Gyung-Jin Park PHD (Auth.) - Analytic Methods For Design Practice-Springer-Verlag London (2007)

Analytic Methods for Design Practice
Gyung-Jin Park
Analytic Methods
for Design Practice
123
Gyung-Jin Park, PhD
Department of Mechanical Engineering
College of Engineering Science
Hanyang University
Republic of Korea
British Library Cataloguing in Publication Data

Park, Gyung-Jin
Analytic methods for design practice
1. Engineering design
I. Title
620’.0042
ISBN-13: 9781846284724
Library of Congress Control Number: 2007920841
ISBN 978-1-84628-472-4 e-ISBN 978-1-84628-473-1 Printed on acid-free paper

© Springer-Verlag London Limited 2007
ABAQUS is a registered trademark of ABAQUS, Inc., Rising Sun Mills, 166 Valley Street, Providence, RI
02909-2499, USA. http://www.hks.com/
ANSYS is a registered trademark of ANSYS, Inc., Southpointe, 275 Technology Drive, Canonsburg, PA 15317,
USA. http://www.ansys.com/
DOT, GENESIS, and VisualDoc are registered trademarks of Vanderplaats Research & Development, Inc., 1767
S. 8th St., Ste. 100, Colorado Springs, CO 80906, USA. http://www.vrand.com/
I DEAS and Unigraphics are registered trademarks of Electronic Data Systems Corp., 5400 Legacy Drive,
Plano, TX 75024, USA. http://www.eds.com/
iSIGHT is a trademark of Engineous Software, Inc., 2000 CentreGreen Way, Suite 100, Cary, NC 27513, USA.
http://www.engineous.com/
LS-DYNA and LS/INGRID are registered trademarks of Livermore Software Technology Corp., 7374 Las
Positas Road, Livermore, CA 94551, USA. http://www.lstc.com/
MADYMO is a registered trademark of TNO Automotive Safety Solutions, Schoemakerstraat 97, 2628 VK
Delft, The Netherlands. http://www.madymo.com/
ModelCenter is a registered trademark of Phoenix Integration, Inc., 1715 Pratt Drive, Suite 2000, Blacksburg,
VA 24060, USA. http://www.phoenix-int.com/
Oracle is a registered trademark of Oracle Corp., 500 Oracle Parkway, Redwood Shores, CA 94065, USA.
http://www.oracle.com/
Starbase is a trademark of Hewlett-Packard, Inc., 3000 Hanover Street, Palo Alto, CA 94304-1185, USA.
www.hp.com/
X Window System is a trademark of X Consortium, Inc., http://www.x.org/
Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted
under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or
transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the
case of reprographic reproduction in accordance with the terms of licences issued by the Copyright Licensing
Agency. Enquiries concerning reproduction outside those terms should be sent to the publishers.
The use of registered names, trademarks, etc. in this publication does not imply, even in the absence of a
speciﬁc statement, that such names are exempt from the relevant laws and regulations and therefore free for
general use.
The publisher makes no representation, express or implied, with regard to the accuracy of the information
contained in this book and cannot accept any legal responsibility or liability for any errors or omissions that
may be made.
987 65432 1
Springer Science+Business Media
springer.com
To my wife MiSun
Preface
This book is a compilation of what I have learned and studied so far. For the past
twenty years, I have been teaching mechanical design. Most of the contents of this
monograph was created during the past fifteen years of my professorship at
Hanyang University, Ansan, Korea. Mechanical design is a process for
incorporation of the results of theoretical calculation into design. Generally, many
resources are invested in mechanical analyses, however, the incorporation process
is carried out by experiences or engineering intuition. Although it may seem
inefficient, the process shows the characteristics of design well. Since the design
process is more complicated than the analysis process, the designer’s decision
made based on engineering intuition may be more acceptable.
Recently, engineering systems have tended to become quite large. When a
system is large, it is not easy to design it by subjective judgement. Instead, a
design method that is objectively validated is needed. Many design methods have
been developed to this end. A decision making process is required in design,
therefore, the designer inevitably contemplates the solution. One may not even be
able to make the final decision. In this case, a design method with definite logic is
needed. That is, a design tool is exploited to facilitate the decision making process.
Design is classified into conceptual design, which determines the overall
characteristics of a product, preliminary design where selection and distribution of
parts are determined, and detailed design, which fixes the final dimensions. A
design generally proceeds in the order of conceptual, preliminary and detailed
designs. Do we have a universal method that can be applied to all the processes?
It seems that we do not have such a method yet. Instead, different methods are
employed in different steps. From the viewpoint of application of the design
methods, the design process is classified into two processes of conceptual design
and detailed design.
This book treats the design methods for each step of the entire design process.
It explains some analytic design methods and discusses their application. An
analytic design method indicates such a method with a definite theoretical logic.
Each chapter handles a method, and the methods are derived and explained.
Readers can practice the methods with examples. Appendices, which are included
at the end of the chapters, demonstrate case studies. Thus, each chapter is
somewhat independent.
Classes of higher levels of undergraduate or graduate courses can use this book
as a textbook. The students should have basic knowledge of engineering and
viii Preface
mathematics. Also, practical engineers can use it in their design practice. If an

engineer only wants to use a design method in a practice, the theoretical parts can
be skipped. The introduced methods can be exploited in the capstone design or
creative design, which is popular in the engineering education community.
Chapter 1 explains the overall aspects of design and briefly introduces the
analytic methods of the book. Chapter 2 introduces axiomatic design, which is
effective in conceptual design. Various methods for detailed design are introduced
in Chapters 3–7. These include optimization, structural optimization, dynamic
response optimization, design of experiments and robust design. Chapters 8–9
show case studies of the application of the analytic methods to practical problems.
Chapter 10 introduces multidisciplinary design optimization, which uses multiple
analysis and design methods.
Examples and case studies have been developed by Ph.D. and M.S. students at
Hanyang University. Since their contributions are extensively cited throughout the
book, I will omit the specifics here. However, I would like to thank former Ph.D.
students (Prof. B.K. Ko, Dr. Y.S. Park, Prof. K.H. Lee, Dr. S.H. Do, Dr. J.M. Lim,
Dr. K.W. Lee, Dr. W.S. Choi, Dr. M.K. Shin, Dr. K.H. Hwang, Dr. B.S. Kang, Dr.
S.I. Song, Dr. J.W. Shim, Dr. C.M. Park, Dr. J.W. Yi, Dr. J.W. Lee, Dr. K.J. Park,
Dr. J.K. Shin, Dr. G.S. Shin and Dr. S.K. Jeon) and more than sixty former M.S.
students. Especially, Dr. B.S. Kang and Dr. J.K. Shin contributed a great deal.
Special thanks also go to my current students. They are continuing the work
initiated by former students.
I would like to thank Professor B.M. Kwak of KAIST who introduced me to
optimization when I was a Master’s student. Thanks go to Professor J.S. Arora of
the University of Iowa who was my Ph.D. adviser. Also, I would like to thank Dr.
K. Rim who is the chairman of Samsung Advanced Institute of Technology.
Professor N.P. Suh of Massachusetts Institute of Technology taught me axiomatic
design when I was an associate professor and he showed me a new direction. I
would like to thank Professor J.B. Park who is a professor of the University of
Iowa and also happens to be my father-in-law. He gave me valuable advice at
each step of my career and was a role model for me in writing this book. I would
like to thank Professor J.E. Renaud of the University of Notre Dame who
thoroughly reviewed the manuscript and gave me a lot of important advice.
Finally, I would like to thank my family. Thanks go to my parents for their
faith in me. I would especially like to dedicate this book to my father who recently
passed away before seeing the final publication. I would like to thank my wife
who grew up in the U.S. and came to Korea with me. She always supports me
with English corrections and trust. I hope that SangEun, SangHwa and SangWoo
will study their Dad’s book in the future.
Gyung-Jin Park
Contents
1 Introduction......................................................................................................1
1.1 What is Engineering Design? ...................................................................1
1.2 Role and Classification of Design ............................................................4
1.3 Analytic Design Methodology .................................................................6
1.4 Axiomatic Design.....................................................................................7
1.5 Design Optimization.................................................................................8
1.6 Structural Optimization ............................................................................9
1.7 Structural Optimization Under Dynamic Loads .....................................11
1.8 Design with DOE ...................................................................................12
1.9 Robust Design ........................................................................................13
1.10 Multidisciplinary Design Optimization ..................................................13
1.11 Summary.................................................................................................15
2 Axiomatic Design ...........................................................................................17

2.1 Introduction ............................................................................................17
2.2 The Independence Axiom.......................................................................19
2.2.1 The Independence Axiom...........................................................19
2.2.2 Independence..............................................................................25
2.2.3 Physical Integration ....................................................................32
2.3 The Information Axiom..........................................................................33
2.3.1 The Calculation of Information Contents Using Probability......33
2.3.2 Probability Density Function and Information Content .............35
2.3.3 The Calculation of Information Content
for a Decoupled Design ..............................................................38
2.4 The Application of Axiomatic Design....................................................52
2.5 Software Design Using the Axiomatic Approach ..................................64
2.5.1 Software Design .........................................................................64
2.5.2 Conventional Languages and Axiomatic Design .......................64
2.5.3 Object Oriented Programming and Axiomatic Design...............69
2.6 Discussion ..............................................................................................71
2.7 Exercises.................................................................................................72
2.A Corollaries and Theorems.......................................................................76
2.B Axiomatic Design of a Beam Adjuster for a Laser Marker....................82
2.B.1 Problem Description ...................................................................82
x Contents
2.B.2 Axiomatic Analysis of an Existing Design.................................84

2.B.3 The Development of a New Beam Adjuster...............................86
2.B.4 Summary.....................................................................................89
2.C The Development of a Design System for a TV Glass Bulb ..................90
2.C.1 Problem Description ...................................................................90
2.C.2 The Conventional Design Process for a Glass Bulb...................90
2.C.3 Automatic Design Software for Product Design ........................92
2.C.4 Software Development ...............................................................95
2.C.5 Summary.....................................................................................97
2.D The Development of a Design System for the EPS Cushioning
Package of a Monitor .............................................................................99
2.D.1 Problem Description ...................................................................99
2.D.2 The Development of an Automatic Design System
for the EPS Cushioning Package ................................................99
2.D.3 Summary...................................................................................103
3 Design Optimization ....................................................................................107

3.1 Introduction ..........................................................................................107
3.2 Optimization Problems in Engineering ................................................109
3.3 Analytic Optimization Methods ...........................................................115
3.3.1 Graphical Optimization ............................................................116
3.3.2 Unconstrained Optimization.....................................................116
3.3.3 Constrained Optimization.........................................................119
3.4 General Concepts of Numerical Methods in Optimization ..................122
3.5 Linear Programming.............................................................................125
3.5.1 Standardization of Linear Programming ..................................125
3.5.2 Basic Concepts of Linear Programming...................................129
3.5.3 Terminologies of LP.................................................................131
3.5.4 Simplex Method .......................................................................131
3.6 One-dimensional Minimization ............................................................133
3.6.1 Equal Interval Search ...............................................................134
3.6.2 Golden Section Search .............................................................135
3.7 Numerical Methods for Unconstrained Optimization ..........................138
3.7.1 Steepest Descent Method..........................................................138
3.7.2 Conjugate Gradient Method .....................................................141
3.7.3 Quasi-Newton Method .............................................................142
3.8 Numerical Methods for Constrained Optimization ..............................144
3.8.1 General Aspects of the Primal Method.....................................145
3.8.2 Primal Method ..........................................................................149
3.8.3 Transformation Method............................................................153
3.9 Exercises...............................................................................................154
3.A Brief Review of Mathematical Terminologies and Background ..........163
Contents xi
4 Structural Optimization ..............................................................................171

4.1 Introduction ..........................................................................................171
4.2 Finite Element Method .........................................................................172
4.2.1 Stress and Strain .......................................................................173
4.2.2 Formulation of the Finite Element Method ..............................174
4.3 Formulation of Structural Optimization ...............................................187
4.4 Sensitivity Analysis ..............................................................................200
4.4.1 Finite Difference Method .........................................................200
4.4.2 Analytic Method with Discrete Equations................................205
4.4.3 Semianalytic Method ................................................................214
4.4.4 Analytic Sensitivity Analysis Using Continuum Equations .....216
4.5 Methods of Structural Optimization .....................................................221
4.5.1 Direct Method...........................................................................221
4.5.2 Approximation Method ............................................................222
4.5.3 Multiple Loading Conditions and Design Variable Linking ....226
4.6 Application of Structural Optimization ................................................229
4.6.1 Size Optimization .....................................................................229
4.6.2 Shape Optimization ..................................................................231
4.6.3 Topology Optimization.............................................................237
4.7 Exercises...............................................................................................240
4.A Automotive Door Design with the ULSAB Concept
Using Structural Optimization..............................................................244
4.A.1 Problem Description .................................................................244
4.A.2 The Design Process for Weight Reduction...............................244
5 Dynamic Response Optimization................................................................255

5.1 Introduction ..........................................................................................255
5.2 Optimization in the Time Domain........................................................256
5.3 Time-dependent Constraints.................................................................257
5.3.1 Treatment of Time-dependent Constraints
in the Direct Method.................................................................257
5.3.2 Treatment of Time-dependent Constraints
in the Transformation Method ..................................................263
5.4 Sensitivity Analysis ..............................................................................265
5.4.1 Direct Differentiation Method ..................................................265
5.4.2 Adjoint Variable Method..........................................................270
5.4.3 Time Finite Element Method....................................................274
5.5 Approximation......................................................................................276
5.5.1 Global Approximation..............................................................276
5.5.2 Local Approximation................................................................277
5.6 A Method Using Equivalent Static Loads ............................................279
5.6.1 Equivalent Static Loads ............................................................279
5.6.2 Optimization in the Time Domain
Using Equivalent Static Loads .................................................281
5.6.3 Optimization of Flexible Multibody Dynamic Systems ...........286
xii Contents
5.7 Structural Optimization in the Frequency Domain...............................292

5.7.1 Optimization of Harmonic Responses ......................................292
5.7.2 Sensitivity Analysis ..................................................................295
5.7.3 Treatment of Constraints and Minimization
of the Maximum Amplitude .....................................................301
6 Design of Experiments .................................................................................309

6.1 Introduction ..........................................................................................309
6.2 Design of Experiments (DOE) .............................................................311
6.2.1 One-way Factorial Design ........................................................312
6.2.2 Two-way Factorial Design .......................................................317
6.3 Design Using Orthogonal Arrays .........................................................325
6.3.1 Orthogonal Arrays ...................................................................325
6.3.2 Analysis of Experimental Results Using Orthogonal Arrays ...327
6.3.3 Design Using Orthogonal Arrays .............................................336
6.4 Design with Orthogonal Arrays Considering Interactions ...................340
6.5 Design Using Orthogonal Arrays for Constrained Problems ...............351
6.6 Sequential Algorithm with Orthogonal Arrays (SOA).........................357
6.7 Design Using the Response Surface Method .......................................360
6.7.1 Introduction ..............................................................................360
6.7.2 Generation of the Response Surface.........................................362
6.7.3 Optimization Using the Response Surface Method..................368
6.7.4 Selection of Candidate Points and Analysis
of the Optimization Results ......................................................371
6.8 Exercises...............................................................................................372
6.A Statistics................................................................................................382
6.B Orthogonal Arrays................................................................................386
6.B.1 Orthogonal Arrays with Two Levels ........................................386
6.B.2 Orthogonal Arrays with Three Levels ......................................388
6.B.3 Orthogonal Array with Mixed Levels ......................................390
7 Robust Design...............................................................................................393
7.1 Introduction ..........................................................................................393
7.2 Mean and Variance...............................................................................394
7.3 Taguchi Method ...................................................................................399
7.3.1 Introduction ..............................................................................399
7.3.2 The Loss Function and S/N Ratio.............................................401
7.3.3 Parameter Design......................................................................404
7.3.4 Various Methods Using the Taguchi Method...........................412
7.4 Robust Optimization.............................................................................416
7.4.1 Introduction ..............................................................................416
7.4.2 Robustness of the Objective Function ......................................418
7.4.3 Robustness of Constraints ........................................................421
7.4.4 Optimization Methods ..............................................................425
Contents xiii
7.5 Robust Design by Axiomatic Design ...................................................428

7.6 The Robustness Index ..........................................................................429
7.7 Summary...............................................................................................435
7.8 Exercises...............................................................................................435
8 Case Studies with Analytic Design Methods..............................................443

8.1 Introduction ..........................................................................................443
8.2 Design of a Strip Casting Roll..............................................................444
8.2.1 Problem Description .................................................................444
8.2.2 Optimization Formulation for a Strip Casting Roll ..................446
8.2.3 Optimization Results ................................................................448
8.2.4 Summary...................................................................................454
8.3 Design of a Spacer Grid Spring............................................................455
8.3.2 Design Requirements................................................................456
8.3.3 Optimization of the Grid Spring ...............................................457
8.3.4 Summary...................................................................................462
8.4 Design of an Automobile Rearview Mirror..........................................464
8.4.1 Finite Element Modelling of the Rearview Mirror ...................464
8.4.2 Topology Optimization of the Rearview Mirror.......................467
8.4.3 Size Optimization of the Rearview Mirror ...............................468
8.4.4 Design Using Orthogonal Arrays in a Discrete Space..............469
8.4.5 Robust Design of the Rearview Mirror ....................................472
8.5 Structural Analysis and Optimization of a Low Speed Vehicle Body .474
8.5.2 The Aluminum Space Frame and the Low Speed Vehicle .......475
8.5.3 Structural Analysis of the Low Speed Vehicle and
a Comparative Study with a General Passenger Car ................477
8.5.4 Improvement of the Low Speed Vehicle
Using Structural Optimization..................................................489
8.5.5 Summary...................................................................................494
9 Application of Design Methods to Automobile Safety Problems.............497

9.1 Introduction ..........................................................................................497
9.2 Modelling for Safety Simulation ..........................................................499
9.2.1 Modelling for Multibody Dynamic Analysis............................499
9.2.2 Modelling Using Nonlinear Finite Element Analysis...............501
9.3 Design of an Airbag Using Orthogonal Arrays ....................................502
9.4 Design of an Energy Absorbing Steering System
Using Orthogonal Arrays .....................................................................504
9.4.2 Modelling of the Energy Absorbing Steering System ..............505
9.4.3 Validation of the Model Through the Body Block Test ...........506
9.4.4 Parameter Study of the Energy Absorbing Parts ......................508
xiv Contents
9.4.5 Definition of the Objective Function........................................510

9.4.6 Design Using Orthogonal Arrays .............................................511
9.4.7 Summary...................................................................................513
9.5 Design of a Side Impact Beam of an Automobile Door .......................513
9.5.2 Problem Formulation................................................................515
9.5.3 Design of an Impact Beam Using Orthogonal Arrays..............516
9.5.4 Design Using the Response Surface Method............................516
9.5.5 Summary...................................................................................518
9.6 Design of a Motor Driven Tilt/Telescopic Steering System.................519
9.6.2 Conceptual Design Using Axiomatic Design ...........................519
9.6.3 Detailed Design for Safety .......................................................521
9.6.4 Detailed Design Considering Vibration....................................526
9.6.5 Summary...................................................................................530
9.7 Crash Analysis and Design of a Belt Integrated Seat
for Occupant Safety..............................................................................530
9.7.2 Simulation Modelling ...............................................................532
9.7.3 Simulation of Frontal and Rear Impacts ...................................533
9.7.4 Structural Optimization of the BIS ...........................................540
9.7.5 Summary...................................................................................544
9.8 Determination of a Crash Pulse and Optimization of Crash
Components Using Response Surface Approximate Optimization ......545
9.8.2 Response Surface Approximate Optimization..........................545
9.8.3 Determination of a Crash Pulse to Minimize Occupant Injury.547
9.8.4 Optimization Formulation for the Determination
of a Crash Pulse ........................................................................548
9.8.5 Frontal Crash with an Airbag ...................................................550
9.8.6 Rear End Impact .......................................................................552
9.8.7 Design of Safety Devices for a Frontal Crash ..........................553
9.8.8 Seat Design in Rear End Impact...............................................556
9.8.9 Summary...................................................................................558
10 Multidisciplinary Design Optimization......................................................561

10.1 Introduction ..........................................................................................561
10.2 Multidisciplinary Design Optimization ................................................563
10.2.1 Coupling in Analysis ................................................................564
10.2.2 Formulation of MDO................................................................566
10.2.3 Classification of MDO Methods...............................................567
10.3 Basic Theories for MDO Methods .......................................................571
10.3.1 Linear Decomposition and Global Sensitivity Equation ..........571
10.3.2 Optimum Sensitivity.................................................................575
10.4 Multidisciplinary Design Optimization Methods .................................579
Contents xv
10.4.1
Multidisciplinary Feasible (MDF)............................................579
10.4.2
Individual Discipline Feasible (IDF) ........................................580
10.4.3
All-at-once (AAO)....................................................................582
10.4.4
Concurrent Subspace Optimization (CSSO) ............................585
10.4.5
Bilevel Integrated System Synthesis (BLISS) ..........................593
10.4.6
Collaborative Optimization (CO) .............................................597
10.4.7
Multidisciplinary Design Optimization Based
on Independent Subspaces (MDOIS) .......................................602
10.4.8 Comparison of MDO Methods .................................................605
10.5 Discussion ............................................................................................608
10.A Application of an MDO Algorithm to the Design of a Belt Integrated
Seat Taking Crashworthiness into Consideration.................................609
10.A.1 Problem Description .................................................................609
10.A.2 The Interdisciplinary Relationship Between the Disciplines
Involved....................................................................................610
10.A.3 Formulation for the Design of a BIS Taking Crashworthiness
into Consideration ....................................................................610
10.A.4 Application of MDOIS to the Design of a BIS.........................611
Index ....................................................................................................................621
1 Introduction
1.1 What is Engineering Design?

What is design? The answer can vary. In a broad sense, design may mean
developing a plan in ordinary life, creative work that serves as a tool to provide
aesthetic beauty or convenience, etc. The concept of design in this book is, in a
narrow sense, a plan to produce products, which is called mechanical design in the
engineering community. It is comprised of functional requirements and
specifications, decisions for detailed dimensions, mechanical drawings for
products, etc. Mechanical design has been conducted since prehistoric times when
tools for hunting were manufactured. Although the history of mechanical design is
longer than any engineering activity, a precise definition is still unclear because
design is manifold. Nevertheless, engineers tend to consider the planning stage for
products as design.
Let us investigate the design stage in the process of engineering. Figure 1.1
presents the process from the planning stage to the marketing stage for products.
The shadowed area in Figure 1.1 is generally regarded as the design stage. In the
design process, the concepts in the planning process are converted to physical
Production planning
Product design
Feedback
Manufacturing
Feedback
Marketing
Figure 1.1. Flow of product design and manufacturing

2 Analytic Methods for Design Practice
Figure 1.2. An example of a drawing
engineering parts, and various aspects are defined for manufacturing. Therefore,
the subjects in the planning process are thoroughly examined and details of parts,
materials, dimensions, etc. are determined. In general, the output from the design
process is expressed by drawings. An example of a drawing is illustrated in Figure
1.2. The role of the design process is a bridge between the stages of the planning
and the manufacture of the final products.
A designer typically goes through decision-making processes. What would be
rational and logical decisions? Who is a good designer and who is not? Can the
characteristics of good decisions be generalized? Can a good design of a specific
product also be good for other products? Design activities can be viewed from
different aspects depending on the function of each part of a product. According
to the manner of the decision making process, there are two viewpoints for the
definition of design. In some parts aesthetic beauty is emphasized. An example is
the design of the external shape of an automobile. In other parts, the engineering
functionality is more important. The engine of an automobile is a prime example.
That is, the design process should be based on objective logic and a general design
principle. While the definition of design may include many views, Figure 1.3 and
Figure 1.4 contrast the different design requirements and objectives in automobile
design. External appearance is mostly considered in Figure 1.3. On the other hand,
strength and manufacturability are important in the design of Figure 1.4. If we
find a general principle for good designs, general design methodologies can be
developed. Recently, design methods have begun to emerge by the finding of
common aspects of good designs. An analytic design method is a design
methodology that has a certain process for design practice.
There are many excellent designs around us. For example, there are numerous
bicycles, automobiles, electrical devices, etc. Some of them have been created by
inspiration from unknown sources while others may have been improved gradually.
Excellent patents are being created in these ways as well. Can we derive specific
Introduction 3
Figure 1.3. A car with an aesthetic shape
rules or general principles for an excellent design? Engineering designers

contemplate the existence of such principles. They also wonder whether a
standard process exists for a good design when artistic aspects are important.
Actually, it seems that adoption of a theory is not logical in order to explain
excellent products from an artistic viewpoint. However, it might be possible to
adopt a logical theory for excellent products in usage and in an engineering sense.
That is, the engineering design process can be analyzed and common aspects of
good products can be derived. Design theories are established from such efforts.
This book only covers design theories that are established and usable in design
practice. It excludes theories that pursue aesthetic aspects because they may lack
objectiveness. Therefore, theoretical approaches for engineering performance are
considered. In other words, this book only explains performance-based
engineering design theories.
Figure 1.4. Inside design of a car

1.2 Role and Classification of Design

Figure 1.1 presents the flow of engineering activities. The concept of a target
product is defined at the beginning of the flow. The process may determine if the
target product is needed in the market through surveys. It also determines
functions and applications of the product, predicted profit, etc. If the concept is
established, detailed characteristics of the product are defined. This procedure can
be called design. However, engineering design means the post process of
conceptualization. That is, physical specification is defined to materialize the pre-
determined concept. This is called the design process. Details of the flow are out
of the scope of this book.
Design has an extremely large influence on the entire engineering process.
Only 5% of the total cost is spent on the design process; however, the result of the
design process determines most of the total product cost. The design process of
Figure 1.1 does not affect the above processes. However, it considerably impacts
the processes below. When erratic behavior is found, it is corrected by feedback
processes. Feedback processes are expensive and introduce loss to the
manufacturers. Therefore, the loss can be reduced through minimization of errors
in the earlier stages such as the design process.
Design can be classified according to methods and the environment. First, the
shadowed area in Figure 1.1 is divided into conceptual design and detailed design.
These are illustrated in Figure 1.5. Some researchers may use three stages such as
conceptual design, preliminary design and detailed design. They are defined based
on the activities involved in the particular stage. However, two stages are
sufficient from the application viewpoint of design methodologies. Conceptual
design is carried out at the early stage of design. Decisions are made for the
definition of functional requirements, selection and location of parts and the
relation between parts. Most of the important factors are determined in this
process. On the other hand, detailed design determines sizes and shapes for the
concept determined in conceptual design. Experiments and analyses are conducted
to evaluate engineering performances. A typical procedure for a detailed design is
presented in Figure 1.6.
Design can be grouped into system design and element design, based on the
Conceptual design
Overall selection of functions and parts
Detailed design
Determination of sizes of parts
Figure 1.5. Sequence of design

Introduction 5
Definition of problem
Determination of design
Analysis and experiment
Satisfied?
No
Yes
Final design
Figure 1.6. Detailed design process
number of parts involved. In system design, many design factors are

simultaneously determined by considering the design circumstances. That is, a
system is designed from the interrelation and correlation of parts. In element
design, one element is determined. In the design of a lathe, if we design the entire
layout of the machine, it is a system design. However, in element design each part
can be designed. These days, even an element design can adopt a design
methodology for system design. For example, shaft design is carried out by
considering the bearings and keys around a shaft.
In some cases, design is classified into creative design, imitative design and
design improvement according to the intention at the beginning. Creative design is
a design effort to create a nonexisting product. It is mostly adopted by advanced
design studios where new consumer products are conceptualized and developed.
Creative design is difficult and it is not usually chosen for complex engineering
systems such as automobile, aircraft or space systems. However, it is frequently
chosen in the electronic industries. The development of cellular phones, DVD,
MP3, etc. is carried out by creative design. Imitative design is usually utilized by
followers. Designers are trying to avoid patents and make better products than
existing ones. It is easier than creative design; however, the impact is not large.
Design improvement is an activity to improve the performance of an existing
product. Usually, designers try to improve their own products by changing or
adding functions.
Objective and scientific methods tend to be implemented in the various
aforementioned designs. In other words, some definite methods are applied to the
design processes. It is ideal to have a definite method in all the design processes.
However, it seems to be impossible to have such a method. Therefore, an
appropriate method should be chosen according to the design processes or the

design environment.
1.3 Analytic Design Methodology

A design principle is defined as a general and scientific rule, which can be
commonly used in all design activities. A scientific rule should have objective
logic and pertinence. Many researchers have contemplated whether scientific rules
exist in engineering design. Design engineers and design theorists pursue the
search for these rules. Many designers emphasize engineering intuition and
experiences, especially in conceptual design. On the other hand, scientific
principles are frequently utilized in detailed design. The results from scientific
analyses can be incorporated in the design to insure performance criteria are
satisfied. Intuition and experiences are used even in this incorporation process.
Thus, the usage of analysis results depends on the designer’s personal judgment
and it is obvious that this design process is inefficient. Therefore, design engineers
are trying to find design principles and establish scientific ways for design.
Analytic design is a design activity using scientific principles or processes with
rigorous methods to achieve a good design. The utilized principle or method is a
design methodology.
For conceptual design, analytic design methodologies are defined by examining
the common aspects in good designs. Axiomatic design is typical for the analytic
design methodology. Axiomatic design consists of two axioms that can be applied
to all the design activities. They are the Independence Axiom and the Information
Axiom. The Independence Axiom says that design parameters should be selected
to independently satisfy functional requirements. The Information Axiom selects
an optimum design with minimum information content among the designs that
satisfy the Independence Axiom. As the axiom states, if a counterexample is
found, the framework of the axiomatic design is destroyed. Recently, many case
studies have shown that this method is quite useful. TRIZ (Russian acronym for
the theory of inventive solving) is also used as an analytic design methodology to
aid the thinking process. A designer analyzes the flow of a design as TRIZ
indicates. Many patented designs are stored in a database and the designer can
create a similar idea from past work. There are methodologies to define design
processes for each case study. The sequence of a design is logically defined in
these methods.
Analytic design methodologies are most frequently utilized in detailed design.
The use of numerical optimization has been increasingly exploited during the past
decades. In optimization, an objective function is minimized or maximized while
constraint functions are satisfied. Objective and constraint functions are
mathematically defined to express the design objective and specifications,
respectively. Various mathematical algorithms have been developed. With the
improvement of computational performance, the use of optimization has expanded
to a variety of areas. Optimization might be the only mathematically proven
design methodology. However, its difficulties lie in the fact that a design problem
Introduction 7
should be mathematically defined. Many researches have proposed other methods

such as design of experiments (DOE), robust design, genetic algorithms,
multidisciplinary design optimization (MDO). However, the viewpoints are
similar and difficulties and restrictions still exist. It is noted that the impact of the
detailed design is not quite as immense because it is utilized after most design
concepts are determined. Design methodologies handled in this book are briefly
explained in the following sections.
1.4 Axiomatic Design

In axiomatic design, design is defined as an interplay between “what we want to
achieve” and “how we achieve it.” The designer intends to meet the design goal
through the appropriate combination and allocation of functions and capabilities of
both sides. The engineering process is defined by domains as illustrated in Figure
1.7. In the customer domain, what a customer wants is described. The description
is named the customer attributes (CAs) or customer needs. Customer needs are
transformed to functional requirements (FRs) in the functional domain. The
functional requirements are described in engineering terminology. To satisfy the
functional requirements, design parameters (DPs) should be found in the physical
domain. This process is usually called the design process. Production variables
(PVs) in the production domain are found for manufacture of the design
parameters. In axiomatic design, the relation between two neighboring domains is
grasped and the aspects of the following domain are defined according to the
characteristics of the previous domain. It is called the mapping process between
Customer domain G Functional domain Physical domain Production domain
What? How? How? How?

How HowG
What?
What What?
G G
G
Why? G Why?
Constraints
Why?
G
Constraints
G
Customer needs Functional requirements Design parameters Process variables
Figure 1.7. Relationship between domains, mapping and spaces

the two domains.

Design axioms are defined as follows:
Axiom 1: The Independence Axiom
Maintain the independence of the functional requirements (FRs).
Axiom 2: The Information Axiom
Minimize the information content of the design.
The axioms may look simple. However, they imply various meanings. The
first axiom can be an expression of the intuition of engineers. It means that when
we design a complicated machine, the functional requirements should be
independently satisfied by design parameters which are selected by the designer.
The functional requirements defined in the functional domain in Figure 1.7 should
be independently satisfied by adequate design parameters in the physical domain.
The second axiom means that when multiple designs to satisfy the first axiom are
found, the one with minimum information content should be selected. The
information content is quantified by the probability of success. It also corresponds
to the general experience of engineers. The second axiom is related to robust
design, which will be explained later. The axioms are quite simple; however,
application can sometimes be very difficult.
Axiomatic design can be applied to the following areas:
(1) The creative process in conceptual design
(2) The comparison of existing designs
(3) Improvement of the present design
Details will be introduced later. Axioms have mostly been utilized in geometry.
As in geometry, if a counterexample is found, the axioms are nullified. Also,
theorems and corollaries are induced from the two axioms.
1.5 Design Optimization

Optimization is a field of study focussing on numerical solution of a mathematical
design problem formulated as follows:
Find b Rn (1.1a)
to minimize f (b) (1.1b)
subject to hi (b) 0 , i 1, ,l (1.1c)
g j (b ) d 0 , j 1, ,m (1.1d)
bL d b d bU (1.1e)
where b is the design variable vector with n elements, f is the objective function,
hi is the ith equality constraint, g j is the jth inequality constraint, b L and b U are
Introduction 9
the lower and upper bounds of b, respectively, l is the number of equality

constraints and m is the number of inequality constraints.
In optimization, a given objective function is minimized or maximized. When
an optimization theory is employed in engineering design, it is called design
optimization. In design optimization, the objective function in Equation 1.1b is an
engineering attribute targeted for improvement and various specifications are
transformed to Equations 1.1c and 1.1d. Therefore, when optimization is utilized
in engineering design, the design problem should be formulated in the form of
Equation 1.1.
Design optimization is generally used in detailed design since it determines the
detailed values of the design variables. The process of the conceptual design is not
usually expressed in Equation 1.1, because selection or placement of parts is
determined. Therefore, optimization is utilized after the conceptual design is
finished. When optimization is applied, a specific algorithm is employed to solve
the problem in Equation 1.1.
If the problem is well formulated, design optimization provides an excellent
solution. However, when the formulation is unclear or different from the real
problem, we may get a poor or nonoptimal design. The optimum solution may
marginally satisfy some constraints. In this case, the constraints can be violated
with small perturbation of design variables. Therefore, we should be meticulous
when applying optimization to problems where satisfaction of constraints is
important.
Design optimization is frequently misunderstood because the word
optimization seems to imply that a single optimal solution can be found for all
design problems. Optimization is only applicable when a design problem can be
mathematically formulated as in Equation 1.1. It is especially useful in detailed
design. Therefore, it should be noted that the impact is usually smaller than that of
conceptual design.
1.6 Structural Optimization

The area where optimization is applied the most is structural design. Optimization
can be used to determine the detailed shapes or sizes of structures. Recently, the
development of the finite element method (FEM) has been extensive and FEM is
commonly used. The development of structural optimization methods has evolved
and paralleled the development of FEM.
The linear static FEM problem is primarily solved in structural optimization.
The governing equation for a linear static problem is formulated in FEM as
follows:
K (b)z f (1.2)
where K is the stiffness matrix, z is the nodal displacement vector, b is the design
variable vector for sizes and shapes, and f is the external load vector.
Table 1.1. Classification of structural optimization
Structural
Design variable
optimization
Design variables are thickness or section properties. The

Size optimization
domain for FEM is fixed during optimization.
Design variables are the shapes of structures. The domain for
Shape optimization
FEM is changed during optimization.
Configuration The transformation matrices of FEM are functions of design
optimization variables. Sometimes it is considered as shape optimization.
Design variables are the elastic modulus of the elements for

Topology
FEM. The problem can be viewed as a material distribution
optimization
problem.
In structural optimization, the governing equation is computationally solved

and the solution is used for evaluation of the objective function and constraints.
The governing equation is considered as equality constraints in Equation 1.1c.
Structural optimization is formulated as follows:
Find b R n (1.3a)
to minimize f (b, z ) (1.3b)
subject to K (b)z f (1.3c)
g j (b, z ) d 0 , j 1,, m (1.3d)
bL d b d bU (1.3e)
The objective function in Equation 1.3b can be the weight of the structure or a
specific response. The inequality constraints in Equation 1.3d are generally
defined by limit values on displacements, stresses, natural frequencies, etc. In the
optimization process, we need sensitivity information for the function in Equation
1.3 with respect to design variables. It is known that the sensitivity evaluation is
quite expensive. Therefore, efficient calculation of sensitivity information is
significant.
Structural optimization is classified according to the characteristics of the
design variables. The classification is shown in Table 1.1. In structural design,
various analysis methods are employed to examine the status of the structure.
However, research has been concentrated mostly on linear static problems.
Therefore, the application of optimization is still limited. More research is needed
for problems requiring other analysis methods (i.e. nonlinear FEM).
Introduction 11
1.7 Structural Optimization Under Dynamic Loads

Structures under dynamic loads show varying responses with respect to time.
Therefore, dependency upon time should be considered when the responses are
included in the optimization. Optimization with time-dependent properties is
called dynamic response optimization. In the real world, all loads are dynamic.
Therefore, dynamic properties should be considered in structural optimization.
However, structural optimization is generally carried out with static loads. In other
words, structural optimization is mostly a static response optimization. The reason
is that dynamic response optimization is very difficult due to time dependency. In
this section, dynamic response optimization is briefly discussed and the elements
needed for dynamic response optimization are described.
When f(t) is an external load vector and t is time, the governing equation in the
FEM formulation is as follows:
M (b )z K (b ) z f (t ) (1.4)
where M is the mass matrix and z is the acceleration vector. Equation 1.5 is the
general formulation for dynamic response optimization.
Find b Rn (1.5a)
to minimize f (b, z, t ) (1.5b)
subject to M (b)z K (b) z f (t ) (1.5c)
g j (b, z, t ) d 0 , j 1, ,m (1.5d)
bL d b d bU (1.5e)
Equation 1.4 is imposed as equality constraints as shown in Equation 1.5c.

Dynamic response optimization is the process of obtaining the design solution of
Equation 1.5. It is different from the static response optimization in Equation 1.3
in that the time variable is included in the functions.
Usually, numerical methods are utilized to solve the problem in Equation 1.5.
In a numerical method, the time should be discretized and the constraints in
Equations 1.5c and 1.5d are generated for each of the discretized time steps.
Therefore, it is crucial to efficiently handle this large number of constraints.
Various methods have been proposed to manage time-dependent constraints. For
example, functionals by integration in the time domain can be used or only peak
values are considered. These methods can be very complicated for the calculation
of functions and sensitivity analysis. They have been applied to small-scale
problems. However, it is rare to use them for large-scale problems.
In some research, the dynamic equation is converted to a static equation and
optimization is carried out with the static equation in an iterative manner. The
method exploits the well developed static response optimization. A dynamic load
is discretized and transformed to multiple static loads and a structure is optimized

with multiple loading conditions.
1.8 Design with DOE

Experiments are conducted to examine the characteristics of a system. After the
experiments, the results are analyzed according to the objectives of the
experiments. From the analysis, important factors are identified and values for the
factors are sometimes determined. Design of experiments (DOE) is developed to
systematically conduct the experiments and analyze the results. Recently,
researchers have tended to adopt DOE in design practice.
The design method with DOE sets up a design model of a product and finds the
design solution. Real experiments are utilized or computer simulation is employed.
Generally, the metamodel, which is a model of the model, is used. At certain
points, experiments or simulations are carried out and the metamodel is established
by approximation. Optimization is applied to the approximated metamodel. The
cost for experiments or simulations is considerably reduced. The method is
utilized in two ways.
First, the methods in DOE are used directly. Factors in experiments are
considered as design variables and the experimental result is the objective function.
If an experiment is not expensive, the factorial design can be adopted for a full
combination of the factors. On the other hand, when the experiment is expensive,
fractional factorial design can be used. For example, an orthogonal array can be
used for a fractional factorial design. Since orthogonal arrays are easy to use, they
can be exploited for a design in a discrete space.
Second, approximated functions are used. Typically, the response surface
method is employed for the approximation. It is obvious that a metamodel is
utilized in the response surface method. Experimental values are calculated at
some candidate points and the functions are explicitly approximated with the
results of the experiments. The approximated functions are generally quadratic.
The least squares method is usually used for the approximation that minimizes the
difference between the experimental results and the values of the explicit functions.
When the approximated functions are determined, optimization can be employed
to find an optimum solution. It is noted that the final solution can be different
depending on the experimental points used.
A design with DOE is difficult to use when there are many design variables.
To obtain a meaningful solution, the design results should be validated by various
theories of statistics. The method is popular because engineers can easily use it
without thoroughly understanding the statistics. While the method is easy to use,
considerable judgment should be exercised in selecting the proper experimental
design.
Introduction 13
1.9 Robust Design

Designers pursue not only improved product performance but also quality. One of
the efforts in this area is robust design. In robust design, quality improvement is
attained by minimizing the change of product performance when there are
tolerances on design variables or unknown environmental perturbation. Since Dr.
Taguchi proposed a robust design method by using DOE, various methods have
been developed and applied.
In design, nominal values of design variables are usually determined. This is
called nominal design. The performance (objective function in optimization) from
the nominal design is called a nominal value. The nominal values usually have
tolerances (variances). Also, some environmental variables that are difficult to
include in design may have variances. These variances are called noise. If there is
noise, variance on the nominal value occurs. In a sensitive design, the variance on
the nominal value is quite large with respect to the noise. A sensitive design is
considered as a bad design. When the nominal value is distributed closely to a
target with noise, it is called an insensitive design. Robust design pursues the
insensitive design.
There are quite a few methods in robust design. First, the Taguchi method for
DOE is directly used. A response (characteristic function) is evaluated based on
an orthogonal array considering the tolerances and the distribution of the response
is minimized. The loss function or the signal-to-noise (S/N) ratio is employed for
the distribution. Although many drawbacks are pointed out, the method is still
popularly used. Second, the objective function of optimization is replaced by the
sensitivity of the response and the sensitivity is minimized. It can be expensive
because second-order derivatives are needed. Therefore, some modified methods
are adopted to overcome the difficulties. Third, axiomatic design is applied to
robust design. It was found that the information content of axiomatic design is
related to robustness. The information content is based on the probability of
success. When multiple designs are found after the Independence Axiom has been
applied, robust design is accomplished in the process of finding the best design
with the Information Axiom.
1.10 Multidisciplinary Design Optimization

Various methods for detailed design were introduced in the previous sections.
They largely belong to the area of optimization. Numerous cases can easily be
found for the application of optimization. One of the general features in the
application is that only one discipline is involved in the analysis of a system.
However, analyses of multiple disciplines are needed in many cases.
Multidisciplinary design optimization (MDO) has been proposed to solve such
problems.
First, the coupled relationships between disciplines must be identified. The
coupled relationship means that the output of a discipline becomes the input of the
other discipline and vice versa. Complicated relations occur due to the coupled
relations and entire system analysis is extremely difficult. A typical coupled
Aerodynamics
Pressure
Pressure
Deformation Deformation
Bending stress
Structural analysis Control

Hinge
Figure 1.8. Typical coupling relationship between various analysis methods
Start End
Yes
Analysis of multiple disciplines No Integration

Converge?
of
Analysis 1 optimization
Analysis 2 Analysis 3
Optimum sensitivity analysis
Sensitivity analysis Simultaneous optimization
Analysis 1 Discipline 1
Global
sensitivity
Analysis 2 equation Discipline 2
Analysis 3 Discipline 3
Figure 1.9. An example of a definition for MDO

Introduction 15
relationship is illustrated in Figure 1.8.

When we have coupled relationships in the optimization process, the methods
utilized can vary. An example of the method is presented in Figure 1.9 where
many difficulties are apparent. To calculate the responses and sensitivities, the
coupled analysis should be conducted. It is quite expensive to calculate this
information. Some other methods have been proposed to overcome these
difficulties, but a unified theory for MDO has not yet been proposed.
1.11 Summary
Design methodologies are described briefly. As mentioned earlier, axiomatic
design is introduced as a method for conceptual design. There exist some other
methods to enhance efficiency of the design process. They are design for
assembly (DFA), design for manufacturing (DFM), etc. Also, TRIZ can be
exploited to create a design idea. These are the methods for conceptual design,
which however are not handled in this book.
For detailed design, optimization, DOE and robust design are briefly discussed.
Recently many other methods have been developed such as genetic algorithms
(GA), simulated annealing (SA), design with fuzzy set theory, design with neural
networks, particle swarm optimization algorithms, etc. Although they belong to
the area of optimization, they are not included herein.
References
Antonsson EK, Cagan J (2001) Formal Engineering Design Synthesis. Cambridge
University Press, New York
Altshuller G (1996) And Suddenly the Inventor Appeared. Technical Innovation
Center, Massachusetts
Arora JS (1988) Introduction to Optimum Design. McGraw–Hill, New York
Arora JS (1997) Guide to Structural Optimization. The American Society of Civil
Engineers, New York
Braun R (1996) Collaborative Optimization: An Architecture for Large-Scale
Distributed Design. Ph.D. Thesis, Stanford University, CA
Choi WS, Park GJ (2002) Structural Optimization Using Equivalent Static Loads
at All Time Intervals. Computer Methods in Applied Mechanics and
Engineering 191:2077í2094
Dieter GE (1999) Engineering Design: A Materials and Processing Approach. 3rd
ed., McGraw–Hill, New York
Haftka RT, Gürdal Z (1992) Elements of Structural Optimization. 3rd revised and
expanded ed., Kluwer Academic Publishers, Dordrecht
Haugen EB (1980) Probabilistic Mechanical Design. Wiley, New York
Hibbeler RC (2000) Mechanics of Materials. 4th ed., Prentice Hall, Englewood
Cliffs, NJ
Hurst K (1999) Engineering Design Principles. Wiley, New York

Hyman BI (1998) Fundamentals of Engineering Design. Prentice Hall, Englewood
Cliffs, NJ
Kang BS, Park GJ, Arora JS (2006) A Review of Optimization of Structures
Subjected to Transient Loads. Structural and Multidisciplinary Optimization
31(2):81í95
Montgomery DC (1997) Design and Analysis of Experiments. Wiley, New York
Norton RL (2000) Machine Design: An Integrated Approach. 2nd ed., Prentice
Hall, Englewood Cliffs, NJ
Park GJ, Kang BS (2002) Validation of Structural Optimization Algorithm
Transforming Dynamic Loads into Equivalent Static Loads. Journal of
Optimization Theory and Applications 119(1):191í200
Park GJ, Lee TH, Lee KW, Hwang KH (2006) Robust Design: An Overview.
AIAA Journal 44(1):181-191
Phadke MS (1989) Quality Engineering Using Robust Design. Prentice Hall,
Englewood Cliffs, NJ
Rao SS (1996) Engineering Optimization: Theory and Practice. 3rd ed., Wiley,
New York
Sobieszczanski-Sobieski J, Haftka RT (1997) Multidisciplinary Aerospace Design
Optimization: Survey of Recent Developments. Structural and Multidisciplinary
Optimization 14(1):1í23
Suh NP (1990) The Principles of Design. Oxford University Press, New York
Suh NP (2001) Axiomatic Design: Advances and Applications. Oxford University
Press, New York
Taguchi G (1987) System of Experimental Design. Vols. 1 and 2, Kraus, New
York
Taylor WA (1991) Optimization and Variation Reduction in Quality. McGraw–
Hill, New York
Ullman DG (1997) The Mechanical Design Process. 2nd ed., McGraw–Hill, New
York
2 Axiomatic Design
2.1 Introduction
Axiomatic design is a design theory that was created and popularized by Professor
Suh of the Massachusetts Institute of Technology (Suh 1990, 2000). Actually, it is
a general design framework, rather than a design theory. As the word
“framework” indicates, it can be applied to all design activities. It consists of two
axioms. One is the Independence Axiom and the other is the Information Axiom.
A good design should satisfy the two axioms while a bad design does not. It is
well known that the word “axiom” originates from geometry. An axiom cannot be
proved and becomes obsolete when a counterexample is validated. So far, a
counterexample has not been found in axiomatic design. Instead, many useful
design examples with axioms are validated.
Design is the interplay between “what we want to achieve” and “how we
achieve it.” A designer tries to obtain what he/she wants to achieve through
appropriate interplay between both sides. The engineering sequence can be
classified into four domains as illustrated in Figure 2.1. Customer attributes (CAs)
are delineated in the customer domain. In other words, CAs are the customer
needs. CAs are transformed into functional requirements (FRs) in the functional
domain. FRs are defined by engineering words. This is equivalent to “what we
want to achieve.” FRs are satisfied by defining or selecting design parameters
(DPs) in the physical domain. Mostly, this procedure is referred to as the design
process. Production variables (PVs) are determined from DPs in the same manner.
The aspects for the next domain are determined from the relationship between the
two domains, and this process is called mapping. A good design process means an
efficient mapping process.
Design axioms are defined from common principles for engineering activities
as follows:
Axiom 1: The Independence Axiom

Maintain the independence of FRs.
Alternate Statement 1: An optimal design always maintains the
independence of FRs.
Customer domain Functional domain Physical domain Process domain
What? How? How? How?
What? What?
Constraints Constraints
Customer needs Functional requirements Design parameters Process variables
Figure 2.1. Relationship of domains, mapping and design spaces
Alternate Statement 2: In an acceptable design, DPs and FRs are related in

such a way that a specific DP can be adjusted to satisfy its corresponding
FR without affecting other functional requirements.
Axiom 2: The Information Axiom

Minimize the information content of the design.
Alternate Statement: The best design is a functionally uncoupled design
that has minimum information content.
The axioms may look simple. However, they have significant meanings in
engineering. Details of the axioms will be explained later. Axiom 1 is an
expression that design engineers know consciously or subconsciously. When we
design a complex system, the axiom tells us that a DP should be defined to
independently satisfy its corresponding FR. In other words, the FRs of the
functional domain in Figure 2.1 should be independently satisfied by DPs of the
physical domain. Otherwise, the design is not suitable. When multiple designs are
found from Axiom 1, the best one can be chosen based on Axiom 2. That is, the
best design has minimum information content that is usually quantified by the
probability of success. It also corresponds to the engineering intuition that design
engineers usually have in mind. Axiom 2 is related to robust design and it will be
explained later. Although the axioms are expressed simply, real application can be
very difficult.
As explained earlier, axioms are defined in geometry. As in geometry,
theorems and corollaries are derived from axioms (see Appendix 2.A).
Axiomatic Design 19
2.2 The Independence Axiom
2.2.1 The Independence Axiom
The Independence Axiom indicates that the aspects in the proceeding domain
should be independently satisfied by the choices carried out in the next domain.
The domains are illustrated in Figure 2.1. The relationship of FR–DP is defined to
be independent. When plural FRs are defined, each DP should satisfy each
corresponding FR. The relationship can be expressed by a design matrix. Using
vector notations for FRs and DPs, the relationship is expressed as the following
design equation:
FR A DP (2.1)
Matrix A is called a design matrix. The characteristics of matrix A determine if

the Independence Axiom is satisfied. Suppose we have three FRs and DPs.
Matrix A is as follows:
ª FR1 º ª A11 A12 A13 º ª DP1 º

« FR » «A A22 A23 »» «« DP2 »» (2.2)
« 2» « 21
«¬ FR3 »¼ «¬ A31 A32 A33 »¼ «¬ DP3 »¼
FR–DP relationships according to matrix A are shown in Table 2.1. If the

design matrix is a diagonal matrix, it is an uncoupled design. Because each DP
can satisfy a corresponding FR, the uncoupled design perfectly satisfies the
Independence Axiom. When the design matrix is triangular as shown in the
second case of Table 2.1, the design is a decoupled design. A decoupled design
satisfies the Independence Axiom if the design sequence is correct. In the second
row of Table 2.1, DP1 is first determined for FR1 and fixed. FR2 is satisfied by
the choice of DP2 and the fixed DP1. DP3 is determined in the same manner with
the fixed DP1 and DP2 .
When a design matrix is neither diagonal nor triangular, the design becomes a
coupled design. In a coupled design, no sequences of DPs can satisfy the FRs
independently. Therefore, an uncoupled or a decoupled design satisfies the
Independence Axiom and a coupled design does not. If a design is coupled, an
uncoupled or decoupled design must be found through a new choice of DPs. For
the ith FR or DP, the subscript notation is used in this book. FRi is frequently
expressed by FRI. With design matrices, multiplication and addition are
permitted; however, other manipulations such as coordinate transformation are not
permitted.
It is noted that constraints (Cs) exist in the design. Constraints are generally
defined from design specifications and they must be satisfied. Constraints can be
Table 2.1. FR–DP relationship according to the design matrix
Design equation Design process
ª FR1 º ª A11 0 0 º ª DP1 º FR1 A11 u DP1

Uncoupled « FR » «0 A22 0 »» «« DP2 »» FR2 A22 u DP2
design « 2» «
«¬ FR3 »¼ «¬ 0 0 A33 »¼ «¬ DP3 »¼ FR3 A33 u DP3
FR1 A11 u DP1
ª FR1 º ª A11 0 0 º ª DP1 º
« FR » «A FR2 A21 u DP1 A22 u DP2
Decoupled
« 2» « 21 A22 0 »» «« DP2 »»
design FR3 A31 u DP1 A32 u DP2
«¬ FR3 »¼ «¬ A31 A32 A33 »¼ «¬ DP3 »¼
A33 u DP3
FR1 A11 u DP1 A12 u DP2

A13 u DP3
ª FR1 º ª A11 A12 A13 º ª DP1 º
« FR » «A FR2 A21 u DP1 A22 u DP2
Coupled
« 2» « 21 A22 A23 »» «« DP2 »»
design A23 u DP3
¬« FR3 ¼» ¬« A31 A32 A33 ¼» ¬« DP3 ¼»
FR3 A31 u DP1 A32 u DP2
A33 u DP3
defined without regard to independence of FRs and coupled by DPs. As illustrated

in Figure 2.1, the constraints can be defined in the DP or PV domains.
The following example shows an application of the Independence Axiom.
Generally, an imperative sentence is used for the expression of an FR and a noun is
used for a DP.
Example 2.1 [Design of a Refrigerator Door] (NSF 1998, Suh 2000)

Figure 2.2 shows two refrigerator doors that we most frequently encounter. Which
one has the better design? To answer the question, the doors are analyzed based
on an axiomatic design viewpoint. Functional requirements are defined as follows:
FR1 : Provide access to the items stored in the refrigerator.
FR2 : Minimize energy loss.
Solution
Design parameters for the vertically hung door in Figure 2.2a are as follows:
DP1 : Vertically hung door
DP2 : Thermal insulation material in the door
Axiomatic Design 21
(a) Vertically hung door (b) Horizontally hung door
Figure 2.2. Refrigerator doors
The design equation may be stated as
ª FR1 º ªX 0 º ª DP1 º
« FR » «X « » (2.3)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
where an X indicates a nonzero value, and hence a dependence between an FR and

a DP.
From Equation 2.3, the design is a decoupled one and satisfies the
Independence Axiom. However, when we open the door, energy loss occurs due
to the X in the off-diagonal term. Now, the horizontally hung door in Figure 2.2b
is analyzed.
DP1 : Horizontally hung door
DP2 : Thermal insulation material in the door
The design equation is made as follows:
« FR » «0 « » (2.4)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
When we open the horizontally hung door, cold air remains in the refrigerator
and energy loss can be minimized. Therefore, the horizontally hung door has an
uncoupled design and is a better design than the vertically hung door. Is the
horizontally hung door always better? As far as the functional requirements
defined here are kept, it is correct. Suppose that constraints are proposed for the
amount of stored food or convenience to access items. Then the problem will be
different. If a refrigerator with a horizontally hung door violates the constraints, it

cannot be accepted regardless of the satisfaction of the Independence Axiom.
When constraints exist, they should be checked first.
Example 2.2 [Design of a Water Faucet] (Suh 2000)

A faucet is designed. The user should be able to control the temperature and the
running rate of water. Since there are many commercialized faucets, they are
evaluated. The functional requirements of a faucet are defined as follows:
FR1 : Control the flow of water (Q).
FR2 : Control the temperature of water (T).
Solution
Analyzing the product in Figure 2.3a, DPs and the design equation are defined as
follows:
DP1 : Angle I1
DP2 : Angle I 2
Cold water Hot water

I1 I2
I2
(a) Coupled design
Y
I1
(c) Uncoupled design (b) Uncoupled design
Figure 2.3. Example of a water faucet

Axiomatic Design 23
ª FR1 (Q) º ªX X º ª DP1 (I1 ) º

« FR (T )» «X « » (2.5)
¬ 2 ¼ ¬ X »¼ ¬ DP2 (I 2 )¼
As shown in Equation 2.5, the design is coupled. Thus, the design is not
acceptable.
Another example is presented in Figure 2.3b. The design is analyzed as
follows:
DP1 : Angle I1
DP2 : Angle I 2
ª FR1 (Q) º ªX 0 º ª DP1 (I1 ) º

« FR (T )» «0 « » (2.6)
¬ 2 ¼ ¬ X »¼ ¬ DP2 (I 2 )¼
Because the design matrix is diagonal, the design is uncoupled. Therefore, it

satisfies the Independence Axiom and is acceptable.
One more design is illustrated in Figure 2.3c.
DP1 : Displacement Y
DP2 : Angle I
ª FR1 (Q) º ªX 0 º ª DP1 (Y ) º

« FR (T )» «0 « » (2.7)
¬ 2 ¼ ¬ X »¼ ¬ DP2 (I )¼
The design matrix is diagonal; therefore, the design is uncoupled. We have

two uncoupled designs. Which one is better? It is easy to manipulate the one in
Figure 2.3c. This can be explained by the Information Axiom, which will be
introduced later. The design in Figure 2.3c is the best from the viewpoint of the
Information Axiom. Actually, the one in Figure 2.3c is becoming popular. This
conclusion is made based on engineering functional requirements. If aesthetic
aspects are important, different decisions can be made.
When we design a complicated system, a definition of a simple FR–DP
relationship may not be sufficient. Then we can decompose the relationship. As
illustrated in Figure 2.4, a new relationship is defined by a zigzagging process
between the functional and physical domains. The zigzagging process is presented
by the numbers in Figure 2.4. It is noted that DPs are defined according to FRs in
the same level and FRs of the lower level are defined based on the characteristics
of DPs in the upper level. This decomposition process continues until the leaf
(bottom) level is reached. In Figure 2.5, the decomposition process for a lathe is
illustrated (Suh 1999).
FR ཛ DP
ཝ
FR1 FR2 ˎ DP1 DP2 ˎ
ཞ
FR11 FR12 ˎ FR21 FR22 ˎ DP11 DP12 ˎ DP21 DP22 ˎ
(a) Functional domain (b) Physical domain
Figure 2.4. Zigzagging process between domains
Metal removal device
Power Workpiece Speed Workpiece Support Tool

rotation changing support and
supply source device tool holder structure positioner
Tool holder Positioner Support structure
Longitudinal clamp Rotation stop Tool holder
(a) Functional domain
Lathe
Motor Head Gear Tail Bed Carriage

drive stock box stock
Spindle Feed Frame

assembly screw
Clamp Handle Belt Pin Tapered

bore
(b) Physical domain

Figure 2.5. Decomposition process for a lathe using axiomatic design
Axiomatic Design 25
2.2.2 Independence
Using FR–DP coordinates, Figure 2.6 presents diagrams of mapping processes

when the numbers of FRs and DPs are 2. Each design can be expressed by a
design equation as follows (Rinderle and Suh 1982):
ª FR1 º ª A11 0 º ª DP1 º

« FR » « 0 uncoupled design (Figure 2.6a) (2.8)
¬ 2¼ ¬ A22 »¼ «¬ DP2 »¼
ª FR1 º ª A11 A12 º ª DP1 º

« FR » « 0 decoupled design (Figure 2.6b) (2.9)
¬ 2¼ ¬ A22 »¼ «¬ DP2 »¼
ª FR1 º ª A11 A12 º ª DP1 º

« FR » «A coupled design (Figure 2.6c) (2.10)
¬ 2¼ ¬ 21 A22 »¼ «¬ DP2 »¼
An uncoupled design is presented in Figure 2.6a. The point A has ( FR1 ) A and
( FR2 ) A for ( DP1 ) A and ( DP2 ) A , respectively. The points B, C and D have the
same characteristics. If the design is to be changed from A to C, the path A–D–C
or the path A–B–C can be selected. That is, the uncoupled design is independent
of the design path.
The decoupled design in Figure 2.6b is different. Suppose we want to change
the design from A to C. First, DP2 should be changed from ( DP2 ) A to ( DP2 ) E .
In this process, DP1 is fixed and both FR1 and FR2 are changed. Second, DP2
is fixed and DP1 is changed from ( DP1 ) E to ( DP1 ) C . In this process, FR2 is
fixed and FR1 is changed. Thus, the decoupled design relies upon the design path.
That is, DP2 should be determined first and DP1 should be determined later.
Now, look at the coupled design in Figure 2.6c. When the design is changed
from A to C, the effect is the same no matter what design parameter is changed.
Suppose DP1 is changed first. To satisfy FR1 , DP1 can be changed from A to C' ,
and then FR2 is also changed. Thus the design should be changed from C' to C"
to satisfy FR2 . Then FR1 is changed again and DP1 should be changed again.
Therefore, the design process is repetitively performed until the design converges.
This can be quite a complicated process. In particular, convergence may be
impossible when the design is highly nonlinear.
Figure 2.7 briefly presents the above relationships. The characteristics of the
design equations can be expressed by D1 , D 2 and T in Figure 2.7. The ideal
uncoupled design is obtained when D1 D 2 0 and T 90 o. As an index for
coupling, the following index R called “reangularity” is defined:
R sin T (1 cos2 T )1 / 2 (2.11)

FR2
DP2 B C
( DP2 ) A, D A D
( FR2 ) A, D
( DP1 ) A, B DP1
( FR1 ) A , B ( FR1 ) A ( FR1 ) B FR1
(a) Uncoupled design
FR2 DP2
B E C F
( DP2 ) B ,C
( DP2 ) A, D A D
( FR 2 ) A, D
( DP1 ) A, E ( DP1 ) C DP1
( FR1 ) A ( FR1 ) B ( FR1 ) E ( FR1 ) C , D FR

1
(b) Decoupled design
DP2
FR2
( DP2 ) B , C B
( FR2 ) C C
C cc
( DP2 ) A, D A
Cc
( FR2 ) D D
( DP1 ) A, B DP1
( DP1 ) C , D
( FR1 ) A ( FR1 ) C ,C c FR1
(c) Coupled design
Figure 2.6. Mapping process from the FR domain to the DP domain for each design
Axiomatic Design 27
FR2 DP2
DP2 DP2
DP2
D2
T DP1
D1
(d) Coupled design ( D 1 z 0, D 2 z 0 )
DP1
T
T (c) Decoupled design ( D 1 z 0, D 2 0)
T (b) Decoupled design ( D 1 0, D 2 z 0 ) DP1
(a) Uncoupled design ( D 1 D2 0) DP1
FR1
Figure 2.7. Schematic view of each design according to the coupling characteristics
If the numbers of FRs and DPs are n and each element of the design equation is
Aij , R is as follows:
1/ 2
ª n º
« ( ¦ Aki Akj )2 »
k 1
R «1 n n » (2.12)
i 1, n 1 «
j 1 i , n ( ¦ Aki2 )( ¦ Akj2 ) »
«¬ k 1 k 1 »¼
When T 90 q , the DP1 -axis is orthogonal to the DP2 -axis and R = 1.

Reangularity R is not sufficient to show all the cases of coupling. The fact that
R o 1 does not guarantee that D1 o 0 and D 2 o 0 . D 1 D 2 0 means that the
design equation is diagonal and larger diagonal terms make coupling lower.
Therefore, another index called “semangularity” (this means the same angle
quality in Latin) S is defined as follows:
n | A jj |
S n
(2.13)
j 1
( ¦ Akj2 )1 / 2
k 1
Table 2.2. Reangularity and semangularity for each design
Uncoupled design Decoupled design Coupled design
Reangularity 1 R S 1 R z S 1
Angle between
90q T T
column vectors ( T )
Semangularity 1 S R 1 S z R 1
When the design equation is diagonal, D1 D 2 0 and S = 1. Table 2.2 shows

the characteristics of each design for reangularity and semangularity.
Example 2.3 [Reangularity and Semangularity of a Decoupled Design]

Prove that R and S of Equation 2.9 are the same for the decoupled design.
Solution 1
When there are two functional requirements, R and S are as follows using
Equations 2.12 and 2.13:
1/ 2
ª ( A11 A12 A21 A22 ) 2 º
R «1 2 2 2 2
» (2.14a)
«¬ ( A11 A21 )( A12 A22 ) »¼
ª ºª º
S « | A11 | » « | A22 | » (2.14b)
« A2 A2 »« A2 A2 »
¬ 11 21 ¼ ¬ 12 22 ¼
In Equation 2.9, A21 0 and Equation 2.14 becomes
1/ 2 1/ 2
ª 2 2 º ª 2 2 º
A11 A12 A11 A22 | A22 |
R «1 2 2 2 » « 2 2 2 »
(2.15a)
¬« A11 ( A12 A22 ) ¼» ¬« A11 ( A12 A22 ) ¼» 2
A12 2
A22
ª| A | ºª | A | º | A22 |
S « 11 » « 22 » (2.15b)
« A2 » « A2 A2 » 2
A12 2
A22
¬ 11 ¼ ¬ 12 22 ¼
Therefore, R and S are the same.

Axiomatic Design 29
FR2
DP2c2
FR
D2
T
DP1c1 FR1
Figure 2.8. Vector representation of Example 2.3
Solution 2
Solve the problem geometrically. Define c1 [ A11 0]T and c 2 [ A12 A22 ]T . If
the two functional requirements are expressed by a vector FR, then
FR DP1c1 DP2 c 2 . This is geometrically represented in Figure 2.8. From
Figure 2.8, R and S are as follows:
c 1T c 2 A11 A12
cos T (2.16a)
| c 1 || c 2 | 2
A11 2
A12 2
A22
1/ 2
ª 2 2 º
2 1/ 2 A11 A12
R sin T (1 cos T ) «1 2 2 2 »
¬« A11 ( A12 A22 ) ¼»
1/ 2
(2.16b)
ª 2 2 º
A11 A22 | A22 |
« 2 2 2 »
¬« A11 ( A12 A22 ) ¼» 2
A12 2
A22
ª ºª º
S cos D 1 cos D 2 « | A11 | » « | A22 | » | A22 |
(2.16c)
« A2 0 2 »« A2 A2 » 2
A12 2
A22
¬ 11 ¼ ¬ 12 22 ¼
Therefore, R and S are the same.
When the design equation is nonlinear with respect to design parameters, Aij
of Equation 2.2 may not be constant. Thus, although the uncoupled relationship is
satisfied at a design point, it may not be satisfied at other points. In this case, an
approximation by Taylor expansion can be employed. FRi in Equation 2.2 can be
approximated as follows:
3 wFRi
FRi ( FRi ) 0 GFRi ( FRi ) 0 ¦ GDPj (2.17)
j w
1 DPj
where ( FRi ) 0 is the current functional requirement. Using Equation 2.17, the
design equation at ( DP1 , DP2 , DP3 ) is defined as follows:
ª wFR1 wFR1 wFR1 º

ªGFR1 º « wDP1 wDP2 wDP3 » ªGDP1 º
«GFR » «wFR wFR2 wFR2 »«
« 2» « 2
wDP1 wDP2 wDP3 » «
GDP2 »» (2.18)
« » «GDP »
¬«GFR3 ¼» « wFR3 wFR3 wFR3 »¬ 3¼
¬ wDP1 wDP2 wDP3 ¼
As shown in Equation 2.18, the design matrix is a matrix with partial derivatives,
which defines the relationship between increments of FRs and DPs. The effort to
find an uncoupled design is to find a design window where the design matrix is
diagonal. Therefore, although the Independence Axiom is satisfied at a design
point, it is not guaranteed if the design is changed.
Can we consider a design to be uncoupled when the off-diagonal terms are
quite small compared to the diagonal terms? The following equation is an
example:
ª FR1 º ªX x x º ª DP1 º
« FR » «x X x »» «« DP2 »»
« 2» « (2.19)
«¬ FR3 »¼ «¬ x x X »¼ «¬ DP3 »¼
where X !! x . The decision can be made based on the tolerance ranges and the
sizes of X and x. Theorem 2.A.8 in Appendix 2.A provides the reason for this.
Theorem 2.A.8 is as follows:
Theorem 2.A.8 [Independence and Design Range]

A design is an uncoupled design when the designer-specified range is
greater than
n § wFR ·
¦ ¨¨ i ¸ 'DPj
¸
(2.20)
j z i © wDPj ¹
j 1
in which case the off-diagonal elements of the design matrix can be

neglected from the design consideration.
Axiomatic Design 31
When the magnitude of Equation 2.20 is very small, in other words, when x is
considerably small compared to X in Equation 2.19, Equation 2.19 can be regarded
as an uncoupled design.
Suppose that the current design is FR 0 A DP0 . The change of the design
parameters is 'DP . The change of the functional requirements is 'FR and it is
wFRi
obtained by replacing G with ' in Equation 2.18. ('FRi ) diag { 'DPi and
wDPi
('FRi ) diag is the change of FRi by the ith diagonal term with respect to the
change of DPi . Generally, the diagonal term is the largest. Therefore, ('FRi ) diag
has the largest impact on the FRi change. If we exclude ('FRi ) diag from 'FR ,
the remainder is the off-diagonal terms. When the influence from the off-diagonal
terms is very small, we do not need to consider them. This is expressed as
n § wFR ·
¦ ¨¨ i ¸'DPj d ('FRi ) allowable ,
¸
i 1,..., n (2.21)
j z i© wDPj ¹
j 1
where ('FRi ) allowable is the allowable tolerance specified by the designer.

Equation 2.21 means the range where the influence of the off-diagonal terms is
negligible.
Example 2.4 [The Range of DPs to Be Considered as a Decoupled

Design]
Suppose we have the following design:
ª FR1 º ªX x º ª DP1 º
« FR » «y « » (2.22)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
where X !! x, X !! y. The allowable tolerance in Equation 2.22 is

('FRi ) allowable , i 1, 2.
(1) Obtain the range of design parameters with which we can consider the
design as an uncoupled design.
(2) Obtain the range of design parameters with which we can consider the
design as a decoupled design.
Solution
(1) Ranges of 'DP1 and 'DP2 that satisfy x'DP2 d ('FR1 ) allowable and
y'DP1 d ('FR2 ) allowable .
(2) Ranges of 'DP1 and 'DP2 that satisfy x'DP2 d ('FR1 ) allowable or
y'DP1 d ('FR2 ) allowable .
2.2.3 Physical Integration
There is a saying that a simple design is a good one. From this statement, we may
guess that a good design makes one DP satisfy multiple FRs. In other words, a
coupled design is better. This aspect is very confusing in axiomatic design.
However, from an axiomatic design viewpoint, this is the case where multiple DPs
make a physical entity. That is, multiple DPs satisfy FRs of the same number.
This is called “physical integration.” Physical integration is desirable because the
information quantity can be reduced. The following example is a typical example
of physical integration.
Example 2.5 [Bottle–can Opener] (NSF 1998, Suh 1999)

Suppose we need a device that can open bottles and cans. Functional requirements
are defined as follows:
FR1 : Design a device that can open bottles.
FR2 : Design a device that can open cans.
Solution
The device in Figure 2.9 has one physical entity for the bottle opener and can
opener. However, two DPs at both ends independently satisfy the two functional
requirements. Therefore, the design in Figure 2.9 satisfies the Independence
Axiom. If the constraint set includes “both functions should be simultaneously
used,” then a different design should be investigated.
Figure 2.9. Bottle–can opener Figure 2.10. Beverage can

Axiomatic Design 33
Example 2.6 [Beverage Can Design] (NSF 1998, Suh 2001)

Consider an aluminum beverage can that contains liquid as illustrated in Figure
2.10. According to an expert working at one aluminum can manufacturer, there
are 12 FRs for the can. Plausible FRs: contain axial and radial pressure, withstand
moderate impact when the can is dropped from a certain height, allow stacking on
top of each other, provide easy access to the liquid in the can, minimize the use of
aluminum, be printable on the surface, and more. However, these 12 FRs are not
satisfied by 12 physical pieces. The can consists of three pieces: the body, the lid
and the tab opener. There must be at least 12 DPs and they are distributed to these
three pieces. Most of the DPs are associated with the geometry of the can: the
thickness of the body, the curvatures at the bottom, the reduced diameter at the top
to reduce the material used to make the top lid, the corrugated geometry of the tab
opener to increase the stiffness, the small extrusion on the lid to attach the tab, etc.
The complexity is reduced when physical integration is utilized while the
independence is maintained. That is, related information quantity is reduced.
Therefore, physical integration does not violate the Independence Axiom. Instead,
it is recommended.
2.3 The Information Axiom
2.3.1 The Calculation of Information Contents Using Probability
Axiomatic design requires satisfaction of the Independence Axiom. Multiple

designs that satisfy the Independence Axiom can be derived. In this case, the best
design should be selected. The best design is the one with minimum information.
How can we quantitatively define the information measure? The definition varies
according to the situation. Generally, the information is related to complexity.
Then how can we measure complexity? We need a rigorous definition for the
information content. The information content can be differently defined according
to the characteristics of the design. The probability of success has been utilized as
an index of the information content.
Suppose p is the probability of satisfying FRi with DPi . Then the
information content is defined as
Ii log 2 1 / p (2.23)
In Equation 2.23, the reciprocal of p is used to make the larger probability have
less information. Also, the logarithm function is utilized to enhance additivity.
The base of the logarithm is 2 to express the information content with the bit unit.
Suppose we have the following uncoupled design:
ª FR1 º ª A11 0 0 º ª DP1 º

« FR » « 0 A22 0 »» «« DP2 »» (2.24)
« 2» «
«¬ FR3 »¼ «¬ 0 0 A33 »¼ «¬ DP3 »¼
Suppose p1 , p2 and p3 are the probabilities of satisfying FR1 , FR2 and FR3
with DP1 , DP2 and DP3 , respectively. The total information I total is
3 3 § 1 ·
I total ¦ Ii ¦ log 2 ¨¨ ¸¸ (2.25)
i 1 i 1 © pi ¹
It is noted that the information content should only be defined based on the
corresponding functional requirement.
Example 2.7 [An Example of Calculating Information Content]

Information content is calculated for the design problem in Figure 2.9. It is
assumed that the probability of satisfying FR1 with DP1 is 0.9 and the one for
FR2 with DP2 is 0.85. The total information content is as follows:
§ 1 · § 1 ·
I total I1 I 2 log 2 ¨ ¸ log 2 ¨ ¸ 0.1520 0.2345 0.3865 (bits)
© 0. 9 ¹ © 0.85 ¹
(2.26)
Now the reduction of information due to physical integration is explained with

Example 2.5. Without physical integration, two pieces of the two DPs should be
made. If we keep the amount of material constant, the sizes of each piece should
be smaller. Then the use of each piece is inconvenient and the probability of
success is reduced. The result is that the information content is increased.
Therefore, it is inferred that a tool with physical integration has less information
content. However, not much research has been done on quantifying the reduction
of information content from physical integration. We need more research on this
topic.
Example 2.8 [Manufacture of a Bar with a Specified Tolerance]

Another method to calculate the probability of success is introduced. A bar of 1 m
length is to be manufactured. The cases for the tolerance are r 0.00001 m and
r 0.1 m . Calculate the information content for both cases.
Solution
If we use the same machine for both cases, the probability of success is smaller
when the tolerance is small. Also, if the given length (nominal length) is longer,
the ratio of the tolerance to the total length is smaller. Thus, the probability of
success is as follows:
Axiomatic Design 35
§ tolerance ·
p f ¨¨ ¸¸ (2.27)
© nominal length ¹
If we assume that Equation 2.27 is linear, then it becomes as follows:
tolerance
p c (2.28)
nominal length
where c is a constant.
Calculation of the information content for a decoupled design is somewhat

different. Since independence is satisfied by the sequence of the process, the
probability of success of the later process depends on that of the previous one.
Therefore, it is a conditional probability. Suppose we have the following
decoupled design:
« FR » «X « » (2.29)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
If p1 is the probability that DP1 satisfies FR1 , then the probability that
DP2 satisfies FR2 under satisfaction of FR1 by DP1 is a conditional probability.
Suppose it is p 21 . Then the probability of success p that both FR1 and FR2 are
satisfied is
p p1 p 21 (2.30)
The total information content for p is
I log 2 p log 2 ( p1 p 21 ) log 2 p1 log 2 p 21 I1 I 2 (2.31)
The conditional probability is useful for investigating the characteristics of the

Information Axiom. However, it is rarely applied to real problems because p 21 is
not easy to evaluate. Instead, the probability density function is more practical for
application.
2.3.2 Probability Density Function and Information Content
Information content can be calculated by using the probability density function.

Figure 2.11 presents a schematic view of this. The terminologies are as follows:
the design range is the range for the design target, the system range is the operating
range of the designed product and the common range is the common area between
Target
Probability Bias
density Probability density
function function of the system
Design range
Common
range
FR
Variation from the peak value
Figure 2.11. Calculation of the information content using the probability density function
the design range and the system range. The design range is defined by lower and
upper bounds and the system range is defined by a distribution function of the
system performance. A uniform distribution of a system range is illustrated in
Figure 2.12. The design should be directed to increase the common range. The
information content is defined as follows:
ps Acr / Asr (2.32a)
I log 2 ( Acr / Asr ) (2.32b)
where Asr is the system range and Acr is the common range.
Example 2.9 [Calculation of the Information Content Using a

Probability Density Function]
A problem is made to demonstrate an example. A person defines two functional
requirements to buy a house as follows:
FR1 : Let the price range be from 50,000 dollars to 80,000 dollars.
FR2 : Let the commuting time be within 40 minutes.
The person considers a house in city A or city B. Table 2.3 shows the conditions
of both cities. Where should the person buy a house to minimize the information
content?
Solution
The system range is defined from Table 2.3 and the design range is determined
from the functional requirements. It is assumed that all the probability densities
are uniform. Figure 2.12 presents the probability density for the price of the house
in city A. Other items can be illustrated in the same manner. The information
content for city A is as follows:
Axiomatic Design 37
Table 2.3. Conditions for each city
city A city B
Price $45,000–$60,000 $70,000–$90,000
Commuting time 35–50 min 20–30 min
§ 1.5 · § 15 ·
I A1 log 2 ¨ ¸ 0.59, I A2 log 2 ¨ ¸ 1.59 (2.33)
© 1 ¹ ©5¹
IA I A1 I A2 2.18 (bits) (2.34)
In the same manner, the information content for city B is
§2· § 10 ·
I B1 log 2 ¨ ¸ 1.0 , I B 2 log 2 ¨ ¸ 0.0 (2.35)
©1¹ © 10 ¹
IB I B1 I B 2 1.0 (bits) (2.36)
The information content I A for city A is 2.18 and that for city B I B is 1.0.
Therefore, city B has the optimum house from an axiomatic design viewpoint.
The design should be directed to reduce the information content in Equation

2.25. From Figure 2.11, it is effective to reduce the bias that is the difference
between the averages of the system range and design range. After that, the
standard deviation of the system range should be decreased. Then the common
range is increased and the information content is reduced. This aspect is related to
robust design.
Probability
density Probability density Design range
function function
of the system
Common range
1 3 5 7 9 FR
Figure 2.12. Probability density function of a uniform distribution
2.3.3 The Calculation of Information Content for a Decoupled

Design
The information content for an uncoupled design is relatively easy to calculate by
using Equation 2.25. Generally, the information content is not calculated for a
coupled design because it violates the Independence Axiom. As mentioned earlier,
the information content for a decoupled design is obtained by using the conditional
probability. However, when the system range is given by the probability density
function, it is not easy to use. Therefore, specific methods have been developed.
There are two methods according to the distribution and the tolerance: the
graphical method and the integration method. When the probability density
function does not have uniform distribution or there are more than two functional
requirements, the graphical method cannot be used. On the other hand, the
integration method can be used in many cases, but it is difficult to use because
multiple integrals should be solved.
Suppose we have the following decoupled design:
« FR » «X « » (2.37)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
The random variation of a functional requirement ( FRi ) with respect to the

random variation of a design parameter ( DPi ) is as follows:
ª wFR1 wFR1 º
ª GFR1 º « wDP wDP2 » ªGDP1 º ª A11 0 º ªGDP1 º
«GFR » « 1 » (2.38)
¬ 2¼ « wFR2 wFR2 » «¬GDP2 »¼ «A
¬ 21 A22 »¼ «¬GDP2 »¼
«¬ wDP1 wDP2 »¼
The random variation of design parameters is GDP and n is the number of

design parameters. Suppose the tolerance ranges are 'DPi d GDPi d 'DPi ,
i 1,..., n ( 0 d 'DPi ). If the target value of the functional requirements is FR * ,
the success means that GFR resides within the range specified by the designer. In
other words, 'FRi d GFRi d 'FRi , i 1,..., n ( 0 d 'FRi ) is satisfied. Suppose
FR * is satisfied by DP* . If we treat the random variation as random variables,
the probability of success ( p s ) of the decoupled design in Equation 2.38 is as
follows:
p s { p( 'FR1 d GFR1 d 'FR1 ) p ('FR2 d GFR2 d 'FR2 | 'FR1 d GFR1 d 'FR1 )
(2.39)
Axiomatic Design 39
Let us assume that Aij of Equation 2.38 is a positive constant and GDPi is
statistically independent. GFR2 is a statistically dependent random variable with
respect to GFR1. In the DP domain, the condition in Equation 2.39 can be
expressed as
'FR1 d A11GDP1 d 'FR1 (2.40a)
'FR2 d A21GDP1 A22GDP2 d 'FR2 (2.40b)
'DP1 d GDP1 d 'DP1 (2.40c)
'DP2 d GDP2 d 'DP2 (2.40d)
Equation 2.40 can be mapped into the FR domain as follows:
'FR1 d GFR1 d 'FR1 (2.41a)
'FR2 d GFR2 d 'FR2 (2.41b)
A11 'DP1 d GFR1 d A11 'DP1 (2.41c)
A21
A22 'DP2 d GFR 2 GFR1 d A22 'DP2 (2.41d)
A11
If Aij is negative, Equation 2.41 can be different.

The range of GDP satisfying Equation 2.40 is the range satisfying Equation
2.39. In the same manner, the range of GFR satisfying Equation 2.41 satisfies
Equation 2.39. This is similar to the feasible region of the optimization theory.
That is, if we obtain the probability density function in the feasible region of
Equation 2.40 or 2.41, then the probability of Equation 2.39 is calculated.
In the graphical method, the area of the feasible region is calculated from
Equation 2.40 or 2.41. It is utilized when the probability density functions of the
FRs or DPs are uniform. Figure 2.13 represents the range satisfying Equation 2.40.
The probability of success and the information content are as follows:
ps Af / Adp (2.42a)
I log 2 p s (2.42b)
where Af is the feasible region, which is the shadowed area in Figure 2.13 and
Adp is the tolerance for design parameters, which is 4'DP1 'DP2 in Figure 2.13.
'FR2 A21GDP1 § 'FR2 ·

GDP2 ¨¨ 0, ¸¸
A22 GDP2 © A22 ¹
('DP1 , 'DP2 ) 'FR1

GDP1
A11
( DP1* , DP2* ) GDP1
'FR1
GDP1
A11 ('DP1 , 'DP2 )
'FR2 A21GDP1
GDP2
§ 'FR2 · A22
¨¨ 0, ¸¸
© A22 ¹
(a) When the probability of success = 1
'FR2 A21GDP1 GDP2

GDP2
A22
('DP1 , 'DP2 )
§ 'FR2 ·
¨¨ 0, ¸¸
© A22 ¹ GDP1
( DP1* , DP2* )
'FR1
GDP1
A11
'FR1
GDP1
A11 § 'FR2 ·
¨¨ 0, ¸¸
© A22 ¹
'FR2 A21GDP1
GDP2
A22
(b) When the probability of success < 1
Figure 2.13. The probability of success of the decoupled design in the DP range
Axiomatic Design 41
The graphical method in the FR domain is illustrated in Figure 2.14. In the

functional domain, the system range, the design range and the common range are
defined. In the same manner as Equation 2.32, the probability of success and the
information content are defined as follows:
ps Acr / Asr (2.43a)
I log 2 p s (2.43b)
where Asr is the area of the system range, which is the area of the parallelogram in
Figure 2.14 and the design range is the shadowed area of Figure 2.14. The
common range Acr is the common area of the system range and the design range.
This is the same as the feasible region in Equation 2.41. It is noted that the
probability of success for Figure 2.13a is 1, but that for Figure 2.14a is not 1. It is
somewhat complicated to calculate the shadowed area in Figure 2.13b or the
common area of Figure 2.14b. The probability of success for Figure 2.14b is as
follows:
2
A11 § A21 ·
¨ 'FR1 A22 'DP2 'FR2 ¸¸
'FR1 A21 ¨© A11 ¹
ps (2.44)
A11'DP1 4 A11 'DP1 A22 'DP2
The probability of success can also be calculated by multiple integration. It is

conducted in the DP domain. As mentioned earlier, the probability of success is
evaluated for the feasible region, which is the shadowed area in Figure 2.13.
Suppose pGDP1 and pGDP2 are the distribution functions of GDP1 and GDP2 ,
respectively. Then the probability density function in the feasible region :
(probability of success) is
³³ pGDP1 pGDP2 dGDP2 dGDP1 (2.45)

:
When the feasible region is such as the one in Figure 2.13a, the integration is
easy. However, if it is such as the one in Figure 2.13b, the integration is somewhat
more difficult. In that case, we employ the unit step function u (x) as follows:
u ( x x * ) 1 : when x t x *
0 : when x x * (2.46)
Figure 2.15 represents the unit step function. Using the unit step function, the
probability distribution can be defined not only in the feasible region but also in
the entire region as follows:
GFR2
A21 ( A11'DP1 , A21'DP1 A22 'DP2 )
GFR2 GFR1 A22 'DP2
A11
(0, A22 'DP2 )
('FR1 , 'FR2 )
( FR1 , FR2 ) GFR1
(a) The design range resides within the system range
GFR2
(0, A22 'DP2 )
A21 ( A11'DP1 , A21'DP1 A22 'DP2 )
GFR2 GFR1 A22 'DP2
A11
('FR1 , 'FR2 )
( FR1 , FR2 ) GFR1
Common range
System range
(b) The design range crosses the system range
Figure 2.14. The probability of success of a decoupled design in the FR range

Axiomatic Design 43
u( x x * )
x* x
Figure 2.15. Unit step function
p1 pGDP1 >u (GDP1 ('DP1 )) u (GDP1 'DP1 )@ (2.47a)
p2 pGDP2 >u (GDP2 ('DP2 )) u (GDP2 'DP2 )@ (2.47b)
To integrate in the feasible region, the parallelograms in Figure 2.13 are used as
the integration interval. The interval is
§ 'FR1 'FR1 'FR2 A21GDP1 'FR2 A21GDP1 ·

¨ d GDP1 d , d GDP2 d ¸
¨ A11 A11 A22 A22 ¸
© ¹
(2.48)
Since p1 and p 2 are statistically independent, the probability of success is
'FR1 'FR2 A21GDP1

A11 A22
³ ³ p1 p2 dGDP2 dGDP1 (2.49)
'FR1 'FR2 A GDP
21 1
A11 A22
In some cases, we may not satisfy the target value FR * exactly with the design
parameters. In this case, the following equations hold:
FR c ADPc (2.50a)
FR c z FR * (2.50b)
where DPc [ DPc1 , DPc 2 ]T and FR c is the functional requirement vector made by
the current design parameters. The probability of success by the graphical method
is evaluated by transition of the rectangulars and parallelograms in Figures 2.13
and 2.14, so that DPc [ DPc1 , DPc 2 ]T becomes the origin. Then Equation 2.40
yields
FR1* 'FR1 d A11 ( DPc1 GDP1 ) d FR1* 'FR1 (2.51a)
FR2* 'FR2 d A21 ( DPc1 GDP1 ) A22 ( DPc 2 GDP2 ) d FR2* 'FR2
(2.51b)
'DP1 d GDP1 d 'DP1 (2.51c)
'DP2 d GDP2 d 'DP2 (2.51d)
The probability distributions p1 and p 2 in Equation 2.47 can be directly used.

