PhD Preliminary Oral Exam CHARACTERIZATION AND PREDICTION OF CFD SIMULATION UNCERTAINITIES
1st Automotive CFD Prediction Workshopautocfd-transfer.eng.ox.ac.uk/Presentations/007-ANSYS... ·...
Transcript of 1st Automotive CFD Prediction Workshopautocfd-transfer.eng.ox.ac.uk/Presentations/007-ANSYS... ·...
Florian Menter, Chief Scientist Rob Winstanley, Engineering ManagerDomenico Caridi, Senior Regional Product ManagerKrishna Zore, Software Developer IITushar Jadhav, Senior Application Engineer
1st Automotive CFD Prediction Workshop
2
GEKO - New & Flexible RANS Turbulence Model
Motivation
• 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 jet)
‐ Decaying turbulence in freestream
• Coefficients are linked and cannot be changed easily by user
Central Question: Can we do such a simulation with one set of global constants?
Probably not …
GEKO Model: Introducing Free Coefficients
( ) ( )
+
+−=
+
jk
t
j
k
j
j
x
k
xkCP
x
kU
t
k
( ) ( )
+
+
+−=
+
j
t
j
jj
k
j
j
xx
xx
kFFCP
kFC
x
U
t
23
2
2211
( ),
,max Real
tCS
k
=
• CSEP – changes separation behavior
• CMIX – changes spreading rates of free shear flows
• CNW – changes near-wall behavior
• CJET – Optimizes free jet flows
• CCORNER – Affects corner flows
• CCURVE – Curvature Correction
The functions F1, F2, and F3 contain 6 free coefficients:
All coefficients (except CJET) are UDF functions and can be changed locally
𝑢𝑖′𝑢𝑗
′
→ 𝑢𝑖′𝑢𝑗
′ −𝐶𝐶𝑜𝑟𝑛𝑒𝑟1.2 𝑡
𝑀𝐴𝑋 0.3𝜔, (𝑆2 + 2)/2𝑆𝑖𝑘𝑘𝑗 − 𝑖𝑘𝑆𝑘𝑗
Wall Treatment - Comparison
• The formulation of a turbulence model when integrated through the viscous sublayer is a key aspect of turbulence modelling‐ Defines robustness
‐ Defines accuracy
‐ Can cause undesired pseudo-transition
Makes or Breaks a Turbulence Model
Backstep Simulation
• 4x the same k-e model with different near wall treatment– ML – Menter-Lechner low-Re model
– EWT – Enhanced wall treatment built on 2-Layer formulation
– GEKO-1 exact transformation of k-e to k- with k- wall treatment
– V2F - k-e model with V2F ‘elliptic blending’ wall treatment
• Results are vastly different
• GEKO is closest
Wall Shear Stress Wall Heat Transfer
GEKO Model - Switching
• CSEP - active everywhere
• CNW - active everywhere (but only relevant near wall)
• CMIX – activated by blending function
• CJET – sub-model of CMIX( )...1 BlendJETMIXMIX FFCF −=
Wall-Distance Free Variant option available
1=BlendF
• Incompressible flow‐ Re = 107
• Variation of CSEP and CNW
• Model maintains calibration for wide range of coefficient changes
• CMIX and CJET do not affect boundary layer
Flat Plate Boundary Layer
All 4 coefficients can be tuned by user without loss of accuracy for flat plate
Velocity Profiles for CS0 Diffuser: Cmix=0
Variation of main free coefficients
• CNW – affects only near wall – no effect on Cp
• CSEP – affects separation strength
• CMIX – no effect
• Main parameter - CSEP
CNW=0.5
CSEP=1.0
Separated Flow Around a NACA-4412 Airfoil
Flow scheme Incompressible flow Re = U∞ ∙C/ν = 1.64·106
C - airfoil chord
U∞ - freestream uniform velocity
α = 12o – angle of attack
Triangular Cylinder – Variation of CMIX – Fixed FGEKO
Streamwise velocity contours at the midsection
SST
GEKO-2GEKO-1
Ahmed Body
SEPARATION AND REATTACHMENT AFTER EXPANSION
