1

Turbulent Flow Case Studies

Brian Bell, Fluent Inc.
2
Motivation
How do I know which turbulence model and near
wall modeling approach to choose for a given
application?
Understanding of how turbulence modeling issues
affect turbulence model selection and performance
Observation and comparison of behavior of turbulence
models for flows in similar applications
Results will be presented from a variety of flows to help with
this point
3
Outline
Turbulence Model Selection
Turbulence Model Comparisons (1)
Flows of low to moderate complexity
Analysis of Differences Between Turbulence
Models
Treatment of Reynolds Stresses
Near-wall modeling
Turbulence Model Comparisons (2)
Flows of increasing complexity
Advanced Applications
Large Eddy Simulation (LES) and Detached Eddy
Simulation (DES)
4
Turbulence Model Selection
Elements of Turbulent Flows
Overview of Computational Approaches
Opportunities and Challenges
Turbulence Modeling Choices
5
Elements of Turbulent Flows
Feature Space
Thin B.L. flows
Rotating & swirling
flows
Crossflow/Secondary
flows
Rapidly strained flows
Transitional flows &
re-laminarization
Separated &
recirculating flows
Large-scale unsteady structure
Thick BL, mildly
separated flows
Streamwise
vortices
Free shear flows
(BL, mixing layer,
wakes, jets
6
Overview of Computational Approaches
Direct Numerical Simulation (DNS)
Theoretically all turbulent flows can be simulated by numerically solving the
full Navier-Stokes equations. The whole spectrum of scales is resolved and
no modeling is required.
But the cost is too prohibitive! Not practical for industrial flows - DNS is not
available in Fluent.
Large Eddy Simulation (LES)
Solves the spatially averaged (filtered) N-S equations. Large eddies are
directly resolved, but eddies smaller than the mesh sizes are modeled.
Less expensive than DNS, but the amount of computational resources and
efforts are still too large for most practical applications.
Reynolds-Averaged Navier-Stokes (RANS) Equations Models
Solve ensemble-averaged Navier-Stokes equations
All turbulence scales are modeled in RANS.
The most widely used approach for calculating industrial flows.
There is not yet a single turbulence model that can reliably predict all
turbulent flows found in industrial applications with sufficient accuracy.
7
Turbulence Scales and Prediction Methods
8
Turbulence Modeling - Opportunities
Ever-increasing computing power in terms of memory and speed
Numerical error can be made smaller than ever.
Use of several million cells is a norm these days. Tens of
million cells are not uncommon.
We see more and more unsteady RANS (URANS)
simulations , LES
Mesh flexibility allows us to model complex configurations that
could not be modeled previously.
We have a unique opportunity likely to become the first witness
to how different turbulence models work for real-world
problems.
9
Turbulence Modeling - Challenges
There are many other factors affecting CFD predictions.
Choice of solution domain, boundary conditions, numerical error, etc.
User error
Yet turbulence modeling is a pacing item for the fidelity of CFD predictions.
Higher expectation for the fidelity predictions as CFD technology is
matured
Widely varying requirement on accuracy.
No breakthrough in turbulence modeling for industrial flows.
Theres no single, dominantly superior, universally reliable
engineering turbulence model yet.
There are so many models with so many tweaks ...
All this puts a considerable burden on CFD vendors who have to meet the
widely varying needs.
10
FLUENT Suite of Turbulence Models
Core Turbulence Models
Near-wall options
Customization
Auxiliary Models
Spalart-Allmaras one-equation model
Standard k- model
Renomalization-Group (RNG) k- model
Realizable k- model
Wilcox k- model
Menters (SST) k- model
Gibson & Launders RSM
Speziale, Sarkar, and Gatzkis RSM
Detached Eddy Simulation (DES)
Subgrid-scale models for LES
v
2
f model
z Standard wall functions
z Non-equilibrium wall functions
z Enhanced wall treatment
Buoyancy effects
Compressibility effects
Low Re effects
Pressure gradient effects
z Turbulent viscosity
z Source terms
z Turbulence transport equations
11
Turbulence Model Selection
Many factors affect turbulence model
selection
Flow Physics
Computational Resources
Accuracy Requirements
Turnaround Time
Etc.
12
Turbulence Model Comparisons
Part 1: Flows of low to moderate
complexity
Channel flow
Mild adverse pressure gradient, separation and
recirculation
Free shear flows
Low Re flows
13
Channel Flow
Comparison with DNS data of Moser et al. (1999)
for Re

= 395 (Re = 28,600)

DNS data available on web
http://www.tam.uiuc.edu/Faculty/Moser/channel
Calculations performed with k-e, k-w, RSM, V2F
and low-Re models on fine near-wall mesh with
enhanced wall treatment (y+ 1)
Why channel flow?
Relatively easy to run many cases and compare model
results for 2D flow without complexity
14
Results for k- Models
y
+
u
+
k
+
y
+
2

w
t

u
k
k ;

yu
y ;