Using Equations 2.47 and 2.51, the probability of success is calculated as follows:
FR1* A11DPc1 'FR1 FR2* A21DPc1 A22 DPc 2 'FR2 A21GDP1

A11 A22
³ ³ p1 p2 dGDP2 dGDP1 (2.52)
FR1* A11DPc1 'FR1 FR2* A21DPc1 A22 DPc 2 'FR2 A21GDP1
A11 A22
The advantage of the integration method is that the probability of success can
be calculated for many design parameters. If the number of design parameters is n,
the following multiple integration is utilized:
n n 1
FR n* ¦ Ani DPci 'FRn ¦ Ani GDPi
FR1* A11 DPc1 'FR1 i 1 i 1
A11 Ann
³ ³ p1... pn dGDPn dGDP1 (2.53)
n 1
FR1* A11 DPc1 'FR1 n
FR n* ¦ Ani DPci 'FR n ¦ Ani GDPi
A11 i 1 i 1
Ann
The integration method can be defined by the distribution function of FR for the
feasible region in Equation 2.41. Calculation of the information content in the FR
domain is more complicated than calculation in the DP domain, because the
distribution of DP is usually given.
The above methods can be applied to designs with FR–DP hierarchy of many
levels. When we have multilevel hierarchy, we can make an entire design matrix
for the FRs and DPs in the lowest level. We can apply the above methods to the
entire design matrix. The information content can be evaluated for a coupled
design. The method is defined by modification of the above methods. However, it
Axiomatic Design 45
is very complex and the coupled design is not considered in general design.
Therefore, the information content for the coupled design is not explained here.
Example 2.10 [Calculation of Information Content for a Decoupled

Design–1]
We have the following decoupled design:
ª FR1 º ª3 0º ª DP1 º
« FR » «2 5» « DP » (2.54)
¬ 2¼ ¬ ¼¬ 2 ¼
ª DP* º ª1º ª FR1* º ª3º

where « 1* » «1» , « * » «7 » and the tolerances for design parameters are
¬« DP2 ¼» ¬ ¼ ¬« FR2 »¼ ¬ ¼
'DP1 0.3 and 'DP2 0.3, and the allowable tolerance is 'FR 1.5. GDPi has
uniform distribution in the tolerance range.
(1) Calculate the information content in the DP domain by using the graphical
method.
(2) Calculate the information content in the FR domain by using the graphical
method.
(3) Calculate the information content in the DP domain by using the
integration method.
Solution
(1) From Equation 2.40, the feasible region in the DP domain is as follows:
1.5 d 3GDP1 d 1.5 (2.55a)
1.5 d 2GDP1 5GDP2 d 1.5 (2.55b)
0.3 d GDP1 d 0.3 (2.55c)
0.3 d GDP2 d 0.3 (2.55d)
Equation 2.55 is illustrated in Figure 2.16. From Equation 2.42, the

probability of success and the information content are
Af 0.6 u 0.6 0.3 u 0.12 u 0.5 u 2

ps 0.9 (2.56a)
Adp 0.6 u 0.6
1
I log 2 0.152 (bits) (2.56b)
ps
(2) From Equation 2.41, the feasible region in the FR domain is as follows:
GDP2
2
GDP2 GDP1 0.3
5
( 0 , 0.3 )
(0.3,0.18)
( 0 .3 , 0 ) ( 0 .3 , 0 ) GDP1
( 0.3 , 0.18 )
( 0 , 0 .3 )
2
GDP2 GDP1 0.3
5
Figure 2.16. Graphical presentation of Example 2.10(1) in the DP range
1.5 d GFR1 d 1.5 (2.57a)
1.5 d GFR2 d 1.5 (2.57b)
3 u 0.3 d GFR1 d 3 u 0.3 (2.57c)
2
5 u 0.3 d GFR2 GFR1 d 5 u 0.3 (2.57d)
3
Equation 2.57 is illustrated in Figure 2.17. From Equation 2.43, the

probability of success and the information content are
Acr 1.8 u 3 0.9 u 0.6 u 0.5 u 2

ps 0.9 (2.58a)
Asr 1.8 u 3
1
I log 2 0.152 (bits) (2.58b)
ps
(3) From Equation 2.47,
1 1
pGDP1 pGDP2 (2.59a)
2'DP1 0.6
Axiomatic Design 47
2 GFR2
GFR2 GFR1 1.5
3 (0, 1.5)
GFR1 0.9
(í1.5, 0) (í0.9, 0) (0.9, 0) (1.5, 0) GFR

1
GFR1 0.9
(0, í1.5) 2
GFR2 GFR1 1.5
3
Figure 2.17. Graphical presentation of Example 2.10(2) in the FR range
pi 1.667 u (u (GDPi (0.3)) u (GDPi 0.3)), i 1, 2 (2.59b)
From Equation 2.49, the probability of success and the information content
are
1.5 1.5 2GDP1

3 5
ps ³ ³ p1 p2 dGDP2 dGDP1 0.9 (2.60a)
1.5 1.5 2GDP1

3 5
1
I log 2 0.152 (bits) (2.60b)
ps

Design–2]
ª DP º ª0.9º
When « c1 » « » in Equation 2.54, the allowable tolerance for the functional
¬ DPc 2 ¼ ¬1.1¼
ª FR1* º ª3º
requirement is FRi* 'FRi d FRi d FRi* 'FRi . « * » « » and 'FRi 1.5,
¬« FR2 ¼» ¬7 ¼
i 1,2. The tolerance and distribution are the same as in Example 2.10.
(1) Calculate the information content in the DP range by using the graphical
method.
(2) Calculate the information content in the FR range by using the graphical
method.
(3) Calculate the information content in the DP domain by using the
integration method.
Solution
(1) Equation 2.40 is modified to
FR1* 'FR1 d A11 ( DPc1 GDP1 ) d FR1* 'FR1 (2.61a)
FR2* 'FR2 d A21 ( DPc1 GDP1 ) A22 ( DPc 2 GDP2 ) d FR2* 'FR2
(2.61b)
'DP1 d GDP1 d 'DP1 (2.61c)
'DP2 d GDP2 d 'DP2 (2.61d)
Equation 2.61 becomes
3 1.5 d 3(0.9 GDP1 ) d 3 1.5 (2.62a)
7 1.5 d 2(0.9 GDP1 ) 5(1.1 GDP2 ) d 7 1.5 (2.62b)
0.3 d GDP1 d 0.3 (2.62c)
0.3 d GDP2 d 0.3 (2.62d)
Equation 2.62 is illustrated in Figure 2.18. The probability of success and

the information content are
0.6 u 0.6 (0.18 u 0.45 u 0.5 0.15 u 0.06 u 0.5)

ps 0.875 (2.63a)
0.6 u 0.6
1
I log 2 0.193 (bits) (2.63b)
ps
(2) Equation 2.41 is modified to
ª FRc1 º ª3 0º ª DPc1 º ª2.7º

« FR » «2 5» « DP » « 7.3» (2.64a)
¬ c2 ¼ ¬ ¼ ¬ c2 ¼ ¬ ¼
FR1* 'FR1 d FRc1 GFR1 d FR1* 'FR1 (2.64b)

Axiomatic Design 49
GDP2
1.2 2GDP1 5GDP2
(0.15,0.3)
(0.3,0.3) (0.3,0.3)
(0.3,0.12)
GDP1
(0.3,0.24)
(0.3,0.3) (0.3,0.3)
(0.15,0.3)
1.8 2GDP1 5GDP2
Figure 2.18. Graphical presentation of Example 2.11(1) in the DP range
FR2* 'FR2 d FRc 2 GFR2 d FR2* 'FR2 (2.64c)
A11 'DP1 d GFR1 d A11 'DP1 (2.64d)
A21
A22 'DP2 d GFR2 GFR1 d A22 'DP2 (2.64e)
A11
1.2 d GFR1 d 1.8 (2.65a)
1.8 d GFR2 d 1.2 (2.65b)
0.9 d GFR1 d 0.9 (2.65c)
2
1.5 d GFR2 GFR1 d 1.5 (2.65d)
3
Equation 2.65 is illustrated in Figure 2.19. The probability of success and

the information content are
GFR2
(0, 1.5)
(0.9, 2.1)
(í0.45, 1.2) (0, 1.2)

(í0.9, 0.9)
(í0.9, 0.7)
(í2.25, 0) (1.8, 0)
(T1.2, 0) (í0.9, 0) (0.9, 0) GFR1
(1.8, 0.3)
(2.25, 0)
(0.9,Gí0.9)
(0,Gí1.5)
(í0.9,Gí2.1) (0,Gí1.8)
(í0.45,Gí1.8)
Figure 2.19. Graphical presentation of Example 2.10(2) in the FR range
Acr 1.8 u 3 (0.45 u 0.3 u 0.5 1.35 u 0.9 u 0.5)

ps 0.875 (2.66a)
Asr 1.8 u 3
1
I log 2 0.193 (bits) (2.66b)
ps
(3) The integration range is within the parallelogram in Figure 2.18. Therefore,
the information content is calculated as
1.2 2GDP1
0.6 5
ps ³ ³ p1 p2 dGDP2 dGDP1 0.875 (2.67)
0.4 1.8 2GDP1
5
where p1 and p 2 are the same as those in Equation 2.59.

Design–3]
The distribution function pGDPi for Equation 2.54 is as follows:
3 3
pGDPi (GDPi ) 2 , i 1, 2 (2.68)
4('DPi ) 3 4('DPi )
Axiomatic Design 51
(1) In Example 2.10, replace the distribution function with Equation 2.68. The
tolerances for the design parameters are the same. Calculate the
probability of success by the integration method.
(2) In Example 2.11, replace the distribution function with Equation 2.68. The
tolerances for the design parameters are the same. Calculate the
probability of success by the integration method.
Solution
(1) When the distribution has Equation 2.68 in Example 2.10, the probability
distribution pi for the ith design parameter is
ª 3 3 º
pi « (GDPi ) 2 u [u (GDPi ( 'DPi )) u (GDPi ('DPi ))]
¬ 4 u 0.3 3 4 u 0.3 »¼
(2.69)
By substituting Equation 2.69 into Equation 2.59, the probability of

success and the information content are calculated by the following
multiple integrations:
1.5 1.5 2GDP1

3 5
ps ³ ³ p1 p2 dGDP2 dGDP1 0.978 (2.70a)
1.5 1.5 2GDP1

3 5
1
I log 2 0.032 (bits) (2.70b)
ps
(2) When the distribution is as in Equation 2.68 in Example 2.11, Equations

2.69 and 2.70 are used directly. The probability of success and the
information content are as follows:
1.2 2GDP1
0.6 5
ps ³ ³ p1 p2 dGDP2 dGDP1 0.953 (2.71a)
0.4 1.8 2GDP1
5
I log 2 1 / p s 0.07 (bits) (2.71b)
The example shows various cases for calculating the information content
using the probability density function. A practical example is introduced
in Appendix 2.B.
2.4 The Application of Axiomatic Design

In most cases, the information content is reduced if the Independence Axiom is
satisfied. Therefore, it seems that the Information Axiom is dependent on the
Independence Axiom and the Information Axiom is not required. However, some
particular cases can exist. Suppose we find an uncoupled design and a coupled
design that satisfy the given functional requirements. In some cases, the
information content of the coupled design may be smaller than that of the
uncoupled design. Then the question arises can the coupled design be better than
the uncoupled one? The answer is “no.” Actually, this indicates that there should
be an uncoupled or a decoupled design that has less information content than the
coupled design. Therefore, the designer should make an effort to find an
uncoupled or decoupled design. The designer may find multiple uncoupled or
decoupled designs. If they are the same in satisfying the Independence Axiom,
one should select the one with the minimum information content. The flow chart
to apply the two axioms is illustrated in Figure 2.20. Appendix 2.B demonstrates a
typical example for the flow of Figure 2.20.
We will investigate how axiomatic design is applied to a practical design.
Generally, it is applied to the following areas:
(1) Creative design.
(2) Analysis of existing designs.
(3) Design improvement.
Suppose that new functional requirements are defined and that there is no
product that satisfies the functional requirements. The designer will try to find a
new design. In this case, a creative designer generally creates a new idea for a new
product. Axiomatic design can be exploited to materialize the design idea. The
idea is analyzed and selection and allocation of parts are determined by the
axiomatic approach. However, the creation of an entirely new idea is very difficult
and rare in machine design. Therefore, considerable improvement of an existing
design is regarded as a creative design.
Generally, a survey of public opinion can be conducted to evaluate an existing
product. However, axiomatic design can be utilized for evaluation from the
viewpoint of designers. In particular, different products for the same goal can be
evaluated. The goals of the product are the functional requirements. We can
select a better product that satisfies the Independence Axiom. If multiple products
satisfy the Independence Axiom in the same manner, we can select the best one
from the Information Axiom.
Finally, axiomatic design can be used to improve the current design. When the
current design is not sufficiently good or an improved design is needed, the
Independence Axiom is used first. The FRs and DPs are defined and satisfaction
of the Independence Axiom is checked with them. If the Independence Axiom is
not satisfied, an improved design should be made to satisfy the Independence
Axiom. When the Independence Axiom is satisfied, the DPs are defined to
minimize the information content.
Axiomatic Design 53
Analysis of design
Find designs that satisfy the

Independence Axiom
No
Is the no. of designs sufficient?
Yes
Yes
Multiple designs? Find the best design with
the Information Axiom
No
Determine the final design
Figure 2.20. Flow chart of the application of axiomatic design
Example 2.13 [An Example of a Creative Design: Refrigerator Design]

(Lee et al. 1994)
The example of the design of a refrigerator is introduced. In a general refrigerator,
food is frozen for long-term preservation and is maintained at a cold temperature
for short-term preservation. The following two functional requirements are
defined:
FR1 : Freeze food for long-term preservation.
FR2 : Maintain food at a cold temperature for short-term preservation.
To satisfy the two FRs, a refrigerator with two compartments can be designed.
The design parameters are as follows:
DP1 : The freezer section
DP2 : The chiller section
The design matrix in the first level is diagonal; therefore, it is an uncoupled

design. FR1 can be decomposed by the selection of DP1 .
FR11 : Maintain the temperature of the freezer section in the range of 18q C
r 2q C.
FR12 : Maintain a uniform temperature in the freezer section.
FR13 : Control the relative humidity to 50% in the freezer section.
In the same manner, FR2 can be decomposed with respect to DP2 .
FR21 : Maintain the temperature of the chiller section in the range of 2q C

3q C.
FR22 : Maintain a uniform temperature in the chiller section within r 0.5q C of
the preset temperature.
The design parameters for the second level are to be determined. The DPs
must be determined to satisfy the independence of the FRs. It is noted that DPs in
the lower level should be determined so as not to violate the independence of the
upper level.
The FRs of the freezer section can be satisfied by (1) a device pumping chilled
air into the freezer section, (2) a device for circulation of air for a uniform
temperature, (3) a monitoring device to independently control the temperature and
humidity. Therefore, the DPs in the second level are defined as follows:
DP11 : Sensor/compressor system that activates the compressor when the
temperature of the freezer section is different from the preset one
DP12 : Air circulation system that blows the air into the freezer and circulates it
uniformly
DP13 : Condenser that condenses the moisture in the returned air when the dew
point is exceeded
The design is a decoupled one as follows:
ª FR12 º ªX 0 0 º ª DP12 º
« FR » «X X 0 »» «« DP11 »» (2.72)
« 11 » «
¬« FR13 ¼» ¬« X 0 X ¼» ¬« DP13 ¼»
For food storage in the chiller section, the temperature should be maintained in
the range of 2q C 3q C . The chiller section also activates the compressor and
circulates the air. Design parameters for the chiller section are
DP21 : Sensor/compressor system that activates the compressor when the
temperature of the chiller section is different from the preset one
DP22 : Air circulation system that blows the air into the chiller section and
circulates it uniformly
The design equation is a decoupled one as follows:
Axiomatic Design 55
« FR » «X « » (2.73)
¬ 21 ¼ ¬ X »¼ ¬ DP21 ¼
The entire design equation decomposed up to the second level is a decoupled

one as follows:
ª FR12 º ªX 0 0 0 0 º ª DP12 º
« FR » «X « »
« 11 » « X 0 0 0 »» « DP11 »
« FR13 » «X 0 X 0 0 » « DP13 » (2.74)
« » « »« »
« FR22 » «0 0 0 X 0 » « DP22 »
« FR21 » «¬ 0 0 0 X X »¼ «¬ DP21 »¼
¬ ¼
It is noted that the FRs of the lower level still keep the independence of the upper
level in Equation 2.74.
From the design equation in Equation 2.74, one compressor and two fans can
satisfy the FRs. DP11 and DP21 are sensor/compressor systems so that the
compressor is activated by the sensors. However, the fans of DP12 and DP22 will
not be activated unless the temperature is out of the range of the preset one.
Therefore, the design with one compressor and two fans satisfies the Independence
Axiom. An example is illustrated in Figure 2.21. Other designs can be proposed.
If multiple designs are proposed, we can select one that satisfies the Independence
Axiom and controls the temperature and humidity in a wide range. The new
design and the conventional refrigerator are compared.
New cooling system refrigerator
F-fan Freezing
Capillary
tube room
Cold
air
Condenser Evaporator
R-fan Refrigerating
room
Compressor
Two cooling fan types
Figure 2.21. A new design of a refrigerator that satisfies the Independence Axiom
Conventional refrigerator
Capillary Fan Freezing

tube room
Cold
air
Evaporator
Condenser
Damper
Refrigerating
room
Compressor
One cooling fan type
Figure 2.22. Conventional refrigerator
The conventional refrigerator consists of one compressor and one fan. As

illustrated in Figure 2.22, a damper is utilized to cool the refrigerating room.
Therefore, the temperature of the refrigerator is not independently controlled.
When the temperature exceeds 3q C , the damper is opened. However FR21 is not
satisfied unless the compressor and the fan of the freezer section are activated.
According to Corollary 2.3 of Appendix 2.A, if we can satisfy the FRs with
one fan, the design in Figure 2.21 may not be the best. If we can find another
design that satisfies the Independence Axiom, we have to apply the Information
Axiom to select the best one.
Example 2.14 [An Example of Analysis of Existing Designs: Liquid

Crystal Display Holder] (NSF 1998, Suh 2000)
The liquid crystal display (LCD) is a projection display system. Three LCD panels
project the red, green and blue images of a TV signal. The configuration of an
LCD projector is illustrated in Figure 2.23. To display an exact color image by an
LCD projection system, the three panels should be aligned with respect to the blue
image within a tolerance value.
To align the pixels, the projector uses a device that can control the rotation and
translation of the LCD panels. The pixels of one of the three panels are set as a
reference, and the remaining two panels are properly aligned. Each LCD panel is
attached to an adjusting mechanism, which is called an “LCD holder.” For
alignment of the pixels, at least two LCD holders should have three degrees of
Axiomatic Design 57
Figure 2.23. Schematic view of an LCD projector
freedom (translation along the X and Y axes and rotation with respect to the Z axis).
Two products manufactured by Sanyo and Sharp will be compared.
Based on the Independence Axiom, we will select the better one.
Solution
The FR and DP of the highest level are stated as follows:
FR : Align the pixels of the LCD panels.
DP : The LCD holder that can align the pixels of the LCD panels
To align the pixels of all the LCD panels, the functional requirement is
decomposed as follows:
FR1 : Translate along the X axis = T(X).
FR2 : Translate along the Y axis = T(Y).
FR3 : Rotate with respect to the Z axis = R(Z).
The LCD projector uses three panels, and one is used as a reference one.
Therefore, holders with three degrees of freedom are needed for the two panels.
Sanyo Holder
Figure 2.24 shows the Sanyo holder. The holder is composed of three mechanisms
and is attached to each panel of Figure 2.23. All of them are lead screw structures.
If a screw moves, the attached plane moves accordingly. plate A is fixed to the
side frame. The LCD panel is attached to plate C. The three lead screws are
design parameters.
DP1 : The lead screw for conjunction of plate B and screw 1
DP2 : The lead screw for conjunction of plate C and screw 2
Z X
Figure 2.24. The Sanyo LCD holder mechanism
If plate B is rotated, the LCD panel moves along the T(X) and R(Z) axes. Plate C
and the LCD panel move with plate B. If screw 1 is rotated, the LCD panel moves
in T(X) and R(Z). Therefore, DP1 affects FR1 and FR3 . DP2 is composed of
screw 2 and plate C. If screw 2 is rotated, plate C and the LCD panel move in the
Y direction. Since the rotation of screw 2 changes the position of the LCD panel in
the T(X) axis, DP2 only affects FR2 . DP3 has the same function as DP1 . The
rotation of screw 3 moves the LCD panel in the T(X) and R(Z) axes and DP3
affects FR1 and FR3 . The design equation is
ªT ( X )º ªX 0 X º ª DP1 º
« T (Y ) » «0 X 0 »» «« DP2 »» (2.75)
« » «
«¬ R( Z ) »¼ «¬ X 0 X »¼ «¬ DP3 »¼
Because the design matrix in Equation 2.75 is coupled, it violates the

Independence Axiom.
With this product, repeated adjustment should be conducted to align the pixels.
For example, when the angle should be changed, DP1 can be changed. However,
DP1 changes the position of the LCD panel in the X direction and an undesirable
error occurs. Because we do not have a DP that only affects FR1 , a repeated
process with trial and error is needed. If erratic behavior occurs in a part, a
difficult adjustment process occurs.
Axiomatic Design 59
Sharp Holder
The Sharp holder also has three mechanisms as illustrated in Figure 2.25. One is a
simple lead screw and the other two have a guideway and a guide boss. Plate A is
fixed to the side frame. The LCD panel is attached to plate C. The following three
DPs are defined:
DP2 : The lead screw for conjunction of plate C, plate D, boss 1, boss 2 and
screw 2
DP3 : The lead screw for conjunction of plate B, plate E, boss 2 and screw 3
If screw 1 is rotated, plate B moves in the X direction. Since the LCD panel is
attached to plate C, it moves with plate B in the X direction. Thus, DP1 only
affects FR1 . Rotating screw 2 moves plate D in the X direction and the wall of the
guideway in plate D pushes boss 1. As a result, plate C moves along the Y axis
because the vertical groove in plate B guides the movement of boss 1 in the Y
direction. Therefore, rotating screw 2 moves the LCD panel in the Y direction and
DP2 only affects FR2 .
Screw 3 moves plate E in the X direction and the guideway of plate E pushes
boss 2. Since boss 1 does not have directional constraints, plate C rotates with
Boss 2
Boss 1
Guideway
Figure 2.25. The Sharp LCD holder mechanism

respect to boss 2. Rotating screw 3 is projected into the X and Y directions.

Therefore, DP3 affects FR1 and FR3 . The design equation is in Equation 2.76.
ªT ( X )º ª A11 0 A13 º ª DP1 º

« T (Y ) » « 0 A22 0 »» «« DP2 »» (2.76)
« » «
«¬ R( Z ) »¼ «¬ 0 0 A33 »¼ «¬ DP3 »¼
The design matrix is triangular, so it is a decoupled design. If the adjustment

process proceeds as the design matrix indicates, the Independence Axiom is
satisfied. After the LCD panel is aligned by DP3 in the R(Z) axis, DP1 and DP2
should be adjusted.
In the above method, existing designs can be analyzed or compared by using
the Independence Axiom. In this case, it is easy because the Independence Axiom
is violated by one design. However, when both designs satisfy the Independence
Axiom, they can be compared by reangularity (R) and semangularity (S), or by the
information content.
Example 2.15 [An Example of Design Improvement: Parking Mode of

an Automatic Transmission] (NSF 1998, Suh 2000)
An automobile automatic transmission has a parking mode. The parking mode
locks the transmission mechanism when the vehicle is unattended. Thus, it
prevents the vehicle from moving on its own. When an automobile is parked on a
hill, drivers complain that unlocking is difficult. Also, excessive vibration can
occur during the unlocking process. The current design is illustrated in Figure
2.26. Analyze the current design and develop an improved design.
Solution
First, the current design should be analyzed. In Figure 2.26, the pawl is locked in
the sprocket by the shift-linkage and the vehicle is in parking mode. The sprocket
is attached to the automatic transmission. If the shift-linkage is changed to the
parking mode, the detent spring activates the hydraulic system and spring A
attached to the cam is pushed. The shift-linkage develops the spring force in
spring A, the cam is pushed in as illustrated in Figure 2.26, the surface shape
moves the pawl to the engagement position and the sprocket is locked by the pawl.
The vehicle is then in the parking mode.
While the car is in motion, the pawl cannot be engaged with the sprocket. If
the car speed is over 4.8 km/hour, an impact load occurs between the pawl and the
sprocket, and the impact load prevents engagement. When the impact load is
greater than the spring force, the parking mode does not function.
When the car is parked on a hill, the automobile weight exerts a torque on the
sprocket and the torque is transmitted to the cam by the tooth shape and the pawl.
Therefore, to disengage the parking mode, we need more force than the friction
force between the cam and the pawl. If the cam is pulled out, the pawl is released
by the tension spring.
Axiomatic Design 61
Shift
linkage
Connected to
transmission
To valve Sprocket
Detent Tension
spring spring
Cam
Spring A Pawl
Figure 2.26. Schematic drawing of a parking mechanism
Figure 2.27 is the free body diagram of the forces acting on the pawl. FR is
the reaction force between the pawl and the sprocket. FC is the reaction force
between the pawl and the cam, FS is the spring force, FP is the force acting on the
pawl by the pin and P is the friction coefficient between the pawl and the cam.
As the slope of the tooth profile in the pawl increases, FR , FC and PFC increase
in order. FS is constant while the cam is engaged.
The functional requirements of the system are as follows:
FR1 : Engage the pawl in the locked position.
FR2 : Disengage the pawl from the locked position.
FR3 : Prevent accidental engagement.
Force exerted by
sprocket = FR
Pawl
FP
ȝFC
FC FS
Figure 2.27. Free body diagram of the pawl

FR4 : Keep the pawl in the engaged position.

FR5 : Carry the load transmitted by the vehicle.
The current design has the following design parameters:

DP1 : The tapered section of the cam profile
DP2 : Tension spring
DP3 : The tooth profile of the sprocket and the pawl/spring A/shift-linkage/
tension spring
DP4 : The flat surface of the cam
DP5 : The flat surface of the pawl/cam
The design equation is
ª FR1 º ªX 0 0 X X º ª DP1 º
« FR » «X « »
« 2» « X 0 0 X »» « DP2 »
« FR3 » «0 X X 0 0 » « DP3 » (2.77)
« » « »« »
« FR 4 » «0 X 0 X X » « DP4 »
« FR5 » «¬ X X 0 X X »¼ «¬ DP5 »¼
¬ ¼
Therefore, the current design is coupled. The reason is that the vehicle weight
transmitted by the automatic transmission is sustained by the pawl and the cam.
As the slope of the hill increases, the normal and friction forces on the cam
increase and disengagement of the parking mode becomes more difficult.
Newly Proposed Design

A newly proposed design is presented in Figure 2.28. The sprocket of the new
design has a different tooth profile. The tapered section near the outer edge of the
tooth is to prevent accidental engagement of the pawl, and the flat surface of the
tooth profile of the pawl transmits the vehicle weight. The tapered section of the
pawl prevents accidental engagement when the vehicle speed is lower than 4.8
km/hour. The vertical position of the pin in the pawl is the same as the one for the
flat surface of the tooth profile of the pawl. Therefore, FR and FP of Figure 2.28
are of the same height and the force between the pawl and the cam is eliminated.
The pin is in charge of the vehicle weight and the weight is not transmitted to the
cam. FR is almost the same as FP .
The FRs are the same as before and DPs are as follows:
DP1 : The tapered section of the cam profile
DP2 : Tension spring
DP3 : Tooth profiles of the sprocket wheel and the tapered section of the
pawl/spring A/shift linkage system
DP4 : The flat surface of the cam
Axiomatic Design 63
Tooth
profile =DP32
DP51
Pin = DP52
FR FP
Spring A/ Pawl
linkage =DP31 DP4 DP1 FS
Cam Tension
FC spring = DP2
Figure 2.28. Newly proposed design
DP5 : The flat surfaces of pawl/sprocket and pin
The design equation is as follows:
ª FR 4 º ªX 0 0 0 0 º ª DP4 º
« FR » «X « »
« 2» « X 0 0 0 »» « DP2 »
« FR3 » «0 X X 0 0 » « DP3 » (2.78)
« » « »« »
« FR1 » «X 0 0 X 0 » « DP1 »
« FR5 » «¬ 0 0 0 0 X »¼ «¬ DP5 »¼
¬ ¼
The design is a decoupled one.

From the characteristics of DP3 , FR3 is decomposed as follows:
FR31 : Control the force that pushes the pawl into sprockets.
FR32 : Generate the reaction force if the sprocket is turning.
The corresponding DPs are

DP31 : Spring A/linkage
DP32 : Tooth profile of the sprocket
The related design matrix is triangular.

FR5 and DP5 are decomposed as follows:
FR51 : Transmit the force from the sprocket to the pawl.

FR52 : Carry the load transmitted.
DP51 : Nearly vertical surface of the pawl and the sprocket tooth profile
DP52 : Pin located collinearly with the force vector acting on the vertical
surface the pawl
To minimize the reaction force between the cam and the pawl, the reaction
force between the sprocket and the pawl should be close to the horizontal line. For
this, DP51 is nearly vertical. The small slope between the pawl and the sprocket is
used to minimize the reaction of the pawl. The design equation is diagonal. An
improved design is found by using the Independence Axiom. A better idea may be
created with application of the Independence Axiom.
2.5 Software Design Using the Axiomatic Approach
2.5.1 Software Design
The importance of software is being recognized in all engineering fields. Software

is a technology or a methodology to manipulate computers. Software engineering
is a method or a tool to develop reliable software with minimum cost. Generally,
engineering software developers lack understanding in software engineering.
Engineers tend to develop software based on their own methods and experiences,
which is neither systematic nor efficient. Moreover, documentation is not
sufficient during software development. Therefore, further development is needed
for maintenance, modification, extension, etc.
In software engineering, these problems are solved by two approaches. First,
many resources are invested in the early stages. Independent modules are defined
and software is designed based on the modules. Thus later work can be
considerably reduced. Second, new systematic languages such as the object
oriented language can be utilized. Thus the work of the developers can be reduced.
However, although developers use such methods, they still have classical problems
such as debugging, maintenance, modification and extension. The most important
reason is that physically independent modules can be functionally coupled during
the execution of software.
As mentioned earlier, the axiomatic approach is a method to maintain the
independence between functional requirements. It can be applied to software
engineering. In this section, the axiomatic approach is applied to software
development to overcome the intrinsic limits of conventional software engineering.
2.5.2 Conventional Languages and Axiomatic Design
In software engineering, partitioning is frequently used to manage complexity.

That is, a large program is divided into manageable smaller modules. However, if
a small module is not independent or the interactions are not clearly defined,
Axiomatic Design 65
complexity cannot be controlled. Modulation enables easy maintenance and

modification.
From the axiomatic viewpoint, CAs, FRs, DPs and PVs are redefined for
software development as follows:
CAs: Customer requirements or attributes that the software should satisfy
FRs: Functional requirements that software should satisfy in engineering
terminology
DPs: (1) Input data when an algorithm is developed
(2) Signal from the hardware where software is loaded
(3) Program code
PVs: Subroutine, machine language, compiler
The process for software development will be explained based on the above
definitions. The development of a software system for libraries is selected as an
example (Kim et al. 1991, Suh 2000).
Step 1. Definition of FRs for the software system
The functional requirements of the highest level are defined based on the
customer needs. A functional requirement is a function that the software
system intends to carry out. As mentioned earlier, it starts with a verb
because it executes a process with input.
The functional requirements are as follows:
FR1 : Generate the call number and keyword database for new incoming
books.
FR 2 : Provide a list of books that corresponds to subject keywords of a
search query.
Step 2. Mapping between the domains to maintain the independence of FRs
Design parameters are defined in the physical domain. Design parameters
determine how to achieve the functional requirements. In software design,
design parameters correspond to input data and result data from program
execution.
The design parameters of the highest level are as follows:
DP1 : A classification system based on the content of the book
DP2 : A search system based on the set of subject keywords
The FRs and DPs satisfy the Independence Axiom as a decoupled design
because the design matrix is triangular as in Equation 2.79.
ª FR1 º ª A11 0 º ª DP1 º

« FR » «A (2.79)
¬ 2¼ ¬ 21 A22 »¼ «¬ DP2 »¼
An element of the design matrix Aij can be an operation or a calling

function. In this case, the functional requirements can be satisfied by the
modules in Equation 2.80,
FR1 M 1 DP1
(2.80)
FR 2 M 2 DP2
where M 1 and M 2 are modules defined as follows:
M1 A11
DP1 (2.81)
M2 A21 A22
DP2
A module is regarded as an algorithm. It can be a logical operation or a

function representing an independent system.
Step 3. Decomposition of FRs and DPs
It was mentioned earlier that the FRs and DPs are decomposed up to the
lowest level. The DPs of the current level are references for the FRs of the
next level. Therefore, the functional requirements of the lower level FRij
are defined based on DPi of the upper level. The decomposition is carried
out by a zigzagging process.
FR1 of Step 1 is decomposed into FR11 and FR12 as follows:
FR11 : Assign a call number to a new book.
FR DP
FR1 FR2 DP1 DP2
FR11 FR12 FR21 FR22 DP11 DP12 DP21 DP22
FR121 FR122 DP121 DP122
Functional domain Physical domain
Figure 2.29. Hierarchical structure of a software system for libraries

Axiomatic Design 67
S Left C Right Left F Right
Summation junction Control junction Feedback junction

(uncoupled design) (decoupled design) (coupled design)
Figure 2.30. Unit junctions
FR12 : Generate subject keywords for the new book.

DP11 : Information on the title page of the book
DP12 : The table of contents of the book
FR 2 can also be decomposed and decomposition is continued to the lowest

level. The result of the decomposition is illustrated in Figure 2.29.
Step 4. Definition of modules
After decomposition, the modules are defined for all FRs and DPs. Each
module can be independently coded. The entire flow can be schematically
drawn by junctions and modules. Figure 2.30 presents unit junctions.
There are three junctions as follows:
Summation junction (̺ S ): This is for an uncoupled design. An FR of the
upper level is satisfied by summation of results from the modules of
the lower level.
Control junction ( ̺C ): In Figure 2.30, the results of the left hand side
C
M1 M2
C C
M11 M12 M21 M22
M121 M122
Figure 2.31. Module junction structure diagram of Figure 2.29

M121
M11 C S M21 C M22
M122
Figure 2.32. System flow of Figure 2.29
modules are utilized to control the module of the right hand side. This
represents a decoupled design.
Feedback junction (̺ F ): This is for a coupled design. In Figure 2.30, the
results of the right hand side return to the left hand side as feedback.
Thus, many repetitions are needed. When there are many feedback
junctions, the program is not manageable.
With the junctions, the hierarchical structure of FRs and DPs can be
represented by a tree structure. This is called a module junction structure
diagram. The example in Figure 2.29 is modified to that in Figure 2.31.
The module junction structure diagram can be modified to the flow of the
network type. Figure 2.32 shows the flow induced from Figure 2.31.
In this section, the application of axiomatic design is explained for the
development of software using conventional languages.
Customer 7 Software
needs 1 Code with system product
Define FRs architecture
2 6
Establish interfaces
el
Bu (Top
Map to DPs
roa mod
ild
5
the down
ch)
3
app ted
Identify classes
sof ap
-up ien
-
Decompose
tw pro
om ct or
4
are ac
Define
(B e obj
hie h)
rar
modules
th
ott
ch
ild
y
Bu
Identify leaves
(Full design matrix)
Figure 2.33. The axiomatic approach for objected oriented programming (V-model)
Axiomatic Design 69
2.5.3 Object Oriented Programming and Axiomatic Design
Since the 1980s, the object oriented paradigm has received much attention in
software engineering. It is a new approach compared to process oriented
languages such as C, Pascal, Fortran, etc. The object oriented language, which is
popular these days, is appropriate for graphic user interface (GUI). The object
oriented technology provides methods to use existing programs. Common libraries
are prepared and specialized by customization. Also, a large program is divided
into independent objects and objects have relations by well defined interfaces.
Due to the above advantages, object oriented programming (OOP) is frequently
utilized in software development. The V-model has been proposed to exploit the
axiomatic approach in object oriented programming (Do 2000). In the V-model, a
designer defines the functional requirements of the software and establishes
independent modules from zigzagging decomposition. Each module is modified to
a class of the object oriented programming and coded. The process consists of two
steps: construction of the full design matrix with the top-down approach and
a b c d
Data structure list

(attributes)
A Ed
Fd
F*
B
Gd
C G* Operation list
Hd
H*
D
Behavior list
(methods)
Figure 2.34. The full design matrix using object oriented programming
Class FR1 Class FR2
Class A Class B Class C Class D

a
Ed
Class Bd Class B* Class Cd Class C* Class Dd Class D*
b a c a,b d a,b,c
Fd F* Gd G* Hd H*
Figure 2.35. The class diagram for the full design matrix in Figure 2.34
coding the program with the bottom-up approach. The process is illustrated in
Figure 2.33. The top-down approach up to Step 4 is the same as the steps
explained in the previous section. Thus, this section describes the steps after Step
4.
Step 5. Identification of objects, attributes and operations
The full design matrix is constructed after the decomposition process. It
shows all FRs and DPs. An example of the full design matrix is presented
in Figure 2.34. Rows of Figure 2.34 are FRs and the columns are
DPs. ”X” means an algorithm of a logical relation between an FR and a
DP. The logical relation includes not only operators such as ”+” and ” ”
but also control statements such as ”if,” ”for,” etc. In Figure 2.34, a
rectangle with thick lines is an object. The object is composed of attributes
in columns (DPs) and the methods of the operations list within the
rectangle. The method of the object is the module. Therefore, an object
executes a method with attributes and satisfies the functional requirement.
Step 6. Establishment of interfaces between objects
An object is expressed by a class and a class is a template that defines the
format of the object. Classes share attributes that are the data structures
and behaviors. In this step, the relationships between classes are set up.
They are generalization, aggregation and association. Figure 2.35 presents
a class diagram according to the design matrix of Figure 2.34. We can see
the data and their functions in Figure 2.35 and the class diagram shows the
relations of classes. The design process is shown by the aforementioned
flow. Thus, software development easily proceeds with these.
Step 7. Coding with system architecture
Coding is the programming process based on the classes and their
relationships. The flow chart of the design matrix helps with the coding.
Axiomatic Design 71
2.6 Discussion
As explained earlier, the two axioms are independent of each other. Thus, we have
to apply them separately. Generally, the Independence Axiom should be satisfied
first. In many cases, the design is terminated only with the application of the
Independence Axiom. When both of the axioms are utilized, the flow in Figure
2.20 is recommended.
When we apply the Independence Axiom, the ideal design should be kept in
mind. The numbers of FRs and DPs are the same in an ideal design. The design
matrix should be a square diagonal or triangular one. If the numbers are different,
the design is coupled. When the number of DPs is smaller, new DPs should be
added. In a redundant design where the number of DPs is larger, the number
should be reduced or some specific DPs should be fixed.
Suppose we have the following redundant design:
ª DP1 º
ª FR1 º ªX X 0 º«
« FR » «0 DP2 »» (2.82)
¬ 2¼ ¬ X X »¼ «
«¬ DP3 »¼
First, we can fix DP2 . Then the design becomes an uncoupled one. That is,
redundant parameters are fixed to make the design uncoupled or decoupled with
the rest of the parameters.
The Information Axiom is utilized to quantitatively evaluate a design that
satisfies the Independence Axiom. It is especially useful when multiple designs
are compared. When multiple designs, which satisfy the Independence Axiom, are
found, the one with the minimum information content is selected as the final
design.
Basically, the axiomatic design can be exploited in creating a new design or
evaluating existing designs. It is quite useful in the conceptual design of new
products. Although the history of the method is relatively short, the usefulness has
been verified through many examples. There are some common responses from
application designers. First, they tend to easily agree with the axioms and think
that they can use them right away. However, they have difficulties in testing the
axioms with their existing products. In most cases, they tend to look at the designs
with previous concepts, not from an axiomatic viewpoint. Many designers tend to
stop applying axiomatic design at this stage. However, if the designers overcome
this stage, they realize the usefulness of axiomatic design. It is important not to
consider the existing products when the functional requirements are defined.
Instead, designers should think about the functional requirements in a solution
neutral environment. In recent research, axiomatic design is utilized in detailed
designs. Later examples will demonstrate how axiomatic design is applied to the
detailed design process of structures.
2.7 Exercises
2.1 Analyze the design of a CD player with the Independence Axiom.
2.2 Analyze the design of a cellular phone with the Independence Axiom.
2.3 Find a product that uses the idea of physical integration and analyze it with
the Independence Axiom.
2.4 To design an automobile fuel tank, the following functional requirements

are defined.
FR1 : Provide in-flow of gasoline into the tank.
FR2 : Provide a means of stopping the pump when the tank is full.
FR3 : Prevent gasoline from surging back out through the inlet tube as a
result of the vapor pressure of the gasoline when the gasoline level is
higher than the end of the pipe.
FR4 : Control vapor pressure of the gasoline.
Design a fuel tank that satisfies above four FRs. The new tank should cost
less than the current one.
2.5 We have a design with the following FR–DP relation:
ª f1 º ªX 0 0 º ª x1 º
«f » «X X 0 »» «« x 2 »»
« 2» «
¬« f 3 ¼» ¬« X X X ¼» ¬« x 3 ¼»
The detailed relationships are
f1 x12
f2 x12 x1 x 2 x 22
f3 x12 x 22 x 2 x3 x32
(1) At x = (1.0, 1.0, 1.0), obtain the approximated design matrix.

(2) Obtain a condition so that the design is uncoupled in a specific design
window.
2.6 Calculate the reangularity and semangularity of the design matrices and
discuss the characteristics.
Axiomatic Design 73
ª3 1 º ª 8 0.2º ª9.5 2.4º

(a) « » (b) « » (c) « »
¬1 2¼ ¬0.3 4 ¼ ¬3.2 4.8¼
ª5 0º ª5 0º ª3 0 º
(d) « » (e) « » (f) « »
¬1 2¼ ¬ 4 2¼ ¬0 2 ¼
2.7 We have two designs as follows:
ª FR º ª3 1º ª DP1 º
Design 1 « 1 » «2 4» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
Design 2 « 1 » «8 5» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
(1) Draw an FR–DP graph for each design and explain the order of the
design process.
(2) Calculate the reangularity and semangularity of each design and
compare the results.
2.8 Suppose we have the following FR–DP relation:
ª FR1 º ª 5 0.1 0.2º ª DP1 º

« FR » «0.3 7 0.4» « DP »
« 2» « »« 2 »
«¬ FR3 »¼ «¬0.5 0.7 4 »¼ «¬ DP3 »¼
When DP1 DP2 DP3 1, the manufacturing tolerances are 'DP1 ,

'DP2 and 'DP3 .
(1) In Equation 2.21, ('FRi ) allowable is 0.5. Write inequality equations

composed of 'DP1 , 'DP2 and 'DP3 for the condition that the above
design is an uncoupled one.
(2) Designer specified tolerances are 4.8 FR1 5.8 , 7.3 FR2 8.0
and 5.1 FR3 6.3 . Similarly as in (1), write the condition for the
above design to be an uncoupled one.
2.9 We have the following designs:
Design 1 « 1 » «0 4» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
Design 2 « 1 » «2 4» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
Design 3 « 1 » «2 4» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
(1) Which one satisfies the Independence Axiom?

(2) Among the ones that satisfy the Independence Axiom, select the best
one in the context of independence by comparing R and S.
(3) For the designs satisfying the Independence Axiom, 'DPi 0.1 and
'FRi 0.5 , i 1, 2. Calculate the information content at the design
point that satisfies the functional requirements.
(a) When DPs have uniform distribution in the DP range, calculate the
information content by the graphical method and compare them.
(b) Calculate the information content in the same manner as (a) in the
FR range.
(c) Distributions of the design parameters are as follows:
3 3
Distribution of DP1 pGDP1 3
(GDP1 ) 2
4('DP1 ) 4('DP1 )
1 1
Distribution of DP2 pGDP2 2
GDP2 when
('DP2 ) 'DP2
1 1
GDP2 0 , and pGDP2 2
GDP2 when GDP2 t 0
('DP2 ) 'DP2
Calculate the information content by the integration method and

compare the information contents.
2.10 We have a decoupled design and one element of the design matrix is
expressed by an unknown x. The value of x is 0.5, 1 or 2 in the following
design equation:
ª 0 FR1 2 º ª1 0 º ª 0 DP1 1º
«0 FR 2» «1 x » «0 DP 1»
¬ 2 ¼ ¬ ¼¬ 2 ¼
For each x, calculate the following:

(1) The probability of success.
(2) Tolerances to have 100% of the probability of success. Any method
can be used for the evaluation of the probability of success. Discuss
the trend according to x.
2.11 Make up a problem for buying a laptop computer in the same way as in the
house buying problem and solve it.
Axiomatic Design 75
2.12 We have the following designs:
ª FR º ª100 0 º ª DP1 º
Design 1 « 1 » « 0 100» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
Design 2 « 1 » «1 1» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
The targets of FRs are 10 and 20, respectively. The allowable tolerances
are 9.5 FR1 10.5 and 19.5 FR2 20.5 , and the tolerance of each
DP is r 2.0.
(1) Calculate R and S. Which one is better from the viewpoint of

independence?
(2) When functional requirements are satisfied, calculate the information
content by the graphical method and select the better one.
(3) Discuss the results.
2.13 If the off-diagonal terms are small in a decoupled design, it can be

considered as an uncoupled design. Explain this with Equation 2.21 and R.
2.14 Expand the graphical method for a decoupled design to a coupled design.
2.A Corollaries and Theorems

Corollary 2.A.1 [Decoupling of Coupled Designs]
Decouple or separate parts or aspects of a solution if FRs are coupled or
become independent in the designs proposed.
Corollary 2.A.2 [Minimization of FRs]

Minimize the number of FRs and constraints.
Corollary 2.A.3 [Integration of Physical Parts]

Integrate design features in a single physical part if FRs can be
independently satisfied in the proposed solution.
Corollary 2.A.4 [Use of Standardization]

Use standardized or interchangeable parts if the use of these parts is
consistent with FRs and constraints.
Corollary 2.A.5 [Use of Symmetry]

Use symmetrical shapes and/or components if they are consistent with FRs
and constraints.
Corollary 2.A.6 [Largest Design Ranges]

Specify the largest allowable design range in stating FRs.
Corollary 2.A.7 [Uncoupled Design with Less Information]

Seek an uncoupled design that requires less information than coupled
designs in satisfying a set of FRs.
Corollary 2.A.8 [Effective Reangularity of a Scalar]

The effective reangularity R for a scalar coupling “matrix” or element is
unity.
Axiomatic Design 77
Theorem 2.A.1 [Coupling Due to an Insufficient Number of DPs]

When the number of DPs is less than the number of FRs, either a coupled
design results or the FRs cannot be satisfied.
Theorem 2.A.2 [Decoupling of a Coupled Design]

When a design is coupled because of a larger number of FRs than DPs (i.e.,
m>n), it may be decoupled by the addition of new DPs so as to make the
number of FRs and DPs equal to each other if a subset of the design matrix
containing n u n elements constitutes a triangular matrix.
Theorem 2.A.3 [Redundant Design]

When there are more DPs than FRs, the design is either a redundant design
or a coupled design.
Theorem 2.A.4 [Ideal Design]

In an ideal design, the number of DPs is equal to the number of FRs and
the FRs are always maintained independently of each other.
Theorem 2.A.5 [Need for a New Design]

When a given set of FRs is changed by the addition of a new FR, by
substitution of one of the FRs with a new one, or by selection of a
completely different set of FRs, the design solution given by the original
DPs cannot satisfy the new set of FRs. Consequently, a new design
solution must be sought.
Theorem 2.A.6 [Path Independence of an Uncoupled Design]

The information content of an uncoupled design is independent of the
sequence by which the DPs are changed to satisfy the given set of FRs.
Theorem 2.A.7 [Path Dependency of Coupled and Decoupled Design]

The information contents of coupled and decoupled designs depend on the
sequence by which the DPs are changed to satisfy the given set of FRs.
Theorem 2.A.8 [Independence and Design Range]

A design is an uncoupled design when the designer-specified range is
greater than
n § wFR ·
¦ ¨¨ i ¸'DPj
¸
j z i© wDP j ¹
j 1
in which case the off-diagonal elements of the design matrix can be

neglected from the design consideration.
Theorem 2.A.9 [Design for Manufacturability]

For a product to be manufacturable with reliability and robustness, the
design matrix for the product A (which relates the FR vector for the
product to the DP vector of the product), times the design matrix for the
manufacturing process B (which relates the DP vector to the PV vector of
the manufacturing process), must yield either a diagonal or a triangular
matrix. Consequently, when either A or B represents a coupled design, the
independence of FRs and robust design cannot be achieved. When they are
full triangular matrices, either both of them must be upper triangular or
both must be lower triangular for the manufacturing process to satisfy
independence of functional requirements.
Theorem 2.A.10 [Modularity of Independence Measures]

Suppose that a design matrix A can be partitioned into square submatrices
that are nonzero only along the main diagonal. Then the reangularity and
semangularity for A are equal to the product of their corresponding
measures for each of the nonzero submatrices.
Theorem 2.A.11 [Invariance]

Reangularity and semangularity for a design matrix A are invariant under
alternative orderings of the FR and DP variables, as long as the orderings
preserve the association of each FR with its corresponding DP.
Theorem 2.A.12 [Sum of Information]

The sum of information for a set of events is also information, provided
that proper conditional probabilities are used when the events are not
statistically independent.
Axiomatic Design 79
Theorem 2.A.13 [Information Content of the Total System]

If each DP is probabilistically independent of other DPs, the information
content of the total system is the sum of the information of all individual
events associated with the set of FRs that must be satisfied.
Theorem 2.A.14 [Information Content of Coupled Versus Uncoupled Designs]

When the state of FRs is changed from one state to another in the
functional domain, the information required for the change is greater for a
coupled design than for an uncoupled design.
Theorem 2.A.15 [Design–Manufacturing Interface]

When the manufacturing system compromises the independence of the FRs
of the product, either the design of the product must be modified or a new
manufacturing process must be designed and/or used to maintain the
independence of the FRs of the products.
Theorem 2.A.16 [Equality of Information Content]

All information contents that are relevant to the design task are equally
important regardless of their physical origin, and no weighting factor
should be applied to them.
Theorem 2.A.17 [Design in the Absence of Complete Information]

Design can proceed even in the absence of complete information only in
the case of a decoupled design if the missing information is related to the
off-diagonal elements.
Theorem 2.A.18 [Existence of an Uncoupled or Decoupled Design]

There always exists an uncoupled or decoupled design that has less
information than a coupled design.
Theorem 2.A.19 [Robustness of Design]

An uncoupled design and a decoupled design are more robust than a
coupled design in the sense that it is easier to reduce the information
content of designs that satisfy the Independence Axiom.
Theorem 2.A.20 [Design Range and Coupling]

If the design ranges of uncoupled or decoupled designs are tightened, they
may become coupled designs. Conversely, if the design ranges of some
coupled designs are relaxed, the designs may become either uncoupled or
decoupled.
Theorem 2.A.21 [Robust Design when the System Has a Nonuniform pdf]
If the probability distribution function (pdf) of the FR in the design range is
nonuniform, the probability of success is equal to one when the system
range is inside the design range.
Theorem 2.A.22 [Comparative Robustness of a Decoupled Design]

Given the maximum design ranges for a given set of FRs, decoupled
designs cannot be as robust as uncoupled designs in that the allowable
tolerances for DPs of a decoupled design are less than those of an
uncoupled design.
Theorem 2.A.23 [Decreasing Robustness of a Decoupled Design]

The allowable tolerance and thus the robustness of a decoupled design with
a full triangular matrix diminish with an increase in the number of
functional requirements.
Theorem 2.A.24 [Optimum Scheduling]

Before a schedule for robot motion or factory scheduling can be optimized,
the design of the tasks must be made to satisfy the Independence Axiom by
adding decouplers to eliminate coupling. The decouplers may be in the
form of a queue or of separate hardware or buffer.
Theorem 2.A.25 [“Push” System vs. “Pull” System]

When identical parts are processed through a system, a “push” system can
be designed with the use of decouplers to maximize productivity, whereas
when irregular parts requiring different operations are processed, a “pull”
system is the most effective.
Axiomatic Design 81
Theorem 2.A.26 [Conversion of a System with Infinite Time-Dependent

Combinatorial Complexity to a System with Periodic
Complexity]
Uncertainty associated with a design (or a system) can be reduced
significantly by changing the design from one of serial combinatorial
complexity to one of periodic complexity.
2.B Axiomatic Design of a Beam Adjuster for a Laser

Marker
2.B.1 Problem Description

A laser marker is a machine that engraves characters or logos on the surface of
semiconductors. Figure 2.B.1 shows a laser marker and Figure 2.B.2 is a
schematic presentation of the inside. This is the beam scanning type YAG laser. It
engraves the characters with a laser and high speed mirrors as we write with a pen.
The YAG laser is a solid-state laser that uses crystals of yttrium, aluminum and
garnet. As illustrated in Figure 2.B.1, the laser marker consists of a beam
generating part and a scanning head.
In the beam generator, the laser beam is produced and reflected by the mirrors
as illustrated in Figure 2.B.2. One laser beam is divided into two beams by an
optical device. The optical device is a mirror that reflects 50% of the beam and
passes the rest (see Figure 2.B.2). It is efficient in that two semiconductors are
marked with one generator. This type is called a dual laser marker and is widely
used in the field of semiconductor surface marking. In the scanning head, there are
other mirrors controlled by high-speed motors. The fixed beam from the beam
generator can be redirected by these mirrors to mark certain logos. If the beam
direction is determined by the beam generator, the mirrors and motors in the
scanning head make the detailed marks, and the motors are controlled by a
computer program.
Before the real marking process is conducted, many test processes are needed
for trial and error. If we use the YAG laser in this process, the surfaces of the
Beam generator
Scanning head
Figure 2.B.1. A beam scanning type laser marker

Axiomatic Design 83
Mirror YAG laser

Adjuster
50% mirror
Diode laser
Scanning head
Marking position
Figure 2.B.2. Schematic view of the inside of a laser marker
semiconductors are damaged. Therefore, a low-cost simulation is carried out by a

diode laser as illustrated in Figure 2.B.2. The diode laser sheds a weak light beam
and the simulation can be easily carried out.
The simulation process is as follows:
(1) Test plates are placed at the marking positions in Figure 2.B.2. The YAG
laser is turned on. The mirrors in the beam generator are positioned so as
to make the beam go through the scanning head and mark points on the
YAG laser point
Diode laser point
(a) Before alignment
YAG laser point
Diode laser point
(b) After alignment
Figure 2.B.3. The process of beam alignment

plates. The points are starting points of the marking process and illustrated
as hollow points as shown in Figure 2.B.3.
(2) The YAG laser is turned off. The diode laser is turned on. The solid points
in Figure 2.B.3a are the final destinations of the diode laser.
(3) The adjuster of the diode laser is utilized to make two identical points as
illustrated in Figure 2.B.3b. If the two points match, the angles and the
final destinations from the YAG and diode lasers are considered identical.
Now, we are sure that the two lasers have the same routes.
(4) The marking is simulated with the diode laser. That is, the motors in the
scanning head are simulated by a computer program. The program is the
one specifically developed for the marking process. As mentioned earlier,
the marking result is visible.
(5) If the results are validated, the test is terminated.
Many problems occur in the adjuster of the diode laser. Currently, screws are
used for the adjustment. Precise adjustment is difficult to obtain, since tolerances
and human errors are involved. Thus, the adjustment is a long process.
2.B.2 Axiomatic Analysis of an Existing Design
Since the laser marking machine has already been commercialized, there is an
existing design for the diode beam adjuster. Therefore, it is necessary to define the
functional requirements and corresponding design parameters to evaluate the
existing device. The relationship between FRs and DPs can be expressed by a
design matrix. The FRs of the existing device is defined as follows:
FR1 : Align the vertical position of the diode laser beam.
FR2 : Align the vertical angle of the diode laser beam.
FR3 : Align the horizontal position of the diode laser beam.
FR4 : Align the horizontal angle of the diode laser beam.
FR5 : Fix the beam alignment.
Figure 2.B.4 illustrates each functional requirement. The two beams from the
YAG and diode lasers should be properly matched. First, the horizontal and
vertical destinations of the diode laser should be the same as those of the YAG
laser ( FR1 , FR3 ). Second, the angles of the beams must be the same ( FR2 , FR4 ).
Figure 2.B.5 illustrates the existing product. DPs corresponding to FRs are
Origin FR1 FR2 FR3 FR4
Figure 2.B.4. The functional requirements in order

Axiomatic Design 85
DP1
DP2
DP3
Figure 2.B.5. The existing design
defined as follows:
DP1 : Vertically moving component
DP2 : Supporting block
DP3 : Fixing screw
The design matrix is a coupled one as follows:
ª FR1 º ªX 0 0º
« FR » «X
« 2» « 0 0 »» ª DP1 º
« FR3 » «0 « »
X 0 » « DP2 » (2.B.1)
« » « »
« FR 4 » «0 X 0 » «¬ DP3 »¼
« FR5 » «¬ 0 0 X »¼
¬ ¼
The design in Equation 2.B.1 is a coupled design because the number of DPs is
less than the number of FRs. When we move the solid points in Figure 2.B.3a
( DP1 ), the vertical angle also varies because FR1 and FR2 are coupled by DP1 .
In a similar manner, when we move the horizontal position ( DP2 ) the aligned
angle can vary. If a design is coupled in the way of Equation 2.B.1, it can be
decoupled by adding new DPs to make the numbers of FRs and DPs equal.
2.B.3 The Development of a New Beam Adjuster
New Design Using the Independence Axiom

A new design is created with new design parameters to satisfy the Independence
Axiom. If we make a new design considering FR1 and FR2 , which are for the
vertical position and angle, it can be expanded to FR3 and FR4 , which are for the
horizontal position and angle. The design matrix for FR1 and FR2 is stated in
Equation 2.B.2,
ª FR1 º ªX º
« FR » « X »>DP1 @ (2.B.2)
¬ 2¼ ¬ ¼
We can think of a design that has independent design parameters for the
vertical position and angle. As a result, two designs are made. The first one is
illustrated in Figure 2.B.6. The fastener at the back ( DP1 ) controls the vertical
position and the front one ( DP2 ) controls the vertical angle. After the position is
fixed, the fastener is tightened by a screw. The design matrix for Figure 2.B.6 is
decoupled as follows:
« FR » «X « » (2.B.3)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
We can have multiple designs satisfying the Independence Axiom. Another

design is created as illustrated in Figure 2.B.7. Two screws are used at the front
and the back. This design is different from Figure 2.B.6 in that the position and
angle can be controlled very slowly by using the screws. This design is also a
decoupled design as follows:
« FR » «X « » (2.B.4)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
DP2 DP1
DP2 DP1
Figure 2.B.6. Design 1 Figure 2.B.7. Design 2

Axiomatic Design 87
The above two designs show a method to make a decoupled design by defining
new design parameters. The method can be expanded for other DPs.
Figure 2.B.8 is the expansion of Figure 2.B.7. Five FRs are the same as before
and DPs for Figure 2.B.8 are as follows:
DP1 : Upper rear screw
DP2 : Upper front screw
DP3 : Side rear screw
DP4 : Side front screw
DP5 : Fixing screw
DP3 and DP4 are similar to the aforementioned DP1 and DP2 . We can think of
two designs in Figure 2.B.6 and Figure 2.B.7. DP3 and DP4 of Figure 2.B.8 are
selected in the same manner as in Figure 2.B.7. The expanded design is also a
decoupled design as
ª FR1 º ªX 0 0 0 0 º ª DP1 º
« FR » «X « »
« 2» « X 0 0 0 »» « DP2 »
« FR3 » «0 0 X 0 0 » « DP3 » (2.B.5)
« » « »« »
« FR 4 » «0 0 X X 0 » « DP4 »
« FR5 » «¬ 0 0 0 0 X »¼ «¬ DP5 »¼
¬ ¼
Therefore, the new designs satisfy the Independence Axiom. Using various new
ideas, we can create other designs that satisfy the Independence Axiom.
DP1
DP2
DP3
DP4 DP5
Figure 2.B.8. Final design

Selection of the Final Design Using the Information Axiom

The Information Axiom is utilized to select the best design out of multiple designs
satisfying the Independence Axiom. The probability of success is considered as
the information content. The relationship between the FRs and DPs should be
expressed by explicit functions to evaluate the information content. The two
designs in Figure 2.B.9 are compared for the information content.
In model #1 of Figure 2.B.9, the movement of the DP is the same as the
movement of the beam. Therefore, the slope (m) in Figure 2.B.10 is 1. On the
other hand, the beam moves as much as a pitch when the screw rotates once in
model #2. The relationship is
2Sr tan ș p (2.B.6)
where r is the radius of the screw, T is the angle of the screw and p is the pitch.
When the radius is 1.5 mm and the pitch is 1 mm, the slope is 0.106 as shown
in Figure 2.B.10. Considering the environmental and geometrical aspects of a
certain existing design, the design matrices of the two designs are as follows:
ª FR º ª 1 0º ª DP1 º
Model #1: « 1 » « 4 4» « DP » (2.B.7)
¬ FR2 ¼ ¬ ¼¬ 2 ¼
ª FR º 1 ª 1 0º ª DP1 º
Model #2: « 1 » « 4 4» « DP » (2.B.8)
¬ FR2 ¼ 3S ¬ ¼¬ 2 ¼
Suppose that the target FR * [0 0]T is satisfied by DP* [0 0]T . The

information content can be calculated when the target is achieved. The
information content is calculated by the graphical method using Equation 2.40 or
2.41. Assume that 'FR1 0.2 mm and 'FR2 0.3 rad . The tolerance of the
design parameter is the tolerance when it is controlled by hand. Suppose
DP
DP
Model #1 Model #2
Figure 2.B.9. Design parameters for comparison of the information content

Axiomatic Design 89
FR Model #1
(distance)
mm
m=1
Model #2
m = 1/3ʌ
DP (distance) mm
Figure 2.B.10. Slope of DP with respect to FR
'DPi 1 mm, i 1, 2 have uniform distributions. The probability of success can

vary according to the assumptions. In this case, the selection of a design is more
important than the amount of the probability of success. Therefore, the above
assumptions are valid for the selection.
The graphical method can be used in the same fashion as shown in Figure 2.13
or Figure 2.14. We use the integration method here. Using the unit step function
in Figure 2.15, the probability distribution for each model is as
1
pi [u (GDPi ('DPi )) u (GDPi ('DPi ))], i 1, 2 (2.B.9)
2 u1
Substituting Equations 2.B.7í2.B.9 into Equation 2.49, the probabilities of

success p1s and p 2 s and the information contents I 1 and I 2 are
Model #1: p1s 0.015, I1 6.059 (bits)

Model #2: p 2 s 0.582, I2 0.781 (bits)
Therefore, model #2 is better than model #1. It is the same for the design of DP3
and DP4 . In conclusion, the design in Figure 2.B.8 is determined as the final
design.
2.B.4 Summary
The flow of Figure 2.20 is applied to this problem as an example. Multiple

designs are created based on the Independence Axiom and the final design is
selected by the Information Axiom. As a result, an excellent design is made to
overcome the weakness of the existing design (Shin and Park 2004).
2.C The Development of a Design System for a TV

Glass Bulb
2.C.1 Problem Description

A glass bulb is the output device of a TV. It is an element of a TV tube and the
tube is sometimes called a “brown tube.” A tube consists of a shadow mask, an
electron gun, a band and a glass bulb. The glass bulb is composed of the panel
(front glass) and the funnel (rear glass). Figure 2.C.1 presents the shape of the
glass bulb.
In conventional design, the information flow of product design is carried via
drawings. It is also inefficient in that the design processes are performed in
heterogeneous systems. To improve the process, a design software system is
developed based on the axiomatic approach to improve the design process and to
strengthen the information flow.
2.C.2 The Conventional Design Process for a Glass Bulb
The conventional design process is illustrated in Figure 2.C.2. The process is

defined by functional requirements as follows:
FR1 : Construct the basic information of the product.
FR 2 : Establish the product shape.
Body section radii(x/y/size)
Seal edge line

Z height (Major axis)
Top of round Mold match line

Face curvature height Body
Skirt radius
Center face thickness Reference
Z height Inside blend radius height
Top of round line
Outside blend radius Reference line Yoke
Over
Neck seal line
all height Neck seal
Mold match line height
Mold match Contact start point
line height Neck seal (Z height from seal edge)
Seal edge line diameter
(Major axis) Reference Yoke section radii(x/y/size)
Skirt height
Seal edge thickness diameter
Inside blend radius center distance [MAJ]
Top of round
diameter
Section view
Section view (outside)
(a) Panel (b) Funnel
Figure 2.C.1. Shape of the glass bulb

Axiomatic Design 91
Product request: basic data offered
Rough drawing using CAD software
Three-dimensional shape generation
Strength analysis
Drawing generation
Figure 2.C.2. The conventional design process of the glass bulb
FR3 : Verify the characteristics of the product.

FR 4 : Generate the product drawing.
The FRs are mapped into design parameters in the physical domain.
DP1 : A set of basic data
DP2 : The three-dimensional shape structure for the panel and the funnel
DP3 : Loading conditions for the panel and the funnel
DP4 : A set of drawing data
The relationship between FRs and DPs is
ª FR1 º ªX 0 0 X º ª DP1 º
« FR » «X « »
« 2» « X 0 X »» « DP2 »
(2.C.1)
« FR3 » «X X X X » « DP3 »
« » « »« »
¬ FR4 ¼ ¬X X 0 X ¼ ¬ DP4 ¼
FR1 is enabled by the basic data and drawings given by a customer. In the
same manner, FR 4 is enabled by the basic data, the three-dimensional shape data
and drawings. Therefore, FR1 and FR 4 are coupled in the conventional design
process.
2.C.3 Automatic Design Software for Product Design
The FRs are redefined based on the axiomatic approach.

FR1 : Construct the database for a new product.
FR 2 : Establish the product shape.
FR3 : Verify the characteristics of the product.
FR4 : Generate the product drawing.
The corresponding DPs are

DP1 : A set of data for the new product
DP2 : The three-dimensional shape structure for the panel and the funnel
DP3 : Loading conditions for the panel and the funnel
DP4 : A set of accessory drawing data
The design matrix is a decoupled one as follows:
ª FR1 º ªX 0 0 0 º ª DP1 º
« FR » «X « »
« 2» « X 0 0 »» « DP2 »
(2.C.2)
« FR3 » «X X X 0 » « DP3 »
« » « »« »
¬ FR4 ¼ ¬X X 0 X ¼ ¬ DP4 ¼
If DP1 and DP2 are determined FR3 can be accomplished by using a

commercial structural analysis program. Thus, FR3 can be achieved by an
independent module (M3). Other functional requirements can be decomposed
based on the selected DP.
FR1 is decomposed as follows:
FR11 : Assign an ID number to a new product.

FR12 : Construct a set of data for a new product.
FR 2 and FR 4 are decomposed as follows:
FR 21 : Check the curvature (panel: flatness, funnel: axis profile).

FR 22 : Calculate the three-dimensional shape.
FR 23 : Consider the manufacturability.
FR 41 : Represent the shape of the product.

FR 42 : Display the accessory of the drawing.
The selected design parameters and the design matrix are as follows:
Axiomatic Design 93
DP11 : Representative code of the new product

DP12 : A set of specific data for the new product
« FR » «X « » (2.C.3)
¬ 12 ¼ ¬ X »¼ ¬ DP12 ¼
Since the ID number of a new product is used before making the specific
product data, the design matrix in Equation 2.C.3 is a decoupled one. The design
parameters for FR21 , FR22 and FR23 and the design matrix are as follows:
DP21 : Inside/outside curvature of the product

DP22 : The characteristic geometric equation of the product
DP23 : A set of data for the mold
« FR » «X X 0 »» «« DP22 »» (2.C.4)
« 22 » «
«¬ FR23 »¼ «¬ X X X »¼ «¬ DP23 »¼
DP23 for FR 23 is not specific; therefore, it should be decomposed.

The design parameters for FR 41 and FR 42 are as follows:
DP41 : A set of data for product design

DP42 : A set of data for the accessory
The design matrix is
« FR » «0 « » (2.C.5)
¬ 42 ¼ ¬ X »¼ ¬ DP42 ¼
FR 23 can be decomposed as follows:
FR 231 : Check the useful screen dimension for the panel.

FR 232 : Consider the ejectability.
FR 233 : Examine the deflection angle of a scanning line for the funnel.
The corresponding design parameters are as follows:

DP231 : Distance of the blending circle center position
DP232 : Angle of the side wall
DP233 : Inside curvature of the yoke part
FR Top
M1 M2 M3 M4
C C S
M11 M12 M21 M22 M23 M41 M42
M231 M232 M233
Figure 2.C.3. Module junction structure diagram of the design system for the TV glass
bulb
The design matrix is
ª FR231 º ªX 0 0 º ª DP231 º
« FR » «0 X 0 »» «« DP232 »» (2.C.6)
« 232 » «
«¬ FR233 »¼ «¬ 0 0 X »¼ «¬ DP233 »¼
The lower level of FR 23 is a uncoupled design; therefore, the lower level of FR 2

is a decoupled design.
Figure 2.C.3 presents the module junction structure diagram. Figure 2.C.4
shows the information flow for the design. The dotted lines in Figure 2.C.4 do not
represent the flow for feedback. This means a design change when the analysis
results are not satisfactory.
M231
M41
M11 C M12 C M21 C M22 C M232 S M3 C S
M42
M233
M1 M2 M3 M4
Figure 2.C.4. Information flow from Figure 2.C.3

Axiomatic Design 95
2.C.4 Software Development
Different modules should be integrated to make an automatic software system for

product design. It is recommended that a shared library be used to reduce the size
of the execution file. The functions of the software system are explained.
User Interface
The user interface is developed by using X-window (MOTIF), which is a standard
graphics tool of UNIX (Heller 1994). The user interface provides various menu
systems and graphic displays. An example is illustrated in Figure 2.C.5.
Database
A commercial database management system is utilized to handle enormous data
(Oracle Co. 1990). The transaction commands such as insert, delete, update and
inquire, which are offered by the database management system, are utilized. Input
data or design variables are given by the designer and stored in the database. The
design variables can be viewed from the database contents such as tables and
records. Various graphs can help the designer in decision making. The designer
can find the mismatch of data from the graphs. The program can automate the
manual process of the conventional design. In a conventional design, a rough final
drawing is needed for the decision making process.
Three-Dimensional Shape Generation and Display

A solid modeller is used for the three-dimensional shape display. The shape of the
glass bulb consists of several free curved surfaces in the three-dimensional space.
Figure 2.C.5. Menu display of the developed software system

Figure 2.C.6. Display of the three-dimensional shape
An in-house program called BULB-3D is employed for shape generation (Park et

al. 1995). BULB-3D uses geometrical interrelations and some specific numerical
algorithms. From the design variables, this module constructs the geometry of the
glass bulb. To confirm the shape of the glass bulb, the designer may want to see
the shape. The displays with the wire frame and surface modelling are obtained by
a commercial graphics library and special display hardware. The Starbase
Figure 2.C.7. Display of the results from strength analysis

Axiomatic Design 97
graphics library is used for this purpose (Hewlett–Packard Co. 1993). An example
is illustrated in Figure 2.C.6.
Strength Analysis
After the three-dimensional shape is constructed, strength analysis is performed. A
special mesh generation routine is used to generate an input file for strength
analysis. The failure criterion uses the maximum normal stress theory. If the
result of the strength analysis is not acceptable, an iterative process with the three-
dimensional shape generation module is carried out. This process is shown as the
dotted line in Figure 2.C.4. A commercial software system called ANSYS is
employed for the strength analysis (ANSYS Inc. 1993). Results of the strength
analysis are shown in Figure 2.C.7.
Drawing Generation
When all the activities are finished, the results are drawn. A commercial CAD
(computer-aided design) system called Unigraphics is used (Electronic Data
Systems Co. 1993). An example of the final drawing is illustrated in Figure 2.C.8.
2.C.5 Summary
The axiomatic design framework is applied to software development with a

conventional language. The conventional design of the TV glass bulb is analyzed
Figure 2.C.8. An example of the final drawing

and improved by the axiomatic approach. The approach uses the general methods
defined in the axiomatic design, such as the zigzagging process, module junctions
and system architecture for the flow. The software is designed based on the
improved design process at the early stage of software development. It is noted
that the flow of the software execution is the same as the design process.
Axiomatic Design 99
2.D The Development of a Design System for the EPS

Cushioning Package of a Monitor
2.D.1 Problem Description
A monitor is packed by cushioning materials because it may become damaged

during transportation (Yi and Park 2005). The cushioning part of a monitor is
mostly made of expanded polystyrene (EPS). Although it is lightweight, the usage
of EPS considerably increases the volume of the packing box. Therefore,
industries are trying to minimize the volume while maintaining the strength.
Currently, the cushioning package of monitors is designed based on past
experience, not with a systematic approach. When a design is finished, the
strength is validated by drop tests that are very expensive. If the design is not
satisfied, an iterative process with trial and error is carried out. In recent years,
software for computer simulated drop tests is used in the conceptual design stage.
It is well known that flexible use of the software is quite difficult due to the tedious
modelling procedure and tricky analysis skills required. Therefore, we need a
software system that automatically analyzes and designs the cushioning package of
monitors.
A software system is developed to construct the finite element (FE) model, to
perform the simulation of the drop test and to automatically design the cushioning
part. The FE model is automatically made by a commercial software LS/INGRID
and the drop test is simulated by a software system LS/DYNA3D (Livermore
Software Technology Co. 1998, 1999). The design process is established based on
the axiomatic approach and the software system is designed accordingly. The
Independence Axiom is utilized for the sequence of the design process and
software design.
2.D.2 The Development of an Automatic Design System for the

EPS Cushioning Package
The V-model and the steps introduced in Section 2.5.3 are utilized. An automatic
design system is developed for conceptual and detailed designs. First, the
conventional design method is investigated. Customer attributes (CAs) are defined
by interviewing practical designers.
Definition of FRs for the System and Decomposition (Steps 1, 2 and 3)

The design process for an EPS cushioning package is analyzed from an axiomatic
viewpoint. As a result, FRs, DPs and their relationships for the top level are
defined as shown in Table 2.D.1. The design process is a decoupled one because
the design matrix is triangular. Thus, the software design should be carried out
according to the sequence that the design matrix indicates. As mentioned earlier,
the decomposition is continued up to the minimum unit of the algorithms, that is,
the minimum unit of methods.
Table 2.D.1. Top level FRs of the design system for the EPS cushioning package of a
monitor
FRx DM DPx
1 Set up the options X O O O O Option data
Construct the data set for modelling Data for modelling and
2 X X O O O
and simulation drop test
Generate an FEM model of the Design variables of
3 X X X O O
cushioning material cushioning material
Recommend a good design value
4 X X X X O DYNA3D input deck
through simulation analysis
5 Manage the design data X X X X X Data manager
Table 2.D.2. Decomposition of FR2
FR2.x DP2. x
Construct the data set for modelling and DM Data for modelling and
simulation drop test
1 Construct modelling data for monitor X O O Modelling data for monitor
Construct modelling data for cushioning Modelling data for
2 O X O
material cushioning material
3 Construct condition data for drop test O O X Dropping condition
Table 2.D.3. Decomposition of FR 21223 at the leaf level
FR21223. x DP21223. x
Translate the files for nodes and DM21223
Files for nodes and elements
elements into the DYNA format files
1 Read the files X O O O File names
X X O O Numbers of nodes and
2 Calculate the adding quantity
elements
3 Save the files in DYNA format X X X O DYNA format
4 Save the offset-number of nodes O X O X Offset number of nodes
As shown in Table 2.D.1, FR2 is “construct the data set for modelling and
simulation.” DP2 is input data such as modelling of the monitor and input data for
analysis. From DP2 , the detailed operations of FR2 are defined. That is, various
data constructions should be made for the modelling data, analysis data and
Axiomatic Design 101
material data. Therefore, FR2 is decomposed as shown in Table 2.D.2. The

decomposition continues until the leaf level is reached. An example of the bottom
level is shown in Table 2.D.3. The flow of the software system is the same as that
of the design process except for the options of the software system and data
management.
Definition of Modules and Identification of Objects (Steps 4 and 5)

The entire full design matrix is established from the zigzagging process of the
decomposition. The full design matrix is exploited for definition of software
modules and objects. For example, Figure 2.D.1 illustrates the design matrix for
FR3 . The rectangular matrices with thick lines represent independent sub-
matrices. Each FR is defined as a module and each module is defined in the
functional domain, while each object is defined in the physical domain. Therefore,
the design matrix shows the relationship between the functional domain and the
physical domain.
The developed software modules consist of the main module, the data
B A
Figure 2.D.1. Design matrix of FR3

management module, the module for modeling, the module for the drop test
analysis and the module for automatic analysis and design. The main module
controls the graphic user interface for input and the overall design process. The
data management module has the function of managing the data for the system. It
handles the interface with external systems, the files for material properties and the
database for standard orthogonal arrays. The module for modelling generates the
input data for analysis, which are shape sizes and the input file for LS/INGRID
that automatically generates meshes for the finite element analysis. The module
for the drop test generates an input file for the analysis system LS/DYNA3D and
executes the system. The module for analysis and design performs the analysis,
analyzes the results and proposes a new design.
Each module is defined as an object. An object consists of the functions in the
row of the design matrix and the attributes of the column. For example, FR3 in
Figure 2.D.1 is defined by object A, which includes object B for inner shapes and
object C for external shapes. Object B is composed of four objects for four
positions from the user input and one object for common data. In the same manner,
object C has five objects.
Establishment of Interfaces and Coding (Steps 6 and 7)

Classes are defined by the set of objects as illustrated in Figure 2.D.2. The classes
in the low levels are not presented in Figure 2.D.2. The class “PackDesign” is
defined from the relation of the four classes. The class “Option” has a function of
input for initial definition. It is automatically executed when the system starts.
The class “BasicModel” handles the data for the monitor and material properties
CPackDesign
int designStep;
……
FileSave();
…….
COption CBasicModel CCushionModel CIterationAnal
DirInfo optionDir; DesignVarInfo bModel; cushionInfo cushion; iterationInfo opt;

…… …… …… ……
SetIngridDir(); CalculatePosition(); MakeIngridFiles(); SelectFactors();

……….. ……….. ……….. ………..
Figure 2.D.2. Class diagram of the “PackDesign” software system

Figure 2.D.3. The output screen of the “PackDesign”
and the class “CushionModel” handles the shapes of the cushioning materials.
They receive data from a user, perform coordinate transformation and generate the
input file for the finite element analysis. Finally, the class “IterationAnal” selects
an orthogonal array, performs the drop tests according to the orthogonal array and
analyzes the results. If the aforementioned basic model is varied, derived classes
can be made by inheritance from the classes “BasicModel” and “CushionModel.”
Using the above process, a software system is coded. The overall menus are
illustrated in Figure 2.D.3. The left hand side is for input and the right hand side is
for displaying results. The execution of the system is classified into two modes.
One is analysis with given parameters and the other is a design process for
multiple analyses with changing variables.
2.D.3 Summary
The axiomatic approach for software design is demonstrated. Software is

developed based on the V-model which is related to the object oriented
programming concept. The developed software can be easily used by new
engineers. The design results are stored in the database and exploited for later use.
References
Albano LD, Suh NP (1992) Axiomatic Approach to Structural Design. Research in
Engineering Design l(4):171í183
Albano LD, Suh NP (1993) The Information Axiom and Its Implication. DE–Vol
66, Intelligent Concurrent Design: Fundamentals, Methodology, Modeling and
Practice, ASME
ANSYS Inc. (2004) ANSYS User’s Manual Version 8.1. Canonsburg, PA
Do SH, Park GJ (2001) Application of Design Axioms for Glass Bulb Design and
Software Development for Design Automation. Journal of Mechanical Design,
Transactions of the ASME 123(3):322í329
Do SH, Suh NP (2000) Axiomatic Design of Software Systems. CIRP Annals
49(1):95í100
Electronic Data Systems Co., (1993) User Function Programming Manual. Livonia,
MI
El-Haik B, Yang K (1999) The Components of Complexity in Engineering Design.
IIE Transactions 31:925í934
Frey DD, Jahangir E, Engelhardt F (2000) Computing the Information Content of
Decoupled Designs. Research in Engineering Design 12:90í102
Gebala DA, Suh NP (1992) An Application of Axiomatic Design. Research in
Engineering Design 3:149í162
Heller D, Ferguson PM (1994) Motif Programming Manual. O’Reilly &
Associates, Setastopol, CA
Hewlett–Packard Co. (1991) Starbase Graphics Techniques. TX
Hwang KH, Lee KW, Park GJ (2001) Robust Optimization of an Automobile
Rearview Mirror for Vibration Reduction. Structural and Multidisciplinary
Kim SJ, Suh NP, Kim SK (1991) Design of Software Systems Based on Axiomatic
Design. Annals of CIRP, 40(1): 165í170
Lee J, Cho K, Lee K (1994) A New Control System of a Household Refrigerator-
Freezer. International Refrigeration Conference, Purdue University, IN
Livermore Software Technology Co. (1998) LS/INGRID Manual Ver. 3.5.
Livermore, CA
Livermore Software Technology Co. (1998) LS/DYNA Theory Manual.
Livermore, CA
Livermore Software Technology Co. (1999) LS/DYNA User’s Manual Ver. 950.
Livermore, CA
Mood AM, Graybill FA, Boes DC (1963) Introduction to the Theory of Statistics.
3rd ed. McGraw–Hill, New York
NSF (1998) Axiomatic Design Workshop for Professors. MIT, Cambridge, MA
Oracle Co., (1990) ORACLE 3GL Programmers Guide. Redwood Shore, CA
Park GJ, Do SH, Lee JW, (1995) Construction of Automatic Design Systems.
Final Report, Samsung Corning Co. (in Korean)
Rinderle JR, Suh NP (1982) Measures of Functional Coupling in Design.
Transactions of ASME, Journal of Engineering for Industry 104:383í388
Shin GS, Park GJ (2004) Supplementary Beam Adjuster for a Laser Device. Patent
No. 10-045398, Korea
Shin MK, Hong SW, Park GJ (2001) Axiomatic Design of the Motor-Driven
Tilt/Telescopic Steering System for Safety and Vibration. Proceedings of the
Institution of Mechanical Engineers, Part D, Journal of Automobile Engineering
215(2):179í187
Suh NP (1990) The Principles of Design. Oxford University Press, New York
Suh NP (1995) Axiomatic Design of Mechanical Systems. Special 50th

Anniversary, Combined Issue of the Journal of Mechanical Design and the
Journal of Vibration and Acoustics, ASME 17:1í10
Suh NP (1995) Designing-in of Quality through Axiomatic Design. IEEE
Transactions on Reliability 449(2):256í264
Suh NP (1995) Design and Operation of Large Systems. Journal of Manufacturing
Systems 14(3):203í213
Suh NP (1998) Axiomatic Design Theory for Systems. Research in Engineering
Design 10:189í209
Suh NP (1999) A Theory of Complexity, Periodicity and the Design Axioms.
Research in Engineering Design 11:116í131
Press, New York
Suh NP, Bell AC, Gossard DC (1978) On an Axiomatic Approach to
Manufacturing and Manufacturing Systems. Journal of Engineering for Industry
100(2):127í130
Suh NP, Sekimoto S. (1990) Design of Thinking Design Machine. Annals of the
CIRP 39(1):145í148
Suh NP, Wilson DR, Tice WW, Yasuhara M, Bell AC (1979) Application of
Axiomatic Design Techniques to Manufacturing. Winter Annual Meeting,
ASME 79-WA/Prod-25, New York, December 2í7
Taguchi G (1987) Systems of Experimental Design. Kraus International
Publications, New York
Yi JW, Park GJ (2005) Development of a Design System for EPS Cushioning
Package of a Monitor Using Axiomatic Design. Advances in Engineering
Software 36:273í284
3 Design Optimization
3.1 Introduction
Optimization was originally created by mathematicians out of curiosity. Basically,
a given mathematical function is minimized or maximized. Mathematical
optimization methods have been developed to solve a variety of different
optimization problem formulations. For example, finding the firing angle of a
cannon can be viewed as an optimization problem to obtain the maximum distance.
The flying distance can be expressed by the function of the angle. In optimization,
optimum values of the variables are found to minimize or maximize a function that
is expressed by the variables.
Optimization originates from the variational method, which finds an optimum
function for a given objective functional. In the variational method, a functional is
defined by the total potential energy of a system, and minimization of the
functional results in the governing equation of the system. When there are
constraints, they are multiplied by Lagrange multipliers and added to the objective
functional as penalties. This concept is utilized by modern optimization as well.
The Brachistochrone problem is the well known problem of minimizing a
functional. It calculates the shortest path where we move under gravity. Also, it
can be proved that the shortest path between two points is a straight line. Algebra
and variational calculus are adopted to solve such minimization or maximization
problems.
The variational methods in the early stage are mainly utilized to solve
mechanics problems using energy methods. Modern optimization methods have
been established as engineers apply them to design. In optimization, algebraic
functions are utilized instead of functionals and constraints are emphasized due to
the characteristics of design. Optimization theory is used to solve problems
formulated as follows:
Find b Rn (3.1a)
subject to hi (b) 0 , i 1, , l (3.1c)

g j (b) d 0 , j 1, , m (3.1d)
bL d b d bU (3.1e)
where b is a design variable vector with n components, f is the objective function,

hi is the ith equality constraint, g j is the jth inequality constraint, l is the number
of equality constraints, m is the number of inequality constraints, and b L and b U
are the lower and upper bound vectors for b, respectively.
The concept of optimization matches that of the engineering design process,
which minimizes or maximizes a certain engineering performance. When
optimization is employed in an engineering design, it is called optimal design or
design optimization. In optimal design, the objective function in Equation 3.1b is
the objective engineering performance and Equations 3.1c and 3.1d include
various design conditions. Therefore, an engineering design problem should be
formulated as the above mathematical expressions. As mentioned earlier,
engineering design is classified into conceptual design and detailed design.
Optimal design is generally utilized in the detailed design stage in that detailed
values are determined for design variables.
Because the conceptual design determines the overall configuration and
selection of parts, it is difficult to formulate the design problem in the above
equations. Therefore, optimal design is generally used to determine specific values
after the overall product concept is developed. If an optimization problem is
defined, specific algorithms are employed to solve the formulated problems.
Figure 3.1 presents the conventional method and Figure 3.2 shows the
optimization method. Generally, the two methods are similar except for the
Definition of the design problem
Initial design
Performance analysis
Yes
Satisfied? End
No
Design change based on
experience and intuition
Figure 3.1. Conventional design process

Design Optimization 109
Definition of the design problem

Design variables
Objective function
Constraints
Initial design
Application of an optimization algorithm
Yes
Are convergence criteria satisfied? End
No
Figure 3.2. Optimization process
decision making processes, where the decisions are made by either the designer or
the computer.
Optimization gives an excellent solution when the formulation is accurate. In
other words, a mathematical optimum solution is obtained. However, when the
formulation is not appropriate, the optimum solution can be obsolete in real design
terms because the solution is obtained only from the formulated problem.
Meanwhile, a mathematically excellent solution may satisfy constraints at the
utmost limits. In this case, if tolerances or slight changes exist in design variables,
the constraints can be easily violated. Although many methods are proposed to
overcome these problems, optimization still has numerical difficulties in solving
mathematically sensitive problems.
Optimal design is the terminology for a technique to solve the problems
formulated as Equation 3.1. As mentioned earlier, it can be efficiently utilized in
detailed design. Also, it should be remembered that the impact of detailed design
is generally smaller than that of conceptual design. Therefore, optimization should
be carefully applied to appropriate cases.
3.2 Optimization Problems in Engineering

The formulation in Equation 3.1 should be developed to use an optimization theory
in engineering design. The formulation process generally proceeds as follows:
Step 1. Design variables in Equation 3.1a are defined. Design variables are
parameters that are to be determined in the design process. In structural
design, they can be section properties, thickness and shapes, etc.
Step 2. The objective function in Equation 3.1b is defined. The objective
function is a performance measure that is to be improved (e.g., min or
max) in design. In structural design, the objective function can be the
weight of a structure, a specific response, etc.
Step 3. The constraints in Equations 3.1c and 3.1d are defined. Generally,
design is performed based on some design specifications and conditions.
They should be transformed to constraint functions.
The problem formulation in Equation 3.1 is referred to as the standard form for
constrained optimization. Problems with different forms can be modified to the
standard one. For example, a maximizing problem is transformed to a minimizing
problem as follows:
Maximize F (b) minimize( F (b)) { minimize f (b) (3.2)
where
F (b) { f (b) (3.3)
When a constraint in Equation 3.1d has the greater than type inequality, it can be
transformed as follows:
g j (b) t 0 o g j (b) d 0 (3.4)
All the optimization problems can be transformed to the standard form in

Equation 3.1. From now on, it is assumed that all the optimization problems have
the standard form. Three examples will demonstrate how to formulate
optimization problems.
Example 3.1 [Optimization of a Symmetric Three Bar Truss] (Arora

2004)
The three bar truss in Figure 3.3 is to be optimally designed. It is a simple,
statically indeterminate problem. It has long been a standard problem for the
evaluation of new optimization algorithms. When an external load P is imposed at
the tip, the weight is to be minimized, while the stresses are to be less than the
yield stress and the displacement at the tip is less than the allowable value. Also,
the symmetry of the structure should be maintained.
Solution
The optimization formulation is performed according to the previous sequence.
Step 1. Design variables are defined by the cross sectional area of each
member. They are b1 ,b2 and b3 . From the symmetry condition, b1 { b3 .
l l
ཛ ཛྷ ཝ
l 1 2 3
x
T
y
P
Figure 3.3. Three bar truss
Step 2. The objective function is defined by the weight of the structure. The
objective function is
f (b) Ugl (2 2b1 b2 ) (3.5)
where U is the density and g is the gravitational acceleration. It is noted

that the objective function should be a function of the design variables.
Step 3. Constraint functions are defined. From the condition that the stress
on the ith member (V i , i 1, 2, 3) should be less than the yield stress (V all ) ,
the following three stress constraints are defined:
gi | V i | V all d 0 , i 1, 2, 3 (3.6)
From the condition that the lateral (G x ) and vertical (G y ) displacements

should be within the allowable ranges (G x, all , G y , all ) , the following
constraints are defined:
g4 G x G x, all d 0 (3.7)
g5 G y G y , all d 0 (3.8)
The optimization problem is formulated as above. The constraints should be

functions of design variables. They can be modified by using the equilibrium and
compatibility equations and Hook’s law. The displacements G x and G y at the tip
are derived as follows:
2lPx
Gx (3.9)
b1 E
2lPy
Gy (3.10)
(b1 2b2 ) E
where E is Young’s modulus, and Px ( P cos T ) and Py ( P sin T ) are the lateral
and vertical components of the external load P, respectively. The stresses at the
members are as follows:
1 §¨ Px Py ·
¸
V1 (3.11)
¨
2 © b1 b1 2b2 ¸
¹
2 Py
V2 (3.12)
b1 2b2
1 §¨ Py P ·
V3 x¸ (3.13)
2 ¨© b1 2b2 b1 ¸¹
If Equations 3.9í3.13 are substituted into Equations 3.6í3.8, the constraints

become functions of design variables. Also, the lower and upper bounds of the
design variables are added to the constraint set. Optimization is a process to find
an optimum solution from the formulated problem.
Example 3.2 [Minimum Weight Tubular Column Design] (Arora 2004)

Optimization is applied to the design of the straight column illustrated in Figure
3.4. The design goal is minimization of the weight by determining the length l
while constraints are imposed on the buckling and stresses.
Solution
An optimization problem is formulated according to the sequence as follows:
Step 1. Design variables are defined. The average of the inner and outer
radii R and the thickness t are selected as design variables. The design
variables may alternatively be defined as the inner and outer radii.
Step 2. The objective function f is defined by the total weight of the structure.
f Ug (lA) 2 UglSRt (3.14)
where A is the area of the cross section and has 2SRt.

Step 3. Constraints are defined. Stress constraints are defined not to exceed
the allowable stress as follows:
2R
Figure 3.4. A tubular column
P P
d V all (3.15)
A 2SRt
Also, the following buckling constraint is defined from the condition that
the compressive stress should be within the allowable range:
S 2 EI S 3 ER 3 t
Pd (3.16)
4l 2 4l 2
where I is moment of inertia with I SR 3t and E is Young’s modulus.

Step 4. The upper and lower bounds are defined for the design variables,
R min d R d R max (3.17)
t min d t d t max (3.18)
Example 3.3 [Spring Design Problem] (Arora 2004)

Another popular standard problem is the spring design problem. The spring in
Figure 3.5 is a simple one. It is frequently utilized for the evaluation of
optimization algorithms due to its high nonlinearity. First, the design specification
is investigated. Variables are defined as d wire diameter (mm), D mean coil
diameter (mm) and N number of active coils. The following equations are in the
design specification:
D
P P
d G
Figure 3.5. Spring
Load-displacement equation: P KG (3.19)
d 4G
Spring constant: K (3.20)
8D 3 N
8kPD
Shear stress: W (3.21)
Sd 3
4 D d 0.615d
Wahl stress concentration factor: k (3.22)
4 D 4d D
d G
Frequency of surge waves: Z 2
(3.23)
2SD N 2U
where G is the shear modulus, U is the density of the material and k is the Wahl
stress concentration factor, which is determined from the experiments. Now
formulate an optimization problem.
Solution
An optimization problem is formulated as follows:
Step 1. Design variables are defined as d wire diameter, D mean coil
diameter and N number of active coils.
Step 2. The objective function is the mass as
1
f (b) NS 2 Dd 2 U (3.24)
4
Step 3. Constraints are defined by using the above design equations. The
displacement G should be larger than a certain value ' to store enough
energy. From Equations 3.19 and 3.20, the following constraint is defined:
P 8 PD 3 N
' G ' ' d0 (3.25)
K d 4G
The shear stress W should be less than the allowable value W all . From
Equations 3.21 and 3.22,
8PD § 4 D d 0.615d ·
3 ¨
¸ W all d 0 (3.26)
Sd © 4 D 4d D ¹
To avoid resonance, the natural frequency should be larger than a certain

value Z 0 . From Equation 3.23,
d G
Z0 Z Z0 2
d0 (3.27)
2SD N 2U
A bound is imposed on the outer diameter as
D d Dlim d 0 (3.28)
Step 4. The lower and upper bounds on design variables are defined as
d min d d d d max (3.29)
D min d D d D max (3.30)
N min d N d N max (3.31)
As shown from the examples, problem formulation is the first and the most
important step in engineering optimization. The sequence of the formulation is the
definition of design variables, the objective function and constraints. The
objective function and constraints should be expressed by functions of design
variables. Although small-scale problems are demonstrated in the examples, large-
scale problems also follow the same procedure.
3.3 Analytic Optimization Methods

As mentioned in Section 3.1, the optimum solution is determined as the solution of
the formulated problem. Optimization is a process to find the solution. If all the
functions in Equation 3.1 are mathematically expressed, then a mathematical
solution can also be obtained.
3.3.1 Graphical Optimization
Graphical optimization finds an optimum solution by drawing the objective and the
constraint functions on a plane. Since the graphics can be drawn on a two-
dimensional plane, problems with two design variables can only be solved.
Therefore, it is used for the explanation of an optimization method. The optimum
solution is obtained by the following sequence:
Step 1. Draw constraints on a two-dimensional plane and find the feasible
region.
Step 2. Draw the iso-objective contours.
Step 3. Find a design point that makes the objective function a minimum in
the feasible region.
Example 3.4 [Graphical Optimization]

Using graphical optimization, find the optimum point of the following problem:
Find b [b1 b 2 ]T (3.32a)
to minimize f (b) (b1 1.5) 2 (b2 1.5) 2 (3.32b)
subject to g 1 (b) b1 b2 2 d 0 (3.32c)
g 2 (b) b1 d 0 (3.32d)
g 3 (b) b2 d 0 (3.32e)
Solution
The graph of each function is illustrated in Figure 3.6. The feasible region is
represented by three constraints. The concentric circles represent iso-objective
contours. In the feasible region, the optimum point is b [1 1]T . It is noted that
the constraint g1 is the most important for the optimum solution and it is a tangent
line for an iso-objective contour. At the optimum point, g1 is active.
3.3.2 Unconstrained Optimization
Mathematical optimization is explained first for optimization problems without

constraints. These are called unconstrained problems. Suppose we have the
following problem:
Minimize f (b) (3.33)

b2
f 0.75 Iso-objective contour
2 Optimum
f 0.5
(1.5, 1.5) b* (1, 1)
*
f (b ) 0.5
1
Feasible g1 b1 b2 2 0
region
b1
1 2
Figure 3.6. An example of graphic optimization
The increment of the function f can be obtained from the linear expansion of
the Taylor series
Gf T f (b * )(b b * ) (3.34)
The condition for a local minimum is expressed as
f (b * ) 0 (3.35)
That is, the gradient of f is the zero vector at b*. In the partial differentiation form,
wf (b * )
0, i 1, , n (3.36)
wbi
The point to satisfy Equation 3.36 is called an extreme point or a stationary

point. If we consider up to the second-order expansion term in the Taylor series in
Equation 3A.9, the following inequality should be satisfied to have a positive value
for Gf :
(b b * ) T H (b b * ) ! 0 (3.37)
Equation 3.37 is satisfied when the Hessian matrix H is positive definite.