MidsectionSlant
Incompressible flow
• All the models fail to predict both separation and reattachment on the slant
• Results of GEKO-1 and GEKO-2 are close to the results of their analog
✓ GEKO-1 is similar to k−ε
✓ GEKO-2 is similar to SST
• Results of GEKO-1 and k-e models fit experimental data better than other models
k-ε Std
Best Practice Document - GEKO
https://www.ansys.com/-/media/ansys/corporate/resourcelibrary/technical-paper/geko-tp.pdf
Use 2nd order turbulence when feasible
Summary - GEKO
• A new Generalized k- (GEKO) model has been developed
• It allows optimization of free coefficients over a wide range of applications
• Instead of switching between different models, users can now adjust a single model to their application
• Good chance of consolidation of two-equation models into one optimal format
• Further free coefficients will be added
• Strong defaults
• Coefficients can be changed locally via UDF
• Already successfully used in industrial applications
• Implementation in Fluent (planned for CFX R20)
25
Tuning the GEKO Turbulence Model for Case 2a & 2b
Tuning the GEKO turbulence model using Design of Experiment
• Goal is tuning the GEKO‐ To improve the prediction of drag and lift on two and
eventually on more car models
‐ Using main driving parameters and zonal approach
• Car Models used‐ DrivAer Fastback and DrivAer Estate – Corse Ansa Mesh
• Solver Set up‐ Coupled solver, 2nd order Upwind, LSQ, Pseudo Transient
• Parameters used for this study‐ Csep global
‐ Csep local in the wheel MRF zone
‐ Cmix global
26
Design of Experiment and Optimization using Workbench DX
• DOE main set up‐ Optimal Space Filling Design
‐ 20 samples
‐ Csep range: [1-2]
‐ Cmix range: [0.3-4]
• Input parameters‐ Csep global, Cmix global,
Csep local (wheel MRF)
• Output parameters‐ dCD, dCL for Fastback and Estate, dCD Fastback-Estate, Mean Square Error
• Total time for one model DOE (20 sim) about 6000 CPU hours‐ Comparable with one scale resolved simulation
• Neural Network Response Surface
27
Results
29
• Multi Objective Genetic Algorithm‐ Seek for 0 delta for Drag, Lift on both models
‐ Higher priority for Drag
‐ Minimize Mean Square error
‐ Keep same drag trend between two models
Very good trade off improvement!
Case 2a Coarse – GEKO Csep 1.75 Vs Optimised
Csep 1.75 Optimised
Contours of X Velocity at Plane Y = 0
0 5 15 2010
Case 2a Coarse – GEKO Csep 1.75 Vs Optimised
Csep 1.75 Optimised
Contours of X Velocity at Plane Z = 0
0 5 15 2010
Case 2a Coarse – GEKO Csep 1.75 Vs Optimised
Contours of Pressure Coefficent
-0.9 0.90
Csep 1.75 Optimised
-1.00E+00
-5.00E-01
0.00E+00
5.00E-01
1.00E+00
1.50E+00
2.00E+00
-2.00E+00
-1.50E+00
-1.00E+00
-5.00E-01
0.00E+00
5.00E-01
1.00E+00
-1.00E+00 -5.00E-01 0.00E+00 5.00E-01 1.00E+00 1.50E+00 2.00E+00 2.50E+00 3.00E+00 3.50E+00 4.00E+00
Optimised-pressure-coefficient Csep 1.75-pressure-coefficient z-coordinate
Case 2a Coarse – GEKO Csep 1.75 Vs Optimised
Case 2b Coarse – GEKO Csep 1.75 Vs Optimised
Csep 1.75 Optimised
Contours of X Velocity at Plane Y = 0
0 5 15 2010
Case 2b Coarse – GEKO Csep 1.75 Vs Optimised
Csep 1.75 Optimised
Contours of X Velocity at Plane Z = 0
0 5 15 2010
-1.00E+00
-5.00E-01
0.00E+00
5.00E-01