u ;
u
u
u = = + = =
+ +
Results normalized by:
RNG-DV: RNG model with differential viscosity option enabled
This does not appear to have noticeable effect for this flow
Models predict similar velocity profiles, peak tke values
15
Results for k- and V2F Models
All models predict similar velocity profiles
SKO and V2F predict TKE better than k- models for this flow
SST model calculations performed without transitional flows
option. Would this have helped with TKE?
y
+
u
+
k
+
y
+
16
RSM Results
y
+
u
+
k
+
y
+
Calculations performed with default pressure strain term (GL) and quadratic pressure strain term
(SSG) using wall boundary conditions obtained from the k-equation and from the individual Reynolds
stresses (BC)
The wall boundary condition treatment does not appear to have much effect for this flow
The quadratic pressure strain model is not intended for use in the viscous sublayer.
Results from V2F and k- model appear to be more accurate for this flow
17
Results for Low-Re Models
y
+
u
+
k
+
y
+
Calculations performed with Lam Bremhorst and Launder-Sharma low Re k-
models
The Lam Bremhorst model appears slightly more accurate than the other
variations of k- models shown in a previous slide
The Launder-Sharma model does not appear to have been calibrated for this type
of flow
18
Are Results Grid-Independent?
Results shown for RSM and V2F. Similar agreement seen for standard k-
and standard k- models
y
+
u
+
k
+
y
+
y
+
u
+
k
+
y
+
19
What About Triangular Cells?
Standard k- model: For quadrilateral cells
and boundary layer mesh, y+ = 1. For
triangular cells, 0.35 < y+ < 0.7.
With sufficient mesh resolution, results are
nearly identical for quad and tri meshes
y
+
u
+
k
+
y
+
20
2d Backstep
Experiments conducted at NASA Ames (Driver and
Seegmiller, 1985); Re
H
= 3.74 x 10
4
, = 0 deg.
The flow features re-circulation, reattachment, and re-
developing BL.
Computed using SKE, RNG, RKE, and k- models on a
fine mesh.
21
Std. k- Real. k- SST k- Wilcox k- Measured
x
r
/H 5.8 6.6 6.6 7.3 6.4
Predicted reattachment lengths
Skin Friction Coefficient
22
The 2-D back-step of Driver and Seegmiller was computed
using five different near-wall mesh resolutions with the
standard wall functions (SWF) and the enhanced wall
treatment (EWT).

2D Backstep: Mesh (y
+
) Dependency
23
Symmetric Diffuser
Measured by Reneau et al.(1967)
Flow goes from attached to stalled as
included angle, , increases
24
Flow Near End of Diverging Section
RSM, 2 = 12
SKO, 2 = 12
SKO, 2 = 16
RSM, 2 = 16
25
Pressure Recovery Results
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25
2
C
p
2

-

C
p
1
Reneau et al. (1967)
SKO
SST
RSM
SKE
RKE
RNG
SA
SKO (w/o transitional flows option)
26
Summary
Standard k-omega model results most
closely match data
Quality of standard k-omega results
decreases without transitional flows options
active
Results are similar for different models for
small included angle, but differ significantly
after stall begins
27
Measured by Obi (1993), Bruice and
Eaton (1997) - ERCOFTAC test case)
Incompressible, moderately high-Re flow
(Re
H
= 20,000 at the inlet channel) with
separation
Computed using various k- models and
k- models on a fine near-wall mesh (y+
< 1)
Asymmetric Planar Diffuser
28
Comparison with Data
Y(m)
Y(m)
29
Skin-friction predictions
Asymmetric Planar Diffuser
30
2-D Hill
Measured by Baskaran et al. (JFM, Vol. 182, 1987)
High-Re (Re
L
= 1.33 x 10
6
/m) incompressible BL
subjected to pressure gradient, streamline curvature
The main interests are the skin-friction, static pressure, and
extent of the BL separation (x=1.1 m).
Computed using SA, SKE, RKE, and k- models.
31
Results from 2D Hill
Pressure distribution
The k- models predict the
C
p
plateau very closely.
Skin-friction distribution
The k- models give an earlier
and larger separation than other
models.
32
Axisymmetric Bump
Measured by Bachalo and Johnson (1986).
Transonic BL flow with a standing shock and a pocket of
BL separation behind the shock.
Ma = 0.875, Re
c
= 13.6 x 10
6
at freestream.
Computed using S-A, SKE, RKE, KO, SST models.
33
Axisymmetric Bump (2)
Wall pressure predictions
34
RAE 2822 Airfoil
RAE2822 Transonic airfoil
Measured by Cox (1981) (Case 9 in Stanford
database)
The corrected = 2.79 deg., Ma = 0.73, Re = 6.5
x 10
6
Computed using SA, SKE, RKE, and k- models
on a wall function (coarse) mesh
35
RAE 2822 Airfoil
C
p
Predictions
36
RAE 2822
C
f
Predictions
37
RAE 2822 Airfoil Summary
Forces and moment predictions
( = 2.79, Re = 6.5 x 10
6
, Ma = 0.73)
The shock location predicted by the k- models is slightly
upstream of the measured one and the prediction by other
models.
The two k- models give a slightly lower lift coefficient,
but their results are almost identical.
Flow S-A SKE RKE SST k- Wilcox k- Exp.
CL 0.811 0.835 0.820 0.772 0.774 0.803
CD 0.0180 0.0198 0.0189 0.0172 0.0172 0.0168
CM -0.1093 -0.1063 -0.1092 -0.1068 -0.1072 -0.099
38
Axisymmetric Underwater Body
Experiments conducted (Huang et al., 1976) at DTNSRDC
High-Re (ReL= 5.9 x 106), incompressible BL flow with a separation
at around x/L = 0.92, and reattachment at x/L = 0.97.
SKE, RNG, RKE, SA, SKO, SST, RSM and Low Re models tried.
Different near-wall treatments tried.
Modified hull form
39
Axisymmetric Afterbody
Spalart-Allmaras
model (fine mesh)
Std. k- model +
2-layer (fine mesh)
No separation
on afterbody
40
Axisymmetric Afterbody
Model Separates?
Std k- n
RNG k- n
Real. k- y
RSM y
S-A y
Cp
Pressure coefficient on
coarse mesh (y
+
~ 40)
using wall functions
Position (m)
41
Axisymmetric Afterbody
Cp
Pressure coefficient on
fine mesh (y
+
~ 0.5)
using two-layer model
Model Separates?
Std k- n
RNG k- n
Real. k- y?
RSM y
S-A y
42
Axisymmetric Underwater Body
Pressure (C
p
) predictions
Static pressure in the separated
region is over-predicted by k-
models.
Skin-friction predictions
The experiment shows the flow
separates at x/L = 0.92 and
reattaches at x/L = 0.97
k- models gives too large a
separation.
43
Axisymmetric Afterbody
Spalart-Allmaras gives consistent results on both meshes
Separation not predicted by Standard k- on either mesh
RSM separates on both meshes
C
p
on body somewhat overpredicted on coarse mesh
Wall reflection term, or quadratic pressure-strain
term, necessary to obtain coarse mesh separation
Subtle separation illustrates effect of near-wall treatment
Realizable k- has smaller separation bubble on fine
mesh
Difficult to get grid-independent solutions using wall
functions --- would a low-Re formulation work?
44
Axisymmetric Afterbody
Cp
Position (m)
Model Separates?
V2F y
Abid n
Launder-Sharma n
Yang-Shih n
Abe-Kondo-Nagano n
Chang-Hsieh-Chen n
Pressure coefficient on
fine mesh using Low-Re
models
45
Axisymmetric Afterbody
Low-Re models using damping
functions do not predict the separation
Durbins V2F (4-equation) model
predicts separation
46
Low-Re Backstep
Re = 5,100
Comparison with DNS data of Le and Moin
(1994)
Comparison of Standard k- + 2-layer, Yang-Shih
low-Re model and V2F low-Re model
47
Low-Re Backstep
C
p
C
fx
Pressure coefficient and x-component of skin friction
2-layer model less accurate than V2F and Yang-Shih
48
X
Velocity
X
Velocity
X
Velocity
X
Velocity
X/h = 1 X/h = 3
X/h = 5
X/h = 7
Low-Re Backstep
49
X/h = 1
X/h = 3
X/h = 5 X/h = 7
Y
Velocity
Y
Velocity
Y
Velocity
Y
Velocity
Low-Re Backstep
50
Low-Re Backstep
Contours of Re
y
< 200
For 2-layer model where Re
y
< 200, and
t
are
prescribed algebraically. Much of the flow is in
this region
2-layer model is not always a good substitute for a
low-Re model
51
Measured by Graziani (1980) -
P&W Aircraft Group in UTC
The local heat transfer rate was
measured.
A 2-D model of the original (3d-
D) configuration at the mid-span
The suction side flow undergoes a
laminar-to-turbulent transition.
Several near-wall models and
low-Re models were tested
Two-layer zonal model
k- models with and without
the transitional flow option
2D Turbine Blade
52
The k- models with the transitional flow
option give much better results than other models
on the suction side.
Results for 2D Turbine Blade
53
Ota & Kan 151x75 quad mesh
Impinging Flow Over a Blunt Plate
54
Standard k- model
Reynolds-Stress model
(exact)
Contours of TKE production
Blunt Plate
The standard k- model gives spuriously large turbulent
kinetic energy on the front face, underpredicting the size of
the recirculation.
55
Results: Blunt Plate Skin Friction
Standard k-
Realizable k-
Experimentally observed
reattachment point is at x/d = 4.7
Predicted separation bubble
56
Compressible Mixing Layer
A
B
STREAM A
Total Pressure: 487 kPa
Static Pressure: 36 kPa
Total Temperature: 360 K
Mach Number: 2.35
k: 74 m/s
: 62,300 m/s
STREAM B
Total Pressure: 38 kPa
Static Pressure: 36 kPa
Total Temperature: 290 K
Mach Number: 0.36
k: 226 m/s
: 332,000 m/s
300 mm
72 mm
Comparison with experimental data of Goebel and Dutton (1991)
x
=