Therefore, a theorem is derived for the necessary and sufficient conditions for a
local minimum.
Theorem 3.1 [Necessary and Sufficient Conditions for the

Local Minimum]
Necessary condition: In a single variable problem, if f(b) has a local
minimum at b* , then f c(b* ) 0 . In a multivariable problem, if f(b) has a
local minimum at b* , then
f (b * ) 0 (3.38)
Sufficient condition: If the Hessian matrix H (b* ) is positive definite at the

stationary point b * , then b* is a local minimum point for the function f(b).
In a single variable problem,
f cc(b * ) ! 0 (3.39)
To obtain a local minimum for a function f(b): (1) We obtain a solution of the
simultaneous equation of Equation 3.35 or 3.36. (2) We check whether the
Hessian matrix is positive definite at the solution of the first step. When the
Hessian matrix is positive semidefinite, we need to check for a higher order.
Details are given in the references (Arora 2004, Rao 1996).
Example 3.5 [Necessary and Sufficient Condition for a Local

Minimum]
Find a local minimum for the following function:
f (b) b12 5b22 4b1b2 2b2 2 (3.40)
Solution
Necessary condition:
wf
2b1 4b2 0 (3.41)
wb1
wf
10b2 4b1 2 0 (3.42)
wb2
The solution satisfying Equations 3.41 and 3.42 is

b1 2, b2 1 (3.43)
Sufficient condition:
ª w2 f w2 f º
« 2 »
« wb1 wb1 wb2 » ª 2 4º
H « 4 10 » (3.44)
« w2 f w2 f » ¬ ¼
« »
«¬ wb2 wb1 wb22 »¼
b T Hb 2(b1 2b2 ) 2 2b22 t 0 (when b1 b2 0 , b T Hb 0 ) (3.45)
The Hessian matrix in Equation 3.44 is positive definite. Thus, the solution in
Equation 3.43 satisfies the necessary and sufficient conditions for the local
minimum.
3.3.3 Constrained Optimization
Generally, most optimization problems have constraints. Constrained problems

are formulated as Equation 3.1. An optimum point for a constrained problem is
mathematically obtained from the following theorem:
Theorem 3.2 [Karush–Kuhn–Tucker (KKT) Condition]
In the constrained optimization problem in Equation 3.1, let b* be a

regular point in the feasible region and a local minimum point of the
objective function f(b). Define the Lagrangian function for the problem as
l m
L(b, v, u, s) f (b) ¦ vi hi (b) ¦ u j ( g j (b) s 2j ) (3.46)
i 1 j 1
Then there exists Lagrange multiplier vectors v * R l and u* R m such

that
wL wf l whi m wg j
{ ¦ v*i ¦ u *j 0, k 1, , n (3.47a)
wbk wbk i 1 wbk j 1 wbk
hi (b* ) 0, i 1, , l (3.47b)
g j (b* ) s 2j 0, j 1, , m (3.47c)
u *j s j 0, j 1, , m (3.47d)
u *j t 0, j 1, , m (3.47e)
where all derivatives are evaluated at b* .

A local minimum is a root of the simultaneous equations of Equation 3.47. At

a local minimum, the gradient vector of the Lagrangian function (Equation 3.47a)
becomes the zero vector. It is noted that constraints are added to the objective
function by multiplying each by a unique Lagrange multiplier. As shown in
Equation 3.47e, the condition that the Lagrange multipliers for inequality
constraints must be nonnegative makes the mathematical calculation difficult.
Readers are referred to the references (Arora 2004, Rao 1996) for other aspects
such as the KKT condition at nonregular points, sufficient conditions using the
Hessian matrix, higher order necessary conditions, etc. Theorem 3.A.4 introduces
the global minimum for a convex function. There is a theorem for the sufficient
condition as follows:
Theorem 3.3 [Sufficient Condition for a Convex Problem]

If a function f(b) is an objective function defined on a convex feasible set
and a convex function, the KKT condition is a necessary and sufficient
condition for a global minimum.
Example 3.6 [KKT Necessary Condition]

Find a point that satisfies the KKT necessary condition from the following
problem:
Minimize f (b) (b1 2b2 ) 2 (b2 1) 2 1 (3.48a)
subject to b1 b2 0 (3.48b)
b2 3b1 t 0 (3.48c)
Solution
Since the problem has constraints, the Lagrangian function is defined as
L(b, v, u, s) (b1 2b2 ) 2 (b2 1) 2 1 Q (b1 b2 ) u (b2 3b1 s 2 )

(3.49)
The values of b1 , b2 , u, v and s are calculated to satisfy the following KKT

condition:
wL
2b1 4b2 v 3u 0 (3.50a)
wb1
wL
10b2 4b1 2 v u 0 (3.50b)
wb2
b1 b2 0 (3.50c)
b2 3b1 s 2 0 (3.50d)
us 0 (3.50e)
ut0 (3.50f)
Two cases are made from Equation 3.50e.

(1) When u 0
2b1 4b2 v 0 (3.51a)
10b2 4b1 2 v 0 (3.51b)
b1 b2 0 (3.51c)
b2 3b1 s 2 0 (3.51d)
The solution of Equation 3.51 is b1 0.1, b2 0.1, v 0.6 and s 2 0.4 .

(2) When s 0
2b1 4b2 v 3u 0 (3.52a)
10b2 4b1 2 v u 0 (3.52b)
b1 b2 0 (3.52c)
b2 3b1 0 (3.52d)
The solution of Equation 3.52 is b1 0, b2 0, v 1.5 and u 0.5. If

we consider the condition of Equation 3.50f, the solution of case (2) is not
appropriate. Therefore, the solution is case (1) as follows:
ª b1 º ª 0.1º
«b » « 0.1 »
« 2» « »
«v» « 0.6 » (3.53)
« 2» « »
«s » « 0.4 »
«¬ u »¼ «¬ 0 »¼
3.4 General Concepts of Numerical Methods in

Optimization
An optimum solution can be obtained from mathematical theorems. In reality, it is
quite difficult to find an optimum solution from mathematical calculation. A
solution of an unconstrained problem is found from simultaneous equations where
the gradient vector of the objective function is the zero vector. Since most
problems are nonlinear, it is often very difficult to solve a simultaneous equation.
Besides, the solution process for a constrained problem should evaluate many
cases for an inequality constraint whether or not the constraint is zero and whether
or not the Lagrange multiplier is nonnegative. If the number of constraints
increases, the problem becomes considerably more complicated. There are many
constraints in engineering optimization, so that it is almost impossible to solve
engineering optimization with mathematical calculation.
The above difficulties can be overcome by numerical methods using computers.
Root finding algorithms can be employed to find an optimum for an unconstrained
problem. However, since these algorithms are not efficient, various numerical
methods have been developed specifically for optimization applications. In
general, those numerical methods are utilized in optimization with computers.
Numerical methods have been developed according to the characteristics of the
problems involved. As mentioned earlier, optimization problems are classified by
the existence of constraints, the characteristics of design variables, etc. For
example, optimization problems are characterized as unconstrained or constrained
problems, continuous or discrete design variable problems, linear or nonlinear
problems, single or multiple objective problems, etc.
In numerical methods, a specific algorithm is developed for each problem.
When all the objective and constraints are linear functions, a group of algorithms
called linear programming (LP) methods is utilized. In the same manner, various
nonlinear programming (NLP) methods are employed for problems with nonlinear
functions. Popular methods in engineering have been introduced.
In engineering optimization, it is assumed that there are constraints, that design
variables are continuous and that the functions are nonlinear. Nonlinear
programming (NLP) methods are the most frequently used numerical methods.
The methods are basically different from mathematical methods. While
mathematical methods find an optimum solution from simultaneous equations
made from the optimality condition, numerical methods start from an initial point
and find an optimum in an iterative fashion. At each iteration, a new design is
found. The new design point is considered as an optimum when the convergence
criteria are satisfied. The design point is improved at each iteration using the
following equation:
b ( k 1) b ( k ) 'b ( k ) , k 0, 1, 2, (3.54)
where k is the iteration number. At the kth iteration, the change of the design
variable vector 'b ( k ) is evaluated based on the characteristics of the algorithm and
added to the current design b ( k ) , and a new design b ( k 1) is obtained. The change
'b ( k ) consists of two parts as follows:
'b ( k ) D k d (k ) (3.55)
where d ( k ) is the direction vector to move b ( k ) in the design space and D k is a

scalar stepsize in the d ( k ) direction.
The schematic flow of the design change is illustrated in Figure 3.7. In the kth
iteration of Figure 3.7, the direction vector d (k ) is obtained first and the stepsize
D k is determined. The method to evaluate d (k ) and D k is dependent on the
algorithm utilized.
The overall steps are as follows:
Step 1. Estimate an initial design b ( 0) and set the iteration number k 0 .

Step 2. Determine the direction vector d ( k ) . The method in this process
varies according to the characteristics of the problem and the algorithm
utilized.
Step 3. Check the convergence. The convergence criteria depend on the
algorithm utilized. If the convergence criteria are satisfied, terminate the
process. Otherwise, go to the next step.
Step 4. Determine the stepsize D k . Different methods can be used in
different algorithms.
Step 5. A new design is found from Equations 3.54 and 3.55. Change the
iteration number k k 1 and go to Step 2.
The above steps are general ones and some methods can use different
d ( k 1)
D k d (k ) d (k )
D k 1d ( k 1) (k )
b ( k 1)
b
d ( k 1)
b ( k 1)
Figure 3.7. Change of the design point during iteration

procedures. The direction vector d (k ) and the stepsize D k are determined to

reduce the objective function f(b) when the current design b ( k ) is not an optimum.
The condition can be expressed as
f (b ( k 1) ) f (b ( k ) D k d ( k ) ) f (b ( k ) ) (3.56)
where constraints are not considered. When there are constraints, active
constraints are modified to the penalty function and added to the objective function.
Then the augmented objective function is treated as an unconstrained objective
function. This aspect will be explained later. The condition in Equation 3.56 is
called a descent condition. If we expand it with a Taylor series expansion
T
f (b ( k 1) ) # f (b ( k ) ) D k c ( k ) d ( k ) f (b ( k ) ) (3.57)
where c ( k ) f (b ( k ) ) . Since D k is a positive number, the vector d (k ) to make

T
c ( k ) d ( k ) 0 is the descent direction. Theoretically, when d ( k ) f (b ( k ) ) , the
reduction of the function is the maximum in the first-order and is called the
steepest descent direction.
The performance of an algorithm is evaluated by the convergence property and
the rate of convergence. The convergence property is whether or not any initial
design can converge to an optimum point. There is frequent confusion between
the global minimum and the global convergence. The global minimum indicates
the best point in the feasible region. However, the global convergence means that
any initial design can converge to a local minimum point. In general, optimization
algorithms do not find the global minimum. Instead, many excellent algorithms
have global convergence. When the feasible region is closed and bounded, an
algorithm with the global convergence should have the following condition: There
should be a descent function. The descent function is reduced as the iteration
proceeds. In unconstrained problems, the objective function is usually utilized as
the descent function. In constrained problems, a compounded function of the
objective and penalty functions can be used as the descent function. It is a
condition that the iteration is directed to an optimum point.
The rate of convergence is the speed of convergence. If b* is an optimum
point, the rate is
|| b ( k 1) b* ||
d K , K t 0, p t 1 (3.58)
|| b ( k ) b* || p
When p 1 and 0 d K d 1 , the rate is linear. If p 2, then the rate is

quadratic. In an engineering algorithm, it may not be easy to determine the indices
p and K rigorously. In this case, the rate can be replaced by the number of
iterations and the number of function calculations. The number of function
calculations will be explained later. Mathematically excellent algorithms often do

not exhibit good performance in application. The reason is either that the reality
can be different from the theory or the coding may not be good.
When we use a numerical method, we need sensitivity information. The
sensitivity information is the (partial) derivative of the objective and constraint
functions with respect to the design variables. The sensitivity shows how sensitive
a function is with respect to a design variable and it is expressed by the gradient
vector. Sensitivity analysis is the process that calculates sensitivity information.
Sensitivity information can be evaluated in two ways. One is the analytic method,
which directly uses mathematical derivative functions in computer programming.
The other is the finite difference method, which uses the following equation:
wf f (b1 , , bi 'bi , , bn ) f (b1 , , bi , , bn )

(3.59)
wbi 'bi
where 'bi is a small perturbation of design variable bi . There is no general

theory on how large it should be. It is well known that the analytic method is
accurate, efficient and difficult to code, and that the finite difference method is the
opposite.
3.5 Linear Programming
3.5.1 Standardization of Linear Programming
The linear programming problem is an optimization problem where the objective

and constraint functions are linear functions in Equation 3.1. It is also called a
linear optimization problem, a linear programming problem or LP. It frequently
appears in industrial engineering. It is rarely used in engineering design, but linear
programming problems are found in the iterative process where a nonlinear
problem is linearized. If one or more functions of the objective and constraint
functions are nonlinear, the optimization problem is called a nonlinear
programming problem and different algorithms are employed. The linear
programming method is a numerical method to find an optimum of the LP.
A linear programming problem is standardized as
Find b [b1 b2 bn ]T R n (3.60a)

n
to minimize f c1b1 c 2 b2 c n bn ¦ c i bi (3.60b)
i 1
n
subject to a11b1 a12 b2 a1n bn ¦ a1i bi e1 t 0
i 1
n
a 21b1 a 22 b2 a 2 n bn ¦ a 2i bi e2 t 0
i 1

(3.60c)

n
a m1b1 a m 2 b2 a mn bn ¦ a mi bi em t 0
i 1
bi t 0, i 1, , n (3.60d)
If a linear programming problem can be transformed to the above standard

problem, we can develop algorithms to solve the problem in Equation 3.60. The
objective function in Equation 3.60b can include any form of a linear function.
However, constraint functions can have various forms that are different from
Equations 3.60c and 3.60d. ei can be negative, inequality constraints can exist
and the lower or upper bound of bi can be negative.
Now, it is shown that all cases can be transformed to the standard form in
Equation 3.60.
(1) When ei is less than zero
If ei is positive, the condition matches with Equation 3.60c. When it is
negative, í1 is multiplied to both sides of Equation 3.60c as follows:
(a i1b1 a i 2 b2 a in bn ) e i t 0 (3.61)
Equation 3.61 shows that the right hand side can always be zero or positive.
(2) When Equation 3.60c is an inequality constraint
When Equation 3.60c is a less than type (<) function, a slack variable si is
added as follows:
a i1b1 a i 2 b2 a in bn s i ei (3.62a)
si t 0 (3.62b)
If it is greater than type (>), a surplus variable si is added as
a i1b1 a i 2 b2 a in bn s i ei (3.62c)
si t 0 (3.62d)
An inequality constraint can be transformed into the standard form through

the addition of a variable.
(3) Lower and upper bounds of design variables
A design variable should be zero or positive. In real problems, a design

variable can have any real value. In other words, the lower or upper bound
can have a real number. It also can be modified to fit into the standard
problem. A constraint bi t z (real number) is transformed to bi {
bi z t 0 by defining a new variable bi . When a constraint has the form
of bi d z , the new design variable is defined by bi { (bi z ) t 0 .
Now the new design variables are appropriate for the standard problem.
(4) Free variable
Some problems have a free variable that does not have bounds. The free
variable is replaced by two virtual variables as follows:
bi bi bi (3.63)
where
bi t 0, bi t 0 (3.64)
At the optimum, if bi t bi , then bi t 0 and if bi d bi , then bi d 0 .
Any type of problem can be transformed to a standard problem. LP algorithms

are developed to solve the standard problem. Thus, only a standard problem is
considered here. Using the matrix notation, the standard problem can be expressed
as
Find b (3.65a)
to minimize f cT b (3.65b)
subject to Ab e (3.65c)
bt0 (3.65d)
Example 3.7 [Standard Linear Programming Problem]

Convert the following problem to a standard linear programming problem:
Minimize f (b) b1 2b2 (3.66a)
subject to 2b1 3b2 t 8 (3.66b)
5b1 2b2 d 1 (3.66c)
b2 t 0 (3.66d)
Solution
Step 1. Since b1 is a free variable, virtual variables are defined as follows:
b1 b1 b1 (3.67)
b1 t 0, b1 t 0 (3.68)
Step 2. With the virtual variables, the problem is rewritten as follows:
Minimize f (b) b1 b1 2b2 (3.69a)
subject to 2(b1 b1 ) 3b2 t 8 (3.69b)
5(b1 b1 ) 2b2 d 1 (3.69c)
b1 t 0, b1 t 0, b2 t 0 (3.69d)
Step 3. If slack and surplus variables are added, Equation 3.69 is converted
to the standard form as follows:
Minimize f (b) b1 b1 2b2 (3.70a)
subject to 2(b1 b1 ) 3b2 s1 8 (3.70b)
5(b1 b1 ) 2b2 s2 1 (3.70c)
b1 t 0, b1 t 0, b2 t 0, s1 t 0, s 2 t 0 (3.70d)
Step 4. With b1 z1 , b1 z 2 , b2 z 3 , s1 z 4 , and s 2 z 5 , Equation

3.70 is rewritten as
Minimize f (z ) z1 z 2 2 z 3 (3.71a)
subject to 2 z1 2 z 2 3 z3 z 4 8 (3.71b)
5 z1 5 z 2 2 z3 z5 1 (3.71c)
z i t 0, i 1, , 5 (3.71d)
3.5.2 Basic Concepts of Linear Programming
Since all the functions are linear in a linear programming problem, the feasible set
is convex from Theorem 3.A.3. Therefore, a local minimum is the global
minimum from Theorem 3.A.4. In Equation 3.60c, the number of linearly
independent constraints m must be less than the number of design variables n. If
m ! n , there is no solution satisfying constraints, and if m n only one solution
exists for the constraints and this is the optimum solution. When there are
inequality constraints, the optimum solution resides on the boundary of the feasible
set. If the optimum solution is inside the boundaries, the problem is the same as an
unconstrained problem. If the objective function is linear without constraints, the
problem is unbounded. Therefore, the optimum solution should be on a boundary.
From the condition m n, the number of solutions satisfying Equation 3.60c is
infinite and the one with the minimum objective function is the optimum out of the
infinite solutions. The above concept can be assured from the following example:
Example 3.8 [Profit Maximization]

Find the optimum by the graphical method for the following profit maximization
problem:
Maximize f (b) 400b1 600b2 (3.72a)
subject to b1 b2 d 16 (3.72b)
1 1
b1 b2 d 1 (3.72c)
28 14
b2
25 G
b1 b2
20 1
14 24
F
b1 b2 16
15
E
f= D
–7 I
10 36
0
b1 b2
f= 1
–4 28 14
5 800 C
f=
–2 f=
400 –8
800
B
J H
b1
A
5 10 15 20 25 30
Figure 3.8. Graphical presentation of Example 3.8
1 1
b1 b2 d 1 (3.72d)
14 24
b1 , b2 t 0 (3.72e)
Solution
Figure 3.8 illustrates each function in a plane. The feasible region is presented and
the dotted lines are the iso-objective contours. Since the objective function has the
maximum at point D, it is an optimum point. As shown in the graphical method,
the optimum is on a boundary of the feasible set. Therefore, the optimum can be
obtained from comparison of the vertex points. Now Equation 3.72 is transformed
to the standard problem. The objective function is to be minimized by multiplying
–1 and a slack variable is added to each inequality constraint.
The problem is transformed as follows:
Minimize f (b) 400b1 600b2 (3.73a)
subject to b1 b2 b3 16 (3.73b)
1 1
b1 b2 b4 1 (3.73c)
28 14
1 1
b1 b2 b5 1 (3.73d)
14 24
bi t 0, i 1, ,5 (3.73e)
Table 3.1. Ten basic solutions for Example 3.8
Position in
No. b1 b2 b3 b4 b5 f
Figure 3.1
1 0 0 16 1 1 0 A
2 0 14 2 0 5/12 í8400 E
3 0 16 0 í1/7 1/3 í F (infeasible)
4 0 24 í8 í5/7 0 í G (infeasible)
5 16 0 0 3/7 í1/7 í J (infeasible)
6 14 0 2 1/2 0 í5600 B
7 28 0 í12 0 í1 í H (infeasible)
8 4 12 0 0 3/14 í8800 D
9 11.2 4.8 0 9/35 0 í7360 C
10 140/17 168/17 í36/17 0 0 í I (infeasible)
where b3 ,b4 and b5 in Equations 3.73bí3.73d are slack variables. The equations
in Equations 3.73bí3.73d are linearly independent. Thus, there is an infinite
number of solutions for those equations. Let us define parameter p as
p n m (n ! m). Then p 2 in this problem. At the vertices of the feasible set,
p variables have zero and they are called basic solutions. If we set p variables as
zero in Equations 3.73bí3.73d, a unique solution exists. We have ten basic
solutions and they are analyzed in Table 3.1. From the basic solutions, the
optimum solution is the one with minimum objective while the constraints are
satisfied.
3.5.3 Terminologies of LP
The terminologies of LP are explained below.

Vertex (extreme) point: This is a point in the feasible set that does not lie on a
line segment joining two other points of the set. All the points in a circle
or vertices of a polygon are examples of the vertex point.
Feasible solution: A point that satisfies the constraints. In Figure 3.8, feasible
solutions are all the points within the polygon made by points A, B, C, D
and E.
Basic solution: A point satisfying constraints while p (n m) variables are
zero.
Basic feasible solution: A basic solution that satisfies constraints. Points A, B,
C, D and E in Figure 3.8 or 1, 2, 6, 8 and 9 in Table 3.1 are basic feasible
solutions.
Degenerate basic solution: A basic solution when one or more basic variables
are zero.
Degenerate basic feasible solution: A basic feasible solution when one or more
variables are zero.
Optimum solution: Out of basic feasible solutions, a solution with minimum
objective function is the optimum solution. Point D in Figure 3.8 or the
8th solution in Table 3.1 is the optimum solution.
3.5.4 Simplex Method
The simplex method is an algorithm to solve linear programming problems. The

basic feasible solutions are searched until an optimum solution is obtained. A
simplex in the n-dimensional space is a convex hull of any (n 1) points that do
not lie on one hyperplane.
Example 3.9 [The Simplex Method]

Solve the following problem by using the simplex method:
Minimize f (b) 3b1 b2 3 (3.74a)

subject to b1 b2 d 4 (3.74b)
2b1 b2 d 2 (3.74c)
b1 t 0, b2 t 0 (3.74d)
Solution
First, the given problem is converted to a standard form as follows:
Minimize f (b) 3b1 b2 3 (3.75a)
subject to b1 b2 b3 4 (3.75b)
2b1 b2 b4 2 (3.75c)
bi t 0, i 1, , 4 (3.75d)
A simplex tableau is used as shown in Table 3.2. The constraints in Equation

3.75 are shown in the first two rows of Table 3.2. The table is said to have a
canonical form. First, a basic feasible solution is obtained. In this problem, (the
number of design variables – the number of constraints) is two, thus, two design
variables are set to zero. Since the number of cases where two design variables are
zero, is six, there are six basic feasible solutions. Therefore, the maximum number
of searches is six, because the search is carried out with the basic feasible solutions.
One of the basic feasible solutions is b3 4 and b4 2. The other variables (b1
and b2 ) are set to zero and the objective function is 3.
Table 3.2. Initial simplex tableau for Example 3.9
b1 b2 b3 b4
1 1 1 0 = 4
2 í1 0 1 = 2
í3 1 0 0 = f í3
Table 3.3. The simplex tableau after one iteration in Example 3.9
b1 b2 b3 b4
0 3/2 1 í1 = 3
1 í1/2 0 1/2 = 1
0 í1/2 0 3/2 = f
Table 3.4. The simplex tableau after two iterations in Example 3.9
b1 b2 b3 b4
0 1 2/3 í1/3 = 2
1 0 1/3 1/3 = 2
0 0 1/3 4/3 = f+1
The third row of the tableau is for the objective function. The least negative
coefficient is found from the third row. It is í3 in Table 3.2. The negative value
means that the objective function can be reduced. Thus, a new basic feasible
solution having nonzero b1 is selected. Since we can have two nonzero variables,
one of the nonzero variables should be changed to zero. If we investigate the
column with í3 in the third row, there are positive numbers 1 and 2 in the other
rows. We evaluate the ratios between these numbers and the numbers in the right
most column and compare them. In this case, 4/1 > 2/2. The nonzero variable in
the row with the smaller rate is changed to zero. That is, b4 is changed to zero
and b1 is changed to nonzero.
A simplex tableau is again made with the new basic feasible solution. A new
tableau with the canonical form is shown in Table 3.3. The pivoting process is
already applied to Table 3.3. The solution from the tableau is b1 1, b3 3,
b2 0, b4 0 and f 0. Since there is a negative value in the third row, the
objective function can be further reduced. The previous process is applied to
Table 3.3. If we investigate the column with í1/2 in the third row, the positive
coefficient is 3/2. Therefore, b3 is changed to zero and b2 is changed to nonzero.
The simplex tableau after two iterations is shown in Table 3.4. The basic
feasible solution is b1 2, b2 2, b3 0 and b4 0 , and the objective function
is í1. Since there is no negative value in the third row of Table 3.4, the objective
function cannot be reduced and the solution is the optimum solution.
3.6 One-dimensional Minimization

One-dimensional minimization with a single variable is rare in real design
problems. It may seem meaningless to develop algorithms for one-dimensional
minimization since they are not directly utilized in practical optimization.
However, the stepsize evaluation process is a problem with a single variable.
Therefore, one-dimensional minimization algorithms are used for the stepsize
calculation. As mentioned earlier, the following function is used for the stepsize:
f (b ( k 1) ) f (b ( k ) Dd ( k ) ) f (D ) (3.76)
The stepsize D , which makes the function f minimum, is the best value. In
df (D )
mathematics, D should satisfy the condition 0. Generally, the process
dD
for the stepsize calculation is in the middle of an optimization algorithm. Thus, D
df (D )
is calculated in an iterative manner instead of calculating 0. A few
dD
algorithms are introduced below.
3.6.1 Equal Interval Search
One simple numerical method is equal interval search. Suppose we have a

function f (D ) as illustrated in Figure 3.9. The function is calculated at the points
with equal intervals, the function values are compared and D with minimum
f (D ) is found. First, a certain small value G is selected and function values are
calculated at D G , 2G , 3G ,, qG , (q 1)G . The function values are compared at
adjacent points qG and (q 1)G as q becomes larger. If f (qG ) ! f ((q 1)G ),
then the search does not pass the minimum point. On the other hand, if
f (qG ) f ((q 1)G ), then the search passes the minimum point. The minimum
point is between (q 1)G and (q 1)G . Then the interval of uncertainty I is 2G .
The interval of uncertainty is searched again with rG where r 1 . The interval
of uncertainty is successively reduced and the process terminates when it is less
than a sufficiently small H .
f (D ) f (D )
rG
G
D D
G 2G qG D * (q 1)G D Dl D* Du
(a) Initial bracketing of the minimum (b) Reduction of the interval of uncertainty
Figure 3.9. Equal interval search process

3.6.2 Golden Section Search
The equal interval search has weak points in that: (1) It is difficult to estimate the
size of G appropriately and (2) the number of calculations of f (D ) can be
extremely large. Function calculation is quite costly in engineering, so it should be
reduced. Various algorithms have been developed to overcome these difficulties.
The golden section search method is introduced here. In the golden section search
method, G is increased by multiplying the golden section ratio 1.618 as follows:
q
j
Dq ¦ G (1.618) , q 0, 1, 2, (3.77)
j 0
First, we need to understand the following Fibonacci sequence:
F0 1, F1 1, Fn Fn 1 Fn 2 , n 2, 3, (3.78)
Fn
lim 1.618 (3.79)
nof Fn 1
As n goes to infinity in the Fibonacci sequence, the ratio of the consecutive

numbers is 1.618. In the golden section method, the points are searched for in a
pattern similar to the Fibonacci sequence.
Figure 3.10 shows the initial process to evaluate the minimum point by using
the golden section method. When f (D q 1 ) f (D q 2 ) and f (D q 1 ) f (D q ), the
search passes the minimum point. The minimum point is located between D q 2
and D q . Thus, the initial interval of uncertainty is as follows:
ƒ(Į)
q =0 1 2 3 ···
Į
0 į 2.618į 5.236į Į* 9.472į
Figure 3.10. Initial bracketing of the minimum point in the golden section method
ĲI (1íĲ)I Įu
(a) Įl Įa
Įb
(1íĲ)I ĲI
I’
Į'b
Į'l Į'u
(b)
ĲI’ (1íĲ)I’
Figure 3.11. Golden section partition
q q2
j j
I Du Dl D q D q2 ¦ G (1.618) ¦ G (1.618) (3.80)
j 0 j 0
If the interval of uncertainty I is determined, then it is subdivided in the next

process. As illustrated in Figure 3.11, D a and D b are defined at the points of WI
and (1 W ) I , where W 0.618 .
The steps of the golden section search are as follows:
Step 1. A small increment G is defined and Equation 3.77 is evaluated.
Determine q so that f (D q 1 ) f (D q 2 ) and f (D q 1 ) f (D q ). The
interval of uncertainty is obtained from Equation 3.80. Define a
sufficiently small number H , which is used for the convergence criterion.
Step 2. Determine D a and D b as illustrated in Figure 3.11. Since D a D q 1 ,
f (D a ) is already calculated in Step 1.
Step 3. Compare f (D a ) and f (D b ) , and go to (1), (2) or (3).
(1) If f (D a ) f (D b ), the optimum point D * is within D l D * D b . As

illustrated in Figure 3.11b, the new limits of uncertainty are D lc D l
and D uc D b . A variable is defined as D bc D a . Calculation of f (D bc )
is not needed because f (D bc ) f (D a ). f (D ac ) is calculated from
D ac D lc 0.328(D uc D lc ). Go to Step 4.
(2) If f (D a ) ! f (D b ), then the optimum point D * resides within

D a D * D u . New limits of uncertainty are defined by D lc D a ,
D uc D u and D ac D b. As mentioned earlier, the calculation of
f (D ac ) is not needed. Set D bc D lc 0.618(D uc D lc ), calculate f (D bc )
and go to Step 4.
(3) If f (D a ) f (D b ), then set D l D a , D u D u or D l D l , D u D b and
go to Step 2.
Step 4. If I c( D uc D lc ) H , then the process is considered to converge. The
optimum point is D * (D uc D lc ) / 2 and the process terminates. Otherwise,
eliminate the prime symbols and go to Step 3.
Example 3.10 [Golden Section Search]

Find a minimum point for the following problem with the golden section search:
f ( x) x2 5x 4 (3.81)
G = 1.
Solution
x0 1
x1 1 1.618 2.618
x2 1 1.618 1.618 2 5.2359
Iteration 1:
(1) f ( x0 ) 0, f ( x1 ) 2.2361, f ( x2 ) 5.2351
0
I 4.2359, xu 5.2359, xl 1
(2) xa 2.618
xb 1 0.618, I 0 3.6178
(3) f ( x a ) 2.2361, f ( x b ) 1.0005
f ( xa ) f ( xb ), thus, the range of the minimum point x * is 1 d x*
d 3.6178.
(4) xl' xl 1, xu' xb 3.6178, xb' xa 2.618
xa' xl' 0.382( xu' xl' ) 2
(5) I xu' xl' 2.6178
Iteration 2:
(1) x a' o x a , x u' o x u , x b' o x b , x l' o x l

(2) f ( x a ) f (2) 2, f ( x b ) f (2.618) 2.2361
f ( xa ) ! f ( xb ), thus, the range of the minimum point x * is 2 d x*
d 3.6178. .
(3) xl' xa 2, xu' xu 3.6178, xa' xb 2.618
xb' xl' 0.618( xu' xl' ) 2.9998
(4) I xu' xl' 3.6178 2 1.6178
If the iteration is continued, the minimum point is found.
Only representative methods are introduced. Some methods use the

interpolation method to reduce the number of calculations for function f.
3.7 Numerical Methods for Unconstrained

Optimization
Unconstrained optimization only has the objective function without constraints. If
the objective function is a linear function, the problem is unbounded. Therefore,
the objective function has to be nonlinear and the optimization problem is
formulated as follows:
Minimize f (b) (3.82)
As mentioned earlier, the optimum solution b* is the solution of the necessary

condition made of simultaneous equations f (b) 0 , which means that the
gradient is the zero vector. Numerical rootfinding methods can be utilized but this
can be costly. Moreover, the solution does not guarantee the sufficient condition
that the Hessian matrix H (b* ) is positive definite. Thus, the rootfinding methods
are not generally utilized in optimization. Various methods have been developed
specifically for unconstrained optimization. The flow of the methods proceeds
iteratively as explained in Section 3.4. After a direction vector is evaluated, the
stepsize is calculated and the design point is updated until the convergence criteria
are satisfied.
3.7.1 Steepest Descent Method
A representative algorithm is the steepest descent method. It changes the design in

the steepest descent direction and gives a basic idea to most methods. The first
step is to find the direction vector d in the steepest descent direction. The descent
condition is explained in Equations 3.56 and 3.57. At the kth iteration, the change
T
of the objective function in the direction of d ( k ) is D k (c ( k ) d ( k ) ) in the linear
expansion of the Taylor series. Because stepsize D k is a positive number, the
vector d ( k ) to reduce the objective function the most is c ( k ) , as seen in the linear
expansion. The vector c( k ) is the gradient vector and d ( k ) f (b ( k ) ) yields the
steepest descent vector.
The steps of the method are as follows:
Step 1. Estimate an initial design b ( 0) . Set the iteration number k 0 and a

small convergence parameter H ! 0.
Step 2. At design point b ( k ) , evaluate the gradient vector c ( k ) f (b ( k ) ). If
|| c ( k ) || H , terminate and the optimum solution b* is b ( k ) . Otherwise, go
to Step 3.
Step 3. Evaluate the direction vector d ( k ) c ( k ) .
Step 4. Calculate stepsize D k that minimizes f (b ( k ) D k d ( k ) ).
Step 5. Update the design with b ( k 1) b (k ) D k d (k ) . Set the iteration
number by k k 1 and go to Step 2.
Now the steepest descent method is investigated. The gradient vector is
utilized in Step 2. The process to calculate the gradient vector is referred to as
sensitivity analysis and it is obtained from an analytic method or the finite
difference method. The vector d ( k ) c ( k ) in Step 3 is defined in the direction
where the objective function is reduced the most linearly. The stepsize in Step 4
can be obtained from one-dimensional optimization because only D k is a variable
and other parameters are constant in function f (b ( k ) D k d ( k ) ) . As the iteration
proceeds, the objective function is continually reduced and the process terminates
when the objective function can no longer be reduced. Therefore, the necessary
and sufficient condition is satisfied.
It is noted that the steepest descent directions of adjacent iterations are
orthogonal to each other as follows:
T
c ( k ) c ( k 1) 0 (3.83)
The minimum condition in Step 4 is df / dD k 0 . Using the chain rule,
T
df (b ( k 1) ) § wf (b ( k 1) ) · db ( k 1)
¨ ¸
¨ ¸ dD 0 (3.84)
dD k © w b ¹ k
From Equation 3.84, if we use the definition of the gradient vector
wf (b ( k 1) )
c ( k 1) (3.85)
wb
and
db ( k 1) d
(b ( k ) D k d ( k ) ) d(k ) c ( k ) (3.86)
dD k dD k
Using Equations 3.85 and 3.86, Equation 3.84 becomes
T T
c ( k 1) d ( k ) c ( k 1) c ( k ) 0 (3.87)
Thus, Equation 3.83 is satisfied.

The condition in Equation 3.83 is the drawback of the method. The iso-
objective contours for an objective function can be closely distributed as illustrated
in Figure 3.12. In other words, the partial differentiation value for a certain
variable is much larger than the others. Because the direction vector of an iteration
should be orthogonal to the one in the previous iteration, the speed of the
convergence can be very slow. The method shows quite inefficient performance
for many highly nonlinear problems. Thus, the method is rarely used, although the
idea is excellent in that the gradient vector is exploited.
Example 3.11 [Steepest Descent Method]

Repeat two iterations of the following problem by using the steepest descent
method:
f (b) b12 5b22 4b1b2 4b2 4
Solution
Iteration 1:
(1) b ( 0) [0 0]T
(2) f [2b1 4b2 10b2 4b1 4]T [0 4]T
(3) direction vector d f [0 4]T
(4) b (1) b ( 0 ) Dd [0 0]T D [0 4]T [0 4D ]T
f (b ( 0) Dd) 80D 2 16D 4
D* 0.1
b2
Iso-objective line of f (b1 , b2 )
b1
Figure 3.12. Orthogonal steepest descent paths when the difference of partial derivatives is
large
(5) b (1) b ( 0) D *d [0 0]T 0.1 u [0 4]T [0 0.4]T
Iteration 2:
(1) f [ 2b1 4b2 10b2 4b1 4]T [ 1. 6 0] T

(2) direction vector d f [1.6 0]T
(3) b ( 2) b (1) Dd [0 0 .4]T D [1.6 0]T [1.6D 0.4]T
64 2 64 80
f (b (1) Dd) D D
25 25 25
*
D 0.5
(4) b ( 2) b (1) D *d [0 0.4]T 0.5 u [1.6 0]T [0.8 0.4]T
3.7.2 Conjugate Gradient Method
The conjugate gradient method is modified from the steepest descent method and
shows good performance. It was developed by Fletcher and Reeves and the
convergence rate was improved considerably through modification of the
orthogonal condition in Equation 3.83. It is exactly the same as the steepest
descent method at the initial iteration (k 0) . From the second iteration (k 1),
all the processes are the same except for the evaluation of direction vector d ( k ) in
Step 3. The direction vector is
d (k ) c ( k ) E k d ( k 1) (3.88)
Ek (|| c ( k ) || / || c ( k 1) ||) 2 (3.89)

The convergence rate is improved because direction vector d ( k ) resides between

vectors c ( k ) and d ( k 1) .
3.7.3 Quasi-Newton Method
Newton’s method is also frequently used for unconstrained optimization. The

rootfinding Newton’s algorithm is exploited in finding direction vector d. By the
Taylor series, function f(b) can be expanded up to the second-order as follows:
f (b 'b) f (b) c T 'b 0.5'b T H'b (3.90)
If matrix H is positive definite, Equation 3.90 is convex and 'b for the minimum
of Equation 3.90 is unique. The condition is
df
c H'b 0 (3.91)
d('b)
Unless H is singular, direction vector d is replaced by 'b in Equation 3.91
d 'b H 1c (3.92)
All the processes of the algorithm are the same as the steepest descent method,
except for the evaluation of the direction vector in Step 3.
Newton’s method is a theoretically excellent method in that the order of
convergence is two. However, the method needs the Hessian matrix, which is very
expensive to evaluate. Moreover, when the Hessian matrix is not positive definite,
convergence is not guaranteed. If a design iteration comes to a point where the
Hessian is indefinite in the iterative process, the next design can be extremely
deteriorated. Therefore, it is not utilized in practice, although provides ideas for
quasi-Newton methods.
Quasi-Newton methods use an approximation of the Hessian matrix H or its
inverse H 1. They do not require the condition of positive definiteness of the
Hessian matrix or calculation of the second derivatives. It is known that the
convergence rate of the methods is inferior to Newton’s method and superior to the
conjugate gradient method. This is evidenced from numerical research and not
from rigorous mathematics. The Hessian matrix is assumed to be a unit matrix at
the initial iteration and is updated as the iteration proceeds. Two methods exist
and their characteristics are as follows:
DFP (Davidon–Fletcher–Powell) Method

The inverse of the Hessian matrix is approximated in this method. It was proposed
by Davidon and modified by Fletcher and Powell. The method is called the DFP
method. The algorithm is the same as in the steepest descent method, except for
two aspects.
(1) In Step 3 of the steepest descent method, the direction vector is evaluated
as follows:
d (k ) A (k ) c (k ) (3.93)
where A (k ) is an approximated matrix of the inverse Hessian. At the first

iteration, it can be presumed by some positive definite matrix. However,
an initial guess of A ( 0) I is usually used.
(2) After the design is changed, the inverse Hessian matrix is updated as
follows:
A ( k 1) A (k ) B (k ) C (k ) (3.94)
where matrices C(k ) and B (k ) are defined as
T
(k ) s (k ) s (k )
B T
(3.95a)
s (k ) y (k )
T
z (k ) z (k )
C (k ) T
(3.95b)
y (k ) z (k )
s (k ) D k d ( k ) (change of design) (3.95c)
y (k ) c ( k 1) c ( k ) (change of gradient vector) (3.95d)
c ( k 1) f (x ( k 1) ) (3.95e)
z (k ) A (k ) y (k ) (3.95f)
Fletcher and Powell proved that matrix A (k ) is always positive definite and it
converges to the inverse Hessian when the objective function is a positive definite
quadratic function.
BFGS (Broyden–Fletcher–Goldfarb–Shanno) Method

The Hessian matrix is directly approximated in this method. The method is
commonly referred to as the BFGS method. All the processes are the same as in
the DFP method, except for the following two steps:
(1) The direction vector is evaluated from a linear system of equations
H (k ) d (k ) c ( k ) (3.96)
where H (k ) is an approximated matrix for the Hessian at the kth iteration.

H ( 0) I is usually utilized.
(2) The Hessian matrix is updated as follows:
H ( k 1) H (k ) D (k ) E (k ) (3.97)
where the matrices D(k ) and E(k ) are defined as
T
y (k ) y (k )
D (k ) T
(3.98a)
y (k ) s (k )
T
c (k ) c (k )
E (k ) T
(3.98b)
c (k ) d (k )
s (k ) D k d (k ) (change of design) (3.98c)
y (k ) c ( k 1) c ( k ) (change of gradient vector) (3.98d)
c ( k 1) f (x ( k 1) ) (3.98e)
The positive definiteness of H ( k ) is guaranteed in this method as well.
3.8 Numerical Methods for Constrained Optimization

General engineering optimization imposes constraints such as various design
specifications and environmental restrictions. Optimization problems with
constraints are called constrained optimization and formulated as in Equation 3.1.
Once the formulation is established, the engineering problem can be regarded as a
mathematical problem. Various numerical methods have been developed to solve
the mathematical problem. Some popular methods are introduced in Table 3.5.
The indirect method tries to solve the Karush–Kuhn–Tucker necessary conditions.
It has been verified to be numerically inefficient. The method is rarely used except
for topology optimization methods, which have an extremely large number of
design variables. The use of transformation methods, while popular in 1960s and
1970s, has seen a decline. The primal method is generally used in engineering
optimization and this chapter explains the basis of most primal methods.
Table 3.5. Classification of numerical methods for constrained optimization
Classification Algorithms Remarks

A solution of an
Sequential linear programming (SLP)
approximated
Gradient projection method (GRP) subproblem is
Primal
Feasible directions method (FDM) obtained and the
method
optimum solution is
Sequential or recursive quadratic
found in an iterative
method (SQP or RQP)
Direct process
method A constrained
Sequential unconstrained minimization
problem is
technique (SUMT)
Transfor- transformed to an
mation Penalty function method unconstrained
method Barrier function method problem and the
solution is found in
Augmented Lagrangian method
an iterative manner
The KKT
conditions of the
Indirect formulated problem
Optimality criteria method
method are used to drive an
iterative solution
strategy
3.8.1 General Aspects of the Primal Method
In the primal method, a linearized subproblem is defined from a nonlinear

optimization problem. Linearization of a nonlinear function is performed. A
direction vector is evaluated from the subproblem, a stepsize is calculated and the
next design point is found in an iterative manner. At the current design point, a
subproblem is established by linearization of functions in Equation 3.1.
A linearized subproblem in the kth iteration is as follows:
Find Gb ( k ) (3.99a)
to minimize T f (b ( k ) )Gb ( k ) (3.99b)
subject to hi (b ( k ) ) T hi (b ( k ) )Gb ( k ) 0, i 1, , l (3.99c)
g j (b ( k ) ) T g j (b ( k ) )Gb ( k ) d 0, j 1, , m (3.99d)
Gb (Lk ) d Gb ( k ) d Gb U( k ) (3.99e)
The above subproblem is interpreted as follows:

(1) Vector Gb ( k ) is the design variable vector of the subproblem and is the
design change from the current design point.
(2) Equation 3.99b is the linear term of the Taylor expansion and the change of
the objective function when the design variable vector b ( k ) is changed by
Gb ( k ) . It is minimized to reduce the objective function.
(3) Equation 3.99c requires the equality constraints to be satisfied. Suppose
the ith constraint is satisfied at the current point by hi (b ( k ) ) 0. Then
Equation 3.99c changes the constraint to satisfy T hi (b ( k ) )Gb ( k ) 0.
When the ith constraint is violated by hi (b ( k ) ) z 0, Equation 3.99c tries to
satisfy the ith constraint at the next iteration.
(4) An inequality constraint can be corrected by Equation 3.99d. In any cases,
Equation 3.99d tries to satisfy the jth inequality constraint by
g j (b ( k ) ) T g j (b ( k ) )Gb ( k ) d 0 at the next iteration.
(5) Equation 3.99e restricts the size of Gb ( k ) .
The subproblem is a linear programming problem, thus the optimum solution
of Gb ( k ) can be obtained by a linear programming method. In general, algorithms
of the primal method proceed as follows:
Step 1. An initial design b (0) is estimated. The iteration number is set by

k 0. Small numbers H 1 and H 2 are defined.
Step 2. The subproblem in Equation 3.99 is constructed at the current design
point and Gb ( k ) is determined. The direction vector is set by d ( k ) Gb ( k ) .
Step 3. The convergence is checked. General convergence criteria are that
the norm of the direction vector || d ( k ) || is less than H 1 and the maximum
of constraint violation is less than H 2 . When the convergence criteria are
satisfied, the process terminates.
Step 4. Stepsize D k is evaluated. The method depends on the method
utilized for the one-dimensional minimization.
Step 5. The design is updated using Equations 3.54 and 3.55. The iteration
number is set by k k 1 . Go to Step 2.
It is noted that one needs gradients in the construction of the subproblem. The
calculation of the gradients is called a sensitivity analysis and it can be quite an
expensive process in engineering.
Example 3.12 [Linearized Subproblem]

Make a linearized subproblem of the following problem:
Minimize f (b) b13 b2 (3.100a)
subject to b22 2b1b2 3 d 0 (3.100b)

b1 3 t 0 (3.100c)
1
t0 (3.100d)
b2 1
The current design is b 0 [ 2 1]T .
Solution
f [3b12 1]T , f 9 , f [12 1]T
g 1 [2b2 2b2 2b1 ]T , g1 0 , g 1 [2 2]T
g 2 [1 0]T , g 2 5
T T
ª 1 º 1 ª 1º
g 3 «0 » , g3 , g 3 «0
«¬ b2 1 »¼
2 2 ¬ 4 »¼
If we substitute the above information into Equation 3.99, then the following
linearized subproblem is obtained:
Find [Gb1 Gb2 ]T (3.101a)
to minimize [12 1][Gb1 Gb2 ]T (3.101b)
subject to 0 [2 2][Gb1 Gb2 ]T d 0 (3.101c)
5 [1 0][Gb1 Gb2 ]T t 0 (3.101d)
1 ª 1º
0 [Gb1 Gb2 ]T t 0 (3.101e)
2 «¬ 4 »¼
In engineering, the number of inequality constraints m in Equation 3.99d is
often extremely large. Then the size of the subproblem becomes very large.
Frequently some inequality constraints satisfied in an iteration are satisfied again
in the next iteration. Then those constraints have no roles in the iterative process.
In other words, those constraints do not need to be considered in the subproblem.
Therefore, if a constraint is sufficiently satisfied at the kth iteration, the
subproblem of the kth iteration does not include the constraint. Then the size of
the subproblem is reduced and the cost for the sensitivity analysis for the constraint
is saved. This is called the active set strategy.
When a satisfied constraint is close to zero, satisfaction and violation can
alternate between feasibility and infeasibility. The İ-active set strategy is utilized
to prevent it. Although the constraints are satisfied, the ones with the following
condition are included in the active set:
g j (b) t H , H : small positive number (3.102)
A feasible region is presented in Figure 3.13. If we use the İ-active set strategy,
the feasible region is reduced. With the active set strategy (not the İ-active set
strategy), when the design points such as B and C in Figure 3.13 are close to the
boundary, the constraint g j (b ( k ) ) is not included in the subproblem and the next
design can have a large probability of moving into the infeasible region. However,
if we use the İ-active set strategy, the feasible region is reduced, because the inside
region of the dotted lines yields the feasible region.
When the active set strategy is employed and the design point is at B or C, the
constraint g j (b) is excluded in the active constraint set and the design may move
to an infeasible point such as D. However, when the İ-active set strategy is
utilized, the constraints are active and the design point remains in the feasible
region. Because the design point frequently exists around the boundary of a
constraint, the İ-active set strategy works quite well. When the number of
constraints is reduced, the subproblem size is reduced and sensitivity analysis of
the excluded constraints is not needed. Of course the equality constraints are
included in the active set all of the time.
In the İ-active set strategy, active constraint set I k is defined as
Ik {i | hi (b) 0, i 1, , l} { j | g j (b) H ! 0, j 1, , m} (3.103)
The set I k is a list of the active constraints. The strategy of using I k in a

subproblem is called the İ-active potential set strategy or the İ-active set strategy.
The constrained deletion strategy is another name (Vanderplaats 1984). Most
commercial software systems use the İ-active set strategy. When the İ-active set
strategy is used, it is noted that the constraints of the İ-active set are changed at
each iteration.
D•
•
C
B•
İ
g j (b) 0
•A
Feasible region g j (b) H 0
Figure 3.13. Schematic of the İ-active set strategy

Example 3.13 [İ-active Set Strategy]

Obtain the İ-active set from the following constraints, b 0 [0.56 2.5]T and
H 0.1 ,
b1 b22
g1 1 d 0 (3.104)
3 3
g2 b1 1 d 0 (3.105)
b1b2
g3 1 d 0 (3.106)
4
3
g4 1 d 0 (3.107)
2b1 2
3
g5 1 d 0 (3.108)
2b2
Solution
g1 0.8967 o not active (3.109)
g2 1.56 o active (3.110)
g3 0.65 o not active (3.111)
g4 0.0385 o İ-active (3.112)
g5 0.4 o not active (3.113)
Therefore, the İ-active set is I k {g 2 , g 4 } .
3.8.2 Primal Method
Primal methods have been developed based on the steps in Section 3.8.1. Each
algorithm includes slight modifications of the baseline steps. In this section,
representative algorithms are introduced by virtue of popularity and applicability.
Sequential Linear Programming (SLP)

Sequential linear programming (SLP) is a primal method that has almost the same
steps as those in Section 3.8.1. SLP utilizes the subproblem in Equation 3.99,
whereas the move limit on the direction vector Gb ( k ) in Equation 3.99e is defined
as
'(ik ) d Gbi( k ) d '(ik ) , i 1, , n (3.114)
Equation 3.99e is replaced by Equation 3.114 and LP methods can be used to

solve the subproblem, because Equation 3.114 is a linear equation. It is noted that
the move limit can be varied at each iteration. Therefore, the move limit strategy
can determine the efficiency of the method. When the move limit strategy is
excellent, the stepsize to update design in Equations 3.54 and 3.55 can be 1. This
means Step 4 to calculate the stepsize in Section 3.8.1 can be omitted and the
expensive one-dimensional line search is not required. However, the method
encounters difficulties because an efficient move limit strategy is difficult to define.
Nevertheless, the method is quite popular due to its simplicity and numerical
efficiency.
Sequential Quadratic Programming (SQP)

As an advancement of SLP, we have the sequential quadratic programming (SQP)
method. It is also called the recursive quadratic programming (RQP) method. The
subproblem is defined by the quadratic programming (QP) problem. The objective
function is a quadratic function and the constraints are linear in the QP.
First, we will investigate the quadratic subproblem. Equation 3.99e can also be
expressed as
|| Gb ( k ) || d [ 2 (3.115)
The norm in Equation 3.115 is expanded as
n
0.5 ¦ (Gbi( k ) ) 2 d [ 2 (3.116)
i 1
Equation 3.116 is modified as

T
0.5Gb ( k ) WGb ( k ) d [ 2 (3.117)
where W is an identity matrix. However, W may not be an identity matrix

according to the bounds in Equation 3.99e. If we substitute Equation 3.99e into
Equation 3.117, the subproblem yields the following:
Find Gb ( k ) (3.118a)
to minimize T f( b ( k ) )Gb ( k ) (3.118b)

T
0.5Gb ( k ) WGb ( k ) d [ 2 (3.118e)
Since Equation 3.118e is nonlinear, the algorithm for linear programming cannot
be used. The KKT condition of Equation 3.118 is equivalent to that of Equation
3.119:
Find Gb ( k ) (3.119a)
T
to minimize T f( b ( k ) )Gb ( k ) 0.5Gb ( k ) WGb (3.119b)
Equation 3.119 is a quadratic programming (QP) problem. A QP problem has a

quadratic objective function and linear constraints. The sequential quadratic
programming algorithm solves the subproblem in Equation 3.119. A quadratic
programming algorithm is employed to solve the subproblem. It is noted that
Equation 3.119b can be interpreted as the second-order Taylor series expansion of
the objective function.
The direction vector Gb ( k ) is evaluated from the quadratic programming
problem in Equation 3.119. Using the above subproblem, the flow of the method
follows the steps in Section 3.8.1. After the direction vector is obtained, the
stepsize is determined by a descent function. That is, the design moves to a new
design that minimizes the descent function in the direction of Gb ( k ) . For instance,
the descent function in the kth iteration is defined as
l
I f (b( k ) D k d(k ) ) ¦ vi | hi (b(k ) D k d( k ) ) |
i 1
(3.120)
m
¦ u j (max[0, g j (b(k ) D k d( k ) )])
j 1
The descent function in Equation 3.120 is made by adding the objective

function and the penalty functions representing the constraint violation. A penalty
function is defined by multiplying Lagrange multipliers and constraint violations.
It has the same unit as that of the objective function. Stepsize D k is the only
variable in Equation 3.120 and is determined to minimize the descent function I .
A one-dimensional search method can be employed. The number of calculations
for the functions f, h and g can be very large. Therefore, selection of an efficient
one-dimensional search method is quite important. Pshenichny’s method is one of
the original SQP methods. It uses a specific active set strategy with W I and a
specific method for determination of D k (Arora 2004).
There is an SQP method that exploits the idea of Newton’s method for
unconstrained problems. The matrix W is replaced by the Hessian matrix of the
Lagrangian function in Equation 3.46 as
W 2L (3.121)
This method is quite efficient in that it has a second-order convergence rate.

However, it is not practical because the Hessian matrix is very costly to evaluate
and the convergence is not guaranteed when W is not positive definite.
To overcome the above difficulties, constrained quasi-Newton methods have
been developed by using the approximation of the Hessian matrix for W. They
are similar to the unconstrained quasi-Newton methods. Now, another method is
introduced. In every iteration, matrix W 2 L is updated by using the Broyden–
Fletcher–Goldfarb–Shanno (BFGS) method as follows:
s (k ) D k d (k ) (3.122a)
z (k ) W (k ) s (k ) (3.122b)
y (k ) L(b ( k 1) , u ( k ) , v ( k ) ) L(b ( k ) , u ( k ) , v ( k ) ) (3.122c)

T
[1 s (k ) y (k ) (3.122d)
T
[2 s (k ) z (k ) (3.122e)
0.8[ 2
if [1 t 0.2[ 2 T 1 , otherwise T (3.122f)
[ 2 [1
p( k ) șy ( k ) (1 T )z ( k ) (3.122g)
T
[3 s (k ) p (k ) (3.122h)
1 T
D (k ) p (k ) p (k ) (3.122i)
[3
1 T
E (k ) z (k ) z (k ) (3.122j)
[2
W ( k 1) W (k ) D(k ) E (k ) (3.122k)
There are many other methods according to constraint treatment, stepsize

evaluation, etc. (Arora 2004, Rao 1996)
3.8.3 Transformation Method
It is well known that many excellent methods have already been developed for
unconstrained problems. If constrained problems can be transformed to
unconstrained problems, the existing unconstrained optimization methods can be
exploited. Numerous research investigations have been carried out to explore this
idea. Generally, the constraint functions are transformed to a penalty function and
added to the objective function. The transformed functions are usually expressed
as follows:
< (b, r ) f (b) P(h(x), g(b), r ) (3.123)
where < is the transformed objective function, P is the penalty function and r is a
vector for the penalty parameters. P is a real function and the magnitude of the
penalty is controlled by r. Based on the definition of <, an unconstrained
optimization method finds an optimum for < with initial values of b (0) and r.
During the optimization process, b is the design variable vector and r is considered
as a constant vector. The optimum values yield the initial design of the next
iteration and the new iteration finds an optimum of < with an adjusted r. The
iterative process terminates if no improvement is observed.
The sequential unconstrained minimization technique (SUMT) is the
representative method for the transformation methods. It is generally called
SUMT. The method is varied by virtue of the definitions of the penalty function
and the penalty parameter r. One well known penalty function is the quadratic loss
function below
l m
P(h(b), g(b), r ) r{ ¦ [ hi (b)] 2 ¦ [ max (0, g j (b))] 2 } (3.124)
i 1 j 1
where r is a scalar property. This is called an exterior penalty function method

because the solutions in the iteration pass through the infeasible region.
Another well known method of the penalty function is the barrier function
method, which is called the interior penalty function method. It can only be used
for inequality constraints and the following two forms are defined:
m
P (g(b), r ) (1 / r ) ¦ [1 / g j (b)] : Inverse barrier function (3.125)
j 1
m
P (g(b), r ) (1 / r ) ¦ log( g j (b)) : Log barrier function (3.126)
j 1
Because these functions set a large barrier around the feasible region, a feasible
solution cannot jump to the infeasible region. When a constraint is active, P is
theoretically infinite. Also, the algorithm needs the condition that the initial design
should be feasible. The parameter r should be increased as the iterations proceed.

When r becomes infinite, the optimum of Equation 3.123 converges to the
optimum of the original problem.
The augmented Lagrangian method is another transformation method. It is also
called the multiplier method. Because it does not use the aforementioned r, poles
do not appear in the transformation function <. The penalty function is defined
as
l m
P (h(b), g(b), r ) 0.5 ¦ ri' (hi T i' ) 2 0.5 ¦ r j [( g j T j ) ] 2 (3.127)
i 1 j 1
where ( g j T j ) max(0, g j T j ) . T i' and ri' are parameters related to the ith
equality constraint. T j ! 0 and r j ! 0 are related to the jth inequality constraint.
If T j T i' 0 and ri ri' r , Equation 3.127 is the same as the quadratic loss
function as in Equation 3.124. The parameter r goes to infinity in the quadratic
loss function, whereas it has finite values here. The optimum of Equation 3.123
should be found by substituting Equation 3.127 into Equation 3.123 in an iterative
manner. The parameters are updated using various rules.
3.9 Exercises
3.1 The following statements are made to formulate engineering optimization

problems. Each problem is to be changed to an optimization formulation.
However, the statements are not sufficient for formulation. Indicate what
is needed for the optimization formulation.
(1) Optimize a pump.
(2) In the optimization of a pump, we can change the outer radius and the
number of blades.
(3) Maximize the flow rate of the pump.
(4) To maximize the flow rate of the pump, we can change the outer
radius and the number of blades.
3.2 A cantilever beam has a vertical load P at the tip and the length is L. The
vertical displacement at the tip (d) should be less than 10 mm and the
maximum stress at the clamped end should be less than 100 MPa.
Formulate the optimization problem to minimize the weight of the beam.
Determine the cross section. (Young’s modulus is E, the moment of inertia
is I, the density is U , and the width and the height of the cross section are
b and h, respectively.)
s wG
h
d b
3.3 In the following spring system, we have external loads P1 and P2 .

Formulate an optimization problem to find Q x and Q y , which are
displacements at the point Q. (Hint: minimize the total potential energy of
the system.)
k1
l1
Qx P2
P2 Qy P1
Q
Q
P1
l2 k2
3.4 From occupant analysis, a force–time graph r(t) is obtained as the

following figure. An approximated polynomial p(t) is to be evaluated. p(t)
should satisfy the following conditions:
p(t)
r(t)
(1) The difference between the momentum by p(t) and that of r(t) should
be minimized.
(2) The differences between p(t1) and r(t1) and p(t2) and r(t2) should be
less than 0.5.
(3) p(t) is a polynomial equation of the 6th degree.
Formulate an optimization problem to obtain p(t) while the given
conditions are satisfied.
3.5 Solve the following optimization problems by using graphical

optimization:
(1)
Minimize f (b) b22 12b1 4
subject to g 1 (b) b1 1 d 0
g 2 (b) b12 b22 3 d 0
(2)
Minimize f (b) b1 b2
subject to g1 (b) b1 3b 2 5 t 0
g 2 (b) 2b1 10b 2 2 t 0
g 3 (b) b1 t 0
3.6 Find a local optimum point of the following functions:
(1) f (b1 , b2 ) (b1 b2 ) 2 (b2 1) 2

(2) f (b1 , b2 ) b12 b1b2 3b22 4b2
(3) f (b1 , b2 ) 2b1 3 /(b1b2 ) 3b2
3.7 Find a local optimum point of the following problems by using the KKT
necessary condition:
(1)
Minimize f (b) 10b12 2b22 5b1b2 3b1
subject to g 1 (b) b1 b2 4
(2)
Minimize f (b) 3b12 16b1 16b22 7
subject to g (b) b12 b22 4b1 10
(3)
Minimize f (b) 2b1 3b 2 b12 b22
subject to g 1 (b) b1 3b2 d 5
g 2 (b) 2b1 5b2 d 8

b1 , b2 t 0
3.8 Minimize the following function:
f (b) 3b 2 6b 7 , b (0) 2
Repeat the five steps of Section 3.4 twice. The direction vector
d f and the stepsize D 0.2 .
3.9 Calculate the sensitivity of the following function by the finite difference
method with the following finite differences. Compare the values with
those from the analytic method. The current design point is b ( 0) 1 .
f (b) e x sin x, 'x 1, 0.1, 0.001, 0.00001, 1 u 10 4
3.10 Transform the following problems to standard linear programming

problems:
(1)
Minimize f (b) 10b1 4b2 2b3
subject to b1 b2 2b3 t 2
5b1 2b2 3b3 t 4
b1 , b2 t 0
(2)
Minimize f (b) b1 3b2
subject to b1 3b2 d 10
b1 4b2 d 5
b1 2b2 t 6
b1 , b2 t 0
(3)
Maximize f (b) 3b1 4b2 5b3 5b4
subject to 5b1 3b2 7b3 d 8
2b1 3b2 3b3 b4 t 3

3b1 9b2 6b3 7b4 15
b1 , b2 , b3 , b4 t 0
(4)
Minimize f (b) 10b1 b2
subject to b1 5b2 t 8
4b1 b2 t 10
b1 d 0
3.11 If the following problems have solutions, obtain them by the graphical
method. Otherwise, explain why there are no solutions.
(1)
subject to 10b1 3b2 d 3
5b1 7b2 1
b1 3b2 7
b1 , b2 t 0
(2)
b1 , b2 t 0
(3)
subject to 2b1 6b2 14
b1 3b2 7
b1 5b2 d 10
3.12 Find the basic solutions to the following problems and show them by the
graphical method:
(1)
Maximize f (b) b1 2b2
subject to 2b1 b2 t 4
b1 2b2 d 3
2b1 3b2 5
b1 , b2 t 0
(2)
Maximize f (b) 3b1 2b2
subject to 7b1 b2 t 14
b1 4b2 t 5
b1 t 0
(3)
Minimize f (b) 9b1 8b2 2b3
subject to b1 5b2 4b3 d 5
b1 2b2 3b3 t 7
b1 , b2 , b3 t 0
3.13 Find the minimum point of f (b) by the equal interval search. The interval
of uncertainty for convergence is 0.001 and G 0.02 . b 1 is the initial
design. Use programming languages such as C or FORTRAN if needed.
f (b) 4b 2 5b 9
3.14 Solve Exercise 3.13 by using the golden section method.
3.15 Convert the following functions to functions of D in order to perform a

one-dimensional optimization for D :
(1) f (b) 2b12 2b1 b23 1 , b [0 1]T , d [ 2 3]T

(2) f (b) cos b1 b22 4b1 , b [S 3]T , d [ 4 6]T
(3) f (b) b12 3b1 7b23 3b2 1 , b [1 0]T , d [1 3]T
(4) f (b) (b1 4) 2 (b2 3) 2 (b3 2) 2 , b [0 0 0]T ,

d [8 6 4]T
3.16 Solve the problems in Exercise 3.15 by using the golden section method.
The solution process can be programmed. G 0.02 and the interval of
uncertainty for convergence is 0.001.
3.17 Repeat two iterations of the steepest descent method to minimize the
following functions:
(1) f (b1 , b2 ) b12 2b22 2b1b2 4b1 1 , b ( 0) [1 0]T

(2) f (b1 , b2 ) (b1 b22 ) 2 (5b2 1) 2 , b ( 0) [1 1]T
(3) f (b1 , b2 ) (b1 1) 2 / 36 (b2 3) 2 / 4 , b ( 0) [ 1 1]T
3.18 Solve Exercise 3.17 by using the conjugate gradient method.
3.19 Solve Exercise 3.17(3) by using the steepest descent method and the
conjugate gradient method. Continue the iteration until the interval of
uncertainty is less than 0.001. Compare the histories of the two processes.
3.20 Solve Exercise 3.17 by using the DFP method.
3.21 Solve Exercise 3.17 by using the BFGS method.
3.22 Normalize the following constraints and obtain the İ-active constraints.
Use H 0.1 and the current design point is b [3.3 3.3]T .
g 1 (b) b1 3.3 d 0
g 2 (b) b2 3.5 d 0
g 3 (b) b12 12.25 d 0
g 4 (b) b1b2 12 d 0
g 5 (b) b2 d 0
g 6 (b) b1 b22 10 d 0
g 7 (b) b2 b1 2 d 0
3.23 In the following problems, make a linearized subproblem at the given

point:
(1)
Minimize f (b) 2b1 b2
subject to b12 3b1b2 2b22 1 t 0
b [3 1]T
(2)
Minimize f (b) b12 2b22 4b1 b2 1
3b1 7b2 d 10
b1 2b2 t 0
b [1 1]T
(3)
Minimize f (b) (b1 1) 2 (b2 1) 2
subject to b1 b22 2 d 0
b1 2b2 d 0
b [4 5]T
3.24 Make subproblems for Exercise 3.23 to use the sequential linear
programming (SLP) algorithm. The move limit is r 0.5.
3.25 Make subproblems to use the sequential quadratic programming (SQP)

algorithm for the following problems:
(1)
Minimize f (b) (b1 2) 2 (b2 3) 2
subject to b1 b22 d 6
b1 , b2 d 0
b ( 0) [0 1]T
(2)
Minimize f (b) b12 2b22 4b1b2 3b2 1
subject to 4b13 b22 d 0
b1 2b22 d 0
b (0) [1 3]T
3.26 Define the inverse barrier function to solve the following problem by the
transformation method. Draw the contour of the inverse barrier function
according to the parameter r. Explain the phenomena around the
boundaries of the design variable.
Minimize f (b) b 2 2b 5
subject to 3dbd7
3.A Brief Review of Mathematical Terminologies and

Background
Feasible region: The feasible region is the design range where the constraints in
Equations 3.1cí3.1e are satisfied within the design space. It is expressed by the
set S given as
S {b | hi (b) 0, i 1, , l g j (b) 0, j 1, , m} (3.A.1)
A design solution must be in the feasible region and thus the optimum solution
should be found in the feasible region. That is, the optimum solution makes the
objective function have the minimum within the feasible region. Suppose we have
the following constraint set:
S1 {b | g1 b1 b2 1 d 0 g 2 b12 b22 2 d 0} (3.A.2)
Figure 3.A.1 illustrates the above constraint functions. The region satisfying each
constraint is the opposite side of the oblique lines. Therefore, the feasible region is
the shadowed area. In Figure 3.A.1, the design points A, B and C are in the
feasible region and the points D, E and F are outside the feasible region.
Global (absolute) minimum: At a certain design point b* , the objective function
has the global (absolute) minimum if the following condition is satisfied for any
design point b in the feasible region:
b2
2 F
1 E
g 2 { b12 b22 4 0
C
1 2 b1
B A
g1 { b1 b2 1 0
Figure 3.A.1. An example of the feasible region

f (b* ) d f (b) (3.A.3)
where, if the inequality (<) is always satisfied, the minimum is called the strict
global minimum.
Theorem 3.A.1 [Weierstrass Theorem – The Existence of

the Global Minimum]
If f (b) is continuous on a nonempty feasible set S, which is closed and
bounded, then f (b) has a global minimum in S.
Local minimum: The neighborhood N of b * in a feasible region S is defined by
the following set of points:
N {b | b S , | b b * | G } (3.A.4)
where G ! 0 is some small number. If Equation 3.A.3 is satisfied in the

neighborhood N in Equation 3.A.4, b * is called a local minimum. Figure 3.A.2
presents local minima and maxima for a single design variable. In the feasible
region, the points A, C and E are local maxima, and the points B, D and F are local
minima. The point C is the global maximum and the point F is the global
minimum.
Gradient: The gradient of a function f (b) ( f ) is defined as
T
ª wf wf wf º
f « » (3.A.5)
¬ wb1 wb2 wbn ¼
where T indicates the transpose vector. Geometrically, the gradient vector is a
f(b) C
A
B
D
F
Feasible region
Figure 3.A.2. Maxima and minima

normal vector to the tangent plane of a hyper-plane made from f constant.

Suppose we have the following function:
f b12 b22 2 0 (3.A.6)
The gradient vector is
T
ª wf wf º
f « » [2b1 2b2 ]T (3.A.7)
¬ wb1 wb2 ¼
The geometric representation is illustrated in Figure 3.A.3. Figure 3.A.3

presents the tangent line and the normal vector at (b1 , b2 ) (1.0, 1.0) when f 0.
Hessian matrix: The Hessian matrix is a matrix with second-order derivatives as
follows:
ª w2 f w2 f w2 f º
« 2
»
« wb1 wb1 wb2 wb1 wbn »
« w f
2
w2 f w2 f »
w2 f « »
H « wb2 wb1 wb22 wb2 wbn » (3.A.8)
wbwb
« »
« 2 2 2 »
« w f w f

w f »
« wbn wb1 wbn wb2 wbn2 »¼
¬
The matrix is symmetric due to the characteristics of the partial derivatives.

Taylor series: The Taylor series is a series of polynomial functions that
b2
f [ 2 2] T
1
b1
1 Tangent line (plane)
g2 b12 b22 2 0
Figure 3.A.3. Gradient vector

approximate a function f around b * as follows:
1
f (b) f (b * ) T f (b * )(b b * ) (b b * ) T H(b b * ) R (3.A.9)
2
where R is the remainder that is considered as an error. The Taylor series is

utilized to approximate a nonlinear function as a polynomial function. Mostly,
linear or quadratic approximation is used. For example, the linear approximation
is as follows:
f (b) # f app (b) f (b * ) T f (b * )(b b * ) (3.A.10)
where Gf f T (b * )(b b * ) is the linear increment of f according to the change

of b. The increment has the maximum when the directions of T f (b * ) and
(b b * ) are identical. Therefore, the direction of the gradient vector will
produce a maximum. It is noted that the direction in T f (b * ) is the minimum
increment or maximum reduction direction.
Look at the characteristics of the Hessian matrix H. In Equation 3.A.9, the
matrix is called positive definite if x T Hx ! 0 is satisfied at any x ( x z 0 ). If there
is an x ( x z 0 ) that makes x T Hx 0 and xT Hx t 0 for any x, then it is called
positive semidefinite. This aspect is very important in mathematics as well as in
mechanics.
Convex set: A convex set S in the n-dimensional space is a collection of points
having the following property: If two points x1 and x 2 are any points in S, then
the line segment between the two points x ( x Dx1 (1 D )x 2 (0 d D d 1) ) is in S.
Examples of convex sets are illustrated in Figure 3.A.4.
Convex function: A convex function is defined on a convex set. For x in a
convex set, if the following inequality condition is satisfied for 0 d D d 1, then f(x)
is a convex function:
f (Dx1 (1 D )x 2 ) d Df (x1 ) (1 D ) f (x 2 ) (3.A.11)
The following theorem is derived from convexity:
Theorem 3.A.2 [Convexity of a Function]
A function of x R n , f(x) defined on a convex set S is convex if and only

if the Hessian matrix of the function is positive semidefinite or positive
definite at all points in the set S. If the Hessian matrix is positive definite
for all points in the set, then f is called a strictly convex function. (Note that
the converse of this is not true, i.e. a strictly convex function may only have
a positive semidefinite Hessian at some points, e.g. f ( x) ( x 1) 4 is a

strictly convex function but its second derivative is zero at x 1. )
Theorem 3.A.3 [Convex Function and Convex Set]

Let a set S be defined with constraints of the general optimization problem
of Equation 3.1 as
S {b | hi (b) 0, i 1, , l g j (b) d 0, j 1, , m} (3.A.12)
Then S is a convex set if functions g j are convex and hi are linear.
Theorem 3.A.4 [Global Minimum]
If f(b) has a local minimum at b * and f (b) is a convex function defined

on a convex set S, then it is also a global minimum.
Active constraint: An inequality constraint g j (b * ) d 0 is said to be active at a

design point b * if it is satisfied at equality g j (b * ) 0 . For a feasible design, an
inequality constraint may or may not be active. However, all equality constraints
are active for all feasible designs. In Figure 3.A.1, constraint g 2 is active at the
point B and g1 is active at the point C.
x2 x2
1
x1
x
(a) Convex set
x1 x2
x1 x2
(b) Nonconvex set
Figure 3.A.4. Convex and nonconvex sets

Regular point: Assume that we have only equality constraints in the optimization
problem in Equation 3.1. If all the equality constraints are satisfied with
hi (b * ) 0, i 1, ..., l at point b * and the gradient vectors of the equality
constraints are linearly independent, then the point b * is said to be a regular point.
When inequality constraints exist, the active constraints are considered as equality
constraints and the gradient vectors of active constraints should be linearly
independent of each other at a regular point.
References
Arora JS (2004) Introduction to Optimum Design. 2nd ed., Elsevier, San Diego,
CA
Bazaraa MS, Sherali HD, Shetty CM (1993) Nonlinear Programming. Wiley, New
York
Belegundu AD, Chandrupatla TR (1999) Optimization Concepts and Applications
in Engineering. Prentice Hall, New Jersey
Dantzig G (1963) Linear Programming and Extensions. Princeton University Press,
Princeton, New Jersey
Fiacco AV, McCormic GP (1968) Nonlinear Programming: Sequential
Unconstrained Minimization Techniques. Wiley, New York
Fletcher R, Reeves CM (1964) Function Minimization by Conjugate Gradients.
Computer Journal 7:149í154
Fletcher R (1970) A New Approach to Variable Metric Algorithms. Computer
Journal 13:317í322
Goldberg DE, Samtani MP (1986) Engineering Optimization via Genetic
Algorithm. Proceedings of the 9th Conference on Electronic Computation,
ASCE, pp 471í482
Haftka RT, GĂrdal Z (1992) Elements of Structural Optimization. Kluwer,
Dordrecht
Haug EJ, Arora JS (1979) Applied Optimal Design. Wiley, New York
Papalambros PY, Wilde DJ (2000) Principles of Optimal Design: Modeling and
Computation. 2nd ed., Cambridge University Press, Cambridge
Parmee IC, Hajela P (eds) (2002) Optimization in Industry. Springer, Berlin
Heidelberg New York
Polak E (1971) Computational Methods in Optimization. Academic Press, New
York
Powell MJD (1964) An Efficient Method for Finding the Minimum of a Function
of Several Variables without Calculating Derivatives. Computer Journal
7:303í307
Rao SS (1996) Engineering Optimization. 3rd ed., Wiley, New York
Rosen JB (1960) The Gradient Projection Method for Nonlinear Programming,
Part I, Linear Constraints. SIAM Journal Appl Math 8:181í217
Rozvany GIN (1997) Topology Optimization in Structural Mechanics. Springer,
Berlin Heidelberg New York
Rutenbar RA (1989) Simulated Annealing Algorithms: An Overview. IEEE

Circuits and Devices:19í26
Schmit LA (1981) Structural Synthesis–Its Genesis and Development. AIAA
Journal 19:1249í1263
Vanderplaats GN (1984) Numerical Optimization Techniques for Engineering
Design with Applications. McGraw–Hill, New York
Venkayya VB (1978) Structural Optimization–Recent Developments and
Applications. Int J Num Meth Eng 13:203í228
Zangwill WI (1969) Nonlinear Programming: A Unified Approach. Prentice-Hall,
New Jersey
Zoutendijk G (1960) Methods of Feasible Directions. Elsevier, Amsterdam,
Netherlands
4 Structural Optimization
4.1 Introduction
In structural optimization, optimization theory is applied to the design of structures.
In Chapter 3, it was emphasized that a design problem should be formulated as in
Equation 3.1 in order to apply optimization theory to practical design. It may not
be easy to precisely define a design problem as in Equation 3.1. Especially, the
equality constraints in Equation 3.1c are difficult to define and manipulate.
Nevertheless, the optimization method is frequently utilized in structural design
because the formulation in Equation 3.1 can incorporate the objective and the
constraint conditions of design.
Structural optimization has been employed to reduce the weight of airplane
structures. Aerospace engineering is the field where structural optimization is
applied most extensively. Recently, application has expanded to other fields.
There are several reasons why optimization is actively applied to structural design.
First, the design process for structures is quite similar to the optimization
process. Reduction of weight is important in structural design and it is equivalent
to the objective function of optimization in that the reduction is similar to
minimization of the objective function. When maximization or minimization of a
specific response is pursued, it can be treated as an objective function in
optimization. Also, various design conditions for structural design can be easily
transformed into constraints in optimization.
Second, the finite element method (FEM) can be exploited in structural
optimization. In optimization, the objective function and constraints should be
defined in definite forms and they must be evaluated precisely. FEM is an analysis
method that facilitates this aspect and it is one of the computational methods that
are well developed for analysis of a system. Since many commercial systems are
available, it is easy to interface an optimization system with the systems of FEM.
The optimization theories in Chapter 3 are directly used in structural
optimization. The process to find optimum sizes of structures is explained by the
primal method in Section 3.8. With the current sizes, we evaluate the total weight
of a structure (objective function) and the stresses (constraints). A linearized
subproblem is defined based on the evaluation. A direction vector is calculated
from the subproblem and the stepsize is determined. Sensitivity information is

needed to make the subproblem. In structural optimization, sensitivity analysis is
carried out by using the equilibrium equation (generally formed by FEM). Details
will be explained later.
There are two approaches in structural optimization. One is the direct method
and the other is the indirect method. In the direct method, an optimization
algorithm directly uses the responses from FEM for the functions in Equation 3.1.
The functions in Equation 3.1 are approximated to explicit functions of design
variables or parameters in the indirect method. An optimization algorithm uses the
approximated functions. The functions in Equation 3.1 are replaced by linear or
nonlinear functions. The approximation process is conducted based on the
characteristics of the responses. Sensitivity information is frequently required in
the approximation process.
As mentioned earlier, FEM is utilized to obtain the objective function and
constraints. FEM can be extremely expensive for large-scale structures, especially
the sensitivity analysis. The process of sensitivity analysis can be directly
implemented in FEM software development. Much effort is needed in the
implementation, but the resulting numerical cost for the analysis is relatively low.
The finite difference method can be used for the sensitivity analysis. In this case,
the implementation cost is low, but the numerical cost is relatively high because
the finite element analysis is repeatedly utilized. In structural optimization, it is
crucial to reduce the numbers of the finite element analyses and sensitivity
analyses.
Structural optimization is classified depending upon the characteristics of the
design variables. They are size optimization, shape optimization and topology
optimization. In size optimization, the domain of FEM is not altered by the design
variables. The design variables for shape optimization are defined by the domain
of FEM. The distribution or existence of materials serves as design variables in
topology optimization. The theories of size and shape optimization are similar, but
topology optimization uses quite a different theory.
Structural optimization can be classified into continuous design and discrete
design based on the characteristics of the design space. In continuous design,
design variables are defined in a continuous space and they can have any real
values in the design range. In practical design, the design variables may only have
some specific set of values. This is referred to as discrete design. In this chapter,
only continuous design is considered and discrete design will be discussed in
Chapter 6.
4.2 Finite Element Method

FEM is an analysis method to solve a governing equation that is usually expressed
by a differential equation. FEM is generally used for stress analysis, vibrational
analysis, etc. in solids. It has been expanded to the areas of heat transfer and fluid
mechanics. FEM is not explained in detail and it is introduced from the viewpoint
of structural optimization.
Structural Optimization 173
Vy
W yz
W yx
W zy W xy
W xz
Vx
Vz W zx
z x
Figure 4.1. Stress components
4.2.1 Stress and Strain
Stress is the reaction force per unit area. Each stress component in Figure 4.1
yields the following stress vector:
ı [V x V y V z W xy W yz W zx ]T (4.1)
At a point in the space, the stresses in Equation 4.1 should satisfy the equilibrium
equation in Equation 4.2
wV x wW xy wW xz
Fx 0 (4.2a)
wx wy wz
wV y wW xy wW yz
Fy 0 (4.2b)
wy wx wz
wV z wW xz wW yz
Fz 0 (4.2c)
wz wx wy
where Fx , F y , and Fz are components of the body force such as the gravitational
force.
Suppose u [u v w]T is a displacement vector at a point. Then the strains
are
wu
Hx (4.3a)
wx
wv
Hy (4.3b)
wy
ww
Hz (4.3c)
wz
wu wv
J xy J yx (4.3d)
wy wx
wv ww
J yz J zy (4.3e)
wz wy
ww wu
J zx J xz (4.3f)
wx wz
In vector form
İ [H x H y H z J xy J yz J zx ]T (4.4)
The stress–strain relationship is
ı Eİ (4.5)
where E is a matrix defined by Young’s modulus and Poisson’s ratio.
4.2.2 Formulation of the Finite Element Method

In the finite element analysis for solids, a continuous body is divided into finite
elements, and displacements and stresses of elements are evaluated by using solid
mechanics. The steps for the finite element method are usually as follows:
Step 1. The body for stress analysis is divided into various elements as
illustrated in Figure 4.2.
Step 2. The equilibrium equation is applied to each element and transformed
to a matrix equation by the displacements of nodes in the element. The
equation for each element is assembled to form a large equation for the
entire structure. This is a simultaneous equation where the unknowns are
the displacements of the nodes in elements in Figure 4.3.
Step 3. The solution of the simultaneous equation is obtained. The solution
includes the displacements at the nodes.
Step 4. Other responses such as stresses are obtained from the displacements.
Other than the above process, there are many other methods or processes in the
finite element method.
Figure 4.2. A structure defined by finite elements
The stiffness matrix is made in Step 2. There are a few methods to make a
stiffness matrix as follows:
(1) The matrix method: A simultaneous equation is defined from the
equilibrium equation.
(2) The principle of minimum potential energy: The principle that the
condition of minimum potential energy satisfies the equilibrium equation is
utilized.
(3) The principle of virtual work: The principle of virtual work is utilized. It
indicates that the sum of the work by infinitesimal virtual displacements
satisfying the boundary conditions and the internal strain energy is zero.
Table 4.1 shows the processes for making the element stiffness matrix by the
above methods. The matrix method can only make stiffness matrices for simple
elements such as truss, beam and triangular elements. The principle of minimum
potential energy or the principle of virtual work should be utilized to attain the
stiffness matrices for modern complicated elements. It is noted that the stiffness
matrices for truss, beam and triangular elements have explicit expressions.
However, the stiffness matrix is generally expressed by functionals as follows:
T
kL ³ B EB dV (4.6)
V
where V is the domain of each element. As shown in Table 4.1, the equilibrium
equation of each element is
k LzL fL (4.7)
(a) Truss or beam element
(b) Rectangular element
(c) Triangular element
(d) Hexahedral solid element
(e) Tetrahedral solid element
Figure 4.3. Elements and nodes

Table 4.1. Formulation process of the finite element method (Choi 2002)
Matrix method Principle of minimum potential energy Principle of virtual work

1. By using the shape function N, the displacement field in an element is approximated as follows:
u Nz L
where z L R ec is a displacement vector at the nodes of an element
2. The above equation is differentiated according to coordinate variables and matrix B is obtained. The strain–displacement
relationship is defined as
İ Bz L
3. By using matrix E, which characterizes the material properties, the stress field is obtained as
ı Eİ EBz L Sz L
4. The relationship 4. The potential of external forces is 4. Using the shape function N, virtual displacements
equation between ı and obtained as and virtual strains are
f L is defined as V zTL P u Nz L , İ B z L
f L Aı 5. The strain energy is 5. The potential of the forces by the virtual
T T displacement is
where ı is the stress field U ³ ı İ dV ³ İ Eİ dV
on the boundary of an V V V z TL f L
element and f L is the 1
z TL ³ B T EB dV z L = zTL k L z L
equivalent nodal force V 2
vector T
where k L ³ B EB dV
V
Structural Optimization
177
178
Table 4.1. Formulation process of the finite element method (Choi 2002) (continued)
5. From the relationship 6. The functional of the total potential 6. The strain energy by the virtual strain is evaluated as
between ı and the nodal energy is obtained from the sum of the U ³ ı T İ dV ³ İ T Eİ dV
displacement vector z L , deformation energy and the potential of V V
the relationship between the external forces z TL ³ B T EB dV z L z TL k L z L

f L and the nodal displace- 7. The functional is minimized. It is V
differentiated with respect to z L as T
ments is obtained as where k L ³ B EB dV
follows: V
f L AEBz L k L z L
w3 P 7. Use the principle of virtual work
where k L AEB 0 U V 0
Analytic Methods for Design Practice
wz L
T
z L k L z L z TL f L 0
dz TL k L z L dz TL f L 0
If z TL is eliminated
If dz TL is eliminated
k LzL fL
k Lz L fL
where z L R l ' is the nodal displacement vector at each element and l ' is the
degrees of freedom for each element. The vector f L is a vector of the external
loads imposed on the nodes.
The equilibrium equation of the entire structure is obtained by assembling the
element equilibrium equation in Equation 4.7. Equation 4.7 is defined in the
global coordinate. If the boundary condition is applied to the assembled equation,
the equation yields
Kz f (4.8)
where z R l is a displacement vector for all the degrees of freedom and l is the
total degrees of freedom. K is an (l u l ) matrix and f is an (l u 1) column vector
and they are the stiffness matrix and the external load vector for the entire structure,
respectively.
The governing equation for vibration of structures can be formulated in the
finite element method as follows:
Ky [My 0 (4.9)
where K is the stiffness matrix element, M is the (l u l ) mass matrix, y is the

(l u 1) eigenvector and [ is the natural frequency or eigenvalue. The method for
obtaining the mass matrix M is not explained.
When the finite element method is numerically implemented, the Galerkin
method, which is a weighted residual method, is frequently used. The Galerkin
method only uses the equilibrium equation (usually a differential equation, not
functionals). Therefore, the Galerkin method is convenient because we do not
have to use the functionals. The results of the Galerkin method are the same as
those from the principle of minimum potential energy or the principle of virtual
work. Details of the finite element method are explained in many books (Bathe
1996, Cook et al. 1989).
Example 4.1 [Symmetric Three Bar Truss Structure]

In Figure 3.3, l 1 m, T 60q and the load P 40,000 N. Young’s modulus is
E 207 GPa. Perform the stress analysis for the three bar truss by the finite
element method. The areas of the cross sections are b1 b3 4.0 u 10 4 m 2 and
b2 2.0 u 10 4 m 2 .
Solution
The stiffness matrix k iL of the ith element in the local coordinate system is
bi E ª 1 1º
k iL « 1 1 » (4.10)
Li ¬ ¼
where Li is the length of the ith member. Equation 4.10 should be transformed to
the stiffness matrix in the global coordinate system. For the transformation, the
following transformation matrix T is utilized:
ªcos D sin D 0 0 º
T « 0 (4.11)
¬ 0 cos D sin D »¼
where D is the angle between a truss element and the x-axis of the global
coordinate system.
Using the transformation matrix T, the matrix in Equation 4.10 is transformed
as follows:
ki TT k iL T (4.12)
where k i is the stiffness matrix in the global coordinate system for the ith element.
Using Equation 4.12, the product of the stiffness matrix and the displacement
vector in the global coordinate system for each member is as follows:
ª § 3S · º
«cos ¨ 4 ¸ 0 »
« © ¹ »
« sin § 3S · »
¨ ¸ 0
b1 E « © 4 ¹ » ª 1 1º
k 1z 1 « »
L1 « § 3S ·» «¬ 1 1 »¼
0 cos ¨ ¸
« © 4 ¹»
« § 3S · »
« 0 sin ¨ ¸»
¬« © 4 ¹ ¼»
ª § 3S · § 3S · º ª u1 º
«cos ¨ 4 ¸ sin ¨ 4 ¸ 0 0 »« v »
« © ¹ © ¹ »« 1 »
« § 3S · § 3S ·» «u 2 »
« 0 0 cos ¨ ¸ sin ¨ ¸» « »
¬ © 4 ¹ © 4 ¹ ¼ ¬ v2 ¼
ª 1 1 1 1º ª u1 º
« »« »
b1 E « 1 1 1 1» « v1 »
(4.13)
2 L1 « 1 1 1 1 » «u 2 »
« »« »
¬ 1 1 1 1 ¼ ¬ v2 ¼
ª § S· º
«cos ¨ 2 ¸ 0 »
« © ¹ »
« sin § S · 0 »
¨ ¸
b2 E « © 2 ¹ » ª 1 1º
k 2z 2 « »
L2 « § S · « 1 1 »¼
0 cos ¨ ¸» ¬
« © 2 ¹»
« § S ·»
« 0 sin ¨ ¸ »
¬« © 2 ¹ »¼
ª § S· § S· º ª u1 º
«cos ¨ 2 ¸ sin ¨ 2 ¸ 0 0 »«v »
« © ¹ © ¹ »« 1 »
« § S· § S ·» «u3 »
« 0 0 cos ¨ ¸ sin ¨ ¸» « »
¬ © 2¹ © 2 ¹¼ ¬ v3 ¼
ª0 0 0 0 º ª u1 º
« »« »
b2 E «0 1 0 1» « v1 »
(4.14)
L 2 «0 »
0 0 0 » «u 3 »
« « »
«¬0 1 0 1 »¼ ¬ v3 ¼
ª § S· º
«cos ¨ 4 ¸ 0 »
« © ¹ »
« sin § S · 0 »
¨ ¸
b3 E « © 4 ¹ » ª 1 1º
k 3z 3 « »
L3 « § S · « 1 1 »¼
0 cos ¨ ¸» ¬
« © 4 ¹»
« § S ·»
« 0 sin ¨ ¸ »
«¬ © 4 ¹ »¼
ª § S· § S· º ª u1 º
«cos ¨ 4 ¸ sin ¨ 4 ¸ 0 0 »« v »
« © ¹ © ¹ »« 1 »
« § S · § S · «u »
« 0 0 cos ¨ ¸ sin ¨ ¸»» « 4 »
¬ © 4¹ © 4 ¹ ¼ ¬ v4 ¼
ª 1 1 1 1 º ª u1 º
« »« »
b3 E « 1 1 1 1» « v1 »
(4.15)
2 L3 « 1 1 1 1» «u 4 »
« »« »
¬ 1 1 1 1 ¼ ¬v4 ¼
where z i is the displacement vector of the ith element in the global coordinate
system. The product of the global stiffness matrix K and the displacement z is
obtained by assembling the element stiffness matrices as follows:
ª b1 b3 b1 b b1 b1 b3 b3 º ª u 1 º
« 2L 0 2L 0 3 00 00
« 1 3 2 L1 2 L3 2 L1 2 L1 2 L3 2 L3 »» « »
« »
« b1 0 b3 b1 b
2 3
b
1
b
1
b
00 0
b2 b3 b
3 »» « v1 »
« 2L 2 L3 2 L1 L 2 2 L3 2 L1 2 L1 L2 2 L3 2 L3 « »
« 1 »« »
« b b b1 b1
1
1 0 0 0 0 » «u 2 »
« 2 L1 2 L1 2 L1 2 L1 »« »
« b b b1 b1 »« »
« 1 1 0 0 0 0 »« »
Kz E« 2 L1 2 L1 2 L1 2 L1 » «v 2 »
« 00 00 0 0 0 0 0 0 » «u 3 »
« b b2 »« »
« 00 0 2 0 0 0 0 0 » « v3 »
« L2 L2 »« »
« b b3 b3 b3 » « »
« 3 0 0 0 0
2 L3 » « 4 »
u
2 L3 2 L3 2 L3
« »
« b3 b b3 b3 » «« »»
3 0 0 0 0
« 2 L3 2 L3 2 L3 2 L3 »¼ ¬« v 4 ¼»
¬
(4.16)
If we apply the boundary conditions, components are eliminated by lines as

shown in Equation 4.16. Then Equation 4.16 yields
ª2 0 º ªu1 º
Kz 2 u 10 4 u 207 u 10 9 « »« » (4.17)
¬0 2 2 ¼ ¬ v1 ¼
The external load f is
ª 40000 u 1 / 2 º
f «40000 u 3 / 2» (4.18)
¬ ¼
Substituting Equations 4.17 and 4.18 into Equation 4.8, the following equilibrium
equation is obtained:
ª2 0 º ªu1 º ª 40000 u 1 / 2 º
2 u 10 4 u 207 u 10 9 « « » (4.19)
¬ 0 2 2 »¼ ¬ v1 ¼ «40000 u 3 / 2»
¬ ¼
The simultaneous equation in Equation 4.19 yields
ªu1 º ª3.4160 u 10 4 º
z «v » « 4 »
(4.20)
¬ 1¼ «¬3.4659 u 10 »¼
The stress ı in a member is defined as

ı EBz L (4.21)
where B is the matrix with a differential operator and z L is the displacement

vector of a member in the local coordinate system. The stress of a truss member
can also be obtained by the following equation:
V p/b (4.22)
where b is the area. Equation 4.22 gives the same solution as Equation 4.21.
Equation 4.22 is utilized here for stress calculation.
Equation 4.23 is the equilibrium equation in the local coordinate for the truss
element in Figure 4.4
ª p1 º bE ª 1 1º ª z L1 º
«p » « »« » (4.23)
¬ 2¼ L ¬ 1 1 ¼ ¬ z L 2 ¼
Equation 4.23 is solved as
bE
p2 p1 [ 1 1]z L (4.24)
L
If we substitute Equation 4.24 into Equation 4.22, the stress is
E
V [ 1 1]z L (4.25)
L
Equation 4.25 is a detailed expression of Equation 4.21. Equation 4.20 shows

the displacement vector z in the global coordinate system. We need the
displacement vector z L in the local coordinate system in Equation 4.25. The
relationship between z and z L is
zL Tz (4.26)
where the transformation matrix T represents the relationship between the two
coordinate systems as shown in Equation 4.11. By substituting Equation 4.26 into
Equation 4.25, the stress can be obtained by the displacement vector in the global
zL2 z L1
p2 p1
Figure 4.4. Degrees of freedom for a truss element

coordinate system as follows:
E
Vi [ 1 1]Tz i (4.27)
L
Using Equation 4.27, the stresses of the members are
4
ª § 3S · § 3S · º ª3.4160 u 10 º
E «cos ¨ 4 ¸ sin ¨ 4 ¸ 0 0 » «3.4659 u 10 4 »
V1 [ 1 1]« © ¹ © ¹ »« »
L1 « § 3S · § 3S ·» « 0 »
« 0 0 cos ¨ ¸ sin ¨ ¸» « »
¬ © 4 ¹ © 4 ¹¼ ¬« 0 ¼»
71.227 MPa (4.28)
4
ª § S· § S· º ª3.4160 u 10 º
E
cos
« ¨ 2¸ sin ¨
2
¸ 0 0 «
» 3.4659 u 10 4 »
V2 [ 1 1]« © ¹ © ¹ »« »
L2 « § S· § S ·» « 0 »
« 0 0 cos ¨ ¸ sin ¨ ¸» « »
¬ © 2¹ © 2 ¹¼ ¬« 0 ¼»
71.744 MPa (4.29)
4
ª § S· § S· º ª3.4160 u 10 º
E « cos ¨ ¸
4¹
sin ¨ ¸ 0 0 «
» 3.4659 u 10 4 »

V3 [ 1 1]« © © 4¹ »« »
L3 « § S· § S ·» « 0 »
« 0 0 cos ¨ ¸ sin ¨ ¸» « »
¬ © 4 ¹ © 4 ¹¼ «¬ 0 »¼
516.5 kPa (4.30)
The aforementioned analysis details the process for obtaining the stresses for a
truss structure. Generally, the first process in the finite element method is the
calculation of the displacements. Other responses such as strains and stresses are
obtained from the displacement information.
Example 4.2 [A Structure with Elements for Plane Stress]

Figure 4.5 shows a four node element for plane stress. Loads are imposed as
illustrated in Figure 4.5. Calculate the stress of the element. The thickness is
h 1 mm, Young’s modulus is E 200 GPa and Poisson’s ratio is Q 0.25.
Solution
The stiffness matrix of the four node element with thickness h is
kL h ³ B T EB dxdy (4.31)
100kN
4 3 100kN
y 2m
x
1 2
2m
Figure 4.5. Four node element
The matrix E in Equation 4.31 is
ª º
«1 v 0 » ª1.06667 0.26667 0 º
E «
E v 1 0 » E ««0.26667 1.06667 0 »» (4.32)
1 v2 « 1 v» «¬ 0
«0 0 » 0 0.4»¼
¬ 2 ¼
Using the s–t natural coordinate system within an element, Equation 4.31 is
11
kL h ³ ³ BT ( s, t )EB( s, t ) | J ( s, t , x, y ) | dsdt (4.33)
11
where the natural coordinate system is an artificial coordinate system to define the
shape function in an isoparametric formulation. The matrix J is the Jacobian
matrix between the s–t coordinate and the x–y coordinate systems. Details are
explained in the references for the finite element method. Using four point
Gaussian quadrature, the integration of Equation 4.33 is approximated as
4
k L # h ¦ B T ( s i , t i )EB( s i , t i ) J ( s i , t i , x, y ) (4.34)
i 1
From Equation 4.34, it is noted that thickness h is the only property that can be a
design variable in size optimization. It is separated from the other terms.
Therefore, partial differentiation with respect to h is easy and the sensitivity
analysis is relatively simple.
For this problem, the element stiffness matrix is
ª 0.488667 0.166667 0.288667 0.033 0.244333 0.166667 0.044333 0.033 º

« 0.166667 0.488667 0.033 0.044333 0.166667 0.244333 0.033 0.288667»»
«
« 0.288667 0.033 0.488667 0.166667 0.044333 0.033 0.244333 0.166667 »
« »
« 0.033 0.044333 0.166667 0.488667 0.033 0.288667 0.166667 0.244333»
kL Eh
« 0.244333 0.166667 0.044333 0.033 0.488667 0.166667 0.288667 0.033 »
« »
« 0.166667 0.244333 0.033 0.288667 0.166667 0.488667 0.033 0.044333 »
« 0.044333 0.033 0.244333 0.166667 0.288667 0.033 0.488667 0.166667»
« »
«¬ 0.033 0.288667 0.166667 0.244333 0.033 0.044333 0.166667 0.488667 »¼
(4.35)
The stiffness matrix in the global coordinate system can be obtained by using
Equation 4.12. Since the transformation matrix T is the unit matrix and we have
one element in this problem, the element stiffness matrix in Equation 4.35 is
identical to the assembled matrix in the global coordinate system. If we apply the
boundary condition, the final stiffness matrix is
ª 0.488667 0.166667 0.044333 0.033 º

« 0.166667 0.488667 0.033 0.288667 »»
K Eh « (4.36)
« 0.044333 0.033 0.488667 0.166667 »
« »
¬ 0.033 0.288667 0.166667 0.488667 ¼
Now we solve the following equilibrium equation to obtain displacements:
Kz f (4.37)
where K is the matrix from Equation 4.36 and
z [u2 v2 u3 v3 ]T (4.38)
f [0 0 100000 100000]T (4.39)
From Equation 4.37, the displacements are obtained as
z [4.0419 9.9649 4.0759 15.0012]T u 10 4 m (4.40)
Plane stresses are obtained. Stresses of different places in an element are

different for a plate or a shell element. Therefore, the stresses at all the nodes are
evaluated from Equation 4.21 and the average value is considered as the final
stress. The stress is frequently adopted from the stress at the center of an element
(s = t = 0). The latter one is used here. In this case, Equation 4.21 becomes
ı EBz (4.41)
If we substitute B(0,0) at the center, Equation 4.32 and Equation 4.40 into
Equation 4.41, the following stresses are obtained:
ª1.06667 0.26667 0 º
ı E ««0.26667 1.06667 0 »»
«¬ 0 0 0.4»¼
ª 0 º
« 0 »
ª 1 1 1 1 º« »
« 4 0 0 0 0 » « 4.0419 »
4 4 4 « »
« 1 1 1 1 » « 9.9649 »
« 0 0 0 0 » u 10 4 (4.42)
« 2 2 2 2 » « 4.0759 »
« 1 1 1 1 1 1 1 1 »« »
«¬ 2 » «15.0012»
4 2 4 2 4 2 4 ¼«
0 »
« »
«¬ 0 »¼
From Equation 4.42, the stresses are
ı [56.726 64.545 50.068]T MPa (4.43)
4.3 Formulation of Structural Optimization

To use the optimization theory in a design problem, the design problem should be
formulated as in Equation 3.1. In structural optimization, Equations 4.8 and 4.9
are treated as equality constraints. A structural design problem is formulated as
follows:
Find b R n , z R l , [ R1 (4.44a)
to minimize f (b, z, [ ) (4.44b)
subject to K (b)z f (4.44c)
K (b)y [M (b)y 0 (4.44d)
g j (b, z, [ ) d 0, j 1, , m (4.44e)
bL d b d bU (4.44f)
where Equation 4.44c is the equilibrium equation in the finite element method and
it is assumed to be linear. Equation 4.44d is the eigenvalue equation. In Equation
4.44d, M is the mass matrix in a vibration problem or the geometric matrix in a

buckling problem. Generally, it is normalized as follows:
y T My 1 (4.45)
Using an optimization algorithm, we can find the solution of a structural

optimization problem formulated as in Equation 4.44. The primal method in
Section 3.8.1 is generally used. Thus, structural optimization is described based on
the application of the primal method. However, aspects such as sensitivity analysis
are required in the other methods as well. z R l and [ R1 in Equation 4.44
are called state variables that can be obtained as responses from the finite element
method. They are different from the design variable vector b R n . In many cases,
all the variables of b, z and [ are treated in the same manner. Then the number of
design variables is (n l 1).
A linearized subproblem in Equation 3.99 should be derived in the primal
method. The characteristics of structural optimization are exploited in this process.
The subproblem in the kth iteration is as follows:
Find Gb ( k ) , Gz ( k ) , G[ ( k ) (4.46a)
to minimize Tb fGb ( k ) Tz fGz ( k ) T[ fG[ ( k ) (4.46b)
subject to
w w
(Kz f ) (Kz f )Gb ( k ) (Kz f )Gz ( k ) 0 (4.46c)
wb wz
w w
(Ky [My ) (Ky [My )Gb ( k ) (Ky [My )G[ ( k ) 0 (4.46d)
wb w[
g j Tb g j Gb ( k ) Tz g j Gz ( k ) [T g j G[ ( k ) d 0, j 1, , m (4.46e)
Gb (Lk ) d Gb ( k ) d Gb (Uk ) (4.46f)
where Tb { w / wb , Tz { w / wz and [ { w / w[ . It is noted that b or z is a

column vector and w / wb or w / wz is a row vector.
Equations 4.46c and 4.46d represent n equality constraints, respectively. When
l is large, the number of design variables is quite large and the problem in Equation
4.46 becomes large. Thus, solving Equation 4.46 directly can be inefficient. In
structural optimization, Equation 4.46 is modified to have a small number of
design variables by eliminating the equality constraints. In other words, Gz ( k ) and
G[ ( k ) are expressed by Gb ( k ) from Equations 4.46c and 4.46d, and the results are
substituted into the inequality constraints in Equation 4.46e. The equality
constraints are eliminated by this process and the number of design variables is
reduced from (n + l + 1) to n.
We will investigate the elimination process of the equality constraints. At the
kth iteration, the solution z (k ) of the finite element method satisfies
Kz (k ) f 0 (4.47)
Since K and f are functions of the design variable vector b, the second term of
Equation 4.46c becomes
w § w (K~z ) wf · ( k )
(Kz f )Gb ( k ) ¨ ¸Gb (4.48)
wb © wb wb ¹
where ~z is a constant vector and the same as z (k ) obtained from Equation 4.47.
The third term of Equation 4.46c becomes
w
(Kz f )Gz ( k ) KGz ( k ) (4.49)
wz
From the rule of differentiation,
dz
Gz Gb (4.50)
db
If we substitute Equations 4.47í4.50 into Equation 4.46c,
§ w (K~z ) wf · ( k ) dz ( k )
¨ ¸Gb K Gb 0 (4.51)
© wb wb ¹ db
The following equation is obtained from Equation 4.51:
w (K~z ) wf dz
K 0 (4.52)
wb wb db
If we calculate dz/db (sensitivity) in Equation 4.52, the sensitivity analysis is

finished. If Equation 4.52 is expressed with respect to the ith design variable,
wK ( k ) wf dz
z K 0, i 1, , n (4.53)
wbi wbi dbi
As shown in Equation 4.53, sensitivity analysis requires wK / wbi . When there are
no body forces and wf / wbi 0 , Equation 4.53 yields
dz wK ( k )
K z , i 1, , n (4.54)
dbi wbi
Equation 4.46c is the same as Equation 4.54.

