Multi-scale :

Turbulence modeling
Computational Fluid Dynamics

* Covered partly in the exam.

Primary references :
• Pope, Turbulent flows
• 허남건, CFD class
• 이정일, 2013년 대한기계학회 추계학술대회
 Energy flow by region
Energy cascade behind an airfoil

Source: Van Dyke, Album of Fluid Motions

Source: Applied Computational Fluid Dynamics, Lecture 10 –Turbulence models, André Bakker (2002- 2005), Fluent Inc.(2002)
 Turbulent flows : DNS, LES & RANS
To simulate turbulent flows : Based on resolved scales (~ filter threshold)

l h = l/ReL3/4

Direct numerical simulation (DNS)

Large eddy simulation (LES)

Reynolds-averaged Navier-Stokes (RANS)

Source: Applied Computational Fluid Dynamics, Lecture 10 –Turbulence models, André Bakker (2002- 2005), Fluent Inc.(2002)
 Turbulent flows : DNS, LES & RANS
To simulate turbulent flows : Level of low-pass filtering
DNS (Direct numerical simulation) : no filter ~ experiment
LES (Large eddy simulation) : moderate filter size
(U)RANS (Reynolds-averaged Navier-Stokes) : relatively large filter
  

DNS – LES – RANS로 갈수록 mean flow field에 근접

 Direct numerical simulation (DNS)
No (artificial) model + Resolving all scales with grid

Use the fundamental governing equations in textbooks

Ex) Navier-Stokes equation
No model or model parameter except for physical ones

Grid size ≈ Smallest physical scale

Solution quality depends significantly on

Numerical methods : Accuracy (+ stability)
Grid resolution : Truncation error

Main advantage
Low error ~ Experiment

Main disadvantage
(Very) High computational cost : Need relatively long time
DNS of a turbulent channel flow (betw'n 2 flat plates)
A classic study by Kim, Moin & Moser (1987 JFM)
It shows that a DNS study can virtually replace an experiment
Particle trace and mean velocity profile

Courtesy of S. Hahn, SNU

 Examples of DNS
Heat transfer, combustion, multi-phase flows
Setting Δx for DNS of wall-bounded flows
Grid size condition for near-wall grid resolution
 ~ Smallest (Kolmogorov) length scale
∆  ∆
∆    1
 
Universal

Smallest (Kolmogorov) length scale away from wall

In fully-developed wall-bounded flows
Kolmogorov length scale      /
for    50 (semi-empirical)
 : 0.41
  
Length in wall unit :     
 

Velocity in wall unit :  

 Introduction to RANS

Computational Fluid Dynamics

Reynolds-averaged Navier-Stokes (RANS)
modeled
Reynolds decomposition :  , !   , ! " # , ! resolued -

Instantaneous = mean + fluctuation components


Ensemble average :  , !  lim ∑' +, 
-+. , !
'→) '
Mean properties
Commute with differentiation
#  0 /    / / " 0  / " 0

Mass conservation for incompressible flows

34 34 346
∇⋅   0 => mean :  0 & fluctuation : 0
354 354 354

Momentum equation for incompressible flows

37 3  3< 3 > 7
" 9 +  : " =
38 354 ; 357 354 354

37 3  3< 3 > 7 376 46

Mean : " 35 9 +  : ; 35 " = 35 35 :
38 4 7 4 4 354
6
37
3 # # # # # #  3<6 3> 76
Fluctuation :
38
" 354
9  + " 9  + "  + 9 :  + 9  : ; 35 " = 35
7 4 354
3? 3? 3@4
TKE(k) equation : " 9  : "A:B
38 354 354
 RANS governing equations
Reynolds equation for mean velocity
34 37 34 7  3< 3> 7 376 46
0 & "  : " = :
354 38 354 ; 357 354 354 354

2 equations + 3 unknowns (+ , A, +# 9# )

One more condition is necessary for solution

Boussinesq approximation
37
+# 9# ~ + or : mean strain rate induces turbulence
354

Analogous to relations between viscous stress and strain rate

37 34 G 
:+# 9#  =8 " : I+9 * I  +# +# : TKE
354 357 H G
=8 : turbulent (eddy) viscosity
37
A limitation : no time lag between +# 9# and is allowed
354
1, 1P32
+ Koj
34 37 34 7  3< ∗ 3 37 34
0 & "  : " = " =8 "
354 38 354 ; 357 354 354 357
 Model approach of RANS
Requirement : Mean flow field => =8

A simple but widely-used assumption

=8 ≃ L ∗ ∗
Dimension of =8 : LG /T  L O L/T
Using turbulence length scale L ∗ and velocity scale ∗

Many models were derived with different L ∗ & ∗


Ex) ∗ ~I/G * I  +# +# : TKE
G
 Ref.) TKE model
Turbulent kinetic energy (TKE) model (Prandtl, 1945) 
* I  +# +# : TKE
G
Velocity scale :  ∗ from TKE: ∗  QI/G
Derived model : =8   ∗ LR  QI/G LR ; Q  0.55 -empirical.
-

empirical
Original TKE equation (no assumption)
3? 3? 3@4
38
" 9 35  : 35 " A : B
4 4

Assumptions (physical and empirical)

 ^ 3?
Y+  +# 9# 9# " +# A′/[ : 2=9# ]+9# ≃ : = " ; a?  1.0 from exp.
G _` 357
# # 37 37
Turb. production : A  :+ 9 ≃ 2=8 ]+9
35 4 354
f
# # d? e/> ? f/>
Turb. dissipation : B  2=]+9 ]+9 ≃ gh
 ij
gh

Modeled TKE eq.

