INTRODUCTION TO

DESIGN OPTIMIZATION

Ranjith Dissanayake
Structures Laboratory
Dept. of Civil of Engineering
Faculty of Engineering
University of Peradeniya
VR&D

OPTIMIZATION

VR&D

VR&D

G. N. V.

REDUCE DESIGN
TIME

TASK 1
TASK 2
DEADLINE

IMPROVE DESIGN
QUALITY

FREE THE ENGINEER

FOR CREATIVE WORK
VR&D

THE PHYSICAL PROBLEM

BET I CAN
FIND THE
TOP OF THE
HILL!

YOU CAN
TRY, BUT
STAYINSIDE
THE FENCES

G. N. Vanderplaats

OBJECTIVE: FIND THE HIGHEST POINT

DESIGN VARIABLES: LONGITUDE AND LATITUDE
CONSTRAINTS: STAY INSIDE THE FENCES
YOU MAY START OUTSIDE THE FENCES
VR&D

THE ENGINEERING PROBLEM

G. N. V.

FENCE NO. 1
F1 = f1(X1, X2)
< 0 INSIDE
> 0 OUTSIDE

FENCE NO. 2
HILL
Y = f (X1, X2)

F2 = f2(X1, X2)
< 0 INSIDE
> 0 OUTSIDE

THE PHYSICAL PROBLEM IS NOW DEFINED

MATHEMATICALLY
BY CONVENTION, < MEANS INSIDE THE FENCES
VR&D

TH E OPTIMIZ ATION PROBLEM

MAX IMIZ E Y = f ( X 1, X 2)

OBJECTIVE

SUBJECT TO :
F1 = f1( X 1, X 2) <_ 0
CON STRAIN TS
F2 = f2( X 1, X 2) <_ 0

DESIGN VARIABLES

OPTIMIZATION IS A VERY SIMPLE EXTENSION

OF THE ENGINEERING PROBLEM
VR&D

THE OPTIMIZATION PROCESS

S3
X2
S2
X1
S1

G. N. Vanderplaats

FIND A SEARCH DIRECTION THAT WILL IMPROVE THE

OBJECTIVE WHILE STAYING INSIDE THE FENCES
SEARCH IN THIS DIRECTION UNTIL NO MORE
IMPROVEMENT IS MADE
REPEAT TO CONVERGENCE
VR&D

ANALYSIS VERSUS DESIGN

ANALYSIS: Given a Component or
System, Together With Loads, Materials,
etc.
Calculate the Responses to See if They Satisfy the Requirements

DESIGN: Given A Set of Requirements

Find the Component or System that Satisfies the Requirements with
Minimum Mass or Cost, Maximum Reliability, Maximum
Performance, etc.

ANALYSIS IS A SUB-PROBLEM OF DESIGN

VR&D

DESIGN
NONLINEAR, CONSTRAINED, OPTIMIZATION TASK
Find the Set of Design Variables, X, that will
Minimize F(X)
Subject to (Such That);

g j(X ) 0
X iL X i X iU

VR&D

j = 1, M
i = 1, n

10

10

DESIGN
For Example
Minimize Structural Mass
Subject to Stress Limits;

ijk
g j(X ) =
0

Objective Function

Inequality Constraints

i = Load Condition
j = Stress Calculation Point
k = Stress Component

VR&D

11

11

OPTIMIZATION ALGORITHMS
1948: SIMPLEX Method for Linear Programming
1950s: Various Random Methods. Gradient Based
Methods Developed in the Late 1950s
1960s: Sequential Unconstrained Minimization
Techniques, Sequential Linear Programming, Feasible
Directions Methods
1970s: Enhanced Feasible Directions Methods, Multiplier
Methods, Reduced Gradient Methods
1980s: Variable Metric Methods, Sequential Quadratic
Programming Methods
VR&D

