Sunteți pe pagina 1din 8

Closure Modeling of Low Dimensional Models

using LES Analogy


Haroon Imtiaz & Imran Akhtar
Department of Mechanical Engineering
NUST College of Electrical & Mechanical Engineering
National University of Science and Technology Islamabad, Pakistan 44000.

Abstract—Proper orthogonal decomposition (POD) is an im- are modeled through Boussinesq hypothesis which is simple
portant reduced order modeling technique in fluid mechanics. relationship between Reynolds stresses and velocity gradi-
The first step in POD is the collection of flow field data at different ents through the eddy viscosity. Turbulence models including
time steps from Direct numerical simulation. The singular value
decomposition extracts the most dominant modes from the col- Prandtl ’s mixing length model, K-ω model, Spalart-Allmaras
lected data. The implementation of Galerkin procedure produces model, K-∈ model and Smagronisky model are used to model
a reduced order model of considered flow field named as POD- the turbulent eddy viscosity.
ROM (ROM stands for reduced order model). In this paper, The schematic of energy distribution in turbulence modeling is
We have considered the one-dimensional Burgers equation and shown in Fig. 1a while Fig. 1b shows the energy distribution in
developed its reduce order model named as POD-G ROM. Two
closure models of POD-ROM have also been implemented on one- POD closure modeling. In turbulence modeling, Fig. 1a shows
dimensional Burgers equation. This work focuses on Smagronisky the shaded area which is modeled through different models
(POD-S) reduced order model and Dynamic subgrid-scale (POD- of RANS. Fig. 1b shows the shaded area which is modeled
D) reduced order model. We have further analyzed the effect of through different techniques of POD closure modeling. In
modes on accuracy of solution obtained through POD-D, POD- POD-G ROM, we consider M modes and reduced order model
S and POD-G ROMs. It was concluded that accuracy obtained
depends upon number of considered modes for all three ROMs is developed. The higher modes after M i.e. (M+1,M+2,...,N)
(POD-G, POD-S and POD-D). POD-S ROM performs same as contain relatively low energy but in turbulent flows discarded
POD-D ROM but needs a lot of iteration for Cs optimization. modes play a vital role. There is a need to model the
The constraint of Cs estimation through hit and trial is removed discarded modes through different modeling techniques for
in POD-S ROM. This model performs better than POD-S ROM better performance of reduced order models. The effect of
but it is computationally expensive.
discarded modes are modeled through POD closure modeling.
Closure modeling in POD-ROM is inspired from the modeling
I. I NTRODUCTION
techniques of turbulence. Few people have done work on
Realistic 3D flows are solved through direct numerical closure modeling of POD ROM [8]–[16]. This is in contrast
simulation which usually requires millions and sometimes with LES modeling of turbulence in which hundred of papers
billions degrees of freedom. Numerical simulation of fluid have already been published. There are numerous models in
flow through DNS is a computationally expensive endeavour literature including Mixing length, Smagorinsky, Variational
and its control is not practically feasible. The control in multiple scale and Dynamic subgrid-scale reduced order POD
real flows requires to update the plant output based upon model [7]. This work will focus on Smagornisky and Dynamic
feedback signal which needs numerous quick simulations. subgrid-scale reduced order models and compare the results
This motivates the engineers and scientists towards model with POD-G ROM. The symbols used for Smagornisky and
reduction techniques. Model reduction techniques based on Dynamic Subgrid-scale reduced order POD models are POD-S
proper orthogonal decomposition (POD) have successfully ROM and POD-D ROM respectively. In this work, the one-
been applied to many industrial problems. These applications dimensional Burgers equation is used for proof of concept. The
involve fluid flow control [1]–[4], image processing [5] and boundary conditions for one-dimensional Burgers equation are
pattern recognition. The simple reduced order model based on homogenous, i.e. at boundary u is set zero.
POD and Galerkin projection is called POD-G ROM, where
G stands for Galerkin projection. The POD method provides II. POD R EDUCED O RDERED M ODELS
optimal basis (modes) for dynamical system. Akhtar et al. [6] The POD reduced order model can be constructed in two
discussed that the POD-G ROM produces erroneous results steps, computation of POD modes from ensembled flow field
for turbulent flow. In this case the truncated modes do not data and Galerkin projection of Burgers equation onto a space
have significant amount of energy but have vital role in overall spanned by a number of POD modes. Let H be the Hilbert
dynamics of system [7]. space endowed with the inner product (., .)H and u(., t) ∈ H,
In turbulence modeling, closure problem arises when t ∈ [0, T ] be the state variable of a dynamical system.
Reynolds-Averaged Naiver-Stokes equation (RANS) contains Taking snapshots at different time instances t1 , t2 , ..., tN ∈
a term known as Reynolds stresses. These Reynolds stresses [0, T ] and ensembling these snapshots into a larger matrix will