3? 3? 3 ^ 3? 37 ? f/>
" 9  =" " 2=8 ]+9 : ij
38 354 354 _` 354 354 gh
 Ref.) TKE model
f/>
# # ?
Validity of model assumption : ε  2=]+9 ]+9 ≃ ij
gh

? f/>
Length scale : LR  ij
l

/G ? f/> /G ?> ^ l

Outcome : =8  QLR I  Q ij I  Qij ⇒  Qij  Qnop!
l l ?>

Source: “ Turbulent flows” by Pope

 Ref.) TKE-based RANS models
TKE model
3? 3? 3 ^ 3? 37 ? f/>
" 9  =" " 2=8 ]+9 : ij
38 354 354 _` 354 354 gh

=8  iq I/G LR
LR needs a better model than a constant

Empirical LR model (e.g. simple wall-bounded flow)

Using B -equation (turbulent dissipation)

ㅡ

f
?>
LR  (dimensionally)
l
Outcome : =8  iq I/G LR  iq I G /B : I : B model
~

Using r-equation (turbulent frequency) T

~
ㅡ ,

e
? >
LR 
s>
Outcome : =8  iq I/G LR  iq I/r : I : r model
~

ㅡ
 Turbulence
Ref.) RANS models Modeling
in practice
I : B model (Launder and Spalding, 1974)
IG
Eddy (turbulent) viscosity t8  [iq
B
Transport equations for k and ε

u[I t8
" ∇ ⋅ [v I  ∇ ⋅ t" ∇I " A? : [B TKE eq.
u! a?

u[B t8 B BG homogeneous
" ∇ ⋅ [vB  ∇ ⋅ t" ∇B " il A? : ilG [
u! al I I turbulence

iq  0.09 a?  1.0 al  1.3 il  1.44 ilG  1.92

uv+ 2 u? 2 1 u+ u9
A?  z+9 z+9  t8 2]+9 :  : [I+9 ]+9  "
u9 3 u? +9 3 2 u9 u+

Transport Transport Rate of Rate of

Rate of change
+ of k or ε = of k or ε + production - destruction
of k or ε
by convection by diffusion of k or ε of k or ε
 Ref.) RANS models in practice
I : B model (Launder and Spalding, 1974)

• Advantages:
• Relatively simple to implement
• Leads to stable calculations that converge relatively easily
• Reasonable predictions for many flows
• Disadvantages:
• Poor predictions for:
• swirling and rotating flows
• flows with strong separation
• axisymmetric jets
• certain unconfined flows, and
• fully developed flows in non-circular ducts
• Valid only for fully turbulent flows
• Simplistic ε equation

Ref. : Applied Computational Fluid Dynamics, Lecture 10 –Turbulence models, André Bakker (2002- 2005), Fluent Inc.(2002)
https:/ / www.cfd- online.com/ Wiki/ Standard_k- epsilon_model
 Turbulence
Ref.) RANS models Modeling
in practice
RNG k-ε model (Yakhot, 1992)
: RNG procedure systematically removes the small scales of motion from
the governing equations by expressing their effects in terms of larger scale
motions and a modified viscosity
IG
t8  [iq t|}}  t + t8
B
Transport equations for k and ε
u[I
" ∇ ⋅ [vI  ∇ ⋅ {? t|}} ∇I " A? : [B
u!
u[B B B G
∗
" ∇ ⋅ [v B  ∇ ⋅ {l t|}} ∇B " il A? : ilG [
u! I I

iq  0.0845 {?  {l  1.39 il  1.42 ilG  1.68

iq  H 1 :  ⁄~ I
∗
ilG  ilG "  2]+9 ]+9 ~  4.38 Ä  0.012
1 " Ä H B
Yakhot, V., Orszag, S.A., Thangam, S., Gatski, T.B. & Speziale, C.G. (1992), "Development of turbulence models for shear flows by
a double expansion technique", Physics of Fluids A, Vol. 4, No. 7, pp. 1510- 1520.
 Ref.) RANS models in practice
RNG k-ε model (Yakhot, 1992)
• I : B equations are derived from the application of a rigoro
us statistical technique (Renormalization Group Method) to
the instantaneous Navier-Stokes equation.
V

• Similar in form to the standard I : B equations but includes:

• Additional term in ε equation for interaction between turbulence dissipation
and mean shear.
• The effect of swirl on turbulence.
• Analytical derivation for turbulent Prandtl number.
• Differential formula for effective viscosity.
• Improved predictions for:
• High streamline curvature and strain rate.
• Transitional flows.
• Wall heat and mass transfer.
• But still does not predict the spreading of a round jet correctly.
Ref. : Applied Computational Fluid Dynamics, Lecture 10 –Turbulence models, André Bakker (2002- 2005), Fluent Inc.(2002)
https:/ / www.cfd- online.com/ Wiki/ RNG_k- epsilon_model
 Ref.) RANS models in practice
Realizable k-ε model (Shih, 1995)*
• Shares the same turbulent kinetic energy equation as the
standard I : B model
• Improved equation for ε IG
t8  [iq
• Variable iq instead of constant B
• Improved performance for flows involving:
• Planar and round jets (predicts round jet spreading correctly)
• Boundary layers under strong adverse pressure gradients or separation
• Rotation, recirculation
• Strong streamline curvature

