More Related Content Similar to Automotive aerodynamics-optimization---2013-07-17 (20) Automotive aerodynamics-optimization---2013-07-171. © 2011 ANSYS, Inc. July 17, 20131
Details of the
Automotive
Aerodynamics
Optimization work
featured in this
Bloomberg
Businessweek article
in the March 11-17,
2013 issue, are
presented here
2. © 2011 ANSYS, Inc. July 17, 20132
Automotive Aerodynamics Optimization
Sandeep Sovani, Ph.D.
Global Automotive Strategy and Marketing
ANSYS Inc, Detroit, USA
July 17, 2012
Ashok Khondge
Lead Technology Specialist
ANSYS India Pvt Ltd, Pune India
3. © 2011 ANSYS, Inc. July 17, 20133
Aerodynamics Optimization Methods
Parametric Method – 50:50:50 Method
Gradient Based Method – Adjoint Method
Agenda
4. © 2011 ANSYS, Inc. July 17, 20134
• Method 1 – Parametric Method
– Parameterize vehicle shape
– Change shape parameters and run numerous simulations
– Generate response surfaces and optimize vehicle shape
• Method 2 – Gradient Based Method
– Solve for flowfield of baseline shape
– Perturb shape and calculate derivative of drag w.r.t. to shape
– Identify shape changes yield most improvement in drag
Aerodynamics Optimization Methods
5. © 2011 ANSYS, Inc. July 17, 20135
Aerodynamics Optimization Methods
Parametric Method – 50:50:50 Method
Gradient Based Method – Adjoint Method
Agenda
6. © 2011 ANSYS, Inc. July 17, 20136
• Aerodynamic Optimization via shape exploration
– Parameterize vehicle shape
– Change shape parameters and run numerous simulations
– Generate response surfaces and optimize vehicle shape
• Ideal Aerodynamics Optimization Process
– Ability (to explore a large design space)
– Automatic (with least human effort)
– Fast (Fits in the vehicle development process)
– Accurate (High Fidelity Meshes, Physics Models)
Introduction
7. © 2011 ANSYS, Inc. July 17, 20137
7
Three
Essentials
PAPER 2012-01-0174
Introduction
8. © 2011 ANSYS, Inc. July 17, 20138
Introduction
The 50:50:50 Method
50 50 design points in the design space EXTENT
50
50 million cells used in CFD simulation
of each design point
ACCURACY
50
50 hours total elapsed time to
simulate all the design points
SPEED
“One – Click” – Entire design space is simulated and post-processed
completely automatically after the initial baseline case setup
9. © 2011 ANSYS, Inc. July 17, 20139
Methodology
Prepare Meshed Model for
Baseline Vehicle Shape
CFD Solver Setup,
Define Shape Parameters
Generate DOE using
Input Shape Parameters
Collate Data,
Perform Optimization
Morph Vehicle Shape
Run CFD Simulation
STEP 1
STEP 2
STEP 3
STEP 4
STEP 5
10. © 2011 ANSYS, Inc. July 17, 201310
• To Demonstrate 50:50:50
Method
– Volvo XC60 vehicle model
– Four shape parameters
– RBF Morph (Integrated within
FLUENT) to define shape
parameters
• ANSYS WorkBench (Frame Work
to Automate Process)
– To drive shape parameters
– To create DOE
– To perform Goal Driven
Optimization
Test Case Description
11. © 2011 ANSYS, Inc. July 17, 201311
Methodology – ANSYS WB
STEP 3
STEP 5
STEP 2
STEP 4
12. © 2011 ANSYS, Inc. July 17, 201312
• Surface mesh – ANSA
• Surface mesh size
– Front facia : 3.0 to 4.0 mm,
– Windshield : 4.0 to 5.0 mm
– Doors &windows : 5.0 to 6.0 mm
– Roof : 6.0 to 8.0 mm
– Rear : 4.0 to 5.0 mm
– Underbody : 5.0 to 6.0 mm
Step #1 : Baseline Model Creation
13. © 2011 ANSYS, Inc. July 17, 201313
13 PAPER 2012-01-0174
Prism Layer
• 10 Layers (First Aspect Ratio 10, Growth 1.1)
• 24.4 million cells (about half of total cells are in prism layer
Cell size
needed if
using
cartesian cells
with same cell
count
Number of cartesian cells needed to achieve same
near wall resolution 550 million!