E(k) Where S, and V contain the POD basis, singular values
of W and eigenvectors of W respectively. In this work, a
truncated POD basis S M =span{ϕ1 , ϕ2 , ϕ3 , ..., ϕM } is obtained
from snapshots ensembled data {u(., t1 ), u(., t2 ), ..., u(., tN )}
through SVD.
The one-dimensional Burgers equation
{
ut − vuxx + uux = f
(6)
u(x, 0) = u0 x
Turbulence
model A. POD Galerkin (POD-G) Reduced Order Model
LES The second step for obtaining a reduced order model is
Galerkin projection of Burgers equation on POD modes. This
(a) kc k reduced order model is without closure term and known as
POD-G. Taking the dot product of Eq. (5) with ϕ results in
E(m) the following equation.
{ }
ut , ϕ − vuxx , ϕ + uux , ϕ = f, ϕ ∀ϕ ∈ S M (7)
∑M
Here f =0 and u is substituted with i=1 ϕi (x)qi (t) which
results in

M ∑
M
q̇k (t) = bk + Akm qm (t) +
m=1 m=1
(8)

M
Bkmn qm (t)qn (t)
LES Inspired POD n=1
model
where bk = f, ϕk
(b) M m Akm = −νϕm,xx , ϕk
Bkmn = ϕn ϕm,x , ϕk
Fig. 1. Energy distribution versus (a) Wave number (b)Modes
B. The Smagronisky POD (POD-S) reduced order model
The dynamical system obtained from POD-G reduced order
result in model is accurate for laminar flow. In case of turbulent flows,
it is not recommended because Galerkin truncation does not
W = span{u(., t1 ), u(., t2 ), ...u(., tN )} (1)
produce accurate results despite having the most of energy
Let the time instances are tk = kδt, k = 1, 2, 3, ..., N having in captured modes [7]. The Smagronisky POD-ROM is LES
time step δt = NT−1 . The correlation matrix can be defined as inspired POD closure model which is based on concept of
energy cascading. In this method, energy of discarded modes is
Kij = (ui , uj ) (2) modeled through addition of extra term containing a constant
∫ coefficient of eddy viscosity and this model is named as POD-
where inner product is defined as (a, b) = Ω
a.b dΩ S ROM.
Kν = λν (3) ∑
M ∑
M
q̇k (t) = bk + Akm qm (t) +
νk (k=1,2,...,N) are eigenvectors , K ∈ RNXN is snapshot m=1 m=1
(9)
correlation matrix and λ1 , λ2 , λ3 , ..., λN > 0 are positive ∑
M ∑
M

eigenvalues. Bkmn qn (t)qm (t) + Dkm qm (t)