Ref. : Applied Computational Fluid Dynamics, Lecture 10 –Turbulence models, André Bakker (2002- 2005), Fluent Inc.(2002)
 Ref.) RANS models in practice
Realizable k-ε model (Shih, 1995)
Distinctions from standard I : B model:
– Alternative formulation for turbulent viscosity:
IG 1
t8  [iq iq 
B where É~ " ÉÑ
IÖ ∗
B
• (As and U* are functions of velocity gradients)
• Ensures positivity of normal stresses: u 2 ≥0
• Ensures Schwarz’s inequality: (u u )2 ≤ u2 u 2
i j i j

– New transport equation for dissipation rate, ε:

u[B t8 BG B
" ∇ ⋅ [vB  ∇ ⋅ t" ̅
∇B " [i ]B : [iG " il iHl Aá
u! al I " =B I

Diffusion Generation Destruction Buoyancy

Ref. : Applied Computational Fluid Dynamics, Lecture 10 –Turbulence models, André Bakker (2002- 2005), Fluent Inc.(2002)
 Ref.) RANS models in practice
Realizable k-ε model (Shih, 1995)
IG 1
Eddy viscosity computed from t8  [iq , iq 
B IÖ ∗
É~ " ÉÑ
B
Ö∗  ]+9 ]+9 " Ω+9 Ω+9
1
É~  4.04, ÉÑ  6 cos ç, ç  cos é 6à
3
]+9 ]9? ]?+ 1 u+ u9
à ]â  ]+9 ]+9 ]̅  2]+9 ]+9 Ω+9  :
]â H 2 u9 u+
 I
i  max 0.43,  2]+9 ]+9
"5 B

il  1.44 iG  1.9 a?  1.0 al  1.2

Ref. : Applied Computational Fluid Dynamics, Lecture 10 –Turbulence models, André Bakker (2002- 2005), Fluent Inc.(2002)
https:/ / www.cfd- online.com/ Wiki/ Realisable_k- epsilon_model
https:/ / www.afs.enea.it/ project/ neptunius/ docs/ fluent/ html/ th/ node60.htm
 Turbulence
Ref.) RANS models Modeling
in practice
k-ω model (Wilcox, 1988)
NPV는
I
Eddy (turbulent) viscosity t8  [
r
l js s j? s jl
Transport equations for k and ω * r≡ ⇒  : "
? j8 ? j8 l j8

u[I u[v9 I u uI
"  t" a ∗t 8 " A? : Ä ∗ [Ir
u! u9 u9 u9

u[r u[v9 r u ur r
"  t " at8 " { A? : Ä[rG
u! u9 u9 u9 I

{  5⁄9 Ä  3⁄40 Ä ∗  9⁄100 a  1⁄2 a ∗  1⁄2

uv+ 2 uv+ 1 u+ u9
A?  z+9 ̅ G ë G
nê t8 ] nê t8 Ω : [I+9 ë
Ω  2Ω+9 Ω+9 Ω+9  :
u9 3 u9 2 u9 u+
B  Ä ∗ rI
Wilcox, D.C. (1988), "Re- assessment of the scale- determining equation for advanced turbulence models", AIAA Journal, vol. 26, no.
11, pp. 1299- 1310.
https:/ / turbmodels.larc.nasa.gov/ wilcox.html
 Ref.) RANS models in practice
k-ω model (Wilcox, 1988)

• This is another two equation model. In this model ω is an

inverse time scale associated with the turbulence.
• This model solves two additional PDEs:
• A modified version of the k equation used in the k-ε model.
• A transport equation for ω.

• Its numerical behavior is similar to that of the k-ε models.

• In boundary layer flows, k-ω model is more accurate (than k-ε
model) for
• Viscous near-wall region
• Presence of streamwise pressure gradient (acceleration,
deceleration)

Ref. : Applied Computational Fluid Dynamics, Lecture 10 –Turbulence models, André Bakker (2002- 2005), Fluent Inc.(2002)
 Ref.) RANS models in practice
k-ω SST model (Menter, 1994)*
The k-ω based SST model accounts for the transport of the turbulent shear str
ess and gives highly accurate predictions of the onset and the amount of flow
separation under adverse pressure gradients.
The baseline model combines the advantages of the Wilcox (near-wall) and t
he k-ε models, but still fails to properly predict the onset and amount of flow
separation from smooth surfaces
The main reason is that both models do not account for the transport of the
turbulent shear stress. This results in an over-prediction of the eddy-viscosi
ty.

The proper transport behavior can be obtained by a limiter to the formulation

of the eddy viscosity:

Eddy (turbulent) viscosity íG  tanh arg GG

? ô~~
[/ I [/ I arg G  max 2 ,
t8  ⇐ ò ∗ s  > s
max / r, ]̅íG ë íG
max / r, Ω /  0.31 Ä ∗  0.09
 Ref.) RANS models in practice
k-ω SST model (Menter, 1994)
F2 is a blending function similar to F1, which restricts the limiter to the wall boun
dary layer, as the underlying assumptions are not correct for free shear flows.
S is an invariant measure of the strain rate.

