DNV GL © 2014 SAFER, SMARTER, GREENER DNV GL © 2014 Efficient Optimization of Flow-Exposed Geometries 1 by mapping adjoint shape sensitivities to CAD model parameters Mattia Brenner, Andre Zimmer, Sebastian Weickgenannt
Aug 19, 2014
DNV GL © 2014 SAFER, SMARTER, GREENER DNV GL © 2014
Efficient Optimization of Flow-Exposed Geometries
1
by mapping adjoint shape sensitivities to CAD model parameters
Mattia Brenner, Andre Zimmer, Sebastian Weickgenannt
DNV GL © 2014
Optimization Task
Rear wing of a hyper sports car should be
optimized with respect to the objectives
drag and downforce
iconCFD is used for adjoint CFD
computations
CAESES is used for parametric
modeling, mapping of the
adjoint sensitivities to the
CAD model parameters
and subsequent
geometry modification
2
With kind permission of: Koenigsegg Automotive AB
www.koenigsegg.com
DNV GL © 2014
CAESES
Specifically geared for automated CFD-driven design
Focus on complex free-form surfaces and shapes, difficult to parameterize
Models defined and varied according to “smart parametric” descriptions
Reduced degrees-of-freedom; constraints can be built in to the description
Complex models and their variants maintain high-fidelity and fairness
3
Profile defined using specialized curve types and controlled by user-defined parameters
Initial profile is transformed along a specified path, and its parameters are varied based on functional distributions
Proprietary meta-surface technology creates complex surfaces with intelligent parameterization and high quality
DNV GL © 2014
Parametric Model
4
DNV GL © 2014
Parametric Model
5
DNV GL © 2014
Parametric Model
19 design variables control the shape with the help of spanwise distributions of
profile parameters
Geometry constraints regarding span and chord length are automatically met
6
DNV GL © 2014
Optimization with a Parametric Model
Usually, a high number of parameters defines a parametric model (in this case 19,
often 30-50 for a typical CAESES model)
Way too much effort to involve all of them in a conventional optimization
process…
How to select the most suitable parameters for the optimization task at hand?
Normally, the designer selects or specifically creates a small number of
parameters based on experience and engineering judgment
– This reduces the design space for the optimization
– The designer might not have enough experience to make a good selection
– Especially difficult if the model was created by someone else
Solution: adjoint analysis
7
DNV GL © 2014
What is Adjoint Analysis?
Comparison to direct gradient-based method:
8
Adjoint method
Question: what changes to the design parameters are
necessary to get an optimal product performance?
1 primal + 1 adjoint simulation run for each objective
necessary
Direct method
Question: how do changes of the design parameters influence the product‘s
performance?
n+1 simulation runs necessary
nn J
J
J
J
33
22
11
optn J ,,,, 321
J: objective function, α: parameter
DNV GL © 2014
Adjoint Analysis Results
In shape optimization: shape
sensitivity (change of objective
function J due to normal displacement
of cells on the design boundary)
A positive shape sensitivity means that
the boundary should be moved in
positive normal direction
A negative shape sensitivity calls for
boundary movement in negative
normal direction
9
kn
J
With kind permission of: Koenigsegg Automotive AB
www.koenigsegg.com
Push inwards Pull outwards
DNV GL © 2014
How to Use the Adjoint Sensitivities
Adjoint shape sensitivity values can be used to displace the surface cells directly
and to morph the shape, e.g. in a CAD independent approach
Downside is that the shape changes cannot easily be fed back into the design
workflow, geometry constraints may be violated
Solution: map shape sensitivities to CAD model parameters
10
.avg
k
k n
k
kn A
An
n
JJ
Adjoint shape sensitivity
Normal displacement of model boundary due to CAD parameter change: „design velocity“
Relative local cell size
Parametric sensitivity
DNV GL © 2014
Design Velocities
Normal displacement of the boundary due to changes of CAD parameter values
Determined by perturbing the parameters and measuring the normal
displacement of given positions on the model boundary (surface tessellation
nodes)
CAESES Sensitivity Computation
Input: model boundary as surface geometry
Automatically determines all design variables
that are suppliers to the boundary
User can deselect design variables and set
individual deltas
Result: design velocity maps for all selected
parameters
11
DNV GL © 2014
Parametric Sensitivities
Also computed by Sensitivity Computation in CAESES
Adjoint shape sensitivities are mapped to the positions on the boundary used to
determine the normal displacement (tessellation nodes) by probing the result
data set of the adjoint computation
For each parameter adjoint sensitivities are locally multiplied with the design
velocities and summed up, weighted with the relative area of the local element
Result: one scalar parametric sensitivity for
each parameter
12
DNV GL © 2014
Parametric Sensitivities
From the list of parametric sensitivities select the parameters with the biggest
influence on the objective function manually change values or follow up with
conventional optimization
or
Multiply the vector of the parametric sensitivities with a step size factor and add
to the parameter values determine the right step size with 1-dimensional
search using primal CFD computations
or
…?