12

12

OPTIMIZATION ALGORITHMS
1990s: Genetic Algorithms, Simulated Annealing, New
Interest in Sequential Unconstrained Minimization
Techniques
2000s: Particle Swarming, Advanced Sequential
Unconstrained Minimization Techniques
Largest Known Test Example
250,000 Variables With 250,000 Active Constraints

Largest Known Real Structural Optimization Problem

190,000 Thickness Variables with Frequency Constraints
VR&D

13

13

OPTIMIZATION PROBLEM SIZE

BIGDOT
250,000
VARIABLES
100,000

10,000
# Des. Var.
1,000

100

0
1960

VR&D

1970

1980

1990
Year

2000

2010

14

14

GENERAL OPTIMIZATION
PROBLEM STATEMENT
Minimize F(X)

Objective Function

Subject to (Such That);

g j (X ) 0

j = 1, M

Inequality Constraints

hk ( X ) = 0

k = 1, L

Equality Constraints

X iL X i X iU

i = 1, N

Side Constraints

F(X), gj(X) and hk(X) May be Linear, Nonlinear, Explicit,

Implicit, but Should be Continuous with Continuous First
Derivatives
VR&D

15

15

GENERAL STRATEGY
Given X0
At Iteration q, Update X by
Xq = Xq-1 + Sq
Sq = Vector Search Direction

= Step Size

Calculation of Sq Requires Gradients

Calculation of (One-Dimensional Search)
Requires Several Function Evaluations

Some Methods Dont Use a One-Dimensional Search

VR&D

16

16

THE OPTIMIZATION PROCESS

Given Xq
Update the Design by
Xq = Xq-1 + Sq Xq-1 + X

Note that this is Very Close to the Traditional

Design Process of Beginning with a Design and
Modifying it

VR&D

17

17

THE OPTIMIZATION PROCESS

The Best Search Direction Is One That Will
Minimize

F ( X q )T S q

Sq is Usable

Subject to;
g j ( X )T S q 0

jJ

Sq is Feasible

where J = the Set of Active (Critical) Constraints

This Requires Calculating the Gradient Vector of

the Objective and Active Constraints
VR&D

18

18

PHYSICAL INTERPRETATION
X2

F
F = CONSTANT

FEASIBLE

S
INFEASIBLE

g =0
X1

During the Search in Direction S, We Adjust the Variables

X1 and X2 to Move Back to the Feasible Region
VR&D

19

19

GRADIENT CALCULATIONS
By First Forward Finite Difference
F ( X + X1 ) F ( X )

X1

F(X + X ) F(X )
2

X2

......

......

F
(
X
+

X
)

F
(
X
)
N

XN

The Finite Difference Step Size Must be Chosen to Avoid

Errors From Noise and/or Round-Off
Central Difference is More Accurate but Twice as Costly
VR&D

20

20

GRADIENT CALCULATIONS
Analytically in Structural Optimization
KU = P
U
1 P K
=K

U
X
X X
K
X

Is the Sum of Derivatives of Element Matrices, ki

K
X

May be Calculated by Finite Difference

Gradients of Stress are Calculated From This

Gradients of Other Responses (Eigenvalues, Dynamic
Response, etc.) May Also be Calculated in This Way
VR&D

21

21

USEFUL DENINITIONS
Design Variables: Those Parameters to be Changed to
Improve the Design
Objective Function: The Function of the Design
Variables to be Minimized or Maximized
Inequality Constraints: One Sided Conditions that Must
be Satisfied for the Design to be Acceptable
Equality Constraints: Precise Conditions that Must be
Satisfied for the Design to be Acceptable
Side Constraints: Bounds on the Design Variables that
Limit the Region of Search for the Optimum

VR&D

22

22

USEFUL DENINITIONS
Feasible Design: One that Satisfies All Constraints
Infeasible Design: One that Violates One or More
Constraints
Optimum Design: The Set of Design Variables and the
Corresponding Minimum (Maximum) Objective Satisfying
All Constraints
Kuhn-Tucker Conditions: Necessary Mathematical
Conditions that Must be Satisfied for a Design to be
Optimum
Two-Variable Function Space: Geometric
Representation of a Two-Variable Design Problem
VR&D

