RANS Turbulence

Modeling and Wall

Treatment
Dr. Florian Menter
Chief Scientist

1
Turbulence

• 3-dimensional, vorticity
dominated , dissipative, stochastic
• Spectrum of sizes in space and
time depends strongly on Re U D
Re 

• Is described by Navier-Stokes
equations
• Direct Numerical Simualtion (DNS)
not feasible as it scales with CPU
~ Re3
• Classical multi-scale problem
Web Seite KTH
Averaging Procedures

• In order to avoid the resolution of

the turbulent scales – the
equations are time averaged.
• The resulting equations are called
Reynolds Averaged Navier Stokes
(RANS) equations.
• Turbulence models required to U ( x, t )  U ( x )  u ' ( x, t )
consider effect of turbulence on
mean flow. u ' ( x, t )  0
t  t
1
• Time average: U ( x) 
t  U ( x, t )dt
t
RANS Equations

• Reynolds (Time) Averaged Navier Stokes (RANS) Equations

 (U i )  (U jU i ) P  ( ij  u 'i u ' j )
  
t x j xi x j
• Reynolds Stresses unknown:

 u ' u ' u ' v' u ' w' 

 
 u 'i u ' j   u ' v' v' v' v' w' 
 u ' w' v' w' w' w' 
 
Molecular Viscosity
y U
• Mixing Layer inviscid:
– No random motion
– No momentum transfer
across layer x
– Jump in U(y)

• Mixing Layer laminar:

– Random motion (Brown‘s
motion)
– Momentum transfer across layer

• Momentum transfer through viscous  U i U j  dU

stress tensor:  ij        12  
 x xi dy
 j 
Eddy Viscosity Concept
• Eddies are started by instability L(x)
(Kelvin-Helmholtz) T(x)
• Large eddies interact and
generate smaller eddies
(Turbulence)
• The large eddies cause efficient
mixing across layer

• Eddy Viscosity Assumption:

 U i U j 
 u 'i u ' j   turb
 t      12turb   t U with  t ~ C
L2
ij  x xi  y T
 j 
Eddy Viscosity Models
• Reynolds Averaged Navier-Stokes (RANS) Equations:

 (U i )  (U j U i ) 1 P 
    ijmol   ijturb 
t x j  xi x j
 U i U j 
• Stokes‘ stress tensor:  ij   
mol
 
 x xi
 j 

• Reynolds‘ stress tensor:  U i U j  2

 turb
 t     k ij
ij  x xi
 j  3
2
L ( xi )
• Eddy viscosity:  t  Const.
T ( xi )
k-Equation – Derivation
U i  U i  u 'i
• Definition :
k
1

2 i 1,3
u 'i u 'i

1
2
u 'i u 'i


1 2
2
u '  v'2  w'2 
 U i u 'i 
 
Navier-Stokes minus RANS  t  t  ...  0 
equation gives equation gives  
eqation for fluctuating  U i   u 'i 
velocities    ...  0     ...  0 
 t   t 

Multiplication be u‘i  u 'i   1 u 'i u 'i 

u 'i   ...  0     ...  0 
 t   2 t 

 1 u 'i u 'i   k 
Time averaging   ...  0   ...  0  …= All Other Terms
 2 t   t 
k-Equation – exact Form
Change in Time Convective Transport


 
k  U j k
 u 'i u ' j
U i Pk=Production
t x j x j
u 'i u 'i
 e=Dissipation
xi xi
  1 
 u 'i  u 'i u ' j  p ij  turbulent Diffusion
x j
  2 
 2k
 2 molecular Diffusion
x j
Modelling of k-Equation
• The equation system consists of (incompressible):
‐ Equations for Ui , P
‐ Equations for k und e.
• Molecular Diffusion is exact. Can be neglected at high Reynolds numbers.
• Unknown correlations have to be modelled based on known quantities Ui, k and e.