Step #1 : Baseline Model Creation
14. © 2011 ANSYS, Inc. July 17, 201314
• Volume Mesh – TGrid
• Cell Count : 50.2 Million Cells
• Prism Layers : 10 (First Aspect Ratio 10,
Growth 1.1)
• Prism Count : 24.4 Million Cells
• Skewness < 0.9
Step #1 : Baseline Model Creation
Prism Layers
Cut Plane Y=0
Cut Plane Z = 1.4 m
15. © 2011 ANSYS, Inc. July 17, 201315
• Boundary Conditions
– Inlet : Velocity Inlet 100 kmph
– Outlet : Pressure Outlet, 0 Pa (Gage)
– Side walls : Wall, no-slip
– Top wall : Wall, no-slip
• Solver Settings
– Steady, PBCS, Green Gauss Node
Based Gradient
– Fluid : Incompressible air,
– Density = 1.225 kg/m3
– Turbulence : Realizable K-epsilon,
Non-equilibrium wall treatment
– Discretization :
• Pressure – Standard
• Momentum, TKE, TDR – 2nd Order
Step #2 : CFD Setup
• Solution Controls
– Courant Number = 200
– ERF
Momentum, Pressure = 0.7
– URFs
Density = 1.0, Body Forces = 1.0
TKE, TDR = 0.8
TR = 1.0
16. © 2011 ANSYS, Inc. July 17, 201316
Step #2 : RBF Morph
• RBF Morph : Add-On Module
– Fully Integrated within ANSYS
FLUENT with GUI/TUI
• Uses Radial Basis Function
Technique for Mesh
Morphing
– System of radial basis function
is used to produce solution for
mesh movement
– List of source points and their
displacements are used as
input
– Valid for both surface shape
change and volume mesh
smoothing
• Developed by RBF Morph
http://www.rbf-morph.com/
RBF GUI
17. © 2011 ANSYS, Inc. July 17, 201317
Step #2 : Shape Parameters - Definition
1. Boat Tail Angle (P2)
Constraint :
Point “B” to move upto 20 mm in
+ve /-ve Y-direction about the
Pivot axis
2. Long Roof Drop Angle (P3)
Constraint :
Rear edge to move
Upto 30 mm in +ve Z-direction
Upto 45 mm in -ve Z-direction
about the Pivot axis
18. © 2011 ANSYS, Inc. July 17, 201318
Step #2 : Shape Parameters - Definition
3. Green House (P4)
Constraint:
Point “A” to move 20 mm in
+ve /-ve Y-direction about the
Pivot axis
4. Front Spoiler Angle (P5)
Constraint:
Point “C” to move 30 mm in
+ve Z direction
19. © 2011 ANSYS, Inc. July 17, 201319
Axis about which selected
surface set gets morphedEncapsulation RegionSurface set selection
Step #2 : Setup – Boat tail angle (P2)
20. © 2011 ANSYS, Inc. July 17, 201320
Step #2 : Setup – Long roof drop angle (P3)
Encapsulation RegionSurface set selection Axis about which selected surface set gets morphed
25. © 2011 ANSYS, Inc. July 17, 201325
• RBF Morph shape parameters
– Define in FLUENT
– Available in WB for
• Input Shape Parameters
– P2 : Boat tail angle
– P3 : Long roof drop angle
– P4 : Green house
– P5 : Front spoiler
• Output Parameter
– P1: Drag Force on vehicle
• DOE Algorithm
– Central Composite Design
– Design Type : Face Centered with
Enhanced Template
– 49 DOE Points Generated
Step #3 : Design Space
Design Space Bounds
Parameter Min Baseline Max
Boat tail angle - 1.85 0.0 + 1.85
Long roof drop angle - 2.30 0.0 + 1.50
Green House Angle - 0.70 0.0 + 0.70
Front Spoiler Angle 0.0 + 3.80
27. © 2011 ANSYS, Inc. July 17, 201327
Step #4 : Running Simulations
• Current Study
– Simulations were run outside of
WorkBench using journal file in
batch mode
– Output the drag force
– DOE table updated by
importing output parameters
• Five Runs Using
– 768, 384, 288, 240, 144 Cores
• Convergence Monitors
– Drag force
– Pressure /velocity at few points
in wake
Design point # 1
Design point # 2
Design point # 50
28. © 2011 ANSYS, Inc. July 17, 201328
Step #4 : Running Simulations
768 Cores 384 Cores 288 Cores 240 Cores 144 Cores
Task Time (Seconds) Time (Seconds)
Time
(Seconds)
Time
(Seconds)
Time
(Seconds)
Baseline Case (i.e. Design Point 1)
Read volume mesh of baseline
case into the CFD solver and
apply solver settings
225 340 365 481 228
CFD Solution 6979 11153 14409 17256 27246
Writing CFD data file 681 538 558 600 532
Each Subsequent Design Point
Morph vehicle shape 84 59 65 69 100
CFD Solution 1284 1754 2208 2630 4100
Writing CFD data file 734 559 572 621 532
Total Run Time (Wall Clock)
Needed for All 50 Design
Points (Hours)
30.80 35.63 42.98 50.28 72.19
If data file is not written at each data point, then 50
hours target is achieved in less than 200 cores
Same study was repeated with newer hardware.