The blending functions are critical to the success of the method.

Their formulation is based on the distance to the nearest surface and on the flo
w variables.
í  tanh arg
ç  í ç " 1 : í çG I 500= 4[asG I
arg  min max ∗ , ,
Ä r  G r CD?s  G

where y is the distance to the nearest wall and ν is the kinematic viscosity and:
1 uI ur
CD?s  max 2[asG , 10é~
r u9 u9
Ä as  G
I : r closure : a?  0.85 aú  0.5 {  5/9 Ä  0.075 û  ∗ :
Ä Ä∗
(near- wall)
ÄG asG  G
I : B closure : a?G  1.0 aúG  0.856 {G  0.44 ÄG  0.0828 ûG  ∗ :
Ä Ä∗
 Turbulence
Ref.) RANS models Modeling
in practice
k-ω SST model (Menter, 1994)
Transport equations for k and ω
u[I u[v9 I u uI
"  t " a? t8 " A? : Ä ∗ [Ir
u! u9 u9 u9
u[r u[v9 r u ur û G
[asG uI ur
"  t " as t8 " A? : Ä[r " 2 1 : í
u! u9 u9 u9 =8 r u9 u9
uv+ 2 uv+
A?  z+9 ̅ G ë G
nê t8 ] nê t8 Ω : [I+9 Prod. limiter : A? min A? , 10Ä ∗ [rI
u9 3 u9
2 u? 2 1 u+ u9
z+9  t8 2]+9 : +9 : [I+9 ]̅  2]+9 ]+9 ë
Ω 2Ω+9 Ω+9 Ω+9  :
3 u? 3 2 u9 u+

Boundary conditions
†
Wall : Iúügg  0 rúügg  10
òe °e >
>
~¢£ §• >
~.§• §• §•
Far- field :  Ifarfield   rfarfield  10
¶|ß ¶|ß © ©

Menter, F. R. (1994), "Two- Equation Eddy- Viscosity Turbulence Models for Engineering Applications", AIAA Journal, vol.
32, no 8. pp. 1598- 1605.
https:/ / www.cfd- online.com/ Wiki/ SST_k- omega_model / https:/ / turbmodels.larc.nasa.gov/ sst.html
https:/ / en.wikipedia.org/ wiki/ Menter%27s_Shear_Stress_Transport
 Turbulence
Ref.) RANS models Modeling
in practice
Reynolds stress models (Launder et al., 1975)
1 2 3 4 5 Reynolds stress
´™+9
 A+9 " ´+9 : B+9 " Π+9 " Ω+9 ™+9  : z+9 ⁄[  ′+ ′9
´!
uv9 uv9
1 Rate of production of ™+9 A+9  : ™+R " ™9R
uR uR

u =8 u™+9 =8
2 Transport of ™+9 by diffusion ´+9  ∇⋅ ∇ ™+9
uR a? uR a?
2
3 Rate of dissipation of ™+9 B+9  B+9
3

4 Transport of ™+9 due to turbulent pressure-strain interactions

B 2 2
Π+9  :i ™ : I : iG A+9 : A+9 C1 = 1.8 C2 = 0.6
I +9 3 +9 3

5 Transport of ™+9 due to rotation Ω+9  :2r? ™9R +?R : ™+R 9?R
 Turbulence
Ref.) RANS models Modeling
in practice
Algebraic stress equation models (Naot and Rodi, 1982)
: Assume that the sum of the convection and diffusion terms of the Reynolds
stresses is proportional to the sum of the convection and diffusion terms of
turbulent kinetic energy

´′+ ′9 ′+ ′9 ´I

: ´+9  ⋅ : I : !ê/opÆnê!_!êp
´! I ´!
#7 #4
= ⋅ :′+ ′9 ⋅ ]+9 : B
?

2 ij 2 I
Algebraic stress model ™+9  ′+ ′9  I+9 " A+9 : A+9
3 i : 1 " A⁄B 3 B

ij  0.55 i  2.2

Ref. : Applied Computational Fluid Dynamics, Lecture 10 –Turbulence models, André Bakker (2002- 2005), Fluent Inc.(2002)
 Ref.) RANS models in practice
Spalart-Allmaras model (1992)
Eddy (turbulent) viscosity t8  [=̃8 µ
" fasti less acurate
Transport equations for ¥ (1-equation) -

G G
u[=̃8 u=̃8 1 u u=̃8 iáG [ u=̃8 iá =̃8
" [9  t|}} " " iá 1 : 8G []â=̃8 : [ is s : G 8G
u! u9 a u9 u9 a u9  
=̃8 H =̃8 
t|}}  t + [=̃8  µ  H ]â  ] "  µG  1 :
= H
 " iµ  G  G µG 1 " µ
1 u+ u9
] 2Ω+9 Ω+9 Ω+9  : 8G  i8H Æ :i8  G
2 u9 u+
† /†
iá 1 " iáG 1 " isH =̃8
is  G " s   †   ê " isG ê † : ê ê
 a †
 " isH ]â G  G


