel flow at Res = 395 on a regular grid obtained using a
second-order numerical scheme. Solid lines, DGMM; dashed lines, DGM; dotted
Fig. 9. Profiles of the mean eddy-viscosity stress (solid lines) and the mean lines: constant coefficient Smagorinsky model; symbols, non-filtered DNS (Moser
modified Leonard stress (dashed lines) predicted by DGMM for turbulent channel et al., 1999).
flow at (a) Res = 180 and (b) Res = 395.

Figs. 3–9 show solutions obtained with the fourth-order finite-


difference scheme. Figs. 3 and 4 show the streamwise mean veloc-
ity and root-mean-square of velocity fluctuations predicted by
DGMM and DGM at Res = 180 and 395, respectively. LES solutions
are compared with the non-filtered DNS data of Moser et al.
(1999). DNS data filtered on the same resolution as that employed Fig. 11. A multi-resolution grid for simulations of turbulent channel flow.
for present LES, are not available. However, it is shown in the pre-
vious work of the corresponding author (Bose et al., 2010) that the
near wall peak values of the mean streamwise velocity and rms Table 2
velocity fluctuations from the filtered DNS solution only slightly Grid parameters for large-eddy simulation of turbulent channel flow at Res = 180 on a
multi-resolution grid. Lx, Ly, and Lz are the streamwise, wall-normal, and spanwise
decrease. domain sizes in each grid block, respectively. d is the channel half height. Nx(y, z) and
On the grid resolution shown in Table 1, DGMM is found to bet- Dx(y, z)+ are the number of grid points and resolution in wall units, respectively.
ter predict the mean streamwise velocity and the root-mean-
Lx Ly Lz Nx Ny Nz Dx+ Dy+ Dz+
square fluctuations of the streamwise velocity than DGM. DGMM
and DGM show similar predictions for v þ þ Block 1 pd 2d 4
pd 12 48 3 47 0.32–18 24
rms and wrms . Errors in total 3

flow rates predicted using DGMM are less than 1% at both Block 2 2pd 2d 4
3 pd 24 64 64 47 0.24–14 12

Res = 180 and 395 (Figs. 3 and 4). These predictions are better than Block 3 pd 2d 4
3 pd 12 48 32 47 0.32–18 24

results presented in the literature for similar grid resolution (Park


et al., 2006; Vreman, 2004; You and Moin, 2007).
Figs. 5 and 6 show dynamic model coefficients and eddy viscos- et al. (1999). In DGMM, the modeled stress is composed of the
ity, respectively, obtained using DGM and DGMM at Res = 395. As modified Leonard stress Lm 12 and the eddy-viscosity stress
expected, DGMM results in a lower model coefficient and lower 2mT S12 . At both Reynolds numbers, DGMM predicts higher levels
eddy viscosity compared to those obtained using DGM. The model of the modeled stress (Figs. 7 and 8a) leading to improved compar-
coefficient is nearly constant in time and slightly fluctuates about a ison against DNS data for the total turbulent stress as shown in
mean value of 0.075 for DGM, and about a mean value of 0.057 for Figs. 7 and 8b.
DGMM. A lower model coefficient and eddy viscosity for DGMM Fig. 9 shows the mean modified Leonard stress Lm 12 and the
indicate less reliance of the model on the eddy-viscosity SGS stress. eddy-viscosity stress predicted by DGMM. At both Reynolds num-
Figs. 7 and 8 show the mean modeled stress s12 and the total bers, the peaks of the modified Leonard and eddy-viscosity stresses
stress (sum of the modeled stress and resolved stress) at are of similar magnitudes, while the location of the peak magni-
Res = 180 and 395, respectively, along with DNS data of Moser tude of the eddy-viscosity stress is more towards the wall.
100 S. Singh, D. You / International Journal of Heat and Fluid Flow 42 (2013) 94–104

(a) (a)

(b) (b)
Fig. 12. Profiles of (a) the mean streamwise velocity and (b) rms velocity Fig. 13. Profiles of (a) the mean streamwise velocity and (b) rms velocity
fluctuations for channel flow at Res = 180 on a coarse block of a multi-resolution fluctuations for channel flow at Res = 180 on a fine block of a multi-resolution
grid shown in Fig. 11. Solid lines, DGMM; dashed lines, DGM; symbols, non-filtered grid shown in Fig. 11. Solid lines, DGMM; dashed lines, DGM; symbols, non-filtered
DNS (Moser et al., 1999). DNS (Moser et al., 1999).