5
0
m
m
x
=

1
0
0
m
m
x
=

1
5
0
m
m
x
=

1
7
5
m
m
57
Velocity Predictions
X= 50mm
X= 100mm
X= 150mm
X= 175mm
58
TKE Predictions
X= 50mm
X= 100mm
X= 150mm
X= 175mm
59
Conclusions from Mixing Layer
RNG and Realizable k-epsilon models more
accurately predict velocity profiles in mixing layer
RNG and Realizable k-epsilon models reasonably
accurate in predicting tke in low-speed layer, but
overpredict tke in high-speed layer
RSM and standard k-epsilon results very similar in
this case
60
Confined Swirling Coaxial Jet
Inner
Jet
Swirler
Computational
Domain
Swirling Outer
Jet
An axisymmetric representation of the geometry [Roback, R. and Johnson, B.V., 1983]
Calculations performed on fine mesh with y+ ~ 1
Velocity and turbulence
profiles specified at inlet to
computational domain
X=5mm X=25mmX=51mm X=102mm X=203mm
Inner injector
Annular injector
61
Velocity and Stream Function
62
Results at x = 5 mm Section
-0.5
0
0.5
1
1.5
2
2.5
0 0.2 0.4 0.6 0.8 1
y/R0
A
x
i
a
l

v
e
l
o
c
i
l
y

u

(
m
/
s
)
exp
u-SKE
u-RKE
u-RNG
u-RSM
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1
y/R0
S
w
i
r
l

v
e
l
o
c
i
l
y

w

(
m
/
s
)
exp
w-SKE
w-RKE
w-RNG
w-RSM
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0 0.2 0.4 0.6 0.8 1
y/R0
R
a
d
i
a
l

v
e
l
o
c
i
l
y

v

(
m
/
s
)
exp
v-SKE
v-RKE
v-RNG
v-RSM
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1
y/R0
I
n
n
e
r

j
e
t

m
o
l
e

f
r
a
c
t
i
o
n
exp
Xjet-SKE
Xjet-RKE
Xjet-RNG
Xjet-RSM
63
Results at x = 25 mm Section
-1
-0.5
0
0.5
1
1.5
2
0 0.2 0.4 0.6 0.8 1
y/R0
A
x
i
a
l

v
e
l
o
c
i
l
y

u

(
m
/
s
)
exp
u-SKE
u-RKE
u-RNG
u-RSM
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 0.2 0.4 0.6 0.8 1
y/R0
R
a
d
i
a
l

v
e
l
o
c
i
l
y

v

(
m
/
s
)
exp
v-SKE
v-RKE
v-RNG
v-RSM
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 0.2 0.4 0.6 0.8 1
y/R0
S
w
i
r
l

v
e
l
o
c
i
l
y

w

(
m
/
s
)
exp
w-SKE
w-RKE
w-RNG
w-RSM
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1
y/R0
I
n
n
e
r