U i  U i U j  U i
 u 'i u ' j  t   
Eddy-Viscosity
x j   x
 x j xi  j
  1     t k 
Gradienten diffusion u 'i  u 'i u ' j  p ij    
x j   2  x j  s x 
 k j

New Constant sk
e-Equation for high Re Numbers
• Exact e-equation can be derived based on its definition from the Navier-Stokes
equations.
• This equation is very complex and contains many higher order correlations.

 (e )  (U j e )
  (Production - Destruction) + Diffusion
t x j

e e2    t e 
ce 1 Pk  ce 2   
k k x j  s x 
 e j 
• Modelling based on dimensional analysis and in analogy to k-equation.
• New empirical constants:
ce 1 ; ce 2 ; s e
ke Model for high Ret-Zahlen

• Closed set of equations  U i U j  2

which can be solved  ui u j   t    - k ij ;
 x xi  3
numerically  j 
• Coefficients/Constants  ( k )  (U j k )    t k 
need to be determinde   Pk  e   
t x j x j  s x 
from experinents or DNS  k j 

 (e )  (U j e ) e e2    t e 
  ce 1 Pk  ce 2   
t x j k k x j  s x 
 e j 

U i  t  c
k2
Pk  ui u j e
x j

• Empirical Constants: c ; ce 1 ; ce 2 ; s k ;s e
Determination of Constants, 1
• Empirical constants are best determined based on simple and accurate
experiments.
• k- and e–equations can be simplified, depending on experiment.
• For flows in local equilibrium one gets Pk  e .

Pk  e ; 
2 2
 dU  k 2
 dU   k2 
Pk   t   ;
 e  c ;   t     c
  
 dy  t  dy   t 
2
 dU 
2
uv
 t2    uv  c k 2 ;
2
  c  2
 dy  k

• Form measurements: c0,09.

Determination of Constants, 2
• Simplest turbulent flow: isotropic turbulence in space (grid turbulence);

dk de e2
U  e ; U  ce 2
dx dx k
 1 
  
 e0x   ce 2 1 
• Analytical solution: k  k 0 1  ce 2  1  
 k 0 U 
• In comparison with experiments ce2  1,8-2.

• For the near wall logarithmic layer (local equlibrium) one gets from the e-
equation:

2
ce 1  ce 2 
s e c
Standard Set of Constants
• The missing diffusion constants sk and se as well as the final values of ce1 and ce2
are determined through computer optimizations and detailed comparison with
additional experiments (mixing layer, jets etc.):

c ce1 ce2 sk se
0,09 1,44 1,92 1,0 1,3

• Modification of constants required for :

– Pk and e stark strongly different (c)
– Axi-symmetric jet (ce2, c)
t
– Small turbulent Reynolds numbers Ret: Re t ~ ~ o(0  10)

w-Equation

• Standard model of Wilcox .

• The w-equations is solved instead of the e-equation. 1 k 1
• Characteristic Frequency of energy containing eddies. w ~
c e T
• No fundameltal difference in derivation and calibration.
Turbulent Frequency:
• Differences however for non-equilibrium flows (outside
calibration range)
• Main advantage is robust and accurate near wall treatment.
k- w Modell (Wilcox)

 ( k )  (U j k )    t k 
  Pk  C  kw   
t x j x j  s x 
 k j 

 (w )  (U jw ) w    t w 
  cw1 Pk  cw 2w 
2  
t x j k x j  s x 
 w j 

k C Cw1 Cw2 sk sw
t 
w 0,09 5/9 0.075 2,0 2,0
• Advantage:
– Reduced length scale near walls and pressure gradients.
– Robust near wall formulation (viscous sublayer).
• Problem:
– “Freestream sensitivity”.
k- e and k- w Two-Equation Models
• Solve two equations to get two  ( k )  (U j k )    t k 
independent scales   Pk  e  DkLR   
t x j x j  s x 
 k j 
• Compute eddy viscosity from
the two scales  (e )  (U j e ) e e2    t e 
  f1ce 1 Pk  f 2 ce 2  DeLR   
t x j k k x j  s x 
• Scale equation (w or e) is  e j 
modelled in analogy with the k2
 t  c f 
k-equation e
• Differences in Low Re (viscous
sublayer) and non-equilibrium
 
    t  k
flows  ( k )  (U j k )  
  Pk  C  kw  
t x j x j  s  x 
  k  j 
 
 (w )  (U jw ) w     t  w
 
  cw1 Pk  cw 2w 2  
t x j x j  s w  x j 
k  
k
t 
w
RANS vs LES/DNS
Turbulent structures RANS - L