Fig. 10 shows the mean streamwise velocity and root-mean- should respond properly to the change in local filter resolution.
square of velocity fluctuations predicted by DGMM, DGM, and This is particularly important for SGS models that use a spatially
the constant coefficient Smagorisnky model (Smagorinsky, 1963) uniform model coefficient, such as DGM and DGMM. In order to as-
at Res = 395 using a second-order finite-volume method (Singh sess the capability of DGMM for LES on inhomogeneously refined
and You, 2011). Grid resolution is kept the same as the one used grids, large-eddy simulations of turbulent flow through a straight
for the fourth-order method (table I). Again, LES solutions are com- channel at Res = 180 are performed using a multi-resolution grid
pared with non-filtered DNS data of Moser et al. (1999). For the as shown in Fig. 11. The computational time-step size is chosen
Smagorinsky model, a model coefficient of 0.17 is used. On the grid to satisfy the CFL number of 1.0.
resolution shown in table I, DGMM is found to better predict the The computational domain size and grid parameters used for
mean streamwise velocity and root-mean-square of the stream- simulations are summarized in Table 2. A central block with finer
wise velocity fluctuations than DGM. Predictions of v þ þ
rms and wrms grid resolution is surrounded by two outer blocks with coarser grid
are slightly better by DGM. Both DGM and DGMM predict better resolution. Hexahedral cells in the adjacent blocks are joined by a
the mean streamwise velocity and root-mean-square fluctuations single layer of unstructured cells at the interface. Grid stretching is
of all three velocity components, when compared with those pre- applied in the wall-normal directions using the hyperbolic tangent
dicted by the constant coefficient Smagorinsky model. It was also function described in Eq. (19). A second-order accurate unstruc-
reported by You and Moin (2007) that DGM performs better than tured-grid finite-volume code (You et al., 2008) is employed. Peri-
the dynamic Smagorinsky model for prediction of turbulent chan- odic boundary conditions are used in the streamwise and spanwise
nel flow. directions. A fixed pressure gradient is imposed to drive the flow in
Due to test filtering, DGM and DGMM require about twice more the streamwise direction. The flow field is initialized using a uni-
computational time than the constant coefficient Vreman and con- form velocity profile with 10% random perturbations in all three
stant coefficient Smagorinsky models for calculating the eddy vis- velocity components, and is allowed to evolve for 120d/us. Thereaf-
cosity. However, both DGM and DGMM require about the same ter, each case is integrated for another 120d/us to produce a statis-
computational time as that of the dynamic Smagorinsky model. tically converged solution for the mean streamwise velocity and
root-mean-square fluctuations of three velocity components.
3.3. Turbulent channel flow on a multi-resolution grid Figs. 12 and 13 show the streamwise mean velocity and root-
mean-square velocity fluctuations predicted using DGMM and
In complex flow configurations, the computational grid is often DGM at Res = 180 on the coarse grid block (block 1 in Fig. 11)
inhomogeneously refined, either due to limitations of the grid gen- and fine grid block (block 2 in Fig. 11), respectively. LES predictions
eration technique or due to the necessity of a variable grid resolu- are compared with the non-filtered DNS data of Moser et al.
tion among various flow regions. In large-eddy simulation of (1999). On both coarse and fine blocks, DGMM is found to produce
turbulent flows, the difference in the grid spacing translates di- slightly improved mean velocity profiles compared to those pre-
rectly into the difference in LES filter resolution. An SGS model dicted by DGM, as shown in Figs. 12 and 13a. DGMM is found to
S. Singh, D. You / International Journal of Heat and Fluid Flow 42 (2013) 94–104 101

(a) (a)

(b)
(b)
Fig. 15. Profiles of the mean eddy-viscosity shear stress (solid lines) and the
Fig. 14. Profiles of the mean modeled shear stress s12 for channel flow at Res = 180 modified Leonard stress (dashed lines) in DGMM for channel flow at Res = 180 on
on (a) the coarse block and (b) the fine block of a multi-resolution grid shown in (a) the coarse block and (b) the fine block of a multi-resolution grid shown in
Fig. 11. Solid lines, DGMM; dashed lines, DGM. Fig. 11.