0.1355 0.622 0.666 7.1 0.3 2.0 0.41
i8 i8G i8H i8
1 2 1.1(1.2) 2(0.5)
Ref: https:/ / www.cfd- online.com/ Wiki/ Spalart- Allmaras_model
https:/ / en.wikipedia.org/ wiki/ Spalart%E2%80%93Allmaras_turbulence_model
 Ref.) RANS models in practice
Spalart-Allmaras model (1992)
• Solves a single conservation equation (PDE) for the turbulent
viscosity:
• This conservation equation contains convective and diffusive transport
terms, as well as expressions for the production and dissipation of νt.
• Developed for use in unstructured codes in the aerospace industry.
• Economical and accurate for:
• Attached wall-bounded flows.
• Flows with mild separation and recirculation.
• Weak for:
• Massively separated flows.
• Free shear flows.
• Decaying turbulence.
• Relatively narrow use
 Comparison of RANS models
There are many (case-specific) previous comparison studies
Common sense : the more recent and more complex, the better

Ex) Heated pipe (김영인 et al., 한국원자력학회 2003)

 Comparison of RANS models
Ex) Turbulent swirling flow inside annular pipe at ™=25,000

Swirl number profiles along the streamwise direction

 Summary of RANS
Energyy
Reynolds decomposition :  , !   , ! " # , !
Instantaneous = mean + fluctuation components

KFJUi +☆ U
블

( 5+2
UI
Features ,
Assume turbulent fluctuation ~ mean strain rate
Boussinesq approximation
Analogous to viscous stress ~ strain rate
It can predict mean flow fields ( Uz p ).
-

Most of the energy is carried by largest scales

Very fast : 10 min. ~ 1 day
Easy to set boundary conditions ( iP
U ) 발

For mean flow fields

Accurate for
Large-scale dominant problems (i
U P)
발

Inaccurate for
Small-scale dominant problems (U: p
'
)
,

Large turbulence model effect

 On application of RANS
RANS is popular for
Single-phase turbulent flows modelsare
. funed to

0
Large-scale dominant problems
Fluid machinery
Heat transfer 0
Bubbly flows with many bubbles

simalatione
…

CostN

( UZ P )
.
 On application of RANS
RANS is less favored for
Any problem for which small-scale physics is important
Ex) Turbulent combustion
~

Mixing near thin flame front matters

*
smal
scalematters
Ex) Flow-generated acoustic noise
Density and pressure waves are easily damped out
~

snalt
- scales

Freund et al. (2000)

Ex) Interface phenomena in two-phase flows

 Some thoughts on RANS
Why RANS flow fields look like a laminar flow?
RANS : =8 ~ 10~ O =Rg (LES : =8 ~=Rg )
⇒ large model effect7 large lt
-

Solution quality depends mainly on model (=8 ) >> Numerical error

…

Relatively easy to achieve grid-independence

Numerical method is less important ≠ DNS
*
Uncertainty of RANS
Especially for unsteady RANS
RANS is targeted for mean flow fields
평
P
Vorticity filed of 2-D k-ԑ standard

Vorticity filed of 2-D k-ω SST Vorticity filed of 2-D Spalart-Allmaras

 Introduction to
large eddy simulation

Computational Fluid Dynamics

 Idea of large eddy simulation (LES)
Based on division of scales in LES (based on filter width Δ)

* Courtesy of J. Lee, Ajou University

Large scales
Has characteristic features (←geometry) of turbulent flows
Has a majority of total kinetic energy
Usually, anisotropic
Small scales
Has a relatively smaller portion of kinetic energy
Nearly isotropic, modeling is relatively easy
 How LES works for turbulent flows
Small scales are modeled
More accurate modeling due to local isotropy
Log(Energy)

O
Δ ghid
=
size

Large
scales
Small
scales
Log(k)

o
Solve directly Modeling * Courtesy of J. Lee, Ajou University

Large scales are resolved by accurate numerical methods

Large scales contains most of energy

Solution quality depends on numerical method, grid resolution (+ SGS model)

It can use much coarser grid than DNS

In conclusion, LES achieves accuracy (compared to RANS) and efficiency

(compared to DNS) simultaneously
 LES governing equations
Spatial filtering as opposed to Reynolds (ensemble) averaging (RANS)
+ ⃗ ≡ ∭ æ ê⃗, ⃗ + ⃗ : ê⃗, ! øê⃗

Filter kernel
∭ æ ê⃗, ⃗ øê⃗  1


  ƒ |ê|  1/2
Box filter in 1D : æ-ê.  ¡ Δ : |ê|  ¿ * Δ: filter width
ㅡ

¿ G
0 n!ℎ꫃p
Outcome
+ ⃗  + ⃗ " +# ⃗
instantaneous = filtered + fluctuation
+# ≠ 0 RAWs
에
uiuy

&
.