j
e
t

m
o
l
e

f
r
a
c
t
i
o
n
exp
Xjet-SKE
Xjet-RKE
Xjet-RNG
Xjet-RSM
64
Results at x = 51 mm Section
-1
-0.5
0
0.5
1
1.5
0 0.2 0.4 0.6 0.8 1
y/R0
A
x
i
a
l

v
e
l
o
c
i
l
y

u

(
m
/
s
)
exp
u-SKE
u-RKE
u-RNG
u-RSM
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0 0.2 0.4 0.6 0.8 1
y/R0
R
a
d
i
a
l

v
e
l
o
c
i
l
y

v

(
m
/
s
)
exp
v-SKE
v-RKE
v-RNG
v-RSM
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0 0.2 0.4 0.6 0.8 1
y/R0
S
w
i
r
l

v
e
l
o
c
i
l
y

w

(
m
/
s
)
exp
w-SKE
w-RKE
w-RNG
w-RSM
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 0.2 0.4 0.6 0.8 1
y/R0
I
n
n
e
r

j
e
t

m
o
l
e

f
r
a
c
t
i
o
n
exp
Xjet-SKE
Xjet-RKE
Xjet-RNG
Xjet-RSM
65
Results at x = 102 mm Section
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1
y/R0
A
x
i
a
l

v
e
l
o
c
i
l
y

u

(
m
/
s
)
exp
u-SKE
u-RKE
u-RNG
u-RSM
-0.2
-0.15
-0.1
-0.05
0
0.05
0 0.2 0.4 0.6 0.8 1
y/R0
R
a
d
i
a
l

v
e
l
o
c
i
l
y

v

(
m
/
s
)
exp
v-SKE
v-RKE
v-RNG
v-RSM
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0 0.2 0.4 0.6 0.8 1
y/R0
S
w
i
r
l

v
e
l
o
c
i
l
y

w

(
m
/
s
)
exp
w-SKE
w-RKE
w-RNG
w-RSM
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0 0.2 0.4 0.6 0.8 1
y/R0
I
n
n
e
r

j
e
t

m
o
l
e

f
r
a
c
t
i
o
n
exp
Xjet-SKE
Xjet-RKE
Xjet-RNG
Xjet-RSM
66
Results at x = 203 mm Section
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 0.2 0.4 0.6 0.8 1
y/R0
A
x
i
a
l

v
e
l
o
c
i
l
y

u

(
m
/
s
)
exp
u-SKE
u-RKE
u-RNG
u-RSM
-0.16
-0.14
-0.12
-0.1
-0.08
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0 0.2 0.4 0.6 0.8 1
y/R0
R
a
d
i
a
l

v
e
l
o
c
i
l
y

v

(
m
/
s
)
exp
v-SKE
v-RKE
v-RNG
v-RSM
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0 0.2 0.4 0.6 0.8 1
y/R0
S
w
i
r
l

v
e
l
o
c
i
l
y

w

(
m
/
s
)
exp
w-SKE
w-RKE
w-RNG
w-RSM
0
0.02
0.04
0.06
0.08
0.1
0.12
0 0.2 0.4 0.6 0.8 1
y/R0
I
n
n
e
r

j
e
t

m
o
l
e

f
r
a
c
t
i
o
n
exp
Xjet-SKE
Xjet-RKE
Xjet-RNG
Xjet-RSM
67
Conclusions
Johnson-Roback test case was run using the k- turbulence
models and the Reynolds Stress model (RSM)
Velocities (axial, radial and swirl) showed good agreement
with data
RNG k- model performed the best in predicting velocities and
mixing
Mixing results were poor downstream (x > 25 mm)
A possible cause for this behavior is the presence of large,
unsteady flow structures that cannot be captured in a RANS
framework.
68
Analysis of Differences Between
Turbulence Models
Treatment of Reynolds stresses
Treatment of terms in model equations
Treatment of wall boundary conditions
Near-wall modeling
69
RANS Equations
Reynolds Averaged Navier-Stokes equations:
How to model the Reynolds Stresses, R
ij
= ?
1. Boussinesq hypothesis
Isotropic eddy viscosity based on dimensional analysis
2. Reynolds stress transport equations
No assumption of isotropy, but more computationally
expensive and requires additional modeling
( )
j
j i
j
i
j i k
i
k
i
x
u u
x
U
x x
p
x
U
U
t
U

+
|
|
.
|

\
|

=
|
|
.
|

\
|

j i
u u
70
The Boussinesq Approach
Relates the Reynolds stresses to the mean flow by a turbulent
(eddy) viscosity, m
t
Relation is drawn from analogy with molecular transport of
momentum.
Assumptions valid at molecular level, not necessarily valid
at macroscopic level
m
t
is a scalar (R
ij
aligned with strain-rate tensor, S
ij
)
Taylor series expansion valid if l
mfp
|d
2
U/dy
2
| << |dU/dy|
Average time between collisions l
mfp
/ v
th
<< |dU/dy|
-1
|
|
.
|

\
|

= =
i
j
j
i
ij ij ij
k
k
ij j i ij
x
U
x
U
S k
x
U
S u u R
2
1
;
3
2
3
2
2
t t

ij xy
S v u t 2 =

=
71
Modeling
t
Oh well, focus attention on modeling
t
anyways.
Basic approach made through dimensional arguments
Units of
t
=
t
/ are [m
2
/s]
Typically one needs 2 out of the 3 scales:
velocity - length - time
Models classified in terms of number of transport
equations solved, e.g.,
zero-equation
one-equation
two-equation
72
K- Model of Wilcox (1998)
Originally conceived by Kolmogorov (1942) - The first
two-equation model
Based on Kolmogorov-Prandtl relation:
Turbulent viscosity
The dependency of * upon Re
T
was designed to
recover the correct asymptotic values in the limiting
cases.
k
k k
k k
t

where
, , l

k
t
*
=

k
R
R
R
T k
i
i
k T
k T
= =
= =
+
+
=
Re , 6
125
9
,
3
,
Re 1
Re
*
0
*
0 *
bulent) (fully tur as 1
*