n=1 m=1
1 ∑
N
Where Akm , bk and Bkmn are same as in Eq. (8). The Dkm
ϕk (.) = √ (νk )j u(., tj ), 1 ≤ k ≤ M (4)
λk j=1 can be defined as
Dkm = (Cs ∥SuL ∥ϕm,xx , ϕk ) (10)
This is the mathematical formulation for the POD modes in
which M is number of considered modes. In other approach, In Eq. (10), the ∥SuL ∥ is Frobenius norm of ux and Cs is
it can also be found out using SVD (singular value decompo- constant. The value of Cs is estimated and it remains constant
sition). In SVD, W is decomposed as during the simulation. The Cs estimation is based on a hit
∑ and trial procedure through many numerical experiments and
W =S VT (5) computationally expensive.
C. The Dynamic subgrid-scale POD (POD-D) reduced order This second POD filter is applied to Eq. (13) which results in
model following equation
The Cs estimation is computationally expensive in POD- ˜
∂ ū
S ROM so there is a need for modeling of Cs . In closure − v∆ū
˜ + ∇.(ū
˜ū˜) + ∇.Tr = 0 (19)
∂t
modeling of turbulent flows, the Cs is estimated through
where
two methods. The first method involves estimation through fu − (ū
˜ū˜)
Tr = u (20)
available literature data. The Cs depends upon the geometry
and type of turbulence. The Cs is 0.23 for homogeneous f − ū
Tr = uu ˜ū f̄ − ū
˜ = (uū ˜ū f − uū)
˜) + (uu f̄
(21)
isotropic turbulence [17], 0.1 for turbulent channel flow [18] = Lr + Ter
and 0.15 for a turbulent mixing layer [19]. The value of Cs
= 0.1 is widely used for turbulent channels, pipe flows, and f̄ − ū
where Lr = (uū ˜ū˜) and Ter = (u
fu − uū)

backward-facing step flow [20]–[22]. It is modeled as
In second method, the Cs is calculated by using Dynamic 2
˜ ∥Dū
subgrid-scale model [23]–[25]. The Cs estimation in ROM Tr := −2(Cs δ) ˜∥Dū
˜ (22)
is inspired from this method. In POD-D ROM, POD filter is Where δe is filter radius used in second filtering operation Eq.
used instead of spatial filter which is used in conventional LES (18) and Eq. (16) result in
modeling. In POD-D ROM, the value of Cs is dynamic and
changes with time and space. Ter := −2(Cs δ)g s
g
2 ∥Dū∥Dū := −2(C δ)2 ∥Dū∥Dū (23)
Let X be a appropriate Hilbert space and X r be subspace
spanned by r POD basis functions. For all u ∈ X, the Galerkin After using Eqs. (21 & 22), The Eq. (20) results in
projection u ∈ X r is solution of following equation.
˜ 2 ∥Dū
−2(Cs δ) ˜∥Dū f̄ − ū
˜ = (uū ˜ū g
˜) − 2(Cs δ)2 ∥Dū∥Dū (24)
(u − u, ϕ) = 0 ∀ϕ ∈ X r (11)
f̄ − ū
(uū ˜ū˜)
The length scale is used in LES Dynamic subgrid-scale model Cs2 (x, t) = (25)
which can also be defined for POD-S ROM g − 2(δ)
2(δ)2 ∥Dū∥Dū ˜ 2 ∥Dū
˜∥Dū
˜
( 1 ∫ X2 (u u ) ) 12
i> i> III. RESULTS AND DISCUSSION
l> := ( ) dx (12)
X2 0 ui>,j ui>,j In this problem, following input parameters are used ∆x =
POD filter and length scale Eqs. (11-12) was used by Ref.
1
8193 , ∆t = 999 1
and ν = 10−3 . DNS solution is obtained
[11]. This is general form of Burgers equation using Implicit and second order central difference scheme.
The Newton’s method is used for solving the nonlinear term
∂u
− v∆u + ∇.(uu) = 0 (13) in the Burgers equation. The DNS results are compared with
∂t available numerical results in literature. The most relevant
The following derivation for Cs is inspired from the LES work in literature was done by Kunisch and Volkwein [27].
derivation in Ref. [26]. In POD-D ROM spatial filter of LES is Fig. 2 shows a good quantitative match between the present
replaced with POD filter. The POD filter is Galerkin projection work and the results obtained by Kunisch and Volkwein [27].
of specific number of POD modes. After applying the POD After establishing the accuracy of DNS solution, snapshots
filter to Burgers equation, the following equation is obtained are captured after each time step as solution is marched from
∂u t=0 to 1. These ensembled snapshots are 1000 which are used
− v∆u + ∇.(uu) = 0 (14) to obtain the POD modes. First 10 modes capture the most
∂t
POD filtered equation can be written as of energy of the system which is 99.14%. The energy of
discarded modes at M=10 (M stands for number of modes)
∂u
− v∆u + ∇.(ūū) + ∇.Tr = 0 (15) is 0.86% as shown in Table 1. It shows the energy distribution
∂t at various modes. It can also be deduced that cumulative
where energy increases and energy of discarded modes decreases
Tr = ∇.uu − ∇.(ūū) (16) with increase in number of modes. The energy of discarded
In POD-G, this term Eq. (16) is not modeled and assumed as modes at M=2 is 8.98% which is more than 10 times than the
zero. In POD-D, it can be written as energy of discarded modes at M=10.
These POD modes are used in POD-G ROM for POD basis
Tr := −2(Cs δ)2 ∥Dū∥Dū (17) coefficients qi (i=1,2,...,10). The solution obtained from POD-
G ROM is compared with the available work in literature
Here δ is filter radius used in first filtering operation Eq. (11)
which was done by Akhtar et al. [6]. For comparison, it is
and Cs is a function of time and space i.e. Cs (x,t). For all
needed to reconstruct the u from q’s of POD-G ROM and
u ∈ X, the second test Galerkin projection u e ∈ X R is the
POD modes. The comparison between the present work and
solution of the following equation
correspondent results obtained by Akhtar et al. [6] is made
(u − ũ, ϕ) = 0 (18) by comparing the u at t=1. Fig. 3 shows a good quantitative
TABLE I
E NERGY DISTRIBUTION IN PERCENTAGE