significantly improve the prediction of the streamwise velocity


fluctuations as shown in Figs. 12 and 13b, while DGMM slightly
under-predicts fluctuations of the cross-stream velocity compo-
nents (v þ þ
rms and wrms ) than DGM.
However, it is found that close to 20% errors in the skin friction
are produced by both models. The errors in the skin friction pre-
dicted on the present multi-resolution grid are similar to or a little
worse than those reported by other researchers in large-eddy sim-
ulations of turbulent channel flow using dynamic Smagorinsky
model and in zonal detached eddy simulations of developing tur-
bulent boundary layer flow on a single-block grid of which resolu-
tion is similar to that of the coarse block of the multi-resolution
grid (Meyers and Sagaut, 2007; Deck et al., 2011). In the present
simulations, transition of mesh resolution in the wall-normal
direction and spanwise direction is found to adversely affect the
prediction of the skin friction.
Fig. 14 shows the mean modeled stresses s12 predicted by
DGMM and DGM on the coarse and fine grid blocks. The modeled
Fig. 16. Configuration for simulations of flow in a lid-driven cubical cavity.
stress is composed of the modified Leonard stress and the eddy-
viscosity stress in DGMM, while the modeled stress corresponds
to the eddy-viscosity stress in DGM. Compared to DGM, DGMM
is found to predict significantly higher magnitudes of the modeled to be less sensitive to the change of grid resolution. The magnitude
stress. The magnitude of the modeled stress predicted by DGM is of the modified Leonard stress is higher than the magnitude of the
predicted to be relatively insensitive to the change of grid resolu- eddy-viscosity stress throughout the domain. The modified Leon-
tion, while the magnitude of the modeled stress predicted by ard stress is found to properly adjust to the resolution change. As
DGMM clearly reflects the change of grid resolution (i.e., coarse the grid resolution is changed from coarse resolution (block 1) to
resolution (Fig. 14a) to fine resolution (Fig. 14b)). fine resolution (block 2), the magnitude of the modified Leonard
Fig. 15 shows the mean modified Leonard stress Lm 12 and the stress is decreased. This feature implies that the present mixed
mean eddy-viscosity stress predicted by DGMM on the coarse model has an improved capability for adjusting to the grid resolu-
and fine grid blocks. As in DGM, the eddy-viscosity stress is found tion change compared to DGM.
102 S. Singh, D. You / International Journal of Heat and Fluid Flow 42 (2013) 94–104

(a)
Fig. 17. Profiles of the mean velocity components along symmetry axes obtained
using DGMM (solid lines) and DGM (dashed lines) for flow in a lid-driven cubical
cavity at Re = 10,000. Symbols, experimental measurements of Prasad and Koseff
(1989).

(b)
Fig. 19. Profiles of the mean Reynolds stress obtained using DGMM (solid lines) and
DGM (dashed lines) for flow in a lid-driven cubical cavity at Re = 10,000. (a) Along x/
h = 1 axis and (b) along y/h = 1 axis. Symbols, experimental measurements of Prasad

(a) and Koseff (1989).

To cluster grid lines near the walls, the following grid stretching
function is employed in directions normal to the wall (Tannehill
et al., 1984):