In the same manner, the last term of Equation 4.46d can be eliminated. In the
kth iteration,
Ky [ ( k ) My 0 (4.55)
The second term of Equation 4.46d can be expressed with respect to the ith design
variable bi as follows:
w ~ dy wK ~ wM
(Ky [My ) (K [ M ) y [ y, i 1, , n (4.56)
wbi dbi wbi wbi
~
where [ is a constant and the same as [ ( k ) . The third term of Equation 4.46d is
w
(Ky [My ) My (4.57)
w[
From the rule of differentiation,
d[
G[ Gb (4.58)
db
If we substitute Equations 4.55í4.58 into Equation 4.46d, Equation 4.46d yields
~ dy wK ~ wM d[
(K [ M ) y [ y My 0, i 1, , n (4.59)
dbi wbi wbi dbi
where y is the eigenvector. If we premultiply y T to Equation 4.59, the first term

of Equation 4.59 is eliminated and
§ wK ~ wM ·
y T ¨¨ [ ¸y
d[ © wbi wbi ¸¹
, i 1, , n (4.60)
dbi y T My
Equation 4.46d is the same as Equation 4.60. The above process is derived based
on the assumptions that the eigenvalues are not multiple roots and the eigenvectors
are not included in the objective function and constraints. The eigenvalue [
physically the natural frequency or the buckling load. When the eigenvalue has a
multiple root, it is known that Equation 4.60 is defined by directional derivatives.
Details are not introduced here.
Now the inequality constraints in Equation 4.46e are considered. If we
substitute Equation 4.50 and Equation 4.58 into Equation 4.46e,
wg j § wg j dz · ( k ) § wg j d[ · ( k )
Gb ( k ) ¨¨ ¸Gb ¨
¸
¸
¨ w[ db ¸Gb d g j , j 1, , m (4.61)
wb © wz db ¹ © ¹
If we substitute dz/db and dȟ/db from Equations 4.54 and 4.60 into Equation
4.61, then the equality constraints in Equations 4.46c and 4.46d are eliminated and
the inequality constraints in Equation 4.46e are functions of Gb ( k ) only. It is
essential to obtain the roots (dz/db and dȟ/db) of Equations 4.54 and 4.60 in the
optimization process. This is called sensitivity analysis and is a fairly expensive
process. The detailed process will be introduced later.
The sensitivity of eigenvectors can be derived as well. If we differentiate
Equation 4.45,
dy 1 wM
yT M yT y (4.62)
dbi 2 wbi
From Equations 4.59 and 4.62, the following relationship is derived:
ª dy º ª § wK wM · º
ªK [M My º « dbi » « ¨¨ [ ¸y »
« » « © wbi wbi ¸¹ »
« yT M (4.63)
¬ 0 »¼ « d[ » « 1 T wM »
«¬ dbi »¼ « 2 y wb y »
¬ i ¼
Sensitivity information for the eigenvectors is obtained by solving the

simultaneous equation in Equation 4.63. There are several other methods for
sensitivity analysis of the eigenvectors, but they are not introduced here.
The process introduced for sensitivity analysis is carried out after the discrete
formulation of the finite element is made for the governing equation. However,
sensitivity analysis can be performed before the process of the finite element
method. In this case, functional analysis is utilized and this will be explained later.
Example 4.3 [A Mathematical Problem Similar to Structural

Optimization]
The optimization problem in Equation 4.64 is to be solved by the primal method.
The initial values are b 0 1 , z 0 1 . At the first iteration, establish a
subproblem where the equality constraints are eliminated.
Find b, z (4.64a)
to minimize f b2 z (4.64b)
subject to g1 b 4z2 2 t 0 (4.64c)
h1 bz 0 (4.64d)
Solution
Using Equation 4.50, a subproblem to find Gb is established in the manner of
Equation 4.46.
Find Gb (4.65a)
§ dz ·
to minimize ¨ Tb f Tz f ¸Gb (4.65b)
© db ¹
§ dz ·
subject to g1 ¨ Tb g1 Tz g1 ¸Gb t 0 (4.65c)
© db ¹
§ dz ·
h1 ¨ Tb h1 Tz h1 ¸Gb 0 (4.65d)
© db ¹
In Equation 4.65d, h1 0 and Gb z 0 . Therefore,
§ T dz ·
¨ b h1 Tz h1 ¸ 0 (4.66)
© db ¹
dz
is calculated from Equation 4.66 as follows:
db
dz Tb h1 wh1 wh1

1 (4.67)
db Tz h1 wb wz
If we substitute 4.67 into Equations 4.65b and 4.65c, the following subproblem is
obtained:
Find Gb (4.68a)
to minimize ( Tb f Tz f )Gb (4.68b)
subject to 1 ( Tb g1 Tz g1 )Gb t 0 (4.68c)

Example 4.4 [Optimization of the Three Bar Truss]

The three bar truss in Example 3.1 is recalled. The objective function is the mass
of the structure and the stress of each member should not exceed 165 MPa .
Design variables are the cross sectional areas of the members. Formulate the
optimization problem by using the finite element method. To solve the problem
with the primal method, establish the subproblem of the first iteration by
eliminating the equality constraints. In this example, the structure does not
maintain symmetry and the density ( U ) is 7830kg/m 3 . The finite element
analysis of the structure is introduced in Example 4.1.
Solution
The optimization problem is formulated as follows:
Find b [b1 b2 b3 ]T (4.69a)
to minimize f Ul ( 2b1 b2 2b3 ) (4.69b)
subject to g1 V 1 d V all (4.69c)
g2 V 2 d V all (4.69d)
g3 V 3 d V all (4.69e)
Kz f (4.69f)
where
ª b1 b3 b1 b º
« 2L 2L 3
2 L1 2 L3 »»
K E« 1 3 (4.69g)
« b1 b3 b1 b b
2 3 »
« 2L 2 L1 L2 2 L3 »¼
¬ 1 2 L3
z [u1 v1 ]T (4.69h)
f [ f cos T f sin T ]T (4.69i)
V all 165 MPa (4.69j)
When the linearized subproblem is defined, Equation 4.54 is used for Equation
4.69f. Equation 4.54 is recalled as
dz wK ( k )
K z , i 1, , n (4.70)
dbi wbi
To calculate the right hand side of Equation 4.70, Equation 4.69g is differentiated
as
wK ª1 / 2 L1 1 / 2 L1 º
E« » (4.71a)
wb1 ¬1 / 2 L1 1 / 2 L1 ¼
wK ª0 0 º
E« (4.71b)
wb2 ¬ 0 1 / L2 »¼
wK ª 1 / 2 L3 1 / 2 L3 º
E« (4.71c)
wb3 ¬ 1 / 2 L3 1 / 2 L3 »¼
Equations 4.69g, 4.71 and 4.20 are substituted into Equation 4.70, and the values
of parameters are used. Then Equation 4.70 is solved to obtain dz/dbi as follows:
dz
[ 0.86024 0.50392]T (4.72a)
db1
dz
[ 0.0 0.71781]T (4.72b)
db2
dz
[0.00624 0.00365]T (4.72c)
db3
Since the solution of Equation 4.72 is obtained from Equation 4.70, the
equality constraints can be eliminated. Now the subproblem is made by Equations
4.69aí4.69e by substituting Equation 4.72 into the subproblem. The subproblem
is as follows:
Find Gb [Gb1 Gb2 Gb3 ]T (4.73a)
to minimize Tb fGb Ul[ 2 1 2 ]Gb (4.73b)
§ wV 1 wV 1 dz1 ·
subject to V 1 ¨¨ ¸Gb d 0 (4.73c)
© wb wz1 db ¸¹
§ wV 2 wV 2 dz 2 ·
V 2 ¨¨ ¸Gb d 0 (4.73d)
© wb wz 2 db ¸¹
§ wV 3 wV 3 dz 3 ·
V 3 ¨¨ ¸Gb d 0 (4.73e)
© wb wz 3 db ¸¹
where z i is the displacement vector of the ith truss element in the global
coordinate system.
The equation for the stress in a truss element Equation 4.27 is recalled as
E
Vi [ 1 1]Tz i (4.74)
L
Equation 4.74 is differentiated with respect to b and z as follows:
wV i
0, i 1, ,3 (4.75)
wb
wV i E
[ 1 1]T, i 1, , 3 (4.76)
wz i L
If we substitute Equations 4.75 and 4.76 into Equation 4.73, we obtain the
following subproblem:
§ E ª dz dz1 dz1 º ·¸
subject to V 1 ¨¨ 0 [ 1 1]T « 1 » Gb d 0 (4.77c)
© L1 ¬ db1 db2 db3 ¼ ¸¹
§ E ª dz dz 2 dz 2 º ·¸
V 2 ¨¨ 0 [ 1 1]T « 2 » Gb d 0 (4.77d)
© L2 ¬ db1 db2 db3 ¼ ¸¹
§ E ª dz dz 3 dz 3 º ·¸
V 3 ¨¨ 0 [ 1 1]T « 3 » Gb d 0 (4.77e)
© L3 ¬ db1 db2 db3 ¼ ¸¹
Equation 4.72 and the transformation matrix are substituted into Equation 4.77
and the values of parameters are used. Then the following subproblem is obtained:
subject to
§ ª § 3S · § 3S · º
¨ 9 «cos ¨ ¸ sin ¨ ¸ 0 0 »
¨ 207 u 10 4 ¹ © 4 ¹
71.227 u 10 6 ¨ 0 [ 1 1]« © »
2 « § 3S · § 3S ·»
¨¨ « 0 0 cos ¨ ¸ sin ¨ ¸»
© ¬ © 4 ¹ © 4 ¹¼
ª 0.86024 0.0 0.00624 º ·
« 0.50392 0.71781 0.00365» ¸
« » ¸Gb d 0
« 0 0 0 »¸
« » ¸¸
¬ 0 0 0 ¼¹
(4.78c)
§ ª § S· § S· º
¨ «cos ¨ ¸ sin ¨ ¸ 0 0 »
¨ 207 u 109 2 2
6
71.744 u 10 ¨ 0 [ 1 1]« © ¹ © ¹ »
1 « § S· § S ·»
¨¨ « 0 0 cos ¨ ¸ sin ¨ ¸»
© ¬ © 2¹ © 2 ¹¼
ª 0.86024 0. 0 0.00624 º ·
« 0.50392 0.71781 0.00365» ¸
« » ¸Gb d 0
« 0 0 0 »¸
« » ¸¸
¬ 0 0 0 ¼¹
(4.78d)
§ ª § S· § S· º
¨ «cos¨ 4 ¸ sin¨ 4 ¸ 0 0 »
3 ¨ 207 u 109 © ¹ © ¹
516.5 u 10 ¨ 0 [ 1 1]« »
2 « § S· § S ·»
¨¨ « 0 0 cos¨ ¸ sin ¨ ¸»
© ¬ © 4¹ © 4 ¹¼
ª 0.86024 0.0 0.00624 º ·
« 0.50392 0.71781 0.00365» ¸
« » ¸Gb d 0
« 0 0 0 »¸
« » ¸¸
¬ 0 0 0 ¼¹
(4.78e)
Equation 4.78 is summarized as
subject to
71.227 u 10 6 [ 1.4119 u 1011 7.4293 u 1010 2.6741u 10 8 ]Gb d 0

(4.79c)
71.744 u 10 6 [ 1.0431u 1011 1.4859 u 1011 7.5635 u 10 8 ]Gb d 0

(4.79d)
516.5 u 10 3 [3.6879 u 1011 7.4293 u 1010 1.0238 u 10 9 ]Gb d 0

(4.79e)
Example 4.5 [Optimization Formulation of the Two Member Beam]

A beam with two beam elements is illustrated in Figure 4.6. The sizes of the cross
sections are the design variables. The objective function is the mass and
constraints are imposed on the vertical displacement and stresses. Formulate an
optimization problem. The density is U , Young’s modulus is E and the product of
the element stiffness matrix and the displacement vector is
ª AL2 AL2 º
« 0 0 0 0 » ª u1 º
« 2I 2I »« »
« 0 6 3L 0 6 3L » « v1 »
k i z iL
2 EI « 0 3L 2 L2 0 3L L2 » «T 1 » (4.80)
« 2
AL2 » «u »
L3 « AL 0 0 0 0 »« 2 »
« 2I 2I » «v 2 »
« 0 6 3L 0 6 3L » « »
« 0 ¬«T 2 ¼»
¬ 3L L2 0 3L 2 L2 »¼
where I is the moment of inertia, A is the area of the cross section, L is the length
of the element and the transformation matrix is
ª cos D sin D 0 0 0 0º
« sin D cos D 0 0 0 0»»
«
« 0 0 1 0 0 0»
T « » (4.81)
« 0 0 0 cos D sin D 0»
« 0 0 0 sin D cos D 0»
« »
¬« 0 0 0 0 0 1¼»
y P
L1 L2
x h
3
1 ཛ 2 ཛྷ
b
Figure 4.6. A beam with two beam elements

Solution
Since the global and the local coordinate systems are identical, the transformation
matrix is T I. Thus the entire stiffness matrix K in the global coordinate system
is
ª A1 E1 A1 E1 º
« L 0 0 0 0 0 0 0 »
« 1 L1 »
« 0 12 E1 I 1 6 E1 I 1 12 E1 I 1 6 E1 I 1 »
0 0 0 0
« L13 L12 L13 L12 »
« »
« 0 6 E1 I 1 4 E1 I 1 6 E1 I 1 2 E1 I 1 »
0 0 0 0
« L12 L1 L12 L1 »
« »
« A1 E1 0 0
A1 E1 A2 E 2
0 0
A2 E 2
0 0 »
« L1 L1 L2 L2 »
« 12 E1 I 1 6 E1 I 1 12 E1 I 1 12 E 2 I 2 6 E1 I 1 6E 2 I 2 12 E 2 I 2 6E 2 I 2 »
K « 0 0 0 »
« L13 L12 L13 L32 L12 L22 L32 L22 »
« 6 E1 I 1 2 E1 I 1 6 E1 I 1 6E 2 I 2 4 E1 I 1 4 E 2 I 2 6E 2 I 2 2E 2 I 2 »
« 0 0 0 »
« L12 L1 L12 L22 L1 L2 L22 L2 »
« A E A2 E 2 »
« 0 0 0 2 2 0 0 0 0 »
« L2 L2 »
« 12 E 2 I 2 6E 2 I 2 12 E 2 I 2 6E 2 I 2 »
« 0 0 0 0 0 »
« L32 L22 L32 2
L2 »
« 6E 2 I 2 2E 2 I 2 6E 2 I 2 4E 2 I 2 »
« 0 0 0 0 0
¬ L22 L2 L22 L 2 »¼
(4.82)
Using boundary conditions, matrix K is reduced to
ª A1 E1 A2 E 2 A2 E 2 º
« L L 0 0
L2
0 »
« 1 2 »
« 12 E1 I 1 12 E 2 I 2 6 E1 I 1 6E 2 I 2 6E 2 I 2 »
0 0
« L13 L32 L12 L22 L22 »
« »
« 6 E1 I 1 6E 2 I 2 4 E1 I 1 4 E 2 I 2 2E 2 I 2 »
K 0 0
« L12 L22 L1 L2 L2 »
« »
« A2 E 2 0 0
A2 E 2
0 »
« L2 L2 »
« 6E 2 I 2 2E 2 I 2 4E 2 I 2 »
« 0 0 »
¬« L22 L2 L 2 ¼»
(4.83)
Find b1 , h1 , b2 , h2 (4.84a)
to minimize U (b1 h1 L1 b2 h2 L2 ) (4.84b)
subject to g1 v2 d vall (4.84c)
g2 V 1 d V all
g3 V 2 d V all
Kz f (4.84d)
where
ª b1 h1 E b2 h2 E b2 h2 E º
« L L 0 0
L2
0 »
« 1 2
3
»
« 12 E b1 h1 12 E b2 h23 3
6 E b1 h1 6 E b2 h2
3
6E b2 h23 »
« 0 3 2 0
«
3
L1 12 L2 12 L12 12
L2 12 L22 12 »»
« 6 E b1 h13 6 E b2 h23 4 E b1 h13 4 E b2 h23 2E b2 h23 »
K « 0 2 2 0 »
« L1 12 L2 12 L1 12 L2 12 L2 12 »
« b2 h2 E b2 h2 E »
« 0 0 0 »
« L2 L2 »
3 3 3
« 6 E b2 h2 2 E b2 h2 4E b2 h2 »
« 0 0 »
¬« L22 12 L2 12 L2 12 ¼»
(4.84e)
z [u 2 v2 T2 u 3 T 3 ]T (4.84f)
f [0 P 0 0 0]T (4.84g)
Example 4.6 [Optimization Formulation for a Structure of Plane

Stress]
Formulate an optimization problem to minimize the weight of the structure in
Example 4.2. The design variable is the thickness and constraints are imposed on
the displacements. Use the data of Example 4.2, the density is U .
Solution
Find h (4.85a)
to minimize Uh (4.85b)
subject to u 2 d u 2, all (4.85c)
v2 d v2, all (4.85d)
u3 d u3, all (4.85e)
v3 d v3, all (4.85f)
Kz f (4.85g)
where
ª 0.488667 0.166667 0.044333 0.033 º

« 0.166667 0.488667 0.033 0.288667»»
K Eh « (4.85h)
« 0.044333 0.033 0.488667 0.166667 »
« »
¬ 0.033 0.288667 0.166667 0.488667 ¼
z [u 2 v2 u3 v 3 ]T (4.85i)
f [0 0 100000 100000]T (4.85j)
4.4 Sensitivity Analysis

Sensitivity analysis requires the most resources in structural optimization. Thus,
much research has been carried out on the efficiency of sensitivity analysis. There
are generally three methods for sensitivity analysis. They are the finite difference
method, the analytic method with discrete equations and the analytic method with
continuum equations. The analytic method with discrete equations is classified
into the analytic method and the semianalytic method. The method with
continuous equations is basically an analytic method. It is classified into two
methods. In one method, the continuum equation is mathematically differentiated.
In the other method, the continuum equation is differentiated and the sensitivity
information is obtained from the finite element method, which discretizes the
continuum equation. The analytic methods consist of the direct differentiation
method and the adjoint variable method. Although the methods are different in
derivation, they are commonly used for linearization of the objective function and
constraints in the optimization process. The classification of the methods and their
characteristics are shown in Tables 4.2 and 4.3.
4.4.1 Finite Difference Method
The derivatives of the objective function and constraints can be obtained by the
finite difference method. It can be applied to Equation 4.44b or 4.44e. The
method is explained in Equation 4.44e. Using the forward difference method, the
sensitivity information of Equation 4.44e with respect to bi is
wg j g j (b ('b) i , z (b ('b) i ), [ (b ('b) i )) g j (b, z , [ )

(4.86)
wbi 'bi
z (b ('b) i ) and [ (b ('b) i ) in Equation 4.86 are solutions of
K (b ('b) i )z (b ('b) i ) f (b ('b) i ) (4.87)

Table 4.2. Classification of sensitivity analysis according to the method of approach
Method Calculation Remarks

Forward
difference method
Backward
Finite difference method
difference method
Central difference
method
Semianalytic Direct
Discrete equation
Analytic Adjoint variable
Continuum–
continuum Direct
Continuum equation
Continuumˀ Adjoint variable
discrete
Table 4.3. Characteristics of the methods for sensitivity analysis
Method Advantages Disadvantages Remarks
Precise information is Recommended for

provided commercial software
Analytic Programming is
where sensitivity
method Calculation time is difficult
analysis is carried out
small repeatedly
Error can occur
Programming is easy Error may be large
Semianalytic depending on the
when a finite element for shape
method size of the finite
analysis code exists optimization
difference
It is difficult to de-
termine the size of
the finite Recommended when
Finite difference
Programming is easy difference a specific problem is
method
solved
Calculation time is
large
K (b ( 'b) i )y (b ( 'b) i ) [ (b ('b) i )M (b ('b) i )y (b ('b) i ) 0

(4.88)
where ('b) i is a vector of [0 0 'bi 0 0]T . Only the ith element is

'bi and the remaining elements are zeros (0). As shown in Equations 4.87 and
4.88, the forward difference method needs to calculate z and [ again when the
design variable vector is b ('b) i . The finite element analysis should be

performed again for this calculation. The equality constraints in Equations 4.44c
and 4.44d are automatically satisfied in this process.
In the finite difference method, we also have the backward difference method
in Equation 4.89 and the central difference method in Equation 4.90,
wg j g j (b, z, [ ) g j (b ('b) i , z (b ('b) i ), [ (b ('b) i ))

wbi 'bi
(4.89)
wg j g j (b ('b / 2) i , z (b ('b / 2) i ), [ (b ('b / 2) i ))
wbi 'bi
g j (b ('b / 2) i , z (b ('b / 2) i ), [ (b ('b / 2) i ))

(4.90)
'bi
The accuracy of the finite difference method depends on the size of 'bi . The
forward and backward difference methods have the same accuracy. However, the
central difference method has better accuracy, even though it requires more
calculation. For accuracy, it is important to have an appropriate size for 'bi .
Although much research has been conducted, a general rule does not exist for
determination of 'bi . The size of 'bi tends to be reduced due to the enhanced
power of the computer. It is noted that small 'bi can increase the round-off error.
The finite difference method is numerically simple to implement. Moreover, it
has the advantage that a commercial software system can be treated as a black box
in the calculation process. It has the disadvantage that the numerical cost is quite
high. When a number of function calculations are repeatedly required, this method
may not be practical.
Example 4.7 [Sensitivity Analysis for the Three Bar Truss: Finite
Difference Method]
We have the symmetric three bar truss of Example 4.1. Evaluate the sensitivity of
the stress of member 1 with respect to the area by using the forward difference
method and the central difference method. Set the perturbation as 'b 10 8 .
Solution
Using Equation 4.69g, which is the stiffness matrix of the three bar truss,
ª b1 b3 b1 b º
« 2L 2L 3
2 L1 2 L3 »
K E« 1 3 » (4.91)
« b1 b3 b1

b2

b3 »
«¬ 2 L1 2 L3 2 L1 L2 2 L3 »¼
From Equation 4.18, the external load vector is
ª 40000 u 1 / 2 º
f «40000 u 3 / 2» (4.92)
¬ ¼
The displacement vector for the current design is as in Equation 4.20. The stress
of member 1 is as in Equation 4.28 by substituting the displacements in Equation
4.20 into Equation 4.27. The stress for the area increased by ('b) can be obtained
in the same way.
(1) Sensitivity using the forward difference method
Using Equation 4.86
dV 1 V 1 (b1 'b, b2 , b3 , z (b1 'b, b2 , b3 )) V 1 (b1 , b2 , b3 , z (b1 , b2 , b3 ))

db1 'b
(4.93a)
dV 1 V 1 (b1, b2 'b, b3 , z (b1 , b2 'b, b3 )) V 1 (b1 , b2 , b3 , z (b1 , b2 , b3 ))
db2 'b
(4.93b)
dV 1 V 1 (b1, b2 , b3 'b, z (b1 , b2 , b3 'b)) V 1 (b1 , b2 , b3 , z (b1 , b2 , b3 ))
db3 'b
(4.93c)
Equation 4.93 yields
dV 1
[ 1.4119 u 1011 7.4292 u 1010 2.6748 u 108 ] (4.93d)
db
(2) Sensitivity analysis using the central difference method

Using Equation 4.90
dV 1
[V 1 (b1 'b / 2, b2 , b3 , z (b1 'b / 2, b2 , b3 ))
db1 (4.94a)
V 1 (b1 'b / 2, b2 , b3 , z (b1 'b / 2, b2 , b3 ))] / 'b
dV 1
[V 1 (b1 , b2 'b / 2, b3 , z (b1 , b2 'b / 2, b3 ))
db2 (4.94b)
V 1 (b1 , b2 'b / 2, b3 , z (b1 , b2 'b / 2, b3 ))] / 'b
dV 1
[V 1 (b1 , b2 , b3 'b / 2, z (b1 , b2 , b3 'b / 2))
db3 (4.94c)
V 1 (b1 , b2 , b3 'b / 2, z (b1 , b2 , b3 'b / 2))] / 'b
dV 1
[ 1.4119 u1011 7.4293 u1010 2.6748 u108 ] (4.94d)
db
Example 4.8 [Sensitivity Analysis of the Elements for the Plane

Stress: Finite Difference Method]
We have a structure as shown in Example 4.2. Evaluate the sensitivity of the
stress with respect to the thickness by using the forward difference method and the
central difference method. Set the perturbation as 'h 10 4 mm.
Solution
(1) Sensitivity analysis using the forward difference method

The sensitivity by the forward difference method is
dV V (h 'h, z (h 'h)) V (h, z (h))

(4.95)
dh 'h
The stress can be evaluated by substituting the displacements from

Equation 4.37 into Equation 4.41. Equation 4.95 yields
dV x
5.6720 u 1010 (4.96a)
dh
dV y
6.4538 u1010 (4.96b)
dh
dW xy
5.0063 u 1010 (4.96c)
dh
(2) Sensitivity analysis using the central difference method

The sensitivity by the central difference method is
dV V (h 'h / 2, z (h 'h / 2)) V (h 'h / 2, z (h 'h / 2))

(4.97)
dh 'h
Substituting the stresses from Equations 4.37 and 4.41 into Equation 4.97,
the sensitivity information is
dV x
5.6725 u 1010 (4.98a)
dh
dV y
6.4545 u1010 (4.98b)
dh
dW xy
5.0068 u 1010 (4.98c)
dh
4.4.2 Analytic Method with Discrete Equations
In the analytic method, the differentiation process in Equations 4.53 and 4.60 is
directly programmed. It is the best method when the programming is possible,
since the calculation cost is the lowest. However, the programming is generally
extremely complicated and we have to use the source code of the finite element
method. Therefore, it is almost impossible to use it with commercial software
systems. The analytic method with continuum equations can be utilized to
overcome these difficulties. This will be introduced later. In this section, the
analytic method with discrete equations is introduced.
The governing equation in a discrete form comes from the algebraic equations
in Equation 4.44c or 4.44d. Equation 4.46e becomes Equation 4.61 by using
Equations 4.53 and 4.60. Generally, z and [ of Equation 4.46e do not appear
simultaneously in an equation. Therefore, the differentiation in Equation 4.61
results in either of the following two equations:
dg j wg j wg j dz
(4.99)
dbi wbi wz dbi
dg j wg j wg j d[
(4.100)
dbi wbi w[ dbi
The calculation method for Equation 4.99 is explained here. Equation 4.100
can be calculated in a similar way by using Equation 4.60. The complex
wg j dz
calculation in Equation 4.99 is the second term . Especially the
wz dbi
dz
calculation of is very expensive. There are two methods for the calculation
dbi
dz
of . These are the direct differentiation method and the adjoint variable
dbi
method.
In the direct differentiation method, the following simultaneous equation,
which is derived from Equation 4.53, is solved:
dz
K v i , i 1, , n (4.101a)
dbi
where
wK ( k ) wf
vi z , i 1, , n (4.101b)
wbi wbi
In Equation 4.101, v i can be considered as an external load vector that

dz
generates a displacement vector . Thus, Equation 4.101b is called the pseudo
dbi
load. In Equation 4.101, the simultaneous equation should be solved n (the
dz
number of design variables) times to obtain .
dbi
dz
In the adjoint variable method, derived from Equation 4.53 is substituted
dbi
into Equation 4.99 as follows:
wg j dz wg j § wK ( k ) wf ·
K 1 ¨¨ z ¸
¸ (4.102)
wz dbi wz © wbi wbi ¹
An artificial vector Ȝ j is defined by the following adjoint equation:
wg j
Ȝ Tj K 1 (4.103)
wz
Equation 4.103 can be modified as
T
§ wg j ·
KȜ j ¨ ¸ , j 1,, m (4.104)
¨ wz ¸
© ¹
Ȝ j can be considered as a displacement vector when an external load vector

T
§ wg j ·
¨ ¸
¨ wz ¸ is imposed on a structure. The simultaneous equation in Equation 4.104
© ¹
should be solved m (the number of constraints) times. In reality, m is the number
of İ-active constraints. If we substitute Equation 4.103 into Equation 4.102,
wg j dz § wK ( k ) wf ·
Ȝ Tj ¨¨ z ¸
¸ (4.105)
wz dbi © wbi wbi ¹
Sensitivity analysis is classified into the direct differentiation method to solve

Equation 4.101 and the adjoint variable method to solve Equation 4.104. The
efficiency of the two methods is related to the number of simultaneous equations to
be solved. In the direct differentiation method, the number is the same as the
number of design variables n, while the number in the adjoint method is the
number of İ-active constraints m. In structural optimization the number of
inequality constraints is a lot bigger than the number of design variables. However,
the number of İ-active constraints can be smaller than the number of design
variables. Therefore, an appropriate method should be chosen depending on the
problem.
Example 4.9 [Sensitivity Analysis for the Three Bar Truss: Analytic
Method]
In Example 4.1, obtain the sensitivity information of the stress in member 1 with
respect to the area by using the analytic method.
Solution
Equation 4.99 can be modified as
dV 1 wV 1 wV 1 dz
(4.106)
db wb wz db
In the analytic method, Equation 4.106 is analytically obtained.

(1) The direct differentiation method
In Equation 4.27, the stress is
V1 E / L[1 1]Tz (4.107)
Thus
wV 1
0 (4.108)
wb
wV 1 E
[ 1 1]T (4.109)
wz L
dz
is obtained from Equation 4.101. In Equation 4.101, the
db
differentiation of the stiffness matrix should be obtained analytically and it
dz
has been derived in Equation 4.71. The solution of Equation 4.101 is
db
obtained in Equation 4.72. If we substitute Equations 4.72 and 4.108í
4.109 into Equation 4.106, then the sensitivity with respect to the area is
ª § 3S · § 3S · º
dV 1 E «cos ¨ 4 ¸ sin ¨ 4 ¸ 0 0 »
0 [ 1 1]« © ¹ © ¹ »
db L « § 3S · § 3S ·»
« 0 0 cos ¨ ¸ sin ¨ ¸»
¬ © 4 ¹ © 4 ¹¼
ª 0.86024 0.0 0.00624 º

« 0.50392 0.71781 0.00365»
« »
« 0 0 0 »
« »
¬ 0 0 0 ¼
[ 1.4119 u 1011 7.4293 u 1010 2.6741 u 10 8 ] (4.110)
(2) The adjoint variable method

If we use Equation 4.102 in Equation 4.106, the following equation is
obtained:
dV 1 wV 1 wV 1 1 § wK wf ·
K ¨ z ¸ (4.111)
db wb wz © wb wb ¹
The adjoint equation is defined as
wV 1 1
ȜT K (4.112)
wz
Since the stiffness matrix K is symmetric, Equation 4.112 is
T
§ wV 1 ·
KȜ ¨ ¸ (4.113)
© wz ¹
If we substitute Equation 4.91 and Equation 4.109 with current values into
Equation 4.113, the following equation is obtained:
T
§ ª § 3S · § 3S ·º ·¸ (4.114)
9 ª2 0 º ¨ 207 u 109 cos ¨ ¸ sin ¨ ¸»
2 u 10 4
u 207 u 10 « »Ȝ ¨ [ 1 1] « © 4 ¹ © 4 ¹» ¸
¬0 2 2 ¼ ¨ 2 « ¸
© ¬« 0 0 ¼» ¹
It is noted that the boundary condition is already applied in Equation 4.114.

The solution of Equation 4.114 is
Ȝ [1767.78 1035.53]T (4.115)
If we substitute Equation 4.112 into Equation 4.111,

dV 1 wV1 § wK wf ·
ȜT ¨ z ¸ (4.116)
db wb © wb wb ¹
wK
in Equation 4.116 is in Equation 4.71. Equation 4.71 with current
wb
values is
wK 207 u 10 9 ª1 1º
« » (4.117a)
wb1 2 2 ¬1 1¼
wK ª0 0º
207 u 10 9 « » (4.117b)
wb2 ¬0 1 ¼
wK 207 u 10 9 ª 1 1º
« » (4.117c)
wb3 2 2 ¬ 1 1 ¼
If we substitute Equations 4.115, 4.117 and 4.20 into Equation 4.116, then
the following solutions are obtained:
dV 1 § 207 u109 ª1 1º ª3.4160 u10 4 º ·

0 [1767.78 1035.53]¨ «1 1» « 4 »
0¸
db1 ¨ 2 2 ¬ ¼ «3. 4659 u 10 » ¸
© ¬ ¼ ¹
1.4119 u 1011 (4.118a)
dV 1 § ª0 0º ª3.4160 u 10 4 º ·
0 [1767.78 1035.53]¨ 207 u 109 « » « » 0 ¸¸
db2 ¨ 4
¬0 1¼ «¬3.4659 u 10 »¼
© ¹
7.4293 u 1010 (4.118b)
dV 1 § 207 u 109 ª 1 1º ª3.4160 u 104 º ·

0 [1767.78 1035.53]¨ « 1 1 » « 4 »
0¸
db1 ¨ 2 2 ¬ ¼ «¬3.4659 u 10 »¼ ¸
© ¹
2.6741u 10 8 (4.118c)
Example 4.10 [Sensitivity Analysis of Elements for a Plane Stress:

Analytic Method]
For the structure in Example 4.2, obtain the sensitivity of the stress with respect to
the thickness by the analytic method.
Solution

The sensitivity of the stress with respect to the thickness is
dı wı wı dz
(4.119)
dh wh wz dh
Since the plane stress is calculated from Equation 4.41, each term in
Equation 4.119 is
wı
0 (4.120a)
wh
ª 1 1 1 1 º
ª º« 4 0 0 0 0 »
1 v 0 » 4 4 4
wı E « « 1 1 1 1 »
«v 1 0 »« 0 0 0 0 »
wz 1 v2 « 1 v »« 2 2 2 2 »
«0 0 » 1 1 1 1 1 1 1 1
¬ 2 ¼ «« »»
¬ 2 4 2 4 2 4 2 4¼
(4.120b)
dz
Using the stiffness matrix in Equation 4.36, of Equation 4.119 is
dh
obtained by solving Equation 4.101 as follows:
ª 0.488667 0.166667 0.044333 0.033 º

« 0.166667 0.488667 0. 033 0 .288667 »» dz
Eh «
« 0.044333 0.033 0.488667 0.166667 » dh
« »
¬ 0.033 0.288667 0.166667 0.488667 ¼
ª 0.488667 0.166667 0.044333 0.033 º

« 0.166667 0.488667 0.033 0.288667 »»
E « z (4.121)
« 0.044333 0.033 0.488667 0.166667 »
« »
¬ 0.033 0.288667 0.166667 0.488667 ¼
wK
where it is noted that is analytically calculated. In Example 4.12 it
wh
will be calculated by the semianalytic method.
dz z
[ 4.0419 9.9649 4.0759 15.0012]T u 10 1 (4.122)
dh h
Substituting Equations 4.120 and 4.122 into Equation 4.119, the sensitivity
with respect to the thickness is
dı wı wı dz

dh wh wz dh
ª 0 º
« 0 »
ª 1 1 1 1 º« »
ª º« 4 0 0 0 0 » «0.40419»
1 v 0 »« 4 4 4 « »
E «« 1 1 1 1 » «0.99649»
0 v 1 0 »« 0 0 0 0 »
1 v2 « 1 v »« 2 2 2 2 » «0.40759»
«0 0 »« 1 1 1 1 1 1 1 1 »« »
¬ 2 ¼ « » «1.50012 »
¬ 2 4 2 4 2 4 2 4 ¼«
0 »
« »
«¬ 0 »¼
(4.123)
If we calculate Equation 4.123, the sensitivity information is
dV x
5.6725 u 1010 (4.124a)
dh
dV y
6.4545 u 1010 (4.124b)
dh
dW xy
5.0068 u 1010 (4.124c)
dh

Substituting Equation 4.102 into Equation 4.119, the sensitivity is
dı wı wı 1 § wK wf ·
K ¨ z ¸ (4.125)
dh wh wz © wh wh ¹
The adjoint equation is defined as
wı 1
ȜT K (4.126)
wz
Since matrix K is symmetric, Equation 4.126 is
T
§ wı ·
KȜ ¨ ¸ (4.127)
© wz ¹
From Equation 4.41
wı
EB (4.128)
wz
Thus, if we substitute Equation 4.128 into Equation 4.127, then
KȜ (EB ) T (4.129)
In the process of calculating Equations 4.41 and 4.42, (EB) T is
ª 5.3333 1.3333 4º
« 2.6667 10.667 2 »
(EB) T « » u 1010 (4.130)
« 5.3333 1.3333 4»
« »
¬ 2.6667 10.667 2¼
where Equation 4.130 is the result after the boundary condition is applied.
If we substitute Equations 4.36 and 4.130 into Equation 4.129, the adjoint
variable vector (matrix) Ȝ is obtained as
ª 477.91 49.103 408.84º

« 89.348 694.55 91.837 »
Ȝ « » (4.131)
« 477.91 49.103 408.84 »
« »
¬ 89.348 694.55 91.837 ¼
Substituting Equation 4.126 into Equation 4.125, we obtain
dı wı § wK wf ·
Ȝ¨ z ¸ (4.132)
dh wh © wh wh¹
From Equation 4.41
wı
0 (4.133a)
wh
and from Equation 4.36

ª 0.488667 0.166667 0.044333 0.033 º

« 0.166667 0.488667 0.033 0.288667»»
wK
E« (4.133b)
wh « 0.044333 0.033 0.488667 0.166667 »
« »
¬ 0.033 0.288667 0.166667 0.488667 ¼
Since the external load is not a function of the thickness
wf
0 (4.133c)
wh
If we substitute Equations 4.40, 4.131 and 4.133 into Equation 4.132, then
the sensitivity is
T
ª 477.91 49.103 408.84º
« 89.348 694.55 91.837 »
dı
0« »
dh « 477.91 49.103 408.84 »
« »
¬ 89.348 694.55 91.837 ¼
§ ª 0.488667 0.166667 0.044333 0.033 º

¨ « 0.166667 0.488667
¨ 9 « 0.033 0.288667»»
¨ 200 u 10 u «
¨ 0.044333 0.033 0.488667 0.166667 »
¨ « »
© ¬ 0.033 0.288667 0.166667 0.488667 ¼
ª 4.0419 º ·
« 9.9649 » ¸
« » u 10 4 0 ¸ (4.134)
« 4.0759 » ¸
« » ¸
¸
¬15.0012¼ ¹
From Equation 4.134, the final results are obtained as follows:
dV x
5.6725 u 1010 (4.135a)
dh
dV y
6.4545 u1010 (4.135b)
dh
dW xy
5.0068 u 1010 (4.135c)
dh
4.4.3 Semianalytic Method
In the analytic method, the differentiation of the stiffness or mass matrix is the
most difficult task in numerical implementation (programming). The semianalytic
method simplifies the process. The differentiation of the matrices is carried out by
the finite difference method. In other words, wK / wbi and wM / wbi in Equations
4.53 and 4.60 are calculated as follows:
wK K (bi 'bi ) K (bi )

# (4.136)
wbi 'bi
wM M (bi 'bi ) M (bi )

# (4.137)
wbi 'bi
Since the amount of calculation in Equations 4.136 and 4.137 is not great, the
central difference method is frequently employed for accuracy. The results of
Equations 4.136 and 4.137 are substituted into Equations 4.53 and 4.60. The
remaining procedures are exactly the same as the one used in the analytic method.
Since the numerical implementation of the semianalytic method is handy, it is
frequently used in commercial software systems. It is noted that the accuracy
depends on the size of 'bi and the error can be enlarged in the later processes.
There are reports that the error is large for shape design variables.
Example 4.11 [Sensitivity Analysis for the Three Bar Truss:

Semianalytic Method]
In Example 4.1, obtain the sensitivity information of the stress in member 1 with
respect to the area by using the semianalytic method. Set the perturbation as
'b 10 8 m2.
Solution
The solution process is the same as that of Example 4.9, except for calculation of
the sensitivity of the stiffness matrix by the finite difference method.
dz
In Example 4.9, in Equation 4.106 is obtained by solving Equation
db
4.101 and the sensitivity of the stiffness matrix is obtained analytically. In
this example, the sensitivity of the stiffness matrix in Equation 4.101 is
obtained by the finite difference method as follows:
wK K (b1 'b, b2 , b3 ) K (b1 , b2 , b3 ) ª1 1º

# 7.3186 u 1010 « » (4.138a)
wb1 'b ¬1 1¼
wK K (b1 , b2 'b, b3 ) K (b1 , b2 , b3 ) ª0 0º

# 2.0700 u 1011 « » (4.138b)
wb2 'b ¬0 1¼
wK K (b1 , b2 , b3 'b) K (b1 , b2 , b3 ) ª 1 1º

# 7.3186 u 1010 « » (4.138c)
wb3 'b ¬ 1 1 ¼
The stiffness matrix in Equation 4.91 is utilized. Using Equation 4.138 the
solution of Equation 4.101 is
dz
[ 0.86024 0.50392]T (4.139a)
db1
dz
[ 0.0 0.71781]T (4.139b)
db2
dz
[0.00624 0.00365]T (4.139c)
db3
The results in Equation 4.139 are the same as those in Equation 4.72.
Therefore, the remaining process is the same as that of Example 4.9.
wK
from the semianalytic method is shown in Equation 4.138. Equation
wb
4.138 is used instead of Equation 4.117 in Example 4.9. The remaining
process is the same as that of Example 4.9.
Example 4.12 [Sensitivity Analysis of Elements for a Plane Stress:

Semianalytic Method]
For the structure in Example 4.2, evaluate the sensitivity of the stress with respect
to the thickness by using the semianalytic method. Set the perturbation as
'h 10 4 mm.
Solution

wK
In the process of Example 4.10, for Equation 4.121 is obtained by the
wh
semianalytic method. The other processes are the same as those of
Example 4.10.
ª 9.7733 3.3333 0.8867 0.6600º

« 3.3333 9.7733 0.6600 5.7733»»
wK K (h 'h) K (h) «
# u 1010
wh 'h « 0.8867 0.6600 9.7733 3.3333 »
« »
¬ 0.6600 5.7733 3.3333 9.7733 ¼
(4.140)
Substituting Equation 4.140 into Equation 4.54 results in
ª 0.488667 0.166667 0.044333 0.033 º

« 0.166667 0.488667 0.033 0.288667»» dz
200 u 109 u 10 3 «
« 0.044333 0.033 0.488667 0.166667 » dh
« »
¬ 0.033 0.288667 0.166667 0.488667 ¼
ª 9.7733 3.3333 0.8867 0.6600º

« 3.3333 9.7733 0.6600 5.7733»»
« u 1010 z (4.141)
« 0.8867 0.6600 9.7733 3.3333 »
« »
¬ 0.6600 5.7733 3.3333 9.7733 ¼
dz
of Equation 4.141 is
db
dz
[ 4.0419 9.9649 4.0759 15.0012]T u101 (4.142)
dh
The remaining process is the same as that of Example 4.10.

In Example 4.10, Equation 4.133b is calculated by the finite difference
method. The results are in Equation 4.140. The remaining process is the
same as that of Example 4.10.
4.4.4 Analytic Sensitivity Analysis Using Continuum Equations
So far, sensitivity analysis has been derived by using the discrete equations for the
finite element method. In this section, it is derived based on the continuum
equations. The objective or a constraint function using a continuum equation can
be written as
\ ³ g (z, b)dȍ (4.143)

ȍ
To use Equation 4.143 in the optimization process, the following sensitivity

equation should be obtained:
G\ ³ ( g zGz g bGb)dȍ (4.144)

ȍ
wg wg dz
where g z , gb and Gz Gb . Calculating Gz is the most important
wz wb db
task in Equation 4.144.
Using the principle of virtual work, the discrete equation in Equation 4.8 yields
z T Kz zT f fTz (4.145)
where z is the infinitesimal virtual displacement, which is kinematically

admissible. In a continuum equation, Equation 4.145 is
ab (z, z ) lb ( z ) (4.146)
Equation 4.146 is explained by using a linear operator. Suppose the following

equation is satisfied in the domain ȍ with the boundary ī :
Ab z f (4.147)
where Ab is a linear operator. For kinematically admissible z , the following

equation is satisfied:
T
³ z ( Ab z f )dȍ 0 (4.148)
ȍ
Equation 4.146 and Equation 4.148 are identical from the following relationships:
ab (z, z ) { ³ z T Ab zdȍ (4.149a)

ȍ
lb ( z ) { ³ z T fdȍ (4.149b)
ȍ
If we take a variation of Equation 4.146, this yields
ab (Gz, z ) aGc b (~z , z ) lGcb ( z ) (4.150)
wa wl
where aGc b Gb and lGcb Gb . ~z is the current value and is considered
wb wb
constant. If we compare Equation 4.150 with Equation 4.53, which has a
discretized form, then
§ dz wK ~z ·¸Gb § wf ·
zT ¨ K z T ¨ ¸Gb (4.151)
© db wb ¹ © wb ¹
Sensitivity analysis is the process of obtaining Gz in Equation 4.150 and Gz is

substituted into Equation 4.144. This is the same as the direct differentiation
method in Section 4.3.
The adjoint variable method is explained. The following adjoint variable
equation is defined:
ab ( Ȝ , Ȝ ) ³ g z Ȝdȍ (4.152)
ȍ
where Ȝ is a kinematically admissible vector. The adjoint variable vector Ȝ is

obtained from Equation 4.152. Equation 4.152 is the same as the original
governing equation in Equation 4.146, except for the term on the right hand side.
If we replace Ȝ in Equation 4.152 with Gz , then
ab (Ȝ , Gz ) ³ g zGzdȍ (4.153)
ȍ
If we replace z in Equation 4.150 with Ȝ , then
ab (Gz, Ȝ ) aGc b (~z , Ȝ ) lGcb (Ȝ ) (4.154)
From Equations 4.153 and 4.154,
³ g zGzdȍ aGc b (~z , Ȝ ) lGcb (Ȝ ) (4.155)

ȍ
Therefore, if we substitute Equation 4.155 into Equation 4.144, then
~
G\ ³ g bGbdȍ lGcb (Ȝ ) aGc b ( z , Ȝ ) (4.156)
ȍ
It is noted that the adjoint variable vector Ȝ in Equation 4.156 is the one obtained
from Equation 4.152.
The method with the continuum equations is mathematically well verified. It is
explained in detail in the references (Choi and Kim 2004a, 2004b). In the direct
differentiation method, the sensitivity information is obtained by substituting Gz
of Equation 4.150 into Equation 4.144. On the other hand, the sensitivity
information is obtained from Equation 4.156 in the adjoint variable method. The
difference from the method with discrete equations is that wK / wbi does not need
to be obtained in the method with continuum equations. The reason is that the
continuum equation is differentiated first and then discretized. A simple example
will be presented later.
The method with continuum equations is classified into two methods as shown
in Table 4.2. They are the continuum–continuum method and the continuum–
discrete method. The continuum–continuum method is used when Gz of Equation
4.150 can be obtained exactly. It can only be applied to simple problems. In
practical problems, the continuum–discrete method is generally employed. That is,
Equation 4.150 or Equation 4.156 is discretized by approximation and the
sensitivity information is evaluated.
Example 4.13 [Sensitivity Analysis Using the Continuum–Continuum

Method]
A clamped beam with a uniform load is illustrated in Figure 4.7. In Figure 4.7,
l 1 m, A 0.005 m 2 , E 200 GPa, D 1 / 6, I DA2 , F 49.61 kN/m and
J 77.126 N/m. The external force is f F AJ and the displacement curve is
z 2.5 u 10 3 [ x 2 (1 x) 2 ] . The design variable vector is b [ E A]T . The
compliance of the structure is
l l
\ ³ fzdz ³ ( F JA) zdx (4.157)
0 0
Obtain the sensitivity of Equation 4.157 with respect to the design variables by
using the direct differentiation method and the adjoint variable method. Use the
continuum–continuum method.
Solution
Using the variational method, the governing equation for a beam is
l l l
ab ( z , z ) { ³ EDA2 z xx z xx dx ³ fz dx ³ ( F JA) z dx { lb ( z ) (4.158)
0 0 0
w2z
where z xx . Differentiation of Equation 4.157 yields
wx 2
x h
z l w
Figure 4.7. A clamped beam with a uniform load

l
G\ ³ [( F JA)Gz JzGA]dx (4.159)
0

Gz in Equation 4.159 is obtained from Equation 4.150. Each term of
Equation 4.150 is
l
2
ab (Gz , z ) ³ EDA Gz xx z xx dx (4.160)
0
l
lGcb ( z ) ³ Jz dxGA (4.161)
0
l l
aGc b (~z , z ) 2
³ DA z xx z xx dxGE ³ 2 EDAz xx z xx dxGA
0 0
1l 2 2l 2
³ EDA z xx z xx dxGE ³ EDA z xx z xx dxGA (4.162)
E0 A0
From Equation 4.158, Equation 4.162 yields
1l 2l
aGc b ( z, z ) ³ ( F JA) z dxGE ³ ( F JA) z dxGA (4.163)
E0 A0
Substituting Equations 4.160–4.161 and 4.163 into Equation 4.150 yields
l l 1l 2l
2
³ EDA Gz xx z xx dx ³ Jz dxGA ³ ( F JA) z dxGE ³ ( F JA) z dxGA
0 0 E0 A0
(4.164)
From Equation 4.164, Gz is the displacement of Figure 4.7 when the

1 2
external force JGA ( F JA)GE ( F JA)GA is imposed on the
E A
beam.
Therefore, Gz in Equation 4.164 is expressed as
x 2 (1 x 2 ) ª 1 2 º
Gz 2 «
JGA ( F JA)GE ( F JA)GA» (4.165)
24 EDA ¬ E A ¼
Substituting Equation 4.165 into Equation 4.159 with real values, the
sensitivity is
G\ 1653.75GA 2.08 u 10 11 GE (4.166)

From Equation 4.152, the adjoint equation is
ab (O , O ) ³ ( F JA)O dx (4.167)
Equation 4.167 is the same as Equation 4.158, which is the governing

equation of the beam. Therefore, the adjoint variable is O z. From
Equation 4.156, the sensitivity is
~
G\ ³ g bGbdȍ lGcb ( z ) aGc b ( z , z )
ȍ
l l
2 2 2
³ [2Jz 2 EDA( z xx ) ]GAdx [ ³ DA ( z xx ) dx]GE
0 0
ªl ª 25,000 º º
2 2
« ³ «385.6 x (1 x) (6 x 2 6 x 1) 2 » dx »GA 2.08 u 10 11GE
¬0 ¬ 3 ¼ ¼
1653.75GA 2.08 u 10 11 GE (4.168)
4.5 Methods of Structural Optimization

In structural optimization, design is carried out by integrating the finite element
analysis, the sensitivity analysis and an optimization algorithm. The method for
structural optimization is classified into two methods: the direct method and the
approximation method. An optimization algorithm controls the entire process of
optimization in the direct method, while the algorithm controls some parts of the
process in the approximation method.
4.5.1 Direct Method
A structural optimization problem is formulated as Equation 4.44 and a nonlinear

programming algorithm finds the optimum solution. When a primal method of the
nonlinear programming algorithm is utilized, a reduced problem of Equation 4.46
is solved in an iterative fashion. The direct method finds the solution of Equation
4.46, which is the direction vector Gb and determines the stepsize through a line
search. A new design is made by the direction vector and the stepsize for a new
iteration. One of the nonlinear programming algorithms in Chapter 3 is directly
applied, and therefore, the performance of the algorithm is significant. The flow of
the process is presented in Figure 4.8.
Given problem
Linearization
Subproblem
LP or QP algorithm
(LP, QP)
No
Check convergence criteria
Yes
New design
End
Figure 4.8. Flow chart of the direct method

In the direct method, the process of structural optimization is the same as that
of general optimization. However, convergence may be difficult for large-scale
problems. The main reason is that the direction vector can be different from the
overall design trend. Also, a number of finite element analyses can be carried out
to evaluate the stepsize and significant computational effort is required for the
finite element analyses. The convergence criteria of the direct method are the
same as those of the nonlinear programming algorithms in Chapter 3. That is, the
norm of the direction vector is small and the constraints are satisfied. The solution
of the direct method generally satisfies the Karush–Kuhn–Tucker (KKT) condition
and an excellent design is provided.
4.5.2 Approximation Method
The approximation method has been proposed to overcome the drawbacks of the
direct method. Approximated functions are generated by considering the
characteristics of structural design. The direct method is also an approximation
method in that the original functions are reduced to linear or quadratic functions.
However, the functions can be approximated to any nonlinear functions in the
approximation method. The functions in structural optimization are frequently
implicit functions with respect to design variables. The implicit functions are
approximated to explicit nonlinear functions.
An optimization problem is defined with the approximated functions. A
nonlinear programming algorithm finds an optimum of the approximated problem.
A typical flow of the approximation method is illustrated in Figure 4.9. The
optimization process of the approximated problem does not find the direction
vector; it finds a new design. When the process does not converge, the
approximated problem is redefined with the new design. The process is not called
Given problem
Approximation
Approximated problem LP or NLP algorithm
New design
No
Check convergence criteria
Yes
End
Figure 4.9. Flow chart of the approximation method
an iteration, which is the term used in the direct method, but is called a cycle.
Therefore, a nonlinear programming algorithm is used once at each cycle.
A new design is found at each cycle in the approximation method. A nonlinear
programming algorithm finds the new design with explicit functions. Therefore,
the performance of the algorithm does not have a large impact on the overall
efficiency. Instead, the quality of the approximation determines the efficiency.
Generally, the convergence criterion is that the objective function is not changed
while all the constraints are satisfied. It is not guaranteed that the final solution
satisfies the KKT condition. It is noted that the İ-active set strategy is also used in
the approximation method. That is, only violated constraints are approximated.
Generally, commercial software systems are coded by using the approximation
method. The functions can be approximated in various manners according to the
application area. Therefore, there are many approximation methods. In this
section, the basic concept is introduced with the results of the early research.
Although Equation 4.46e is used in the direct method, it is a linear
approximation in some sense. From an approximation viewpoint, Equation 4.46e
is
n § wg ·
gL g (b 0 ) ¦ (bi bi 0 )¨¨ ¸
¸ (4.169)
i 1 © wbi ¹ b0
where g L is a linearly approximated function.

Instead of the design variable vector b of 4.169, the parameters defined in
Equation 4.170 can be used:
c c(b) (4.170)
Actually, a parameter is usually defined for a corresponding design variable as
ci ci (bi ), i 1,, n (4.171)
Using the parameters, Equation 4.169 is expressed as
n wg wbi
g A (b) g (b 0 ) ¦ (c i (bi ) c i (b0i )) (4.172)
i 1 wbi wc i b0
For example, a parameter defined as ci 1 / bi is often used. Then Equation 4.172

yields
n b0i wg
g A (b) g (b 0 ) ¦ (bi b0i ) (4.173)
i 1 bi wbi b0
Many other approximation methods are proposed. They are usually defined as in
the above concept (Haftka and GĂrdal 1992).
There is a particular approximation method for a stress constraint function.
Suppose the stress in a beam element of the finite element method is as follows:
F (b) Mc(b) Tr (b)

g V all 0 (4.174)
A(b) I (b) J (b)
where A, I, J, F, M, T, c and r are the area, the moment of inertia, the polar
moment of inertia, the normal force on the area, the moment, the torsion, the
distance from the neutral axis and the distance from the center, respectively.
Equation 4.174 can be approximated to a nonlinear function of b as follows:
wF § wM · § wT ·
F Gb ¨ M Gb ¸c(b) ¨ T ¸r (b)
wb © w b ¹ © wb ¹
g V all 0 (4.175)
A(b) I (b) J (b)
This approximation method is called force approximation (Vanderplaats 1984). In

Equation 4.175, the stress is not directly approximated. Instead, the internal force
or moment is approximated.
So far, the approximation is carried out based on one design point. A function
can be approximated by considering a wide range such as two points. The two-
point exponential approximation is a representative method. An approximation
function
pi
ci bi (4.176)
is used. Using Equation 4.176, the function g is linearly expanded by the Taylor
series as follows:
p
n ª§ b · i º§ b ·§ wg ·
gA g (b 0 ) ¦ «¨¨ i ¸¸ 1»¨¨ 0i ¸¸¨¨ ¸¸ (4.177)
i 1«© b0i ¹ »© pi ¹© wbi ¹b 0
¬ ¼
If we use the condition that the derivatives of g and g A are the same at an another
design point b1 , then p i is
ª§ wg · § wg · º
log «¨¨ ¸
¸
¨
¨ wb
¸ »
¸ »
«© wbi ¹ b1 © i ¹ b0 ¼
¬
pi 1 (4.178)
§b ·
log¨¨ 1i ¸
¸
© b0 i ¹
where p i has a value from –1 to 1. When the value in the logarithm is negative,
then p i 1 . If the result of Equation 4.178 is that p i 0 , the following equation
is used by using the L’Hospital theorem:
ª§ b pi º
·
«¨ i ¸ 1»
«¨© b0i ¸
¹ » §b ·
¬ ¼
lim log¨¨ i ¸
¸ (4.179)
pi o 0 pi © b0 i ¹
The result of Equation 4.179 is substituted into Equation 4.177.

As mentioned earlier, there are many other methods for the two-point
exponential approximation. Most of them are similar to the above method. They
are usually defined by modification of Equation 4.177. Approximations can be
developed using more than two points. They are generally classified into two
methods. One approximates functions with multiple points cumulatively and the
other uses a regression model such as the response surface method. The response
surface method will be introduced in Chapter 6.
The user chooses between the direct method and the approximation method.
Generally, the direct method is recommended to simply solve one or two problems.
On the other hand, the approximation method is recommended to repeatedly solve
large-scale problems or to develop a generalized system. The two methods are
compared in Table 4.4 (Park et al. 1994).
Table 4.4. Comparison of the direct and approximation methods
Direct method Approximation method

Important Efficiency of the NLP algorithm Accuracy of the
aspect used approximation
The number of finite element
Calculation One finite element analysis is
analyses can be increased due to
time carried out in each cycle
the line search
Quality of A mathematical optimum is The solution may not be a ma-
the solution obtained. thematical optimum
Approximation can be defined
Any type of the problem can be
Application differently according to the
solved in the same fashion
problem
Programming for the
The amount of programming is
Programming approximation process is
relatively small
required
4.5.3 Multiple Loading Conditions and Design Variable Linking
Two techniques are often used in structural optimization. They are the application
of multiple loading conditions and design variable linking. In some cases, we have
many loading conditions. When there are multiple loading conditions, the same
number of finite element analyses is performed and the results are incorporated in
the optimization process.
Suppose that the objective function is a function of only design variables and
that constraints are imposed on the displacements and the stresses. Then the
optimization formulation with multiple loading conditions is
Find b R n , z Rl (4.180a)
subject to K (b )z J fJ , J 1, , NLC (4.180c)
gj max{g j (b, z J ) d 0 }, J 1, , NLC, j 1, , m

(4.180d)
b L d b d bU (4.180e)
where NLC is the number of multiple loading conditions. Since the natural
frequency is not dependent on the external loads, it is excluded in Equation 4.180.
In Equation 4.180c, it seems that a finite element analysis is carried out for each of
the multiple loading conditions. However, it should be noted that the external load
vectors appear only on the right hand side. Therefore, the computational effort is
almost the same as that with the single loading condition because all the loading
conditions are simultaneously considered in the process of solving the
simultaneous equations. Since the worst case is taken into account in Equation
4.180d, the number of constraints does not increase compared to the single loading
case.
A problem can include the displacement vector z in the objective function. In
this case, an artificial variable E is adopted and Equation 4.180b is modified to
Minimize E (4.181a)
subject to max{ f (b, z J )} E , J 1, , NLC (4.181b)
That is, the objective function is changed to an objective function and constraints.
In structural design, some design variables must have the same value. The
following equation is employed for design variable linking:
b Ab el (4.182)
where b el is the vector for all the variables and A is the matrix for linking them.
Design variable linking not only enables a meaningful design in an engineering
sense but also reduces the number of design variables in the optimization process.
Because the size of optimization depends on the number of design variables,
design variable linking is fairly useful.
The İ-active set strategy in Chapter 3 excludes constraints that are sufficiently
satisfied. Generally, the number of displacement constraints is not very large.
However, because the stress constraints are defined at all the elements of the finite
element analysis, the number of stress constraints can be extremely large. The İ-
active set strategy can be exploited in this case. With the İ-active set strategy, the
number of constraints in the subproblem is reduced and the effort for sensitivity
analysis is significantly reduced.
Example 4.14 [Optimization of a Beam Using the Direct and

Approximation Methods]
A cantilever beam is illustrated in Figure 4.10: l 1 m, w 0.003 m, h 0.005 m,
l w
Figure 4.10. A cantilever beam of a rectangular cross section with a load at the tip
E 200 GPa, D DA 2 and P 10 7 N . A stress constraint is defined at

1 / 6, I
the top of the cross section at the fixed end as g V / V all 1 d 0 . Approximate
the constraint with respect to w by using the linear approximation and the force
approximation methods.
Solution
(1) Linear approximation

The stress is
h h
Pl Pl
2 2 Pl
V 2 2
6.67 u 1014
I DA 2Dw h
Each term of Equation 4.169 is
V 6.67 u 1014
g (b 0 ) = 1 1 (4.183)
V all V all
wg Pl 4.44 u 1017
(4.184)
ww Dw3hV all V all
wg Pl 1.33 u 1017
(4.185)
wh 2Dw 2 h 2V all V all
If we substitute Equations 4.183–4.185 into Equation 4.169, the following

linear equation is obtained:
6.67 u 10 4 4.44 u 1017 u ( w 0.003) 1.33 u 1017 u (h 0.005)

gL 1
V all V all V all
(4.186)
(2) Force approximation

The moment at the clamped end is M Pl . The differentiation of M is
wM
0 (4.187)
wb
From Equation 4.175, the following equation is obtained:
h
M
2 Plh
gA d1 (4.188)
IV all 2Dw 2 h 2V all
Table 4.5. Examples of size variables for elements of the FEM
Finite element Size variable
Truss element Area, density, elastic modulus

Area, cross sectional properties, moment of inertia,
Beam element
density, elastic modulus
Membrane element, plate
Thickness, density, elastic modulus
element, shell element
4.6 Application of Structural Optimization

The application of structural optimization is classified into three areas. They are
size optimization, shape optimization and topology optimization. The
classification is made by the changing pattern of the domain in the finite element
analysis. In size optimization, the domain is not changed. In other words, the
location of the nodes and the element shape are fixed. In shape optimization, their
location changes because they are the design variables. Existences of materials in
the domain of the finite element analysis comprise the design variables in topology
optimization. Topology optimization is used to optimally distribute material
within the design domain. The techniques involved in topology optimization are
quite different from those involved in size or shape optimization. Methods for
each optimization are briefly explained in this section. Details are given in related
references.
4.6.1 Size Optimization
The sizes of structures are the design variables in size optimization. The
optimization process starts from design parameterization. Design variables can be
selected from any variables that do not change the domain of the finite element
analysis. For example, they are the cross sectional area of the truss element, cross
sectional properties of the beam element and the thickness of the shell element.
Examples of design variables are shown in Table 4.5. The theories in Sections
4.3í4.5 can be directly applied to size optimization. Thus, a detailed explanation
is omitted and instead, some examples are given.
Example 4.15 [Optimization Results of the Three Bar Truss]

Optimize the three bar structure of Example 4.4.
Solution
Table 4.6. History of the optimization process for the three bar truss
Itera- Objective Active

b1 b2 b3 constraints
tion function
1 4.000000Eí4 2.000000Eí4 4.000000Eí4 1.042463E+1
2 4.000000Eí6 4.000000Eí6 4.000000Eí6 1.199063Eí1 1, 2, 3
3 9.491768Eí5 1.061901Eí4 2.609107Eí5 2.171434E+0 1, 2
4 2.063529Eí4 4.112808Eí5 4.000000Eí6 2.651332E+0 2, 3
5 1.667416Eí4 9.626272Eí5 4.175927Eí6 2.646356E+0 1, 2
6 1.719769Eí4 8.896829Eí5 4.000000Eí6 2.645265E+0 1, 2
Table 4.7. History of the optimization process for the member beam
Active
Itera- Objective
b1 h1 b2 h2 constr-
tion function
aint
1 5.0000Eí2 5.0000Eí2 5.0000Eí2 5.0000Eí2 3.9450E+1 2, 3
2 5.3312Eí2 1.4999Eí1 5.4378Eí2 1.5000Eí1 1.2745E+2 2, 3
3 2.5302Eí2 2.4990Eí1 1.6513Eí2 2.5000Eí1 8.2461E+1 2, 3
4 5.0000Eí3 4.1650Eí1 9.2194Eí3 4.1667Eí1 4.6740E+1 2
5 5.0000Eí3 5.4602Eí1 5.0003Eí3 4.6454Eí1 3.9867E+1 2, 3
6 5.0000Eí3 5.5591Eí1 5.0000Eí3 4.6879Eí1 4.0425E+1 2, 3
Based on the results of Example 4.4, the history of the optimization process is
shown in Table 4.6. In Table 4.6, it is noted that the design variable b3 converges
to the lower limit. Thus, we do not need the member of the right hand side.
Example 4.16 [Optimization Results of the Two Member Beam]

Optimize the two member beam of Example 4.5. The allowable value for the
vertical displacement of node 2 is 0.01 m, the allowable stress is 165 GPa , the
initial value of each design variable is 0.05 m, Young’s modulus E is 200 GPa ,
the density is 7890 kg / m 3 , the load is P 100 kN, and the length of each
member L is 1 m.
Solution
The history of the optimization process is shown in Table 4.7. The height of the
cross section is increased and the width is reduced to the lower limit. Overall, the
mass is reduced and the moment of inertia is increased.
r1
L d e
r2
(a) Circle (b) Domains with some definite shapes (c) Arbitrary shape
Figure 4.11. Examples of shape design variables
4.6.2 Shape Optimization
The domain of the finite element method is the design variable in shape
optimization. As in size optimization, shape design parameterization is the first
step in shape optimization. The optimization problem is formulated as Equation
4.44 and the subproblem is the same as in Equation 4.46. The obvious difference
from size optimization is the process of sensitivity analysis.
Shape Design Parameterization

As mentioned earlier, when design variables change, the domain of the finite
element method changes. Some shape variables are illustrated in Figures 4.11a
and 4.11b. If Figures 4.11a and 4.11b illustrate cross sections of beam elements,
r , r1 , r2 , d and L are size variables. However, when the cross sections are
modelled by many finite elements (mesh), then they become shape variables. The
reason is because the domain changes due to the change of the design variables.
The boundary of the shape in Figure 4.11c can be a design variable and it is also a
shape variable. In general, the boundary is expressed by polynomials or spline
functions and the coefficients of the functions are design variables.
Sensitivity Analysis
In shape optimization, sensitivity analysis is performed as well to establish the
subproblem in Equation 4.46. In general, the shape optimization problem is larger
than the size optimization problem. Therefore, the finite difference method is
rarely used in sensitivity analysis due to cost. The semianalytic method in
Equations 4.136 and 4.137 is usually adopted in commercial software systems. In
shape optimization, numerical error of the semianalytic method can be very large
according to the perturbation of the design variables. In this section, the method
WV (x )
īĲ
x xW
ȍ ȍW
Figure 4.12. Velocity field according to the change of the domain
with continuum equations is introduced. It is explained well in the references

(Choi and Kim 2004a, 2004b).
Since the domain is changed in shape optimization, sensitivity analysis is
performed by using the material derivative or the control volume method. It has
been reported that the two methods are basically the same. The method of the
material derivative is briefly explained here. When the design variable is changed,
the points in the domain move as illustrated in Figure 4.12. A time-like variable W
is defined and the velocity field V represents the movement. ȍ and ī are the
initial domain and boundary, while ȍ Ĳ and ī Ĳ are the changed domain and
boundary, respectively.
The material derivative of z is z , which is calculated as
z ( x) z ' (x) T zV ( x) (4.189)
where
ª z ( x ) z ( x) º
z ' ( x) lim « W » (4.190)
W o0 ¬ W ¼
Using the principle of virtual work and Equation 4.146, the governing equation
in the domain is
a ȍW ( z, z ) l ȍĲ ( z ) (4.191)
where z is the kinematically admissible displacement vector.

Suppose we have the following formula:
\ ³³ fW (xW )dȍW (4.192)

ȍĲ
The result of the material derivative of Equation 4.192 is
T
\ ³³ f ' ( x)dȍ ³ f ( x)(V n)dī ³³ [ f ' (x) div ( f (x)V (x))]dȍ (4.193)
ȍ ī ȍ
where n is the normal vector at the boundary and div means the operator of
wf wf wf
divergence and div f .
wx1 wx2 wx3
Equation 4.191 is
aȍW (zW , zW ) { ³³ c(zW , zW )dȍW ³³ fzW dȍW { lȍĲ ( zW ) (4.194)

ȍĲ ȍĲ
The material derivative of Equation 4.194 is
a ȍ (zW , zW ) a'V ( z, z ) aȍ (z , z ) lV ( z ) (4.195)
where
a' V ( z, z ) ³³ [c( z, T zV ) c( T zV , z )]dȍ ³ c(z, z )(V T n)dī (4.196)

ȍ ī
l V (z ) ³³ f ( T zV )dȍ ³ fz (V T n)dī (4.197)

ȍ ī
aȍ (z , z ) l 'V ( z ) a'V (z, z )

T T T
³³ [c(z, zV ) c( zV, z ) f ( zV )]dȍ
ȍ
³ [ fz c(z, z )](VT n)dī (4.198)

ī
Suppose we have the following formula for shape optimization:
\ ³³ g (b, z )dȍW (4.199)

ȍĲ
The material derivative of Equation 4.199 is
T
\ ³³ ( g z z ' g bGb)dȍ ³ g (V n)dī
ȍ ī
T T
³³ g z z g z ( zV ) g b Gb)dȍ ³ g (V n)dī (4.200)
: ī
In the direct differentiation method, z is obtained from Equation 4.198 and

substituted into Equation 4.200. In the adjoint variable method, z is eliminated in
Equation 4.200 by substituting Equations 4.201í4.203 into Equation 4.200,
aȍ ( Ȝ , Ȝ ) ³³ g z Ȝdȍ (4.201)
ȍ
aȍ (Ȝ , z ) ³³ g z z dȍ (4.202)
ȍ
aȍ (z , Ȝ ) ³³ g z z dȍ l 'V ( Ȝ ) a ' V ( z , Ȝ ) (4.203)

ȍ
where Ȝ is an adjoint variable vector that has a similar concept as in Equation

4.152.
\ l 'V (Ȝ ) a 'V (z, Ȝ ) ³³ g z (z T V )dȍ ³ g (V T n)dī

ȍ *
T T T T
³³ [c(z, ȜV ) f ( ȜV) c( zV , Ȝ ) g z ( zV )]dȍ
ȍ
³ [ g fȜ c(z, Ȝ )](V T n)dī (4.204)

ī
After the adjoint variable vector Ȝ is obtained from Equation 4.201, Equation
4.204 is calculated. It can be analytically derived in simple problems. In most
problems, it is obtained from a process similar to the assembling process of the
finite element method. The sensitivity in Equation 4.204 can be calculated in the
domain or the boundary. Green’s theorem is utilized in this process.
The results of Equation 4.204 depend on the velocity field V; therefore, the
definition of the velocity field is important. It is related to the design variables.
The boundary of the domain can be defined by certain vectors or functions. A
change of the mesh for the finite element follows and the velocity field is
determined. Details are explained in the references (Choi and Kim 2004a, 2004b).
Example 4.17 [Sensitivity Information for a Shape Variable in a

Clamped Beam]
A clamped beam is illustrated in Figure 4.7. The length l is a shape variable and
the uniform distributed load is f. The compliance function is
l
\ ³ fzdx (4.205)
0
Using the continuum–continuum method, obtain \ by (1) the direct method and
(2) the adjoint variable method.
Solution
(1) The direct method

\ can be obtained by substituting z from Equation 4.198 into Equation
4.200; however, this is a complicated procedure. A simple method is
presented here. The displacement curve of a clamped beam is
f
z ( x) ( x 2 (1 x) 2 ) (4.206)
2 EI
Substitution of Equation 4.206 into Equation 4.205 yields
f 2l 5
\ (4.207)
720 EI
Differentiation of Equation 4.207 with respect to l is
f 2l 4
\ Gl (4.208)
144 EI
The solution of the direct method in Example 4.13 can be obtained in the
same way.
As Example 4.13, O z . Equation 4.204 yields
l
\ ³ [2 EI ( z xV ) xx z xx 2 f ( z xV )]dx (4.209)
0
If we apply the integration by parts to Equation 4.209, then
l l l
\ 2 EIz xx ( z xV ) x 0 2( EIz xx ) x ( z xV ) 0 (2 fz EI ( z xx ) 2 )V (4.210)
0
The boundary conditions of Equation 4.210 are z (0) z (l ) z x ( 0) z x (l )

= 0. Then
l
\ EI ( z xx ) 2 V (4.211)
0
Gl Gl
In Equations 4.211 and Equation 4.205, V (l ) and V (l ) are
2 2
applied, and then
f 2l 4
\ Gl (4.212)
144 EI
It is noted that Equation 4.212 is the same as Equation 4.208.
Example 4.18 [Shape Optimization of a Connecting Rod]

The half model of a connecting rod is illustrated in Figure 4.13. A uniform load is
imposed on the surface of the circle of the left hand side. Find the shape that
minimizes the volume while the von Mises stress is less than the allowable value.
Solution
It is important to define shape variables because different design variables generate
different designs. Shape variables are defined as illustrated in Figure 4.13a. Shape
variables are defined by the thicknesses of the rings of both sides and the thickness
of the middle part that connects the rings. The optimization results are illustrated
in Figure 4.13b. It is found that all the thicknesses are reduced while the basic
shape is maintained.
The basic shape of the connecting rod is predetermined by the shape variables.
Thus, an abnormal shape (such as the shape of a saw) is not obtained. In most
software systems, the boundary is changed according to the predetermined curves
such as polynomials or spline functions. Then a realistic optimum shape can be
obtained.
(a) Connecting rod and shape variables
(b) Optimized shape
Figure 4.13. Shape optimization of a connecting rod

4.6.3 Topology Optimization
The procedure of topology optimization is different from that of size or shape

optimization. The distribution of materials is determined in topology optimization.
In other words, the existence of an element of the finite element method is
determined. Figure 4.14 presents the topology optimization results of an
automotive wheel cap. Dark areas mean that the elements exist. That is, the
optimum topology is obtained by allocating materials in the dark areas. Topology
optimization is carried out in two ways. First, it is applied to continuum structures
as in the case of Figure 4.14. Second, it is applied to discrete structures with truss
or beam elements. The nodes of truss or beam elements are fixed and the
connectivity is determined. In this section, the first case is explained.
There are two methods in topology optimization and they are the
homogenization method and the density method. The homogenization method
starts from the governing equation of a porous material with infinitely many holes
Figure 4.14. Results of topology optimization of an automotive wheel cap
1 b
a
1
(a) Porous material (b) A hole in a finite element
Figure 4.15. Holes in the homogenization method

in it. It is assumed that the holes are evenly distributed as illustrated in Figure
4.15a. In the average sense, the porous material can be modelled as Figure 4.15b
where a large hole exists in a finite element. If the hole is larger than a certain
value, the element is considered as an empty one. In the homogenization method,
the perturbation theory with the multiscale method is exploited. The method is
mathematically well established and it is reliable. The density method regards the
density of an element as a design variable. The density is used to calculate the
modulus of each element. If the density is close to zero, the corresponding
element is considered empty. The density method is slightly weaker in
mathematics than the homogenization method; however, it has been popularized
due to simplicity of the theory and numerical implementation. The density method
is introduced here. Details are given in the references (BednsIe 1998, 2003,
Hassani 1999, Mlejnek 1990).
In general, the stiffness of a structure is maximized in topology optimization.
Maximization of the stiffness is equivalent to minimization of the compliance. The
goal of topology optimization is generally minimization of the compliance. If we
substitute z into z in Equation 4.191, then
a(z, z ) l (z) (4.213)
The optimization problem to minimize the average compliance is
Minimize l (z ) (4.214)
Considering Equation 4.213, Equation 4.214 is a problem of minimizing the strain

energy from the total potential energy in Section 4.2.
If the material is assumed to be isotropic and Young’s modulus is virtually
defined as follows:
E U p E0 (4.215)
where U is the density, which is a design variable and E0 is the real Young’s
modulus. The coefficient p is defined as p ! 1 and Young’s modulus depends on
the density.
Using Equation 4.125, the optimization formulation of topology optimization is
as follows:
Find U (4.216a)
to minimize l (z ) (4.216b)
subject to a(z, z ) l (z ) (4.216c)
³ U dȍ d V , 0 U min d U 1 (4.216d)
ȍ
where V is the given limit of the volume. The problem of Equation 4.216 has
many design variables and a small number of constraints. Thus, the optimality
criterion method is often employed to find the optimum solution. It numerically
solves the KKT condition.
Details are not explained here. The process to update design variables is
max{(1 9 ) U k , U min }, if U k Bk d max{(1 9 ) U k , U min }

°
U k 1 ® min{(1 9 ) U k ,1}, if min{(1 9 ) U k ,1} d U k BKk (4.217)
° U BK , otherwise
¯ k k
where k is the iteration number, K is a tuning parameter and 9 is the move limit.
Bk is calculated from the Lagrange multiplier / k as follows:
pU kp 1 EH (z k )H (z k ) /k (4.218)
Bk /k1 pU p 1EH (z k )H (z k ) (4.219)
Recently, the technique of size optimization has been utilized to solve the
topology optimization problem. Using the finite element equation and Equation
4.215, topology optimization is formulated as
Find Ui , i 1,, NE (4.220a)
to minimize f T z (4.220b)
NE
p
subject to ¦ Ui k i f (4.220c)
i 1
NE
¦ vi U i d V , 0 U min d U i d 1, i 1, , NE (4.220d)
i 1
where vi is the volume of the ith element and NE is the number of finite elements.
Sensitivity analysis requires the use of the technique of size optimization. As
shown in Equation 4.220c, the design variables (density) are separated from the
finite element equation. Therefore, sensitivity analysis is fairly simple. Since
topology optimization has many design variables, the transformation method in
Section 3.8.3 is often used rather than the primal method in Section 3.8.2.
Example 4.19 [Michell Structure]

A cantilever plate is illustrated in Figure 4.16a. The circumference of the circle is
fixed and a concentrated load is imposed at the tip. This structure is called the
Michell structure and is often used for evaluation of topology optimization
methods. The weight is to be minimized while 20% of the mass is maintained.
Perform topology optimization of the structure.
(a) A cantilever plate with a concentrated load
(b) Results of topology optimization
Figure 4.16. Topology optimization of the Michell structure
Solution
Figure 4.16b presents the results of topology optimization by using the density
method. Materials should be kept in the dark areas.
4.7 Exercises
4.1 A vertical load is imposed at the tip of a structure with two truss elements.
Evaluate the displacement of the tip. A 200 m 2 , E 200 GPa ,
L 1.0 m and P 100 kN.
L
$
60
60$ P
4.2 Solve Exercise 4.1 when each member has two truss elements as
illustrated in the following figure. Compare the results with those of
Exercise 4.1 and analyze the results.
L
$
60
60$ P
4.3 Eliminate the equality constraints and make subproblems for the following
problems. The design variable is b and the initial value is b 0 1.
(1)
Find b, z
to minimize f b2 z 2
subject to g1 b2 z 1 t 0
h1 5z b 2 4 0
(2)
Find b, z
to minimize f b2
subject to g1 b2 z 2 1 t 0
g2 4b z 3 10 d 0
h1 z 2b 2 2 0
4.4 Make a subproblem for Exercise 4.1. The cost function is the volume of
the structure, the design variables are the areas of the members and a
displacement constraint is applied at the tip.
4.5 A clamped beam has a concentrated load at the center. The volume of the
structure is to be minimized while a displacement constraint is defined at
the center. Design variables are the width and the height of the cross
section. Make a subproblem. The structure is modelled by two beam

elements. Refer to Example 4.5 for the element stiffness matrix.
E 200 GPa , l 1.0 m, P 200 kN and the initial values of the design
variables are w0 h0 0.01 m.
l/2
P
l w
4.6 In Exercise 4.1, evaluate the sensitivity of the displacement at the tip with
respect to the area by using the following methods when the perturbation
of the area is 10 3 m 2 :
(1) Forward and central difference methods of the finite difference
method.
(2) Direct and adjoint variable methods of the analytic method.
(3) Semianalytic method.
4.7 Evaluate the sensitivity of the displacement with respect to the width and
the height of the cross section for the following structures. Use the direct
and the adjoint variable methods with the continuum–continuum method.
Refer to the data of Example 4.13.
(1)
x h
z l w
(2)
z x h
l w
4.8 Solve Example 4.14 for the following structures:

(1)
l/2
P
z x h
l w
(2)
l/2
P
z x h
l w
4.9 Perform optimization using Exercise 4.1. The objective function is the
total volume of the structure, the design variables are the cross sectional
areas of the members and the tip displacement should be less than 0.05 m.
The upper and lower limits of the design variables are 10 1 m 2 and
10 8 m 2 , respectively.
4.10 In Exercise 4.7, evaluate the sensitivity of the displacement with respect to
the shape variable l by using the direct and adjoint variable methods. Use
the data of Example 4.13 and refer to Example 4.17 for the derivation
process.
4.A Automotive Door Design with the ULSAB Concept

Using Structural Optimization
4.A.1 Problem Description

The concept of the ultra light steel auto body (ULSAB) has been proposed to
reduce the weight of automobile structures. The concept of ULSAB consists of the
holistic design, the hydro-forming process and the tailor welded blank (TWB)
technology. The tailor welded blank technology can decrease the manufacturing
cost by reduction of the number of parts. Also, the elimination of unnecessary
parts reduces the weight of the structure. TWB technology can be applied to
automobile closures such as the hood, the door and the trunk lid, etc. TWB
technology is applied to the inner panel of the front door to reduce the weight.
First, the existing reinforcement is eliminated from the inner panel. The inner
panel is replaced by a few plates. The thicknesses of the plates are determined and
the plates are welded. The final product is manufactured by stamping of the
welded structure. Optimization is adopted for the determination of the thickness.
As mentioned earlier, the inner reinforcements of the inner panel of the front
door are removed. In TWB, the role of the inner reinforcements is transferred to
the thickness distribution. The finite element model of the entire door is
established and design is conducted on the inner panel. In conceptual design,
topology optimization is performed to determine the thickness distribution, the
number of parts and the rough welding lines. The designer can base these
decisions on the results of topology optimization. Size optimization is carried out
to find out the thickness of each part under various constraints. The thickness of
each part can have a standard value from an existing design code. The welding
lines are fine-tuned by shape optimization. Since shape optimization is applied to
an uneven surface, a special method is developed for the welding lines. Size and
shape optimization can be utilized simultaneously. Multiple loading conditions are
imposed according to the design specification. Structural optimization is carried
out by a commercial code called GENESIS. In this research, various optimization
techniques are involved. Since GENESIS has all the capabilities, a single finite
element model can be commonly used (Vanderplaats Research & Development,
2000).
4.A.2 The Design Process for Weight Reduction
Design Process for Weight Reduction Using TWB

Figure 4.A.1 presents automobile closures. The structures have reinforcements
and the TWB technology replaces the reinforcements with various plates. The
door is chosen to apply the TWB technology. A designed product is illustrated in
Figure 4.A.2. The inner panel of a door with TWB is illustrated in Figure 4.A.2a.
Door Hood Deck lid Hatch
Figure 4.A.1. Automobile closures
(a) Blank (b) After stamping
Figure 4.A.2. Application of the TWB technology
The panel typically consists of two blank pieces and the pieces have different
thicknesses (Part 1 and Part 2). The door in Figure 4.A.2b is manufactured by a
stamping process.
A design process is proposed in Figure 4.A.3. First, the design domain is
selected. The domain may be the entire area or in part of the structure. In
conceptual design, topology optimization is performed for several mass constraints
to determine the tendency of the stiffest structure. The objective is to minimize the
strain energy for the given loading conditions. According to the topology
optimization results, the design domain is divided into a few parts with different
thicknesses. The divided lines are equivalent to the welding lines.
In detailed design, size and shape optimizations are performed. Size
optimization begins with the results from the above part selection process. The
thickness of each part is determined. The final welding lines are rendered by
shape optimization. Sometimes, it is extremely difficult to use shape optimization
for a curved boundary such as the inner panel. Also, the design boundaries of
shape optimization exist within the domain of the finite element method. A
method for the problem is developed for usage with a commercial code.
In many applications, size and shape optimizations can be used simultaneously.
In this problem, there are large differences between size and shape sensitivities.
The order of the size sensitivity is considerably larger. The welding lines do not
move in simultaneous application. Therefore, they are utilized separately as
illustrated in Figure 4.A.3.
The Optimization Method

The formulation of topology optimization in Equation 4.216 is adopted to
minimize the compliance. It is equivalent to the maximization of the stiffness.
Since topology optimization does not give detailed dimensions, it is used in the
conceptual design process. The density method in Section 4.6.3 is employed. In a
detailed design, size and shape optimizations are used. The design variables in
size optimizations are the dimensions of height, width or thickness of the cross
sections. In shape optimization, the design variables are the perturbations of nodes
at the boundary. The thicknesses of the parts are found from size optimization and
the welding lines are determined from shape optimization. During shape
optimization, it is quite important to avoid distortion of the finite element mesh of
the structure. Moreover, the perturbation vector for shape change must be defined
appropriately. Special care is needed when the boundary for shape optimization is
within the domain of the finite element analysis. Therefore, conventional shape
optimization cannot be used directly and an element-based shape optimization
technique is proposed. A shape optimization process is defined as a cycle.
In the TWB problem, the welding lines should be determined while the overall
shape is maintained. If the shape variable changes drastically, the finite elements
can be distorted. Therefore, the upper and lower bounds are defined in the middle
of the adjacent elements as illustrated in Figure 4.A.4a. If the shape change hits
the bounds, then the welding lines are moved to the next elements in the
subsequent cycle. This is illustrated in Figures 4.A.4b and 4.A.4c. If the optimum
shape is in the middle of the bounds, the cycle terminates. The termination criteria
for the shape optimization are as follows:
Begin design (design domain)
Topology optimization (thickness distribution)
Part selection (designer’s decision-making)
Size optimization (optimum thickness)
Shape optimization (optimum welding line)
End design
Figure 4.A.3. Design process of the door using various optimization methods
(a) ith cycle
(b) (i+1)th cycle (case 1)
(c) (i+1)th cycle (case 2)
Figure 4.A.4. Shape optimization to determine the welding lines
(1) When the welding lines fluctuate, the lines of the current cycle are
considered as the final design.
(2) When there are no constraint violations at the ith cycle and some constraint
violations at the (i+1)th cycle, the final design is determined by the ith
cycle.
The design time increases with this method because the design change is small
in a cycle. Moreover, a better design may exist, which can be exploited when the
shape sensitivity is small and the shape boundary is within the domain.
Front Door Design Using the TWB Method

An automobile front door consists of several parts. An inner panel resides in the
front door and has a large portion of the total weight. The methods mentioned
earlier are applied to the inner panel. The design specifications for structural
performance of a new door design include frame rigidity and door sag. The
conditions for frame rigidity and door sag are defined to make the door frame
resist the bending effect due to negative pressure at high speeds and the sag from
the door weight, respectively.
x
Load case 1 Load case 2 Load case 3
Figure 4.A.5. Finite element model and loading conditions for the front door
The finite element model is illustrated in Figure 4.A.5 and represents boundary
conditions for analysis, the frame rigidity conditions (load cases 1 and 2) and the
sag condition (load case 3). The number 12345 means that only the rotation with
respect to the z-axis is free and 2356 means that the translation and rotation in the
x-axis are free. The number 2 means that only the translation in the y-axis is fixed.
The finite element model includes the inner and outer panels and the model is
verified via experiments.
The design domain is defined on the inner panel of the front door. The reasons
are that the TWB technology can be easily applied, and the additional effect of
reduction of both manufacturing processes and weight can be obtained by
elimination of reinforcements and some parts. Topology optimization is
performed to determine the number of parts and rough welding lines in conceptual
design. As mentioned earlier, the inner reinforcements are removed to apply the
TWB technology. This is illustrated in Figure 4.A.6. The stiffest structure is
found through topology optimization. A mass fraction constraint is defined for
three cases: 50%, 60% and 70% of the inner panel mass. The objective function is
the linear sum of strain energies under three loading conditions.
The results are illustrated in Figure 4.A.7. The inner panel is divided into three
parts based on the results of the topology optimization. The divided three parts are
Inner reinforcement
Inner reinforcement Inner panel
Figure 4.A.6. Elimination of reinforcements in the inner panel

(a) 50% (b) 60%
(c) 70%
Figure 4.A.7. The results of the topology optimization
represented by X 1 , X 2 and X 3 as illustrated in Figure 4.A.8.