T
Re

1

k

specific dissipation rate

(SDR)
Eddy turn-over frequency
73
TKE Equation for k- Model
4 4 4 3 4 4 4 2 1
43 42 1
3 2 1
k of Diffusion
k of rate n Dissipatio
k of production
(
(

|
|
.
|

\
|
+

=
j k
t
j j
i
ij
x
k
x
k f
x
U
Dt
Dk

*
*
( )
( )
( )
0 . 2
8 ,
Re 1
Re 15 4
100
9
2
3
1
4
4
*
* *
=
=
+
+
=
(

+ =
k
T
T
i
t i
R
R
R
M F

( )
4 4 3 4 4 2 1
parameter diffusion - cross
j j
k
k
k
k
k
t t
t t t t
t t
t
x x
k
f
RT a M
a
k
M
M M M M
M M
M F

>
+
+

=
= = =

>

=

3
2
2
0
2
2
0
2
0
2
0
1
,
0
400 1
680 1
0 1
,
4
1
,
2
0
*
Note the dependence upon Re
T
, M
t
, and
k
.
Dilatation dissipation is accounted for via M
t
term. Improves high-Mach
number free shear and boundary layer flow predictions - reduces spreading rates
The cross-diffusion parameter (
k
) is designed to improve free shear flow predictions.
Transitional Flows option controls all Re
T
Terms, Shear Flow Corrections option
controls cross-diffusion, parameter, Compressibility Effects option controls M
t
terms.
74
SDR Equation for k- Model
(
(

|
|
.
|

\
|
+

=
j
t
j j
i
ij
x x
f
x
U
k Dt
D

2
( )
( )
|
|
.
|

\
|

=
|
|
.
|

\
|

=

=
+
+
=
(

+ =
= = = =
+
+
=

i
j
j
i
ij
i
j
j
i
ij
ki jk ij
t
i
i
i
T
T
x
U
x
U
x
U
x
U
S
S
f M F
R
R
R
2
1
,
2
1
,
80 1
70 1
,
2
3
1
0 . 2 , 95 . 2 ,
9
1
,
25
13
,
Re 1
Re
3
*
*
0
0
*

Note the dependence upon Re

T
, M
t
, and

.
Vortex-stretching parameter (

) designed to remedy the plane/round-jet anomaly

Transitional Flows option controls all Re
T
Terms, Shear Flow Corrections option
controls vortex stretching parameter, Compressibility Effects option controls M
t
terms
75
Faults in the Boussinesq Assumption
Boussinesq: R
ij
= 2
t
S
ij
Is simple linear relationship sufficient?
R
ij
is strongly dependent on flow conditions and history
R
ij
changes at rates not entirely related to mean flow
processes
R
ij
is not strictly aligned with S
ij
for flows with:
sudden changes in mean strain rate
extra rates of strain (e.g., rapid dilatation, strong
streamline curvature)
rotating fluids
stress-induced secondary flows
Modifications to two-equation models cannot be
generalized for arbitrary flows.
76
0 ) ( ) ( =

i j j i
u NS u u NS u
Reynolds Stress Models
Starting point is the exact transport equations for the
transport of Reynolds stresses, R
ij
.
six transport equations in 3d
Equations are obtained by Reynolds-averaging the product
of the exact momentum equations and a fluctuating velocity.

The resulting equations contain several terms that must be

modeled, including pressure strain redistribution, turbulent
diffusion and dissipation.
Additional modifications may be required for these
terms to ensure correct behavior in the near-wall region
Because there is no assumption of isotropy, Reynolds Stress
Models have significant advantages over EVM for flows
described on previous slide
77
Reynolds Stress Transport Equations
k
ijk
ij ij ij
ij
x
J
P
Dt
DR

+ + =
Generation
|
|
.
|

\
|

k
i
k j
k
j
k i ij
x
U
u u
x
U
u u P
|
|
.
|

\
|

i
j
j
i
ij
x
u
x
u
p
k
j
k
i
ij
x
u
x
u

2
Pressure-Strain
Redistribution
Dissipation
Turbulent
Diffusion
(modeled)
(related to )
(modeled)
(computed)
(incompressible flow w/o body
forces)
Reynolds Stress
Transport Eqns.
43 42 1
43 42 1 4 4 3 4 4 2 1
) (
j i
k
k j i ik j jk i ijk
u u
x
u u u u p u p J

Pressure/velocity
fluctuations
Turbulent
transport
Molecular
transport
78
Importance of Near-Wall Turbulence
Walls are main source of vorticity and turbulence.
Accurate near-wall modeling is important for most
engineering applications.
Successful prediction of frictional drag for external
flows, or pressure drop for internal flows, depends on
fidelity of local wall shear predictions.
Pressure drag for bluff bodies is dependent upon extent
of separation.
Thermal performance of heat exchangers is determined
by wall heat transfer whose prediction depends upon
near-wall effects.
79
Near-Wall Modeling Issues (1)
k- and RSM models are valid in the turbulent
core region and through the log layer.
Some of the modeled terms in these equations are based
on isotropic behavior.
Isotropic diffusion (
t
/)
Isotropic dissipation
Pressure-strain redistribution
Some model parameters based on experiments of isotropic
turbulence.
Near-wall flows are anisotropic due to presence of
walls.
Special near-wall treatments are necessary since
equations cannot be integrated down to wall.
80
Near wall modeling issues (2)
K-, Spalart-Allmaras and V2F models require no
special near-wall treatment.
Designed to predict correct behavior when integrated to
the wall (first grid point in viscous sublayer)
FLUENTs implementation of these models is
sufficiently robust for use on coarse meshes (first grid
point in log-law region)
Low Reynolds number variations of standard k-
models use damping functions to attempt to
reproduce correct near wall behavior
81
Flow Behavior in Near-Wall Region
Velocity profile exhibits layer structure identified
from dimensional analysis
Inner layer
viscous forces rule, U = f(,
w
, , y)
Outer layer
dependent upon mean flow
Overlap layer
log-law applies
k U/u