ðb þ 2aÞ½ðb þ 1Þ=ðb  1ÞðgaÞ=ð1aÞ  b þ 2a


y¼h ; ð20Þ
ð2a þ 1Þf1 þ ½ðb þ 1Þ=ðb  1ÞðgaÞ=ð1aÞ g

where y and g are the normal coordinates in the physical and com-
putational space, respectively, and a and b are grid stretching
parameters, set equal to 0.5 and 1.2, respectively. The computa-
tional grid consists of 64 cells in each direction. The Reynolds num-
ber Re ¼ U2h
m is based on the length of each side of the cavity 2h and
the velocity of the top lid U. A second-order accurate structure-grid
finite-volume code (Singh and You, 2011) is employed. The compu-
tational time-step size is adjusted during the simulations to satisfy
(b) the CFL number condition of 1.0. The flow field is first allowed to
evolve for t = 1000h/U and flow statistics are then collected from
Fig. 18. Profiles of root-mean-squared velocity fluctuations obtained using DGMM t = 1000h/U to t = 1500h/U. The present LES predictions are com-
(solid lines) and DGM (dashed lines) for flow in a lid-driven cubical cavity at
pared with experimental measurements of Prasad and Koseff
Re = 10,000. (a) urms along x/h = 1 axis and (b) vrms along y/h = 1 axis. Symbols,
experimental measurements of Prasad and Koseff (1989). (1989) on a symmetry plane at z/h = 1.
Fig. 17 shows comparisons of the mean velocity profiles pre-
dicted by the present LES and experimentally measured velocity
profiles along the horizontal (y/h = 1) and vertical (x/h = 1) symme-
try axes. LES predictions agree quite well with experimental mea-
3.4. Turbulent flow in a lid-driven cubical cavity surements. The location of the center of a large scale motion, as
indicated by the zero crossing of the horizontal and vertical veloc-
The predictive capability of DGMM for fully three-dimensional ity components near the center of the cavity, is well predicted by
inhomogeneous flow is assessed in simulations of turbulent flow both DGMM and DGM. Predictions of the two models are slightly
in a lid-driven cubical cavity at Re = 10,000. The geometry and different in the vicinity of the right-hand-side wall. DGMM pre-
boundary conditions for the flow configuration are illustrated in dicts a slightly thicker wall jet parallel to the right-hand-side wall
Fig. 16. compared to DGM.
S. Singh, D. You / International Journal of Heat and Fluid Flow 42 (2013) 94–104 103

the right-hand-side wall (x/h = 1) is also better predicted by


DGMM.
In Fig. 19a and b, the mean Reynolds stress predicted by the
present LES is compared with experimental data along the vertical
(x/h = 1) and horizontal (y/h = 1) symmetry axes, respectively. For
LES, the Reynolds stress is a sum of the resolved stress
v
(hu  ihv
 i  hu  i) and modeled stress (s12). As shown in Fig. 19a,
the peak Reynolds stress near the bottom wall and its decay away
from the wall is better predicted with DGMM. Near the moving
wall (y/h = 1), DGMM predicts higher stress compared to DGM.
Along the horizontal symmetry axis (y/h = 1), the stress is nearly
zero except inside the wall jet near the right-hand-side wall (x/
h = 1). Within the wall jet, DGMM slightly over-predicts the Rey-
nolds stress while DGM slightly under-predicts it.
The k2 eigenvalue (Jeong and Hussain, 1995) is computed for
both models to visualize the flow field. Fig. 20 shows iso-surface
of k2 = 0.5, colored by the vorticity magnitude. Qualitatively, both
models predict vortex cores to be present primarily in the lower-
right corner of the cavity, where the vertical-wall jet meets the
bottom-wall of the cavity. Vorticity magnitudes are found to be
grossly similar to each other for the two models.