M Cumulative Cumulative
energy of energy of
considered modes discarded modes
2 91.01 8.98
3 95.21 4.78
4 96.82 3.18
5 97.64 2.35
6 98.17 1.83
7 98.53 1.46
8 98.80 1.20
9 98.99 1.00
10 99.14 0.86

agreement between the present and the results of Ref. [6].


Fig. 4 shows the DNS solution while the Fig. 5 presents the
u obtained from POD-G ROM. It can be inferred from Figs.
(4 & 5) that POD-G ROM does not give accurate results. The
relative error is defined as Fig. 2. DNS for one-dimensional Burgers equation at t=1
∑N
N ∗ k=1 ∥u (x, tk ) − uDN S (x, tk )∥20
1 P OD−ROM
Error = ∑N
k=1 ∥u
DN S ∥2 b) The Cs estimation is a time consuming activity
o
(26) c) POD-S assumes constant value of Cs throughout the
numerical experiment. In complex flows, constant value
Where ∥.∥0 is called L-2 norm. Fig. 6 shows the relative error of Cs will not produce an accurate solution. In POD-
of POD-G ROM at wide range of modes. It shows that the S ROM for complex flows, the Cs is required to be
relative error decreases with increase in number of modes as estimated for each mode through hit and trial. The
shown in Fig. 6 . POD-G ROM works better at higher number activity of Cs estimation for each mode increases the
of modes but it is computationally expensive and loses its time required for simulation through POD-S ROM. The
feasibility in complex flows. Fig. 9 shows the Cs variation at various modes which
Closure modeling of reduced order models through POD-S presents that Cs is relatively higher at M=2 than at
ROM have also been discussed by Refs. [6], [7]. In the same M=10. At low number of modes, larger modeling term
way, POD basis coefficients for POD-S ROM are obtained require greater value of Cs as can be seen in Fig. 9.
by using Eq. (9). The value of Cs is estimated through hit d) Least square method is not effective in Cs estimation.
and trial and optimal solution is obtained for POD-S ROM. In the above discussion, one thing is evident that Cs mea-
Two options can be used for assessment of optimial POD-S surement is time consuming activity and it is required to
ROM solution. One benchmark for POD-S ROM solution is measure Cs automatically. In turbulence modeling through
POD basis coefficients obtained through DNS projection. The LES, Germano et al. [23] has proposed the Dynamic subgrid-
value of Cs is varied until a minimum difference is observed scale model in which Cs is calculated at each point in space
between POD basis coefficients obtained from POD-S ROM and time. In model reduction through POD, the Dynamic
and through DNS projection onto POD modes. subgrid-scale ROM is proposed by Wang [7]. This model
Fig. 7 provides a good comparison of POD basis coefficients is inspired from Dynamic subgrid-scale model developed by
obtained from POD-G ROM, POD-S ROM and DNS pojec- Germano et al.[23]. The mathematical formulation for POD-D
tion. The relative error is a second option for comparison ROM has been explained above in section (II-C).
between POD-G, POD-S and POD-D ROMs. The minimum The POD-D ROM has addressed the limitations of POD-S
value of relative error indicates the optimal value of Cs . The ROM. In this study, POD-D ROM (having Cs term dependent
