Multi-Objective Design Exploration (MODE) - Aerodynamic Applications

at Tohoku University

Dr. Daisuke Sasaki

Aerodynamic Design Lab.Department of Aerospace Engineering

Tohoku Universitysasaki@ad.mech.tohoku.ac.jp

Contents• GCOE• MODE Application• Data Mining for Aerodynamics• Conclusion

GCOE: Overview• Global COE Program at Institute of Fluid Science and Dep’t of Aerospace

Engineering at Tohoku University.• Integration of Knowledge: Establishment and Expansion of International

Education and Research Center for Flow Dynamics.

GCOE: Activities• International Conference on Flow Dynamics (Nov. 4-6, 2009)• Internship Internship project• International Joint Laboratory – Design Exploration

MODE: Introduction• Progress of Computational Fluid Dynamics (CFD) for an aircraft

enables reliable prediction of aerodynamic performance at steady state.• Aerodynamic optimization has been frequently conducted.• We have proposed Multi-Objective Design Exploration for further

improvement of designs through knowledge discovery:– Multi-Objective Optimization, Data Mining, Visualization, etc.

MODE: importance• Multi-Objective Design Exploration

– Design exploration rather than optimization• Knowledge discovery from optimization results is more important than

obtaining optimal solutions– Optimization results contain

• Relationship between design variables and objective function• Importance of design variables• Trade-off between multiple objective functions

– Our approaches are • Efficient search - Kriging approximation model• Multi-Objective Optimization – Genetic Algorithms• Data Mining – ANOVA, SOM, visualization, etc.

f2

f1

Trade-off

MODE: Procedure• Flowchart

– Construct initial samplings for approximation model using Latin Hypercube Sampling

– Build Kriging approximation model for each objective

– Search non-dominated solutions on approximation model with MOGA by considering objective and its EI value

– Analyze additional sample points to update accuracy of approximations

– Data Mining based on ANOVA

CFD

Hybrid Mesh Generation

START

END

Generation of initial sample pointsby the Latin Hypercube Sampling

Update of the Kriging modelSelection

Generation of initial populationand evaluate each individual

Construction of initial Kriging models

Selection of additional sample pointsfrom non-dominated solutions

Evaluate new individuals

Crossover and mutation

NGEN>100

Continue?

Generation of a fixed-slat airfoil shape by Free-Form Deformation

CFD

Data Mining

Initial Model

Model Update

Application: PAV

• Application of Multi-Objective Design Exploration method to mild-stall wing for PAV (Personal Air Vehicle) or UAV (Unmanned Aerial Vehicle) use– Multi-Objective Optimization by MOGA– Approximation objectives by Kriging Method– Functional Analysis of Variance (ANOVA) for Data Mining

© Terrafugia Inc. 2006http://www.cafefoundation.org/v2/pav_home.php

PAV Wing: Introduction• PAV (Personal Air Vehicle) for leisure and future commuter

– 2-6 seats– 150-200 mph– akin to drive a car

• Demands of PAV for popularity– Operability: simple – Comfort: noise, quiet– Efficiency: performance– Safety: stall characteristics– Reliability: – Cost:

Stall characteristics is quite important for small vehicles as stall often leads flight accidents: taking-off and landing failure, spin in flight

Mild stall

Sudden drop

CL

Angle of attack

PAV Wing: Introduction 2• Targets of our research are to

– Accomplish mild-stall characteristics for safety enhancement– Achieve high Cl at high AOA for landing and taking off– Improve cruising performance for practical use

Idea to realize mild-stall characteristics is to prevent sudden drop of CL by making use of two airfoils having different stall characteristics

NACA4415SLAT4415

Tip Root

2/3 Original

PAV Wing: Introduction 3• Original – slotted wing

– Inboard: NACA4415– Outboard: SLAT4415

• Simple – rectangular wing– NACA4415

Tip

Mach contours at 20% and 80% span station at aoa of 24

TipSimple Original

NACA4415SLAT4415

TipRoot 2/3

Due to the different stall characteristics, slotted wing achieved mild stall characteristics.