4. Concluding remarks

A new dynamic global-coefficient mixed model (DGMM) is


developed for large-eddy simulation of turbulent flows. In the
newly proposed model, the eddy-viscosity model proposed by Vre-
man (2004) is combined with the scale-similarity model of Bardina
(1983) to form a mixed model. The dynamic procedure for deter-
mining the model coefficient is based on the global-equilibrium
between the SGS dissipation and the viscous dissipation as devel-
oped by You and Moin (2007). The present model does not require
averaging of model coefficient along homogeneous directions,
which is a highly desirable feature for LES in complex flow config-
urations. The present model also ensures a positive model coeffi-
cient avoiding any ad hoc clipping procedures. The new model
has been evaluated in large-eddy simulations of decaying homoge-
neous isotropic turbulence, turbulent flow in a straight channel on
a regular grid and a multi-resolution grid, and turbulent flow in a
lid-driven cubical cavity.
For decaying homogeneous isotopic turbulence, DGMM and
DGM have been found to equally well predict the energy decay rate
and energy spectrum of DNS. For highly anisotropic turbulent
flows such as in a channel and a lid-driven cubical cavity, DGMM
clearly demonstrated a superior capability for prediction of the
mean velocity and velocity fluctuations. This is likely due to the
incorporation of a scale-similarity term which improves the mod-
Fig. 20. Iso-surfaces of k2 colored with the vorticity magnitude predicted by (a) eling of anisotropy in SGS turbulence.
DGM (b) DGMM for flow in a lid-driven cubical cavity at Re = 10,000. Vorticity DGMM also shows improved adaptability to local characteris-
magnitudes in the range from 10 to 10 and the k2 value of 0.5 are shown. tics of the computational configuration. For example, in large-eddy
simulation of turbulent channel flow on a multi-resolution grid,
the modeled stress predicted by DGMM is found to clearly reflect
the change of grid resolution, while the modeled stress predicted
by DGM is found to be relatively insensitive to the change of grid
In Fig. 18a and b, root-mean-square fluctuations of velocity
resolution.
components predicted by the present LES are compared with
experimental data. Both DGMM and DGM predict a spike in urms
near the moving wall (y/h = 1), as also observed in the experiments. Acknowledgments
Near the moving wall, experimentally measured urms does not be-
come zero at the wall. This could be due to difficulties in measuring The corresponding author acknowledges the support of the
velocities very close to the moving surface. Near the bottom wall Army Research Office under Grant No. W911NF1010348, with Dr.
(y/h = 0), similar to the experiments, both models predict two dis- Frederick Ferguson as the program manager and the support of
tinct peaks in urms. While moving away from the bottom wall to- the National Research Foundation of Korea and the Brain Korea
wards the center of the cavity, DGMM predicts a faster decay in 21 Program of the Korea Research Foundation. Computing time
urms compared to DGM, giving a better agreement with experimen- was provided by the National Science Foundation TeraGrid under
tal measurements. As shown in Fig. 18b, the peak value of vrms near Grant No. ASC110028 through the Pittsburgh Supercomputing
104 S. Singh, D. You / International Journal of Heat and Fluid Flow 42 (2013) 94–104

Center. The authors are grateful to Dr. Chumakov for the provision where d = jxi  xi1j and d+ = jxi+1  xij. A three-dimensional filter-
of a pseudo spectral code. Authors would also like to thank under- ing operation is performed by a tensor product of the one-dimen-
graduate student Kyung Jae Lee for help with computational mesh sional filters.
generation.
References
Appendix A
Bardina, J., 1983. Improved Turbulence Models based on Large Eddy Simulation of
Homogeneous, Incompressible, Turbulent Flows. Ph.D. Thesis, Stanford
 with a
The filtering operator applied to a grid-filtered function / University, Stanford.
filter kernel G can be expressed as follows: Bose, S.T., Moin, P., You, D., 2010. Grid-independent large-eddy simulation using
explicit filtering. Physics of Fluids 22, 105103.
Z X
L Deck, S., Weiss, P.-E., Pamies, M., Garnier, E., 2011. Zonal detached eddy simulation
 0
 0 Þdx ¼   jDÞ;
/ðxÞ ¼ Gðx0 ; xÞ/ðx wj /ðx ðA:1Þ of a spatially developing flat plate turbulent boundary layer. Computers and
j¼K Fluids 48, 1–15.
Germano, M., 1986. A proposal for a redefinition of the turbulent stresses in the
where wj are the discrete filter weights over the interval filtered Navier–Stokes equations. Physics of Fluids 29 (7), 2323–2324.
Germano, M., Piomelli, U., Moin, P., Cabot, W.H., 1991. A dynamic subgrid-scale
x0 2 [x  LD, x + KD], and D is a measure of the filter width. eddy viscosity model. Physics of Fluids 3 (7), 1760–1765.
Ghosal, S., Lund, T.S., Moin, P., 1994. A dynamic localization model for large-eddy
A.1. Filtering operators used in Section 3.2 simulation of turbulent flows. Journal of Fluid Mechanics 282, 1–27.
Jeong, J., Hussain, F., 1995. On the identification of a vortex. Journal of Fluid
Mechanics 285, 69–94.
Fourth-order commuting discrete filters proposed by Vasilyev Machiels, L., 1997. Predictability of small-scale motion in isotropic fluid turbulence.
et al. (1998) are used. The filtering operation is performed for a Physical Review Letters 79 (18), 3411–3414.
Meneveau, C., Lund, T.S., Cabot, W.H., 1996. A Lagrangian dynamic subgrid-scale
uniform grid spacing in the mapped space. Discrete one-dimen-
model of turbulence. Journal of Fluid Mechanics 319, 353–385.
sional filtering operators used in simulations of turbulent channel Meyers, J., Sagaut, P., 2007. Evaluation of Smagorinsky variants in large-eddy
flow on a regular grid are given as follows: simulations of wall-resolved plane channel flows. Physics of Fluids 19, 095105.
Morinishi, Y., Lund, T.S., Vasilyev, O.V., Moin, P., 1998. Fully conservative higher
 ¼ 0:0095/
/  i2 þ 0:24/
 i3  0:081/  i1 þ 0:66/
 i þ 0:24/
 iþ1 order finite difference schemes for incompressible flow. Journal of
Computational Physics 143, 90–124.
 