50 hours target is achieved in roughly 150 cores
29. © 2011 ANSYS, Inc. July 17, 201329
Step #4 : Running Simulations
Compute Cluster Details
1. Intel’s Endeavor Cluster
2. Intel Xeon X5670 (dual socket)
3. Clock speed 2.93 GHz
4. Six cores per socket
(12 cores per node)
5. 24 GB RAM @ 1333 MHz, SMT
ON, Turbo ON
6. QDR Infiniband
7. RHEL Server Release 6.1
31. © 2011 ANSYS, Inc. July 17, 201331
• Response Surface Analysis
– Non Parametric Regression (NPR) Algorithm
– Plots
• Goodness of fit
• 2d / 3d Response surface plots
• Sensitivity plots
• Parallel Co-ordinates (Pareto) Plots
• Trade-Off plots
• Optimization Study
– Goal driven optimization
– Screening algorithm (no of samples = 5000)
– NLPQL (Non-Linear Programming by Quadratic Lagrangian)
• Flow Results
Results
32. © 2011 ANSYS, Inc. July 17, 201332
Results : Response Surface Plots
33. © 2011 ANSYS, Inc. July 17, 201333
Results : Response Surface Plots
34. © 2011 ANSYS, Inc. July 17, 201334
Results : Response Surface Plots
35. © 2011 ANSYS, Inc. July 17, 201335
Results : Sensitivity Plots
Local Sensitivity Global Sensitivity
36. © 2011 ANSYS, Inc. July 17, 201336
Results : Parallel Co-ordinates (Pareto) Plot
38. © 2011 ANSYS, Inc. July 17, 201338
Results : Goal Driven Optimization
39. © 2011 ANSYS, Inc. July 17, 201339
Results : Goal Driven Optimization
40. © 2011 ANSYS, Inc. July 17, 201340
Flow Results
Design
Points
Boat Tail Angle
(P2)
Long Roof Angle
(P3)
Green House
(P4)
Front Spoiler
Angle (P5)
Drag Force (N)
(P1)
1 0.000 0.000 0.000 0.000 388.01
9 0.000 1.500 0.000 1.900 393.01
19 1.850 -2.300 -0.700 0.000 372.30
25 -1.850 1.500 -0.700 0.000 397.33
• Flow Results Discussion
– Design point 1, 9, 19 & 25
– Velocity contours
– Iso-surface of total pressure = 0.0
41. © 2011 ANSYS, Inc. July 17, 201341
Summary of 50:50:50 Method
The 50:50:50 Method
50 50 design points in the design space EXTENT
50
50 million cells used in CFD simulation
of each design point
ACCURACY
50
50 hours total elapsed time to
simulate all the design points
SPEED
“One – Click” – Entire design space is simulated and post-processed
completely automatically after the initial baseline case setup
42. © 2011 ANSYS, Inc. July 17, 201342
• The 50:50:50 Method
– An extensive, fast, and accurate DOE method
– In case study: 4% drag reduction achieved in 1 week of work
• Fully automated workflow (after baseline case setup)
using industry leading technologies
– FLUENT Solver (With High Performance Computing)
– RBF Morph (Fast, User-Friendly, Accurate Mesh Morpher)
– ANSYS WorkBench (Integration Platform)
– Design Xplorer (Optimization)
– CFD Post (Flow Visualization)
Summary of 50:50:50 Method
43. © 2011 ANSYS, Inc. July 17, 201343
Aerodynamics Optimization Methods
Parametric Method – 50:50:50 Method
Gradient Based Method – Adjoint Method
Agenda
44. © 2011 ANSYS, Inc. July 17, 201344
Overview
Application Areas and Associated
Technologies
R14.5 New Features
Solved Examples using R14.5
Summary
Agenda
45. © 2011 ANSYS, Inc. July 17, 201345
What is It ?