Governing equation for filtered velocity (for incompressible flow)

∗ … ¶ 1 ¶
34 37 34 7  3< 3>  7 3»74 z+9  z+9 : z?? +9
0 & " : "= : ¶
3
354
~
38 354 ; 357 354 354 354 z+9 ≡ + 9 : + 9
~
Smagorinsky subgrid-scale (SGS) model
RA 1
-
소
The first LES (SGS) model 인
37
"
34 7
 :
 3<
∗
" =
3 > 7
:
V …
3»74
38 354 ; 357 354 354 354

Assumption #1 : ~ Boussinesq approximation

37 34
Residual stress z+9  :=8 "  :2=8 ]+9
354 357
…
∗
u+ u+ 9 1 uA u u+ u9 Joseph Smagorinsky
⇒ " : " = " =8 " = RANS
u! u9 [ u+ u9 은
u9 u+ (American meteorologist)

ë .G ]
Assumption #2 : Eddy viscosity =8  LÀ ¿#  LÀG ]  -iÀ Δ
ë : Length scale ~ local grid size Δ
LÀ  iÀ Δ ë * ] 2]+9 ]+9
Coefficient iÀ must be specified
f/>
lengthtengx
백타 pls
5 =

ë
LÀ ~Δ (LES) ≪ LR ~
?
ë (LES) ≪ ∗ ~I /G (RANS)
(RANS), ¿# ~LÀ ~Δ
l

ë ) decreases
ë G ) decreases as the grid size (Δ
In LES, the SGS stress (z+9 ~=8 ~Δ
Grid resolution takes over G .f t ,Vto ( DNS )
신

In RANS, the Reynolds stress (:+# 9# ~=8 ~I G /B) remain significant although
ë becomes small
the grid size Δ
 Smagorinsky SGS model
G
 Role of model coefficient Cs z+9 ~ iÀ Δ ] ]+9

1. If Cs is too small 7 ⑦7 Vt - 0 2. If Cs is too big → Ut 범 RANS

Energy Energy

Too
Large No Large much
scales small scales small
scales scales

k k
τ ij ≈ 0 τ ij is too big

Actually, this is coarse DNS Energy in turbulence is dissipated

* Courtesy of J. Lee, Ajou University

No turbulence
Model coefficient Cs should have appropriate value
Optimal Cs may vary over space
 Dynamic Smagorinsky model (DSM)
Optimal iÀ from DSM varies over space and time

Energy Dynamic determination of Cs

Appropriate energy

Large
scales
Small
scales
k
lengthscale
* Courtesy of J. Lee, Ajou University

Cited more than 2000 times

One of the most important contributions to CFD
Ref.) Dynamic Smagorinsky model (DSM)
Dynamic Smagorinsky model (Germano et al., 1991; Lilly, 1991)
Model coefficient iÀ is dynamically adjusted by introducing test-filtering
Grid filter : size Δ ; Test filter : size Δ  2Δ
Õ

Assume an identity in iÀ between two different filtered scales

Õ levels)
Germano identity (between Δ and Δ
∗
u+ u+ 9 1 uA u G G
u+ u9
" : " =" iÀ Δ ] " for grid filter level
u! u9 [ u+ u9 u9 u+
uvŒ+ uvŒ+ vŒ9 1 uAvŒ∗ u G
GÕ Œ
uvŒ+ uvŒ9
" : " =" iÀ Δ ] " for test filter level
u! u9 [ u+ u9 u9 u+
iÀ can be obtained by introducing one additional eqn. from test-filtering
œ+9 ]+9 Germano identity
Germano DSM : iÀG 
—+9 ]+9
œ+9  – ÕÕ
+ 9 : + 9
œ+9 —+9 G
—  2Δ ] ] : 2Δ ]Œ ]Œ
G– Õ
Lilly DSM : iÀG  +9 +9 +9
—+9 —+9
 Ref.) Global SGS model
Distinguish laminar/turbulent flows mathematically based on flow info.

“ “ H⁄G
WALE model (Nicoud & Ducros, 1999) G
]+9 ]+9
=8  iΔ
- The first global SGS model ̅ ô⁄G " ]+9
̅ ]+9
]+9 “ “ ô⁄
]+9
- Determination of laminar/turbulent flow is
sometimes incorrect

Vreman model (Vreman, 2004) ””‘

=8  i
- Guarantees that zero SGS dissipation for ë+9 α
α ë+9
laminar flows uv9
H
G
α+9  , ””‘  ÷ ΔR αR+ {R9
u+
R,

Sigma model (Nicoud et al., 2011)

σH σ : σG σG : σH
- Guarantees that zero SGS dissipation for =8  iΔG
σG
laminar flows
σi is eigenvalue of αij
- Correct wall limiting behavior of νT at y+<5.

Dynamic determination of C is another issue Dynamic Global model

 Examples of LES : Fin-tube HX
Visualization of 3-D vortical structures

Inlet
Start of the
streamline
Helical flows
near the outlet

y-velocity
Streamwise
direction

Helicity -⁄¤‹›fifl¥‡ ·¤fi¥›‚ · ·›‚¥flfifl¥‡ ·¤fi¥›‚.