23

23

EXAMPLE
Minimize

F ( X ) = X1 + X 2

Subject to;

g(X ) =

1
1
+
0
X1 X 2

X1 0

X2 0

X2
4
F=4
3
FEASIBLE REGION

3
2
2

OPTIMUM

1
1
g=0

VR&D

X1

24

24

THE KUHNKUHN-TUCKER
CONDITIONS FOR OPTIMALITY
X* is Feasible
jg j = 0

F ( X ) =

j = 1, M
M

M +L

j =1

k = M +1

jg j ( X ) + k hk ( X ) = 0

j 0, j = 1, M

k Unrestricted in Sign, k=M+1,L

This States that the Vector Sum of the Gradients of the

Objective and the Critical Constraints (Scaled by Their
Lagrange Multipliers) Must Add to Zero at the Optimum
VR&D

25

25

KUHN--TUCKER CONDITIONS
KUHN
Geometric Interpretation
X2

g3 = 0

3g3
F

1g1
F = CONSTANT

g1

g2 = 0

g 3

g1 = 0
X1

Note that Constraint Number 2 is Not Active, so it is Not

Included in the Vector Sum
VR&D

26

26

TYPICAL APPLICATIONS
Structures
Trusses, Panels (Isotropic, Composite), Shells, Pressure Vessels,
Frames
Automotive, Aerospace, Space, Ship, Rail

Mechanical Components
Shafts, Gears, Vibration Isolation, Piping Systems, Flyweels

Thermal Systems, Injection Molding, Heat Exchangers,

Steam Condensers, Heat Sinks, Combustion Cycles
Systems
Aircraft, Space, Automobile, Ships

VR&D

27

27

TYPICAL APPLICATIONS
Computational Fluid Dynamics, Combustion, Acoustics
Airfoils, Ducting, Noise Minimization, etc.

Other Applications

Control Systems
System Identification
Curve Fits
Lens Optics
Portfolio
Other

Applications are Limited Only by Our Ingenuity

VR&D

28

28

THE COST OF OPTIMIZATION

(BLACK BOX)
Criteria
Find a Near Optimum Quickly
Use as Few Function Evaluations as Possible

Basis for Criteria

Precise Optimum is Seldom Meaningful
Loads, Material Properties, Heat Transfer Coefficients, etc. are Seldom
Known to Within a Few Percent
Each Analysis can be Expensive

Cost
About 10N + 40 Times the Cost of One Analysis
This Estimate Assumes Finite Difference Gradients

VR&D

29

29

OPTIMIZATION WORKS
1975

SUPERSONIC CRUISE AIRCRAFT

COMBAT

MISSION

INITIAL

OPTIMUM

SOLVED BY THE ACSYNT PROGRAM

VR&D

5 DESIGN VARIABLES, 2 PERFORMANCE CONSTRAINTS

30

30

OPTIMIZATION WORKS
1975

SUPERSONIC CRUISE AIRCRAFT

1.4

RELATIVE
E MASS

CONVENTIONAL
1.2

NOMINAL
DESIGN

1.0

1.0

TECHNOLOGY
0.9
FACTOR
0.8

ADVANCED

0.8

0.6

SUSTAINED LOAD FACTOR AT M = 0.9

VR&D

TRADE - OFF STUDY

31

31

OPTIMIZATION WORKS
1976: A Two Hour Study
STOL AIRCRAFT TAKEOFF

CONVENTIONAL: W =
G W

VR&D

SKI JUMP: W =
G 1.2W

32

32

OPTIMIZATION WORKS
1978: Today Called Response Surface Method
HIGH SPEED AIRFOIL OPTIMIZATION