The thickness of each part is determined by size optimization and the
formulation is as follows:
Find b [b1 b2 b3 ]T (4.A.1a)
to minimize weight (4.A.1b)

subject to G TWB
j G BASE
j d 0 . 0 , j = load case 1, 2, 3 (4.A.1c)
Figure 4.A.8. Division of the inner panel into three plates

0.5 d bi d 2.1 , i = load case 1, 2, 3 (4.A.1d)
bi / 2b j 1.0 d 0.0 , i, j parts 1, 2, 3, i z j (4.A.1e)
where b is the design variable vector for the thicknesses of the plates for the inner
panel, G BASE
j
is the displacement of the existing model with reinforcements where
the jth load is imposed, G TWB
j
is the displacement of the model with TWB at the
same place of G BASE
j
.
The starting values of b are set to 0.7 mm. It is noted that the thickness
difference between adjacent parts should not be very large in the TWB application.
Therefore, Equation 4.A.1e is defined so as to make one thickness not larger than
twice the size of the other. The constraint is considered as the formability of TWB.
The initial and optimum designs are shown in Table 4.A.1. The results in
Table 4.A.1 are obtained in a continuous space. However, they are not usable in
practical design and therefore, they should be modified to have discrete values.
An available rounded-up value is given to each thickness: b1 = b3 = 0.65 mm and
b2 = 1.3 mm. The rounded-up values are checked for the constraints in Equation
4.A.1. Since all the constraints are satisfied, they are taken as the final design.
Shape optimization is performed to determine the location of the welding lines.
It is noted that shape optimization can reduce the increased weight of the inner
panel due to the rounded-ups of the thickness. The formulation is as follows:
Find bi , i 1,..., 11 (4.A.2a)

subject to TWB BASE
G j G j d 0 . 0 , j = load case 1, 2, 3 (4.A.2c)
11
{ XYZ} { XYZ}0 ¦ bi { XYZP}i (4.A.2d)
i 1
Table 4.A.1. The results of size optimization
Initial value Optimum in Rounded up

with continuous space value with
reinforcement with TWB TWB
Objective 17.009 kg 16.777 kg 16.815 kg
b1 0.700 mm 0.647 mm 0.650 mm
b2 0.700 mm 1.290 mm 1.300 mm
b3 0.700 mm 0.647 mm 0.650 mm

Table 4.A.2. The results of shape optimization
Initial value Optimum at each cycle

1st cycle 16.815 kg 16.778 kg
2nd cycle 16.609 kg 16.569 kg
3rd cycle 16.413 kg 16.459 kg
3.0 d bi d 3.0, i 1, ,11 (4.A.2e)
where { XYZP}i are 11 perturbation vectors to change the domain of X 2 plate and
bi are 11 design variables that determine the amount of perturbations. The
welding lines in shape optimization should maintain the initial shape of the
structure. Thus, the element-based shape optimization technique is utilized and the
final welding line is identical to the boundary of X 2 .
Table 4.A.2 shows the optimization cycles. The lower and upper limits are
defined as Figure 4.A.4. The process terminates in three cycles. The optimization
results are illustrated in Figure 4.A.9. The shape sensitivities are smaller than the
size sensitivities. Therefore, drastic change is not made in the shape optimization
process. From the optimization process, the weight of the inner panel is reduced
by 8.72% compared to the existing one with reinforcements. Since the outer panel
is not included in the optimization process the weight of the entire door is reduced
by 3.23% compared to the original structure. The reduction may not seem to be
very large. However, it is quite a large amount because a car has four doors. Also,
it is noted that a better design exists with the TWB technology, which can reduce
manufacturing cost considerably.
Figure 4.A.9. Final welding lines

Conclusions
The optimization method for a lightweight TWB door has been investigated. The
design flow is defined with topology, size and shape optimizations. The sequence
of using optimization methods is appropriately defined for the door design. It is
found that the door with TWB is a typical industrial example where various
optimization methods can be utilized. In conceptual design, topology optimization
is performed to determine the number of TWB parts and rough welding lines. Size
and shape optimizations are conducted to find the optimum thickness of each part
and the optimum welding line locations, respectively. The inner panel of the door
is optimized, and the weight of the TWB inner panel is reduced by 8.72%
compared to the original inner panel.
References
Arora JS (1988) Introduction to Optimum Design. McGraw–Hill, New York
Arora JS (1997) Guide to Structural Optimization. The American Society of Civil
Engineers, New York
Arora JS, Haug EJ (1979) Methods of Design Sensitivity Analysis in Structural
Optimization. AIAA Journal 17:970í974
Bathe KJ (1996) Finite Element Procedure. Prentice Hall, Englewood Cliffs, NJ
BednsIe MP, Diaz A, Kikuchi N (1998) Topology and Generalized Layout
Optimization of Elastic Structures. Computer Methods in Applied Mechanics
and Engineering 71:197í224
BednsIe MP, Sigmund O (2003) Topology Optimization. Springer, Berlin
Heidelberg New York
Choi CK (2002) Finite Element Method. Techno Press, Daejon, Korea (in Korean)
Choi KK, Kim NH (2004a) Structural Sensitivity Analysis and Optimization I:
Linear Systems. Springer, Berlin Heidelberg New York
Choi KK, Kim NH (2004b) Structural Sensitivity Analysis and Optimization II:
Nonlinear Systems. Springer, Berlin Heidelberg New York
Cook RD, Malkus DS, Plesha ME (1989) Concepts and Applications of Finite
Element Analysis. 3rd ed., Wiley, New York
Dordrecht
Haftka RT, Adelman HM (1989) Recent Developments in Structural Sensitivity
Analysis. Structural and Multidisciplinary Optimization 1:137í151
Haftka RT, Grandhi RV (1985) Structural Shape Optimization–A Survey.
Computer Methods in Applied Mechanics and Engineering 57:99í106
Haug EJ, Arora JS (1979) Applied Optimal Design: Mechanical and Structural
Systems. Wiley Interscience, New York
Haug EJ, Choi KK, Komkov V (1986) Design Sensitivity Analysis of Structural
Systems. Academic Press, Orlando, FL
Hassani B, Hinton E (1999) Homogenization and Structural Topology
Optimization. Springer, Berlin Heidelberg New York
Lee KW, Park GJ (1997) Accuracy Test of Sensitivity Analysis in the

Semianalytic Method with Respect to Configuration Variables. Computers &
Structures 63:1139í1148
Logan DL (2002) A First Course in the Finite Element Method. 3rd ed.,
Brooks/Cole, Pacific Grove, CA
Mlejnek HP, Schirrmacher R. (1990) An Engineers Approach to Optimal Material
Distribution and Shape Finding. Computer Methods in Applied Mechanics and
Park GJ (1993) Accuracy Tests for Various Sensitivity Analysis Methods with
Respect to Shape Variables in Planar Cantilever Beams. Computers &
Structures 47:1057í1063
Park YS, Lee SH, Park GJ (1994) A Study of Direct vs. Approximation Methods
in Structural Optimization. Structural Optimization 10:64í66
Rao SS (1996) Engineering Optimization. 3rd ed., Wiley, New York
Rozvany GIN (1997) Topology Optimization in Structural Mechanics. Springer,
Schmit LA (1981) Structural Synthesisa–Its Genesis and Development. AIAA
Journal 19:1249í1263
Shin JK, Lee KH, Song SI, Park GJ (2002) Automotive Door Design with the
ULSAP Concept using Structural Optimization. Structural and Multidisciplinary
Optimization 23:320í327
Design with Applications. McGraw Hill, New York
Vanderplaats GN (1993) 30 Years of Modern Structural Optimization. Advances
in Engineering Software 16:81í88
Vanderplaats GN, Salajegheh E (1989) New Approximation Method for Stress
Constraints in Structural Synthesis. AIAA Journal 27:352í358
Vanderplaats Research & Development, Inc. (2000) GENESIS User's Manual. Ver.
6.0., Colorado Springs, CO
5 Dynamic Response Optimization
5.1 Introduction
In Chapter 4, structural optimization was introduced and the external loads were
imposed in a static manner. In this chapter, structural optimization when the
imposed loads have dynamic characteristics is explained. A dynamic load is a load
that has varying magnitude and direction with respect to time. Dynamic response
optimization is structural optimization based on the dynamic loads.
It is fairly important to precisely identify the external loads in structural design.
However, identifying the loads is difficult in most cases. Generally, the loads are
assumed to be static with some specific values and the design process is performed
with the static loads. In reality, loads are dynamically imposed on structures. A
dynamic load is transformed to a static load with a dynamic factor and a structure
is designed based on the transformed static load. However, the structure can be
over-designed or under-designed. Therefore, dynamic loads should be considered
directly in the design process. Dynamic response optimization directly handles
dynamic loads.
Structural optimization is classified into methods using sensitivity information
and methods without sensitivity information. In this chapter, the optimization
methods are assumed to use sensitivity information and the utilized analysis is
linear. Dynamic response optimization is classified into optimization in the time
domain and optimization in the frequency domain. The optimization methods
utilized are basically the same as those used in static response optimization. In
particular, the optimization process in the frequency domain is the same as that of
static response optimization. Since optimization in the time domain solves a
second-order differential equation, it is complicated. This chapter mostly deals
with optimization in the time domain. Optimization in the frequency domain will
be briefly explained at the end of this chapter.
5.2 Optimization in the Time Domain

In the 1970s, comprehensive studies began on optimization of structures subjected
to transient loads (Afimiwala and Mayne 1974, Fox and Kapoor 1970, Willmert
and Fox 1972). Before the early 1970s, research regarding frequency response
optimization was predominant. Some research on frequency response optimization
was published in the early 1970s in a reference (e.g. Pierson 1972). Optimization
in the time domain is formulated as follows:
Find b (5.1a)
subject to Mz Kz f (t ) (5.1c)
g j (b, z (t ), z (t ), z(t ), t ) d 0, j 1, , m (5.1d)
where t is time, M is the mass matrix, K is the stiffness matrix, f is the external
force vector, z is the displacement vector and z and z are velocity and
acceleration vectors, respectively. Other notations are the same as those of
Equation 3.1. Optimization in the time domain is attributed to structural
optimization; therefore, the contents in Chapter 4 are valid here. The difference is
that the equality constraints in Equation 5.1c yield a second-order differential
equation.
A typical objective function in the time domain is a min-max problem as
follows:
Minimize f max c 0 (b, z (t ), z (t ), z(t ), t ) (5.2)
A representative example of Equation 5.2 is to minimize the maximum

acceleration. Equation 5.2 can be modified by an artificial variable E as follows:
Minimize f E (5.3a)
subject to g max c 0 (b, z (t ), z (t ), z(t ), t ) d E (5.3b)
Equation 5.3 is equivalent to Equation 5.1b. Therefore, most optimization

problems in the time domain can be expressed as in Equation 5.1.
There are three topics in optimization in the time domain. They are the
treatment of time-dependent constraints, sensitivity analysis and approximation.
Most research in this area is focused on providing methods for the solution.
Dynamic Response Optimization 257
5.3 Time-dependent Constraints

Since the constraints in Equation 5.1d vary with respect to time, they are called
time-dependent constraints. An example of 5.1d is the stress constraint, which
should be less than the allowable stress during a time duration. Equation 5.1 is
stated in a continuous time domain. In most engineering practices, Equation 5.1 is
treated in a discrete time domain, since numerical methods are employed to solve it.
In a discrete time domain, Equation 5.1 yields the following:
Find b (5.4a)
subject to Mz(t u ) Kz (t u ) f (t u ), u 1, , q (5.4c)
g j (b, z (t u ), z (t u ), z(t u ), t u ) d 0, j 1, , m, u 1, , q
(5.4d)
Each constraint in Equation 5.1d is modified to q constraints as in Equation

5.4d. q is the number of time grid points in the time domain. As explained in
Chapter 3, the optimization method is classified into the direct method where each
constraint is separately considered and the transformation method where a
constraint is defined for all the constraints. Treatment of time-dependent
constraints with the two methods will be explained in subsequent sections.
5.3.1 Treatment of Time-dependent Constraints in the Direct

Method
A significant feature of the direct method is that a subproblem is to be solved. In
order to solve it, gradients of the objective function and constraints are required.
Because there are many constraints in a discrete time space, it is important to
reduce the number of constraints for gradient calculation.
A Method Using an Equivalent Functional

The constraints in Equation 5.4d can be transformed to an equivalent functional as
follows:
T
³ g j (b, z (t ), z (t ), z(t ), t ) ! dt d 0 (5.5)
0
where
0 if g j 0
g j (b, z (t ), z (t ), z(t ), t ) ! ® (5.6)
¯g j if g j t 0
0 t11 t12 t21 t22 t31 t32 Time
Figure 5.1. Constraint violation in multiple sections
where T is the duration of the time. The basic concept of this transformation is to
reduce many active or violated constraints to one constraint. Satisfaction of
Equation 5.5 is equivalent to satisfaction of Equation 5.1d and vice versa. The
transformation is reasonable in that the number of constraints is reduced
significantly.
When a constraint is violated in multiple sections as shown in Figure 5.1, the
number of the violated sections can be changed. Then Equation 5.5 becomes
nondifferentiable and it may be difficult to find an optimum. Moreover, when an
equivalent functional is used, we need a premise that the Lagrange multiplier is
time invariant during [0, T]. However, it is obvious that the Lagrange multiplier is
time variant. Therefore, the KKT necessary conditions are not generally the same
when Equation 5.1d is directly considered and an equivalent functional is used.
Each violated section can be separated, and a corresponding constraint is
defined. Equation 5.7 shows this method
tk 2
³ g j (b, z (t ), z (t ), z(t ), t ) ! dt d 0, j 1, , m (5.7)
tk 1
where tk1 is the starting point of the kth violated section and tk2 is the ending
point. This is illustrated in Figure 5.1.
Worst Case Approach

A method to overcome the above difficulty is to use pointwise constraints. A
pointwise constraint means a constraint that is defined at each time grid point in a
0 Time
Figure 5.2. Worst case approach

discrete time domain. The simplest method with this is the worst case approach,
which is illustrated in Figure 5.2.
This approach only takes the maximum violated response as a constraint. The
number of constraints is reduced from the number of time grid points to one,
unless the same valued multiple peaks exist along the time axis. However, the
time when the maximum violated response occurs generally varies as the
optimization process proceeds. A response in the time domain, which is not the
worst case in a previous iteration, can become the worst case in the next iteration.
Thus, the convergence can be slow or the problem can diverge.
An Approach Considering All the Local Maxima

In order to overcome the disadvantage of the worst case approach, one can
consider all the local maximum responses as constraints. Figure 5.3 represents this
treatment. It requires additional effort to find the times when the local maximum
responses occur. When an integration algorithm with a variable time step is used,
the search may not be easy.
When the time step D is constant and the number of time steps is sufficient,
this method is efficient. The performance of the algorithm to find the maxima is
important. An algorithm is introduced in a reference (Grandhi et al. 1986). In the
range where the violation is small, the search interval is increased because the
probability of the existence of a maximum is low.
The steps of the algorithm are as follows:
Step 1. Define various coefficients. nmin and nmax are used to determine the
smallest and largest intervals, respectively. t f is the ending point. g 0 is
defined as a basis value for increasing the search interval. Evaluate g(t) at
t t 0 . If g (t 0 ) g 0 (the violation is sufficiently small), then set n n min
and go to Step 2. Otherwise, evaluate g(t) at t t 0 D . If g (t ) g (t 0 ) ;
in other words, if g (t 0 ) is a local maximum, then go to Step 5. Otherwise,
search g(t) at t t 0 2D , t t 0 3D , until g (t ) g (t D ) . As a result,
g (t D ) is a local maximum. Go to Step 5.
Step 2. Set t t nD . If t t t f , stop. Otherwise, go to Step 3 to evaluate
g (t ).
g0
0 Time
Figure 5.3. The approach considering all the local maxima

Step 3. If g (t ) g 0 , then set n min (n 1, nmax ) and go to Step 2.

Otherwise, go to Step 4.
Step 4. The constraints are evaluated at the time steps surrounding the
current one. The maximum out of the three is found as follows:
(1) If the constraint at the first time step is the largest, a search is
performed in the opposite direction until g (t 2D ) g (t D ) and
g (t ) g (t D ). The stepsize is D . Then g (t D ) is a local
maximum.
(2) If the second time step is the largest, it is a local maximum.
(3) If the third one is the largest, the search is continued until g (t 2D )
g (t D ) and g (t ) g (t D ). The stepsize of the search is D . Then
g (t D ) is a local maximum. If the search continues until t t t f ,
then g (t f ) is a local maximum.
Step 5. Set n n min .

Step 6. Set t t nD . If t t t f , stop.
Step 7. If g (t ) g 0 , go to Step 3. If g (t ) d g (t nD ), go to Step 6.
Otherwise, set t t nD and go to Step 4(3).
Pointwise Constraints at Time Grid Points near the Local Maxima

Even though all the local maxima are considered for constraints, it is not
guaranteed that the local maximum points will not vary as the design proceeds. In
0 Time
Figure 5.4. Pointwise constraints near the local maxima
0 Time
Figure 5.5. Multiple equivalent functionals for Figure 5.4

0 Time
Figure 5.6. Pointwise constraints at all time grid points with an active set strategy
other words, the local maxima of the current iteration may not be local maxima in
the next iteration. Thus, a few points near the maximum can be included in the
constraint set. The method handles relatively many constraints and is illustrated in
Figure 5.4. The method in Figure 5.5 can be used. The points near the maximum
are considered in an equivalent functional.
Pointwise Constraints at All Time Grid Points with an Active Set Strategy
The simplest treatment of time-dependent constraints would be to consider
responses at all time grid points with an active constraint set strategy. Figure 5.6
represents this method and moreover, deals with the most constraints of all the
methods. This one is simple to implement. If a constraint is not violated at most
time grid points, it can be efficient, since the modern computing environment can
cover more constraints than ever.
Example 5.1 [Searching Local Maxima]

Find all the local maxima of the following function. The time step is 0.1 and
g 0 0,
g (t ) sin t sin 5t , 0 d t d 2
Solution
The time step D 0.1 , t 0 0, tf 2 , n min 1 and nmax 3. Table 5.1 shows
the calculation results.
The procedure is solved as follows:
(1) Since g (t 0 0) t g 0 g(t) is evaluated at t 0 0.1 0.1. Since g(t) =
0.57926, g(t) is evaluated until g (t ) g (t 0.1) at t = 0.2, t = 0.3,
(2) When t = 0.5, g (t ) g (t 0.1) . Thus, g (t 0.1) is a local maximum and
g(0.4) = 1.29872.
(3) Set t = t + 0.1 = 0.5. g(t) is evaluated at t = 0.4, t = 0.6, which are the
neighboring points of t = 0.5. The function is the largest at t = 0.4. From t
= 0.7, the function is searched with the time step until
Table 5.1. Function values with D 0.1
t g(t) t g(t)
0 0.0 1.1 0.18567
0.1 0.57926 1.2 0.65262
0.2 1.04014 1.3 1.17868
0.3 1.29302 1.4 1.64244
0.4 1.29872 1.5 1.93549
0.5 1.07790 1.6 1.98893
0.6 0.70576 1.7 1.79015
0.7 0.29343 1.8 1.38597
0.8 T0.03945 1.9 0.87115
0.9 T0.19420 2.0 0.36528
1.0 T0.11745
g (t 0.2) g (t 0.1) and g (t ) g (t 0.1) . The condition is satisfied

when t = 1.7. Thus g(t í 0.1) = g(1.6) = 1.98893 is a local maximum.
(4) Set n n min 1 and t = t + 0.1.
(5) At t = 1.8, g(t) = 1.38597 and g (t 0.1) 1.79015. Since g (t ) d g (t
nD ), set t t 0.1 1.9 .
(6) At t = 1.9, g (t ) d g (t nD ) . Thus, t t 0.1 2.0 .
(7) t 2.0 t f ; therefore, stop.
The local maxima are
g(0.4) = 1.29872, g(1.6) = 1.98893
The function is illustrated in Figure 5.7.
2.1 g (t )
1.6
1.1
0.6
0.1
-0.4 t
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Figure 5.7. Graph of g (t ) sin t sin 5t

5.3.2 Treatment of Time-dependent Constraints in the

Transformation Method
The key feature of the transformation method is to transform a constrained
problem into an unconstrained problem. Thus, we minimize only one function in
the transformation method. This is an attractive feature in that many time-
dependent constraints and an objective function can be merged into a single
function. Although there is an example that adopted the exterior penalty function
method, a representative of the transformation method is the augmented
Lagrangian (or Lagrange multiplier, ALM) method. The ALM method in dynamic
response optimization is discussed in this section.
There can be various forms of the augmented Lagrange function. The rules for
updating Lagrange multipliers and penalty parameters are dependent on the form
of the augmented Lagrange function. One of them is as follows (Paeng and Arora
1989):
1Tª m 2º
< (b, ș, r, T ) f (b) ³ « ¦ r j {max(0, g j (b) T j )} »dt (5.8)
2 0¬j 1 ¼
where r j T j { u j is the Lagrange multiplier corresponding to g j (b) and is a

continuous function with respect to time. T j is a positive coefficient for g j (b)
and r j is the penalty parameter for g j (b) . The second term of Equation 5.8 is
defined from Equation 5.1d.
The algorithm using Equation 5.8 is explained. If the maximum violation of
constraints is less than a certain value, the process terminates.
Step 1. Set k = 0 and V prev f . Define b ( 0) , ș (0) , r, D ! 1, E ! 1, H 1 ! 0

and H 2 ! 0 .
Step 2. k = k + 1
Step 3. Minimize the function in Equation 5.8. Suppose b (k ) is the optimum
of Equation 5.8.
Step 4. The maximum violation of constraints V is calculated as
V max§¨ max | max( g j , T j ) | ·¸ (5.9)

j © t ¹
If V H 1 and || < || H 2 , then stop. Otherwise, make the following set:
I { j : max | max( g j , T j ) | ! K / D , j 1, , m}
t
I is the set with constraints whose violations are larger than 1 / D times the
maximum violation of the previous iteration.
Step 5. If V t V prev , which means that the violation is not reduced, update
the parameters as follows:
rj E r j , T (j k 1) T (j k ) / E (5.10)
where j I . The Lagrange multiplier is not changed and the penalty

parameter is increased. Go to Step 2. If V V prev , which means that the
maximum violation is reduced, go to Step 6.
Step 6. T (k )
j is updated as follows:
T (j k 1) T (j k ) max( g j (b ( k ) ), T ( k ) ) (5.11)
It is noted that this step is carried out when the constraint violation is
improved. If V d V prev / D , set V prev V and go to Step 2. Otherwise, go
to Step 7.
Step 7. For j I , the parameters are updated as follows:
rj E r j , T (j k 1) T (j k 1) / E (5.12)
Go to Step 2. It is noted that this step is carried out when the constraint
violation is larger than 1 / D times the maximum violation of the previous
iteration.
The initial value of T j is usually set to 0. When D has a value from 2 to 3 and E
has a value from 6 to 10, the optimization process is stable.
In Equation 5.8, max(0, g T ) only accounts for the violated constraints at all
the time steps. This process has the effect that the active set strategy is employed.
The termination criteria in Step 4 guarantee that all the constraints are satisfied in
the time domain. Meanwhile, the adjoint variable method is effective for
sensitivity analysis of the augmented Lagrange function. This will be explained
later. In this method, the Lagrange multipliers should be updated carefully. There
are a few studies that investigate the update rules (Chahande and Arora 1993, Kim
and Choi 2000).
5.4 Sensitivity Analysis

One of the crucial aspects in the direct method of structural optimization is to
determine the direction vector precisely. The direction vector is usually evaluated
from a subproblem as shown in Equation 4.46. To make a subproblem, gradient
vectors such as f and g are required. The constraints to consider are
determined by the method in Section 5.3.1. The gradient vectors are required in
the transformation method as well. As explained in Chapter 4, when the analytic
or semianalytic method is used for sensitivity analysis, the sensitivity analysis
method is classified into the direct differentiation method and the adjoint variable
method. Sensitivity analysis in the time domain is basically the same as that of
static response optimization. It is different in that the governing equation is a
differential equation and the sensitivity analysis is derived by using the differential
equation.
The gradient vectors of the functions in Equation 5.1 are calculated as follows:
wf
T f (5.13a)
wb
wg j wg j dz wg j dz wg j dz
T g j (5.13b)
wb wz db wz db wz db
dz dz dz
The difficult terms to calculate in Equation 5.13 are , and . They are
db db db
obtained from the equilibrium (governing) equation in Equation 5.1c. Sensitivity
analysis with Equation 5.1c is discussed.
5.4.1 Direct Differentiation Method
The direct differentiation method is classified into two methods: the direct
integration method and the modal approximation method. In the direct integration
method, Equation 5.1c is differentiated with respect to design variables and then
integrated. In the second method, the size of Equation 5.1c is reduced by modal
approximation, the reduced equation is differentiated with respect to design
variables and then integrated. The integration method is the same as that for
second-order differential equations.
The Direct Integration Method

Differentiation of Equation 5.1c with respect to the ith design variable yields
d 2 § dz · § dz · df (t ) dM dK
M ¨ ¸ K¨ ¸ z z (5.14)
dt 2 ¨© dbi ¸
¹
¨ db
© i
¸
¹ dbi dbi dbi
dz
Equation 5.14 is a second-order differential equation of . A numerical method
dbi
such as the Newmark method can be employed to integrate Equation 5.14. Using
the Newmark method, the acceleration, velocity and displacement vectors can be
calculated with the time step 't as follows:
dz(t 't )
[M E ('t ) 2 K ] 1[ A1z (t ) A 2 z (t ) A 3z(t )
dbi
dz (t ) dz (t ) dz(t )
K A4 A5 A 6z(t 't )] (5.15a)
dbi dbi dbi
dK
A1 (5.15b)
dbi
dK
A2 't (5.15c)
dbi
§1 · 2 dK
A3 ¨ E ¸('t ) (5.15d)
©2 ¹ dbi
A4 't K (5.15e)
§1 · 2
A5 ¨ E ¸('t ) K (5.15f)
©2 ¹
dM dK
A6 E ('t ) 2 (5.15g)
dbi dbi
dz (t 't ) dz (t ) 't ª dz(t ) dz(t 't ) º

« » (5.16)
dbi dbi 2 ¬ dbi dbi ¼
dz (t 't ) dz (t ) dz (t ) § 1 · dz(t ) dz(t 't )

't ¨ E ¸('t ) 2 E ('t ) 2
dbi dbi dbi ©2 ¹ dbi dbi
(5.17)
where
d 2 § dz · dz d § dz · dz
¨ ¸ , ¨ ¸ (5.18)
dt 2 ¨© dbi ¸
¹ dbi dt ¨© dbi ¸
¹ dbi
Numerical stability is assured when E has a value between 0 and 1/6. Equations
5.15í5.17 are utilized in a subproblem and a direction vector is evaluated from the
dM dK
subproblem. If and are analytically evaluated, it is the analytic method.
dbi dbi
When they are evaluated by the finite difference method, it is the semianalytic
method.
When sensitivity analysis is performed by using the direct integration method,
the number of differential equations to be solved is the same as that of the design
variables. Thus, the method may not be efficient for problems with many design
variables.
The Mode Superposition Method

Equation 5.14 is similar to Equation 5.1c. Thus, a method to solve Equation 5.1c
can be utilized to solve Equation 5.14. Here the mode superposition method is
utilized. The solution of Equation 5.1c is approximated as follows:
z (t ) ĭq(t ) (5.19)
where ĭ is the modal matrix and q is the generalized coordinate vector. Equation
5.19 is differentiated by the ith design variable as follows:
dz dĭ dq
qĭ (5.20a)
dbi dbi dbi
dz dĭ dq
q ĭ (5.20b)
dbi dbi dbi
dz dĭ
dq
ĭ
q (5.20c)
dbi dbi dbi
dĭ
in Equation 5.20 is the sensitivity of the modal matrix with respect to the ith
dbi
design variable and it is evaluated from Equation 4.63. q, q and q in Equation
5.20 are obtained from the following equation:
ĭT Mĭq
ĭT Kĭĭ ĭT f (5.21)
dq
Equation 5.21 is derived by substituting Equation 5.19 into Equation 5.1c. ,
dbi
dq
dq
and in Equation 5.20 can be obtained from the following differential
dbi dbi
equation:

dq dq wf
ĭ T Mĭ ĭ T Kĭ ĭT B 2 q
B1q (5.22a)
dbi dbi wbi
ª wM dĭ º
B1 «ĭ T ĭ ĭT M » (5.22b)
¬ wbi dbi ¼
ª wK dĭ º
B2 «ĭ T ĭ ĭT K » (5.22c)
¬ wbi dbi ¼
Equation 5.22 is made by substituting Equation 5.20 into Equation 5.14. Generally,
the size of Equation 5.22 is very small compared to Equation 5.14. That is, the
dimension of q is a lot smaller than that of z. Therefore, the CPU time for
Equation 5.22 is less than that for Equation 5.14 when a numerical method is used
to solve them.
Generally, the sensitivity of eigenvectors with respect to design variables is
quite small. Therefore, it can be assumed to be zero as follows:
dĭ
#0 (5.23)
dbi
Substituting Equation 5.23 into Equations 5.20 and 5.22 yields
dz dq
ĭ (5.24a)
dbi dbi
dz dq
ĭ (5.24b)
dbi dbi
dz
dq
ĭ (5.24c)
dbi dbi

dq dq wf wM wK
ĭT Mĭ ĭ T Kĭ ĭT ĭT ĭ T
ĭq ĭq (5.25)
dbi dbi wbi wbi wbi
Some commercial structural optimization software systems use the above method
(Moore 1994).
Example 5.2 [Sensitivity Analysis: Direct Differentiation Method]

Using the direct differentiation method, evaluate the sensitivity dz/db of the
spring–mass system with one degree-of-freedom as illustrated in Figure 5.8. The
current design is b = 1 and the initial conditions are z(0) = 0, z (0) 0, dz/db(0) =
0 and dz / db(0) 0 .
z
k = 4b
f(t) = sin t
m b2
Figure 5.8. Spring–mass system with one DOF
Solution
The governing equation of the system is
b 2 z 4bz sin t (5.26)
The homogeneous solution has the following form:
zh A sin 2t B cos 2t (5.27)
The particular solution of Equation 5.26 is
1
zp sin t (5.28)
3
Therefore, the solution is
1
z zh z p A sin 2t B cos 2t sin t (5.29)
3
From the initial conditions, the solution is
1 1
z sin 2t sin t (5.30)
6 3
If we use Equation 5.14, the sensitivity dz/db should satisfy the following
differential equation:
d 2 dz dz
2 db
4 0 2z 4 z (5.31)
dt db
If we set y { dz / db , then Equation 5.31 yields
y 4 y 2z 4 z (5.32)

0
dz/db
-0.1
-0.2
-0.3
-0.4
-0.5
-0.6 t
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Figure 5.9. Sensitivity of the one DOF spring-mass system
The homogeneous solution of Equation 5.32 is
yh A sin 2t B cos 2t (5.33)
The particular solution of Equation 5.32 is
yp Jt sin 2t Kt cos 2t L sin t M cos t (5.34)
Then the solution is
1 2
y A sin 2t B cos 2t t cos 2t sin t (5.35)
6 9
Using the initial conditions, the solution is
dz 1 1 2
(t ) sin 2t t cos 2t sin t (5.36)
db 36 6 9
Equation 5.36 represents the sensitivity. It is graphically illustrated in Figure

5.9. As time increases, the absolute value of the sensitivity becomes larger. The
reason is that some terms are functions of time. In Equation 5.34
Jt sin 2t Kt cos 2t are such terms.
5.4.2 Adjoint Variable Method
In the adjoint variable method, an adjoint variable is defined in the process of

sensitivity analysis. There are various methods for this. In this section, an
augmented response function is defined and an adjoint variable is employed for the
augmented response function.
In a structure with dynamic loads, suppose we have the following constraint:
T
g ³ h(z, z , z, b, t )dt (5.37)
0
Using an arbitrary adjoint variable vector Ȝ , an augmented response function A of

Equation 5.37 is defined as follows:
T T
T
A ³ h(z, z , z, b, t )dt ³ Ȝ (Mz Kz f )dt (5.38)
0 0
where
Mz Kz f 0 (5.39)
Thus, although we have Ȝ in Equation 5.38, Equation 5.37 is equivalent to

Equation 5.38. Since Equation 5.39 is always satisfied, Equation 5.40 is also
always satisfied,
dg dA
(5.40)
db db
Therefore, the sensitivity of the constraint is the same as the sensitivity of the
augmented response function.
Differentiation of Equation 5.38 with respect to the ith design variable is
dA T wh T § wM wK wf ·
³ dt ³ Ȝ T ¨¨ z z ¸¸dt
dbi 0 wbi 0 w
© ib wbi wbi ¹
T § wh dz wh dz wh dz · T § dz dz ·
³ ¨¨ ¸¸dt ³ Ȝ T ¨¨ M K ¸dt (5.41)
0 © wz dbi wz dbi wz dbi ¹ 0 © dbi dbi ¸¹

X
An arbitrary adjoint variable vector Ȝ is determined so that X in Equation 5.41

satisfies the following relationship:
T § wh dz wh dz wh dz · T § dz dz ·

X ³ ¨¨ ¸¸dt ³ Ȝ T ¨¨ M K ¸¸dt 0 (5.42)
0 © wz dbi wz dbi wz dbi ¹ 0 © dbi dbi ¹
Equation 5.42 is called an adjoint variable equation.

Ȝ is obtained from Equation 5.42. Using the integration by parts, Equation
5.42 yields
T § wh d wh d 2 wh T · dz
³ ¨¨ 2 Ȝ M Ȝ T K ¸¸ dt
0 © wz dt wz dt wz ¹ dbi
T T
§ wh · dz § wh d wh · dz
¨ ȜT M ¸ ¨ Ȝ T M ¸ 0 (5.43)
© wz ¹ dbi 0

© wz dt wz ¹ dbi 0
In Equation 5.43, the following conditions should be satisfied:
wh d wh d 2 wh T
Ȝ M ȜT K 0 (5.44a)
wz dt wz dt 2 wz
T
§ T wh · dz
¨Ȝ M ¸ 0 (5.44b)
© wz ¹ dbi 0
T
§ wh T d wh · dz
¨ Ȝ M ¸ 0 (5.44c)
© wz dt wz ¹ dbi 0
Suppose we have the initial conditions as follows:
dz dz
(0) 0, (0) 0 (5.45)
dbi dbi
Then Equation 5.44 yields
T
§ wh d wh d 2 wh ·
MȜ KȜ ¨ ¸
¨ wz dt wz dt 2 wz ¸ (5.46a)
© ¹
wh wh d wh
Ȝ T (T )M (T ) 0 , (T ) Ȝ T (T )M (T ) 0 (5.46b)
wz wz dt wz
where M and K are assumed to be symmetric. Equation 5.46 is a second-order

differential equation of Ȝ . It is a terminal value problem where the terminal
conditions are given. Suppose Ȝ * is the solution of Equation 5.46. Substituting
Ȝ * into Equation 5.41 makes X in Equation 5.41 disappear and Equation 5.41 is
simplified as
dA T wh T § wM wK wf ·
³ dt ³ Ȝ *T ¨¨ z z ¸¸dt (5.47)
dbi 0 wbi 0 w
© ib wbi wbi ¹
Using Equations 5.40, Equation 5.47 is

dg T wh T § wM wK wf ·
³ dt ³ Ȝ *T ¨¨ z z ¸¸dt (5.48)
dbi 0 wbi 0 © wbi wbi wbi ¹
Equation 5.48 is the sensitivity of the constraints, which uses the adjoint variable
method.
The adjoint variable method can be applied to the constraint of the integral
form as shown in Equation 5.37. The constraint in Equation 5.5, which is an
equivalent functional, has an integral form. Therefore, the adjoint variable method
can be utilized in calculating sensitivity for the constraints of Equation 5.5 (Feng et
al. 1977). Moreover, the adjoint variable method is more efficient than the direct
differentiation method for problems with many design variables and a small
number of active constraints.
However, the adjoint variable method is not appropriate for constraints with a
nonintegral form. A representative case of the nonintegral form is Equation 5.4d
where the constraints are defined at particular times. The Dirac delta function can
be adopted in order to use the adjoint variable method for such constraints. A
constraint defined at a particular t i is
g h(b, z, z , z, t ) t (5.49)

i
Equation 5.49 can be modified to
T
g ³ h(b, z, z , z, t )G (t t i )dt (5.50)
0
Since Equation 5.50 has an integral form, the adjoint variable method can be used.
If many constraints such as those in Equation 5.49 are considered at many time
grid points, the above method can be inefficient. If we have q constraints such as
in Equation 5.50, we will have q augmented response functions. Then we have to
solve the differential equation in Equation 5.50 q times. If q is large, the adjoint
variable method does not have any advantages compared to the direct
differentiation method. Therefore, the adjoint variable method is not usually used
for constraints such as those in Equation 5.49.
Example 5.3 [The Adjoint Variable Method for the Transformation

Method]
Using the adjoint variable method, obtain the sensitivity of the augmented
Lagrange function in Equation 5.8.
Solution
In Equation 5.48
1Tª m 2º
h ³ « ¦ r j {max(0, g j (b, t ) T j (t ))} »dt (5.51)
2 0¬j 1 ¼
Thus, the sensitivity of Equation 5.8 with respect to the ith design variable is
d< wf T wh T § wM wK wf ·
³ dt ³ Ȝ T ¨¨ z z ¸¸dt (5.52)
dbi wbi 0 wbi 0 w
© ib wbi wbi ¹
Equation 5.52 consists of the sensitivity of the objective function f(b) and that of
the penalty term. The sensitivity of the objective function can be easily obtained.
We need an adjoint variable vector Ȝ for the sensitivity of the penalty term. The
adjoint variable vector is obtained from the following adjoint variable equation:
ª m wg j d m wg j ½
MȜ KȜ « ¦ r j max(0, g j T j ) ® ¦ r j max(0, g j T j ) ¾
¬« j 1 wz dt ¯ j 1 wz ¿
T
d2 m wg j ½º
2 ®
¦ r j max(0, g j T j ) ¾» (5.53a)
dt ¯i 1 wz ¿¼»
m wgi
Ȝ T (T )M ¦ r j max(0, g j (b, z, z , z, T ) T j (T )) (T ) 0 (5.53b)
j 1 wz
m wg j
Ȝ T (T )M ¦ r j max(0, g j (b, z, z , z, T ) T j (T )) (T )
j 1 wz
d m wg j
¦ r j max(0, g j (b, z, z , z, T ) T j (T )) (T ) 0 (5.53c)
dt j 1 wz
If we substitute Ȝ of Equation 5.53 into Equation 5.52, the sensitivity of the

augmented Lagrange function is obtained.
As shown in Equation 5.53, we need to solve the differential equations for the
equilibrium equation and the adjoint variable equation. It is fairly efficient to
combine the augmented Lagrangian method and the adjoint variable method.
Therefore, many research articles conclude that the transformation method is
efficient for optimization in the time domain.
5.4.3 Time Finite Element Method
The time finite element method is a method to solve the differential equation of
Equation 5.1 by transforming it to an algebraic equation. When the time finite
element method is utilized, sensitivity is evaluated by direct differentiation of the
algebraic equation. Suppose Equation 5.1c has one degree of freedom. The
solution of Equation 5.1c is assumed to be as follows:
N
z (t ) ¦ q j I j (t ) (5.54)
j 1
where q j is a coefficient to be obtained and I j (t ) is a known basis function. In

the general finite element method, q j is time variant and I j is time invariant.
However, the opposite assumption is made here.
If we integrate Equation 5.55 with Equation 5.54, then the algebraic equation
in 5.56 is obtained,
T
³Q ( Mz Kz f )dt 0 (5.55)
0
ª B M ȥ (T )º ªq j º ª a* º
«ĳ(T ) T « » (5.56a)
¬ 0 »¼ «¬ P »¼ «¬ z (T )»¼
where
T
³ ( K\ iI j M\ iI j )dt , ȥ
T
B [\ 1 \ 2 \ n ] (5.56b)
0
T
a* M\ i (0) z (0) ³ f\ i dt , ĳ T [I1 I 2 I n ] (5.56c)
0
P z (T ) (5.56d)
\ i is an admissible function and Q (t ) Gq i\ i (t ), 1 d i d N . P is an arbitrary

Lagrange multiplier. A detailed derivation is given in a reference (Kapania and
Park 1996). If the solution of Equation 5.56 is substituted into Equation 5.54, the
solution of Equation 5.1c is obtained.
If we differentiate Equation 5.56, the following algebraic equation is obtained:
ª wq j º ª wa * º
ª B M ȥ (T )º « wb » ª wB wM º ªq º « »
ȥ (T )» j wb
«ĳ(T ) T « » « wb wb « P » « wz (T ) » (5.57)
¬ 0 »¼ « wP » « 0 0 »¬ ¼ « »
¬ ¼
¬« wb ¼» ¬« wb ¼»
where the elements of the right hand side can be easily obtained by partial
differentiation. This is well explained in a reference (Kapania and Park 1996).
Sensitivity information is evaluated by solving the algebraic equation in

Equation 5.57. The time finite element method is attractive in that the solution of
the differential equation is obtained by solving the algebraic equation once.
However, the conditions for the transformation are not precisely defined and
general application to systems with multidegrees of freedom is rare. Thus, more
research into this method is required.
5.5 Approximation
Transient response analysis is required for optimization in the time domain. It is
well known that the analysis is costly. Therefore, many researchers have tried to
approximate the dynamic response. The approximation is classified into global
approximation and local approximation. The two methods are briefly explained
below.
5.5.1 Global Approximation
The representative method for global approximation is the response surface

method (RSM) (Myers and Montgomery 1995). RSM generates approximated
response surfaces from candidate values. The generated surfaces are usually
quadratic functions. The method is explained in Chapter 6. When RSM is used
for global approximation of dynamic responses, a number of dynamic analyses are
required. Then it may not be of advantage to use the approximation method.
Therefore, RSM is utilized for global approximation when the range for
approximation is not large.
RSM has been used in the area of crashworthiness (Hong and Park 2003,
Kurtaran et al. 2001, 2002, Marklund and Nilsson 2001, Craig et al. 2003, Bennett
and Park 1995). The functions involved in crashworthiness are usually highly
nonlinear and quite noisy. For a gradient-based optimization method, these
functions should be approximated by some smooth functions. RSM can be a good
candidate for the approximation. The worst case method is generally adopted for
time variant constraints.
The following formulation is typical to minimize the head injury while some
constraints are satisfied:
Find design variables (5.58a)

§§ 2.5 ·
¨ ¨ 1 t2 · ¸
to minimize sup¨ ³ a (t )dt ¸ (t 2 t1 ) ¸ (5.58b)
¨
t1 ,t2 ¨ © t 2 t1 t1
¸ ¸
© ¹ ¹
subject to maximum intrusion < 400 mm (5.58c)
where a(t ) is the head acceleration. The formulation is explained in detail in

Chapter 9. The objective function is to minimize a maximum property. The
Start
Generation of response surfaces for

the objective function and
constraints
Optimization with approximated

functions
End
Figure 5.10. Flow chart of optimization using RSM in the time domain
functions in Equations 5.58b and 5.58c are approximated by RSM and

optimization is applied to the approximated functions. The worst case (maximum)
is obtained at each candidate point and the functions are approximated by the
worst cases. Some examples are introduced in Chapter 9. Figure 5.10 shows the
process.
The candidate points should be selected by legitimate rules (e.g. D-optimality
or orthogonal arrays) and analysis is performed for the candidate points. If the
number of analyses is large, RSM may not be useful. Therefore, an effort is made
to reduce the number of analyses. In a typical method, the response surfaces are
roughly made at the beginning of the process and the detailed response surfaces
are defined as the design proceeds (Hong and Park 2003, Kurtaran et al. 2002).
Section 9.8 shows an example.
5.5.2 Local Approximation
Modern structural optimization approximates the optimization problem and the

approximated problem is solved in a cyclic manner. In the approximation, the
objective function and constraints are approximated to explicit functions of the
design variables. Optimization in the time domain utilizes this concept. The basic
strategy is that the sensitivity is precisely evaluated and the approximation is
performed based on the sensitivity. Since the sensitivity is generally obtained
through first-order differentiation, the approximation is valid within some move
limit, which should be appropriately defined.
In the direct method, the following approximated problem is solved in the kth
cycle (Etman and Van 1998):
Find b (5.59a)
~
to minimize f ( k ) (b) (5.59b)
subject to g~ (jk ) (b, t ) g~ (jk ) (~

r (b, t ), b, t ) d 0, j 1, , m (5.59c)
s L d b b ( k 1) d s U (5.59d)
where ~ means approximation and s U and s L are the upper and the lower bounds
of the move limit, respectively. The approximated ~r is a vector of the explicit
functions of intermediate variables (N) or design variables. The intermediate
variables are used as follows:
~ wr
r ( N, t ) r0 (N, t ) 'N (5.60)
wN
N N(b) (5.61)
g~ (jk ) of Equation 5.59c is an explicit function of approximated responses and

design variables. In many cases, the responses can be taken as the constraints
directly.
In the transformation method, the augmented Lagrange function can be
approximated by various methods. The proposed augmented Lagrange function is
(Kim and Choi 1997)
ª 2
T m
« § P j (t ) · 1 ° § P j (t ) ·½° º»
< f (b) ³ ¦ P j (t ) max ¨ ¸
, g j r j ®max ¨ ¸
, g j ¾ dt
0j 1
« ¨ rj ¸ 2 °̄ ¨ rj ¸° »
¬« © ¹ © ¹¿ ¼»
(5.62)
where P (t ) is a Lagrange multiplier and r j is a penalty parameter. At the kth

iteration, D , which makes the augmented Lagrange function in Equation 5.62
minimum, is determined from the line search and this process is expensive.
Therefore, the line search does not use Equation 5.62 directly. Instead, the
following approximated function is minimized:
ª 2
~ ~ T m
« § P j (t ) ~ · 1 ° § P j (t ) ~ ·½° º»
< (D ) f (D ) ³ ¦ P j (t ) max¨ , g j ¸ r j ®max¨ , g j ¸¾ dt
0j 1
« ¨ rj ¸ 2 °̄ ¨ rj ¸° »
¬« © ¹ © ¹¿ ¼»
(5.63a)
where
~
f (D ) f (b ( k ) ) T f (b ( k ) )(Ds ( k ) ) (5.63b)
g~ j (D , t ) g j (b ( k ) , t ) T g j (b ( k ) , t )(Ds ( k ) ) (5.63c)
Vector s ( k ) is the direction vector to minimize Equation 5.62.

The basic idea of Equation 5.63 is that the sensitivity of the augmented
Lagrange function is exactly evaluated and the line search is performed with an
approximated function. Since the approximation is carried out by first-order
differentiation, the move limit should be appropriately defined. Intermediate
variables can be adopted for the approximation in Equation 5.63b or Equation
5.63c.
5.6 A Method Using Equivalent Static Loads

The methods introduced so far directly handle dynamic loads for dynamic response
optimization. These methods should solve differential equations to evaluate
responses and sensitivity information. When a problem has many design variables
and constraints, a number of differential equations should be solved. In structural
optimization, the number of stress constraints is fairly large because they are
defined on all the elements of the finite element method. Therefore, the
optimization process is very costly
In static response optimization, algebraic equations are solved for responses
and sensitivity information. Thus, it costs less as compared to dynamic response
optimization. Moreover, many validated commercial software systems are
available. If static response optimization can be used for dynamic response
optimization, dynamic systems will be easily optimized. In this section, a method
using equivalent static loads is introduced; this is a method for dynamic response
optimization. Dynamic loads are transformed to equivalent static loads and a
method for static response optimization is cyclicly employed.
5.6.1 Equivalent Static Loads
From the equilibrium equation of a dynamic system, an equivalent static load

vector f eq is defined as follows:
f eq { Kz f Mz (5.64)
When a dynamic load is applied to a structure, the equivalent static load is defined
as the static load that makes the same displacement field as a dynamic load at an
arbitrary time. After z is calculated from Equation 5.1c, f eq is evaluated by
multiplying the stiffness matrix and z as shown in Equation 5.64. The equivalent
static loads are calculated at all the time grid points.
In engineering practice, dynamic factors are usually used to transform dynamic
loads into static loads. This may generate a dangerous design. Figure 5.11 shows
the behavior of a structure under a dynamic load. If a dynamic load is imposed as
(a) (b) (c)
Figure 5.11. Behavior of a structure under a dynamic load
(a) (b)
Figure 5.12. A dynamic load (a) and the corresponding equivalent static loads (b)
illustrated in Figure 5.11a, the structure vibrates and has the modes shown in
Figures 5.11b and c. The design with a dynamic factor only accounts for the mode
in Figure 5.11b. However, it is obvious that the mode in Figure 5.11c should be
considered. Since the equivalent static loads are calculated at all the time grid
points, all the modes of the behavior are considered.
It is noted that although a dynamic load is imposed on a node as illustrated in
Figure 5.12a, the equivalent loads are generated for all nodes as shown in Figure
5.12b. The reason is that the equivalent static loads include the external force as
well as the inertia forces at all nodes. In Equation 5.64, it is noted that the
elements of the inertia force are not usually zeros and the corresponding elements
of the equivalent static load f eq are not zeros.
Equation 5.64 defines the equivalent static loads with respect to displacements.
In a similar way, the equivalent inertia load vector with respect to accelerations is
defined as follows:
p eq { Mz f Kz (5.65)
The equivalent inertia load is defined as the loads that make the same acceleration
field as a dynamic load at an arbitrary time. In real implementation, z is
calculated from Equation 5.1c and p eq is evaluated by multiplying the mass
matrix and z as shown in Equation 5.65. The equivalent inertia loads are utilized
when a constraint for acceleration is included in the optimization problem.
5.6.2 Optimization in the Time Domain Using Equivalent Static

Loads
A structure under dynamic loads can be optimized by using the equivalent static
loads. When an optimization problem is given as Equation 5.1, the steps of the
algorithm using equivalent static loads are as follows:
Step 1. Set p = 0 and b p b0.
Step 2. In Equation 5.1c, transient (dynamic) analysis is performed with b =
bp .
Step 3. At all the time grid points, the equivalent static loads and the
equivalent inertia loads are calculated as shown in Equations 5.66 and 5.67.
Since the number of time grid points is q, q equivalent static load sets and
q equivalent inertia load sets are generated,
u
f eq K (b p )z u , u 1, , q (5.66)
p ueq M (b p )zu , u 1, , q (5.67)
Step 4. The following static optimization problem is solved:
Find b p 1 (5.68a)
to minimize f (b p 1 ) (5.68b)
u
subject to K (b p 1 )y u f eq , u 1, , q (5.68c)
M (b p 1 )a u p ueq , u 1, , q (5.68d)
g ju (b p 1 , z u , a u ) d 0, u 1, , q, j 1, , m (5.68e)
Step 5. The convergence criterion is checked. If
|| b p 1 b p ||
dH (5.69)
|| b p ||
the optimization process terminates. H is a positive number that is

sufficiently small. Otherwise, go to Step 6.
Step 6. Set p = p + 1 and go to Step 2.
The constraints on velocity are not considered because structural optimization
usually does not have such constraints.
The procedure from Step 2 to Step 6 forms a cycle. The equivalent static and
inertia loads are constant during the optimization process in Equation 5.68. The
equivalent static and inertia loads are updated by the optimum solution of Equation
5.68.
If the design variables are not changed in the process of Equation 5.68, then the
equivalent static and inertia loads are not changed. Therefore, when the design
variables are not changed, the optimization process terminates and the design
variables are considered as the optimum. It has been proved that the KKT
necessary condition of Equation 5.68 in the last cycle is the same as that of
Equation 5.1. The proof is shown in a reference (Park and Kang 2003).
In Step 3, the equivalent static and inertia loads are obtained at q time grid
points. It is noted that the total number of time grid points is q. They are treated
as q multiple loading conditions in the static response optimization process as
shown in Equations 5.68c and 5.68d. It is well known that the multiple loading
conditions are easily handled with low cost in structural optimization because only
the worst cases are considered in the optimization process.
The advantage of the proposed method is that the well developed static
response optimization is exploited. Modern structural optimization software
systems can easily handle large-scale structures. Therefore, the method using
equivalent static loads can easily optimize large structures in the time domain
(Choi and Park 2002, Kang et al. 2003).
Example 5.4 [Dynamic Factor and Equivalent Static Loads]

A connecting rod is shown in Figure 5.13a and a dynamic load is shown in Figure
5.13b. The shape of the rod is to be found to minimize the weight, while the
stresses in the rod are less than 250 MPa. Optimize the structure with a dynamic
factor and the equivalent loads and compare the results.
Solution
Optimization is performed with a static load which uses 2 for the dynamic factor.
Twice the dynamic load (23 kN) is imposed in the direction of the dynamic load in
Figure 5.13a. The optimum shape is shown in Figure 5.14.
Transient analysis is performed with the optimum shape in Figure 5.14. The
maximum stress occurs at point A in Figure 5.14 and the magnitude is about 300
MPa. Therefore, the optimum does not satisfy the constraints. If the dynamic
factor is larger than 2, the constraints can be satisfied; however, the design can be
over designed.
f(t)
11.5 kN
f(t)
0.01 s
(a) (b)
Figure 5.13. A connecting rod and a dynamic load profile

Figure 5.14. The optimum shape using the dynamic factor
Figure 5.15. The optimum shape using equivalent static loads
Optimization is carried out with equivalent static loads. The results are shown
in Figure 5.15. When transient analysis is performed, the maximum stress occurs
at point A of Figure 5.15 and the magnitude is about 250 MPa. It is noted that the
constraints are satisfied.
Example 5.5 [Optimization of an 18 Member Truss Structure: Stress

and Displacement Constraints]
The 18 member truss structure and the dynamic load of a sine curve are illustrated
in Figure 5.16. The dynamic load acts at the upper nodes as shown in Figure 5.16.
Young’s modulus and the density are 69 GPa and 2765 kg/m 3 , respectively.
Design variables are the cross sectional areas and the coordinates of nodes 3, 5, 7
and 9. Minimize the mass by using the equivalent static loads. The stresses should
be less than 138 MPa and all the nodal displacements should be less than 20.3 cm.
Solution
Since there are many cross sections, the following design variable linking is used
5×6.35 m
y
f(t) f(t)
(10) 16 (8)12 (6) 8 (4) 4 (2) 1 (1) 59.3 kN
17 15 13 11 9 7 5 3 2 6.35 m
18 14 10 6
2s
(11) (9) (7) (5) (3) x
Figure 5.16. 18 member truss structure and a dynamic load profile

to reduce the number of design variables.

Cross sections of the upper members: A1 = A4 = A8 = A12 = A16
Cross sections of the lower members: A2 = A6 = A10 = A14 = A18
Cross sections of the vertical members: A3 = A7 = A11 = A15
Cross sections of the diagonal members: A5 = A9 = A13 = A17
Thus, we have four design variables. The number of time grid points is 120 from
1.5 seconds to 4.0 seconds. 120 sets of equivalent static loads are evaluated. The
process terminates in five cycles. The optimum solution is presented in Figure
5.17 and Table 5.2.
The tip displacement is active at the optimum. Six transient analyses are
Table 5.2. Optimum results of the 18 member truss structure
Cycle
Design Initial value 1 2 3 4 5
variable
Upper members
64.52 113.34 129.03 129.03 129.03 129.03
(cm2)
Lower members
139.68 121.22 149.16 152.60 168.74 143.08
(cm2)
Vertical
80.65 29.21 40.22 34.91 27.72 44.35
members (cm2)
Diagonal
45.61 53.11 64.51 64.01 62.97 64.52
members (cm2)
x3 (cm) 0.00 í508.00 í494.03 í495.30 í450.85 í500.38
y3 (cm) 0.00 204.47 185.42 187.96 250.19 138.43
x5 (cm) 0.00 í374.65 í294.64 í285.75 í232.41 í303.53
y5 (cm) 0.00 101.60 74.93 102.87 175.26 53.34
x7 (cm) 0.00 í222.25 í218.44 í210.82 8.89 í254.00
y7 (cm) 0.00 24.13 2.54 20.32 85.09 í12.70
x9 (cm) 0.00 í194.31 í220.98 í189.23 í106.68 í226.06
y9 (cm) 0.00 í30.48 í22.86 í33.02 í22.86 í27.94
Mass (kg) 2916.78 2688.72 3260.07 3241.71 3313.62 3260.90
Tip
displacement í27.4 í24.7 í20.3 í20.4 í20.3 í20.3
(cm)
í156.5 í133.1 í111.7 í112.7 í113.2 í106.5
Stress (MPa)
(16) (15) (15) (15) (15) (8)
*The number of the stress row indicates the member where the maximum stress occurs
y
x
Figure 5.17. Optimum shape of the 18 member truss structure
performed to obtain the optimum. The approximation method in Chapter 4 is

employed for static response optimization. The analytic method is used for the
sensitivity of the cross sections and the semianalytic method is utilized for the
sensitivity of coordinates (Park et al. 2005).
Example 5.6 [Optimization of a Cantilever Beam: Acceleration

Constraint]
At the tip of the cantilever beam as illustrated in Figure 5.18, the following vertical
dynamic load acts:
100t N, 0.0 d t d 0.005 s

f (t ) ® (5.70)
¯ 0 N, 0.005 t d 0.1 s
Minimize the volume of the beam by using the equivalent inertia loads while the
tip acceleration is less than 40m/s 2 . The acceleration from 0.02 seconds to 0.1
seconds is considered and the number of time grid points is less than 4000. Since
the acceleration is drastically changed in the time domain, there should be many
time grid points. Design variables are the width (b) and the height (h) of the cross
section. The initial values are 0.01 m and 0.02 m, respectively. Young’s modulus
and the density are 300 MPa and 1700 kg/m 3 , respectively.
f(t)
b
1m
Figure 5.18. A cantilever beam under a dynamic load

Table 5.3. Optimization results of the cantilever beam
Initial Optimum
Width (m) 0.01 0.01195
Height (m) 0.02 0.02300
Volume (m3) 0.2 0.27483
No. of cycles 6
60
40
20
Acceleration (m/s2)
–20
–40
Time (s)
–60
0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Figure 5.19. The tip acceleration of the initial design
Solution
The volume is multiplied by 1000 and used as the objective function as follows:
f (b, h) 1000 u b u h u 1.0 (5.71)
The first natural frequency of the initial design is 1.356 Hz. The results of the
transient analysis for the initial design are illustrated in Figure 5.19. The
acceleration constraint is violated.
Table 5.3 shows the optimization results by using the equivalent inertia loads.
The tip acceleration is illustrated in Figure 5.20. It is noted that the acceleration
constraint is satisfied.
5.6.3 Optimization of Flexible Multibody Dynamic Systems
In flexible multibody dynamics, the elements of the system are considered as

flexible bodies while they are modelled as rigid bodies in general dynamics. The
analysis methods for flexible multibody dynamics can be found in the reference
(Shabana 1998). The finite element method is adopted to evaluate the deformation
of the flexible bodies. Although the boundary conditions between flexible links
60
40
Acceleration (m/s2)
20
–20
–40
Time (s)
–60
0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Figure 5.20. The tip acceleration at the optimum
and rigid joints are not yet clearly defined, flexible body dynamics is exploited for
systems with high speeds.
The loads in flexible multibody dynamics are external loads, reaction forces of
the joints, centrifugal forces of the elements, etc. They are dynamic and thus,
optimization of flexible multibody dynamic systems is attributed to dynamic
response optimization. The method using equivalent static loads is utilized for
optimization of flexible multibody dynamic systems (Kang et al. 2005).
Equivalent Static Loads for Flexible Multibody Dynamic Systems

The equivalent static loads are evaluated from the governing equation. The
governing equation for the flexible multibody dynamic systems is as follows:
Mz Kz fc fe fv (5.72a)
ªm RR m RT m Rf º
« »
M « mTT mTf » (5.72b)
« sym m ff »
¬ ¼
ª0 0 0 º
« »
K «0 0 0 » (5.72c)
«0 0 K ff »
¬ ¼
ªR º
« »
z «ș » (5.72d)
«q f »
¬ ¼
ª C TR º
« »
fc « CTT »Ȝ (5.72e)
«C T »
¬ qf ¼
ª(Q e ) R º
« »
fe « (Q e ) T » (5.72f)
«(Q e ) f »
¬ ¼
ª(Q v ) R º
« »
fv « (Q v ) T » (5.72g)
«(Q v ) f »
¬ ¼
The process to derive Equation 5.72 is explained in detail in a reference (Shabana

1998).
Equation 5.72b represents the mass matrix. The off-diagonal terms of the mass
matrix are relationships between the rigid body motion and the elastic deformation.
When the rigid body motion and the elastic deformation are considered separately,
the off-diagonal terms are zeros. Equation 5.72c is the stiffness matrix of the
flexible bodies when the elastic deformation is calculated. It can be expressed by
the stiffness matrix of the finite element method.
q f of Equation 5.72d indicates the nodal displacements of the finite element
model, R represents the coordinates of the origin in the floating body reference
frame and ș shows the angles of the floating body reference frame. Equation
5.72e is the reaction force vector of the joints. The vector in Equation 5.72f is for
the external loads. Equation 5.72g represents the Coriolis force and the centrifugal
forces. These forces are functions of time.
First, Equation 5.72 should be solved to obtain the equivalent static loads.
Suppose we have solved the transient analysis problem in Equation 5.72. Equation
5.72 is divided into three groups of equations as follows:
i mi și mi qi
miRRR CTRi Ȝ (Qie ) R (Qiv ) R (5.73a)
RT Rf f
i mi și mi qi
miRT R CTT i Ȝ (Qie )T (Qiv )T (5.73b)
TT Tf f
i mi și mi q
miRf R i i i
CTqi Ȝ (Qie ) f (Qiv ) f
fT ff f K ff q f f
(5.73c)
Equation 5.73c yields
i mi și mi q
Kiff qif miRf R i T i i
fT ff f Cqi Ȝ (Qe ) f (Qv ) f
f
(5.74)
Equation 5.74 has the displacement information of the flexible multibodies.

Therefore, the equivalent static load vector f eq is
f eq { K iff q if i m i ș i m i q
m iRf R i T i i
fT ff f C q i Ȝ (Q e ) f (Q v ) f
f
(5.75)
The first two terms on the right hand side in Equation 5.75 represent the influence
from the interrelation between the rigid body motion and the elastic deformation.
The third term is the inertia force from the deformation of the flexible bodies.
These three terms are developed by the flexible bodies and other terms are from
the rigid bodies.
Structural Optimization of Flexible Multibody Systems Using Equivalent

Static Loads
The method in Section 5.6.2 can be utilized directly. Equation 5.72 is solved as
transient analysis and Equation 5.75 is used to obtain the equivalent static loads. It
is noted that we need eigenvectors to solve Equation 5.72. Therefore, we need a
systematic interface of software systems for structural optimization, flexible
multibody dynamics and modal analysis.
The steps of the algorithm are as follows:
Step 1. Set p = 0 and b p b0 .
Step 2. Update the design variables and perform modal analysis to obtain the
eigenvectors.
Step 3. With b b p , solve Equation 5.72 by using a software system for
flexible multibody systems. The solution of the flexible bodies is
approximated by the eigenvectors of Step 2.
Step 4. Calculate equivalent static loads at q time grid points as shown in
Equation 5.76. As a result, we have q sets of equivalent static loads,
u
f eq K ff (b p )q uff , u 1, , q (5.76)
Step 5. Using the equivalent loads in Step 4, the static response optimization
problem in Equation 5.77 is to be solved. Structural optimization software
can be employed.
Find b p 1 (5.77a)
to minimize f (b p 1 ) (5.77b)
u
subject to f eq K ff (b p 1 )z uff , u 1, , q (5.77c)
g ju (b p 1 , z uff ) d 0, u 1, , q, j 1, , m (5.77d)
Step 6. The convergence criterion is checked. If
|| b p 1 b p ||
dH (5.78)
|| b p ||
terminate the process and the solution of Equation 5.77 is the optimum. H
is a sufficiently small number. Otherwise, go to Step 7.
Step 7. Set p = p + 1 and go to Step 2.
Example 5.7 [Optimization of Cross Sections of a Robot Arm]

The two links of a robot arm in Figure 5.21 have circular hollow cross sections.
The length of each link is 0.6 m and the design variables are the outer radii of the
links. The thickness of a cross section is one tenth of the outer radius. Young’s
modulus is 72 GPa and the density is 2700 kg/m3. A is a revolute joint and the
mass of A is 2 kg. The mass of the end effector E is 1 kg. The objective function
is the weight of the arm and the route of the end effector is (Oral and Ider 1997)
0.5 T 2St
'X E 'YE (t sin ), t [0, 0.66] (5.79)
T 2S T
where T is 0.5 seconds. The constraints are as follows:
75.0 d V i d 75.0 MPa , i = 1, 2, 3, 4 (5.80)
(Gy a2 Gy b2 ) 0.5 d 0.001 m (5.81)
Equation 5.80 represents stress constraints and Gy a and Gy b in Equation 5.81 are
vertical displacements at nodes a and b, respectively. The initial value of each
design variable is 50 mm. Optimize the system.
a
D12 D21
D22
D11
b
E
Figure 5.21. A robot arm with flexible bodies

Table 5.4. Optimization results of the robot arm using equivalent static loads (mm, N)
Design variable and the

Optimum result
objective function
D11 54.500
D12 38.525
D21 30.400
D22 22.700
Weight 13.229
Solution
The problem converges in six cycles and the optimum solution is shown in Table
5.4. The number of modes for transient analysis is three. The first two modes are
for bending and the third one is for the deformation in the axial direction. Since
the number of time grid points is 300, 300 sets of equivalent loads are utilized in
Equation 5.80 as multiple loading conditions. The active set strategy and the
approximation by intermediate variables are adopted in static response
optimization in Equation 5.77.
Example 5.8 [Optimization of Cross Sections of a Four Bar Linkage]

The four bar linkage in Figure 5.22 is modelled using finite element analysis.
Each link consists of six beam elements and has a circular solid cross section.
Young’s modulus is 68.95 GPa and the density is 2757 kg/m 3 . The lengths of the
links from number 1 to number 4 are 0.3048 m, 0.9144 m, 0.7620 m and 0.9144 m,
respectively. Link 1 rotates at an angular speed of 10S / s. The design variables
are the diameters of the links. The bending stress of the kth element of the ith link
is
2
d2
1 d1
3 d3
Figure 5.22. A flexible four bar linkage

Table 5.5. Optimization results of the four bar linkage (mm, kg)
Design variable and the

Optimum result
objective function
d1 33.5
d2 23.2
d3 16.9
Mass 2.28
4 S
V ik 3
M ik (5.82)
Ai 2
where Ai is the area of the ith link and M ik is the bending moment at the kth node
of the ith link. The objective function is the total mass of the system. Determine
the design variables while the bending stresses in Equation 5.82 are less than 27.58
MPa in a steady state. The initial value of each design variable is 356.8 mm.
Solution
Table 5.5 shows the optimization results. This problem has many local optimum
values (Etman and Van 1998). Therefore, the optimum in Table 5.5 is one of
many local optima. Transient analysis is performed using the first three modes.
The problem converges in 16 cycles. The number of time grid points is 500.
5.7 Structural Optimization in the Frequency Domain

Structural optimization in the frequency domain is explained here. It is classified
into optimization with the eigenvalue problem, optimization based on the random
loads and optimization of the harmonic responses. Optimization with the
eigenvalue problem is explained in Chapter 4. Analysis with random loads has
been studied extensively. However, research on optimization has not been widely
conducted. This section handles optimization of harmonic responses.
5.7.1 Optimization of Harmonic Responses

The harmonic response is the behavior of a structure when a harmonic force is
imposed it. The harmonic response is the solution of the following differential
equation:
Mz Kz p sin Zt (5.83a)

p [ p1 p n ]T (5.83b)
Suppose the solution of Equation 5.83 in the steady state is
z x sinZt (5.84)
Substituting Equation 5.84 into Equation 5.83 yields
(Z 2 M K )x p (5.85)
In harmonic analysis, Equation 5.84 is calculated by x obtained from Equation

5.85 when a harmonic force is imposed on a structure. The magnitude of the force
is p and the frequency is Z. In many cases, only the magnitude in Equation 5.84
is meaningful. Thus, attention is given to x in Equation 5.84. In reality, harmonic
analysis is performed for the range of Z. It is similar to transient analysis
performed for some range of time. Thus, some methods to handle time-dependent
constraints can be used for the frequency-dependent constraints.
Optimization of the harmonic responses is to obtain a solution from the
following formulation:
Find b R n (5.86a)
subject to (Z i2 M K )x i p i , i 1, , r (5.86c)
g j (b, x i , Z i ) d 0, i 1, , r , j 1, , m (5.86d)
bL d b d bU (5.86e)
where r is the number of frequencies of interest and the number of multiple

loadings. The equality constraints in Equation 5.86c are given for harmonic
analysis of an algebraic equation. It is noted that finite element analysis is
represented by equality constraints of an algebraic equation in Chapter 4.
Therefore, the various methods for static analysis in Chapter 4 can be used for
optimization of harmonic responses. Physical phenomena are dynamic in
optimization of harmonic responses. However, the techniques for static response
optimization can be used for optimization of harmonic responses.
Example 5.9 [Harmonic Response Analysis]

Perform harmonic response analysis for the 2 DOF spring–mass system in Figure
5.23. m = 1, k = 2, p = 100 and the range of Z is 1 d Z d 10.
Solution
The equation of motion is
Figure 5.23. Two DOF spring-mass system
ªm 0 º ª z1 º ª 2k k º ª z1 º ª pº
« 0 m» « z » « k « »
2k »¼ ¬ z 2 ¼ « 0 » sin Zt (5.87)
¬ ¼¬ 2 ¼ ¬ ¬ ¼
Suppose the solution of Equation 5.87 is
ª z1 º ª x1 º
«z » « x » sin Zt (5.88)
¬ 2¼ ¬ 2¼
§ ªm 0 º ª 2k k º · ª x1 º ª pº
¨Z 2 « »« ¸« » (5.89)
¨
© ¬ 0 m¼ ¬ k 2k »¼ ¸¹ ¬ x 2 ¼ «0»
¬ ¼
(2k mZ 2 ) p
x1 (5.90a)
§ k ·2
2 ¨§
·§ § 3k · 2 ·
m ¨ ¸ Z 2 ¸¨ ¨ ¸ Z 2 ¸
¨© m ¹ ¸¨ © m ¹ ¸
© ¹© ¹
kp
x2 (5.90b)
§ k· 2
·§ § 3k · 2 ·
2 ¨§
m ¨ ¸ Z 2 ¸¨ ¨ ¸ Z 2 ¸
¨© m ¹ ¸¨ © m ¹ ¸
© ¹© ¹
If we substitute the given data into Equation 5.90, the solution is
10
x1 (5.91a)
36 Z 2
20
x2 (5.91b)
(4 Z 2 )(36 Z 2 )
X
90
70 xx1
1
50 xx2
2
30
10
Z
Ṗ㰚㭒䕢㑮
-10 0 1 2 3 4 5 6 7 8 9 10
-30
-50
-70
-90
Figure 5.24. Harmonic response of the 2 DOF spring–mass system
The graph of Equation 5.91 with respect to the range of Z is illustrated in Figure
5.24.
5.7.2 Sensitivity Analysis
Sensitivity analysis of Equation 5.86 is required for optimization. In the direct

method, sensitivity information is utilized to generate a subproblem. Here,
sensitivity analysis for Equation 5.86c will be explained. Sensitivity analysis for
other equations is explained in Chapter 4. In Chapter 4, it is mentioned that the
equality constraints in Equation 5.86c are eliminated in the process to generate a
subproblem. The sensitivity analysis is described below.
Direct Differentiation
It is assumed that the external forces are not functions of design variables.
Differentiation of Equation 5.86c with respect to the ith design variable yields
dx § 2 dM dK ·
(Z 2 M K ) ¨¨ Z ¸x (5.92)
dbi © dbi dbi ¸¹
Sensitivity information dx / dbi is obtained by solving the algebraic equation in

Equation 5.92.
Mode Superposition Method

The solution of 5.86c can be approximated as follows:
x ĭq (5.93)
where ĭ is the modal matrix. It is noted that the dimension of q is less than that
of x. Differentiation of Equation 5.93 with respect to the ith design variable yields
dx dĭ dq
qĭ (5.94)
dbi dbi dbi
dĭ dq
where is obtained from the sensitivity analysis of eigenvectors. is
dbi dbi
calculated from the following equation:
dq dĭ § wM wK ·
(Z 2 ĭT Mĭ ĭT Kĭ) ĭT (Z 2 M K ) q ĭT ¨¨ Z 2 ¸ĭq
dbi dbi © wbi wbi ¸¹
(5.95)
Equation 5.95 is obtained by multiplying the results of substituting Equations 5.93

dĭ
and 5.94 into Equation 5.92 by ĭT . If is small enough to be negligible,
dbi
Equations 5.94 and 5.95 become
dx dq
#ĭ (5.96)
dbi dbi
dq § wM wK ·
(Z 2 ĭ T Mĭ ĭ T Kĭ) # ĭ T ¨¨ Z 2 ¸ĭq (5.97)
dbi © wbi wbi ¸¹
dq
Equation 5.96 represents sensitivity information and is calculated from
dbi
Equation 5.97.
Adjoint Variable Method

dg j
The adjoint variable method is employed to derive the sensitivity of the
dbi
dg j
function g j of Equation 5.86d. is derived as
dbi
dg j wg j wg j dx
(5.98)
dbi wbi wx dbi
Equation 5.92 is modified as
dx § wM wK ·
(Z 2 M K ) 1 ¨¨ Z 2 ¸x (5.99)
dbi © wbi wbi ¸¹
dg j wg j wg j § wM wK ·
(Z 2 M K ) 1 ¨¨ Z 2 ¸x (5.100)
dbi wbi wx © wbi wbi ¸¹
In Equation 5.100, an adjoint variable vector Ȝ is defined as follows:
wg j
ȜT (Z 2 M K ) 1 (5.101)
wx
The following adjoint variable equation is obtained by modifying Equation 5.101:
T
§ wg j ·
(Z 2 M K )Ȝ ¨
¨ wx
¸
¸ (5.102)
© ¹
where the mass matrix M and the stiffness matrix K are assumed to be symmetric.
Using Equation 5.101, Equation 5.100 is
dg j wg j § wM wK ·
Ȝ T ¨¨ Z 2 ¸x (5.103)
dbi wbi © wbi wbi ¸¹
The adjoint variable vector Ȝ is obtained from Equation 5.102 and substituted into
dg j
Equation 5.103. Then is obtained.
dbi
Example 5.10 [Sensitivity Analysis of Harmonic Responses]

A load P 40,000 u sin 20t N is imposed on the three bar truss of Example 4.1.
dx
Obtain by the direct differentiation, the mode superposition method and the
db1
adjoint variable method. The density is U 7830kg / m 3 .
Solution
The mass matrix and the stiffness matrix in the global coordinate system are as
follows:
ª b1 b3 b1 b º
« 2L 2L 3
2 L1 2 L3 »»
K E« 1 3 (5.104)
« b1 b3 b1 b b
2 3 »
« 2L 2L 2 L1 L 2 2 L3 »¼
¬ 1 3
ª b1 L1 b3 L3 b1 L1 b3 L3 º
« »
M U« 6 6 6 6 (5.105)
« b1 L1 b3 L3 b1 L1 b2 L 2 b3 L3 »»

¬« 6 6 6 3 6 ¼»
With the given data, Equations 5.104 and 5.105 become
ª2 0 º
K 2 u 10 4 u 207 u 10 9 « » (5.106)
¬0 2 2 ¼
2 u 7830 u 10 4 ª4 2 0 º
M « » (5.107)
3 «¬ 0 1 4 2 »¼
The external load vector is
ª 40000 u 1 / 2 º
p «40000 u 3 / 2» (5.108)
¬ ¼
Substituting Equations 5.106í5.108 into Equation 5.85, the magnitude of the

harmonic response x is obtained as follows:
x [0.00034160 0.00034659]T (5.109)
(1) Direct differentiation

Equations 5.104 and 5.105 are differentiated with respect to the design
variables as follows:
wK E ª1 1º 207 u 10 9 ª1 1º
«1 1» « » (5.110)
wb1 2 L1 ¬ ¼ 2 2 ¬1 1¼
wM UL1 ª1 1º 7830 u 2 ª1 1º
« » «1 1» (5.111)
wb1 6 ¬1 1¼ 6 ¬ ¼
Substituting Equations 5.106í5.111 into Equation 5.92 yields

§ 0 º 2 u 7830 u 10 4 ª4 2 0 º ·¸ dx
¨ 20 2 u 2 u 10 4 u 207 u 10 9 ª2
¨ « 0 2 2 » « »¸
© ¬ ¼ 3 ¬« 0 1 4 2 ¼» ¹ db1
§ 9 ·
= ¨ 20 2 u 7830 u 2 ª«1 1º» 207 u 10 ª«1 1º» ¸ ª«0.00034160º» (5.112)
¨ 6 ¬1 1 ¼ 2 2 ¬1 1 ¼ ¸¹ ¬0.00034659¼
©
dx
[ 0.86026 0.50392]T (5.113)
db1
(2) The mode superposition method

Assuming that the sensitivity of the eigenvectors is negligible, Equations
5.96 and 5.97 are adopted for sensitivity evaluation. The modal matrix is
as follows:
ĭ [v1 v2] (5.114a)
v1 [1 0]T (5.114b)
v2 [0 1]T (5.114c)
In this problem, the first eigenvector (horizontal direction) and the second
one (vertical direction) are equally important. Therefore, the modal matrix
for all the DOF’s should be used. However, only the first eigenvector is
utilized to show approximation. Therefore, x is approximated by q as
follows:
x v1q1 (5.115)
At the given design point, Equation 5.115 is substituted into Equation 5.85
as follows:
§ 0 º 2 u 7830 u 10 4 ª4 2 0 º ·¸ ª1º
¨ 20 2 u 2 u 10 4 u 207 u 10 9 ª2
¨ « » « » « » q1
© ¬0 2 2 ¼ 3 «¬ 0 1 4 2 »¼ ¸¹ ¬0¼
ª 40000 u1 / 2 º (5.116)
«40000 u 3 / 2»
¬ ¼
q1 calculated from Equation 5.116 is substituted into Equation 5.115.

