Transcript FlavienBillard_EDF_MAN_2008.ppt

Code_Saturne at
Manchester
An overview of the current
research activities
1
The
2
v− f
model and Code_Saturne
Accuracy

A low-Reynolds (near-wall integration) eddy viscosity model
derived from second moment closure models

No damping functions, no wall functions, less empirical
assumptions
Stanford 1991

Best results on range of test cases, heat transfer and
natural convection in particular.
TU-Delft 2004

The original model is stiff (requires coupled solver or very
small time-step)
Manchester 2004

Degraded version available in StarCD, Fluent, NUMECA..

Long collaboration Stanford, Delft, Chatou, Manchester
(Durbin, Parneix, Hanjalic, Manceau, Uribe)
=> “several code friendly” versions since 1995.
Stanford 1996
(Fluent, STAR-CD)

Present: Reconsider all historical choices with numerical
stability and known asymptotic states as principal objectives
Robustness
2
The    model

v2

k
Using the elliptic blending of Manceau and Hanjalic 2002 (Re Stress model only) :
L     = 1
2
2
w  0

D
 2 t 
t
3
3
=  f hom + 1  f w  P +
 k +(( + ) )
Dt
k
k

f w  

y2
f w    2  o(1)

Successfully tested on channel flows for many Re numbers, flow around
airfoil trailing edge, heated pipe, heated channel flow, heated cavity
 Normal time-step (external flow CFL values as for k-omega)
 Unlike, code friendly Stanford model, no term has been neglected here
2
 Unlike UMIST and Delft model, the correct asymptotic behaviour of v and
is accurately predicted without impairing the numerical robustness
t
3
Results (Channel flow, Re*=395) (1)
4
Results (Channel flow, Re*=395) (2)
5
The kink of epsilon
D C1P  C 2

 Diff (  t /  )
Dt
T

The epsilon equation
A closer look at the constitutive relation
uv
t 
dU /dy
Durbin’s formula
K-epsilon, …

 t  C v T
 t  f  C kT
t
2
T  k /
dU 1 y

dy    t
U
Results
Results
Parametric tests using Python and a 1D code
Devising a genetic algorithm for parameter optimization
(C 1,C 2 ,C ,CL ,C ,CSSG,...)

Optimizing the set

An integer is coded as a sequence of bits
Cmin

Cmax
0
1
0
11
00
0
1

1
0100

The set of parameter is coded as a vector of bits
01110110110101011111101101101101110111011111101101101101101101011010101111
C1,C 2 ,C ,CL ,C ,CSSG
GA Description

Parameters: NbPop, NbPopMax, Bounds of each parameter…
Initial population (NbPop)

Initialization: NbPop initial guesses

Iteration i
EVOLUTION
Population (n) = Parents (NbPop)
Children (NbPopMax-NbPop)
1001000010111010111101101
0101111011101011011011011
0111011011010101111110110
N
0
Mutation
1pt-CrossOver
generations
(1 parents, 1 child)
(2 parents, 2 children)
SELECTION
2
1
0101111011101011011011011
U
CFD Code
DNS
(y)  U output(y) dy
y 0
1D Channel flow
Too many individuals… Drop the weakest ones

= Fitness
Conclusion

Prediction of transition (good results given by the Launder
and Sharma model)

Source terms in the
layer)

Devising a good near-wall low-Reynolds RANS model
suitable for RANS/LES coupling

 equation (prediction of the defect