You are on page 1of 64

200

Overview of Turbulence Models


for Industrial Applications:
Professor Ismail B. Celik
West Virginia University
icelik@wvu.edu ; (304) 293 3111
Part -II: Advanced Turbulence Modeling
201
Outline: Part-II
Introduction: laminar versus turbulent flow
Governing equations
Averaging techniques
Two equation models
Compressibility effects
Reynolds-stress-transport models (RSTM)
Algebraic stress models (ASM)
Large eddy simulations (LES)
Summary and conclusions
202
Introduction: Flow Regimes
Steady and Unsteady Laminar and Turbulent Flow
203
Introduction: What is Turbulence?
What is turbulence?
Fluid flow occurs primarily in two regimes: laminar and
turbulent flow regimes.
Laminar flow:
smooth, orderly flow restricted (usually) to low values of
key parameters- Reynolds number, Grashof number,
Taylor number, Richardson number.
Turbulent flow:
fluctuating, disorderly (random) motion of fluids
204
Turbulent fluid motion is an irregular condition of flow in
which various quantities show a random variation with time
and space coordinates, so that statistically distinct average
values can be discerned. (Hinze, 1975)
Beyond the critical values of some dimensionless
parameters (e.g. Reynolds number) the laminar flow
becomes unstable and transitions itself into a more stable
but chaotic mode called turbulence characterized by
unsteady, and spatially varying (three-dimensional) random
fluctuations which enhance mixing, diffusion, entrainment,
and dissipation.
Introduction: What is Turbulence?
205
Laminar Flow Examples
(Woods et al., 1988) (Van dyke, 1982)
206
Turbulent Flow Examples
(Van Dyke, 1982) (Van Dyke, 1982)
207
Turbulence Scales
Velocity (fluctuations): u
Length (eddy size):
Time, = /u
Turbulence Reynolds
number
Re
t
= u /
Turbulent kinetic energy:
k~ 3u
2
/2
Dissipation rate: ~ u
3
/
Kolmogorov scales:

K
= (/)
1/2

K
= (
3
/)
1/4
u
K
= ()
1/4
Large eddies in a turbulent
boundary layer (Tennekes
and Lumley, 1992):
~L
t
=boundary layer thickness
208
Governing Equations
Conservation of Mass:
Conservation of Momentum (Navier-Stokes Equations):
Conservation of Energy:
0 ) U (
x t
i
i
=


j
ij
i
i j
i
i
x x
P
) U U (
x
) U (
t

+

direction) - j in flux (heat


x
t
k q
Dt
DP
x
q
x
) h U (
) h (
t
j
j
j
j
j
j

=
+ +


e, temperatur T enthalpy, h = =
209
Governing Equations (Continued)
Stress-strain relation
Viscous dissipation
Equation of state
; Kroneckers delta
j
i
ij
x
U

=
ij
~ ~
j
i
i
j
ij
U
3
2
x
U
x
U

=

) T , ( func P =
{
j i if 1
j i if 0 ij
=

=
210
Stationary Turbulence
Averaging Techniques: Reynolds Averaging
Unstationary Turbulence
211
U = <U> + u; Notation u = u = fluctuating component of U(x,t)
Time average:
Ensemble average:
Phase Averaging:
t = window width
>= =<
+
t as Limit ; dt ) t ( U
t
1
U U
t t
t
0
0



( ) Large N ; t , x U
N
1
U
N
1 i
i
>= <
=
( ) + >= <

2
t
2
t
d ) t , x ( U
t
1
t , x U



Averaging Techniques: Reynolds Averaging
212
For flows with significant variations in fluid properties,
fluctuations in density and viscosity etc., can not be neglected.
In such flows a density weighted average (Favre average) is
more appropriate.
U = {U}+ u; decomposition of U(x,t)
Definition {U}= <U>/< >, {}= Favre average
note: <u> = -< u>/< > = - < u>/< > 0,
but < u> = 0.
{uv}= <uv> + < uv>/< > - < u
v>/< >
2
Averaging Techniques: Favre Averaging
213
< U + V > = <U> + <V>; < <U> > = <U>; <U><V>> = <U><V>
<dU/dt> = d(<U>)/dt; <d(UV)/dx> = d (<UV>)/dx
average of a derivative = derivative of the average
<u> = 0; average of the fluctuations is zero , (not for Favre
averaging)
<UV> = <U><V> + <uv> ; <uv> 0. (non linear terms!)
Comment: Average of linear terms is the same with the averaged
quantities substituted, Non-linear terms, e.g. d(UV)/dx, lead to extra
terms that need to be calculated separately.
Averaging Rules
214
Two- Equation Models: Exact k- Equations
Exact equation can be derived for turbulent kinetic energy, k,
and its dissipation rate, , from Naiver-Stokes Equation.
k-Equation
; P Diff
Dt
Dk
k k
+ =
; diffusion diffusion
Turbulent Laminar
; u ' p u u u
2
1
x
k
q
j j i i
j
k
j