Then
x [0.00034160 0]T (5.117)
According to Equation 5.96, the sensitivity is approximated as

dx dq
# v1 1 (5.118)
db1 db1
dq1
After using Equation 5.97, is calculated from the following equation:
db1
dq1 § wM wK ·
v1T (Z 2 M K ) v1 v1T ¨¨ Z 2 ¸ v1q1 (5.119)
db1 © wb1 wb1 ¸¹
If we use the given data for Equation 5.119, the following is obtained:
dq1
0.4270 (5.120)
db1
dx
# [ 0.4270 0]T (5.121)
db1
Because the problem does not use a sufficient number of eigenvectors,

Equation 5.121 is very different from Equation 5.113.
dx1 dx2
In this problem g j x j . Applying Equation 5.103 to and
db1 db1
yields
dx j § wM wK ·
Ȝ Tj ¨¨ Z 2 ¸x, j 1, 2 (5.122)
db1 © wb1 wb1 ¸¹
where the adjoint variable vector Ȝ j is obtained from Equation 5.102 as

follows:
§ 0 º 2 u 7830 u 10 4 ª4 2 0 º ·¸ T
¨ 20 2 u 2 u 10 4 u 207 u 10 9 ª2 « » ¸Ȝ1
¨ « »
© ¬0 2 2 ¼ 3 ¬« 0 1 4 2 »¼ ¹
ª1º (5.123)
«0 »
¬ ¼
§ 0 º 2 u 7830 u 10 4 ª4 2 0 º ·¸ T
¨ 20 2 u 2 u 10 4 u 207 u 10 9 ª2
¨ « 0 2 2 » « » ¸Ȝ 2
© ¬ ¼ 3 ¬« 0 1 4 2 ¼» ¹
ª0 º (5.124)
«1»
¬ ¼
From Equations 5.123 and 5.124, the following adjoint variable vectors are
calculated:
Ȝ 1T [1.7080 0] u 10 8 (5.125)
Ȝ T2 [0 1.0005] u 10 8 (5.126)
Substituting Equations 5.125 and 5.126 into Equation 5.122, the following
sensitivities are obtained:
dx1 / db1 0.86026 (5.127)
dx2 / db1 0.50392 (5.128)
5.7.3 Treatment of Constraints and Minimization of the

Maximum Amplitude
The constraints in optimization of the harmonic responses are dependent on the
frequencies. Thus the constraints are expressed as Equation 5.86d, which is
rewritten as follows:
g j (b, x i , Z i ) d 0, i 1, , r , j 1, , m (5.129)
As shown in Figure 5.25, a constraint in Equation 5.129 is changed to multiple

constraints by r multiple loading conditions. The number of Z i is the number of
multiple loadings. If we consider Z i as time, Equation 5.129 is similar to the
constraints that are dependent on time. Therefore, treatment of Equation 5.129 is
similar to the treatment of the time-dependent constraints. Here, algebraic
equations are solved for the responses and their sensitivities as opposed to
differential equations. Treatment of the time-dependent constraints was explained
earlier.
In many cases, the maximum amplitude is to be minimized. This problem can
be formulated as follows:
Find b (5.130a)
to minimize the maximum amplitude at a DOF (5.130b)

0 Frequency
Z1Z 2 Z 3 … … Zr
Figure 5.25. Frequency-dependent constraint
g j (b, x i , Z i ) d 0, i 1, , r , j 1, , m (5.130d)
bL d b d bU (5.130e)
The methods to solve Equation 5.130 are classified into two types. One method is
to use an artificial variable and the other is a method to minimize the area under
the amplitude versus the frequency curve.
A Method to Use an Artificial Variable

Using the technique in Equation 5.3, Equation 5.130 is modified by an artificial
variable E as follows:
Find b, E (5.131a)
to minimize E (5.131b)
g j (b, xi , Z i ) d 0, i 1, , r , j 1, , m (5.131d)
all amplitudes at a DOF in specified range of frequency d E

(5.131e)
bL d b d bU (5.131f)
In Equation 5.131, the artificial variable E is minimized and the constraint in

Equation 5.131e is added. This is illustrated in Figure 5.26.
Minimization of the Area

Another method is to solve the following formulation:
Find b, x (5.132a)
x Maximum amplitude
0 Frequency
Figure 5.26. Schematic view of the artificial variable
x Z2
³ xdZ
Z1
0 Frequency
Figure 5.27. Area made by the amplitude
Z2
to minimize ³ xdZ (5.132b)
Z1
g j (b, x i , Z i ) d 0, i 1, , r , j 1, , m (5.132d)
In Figure 5.27, the area under the amplitude versus the frequency curve as
illustrated. Equation 5.132 minimizes the area. Equation 5.132 is not the same as
Equation 5.130. However, when the area is reduced, the maximum amplitude
tends to be reduced. Thus, the formulation in Equation 5.132 is utilized to
minimize the maximum amplitude in an engineering sense.
Example 5.11 [Minimization of the Maximum Amplitude]

A cantilever is shown in Figure 5.28. The cross section is rectangular and the
length is 3 m. Minimize the maximum vertical amplitude at the tip. The volume
should be less than 15u 10 6 mm 3 and the frequency should have the range of
9 Hz d f d 57 Hz. Young’s modulus is E = 207 GPa and the density is U 8000
kg / m 3 . The width and height of the cross section are the same and the initial
values are 100 mm.
f (Z ) 100 sin (Zt )
h
1 2 3 4 5 b
Figure 5.28. A cantilever under a dynamic load of the sine function
Solution
(1) Method using an artificial variable

The frequency range of 9 Hz d f (Z / 2S ) d 57 Hz is discretized by 1 Hz.
The formulation in Equation 5.131 is utilized. The amplitude of the initial
design is illustrated in Figure 5.29. The maximum amplitude is 18 mm and
the volume is 30 u 10 6 mm 3 .
Using an artificial variable E , the optimization problem is formulated as
follows:
Find b, x, E (5.133a)
volume d 15 u 10 6 (5.133d)
all amplitudes at x5 y in a given range of frequency d E
(5.133e)
20 Amplitude
18
16
14
12
10
8
6
4
2
0
9 14 19 24 29 34 39 44 49 54 f (Hz)
Figure 5.29. Harmonic response of the initial design

Table 5.6. Optimization results: the method using an artificial variable
b1 (mm) b2 (mm) b3 (mm) b4 (mm) 3

Volume ( mm )
74.192 59.7001 65.563 81.074 14,955,030
1.6 Amplitude
1.4
1.2
1
0.8
0.6
0.4
0.2 f (Hz)
0
9 15 21 27 33 39 45 51 57
Figure 5.30. Harmonic response of the optimum results–the method with an artificial
variable
bL d b d bU (5.133f)
The optimization results are shown in Table 5.6. The harmonic response
from the optimum in Table 5.6 is illustrated in Figure 5.30. The maximum
amplitude (1.48 mm) is significantly reduced and the volume is also
reduced by half of the initial design.
(2) Minimization of the area
The maximum amplitude is indirectly reduced by minimizing the area
between the amplitude curve and the frequency axis. The following
optimization problem is solved:
Find b, x (5.134a)
( x5 y1 x5 y n ) n 1
to minimize 'f ¦ | x5 y i | 'f (5.134b)
2 i 2
volume d 15 u 10 6 (5.134d)
Table 5.7. Optimization results: minimization of the area
b1 (mm) b2 (mm) b3 (mm) b4 (mm) Area (mmHz) 3

Volume ( mm )
72.861 65.154 77.723 66.060 17.9989 14,968,960
3 Amplitude
2.5
1.5
0.5 f (Hz)
0
9 15 21 27 33 39 45 51 57
Figure 5.31. Harmonic response of the optimum results: minimization of the area
The area is calculated by the trapezoid rule for numerical integration. The
area of the initial design is 17.9062. The frequency domain is discretized
by 1 Hz. The optimization results are shown in Table 5.7 and the harmonic
response of the optimum is illustrated in Figure 5.31.
References
Afimiwala KA, Mayne RW (1974) Optimal Design of an Impact Absorber.
Journal of Engineering for Industry, Transactions of the ASME 96:124í130
Barthelemy JFM, Haftka RT (1993) Approximation Concepts for Optimum
Structural Design–A Review. Structural Optimization 5:129í144
Bennett JA, Park GJ (1995) Automotive Occupant Dynamics Optimization. Shock
and Vibration 2:471í479
Chahande AI, Arora JS (1993) Development of a Multiplier Method for Dynamic
Response Optimization Problems. Structural Optimization 6:69í78
Choi WS, Park GJ (2002) Structural Optimization Using Equivalent Static Loads
at All the Time Intervals. Computer Methods in Applied Mechanics and
Craig KJ, Stander N, Balasubramanyam S (2003) Worst-Case Design in Head
Impact Crashworthiness Optimization. International Journal for Numerical
Methods in Engineering 57:795í817
Etman LFP, Van CDH (1998) Design Optimization of Multibody Systems by
Sequential Approximation. Multibody System Dynamics 2:393í415
Feng TT, Arora JS, Haug JR (1977) Optimal Structural Design under Dynamic
Loads. International Journal for Numerical Methods in Engineering 11:39í52
Fox RL, Kapoor MP (1970) Structural Optimization in the Dynamic Regime: A
Computational Approach. AIAA Journal 8:1798í1804
Grandhi RV, Haftka RT, Watson LT (1986) Design-oriented Identification of
Critical Times in Transient Response. AIAA Journal 24:649í656
Greene WH, Haftka RT (1991) Computational Aspects of Sensitivity Calculations
in Linear Transient Structural Analysis. Structural Optimization 3:176í201
Gu YX, Chen BS, Zhang HW, Grandhi RV (2002) A Sensitivity Analysis Method
for Linear and Nonlinear Transient Heat Conduction with Precise Time
Integration. Structural and Multidisciplinary Optimization 24:23í37
Dordrecht
Hong UP, Park GJ (2003) Determination of Crash and Optimization of Crash
Components Using Response Surface Approximate Approximation. Journal of
Automobile Engineering 217:203í214
Kang BS, Choi WS, Park GJ (2001) Structural Optimization under Equivalent
Static Loads Transformed from Dynamic Loads Based on Displacement.
Computers and Structures 79:145í154
Kang BS, Park GJ, Arora JS (2005) Optimization of Flexible Multibody Dynamic
Systems Using the Equivalent Static Load. AIAA Journal 43:846í852
Kang BS, Park GJ, Arora JS (2006) A Review of Optimization of Structures
Subjected to Transient Loads. Structural and Multidisciplinary Optimization 31:
81í95
Kapania RK, Park S (1996) Nonlinear Transient Response and Its Sensitivity
Using Finite Elements in Time. Computational Mechanics 17:306í317
Kim MS, Choi DH (1997) Multibody Dynamic Response Optimization with ALM
and Approximation Line Search. Multibody System Dynamics 1:47í64
Kim MS, Choi DH (2000) A New Penalty Parameter Update Rule in the
Augmented Lagrange Multiplier Method for Dynamic Response Optimization.
KSME International Journal 14:1122í1130
Kurtaran H, Eskandarian A (2001) Design Optimization of Multibody Systems
under Impact Loading by Response Surface Methodology. Proceedings of the
Institution of Mechanical Engineers, Part K: Journal of Multibody Dynamics
215:173í185
Kurtaran H, Eskandarian A, Marzougui D, Bedewi NE (2002) Crashworthiness
Design Optimization Using Successive Response Surface Approximations.
Computational Mechanics 29:409í421
Logan DL (2002) A First Course in the Finite Element Method. Brooks/Cole,
Pacific Grove, CA
Marklund PO, Nilsson L (2001) Optimization of a Car Body Component Subjected
to Side Impact. Structural and Multidisciplinary Optimization 21:383í392
Moore GJ (1994) MSC/Nastran Design Sensitivity and Optimization. The
MacNeal–Schwendler Corporation, Los Angeles, CA
Myers RH, Montgomery DC (1995) Response Surface Methodology. Wiley, New
York
Oral S, Ider SK (1997) Optimum Design of High-speed Flexible Robotic Arms

with Dynamic Behavior Constraints. Computers and Structures 65:255í259
Paeng JK, Arora JS (1989) Dynamic Response Optimization of Mechanical
Systems with Multiplier Methods. Journal of Mechanisms, Transmissions, and
Automation in Design, Transactions of the ASME 111:73í80
Park GJ, Kang BS (2003) Validation of a Structural Optimization Algorithm
Transforming Dynamic Loads into Equivalent Static Loads. Journal of
Optimization Theory and Applications 118:191í200
Park KJ, Lee JN, Park GJ (2005) Structural Shape Optimization Using Equivalent
Static Loads Transformed from Dynamic Loads. International Journal for
Numerical Methods in Engineering 63:589í602
Pierson BL (1972) A Survey of Optimal Structural Design under Dynamic
Constraints. International Journal for Numerical Methods in Engineering 4:
491í499
Sepulveda AE, Thomas HL (1996) Improved Transient Response Approximation
for General Damped Systems. AIAA Journal 34:1261í1269
Shabana AA (1998) Dynamics of Multibody Systems. Cambridge University,
Cambridge
Wang S, Choi KK (1992) Continuum Design Sensitivity of Transient Response
Using Ritz and Mode Acceleration Method. AIAA Journal 30:1099í1109
Willmert KD, Fox RL (1972) Optimum Design of a Linear Multidegree-of-
freedom Shock Isolation System. Journal of Engineering for Industry,
Transactions of the ASME 94:465í471
6 Design of Experiments
6.1 Introduction
Experiments are performed to investigate characteristics of a system. Responses or
characteristics are the results evaluated from experiments. The responses should
be analyzed appropriately according to the objectives of the experiments. The
factors are the sources that affect the experiments. It may not be easy to identify
all the factors and give right values because there may be many factors and they
may have distributions and be affected by the variations of the environment.
Generally, the objectives of experiments are as follows:
(1) To measure or evaluate characteristic values without analyzing them
statistically.
(2) Identification through experiment of factors that are significant to the
responses and determination of how large their impact is. The results of
experiments are analyzed statistically.
(3) Statistical identification of factors with little influence.
(4) Determination of the values of significant factors. Then, the optimum
condition can be found.
Design of Experiments (DOE) determines the allocation and method of
experiments to satisfy the objectives. Various factors are determined by analysis
of the experimental results. Generally, DOE proceeds as illustrated in Figure 6.1.
As mentioned earlier, DOE has been utilized for efficient experiments and analysis
of the results. Recently, this method has been applied to design. In design, the
fourth method is adopted to determine optimum values. The characteristic value
and factors of DOE are equivalent to the objective function and design variables of
the design, respectively. Many errors are involved in the experiments. When the
DOE method is utilized in design, the functions are usually evaluated numerically.
The numerical method does not have such errors and is a deterministic method.
Design using DOE can be regarded as a method using metamodels. The
metamodel is the model of a model. A function can be approximated by some
specific function values. The approximated function is a metamodel. Suppose we
have a function y with design variable vector b as follows:
Definition of the objective experiments
Selection of characteristic values

Selection of factors and levels
Allocation of experiments
Randomization of the order of the experiments
Operation of experiments
Analysis of results
Incorporation of the analysis
Figure 6.1. General DOE process
y f (b) (6.1)
where y is a model for a specific characteristic. When it is impossible to obtain the

explicit expression of f, an approximated explicit function can be used in the
design process. For example, the approximated function can be calculated at
b i , i 1,..., k , as follows:
yi f (b i ), i 1,..., k (6.2)
A metamodel ŷ is an approximated function for y with b i and y i using a

specific approximation method as follows:
yˆ g (b) (6.3)
In the real design process, the metamodel ŷ is used. When the error in the
approximation process is H , the relation between y and ŷ is
Design of Experiments 311
y yˆ H (6.4)
In a design using DOE, well developed DOE methods are employed to define
ŷ , and the design variable vector b is obtained. The design problem in this case
can also be expressed as the optimization in Equation 3.1 Generally, a design
using DOE is not mathematically complicated; however, somewhat inferior
solutions are obtained compared to optimization. Two methods are usually used:
(1) the orthogonal arrays method and (2) the response surface method, which uses
the metamodel of explicit functions for approximation.
6.2 Design of Experiments (DOE)

In DOE, the conditions for experiments are systematically allocated, experiments
are performed and the results are statistically analyzed. Generally, we need
equipment and resources for experiments. Therefore, it is quite important to
perform a minimum number of experiments and to obtain maximum information.
There are many mathematical and statistical theories for DOE. In this section,
basic DOE theories applied to design are introduced.
Some terminologies in DOE are equivalent to those in design, for example,
characteristic function, factor, level, etc. These are shown in Table 6.1. The
characteristic function is an experimental result that is a response of the system for
the experiment. Factors are sources that affect the experiment. When conditions
vary, the conditions are factors. The factors are similar to design variables in
design. If an experiment is performed under the temperatures of 50 degrees, 100
degrees and 150 degrees, then the temperature is a factor and the values are levels.
Sophisticated mathematical as well as statistical aspects are utilized in DOE;
however, only simple theories are generally employed in design. See Appendix
6.A. References are recommended for details (Montgomery 1991).
Table 6.1. Terminologies in DOE and design
Terminologies Terminologies
Description
in DOE in design
Design Sources that influence the characteristic function or

Factor the objective function. They can be changed to
variable improve the performance of the system
Values of de-
Level The values that a factor or a design variable can have
sign variables
Characteristic Objective The responses of the system, which are to be
function function maximized or minimized
6.2.1 One-way Factorial Design
One of the simplest DOE theories is the one-way factorial design. It is a method
for allocation of experiments when the number of factors is one and the number of
levels is more than one. For each level, experiments are repeatedly performed.
The number of repetitions (observations) does not need to be the same for each
level. In this section, the case when the number of repetitions is the same is
explained. The data for the one-way factorial design are shown in Table 6.2 for
the case when the number of levels for factor A is l ( A1 , A2 , ..., Al ) and the number
of repetitions is m. In Table 6.2
m
Ti. ¦ x ij , x i. Ti. / m , i 1, 2 , ... , l (6.5a)
j 1
l
T ¦ Ti. , x T / lm (6.5b)
i 1
where xij is the characteristic value (result of experiment) of the jth experiment
when A is Ai . The population exists for each level, the mean is P i and the
variance of the error is V E 2 . Then the mean of the entire population P is
l
P ¦ Pi / l (6.6)
i 1
Using the additive (linear) statistical model, xij is
x ij P i eij P ( P i P ) eij P a i eij (6.7)
where the error eij has a normal distribution N (0, V E 2 ) and is independent.
l
¦ ai 0 should be satisfied because it is an additive model. a i is the main effect
i 1
of A at the ith level. In Equation 6.7, the effect of each level is independently
added. Assumption of the additive model is exploited in DOE.
The analysis of variance (ANOVA) is explained for the one-way factorial
design. ANOVA is a frequently used for statistical analysis. The distribution of a
characteristic function is expressed by the sum of squares (This is called
“variance.”). The sum of squares is decomposed into the sum of squares for each
source and the significant sources are identified. The source means the source of
variation and the variation is the sum of squares of a characteristic function. For
instance, factors, errors and interactions can be sources.
From the data in Table 6.2, the deviation between xij and x is decomposed as
Table 6.2. Data arrangement of a one-way factorial design
Levels of the factor

A1 A2 A3 … Al
x11 x 21 x31 … x l1
x12 x 22 x32 … xl 2
Repetition of x13 x 23 x33 xl 3
…
experiments

x1m x 2m x3 m … xlm
Total T1. T2. T3. … Tl . T
Mean x1. x 2. x3. … xl . x
( x ij x ) ( x ij x i. ) ( x i. x ) (6.8)
Squaring Equation 6.8 yields
l m
2 2 2
¦ ¦ ( x ij x ) ¦ ¦ ( x ij x i. ) ¦ ¦ ( x i. x ) 2¦ ¦ ( x ij x i. )( x i. x )
i 1j 1 i j i j i j
(6.9)
The third term of the right hand side of Equation 6.9 is zero. Thus, Equation 6.9
becomes
2 2 2
¦ ¦ ( xij x ) ¦ ¦ ( x ij x i. ) ¦ ¦ ( x i. x ) (6.10)
i j i j i j
Therefore, the total sum of squares S T is decomposed into the sum of squares S A
due to A and the sum of squares S E due to errors as follows:
ST S A SE (6.11a)
l m 2
2
ST ¦ ¦ ( xij x ) ¦ ¦ xij CT (6.11b)
i 1j 1 i j
Ti. 2
SA ¦ ¦ ( xi . x )
2
m ¦ ( xi . x ) 2 m¦ xi. 2 CT ¦ CT (6.11c)
i j i i i m
2 2 2
SE ¦ ¦ ( xij xi. ) ¦ ¦ xij m ¦ xi.
i j i j i
§ · § T 2 ·
¨¨ ¦ ¦ xij 2 CT ¸¸ ¨ ¦ i. CT ¸ ST S A (6.11d)
©i j ¨
¹ ©i m ¸
¹
T2
where CT is the correction term.
lm
Next, the degree of freedom (DOF) is derived. Subtraction of the number of
constraints from the number of deviations for squares yields the DOF of the sum of
squares. Thus, each DOF is
IT lm 1 (6.12a)
IA l 1 (6.12b)
IE lm l l (m 1) (6.12c)
and
IT I A IE (6.13)
The ANOVA table shows the above results in a compact form. The mean
square V is defined as follows:
Table 6.3. ANOVA table for a one-way factorial design
Factor Sum of squares DOF Mean square F0
l SA VA
A SA m ¦ ( x i. x ) 2 IA l 1 VA F0
i 1 IA VE
SE
E SE ST S A IE l (m 1) VE
IE
l m
2
T ST ¦ ¦ ( x ij x ) IT lm 1
i 1j 1
S
V (6.14)
I
where S is the sum of squares and I is the degree of freedom. The ANOVA table
for the one-way factorial design is shown in Table 6.3.
From the ANOVA table, the population mean at each level P i P ai is
estimated. The population mean is the characteristic value at each level. m data of
the ith level are regarded as random samples with a normal distribution
N ( Pi , V E 2 ). The unbiased estimate for V E 2 is VE . Using Equations
6.A.20í6.A.22 of Appendix 6.A, the estimate of P i with 100(1 D )% confidence
interval is as follows:
§ D· V
xi. r t ¨ I E ; ¸ E (6.15)
© 2¹ m
SE
where V E and the degree of freedom I E for V E is l (m 1) . The t-
IE
distribution in Equation 6.15 is briefly introduced in Appendix 6.A (Walpole
SA
1982). If V A , F0 is defined to compare the mean squares of the variations,
IA
VA
F0 (6.16)
VE
The value in Equation 6.16 is used in an F test and it is the relative ratio of the
sum of squares due to a factor with respect to the sum of squares due to error.
When we have multiple factors, the factors with low F values can be ignored.
Table 6.4. Results of Example 6.1
Level of the factor

A1 A2 A3
7.45 8.21 9.78
Response
8.24 7.82 8.82
values
6.54 9.21 8.54
Total T1. = 22.23 T2. = 25.24 T3. = 27.14 T = 74.61
Mean x1. = 7.41 x2. = 8.41 x3. = 9.05 x = 8.29

Table 6.5. ANOVA table for Example 6.1
Factor Square sum S DOF I Mean square V F
A S A = 4.0992 IA = 2 V A = 2.0496 V A V E =3.7025
E S E =3.3214 IE = 6 V E = 0.5536
T S T =7.4078 IT = 8
Also, the F test is used for statistical hypothesis to accept or reject a hypothesis.
However, since it is rarely used in design, it is not explained here.
Example 6.1 [ANOVA Table for a One-way Factorial Design]

The results in Table 6.4 were obtained from an experiment. Make an ANOVA
table of the results.
Solution
From Equation 6.11
3 3
2
ST ¦ ¦ ( xij x )
i 1j 1
(7.45 8.29) 2 (8.24 8.29) 2 (8.54 8.29) 2
7.4078 (6.17a)
3
SA m ¦ ( x i. x ) 2
i 1
3[(7.41 8.29) 2 (8.41 8.29) 2 (9.05 8.29) 2 ]
4.0992 (6.17b)
3 3
2
SE ¦ ¦ ( xij xi. )
i 1j 1
(7.45 7.41) 2 (8.24 7.41) 2 (8.54 9.05) 2
3.3214 (6.17c)
The above results can be obtained by using CT in Equation 6.11. The ANOVA
table is shown in Table 6.5.
Example 6.2 [Estimation of the Confidence Interval for a One-way

Factorial Design]
For Example 6.1, estimate the 95% confidence interval for the population mean at
Ai , i 1, 2, 3.
Solution
Equation 6.15 is utilized for the calculation of the confidence interval. From the
95% confidence interval, D = 0.05. From Tables 6.4 and 6.5, m 3, x1. 7.41,
x2. 8.41, x3. 9.05, V E 0.5536 and I E 6. Substituting these values into
Equation 6.15, the confidence interval is
VE 0.5536
t (I E ; 0.025) t (6; 0.025) (6.18)
m 3
where t (I E ; 0.025) 2.447. This can be found in the t-distribution table in a

general textbook for statistics (Walpole 1982). Therefore, the confidence intervals
of the population means at the levels are as follows:
P1 : 7.41 r 1.0511 (6.36, 8.46) (6.19a)
P 2 : 8.41 r 1.0511 (7.36, 9.46) (6.19b)
P 3 : 9.05 r 1.0511 (8.00,10.10) (6.19c)
6.2.2 Two-way Factorial Design

Now we consider a two-way factorial design without repetitions. The two-way
factorial design is an analysis method for experimental results with two factors.
Suppose the two factors are A and B, and the numbers of levels are l and m,
respectively. The data from an experiment are given in Table 6.6 where a factorial
design that performs experiments for all combinations is shown.
The additive statistical model is
xij P ai b j eij , i 1,..., l , j 1,..., m (6.20)
where eij has a normal distribution N (0, V E 2 ). xij is the measured value for Ai
and B j , P is the population mean, ai is the main effect of Ai , b j is the main
effect of B j and eij is the measuring error for Ai and B j . The interaction cannot
be measured from the two-way factorial design without repetitions. In other words,
the interaction effects are confounded with the error eij . The interaction will be
explained later.
Table 6.6. Data of a two-way factorial design
Factor A
A1 A2 … Al Total Mean
Factor B
B1 x11 x21 … xl1 T.1 x.1
B2 x12 x22 xl 2 T.2 x.2

Bm x1m x2 m … xlm T.m x.m
Total T1. T2. … Tl . T
Mean x1. x 2. … xl . x
From Table 6.6, the following equations are derived:
m
Ti. ¦ xij , xi. Ti. / m, i 1,..., l (6.21a)
j 1
l
T. j ¦ xij , x. j T. j / l , j 1,..., m (6.21b)
i 1
Data in Table 6.6 are modified as follows:
( x ij x ) ( xi. x ) ( x. j x ) ( xij x i. x. j x ) (6.22a)
l m
2 2 2
¦ ¦ ( x ij x ) ¦ ¦ ( x i . x ) ¦ ¦ ( x. j x )
i 1j 1 i j i j
¦ ¦ ( x ij x i. x. j x ) 2 (6.22b)
i j
l m 2
2
ST ¦ ¦ ( x ij x ) ¦ ¦ x ij CT (6.22c)
i 1j 1 i j
2 Ti. 2
SA ¦ ¦ ( x i. x ) m ¦ ( x i. x ) 2 ¦ CT (6.22d)
i j i i m
2
T. j 2
SB ¦ ¦ ( x. j x ) l ¦ ( x. j x ) 2 ¦ CT (6.22e)
i j j j l
Table 6.7. ANOVA table for a two-way factorial design without repetitions
Factor S I V E (V ) F0 F (D )
VA
A SA l 1 VA V E 2 mV A 2 F (I A , I E ;D )
VE
VB
B SB m 1 VB V E 2 lV B 2 F (I B , I E ;D )
VE
E SE (l 1)(m 1) VE V E2
T ST lm 1
2 2
SE ¦ ¦ ( xij xi. x. j x ) ¦ ¦ {( xij x ) ( xi. x ) ( x. j x )}
i j i j
2 2 2
¦ ¦ ( xij x ) ¦ ¦ ( xi. x ) ¦ ¦ ( x. j x ) S T S A S B (6.22f)
i j i j i j
The corresponding ANOVA table is shown in Table 6.7. Using the

aforementioned formulae for estimation, the 100(1 D )% confidence intervals for
P ( Ai ), P ( B j ) and P ( Ai , B j ) are
§ D· V
xi . r t ¨ I E ; ¸ E (6.23a)
© 2¹ m
§ D· V
x. j r t ¨ I E ; ¸ E (6.23b)
© 2¹ l
§ D· V
( xi. x. j x ) r t ¨ I E ; ¸ E (6.23c)
© 2 ¹ ne
lm
respectively. ne is the effective number of repetitions and ne . The
l m 1
effective number of repetitions is the number of repetitions to which a specific
factor contributes at a specific level. For example, m is the number of repetitions
in Equation 6.15. In a two-way factorial design, when a factor is repeated, the
levels of other factors are changed. Therefore, the number of repetitions of a
factor does not directly contribute to the errors. The effective number of
repetitions includes this phenomenon. The optimum levels for the population
K K B3 K
B3
B3 B2 B2
B2 B1 B1
B1
A1 A2 A3 A1 A2 A3 A1 A2 A3
(a) (b) (c)
Figure 6.2. Interaction: (a) No interaction, (b) synergistic interaction, (c) antisynergistic
interaction
mean are the levels for the optimum ( xi. x. j x ) . This is further explained in
(Montgomery 1991).
When experiments are performed with two factors, experimental results can be
measured repeatedly for specific levels. This is called a “two-way factorial design
with repetitions” and the cost increases due to the increments of the number of
experiments. However, the interactions can be measured. When factor A has l
levels, factor B has m levels and the number of repetitions is r, the additive
statistical model is
xijk P ai b j (ab) ij eijk , i 1,..., l , j 1,..., m, k 1,..., r (6.24)
where eijk has a normal distribution N (0, V E 2 ) . (ab) ij represents the interaction
effect of factors A and B in the ith and the jth levels, respectively. The interaction
is the effect due to a combination of factors. Figure 6.2a does not have
interactions. Figures 6.2b and 6.2c show interaction effects. Since Figure 6.2b
shows the synergistic interaction, it can be ignored. However, the effect of the
antisynergistic interaction as illustrated in Figure 6.2c is significant and the
interactions should be considered in the design process.
Data from a two-way factorial design with repetitions are shown in Table 6.8.
The following equations are obtained:
m r
Ti.. ¦ ¦ xijk , xi.. Ti.. / mr , i 1,..., l (6.25a)
j 1k 1
l r
T. j. ¦ ¦ xijk , x. j. T. j. / lr , j 1,..., m (6.25b)
i 1k 1
l m
T ¦ Ti.. ¦ T. j. , x T / lmr (6.25c)
i 1 j 1
Table 6.8. Data for a two-way factorial design with repetitions
Factor A
A1 A2 … Al Total Mean
Factor B
x111 x 211 xl11
x112 x 212 xl12
B1 … T.1. x.1.

x11r x 21r xl1r
x121 x 221 xl 21
x122 x 222 xl 22
B2 … T.2. x.2.

x12 r x 22 r xl 2 r

x1m1 x 2 m1 xlm1
x1m 2 x 2m 2 xlm 2
Bm … T.m. x.m.

x1mr x 2 mr xlmr
Total T1.. T2.. … Tl .. T
Mean x1.. x 2.. … xl .. x
r
Tij. ¦ xijk (6.25d)
k 1
Tij.
xij. (6.25e)
r
At the levels of Ai and B j , the sum of r data is
r
Tij. ¦ xijk (6.26)
k 1
The experimental data yields
( xijk x ) ( xi.. x ) ( x. j. x ) ( xij . xi.. x. j. x ) ( xijk xij. ) (6.27)

Equation 6.27 can be modified as follows:
l m r
2 2 2
¦ ¦ ¦ ( xijk x ) ¦ ¦ ¦ ( xi.. x ) ¦ ¦ ¦ ( x. j. x )
i 1 j 1k 1 i j k i j k
¦ ¦ ¦ ( xij . xi.. x. j. x ) 2 ¦ ¦ ¦ ( xijk xij . ) 2 (6.28)

i j k i j k
Each term of Equation 6.28 is the total sum of squares ( ST ) , the sum of squares
due to A ( S A ) , the sum of squares due to B ( S B ) , the sum of squares due to the
interaction between A and B ( S AuB ) and the sum of squares due to error ( S E ).
Actually, it is convenient to use Equation 6.29 instead of Equation 6.28,
2
ST ¦ ¦ ¦ xijk CT (6.29a)
i j k
Ti..2
SA ¦ CT (6.29b)
i mr
T. j.2
SB ¦ CT (6.29c)
j lr
S AuB S AB S A S B (6.29d)
SE ST S AB (6.29e)
Tij .2
S AB ¦¦ CT (6.29f)
i j r
T2
where CT . The degrees of freedom for above sums of squares are IT
lmr
lmr 1, I AB lm 1, I A l 1, I B m 1, I AuB (l 1)(m 1) and I E lm
(r 1), respectively. The ANOVA table is shown in Table 6.9.
The (1 D ) confidence intervals for the combination of Ai B j are
§ D· V
xi.. r t ¨ I E ; ¸ E (6.30a)
© 2 ¹ mr
§ D· V
x. j. r t ¨ I E ; ¸ E (6.30b)
© 2 ¹ lr
§ D· V
xij . r t ¨ I E ; ¸ E (6.30c)
© 2¹ r
Table 6.9. ANOVA table for a two-way factorial design with repetitions
Factor S I V E (V ) F0 F (D )
VA
A SA l 1 VA V E 2 mrV A 2 F (I A , I E ; D )
VE
VB
B SB m 1 VB V E 2 lrV B 2 F (I B , I E ; D )
VE
V Au B
Au B S Au B (l 1)(m 1) V Au B V E 2 rV AuB 2 F (I Au B , I E ; D )
VE
E SE lm(r 1) VE V E2
T ST lmr 1
respectively. The optimum values are the levels that make the optimum of xij . .
The design of experiments (DOE) is explained briefly. As mentioned earlier,
experiments are appropriately allocated in DOE. Through statistical analysis,
values are estimated or optimum values are derived. When there are more than
two factors, a multiway factorial design must be utilized. Since the number of
experiments is considerably increased, it is rarely used. Moreover, there are many
other statistical theories for estimation, testing, error analysis, etc. However,
because those theories are not generally used for design, they are not explained
here.
Example 6.3 [ANOVA Table for a Two-way Factorial Design]

In an experiment, suppose that factor A has three levels and factor B has four levels.
A two-way factorial design is performed with two repetitions for each case. The
results are shown in Table 6.10. Obtain the ANOVA table for the results.
Solution
From Table 6.10, the sum Tij . and the mean xij . are derived. From Equation 6.29,
T2
CT 630.38 (6.31a)
lmr
2
ST ¦ ¦ ¦ xijk CT 787 630.38 156.62 (6.31b)
i j k
Ti..2
SA ¦ CT 33.25 (6.31c)
i mr
Table 6.10. Experimental results of Example 6.3
Factor A
A1 A2 A3 T. j. x. j .
Factor B
2 3 4 18 3
B1
1 5 3
3 7 6 28 4.67
B2
4 3 5
6 10 7 48 8
B3
8 12 5
4 8 6 29 4.83
B4
2 5 4
Ti.. 30 53 40 123
x i.. 3.75 6.63 5 x 5.13
T. j.2
SB ¦ CT 78.45 (6.31d)
j lr
Tij.2
S AB ¦¦ CT 130.12 (6.31e)
i j r
S AuB S AB S A S B 18.42 (6.31f)
SE ST S AB 26.5 (6.31g)
From Equation 6.31, the ANOVA table in Table 6.11 is obtained.
Table 6.11. ANOVA table of Example 6.3
Square
Factor DOF I Mean square V F
sum S
A 33.25 IA l 1 2 16.63 V A VE 7.52
B 78.45 IB m 1 3 26.15 VB VE 11.84
Au B 18.42 I AuB (l 1)(m 1) 6 3.07 V Au B V E 1.39
E 26.5 IE lm(r 1) 12 2.21

T 156.62 lmr 1 23
Table 6.12. The characteristic function for each factor of Example 6.3
Factor A A1 A2 A3
Factor B Total Mean Total Mean Total Mean
B1 3 1.5 8 4 7 3.5
B2 7 3.5 10 5 11 5.5
B3 14 7 22 11 12 6
B4 6 3 13 6.5 10 5
Example 6.4 [Optimization of the Characteristic Function and

Estimation of the Confidence Interval of the Population Mean in a
Two-way Factorial Design]
In Example 6.3, obtain the levels that maximize the characteristic function and
estimate the 95% confidence interval of the population mean.
Solution
From Table 6.12, the combination that maximizes the characteristic function is
A2 B3 . From Equation 6.30c, the 95% confidence interval for the population mean
is
§ D· V 2.21
x23. r t ¨ I E ; ¸ E 11 r t (12; 0.025)
© 2¹ r 2
11 r 2.179 u 1.051 11 r 2.29 (6.32)
t (12; 0.025) can be found in the table for t-distribution (Walpole 1982).
6.3 Design Using Orthogonal Arrays
6.3.1 Orthogonal Arrays
The aforementioned DOE is the factorial design. Although the factorial design
provides a lot of precise data, it is an expensive method due to the large number of
experiments required. The fractional design is utilized to save cost. In fractional
design, the metamodel for the factorial design is derived and subsequently the
metamodel is analyzed. Therefore, allocation of factors and the analysis of the
DOE results are quite important.
There are many methods in fractional design. In this section, the method using
orthogonal arrays is explained. Orthogonal arrays are generally utilized when
fractional design is directly applied to design. Other than this method, fractional
design is usually utilized to generate the design surfaces in the response surface
method. Fractional design is indirectly used in the response surface method. An
orthogonal array is an array that easily enables the design of experiments. When
there are many factors, each factor and the plausible interactions between factors
are allocated to columns of an orthogonal array. Since their introduction by Rao in
1946, orthogonal arrays have been widely used in many fields. They are mostly
used in the statistics area (Hedayat et al. 1999) where there are many theories
concerning them. However, aspects related to design are briefly discussed in this
chapter.
An orthogonal array is represented by a matrix. There are a few methods to
express them and we will use Taguchi’s method as follows (Taguchi 1987):
n
LN ( sik i ) (6.33)
i 1
where N is the number of rows for an orthogonal array, n is the number of factors
with a specific level, si is the number of levels, ki is the number of factors with
si levels. For example, L18 (2137 ) is an orthogonal array having 18 rows, one
factor with 2 levels and 7 factors with 3 levels.
The orthogonal array L9 (3 4 ) is shown in Table 6.13. It enables an experiment
of up to four factors with three levels. In a factorial design, 34 81 times the
experiments are required. Only nine times the experiments are required with the
orthogonal array. At each column, each level appears at the same time. Each
column is a vector with nine components. Substitute the components 1, 2 and 3 by
í1, 0 and 1. If we make inner products with two columns, they are zeros. In other
words, the vectors are orthogonal to each other. This orthogonality is the basis of
its name. Various orthogonal arrays are shown in Appendix 6.B.
Example 6.5 [Orthogonality of the Orthogonal Array]

Prove that the columns of L9 (3 4 ) in Table 6.13 are orthogonal.
Solution
Let each column of Table 6.13 n i and substitute 1, 2 and 3 in Table 6.13 by í1, 0
and 1 as follows:
n1 [ 1 1 1 0 0 0 1 1 1]T (6.34a)
n2 [ 1 0 1 1 0 1 1 0 1]T (6.34b)
n3 [ 1 0 1 0 1 1 1 1 0]T (6.34c)
Table 6.13. The L9 (3 4 ) orthogonal array
Experiment Factor and level Characteristic

number A B C D function
1 1 1 1 1 K1
2 1 2 2 2 K2
3 1 3 3 3 K3
4 2 1 2 3 K4
5 2 2 3 1 K5
6 2 3 1 2 K6
7 3 1 3 2 K7
8 3 2 1 3 K8
9 3 3 2 1 K9
n4 [ 1 0 1 1 1 0 0 1 1]T (6.34d)
The inner product of n1 and n 2 is
n1T n 2 (1)(1) (1)(0) (1)(1) (0)(1) (0)(0) (0)(1)
(1)(1) (1)(0) (1)(1) 1 1 1 1 0 (6.35)
In the same manner,
n1T n 3 n1T n 4 n 2T n 3 n 2T n 4 n 3T n 4 0 (6.36)
Therefore, the columns are orthogonal to each other.
6.3.2 Analysis of Experimental Results Using Orthogonal Arrays
An experiment using an orthogonal array is explained for an example with four

factors of three levels. The data for the factors are shown in Table 6.14. Each row
of Table 6.13 is a case of the experiment. The characteristic function K of Table
6.13 is summarized in Table 6.15. The experimental results should be
systematically analyzed. Suppose the goal of the experiment is to determine the
levels that minimize the characteristic function K . Moreover, the ANOVA table
Table 6.14. An example of factors and levels
Level
Factor
1 2 3
A A1 A2 A3
B B1 B2 B3
C C1 C2 C3
D D1 D2 D3
Table 6.15. Characteristic functions for each experiment
Experiment
K1 K2 K3 K4 K5 K6 K7 K8 K9
number
Characteristic
20 50 30 25 45 30 45 65 70
function
can be made from the results of the orthogonal array. This will be shown in
Example 6.8.
From Table 6.15, the mean of the characteristic function is
1 9
m ¦ Ki (6.37)
9i 1
The effect by the levels of each factor is investigated. m A3 is the mean when
factor A has A3 as follows:
1
m A3 (K 7 K8 K9 ) (6.38)
3
Therefore, the effect of the level A3 is (m A3 m) . In the same manner, m A1 and

m A2 are
1
m A1 (K1 K2 K3 ) (6.39a)
3
1
m A2 (K 4 K5 K 6 ) (6.39b)
3
Ș
80
70
60
50
40
30
20
10
A1 A2 A3 B1 B2 B3 C1 C2 C3 D1 D2 D3
Figure 6.3. Graphs for the results of each factor in Table 6.15
The means for factors B, C and D are obtained by the mean with specific levels
as well. Results derived from Table 6.15 are illustrated in Figure 6.3. The levels
minimizing K are considered optimum. Factor A has the same results for the first
and second levels. Therefore, the optimum levels are A1 B1C1 D3 or A2 B1C1 D3 .
This process is called the “analysis of means (ANOM).” Instead of Figure 6.3, the
one-way table in Table 6.16 can be utilized. This will be explained later.
The design process from ANOM is performed based on the assumption that the
statistical model for K satisfies additivity as follows:
K ( Ai , B j , Ck , Dl ) P ai b j c k d l e (6.40)
where P is the population mean. ai , b j , ck and d l are the deviations from the
population mean when factors A, B, C and D have the i, j, k and lth level,
respectively. Error e includes the errors from the experiments and the additive
model. From Equation 6.40, the following equations are derived:
a1 a 2 a3 0 (6.41a)
b1 b2 b3 0 (6.41b)
c1 c2 c3 0 (6.41c)
d1 d 2 d 3 0 (6.41d)
From the additive model of Equation 6.38, m A3 is as follows:

Table 6.16. An example of the one-way table for an orthogonal array
Level
Factor
1 2 3
K1 K 2 K 3 K 4 K5 K 6 K 7 K8 K 9
A m A1 m A2 m A3
3 3 3
K1 K 4 K 7 K 2 K 5 K8 K3 K6 K9
B m B1 m B2 m B3
3 3 3
K1 K 6 K8 K 2 K 4 K9 K3 K5 K7
C mC1 mC2 mC3
3 3 3
K1 K 5 K 9 K2 K6 K7 K 3 K 4 K8
D m D1 m D2 m D3
3 3 3
1
m A3 (K 7 K8 K 9 )
3
1
[( P a3 b1 c3 d 2 e7 ) ( P a3 b2 c1 d 3 e8 )
3
( P a 3 b3 c 2 d1 e9 )]
1 1 1
(3P 3a3 ) (b1 b2 b3 ) (c1 c 2 c3 )
3 3 3
1 1
(d1 d 2 d 3 ) (e7 e8 e9 )
3 3
1
( P a3 ) (e7 e8 e9 ) (6.42)
3
If we ignore the error in Equation 6.42, m A3 is made by adding a3 to the

population mean. In the rows with a specific level of a factor, each level for other
factors appear the same time and their effect is zero as shown in Equation 6.41. It
is noted that 1/3 is multiplied to each error in Equation 6.42 and the error variance
1
becomes V e2 . Therefore, DOE using orthogonal arrays efficiently evaluates the
3
effect of each factor with a small number of experiments.
The optimum levels from DOE using orthogonal arrays may or may not be a
row in the orthogonal array. In the previous example, suppose that the optimum
combination is A1 B1C 2 D2 , which is different from any rows in the orthogonal

array L9 (3 4 ) . In this case, we can estimate the characteristic function K opt
pred
from
the additive statistical model as follows:
pred
K opt m (m A1 m) (mB1 m) (mC 2 m) (mD2 m) (6.43)
When a term in Equation 6.43 is not significant, it can be pooled to the error
term. For example, if the effects of C2 and D2 are negligible, the prediction can
be made by Equation 6.44,
pred
K opt m (m A1 m) (mB1 m) (6.44)
The decision can be made from the ANOVA table introduced in Example 6.8.
When the optimum level is different from any rows of the orthogonal array and
the characteristic function is estimated by Equation 6.43, a confirmation
experiment with the optimum levels should be performed. If the results of the
confirmation experiment are similar to the estimated value, the process is correct.
On the other hand, if the difference is large the additive model is not appropriate.
That is, the interaction effect compounded to the error is considerably large. It is
not easy to consider the interactions; this will be explained later. Selection of an
orthogonal array is quite important. Example 6.6 illustrates the method.
Example 6.6 [Selection of an Orthogonal Array]

In an experiment, factor A has two levels and factors B, C, D, E and F have three
levels. Moreover, the interaction A u B exists. Select an appropriate orthogonal
array.
Solution
First, the degree of freedom is evaluated as in the table below. The total DOF is
14. We have to select an orthogonal array that has at least one column of two
levels, at least five columns of three levels and at least fourteen rows. Generally,
Factor/Interaction Degrees of freedom
Mean 1
A 2 (levels) – 1 = 1
B,C,D,E,F 5 u (3 (levels) –1) = 10
Au B (2 – 1) u (3 – 1) = 2
Total 14
the smallest one is selected. When the second smallest one is utilized, we may
obtain better results. On the other hand, then the number of experiments is
increased.
Example 6.7 [Optimization of a Mathematical Problem Using an

Orthogonal Array]
We have an optimization problem as follows:
minimize f A 2 ( B 1.5) 2 (C 2.5) 2 (6.45)
Design variables A, B and C have discrete values as shown in Table 6.17.

(1) Obtain a mathematical optimum in a continuous space.
(2) Obtain a discrete optimum from the factorial design.
(3) Obtain a discrete optimum by using an L9 (34 ) orthogonal array. Compare
the solution with the one from the factorial design.
Solution
(1) The function consists of perfect square terms. Thus, the optimum solution
is A = 0, B = 1.5 and C = 2.5.
(2) The factorial design is shown in Table 6.18 where function f is the smallest
when the design point is A1B3C2 . The values of the design point are (í0.2,
2.0, 2.8) and the function value is f = 0.38.
(3) The function value of Equation 6.45 is calculated according to the rows of
the orthogonal array in Table 6.13. They are
f1 12.54, f2 2.38, f3 1.5, f4 12.7
f5 3.82, f6 0.86, f7 15.42, f8 4.46, f9 2.3
In Table 6.16, K is replaced by f and the one-way table in Table 6.19 is set up.
Table 6.17. Factors and levels of Example 6.7
Level
Factor
1 2 3
A –0.2 0.6 1.4
B –2.0 0 2.0
C 2.0 2.8 3.6

Table 6.18. Factorial design of Example 6.7
Exp.
A B C f
no.
1 –0.2 –2.0 2.0 12.54
2 –0.2 –2.0 2.8 12.38
3 –0.2 –2.0 3.6 13.5
4 –0.2 0 2.0 2.54
5 –0.2 0 2.8 2.38
6 –0.2 0 3.6 3.5
7 –0.2 2.0 2.0 0.54
8 –0.2 2.0 2.8 0.38
9 –0.2 2.0 3.6 1.5
10 0.6 –2.0 2.0 12.86
11 0.6 –2.0 2.8 12.7
12 0.6 –2.0 3.6 13.82
13 0.6 0 2.0 2.86
14 0.6 0 2.8 2.7
15 0.6 0 3.6 3.82
16 0.6 2.0 2.0 0.86
17 0.6 2.0 2.8 0.7
18 0.6 2.0 3.6 1.82
19 1.4 –2.0 2.0 14.46
20 1.4 –2.0 2.8 14.3
21 1.4 –2.0 3.6 15.42
22 1.4 0 2.0 4.46
23 1.4 0 2.8 4.3
24 1.4 0 3.6 5.42
25 1.4 2.0 2.0 2.46
26 1.4 2.0 2.8 2.3
27 1.4 2.0 3.6 3.42
In each row of Table 6.19, the level for each factor that makes the objective
function the smallest is chosen. Therefore, the optimum levels are A1 B3C 2 and
their values are (í0.2, 2.0, 2.8). This is the same as the case of the factorial design
because there are no interactions.
Using Equation 6.43, the function is estimated at the optimum point. The mean
m is
Table 6.19. One-way table of the orthogonal array of Example 6.7
Design Level
variable 1 2 3
A m A1 5.47 m A2 5.79 m A3 7.39
B m B1 13.55 m B2 3.55 m B3 1.55
C mC1 5.95 mC2 5.79 mC3 6.91
9
¦ fi
i 1
m 6.22 (6.46)
9
Substituting the values in Table 6.19 into Equation 6.43, the estimated value is
pred
f opt 6.22 (5.47 6.22) (1.55 6.22) (5.79 6.22) 0.37 (6.47)
The confirmation experiment at the optimum point is f opt 0.38 .

Since there are no interactions in Equation 6.45, additivity is satisfied.
Therefore, the solution from the orthogonal array is the same as the case of the
pred
factorial design. The error between f opt and f opt comes from the round-off
process.
Example 6.8 [ANOVA Table of Example 6.7]

In Example 6.7, make an ANOVA table for the results of the orthogonal array.
Solution
Considering the levels and function values of Example 6.7, yijk is defined as
follows:
y111 f1 , y122 f2 , y133 f3
y 212 f4 , y 223 f5 , y 231 f6
y313 f7 , y321 f8 , y332 f9

2
§9 ·
¨ ¦ fi ¸
T2 ©i 1 ¹ 55.98 2
CT 348.20 (6.48)
9 9 9
The sum of squares due to factor A is made from the first column of the
orthogonal array in Table 6.13.
3Ti..2
SA ¦ CT
i 1 3
1
(T1.. 2 T2.. 2 T3.. 2 ) CT
3
1
[( y111 y122 y133 ) 2 ( y 212 y 223 y 231 ) 2
3
( y313 y321 y332 ) 2 ] CT
1
[(12.54 2.38 1.5) 2 (12.7 3.82 0.86) 2
3
(15.42 4.46 2.3) 2 ] 348.20 6.34 (6.49a)
where Ti.. is the sum of f for the ith level ( Ai ) of factor A. In the same manner,
3T.i.2
SB ¦ CT
i 1 3
1
(T.1. 2 T.2. 2 T.3. 2 ) CT
3
1
[( y111 y 212 y313 ) 2 ( y122 y 223 y321 ) 2
3
( y133 y 231 y332 ) 2 ] CT
248 (6.49b)
3 T..i 2
SC ¦ CT
i 1 3
1
(T..12 T..2 2 T..3 2 ) CT
3
1
[( y111 y 231 y321 ) 2 ( y122 y 212 y332 ) 2
3
( y133 y 223 y313 ) 2 ] CT
2.20 (6.49c)
The sum of squares due to error is calculated from column D of Table 6.13,
Table 6.20. ANOVA table of Example 6.8
Factor S I V
A 6.34 2 3.17
B 248 2 124
C 2.20 2 1.10
Error (E ) 0 2
T 256.54 8
1
SE [( y111 y 223 y332 ) 2 ( y122 y 231 y313 ) 2
3
( y133 y 212 y321 ) 2 ] CT 0 (6.49d)
S E 0 because there are no interactions, and no measuring errors. Therefore, the

total sum of squares is
ST S A S B SC S E 256.54 (6.50)
In another way,
2
ST ¦ ¦ ¦ y ijk CT 604.74 348.20 256.54 (6.51)
i j k
The ANOVA table is shown in Table 6.20.
6.3.3 Design Using Orthogonal Arrays
Orthogonal arrays can be directly used in design. An algorithm has been

developed called “discrete design using orthogonal arrays (DOA).” Constrained
problems will be explained later. Unconstrained problems are considered in the
literature (Lee et al. 1996). Suppose we have the following problem:
Find b Rn (6.52a)
where the number of design variables n = 4 and the number of levels for each
design variable is three. The DOE notation is somewhat different from the design
notation. For instance, the objective function f in design is equivalent to the
characteristic function K in DOE. The design variables b1 , b2 , b3 and b4 in design

are the same as the factors A, B, C and D in DOE. The steps in DOA are as
follows:
Step 1. An appropriate orthogonal array is selected according to the number
of design variables and the number of levels. The selection of the L9 (34 )
orthogonal array is shown in Table 6.13.
Step 2. The objective function f or K is calculated for each row of the
selected orthogonal array. This process is called a matrix experiment.
Step 3. The one-way table shown in Table 6.16 is made for each design
variable bi , i 1, ..., 4 or A, B, C and D.
Step 4. For each design variable bi , i 1, ...,4, find the indices oi , i 1, ...,4,
for the smallest mbij , j 1, ...,3.
Step 5. The solution is the oi th level of bi , i 1, ...,4. For the confirmation
experiment, the objective function is evaluated with the solution. If it is
better than f1 , f 2 , ..., f 9 of the orthogonal array, the solution is the final
solution. Otherwise, the best one from f1 , f 2 , ..., f 9 is chosen and the
corresponding levels are the final solution.
The above algorithm is a method that can find a solution with a small number
of experiments (function calculation). Moreover, a discrete design is obtained.
When interactions exist, Step 4 may not give a good solution. A row of the
orthogonal array can be better than the one from Step 4. Step 5 is added to
consider this phenomenon. The algorithm can be exploited without considering
interactions that are very expensive to identify. When factors are allocated to
levels, it is recommended that the current design be allocated to the first levels.
The current design appears in the first row of the orthogonal array. Then, at least
an improved solution from the current design is guaranteed. It is noted that the
order of the levels does not need to be in the order of the magnitude of a design
variable.
In general optimization or the response surface method, an optimum solution is
found in a continuous space. Available solutions in design practice frequently
exist in a discrete space. Rounded up values are chosen as the final solution of
optimization in the continuous space. However, a rounded up solution may not be
good when the constraints are sensitive with respect to design variables. The
constraints can be violated with the rounded up solution. Other methods for
discrete design are integer programming (Korte and Vygen 2003), genetic
algorithms (Goldberg 1989), simulated annealing (Osman and Kelly 1996), etc.
These methods are not explained here. They are mathematically excellent;
however, a large number of function evaluations are required. Therefore, the
method in this section excels in that a better solution than the current design can be
obtained with low cost.
Table 6.21. Levels for the factors of Example 6.9
Level
Factor
1 2
A – 0.5 0.3
B – 0.6 0.4
C – 0.9 1.0
D – 0.5 0.2
Example 6.9 [Application of DOA to a Mathematical Problem]

We have the following mathematical problem:
Minimize f ( A B) 2 2 B 2 (2 A C ) 2 ( D 0.3) 2 (6.53)
Each factor has two levels as shown in Table 6.21.

(1) Using the DOA method, obtain the solution of Equation 6.53.
(2) Obtain the solution by using the factorial design. Compare the solution
with that of (1).
Solution
(1) The levels of each factor in Table 6.21 are substituted into the L8 (2 7 )
Table 6.22. The L8(27) orthogonal array
Factor
Exp. no.
1 2 3 4 5 6 7
1 1 1 1 1 1 1 1
2 1 1 1 2 2 2 2
3 1 2 2 1 1 2 2
4 1 2 2 2 2 1 1
5 2 1 2 1 2 1 2
6 2 1 2 2 1 2 1
7 2 2 1 1 2 2 1
8 2 2 1 2 1 1 2
Interaction A B Au B C Au C D e
Table 6.23. The function value of each row of L8 ( 2 7 ) in Example 6.9
f1 f2 f3 f4 f5 f6 f7 f8
4.96 0.72 1.04 2.88 3.13 0.15 0.01 3.11
Table 6.24. One-way table of Example 6.9
Factor Level 1 Level 2
f1 f 2 f 5 f 6 f3 f 4 f 7 f8
A 2.24 1.76
4 4
f1 f 2 f 7 f 8 f3 f 4 f5 f6
B 2.2 1.8
4 4
f1 f 3 f 5 f 7 f 2 f 4 f6 f8
C 2.29 1.72
4 4
f1 f 3 f 6 f 8 f 2 f 4 f5 f7
D 2.32 1.69
4 4
orthogonal array in Table 6.22. Factors A, B, C and D are allocated to

columns 2, 3, 4, and 5 of Table 6.22 and the function is calculated for each
row. f i is the function value of the ith row as shown in Table 6.23. From
the one-way table in Table 6.24, the smallest level is chosen. The solution
exp
is A2 B2 C 2 D2 and the confirmation experiment gives f opt 2.88.
Comparing this with the rows of the orthogonal array, the seventh row of
the orthogonal array gives a better solution because the interaction is
significant. Therefore, the final result is A2 B1C1 D2 .
(2) Table 6.25 shows the results of the factorial design. The 8th row gives the
solution. Therefore, the final solution is A1 B2 C 2 D2 and f 0.48 . This
is somewhat different from the solution of (1) because the interactions are
not considered in (1). The solution of (1) can be spontaneously the same
as the one from the factorial design. However, it is not guaranteed. The
factorial design always gives the best solution. Generally, it is an
extremely expensive method in design.
Table 6.25. The factorial design of Example 6.9
Exp. no. A B C D f
1 – 0.5 (1) – 0.6 (1) – 0.9 (1) – 0.5 (1) 4.96
2 – 0.5 (1) – 0.6 (1) – 0.9 (1) 0.2 (2) 4.33
3 – 0.5 (1) – 0.6 (1) 1.0 (2) – 0.5 (1) 1.35
4 – 0.5 (1) – 0.6 (1) 1.0 (2) 0.2 (2) 0.72
5 – 0.5 (1) 0.4 (2) – 0.9 (1) – 0.5 (1) 3.76
6 – 0.5 (1) 0.4 (2) – 0.9 (1) 0.2 (2) 3.13
7 – 0.5 (1) 0.4 (2) 1.0 (2) – 0.5 (1) 0.15
8 – 0.5 (1) 0.4 (2) 1.0 (2) 0.2 (2) -0.48
9 0.3 (2) – 0.6 (1) – 0.9 (1) – 0.5 (1) 0.64
10 0.3 (2) – 0.6 (1) – 0.9 (1) 0.2 (2) 0.01
11 0.3 (2) – 0.6 (1) 1.0 (2) – 0.5 (1) 3.11
12 0.3 (2) – 0.6 (1) 1.0 (2) 0.2 (2) 2.48
13 0.3 (2) 0.4 (2) – 0.9 (1) – 0.5 (1) 1.04
14 0.3 (2) 0.4 (2) – 0.9 (1) 0.2 (2) 0.41
15 0.3 (2) 0.4 (2) 1.0 (2) – 0.5 (1) 3.51
16 0.3 (2) 0.4 (2) 1.0 (2) 0.2 (2) 2.88
6.4 Design with Orthogonal Arrays Considering

Interactions
When we design using orthogonal arrays, strong interactions make the problem
complicated. That is, the characteristic function depends on the combination of
factors. In other words, the (ab) ij term in Equation 6.24 is large. In Figure 6.2,
the interactions are schematically illustrated. Figure 6.2a has no interactions. The
cases in Figures 6.2b and 6.2c cannot ignore the interactions. In the previous
section, the statistical model was defined in Equation 6.40 without considering the
interactions. If we have the synergistic interaction as illustrated in Figure 6.2b, the
model is approximately close to the true trend. However, if we have the
antisynergistic interaction as illustrated in Figure 6.2c, the model may be far from
the true trend.
When a design using orthogonal arrays is performed, the following two aspects
should be kept in mind:
(1) The interactions between factors are identified.
(2) If an interaction is identified, it should be allocated to a column of an

orthogonal array as a factor.
Graphs in Figure 6.2 are utilized for the identification. Although interactions
do not exist in a specific range, they can appear in other ranges. It is quite
expensive to use the method in Figure 6.2 when there are many design variables.
Experience or engineering intuition is frequently utilized for the identification.
A method to briefly identify interactions is explained. Interactions are ignored
and factors are allocated to an orthogonal array. After the solution is obtained by
DOA, the following steps should be taken:
pred
(1) Evaluate the estimation of function f opt as in Equation 6.43 and perform
pred exp exp exp
a test to see if | f opt f opt | / | f opt | has a large value. f opt is the
function value at the optimum and it is calculated by a confirmation
experiment.
(2) Perform a test to see if a row of the orthogonal array gives a better solution
than the one from the one-way table.
(3) Perform a test to see if VE of the ANOVA table has a large value.
If one of the above three conditions is satisfied, the effect of interaction is large.
However, interactions can still exist even though all the conditions are not satisfied.
Therefore, the conditions are sufficient, but not necessary. Although the existence
of interactions can be verified, the factors interacting with each other should be
identified through the process shown in Figure 6.2. The identification of the
interactions is a complicated and expensive process. Therefore, if we use the
algorithm of the previous section, a good solution can be found in an engineering
sense, although it is not the best one.
As explained above, it is not easy to identify the interactions but they should be
identified before DOE is carried out. What if the interactions are identified
beforehand? In this case, they should be allocated to a column of the orthogonal
array. For example, we have four factors of two levels (A, B, C and D) and the
interaction between A and B. If we use the L8 (2 7 ) orthogonal array in Table 6.22,
A, B are allocated to the first two columns, respectively. The interaction A u B is
allocated to the third column. Moreover, C and D are allocated to the fourth and
fifth columns, respectively. No factors are allocated to the other columns. Details
about allocation are given in a reference (Phadke 1989).
Example 6.10 [Mathematical Problem for Identification of Interactions]

Examine if there are interactions in Example 6.9.
Solution
The results of Example 6.9 are shown in Table 6.26. The ANOVA table is
established as shown in Table 6.27. The aforementioned methods are applied.
(1) Using Equation 6.43
Column no. 1 2 3 4 5 6 7
Factor e A B C D e e
Level 1 2 1 2 1 2 1 2 1 2 1 2 1 2
4.96 3.13 4.96 1.04 4.96 1.04 4.96 0.72 4.96 0.72 4.96 0.72 4.96 0.72
0.72 0.15 0.72 2.88 0.72 2.88 1.04 2.88 1.04 2.88 2.88 1.04 2.88 1.04
1.04 0.01 3.13 0.01 0.01 3.13 3.13 0.15 0.15 3.13 3.13 0.15 0.15 3.13
2.88 3.11 0.15 3.11 3.11 0.15 0.01 3.11 3.11 0.01 3.11 0.01 0.01 3.11
(1) Total 9.6 6.4 8.96 7.04 8.8 7.2 9.14 6.86 9.26 6.74 14.08 1.92 8 8
(2) Sum of
16 16 16 16 16 16 16
(1)
(3)
Difference – 3.2 – 1.92 – 1.6 – 2.28 – 2.52 – 12.16 0
of (1)
(4) (3)2 10.24 3.69 2.56 5.20 6.35 147.87 0
(5) (4)/8 1.28 0.46 0.32 0.65 0.79 18.48 0
pred
f opt m (m A2 m) (mB2 m) (mC 2 m) (mD2 m)
0.96 (6.54a)
exp
f opt 2.88 (6.54b)
pred exp exp

| f opt f opt | / | f opt | = 0.6667 = 66.67% (6.54c)
The difference in Equation 6.54 is large. Therefore, there are interactions.

(2) In Example 6.9, the objective function in the 7th row of the orthogonal
array is better than the solution from the one-way table. Therefore, there
are interactions.
(3) From Table 6.27, VE 6.59 is larger than the other V’s. Therefore, there
are interactions.
Table 6.27. ANOVA table from Table 6.26
Factor S I V F
A 0.46 1 0.46 0.07
B 0.32 1 0.32 0.05
C 0.65 1 0.65 0.1
D 0.79 1 0.79 0.12
Error 19.76 3 6.59
T 21.98 7
Figure 6.4 shows the graphs for this problem for the interactions. From Figure
6.4, we can identify that the interactions A u B and A u C are large.
f
C = – 0.9
5 D = – 0.5
B = 0.4
B = – 0.6
– 0.5 0.3 A
(a) Interaction between A and B
f
B = – 0.6
5 D = – 0.5
C = 1.0
C = –0.9
– 0.5 0.3 A
(b) Interaction between A and C
f
A = – 0.5
5 D = – 0.5
C = – 0.9
C = 1.0
– 0.6 0.4 B
(c) Interaction between B and C
Figure 6.4. Identification of interactions for Example 6.10

Example 6.11 [Solving Example 6.10 by Considering Interactions]

Solve Example 6.10 with the L8 (2 7 ) orthogonal array by considering the
interactions.
Solution
In Example 6.10, the interactions A u B and Au C are identified to be large. The
factors A, B, C, and D and interactions are allocated as shown in the last row of
Table 6.22. The seventh column of Table 6.22 is empty because it is allocated for
error. The objective function for each row is shown in Table 6.28. The results are
summarized in Table 6.29 and the ANOVA table is shown in Table 6.30.
In Table 6.30 the sums of squares due to interactions A u B and A u C are
relatively large. In this case, a two-way table should be employed to find the
solution. A two-way table is shown in Table 6.31a and b. Table 6.31a is utilized
because the interaction between A and B is large. The values corresponding to
A1 B1 are made from the function values when A and B have the first levels in
Table 6.22. In the first and the second rows of Table 6.22, A and B have the first
levels. Thus, the function values of the first and second rows are added, which is
shown in Table 6.31a. The values of other conditions A1 B2 , A2 B1 and A2 B2 are
evaluated in the same manner. The smallest one is chosen out of the four values in
Table 6.28. The objective function of Example 6.11 for each row of the L8 (2 7 ) orthogonal
array
f1 f2 f3 f4 f5 f6 f7 f8
4.96 0.72 3.13 0.15 0.64 2.48 0.41 3.51
Table 6.29. Summary of the results for Example 6.11
Column no. 1 2 3 4 5 6 7
Factor A B A uB C A uC D e
Level 1 2 1 2 1 2 1 2 1 2 1 2 1 2
4.96 0.64 4.96 3.13 4.96 3.13 4.96 0.72 4.96 0.72 4.96 0.72 4.96 0.72
0.72 2.48 0.72 0.15 0.72 0.15 3.13 0.15 3.13 0.15 0.15 3.13 0.15 3.13
3.13 0.41 0.64 0.41 0.41 0.64 0.64 2.48 2.48 0.64 0.64 2.48 2.48 0.64
0.15 3.51 2.48 3.51 3.51 2.48 0.41 3.51 3.51 0.41 3.51 0.41 0.41 3.51
(1) Total 8.96 7.04 8.8 7.2 9.6 6.4 9.14 6.86 14.08 1.92 9.26 6.74 8 8
(2) Sum of
16 16 16 16 16 16 16
(1)
(3)
Difference – 1.92 – 1.6 – 3.2 – 2.28 – 12.16 – 2.52 0
of (1)
(4) (3)2 3.69 2.56 10.24 5.20 147.87 6.35 0
(5) (4)/8 0.46 0.32 1.28 0.65 18.48 0.79 0

Table 6.30. ANOVA table for Example 6.29
Factor S I V
A 0.46 1 0.46
B 0.32 1 0.32
Au B 1.28 1 1.28
C 0.65 1 0.65
Au C 18.48 1 18.48
D 0.79 1 0.79
Error 0 1
T 21.98 7
Table 6.31a. In this case, A2 B1 is the solution. The selection of A and C is shown
in Table 6.31b. The solution is A1C2 . The solution for D is the second level from
the one-way table in Table 6.31c, since there are no interactions due to D.
When a factor has interactions with the other two factors, two two-way tables
may give different solutions. Then a one-way table is made for A. From the one-
way table of Example 6.9, the solution of A is A2 . Therefore, the design solution
is A2 B1C2 D2 . In this problem, the interaction of A u C is much greater than the
interaction of A u B from Table 6.30. From Table 6.31b, A1C2 is determined as
the design solution. When A A1 , the good solution from Table 6.31a is A1B2 .
Thus, the final solution is A1B2C2 D2 .
The objective function value is estimated by using the solution determined
above. Using Equation 6.43, the estimated value is
pred
f opt m (m A1B2 m) (m A1C2 m) (m D2 m) 0.24 (6.55)
By using the additive model in Equation 6.43, the components of Equation 6.55
can be expressed as
m A1B2 m a1 b2 (ab)12 (6.56a)
m A1C2 m a1 c2 (ac)12 (6.56b)
mD2 m d2 (6.56c)
Table 6.31. One-way and two-way tables for Example 6.11
(a) Two-way table for factors A and B
A1 A2
B1 f1 f 2 5.68 f5 f6 3.12
B2 f3 f4 3.28 f 7 f8 3.92
(b) Two-way table for factors A and C
A1 A2
C1 f1 f 3 8.09 f5 f7 1.05
C2 f2 f4 0.87 f 6 f8 5.99
(c) One-way table for factor D
D1 D2
f1 f 4 f 5 f 8 9.26 f2 f3 f6 f7 6.74
Thus, a1 is included twice in Equation 6.55. The precise expression of Equation

6.55 is
pred
f opt m (m A1 B2 m) (m A1C 2 m) (mD2 m) a1 0.48 (6.57)
a1 in Equation 6.57 is calculated from (1) of the first column in Table 6.29 as
follows:
(8.96 7.04) / 4
a1 0.24 (6.58)
2
exp
The value of the confirmation experiment is f opt 0.48 . Since this problem
does not have errors, Equation 6.57 predicts the solution precisely.
Example 6.12 [Identification of Interactions for a Five Bar Truss]

A five bar truss is presented in Figure 6.5. An optimization problem is formulated
to maximize the natural frequency of the first mode,
9.144m
A
D
C 9.144m
Figure 6.5. Five bar truss
Table 6.32. Factors and levels of Example 6.12
4
Factor ( u 10 m2 )
Level
A B C D E
1 3.00 8.00 10.10 10.00 6.00
2 17.00 17.00 11.00 25.00 18.00
3 30.00 26.00 20.00 40.00 30.00
Maximize [ (6.59)
where [ is the natural frequency of the first mode. Factors (design variables) are
the areas of the cross sections. The discrete values of the factors are shown in
Table 6.32. Identify the interactions between the factors.
Solution
Since we have five factors with three levels, the orthogonal array L18 in Table
6.33 is chosen. The interactions are evenly distributed in L18 . Thus, it is
frequently chosen when DOE is performed regardless of the interactions. All the
factors are allocated to columns of L18 . The results are shown in Tables 6.33 and
6.34. The ANOVA table is shown in Table 6.34. The mean for the levels of each
factor is shown in Table 6.34 and it is the same as the solution from the one-way
exp
table. The solution is A3 B3C 2 D2 E3 from Table 6.34 and [ opt 42.806. The
methods mentioned before can be used to identify interactions. The mean square
is employed for this problem and VE 6.5563 30.8026 14.6699 52.0288.
The mean square is somewhat large. However, since the interactions are evenly
exp
distributed in L18 , the solution is considered as the final one. It is noted that [ opt
Table 6.33. Experiments of Example 6.12 using the L18 ( 21 u 3 7 ) orthogonal array
Column and factor
Exp. No. 1 2 3 4 5 6 7 8 [ (Hz)

e A B C D E e e
1 1 1 1 1 1 1 1 1 34.40398
2 1 1 2 2 2 2 2 2 34.80376
3 1 1 3 3 3 3 3 3 31.75859
4 1 2 1 1 2 2 3 3 38.07272
5 1 2 2 2 3 3 1 1 40.39151
6 1 2 3 3 1 1 2 2 29.83928
7 1 3 1 2 1 3 2 3 41.13371
8 1 3 2 3 2 1 3 1 32.97368
9 1 3 3 1 3 2 1 2 42.08462
10 2 1 1 3 3 2 2 1 27.82737
11 2 1 2 1 1 3 3 2 28.18135
12 2 1 3 2 2 1 1 3 39.06658
13 2 2 1 2 3 1 3 2 31.56198
14 2 2 2 3 1 2 1 3 34.98063
15 2 2 3 1 2 3 2 1 42.04833
16 2 3 1 3 2 3 1 2 36.60479
17 2 3 2 1 3 1 2 3 36.44593
18 2 3 3 2 1 2 3 1 37.88146
Table 6.34. Mean and analysis of the variance of Example 6.12
Column Mean for each level

Factor DOF Sum of squares Mean square
number 1 2 3
1 error 36.1624 34.9554 1 6.5563 6.5563
2 A 32.6736 36.1491 37.8540 2 83.6452 41.8226
3 B 34.9341 34.6295 37.1131 2 22.0193 11.0097
4 C 36.8728 37.4732 32.3307 2 94.8716 47.4358
5 D 34.4034 37.2616 35.0117 2 27.2039 13.6019
6 E 34.0486 35.9418 36.6864 2 22.1933 11.0967
7 error 37.9220 35.3497 33.4050 2 61.6051 30.8026
8 error 35.9211 33.8460 36.9097 2 29.3397 14.6699
is better than any row in Table 6.33. Therefore, a good solution is obtained
without considering interactions, although they exist.
Example 6.13 [Design Considering Interactions in Example 6.12]

Design the problem of Example 6.12 by considering interactions.
Solution
First, the interactions should be identified. It is found that interactions A u
B, A u C , A u D and A u E exist. Thus, the orthogonal array L27 (313 ) in Table
6.35 is chosen. The characteristic function for each row is shown in Table 6.35.
The mean and the sum of squares for each row are shown in Table 6.36. Since the
interaction A u E is large in Table 6.36, a two-way table is made for A and E as
shown in Table 6.37. One-way tables are made for the other factors. The solution
Table 6.35. Experiments with the L27(313) orthogonal array in Example 6.13
Column and factor

Exp. no. 1 2 3 4 5 6 7 8 9 10 11 12 13 [ (Hz)
A E AE AE2 D AD AD2 B AB AB2 C AC AC2
1 1 1 1 1 1 1 1 1 1 1 1 1 1 34.4040
2 1 1 1 1 2 2 2 2 2 2 2 2 2 38.1740
3 1 1 1 1 3 3 3 3 3 3 3 3 3 36.6973
4 1 2 2 2 1 1 1 2 2 2 3 3 3 27.6969
5 1 2 2 2 2 2 2 3 3 3 1 1 1 35.8637
6 1 2 2 2 3 3 3 1 1 1 2 2 2 30.1594
7 1 3 3 3 1 1 1 3 3 3 2 2 2 27.6409
8 1 3 3 3 2 2 2 1 1 1 3 3 3 26.6787
9 1 3 3 3 3 3 3 2 2 2 1 1 1 32.2225
10 2 1 2 3 1 2 3 1 2 3 1 2 3 34.6604
11 2 1 2 3 2 3 1 2 3 1 2 3 1 38.0122
12 2 1 2 3 3 1 2 3 1 2 3 1 2 36.7979
13 2 2 3 1 1 2 3 2 3 1 3 1 2 34.9806
14 2 2 3 1 2 3 1 3 1 2 1 2 3 42.0869
15 2 2 3 1 3 1 2 1 2 3 2 3 1 34.7767
16 2 3 1 2 1 2 3 3 1 2 2 3 1 38.6751
17 2 3 1 2 2 3 1 1 2 3 3 1 2 35.4017
18 2 3 1 2 3 1 2 2 3 1 1 2 3 40.6838
19 3 1 3 2 1 3 2 1 3 2 1 3 2 32.4918
20 3 1 3 2 2 1 3 2 1 3 2 1 3 36.1618
21 3 1 3 2 3 2 1 3 2 1 3 2 1 35.5968
22 3 2 1 3 1 3 2 2 1 3 3 2 1 35.0717
23 3 2 1 3 2 1 3 3 2 1 1 3 2 41.6593
24 3 2 1 3 3 2 1 1 3 2 2 1 3 34.7589
25 3 3 2 1 1 3 2 3 2 1 2 1 3 40.1213
26 3 3 2 1 2 1 3 1 3 2 3 2 1 36.6048
27 3 3 2 1 3 2 1 2 1 3 1 3 2 41.5861
Table 6.36. The mean and the analysis of variance for Example 6.13
Exp. Mean for each level Sum of

Factor Mean DOF
no. squares
1 2 3
1 A 32.1708 37.3417 37.1169 35.5432 2 153.7579
2 E 35.8885 35.2282 35.5128 35.5432 2 1.9741
3, 4 Au E 4 125.7039
5 D 33.9714 36.7381 35.9199 35.5432 2 36.3626
6, 7 Au D 4 3.3840
8 B 33.3263 36.0655 37.2377 35.5432 2 72.5301
9, 10 Au B 4 1.9863
11 C 37.2954 35.3867 33.9474 35.5432 2 50.7718
12, 13 Au C 4 0.7570
Table 6.37. Two-way table for Example 6.13
Level of factor A
A1 A2 A3
E1 36.4251 36.4902 34.7501
Level of
factor‫ٻ‬E‫ٻ‬
E2 31.2400 37.2814 37.1633
E3 28.8474 38.2535 39.4374
Table 6.38. Comparison of the results of Example 6.13
Factor A B C D E [ (Hz)
Factorial 3 3 1 3 3 43.315
Solu-
L18 3 3 2 2 3 42.806
tion
L27 3 3 1 2 3 43.140
exp
is A3 B3C1 D2 E3 and [ opt 43.140. In Table 6.38, the results from L18 and L27
are compared with the results from the factorial experiments. In an engineering
sense, the two results from the orthogonal arrays are sufficiently excellent as
compared to the factorial solution. Although the interactions are ignored, a good
solution can be obtained by using the orthogonal L18 where the interactions are
evenly distributed.
6.5 Design Using Orthogonal Arrays for Constrained

Problems
So far, unconstrained problems have been considered. In other words, constraints
have been excluded. Constraints do not usually exist in experiments. In design,
constraints in Equation 3.1 generally exist. When we have constraints, they are
indirectly considered. Constraints must be satisfied before the objective function
is minimized. The previous methods are not sufficient to solve constrained
problems. For a constrained problem, a characteristic function < is defined as
follows (Lee and Park 2002):
< (b ) f (b) P (b) (6.60)
where b is the design variable vector and f is the objective function. Using the
notation of Equation 3.1, P is the penalty function defined by the maximum
violation of the constraints,
P (b) max(0, | h |, g ) u s (6.61)
Equation 6.61 is obtained by multiplication of the scale factor s and the maximum
violation. The influence of the constraints depends on s. It is recommended that
the constraints be normalized.
The characteristic function < in Equation 6.60 is equivalent to K in Section
6.3.3. Therefore, the process of Section 6.3.3 can be directly utilized with <.
Some constraints may represent the overall characteristics of a system. They are
usually functions of most design variables. It is noted that the interactions become
important in these cases.
The above method is used to find an optimum from ANOM of <. The
solution from an orthogonal array (1) may not be a good one when the interactions
are ignored or (2) may violate constraints. To overcome (1), after a confirmation
experiment is performed, the best one is picked from the rows of the orthogonal
array and the confirmation experiment. This method does not provide the best
solution. However, a good solution can be obtained with low cost. Therefore, the
method in Section 6.3.3 is utilized for a constrained problem. However, we still
have the possibility of a constraint violation. It should be noted that each row of
the orthogonal array can be evaluated with a different constraint because only the
maximum constraint violation is considered. Therefore, there is a strong
possibility that the solution of Section 6.3.3 violates the constraints.
To overcome this difficulty, two methods are proposed as follows:
(1) A method that directly uses an orthogonal array
The method in Section 6.3.3 is directly used and the constraint violation is
checked. First, the cases satisfying the constraints are picked from the
confirmation experiment and the rows of the orthogonal array. Second, the
one with the best < is picked as the final solution. In other words, K is
replaced by < in Table 6.13, <i is obtained for each row, the
confirmation experiment and the rows of the orthogonal array are
compared, and the best one is picked from those satisfying the constraints.
(2) A method to arrange the estimated values in ascending order
In Equation 6.43, the estimated value is obtained when the interactions are
ignored. The estimation is performed for the full combination. That is,
< pred is estimated for a full combination. These are approximated values
calculated from the matrix experiment with an orthogonal array. The
objective function and the penalty term for constraints are estimated,
respectively. < pred is derived by addition of the estimated objective
function and the penalty term. The full combination of < pred is arranged
in ascending order for the magnitude of < pred . An example with the
L9 (3 4 ) orthogonal array is shown in Table 6.39. The experiment number
in Table 6.39 is the ranking for the magnitude of < pred . That is,
<ipred <ipred
1 , i 1, 2, ... , 80. Confirmation experiments are performed
for the cases of Table 6.39. The first experiment that satisfies the
constraints is picked as a candidate design. In Table 6.39, the fifth one is
the candidate design. The candidate design is compared with all the rows
of the orthogonal array. The final design is the best one out of the
Table 6.39. Arrangement of experiments in ascending order for the L9 (3 4 ) orthogonal

array
Factor and level Estimation Constraint

Exp. no.
A B C D of < violation
1 <1 Violated
2 <2 Violated
3 <3 Violated
4 <4 Violated
5 <5 Satisfied
The ith row for each factor is the level of
6 the factor <6 Violated
. . .
. . .
. . .
81=34 <81 .
candidate designs and the rows of the orthogonal array.

When the values are estimated, the characteristic function in Equation 6.60 can
be directly calculated. It is simple and inexpensive compared to the method where
the objective function and constraints are estimated separately. However, the
estimated value may be considerably different from the actual value because the
mean in Equation 6.37 is evaluated as the objective function and a multiplication
of the scale factor and constraint violation. Therefore, the ranking in the ascending
order can be different. This method can be exploited if a precise design is not
required.
The above two methods do not come up with a mathematical optimum.
However, they are practical methods in that a discrete solution is obtained with
low cost.
Example 6.14 [A Mathematical Problem with Constraints]

Find b1 , b2 , b3 , b4 (6.62a)
to minimize f (b) b12 b2 2 2b3 2 b4 2 5b1 5b2 21b3 7b4 (6.62b)
b1 2 b2 2 b3 2 b4 2 b1 b2 b3 b4
subject to 1 d 0 (6.62c)
8
b1 2 2b2 2 b3 2 2b4 2 b1 b4
1 d 0 (6.62d)
10
2b1 2 b2 2 b3 2 2b1 b2 b4
1 d 0 (6.62e)
5
Each design variable has three levels of –1, 0 and 1.

(1) Obtain a design solution with only an orthogonal array.
(2) Obtain a design solution by arranging the estimated values in ascending
order.
Solution
(1) Using Equation 6.60, < is defined. The scale factor of Equation 6.61 is
arbitrarily defined as 100. Since the problem has four design variables
with three levels, the L9 (34 ) orthogonal array is chosen. The experiments
are performed as shown in Table 6.40. The seventh row violates the
constraints and the other rows satisfy all the constraints. Based on the
results in Table 6.40 the one-way table for < is made as shown in Table
6.41. The design solution from Table 6.41 is (b1 ) 2 (b2 )3 (b3 )3 (b4 )1 . If we
substitute the design solution into Equation 6.62 for the confirmation
Table 6.40. Experiments with the L9 (3 4 ) orthogonal array for Example 6.14
Design variable and level

Exp. no. 1 2 3 4 <
b1 b2 b3 b4
1 1 1 1 1 29 ( <1 )
2 1 2 2 2 6 ( <2 )
3 1 3 3 3 –9 ( <3 )
4 2 1 2 3 14 ( <4 )
5 2 2 3 1 –25 ( <5 )
6 2 3 1 2 19 ( <6 )
7 3 1 3 2 23 ( <7 )
8 3 2 1 3 27 ( <8 )
9 3 3 2 1 –14 ( <9 )
Table 6.41. One-way table for Example 6.14
Level
Design
variable
1 2 3
b1 8.67 2.67 12
b2 22 2.67 –1.33
b3 25 2 –3.67
b4 –3.33 16 10.67
experiment, the objective function is 29 and all the constraints are

satisfied. This solution is compared with all the rows of Table 6.40. The
final solution is (b1 ) 2 (b2 ) 3 (b3 ) 3 (b4 )1.
(2) Using Equation 6.43, the estimated values of < for all combinations are
calculated. The number of estimations is 34 = 81. Table 6.42 shows the
arrangement in ascending order. In Table 6.42, the first design satisfying
Table 6.42. Arrangement of estimated values in ascending order for Example 6.14
Design variable and level Estimation Constraint

Ranking
b1 b2 b3 b4 of < violation
1 2 3 3 1 – 29 Satisfied
2 3 3 3 2 – 27
3 2 2 3 1 – 25
4 1 3 3 1 – 23
. . . . . .
. . . . . .
. . . . . .
78 3 1 1 3 43
79 3 1 1 3 53
80 3 1 1 2 65
81=34 3 1 1 1 79
the constraints is (b1 ) 2 (b2 )3 (b3 ) 3 (b4 )1 and the objective function is – 29. It
is compared with the fifth row of Table 6.40. The final solution is the
same as that of (1).
Example 6.15 [Structural Design Problem]

In Figure 3.3, l = 1 m, T 60q and the load is P = 40,000 N. Determine the areas
b1 , b2 and b3 to minimize the mass m. Young’s modulus (E) is 207 GPa and the
density ( U ) is 2770 kg/m3. The problem formulation is as follows:
Find b1 , b2 , b3 (6.63a)
to minimize m(b1 , b2 , b3 ) ( 2b1 b2 2b3 ) u U (6.63b)
subject to 100 MPa d V i d 140 MPa (i 1, 2, 3) (6.63c)
where V i (i = 1,2,3) is the stress of the ith member. The design variables can have
the levels in Table 6.43.
(1) Find a solution by using an orthogonal array.
(2) Find a solution by arranging the estimated values in ascending order.
Table 6.43. Levels of design variables for Example 6.15 (unit: m2)
Design variable
Level
b1 b2 b3
1 0.00018 0.0001 0.00002
2 0.0002 0.00012 0.00004
3 0.00022 0.00014 0.00006
Table 6.44. Results from the orthogonal array of Example 6.15
Design variable and level

Exp. no. 1 2 3 4 <
b1 b2 Ignored b3
1 1 1 1 1 12.689
2 1 2 2 2 10.141
3 1 3 3 3 7.115
4 2 1 2 3 4.467
5 2 2 3 1 1.194
6 2 3 1 2 1.328
7 3 1 3 2 1.748
8 3 2 1 3 1.429
9 3 3 2 1 1.328
Solution
(1) The number of design variables is three and the number of levels for each
design variable is three. Thus, the L9 (3 4 ) orthogonal array is selected.
The results of the experiments are shown in Table 6.44. The scale factor is
set as s = 100 and the process of Example 6.14 is performed. The results
from the one-way table for < in Equation 6.60 are shown in Table 6.45.
From Table 6.45, the design solution is (b1 )3 (b2 )3 (b3 )3 . The objective
function is 1.485 from the confirmation experiment and the constraints are
satisfied. Comparing the results in Table 6.44 and the solution from the
one-way table, the final solution is (b1 ) 2 (b2 ) 2 (b3 )1 , which is the fifth row
of the orthogonal array. Therefore, the design variables are changed from
b [0.0002 0.00012 0.00004]T to b [0.0002 0.00012 0.00002]T
and the mass is reduced from 1.273 kg to 1.194 kg.
Table 6.45. The one-way table for Example 6.15
Level
Design
variable
1 2 3
b1 9.982 2.330 1.502
b2 6.301 4.255 3.257
b3 5.070 4.406 4.337
Table 6.46. Arrangement of the estimated values of Example 6.15
Design variable and level Estimation of Constraint

Ranking
b1 b2 b3 < violation
1 3 1 1 1.2172 Violated
2 2 3 1 1.2496 Satisfied
. . . . . .
. . . . . .
(2) The estimated values for < are arranged in ascending order as shown in
Table 6.46. In Table 6.46, b [0.0002 0.00014 0.00002]T is the first
one that satisfies the constraints. Therefore, the design solution is (b1 ) 2
(b2 ) 3 (b3 )1 and the objective function is 1.2496. However, the solution of
(1) is better. Therefore, the design solution of the fifth row of the
orthogonal array is selected as the final solution.
6.6 Sequential Algorithm with Orthogonal Arrays

(SOA)
The aforementioned method using orthogonal arrays sets up an orthogonal array
around a specific point. SOA is an extension of the method and the final design is
found in an iterative fashion. Some researchers define an orthogonal array in a
wide range and narrow the design range (Ku et al. 1998). SOA defines levels in a
narrow range of the initial design and applies the previous method. If a design is
found using an orthogonal array locally, an orthogonal array is set up again around
the new design. The iterative process continues until the convergence criteria are
satisfied. The overall flow is illustrated in Figure 6.6 and the steps are as follows
(Lee et al. 2003):
Step 1. (Definition of the problem) The objective function and design
variables are defined. Candidate values are defined in a discrete space.
The candidate values are the values that the design variables can have.
When we define the initial design b (0) , the existing design is usually used
for the initial design to improve the existing design. The iteration number
is set to k 0 and an appropriate orthogonal array is selected. The
number of levels is usually defined as three because three levels easily
include a larger value, a smaller value and the curvature. A larger number
Start
Selection of discrete values for design variables
Selection of an orthogonal array, k = 0
Start the kth iteration

Selection of the levels from the candidate values
First level: The initial value
Second and third levels: Neighboring values of the first level
Matrix experiment
Calculation of the characteristic function
k=k+1
ANOM for the characteristic function
Selection of the initial solution
Comparison of the initial solution and

the best one from the orthogonal array
No
Convergence criteria satisfied?
Yes
End
Figure 6.6. SOA flow chart

Design variable & Optimum or Levels in the next iteration

candidate value next initial
Design variable
design values Level
A1 > A2 > A3 > A4 > A5 A3, B2, C4 A B C
1 A3 B2 C4
B1 > B2 > B3 > B4 > B5
2 A2 B1 C3
C1> C2 > C3 > C4 > C5
3 A4 B3 C5
Figure 6.7. Allocation of levels of design variables
of levels is not recommended because the number of function calculations

is large. The orthogonal array is selected based on the numbers of design
variables and levels.
Step 2. (Determination of the values for levels) Discrete values of design
variables are allocated to the levels of the orthogonal array. The values are
chosen from the candidate values. Each component of b ( k ) is allocated to
the first level and the neighboring values are allocated to the second and
the third levels. When a component of b ( k ) is in the upper or lower limits,
the current value is set to the first level and the neighboring two values are
allocated to the remaining levels. An example of the allocation is
presented in Figure 6.7. The reason why the current design is allocated to
the first level is that the current design should be always included in the
matrix experiment. In this case, SOA can always find an improved design
at each iteration.
Step 3. (Matrix experiment) The objective function and constraints are
evaluated for each row of the orthogonal array. The methods in Sections
6.3.3 and 6.5 are utilized for unconstrained and constrained problems,
respectively. The iteration number is set as k k 1 and the design is
updated to b ( k ) .
Step 4. (Check the convergence) When the design variables are not changed
the process terminates. Otherwise, go to Step 2.
Example 6.16 [A Mathematical Problem Using SOA]

Obtain the design solution of Example 6.14 by using SOA. The candidate values
of the design variables are integers of í10, í9, í8, í7, …, 7, 8, 9 and 10.
Solution
In the first iteration, b [0 1 1 1]T , which is the solution of Example 6.14, is

allocated to the first levels and the surrounding values are allocated to the second
and third levels. The iterative process is performed until the convergence criteria
are satisfied. The process terminates after two iterations. The number of total
Table 6.47. Levels of design variables at the first iteration of Example 6.17 (unit: m2)
Design variable
Level
b1 b2 b3
1 0.00026 0.0001 0.00014
2 0.00024 0.00008 0.00012
3 0.00028 0.00012 0.00016
function calculations is 19. The final solution is b [0 0 2 0]T and the

objective function is í34.
Example 6.17 [A Structural Design Problem Using SOA]

Find a design solution of Example 6.15 by using the SOA. The initial levels are
defined as shown in Table 6.47 and the candidate values are defined as the
following set:
b1 , b2 , b3 (u10 5 m 2 ) {2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24,
26, 28, 30, 32, 34, 36, 38, 40, 42, 44, 46, 48, 50}
Solution
For the first iteration, the levels in Table 6.47 are utilized. The process terminates
after six iterations. The total number of function calculations is 42. The final
design is b [0.0002 0.00012 0.00004]T . The mass is reduced by 31% from
1.844 to 1.273.
6.7 Design Using the Response Surface Method
6.7.1 Introduction
Suppose a design related function f is a function of design variable vector b as

follows:
f f (b) (6.64)
The function f can be considered as a response function with respect to the

design variable vector b . If the number of components in b is two, the function
can be expressed by the graph in Figure 6.8. The plane in Figure 6.8 is a response
surface. As mentioned earlier, a number of function calculations are needed in
optimization and the function calculation is generally quite expensive in practical
Figure 6.8. An example of a response surface in theGthree-dimension
problems. In particular, sensitivity analysis is extremely expensive. Therefore, the

function f is frequently approximated to an explicit function as in Equation 6.3 to
overcome this difficulty. That is, a metamodel is used and the cost for function
evaluation can be considerably reduced. The response surface method (RSM) is
an approximation method (Myers and Montgomery 1995).
RSM is usually utilized in statistics. Suppose it is impossible to express the
function f in Figure 6.9 in an explicit form. To approximate it as an explicit
function, the function value is calculated at x1 , x2 and x3 of the independent
variable x. The function f can be approximated to f1 , f 2 , f 3 , etc. in Figure 6.9
by using different methods. The real response surface is f . However, an
approximated function is used in the design. It is noted that the function f is
approximated by three function calculations. Since the approximated function is
used in the optimization process and it has an explicit form, the effort for function
f(x)
f(x)
f1
f2
f3
x1 x2 x3 x
Figure 6.9. Examples of approximation for a response surface

Selection of the candidate points to generate the response surfaces
Selection of a method to generate the response surfaces
Is the response surface

No sufficiently precise?
Yes
Optimization with the response surfaces
Is the optimization satisfied?

No
Yes
End
Figure 6.10. Flow chart of RSM in the optimization process
calculation and sensitivity analysis is negligible.

RSM is utilized to give various statistical data in statistics. In design, the
approximated functions are used in optimization. RSM is explained here for use in
the optimization process. First, the following aspects should be determined:
(1) Selection of the candidate points to generate the response surface.
(2) The method to generate the response surface with the candidate points.
(3) The optimization method to use with the response surface.
The application of RSM is illustrated in Figure 6.10.
6.7.2 Generation of the Response Surface
A method to generate the response surface is explained. In the flow chart of

Figure 6.10, the selection method for candidate points should be explained first.
However, generation of the response surface is investigated first. Suppose we
have k candidate points for the design variable vector b R n as shown in Table
6.48. A function f is evaluated at the candidate points. Table 6.48 shows
f i , i 1,..., k for bij , i 1,..., k , j 1,..., n. bij is the value of the jth design
variable to calculate f i . Generally, the function f is approximated by a linear or
quadratic equation. The function can be approximated by a cubic, a higher
polynomial or a special function; however, they are not usually used due to cost.
An approximated function in a quadratic equation is as follows:
Table 6.48. The function value of f with respect to k candidates
[1 [2 … [n [ n 1 [ n2 … [l
f
b1 b2 … bn b12 b22 … bn2 b1b2 … bn 1bn
2 2
f1 b11 b12 b1n b11 b12 b12n b11b12 b1, n 1b1n
2 2
f2 b21 b22 b2 n b21 b22 b22n b21b22 b2,n 1b2 n

fk bk1 bk 2 bkn bk21 bk22 2
bkn b k 1b k 2 bk ,n 1bkn
f a 0 a1b1 a 2 b2 ... a n bn a11b12 ... a nn bn2
a12 b1b2 ... a n 1, n bn 1bn H (6.65)
where H is the error in the approximation process. In Equation 6.65, the linear
and quadratic terms are treated in the same manner. l n(n 1) / 2 n is set and
the coefficients are substituted by a0 o c0 , a1 o c1 , ..., a11 o cn 1 ,..., an 1, n o cl .
As shown in Table 6.48, the variables are substituted by b1 o [1 , ..., b12 o [ n 1 , ...,
bn 1bn o [ l . Then Equation 6.65 yields
f c0 c1[1 c2[ 2 ... cl[ l H (6.66)
An approximation function ŷ is defined as
f # yˆ c0 c1[1 c2[ 2 ... cl[ l (6.67)
Table 6.49. Transformed data from Table 6.48
f x1 x2 x3 … xl
f1 x11 x12 x13 … x1l
f2 x 21 x 22 x 23 … x 2l
…
fk x k1 xk 2 xk 3 … x kl
bij in Table 6.48 is replaced by [ ij . To make the maximum and minimum of

each [ i 1 and í1, respectively, [ ij is transformed as
[ ij (max ([ mj ) min ([ mj )) / 2
m m
xij (6.68)
(max ([ mj ) min ([ mj )) / 2
m m
The data in Table 6.48 are shown in Table 6.49 with xij and Equation 6.67 is
modified to
f i # yˆ i c0 c1 xi1 c2 xi 2 ... cl xil , i 1,..., k (6.69)
fi yˆ i H i (6.70)
To determine ŷ , L is defined by the data in Table 6.49 and the error at each
candidate point ( H i ) as follows:
k k k l
2 2 2
L ¦Hi ¦ ( f i yˆ i ) ¦ ( f i (c0 ¦ c j xij )) (6.71)
i 1 i 1 i 1 j 1
Coefficients c 0 , c1 ,..., cl are determined to minimize L. The condition is
wL
0, i 0,1, 2, ... , l (6.72)
wci
The condition in Equation 6.72 results in a linear simultaneous equation as

follows:
ª k k k º ª k º
« k ¦ xi1 ¦ xi 2 ¦ xil » « i¦1 f i »
i 1 i 1 i 1 ªc º
«k k k k »« 0 » «k »
« ¦ xi1 2
¦ xi1 ¦ xi1 xi 2 ¦ xi1 xil » « c1 » « ¦ xi1 f i »
(6.73)
«i 1 i 1 i 1 i 1 »« » «i 1 »
« »« » « »
«k k k k
2 » c «k »
« ¦ xil ¦ xil xi1 ¦ xil xi 2 ¦ xil » ¬ l ¼ « ¦ xil f i »
¬i 1 i 1 i 1 i 1 ¼ ¬i 1 ¼
The coefficient vector c [c 0 c1 ... cl ]T can be obtained from Equation 6.73.