The number of grids : 18,507,282

Time step size : 0.0001s
CFD code : Star-CCM+ 6.0
 Comparison of results from LES, MRV and RANS (I-B)
The mean velocities in the streamwise direction along the x-line

Bottom tube Middle tube Top tube

The results of the LES turbulent model show a better agreement with
MRV measurement compared to RANS in the entire region
 Ex) 교류팬 난류모델 유동장 해석 결과
• 난류모델별 와도장 (998rpm)
– 2-D URANS (k-ɛ standard model)
3-D DES
3-D LES

– 3차원 LES 와도장에서 가장

다양한 스케일의 와도 관찰
– 실제적인 와도장과 흡사 Vorticity filed of 3-D LES D
– 광대역 소음 예측에 유리

O
Vorticity filed of 2-D URANS k-ԑ standard Vorticity filed of 3-D DES
e
Ex) LES of flows in energy conversion devices
Examples from my portfolio
 Ex) LES of Valeo airfoil
Instantaneous x-velocity and x-vorticity
A very small separation bubble near the leading edge prompts
transition to turbulence

Contours of x and z-velocity at z-y planes

ux uz ux uz ux uz ux uz ux uz
 Ex) LES of Valeo airfoil
Wall-pressure spectra at the trailing edge

rexP

LES
 Summary of LES
Turbulence model effect : RANS >> LES >> DNS

또나
Features of LES (compared to RANS)
Realistic representation of flow physics
More accurate prediction of mean flow field
Even more accurate when the role of small-scales are important
E.g., turbulent combustion / flow-generated noise / interface two-phase
phenomena
Usually easier to use with less user parameters
Numerical method & grid resolution are more important than modeling

Barriers in practical applications of LES

Computational cost
Complex geometry
Truncation errors (e.g. dispersion, diffusion) → ghid가 fine 해야함
,

Good discretization methods are necessary

Multi-physics problem
E.g., droplet break-up and coalescence
Grid resolution & Wall modeling
for LES and RANS

Computational Fluid Dynamics

Setting Δx for DNS of wall-bounded flows
Related non-dimensional variables : wall-units
  
Friction Reynolds number : ™»  
 
  
Length in wall unit :     
 

Velocity in wall unit :  


Condition for near-wall grid resolution

 ~ Kolmogorov length scale
∆
∆   1

Universal

Condition away from wall

In fully-developed wall-bounded flows
Kolmogorov length scale      / for    50 (semi-empirical)
 : 0.41
∆    
 Setting Δx for LES
Not every scale needs to be resolved in LES ≠ DNS
∆+ can be larger
But at the wall, SGS model is inactive (=8 → 0) and every scale must
기
be resolved

∆úügg  1  DNS 양 O

Grid requirement is less strict in wall-tangential (i.e. x & z) directions

In practice, ∆  , ∆‰   5~100
Large aspect ratio near the wall

Away from the wall, grid becomes coarser and more isotropic
 Estimating grid resolution of LES
Using sub-grid scale TKE / resolved TKE : the smaller, the better
^ G
Sub-grid scale TKE : IÀÂÀ 
ÊÁ ¿

Resolved TKE : I |Ñ  +# +#
G

LES quality index


Celik et al. (2009) : ”ËÀÂÀ   Í Î
> 0.8 for "good" LES
È ^


{  0.05, ‰  0.53

Using filter width / Kolmogorov length scale

~.Gô
Celik et al. (2009) : Δ/  Δ/ fÌ < 25 for "good" LES
l^Ï^
Resolved B needs to be evaluated

Most conditions have empirical aspects

 Number of grid points for LES
Estimation of number of grid points required for wall-bounded flows

Number of grid points ≲ ™ G

Grid requirement for DNS of homogeneous turbulence : o8  o5 H ~™8 Ô/
Wall-normal direction is the limiting direction
Turbulent boundary layer flow (Piomelli & Balaras, 2002)
 LES at high Re
Estimation of number of grid points required for LES
(turbulent boundary layer flow; Piomelli & Balaras, 2002)
1014 Large eddy simulation
10
13 (LES)
12
10
Chapman Choi & Moin
11 (AIAA J, 1979) (PoF, 2012)
Number 10
of grid pts 10 LES N ~ Re1.8 N ~ Re13/7
10

O
Grid for WM
9 N ~ Re0.4 N ~ Re1
10 near- wall region LES
8 (y+<100)
10
7
Wall-modeled LES
10
(WMLES)
6
10 Grid for
5
far- from- wall Region (y+>100)
10 Re
4 5 6 7 8
10 10 10 10 10
airplane

flow far-from-wall Region (y+>100)

near-wall region (y+<100)

wall
 Detached eddy simulation (DES)
Solution #1 : Detached eddy simulation (DES; Spalart et al., 1997)
- Hybrid RANS/LES

flow far-from-wall Region (y+>100) LES

wall
near-wall region (y+<100) RANS (Reynolds-averaged NS)

- Single turbulence model acts as LES and RANS model

Òf Û^

t8  [=̃8 µ where µ  f ,  Spalart-Allmaras RANS model (1992)
Òf ÊÚe 
G G
´=8 1 u u=8 iáG u=8 =8
 iá ]â=8 " = " =8 " : is s
´! a u9 u9 a u9 øâ d : wall distance
production diffusion destruction
d RANS = d , d = wall distance
d% ≡ min(d RANS , d LES )
d LES = CDES ∆, ∆ = max( ∆x, ∆y, ∆z ), CDES = 0.65