=
j
i
j
i
j
i
j i
j
i
t
ij k
x
u
x
u
x
U
u u
x
U
P

=
j
k
k
x
q
Diff

+ =
Production:
Dissipation:
215
Two- Equation Models: k- Equations
Similarly exact equations can be obtained for any quantity
n
ch
2 / m n
ch
m
ch
k u Z l l = =
(See Rodi, 1980; Menter and Scheueres, 1998, also the work book)
e.g.: : dissipation per unit turbulence kinetic energy (Specific dissipation rate);
t
ch
=()
-1
or rms fluctuation voticity (enstropy)
: Const u/l =c k
1/2
/l=c/k; -Equation(see Wilcox, 1993)

Diss P Diff
Dt
D
+ =

Diss P Diff
Dt
D
+ =
-equation:
(See workbook for details)
216
Two- Equation Models: Modeling assumptions
Turbulent diffusion is proportional to gradient of the mean flow
properties (analogous to heat conduction q
hj
=-k
h
dT/dx
j
)
The principal axes of turbulent stresses and mean-strain rate S
ij
are
aligned (Not valid for many flows)
Small turbulent eddies are isotropic (Valid at high turbulence Reynolds
number)
Turbulence phenomenon is consistent in symmetry, invariance (e.g.
coordinate invariance), permutation, and physical observations
(consistency and reliability)
Turbulence phenomenon can be characterized by one velocity scale
u
ch
k
1/2
, and one length scale, l
ch
.
t
ij

2
3
k
217
Turbulent eddy viscosity; (Dimensional analysis)
Diffusion fluxes
(See e.g. Speziale 1995, Hanjalic &Launder, 1972; p. 168 Shyy et al., 1997) <
add notes from Shyy et al and Ching and Shenq p.38 to work book or lecture notes)
Valid for equilibrium, P
k
(locally) flows at high Reynolds
number, see e.g. Hanjalic and Launder, 1972)
Two- Equation Models: Modeling of k- equations



k
u
t ;
k
; k u ;
k
c ; u c
ch
ch
2 / 3
ch
2 / 1
ch
2
t ch ch t
= = =
= = =
l
l
l
( )
j
eff k
kj
x
k
q

=
ch 2
t / C Diss

=
ch
k
1
t
P
C P

=
218
Turbulent Kinetic Energy budget: Wake flow
(After Menter & Scheuerer, 1998)
219
Two- Equation Models: Variants-1
1.15 0.8927 1.90 1.15 +0.25P/ 0.09
Nonequilibrium
0.7179 0.7179 1.68 1.42 - 0.085 RNG
1.3 1.0 1.92 1.44 0.09 Standard

k
C
2
C
1
C

Model
3
0
1
1


Table 2.0: Comparison of coefficients adopted by different models (Shyy et al., 1997)
( ) 0.015 ; 38 . 4 ; S S 2 S ; S
0
2
1
ij ij = = = =


Low-Reynolds number & near wall effects
2 2 2 1 1 1
C C ; C ; f f C f C C

Additional source term, S

= E; also * = -
0
220
Two- Equation Models: Variants-1
Example: Launder-Sharma Model
(see Wilcox, 1993, and Bardina et al; 1997 for other models)
3 . 1 , 0 . 1 , 09 . 0 C , 92 . 1 C , 44 . 1 C
y
U
2 ;
y
k
2
k
Re ; e 3 . 0 1 f ; 1 f ; e f
k 2 1
2
2
2
t
2
0
t
* 2
t
Re
2 1
50
Re
1 / 4 . 3
2
t
2
t
= = = = =