INITIAL SHAPE

OPTIMUM: MAXIMIZE LIFT WITH DRAG & MOMENT CONSTRAINTS

VR&D

OPTIMUM: MINIMIZE DRAG WITH LIFT & MOMENT CONSTRAINTS

33

33

OPTIMIZATION WORKS
It Has Been Working For Many Years
The Above Examples are 25-28 Years Old!
The Aircraft Example was a 1 Man Month Study, Verified by a
One Year, $250,000 Study by a Commercial Aircraft Company
The Aircraft Take-off Example Solved a Ph.D. Problem that
Took Over a Year and Got the Wrong Answer
The Airfoil Example Produced a Design Almost Identical to a
Multi Year Wind Tunnel Study

It is Not Debatable that Optimization is Useful

OPTIMIZATION IS THE MOST POWERFUL
DESIGN PRODUCTIVITY TOOL AVAILABLE
TODAY
VR&D

34

34

STRUCTURAL OPTIMIZATION
HISTORY
1960 Schmit: Structural Optimization by Systematic
Synthesis
Combined Finite Element Analysis with Numerical Optimization

1974 Schmit & Farshi: Basic Approximation Concepts

Create an Approximation of the Responses and Use this During
Optimization
The Approximation is Based on Physics, Not Just Linearization

1975 Schmit & Miura: Approximation Concepts Refined

Rod and Membrane Elements

VR&D

35

35

STRUCTURAL OPTIMIZATION
HISTORY
1985+ Optimization Added to Commercial Finite Element
Programs
ANSYS, MSC/Nastran, UAI/Nastran, CSA/Nastran, COSMOS, etc.
None used 2nd Generation Approximation Concepts

1986+ Vanderplaats et al: 2nd Generation Approximation

Concepts
Intermediate Variables
Intermediate Responses
Rod Elements, Beam Elements, Plate/Shell Elements, Solid
Elements, Composites, etc.

1992 GENESIS Version 1.0 Released

Full Use of 2nd Generation Methods
VR&D

36

36

PERSPECTIVE
1960 Optimization of the Three-Bar Truss Require Hour
on an IBM 653 Computer. Today it takes Under 1 Second.
1980 Schmit: Only a Congenial Optimist Could
Conclude that Optimization had a Future with Run Times
Like That

A2
A1

A 3=A1

MINIMIZE WEIGHT
2 LOAD CASES
STRESS LIMITS
VR&D

OPTIMUM IS NOT
P1

P2

FULLY STRESSED!
37

37

STRUCTURAL OPTIMIZATION
BEFORE 1974
CONTROL
PROGRAM

FEM
ANALYSIS
OPTIMIZER

SENSITIVITY
ANALYSIS

The Optimizer Often Required Hundreds of Finite Element

Analyses
Too Expensive for Real Design Applications
VR&D

38

38

MODERN STRUCTURAL
OPTIMIZATION
FEM
ANALYSIS
OUTER LOOP

CONSTRAINT
SCREENING

CONTROL
PROGRAM

SENSITIVITY
ANALYSIS

APPROXIMATE
PROBLEM
GENERATOR

APPROXIMATE
ANALYSIS

INNER LOOP

OPTIMIZER

Use Approximations to Avoid Many Calls to the FEA

Optimizer Never Actually Calls the Finite Element Analysis
VR&D

39

39

THE COST OF STRUCTURAL

OPTIMIZATION
Criteria
Find a Very Good Optimum Quickly
Use as Few Full Finite Element Analyses as Possible

Basis for Criteria

Each Analysis Requires a Full Finite Element Solution
This Can be Very Expensive

Cost
About 10-15 Times the Cost of One Analysis
This Estimate Assumes Analytic Gradients are Calculated
It Also Assumes 2nd Generation Approximation Techniques are Used

VR&D

40

40

STRUCTURAL OPTIMIZATION
Modern Structural Optimization Converts the Original
Design Problem to an Approximate Form Before Calling the
Optimizer
Optimizer Calls Approximate Analysis Many Times
Usually About Ten Detailed Finite Element Analyses are Needed
99% of CPU Time is Analysis and Gradient (Sensitivity) Calculations