 0:081/iþ2 þ 0:0095/iþ3 ; Moser, R.D., Kim, J., Mansour, N.M., 1999. Direct numerical simulation of turbulent
channel flow. Physics of Fluids 11, 943–945.
Park, N., Lee, S., Lee, J., Choi, S., 2006. A dynamic subgrid-scale eddy-viscosity model
b
 ¼ 0:031/
/  i2 þ 0:28/
 i3 þ 0:0/  i1 þ 0:5/
 i þ 0:28/
 iþ1 þ 0:0/
 iþ2 with a global coefficient. Physics of Fluids 18 (125109).
Prasad, A.K., Koseff, J.R., 1989. Reynolds number and end-wall effects on a lid-driven
 iþ3 :
 0:031/ cavity flow. Physics of Fluids A 1 (2), 208–218.
Shetty, D.A., Fisher, T.C., Chunekar, A.R., Frankel, S.H., 2010. High-order
incompressible large-eddy simulation of fully inhomogeneous turbulent
A three-dimensional filtering operation is performed by a tensor flows. Journal of Computational Physics 229 (23), 8802–8822.
product of the one-dimensional filters. Singh, S., You, D., 2011. A multi-block ADI finite-volume method for incompressible
Navier–Stokes equations in complex geometries. Journal of Computational
Physics 230 (19), 7400–7417.
A.2. Filtering operators used in Section 3.3 Smagorinsky, J., 1963. General circulation experiments with the primitive
equations. Monthly Weather Review 91, 99.
Flow through a channel resolved with a multi-resolution grid is Tannehill, J.C., Anderson, D.A., Pletcher, R.H., 1984. Computational Fluid Mechanics
and Heat Transfer. McGraw-Hill Book Company, New York.
simulated using a second-order accurate finite-volume method.
 and /^ Vasilyev, O.V., Lund, T.S., Moin, P., 1998. A general class of commutative filters for
Three dimensional filtering operations for both /  are per- LES in complex geometries. Journal of Computational Physics 146, 82–104.
formed using a box filter of which width is identical to twice the Vreman, A.W., 2004. An eddy-viscosity subgrid-scale model for turbulent shear
flow: algebraic theory and applications. Physics of Fluids 16 (10), 3670–3681.
local mesh size.
You, D., Ham, F., Moin, P., 2008. Discrete conservation principles in large-eddy
simulation with application to separation control over an airfoil. Physics of
A.3. Filtering operators used in Sections 3.2 and 3.4 Fluids 20, 101515.
You, D., Moin, P., 2007. A dynamic global-coefficient subgrid-scale eddy-viscosity
model for large-eddy simulation in complex geometries. Physics of Fluids 19
For simulations of channel flow in Section 3.2 and simulations (6), 065110.
of flow inside a three-dimensional lid-driven cavity in Section 3.4, You, D., Moin, P., 2009. A dynamic global-coefficient subgrid-scale model for large-
the following discrete filters are employed: eddy simulation of turbulent scalar transport in complex geometries. Physics of
Fluids 21 (4), 045109.
 þ  þ 2 þ  Zang, Y., Street, R.L., Koseff, J.R., 1993. A dynamic mixed subgrid-scale model and its
 ¼ ð2d  d Þ /
/  i1 þ ðd þ d Þ /  i þ ð2d  d Þ /
 iþ1 ; application to turbulent recirculating flows. Physics of Fluids 5 (12), 3186–
  þ þ
6d 6d d 6d 3196.
 þ  þ 2 þ 
 ¼ ð2d  d Þ /
b
/  i1 þ ðd þ d Þ /  i þ ð2d  d Þ /
 iþ1 ;
  þ þ
4d 8d d 4d

S-ar putea să vă placă și