=
= = = =


Note: This model is asymptotically consistent near the wall.
3
xy
2
2
y ~ ,
y
2
~
k
, y ~ k , o y As

221
Two- Equation Models: Variants-2
Shear-Stress-Transport (SST) effects (see Menter, 1994)
Use standard two-equation (e.g. k-) model outside of the boundary layer (or
shear layer), but inside use
= (dU/dy), F
2
= [0,1] blending function
Dominant Strain-rate Effects
Near stagnation regions dU/dx is large
P
k
~ (u
2
v
2
)(U/x)
Kato-Launder modification:
Curvature, Buoyancy and Rotation effects- Zero Equation Models
l
mix
= l
m0
[ 1 + Ri

]
-1
; Ri = Richardson number (see later)
( )
31 . 0 ;
, max
;
1
2 1
1
1
2
1

= a
F a
k a
k a
P
t



[ ]
i
i
lm ij t k
x
U
k S P

=
3
2
2
2
1
2 2

222
Two- Equation Models: Variants-3
Buoyancy effects: production/destruction of k & is affected by buoyancy forces
Use Boussinesq approximation: => + only in the body force term;
g = gT
Additional source terms in the k- equations
Similarly
Ri

= Ri
g
= -G
k
/P
k
; in 2D Ri
g =
( g/)(/x)/(u/y)
2
C
3
= 1.0 (horizontal layers); C
3
= 0 (vertical layers)
Ri =Ri
f
= - (1/2) G
v
2
/(P
k
+G
k
); Flux Richardson Number, (Rodi, 1980)
Ri
f
= -G
k
/(P
k
+G
k
) Horizontal layers, Ri
f
= 0 (vertical layers); C
3
= 0.8
May also use G

= C
1
C
3
max(G
k
, 0)
If R
i
> 0 (/y>0) stable; otherwise unstable flow
;
y
g ' ' v g G
t
k

= =

( )( ) ( )


k
G P R C 1 C G
k k i 3 1
+ + =
y
g ' ' v g G
t
k

= =


223
Two- Equation Models: Variants-4
Streamline curvature effects: modify production and destruction of k and by
analogy to buoyancy
The Curvature Richardson number can be defined (for various definitions
see notes, also see Sloan et al., 1986) as:
U= velocity tangent to the curved surface, R = radius of curvature
n = direction normal to the curved surface
Rotating Flows: similar to curvature effects use
W= swirl velocity, U=axial velocity, r=radial distance
( )
( )
(

|
.
|

\
|

=
2
2 g
n
U
n
UR
R
U
2 Ri
( )[ ] ( )
2
2
g
r
U
r
W
2
r
U
r
W
2 Ri

=
224
Two- Equation Models: Variants-5
Two-Layer Models [2L]
Use two-eq. model away from the wall
Near the wall (Say for y/ <0.1) use one-eq. Model
Example (Chen and Patel, 1987):
08 . 5 A ; 70 A ;
yk
Re ; c c
A
Re
exp 1 y c ;
k
A
Re
exp 1 y c ; k c
2
1
t
4
3
1
t
1
2
3
t
1
2
1
t
= = = =
(

'

= =
(
(

'

= =

l
l
l l
Comments:
- Saves computer storage and time, increases robustness
- Avoids solving the troublesome and weakest modeled -equation in the
critical near wall regions.
225
Two- Equation Models: Examples
0.086 0.095
0.220 0.175 0.122
Round Jet
0.100 0.110
0.143 0.142 0.102 Plane Jet
0.115 0.109 0.100 0.100
Mix. Layer
0.365 0.339 0.308 0.256 Far Wake
Experimental
SA
RNG K- (SST) Flow
(After Menter and Scheurer, 1998)
Spreading rates for free shear flows: ) 0.5U (U width half ;
dx
d
Rate Spread
2
1
2
1

= = =

(After Chen, et al., 1998)