k production and dissipation are nearly

equal in overlap layer
turbulent equilibrium
dissipation >> production in sublayer
region
82
The goal of near wall modeling is to reproduce flow behavior illustrated on previous
slide. Two choices are available in FLUENT
Wall Functions
In general, wall functions are a collection or set of laws that serve as boundary conditions for
momentum, energy, and species as well as for turbulence quantities.
The Standard and Non-equilibrium Wall Function options
refer to specific sets designed for high Re flows.
The viscosity affected, near-wall region is not resolved.
Near-wall mesh is relatively coarse.
Cell center information bridged by empirically-based wall
functions.
Enhanced Wall Treatment or Low-Re Option
This near-wall model combines the use of enhanced wall
functions and a two-layer model.
Used for low-Re flows or flows with complex near-wall
phenomena.
Generally requires a very fine near-wall mesh capable of
resolving the near-wall region.
Turbulence models are modified for inner layer.
EWT is only option available for k- models, TKE b.c. same as k- model, value in wall-
adjacent cell determined by distance from wall and friction velocity.
Near-Wall Modeling Options
inner
layer
outer
layer
83
Turbulence Model Comparisons
Part 2: Flows of increasing complexity
Streamline curvature
Rotation
Swirl
Impinging flows
Secondary flows
Three dimensional effects
84
Comparison with experimental data of Monson et al. (1990)
2D U-Bend
85
Streamwise Velocity Comparisons
r*
U/U
ref
= 90
= 0
U/U
ref
r*
r*
U/U
ref
= 180
86
C
p
C
p
S/H S/H
Inner
Wall
Outer
Wall
Pressure Coefficients
87
Standard k- Spalart-Allmaras
RNG k- RSM
Stream Function Contours
88
Lessons from 2-D U-Bend
Only the RSM correctly predicts the
effects of streamline curvature
Standard k- does not predict any
separation
RNG k- predicts slight separation
Both RSM and Spalart-Allmaras predict
significant separation
89
Turbulent Vortex Breakdown
Comparison with experimental data of
Sarpkaya (1999)
2D axisymmetric calculation
Simulation courtesy of R. Spall, Utah State
University
90
Axial
Velocity
r/r0
x/r0 = 5
Axial
Velocity
r/r0
x/r0 = 8.3
Comparisons of Axial Velocity Profiles
91
Comparisons of Swirl Velocity Profiles
Swirl
Velocity
r/r0
x/r0 = 5
Swirl
Velocity
r/r0
x/r0 = 8.3
92
Turbulent Vortex Breakdown Summary
k- model cannot predict vortex breakdown
in high strain rates, turbulent kinetic energy
increases and increases turbulent viscosity
RNG k- model is better (additional strain-rate
term and an ad hoc swirl correction reduce the
turbulent viscosity) but not acceptable
RSM results show significant improvement
for this and many other swirling flow cases
93
Non-dimensional
Parameters
Ratio H/D
Reynolds Re
Pr
Calculation of:
h(x)=/(T
p
-T
0
)
Nu=h(x)L/
f
T
0
T
p
or
Example: Impinging Jet
94
Modeling Challenge
Ex: Standard k- model (SKE):
Overestimates Nusselt number (30% - 80%) in the vicinity of the
stagnation point.
Single peak in Nusselt number for H/D < 3
Simulation SKE: H/D=2, Re=70000
Heat transfer calculation
Nu*=Nu/Re
0.7
Exp: Baughn, Shimizu (1989)
95
Modeling Challenge: Complex Flow
Free jet turbulence , stagnation point, boundary layer, strong
streamline curvature, transition ? ...
Free jet
Stagnation zone
Boundary layer & transition
Wall Jet
?
Impinging Jet Flow Characteristics
96
Turbulence intensity and
Nusselt Number (Lytle and Webb (1991))
The second peak in the Nusselt number corresponds to the increase
in turbulence intensity
Laminar/Turbulent transition follows the relaminarization of the
flow after impact => k- model is the model of choice
Effect of Transition
97
Decreased production of turbulence is observed at the stagnation point
Two equation models over-estimate tke production at the stagnation
point
SKE
RKE RNG
=> Can the production of turbulent kinetic energy be reduced?
TKE Production at Stagnation Point
98
Turbulent kinetic energy transport equation:
Modification of production term (F.R. Menter,
92):
P
k
=
t