The process is exactly the same as the least squares method. The amount of
calculation depends on how many terms are considered in Equation 6.65. If l = n,
only the linear terms in Equation 6.65 are considered, and if l 2n the linear
terms and perfect square terms are evaluated. If l n(n 1) / 2 n, all the terms
are included in the approximation. When l is large, more terms are generally
considered and the cost is higher. Therefore, the user should appropriately
determine l according to the environment.
Example 6.18 [A Mathematical Problem for Generation of the

Response Surface]
The following function is to be approximated:
f sin (b1 b2 ) (b1 b2 ) 2 1.5b1 2.5b2 1 (6.74)
Define candidate points with the L9 (3 4 ) orthogonal array by using b1 1.5,

1.25, 4 and b2 3, 0, 3 . Approximate the function by RSM.
Solution
The values for b1 and b2 are allocated to the orthogonal array in Table 6.50. x in
Table 6.50 is the normalized value and the function f at the candidate points is
shown in Table 6.50. Because we have two design variables with three levels, the
experiment with the L9 (34 ) orthogonal array is the same as the factorial
experiment. The approximation function is defined as
f c 0 c1 x1 c 2 x 2 c3 x12 c 4 x 22 c5 x1 x 2 H (6.75a)
f # yˆ c 0 c1 x1 c 2 x 2 c3 x12 c 4 x 22 c5 x1 x 2 (6.75b)
From Table 6.50 and Equation 6.73,
Table 6.50. Allocation of candidates and function calculation of Example 6.18
Observation b1 ( x1 ) b2 ( x 2 ) f
1 –1.5 (–1) –3 (–1) –1.022
2 –1.5 (–1) 0 (0) 4.503
3 –1.5 (–1) 3 (1) 31.997
4 1.25 (0) –3 (–1) 8.704
5 1.25 (0) 0 (0) 1.636
6 1.25 (0) 3 (1) 8.793
7 4 (1) –3 (–1) 37.341
8 4 (1) 0 (0) 10.243
9 4 (1) 3 (1) 4.157
ª9 0 0 6 6 0 º ªc0 º ª 106.35169 º
«0 « »
« 6 0 0 0 0»» « c1 » « 16.26412 »
« »
«0 0 6 0 0 0 » «c 2 » « 0.07552 »
« »« » « » (6.76)
«6 0 0 6 4 0 » « c3 » « 87.21919 »
«6 0 0 4 6 0 » «c 4 » « 89.96951 »
« »« » « »
«¬0 0 0 0 0 4»¼ «¬c5 »¼ «¬ 66.20445»¼
Therefore,
f # yˆ 0.021377 2.710687 x1 0.012587 x 2
16.551112 x1 x 2 8.159028 x12 9.534189 x 22 (6.77)
Example 6.19 [Generation of a Response Surface for Structural

Design]
In Figure 3.3, l 1 m, P 40000 N and T 60q. The stress of member 1 is
1
V (2 2b2 2(1 3 )b3 ) u 10 4 (6.78)
b1b2 2b1b3 b 2 b3
Approximate the function in Equation 6.78 by using RSM. The candidate points
are defined by the central composite design for 0.1 u 10 4 d b1 , b2 , b3 d 5.0 u 10 4
(Myers and Montgomery 1995).
Solution
Table 6.51 shows the allocation of the candidate points made by the central
composite design. x in Table 6.51 is the normalized value and the function f at the
candidate points is shown in Table 6.51. The function in Equation 6.78 is
approximated as
V c 0 c1 x1 c 2 x 2 c 3 x 3 c 4 x1 x 2 c 5 x1 x 3
c 6 x 2 x 3 c 7 x12 c 8 x 22 c 9 x 32 H (6.79a)
V # yˆ c 0 c1 x1 c 2 x 2 c 3 x 3 c 4 x1 x 2 c 5 x1 x 3
c 6 x 2 x 3 c 7 x12 c 8 x 22 c 9 x 32 (6.79b)
From Table 6.51 and Equation 6.73,

Table 6.51. Allocation of candidates and function calculation of Example 6.19 (unit of bi: 10 4 )
Observation b1 ( x1 ) b2 ( x 2 ) b3 ( x3 ) V
1 5 (1) 5 (1) 5 (1) 4.858E+07
2 5 (1) 5 (1) 0.1 (–1) 5.605E+07
3 5 (1) 0.1 (–1) 5 (1) 7.593E+07
4 5 (1) 0.1 (–1) 0.1 (–1) 6.813E+07
5 0.1 (–1) 5 (1) 5 (1) 1.582E+08
6 0.1 (–1) 5 (1) 0.1 (–1) 1.448E+09
7 0.1 (–1) 0.1 (–1) 5 (1) 2.268E+09
8 0.1 (–1) 0.1 (–1) 0.1 (–1) 2.429E+09
9 5 (1) 2.55 (0) 2.55 (0) 5.672E+07
10 0.1 (–1) 2.55 (0) 2.55 (0) 2.971E+08
11 2.55 (0) 5 (1) 2.55 (0) 8.092E+07
12 2.55 (0) 0.1 (–1) 2.55 (0) 1.465E+08
13 2.55 (0) 2.55 (0) 5 (1) 9.262E+07
14 2.55 (0) 2.55 (0) 0.1 (–1) 1.090E+08
15 2.55 (0) 2.55 (0) 2.55 (0) 9.525E+07
ª15 0 0 0 0 0 0 10 10 10º ªc0 º ª 7.430E 09 º

« 0 10 0 0 0 0 0 0 0 0 » « c » « 6.295E 09»
« »« 1 » « »
« 0 0 10 0 0 0 0 0 0 0 » «c2 » « 3.195E 09»
« »« » « »
« 0 0 0 10 0 0 0 0 0 0 » « c3 » « 1.467E 09 »
« 0 0 0 0 8 0 0 0 0 0 » «c 4 » « 3.051E 09 »
« »« » « » (6.80)
« 0 0 0 0 0 8 0 0 0 0 » «c5 » « 1.451E 09 »
« 0 0 0 0 0 0 8 0 0 0 » «c » « 1.145E 09 »
« »« 6 » « »
«10 0 0 0 0 0 0 10 8 8 » «c7 » « 6.906E 09 »
« »« » « »
«10 0 0 0 0 0 0 8 10 8 » « c8 » « 6.779E 09 »
«¬10 0 0 0 0 0 0 8 8 10»¼ «¬c9 »¼ «¬ 6.754E 09 »¼
Therefore, the quadratic response surface for Equation 6.78 is
V # yˆ 108 u (1.245 6.295 x1 3.195 x2 1.467 x3 3.813 x1 x2
1.814 x1x3 1.431x2 x3 3.564 x12 2.932 x22 2.803 x32 ) (6.81)
6.7.3 Optimization Using the Response Surface Method
Functions should be approximated to apply RSM to optimization. The

optimization problem in Equation 3.1 is considered. When approximated
functions for equality constraints are included in the optimization process, it is
difficult to satisfy the original constraints. Therefore, the equality constraints are
usually eliminated and the number of design variables is reduced due to the
elimination. An optimization formulation without equality constraints is as
follows:
Find x Rn (6.82a)
to minimize fˆ (x) (6.82b)
subject to gˆ j (x) d 0, j 1,..., m (6.82c)
xL d x d xU (6.82d)
where x is the design variable vector transformed from the original design
variable vector b, fˆ and ĝ are the objective function and the jth constraint
j
approximated by RSM, respectively. When an original function has an explicit
form, it can be directly used without approximation. Equation 6.82 is a metamodel
of Equation 3.1 and an optimum point is obtained from it. Since only explicit
functions are included in Equation 6.82, it is easy to find an optimum point.
When a highly nonlinear function is approximated in a wide range, the
accuracy can be deteriorated. There are methods to evaluate the accuracy. The
methods are introduced in a reference (Myers and Montgomery 1995).
Example 6.20 [A Mathematical Problem]

Obtain linear response surfaces for the following problem and solve the
optimization problem:
Find b
to minimize f 100(b2 b12 ) 2 (1 b1 ) 2
subject to g1 b1 b22 d 0 (6.83)
g2 b12 b2 d 0
0.5 d b1 d 0.5, 1.0 d b2 d 1.0
Use the same allocation method as Example 6.18 at b1 0.5, 0, 0.5 and b2
1.0, 0, 1.0.
Solution
In the same manner of Example 6.18, the L9 (3 4 ) orthogonal array is employed.

The objective and constraint functions are approximated as shown in Equation
6.84. The effect is the same as when the factorial design is utilized.
Find x
to minimize fˆ 72 x1 33.33 x 2
subject to gˆ 1 0.667 0.5 x1 d 0 (6.84)
gˆ 2 0.167 x 2 d 0
1 d x1 , x2 d 1
The optimum values are f 37.67, x1 1 and x 2 1.
Example 6.21 [Structure Problem]

The three bar structure in Figure 3.3 has l 1 m, P 40000 N and T 60q. The
optimization formulation is as follows:
Find b (6.85a)
to minimize f 2b1 b2 2b3 (6.85b)
2 2b2 2(1 3 )b3

subject to g1 14000 d 0 (6.85c)
b1b2 2b1b3 b 2 b3
2 2b2 2(1 3 )b3

g2 10000 d 0 (6.85d)
b1b2 2b1b3 b 2 b3
2((1 3 )b1 (1 3 )b3 )

g3 14000 d 0 (6.85e)
b1b2 2b1b3 b 2 b3
2((1 3 )b1 (1 3 )b3 )

g4 10000 d 0 (6.85f)
b1b2 2b1b3 b 2 b3
2(1 3 )b1 2 2b2

g5 14000 d 0 (6.85g)
b1b2 2b1b3 b 2 b3
2(1 3 )b1 2 2b2

g6 10000 d 0 (6.85h)
b1b2 2b1b3 b 2 b3
0.1 u 10 4 d b1 , b2 , b3 d 5.0 u 10 4 (6.85i)
For the candidate points and allocation of Example 6.19, approximate the
functions in quadratic forms and obtain the optimum point.
Solution
If the response surface method is performed by normalizing the design variables,
the problem in Equation 6.85 is approximated to the following:
Find x (6.86a)
to minimize f (9.762 3.465 x1 2.450 x2 3.465 x3 ) u 10 4 (6.86b)
subject to g1 2.645 6.295 x1 3.195 x2 1.467 x3 3.813 x1 x2
1.814 x1 x3 1.431x 2 x3 3.564 x12 2.932 x 22
2.803 x32 d 0 (6.86c)
g2 0.2452 6.295 x1 3.195 x2 1.467 x3 3.813 x1 x2
1.814 x1 x3 1.431x2 x3 3.564 x12 2.932 x22
2.803 x32 d 0 (6.86d)
g3 2.391 3.788 x1 4.915 x2 0.2533 x3 4.302 x1 x2
1.006 x1 x3 0.4633 x 2 x3 2.712 x12 2.602 x 22
2.155 x32 d 0 (6.86e)
g4 0.00915 3.788 x1 4.915 x2 0.2533 x3 4.302 x1 x2
1.006 x1 x3 0.4633 x 2 x3 2.712 x12 2.602 x 22
2.155 x32 d 0 (6.86f)
g5 1.146 2.507 x1 1.72 x2 1.214 x3 0.4888 x1 x2
2.82 x1 x3 1.894 x 2 x3 0.852 x12 0.3294 x 22
0.6482 x32 d 0 (6.86g)
g6 1.254 2.507 x1 1.72 x 2 1.214 x3 0.4888 x1 x 2
2.82 x1 x3 1.894 x 2 x3 0.852 x12 0.3294 x 22

0.6482 x 32 d 0 (6.86h)
1.0 d x1 , x2 , x3 d 1.0 (6.86i)
When the initial design is (0.102, í 0.878, í 0.388), the objective function is
f 6.62 u 10 4 and the constraints g1 , g 3 and g 5 are violated. The optimum
point is (0.159, í 0.171, í 0.886). The objective function is f 6.824 u 10 4 and
the constraints g1 and g 3 are active.
6.7.4 Selection of Candidate Points and Analysis of the

Optimization Results
Selection of the candidate points is investigated. Two aspects are considered in the
selection process. First, when we approximate a function in a linear form, there we
have the 2 n factorial design, the fractional design with two levels (orthogonal
arrays are usually used), the simplex design, etc. When the function is
approximated quadratically, the 3n factorial design, the fractional design with
three levels, the central composite design and rotary design are frequently used.
These are introduced in a reference (Myers and Montgomery 1995).
The optimum of Equation 6.82 can be different from that of Equation 3.1. In
particular, it is not guaranteed that the optimum of Equation 6.82 satisfies the
constraints in Equation 3.1. Therefore, the constraints should be tested and it
should be kept in mind that the solution of Equation 6.82 is not the mathematical
optimum of Equation 3.1.
Example 6.22 [Analysis of the Optimization Results of Example 6.20]

Analyze the results of Example 6.20.
Solution
In Example 6.20, the polynomials of the objective and constraint functions are
approximated by linear functions. The original problem has an optimum point at
b* [0.5 0.25]T and the objective function is f 0.25. However, the
approximated problem has the optimum point at b* [0.5 1.0]T and the
objective function is f 37.667. The results have large differences. Because the
original problem is highly nonlinear, the linear functions do not sufficiently
approximate the original functions. Also, the performance of the approximation
depends on the range of the design variables and the allocation method.
Example 6.23 [Analysis of the Optimization Results of Example 6.21]

Analyze the results of Example 6.21.
Solution
The objective function is a linear function and the constraints are rational functions.
The optimum point of the original problem is b* 10 4 u [2.0230 1.0460 0.1]T
and the objective function is f = 4.048 u 10 4. The solution in Example 6.21 is the
normalized one. Two solutions are similar. Moreover, the original constraints are
satisfied at the optimum of the approximated problem. Thus, the performance of
the approximation is fairly good for this problem.
6.8 Exercises
6.1 A university has 20 junior students in the Department of Mechanical

Engineering. Their grades for this semester are as follows:
3.5 2.3 3.6 1.9 4.2 2.1 3.5 2.9 4.2 3.8
2.0 3.3 4.3 1.8 2.7 3.2 3.7 1.6 2.6 3.9
Evaluate the mean and the standard deviation of the grades.
6.2 Regard the grades in Exercise 6.1 as 20 samples of all the alumni.
Estimate the mean and the variance of the population.
6.3 For Exercise 6.2, find the 95% confidence interval for the mean of the
entire population. Use the table of t-distribution.
6.4 The results of a one-way factorial design are as follows:
Level of the factor
A1 A2 A3
97.45 85.89 38.37
68.64 102.35 99.34

Responses for
repeated 72.39 97.37 58.65
experiments
57.98 107.37 63.27
62.38 89.79 46.89
(1) Make an ANOVA table for the experiment.

(2) Find the 90% confidence interval for each level of the factor. Use the
table of t-distribution.
6.5 The results of a two-way factorial design are as follows:
Factor A
A1 A2 A3
Factor B
12.35 17.89 26.89
14.78 12.78 32.24
B1
20.89 25.23 23.24
15.98 18.99 12.38
18.67 23.35 37.27
26.78 23.45 35.89
B2
31.26 35.47 25.67
23.87 42.89 47.64
25.92 23.78 34.78
12.98 32.87 25.98
B3
23.89 45.84 42.34
35.62 59.82 56.73
23.82 48.93 35.89
34.89 34.89 56.83
B4
63.78 63.21 54.23
39.83 33.89 63.67
(1) Make an ANOVA table for the experiment.

(2) Obtain the values that maximize the characteristic function and find
the 95% confidence interval for the mean of the population. Use the
table of t-distribution.
6.6 Show that the columns of the L8 (2 7 ) orthogonal array in Table 6.22 are
orthogonal to each other.
6.7 There are two factors A and B with two levels, four factors C, D, E and F
with three levels. The interactions A u C , A u D and E u F are significant.
Select the minimum orthogonal array for a matrix experiment.
6.8 We have the following optimization problem:
Minimize f 2 A 2 ( B A) 2 2C 2
The levels of each factor are as follows:

Level
Factor
1 2 3
A –1.5 0 1
B –1 0 1
C –1 0 1
(1) With the L9 (3 4 ) orthogonal array, obtain a design solution by using

the method in Section 6.3.3.
(2) Find a design solution with a factorial combination and compare the
solution with the one from (1).
6.9 We have the following optimization problem:
Minimize f ( A 3) 3 (3B 10 A) 2 (2C 15 B ) 2 D 2
The levels of each factor are as follows:
Level
Factor
1 2
A 2 5
B 5 8
C 4 7
D 1 4
(1) With the L8 (2 7 ) orthogonal array, obtain a design solution by using

the method in Section 6.3.3.
(2) Find a design solution with a factorial combination and compare the
solution with the one from (1).
6.10 For Exercise 6.8

(1) Make an ANOVA table and discuss the effect of each factor and the
error.
(2) Using the effect of the error, discuss the effect of the interactions.
(3) If interactions exist, make a design using an orthogonal array by
considering the interactions and compare the results with those of
Exercise 6.8.
6.11 For Exercise 6.9

(1) Make an ANOVA table and discuss the effect of each factor and the
error.
considering the interactions and compare the results with those of
Exercise 6.9.
6.12 For a mathematical problem,
Minimize f (b1 3) 4 (b2 5) 2 b32 0.3(b1b2 b2b3 b3b1 )
The factors have the following levels:
Level b1 b2 b3
1 1.5 2.0 2.0
2 3.0 4.0 3.0
3 4.5 6.0 4.0
(1) Find a design solution by using the L9 (34 ) orthogonal array. Use the
method in Section 6.3.3.
(2) Find a design solution by using the factorial combination and compare
the solution with the one from (1).
(3) Make an ANOVA table and discuss the effects of the factors and the
error.
(4) From the effect of the error, discuss the effect of the interactions.
considering the interactions and compare the results with the one from
(2).
6.13 For a mathematical problem
6
2
f ¦ (bi 2)
i 1
The factors have the following levels:

Level b1 b2 b3 b4 b5 b6
1 1.0 2.0 0.8 2.0 1.5 1.3
2 2.0 3.0 4.0 3.8 2.6 2.4
3 3.1 4.0 5.2 5.6 3.7 3.5
(1) Find a design solution by using the L18 (21 u 3 7 ) orthogonal array.
Use the method in Section 6.3.3.
(2) Find a design solution by using the factorial combination and compare
the solution with the one from (1).
(3) Make an ANOVA table and discuss the effects of the factors and the
error.
considering the interactions and compare the results with the one from
(2).
6.14 A design problem for the three bar truss in Figure 3.3 is as follows:
Maximize [1
where [1 is the first natural frequency. Young’s modulus (E) is 207 GPa,
and the density ( U ) is 2770 kg/m3. The levels of the design variables are
as follows:
Design variables and levels ( u 10 4 m 2 )
Level b1 b2 b3
1 0.1 0.1 0.1
2 2.5 2.5 2.5
3 5.0 5.0 5.0
(1) Select the minimum orthogonal array for the design.

(2) Design the structure by using the method in Section 6.3.3.
(3) Find a solution by using the factorial combination and compare the
results with the one from (2).
(4) Determine if there are interactions.

(5) If interactions exist, design the structure with an orthogonal array by
considering the interactions and compare the solution with the one
from (3).
6.15 A design problem of the ten bar truss below is as follows:
Maximize [1
9.144 m 9.144 m
a b
g
i
e
f 9.144 m
h
j
c d
where [1 is the first natural frequency. Young’s modulus (E) is 207 GPa,
and the density ( U ) is 2770 kg/m3. The levels of design variables are as
follows:
Design variables and levels ( u 10 4 m 2 )
Design variable (factor)

Level
A (a) B (b) E (e) G (g) I (i)
1 6.452 6.452 6.452 6.452 6.452
2 70.968 70.968 70.968 70.968 70.968
3 135.484 135.484 135.484 135.484 135.484
(1) The L18 (21 u 37 ) orthogonal array is selected and the design variables
are allocated as follows:
The L18 ( 21 u 3 7 ) orthogonal array
Allocation of the design variable

Exp. no. 1 2 3 4 5 6 7 8 [1 (Hz)
e A B E G I e e
1 1 1 1 1 1 1 1 1
2 1 1 2 2 2 2 2 2
3 1 1 3 3 3 3 3 3
4 1 2 1 1 2 2 3 3
5 1 2 2 2 3 3 1 1
6 1 2 3 3 1 1 2 2
7 1 3 1 2 1 3 2 3
8 1 3 2 3 2 1 3 1
9 1 3 3 1 3 2 1 2
10 2 1 1 3 3 2 2 1
11 2 1 2 1 1 3 3 2
12 2 1 3 2 2 1 1 3
13 2 2 1 2 3 1 3 2
14 2 2 2 3 1 2 1 3
15 2 2 3 1 2 3 2 1
16 2 3 1 3 2 3 1 2
17 2 3 2 1 3 1 2 3
18 2 3 3 2 1 2 3 1
Perform an experiment for each row and evaluate the characteristic

function.
(2) Design the structure by using the method in Section 6.3.3.
(3) Find a solution by using the factorial combination and compare it with
the one from (2).
(4) Determine if there are interactions.
(5) If interactions exist, design the structure with the orthogonal array
L27 (313 ) by considering the interactions and compare the solution
with the one from (3).
6.16 In Example 3.8, the design variables can have the following discrete
values:
Level
1 2
b1 12 16
b2 0 2
Find the solutions by using Equation (6.60) when the scale factor s is 0.1,
1.0, 10, and 100. Use the method in Section 6.5 with an orthogonal array.
6.17 We have the following mathematical problem:
Find b1 ,b2
to minimize f (b) (b1 2) 2 b22
subject to (1 b1 ) 3 b2 t 0
0 d bi , i 1, 2
The design variables can have discrete values as follows:
Level
1 2 3
b1 –1 1 3
b2 –2 0 2
Find solutions by using Equation (6.60) when the scale factor s is 10 and
50. Use the method in Section 6.5 with an orthogonal array.
Find b1 , b2 , b3
to minimize f (b) b1b2 b3
subject to 72 b1 2b2 2b3 t 0
b1 2b2 2b3 t 0
0 d bi d 42, i 1, 2, 3

Level
1 2 3
b1 14 24 34
b2 2 12 22
b3 2 12 22
Find solutions by using Equation (6.60) when the scale factor s is 5 and 50.
Use the method in Section 6.5 with an orthogonal array.
Find b1 , b2 , b3 , b4
to minimize f (b) b1 b2 b3 b1b3 b1b4 b2 b3 b2 b4
subject to 8 b1 2b2 t 0
12 4b1 b2 t 0
12 3b1 4b2 t 0
5 b3 b4 t 0
8 2b3 b4 t 0
8 b3 2b4 t 0
Level
1 2 3
b1 –2 0 2
b2 1 3 5
b3 –2 0 2
b4 2 4 6
Find solutions by using Equation (6.60) when the scale factor s is 5 and 10.
Use the method in Section 6.5 with an orthogonal array.
6.20 Using SOA, perform two iterations for Exercise 6.16. The candidate
values of the design variables are as follows:
b1 { 10, 11, 12, 13, 14, 15, 16, 17, 18 }

Candidates
b2 { 0, 1, 2, 3, 4 }
s = 10 and the initial design is the solution of Exercise 6.16 when s = 10.
Candidates b1 , b2 { 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5 }
Candidates b1 , b2 , b3 { 0, 2, 4, 6, 8, ..., 36, 38, 40, 42 }
Candidates b1,b2 ,b3 ,b4 { 2.0, 1.5, 1.0, ,8, 8.5, 9 , 9.5, 10 }
6.24 Solve Exercise 6.17 by using RSM. The candidates are those in Exercise
6.17. Use the central composite design. Compare the solution with the
mathematical optimum and the solution from Exercise 6.17.
6.18. Use the L9 (3 4 ) orthogonal array. Compare the solution with the
6.19. Use the central composite design. Compare the solution with the
6.A Statistics
Statistical theories are briefly explained.
Mean: When we have n responses of x1 , x2 ,..., xn , the mean P is as follows:
1 n
P ( x1 x 2 ... x n ) / n ( ¦ xi ) (6.A.1)
n i1
Variance and standard deviation: From Equation 6.A.1, the variance V 2 and the
standard deviation V are defined as:
n
2
¦ ( xi P )
V2 i 1
(6.A.2)
n
Random variable and distribution: The set of all possible outcomes of a

statistical experiment is called the “sample space” and an event is a subset of the
sample space. The probability of event A is the sum of the probabilities of all
sample points. A function whose value is a real number determined by each
element in the sample space is called a “random variable.” A random variable is
denoted by a capital letter, say X. The value of a random variable is denoted by a
small letter, say x. In a design of experiments, the random variable is frequently
denoted by a small letter.
The distribution of x that a random variable X can have is defined by the
function p(x). The function p(x) is called the probability density function (pdf).
The pdf can be obtained for a continuous or a discrete random variable. The pdf
for a continuous random variable is briefly discussed. The pdf for a continuous
random variable X has the following characteristics:
(1) For any x, p( x ) t 0 (6.A.3)

f
(2) ³ p( x)dx 1 (6.A.4)
f
The probability for a random variable X to reside in [a, b] is as follows:
b
P{a d X d b} ³ p( x)dx (6.A.5)
a
The cumulative probability is defined as
F ( x) P{ X d x} (6.A.6)
A representative probability density function is the function for the normal

distribution. When the mean of a random variable X is P and the variance is V 2 ,
the normal distribution is
2
1 § xP ·
¨ ¸
1
f ( x) e 2© V ¹ (6.A.7)
2SV 2
where e is the base of the natural logarithm. The normal distribution is denoted as
N ( P , V 2 ) and the random variable is denoted by Z. The normal distribution
N(0,1) is called the standard normal distribution and the pdf is
1
1 z2
I ( z) e 2 (6.A.8)
2S
In most textbooks on statistics, a table for the following zD , which has 1 D for
the cumulative probability, is included.
zD
1D P{Z d zD } ³ I ( z )dz (6.A.9)
f
The shape of the standard normal distribution is illustrated in Figure 6.A.1.

Ȥ 2 -distribution: If Z1 , Z 2 , ..., Z n are random variables that have independent
standard normal distributions, a random variable F 2 is defined as follows:
F 2 (n) Z12 Z 22 ... Z n2 (6.A.10)
The degrees of freedom of F 2 is n and the distribution is called the F 2 -
1–D
D
0 ZD Z
Figure 6.A.1. Standard normal distribution N (0,1)

distribution. In most textbooks on statistics a table for the following cumulative

probability 1 D of F 2 (n;D ) is included,
1D P{F 2 (n) d F 2 (n;D )} (6.A.11)
t-distribution: When Z and F 2 (n) are independent random variables, the

following random variable (t(n)) has t-distribution:
Z
t ( n) (6.A.12)
F 2 ( n)
n
In most textbooks on statistics, a table for the following cumulative probability

1 D of t (n; D ) is included:
1D P{t (n) d t (n; D )} (6.A.13)
F-distribution: Suppose F12 (n1 ) and F 22 (n2 ) are independent and their degrees
of freedom are n1 and n 2 . When they have F 2 -distributions, a random variable
is defined as follows:
F12 (n1 ) / n1
F (6.A.14)
F 22 (n2 ) / n2
The random variable F has n1 DOF for the numerator, n2 DOF for the
denominator and has F-distribution. In most textbooks on statistics, a table for the
following cumulative probability 1 D of F (n1 , n2 ;D ) is included:
1D P{F (n1 , n2 ) d F (n1 , n2 ;D )} (6.A.15)
Statistical point estimation: Statistical estimation is to make inferences for

parameters of a population from the data of samples. The parameters to estimate
are mean ( P ) , variance (V 2 ) , etc. Suppose n samples are X 1 , X 2 ,..., X n . The
mean and variance of the population are estimated from the samples. The
estimated mean ( Pˆ ) and variance (Vˆ 2 ) are
n
¦ Xi
i 1
P̂ X (6.A.16)
n
n
2
¦(Xi X )
Vˆ 2 S2 i 1
(6.A.17)
n 1
The estimated values P̂ and Vˆ 2 are called “unbiased estimators.” The reason
is that their expected values are as follows:
E (X ) P (6.A.18)
E (S 2 ) V2 (6.A.19)
Statistical interval estimation: As opposed to point estimation, interval

estimation is a probability for a mean or a variance to reside in a specific interval.
The interval estimation for the mean is briefly explained. Suppose n samples have
X and S 2 for their mean and variance. By t-distribution, the population mean
has the following characteristics:
X P
T (6.A.20)
S/ n
§ ·
P¨ t D T t D ¸ 1D (6.A.21)
¨ ¸
© 2 2 ¹
§ S S ·¸
P¨ X t D P X tD 1D (6.A.22)
¨ n n ¸
© 2 2 ¹
Equation 6.A.22 represents the probability for the distribution of the population.
The degrees of freedom of the t-distribution in Equation 6.A.22 is n í 1.
6.B Orthogonal Arrays
6.B.1 Orthogonal Arrays with Two Levels
Table 6.B.1. The L 4 ( 2 3 ) orthogonal array
Column
Exp. no.
1 2 3
1 1 1 1
2 1 2 2
3 2 1 2
4 2 2 1
Table 6.B.2. The L8 ( 2 7 ) orthogonal array
Column
Exp. no.
1 2 3 4 5 6 7
1 1 1 1 1 1 1 1
2 1 1 1 2 2 2 2
3 1 2 2 1 1 2 2
4 1 2 2 2 2 1 1
5 2 1 2 1 2 1 2
6 2 1 2 2 1 2 1
7 2 2 1 1 2 2 1
8 2 2 1 2 1 1 2
11
Table 6.B.3. The L12 ( 2 ) orthogonal array
Exp. Column
no. 1 2 3 4 5 6 7 8 9 10 11
1 1 1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 1 2 2 2 2 2 2
3 1 1 2 2 2 1 1 1 2 2 2
4 1 2 1 2 2 1 2 2 1 1 2
5 1 2 2 1 2 2 1 2 1 2 1
6 1 2 2 2 1 2 2 1 2 1 1
7 2 1 2 2 1 1 2 2 1 2 1
8 2 1 2 1 2 2 2 1 1 1 2
9 2 1 1 2 2 2 1 2 2 1 1
10 2 2 2 1 1 1 1 2 2 1 2
11 2 2 1 2 1 2 1 1 1 2 2
12 2 2 1 1 2 1 2 1 2 2 1
Table 6.B.4. The L16 ( 215 ) orthogonal array
Exp. Column
no. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2
3 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2
4 1 1 1 2 2 2 2 2 2 2 2 1 1 1 1
5 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2
6 1 2 2 1 1 2 2 2 2 1 1 2 2 1 1
7 1 2 2 2 2 1 1 1 1 2 2 2 2 1 1
8 1 2 2 2 2 1 1 2 2 1 1 1 1 2 2
9 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2
10 2 1 2 1 2 1 2 2 1 2 1 2 1 2 1
11 2 1 2 2 1 2 1 1 2 1 2 2 1 2 1
12 2 1 2 2 1 2 1 2 1 2 1 1 2 1 2
13 2 2 1 1 2 2 1 1 2 2 1 1 2 2 1
14 2 2 1 1 2 2 1 2 1 1 2 2 1 1 2
15 2 2 1 2 1 1 2 1 2 2 1 2 1 1 2
16 2 2 1 2 1 1 2 2 1 1 2 1 2 2 1
6.B.2 Orthogonal Arrays with Three Levels
4
Table 6.B.5. The L9 (3 ) orthogonal array
Column
Exp. no.
1 2 3 4
1 1 1 1 1
2 1 2 2 2
3 1 3 3 3
4 2 1 2 3
5 2 2 3 1
6 2 3 1 2
7 3 1 3 2
8 3 2 1 3
9 3 3 2 1
Table 6.B.6. The L27 (313 ) orthogonal array
Exp. Column
no. 1 2 3 4 5 6 7 8 9 10 11 12 13
1 1 1 1 1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 2 2 2 2 2 2 2 2 2
3 1 1 1 1 3 3 3 3 3 3 3 3 3
4 1 2 2 2 1 1 1 2 2 2 3 3 3
5 1 2 2 2 2 2 2 3 3 3 1 1 1
6 1 2 2 2 3 3 3 1 1 1 2 2 2
7 1 3 3 3 1 1 1 3 3 3 2 2 2
8 1 3 3 3 2 2 2 1 1 1 3 3 3
9 1 3 3 3 3 3 3 2 2 2 1 1 1
10 2 1 2 3 1 2 3 1 2 3 1 2 3
11 2 1 2 3 2 3 1 2 3 1 2 3 1
12 2 1 2 3 3 1 2 3 1 2 3 1 2
13 2 2 3 1 1 2 3 2 3 1 3 1 2
14 2 2 3 1 2 3 1 3 1 2 1 2 3
15 2 2 3 1 3 1 2 1 2 3 2 3 1
16 2 3 1 2 1 2 3 3 1 2 2 3 1
17 2 3 1 2 2 3 1 1 2 3 3 1 2
18 2 3 1 2 3 1 2 2 3 1 1 2 3
19 3 1 3 2 1 3 2 1 3 2 1 3 2
20 3 1 3 2 2 1 3 2 1 3 2 1 3
21 3 1 3 2 3 2 1 3 2 1 3 2 1
22 3 2 1 3 1 3 2 2 1 3 3 2 1
23 3 2 1 3 2 1 3 3 2 1 1 3 2
24 3 2 1 3 3 2 1 1 3 2 2 1 3
25 3 3 2 1 1 3 2 3 2 1 2 1 3
26 3 3 2 1 2 1 3 1 3 2 3 2 1
27 3 3 2 1 3 2 1 2 1 3 1 3 2
6.B.3 Orthogonal Array with Mixed Levels
Table 6.B.7. The L18 (21 u 3 7 ) orthogonal array
Column
Exp. no.
1 2 3 4 5 6 7 8
1 1 1 1 1 1 1 1 1
2 1 1 2 2 2 2 2 2
3 1 1 3 3 3 3 3 3
4 1 2 1 1 2 2 3 3
5 1 2 2 2 3 3 1 1
6 1 2 3 3 1 1 2 2
7 1 3 1 2 1 3 2 3
8 1 3 2 3 2 1 3 1
9 1 3 3 1 3 2 1 2
10 2 1 1 3 3 2 2 1
11 2 1 2 1 1 3 3 2
12 2 1 3 2 2 1 1 3
13 2 2 1 2 3 1 3 2
14 2 2 2 3 1 2 1 3
15 2 2 3 1 2 3 2 1
16 2 3 1 3 2 3 1 2
17 2 3 2 1 3 1 2 3
18 2 3 3 2 1 2 3 1
References
Alberto GD, Phillips DT (1994) Principles of Experimental Design and Analysis.
Chapman and Hall, Boca Raton, FL
Box GEP, Draper NR (1987) Empirical Model-Building and Response Surfaces.
Wiley, New York
Box GEP, Hunter WG (1978) Statistics for Experimenters. Wiley, New York
Freund JE, Walpole RE (1987) Mathematical Statistics. Prentice Hall, Englewood
Cliffs, NJ
Goldberg DE (1989) Genetic Algorithms in Search Optimization and Machine
Learning. Addison Wesley, New York
Hedayat AS, Sloane NJA, Stufken J (1999) Orthogonal Array: Theory and
Applications. Springer, Berlin Heidelberg New York
Korte BH, Vygen J (2003) Combinatorial Optimization: Theory and Algorithms.
Springer, Berlin Heidelberg New York
Ku KJ, Rao SS, Chen L (1998) Taguchi-aided Search Method for Design
Optimization of Engineering Systems. Engineering Optimization 30:1í23
Lee KH, Eom IS, Park GJ, Lee WI (1996) Robust Design for Unconstrained
Optimization Problems Using Taguchi Method. AIAA J. 34(4):1059í1063
Lee KH, Park GJ (1996) Discrete Post Process of the Constrained Size
Optimization Using Orthogonal Arrays. Proceedings of 6th AIAA/USAF/
NASA/ISSMO Symposium on Multidisciplinary Analysis and Optimization,
Part 1, AIAA, Sept. 4í6, Bellevue, pp. 56í61
Lee KH, Park GJ (2002) Robust Optimization in Discrete Design Space for
Constrained Problems. AIAA J. 40(4):774í780
Lee KH, Yi JW, Park JS, Park GJ (2003) An Optimization Algorithm Using
Orthogonal Arrays in Discrete Design Space for Structures. Finite Elements in
Analysis and Design 40(1):121í135
Lim JM, Wu, DH, Park GJ (2000) Analysis and Design Consideration of an
Energy Absorbing Steering System using Orthogonal Arrays. International
Journal of Crashworthiness 5(3):271í278
Mood AM, Grabill FA, Boes DC (1974) Introduction to the Theory of Statistics.
McGraw Hill, New York
Myers RH, Montgomery DC (1995) Response Surface Methodology. Wiley, New
York
Osman, IH, Kelly JP (1996) Meta-heuristics: Theory & Application. Kluwer,
Dordrecht
Park GJ, Hwang WJ, Lee WI (1994) A Structural Optimization Post-process Using
Taguchi Method. International Journal of Japan Society of Mechanical
Engineers, Series A 37(2):166í172
Design with Applications. McGraw Hill, New York
Taguchi G (1987) Systems of Experimental Design (Vols. I, II). Kraus
International Publications, New York
Walpole MS (1982) Introduction to Statistics. Macmillan, New York
7 Robust Design
7.1 Introduction
As the engineering environment becomes more and more competitive, quality
products are required in industry. Unexpected deviations from the function that a
designer initially intended are caused by variations. Robust design seeks to
prevent such phenomena and has been developed to improve the quality and
reliability of the product in engineering. Recently, this technology has expanded
to various design areas.
Many researchers have defined robust design. Taguchi, who is the pioneer of
robust design, stated that “robustness is the state where the technology, product, or
process performance is minimally sensitive to factors causing variability (either in
the manufacturing or the user’s environment) and aging at the lowest unit
manufacturing cost” (Taguchi et al. 2000). Suh stated that “robust design is
defined as the design that satisfies the functional requirements even though the
design parameters and the process variables have large tolerances for ease of
manufacturing and assembly. This definition of robust design states that the
information content is minimized” (Suh 2001). Box defined that “robustifying a
product is the process of defining its specifications to minimize the product’s
sensitivity to variation” (Box and Fung 1993). Although different expressions are
used, their meanings are similar. A common aspect of the definitions is that robust
design is insensitive to variations and this concept is accepted in the engineering
community.
Conventional design methodologies have been developed to improve the
performances of the products. If consideration of robustness is added, a new
procedure should be made in addition to the conventional concept. Theories of
robust design have been developed by using existing design theories. The first one
is the Taguchi method and it is the foundation of robust design. Taguchi proposed
methods of determining variables to make a performance insensitive to noises in
the manufacturing process. The concept of robustness was established by the
Taguchi method. In a product design, robustness is obtained by appropriate
modification of the Taguchi method. The second one is robust optimization where
the robustness concept is added to conventional optimization. The objective
function and constraints are redefined with robustness indices. The last one is
axiomatic design, which can be exploited in conceptual design. In axiomatic
design, the Independence Axiom is utilized to find excellent designs. When
multiple designs are found by using the Independence Axiom, the best one is
selected by the Information Axiom and the robustness concept is employed in that
process.
Various robust design methodologies are overviewed based on the above three
methods. The trend of the research is surveyed and the applications are discussed.
However, theoretical aspects are emphasized, rather than the application aspects.
The advantages and disadvantages of the methods are also discussed. In the robust
design area, some terminologies are used in different ways. Thus, those
terminologies are precisely defined and a unified viewpoint of robust design is
proposed.
There are some methods that are similar to robust design such as design
methods considering reliability and uncertainty. Uncertainties or noises are
included in these methods as well. Optimization terminologies can be utilized to
distinguish the methods from robust design. In robust design, low sensitivity of
the objective function is emphasized. In reliability design, reliability of constraints
is important. The reliability of constraints is similar to the robustness of the
constraints in robust design. Design with uncertainties is similar to robust
optimization. However, robustness of the objective function is not considered.
The distinctions are made based on the methods of the early stages. These days,
the methods are often fused; therefore, it may not be possible to distinguish them
in some applications. Methods considering reliability or uncertainties are not
included (Special Supplement on Robust and Reliability-based Design, 2006).
7.2 Mean and Variance

The optimization formulation in Equation 3.1 can be used for a general design.
The functions in Equation 3.1 can be evaluated two ways. One is to use
experiments when the function is not mathematically defined. A number of case
studies using the traditional Taguchi method are attributed to this case. The other
is to use mathematical expressions that are directly calculated or approximated by
numerical approaches. This section is mainly focused on the latter method.
When there are noises or perturbations, the objective and the constraint
functions in Equation 3.1 are modified as follows:
f (b) o f (b z b , p z p ) (7.1a)
g j (b ) o g j ( b z b , p z p ) (7.1b)
where p R o is the design parameter vector, which is usually regarded as a

constant vector in the design process. The noises of design variables and design
Robust Design 395
parameters are represented as z b R n and z p R o , respectively; o is the number

of parameters. The noise factors (z b , z p ) are also called the uncontrollable factors
since they cannot be controlled by the designer. The noise factor may be an
external, an internal or a unit-to-unit noise of a product.
The control factors are selected as independent variables that can be controlled
by the designer and have a significant impact on the problem characteristics. The
control factors correspond to the design variable vector b, while the characteristic
function is equivalent to the objective function f of the optimization problem. The
characteristic function is a function or response that is to be measured and
improved upon.
By substituting Equation 7.1 into Equation 3.1 to consider the robustness in the
design formulation, Equation 3.1 is rewritten as
Find b Rn (7.2a)
to minimize F (b, p, z ) (7.2b)
subject to G j (b, p, z ) d 0 , j 1, ..., m (7.2c)
bL d b d bU (7.2d)
where F and G j are functions considering the noise in f and g j , respectively.
Table 7.1. Formulae of means and variances
Name Symbol Formula
l
1
Sample mean f or (m f ) f
l ¦f
i 1
i
2 1 l 2
Sample variance V or ( s f ) V ¦ ( fi f )
l 1 i 1
Sample standard deviation sf sf V
P 1 l
Population mean P ¦ fi
li 1
1 l
Population variance V2 V2 ¦ ( fi P )
2
li 1
Population standard deviation V V V2

Generally, F and G j are defined as the functions of the mean and the variance of f
and g j , respectively.
Robust design should not be developed by the deterministic concept. It should
be developed by the stochastic concept, since the deterministic approach cannot
include the noises. The performance measures are defined as the data to be
measured experimentally or mathematically, which represent the performances of a
product or a process. The performance measures are usually modified to form a
characteristic function. The performance measures could vary under the same
circumstances due to the distribution of noise factors. Generally, the distribution is
measured by the mean and the variance. The formulae of mean and variance are
summarized in Table 7.1. f i is the ith characteristic function in the experiment or
the ith response function in the design.
The mean and the variance for the design are explained. The characteristic
function f in Table 7.1 is a specific response. A design variable (factor) has a
distribution around a nominal value for many reasons. Thus, the response f also
has a distribution. The distribution is quantified by the variance and the sources
for the variance are shown in Table 7.2. The mean of f is expressed by the mean
P f and the variance is expressed by V 2f . V f is the standard deviation. The
mean P f and the variance V 2f of the characteristic function f are represented as
Table 7.2. Sources of variation
Factor Example
x Temperature and relative humidity

External x Load
factor x Human error
x Dust in the environment
x Weight of each part
Unit-to-unit x Dimension tolerance
factor x Differences by batch-to-batch
x Thickness variations in coated products
x Wear
Internal x Mileage on a car
(aging) x Weathering of paint on a house
factor x Total current passing through a car battery
x Plastic creep
Robust Design 397
b p
Pf E[ f (b)] ³ ³ ³ f (b z , p z )
(7.3a)
u1 ( z1b ) u n ( z nb )v1 ( z1p ) vo ( z op ) dz1b dz nb dz1p dz op
V 2f E[( f (b) P f ) 2 ] b p
³ ³ ³ [ f (b z , p z ) P f ]
2
(7.3b)
u1 ( z1b ) u n ( z nb ) v1 ( z1p ) v o ( z op ) dz1b dz nb dz1p dz op
where E[x] is the expectation value of x, and u i ( z ib ) and vi ( z ib ) are the

probability density functions of noise factors z ib and zip , respectively. In
Equation 7.3, the noise factors are assumed to be statistically independent. If it
follows the Gaussian distribution, the probability density function becomes
1 ª (bi P b ) 2 º
u i ( z ib ) exp « i
» (7.4)
V bi 2S «¬ 2V b2i »¼
where P bi and V bi are the mean and the standard deviation of the ith design
variable bi . The robust design techniques have been developed by replacing the
functions of the mean and variance into the functions of Equation 7.2. In the next
sections, the robust design methodologies are examined.
Example 7.1 [Calculation of the Mean and the Variance]

We have the following equation:
f b1 (b2 2) 2b1b2 (7.5)
b1 and b2 have uniform distributions around (b1 , b2 ) (0.0, 0.0) and the ranges
are 0.5 d b1 d 0.5 and 0.5 d b2 d 0.5 . Obtain the mean and the variance of f at
(b1 , b2 ) (0.0, 0.0) .
Solution
From Equation 7.3a, the mean is
0.5 0.5
Pf ³ ³ (b1 b2 2 2b1b2 ) [u (b1 0.5) u (b1 0.5)]
0.5 0.5
[u (b2 0.5) u (b2 0.5)]db1db2 (7.6)
2
where u is the unit step function enabling the uniform distribution within the
design ranges.
By using Equation 7.3b, the variance is
0.5 0.5
V 2f ³
2
³ [(b1 b2 2 2b1b2 ) (2)] [u (b1 0.5) u (b1 0.5)]
0.5 0.5
[u (b2 0.5) u (b2 0.5)]db1db2 (7.7)
0.1944
Example 7.2 [Three Bar Truss Problem]

In Equations 3.9 and 3.10 of Example 3.1 with the sizes of Example 6.15, the
elastic modulus E has uniform distribution in the range of
203 GPa d E d 209 GPa. Calculate the means and the variances of the horizontal
displacement G x and the vertical displacement G y . The cross sectional areas are b
= [0.0002 0.00012 0.0002]T .
Solution
G x and G y are calculated as
2lPx 1
Gx 1.414 u10 8 (7.8a)
b1 E E
2lPy 1
Gy 1.325 u 10 8 (7.8b)
(b1 2b2 ) E E
According to Equation 7.3a, the means are
209 E 9 1 1
8
PG x ³ 1.414 u 10
203E 9 E 209 u 109 203 u 109
[u ( E 203 u 109 ) u ( E 209 u 109 )]dE (7.9a)
0.000686456
209 E 9 1 1
8
PG y ³ 1.325 u 10
203E 9 E 209 u 10 203 u 109
9
[u ( E 203 u 109 ) u ( E 209 u 109 )]dE (7.9b)

0.000643249
From Equation 7.3b, the variances are

Robust Design 399
209 E 9 1 1
V G2x ³ ( 1.414 u 10
8
0.000686456) 2
203E 9 E 209 u 10 203 u 109
9
[u ( E 203 u 109 ) u ( E 209 u 109 )]dE (7.10a)

11
3.33186 u 10
209 E 9 1 1
V G2y ³ ( 1.325 u 10
8
0.000643249) 2
203E 9 E 209 u 10 203 u 109
9
[u ( E 203 u 109 ) u ( E 209 u 109 )]dE (7.10b)

11
2.92563 u 10
7.3 Taguchi Method
7.3.1 Introduction
By the end of the 1940s the Taguchi method had been developed by Dr. G.
Taguchi for quality improvement. His technique for quality engineering is referred
to as the Taguchi method or robust design (Taguchi 1987). While the Taguchi
method was successfully applied at the Electrical Communications Laboratories of
the Nippon Telephone and Telegraph Company, the AT&T Bell Laboratories
became interested in it. In 1980, Phadke invited Taguchi to the AT&T Bell
Laboratories. After being impressed by the Taguchi method, Phadke published a
textbook on the Taguchi method in 1989 (Phadke 1989).
The Taguchi method has greatly contributed to quality improvement of various
designs. In early case studies, the Taguchi method was applied to the process
design rather than the product design. That is, it was regarded as a method for
design of experiments rather than a design methodology.
In Equation 7.2 of the design formulation, the original Taguchi method does
not deal with the constraints in Equation 7.2. This leads to unconstrained
optimization problems. Although the design parameters p and the noise factors z
are not explicitly defined in the design, the Taguchi method can be used to
determine the settings of control factors for robust design. To explain this in detail,
let us use an example with three distributions of the objective function f as
In Figure 7.1, P A , P B and P C are the means of distributions A, B and C,
respectively. Then, robustness can be measured by the magnitude of variation.
When the design is focused only on the robustness of the performance, the
candidate designs B and C are better than design A. However, the target value of
the performance generally exists in the design. The target value can be zero,
infinite or a specific nominal value. Thus, the Taguchi method suggests a design
Probability
density C B
function
A
PA PC PB
f
m Target value
Figure 7.1. Examples of distribution of the response functions
with minimum variation around the target value (m). From the Taguchi method
viewpoint, design C is the best out of the three candidate designs.
The procedure to obtain a robust design is now explained. The design
candidates D and E with two arbitrary distributions are represented in Figure 7.2.
The design range is equivalent to the allowable range and the means of the two
distributions coincide with the target value m of the design. A small part of
candidate design D is outside the design range while no part of candidate design E
is outside the design range. However, if any unexpected noise factors become
active, the distribution of E has a larger probability to be outside the design range
Design range
Probability D
density
function
E
m Target value f
Figure 7.2. Probability density functions of the two designs with the design range
Robust Design 401
than that of D. For instance, the unexpected noise is the tolerances of product
sizes, variances of the environment such as temperature, humidity, external forces,
manufacturing processes, etc. Thus, it is clear that candidate design D is more
robust than candidate design E.
7.3.2 The Loss Function and S/N Ratio

Taguchi introduced a quadratic loss function to represent robustness as
L( f ) k ( f m) 2 (7.11)
where k is the constant to define the loss and m is the target value. The expected
value of the loss function is defined as
Q E[ L( f )] k[V 2 ( P m) 2 ] (7.12)
where P is the mean of f and V is the standard deviation of f.

The symbol f represents the response or objective function in a general design,
while the symbol y is used for f in the Taguchi method. Robust design is the
design minimizing the average loss, which can select design candidate D in Figure
7.2.
Figure 7.3 is called the P-diagram and is used to explain the procedure of the
Taguchi method from another viewpoint. The P-diagram represents the schematic
relationship between the factors and the product or process. The noise factor is
one that the designer cannot control even though it causes variability. In a general
design, tolerances are considered noise factors. In a parameter design, control
factors are determined to reduce the effect of the noise factors. Therefore, noise
factors are not directly considered. On the other hand, noise factors are directly
controlled in a tolerance design. The parameter design stage is when the quality of
a product or process can be improved without cost increase, while the tolerance
z Noise factors
m f
Product / Process
Target value Response
(Signal factor)
b Design variables
(Control factors)
Figure 7.3. Block diagram of a product/process: P diagram

design increases the cost. The parameter design is dealt with here.
When the target value of a response is given as illustrated in Figure 7.3, the
Taguchi method determines the optimum setting of control factors, so that the
variation of a response is minimized, although uncontrollable factors (noise
factors) exist. Equation 7.12 can be regarded as the index to find a robust design.
Suppose that we have a scale factor s to adjust the current mean to the target value.
The scale factor s is described as
m
s (7.13)
P
When the current mean is adjusted to the target value, the average loss function of
Equation 7.12 is changed to
§§ m 2 2
· §m · · V2
Qa k ¨ ¨¨ P m ¸¸ ¨¨ V ¸¸ ¸ km2 (7.14)
¨© P ¹ © P ¹ ¸¹ P2
©
To enhance additivity of the effect of the control factors, Equation 7.14 is

transformed to
P2
K 10 log10 (7.15)
V2
Equation 7.15 is the ratio of the power of the signal factor P and the power of
noise factors V . Thus, it is called the S/N ratio. Maximizing Equation 7.15 is
equivalent to minimizing Equation 7.14. That is, a robust design is obtained by
maximizing Equation 7.15.
Table 7.3. Examples of S/N ratios
Characteristic type S/N ratio
mf 2
Nominal-the-best with scale factor: K 10 log or K 10 log s f 2
sf 2
ª1 n º
Smaller-the-better K 10 log «
¬« n
¦f
i 1
i
2
»
¼»
ª1 n
1 º
Larger-the-better K 10 log «
«¬ n
¦f
i 1
2
i »
»
¼
Robust Design 403
Table 7.4. The loss function of Example 7.3
Exp. no. b1 b2 f Loss function

1 í0.1 í0.1 í2.18 0.0324
2 í0.1 0.1 í2.02 0.0004
3 0.1 í0.1 í2.02 0.0004
4 0.1 0.1 í1.78 0.0484
As in the previous example of Figure 7.2, the response with target value m is
referred as the nominal-the-best type characteristic. On the contrary, the responses
with target values of zero and infinity are referred as the smaller-the-better type
and the larger-the-better type, respectively. The S/N ratios of the smaller-the-
better type and the larger-the-better type can be derived by a similar procedure as
for the nominal-the-best type. As the index of robustness, the S/N ratio of
Equation 7.15 is preferred to the average loss function of Equation 7.14. The
examples of S/N ratios are summarized in Table 7.3.
Example 7.3 [The Loss Function and S/N Ratio of a Mathematical

Problem]
In Example 7.1, b1 0.1, 0.1 and b2 0.1, 0.1. Calculate f, the loss function
and the S/N ratios for all the combinations. The target value is the mean of
Example 7.1.
Solution
The loss function is the nominal-the-best type and the target value is 2 from
Example 7.1. Table 7.4 shows f’s and the loss functions for all combinations. The
constant k of the loss function is set as 1. The mean and the variance for the loss
function are ȝ = 0.0204 and V 2 = 0.000576. If we use Equation 7.15, the S/N ratio
is 1.41162.
Example 7.4 [The Loss Function and S/N Ratio of the Three Bar Truss]
In Example 7.2, E has the values 203, 208, 209 and 211 GPa. Calculate the loss
functions and the S/N ratios with Equation 7.15. The target values are the means
in Example 7.2.
Solution
The loss functions have the nominal-the-best type and the target values are mG x
0.000686, mG y 0.000643 from Example 7.2. Table 7.5 shows the loss
functions. The constant k for the loss function is 1.
Table 7.5. Loss functions for Example 7.4
Exp. Loss function Loss function

E (GPa) G x (m) G y (m)
no. for G x for G y
1 203 0.000697 1.11 u 10 10 0.000653 9.43 u 10 11

2 208 0.000680 3.83 u 10 11 0.000637 3.58 u 10 11
3 209 0.000677 8.92 u 10 11 0.000634 8.15 u 10 11
4 211 0.000670 2.51 u 10 10 0.000628 2.26 u 10 10
The means of the loss functions are P G x = 1.23 u 10 10 and P G y = 1.09 u 10 10.
The variances are V G2x = 8.32 u 10 21 and V G2y = 6.69 u 10 21 . If we substitute
these into Equation 7.15, the S/N ratios are 2.57 and 2.53.
7.3.3 Parameter Design
The Taguchi method is classified into two submethods. One is the parameter
design, where the loss function is reduced by determining the values of parameters
when noises or tolerances exist. The other is the tolerance design, where the loss
function is reduced by determining the tolerances. Since the tolerance design is
usually expensive, the parameter design is generally utilized. In this section, the
parameter design is explained.
As mentioned earlier, the Taguchi method is generally used in experiments;
however, it is explained here from the viewpoint of application of the method to
the design. The method evaluates the characteristic function, which is equivalent
to the objective function of optimization. The characteristic function is not
evaluated by experiments but is numerically calculated. The Taguchi method has
been applied to various areas. In this section, only the basic aspects are dealt with.
References present many other cases (Taguchi 1987, Phadke 1989).
There are two goals in performing robust design. One is to minimize the
variability produced by the noise factors. The other is to make the mean value
close to the target value. The design that attains one goal is not usually consistent
with the one that attains the other goal. To meet the two goals, Taguchi developed
a two-step optimization strategy. The first step is to reduce the variation, and the
second step is to adjust the mean on the target. The procedure is illustrated in
Figure 7.4. The first step in Figure 7.4a concentrates on minimization of the
variation, while the mean is ignored. In the second step of Figure 7.4b, the mean is
moved to the target value while somewhat sacrificing the improved variation.
The first step is to find the optimum setting of control factors (design variables)
to maximize the S/N ratio or to reduce the variation. Orthogonal arrays are
employed to assign the control factors to an experimental matrix. The best
Robust Design 405
ProbabilityG Low variability

densityG performance
function
Step 1
Target f
(a) The first step
Probability G
densityG Low variability
performance is
function
shifted onto the
target Step 2
Target f
(b) The second step
Figure 7.4. Two steps of the Taguchi method
condition can be selected from the full combinations of control factors. However,
it is efficient to select the smallest orthogonal array because full combinations are
costly in most cases. The orthogonal arrays are explained in detail in Chapter 6.
Experiments are performed repeatedly for each row of an orthogonal array and the
mean and the variance are calculated from the repeated experiments. The noises
are automatically included in the repetition.
The S/N ratio is calculated for each row of an orthogonal array. Then, the
optimum setting of control factors is determined by performing ANOM of the S/N
ratios. A one-way table is utilized in the ANOM process. If the average loss of
Equation 7.12 is defined as the characteristic function to find a robust design, the
effect on the variation is confounded by the effect on the mean. Then the factors
are determined to have the mean close to the target although the variance is large.
This may be opposite to the goal of robust design, which tries to obtain a small
variance. Therefore, it is assumed that there is a scale factor that can move the
mean. Then the S/N ratio, which is the variance when the target is met, is utilized.
It is noted that the interactions should be allocated to the columns of an orthogonal
array. However, it is difficult to identify the existence of the interactions before
experiments. This is an obstacle in using orthogonal arrays; therefore, a
confirmation experiment should be performed with the results from the one-way
table.
A computer simulation model provides the same output with the same input
since it is a deterministic method. While this is a disadvantage, the computer
simulation model has the advantage that it can define the design parameters p and
the noise factors z exactly. In this case, Taylor series expansion, Monte-Carlo
simulation or an outer array can be utilized to obtain the variations. An outer array
is an orthogonal array that is used for a row of the orthogonal array (inner array).
The outer array has multiple cases of experiments for a row; therefore, it
accommodates the repeated experiments of an experiment case.
The second step is to set the mean to the target value by selecting a control
factor called the scale factor. The scale factor should only have an effect on the
mean, while the S/N ratio is maintained. In general experiments, time can be one
of the scale factors. However, it is difficult to find a scale factor in a general
design. Furthermore, it is possible that no scale factor exists and this prevents the
designer from applying the Taguchi method directly to the design. Therefore, each
row can use a function such as the loss function in Equation 7.12, which considers
the mean and the variance.
When it is impossible to find a scale factor, only the first step can be performed
out of the two steps. In this case, the choice of the characteristic function is crucial.
Moreover, the design problem generally has constraints as shown in Equation 3.1.
Thus, the characteristic function should include the constraints. When the
Outer array
Exp. Noise factor
no. z1 z 2 z 3 z 4
2
s f , v, <1
1 1 1 1 1
2 1 2 2 2
Inner array
. . . . .
Exp. Control factor
. . . . .
no. b1 b2 b3 b4
Nout . . . .
1 1 1 1 1
2 1 2 2 2 Exp. Noise factor
. . . . . no. z1 z 2 z 3 z 4
. . . . . 1 1 1 1 1
s 2f , v, < N in
Nin . . . . 2 1 2 2 2
. . . . .
. . . . .
Nout . . . .
Figure 7.5. Inner and outer arrays

Robust Design 407
experiments are performed repeatedly, tolerances should be considered. Since

repeated experiments give the same results in computer simulation, the outer array
is adopted as shown in Figure 7.5. The outer array is an orthogonal array as well.
N in is the number of rows of the inner array and N out is the number of rows of
the outer array. Their usage is explained in an example.
When we have constraints, the characteristic function is defined as
< (b) w1 P f w2V f P (b) (7.16a)
P (b) max(0, h , g ) u s (7.16b)
Equation 7.16 has a similar concept to that of Equations 6.60 and 6.61. P f and
V f are the mean and the standard deviation of the objective function, respectively.
The designer can determine the weighting factors w1 and w2 with values from 0
to 1.
< (b) of Equation 7.16 is considered as the characteristic function and the
design method using orthogonal arrays in Chapter 6 is directly used. When we use
Equation 7.16, it is not easy to obtain a robust design when the constraint violation
is large. Thus, it is recommended that the Taguchi method be used in the feasible
region. It is noted that the characteristics of the method, which are explained in
Chapter 6, are also valid here.
The one step method using calculations or computer simulations is explained.
It is assumed that the number of design variables (control factors) is n = 4 and the
number of levels for each variable is three.
Step 1. An appropriate orthogonal array (inner array) is chosen according to
the number of design variables and levels. It is assumed that the
orthogonal array L9 (3 4 ) in Table 6.13 is chosen.
Step 2. An outer array is chosen by considering the tolerances and noises.
The process is presented in Figure 7.5. The number of total experiments is
N in u N out .
Step 3. By using the outer array, < (b) of Equation 7.16 is evaluated for
each row of the inner array. Define K { < (b) .
Step 4. The one-way table as shown in Table 6.16 is made. For mbij , j = 1, 2,
3 in Table 6.16, the smallest level oi , i=1,…,4 is found for the design
variable bi .
Step 5. The design solution is bi , i = 1, …,4, which has the oi th level. With
these levels, < (b) is calculated again for a confirmation experiment.

Compare this with the best row of the inner array in Step 3. The better one
is the final solution.
It is noted that the estimated value can be utilized as in Section 6.5. That is, the
solution is arranged in ascending order and the best one is included in the set in
Step 5 for comparison.
Example 7.5 [Application of the Taguchi Method to a Mathematical

Problem]
Apply robust design to Example 6.14. All the factors have three levels of 1 , 0
and 1. The tolerances are 0.1 d zi d 0.1 .
(1) Apply the above steps for robust design.

(2) Rearrange the estimated values in ascending order and find the best value.
Compare the solution with that of (1) and find the most robust solution.
Solution
(1) Robust design process

Since there are four design variables with three levels, the L9 (34 )
orthogonal array is chosen for the inner array as shown in Table 7.6. The
L9 (3 4 ) orthogonal array is utilized for the outer array as well. The outer
array for the first each row of the inner array is shown in Table 7.7.
In Step 3, the characteristic function is evaluated. From the outer array,
the mean of the objective function P f and the standard deviation V f are
evaluated and substituted into Equation 7.16. Then the characteristic
function K { < (b) is obtained for the corresponding row of the inner array.
The value of 0.5 is given to the weighting factors w1 and w2 , and the scale
factor s is set by 100. The results for the inner and outer arrays are shown
Table 7.6. The inner array of Example 7.5
Design variable
Exp. no. < (b)
b1 b2 b3 b4
1 í1 í1 í1 í1 17.32
2 í1 0 0 0 5.13
3 í1 1 1 1 í2.88
4 0 í1 0 1 9.25
5 0 0 1 í1 í11.12
6 0 1 í1 0 12.17
7 1 í1 1 0 33.00
8 1 0 í1 1 16.29
9 1 1 0 í1 í5.17
Robust Design 409
Table 7.7. The outer array for the first row of Table 7.6
Tolerance of the design variable

Exp. no.
z1 z2 z3 z4 f (b)
1 í1.1 í1.1 í1.1 í1.1 32.45
2 í1.1 í1.0 í1.0 í1.0 29.71
3 í1.1 í0.9 í0.9 í0.9 27.05
4 í1.0 í1.1 í1.0 í0.9 30.22
5 í1.0 í1.0 í0.9 í1.1 26.03
6 í1.0 í0.9 í1.1 í1.0 30.83
7 í0.9 í1.1 í0.9 í1.0 26.54
8 í0.9 í1.0 í1.1 í0.9 31.34
9 í0.9 í0.9 í1.0 í1.1 27.13
Table 7.8. The one-way table for Example 7.5
Level
Design variable
1 2 3
b1 6.53 3.44 14.71
b2 19.86 3.44 1.38
b3 15.26 3.07 6.34
b4 0.35 16.77 7.56
in Tables 7.6 and 7.7.

The one-way table is made as shown in Table 7.8. The levels (2 3 2 1)
are chosen from the one-way table. The design variables are b =
[0 1 0 1]T . An experiment is carried out for confirmation. In the
confirmation experiment, the mean of the objective function is í 9.97, the
variance is 3.75 and the characteristic function is – 3.11. Constraints are
not violated. The solution is compared with the best row in the orthogonal
array. The fifth row in Table 7.6 is better; therefore, it is chosen for the
final solution. The solution is b = [0 0 1 1]T and the mean and
variance are í 24.97, 2.73, respectively.
(2) Design by the estimated values in ascending order
As explained in Chapter 6, the estimated values for the objective function
and constraints are calculated for all combinations and the characteristic
function is estimated. They are arranged in ascending order and the
confirmation experiments are carried out from the first row of Table 7.9. It
Table 7.9. Table in ascending order for Example
Design variable Estimated Constraint

Order
b1 b2 b3 b4 value violation
1 0 1 1 í1 í13.18 Satisfied
2 1 1 1 0 í12.15 .
. . . . . . .
. . . . . . .
. . . . . . .
81 1 í1 í1 í1 72.17 .
is examined if the constraints are violated. Since the constraints are

satisfied in the first row as shown in Table 7.9, the first row is chosen as
the design solution.
The mean of the objective function is í 28.97 and the variance is 2.61.
Since the design is obtained from the first row, the confirmation
experiment is not performed from the second row. The characteristic
function is 13.18 and it is better than the solution of (1). Therefore, it is
chosen as the final solution.
Example 7.6 [Robust Design of the Three Bar Truss Using the Taguchi
Method]
In Example 6.15, the levels of the design variables are shown in Table 7.10 and the
tolerances are 0.00002 d z1 , z 2 d 0.00002, 0.000002 d z3 d 0.0000015. Solve
the following problems:
(1) Use the above method with five steps for robust design.
(2) Find a design solution from the estimated values in ascending order,
compare it with that from (1), and find the final solution for robust design.
Solution
Table 7.10. Levels of design variables for Example 7.6
Level
Design variable
1 2 3
b1 0.0001 0.0002 0.0004
b2 0.0001 0.0002 0.0003
b3 0.000015 0.00002 0.00003
Robust Design 411
Experiment Design variable

< (b)
number b1 b2 ignored b3
1 í1 í1 í1 í1 91.666
2 í1 0 0 0 80.220
3 í1 1 1 1 67.911
4 0 í1 0 1 4.420
5 0 0 1 í1 0.701
6 0 1 í1 0 0.850
7 1 í1 1 0 0.964
8 1 0 í1 1 1.122
9 1 1 0 í1 1.231
(1) The robust design process

Since we have three design variables with three levels, the L9 (3 4 )
orthogonal array is selected for the inner array and the design variables are
allocated as shown in Table 7.11. Each row has the outer array L9 (3 4 ) for
repeated experiments.
The characteristic function K { < (b) is evaluated for each row of the
inner array. The weighting factors w1 and w2 have 0.5 and the scale
factor s is set by 100. The results are shown in Table 7.11.
Based on the results in Table 7.11, the one-way table is made as shown in
Table 7.12. The levels (3 3 3) are selected. The design variables are b =
[0.0004 0.0003 0.00003]T . A confirmation experiment is carried out with
the selected design variables. The mean and variance of the objective
function are 2.5148 kg and 0.006941, respectively, and the characteristic
function is 1.261. The constraints are not violated. This solution is
Level
Design variable
1 2 3
b1 79.93 1.92 1.11
b2 32.28 27.35 23.33
b3 31.20 27.34 24.41

Table 7.13. Table in ascending order for Example
Design variable Estimated Constraint

Order
b1 b2 ignored b3 value violation
1 0.0004 0.0003 . 0.000015 2.622 Satisfied

2 0.0002 0.0003 . 0.000015 3.437 .
3 0.0004 0.0001 . 0.00003 4.787 .
. . . . . . .
. . . . . . .
27 0.0001 0.0001 . 0.000015 87.855 .
compared with the best one in the orthogonal array. The fifth row is the
best in Table 7.11 and it is better than that from the one-way table.
Therefore, the fifth row of Table 7.11 is selected as the final solution,
which is b = [0.0002 0.0002 0.000015]T . The mean and variance of the
objective function are 1.396 kg and 0.00694, respectively.
(2) Design by the estimated values in ascending order
The characteristic function is estimated in the same way as in Example 7.5.
The estimated values are arranged in ascending order and the constraint
violation is examined from the first row. Since the first row satisfies the
constraints as shown in Table 7.13, it is selected as the design solution.
The mean and variance of the objective function are 1.415 kg and
0.006941, respectively. Since the design is obtained from the first row, the
confirmation experiment is not performed from the second row. The
characteristic function is 0.711. Therefore, the solution of (1) is chosen as
the final solution since it is better.
7.3.4 Various Methods Using the Taguchi Method

The Taguchi method has been widely applied to various fields due to its simple
and explicit procedures. Even beginners can easily use the Taguchi method. Note
that the Taguchi method has been regarded as an excellent tool in design processes
using experimental analysis approaches. Based on the concept of the traditional
Taguchi method, researchers of design methodologies have tried to use Taguchi’s
outstanding ideas in their fields. However, there are some conflicts between the
traditional Taguchi method and the design methodologies. Thus, research has
been carried out to overcome these difficulties.
First, the difficulties with the S/N ratio were identified (Montgomery 1991, Lee
et al. 1996). As shown in Equation 7.15, the variance and the mean are
compounded in the characteristic function of the nominal-the-best type. In this
case, the S/N ratio can be modified as
Robust Design 413
K 10 log P 2 10 logV 2 (7.17)
Equation 7.17 is useful when the factors can be divided into the scale factor that
affects the mean, and the control factors that affect the variance. When this
division is not possible, the solution might not be very good because the effects on
the mean and the variance are compounded in the S/N ratio. That is, the two-step
optimization method proposed by Taguchi is valid when the factors can be divided.
It is noted that the method of the previous section, which uses computer simulation,
uses the one-step method. Phadke defined qualification of the characteristic
function (Phadke 1989). However, it is quite difficult to use in practical design.
The mean and variance are compounded in the S/N ratio in the cases of the
smaller-the-better and the larger-the-better types as well. Therefore, the direct use
of the S/N ratios in Table 7.3 may lack pertinence. It is the same for the
characteristic functions for the smaller-the-better and the larger-the-better types.
Therefore, when the factors are not obviously divided, appropriate solutions may
not be obtained from the S/N ratio. Box proposed a general transformation
approach (Box 1988). However, this solution may not be precise for highly
nonlinear cases. Montgomery (Montgomery 1991) and Leon (Leon et al. 1987)
showed that minimizing the variance is more effective than using the S/N ratio.
Lee (Lee et al. 1996) recommended the use of a multiobjective function with
weighting factors. Wu evaluated robustness for the nominal-the-best type problem
with the indices made by the S/N ratio, mean standard deviation (MSD) and the
variance (Wu and Chen 2002).
Second, the experiment of each row of an orthogonal array is carried out
repeatedly to consider noise. In a design using computer simulation, the outer
array is employed for this. Since many experiments are needed, they can be
inefficient. Welch (Welch et al. 1990) and Wu (Wu and Chu 2002) proposed the
use of a combined array that unifies the inner and outer arrays.
Third, the number of characteristic functions is one in the original Taguchi
method. Phadke proposed a method by using trade-off for the problem with
multiple characteristic functions (Phadke 1989). However, this method can
increase the uncertainty for robustness. Shiau solved the problem with multiple
characteristic functions by defining the weighting factors. It is noted that the
weighting factors are subjectively determined by the user (Shiau 1990). There are
many methods to solve the problem (Tong and Su 1997, Chen 1997, Vining 1998,
Tsui 1999).
Fourth, the Taguchi method cannot consider many constraints that are
generally considered in design. To overcome this problem, Lee proposed the
following steps (Lee and Park 2001): (1) An optimum solution is obtained by a
classical optimization method regardless of the tolerances. It is considered that the
target is obtained. (2) Candidate values are defined around the optimum solution.
An orthogonal array is defined with the candidate values and the method of the
previous section is utilized. Recent research on the Taguchi method has been
carried out to overcome the difficulties of the classical Taguchi method.
Table 7.14. Candidate values of design variables in Example 7.7
Level
Design variable
1 2 3
b1 0.0001 0.0002 0.0003
b2 0.00005 0.0001 0.00015
b3 0.5Eí8 1.0Eí8 1.5Eí8
Example 7.7 [Robust Design of the Three Bar Truss by Using the
Taguchi Method after Optimization]
The optimum of Example 7.6 is b* = [0.0002034 0.0001064 0.00000001]T .
The candidate values around the optimum are defined as shown in Table 7.14. The
tolerances are 0.00002 d z1 , z 2 d 0.00002 , 0.000000002 d z3 d 0.000000002 .
Solve the following problems by using the inner and outer arrays.
(1) Carry out robust design by using the steps in Section 7.3.3.
(2) Arrange the solutions in ascending order by using the estimated values.
Evaluate the best value and obtain the final solution by comparing it with
the solution of (1).
Solution
(1) Robust design process
Table 7.15. The inner array of Example 7.7
Design variable
Exp. no. Ȍ (b)
b1 b2 ignored b3
1 í1 í1 í1 í1 109.425
2 í1 0 0 0 102.359
3 í1 1 1 1 102.418
4 0 í1 0 1 109.576
5 0 0 1 í1 5.112
6 0 1 í1 0 1.617
7 1 í1 1 0 109.778
8 1 0 í1 1 5.300
9 1 1 0 í1 0.799
Robust Design 415
Level
Design variable
1 2 3
b1 104.73 38.77 38.63
b2 109.59 37.59 34.94
b3 38.45 71.25 72.43
Three levels for each design variable are considered. The L9 (34 )
orthogonal array is selected for the inner and outer arrays. The outer arrays
for each row of the inner array are evaluated. The characteristic function
K { < (b) is calculated from the mean P f and the variance V f . The
results are shown in Table 7.15. The weighting factors and the scale factor
are the same as those of Example 7.6.
The one-way table is made as shown in Table 7.16. The solution from
Table 7.16 is (3 3 1) levels. The design variables are b =
[3.0E 4 1.5E 4 5.0E 9]T .
The confirmation experiment is carried out with the determined levels.
The mean is 1.591 kg, the variance is 0.00691 and the characteristic
function is 0.799. The constraints are satisfied. The solution is compared
with the best one from the orthogonal array. The ninth row of the
orthogonal array is the best and it is the same as the one previously
obtained. Thus, it is selected as the solution.
(2) Arranging the solutions in ascending order using the estimated values
The estimated values for the objective function and the maximum
Table 7.17. Arrangement of Example 7.7 in ascending order
Design variable Estimation of the

Constraint
Order characteristic
b1 b2 ignored b3 violation
function
1 0.0002 0.00015 . 0.5E–8 1.62 Violated

2 0.0003 0.00015 . 0.5E–8 5.02 Satisfied
3 0.0002 0.0001 . 1.0E–8 5.11 .
. . . . . . .
. . . . . . .
27 0.0003 0.00005 . 0.5E–8 109.79 .
constraint violation are calculated for the full combination. They are
arranged in ascending order as shown in Table 7.17 and the confirmation
experiment is carried out from the first one. The second row satisfies the
constraints and it is selected as the solution. The mean is 1.591, the
variance is 0.00691 and the characteristic function is 0.799. Since this is
the solution of (1), it is selected as the final solution. The confirmation
experiment is not carried out from the third row of Table 7.17.
7.4 Robust Optimization
7.4.1 Introduction
The aforementioned robust design problems have been solved based on well
developed mathematical optimization techniques during the last decade. Since the
Taguchi method uses an orthogonal array and design variables are defined in a
discrete space, it is difficult to treat a wide design range. Design in a discrete
space may be useful; however, design in a continuous space is often required.
Generally, a design may have many design constraints that the Taguchi method
may not resolve. Because optimization techniques can easily handle the
constraints, robust optimization based on optimization techniques has been studied.
In a general design problem, we can hardly separate the mean and variance of
design variables. Optimization techniques can obtain design variables by
considering the mean and the variance simultaneously. The goal of robust
optimization is to solve these problems.
Robust optimization is formulated as Equation 7.2. The functions F and G in
Equation 7.2 are defined to enhance robustness and the robustness of functions is
defined as follows: (1) Robustness of the objective function can be achieved by
C
B
A
b
Figure 7.6. Robustness of the objective function

Robust Design 417
reducing the change of the objective function with respect to the changes of
tolerances for the design variables. Figure 7.6 presents an objective function that
has a minimum at the peak point A, where the objective function can drastically
change by small variations of the design variables. Thus, the local minimum point
B or C is better for an insensitive (robust) design. (2) Robustness of constraints
means that all the constraints are satisfied within the range of tolerances for the
design variables, i.e., the feasible region is reduced as illustrated in Figure 7.7.
Probabilistic consideration of the objective function and constraints is known as
reliability optimization or stochastic optimization. They will be briefly mentioned
later.
Now let us consider the formulation of robust optimization as follows:
Find b Rn (7.18a)
to minimize [ P f (b, p) V f (b, p)] (7.18b)
subject to g j (b z b , p z p ) d 0 , j 1, ,m (7.18c)
bL d b zb d bU (7.18d)
It is noted that the objective function in Equation 7.18b is defined in terms of the
mean and the standard deviation, i.e., the variance. Equation 7.18c means that all
the constraints should be satisfied despite the noise in design variables and
parameters. Therefore, robust optimization methods have been efficiently
developed to solve the problem defined as Equation 7.18.
g 1, new
g1 Robust
b2 g2,new
feasible region
'b2 g2
b*
'b1
b1
Figure 7.7. Change of the feasible region

7.4.2 Robustness of the Objective Function
Robustness of the objective function has been emphasized more than that of
constraints because an insensitive design of the objective function is pursued in
robust design. The mean P f and the variance V f of the objective function are
approximated as follows:
P f # f (ȝ b , ȝ p ) (7.19a)
2 2
n § wf · 2 o § wf · 2
V 2f # ¦ ¨¨ ¸¸ V b ¦ ¨¨ ¸¸ V p (7.19b)
i 1© wbi i 1© wpi
i i
¹ ¹
where ȝ b is the mean vector of design variable vector b, ȝ p is the mean vector of
the design parameter vector p, V bi is the standard deviation of bi and V pi is the
standard deviation of pi . If b and p are assumed to have normal distribution, the
mean P f of Equation 7.19a can use the nominal value of the objective function
during the optimization process. Note that Equation 7.19 is meaningful when the
interactions of the design variables and the design parameters are negligible, and
their deviations are small.
If we assume that the design variable vector b and the parameter vector p are
symmetrically distributed around the nominal values, then f(b) in Equation 3.1b
can be directly used for the mean P f in Equation 7.19a. Thus, the mean can be
easily included in the optimization process. It has been reported that the variance
is the same as the square mean of the forward difference, backward difference and
central difference, which mean the design sensitivity. It is noted that Equation
7.19 can be used based on the assumption that the design variables and parameters
are statistically independent and their distribution ranges are small.
The variance is made of the sensitivity, i.e. first derivative, as shown in
Equation 7.19b. If the objective function includes the sensitivity, second-order
derivatives of the objective are required in the optimization process. However,
calculation of the second-order derivatives of the objective function is so
expensive that it is not recommended. Since the constraints of Equation 7.18c are
represented in terms of the variance, second-order derivatives of the constraints are
also needed. Thus, robust optimization without the second-order derivatives is
recommended.
In the simplest method, the objective function is defined by the sensitivity and
the second-order derivatives are directly evaluated. This method is quite
expensive but reliable. It has some limits in that the solution may include local
information. In the early research on robust optimization, robustness was obtained
by minimizing the sensitivity information of the objective function. In this case,
we also need the second-order derivatives of the objective function. When the
weight is the objective function in structural optimization, the objective function is
Robust Design 419
an explicit function of the design variables and the second-order derivatives are
easy to evaluate. However, robustness of the weight may not be meaningful in
engineering.
In some research, the second-order derivatives are used for robustness of the
objective function and only first-order derivatives are used for robustness of
constraints. This will be explained later. Calculation of the second-order
derivatives can be viable as computer capacity increases. However, we still have
difficulties with large-scale problems.
Various methods have been proposed in order to avoid calculation of the
second-order derivatives. The sensitivity index (SI), which represents the variance
of the worst case of the current design, has been proposed. The worst case is
found by orthogonal arrays to save computational time (Sundaresan et al. 1995).
The objective function can be approximated by the regression model, which is an
explicit function. In other words, the function is approximated to a quadratic
equation by using RSM. Then evaluation of the second-order derivatives is quite
easy. Many researchers have utilized RSM to approximate the objective function.
When a function is approximated by RSM, the sensitive part (the part where
the first-order derivative is large) can be ignored. In this case, it is noted that the
sensitive part can be determined as the final solution. Robust optimization can be
obtained by the direct use of the Taguchi method, as explained in the previous
section. An outer array is utilized for each row of the inner array to calculate the
mean and the variance. In this method, sensitivity calculation is not required.
Example 7.8 [Robustness of the Objective Function]

In Example 6.14, b1 b2 b3 b4 1 . The tolerances are 0.1 d zi d 0.1 and have
uniform distributions. Calculate the mean and the variance of the objective
function in Equation 7.19.
Solution
This problem does not have parameters. The mean of b1 is
1.1 x
E (b1 ) ³ dx 1 (7.20)
0.9 .2
0
The means of other design variables are obtained in the same manner, ȝ b =
[1 1 1 1]T . Using Equation 7.19a, the mean of the objective function is
P f # f ( P b ) = í19 (7.21)
The variance of each design variable is calculated as follows:
1.1 (bi 1) 2
V b2i ³ dbi 0.00333, i 1, ..., 4 (7.22)
0.9 0.2
The sensitivities are
wf wf
2b1 5 = í3, 2b2 5 = í3, (7.23a)
wb1 wb2
wf wf
4b3 21 = í17, 2b4 7 = 9 (7.23b)
wb3 wb4
Using Equation 7.19b, the variance of the objective function is
2
n § wf · 2
V 2f # ¦ ¨¨ ¸¸ V b = (9 + 9 + 289 + 81)Ý0.00333 = 1.292 (7.24)
i 1© wbi
i
¹
Example 7.9 [Robustness of the Objective Function for the Three Bar
Truss]
Suppose each variable of Example 6.15 has the value of the second level. The
tolerances are 0.00002 d z1 , z 2 d 0.00002 , 0.000002 d z3 d 0.0000015 and
they have uniform distributions. Calculate the mean and the variance of the
objective function in Equation 7.19.
Solution
The design variables are uniformly distributed. If we use the equations in Example
7.8, the mean is ȝ b = [0.0002 0.0002 0.00001975]T . The mean of the
objective function by using Equation 7.19a is P f = 1.415 kg.
The variance of each design variable is
2.2 E 4 ( x 2.0 u 10 4 ) 2
V b21 , V b22 ³ dx 1.333 u 10 10 (7.25a)
1.8 E 4 4.0 u 10 5
2.15 E 5 ( x 1.975 u 10 5 ) 2
V b23 ³ dx 1.021u 10 12 (7.25b)
1.8 E 5 3.5 u 10 6
The sensitivities are
wf wf wf
2 U = 3917.372, U = 2770, 2 U = 3917.372 (7.26)
wb1 wb2 wb3
Using Equation 7.19b, the variance of the objective function is

Robust Design 421
2
n § wf ·
V 2f # ¦ ¨¨ ¸¸ V b2 (7.27)
i 1 wb ©
i
i ¹
= (15345800 + 7672000)Ý1.333 u 10 10 + 15345800Ý1.021 u 10 12

= 0.00308
7.4.3 Robustness of Constraints
As mentioned earlier, the constraints can be transformed into Equation 7.18c in the
case where noise exists. Treatment of constraints varies according to the detailed
definition of Equation 7.18c. First, we can reduce the feasible region. The
feasible region is reduced by the noise as illustrated in Figure 7.7. In other words,
the constraints in Equation 7.18c must be satisfied even though we have z b and
z p . Generally, z b and z p are considered as tolerances. These can be defined by
ranges or distributions.
When noise exists as distribution, Equation 7.18c can be expressed as
g j , new { g j kV g2 j d 0 (7.28)
where k is a user-defined constant depending on the design purpose and V g j

denotes the standard deviation of the constraint g j . V g j can be approximated by
using Equation 7.19b as follows:
2 2
n§ wg j · 2 o § wg j · 2
V g2 j # ¦ ¨¨ ¸ V b ¦¨
¸ i i 1¨ wp
¸ Vp
¸ (7.29)
1© wbi ¹
i
i © i ¹
If the noise is given by a range, the worst case of g j , new and Equation 7.18d
become the following:
n wg j o wg j
g j ,new g j k bj ¦ | zib | k jp ¦ | zip | (7.30a)
i 1 wbi i 1 wpi
b L d b z bmax d b U (7.30b)
where | z ib | and | z ip | denote the maximum values of the tolerance ranges, and
k bj and k jp are user defined constants depending on the design purpose.
Equations 7.29 and 7.30a have the first-order derivatives of constraints. Thus, the
second-order derivatives of constraints are needed if mathematical programming is
employed as an optimizer. This is similar to the case of the objective function. In

some research, first-order derivatives are regarded as constants.
The design point b * of Figure 7.7 is the optimum point when the design
variables have tolerances whose ranges are box-type domains as illustrated in
Figure 7.7. g j , new is defined to satisfy the constraints within the ranges as
follows:
g j , new max{g j (b), b the hypersurface of Figure 7.7} (7.31)
In addition to the methods mentioned above, g j , new can be defined based on the
theory of tolerance design, but the concept is very similar.
Probabilistic analysis is adopted from the theories of reliability optimization. It
is also being used in design that considers uncertainties. Satisfaction of constraints
can be written as probabilistic forms for normal distribution of noise as follows:
P[ g j (b z b , p z p ) d 0] t Pj ,0 j 1,..., m (7.32)
where P[x] denotes probability such that Pj ,0 represents the probability of

satisfaction for constraints, while the noise z b or z p is of normal distribution. If
the probabilistic distributions of noise are given, Equation 7.32 can be rewritten as
follows:
P[ g j (b z b , p z p ) d 0] b p b p
³ d ( z , z ) dz dz , j 1,..., m (7.33)
g j ( b z , p z p )d 0
b
where d (z b , z p ) is the joint probability density function of probabilistic variables

z b and z p . Evaluation of Equation 7.33 is very difficult to handle; therefore, an
approximate equation can be used to reduce the computational cost.
To approximate Equation 7.33, the moment matching formulation can be used.
In the moment matching method, the probability is simplified by assuming that the
constraints have normal distribution. This is one of the most popular methods. If
the constraints have normal distribution, Equation 7.32 becomes
g j (b , p )
Vgj Ij(pj)
1 2 1 2
³ e 2T 2
dT t ³ e 2t 2
dt (7.34)
f 2S f 2S
where I j ( p j ) is the variable for the probability p j of the normal distribution and
is given as
Robust Design 423
g j (b z b , p z p ) g j (b, z )
Ij(pj) (7.35)
V gj
From the inequality of Equation 7.34, Equation 7.32 can be modified as follows:
g j , new { g j I j ( p j )V g2 j d 0 (7.36)
Equation 7.36 provides the basis for Equation 7.28. Equation 7.36 can be
employed in robust optimization; however, Equation 7.28 is used more often.
Example 7.10 [Robustness of Constraints]

In Example 6.14, b1 b2 b3 b4 1 . The tolerances are 0.1 d zi d 0.1 and
have uniform distributions. Evaluate the variance of each constraint in Equation
7.29 and calculate Equations 7.28 and 7.30a with k = 5.
Solution
The constraints are
b1 2 b2 2 b3 2 b4 2 b1 b2 b3 b4
g1 = 1 d 0 (7.37a)
8
b1 2 2b2 2 b3 2 2b4 2 b1 b4
g2 = 1 d 0 (7.37b)
10
2b1 2 b2 2 b3 2 2b1 b2 b4
g3 = 1 d 0 (7.37c)
5
The sensitivities of the constraints are shown in Table 7.18. The variance of
each design variable has been obtained in Example 7.8. Using Equation 7.29, the
variance of each constraint can be obtained as follows:
2
n § wg ·
V g21 # ¦ ¨¨ 1 ¸¸ V b2i = 0.001041 (7.38a)
i 1© wbi ¹
V g2 2 = 0.000999, V g23 = 0.005594 (7.38b)
The new constraints from Equation 7.28 are
g1, new = í0.5+5Ý0.001041 = í0.494795 (7.39a)
g 2, new = í0.6+5Ý0.000999 = í0.595005 (7.39b)

Table 7.18. Sensitivity of Example 7.10
w w w w
Constraint
wb1 wb2 wb3 wb4
g1 0.375 0.125 0.375 0.125
g2 0.1 0.4 0.5 0.3
g3 1.2 0.2 0.4 –0.2
g 3, new = í0.2+5Ý0.005594 = í0.172030 (7.39c)
When the noise has ranges, Equation 7.30a is utilized. The constraints are
g1, new = í0.5+5Ý0.1Ý (0.375+0.125+0.375+0.125) = 0 (7.40a)
g 2, new = í0.1, g 3, new = 0.8 (7.40b)
Example 7.11 [Robustness of Constraints for the Three Bar Truss]

In Example 6.15, suppose each variable has the value of the second level in Table
7.10. The tolerances are 0.00002 d z1 , z 2 d 0.00002 and 0.000002 d z3 d
0.0000015 and have uniform distributions. Obtain the variance in Equation 7.29
for each constraint and calculate Equations 7.28 and 7.30a with k = 5.
Solution
The constraints are
V1 1 3b3 0.5(2 2b2 2b3 )

g1 1 P 1 (7.41a)
140 MPa 140 MPa 2b1b2 2 2b1b3 2b2 b3
V1
g1 d 0 , g 2 = 1 d 0 (7.41b)
100 MPa
V2 V 2
g3 = 1 d 0 , g4 = 1 d 0 (7.41c)
140 MPa 100 MPa
V3 V 3
g5 = 1 d 0 , g6 = 1 d 0 (7.41d)
140 MPa 100 MPa
The sensitivities of the constraints are shown in Table 7.19.