Table 4: Comparison of predicted spread rate for free shear layers
Table 3: Performance of two-equation models for free shear layers
226
Two- Equation Models: Examples
(After Menter and Scheurer, 1998)
(After Menter and Scheurer, 1998)
(After Menter and Scheurer, 1998)
(After Menter and Scheurer, 1998)
227
Two- Equation Models: Examples
(After Bardina et al., 1997)
(After Bardina et al., 1997)
(After Bardina et al., 1997)
228
Axial Dis tance from Head (mm)
R
a
d
i
a
l
D
i
s
t
a
n
c
e
f
r
o
m
A
x
i
s
(
m
m
)
0 10 20 30 40 50
0
10
20
30
k-
RNG k-
2 V
P
0 10 20 30 40 50
Ax ia l D is ta n c e fro m H e a d ( m m )
0
10
20
30
R
a
d
i
a
l
D
i
s
t
a
n
c
e
f
r
o
m
A
x
i
s
(
m
m
)
k-
R N G k-
Lo w - R e k- ( LB )
Mo rs e
2 0 V
P
C A = 9 0
o
Streamlines of the intake case at 90
o
CA Streamlines of the engine case at 180
o
CA
Profiles of axial velocity at 90 CA for various RANS turbulence
models (Yavuz and Celik, 1999)
Intake Engine (Piston-Bowl)
k-
k-
RNG k-
RNG k-
229
Effects of Compressibility
In compressible flows significant density changes occur even if the
pressure changes are small; D/dt 0; .u 0
For shock free, non-supersonic flows the Markovin hypothesis can be
used, i.e. the effect of density fluctuations on turbulence is small if
Favre averaged equations should be used with proper account of
dilation, .u , and the second coefficient of viscosity (see e.g.
Vandromme, 1995)
The k - equations should be modified to account for the dilatation
dissipation as a function of the Mach number (see e.g. Zeman, 1990)
Wall functions should be modified (see e.g. Wilcox, 1993) to include
density changes near the wall, and the Mach number effects on the log-
law coefficients.
1 <<

230
Compressibility Effects: Examples
(After Bardina et al., 1997)
Fig 5.9: Comparison of computed and measured surface
pressure and heat transfer for Mach 9.2 flow past a 40
cylinder flare.
(After Wilcox, 1993)
231
Compressibility Effects: Examples
(After Menter and Scheuerer, 1998)
232
Two- Equation Models: Assessment-1
Simple, robust, and easy to apply to complex industrial flows. No
restriction other than performance and accuracy concerns.
Eddy viscosity usually improves stability and convergence , but -
equation, especially when used with low Re-corrections can cause
convergence problems.
In general, the results can be rated as good to fair except for some
certain cases for which the model variants are calibrated.
Transport effects are partially taken into account via k- type equations,
but the history effects on Reynolds stresses are not.
233
Two- Equation Models: Assessment-2
Deficiencies of Boussinesq Approximation (i.e. eddy viscosity models)
Principal axias of Reynolds stresses are aligned with those of mean stain
rate; not necessarily so in reality (dU/dy = 0 does not always imply
Normal stresses are usually not well predicted; local isotropy assumption
which is implicitly inherent to these models is not always valid
In general these models are not good for flows with extra rate of strains
(rotation, curvature, buoyancy, secondary motion, sudden acceleration etc.)
Remedy: Reynolds Stress Transport, Models (RSTM)
Most of these short falls can be rectified by solving for the Reynolds stresses
explicitly using appropriate transport equations
This is also known as Second Moment Closure Models (SMCM)
0 uv
xy
= =
234
Turbulent scalar fluxes and variance:
For problems involving buoyancy effects or density fluctuations (e.g.
combustion, mixture fraction) turbulent fluxes , and variance (or
rms fluctuations) appear in the equations.
(Rodi, 1980)
( )
" " ! " " "! "
"! "
" " " ! " " "
" " " " ! " " " " "! "
!
isotropy) local in 0 (
destruction viscous
l l
j
correlation gradient
scalar - pressure
i
production
bouyancy
2
i
production field - mean
j
i
j
j
j i
transport diffusive
il l i
l
transport
convective
l
i
l
of change
time rate
i
x x
u
-
x
p
1
g -
x
U
u
x
u u
1
i u
x x
u
U
t
u
i
=
=

( )
" "! " " "! " "! " "! "

j j
field mean by the
production P
j
j
transport
diffusive
2
j
j
convective
transport
j
2
j
change
of rate
2
x x
2 -
x
u 2
u
x x
U
t
= =



Destruction of
2

j
u
2

235
Reynolds Stress Transport Models:
Exact Equations
Diff.
ij
=Diffusion (molecular + turbulent transport)
P
ij
= Production

ij
= redistribution or pressure-strain term

ij
= Dissipation (relation to dissipation rate of k , (3/2)
ij
=
ij
)
P
ij
= ; Production rate by the mean flow
k
j
k
i
ij
x
u
x
u
2 =