2
j
t i i
t
j k j j j i
u
u u Dk k
Dt x x x x x

( | |

| |

= + + +
| (
|
|

(
\ .
\ .
Diffusion production dissipation
Fluent 6.1: define/model/viscous/turbulent-expert/
Impinging Jet Example
99
Modification of TKE production term
Production based on S Production based on :
k- model
Effect of Modified Production Term
100
The following RANS models were evaluated
Standard k- model SKE
RNG k- model: minimizes TKE at stagnation point.
k- model: laminar/turbulent transition in boundary layer.
k- with production based on rotation rate (KWW)
V2F model: effect of near-wall anisotropy accounted for with transport
equation for v
2
Flow Characteristics: Pr=0.7
RE=23 000, H/D=2
RE=23 000, H/D=6
Flow Calculations
101
Results: H/D=2, RE=23 000
Mean velocity
profiles
KWW
RNG
V2F
r/D=1
r/D=2
Impinging Jet: Velocity Profiles
102
Results: H/D=2, RE=23 000
SKE
RNG
KWW
SKE
RNG
KWW
Nu*
Results from Two-Equation Models
TKE*
103
Results: H/D=2, RE=23 000
Comparison of k- and V2F models
Nu*
TKE*
V2F
KWW
V2F
KWW
104
In the vicinity (r<2D) of the stagnation point
H/D=2, Re=23 000
Standard k-e model (SKE) is not suitable
Heat transfer coefficient: up to 80 % error.
30-40 % energy imbalance.
Modified k- model (KWW): good estimation of the
stagnation point and transition zone
Heat transfer coefficient: 25 %
Energy imbalance less than 5%
V2f:
Best results at low Reynolds numbers
Impinging Jet: Conclusions (1)
105
Conclusions:
H/D=6 (not shown):
Two-equation models give similar results (heat transfer coefficient
over predicted by 20%).
The results obtained with the V2F model are still superior to the
results from the two-equation models.
Computational expense:
Two-equation models: Grid independence achieved with 10,000 cells,
acceptable results with 6,000 cells
V2F: 30% - 40% more expensive for similar mesh. 30,000 cells
needed for grid independent results.
Impinging Jet: Conclusions (2)
106
Flow configuration:
Johnston et al. (1972)
Re
H
= 11,500
Ro = 0.21
Flow in a Rotating Channel
Represents flows through
rotating internal passages
(e.g. turbomachinery
applications)
Rotation affects mean axial
momentum equation
through turbulent stresses.
Rotation makes mean axial
velocity asymmetrical.
Computations are carried
out using SKE, RNG, RKE
and RSM models are with
the standard wall
functions.
107
Flow in a Rotating Channel
Predicted axial velocity profiles (Re
H
= 11.500, Ro = 0.21)
Symmetric
profiles
108
Flow in a Cyclone
40,000 cell hexahedral
mesh
High-order upwind
scheme was used.
Computed using SKE,
RNG, RKE and RSM
models with the
standard wall functions
Represents highly
swirling flows (W
max
=
1.8 U
in
)
0.97 m
0.1 m
0.2 m
U
in
= 20 m/s
0.12 m
109
Velocity Profiles in Cyclone
Tangential velocity profile at 0.41 m below the vortex finder
110
Flow in a Triangular Duct
Duct flows exhibit secondary flows caused by
anisotropy of Reynolds stresses
Solved using RSM, SST and RNG with swirl
and differential viscosity options.
Periodic flow with Re = 9870. 14,772 hex
cells, fine near wall mesh (y+ < 3).
111
Streamwise Velocity Contours
Similar streamwise velocity profiles
predicted by all models
RNG SST RSM
112
Transverse Velocity Components
Only the Reynolds stress model predicts flow in
plane normal to streamwise direction
RNG SST RSM
113
Secondary Flow Details
Recirculating secondary flow patterns
caused by anisotropy of Reynolds stresses
RSM SST & RNG RSM
114
Friction Factor
Is prediction of secondary flows important in
predicting pressure drop?
0.0349
0.0252
0.0301
0.0373
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
Braga and
Saboya (1996)
SST RNG RSM
h
2
2
D

v
z p

=
f
115
Measured by Prof. Simpsons group at VPI
Incompressible (Ma = 0.15), high-Re (Re = 4.2 x 10
6
) flow
The most salient features are the cross-flow (open) separation,
stream-wise vortices, and vortex-lift (nonlinear).
Computed using SA, SKE, RKE, k-, and RSM models.
6:1 Prolate Spheroid at Incidence
116
6:1 Prolate Spheroid at Incidence
117
Flow in a Transition Duct
Fully developed inlet profiles, 64,240 hexahedral cells,
Re = 3.9x10
5
.
Calculated with k-, k- and RSM with non-equilibrium wall
functions (30 < y+ < 70)
Measurements by Davis and Gessner (1992) taken at centerline and
locations shown below
Station 5
Station 6
Inlet
Outlet
118
All models predict similar transverse velocity pattern at
station 6
Secondary flow induced by transition from circular to
rectangular duct in this case
Velocity Vectors at Station 6
RSM SST SKO SKE
119
Pressure and Skin-Friction Coefficients
Station 5 Station 5
Station 6 Station 6
120
Centerline Pressure Coefficient
121
Transition Duct Summary
All models considered predict skin friction
and pressure coefficients qualitatively
Except for standard k- model, all models
considered here predict similar results.
Experimental velocity contours (not shown)
suggest that velocity field predicted by SKE
is slightly less accurate than the other
models.
122
Example: Ship Hull Flow
Experiments: KRISOs 300K VLCC (1998)
Complex, high Re
L
(4.6 10
6
) 3D Flow
Thick 3D boundary layer in moderate pressure gradient
Streamline curvature
Crossflow
Free vortex-sheet formation
(open separation)
Streamwise vortices embedded
in TBL and wake
Simulation
Wall Functions used to manage mesh size.
y
+
30 - 80
Hex mesh ~200,000 cells
Contours of axial velocity compared with simulations.
123
Contour Plots of Axial Velocity
SKO and RSM models capture characteristic shape at propeller plane.
SA RKE RNG
SKE SKO RSM
124
0.486
0.482
0.537
0.539 0.538
0.583
0.561 0.56
0.557
0.3
0.35
0.4
0.45
0.5
0.55
0.6
S
-
A
S
K
E
R
N
G
R
K
E
K
O
-
S
S
T
K
O
-
W
i
l
c
o
x
R
S
M
-
G
L
R
S
M
-
S
S
G
E
x
p
.
w
4.051
4.216
4.145 4.149
4.2
4.258
4.048 4.06 4.056
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
S
-
A
S
K
E
R
N
G
R
K
E
K
O
-
S
S
T
K
O
-
W
i
l
c
o
x
R
S
M
-
G
L
R
S
M
-
S
S
G
E
x
p
.
1
0
0
0
x
C
T
,

C
F
,

C
V
P
CT
CF
CVP
Wake Fraction and Drag
Though SKO (and SST)
were able to resolve salient
features in propeller plane,
not all aspects of flow
could be accurately
captured.
Eddy viscosity model
RSM models accurately
capture all aspects of the
flow.
Complex industrial flows
provide new challenges to
turbulence models.
dA
U
u
A
w
P
A P
|
|
.
|