13
DNV GL © 2014
CFD Setup
Mesh size: 18 Mill. cells
Spalart-Allmaras turbulence model
Primal run: 6000 iterations, ~7h on 64 CPUs
Adjoint run: 6000 iterations, ~5h on 64 CPUs
Adjoints computed for drag and downforce
Inlet velocity: 40m/s
turbulence intensity: 0.005
turbulence length scale: 0.01
Wheels with rotating wall
Incompressible, rho=1.204, nu= 1.436877e-05
Three radiators modeled as porous media
Frozen adjoint turbulence
14
With kind permission of: Koenigsegg Automotive AB
www.koenigsegg.com
DNV GL © 2014
Design Velocity Results
Thickness_pressCenter
– Rank 1 for drag
– Rank 2 for downforce
15
Adjoint Sensitivities (drag)
Adjoint Sensitivities (downforce)
Design Velocity
DNV GL © 2014
Design Velocity Results
StepPos_yShift
– Rank 2 for drag
– Rank 1 for downforce
16
Adjoint Sensitivities (drag)
Adjoint Sensitivities (downforce)
Design Velocity
DNV GL © 2014
Design Velocity Results
Camber_pressCenter
– Rank 5 for drag
– Rank 4 for downforce
17
Adjoint Sensitivities (drag)
Adjoint Sensitivities (downforce)
Design Velocity
DNV GL © 2014
Design Velocity Results
Thickness_pressTip
– Rank 3 for downforce
18
Adjoint Sensitivities (drag)
Adjoint Sensitivities (downforce)
Design Velocity
DNV GL © 2014
Design Velocity Results
Thickness_sucCenter
– Rank 3 for drag
– Rank 5 for downforce
19
Adjoint Sensitivities (drag)
Adjoint Sensitivities (downforce)
Design Velocity
DNV GL © 2014
Design Velocity Results
Thickness_sucInner
– Rank 4 for drag
20
Adjoint Sensitivities (drag)
Adjoint Sensitivities (downforce)
Design Velocity
DNV GL © 2014
Parametric Sensitivity Results
Top 5 of most influential parameters
for drag and downforce
Most of these parameters point in
opposite direction for the two
objectives
Three modified geometries were
generated
1. Based on the top 5 parametric
sensitivities for drag
2. Based on the top 5 parametric
sensitivities for downforce
3. Changing all parameters that primarily
affect one objective while having little
influence on the other, in an effort to
improve both objectives
21
DNV GL © 2014
Optimization Results
Baseline geometry
22
With kind permission of: Koenigsegg Automotive AB
www.koenigsegg.com
DNV GL © 2014
Optimization Results
Drag optimized geometry
-0.96% drag
-3.75% downforce
23
With kind permission of: Koenigsegg Automotive AB
www.koenigsegg.com
DNV GL © 2014
Optimization Results
Downforce optimized geometry
-0.03% drag
+3.86% downforce
24
With kind permission of: Koenigsegg Automotive AB
www.koenigsegg.com
DNV GL © 2014
Optimization Results
Combined optimized geometry
-10.58% drag
+2.94% downforce
25
With kind permission of: Koenigsegg Automotive AB
www.koenigsegg.com
DNV GL © 2014
Discussion
Adjoint sensitivities can be mapped to the CAD geometry parameters to allow for
feasible shape modifications
Largest possible design space can be considered
All parameters in the model can be involved in the optimization, without the need
for the user to understand the effect of each parameter
The effort does not scale with the number of parameters
Parametric sensitivities for all parameters in the model can be computer more
quickly than for a small set using the direct approach
Separate adjoint solutions may be used to perform a multi-objective optimization
But:
Predictions based on sensitivities are only valid for small changes in shape
In practice iterative process: new sensitivities should be computed for modified
shape
26
DNV GL © 2014
SAFER, SMARTER, GREENER
www.dnvgl.com
CAESES , Your Upfront CAE System for Shape Optimization
27
Dipl.-Ing. Mattia Brenner
+49 331 96 766 0
Design Solve Optimize