Original

PAV Wing: Introduction 4• A slotted wing (Original wing) has favorable

stall characteristics, but the aerodynamic performance at cruise is seriously low.

170count

To improve cruise performance is necessary to realize efficient PAV

SLAT4415 (based on NACA4415)

Original

Design Problem• Objectives

– Maximization of L/D at Cl of 0.4 (cruise efficiency)– Maximization of Cl at aoa of 16 (take off performance)

• Constraint– Cl > 1.9 at aoa of 16

• Design variables: 30– 15 control points for x,y direction

• Optimizer – MOGA coupled with Kriging model– 120 initial samples by Latin Hyper-cube Sampling– 12 updates for improving Kriging model accuracy

• CFD evaluation– 2 CFDs per design by TAS 2D NS with SA– Mach number: 0.2– Reynolds number: 1.5x106

SLAT4415 (based on NACA4415)

fixed

Optimization Result

• Optimization Result

Mopt1 (L/Dcruise Max)

SLAT4415

NACA4415

airfoil L/D @ Cl0.4 Cl @ aoa16

Mopt1 26.77 1.96

NACA4415 29.99 1.58

SLAT4415 16.21 1.92

cruise

PAV: Sharp nose effect• Sharp nose effect

– ANOVA shows the influence of leading edge shape is small for objective function (L/D @ Cl0.4)

– Mopt1B is modified to blunt nose.– Mopt1S has sharpened leading edge nose.– L/D for 3 shapes are more or less similar.

Mopt1 　 C ｌ　 = 0.4 Mopt1B Mopt1S

airfoil L/D @ Cl0.4 Cl @ aoa16

Mopt1 26.77 1.96

Mopt1B 26.61 1.96

Mopt1S 26.55 1.92

CP9_YCP1_YCP9_XCP2_YCP8_YOthers

PAV: ANOVA • Effect of design variables to L/D at Cl of 0.4 (objective1)

CP9_Y

CP1_Y

To achieve high L/D, leading edge and trailing position of a slat is deformed to flow directions

Positive directionNegative direction

Positive directionNegative direction

DM: Procedure • Data Mining Procedure

– Data Preparation

– Mining Methods & Thresholds Selection

– Mining Result Visualization

– Knowledge Evaluation and Validation• Testing

– Knowledge Fusion• Cross validation, bootstrap

– Decision Strategies

0.006

0.008

0.01

0.012

0.014

0.016

0.018

0.02

0.008 0.01 0.012 0.014 0.016 0.018

CD(supersonic)

CD(transonic)

DM: Data Quality Mining • Data preparation: Low Quality Data Problem

– Inconsistent Data– Duplicate Data– Missing Data– Outliers– Censored and Truncated Data– Time-dependent anomalies

• Data Quality Mining: data quality measurement and improvement through Data Mining techniques– Exploratory Data Analysis– Multivariate Statistics– Classification– Clustering– Visualization– Quantitative Data Cleaning

DM: Visual Analytics• Visual Analytics: Science Discovery Method

– 70s – 80s CAD/CAM, 3D modeling, animation– 80s – 90s Scientific visualization – 90s – 2000s Information visualization– 2000s – Visual Analytics

• Science of analytical reasoning facilitated by interactive visual interfaces

• Interactive Visualization for knowledge discovery in complex data• Tight integration of Visual and Automatic Data Analysis Methods

with Database Technology for a scalable decision support

Data

Visualization

Models

Knowledge

Visual Data Exploration

Automatic Data Analysis

DM: Applications and Future• Potential fields

– Data Quality Management for Approximation model– Data Mining for optimization results

• Classification, pattern analysis, clustering• Association rules, causal analysis• Time-series data feature detection

– Uncertainty assessment– Unsteady data analysis, feature capturing– Decision-Making support

Applications at NASA (Dr. Srivastava, NASA Ames)• Virtual sensors in Earth Science and Astrophysics• Decision-Making support of flight readiness review or

aviation safety (vehicle health management)

• Future ideal framework is to integrate– Physics-based model– Machine Learning-based model

DM: Joint-Lab.• GCOE Joint Laboratory of Design Exploration

“To advance optimization and data mining techniques for real-world applications through international collaboration”– International workshop (MLA09 – 2nd workshop) – Internship program for PhD students– Collaborative Research

• To improve accuracy of approximation model by applying different techniques to various test problems and applications.