RANS – T

RANS - U
RANS – t
Advanced Turbulence Modeling

Flow Separation Flow

Reattachment
Corner
Vortices Mixing
/combustion

Transition Vortical
Flows
NASA X-29

Unsteady Effects
Integration Platform w-equation

2-equation models Transition Model

• g-ReQ model
• k-w, BSL, SST

Unsteady models w-equation

Wall Treatment
• SST-SAS • Automatic wall treatment
• SST-DES

Higher order models Extensions

•Stagnation point
• EARSM – w •Curvature correction
• SMC - w •Rough walls
•Reattachment correction
k-w Model “Freestream” Problem
Velocity Eddy Viscoisty
Wall Distance/

Wall Distance
w w

• Solution depends strongly on freestream (inlet) w

• Solution therefore dependent on User
• Model is not suitable for high accuracy aerodynamic simulations
• Change in Wall Shear stress ~15-20% is too much for accurate aerodynamic
simulation
k-e versus k-w Formulation
• In wall boudary layers there are Boundary k-w k-e

Quality
clear quality gradienst for each Layer
model Sublayer • Robust • Stiff
• Accurate • Less accurate
• k-e: • Complex
– Poor performance near the wall (sublayer and
• Simple
log-layer under dp/dx) resulting in a Logarithmic • Accurate • Large length
robustness probems and delayed separation scales
Layer
– The model avoids the freestream sensitivity
near the boundary layer edge Wake Region • Missing • Missing
transport transport
• k-w: effects effects
– Poor near boundary layer edge (freestream Boundary layer • Free-stream • Well defined

Quality
sensitivity) edge sensitive
– High quality near wall (sublayer and log-layer
under dp/dx)
Combination of models!
Baseline (BSL) k-w Model Background
• The BSL model is a combination of k-e and k-w. It uses the w-equation near the wall and then
blends to the e-equation near the boundary layer edge
• Since it is undesirable to solve both, the w- and the e-equation and then blend the solution,
the e-equation is transformed mathematically to an exact equivalent of in w-formulation and
then blended with the original k-w model
• Blending is achieved through a blending function

k-w model transformed Outer layer

from standard k-ԑ model3 (wake and outward)
k 2
e
e Inner layer
WilcoxWilcox’
k-w model
original k-w model (sublayer, log-layer)

Wall
Blending of k- w and k- e Model

 ( k )  ( U j k )    t k 
  Pk    kw  (   ~ ) 
t x j x j  s  x j 

 ( w )  ( U jw ) w 2 k w   t w 
   Pk  w  (1  F1 ) 
2
 (   ) 
t x j k s w x j x j x j  s w x j 

• Term from k-e model

• Combination of k-e and k-w
advantages
• Near the wall k-w
• Away from the wall k-e
Blending Function F1
Shear Stress Transport (SST) Model

Bradshaw relation: U
 uv   t  a1k
y

Standard model: Pk
 uv  a1 k
e
a1k   uv  a1k
SST model: t 
max S  F2 , a1w 

• Standard models (k-e and k-w, …) overpredict shear stress in adverse pressure
gradient flows – no or delayed separation
• SST model enforces Bradshw relation – accurate separation prediction ->
simplest Explicit Algebraic Reynolds Stress Model (EARSM).
Diffuser Flow
• SST Model is optimized for adverse pressure gradient flows
with separation
Diffuser Flow CS0 (NASA Testcase)
• SST model gives highly accurate separation prediction,
compared to k-e and k-w models
NACA 4412 Airfoil