• Different methods for computing derivatives – sometimes
referred to as sensitivities.
• Consider a high-level view of a fluid system
ADJOINT METHOD
Inputs c Quantities q
Flow Solver Flow variables
Integral quantities
Mesh
Boundary conditions
Material properties
Transfer matrix j
i
c
q
Tangent method
Adjoint method
46. © 2011 ANSYS, Inc. July 17, 201346
Overview of the adjoint method
Workflow
• Solve the flow equations and post-process the results as usual.
• Pick an observation that is of engineering interest.
• Lift, drag, total pressure drop?
• Set up and solve the adjoint problem for this observation
• Define solution advancement controls
• Set convergence criteria
• Initialize
• Iterate to convergence
• Post-process the adjoint solution to view
• Shape sensitivity
• Sensitivity to boundary condition settings
• Contour & vector plots
47. © 2011 ANSYS, Inc. July 17, 201347
Overview of the adjoint method
Shape sensitivity: Sensitivity of the observed value with respect to
(boundary) grid node locations
mesh
nn
xwDrag .)(
Shape sensitivity coefficients:
Vector field defined
on mesh nodes
Node displacement
Visualization of shape sensitivity
• Uses vector field visualization.
• Identifies regions of high and low
sensitivity.
• These are the places where
changes to the shape can have a
big impact on the quantity of
interest.
• The guidance is specific to the
quantity of interest, and the current
flow state.
Drag sensitivity for NACA0012
48. © 2011 ANSYS, Inc. July 17, 201348
Associated Technologies
Mesh Morphing
• Use a Bernstein polynomial-based morphing scheme for freeform mesh
deformation.
• Select portions of the geometry to be modified by specifying a rectilinear
deformation region.
• Define a uniform distribution of control points in space in each coordinate
direction.
• Movement of any control point leads to a smooth deformation field throughout
the deformation region.
• Can be driven by non-gradient based algorithm – e.g. Simplex.
49. © 2011 ANSYS, Inc. July 17, 201349
Increase the downforce on the vehicle
Look for regions of high sensitivity of downforce to shape
Downforce enhancement for a generic race car
50. © 2011 ANSYS, Inc. July 17, 201350
Front wing redesign to generate more downforce
Downforce enhancement for a
generic race car
Downforce (N)
Geometry Predicted Result
Original --- 425.7
Modified 447.4 (+5.1%) 451.1 (+6.0%)
51. © 2011 ANSYS, Inc. July 17, 201351
Rear wing redesign to generate more downforce
Lift enhancement for a generic race car
Downforce (N)
Geometry Predicted Result
Original --- 425.7
Modified 481.3 (+13.1%) 492.5 (+15.7%)
52. © 2011 ANSYS, Inc. July 17, 201352
• What are the major factors affecting
the uniformity in the mass flow rates
at the 4 outlets?
• Material is air
• Solve the flow problem.
• Set up and solve the adjoint problem
with the variance in mass flow rates
as the quantity of interest.
• Post-process the field to identify
important influences.
Robust Design Example: Internal Flow
Total pressure (pascal)
60 cm
1
2
3
4
m1: 0.0020 Kg/s
m2: 0.0023 Kg/s
m3: 0.0028 Kg/s
m4: 0.0025 Kg/s
var: 7.52e-08
4
14
1
i
imm
24
14
1
var
i
i mm
inlet
53. © 2011 ANSYS, Inc. July 17, 201353
• Variance computed to be
7.52e-08 (Kg/s)^2
• Plot the displacements of the
surface that, based on linear
extrapolation, would drive
the variance to zero.
• Geometry far upstream is
dictating the flow split.
• Manufacturing variances of
the order of 3mm in the inlet
region can cause O(1) flow
variance variations.
Robust Design Example: Internal Flow
Surface normal displacements that
Induce O(1) change in variance
Editor's Notes It is important for new users to understand the context in which the adjoint solver is used. In this slide we explain how the adjoint solver fits into an existing CFD workflow. The mystery of what sensitivity to shape means is now explained. A vector of derivative components is associated with each node in the mesh (both interior and boundary). In some places the vectors are large. This signifies that a small movement in the mesh node position results in a relatively large change in the drag. Where the vector is of zero length then to first order there will be no change in the drag if the node is moved. The inwards/outwards orientation of the vector indicates that inwards/outwards motion of the node increases the drag. Introduces the basic idea of mesh morphing without any initial reference to the adjoint. This is a 15 M cell half vehicle case run on 32 processors.