• To investigate the characteristics and usefulness of various data miniig techniques by applying them to optimization results.

• To develop a useful data fusion method for reliable numerical computation and optimization.

– Dataset of aerodynamic optimization results will be available for applying various tools of data mining techniques.

Please contact Sasaki (sasaki@ad.mech.tohoku.ac.jp)

Conclusion• Multi-Objective Design Exploration is important and useful for

aircraft designs to extract important information, particularly when– Design space or problem is unknown – innovative configuration– Design space is large – conceptual design

• There are many Machine Learning techniques useful to improve aerospace design exploration system:– Data Mining– Data Quality Management– Visual Analytics– Etc.

Last Slide• Thank your for your attention and participant

1/2b out

c root

Λ in =automatically

Λ out

c kink1/2b in

γ kink

γ tip

c tip=automatically

t/c root

t/c tipFLOW

Λ in_TR =0 deg

t/c kink

1/2b out

c root

Λ in =automatically

Λ out

c kink1/2b in

γ kink

γ tip

c tip=automatically

t/c root

t/c tipFLOW

Λ in_TR =0 deg

t/c kink

TAS-Code Development

and Its Applications

Next-Generation CFD (Building-Cube Method)

Multi-Disciplinary Analysis and Design Methods

Conceptual Aerodynamic Design of New Aircrafts

Research on CFD

Research activities at Aerodynamic Design Laboratory

Announcement

• Post-workshop proceeding:– Deadline : August 3rd– Maximum 4 pages– To mla09@ad.mech.tohoku.ac.jp

– Proceeding will be sent to you directly, so please write down your name and postal address on the sheet.

• Dinner tonight– If you want to join for dinner, please tell David now.

We will meet at 7:30 pm at hotel lobby.

• GCOE invited speakers– Please give me the documents now for reimbursement..

Optimal Design (Mopt1) • Optimal design (Mopt1)

– As estimated by ANOVA result, L/D max design (Mopt1) is modified to lean the shape for flow direction.

– Separation behind the slat is suppressed, though large separation observed for slat4415.

– Lift is maintained at high angle of attack.

Mopt1 　 C ｌ　 = 0.4 Slat4415 at Cl of 0.4Mopt1 　 aoa　 = 16

airfoil L/D @ Cl0.4 Cl @ aoa16

Mopt1 26.77 1.96

NACA4415 29.99 1.58

SLAT4415 16.21 1.92

Mopt1 and SLAT4415

Characteristics of MOpt1 • Mopt1_single is designed by connecting the outline of Mopt1.• Mopt1 shows higher Cl at low aoa but sharp drop of Cl at high

aoa compared to SLATT4415.• Mopt1_single shows lower Cl compared to NACA4415.

airfoil L/D @ Cl0.4

Cl @ aoa16

Mopt1 26.77 1.96

NACA4415 29.99 1.58

SLAT4415 16.21 1.92

Mopt1 and Mopt1_single

Mopt2 and Mopt1 airfoils • Mopt1_3D did not show mild-stall characteristics.• From the solutions, an airfoil having relatively milder and higher Cl at high

aoa airfoil is chosen (Mopt2).• Mopt2 is superior to SLAT4415. airfoil L/D @ Cl0.4 Cl @ aoa16

Mopt1 26.77 1.96

Mopt2 24.57 1.92

NACA4415 29.99 1.58

SLAT4415 16.21 1.92

Mopt1@24 Mopt2@24

Approach: ANOVA

• The functional analysis of variance (ANOVA )– In order to identify the effect of each design variable to

the objective functions, the total variance of this model is decomposed into the variance component due to each design variable.

＜ Zup ＞