• SST model in comparison with separated velocity profiles compared to

different k-e models and Spalart-Allmaras (SA) model
NASA Transonic Bump Flow

• Transonic bump Ma=0.875 Ma>1

shock

flow separation

• Shock causes
separation
• Separation
prediction
influences shock
location
• SST model gives
proper shock
location
Batchello and Johnson bump
Generalized k-w Model (GEKO)
• Two-equation models are
the work-horse in industrial
CFD
• The have typically 5
coefficients which can be
calibrated to match physics
• They are calibrated for
‐ Flat plate boundary layers (log-layer)
‐ Selected free shear flows (plane mixing layer, plane Central Question: Can we do such a simulation with
jet) one set of global constants?
‐ Decaying turbulence in freestream
Probably not …
• Coefficients are linked and
cannot be changed easily
by user
Status – Multitude of Models in Industrial Codes

• Advantages: • Disadvantages:
‐ Allows covering different flow scenarios – Model selection is confusing to users
(different models for different types of – Often users select non-optimal models
flows)
– Hard to compare different CFD codes
‐ Provides options in case one model is
more robust for a given application/flow – Implementation quality/consistency is hard
to achieve for multiple models:
‐ Users have preferential models – often
• Interaction with large number of sub-models (y+-
for historic reasons (it is hard to insensitive wall treatment, rough walls, buoyancy,
‘eliminate’ models from a CFD code) transition, EARSM/RSM, DES/SBES, numerics …)
‐ For marketing reasons (CFD companies – Huge effort implementation and
like to have their own ‘flavour’ of maintenance for CFD vendors – loss of
modelling) implementation quality

Consolidated Two-Equation Turbulence Models?

GEKO Model: Introducing Free Coefficients

 k   U j k    t  k 
The functions F1, F2, and F3 contain 6   Pk  C kw      
t x j x j  sk 
 j 
x
free coefficients:
 w   U jw  w 2  k w
  Cw1 F1 Pk  Cw 2 F2 w 2  F3
t x j k s w w x j x j
• CSEP – changes separation behavior   t  w 
     
x j  sw 
 j 
• CMIX – changes spreading rates of free x

shear flows k
t   ,
• CNW – changes near-wall behavior maxw , S CRe al 

• CJET – Optimizes free jet flows

All coefficients (except CJET) are UDF functions
• CCORNER – Affects corner flows and can be changed locally
• CCURVE – Curvature Correction
Backward-facing Step –CNW
• Incompressible flow
– Re = ρ∙Uref ∙H/μ = 2.8·104

CSEP=1
• CNW has a strong
effect on the wall
shear stress (Cf) and
even more on heat
transfer (St) in CSEP=2
reattachment regions
• CNW =0.5 seems a
good value for both
Cf and St.
Flow Around a NACA-4412 Airfoil - CSEP Re = 1.64·106
α = 12o
Flow scheme

• CSEP has a strong effect

on separation
characteristics
• Viable range of values can
cover a wide range of
behavior
• CSEP ~ 1.75-20 seems
optimal for airfoils
Triangular Cylinder – Versatility Tests

• Bluff body separation with vortex shedding

• Steady state RAND known to give overly large separation zones
• Can we mimic SRS with GEKO?
Flow Around Bluff Body CMIX=0.3

• Real flow has vortex shedding

• GEKO run with high mixing
parameter CMIX in steady state CMIX=5.0
mode
• Massive reduction in re-circulation
zone
• Better agreement with data
• Imagine flow behind a wheel of a
Formula 1 car
‐ Downstream parts of the car see much
more realistic flow
 Ahmed Body
Incompressible flow
Midsection
Slant