optimal value of Cs is 4x10−4 at M=10. The accuracy of this on space and time) has been implemented on one-dimensional
POD-S ROM is assessed with results of Akhtar et al. [6]. Burgers equation. In POD-D ROM, we have two filters level,
This quantative comparison between the present work and the first filter is applied at r (r=M i.e. number of considered modes)
results obtained by Akhtar et al. [6] is shown in Fig. 3 which while other filter is applied at R. In this case r is varied from
looks good. Fig. 8 shows the u obtained from POD-S ROM 8 to 3 while R is set at 1. The energy difference between these
and it can also be inferred from these results that POD-S ROM two levels is used for calculation of Cs .
is better than POD-G ROM (see Fig. 4 and Fig. 5). Fig. 10 shows a comparison between the results obtained from
There are certain limitations associated with POD-S ROM that POD-G, POD-S and POD-D ROMs. The relative error for
needs a little consideration. POD-G ROM is more than POD-S and POD-D ROMs which
a) Mathematical procedure for calculation of Cs is not shows that POD-S and POD-D ROMs are more accurate than
available POD-G ROM. It also shows that relative error decreases as
Fig. 3. u of one-dimensional Burgers equation at t=1 using POD-G and Fig. 5. POD-G ROM for one-dimensional Burgers equation
POD-S ROMs

Fig. 6. Relative error of POD-G ROM at wide range of modes

Fig. 4. DNS for one-dimensional Burgers equation

POD-D ROM performs well at M=7 and time requirement


number of modes increases in POD-S and POD-D ROMs. for numerical experiment is 18.60s which is approximately 50
The POD-G ROM is not predictable with increase in number times less than DNS time of numerical experiment.
of modes (M is varied from 1 to 10) as shown in Fig. 6. Fig. The POD-D ROM has addressed most of limitations of POD-
10 also presents that the accuracy of POD-D ROM is same as S ROM. POD-D ROM is based on mathematical formulation
POD-S ROM. for Cs measurement. There is no need to use hit and trial
Table 2 shows the time requirement for numerical experiments method for Cs at various modes because the mathematical
using POD-G, POD-S and POD-D ROMs. The time require- formulation automatically calculates Cs at any number of
ment for DNS solution is 915.08 sec. It can be inferred that the considered modes. The implementation of POD-D ROM is
time required for DNS is maximum while the time requirement expensive but it is more feasible than POD-G and POD-S
for POD-G ROM is minimum. The POD-G ROM results are ROMs. This is simple problem in which Cs variation with time
not very accurate so POD-G model loses it feasibility despite and space is low. In case of complex flows, it is anticipated
having low computational cost. The time requirement for that POD-D ROM should perform better than POD-S ROM
POD-S ROM is less than POD-D ROM and more than POD- due to high variation of Cs with time and space. For future,
G ROM. The time mentioned here for POD-S ROM is time it is suggested to modify the POD-D ROM or develop a new
for final simulation but time requirement for Cs optimization model which should have capability to address the both issues
through hit and trial is higher than mentioned in Table 2. The i.e. time and accuracy.
Fig. 8. POD-S ROM for one-dimensional Burgers equation