; Dissipation Rate
t
) (
t
ij


= +
k
t
k
x
) (
U
ij


Diff
ij
+P
ij
+

ij
-

ij

+
ijk
k
t
k
C
x
) (
x
ij

jk i ik j j i k ijk
pu pu u u u C + + =
k
j t
k
i t
x
U
x
U
ik jk

+
ij
i
j
j
i
ij
ps 2
x
u
x
u
p =

=
Diff = ;
236
Reynolds Stress Transport Models:
Assumptions
The modeling of various terms in the RST equations is quite
complicated and involves a deeper understanding of physics of
turbulent flows.
A brief discussion is presented in the work book. Interested readers can
find excellent reviews in the following references: Rodi (1980), Shih
(1996), Leschziner (1998), and Speziale (1998)
Here a commonly used RSTM by Launder, Reece and Rodi (1975)
namely LRR model, is given as an example. Most other models are
based on LRR model, they differ primarily in modeling of the pressure
strain term
ij
.
237
Reynolds Stress Transport Models:
Modeled Eqs.
The LRR model (Launder-Reece-Rodi, 1975)
In their original paper, Launder, Reece and Rodi recommend C
1
=1.5, C
2
=0.4, C
s
=0.11, C

=0.15,
C
1
=1.44 and C
2
=1.90. (Wilcox, 1993, pp. 232)
(Note here n is the distance normal to the surface, but P is not the pressure !)
238
Reynolds Stress Transport Models:
Examples.
(Hogg, et. al., 1989)
239
Reynolds Stress Models: Assessment.
Most rigorous of all models
Have great potential for remedying the short comings of Boussinesq
approximation without ad hoc corrections
Physically realistic predictions for flows with curved streamlines, system
rotation, stratification, sudden changes in mean strain rate, secondary
motion and anisotropy.
The most problematic equation is still the -equation
These models are mathematically complex, numerically challenging
and computationally expensive.
Wall functions and viscous damping functions are still necessary for
wall bounded and free surface flows.
Possible remedy(a compromise): non-linear eddy viscosity models
(NLEVM) and algebraic stress models (ASM)
240
Non-linear Eddy Viscosity Models - NLEVM
Assume that the Boussinesq approximation is the first term in a series expansion of
functionals (see Wilcox, 1993; Leschziner, 1997; Speziale, 1998). Here we give as an
example the Shih et al. (1993) model. See also the work book.
( )
( )
ij kl kl
3
3
7 ij kl kl
3
3
6 ij mk lm kl li kl jk lj kl ik
3
3
5 kl il jk jl ik
3
3
4
ij kl kl jk ik
2
2
3 ik jk jk ik
2
2
2 ij kl kl ij ik
2
2
1 ij ij
S
~
k
C S S S
~
k
C S
3
2
S S
~
k
C S S S
~
k
C
3
1
~
k
C S S
~
k
C S S
3
1
S S
~
k
C S
~
k
C 2 a