- Easy to implement from S-A code

Ref. : Spalart, P. R., Jou, W.- H., Stretlets, M., and Allmaras, S. R. (1997), "Comments on the Feasibility of LES for Wings and on the Hybrid
RANS/ LES Approach", Advances in DNS/ LES, Proceedings of the First AFOSR International Conference on DNS/ LES.
Strelets, M. (2001), "Detached Eddy Simulation of Massively Separated Flows", AIAA 2001- 0879.
 Detached eddy simulation (DES)
▶ DES is the most widely used WMLES technique for engineering applications
because of the simplicity of the model implementation

F-15 jet at a 65˚ AOA Ford automobile

(Forsythe et al. 2004) A sharp-edged delta wing
(Mendonca et al. 2002) at 27˚ AOA (Morton, 2003)

Flap edge Ground transportation system Scramjet combustor

(Langtry et al. 2009) (Maddox et al. 2004) (Choi et al. 2007)
 Ex) 교류팬 LES와 DES
• 난류모델별 와도장 (998rpm)

Vorticity filed of 3-D LES

Vorticity filed of 3-D DES

 Issues of detached eddy simulation
▶ Issues in DES
- Many arbitrary coefficients (compared to LES)
χ 3 %
ν
ν T = ν% f v1 where f v1 = 3 3 , χ ≡
χ + cv1 ν
2
Dν% 1 ν% 
Dt
= Cb1 S%ν% +
σ
{ } 2
∇ ⋅ [ (ν + ν% )∇ν% ] + Cb 2 ( ∇ν% ) − Cw1 f w  
 d%  d : wall distance
d RANS = d , d = wall distance
d% ≡ min(d RANS , d LES )
d LES = CDES ∆, ∆ = max( ∆x, ∆y , ∆z )

- DES requires △y+ = O(1) for the first grid size above the wall to provide accurate
wall shear stress, which still requires a significant number of grids in the wall-
normal direction.
LES DES

∆y + ~ O(1) ∆y + ~ O(1)
 On wall layer modeling
Idea : Impose correct mean velocity or wall shear stress using very
coarse grid
Balaras and Benocci (1994)

Ref. : Large Eddy Simulation for Particle- Laden Turbulent Flows, Kyle D. Squires, Dispersion of Particles in Turbulent flow .
 On wall layer modeling
LES (Balaras and Benocci, 1994)
Wall layer modeling for LES

Purpose: relate the wall stress to the outer layer velocity locally and
instantaneously

Ref. : Large Eddy Simulation for Particle- Laden Turbulent Flows, Kyle D. Squires, Dispersion of Particles in Turbulent flow .
 Wall modeling in RANS
High Reynolds No. model : Wall Function
• y+ is the non-dimensional distance from the wall

u
Boundary layer
Wall functions used to
resolve boundary layer
 
Log law :    ln   " ı  ln ˆ 
Ù Ù
y+

• Wall functions are only valid within specific y+ values

• If y+ is too high the first node is outside the boundary layer and wall functions
will be imposed too far into the domain
• If y+ is too low the first node will lie in the laminar (viscous) part of the boundar
y layer where wall functions are not valid
Ref. : ANSYS, Inc. Proprietary © 2009 ANSYS, Inc. All rights reserved.
 Wall modeling in RANS
High Reynolds No. model : Wall Function

p
LES with wall shear stress boundary condition
Solution #2 : LES with wall shear stress boundary condition
LES with
LES
wall shear stress B.C.

Δ  ~˜ 1 Δ  ~
˜ 10~100
wall wall z5,ú , ¯˘  0, z˙,ú
ú
 
▶ Schumann (1975) Log law :    ln   " ı  ln ˆ 
Ù Ù

u ( x , y2 , z , t ) τ w is prescribed from the mean pressure gradient.

τ xy , w ( x, z, t ) = τw
u ( y2 )
y2: location of the first off-wall grid point
vw = 0
2 w( x, y2 , z , t ) • : plane-averaged value
τ zy , w ( x, z , t ) =
Reτ y2 u, w: filtered velocity components in x, z directions
LES with mean wall shear stress BC (Lee et al. 2013)
 LES of turbulent channel flow
 Mean streamwise velocity profiles at ™˚  2 O 10 ~2 O 10¸ 64(x)×33(y)×32(z)

60
(1/0.37)ln y++3.7 (Nagib et al., 2008)
LES (present)
50
Reτ = 2 ×108

Reτ = 2 ×107
40
Reτ = 2 ×106
u+ 30 Reτ = 2 ×105

Reτ = 2 ×104
20
Reτ = 2 ×103

10

0
101 102 103 104 105 106 107 108 109
y+
 Issue of wall function and wall model
Wall function and wall model : based on log law

• In some situations, such as boundary layer separation, wall functions do

not correctly predict the boundary layer profile

Wall function applicable Wall function not applicable

• In these cases, wall functions should not be used
• Instead, directly resolving the boundary layer can provide accurate
results
• Not all turbulence models allow the wall functions to be turned off
Ref. : ANSYS, Inc. Proprietary © 2009 ANSYS, Inc. All rights reserved.
 Setting Δx for RANS
 ∆
The condition (  ∆   1) is still valid only in the wall-normal (y)

direction
Leads to near-wall grids with a very high aspect ratio (> LES)
Away from the wall, the grid size can be very large
Turbulence model > Discretization method

Lack of conditions : Grid convergence study matters

Use of wall model (wall function) also helps a lot to reduce grid points
Similar to LES