Streamwise velocity contours at the midsection

• All the models fail to predict both
separation and reattachment on the
GEKO-1 GEKO-2 slant
• Results of GEKO-1 and GEKO-2 are
close to the results of their analog
SST  GEKO-1 is similar to k−ε
k-ε Std
 GEKO-2 is similar to SST
• Results of GEKO-1 and k-e models fit
experimental data better than other
• S.R. Ahmed, G. Ramm, Some Salient Features of the Time-Averaged Ground Vehicle Wake, SAE-Paper 840300, 1984
• H. Lienhart, S. Becker, Flow and Turbulence Structure in the Wake of a Simplified Car Model, SAE 2003 World
models
Congress, SAE Paper 2003-01-0656, Detroit, Michigan, USA, 2003
RANS Models

• RANS models are the work horse of industrial CFD simulations

• There is a wide variety of RANS models in both ANSYS CFD
codes – and we will maintain a reasonable selection of
different models
• The w-equation (BSL) based models offer significant
advantages over e-equation based models (separation,
robustness, transition modelling, SAS, DES, ...)
• Eddy-viscosity models are most widely used – robsutness is the
main reason
• Higher order models (EARSM, RSM) are available and applied
to specific applications (scondary flows, strong swirl, etc.)
• Tuneble Models (GEKO) will e the next step
Wall Treatment – Turbulent BL Structure

• Structure of turbulent boundary layer:

Couette Flow
U
y x
• Self-similar flow d/dx=0
• Conti:
U V
  0  V  const.  0
x y
U  U ( y)
DU U U U  mol
Dt

t
U
x
V
y
0  0
y
 12   12turb  
• Momentum:
  12mol   12turb  const.   wall

U
 mol
 turb
    t      t  S   wall
y
12 12
Couette Flow – Mean Flow U(y)

u y U
Definitions  wall   u ;
2 
y  ; U 

 u

Viscous Sublayer   t

U  wall u y
 mol
   wall  U y  u  u y   U   y 
12
y  
Couette Flow – Mean Flow U(y)
Log Layer   t

Assumtion I: t  u y Lt  cy 

U U U 1 u
 turb
 t   u y   u2  
y y y  y
12

 1
U  log( y  )  C


 f  
dU dU u
Or Assumtion II:  c 
dy dy y
1
U  log( y  )  C

Couette Flow – Turbulent Variables

• Log Layer Pk  t S 2  e | t 

k2
 t S   c
2 2 2
e   2 c k 2
e
 2
turb   c k  const.  
2 2 2
wall    u
2 2

u2
k
c

k2 u3 u4
 t  c  e 
e  y  y 

All variables known in logarithmic layer as function of y+ and u

 Mean Velocity Profile

• Velocity profile in
logarithmic scale Outer region
• Near wall – universal
profiles:
• Outer flow depends on Log. region
 1
problem U  log( y  )  C


U   y
Sub-layer
 Wall Function Boundary Condition

• Why resolve “universal” part

of velocity profile?
• Place first integration (grid)
point in the log. region.
• Specify flux at the surface to
ensure proper wall shear
stress. log.

• Only correct for equilibrium

flows Flux Flux
 Mathematical Formulation

U u y
Compute from old U: u  
y 
1
log( y  )  C 

FU   u
2
Momentum equation:

k-equation: Fk  0

u4
e-equation: e
y 
Invalid for y+ < 10
Singular for y+ 0
Thickness of Boundary Layer

•y+ Thickness of boundary

layer is function of Y+=5000
Re number.
Y+=300
•Extent of log-layer shrinks
with Re number

For low Re number resolution of the boundary

layer not possible with standard wall function
 Fine Grid Errors

Finer Grids

Finer Grids

Errors on fine grids

Solution deteriorates under mesh refinement!
 Y+-insensitive Wall Treatment

• Need for y+
insensitive wall
treatment
• AWT and EWT
smoothly varies
from low-Re to wall
function with mesh
resolution
y+-insensitive Wall Treatment
ke, Grid 1  20
ke,Sep 2 
Grid2006  20
ke,Sep
Grid2006
3   20 Sep 2
0.006 Cf, CFX 0.006 0.006
Cf, Fluent
St, CFX
St, Fluent
Cf, exp
0.004 St, exp 0.004
0.004
y+~1.0 y+~5.0
Cf, St