+ +

+ + + +

+ + +

+ =
ij
j i
ij
k
3
2
k
u u
a
9 . 0 S 25 . 1
3 2
C
+ +
=

( ) ( )
2
2
1
~
3 3 2 1
y
k
2 ;
S 1000
1
19 , 15 , 3 C , C , C

=
+
=
ij ij
S S 2
k
S

=
ij ij
2
k

= ( ) ( )
3
7 6 5 4
C 16 , 16 , 0 , 80 C , C , C , C

=
241
Algebraic Stress Models - ASM
The traditional ASMs can be viewed as implicit NLEVM. The most
commonly used ASM was proposed by Rodi (1976, 1980)
Assumption: Transport of
t
ij
is proportional to transport of k
D(
t
ij
)/Dt - Diff(
t
ij
) =
t
ij
/k [ Dk/Dt - Diff ( k ) ] =
t
ij
/k [P
k
+ G - ]
Result:
a
ij
= F
a
[ P
ij
/ - (2/3)P/
ij
] + (1-c
3
)[ G
ij
/ - (2/3) G/ ]
F
a
= (1- )/[c
1
- 1 + (P+G)/]
P
ij
= (
t
il
U
j
/ x
l
+
t
jl
U
i
/ x
l
)/ ; G
ij
= - [g
i
<u
j
> +g
j
<u
i
>]
P
k
= P
ii
/2 ; G
k
= G
ii
/2
Typical values for the constants are(Rodi, 1980):
= 0.6, c
1
= 1.8, c
3
= 0.5
Since
t
ij
appears in P
ij
and a
ij
this equation needs to be solved iteratively.
242
ASM & k- model: Examples
Swirling flow; (After Sloan et al., 1986)
Swirling flow; (After Sloan et al., 1986)
Fig. 1. Qualitative representation of a combined vortex.
Fig. 2. Qualitative spatial distribution of the stream function as
induced by a strongly swirling flow.
243
ASM & k- model: Examples
(After Sloan et al., 1986)
Fig. 14. Comparison of predicted and measured velocity
profiles for Case 4 (data from Yoon
71
; legend supplied
by table 18)
Fig. 15. Case 4
Fig. 16. Comparison of predicted and measured tangential
velocity profiles for case 4 (data from Yoon
71
; legend
supplied by table 18).
Fig. 3: Cases 3-5
Fig. 3. Case 6
244
ASM & k- model: Examples
(After Wilcox, 1993)
(After Chen et al., 1998)
245
NLEUM & k- model: Examples
(After Apsley et al., 1998)
(After Apsley et al., 1998)
Fig. 1. Plane Channel flow: comparison of solutions with
different models against DNS data ok Kim et al., (1987);
(a) u
2
; (b) v
2
; (c) -uv
Fig. 4. High-lift aerofoil: mean-velocity and Reynolds-stress profiles
at 82.5% chord; (a) U; (b) -uv; (c) u
2
; (d) v
2
.
246
NLEUM & RSTM: Examples
(After Apsley et al., 1998)
Fig. 5. High-lift aerofoil: streamwise normal stress in aerofoil wake.
Fig. 7. Plane asymmetric diffuser: mean-velocity and Reynolds-
stress profiles in the diffuser section (A U; (b) -uv; u
2
; (d) v
2
247
NLEUM & RSTM: Examples
(Ref., Apsley et al., 1998)
Figure 6: Plane asymmetric diffuser: development of the mean velocity
profile along the diffuser
248
Algebraic Stress Models: Assessment
Have the potential of including the extra strain effects, as well as
anisotropy at some cost less than that of RSTMs
Mimic the physical behavior by means of mathematical artifacts
and careful calibration (Apsley et al. 1997)
They need to be modified for low-Re effects and near wall
treatment similar to the two-Eq. models
The advantages seems to be less pronounced in 3D than 2D
flows.
Recommended for problems where anisotropy and certain extra
strain rate effects are known to dominate
249
Influence of Inlet Conditions
(After Sloan, et al., 1986)
(After Hogg, et al., 1989)
Fig. 31. Comparison of predicted and measured centerline
axial velocity profiles for Case 7 based on various inlet
conditions (data from Vu and Gouldin
30
; legend supplied
by Table 19).
250
Initial and Boundary Conditions
Inlet: Prescribe all unknowns from experiments
example: U, V, W, k, etc.
If k, are not available from experiments:
( ) assumed) or given intensity e (turbulenc
U
u
u T ; TuU
2
3
k
inlet
rms
2
inlet
= =
1 . 0 C diameter; Hydraulic D ; D C ;
U
h h
3
rms
= =

l
l
or
Let ( ) model) - (k
k
C 10 - 10
t
2
3 2
t



Outlet: Put outlet boundary away from recirculation regions and set , P =P
ambient
.
0
x
=

Walls: Use wall functions and/or no-slip condition.