Fig. 9. Cs variation in POD-G ROM for one-dimensional Burgers equation

Fig. 7. Time evolution of POD basis coefficients qi 1,2,...,10 for DNS


projection (black), POD-G (red) and POD-S ROMs (green)

Fig. 10. The relative error of POD-G, POD-S and POD-D ROMs
TABLE II R EFERENCES
T IME FOR SIMULATION OF DIFFERENT ROMS IN SECONDS
[1] A. E. Deane and C. Mavriplis, “Low-dimensional de-
M POD-G POD-S POD-D scription of the dynamics in separated flow past thick
2 0.06 0.19 86.18 airfoils,” AIAA J., vol. 6, pp. 1222–1234, 1994.
3 0.08 0.35 43.95 [2] X. Ma and Karniadakis, “A low-dimensional model for
4 0.11 0.57 58.02
5 0.163 0.96 96.45 simulating three-dimensional cylinder flow,” J. Fluid
6 0.26 1.79 78.67 Mech, vol. 458, pp. 181–190, 2002.
7 0.39 2.88 18.60 [3] B. R. Noack, K. Afanasiev, M. Morzynski, and F.
8 0.63 4.73 56.45
9 1.12 6.55 720.81 Thiele, “A hierarchy of low-dimensional models for
10 1.24 9.56 778.27 the transient and post-transient cylinder wake,” J. Fluid
Mech, vol. 497, pp. 335–363, 2003.
[4] I. Akhtar and A. H. Nayfeh., “Model based control
IV. C ONCLUSION of laminar wake using fluidic actuation,” J. Comput.
In this paper, the POD-D ROM on one-dimensional Burgers Nonlin. Dyn., vol. 5(4), p. 041 015, 2010.
equation subjected to homogenous boundary conditions is [5] L. Sirovichs and M. Kirby, “Low-dimensional proce-
implemented. The effect of modes on POD reduced order dure for the characterization of human faces,” J. Opt.
models including POD-G, POD-S and POD-D ROMs is in- Soc. Am. A, vol. 4, no. 3, pp. 529–524, 1987.
vestigated. POD-D and POD-S ROMs perform better than [6] I. Akhtar, Z. Wang, J. Borggaard, and T. Iliescu, “A new
POD-G ROM. The constraint of Cs estimation through hit and closure strategy for proper orthogonal decomposition
trial method is removed in POD-D ROM. In POD-D ROM, reduced-order models,” Journal of Computational and
the Cs can be calculated through defined procedures at any Nonlinear Dynamics, vol. 7, no. 034503, pp. 1–6, 2012.
number of considered modes. This paper shows that relative [7] Z. Wang, “Reduced-order modeling of complex engi-
error decreases with increase in number of modes. The relative neering and geophysical flows: analysis and computa-
of POD-D ROM is same as of POD-S ROM as we vary M tions,” Tech. Rep., 2012, Virginia Polytechnic Institute
from 3 to 8. The time for numerical experiment using POD- and State University.
D ROM is still very large than POD-G ROM. In future, more [8] B. Podvin, “A proper-orthogonal-decomposition based
POD-closure models should be proposed which should address model for the wall layer of a turbulent channel flow,”
the issue of time as well as accuracy. Phys. Fluids, vol. 21, pp. 581–586, 2009.
[9] D. Rempfer and H. F. Fasel, “Dynamics of three-
ACKNOWLEDGMENT dimensional coherent structures in a flat-plate boundary
The first author would like to thank Dr Zhu Wang (Industrial layer,” J. Fluid Mech, vol. 275, pp. 257–283, 1994.
postdoc of the Institute for Mathematics and its Applications [10] W. Cazemier, R. W. Verstappen, and A. E. Veldman,
(IMA) at University of Minnesota, Twin Cities U.S.A.) for his “Proper orthogonal decomposition and low-dimensional