Finite Element Models of the Order of 1,000,000 Degrees of

Freedom are Becoming Common
Problems in Excess of 100,000 Design Variables Have Been
Solved by the GENESIS Program
VR&D

41

41

RECENT EXAMPLES USING

MODERN SOFTWARE
DOT: Basic Optimizer
GENESIS: Structural Analysis & Optimization
VISUALDOC: General Multidiscipline Optimization
VISUALSCRIPT: Integration Tool
BIGDOT: Very Large Scale Optimization
VR&D

42

42

DOT
1987: VR&Ds First Commercial Product
3rd Generation Optimization Software by Vanderplaats
CONMIN 1972, ADS - 1984

Solves General Nonlinear Constrained Optimization

Tasks

Used by VisualDOC, GENESIS, MSC.Nastran,

COSMOS/M, FEM5, LMS/CADSI, BEASY,
MODEL CENTER, TBCAD, SINDA/FLUINT,
MERCURY, DAKOTA, Other
VR&D

43

43

GENESIS
1992: Fully Integrated Finite Element Analysis
and Optimization
Topology, Member Sizing & Shape Optimization

NASTRAN Compatible Analysis Data

Currently About 90% of NASTRAN Capabilities
Solves About 99% of Daily Analysis Tasks

Second Generation Approximation Methods

Typically 10 Finite Element Analyses Optimize
Have Optimized with Almost 200,000 Variables
VR&D

44

44

GENESIS
Truck Frame Topology (10,910 Variables)

VR&D

PACCAR

45

45

GENESIS
STEERING KNUCKLE
8 SHAPE VARIABLES
REDUCED MASS 13%
NO INCREASE IN MAXIMUM STRESS

FORD

VR&D

46

46

GENESIS
Fuel Tank Bead Design (99 Shape Variables)
Increased Stiffness > 50%

VR&D

VISTEON

47

47

GENESIS
Air Cleaner Design
Reduced Radiated Noise 10 db

VR&D

DELPHI

48

48

Femb

Interfaces

EDS/IDEAS

MSC.Patran

VR&D

49

49

VisualDOC/VisualScript
1998: Graphics Based Design Environment
General Optimization (Gradient & Non-Gradient)
Design Of Experiments
Response Surface Optimization

Allows User to Interface Optimization Tools with

Almost Any Analysis
Modules Available as APIs For Easy Integration
with other Control Programs
VR&D

50

50

VisualDOC
Hybrid Electric Vehicle

Optimize Control Strategy

Maximize Economy
Minimize Emissions
Limits on Acceleration
and Grade

Results
3.3% Economy Increase
36% Nox Reduction
14% Particulate Mass Reduction

VR&D

NATIONAL RENEWABLE ENERGY LAB

51

51

VisualDOC
Aircraft Wing MDO
Maximize Range for Fixed Gross Weight
Call Aerodynamic Analysis
Call GENESIS for Structural Mass Sub-Optimization

25% Range Increase

VR&D

NASA/VR&D

52

52

Cantilevered Beam

Number of Design Variables, NDV

VR&D

10,000

50,000

100,000

250,000

CONTINUOUS
OPTIMUM

53,744
(233/43)
[9,995/12]

53,744
(243/46)
[49,979/46]

53,720
(209/38)
[99,927/150]

53,755
(262/49)
[249,919/211]

DISCRETE OPTIMUM

54,864
(80/14)

54,864
(92/38)

54,848
(96/25)

54,887
(143/24)

53

53

Summary
Optimization Technology is Well Developed
For General Applications
We Can Couple Almost Any Analysis With
Optimization

For Structural Optimization

Technology is Very Advanced
Find an Optimum Using Only About 10 Finite Element
Analyses

Optimization is the Most Powerful Design

Improvement Tool Available Today
VR&D

54

54