Cf, St
Cf, St
y+~0.2

0.002 0.002 0.002

0 5E+06 1E+07 0 5E+06 1E+07 0 5E+06 1E+07

ReX ReX ReX
ke, Grid 4  20
ke,Sep
Grid 5 
2006  20 Sep 2006 
0.006 0.006

0.004
y+~10.0 0.004
y+~20.0
Cf, St

Cf, St

0.002 0.002

0 5E+06 1E+07 0 5E+06 1E+07

ReX ReX
Check Boundary Layer Resolution Very fine mesh for transition prediction
EVR
• Boundary layers require a minimum resolution for
accurate results
• Number of cells depends on accuracy requrements
• For aerodynamic flows, one should have more than
10 cells inside the boundary layer – for highly
accurate simulation even up to Ny~30-40.
• For industrial flows around Ny~10 should be the
target
• For complex flows, it is possible that one can only t
EVR 
afford a few prism layers (3-5). In this case accuracy 
can be compromised
• Count prism layers inside boundary layer by plotting EVR typically clearly indicates boundary layer
EVR with mesh on top (the mesh shown here is very as it has maximum in the middle of the layer
fine)
Check Boundary Layer Resolution and Prism Layer Height

Good Not Good

Prism layer should cover entire

Resolution on Airfoil boundary layer
• Good at leading edge (boundary
layer covered by prism layer)
• Boundary layer too thick for
prism layer at mid-chord Bad
• Boundary layer much thicker
than prism layer at trailing edge
Boundary Layers Resolution
• Place a minimum(!) of 10 points inside boundary layer
• Try to resolve viscous sublayer
• Y+ alone does not say much about resolution of boundary layer (strongly Re
number dependent and also stretching factors etc.)
• Estimate boundary layer thickness from flat plate correlation and adjust grid
• Check by looking at eddy viscosity (maximum in middle of boundary layer
(count nodes).
• Don’t be afraid of high aspect ratio cells (both codes can handle wall aspect
ratios of 104 and higher – coupled pressure based solvers recommended).
Recommendations
• For free shear flows:
‐ The differences between the models are fairly small
• For boundary layers in equilibrium (or separation from corners in
geometry etc.):
‐ The differences between the models are fairly small
• For flows with need for high accuracy in boundary layers (separation,
heat transfer):
‐ Use SST model with sufficient boundary layer resolution – model implementation in Fluent 6.4 and CFX 12
will be very close
• For Flows with strong swirl:
‐ Use EARSM or Reynolds Stress model
‐ Use two-equation model with curvature correction
• For highly separated flows:
‐ Future will be unsteady models
Summary
• Wide range of RANS “best-in-class” models available
• We want turbulence modeling to be a generic subject independent of the codes
• The w-equation is an attractive integration platform
‐ BSL, SST model
‐ EARSM-w model
‐ SMC-w models
• Extensions
‐ Stagnation flows
‐ Curvature correction
‐ SAS – unsteady model
‐ Transition …
• Advantage of current strategy is ability to combine all effects in a single model
environment
• We will of course also continue developing the k-e based models in the code
Literature
• Wilcox, D. C.: Turbulence Modeling for CFD, DCW
Industries Inc., 2. Edition (1998).
• Rodi, W. :Turbulence models and their application in
hydraulics. Taylor and Francis 1993
• Durbin, P.A. and B.A. Pettersen Reif.: Statistical Theory
and Modeling for Turbulent Flows, John Wiley and Sons
Ltd 2003.
• Launder B.E.and Spalding D.B.: Lectures in mathematical
models of turbulence. Academic Press, 1972
• Menter, F.R.: Two-equation eddy-viscosity turbulence
models for engineering applications, AIAA Journal, Vol.
32, No. 8, pp. 1598-1605, (1994).