\
|
=
0
1
1
125
Advanced Applications
Large Eddy Simulation (LES) and Detached
Eddy Simulation (DES)
Theory
Applications
126
Large Eddy Simulation (LES)
Recall: Two methods can be used to eliminate the need to resolve
small scales.
Reynolds Averaging Approach
Periodic and quasi-periodic unsteady flows
Filtering (LES)
Transport equations are filtered such that only larger eddies need be
resolved.
Difficult to model large eddies since they are
anisotropic
subject to history effects
dependent upon flow configuration, boundary conditions, etc.
Smaller eddies are modeled.
Typically isotropic and so more amenable to modeling.
Deterministic unsteadiness of large eddy motions can be resolved.
127
In finite-volume schemes, the cell size in a mesh can determine the
filter width.
e.g., in 1-D,
more or less information is filtered as is varied.
In general,
where the subgrid scale (SGS) velocity,
( ) ( ) V d x t u
V
t x u
V
i i



r r
r
r
r
, ; ,
1
,
where V is the volume of cell
u
i
x
i

u
i
u
i
|
.
|

\
|

+

x
x
i
i i
d u
dx
d x u x u
) (
2
1
2
) ( ) (
i i i
u u u =

Resolvable-scale filtered velocity

Filtering
128
The governing equations for LES are obtained by filtering
(space-averaging) Navier-Stokes equations:
The SGS stresses consist of terms that must be modeled:
j
ij
j
i
j i j
j i
i
i
i
x x
u
x x
p
x
u u
t
u
x
u

|
|
.
|

\
|

1
) (
0
j i j i ij
u u u u
j i j i j i j i j i
u u u u u u u u u u

+

+ =
LES - Governing Equations
129
The subgrid-scale stress is modeled by;
The subgrid-scale eddy viscosity

is modeled by:
Smagorinskys subgrid-scale model
RNG-based subgrid-scale model
|
|
.
|

\
|

=

i
j
j
i
ij ij ij kk ij
x
u
x
u
S S
2
1
; 2
3
1

( )
ij ij S
S S C 2
2

( )
ij ij RNG S
eff s
eff
S S C C H 2 , 1
2
3 / 1
3
2

(
(

|
|
.
|

\
|
+ +

Smagorinsky constant C
s
varied from flow to flow
SGS Stress Modeling
130
Iso-surface of instantaneous
vorticity magnitude colored by
velocity angle
LES Example: Dump Combustor
A 3-D model of a lean premixed combustor studied by Gould (1987) at
Purdue University
Non-reacting (cold) flow was simulated with a 170K cell hexahedral
mesh using second-order temporal and spatial discretization schemes.
131
Simulation done for:
Computed using
RNG-based subgrid-
scale model
Mean axial velocity at x/h = 5
( ) 150 Re 10 Re
5
=
d
LES Example: Dump Combustor
Mean axial velocity
prediction at x/h = 5;
132
LES Example: Dump Combustor
RMS velocities predictions at x/h = 10;
133
Detached Eddy Simulation (DES)
Hybrid RANS/LES Modeling Approach
DES approach combines an unsteady RANS version of the Spalart-
Allmaras model with a filtered version of the same model
A practical and computationally efficient alternative to LES for predicting
flow around high-Reynolds-number, high-lift airfoils
To enable DES
1. Activate S-A model in viscous panel
2. In TUI, enter
/define/models/viscous/turbulence-expert/detached-eddy-
simulation? yes
134
DES: Calculation of Turbulent/SGS Viscosity
Recall that for S-A model, the distance from the wall, d, plays a major
role in the terms for production and destruction of turbulent viscosity

This creates two separate regions in the flow calculation

Near walls, the flow calculation reduces to unsteady RANS with
the S-A model
In the high-Re turbulent core region, where large turbulence scales
play a dominant role, DES recovers LES with a one-equation
model for the sub-grid scale viscosity
( ) direction - z or y in x, dimension cell maximum , 65 . 0 C , C d, min d
~
such that d
~
e, lengthscal new a by model A - S in the everywhere replaced is d DES, In
des des
= = =
LES Region
RANS Region
( ) d d
~
=
( ) =
DES
C d
~

135
DES Example: Airfoil at High Incidence
Angle of attack 13.3, Re = 2.1x10
6
360 x 64 x 16 mesh (368,640 cells)
Affordable for real engineering applications
Cell count decreased by order of magnitude compared with
successful LES simulation of same airfoil
136
DES Example: Results
Instantaneous x-vorticity contours Time-averaged velocity vectors at trailing edge
137
DES Example: Airfoil Grid
Grid Detail Instantaneous y+ values
138
DES (3)
P
r
e
s
s
u
r
e

a
n
d

s
k
i
n

f
r
i
c
t
i
o
n

c
o
e
f
f
i
c
i
e
n
t
s
T
i
m
e
-
a
v
e
r
a
g
e
d

a
n
d

r
m
s

v
e
l
o
c
i
t
y

p
r
o
f
i
l
e
s

i
n

t
h
e

w
a
k
e
139
Conclusions
For flows with strong streamline curvature,
rotation, swirl or three-dimensional boundary
layers, RSM results are generally more accurate.
For less complex flows there does not appear to be
a demonstrable advantage to using RSM. SKO,
SST, RNG and RKE demonstrate satisfactory
results for a wide range of flows.
Check comparisons between models to see which is
best for your particular application
140
Conclusions
For flows with separation, transition, low Re
effects, impingement, the standard k- model
(with transitional flows option) appears more
accurate than other 2-equation models
Transitional flows option can also be applied to SST
model, but it is not part of the original formulation
The V2F model was the most accurate model in
all cases where it was considered
Intended for flows where near-wall turbulence is of
primary importance, i.e. separating and recirculating
flows, heat transfer
141
Conclusions
The standard k- model was seen to be less
accurate than the other models considered here for
a wide variety of different flows
Ensure proper near wall grid resolution and near
wall treatment.
If possible, avoid placing the first cell in the buffer
layer
Beware of using a turbulence model for a flow
outside its range of applicability
If in doubt, check with support engineer first.
142
Acknowledgements
The following individuals in Fluent Inc.
contributed material shown in this training
session
Davor Cokljat, Yi Dai, Sung-Eun Kim, Fabrice
Mathey, Carl-Henning Rexroth, Shin Rhee,
Amish Thaker, Xuelei Zhu.
May also be other unknown contributors.
Many Thanks to ALL who helped!