Symmetry Axis: Zero derivatives normal to the axis.
251
Numerical Issues: Iteration Convergence
The CFD solution methodology is usually iterative;

n+1
= [A]
n
+S ; n= number of iterations
Erorr = abs(
exact
-
n)
abs[(
n+1
-
n
)/(1 -
max
)]< ;

max
= Largest eigen value of [A]. The eigen values of [A]
must be less than one for convergence
To monitor only E = abs[(
n+1
-
n
)] may be misleading. Better to
monitor overall convergence of profiles over many iterations
The solution must be fully converged befor any assessment is made
(see Ferziger, 1989 for details)
252
Numerical Issues: Grid Convergence
Numerical solutions use finite elements or volumes (cells), called gird
or mesh to discretize the continuum equations (PDEs), to obtain
difference equations (FDEs).
Discretization error = (exact sol. to PDE) - (sol.to FDE)
=
exact
-
num
;
let h = ( x y z)
1/3
, a typical cell size
As h ==> 0,
num
==>
exact
1st order method: E
h
(
h
-
2h
)
2nd order method: E
h
(
h
-
2h
)/3
E
h
must be calculated and minimized if possible
(see e.g. Ferziger, 1989; Celik and Zhang, 1995 for details)
253
Consistency Checks
Check if the boundary conditions are reasonable and correctly
implemented.
Check if 10 < y
+
< 300 (wall functions), and y
+
< 1 (integration through
the sub-layer)
Make sure that grid convergence and iterative convergence are
achieved or characterized. Note that convergence of turbulence
quantities are much more difficult.
For unsteady flow calculations convergence at every time step must be
ensured.
The integral mass, momentum and energy balances must be satisfied
254
Large Eddy Simulation: Introduction
255
Large Eddy Simulation
256
Large Eddy Simulation: Filtered Equations
257
LES Examples (channel Flow)
258
Fluctuating velocity: (a) computed (440000 nodes), (b) Catania and Spessa (1996)
Absolute value of the velocity vectors at 105
0
crank angle (440000 nodes).
LES Examples
(Ref: Celik et al, 1999)
259
Direct Numerical Simulation- DNS
Navier-Stokes equations are not limited to laminar flows. If they can be
solved accurately as is (DNS) turbulence fluctuations can be captured
and statistics can be obtained via post-processing
Require very accurate numerical schemes, at least 4th order in time
and space, or spectral methods (e.g. Fourier, Chebychev expansions)
Must resolve all scales of turbulence down to Kolmogorov scales.
Hence very large number of grid nodes and very small time steps are
necessary. The higher the Re the smaller is the scales, hence the
larger the computational cost and time.
DNS solutions are not suitable to industrial applications but solutions
exist for low Re, simple flows which can be used to bench mark
turbulence models and even experiments!
260
LES and DNS Examples
Fig. 3: Longitudinal hairpin vortices strained behind a backward-facing
step; simulation without subgrid model.
Fig. 5: Plane-averaged velocity, scaled Smagorinsky model
Fig. 4: Time evolution of the wall shear, scaled Smagorinsky model.
(Zang et al., 1993; Galperin and Orszag, editors)
(Lesieur et al., 1993; Galperin and Orszag, editors)
(Zang et al., 1993; Galperin and Orszag, editors)
Scaled Smagorinsky model.
--------- RNG model
______ Dynamic eddy viscosity model.
fine direct simulation
261
LES and DNS Examples
Fig. 7: Plane-averaged shear Reynolds stress Fig. 6: Plane-averaged rms turbulent fluctuations
(Zang et al., 1993; Galperin and Orszag, editors)
Scaled Smagorinsky model.
--------- RNG model
______ Dynamic eddy viscosity model.
fine direct simulation
262
Summary
A overview of turbulence models for industrial application is presented.
This included most commonly used models staring from zero-equation
models to Reynolds Stress Transport models with an introduction to
LES.
The pros and cons of each model are elucidated to help the CFD users
in selection of an appropriate turbulence model for their application. An
assessment is made with concrete examples.
The boundary conditions, consistency checks and possible pitfalls
particularly w.r.t numerical issues are presented as guidance to model
implementation.
The users are also provided with an extensive list of references for future
reading and as a source of detailed information for numerous models.
263
Concluding Remarks
CFD is still not a mature area which can be used an ordinary software
such asword processing!. It is somewhat of an art. The best method
is the one that is validated for a similar problem being solved!
Validation (the process of testing the performance of a model for the
intended application) is the responsibility of the user. Iteration
convergence, and grid convergence errors must be taken into
account before reaching conclusions.
Verification (the process of ensuring a proper implementation of a
turbulence model into a code) is the responsibility of code developers
but the users must be aware of it.
Best use for CFD is trend analysis and hence reduction in prototpe
laboratory testing in design improvements and in new design concepts.
Some minimal background in the area of fluid mechanics, numerical
methods for partial differential equations, and turbulence, is essential!

You might also like