1.00E+00
1.50E+00
2.00E+00
-2.00E+00
-1.50E+00
-1.00E+00
-5.00E-01
0.00E+00
5.00E-01
1.00E+00
-1.00E+00 -5.00E-01 0.00E+00 5.00E-01 1.00E+00 1.50E+00 2.00E+00 2.50E+00 3.00E+00 3.50E+00 4.00E+00
Optimised-pressure-coefficient Csep 1.75-pressure-coefficient z-coordinate
Case 2b Coarse – GEKO Csep 1.75 Vs Optimised
39
MosaicTM Meshing – Case 2a
Mosaic (Poly-Hexcore) Meshing
Hex Core• High quality• Fast solve time
New: MosaicTM Technology• Unique technology to
conformally connect poly prisms to hex
• High quality transition, with significantly fewer cells than tet transition
• Patent pending
Poly Prism• High quality• Significantly fewer
cells than tri prisms
Mosaic (Poly-Hexcore) Meshing Parallel – F1 Car
• If Fluent Meshing is opened in parallel Distributed Parallel Meshing will auto-enable
• Particular benefit for large meshes or number of prism layers
• Up to 8.1 Million cells/min with 64-way parallel
• Typical memory requirement: <3 GB / Million cells
Mosaic Remeshing of Medium 2a Committee Grid
• All wall tri-surfaces unchanged• Quads on MFR Internal Surfaces
triangulated and Remeshed
• BOI regions and sizing replicated
• Prism Layers‐ Car - 22 Layers, 1.8e-5m first height,
variable growth rate to Last Ratio 40%
‐ Road - 22 Layers, first aspect ratio 100, 1.17 growth rate
Mosaic Remeshing of Medium 2a Committee Grid
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
0 50 100 150 200 250 300 350 400 450
CL
Iteration
ANSA (CL mean = 0.0782)
Mosaic (CL mean = 0.0697)
0.2
0.22
0.24
0.26
0.28
0.3
0.32
0.34
0.36
0.38
0.4
0 50 100 150 200 250 300 350 400 450
CD
Iteration
ANSA (CD mean = 0.2394)
Mosaic (CD mean = 0.2380)
Mosaic Vs ANSA Medium – GEKO Csep 1.75
• Mosaic creates similar spatial resolution mesh 86M vs 165M cells‐ Parallel meshing on 32 cores completes the
volume meshing in 19 minutes
• HPC Comparison‐ Mosaic➢ 280 cores (14 nodes with 2x Xeon E5-2660 v3 2.6GHZ)
➢ 618.5 CPU.Hours
‐ ANSA➢ 224 cores (8 nodes with 2x Xeon E5-2690 v4 2.6GHZ)
➢ 1422.4 CPU.Hours
• Similar accuracy 2x speed up‐ Similar results with ANSYS Hexcore meshes
Mosaic Vs ANSA Medium – GEKO Csep 1.75
Mosaic ANSA
Contours of X Velocity at Plane Y = 0
0 5 15 2010
Mosaic Vs ANSA Medium – GEKO Csep 1.75
Mosaic ANSA
Contours of X Velocity at Plane X = 0
0 5 15 2010
Mosaic Vs ANSA Medium – GEKO Csep 1.75
Mosaic ANSA
Contours of Pressure Coefficent
-0.9 0.90
-1.00E+00
-5.00E-01
0.00E+00
5.00E-01
1.00E+00
1.50E+00
2.00E+00
-2.00E+00
-1.50E+00
-1.00E+00
-5.00E-01
0.00E+00
5.00E-01
1.00E+00
-1.00E+00 -5.00E-01 0.00E+00 5.00E-01 1.00E+00 1.50E+00 2.00E+00 2.50E+00 3.00E+00 3.50E+00 4.00E+00
ANSA-pressure-coefficient Mosaic-pressure-coefficient z-coordinate
Mosaic Vs ANSA Medium – GEKO Csep 1.75
Summary
• Tuning GEKO allows a RANS model to get much closer to experimental results ‐ Running a DoE for two vehicle configurations takes similar CPU resource to a single scale
resolving simulation
• Mosaic Parallel Meshing creates a mesh with similar spatial resolution to traditional Prism-Tet-Hexcore meshes with 40% – 50% less cells‐ This results in a 2x speed up in solution time with not loss of accuracy
• Mosaic Parallel Meshing generates 86 Million cells in 19 Minutes.
57
58
Thank You and Questions