guidance in this research. This research is partially supported models for driven cavity flows,” Phys. Fluids, vol. 10,
by TWAS (The World Academy of Sciences for the developing no. 7, pp. 1685–1699, 1998.
world) under Research Grant Agreement (RGA) No: 11-208 [11] Z. Wang, I. Akhtar, J. Borggaard, and T. Iliescu, “Two-
RG/ENG/AS-C UNESCO FR: 3240262645. level discretizations of nonlinear closure models for
proper orthogonal decomposition,” J. Comput. Phys,
vol. 230, pp. 126–146, 2011.
[12] J. Borggaard, T. Iliescu, and Z. Wang, “Artificial viscos-
ity proper orthogonal decomposition,” Math. Comput.
Model, vol. 53, pp. 269–279, 2011.
[13] I. Akhtar, J. Borggaard, T. Iliescu, and C. J. Ribbens,
“Modeling high frequency modes for accurate low-
dimensional galerkin models,” in Proceedings of the
AIAA 39th Computational Fluid Dynamics Conference,
AIAA Paper, AIAA Paper No. 2009-4202, 2009.
[14] J. S. Smagorinsky, “General circulation experiments
with the primitive equations,” Mon. Weather Review,
vol. 91, pp. 99–164, 1963.
[15] J. G. M. Eggels, “Direct and large eddy simulation of
turbulent flow in a cylindrical pipe geometry,” Tech.
Rep., 1994, Delft University Press, Delft.
[16] L. C. Berselli, T. Iliescu, and W. J. Layton, Mathematics
of large eddy simulation of turbulent flows. Berlin:
Springer-Verlag, 2006.
[17] M. A. Domis, “Large-eddy simulation of a passive
scalar in isotropic turbulence,” J. Fluid Mech, vol. 104,
pp. 55–79, 1981.
[18] J. W. Deardorff, “A numerical study of three-
dimensional turbulent channel flow at large reynolds
number,” J. Fluid Mech, vol. 41, pp. 453–480, 1970.
[19] N. Mansour, J. H. Ferziger, and W. C. Reynolds, Large-
eddy simulation of turbulent mixing layer, Stanford
University Report No TF-11, 1978.
[20] Y. Morinishi and H. Kobayashi, “Large eddy simula-
tion of backward facing step flow,” in Int. Symp. on
Engineering Turbulence Modeling and Measurements
(Dubrovnik, Yugoslavia), 1990.
[21] F. Unger and R. Friedrich, “Large eddy simulation of
fully-developed turbulent pipe flow,” in 18th Symp. on
Turbulent Shear Flows (Munich), 1991.
[22] J. G. M. Eggels and F. T. M. Nieuwstadt, “Large eddy
simulation of turbulent flow in an axially rotating pipe,”
in 19th Symp. on Turbulent Shear Flows (Kyoto, Japan),
1993.
[23] M. Germano, U. Piomelli, P. Moin, and W. H. Cabot,
“Dynamic subgrid-scale eddy viscosity model,” Phys.
Fluids A3 J., vol. 3, pp. 1760–1765, 1991.
[24] P. Moin, K. Squires, W. Cabot, and S. Lee, “A dynamic
subgrid-scale model for compressible turbulence and
scalar transport,” Phys. Fluids, vol. 3, pp. 2746–2757,
1991.
[25] N. M. El-Hady, T. A. Zang, and U. Piomelli, “Appli-
cation of dynamic subgrid-scale model to axisymmetric
transitional boundary layer at high speed,” Phys. Fluids,
vol. 6, pp. 1299–1309, 1994.
[26] P. Sagaut, Large eddy simulation for incompress-
ible flows, Third, ser. Scientific Computation. Berlin:
Springer-Verlag, 2006, p. 556.
[27] K. Kunisch and S. Volkwein, “Control of the burgers
equation by a reduced-order approach using proper or-
thogonal decomposition,” J. Optimization Theory Appl.,
vol. 102, no. 2, pp. 345–371,

S-ar putea să vă placă și