You are on page 1of 28

Application of Surrogate-based Optimization

Techniques to Aerodynamic Design Cases

Emiliano Iuliano and Domenico Quagliarella


CIRA, the Italian Aerospace Research Centre
e.iuliano@cira.it
d.quagliarella@cira.it
Contents

• Background & motivations

• Surrogate-assisted Optimization approach


– POD machinery and POD-RBF surrogate model
– Database enrichment and objective-driven in-fill strategies
– Optimization process

• Aerodynamic design cases


– RAE 2822 airfoil in transonic viscous flow
– DPW-W1 wing in transonic viscous flow

• Conclusions & future prospects


EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 2
Background & Motivations
• Surrogate models represent a viable alternative to expensive simulations in
industrial design at much lower computation cost.

• Surrogate modeling by itself is not enough, the prediction accuracy and


suitability to optimization depends on how much “useful” information is
provided to them
– A key point is the training process
– Need to develop surrogate strategies more than models

• AIAA Aerodynamic Design Optimization Workgroup


– benchmark aerodynamic optimization cases defined and updated for optimization codes
comparison
– Yearly meetings held at AIAA main conferences

• GARTEUR Action Group number 52 launched in 2012 to explore SBGO


methods for aerodynamic shape design. Main objectives:
– Assess & validate surrogate modeling techniques
– Demonstrate CPU time savings by using SBGO methods
– “best practice” guidelines for the industrial use of SBGO methods\
EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 3
Contents

• Background & motivations

• Surrogate-assisted Optimization approach


– POD machinery and POD-RBF surrogate model
– Database enrichment and objective-driven in-fill strategies
– Optimization process

• Aerodynamic design cases


– RAE 2822 airfoil in transonic viscous flow
– DPW-W1 wing in transonic viscous flow

• Conclusions & future prospects


EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 4
POD basics

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 5


POD truncation

  0.99  11 modes left

  0.95  6 modes left

  0.90  5 modes left

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 6


POD content

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 7


POD pseudo-continuous prediction

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 8


Contents

• Background & motivations

• Surrogate-assisted Optimization approach


– POD machinery and POD-RBF surrogate model
– Database enrichment and objective-driven in-fill strategies
– Optimization process

• Aerodynamic design cases


– RAE 2822 airfoil in transonic viscous flow
– DPW-W1 wing in transonic viscous flow

• Conclusions & future prospects


EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 9
Training strategies

• The training strategy is heavily dependent on the type and scope of the
surrogate model and should be tailored on it

• The design space sampling has to be designed to “adapt” itself to the


response function, e.g.
– to improve the overall quality of the model prediction (error-driven) OR
– to minimize the true response function (objective-driven)

• Most of the adaptive sampling methods are designed to balance


exploration and exploitation contributions
– Exploration: search away from available data, do not trust the model
– Exploitation: trust the model, try to improve local accuracy

• A trade-off is needed because:


– Exploration may lead to explore uninteresting regions
– Exploitation may lead to local entrapment

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 10


Example of different training choices

Pure exploration
sample Pure exploitation
sample

Balanced sample

Global optimum

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 11


In-fill criteria

Exploration Exploitation
EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 12
In-fill criteria

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 13


In-fill criteria: example

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 14


Contents

• Background & motivations

• Surrogate-assisted Optimization approach


– POD machinery and POD-RBF surrogate model
– Database enrichment and objective-driven in-fill strategies
– Optimization process

• Aerodynamic design cases


– RAE 2822 airfoil in transonic viscous flow
– DPW-W1 wing in transonic viscous flow

• Conclusions & future prospects


EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 15
Surrogate-assisted Optimization approaches
Once built the surrogate, different
Design space strategies to optimize it
1. Direct Optimization
2. Sequential optimizations with re-injection of optimal
A priori sampling samples
3. Model improvement by in-fill criteria
4. Model adaptation + sequential optimizations
Design database SBO optimal samples

CFD solution Adaptive samples

Obj. function evaluation

Surrogate model build In-fill criteria Budget


check
N
Stop

Budget Y
Optimize
check

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 16


Contents

• Background & motivations

• Surrogate-assisted Optimization approach


– POD machinery and POD-RBF surrogate model
– Database enrichment and objective-driven in-fill strategies
– Optimization process

• Aerodynamic design cases


– RAE 2822 airfoil in transonic viscous flow
– DPW-W1 wing in transonic viscous flow

• Conclusions & future prospects


EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 17
Problem Definition: transonic flow over an airfoil

Baseline airfoil: RAE 2822


Cook, P.H., M.A. McDonald, M.C.P. Firmin, Aerofoil RAE 2822 – Pressure Distributions, and Boundary Layer and Wake
Measurements, Experimental Data Base for Computer Program Assessment, AGARD Report AR 138, 1979.

Design point
• Mach = 0.734
• Reynolds = 6.5 millions
• Fully turbulent flow

Constrained minimization problem:

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 18


Mesh generation & CFD simulation

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 19


Optimization studies

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 20


Optimization results

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 21


Contents

• Background & motivations

• Surrogate-assisted Optimization approach


– POD machinery and POD-RBF surrogate model
– Database enrichment and objective-driven in-fill strategies
– Optimization process

• Aerodynamic design cases


– RAE 2822 airfoil in transonic viscous flow
– DPW-W1 wing in transonic viscous flow

• Conclusions & future prospects


EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 22
Problem Definition: transonic flow over a wing

Baseline wing: 3rd Drag Prediction Workshop wing shape


http://aaac.larc.nasa.gov/tsab/cfdlarc/aiaa-dpw/Workshop3/

Transonic flow conditions: Mach = 0.78, Reynolds = 5 millions, AOA=1°


Fully turbulent flow is assumed.

Constrained minimization problem:

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 23


Parameterization & Mesh generation

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 24


Optimization studies

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 25


Optimization results

Fine mesh results

Inboard section airfoil

EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 26


Conclusions & Prospects

• Three stages (a-priori, adaptive and optimization) surrogate database


construction to reflect different needs in training a surrogate

• Adaptive sampling stage  design of two ad-hoc in-fill criteria to


enrich the surrogate model (exploration/exploitation trade-off)

• Different role of each sampling stage: a-priori to initialize, adaptive to


drive, evolutionary optimization to refine.

• Both 2D and 3D results show that significant reduction of the


computational budget needed to achieve a certain performance level
is possible by employing the proposed technique

• Future prospects
– In-fill criteria specifically designed for multi-modal/discontinuous functions
– Real-world application (wing-body-tail optimization, next AIAA Design
Optimization Group Workshop)
EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 27
Thank you for your kind
attention!

The author has requested enhancement of the downloaded file. All in-text references underlined in blue are linked to publications on ResearchGate.
EUROGEN 2015, 14-16 September 2015 – University of Strathclyde, Glasgow (UK) 28