Robust Design 425
Table 7.19. Sensitivity of Example 7.11
w w w
Constraint
wb1 wb2 wb3
g1 í4463.468 í232.959 í1580.853
g2 6248.855 326.143 2213.195
g3 í553.021 í2562.553 2235.664
g4 774.229 3587.574 í3129.930
g5 3910.447 í2329.594 3816.517
g6 í5474.626 3261.431 í5343.124
The variance of each constraint has been obtained in Example 7.9. Using
Equation 7.29, the variance of each constraint is
2
n§ wg1 · 2
V g21 # ¦ ¨¨ ¸¸ V bi = 0.00267, V g2 = 0.00522 (7.42a)
1© wbi ¹
2
i
V g23 = 0.00092, V g24 = 0.00181, V g25 = 0.00278, V g26 = 0.00544 (7.42b)
When the noise has distribution, the variance of each constraint is
g1, new = í0.0291+5Ý0.00267 = í0.0158, g 2, new = í2.3332 (7.43a)
g 3, new = í0.417, g 4, new = í1.8008 (7.43b)
g 5, new = í1.3786, g 6, new = í0.4233 (7.43c)
When the noise has the range, Equation 7.30a is utilized. The constraints are
g1, new = 0.4564, g 2, new = í1.6797, g 3, new = í0.0877 (7.44a)
g 4, new = í1.3423, g 5, new = í0.7304, g 6, new = 0.4766 (7.44b)
7.4.4 Optimization Methods
The optimization scenario is the key point of robust optimization. Robust

optimization employs nonlinear programming, but second-order derivatives are
needed because of the variance in the objective function. Although second-order

derivatives are quite expensive, we may still use nonlinear programming with
second-order derivatives. However, some methods have been developed to
alleviate the need for second-order derivatives. The equations previously
introduced are directly utilized to handle constraints in robust optimization.
Optimization processes have been explained in Chapters 3 and 6.
It is a somewhat complex task to handle the objective function. As shown in
Equation 7.18b, there are multiple objective functions in robust optimization.
Since general optimization considers one objective function, various approaches
are introduced to define a single objective function from multiobjective functions.
The most common approach in multiobjective optimization is the weighted sum
method. Equation 7.18b can be rewritten as follows:
P f (b, p) V f (b, p)
f new D (1 D ) (7.45)
P *f V *f
where P *f and V *f are the base values for the mean P and standard deviation V
of the objective function, respectively. They are used for normalization and
usually serve as the starting values of the optimization process. The weighted sum
method is known to be inefficient, but it is popular due to its simple application.
One type of multiobjective optimization method is goal programming.
Designers determine the design goal, transform the goal into goal constraints, and
then optimize them with real constraints by means of nonlinear programming. A
goal constraint is another expression of “an objective function must achieve the
design goal.” The compromise decision support problem is a modified version of
goal programming for robust optimization (Mistree et al. 1993). A compromise
decision support problem is a hybrid multiobjective optimization technique that
incorporates both traditional mathematical programming and goal programming.
The system constraints must be retained without compromise while objectives are
properly compromised.
The objective function in a compromised decision support problem is a
deviation function x(d i , d i ) in terms of deviation variables. Therefore, a robust
optimum can be found by minimizing the deviation variables while the constraints
are satisfied. The compromise decision support problem with t multiobjective
functions is formulated as follows:
Find bR n (7.46a)
to minimize x {w1h1 (d1 , d1 ), w2 h2 (d 2 , d 2 ),..., wt ht (d t , d t )} (7.46b)
subject to g i (b, p) 0 , i 1, ,l (7.46c)
f j (b, p) d j d j g jf (7.46d)
Robust Design 427
d j d j 0 (7.46e)
d j t 0, d j t 0 , j 1,..., t (7.46f)
where w j is the weight and g jf is the goal that the jth objective function achieves.
d j and d j are the deviation variables that represent the distance between the
target level and the actual attainment of the goal. d j and d j indicate the degrees
of underachievement and overachievement of the goals, respectively. As a result,
we can achieve the goal as much as possible by minimizing these deviations. In
contrast to traditional goal programming, the compromise decision support
program uses real design constraints as shown in Equation 7.46.
Various deviation functions are defined in a compromise decision support
program. The most general and simplest form is the weighted sum of deviations
according to the level of requirement to achieve each goal. The distance function
considering the weighted deviation vector has been used as a deviation function.
When a compromise decision support program is used in robust optimization,
both the robustness and feasibility of constraints are satisfied and various design
goals are achieved by compromising goals. Moreover, deviation variables provide
the degrees of underachievement and overachievement of the goals. This method
can obtain a Pareto solution for a nonconvex problem. There are some other
methods for dealing with multiobjective functions (Osyczka 1984, Messac 1996).
Various optimization methods have been applied to robust optimization, some
of which are successful. However, a unified method has not yet emerged and
large-scale problems have rarely been solved. More sophisticated case studies are
needed for robust optimization.
Example 7.12 [Simple Robust Optimization Problem]

Conduct robust optimization of Example 6.14 by using the results of Example 7.8
and Example 7.10. Use Equation 7.45 for the objective function and Equation
7.30a for the constraints. D 0.5 in Equation 7.45.
Solution
Robust optimization is carried out by using the mean and the variance of the
objective function from Example 7.8 and the constraints from Example 7.10. An
optimization library DOT 5.4 is utilized for the optimization process (Vanderplaats
Research & Development, Inc. 2004). A simple program is coded and interfaced
with DOT. The modified method of feasible directions (MMFD) algorithm in
DOT is selected. The initial design is b = [1.0 1.0 1.0 1.0]T and the optimum
solution is b = [ 0.408 0.548 1.69 0.455]T . The mean and the variance of
the objective function is 33 and 0.993, respectively.
Example 7.13 [Robust Optimization of the Three Bar Truss]

Conduct robust optimization by using the results of Example 7.9 and Example 7.11.
Use the objective function in Equation 7.45 and constraints in Equation 7.30a.
Use D 0.5 in Equation 7.45.
Solution
Robust optimization is carried out by using the mean and the variance of the
objective function from Example 7.9 and constraints from Example 7.11. The
programming process is the same as that of Example 7.12. The initial design is b
= [0.0002 0.0002 0.00002]T and the initial mean and variance are 1.42 kg and
0.00308, respectively. The optimum solution is b = [0.000276 0.000178
0.000006]T . The mean and the variance of the objective function is 1.64 kg and
0.00308, respectively. The objective function is increased because the constraints
are violated due to noise.
7.5 Robust Design by Axiomatic Design

Axiomatic design, which was introduced in Chapter 2, is used for robust design.
Axiomatic design has two axioms: the Independence Axiom and the Information
Axiom. Robust design is mainly related to the Information Axiom.
Robust design with axiomatic design is conducted after the Independence
Axiom is satisfied. Information of each FR is minimized. Therefore, it is one FR–
one DP problem as follows:
FR A DP (7.47)
Suppose that there are two candidate DPs in Equation 7.47, and the
relationships between the FR and the DP are depicted in Figure 7.8. The gradient
of the design is called the stiffness of the design. The specified tolerance of the FR
can be more easily achieved by lowering the stiffness of the system (design A).
Allowing 'DP to be larger leads to a robust design. Therefore, the process of
robust design is stated as follows:
Step 1. Select the design that has low stiffness in order to minimize the
variation of the system range.
Step 2. Minimize the bias between the target and the mean from the selected
design.
As mentioned earlier, common aspects exist between the Taguchi method and
axiomatic design in that a robust design is induced. However, different
characteristics are discovered as well. The Taguchi method emphasizes how each
parameter is determined after the conceptual design. That is, the value of each
parameter is determined from the S/N ratio of each row in an orthogonal array.
Axiomatic design emphasizes which parameters are chosen in the conceptual stage,
Robust Design 429
FR Design B
Design A
¨DP ¨DP DP
for design B for design A
Figure 7.8. The distribution of FR in two designs
because the choice of the DP is important as shown in Figure 7.8. In other words,
selection of a part out of multiple candidates is determined by axiomatic design.
Robust design in axiomatic design is obtained by minimizing the information
content. The methods to calculate the information content are explained in
Chapter 2. Appendix 2.B shows an example of how to obtain a robust design with
less information content out of the two designs satisfying the Independence Axiom.
7.6 The Robustness Index

There are robustness indices such as the probability of success (or information
content) in axiomatic design and the loss function (or S/N ratio) in the Taguchi
method. The probability of success considers the ranges of the design variables
(design parameters); however, it tends to ignore their distribution. On the other
hand, the loss function of the Taguchi method can include the distributions while it
does not have design ranges. A new robustness index is introduced in this section
(Park 2000).
Design is an activity to select design variables to satisfy the functional
requirements. The distributions of the functional requirements depend on the
selection of the design variables. In Figure 7.9, the distributions A and B are the
distributions of a functional requirement. Suppose A has a normal distribution and
B has a uniform distribution. Each distribution has most of the responses within
the design range. B has a larger probability of success than A. However, A may
be better from the viewpoint of robust design, since the responses are closer to the
target.
Engineering design usually has design ranges. Since the loss function is
evaluated by the bias of the mean from the target, it does not consider the design
ranges. As the bias becomes larger, the loss function becomes larger. Suppose the
Design range
Probability
density
function A
B
FR FR
Figure 7.9. A normal distribution and a uniform distribution in the system range
mean resides at the outside of the design range as illustrated in Figure 7.10. The
mean of A is the same as that of B. Since the standard deviation of A is smaller,
the loss function of A is smaller. However, B has more probability to reside within
the design range. Thus, the loss function ignores the design range. The
probability of success and the loss function may not reach the same conclusion
when two designs are compared.
A robustness index is proposed to overcome the above difficulties: the
weighted robustness index. A weighting function is employed to emphasize the
closeness to the target. The weighted robustness index is composed using the
design range, the weighting function and the probability density function of the
system as follows:
Probability
density Design range
function A
FR FR0 FR
Figure 7.10. The distribution of the system range when the mean is outside of the design
range
Robust Design 431
FR * '
Pr ³ W ( FR)I ( FR)dFR (7.48)
FR * '
where Pr is the weighted robustness index, ' is half of the entire design range,
FR * is the target, I (FR) is the probability density function of the system, and
W(FR) is the weighting function.
The weighting function W(FR) is defined between the upper and lower bounds
and has the maximum at the target. It is assumed to be a linear or a quadratic
function. The linear weighting function is
A1 * * *
° ' ( FR ( FR ')) ( FR ' d FR d FR )
W1 ( FR) ®A (7.49)
° 1 (( FR * ') FR ) ( FR * d FR d FR * ')
¯'
For the ideal case, the probability density function of the system range is
expressed by the Dirac delta function as follows:
I ( FR ) G ( FR FR* ) (7.50)
The Dirac delta function is infinite at the target. If we make the weighted
robustness index have a unit value (1) with the Dirac delta function, A1 in
Equation 7.49 is determined by the following equation:
FR * '
*
Pr ³ W1 ( FR)G ( FR FR )dFR 1 (7.51)
*
FR '
Equation 7.51 is arbitrarily set to 1 to make it similar to the probability density

function. Integrating Equation 7.51, A1 1. Therefore, the linear weighting
function is
1 *
° ' ( FR ( FR ')) ( FR * ' d FR d FR*)
W1 ( FR) ®1 (7.52)
° (( FR * ') FR) ( FR* FR d FR * ')
¯'
A quadratic equation can be used for the weighting function. The quadratic
weighting function W2 ( FR) is calculated in a similar fashion to the case of the
linear weighting function. In Equation 7.53 A2 1 and the quadratic weighting
function is
Design range
Probability A1 or A2
density System pdf
function (b)
(a)
FR í¨ FR FR +¨ FR
Figure 7.11. Weighting functions and the distribution of the system ranges: (a) linear
function (b) quadratic function
A2
W2 ( FR) 2
(( FR FR * ) 2 '2 ) (7.53)
'
Unless the distribution is the Dirac delta function, the integration of Equation 7.51
is not guaranteed to be 1 when A1 1. However, multiple designs can be
compared by Equation 7.51 regardless of the value of A1. If we use the logarithm,
the weighted information index can be defined as follows:
1
I log 2 (7.54)
Pr
The above two weighting functions are illustrated in Figure 7.11. As shown in
Design range
Probability
Design A Design B
density
function
PA m PB FR
Figure 7.12. Distribution of two system ranges when the mean is inside the design range
Robust Design 433
Table 7.20. The mean and the standard deviation of Example 7.14
Uniform
Mean / Standard Normal distribution
distribution
deviation
a b c d
Mean (P) 15 15 30 30
Standard
5.78 4 4 6
deviation (V)
Figure 7.12, suppose the mean is inside the design range and the distributions are
the same (e.g. normal distribution). In this case, the comparison of the two
distributions reaches the same conclusion regardless of the selection of the
probability of success, the loss function and the weighted robustness index.
Example 7.14 [Calculation of the Robustness Indices]

Suppose that the target of a functional requirement (FR) is 15 and the design range
is from 5 to 25. Select a robust system that has the maximum probability of
success and the minimum variance. The system distributions are given in Table
7.20. System a has uniform distribution and systems b, c and d have normal
distributions; they are illustrated in Figure 7.13. Using the probability density,
evaluate the loss functions, the probabilities of success and weighted robustness
indices of the systems for the comparison.
Solution
The probability of success is calculated by integration within the design range.
Equation 7.11 is utilized for the loss function and the constant k is set by 2.0. The
Design range
Probability
density c
b
function
a d
5 15 25 30 FR
Figure 7.13. Distributions of the system ranges for Example 7.14

Table 7.21. The probability of success, the loss function and the weighted robustness index
of Example 7.14
Uniform
Normal distribution
Index distribution
a b c d
Probability
1 0.99 0.11 0.20
of success
Loss function 66.82 32 482 522
Robustness
0.5 0.68 0.02 0.07
index (1st order)
Robustness
0.67 0.84 0.03 0.10
index (2nd order)
weighted robustness index is calculated by using the linear and quadratic equations.
The results are shown in Table 7.21. Table 7.22 shows the probability of success
and the robustness index by using Equation 2.23 and Equation 7.54, respectively.
The probability of success is to be large and the information content and the loss
function are to be small.
In Table 7.22, the information content and the loss function show the
differences in comparisons of systems a, b and c. System a is a perfect system
from the viewpoint of the information content; however, it has twice the loss
compared to b. That is, the information content calculated from the probability of
Table 7.22. The information content, loss function and the weighted information index of
Example 7.14
Uniform
Normal distribution
Index distribution
a b c d
Information
0.00 0.02 3.24 2.31
content
Loss function 66.82 32 482 522
Information
1.00 0.55 5.63 3.93
(1st order)
Information
0.58 0.25 4.89 3.36
(2nd order)
Robust Design 435
success only considers the design range. However, it hardly considers robustness
from the variance. The weighted information index, when the quadratic weighting
function is used, shows that the information content of a (0.58) is 2.3 times of that
of b (0.25).
The weighted information content of c is larger than that of d. This means that
the responses from d reside more in the design range. However, the loss function
of c is less than that of d. In this case, the weighted information index of c (4.89)
is larger than that of d (3.36), since only the distribution within the design range is
considered. The weighted information index is more influenced by the probability
of success than by the deviation. The means of c and d are out of the design range.
Therefore, the loss function is not appropriate in this case and the weighted
information index can prevent this weakness.
7.7 Summary
Methodologies for robust design are classified into three groups. They are
summarized as follows:
(1) The Taguchi method is appropriately defined for the design of
experiments. Many design engineers have made efforts to adopt the
robustness concept in design. However, it is difficult to find a scale
factor and the S/N ratio in design. The method should be developed
further to find a mathematically robust solution.
(2) Robust optimization employs optimization for robust design. However,
expensive processes are required for the multiobjective function and the
second-order differentiation. Therefore, a compromise solution is
obtained or some approximation methods are utilized.
(3) The second axiom of axiomatic design is used for robust design. When
the information content is minimized, robust design is obtained and the
loss function of the Taguchi method is also minimized. However, the
method is mainly applied to conceptual design and needs to be applied to
detailed design.
Each method is excellent and has many case studies. It is also true that these
methods are difficult to use in practical problems. Therefore, we need some type
of a unified method for robust design.
7.8 Exercises
7.1 We have a function f (b1 , b2 ) 3b12 16b1 16b22 7 . At b1 b2 1 , the

variables have uniform distributions in the range of 0.5 d b1 d 1.5 and
0.5 d b2 d 1.5. Calculate the expectation value for the mean and the
variance of f (1, 1).
7.2 We have a function f (b1 , b2 ) 2(b1 z1 ) 3(b2 z 2 ) 2. At (b1 , b2 )

(1, 1), z1 and z 2 have the normal distribution where the mean is 0 and the
standard deviation is 1. Calculate the expectation value for the mean and
the variance of f at (b1 , b2 ) (1, 1) .
7.3 In Example 7.2, l has the mean of 1 m and the standard deviation of 0.05
m. Calculate the expectation value of the means and the variances of the
horizontal displacement G x and the vertical displacement G y .
7.4 In Exercise 7.1, b1 0.9, 1.1 and b2 0.8, 1.2 . Calculate f (b1 , b2 ) for
all combinations and evaluate the loss function and the S/N ratio in
Equation 7.15. The target is the mean of Exercise 7.1.
7.5 In Exercise 7.2, z1 0.1, 0.1 and z 2 0.1, 0.1 . Calculate f (b1 , b2 )
for all combinations and evaluate the loss function and the S/N ratio in
Equation 7.15. The target is the mean of Exercise 7.2.
7.6 In Example 7.2, l = 0.9, 0.95, 1.0, 1.05 m. Calculate the loss function and
the S/N ratio in Equation 7.15 for the horizontal displacement G x . The
target is the mean of Exercise 7.3.
7.7 In the following function:
f (b1 3) 4 (b2 5) 2 b32 0.3(b1 b2 b2 b3 b3 b1 )
the design variables have the following levels:
Level b1 b2 b3
1 1.5 2.0 í0.2
2 3.0 4.0 0.0
3 4.5 6.0 2.0
Use the L9 (3 4 ) orthogonal array as the inner array. The tolerances for
design variables are
Robust Design 437
Level z1 z2 z3
1 í0.1 í0.1 í0.1
2 0 0 0
3 0.1 0.1 0.1
(1) Make the outer array with the L9 (3 4 ) orthogonal array.

(2) Perform robust design in the five steps in Section 7.3.3 ( w1 w2
0.5).
(3) Arrange the solutions in ascending order by using the estimated values,
evaluate the best value and obtain the final solution by comparing it
with the solution of (2).
7.8 The three bar truss in Figure 3.3 has the following design problem:
Maximize [1
where [ 1 is the first natural frequency. Young’s modulus (E) is 207 GPa,
and the density ( U ) is 2770 kg/m3. l = 1 m and T 60 $ in Figure 3.3.
The levels of the design variables are
( u 10 4 m 2 )
Level
b1 b2 b3
1 0.1 0.1 0.1
2 2.5 2.5 2.5
3 5.0 5.0 5.0
Use the L9 (3 4 ) orthogonal array as the inner array. The tolerances for
design variables are
( u 10 4 m 2 )
Level
z1 z2 z3
1 í0.001 í0.002 í0.001
2 0 0 0
3 0.001 0.002 0.001
(1) Make the outer array with the L9 (3 4 ) orthogonal array.

(2) Perform robust design in the five steps in Section 7.3.3 ( w1 w2
0.5) .
7.9 Use Exercise 6.15. The tolerances for the design variables are
( u 10 4 m 2 )
Level
za zb ze zg zi
1 í1 í1 í2 í1 í1
2 0 0 0 0 0
3 1 1 2 1 1
where z a , zb , z e , z g and z i are the tolerances of design variables a, b, e, g

and i, respectively. Use the orthogonal array L18 (2 1 u 3 7 ) as the inner and
outer arrays.
(1) Perform robust design in five steps in Section 7.3.3 ( w1 w2 0.5) .
7.10 In the following optimization problem:
Find b1 , b2
to minimize f (b) (b1 2) 2 b22
subject to (1 b1 ) 3 b2 t 0
0 d bi , i 1, 2
The design variables have the following levels:
Level b1 b2
1 í1.0 í2.0
2 1.0 0.0
3 3.0 2.0
Robust Design 439
The variables have the tolerances z i 0 . 2 , 0 . 0 , 0 . 2 , i = 1, 2.
(1) Make the inner array.

(2) Make the outer array.
(3) Perform robust design in the five steps in Section 7.3.3. The scale
factor s = 50 ( w1 w2 0.5).
Find b1 , b2 , b3
to minimize f (b) b1b2b3
subject to 72 b1 2b2 2b3 t 0
b1 2b2 2b3 t 0
0 d bi d 42, i 1, 2, 3
The design variables have following levels:
Level b1 b2 b3
1 14.0 2.0 2.0
2 24.0 12.0 12.0
3 34.0 22.0 22.0
The design variables have the tolerances z i 0 . 1 , 0 . 0 , 1 . 0 , i = 1, 2,

3.
(3) Perform robust design in the five steps in Section 7.3.3. Compare the
results when the scale factor s = 1 and 100 ( w1 w2 0.5).
Find b1 , b2 , b3 , b4
to minimize f (b) b1 b2 b3 b1b3 b1b4 b2b3 b2b4
subject to 8 b1 2b2 t 0 12 4b1 b2 t 0

12 3b1 4b2 t 0 5 b3 b4 t 0
8 2b3 b4 t 0 8 b3 2b4 t 0
The design variables have the following levels:
Level b1 b2 b3 b4
1 í2 1 í2 í2
2 0 3 0 4
3 2 2 2 6
The design variables have the tolerances z i 0 . 1, 0 . 0 , 0 . 1, i =

1,…,4.
(3) Perform robust design in the five steps in Section 7.3.3. Compare the
results when the scale factor s = 1, 10 and 100 ( w1 w2 0.5).
7.13 In Exercise 7.10, suppose each variable has the value of the second level.
The tolerances are 0 . 2 d z i d 0 . 2 , i = 1, 2 and have uniform
distributions.
(1) Calculate the mean and variance of the objective function in Equation
7.19.
(2) Obtain the variance of each constraint in Equation 7.29 and make
Equations 7.28 and 7.30 when k = 10.
The tolerances are 0 . 1 d z i d 0 . 1 , i = 1, 2, 3 and have uniform
distributions.
7.19.
(2) Obtain the variance of each constraint in Equation 7.29 and make
The tolerances are 0 . 1 d z i d 0 . 1 , i = 1,…,4 and have uniform
distributions.
7.19.
Robust Design 441
(2) Obtain the variance of each constraint in Equation (7.29) and make
7.16 Perform robust optimization by using the results of Exercise 7.13. Use the
objective function in Equation 7.45 and the constraints in Equation 7.30.
D 0.5 in Equation 7.45.
References
Belegundu AD, Zhang S (1992) Robustness of Design Through Minimum
Sensitivity. Journal of Mechanical Design, Transactions of the ASME
114:213í217
Box GEP (1988) Signal-to-Noise Ratios, Performance Criteria, and
Transformation (with Discussion). Technometrics 30(1):1í17
Box GEP, Fung C (1993) Is Your Robust Design Procedure Robust? Quality
Engineering 6(3):503í514
Chen LH (1997) Designing Robust Products with Multiple Quality Characteristics.
Computers Operations Research 24:937í944
Han JS, Kwak BM (2001) Robust Optimal Design of a Vibratory Microgyroscope
Considering Fabrication Errors. Journal of Micromechanics and Micro-
engineering 11:662í671
Lee KH, Eom IS, Park GJ, Lee WI (1996) Robust Design for Unconstrained
Optimization Problems Using the Taguchi Method. AIAA Journal
34(5):1059í1063
Lee KH, Park GJ (2001) Robust Optimization Considering Tolerances of Design
Variables. Computers & Structures 79(1):77í86
Lee KH, Park GJ (2002) Robust Optimization in Discrete Design Space for
Constrained Problems. AIAA Journal 40(4):774í780
Leon RV, Shoemaker AC, Kacker RN (1987) Performance Measures Independent
of Adjustment. Technometrics 29(3):253í285
Mistree F, Hughes OF, Bras BA (1993) The Compromise Decision Support
Problem and the Adaptive Linear Programming Algorithm. In: Kamat MP (ed)
Structural Optimization: Status and Promise. Washington, D.C., pp 247í286
Messac A (1996) Physical Programming Effective Optimization for Computational
Design. AIAA Journal 34:149í158
Osyczka A (1984) Multicriterion Optimization in Engineering in FORTRAN

Programs. Ellis, Horwood, West Sussex
Park GJ (2000) Investigation of Taguchi’s Loss Function and the Probability of
Success for Robust Design and Proposition of a New Index. Internal Report,
Hanyang University, Seoul, Korea
Park GJ, Lee TH, Lee KH, Hwang KH (2006) Robust Design: An Overview.
AIAA Journal 44(1): 181í191
Park SH (2000) Quality Engineering. Min-Young Sa, Seoul, Korea (in Korean)
Parkinson A, Sorensen C, Pourhassan N (1993) A General Approach for Robust
Optimal Design. Journal of Mechanical Design 115(1):74í80
Parkinson A (1995) Robust Mechanical Design Using Engineering Models.
Journal of Mechanical Design 117:48í54
Parkinson DB (1997) Robust Design by Variability Optimization. Quality and
Reliability Engineering International 13(2):97í102
Sepecial Supplement on Robust and Reliability-based Design (2006), Journal of
Mechanical Design, Transactions of ASME, Vol. 128, No. 4.
Shiau GH (1990) A Study on the Sintering Properties of Iron Ores Using the
Taguchi’s Parameter Design. Journal of the Chinese Statistical Association
28:253í275
Press, New York
Sundaresan S, Ishii K, Houser DR (1995) A Robust Optimization Procedure with
Variations on Design Variables and Constraints. Engineering Optimization
24(2):101í117
Taguchi G, Chowdhury S, Taguchi S (2000) Robust Engineering. McGraw Hill,
New York
Taguchi G (1986) Introduction to Quality Engineering: Designing Quality into
Products and Processes. Asian Productivity Organization, Tokyo, Japan
Taguchi G (1987) Systems of Experimental Design (Vol. I, II). Kraus International,
New York
Tong LI, Su CT (1997) Optimizing Multiresponse Problems in the Taguchi
Method by Fuzzy Multiple Attribute Decision Making. Quality and Reliability
Engineering International 13(1):25í34
Tsui KL (1999) Robust Design Optimization for Multiple Characteristic Problems.
International Journal of Production Research 37(2):433í445
Vanderplaats Research & Development, Inc. (2004) DOT Design Optimization
Tools User’s Manual Version 5.4. CO
Vining GG (1998) A Compromise Approach to Multiresponse Optimization.
Journal of Quality Technology 30:309í313
Welch WJ, Yu WK, Kang SM, Sacks J (1990) Computer Experiments for Quality
Control by Parameter Design. Journal of Quality Technology 4(4):361í378
Wu FC, Chyu CC (2002) A Comparative Study on Taguchi’s SN Ratio,
Minimizing MSD and Variance for Nominal the Best Characteristic Experiment.
International Journal of Advanced Manufacturing Technology 20(9):655í659
8 Case Studies with Analytic Design Methods
8.1 Introduction
Various design methodologies have been discussed and each method shown to
have specific procedures. A single method can be employed for an engineering
product. However, multiple methods should be used for some products. In this
case, it is quite important to appropriately define the design scenario according to
the characteristics of the problem.
Application of design methodologies varies according to whether the design
phase is the conceptual design or the detailed design. The design phases can be
divided into conceptual design, preliminary design and detailed design. However,
division into two phases is sufficient from the viewpoint of application. The
design process can be defined differently based on the product and given
conditions. Therefore, a general definition of the design scenario is almost
impossible. In spite of this, examples are solved by using design case studies in
this chapter. The scenarios can be changed if the product is changed. Only the
design methods introduced in the previous chapters are utilized. Therefore,
different scenarios can be defined with different methods.
Axiomatic design can be exploited for conceptual design. When a problem is
identified, the Independence Axiom is applied by defining the functional
requirements and design parameters. If multiple designs, which satisfy the
Independence Axiom, are found, the Information Axiom is applied to find the best
one with the least information content. As explained in Chapter 7, the design with
the least information content is robust. A brief scenario is illustrated in Figure
2.20 for the application of axiomatic design. This scenario can be modified
according to the characteristics of the problem.
When the conceptual design process is finished, a detailed design is carried out
to determine the specific details of design. The overall process is illustrated in
Figure 1.6. In this procedure, various methods can be utilized according to the
environment. The environment can be defined according to whether the design
apace is continuous or discrete, whether the designer pursues optimization or
robust design and the analysis methods utilized. Figure 8.1 shows a scenario for
Characteristics
of the design
General design Robustness

Design optimization Robust design
Design Design
space variable
Continuous Discrete
Optimization DOE
Multiple DPs
Analysis
Axiomatic
design
Statics Dynamics Continuous Discrete
Static Dynamic
response response Robust Taguchi
optimization optimization optimization method
Figure 8.1. A scenario for structural design
structural design using the analytic methods of this book. Examples are introduced
in this chapter and use the framework in Figure 8.1 directly or with modification.
8.2 Design of a Strip Casting Roll
8.2.1 Problem Description
The twin-roll strip casting process is a casting technology under which hot-rolled
coils are directly manufactured bypassing the conventional continuous casting
process and the hot-rolling process. The conventional continuous casting process
casts slabs of 200–250 mm in thickness, and produces coils through the rolling
process. The thin slab casting process has a route similar to the continuous casting
process. However, it is a process technology to cast thin slabs so that they may be
50–75 mm in thickness to decrease the load of the hot-rolling process.
Case Studies with Analytic Design Methods 445
SEN Melt pool

Melt Pool
Shell
Solidification end
Solidification End
Hot deformation
Hot Deformation
Roll nip
Roll Nip Zone
zone
Roll
Cast Strip
Cast strip
Figure 8.2. Schematic representation of the twin-roll strip casting process
This problem is the optimization of a POSCO twin-roll strip casting process. It

directly produces hot coils of 1–6 mm in thickness by feeding molten steel in
between the two rotating rolls. The molten steel is solidified in an extremely short
time (0.2–0.6 seconds) and then exits out from the rolls in the form of a strip.
When the molten steel undergoes such a rapid cooling process, uniform
solidification and the shape of the strip should be acquired over the entire width of
the rolls. Since the strip casting process can manufacture the strip when the
extremely hot molten steel makes contact with the casting rolls, these casting rolls
that determine the shape of the strip are extremely important. This in turn makes
the design of the casting rolls very significant, and the design of the casting rolls is
considered as the core of the strip casting process. Since the rolling process is
eliminated, the strip casting roll significantly reduces the overall cost.
Figure 8.2 presents the process of pouring molten steel in between the two
rotating rolls and rolling the solidified shells near the roll nip after being solidified
onto them. The technologies related to the rolls, the nozzle feeding molten steel,
and the edge dam that prevents the molten steel from being leaked are crucial
elements. Technologies applicable to the accurate control of such mechanical
systems are also important. Out of these elements, the rolls play the most
important part in production of the strips.
The roll is repeatedly heated up and cooled down from cyclic contact with the
molten steel. The cooling process is mainly due to the cooling holes. The thermal
stress varies greatly on the roll surface, which may cause cracks on the surface. If
the crack propagates from the surface onto a cooling channel, an emergent
situation of water leakage occurs during casting. Therefore, the casting rolls
should be designed in an appropriate way in order to prevent the roll from cracking.
8.2.2 Optimization Formulation for a Strip Casting Roll
In the hot rolling process, flatness, bending and shearing of the strip cast are
generated from large external forces imposed on the rolling mill. On the other
hand, the external forces are not large in the twin-roll process because
solidification is made around the roll nip. Therefore, the physical external forces
caused by the rolling step do not have much contribution in the casting roll. Thus,
the external forces are neglected and the research is focused on the thermal stress
on the roll surface.
Figure 8.3 illustrates the axial cross section of the finite element model used to
calculate stresses. In order to reduce calculation time, the entire cross section is
not considered as illustrated in Figure 8.4. Only several parts of the axial cross
section of the roll are selected for analysis. The analysis condition of a selected
model is the plane strain condition.
It is assumed that the roll is deformed radially and is not deformed in the
direction of the circumference. The rotation condition of the roll is changed to the
rotation of the heat flux. When the molten steel comes into contact with the roll, a
thermal loading is imposed on the roll; otherwise, it is cooled down. Then this
condition is equivalent to the rotation of the roll.
When the roll temperature is calculated, initial boundary conditions are twofold.
The adiabatic condition is utilized at the faces except for the roll surface, and heat
x
z
Figure 8.3. The mesh of the cooling roll
Figure 8.4. The analysis model of an axisymmetric section plane

is absorbed into the roll through the contact area with molten steel while it is
radiated into the atmosphere from the noncontact area. Examination is conducted
based on the case where the copper roll, the POSCO pilot caster of 1300 mm in
width and 1250 mm in diameter, is used to cast Type 304 austenitic stainless steel
at a speed of 45 m/min with 6.2 MW/m2 of heat flux based on a total of 20
revolutions. The time interval applied during calculation is to be one step when
the roll turns by 2q.
In order to view the trends of deformation and thermal stress, the results of
analysis are obtained by using the elastic finite element method. Since a high
temperature gradient is expected near the roll nip of Figure 8.2, the effective (von
Mises) stress may exceed the yield stress. Thus, it may be desirable to use elasto-
plastic finite element analysis. The results of analysis by the elasto-plastic finite
element method are compared with those of the elastic finite element method. In
addition, since the rotating roll has different boundary conditions depending on the
positions, transient analysis is performed.
As the molten steel makes direct contact with the casting rolls, cooling of the
rolls is extremely important. Therefore, the important variables of cooling include
the position, arrangement and the diameter of the cooling holes. In conjunction
with this, the diameter and the position of the channel of the rolls are defined as
the design variables to optimize the rolls. In addition, the internal stress of the
casting rolls exposed to a high thermal load is expected to be very high, and
accordingly, the degree of deformation will also be high. Since such stress and
deformation are caused by heat, cooling through the cooling channel reduces the
temperature of the rolls, which in turn reduces the stress and deformation to the
same level. The position and the diameter of the cooling holes are, therefore,
optimized using the stress and displacement obtained from the temperature, and
thermal deformation analysis as the constraints. As the volume of the rolls varies
according to the diameter of the cooling hole, design optimization is carried out by
designating the volume of the rolls as the objective function to make the rolls as
light as possible and to facilitate machining.
Copper alloys and nickels are used as materials. Design optimization is carried
out by applying the variation of the heat flux and the casting speed as load
conditions. The formulation of the optimization procedure is made as follows:
Find D, PCD (8.1a)

to minimize volume (8.1b)
subject to V d V max 3.7 u10 8 (8.1c)
H d H max 3 u 10 2 (8.1d)
G d G max 3 u 10 4 (8.1e)
where D is the diameter of the cooling hole, PCD is the position of the cooling
hole, V is the stress, H is the strain and G is the displacement. They are
calculated from the finite element method. Optimization is performed with the
algorithm in the ANSYS program, which is a modified version of the response

surface method in Section 6.7 (ANSYS 2004).
8.2.3 Optimization Results
In the strip casting process, hot molten steel should be solidified and shaped
uniformly over the width of the roll in the rapid cooling process. The roll for strip
casting generally has a large contact area with molten steel in order to ensure high
productivity. Thus, it has a much greater diameter than that of an ordinary rolling
mill. Two aspects should be considered in the roll design of the strip casting
process. First, cooling channels within the roll should be designed with good
efficiency because the roll plays the role of the mold in this process. Second, the
roll acting as a rolling mill should be considered from the viewpoint of
deformation and thermal expansion during the casting process. However, the
external forces are not large in the twin-roll process because solidification occurs
around the roll nip. Therefore, the physical external forces caused by the rolling
step do not have much contribution to the stress distribution in the casting roll.
Thus, the external forces can be considered negligible and the research is focused
on roll cooling because the internal stress has an extremely close relationship with
the roll cooling process.
Elastic stress and deformation are examined prior to applying design
optimization, and the stress and deformation are also examined after the design
optimization is applied. Moreover, changes in the design variables are also
observed when applying design optimization. Figure 8.5 shows the temperature
distribution on the roll nip after the roll has revolved 20 turns. The copper alloy
hardened by zirconium oxide (ZrO2) precipitate is used for strip casting rolls.
Since this alloy may lose its own properties when exposed to a temperature higher
than 400 o C, the temperature should not exceed 400 o C. The results of calculation
Figure 8.5. Temperature distribution at the roll nip

350
315
280
Temperature 245
210
175
140
105
70
35
0
0 8 16 24 32 40 48 56 64 72 80 88 96 104 110
Time
Figure 8.6. Temperature variation of the roll surface with respect to time
show that the temperatures on the roll nip surface are less than 350q C . In
particular, the temperature of the copper alloy, which is below the nickel plating
layer of the casting rolls, is much lower than the temperature of the surface. Figure
8.6 shows the temperature variation on the surface of the casting roll when the roll
makes 20 turns.
As illustrated in Figure 8.6, the roll nip temperature is approximately 330q C
and the temperature of the location prior to the meniscus is approximately 70q C.
From the results of the temperature variation, we can see that the temperature
sharply changes while the casting rolls are revolving. Figure 8.7 indicates the
temperature distribution near the cooling channel on the roll nip.
Figure 8.7. Temperature distribution around the cooling channel at the roll nip
Table 8.1. Radial displacement due to thermal expansion at the roll nip (20 rev.)
After revolving
Nip (after 20 rev.) Before meniscus
180q
Temperature (o C) 330 104 70
Deformation (mm) 0.282 0.265 0.242

2
Stress (N/m ) 1.33E9 3.02E8 1.95E8
Table 8.1 shows the tabulation of the temperature, deformation and stress of
the casting rolls at the roll nip. The maximum thermal expansion is approximately
0.282 mm in the radial direction, and such thermal expansion directly affects the
strip thickness. During casting, strips, excluding those of initial casting, should
almost have a flat state in the direction of the width. To achieve this, the thermal
expansion of the casting rolls should be estimated in advance, and the thermal
crown is given to the casting rolls. The amount of the crown is equivalent to the
thermal expansion of the casting rolls. The thermal crown means the shape of a
crown, which compensates the difference of expansions in the direction of the
width. The rolls continue to expand. When they reach a steady state, thermal
expansion no longer increases and periodic contraction and expansion are repeated.
Figure 8.8 shows that the position of the cooling hole is moved by the thermal
deformation of the rolls. Such an expansion has maximum deformation at the nip.
It is contracted in a region that is not exposed to the thermal load. Expansion and
contraction occur repeatedly; this phenomenon is depicted in Figure 8.9.
Figures 8.10 and 11 show the effective stress distribution at the roll nip. They
are the results from elastic analysis and elasto-plastic analysis, respectively. The
Figure 8.8. Change of the cooling channel position due to thermal expansion
10-4
2.823
2.701
2.578
2.456
Displacement
2.333
2.211
2.088
1.966
1.843
1.721
1.598
0 8 16 24 32 40 48 56 64 72 80 88 96 104 110
Time
Figure 8.9. Radial displacement with respect to time
Figure 8.10. von Mises stress distribution near the roll surface at the roll nip (elastic
deformation)
stress, which is used to determine whether plastic deformation has occurred, is the
effective stress. If the effective stress is greater than the yield stress, this means
that plastic deformation has occurred. In Figure 8.10, the maximum effective
stress is 1.33E09 N/m2. This is an extremely high value and does not occur in
reality. Such a high effective stress occurred because the elastic model was used.
Since the results of elastic analysis are not actual values, elasto-plastic analysis
must be carried out. Figure 8.11 illustrates the results of calculation using the
elasto-plastic model, and the value does not exceed that of the yield stress.
However, even though deformation will continue to change, there will be no more
changes in the level of deformation when it reaches a steady state. The

deformation will repeat contraction and expansion. The effective stress may occur
in a plastic region due to the rapid temperature gradient on the casting roll surface,
and the casting roll surface may be undergoing plastic deformation. Since this
plastic deformation is caused by thermal deformation and is controlled by the
compressive hoop stress, the level of plastic deformation is expected to be
extremely low.
Figure 8.12 presents the effective stress distribution around the cooling channel.
Since the effective stress around the cooling channel is 1í2.5E8 N/m2 and the
yield stress of the copper alloy is 3.7E8 N/m2, the stress distribution around the
Figure 8.11. von Mises stress distribution near the roll surface at the roll nip (elasto-
plastic deformation)
Figure 8.12. von Mises stress distribution around the cooling channel
cooling channel is below the yield stress. As shown in Table 8.1, the stress
distribution of the casting rolls is the highest at the roll nip and the lowest at the
place in front of the meniscus.
The stress is extremely high at the roll nip due to the thermal load and plastic
deformation that may occur. The stress and deformation can be reduced according
to the cooling process. The diameter and the position of the cooling hole are
important variables for the cooling capacity of the casting rolls.
Since the position and the diameter of the channel change the stress and
deformation, they should be appropriately selected. If the position of the channel
is closer to the surface, the surface temperature decreases and this leads to a
decrease in the stress in the rolls. On the other hand, stress increases as the
channel goes away from the roll surface. If the diameter of the channel is large,
the interval between channels is also large. An uneven temperature and stress
distribution occur. However, a smaller diameter used for the channel leads to
uniform temperature and stress distributions. If the diameter of the channel is too
small, it is difficult to make such a channel and therefore, an appropriate diameter
of the channel should be selected.
In Equation 8.1, constraints are imposed on strains and displacements in the
radial direction. The constraints in Equation 8.1c are not utilized because elasto-
plastic analysis is employed. In elasto-plastic analysis, when the stress reaches the
yield stress the stress is fixed and the strain becomes larger. Therefore, the
constraints in Equation 8.1d are utilized to limit the strains. The stress constraints
are indirectly considered. When the roll is expanded excessively, the size of the
crown should be larger to flatten the strip. Therefore, Equation 8.1e is imposed to
prevent excessive expansion of the roll. The objective function is defined by the
volume of the roll.
Table 8.2 shows the thermal load conditions. Design optimization is carried
out based on these conditions, and the results show that the diameter of the cooling
hole is approximately 0.64% of the roll diameter. The position from the origin of
the roll center is 96% of the radius. These results are presented for Case 1 in Table
8.2. Optimization is carried out based on the conditions in Table 8.2 and the
results are shown in Table 8.3.
Table 8.2. The conditions of the heat load for optimization
2
Speed (mpm) Heat flux (MW/m )
Case 1 45 6.2
Case 2 60 8.4
Case 3 80 9.3
Case 4 100 10.2

Table 8.3. Optimization results for the cases in Table 8.2
Hole
Deformation Channel
Stress (Pa) Strain diameter
(m) position (%)
(%)
Case 1 3.71E8 1.43Eí2 1.03Eí4 0.643 96.003
Case 2 3.75E8 2.47Eí2 1.32Eí4 0.641 96.001
Case 3 3.76E8 2.74Eí2 1.33Eí4 0.641 95.984
Case 4 3.78E8 2.99Eí2 1.37Eí4 0.640 96.000
The velocity of the casting roll and heat flux are loading conditions, and the
stresses and displacements are constrained. The maximum effective stress is the
yield stress. The displacements increase as the loading increases. However, they
are within the ranges of the constraints. It is noted that the optimum variables are
similar in all cases.
In all cases, the optimum values are similar. The diameter and the position of
the cooling hole change at a minor level despite the changes in the heat flux and
the casting speed. As illustrated in Figure 8.13, the objective function also shows
almost the same results.
8.2.4 Summary
Stress analysis is carried out considering heat transfer and plastic deformation.
The results of the design optimization are as follows:
6.2300E-03
(m3)3)
6.2250E-03
Volume (m
Volume
6.2200E-03
6.2150E-03
6.2100E-03
0 1 2 3 4 5
Case
Case
Figure 8.13. The value of the objective function for each case
(1) From the viewpoint of effective stress, the roll surface may undergo plastic
deformation due to the large temperature gradient imposed near the roll nip.
However, since the compressive stresses in the direction of the
circumference are predominant, the amount of plastic deformation is
negligible.
(2) The stress around the cooling channel is lower than the yield stress of the
copper alloy. The region having contact with molten steel from the
meniscus to the roll nip shows plastic deformation and the other region of
the casting roll is within the elastic region.
(3) The diameter and the position of the cooling holes are selected as the
design variables for optimization of the casting rolls.
(4) Each design variable shows only minimal changes with the increase of
thermal loads.
8.3 Design of a Spacer Grid Spring
In pressurized light water reactor (PWR) fuel assemblies, spacer grids are
important components to laterally and vertically support nuclear fuel rods. It is
required that nuclear fuel maintain mechanical soundness for safety and efficiency
during operation. The loss of mechanical soundness of nuclear fuel is generated
by excessive deformation and damage of a spacer grid, fuel rod, etc.
The structural grids support the fuel rods both laterally and axially with a
friction grip. Each cell in the spacer grid employs a fuel rod support system
consisting of two orthogonal sets of two dimples and a spring. The support system
Fuel rod
Grid dimple
Grid spring
Figure 8.14. Schematic view of a spacer grid set

not only allows the fuel rod to move through the cell axially, but also follows fuel
rod diametral changes in operation while maintaining alignment.
A schematic view of a grid set is illustrated in Figure 8.14. A spacer grid for a
nuclear fuel assembly has grid springs and opposing dimples, which contact a fuel
rod and soften any vibration impact. The impact is generated during reactor
operation and shipping. Moreover, external forces are imposed on the springs
during insertion of the fuel rod into the spacer grid. Because the strength of the
grid dimples is relatively large compared to that of the grid spring, the deformation
of the dimples is negligible.
The functions of the spacer grid assembly are as follows: First, the grid
assemblies provide both lateral and axial support for the fuel rods. The support
system allows for thermal expansion and growth of the fuel rods. It also retains
the rods with sufficient spring contact force to prevent coolant flow-reduced
vibration damage. However, it should not overstress locally at the fuel rod. In a
recent design, the spring system mitigates bending of the fuel rods. Second, the
grid maintains its spacing under both accidental and operational loading conditions
to maintain the cooling capacity of the fuel rod lattice. The pitch of the fuel rods
in the core is a parameter that has a major effect on the nuclear and
thermal/hydraulic performance of the core. Third, the grids must support the guide
thimbles to keep them sufficiently straight so as not to impede control rod insertion
under any normal or accidental conditions. Fourth, the grid makes a water channel
between the nuclear fuel rods to improve coolant mixing through the core.
8.3.2 Design Requirements
During operation in a nuclear reactor, the grid springs and dimples are exposed to
intensive radiation, which causes the grid springs to lose initial stiffness. Thus, the
grid spring must permit the fuel rods to vibrate and chatter against the dimples.
Some conventional spacer grids have an additional problem in that the fuel rod
springs and dimples may not accurately locate the fuel rod at the center of a
standard cell.
The conventional design of the grid spring has been conducted for the axial and
lateral forces. Thus, the grid spring was designed to be stiff and large
displacement was not considered. However, the grid spring should satisfy not only
force but also displacement conditions. Hence, the grid design requirements are
defined in two categories: displacement and force conditions.
The Displacement Requirement

During fabrication in Figure 8.15, the rod is pulled into each spacer grid cell
having a height difference of maximum 0.4 mm. The grid spring should have the
same displacement. The height difference is caused by manufacturing tolerances.
After insertion of the rod, the performance of the grid spring can deteriorate if the
grid spring is out of position. The grid spring is designed to have an elastic range
of up to 0.4 mm displacement.
This requirement is defined in a linear range. However, slight nonlinear
deformation is allowed. Later, it will be shown that the requirement cannot be
Spring
0.4 mm
Spacer grids
Figure 8.15. The manufacturing process and schematic of the manufacturing tolerance
satisfied by the given configuration. Thus, the spring is designed to have

maximum flexibility within the given range. Further investigation is required for
the definition.
Force Requirement
Throughout the lifetime of the reactor, the grid spring loses 92% of its strength due
to the radiation-induced relaxation. The changes from the radiation induced
property are strength and ductility change, stress relaxation, and growth. If they
are not considered in the design stages, the integrity of the grid can be jeopardized.
During reactor operation, the unit grid spring endures flow-induced vibration of 2
N. The spring should normally perform up to 25 N. The design of all grid
assembly parts should limit flow-induced vibration; therefore, it will not lead to
vibration-induced fatigue, fretting, and out of position problems. During shipping,
the grid spring should withstand vertical and lateral accelerations without allowing
the fuel rods to shift or the grid spring to deform permanently. Therefore, the
requirement should be larger than 25 N to include the unknown disturbances.
Considering the above design requirements, the force–displacement curve for
the spring has to be linear in the range of 0.4 mm for displacement. Moreover, the
maximum spring force is defined by 40 N.
8.3.3 Optimization of the Grid Spring

A spring in a grid set is analyzed and designed. The flow of the design process is
defined and illustrated in Figure 8.16. Because various aspects are involved, it is
extremely difficult to define each step with an automatic design process. Instead, a
practical design using engineering intuition is very efficient in some steps. The
flow in Figure 8.16 includes those characteristics. Each step is explained below.
First, the spring is designed by engineering intuition. The spring should fit in
the structure in Figure 8.14. It is difficult to have a drastic shape change in
optimization but the modification can be made in this step. Design requirements in
Modelling and nonlinear finite analysis of a

spring with an arbitrary conceptual design
Modification of the
Satisfied? design by engineering
No intuition
Yes
Shape optimization in a linear range
Verification of the optimum design with

nonlinear analysis
Satisfied?
No
Yes
Stop
Figure 8.16. Flow of the defined design process
the previous section are defined in a linear range. The spring just reaches the
plastic (nonlinear) range. Therefore, the conceptual design is carried out with
nonlinear analysis. The design is directed to have maximum flexibility and lower
maximum stress by trial and error. A spacer grid cell with imposed displacement
boundary conditions is modelled for analysis of spacer grid spring characteristics.
The geometric model is created by using I-DEAS (IDEAS 2002). Because of
symmetry, only a quarter of the structure modelled by thin shell elements is
analyzed by using ABAQUS (ABAQUS 2003).
For the spring characteristics satisfying the foregoing proposed design
requirements, the stiffness of the spring is selected as a performance measure. The
stiffness is computed by the ratio of the reaction force developed at the final load
step and the displacement. The linear range of the stiffness is also selected as a
performance measure. The stiffness is related to the maximum stress in a model
with nonlinear material property. The linearity of the stiffness is assured when the
Section point 1
Mises Value
3.06E-02
9.96E+01
1.67E+02
2.38E+02
3.12E+02
Figure 8.17. The von Mises stress contour after conceptual design
maximum stress does not exceed the yield stress of the spring at the load step. The
final model for the conceptual design is illustrated in Figure 8.17.
Nonlinear analysis is carried out in nine load steps and the maximum
displacement at the final step is 0.9 mm. Figure 8.17 shows the plot of the von
Mises stresses. Figure 8.18 presents the force–displacement curve for the grid
spring of the initial design.
The area A in Figure 8.17 is further investigated. The fuel rod vibrates against
area A due to the coolant flow. The relative infinitesimal motion between the two
bodies causes fretting wear on the contact surfaces of the fuel rod and the grid
60
50
Force (N)
40
Force(N)
30
20
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Deflection(mm)
Displacement (mm)
Figure 8.18. The force–displacement characteristic curve of the conceptually designed spring
Value
+0.00E+00
+5.36E+00
+1.07E+01
+1.79E+01
+2.32E+01
(a) Initial shape
Value
+0.00E+00
+9.72E+00
+2.19E+01
+3.16E+01
(b) Improved shape
Figure 8.19. Contact pressure configuration of the grid spring with 0.4 mm
spring. Design should be performed to minimize the wear and this can be achieved
through minimization and uniform distribution of the local contact pressure. This
design process cannot be included in the shape optimization (detailed design) since
nonlinear analysis is involved. Therefore, the design for area A in Figure 8.17 is
carried out in the conceptual design process. The contact pressure is evaluated by
ABAQUS. Figure 8.19a shows the contact pressure for the original shape of area
A. Improved shapes are discovered by performing trial and error analysis. The
final result is illustrated in Figure 8.19b.
The overall shape is determined in the conceptual design process. The detailed
design is carried out by shape optimization. A slight modification of the shape is
made. As mentioned earlier, the optimization is performed in a linear range.
Generally, the objective function in structural optimization is defined by the
weight of the structure. However, it has been proved that the structure does not
have a feasible solution for allowable stress with the given configuration in Figure
8.17. Therefore, the problem is formulated with the maximum stress as the
objective function.
The optimization problem is defined as follows:

to minimize maximum V (8.2b)
subject to G 0.4 0 (8.2c)
Fspring 40 0 (8.2d)
where V is the stress, G is the displacement at the center of the spring and Fspring
is an external force on the spring. The displacement condition in Equation 8.2c is
from the first requirement and the force condition in Equation 8.2d is from the
second requirement. The optimization problem in Equation 8.2 is converted as
follows:

subject to V E d0 (8.3c)
G 0.4 0 (8.3d)
Fspring 40 0 (8.3e)
D
B
C
E
G F
x
Y X H
y
I
Z
z
Figure 8.20. A quarter FE model and design variables

Table 8.4. Results of shape optimization
Initial Optimum
Design variable
value value
Scale factor of the perturbation vector for
b1 0.0 0.01
y-coordinates of nodes A, B, C in Figure 8.20
b2 0.0 0.14
y-coordinates of nodes D, E, F in Figure 8.20
b3 0.0 í0.11
y-coordinate of node G in Figure 8.20
b4 0.0 í0.11
y-coordinate of node H in Figure 8.20
b5 0.0 0.54
z-coordinate of node I in Figure 8.20
b6 0.0 1.61
x-coordinate of node G in Figure 8.20
b7 0.0 0.94
x-coordinate of node H in Figure 8.20
b8 0.0 í2.0
z-coordinate of node H in Figure 8.20
E Bound on stress 2.0 0.371
Objective
2.0 0.371
function
A quarter model for the finite element analysis is illustrated in Figure 8.20.
Eight design variables are selected. The design variables are indicated in Figure
8.20 and Table 8.4. Optimization is carried out by GENESIS (GENESIS 2004).
The initial design for the optimization process is the design result of the conceptual
design. The results of the optimization are shown in Figure 8.21 and Table 8.4.
The maximum stress exceeds the yield stress of 0.382 GPa. Because a little plastic
deformation is allowed, the design is considered acceptable. Nonlinear analysis is
performed with the optimum design and the result is illustrated in Figure 8.22. It is
noted that the optimum design is more flexible than the initial design. The
equivalent plastic stress and the strains of the initial and the optimum designs are
compared in Figure 8.23. The new design is compared to the current existing
designs by nonlinear analysis. The performance of the spring is significantly
improved.
8.3.4 Summary
(1) Design requirements for a spacer grid spring are defined for manufacturing
and working conditions.
x
Y y X
Z
z
Figure 8.21. Results of shape optimization
in it ial design
Initial design oOptimal
p t im al design
design
80
70
60
(N)
Force (N)
50
Reaction force
40
Reaction
30
20
10
0
0 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8
Displacement
Deflect io n (m m(mm)
)
Figure 8.22. Results of the spring characteristics
(2) A design process is proposed for the spacer grid design in a practical way.
Conceptual and detailed designs are defined appropriately with linear and
nonlinear finite element analyses and optimization. Basically, the design is
directed to enhance the flexibility of the spring.
80
70
60
force (N)
new design
New design
50
Reaction Force
inconel AA
Inconel
40 inconel BB
Inconel
Reaction
30 zircalloy A
Zircalloy A
zircalloy B
Zircalloy B
20
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Deflection (mm)
Displacement (mm)
Figure 8.23. Comparison of the new and existing springs
(3) In the conceptual design, the overall shape of the grid spring is determined.
Nonlinear analysis is exploited to help the decision making process, which
is made based on engineering intuition.
(4) Shape optimization is employed for the detailed design. The maximum
stress of the spring is minimized, while the design requirements are
satisfied.
(5) Because the shape optimization is conducted in a linear range, the
nonlinear performance of the optimum design is evaluated. It is found that
the new design is much more excellent than the current designs.
8.4 Design of an Automobile Rearview Mirror

An automobile rearview mirror is attached at the side and lets a driver view the
rear. As shown in Figure 8.24, it consists of many parts to which mirror is
attached. When a car is running, various vibrations occur, which are transferred to
the mirror and disturb the driver’s rear view. Since the vibration of the mirror is
related to safety, it is important to minimize it. A case study is introduced for this
problem.
8.4.1 Finite Element Modelling of the Rearview Mirror
First, vibration analysis is performed for a rearview mirror. Experiments for

vibration analysis are the most precise; however, they are quite costly when the
design is frequently changed. Thus, computational analysis is carried out by the
Housing
Shaft Frame Actuator
Glass holder plate Glass holder

Base plate Glass
Figure 8.24. Parts of the rearview mirror
z
y
Figure 8.25. Finite element model of a rearview mirror
finite element method. Various finite elements should be utilized for the
assembled structure as shown in Figure 8.24. The finite element model is
complicated and the model is validated by real experiments.
The finite element model of a rearview mirror is presented in Figure 8.25. It
mainly consists of shell elements. Parts such as the actuator in Figure 8.24 have
only a mass effect, thus, they are modelled by mass elements. The connections of
the parts are modelled by beam elements and the section properties of the beam
elements are determined by comparison with the experimental results. The
boundary conditions are determined by fixing the base plate in Figure 8.24. In
experiments, the displacement on the surface of the mirror is measured.
The natural frequency of each part is measured and compared with the finite
element model for validation. As illustrated in Figure 8.26, the natural frequency
is measured by an impact hammer and the fast Fourier transform (FFT) analyzer.
Impact hammer
Housing
Base plate
Frame subassembly
Charge amplifier Conditioning amplifier
FFT
Evaluation of natural frequency
Figure 8.26. Schematic view of the experiment for measuring the natural frequency of the
parts
The measured values are incorporated into the finite element model to tune the
values of the material properties, the thickness, etc.
The finite element models for the parts are modelled and the assembled model
is tuned to be close to the experiments. The schematic view of the experiment is
illustrated in Figure 8.27. An experiment of harmonic vibration is carried out. L
in Figure 8.27 is the distance from the mirror to the measuring plate and G is the
displacement on the surface of the mirror. Vibrations of 10í200 Hz are imposed.
A laser beam is projected onto the mirror and the displacement angle and the
Small laser
Laser beam
G
Rearview mirror
Jig
Vibration generator
Direction of vibration
Figure 8.27. Experiments for the forced vibration of the mirror

1.5
Displacement (mm)
0.5
0
0 50 100 150 200
Frequency (Hz)
Figure 8.28. Results of the experiments for harmonic response
displacement G are measured at the measuring plate. The resonant displacements

are measured for the range of 10í200 Hz. The acceleration condition for the
vibration is 1G. The results of the experiment are illustrated in Figure 8.28.
Average resonant frequencies are obtained from the experiments. The first
resonant frequency is 40 Hz, the second one is 60 Hz and the third one is 107.5 Hz.
The ones from the finite element analysis are 39.8 Hz, 61.1 Hz and 107.2 Hz.
Thus, the finite element model is regarded as being validated and the design
process is continued.
8.4.2 Topology Optimization of the Rearview Mirror
The first natural frequency is to be increased to avoid resonance. In other words,

the stiffness of the structure is to be increased. This can be achieved by having
stiffeners at the glass holder plate in Figure 8.24. Topology optimization is
conducted to find the direction of the stiffeners. Some parts are separated for
topology optimization. Figure 8.29 shows the parts for topology optimization. A
commercial software system GENESIS is utilized for topology optimization and
the density method is employed (GENESIS 2004). The formulation for topology
optimization is as follows:

1 1
to minimize 0.7 0.3 (8.4b)
[1 [2
Figure 8.29. Results of topology optimization
subject to final mass ≤ 0.5 × initial mass (8.4c)
where ξ1 is the first natural frequency and ξ 2 is the second natural frequency.
Since ξ1 has more influence on stiffness, the weighting factor is larger. In order to
maximize the natural frequency, the reciprocal is minimized.
The results of topology optimization are presented in Figure 8.29. The dark
areas are recommended for stiffeners. The initial design has ξ1 = 67.20 Hz and
ξ 2 = 82.07 Hz. After topology optimization, ξ1 = 73.03 Hz and ξ 2 = 86.71 Hz.
The displacement on the surface of the mirror changes from 0.396 mm to 0.222
mm. Thus, stiffness is increased. The results in Figure 8.29 cannot be used
directly. The final stiffeners are determined by the designer.
8.4.3 Size Optimization of the Rearview Mirror
Size optimization can be adopted for design of the rearview mirror. The design
variables are defined by the thicknesses of parts. Six of them are selected as
Figure 8.30. Finite element model of the rearview mirror for size optimization
4
Thickness
FRAME_1 FRAME_21 FRAM E_22

1
FRAME_23 FRAME_3 FRAM E_4
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Iteration
Figure 8.31. Change of the design variables in size optimization
shown in Figure 8.30, which also presents the finite element model utilized in the
size optimization. The formulation is as follows:
Find the thickness of each part (8.5a)

to maximize the first natural frequency [1 (8.5b)
subject to mass d initial mass (8.5c)
After size optimization, the natural frequency is increased from [1 67.2 Hz to

[1 90.03 Hz. The histories of design variables are presented in Figure 8.31. In
size optimization, constraints for stresses can be included. However, since the
rearview mirror is not damaged during operation, they are not considered.
8.4.4 Design Using Orthogonal Arrays in a Discrete Space
The rearview mirror is designed by using the method given in Section 6.3. The
design variables are shown in Table 8.5. The displacement on the surface of the
mirror is minimized to maximize stiffness. The formulation is as follows:
Find A, B, C , D, E , F , G, H , I , J , K , L, M (8.6a)
to minimize G mirror (8.6b)
As shown in Table 8.5, the number of design variables is 13; their levels are
shown in Table 8.6. If we use the full factorial design we need 313 = 1,594,323
experiments (finite element analysis). The number can be considerably reduced by
using an orthogonal array. The interactions are ignored. A minimum orthogonal
array is selected by using the method in Section 6.3. Since this orthogonal array is
L27 (313 ), only 27 experiments are required. Equation 8.6b is calculated for each
row of the orthogonal array and the results are shown in Table 8.7. The one-way
Table 8.5. Design variables for the rearview mirror
Variable Part Design variable

A Frame Young's modulus and density
B Neck thickness
C Bottom plate thickness
D Rib effect
E Shaft Plate thickness
F Spring stiffness between the frame and shaft
G Radius of the spring
H Young's modulus and density
I Housing Housing material density
J Frame and housing connection
K Glass and holder Glass material density
L Glass plate thickness
M Spring between the frame and glass
Table 8.6. Levels of the design variables in Table 8.5
Level
Level 1 Level 2 Level 3
Factor
A E = 200.0, ȡ = 1.07Ý10í10 E = 1662.40, ȡ = 1.74Ý10í10 E = 7241.73, ȡ = 2.78Ý10í10
B 1 mm 3 mm 5 mm
C 1 mm 5 mm 9 mm
D 5 mm (horizontal) No change 5 mm (vertical)
E 1mm 3 mm 5 mm
F E = 1.05Ý104 E = 2.1Ý104 E = 3.15Ý104
G 11.25 mm 14.5 mm 17.45 mm
H E = 1662.40, ȡ = 1.74Ý10í10 E = 4698.15, ȡ = 4.238Ý10í10 E = 7733.90, ȡ = 6.735Ý10í10
I 0.4Ý10í10 1.07Ý10í10 1.74Ý10í10
J 5 mm (up and down) No change 5 mm (up, down and outside)
K 6.664Ý10í11 8.33Ý10í11 9.996Ý10í11
L 1 mm 3 mm 5 mm
M 1 (right side) No change 2 (both sides)
Table 8.7. Experiment using the L27 (313 ) orthogonal array
Exp. Design variable

no. A B C D E F G H I J K L M Result (mm)
1 1 1 1 1 1 1 1 1 1 1 1 1 1 0.39476
2 1 1 1 1 2 2 2 2 2 2 2 2 2 0.34191
3 1 1 1 1 3 3 3 3 3 3 3 3 3 0.28839
4 1 2 2 2 1 1 1 2 2 2 3 3 3 0.36118
5 1 2 2 2 2 2 2 3 3 3 1 1 1 0.30609
6 1 2 2 2 3 3 3 1 1 1 2 2 2 0.32529
7 1 3 3 3 1 1 1 3 3 3 2 2 2 0.34880
8 1 3 3 3 2 2 2 1 1 1 3 3 3 0.34688
9 1 3 3 3 3 3 3 2 2 2 1 1 1 0.34139
10 2 1 2 3 1 2 3 1 2 3 1 2 3 0.35816
11 2 1 2 3 2 3 1 2 3 1 2 3 1 0.35264
12 2 1 2 3 3 1 2 3 1 2 3 1 2 0.36845
13 2 2 3 1 1 2 3 2 3 1 3 1 2 0.37198
14 2 2 3 1 2 3 1 3 1 2 1 2 3 0.34462
15 2 2 3 1 3 1 2 1 2 3 2 3 1 0.33676
16 2 3 1 2 1 2 3 3 1 2 2 3 1 0.35074
17 2 3 1 2 2 3 1 1 2 3 3 1 2 0.35754
18 2 3 1 2 3 1 2 2 3 1 1 2 3 0.35462
19 3 1 3 2 1 3 2 1 3 2 1 3 2 0.37722
20 3 1 3 2 2 1 3 2 1 3 2 1 3 0.33820
21 3 1 3 2 3 2 1 3 2 1 3 2 1 0.36415
22 3 2 1 3 1 3 2 2 1 3 3 2 1 0.36814
23 3 2 1 3 2 1 3 1 2 1 1 3 2 0.35458
24 3 2 1 3 3 2 1 3 3 2 2 1 3 0.35871
25 3 3 2 1 1 3 2 3 2 1 2 1 3 0.33419
26 3 3 2 1 2 1 3 1 3 2 3 2 1 0.36348
27 3 3 2 1 3 2 1 2 1 3 1 3 2 0.33105
table is made as shown in Table 8.8. From the table, the solution is
( A1 , B 2 , C 2 , D2 , E 3 , F3 , G3 , H 3 , I 3 , J 3 , K 2 , L3 , M 3 ) where the subscript denotes
pred
the level. From this solution, the value of G opt in Equation 6.43 is estimated as
0.2634. Because the interactions are ignored, the confirmation experiment must be
exp
performed; its result is G opt 0.2880 . Since this is better than any row of the
orthogonal array in Table 8.7, it is selected as the final solution.
Interactions are checked. Using the method in Section 6.4, the following value
is calculated:
Table 8.8. One-way table for Table 8.7
A* B C D E* F G H* I J* K L M
1 0.3395 0.3538 0.3522 0.3512 0.3628 0.3579 0.3571 0.3576 0.3520 0.3555 0.3514 0.3523 0.3531
2 0.3551 0.3474 0.3445 0.3453 0.3451 0.3477 0.3483 0.3512 0.3500 0.3564 0.3430 0.3521 0.3530
3 0.3544 0.3476 0.3522 0.3553 0.3411 0.3433 0.3436 0.3401 0.3469 0.3371 0.3545 0.3444 0.3428
pred exp
| G opt G opt |
exp
0.085 (8.7)
| G opt |
The value in Equation 8.7 is small. Although we do not have a column for error,
the solution from the one-way table is better than any row of the orthogonal array.
Therefore, the interactions can be ignored in this problem.
8.4.5 Robust Design of the Rearview Mirror
Robust design is carried out with respect to the displacement on the surface of the
Table 8.9. Design variables and the ANOVA table
Factor DOF Sum of squares Mean square F

A* 2 1.398 0.699 2.705
B 2 0.236 0.118
C 2 0.357 0.179
D 2 0.478 0.239
E* 2 2.408 1.204 4.658
F 2 1.000 0.500
G 2 0.835 0.417
*
H 2 1.424 0.712 2.755
I 2 0.117 0.058
J* 2 2.136 1.068 4.133
K 2 0.637 0.318
L 2 0.365 0.183
M 2 0.626 0.313
Error 0 0 -
Total 26 12.018 0.4622
(Error) (18) (4.652) (0.258)
3
Sum of squares (10í3)
2.5
1.5
0.5
0
A B C D E F G H I J K L M
Design variables
Figure 8.32. Sensitivity of the design variables in Table 8.9
rearview mirror. The outer array is adopted. If we assume that each variable has
three levels, the L27 (313 ) orthogonal array can be utilized for the inner and outer
arrays. Then we need 27 u 27 729 analysis. Since the number of analyses is
very large, only sensitive variables are selected. The ANOVA table is utilized for
the selection and the results are given in Table 8.9. The sensitivity is illustrated in
Figure 8.32. Sensitive variables A,E,H and J are to be used in robust design.
Other variables have the values given in the previous section.
The number of design variables is four and the tolerances are shown in Table
8.10. The following S/N ratio is used:
Ki 10 log(ĭ i2 ) (8.8)
where ĭ i is the standard deviation V i of the ith row of the orthogonal array.
Since we have four design variables with three levels, the L9 (3 4 ) orthogonal array
is used for the inner and outer arrays. When the outer array is defined, the lower
Table 8.10. Tolerances of the selected variables
Variable Part Definition Tolerance

E : 50 kgf/mm2
A Frame Young's modulus and density
U : 3.5u10í11 kgfs2/mm4
E Shaft Plate thickness 0.4 mm
E : 607 kgf/mm2
H Shaft Young's modulus and density
U : 4.0u10í11 kgfs2/mm4
J Connection Frame and housing connection 1 mm
Table 8.11. Characteristic value of the inner array for the L9 (3 4 ) orthogonal array
Experiment Allocation of design variables

S/N ratio
number A E H J
1 1 1 1 1 3.89E1
2 1 2 2 2 5.47E1
3 1 3 3 3 4.72E1
4 2 1 2 3 4.04E1
5 2 2 3 1 4.86E1
6 2 3 1 2 4.84E1
7 3 1 3 2 4.29E1
8 3 2 1 3 4.79E1
9 3 3 2 1 4.91E1
Table 8.12. One-way table of Table 8.11
Level / Factor A E H J
1 46.939 40.727 45.071 45.522

2 45.770 50.404 48.055 48.677
3 46.655 48.233 46.238 45.166
value has level 1 and the upper one has level 3, and the current value has level 2.
Table 8.11 shows the standard deviation of the inner array. The one-way table
for robust design is shown in Table 8.12. This procedure is explained in Section
7.3.3. From Table 8.12, the solution is A1 E 2 H 2 J 2 , which is the same as the
second row of Table 8.11. Thus, A1 E 2 H 2 J 2 is the final solution.
8.5 Structural Analysis and Optimization of a Low

Speed Vehicle Body
In the automobile industry, lightweight material and/or new manufacturing and

assembly techniques are adopted to reduce the weight of a car. One of these
efforts is the adoption of the aluminum space frame (ASF) for the electrical car.
There are various electrical cars. Among them, the low speed vehicle (LSV) is
used for passenger transportation in neighborhoods. Because of the demand for

the LSV, the National Highway Traffic Safety Administration has defined the LSV
as a group and enacted safety laws for the vehicle. LSV is excluded from the
general passenger car group. It is noted that the crashworthiness related law has
still not been established.
A two-passenger LSV, which is currently under development, is investigated.
In many cases, there exists a reference model for automobile development. A new
model is to be designed to have better performance than the previous or reference
model. However, we do not have a reference model for the LSV because this type
of vehicle has just recently been developed. Structural performance of the LSV is
investigated by structural analysis. The results are compared to general passenger
cars. In other words, existing passenger cars are utilized as pseudo reference
models. The targets for structural performance of the LSV are defined based on a
comparative study with general passenger cars.
A regular design methodology is more useful than a simple parameter study to
achieve structural improvement. Structural optimization is a fairly good method
and has been extensively utilized for the design of structures. A reinforcement
such as a rib is employed for the cross section of an extruded part. For rib design,
topology and size optimizations are sequentially employed in the conceptual and
detailed design stages, respectively. Topology optimization is performed for the
entire LSV structure, and one part that needs to be reinforced is selected.
Topology optimization is performed again for the selected part and candidates for
rib design are found from the results of topology optimization. Structural analysis
for each rib design is performed, and a final design for the rib is selected by using
the designer’s judgment. Thicknesses of some parts including the newly designed
rib are optimized by size optimization to satisfy the targets for structural
performance. As mentioned earlier, the targets are defined from passenger cars. A
structurally improved LSV body is obtained through a series of topology and size
optimizations. The commercial software system GENESIS is used for the
optimization process (GENESIS 2004).‫ٻ‬
8.5.2 The Aluminum Space Frame and the Low Speed Vehicle
Aluminum is completely recyclable as opposed to steel. A significant proportion

of efforts is currently being directed toward the substitution of aluminum for steel
in the body structure. However, substitution of material is not straightforward.
Therefore, as illustrated in Figure 8.33, the space frame is used for the aluminium
structure while the monocoque type is used for the steel structure. The monocoque
type is a structure consisting of a torsionally rigid shell. The space frame structure
is a three-dimensional structural framework that serves as a chassis and anchor for
body panels. It is an alternative to the monocoque or conventional chassis
structure. When the aluminum space frame (ASF) structure is used instead of the
steel monocoque, mass reduction can be obtained while the original stiffness is
maintained.
Although the section thickness is generally increased to satisfy the stiffness
compared to steel structures, structural weight is reduced due to the low value of
Figure 8.33. Monocoque (left) and space frame (right) structures
the density of aluminum. Therefore, ASF body structures substantially reduce

weight compared to traditional steel monocoques, while maintaining essential
strength and stiffness. This is especially useful for the structure of an electrical
vehicle where lightweight design is essential.
A neighborhood electric vehicle (NEV) means any four-wheeled electric
vehicle whose maximum speed is not greater than 25 miles per hour. Some NEVs
look more like a passenger car than a conventional golf car whose maximum speed
is under 15 mph. A low speed vehicle (LSV) is any four-wheeled motor vehicle
whose maximum speed is greater than 20 miles per hour, but not greater than 25
miles per hour. As illustrated in Figure 8.34, the group of LSV includes NEV and
speed-modified golf cars whose maximum speed is greater than 20 miles per hour,
but not greater than 25 miles per hour. The purpose of the LSV is transportation in
closed public facilities such as golf courses, parks or airports. Power is usually
supplied by charged batteries.‫ٻ‬
The computer aided design (CAD) model is illustrated in Figure 8.35a. The
finite element method (FEM) is employed for structural analysis and the finite
element model is illustrated in Figure 8.35b. The coordinate system is presented in
Figure 8.35b. The finite element model of the LSV is made by using shell
Sub-25 mph vehicle group
Low speed vehicle

Neighborhood electric vehicle
- below 25 mph
- looks like a passenger car
Conventional golf car
- below 15 mph Speed-modified golf car
- fleet or personal golf car - 20–25 mph
Figure 8.34. Definition of the sub-25 mph vehicle group

Table 8.13. Material properties for T6-6061 aluminum
Young’s modulus 73.1 GPa

Poisson’s ratio 0.33
Mass density 2.7x10í6 kg/mm3
Yield strength 275 MPa
Ultimate strength 310 MPa
z

y
x
(a) CAD model (b) Finite element model
Figure 8.35. Schematic view of a low speed vehicle
elements for the thin space frame parts and solid elements for thick ones. Some
parts are connected to other ones through welding. For these joint parts, adjacent
nodes are shared or constrained. Material properties of the joints may be varied
due to the welding condition and can be obtained from experiments. However,
this variant is not considered here because the LSV model is under development.
Aluminum T6-6061 is used and the details of the material properties are
summarized in Table 8.13.
8.5.3 Structural Analysis of the Low Speed Vehicle and a

Comparative Study with a General Passenger Car
For structural analysis of the automobile, stiffness analysis, strength analysis and
crashworthiness analysis are usually performed. The purpose of stiffness analysis
is to improve comfort when a car is running, and the stiffness analysis is performed
by evaluating the bending and torsional characteristics. Strength analysis pursues
automobile durability. Crashworthiness is evaluated to see whether or not the
regulations are satisfied.
For a general passenger car, the loading conditions are well defined, but the
LSV does not have appropriate metrics. Stiffness and strength analyses are
performed for the LSV based on the loading conditions of a general passenger car.
Crashworthiness aspects are not considered here. The analyses are performed with
the current model. The current model is the initial design of the structure. Only
body in white (BIW) is used for structural analysis. Analysis is performed for the
initial design model in Figure 8.35.
Stiffness Analysis
The automobile body is deformed due to the energy from the engine and the
external forces from the surface of the road. The engine makes the body vibrate.
These deformation patterns are mainly bending or torsional types and cause the
passenger to feel uncomfortable. In the design stage, these problems should be
considered carefully. Static displacement responses are analyzed to find the
rigidity under a given load that causes the body to be bent or twisted. Similarly,
modal analysis finds the natural frequency of the automobile body.
In static stiffness analysis, test loads are imposed on the BIW to check bending
or torsional rigidity. Bending stiffness provides structural rigidity against a force
loaded to the side sill, which is located at the lower center parts between the front
and the rear mounts. Bending stiffness K B is obtained as follows:
| kG
y x
(a) Bending deformation
|
z kG
y
x
(b) Torsional deformation
Figure 8.36. Results of static analysis

z z
y
y x x
(a) First bending mode (b) First torsional mode
Figure 8.37. Results of the modal analysis
KB P /Gz (8.9)
where P is a force to bend the BIW globally and G is the displacement in the
direction of the force ( z direction in this LSV model).
Torsional stiffness K T is obtained when torsion is applied to the front mounts
while the rear mounts are fixed. It is
KT T /Tx (8.10)
where T is the torsion to twist the BIW globally and T is a twisting angle at the
front mount in the x direction. Figure 8.36 presents the deformed shape of the
current LSV. The initial bending stiffness is 3420 N/mm and the initial torsional
stiffness is 3455 Nm/deg.
Modal analysis finds the vibration of the BIW with the corresponding natural
frequency. Global patterns such as bending and torsional modes are examined. In
general, the first bending and torsional modes should occur above at least 25 Hz
and 40 Hz, respectively. In our example, the first bending and torsional modes
occur at 25.6 Hz and 41.4 Hz, respectively. Therefore, the current LSV satisfies
the design requirements. Each mode shape is illustrated in Figure 8.37.
Strength Analysis
There are no rigorous loading conditions for the low speed vehicle yet. Therefore,
the loading conditions for general passenger cars can be utilized with some
modification. There are several load cases for strength analysis of general
passenger cars. These simulate various cases when an automobile is running on
the road. Actually, an automobile receives loads from the road through the
suspension system. This process is omitted and the loads are directly imposed on
the front and rear mounts. The BIW can be considered simply as a cantilevered
Table 8.14. Specifications for a low speed vehicle
Total vehicle weight* 730 kg

Wheelbase 1565 mm
Track front 1205 mm
Track rear 1205 mm
Center of mass height 520 mm
Distance from front axle to mass center in x direction 782.5 mm
*weight of the assembled car with some loads and two passengers
beam. This assumption makes the analysis very easy. However, since it does not
have real boundary conditions, the analysis results can be incorrect. Therefore, the
inertia relief technique is employed to evaluate the stresses in the BIW.
Basic loadings are defined to make various loading conditions. The loading
conditions are made by multiplying the basic loadings by a constant. Five of the
basic loadings are defined and they are the reaction forces for the stopping and
turning situations. It is assumed that the center of mass is located at 782.5 mm
from the front axle and 520 mm from the ground. The acceleration of gravity g is
9.81 m/s 2 . The other sizes are shown in Table 8.14. Basic loadings are calculated
from the static equilibrium and called Ri (i 1,...,5) .
First, consider the situation when a vehicle moves forward or backward and
stops at the acceleration of 1.0 g and 0.8 g, respectively. For these cases, a free
body diagram of the left hand side of the LSV is illustrated in Figure 8.38.
Therefore, 2 R1 and 2 R2 are reaction forces of the two wheels at the front and the
730 kg
1.0 g 0.8 g
520 mm
x
782.5 mm 782.5 mm
Figure 8.38. Free body diagram of the left hand side of the LSV in a forward/reverse brake
situation
Table 8.15. Reaction forces at wheel centers for the forward/reverse brake
R1 R2
Forward braking at 1.0 g 600 N 2980 N
Reverse braking at 0.8 g 2742 N 838 N
rear axle, respectively. At the center of mass, 730 kg is loaded in the íz direction.
Using moment equilibrium at the front and the rear mount, respectively, reaction
forces are obtained as shown in Table 8.15. Two pairs of R1 and R2 are
calculated. One pair is from the forward braking at 1.0 g and the other is from the
reverse braking at 0.8 g. The reaction forces R1 and R2 are directly applied to the
mounts and are illustrated by dotted arrows in Figure 8.38.
Second, consider the situation when a vehicle turns left with the acceleration of
1.0 g. To find the reaction forces acting at the left/right wheel, a free body
diagram is drawn as illustrated in Figure 8.39. Because there are two wheels on
the left and right sides of the vehicle, 2 R3 and 2 R4 are used instead of R3 and
R4 , respectively. Figure 8.39 represents the front side. The reaction forces for
this case are obtained by using equilibrium of the moments. When the vehicle
turns right with 1.0 g, the reaction forces are similarly obtained. Therefore, two
pairs of R3 and R4 are calculated and they are summarized in Table 8.16. Finally,
when a vehicle does not move, the reaction force at each wheel R5 is one fourth of
the total weight of 730 kg.
1.0 g 730 kg
520 mm
602.5 mm 602.5 mm
z
y
Figure 8.39. Free body diagram of the front side of the LSV for the left cornering at 1.0g
Table 8.16. Reaction forces at wheel centers for left/right cornering at 1.0 g
R3 R4
Left cornering at 1.0 g 3335 N 245 N
Right cornering at 1.0 g 245 N 3335 N
Strength analysis is a durability test for various loading conditions when a car
is running on the road. Pothole brake simulates the situation when the front
wheels fall into a pothole after a sudden stop. Pothole corner (left hand/right hand)
is the case when the wheels on the left or right hand side of the vehicle fall into it.
The ultimate vertical load case is when an automobile passes an elevated place,
and the reverse brake is for a sudden backward stop. When an automobile hits a
Table 8.17. Loading conditions of strength analysis
Load case Pothole corner (LHS) Pothole corner (RHS)

Wheel x y z x y z
Front left 0 3335 10005 0 í 245 245
Front right 0 245 245 0 í 3335 10005
Rear left 0 3335 10005 0 í 245 245
Rear right 0 245 245 0 í 3335 10005
Load case Lateral curb strike (LHS) Lateral curb strike (RHS)
Wheel x y z x y z
Front left 0 6670 3335 0 í 245 245
Front right 0 245 245 0 í 6670 3335
Rear left 0 6670 3335 0 í 245 245
Rear right 0 245 245 0 í 6670 3335
Load case Pothole brake Reverse brake
Wheel x y z x y z
Front left 2980 0 8940 í 670 0 838
Front right 2980 0 8940 í 670 0 838
Rear left 600 0 1800 í 2193 0 2742
Rear right 600 0 1800 í 2193 0 2742
Load case Ultimate vertical Remarks:
Wheel x y z Each wheel location is viewed from
Front left 0 0 5370 the top of the vehicle
Front right 0 0 5370 LHS = left hand side
Rear left 0 0 5370 RHS = right hand side
Rear right 0 0 5370 Unit: Newton (N)
curb, it might climb up the curb. This is called a lateral curb strike (left hand/right
hand).‫ٻ‬
Using the basic loadings, the loading conditions are defined to simulate the
running situation and are summarized in Table 8.17. In the pothole brake
condition, loadings are applied one time in the x direction and three times in the z
direction with R1 and R2 of the forward brake situation. For both the pothole
corner and lateral curb strike conditions, basic loadings R3 and R4 are used. Two
or three times the basic loadings are applied to the specific wheels, which may
mainly resist the given loading conditions.
In the case of the left hand side pothole corner side situation, basic loadings R3
and R4 for right cornering at 1.0 g are used. The reason is as follows: when a
driver finds a pothole in the left front of the LSV, she/he actually tries to avoid it
with right cornering. However, the vehicle lands in the pothole anyway and the
left wheels fall into it eventually. Therefore, three times the reaction force R3 is
applied in the z direction to the front and the rear left wheels. Three times the
reaction force R5 is applied as the ultimate vertical load case. Finally, for the
reverse brake condition, R1 and R2 in the reverse brake situation are applied in
the z direction, and 0.8 times these values in the x direction. It is noted that the
negative sign is purposely used for the applied loading to be directed toward the
center of mass in the vehicle.
Strength analysis is performed and the results are summarized in Table 8.18.
Evaluation criteria should be used to evaluate the stress results. The yield strength
of the material is used for the pothole brake, ultimate vertical and reverse brake
loading conditions. On the other hand, the ultimate strength is compared with the
results of the pothole corner and lateral curb strike analysis. It is noted the current
LSV is safe in a running situation.
Although the vehicle is safe at all loading conditions, it is important to find
parts that have relatively high stresses. Relatively high stress occurs in the pothole
brake. In particular, the highest stress occurs in the place where the A-pillar in
Figure 8.35 meets the mount. The second highest stress occurs in the pothole
corner loading condition. The same area is weak. Additionally, the lower area of
the B-pillar should be carefully designed. In Figures 8.40 and 8.41, the
Table 8.18. Maximum stress results of strength analysis
Loading condition Maximum von Mises stress Criterion Evaluation

Pothole brake 205 Yield Safe
Pothole corner 195 (LHS), 195 (RHS) Ultimate Safe
Ultimate vertical 138 Yield Safe
Reverse brake 105 Yield Safe
Lateral curb strike 110 (LHS), 110 (RHS) Ultimate Safe
Remarks: Yield = yield strength, Ultimate = ultimate strength, unit: MPa
z
y
Figure 8.40. Deformation and stress concentration for the pothole brake condition
Figure 8.41. Deformation and stress concentration for the left corner loading conditions
deformation and stress concentration for these loading conditions are illustrated
and the areas with high stress are circled.
In Figure 8.42, the maximum stress for each loading condition is drawn by a
bar chart, and the allowable stresses for each safety factor are represented by lines.
At present, there are no standard methods of stress evaluation for the LSV. The
stresses are evaluated based on the safety factors. If the safety factor is 1.0, all the
maximum stresses of the current design are within the allowable ranges. However,
if the safety factor is set to 1.5, the current design does not satisfy the stress
constraint in the case of the pothole brake. When the safety factor is 2.0, the
ͤͦ͡
ͤ͡͡
;ΒΩΚΞΦΞ͑ΧΠΟ͑;ΚΤΖΤ͙͑;΁Β͚
ͣͦ͡
ͣ͡͡
ͦ͢͡
͢͡͡ ͲΟΒΝΪΤΚΤ
΄ΒΗΖΥΪ͑ΗΒΔΥΠΣ͑ͮ͑͢͟͡
ͦ͡ ΄ΒΗΖΥΪ͑ΗΒΔΥΠΣ͑ͮ͑ͦ͢͟
΄ΒΗΖΥΪ͑ΗΒΔΥΠΣ͑ͮ͑ͣ͟͡
͡
΁ΠΥΙΠΝΖ ΁ΠΥΙΠΝΖ ΁ΠΥΙΠΝΖ ΆΝΥΚΞΒΥΖ ΃ΖΧΖΣΤ Ζ ͽΒΥΖΣΒΝ ͽΒΥΖΣΒΝ
ΓΣΒΜΖ ΔΠΣΟΖΣ ΔΠΣΟΖΣ ΧΖΣΥΚΔΒΝ ΓΣΒΜΖ ΔΦΣΓ͑Τ ΥΣΚΜΖ ΔΦΣΓ͑Τ ΥΣΚΜΖ
͙ͽ͹΄͚ ͙΃͹΄͚ ͙ͽ͹΄͚ ͙΃͹΄͚
ͽΠΒΕΚΟΘ͑ΔΠΟΕΚΥΚΠΟΤ
Figure 8.42. Comparison of strength analysis results
current design does not satisfy four conditions. Figure 8.42 shows the safety
margin for each case.
Stiffness Comparison of the LSV to a General Passenger Car

Based on the static stiffness analysis of the LSV, the results are compared with the
stiffness of the general passenger car group, and the target values for performance
are determined from the comparison. Because the LSV does not have a
competitive or a reference model, data for general passenger cars are used. The
stiffness with respect to vehicle weight and wheelbase is used for the comparison.
First, the trend of the stiffness is identified for general passenger cars with respect
to the weight and the wheelbase. Second, the stiffness levels of the LSV model are
calculated and the results are compared to those of general passenger cars. The
targets for the LSV model are derived from the trend for general passenger cars.
The targets are determined from the stiffness values of general passenger cars
when a passenger car has the weight and the wheelbase of the LSV. The amount
of improvement is evaluated from the difference between the targets and the
current values of the LSV.
Equations 8.9 and 8.10 can be used to compare the stiffness performance. Also,
Equations 8.11 and 8.12 are often used for this purpose as well and are utilized in
this study. For bending stiffness, it is assumed that a vehicle is considered as a
simply supported beam with a concentrated load. The flexural rigidity EI is
defined as
PL(3L2 4 x 2 )
EI (8.11)
48G
where P is a concentrated load (9800 N), L is the wheelbase length (1565 mm), x is
the distance (778.5 mm) from the rear axle to the position of P, and G is the
displacement of the point where P is imposed.
For torsional rigidity GJ, the vehicle is assumed to be a cantilevered beam with
torsion. The value of GJ is defined as
TL
GJ (8.12)
T
where T is a torque (980 Nm) and T is the twisted angle where T is imposed.
The stiffness data for general passenger cars are surveyed and they are
expressed by Equations 8.11 and 8.12. The survey is conducted for Korean
passenger cars whose engine sizes are less than 2000 cc. Using Equations 8.11
and 8.12, bending and torsional stiffnesses with respect to vehicle weight and
wheelbase for general passenger cars are illustrated in Figures 8.43 and 8.44.
Stiffnesses for several models are marked by diamonds. Based on these marks,
solid lines are approximated to represent the trend of stiffness with respect to the
weight and the wheelbase.
The rigidity expressions for bending ( K BS ) and torsion ( KTS ) can be
individually established with respect to the weight and the wheelbase. However,
the average values are used to define the target values for the current LSV model
as follows:
KW B
BS K BS 417263e 0.0016 X 13351e 0.002Y
K BS (8.13)
2 2
W B
KTS KTS 580134e 0.0004 X 230820e 0.005Y
KTS (8.14)
2 2
where superscripts W and B are the vehicle weight and the wheelbase,
respectively. Subscripts BS and TS are the bending stiffness and the torsional
stiffness, respectively. X and Y are the vehicle weight (kg) and the wheelbase
length (mm). Equations 8.13 and 8.14 are derived from the solid lines in Figures
8.43 and 8.44.
The rigidity for the current LSV is marked by a rectangle in Figures 8.43 and
8.44. According to Equations 8.13 and 8.14, the target stiffness for the current
LSV is determined and marked as “x” in Figures 8.43 and 8.44. It is found that the
Table 8.19. Target rigidity, stiffness and displacement for the LSV
Loading condition K BS or K TS K B or K T Displacement
Bending 697577 Nm2 4345 N/mm 2.255 mm

Torsional 621141 Nm2/rad 6927 Nm/deg 0.1415 degree
flexural rigidity with respect to the wheelbase for the LSV is higher than the target
value, and the other rigidities are lower than the target values. Therefore, the
current LSV should be improved to some extent. The LSV model has 600 kg for
the vehicle weight, 1565 mm for the wheelbase length, 9800 N for a concentrated
load, and 980 Nm for a torque. Target values are calculated and shown in Table
1.E+07
Bending stiffness (Nm2)
1.E+06
General passenger
General passengercars
cars
LSV
LSV
Target value
Target valuefor
for LSV
LSV
Approximation considered
Approximation considering general passenger cars
general passenger cars
1.E+05
500 600 700 800 900 1000 1100 1200 1300 1400
Net vehicle weight (kg)
(a) Flexural rigidity according to the net vehicle weight

1.E+07
Bending stiffness (Nm )
2
1.E+06
General passenger
General passengercars
cars
LSV
LSV
Targetvalue
Target valuefor
for LSV
LSV
Approximation considering
Approximation considered only 4-wheeled
general cars
passenger cars
1.E+05
1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500 2600 2700
Wheelbase (mm)
(b) Flexural rigidity according to the wheelbase
Figure 8.43. Comparison of the flexural rigidity of the LSV and general passenger cars
8.19.
The target stiffness is obtained from the assumption that this low speed vehicle
is in the general passenger car group. The bending and torsional stiffness levels
for the LSV model are 79% and 50% of the target values, respectively. Structural
optimization is employed to achieve the targets.
1.E+07
General passenger cars

General cars
Torsional stiffness (Nm2/rad)
LSV
LSV
Target value
value for
forLSV
LSV
Approximation consideringonly
Approximation considered general passenger
general cars
passenger cars
1.E+06 1.E+05
500 600 700 800 900 1000 1100 1200 1300 1400
Net vehicle weight (kg)
(a) Torsional rigidity according to the net vehicle weight

1.E+07
General passenger cars

General cars
Torsional stiffness (Nm2/rad)
LSV
LSV
Target value
value for
forLSV
LSV
Approximationconsidered
Approximation consideringonly
general passenger
general cars cars
passenger
1.E+06
1.E+05
1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500 2600 2700
Wheelbase (mm)
(b) Torsional rigidity according to the wheelbase
Figure 8.44. Comparison of the torsional rigidity of the LSV and general passenger cars
8.5.4 Improvement of the Low Speed Vehicle Using Structural

Optimization
In Section 8.5.3, the target stiffness values for the current LSV model are
determined. To increase the stiffness, the moment of inertia for each part should
be increased. If the thickness of the extruded part in the LSV model is simply
increased, these target values may not be achieved because they are too large. The
stiffness can be considerably improved by using a reinforcement such as a rib for
the hollow extrusion members.
First, topology optimization for the entire LSV body is performed. Topology
optimization results in the stiffest shape of the design domain. Thus, the result
shows effective parts to increase the stiffness of the structure. When a part is
selected for stiffness improvement, topology optimization is again performed on its
cross section to find the candidates for rib design. For multiple rib candidates,
structural analysis is individually performed to evaluate them. An effective rib is
selected out of the candidates and size optimization is performed for the rib.
Optimum thicknesses of the LSV parts including the selected rib design are
obtained.
To find parts that are effective for stiffness improvement, the following
topology optimization is conducted:
Find parts (8.15a)

to minimize E bending E torsion (8.15b)
subject to mass d 0.3 mass initial (8.15c)
where Ebending and Etorsion are the strain energies for bending and torsional loading
conditions, respectively. mass and massinitial are the vehicle mass and the initial
z
y
x
Figure 8.45. Result of topology optimization for the entire model

vehicle mass, respectively. The result of this topology optimization is illustrated in

Figure 8.45. Note that the areas with circles in the model should be improved
more than the other parts. The A-pillar is selected to implement the rib.
To determine the shape of the initial rib, topology optimization is performed
for both A-pillars. Similar to other parts in the LSV model, the A-pillar is a
hollow extruded frame structure. The original section of the A-pillar, which is
modelled by shell elements in Figure 8.46a, is shown. The rib is located in the
hollow area. Before performing topology optimization, the hollow section is filled
with finite elements. A modified section of the A-pillar is illustrated in Figure
8.46a. Using solid elements, both A-pillars are remodelled and the pillars are
replaced by the remodelled ones.
The circumferential part of the A-pillar section is excluded in the design
domain of topology optimization because this area means the thickness (t) of the
original A-pillar section. The A-pillar is a long slender member; therefore, it is
difficult to model it with a long slender solid element. The ASF is made by
extrusion and its section is uniform with respect to length. When the entire A-
pillar section is selected as one design domain, it is not easy to determine the rib
candidates from topology optimization. Therefore, each A-pillar is divided into 65
design domains and some of these are shown in Figure 8.46b.
Among the results of 130 design domains, some are to be selected to make the
Nondesign
domain
Design domain
Shell element
with thickness t
(a) The original A-pillar section and the modified A-pillar section for topology optimization
(b) Nondesign domain and some design domains of the A-pillar for topology optimization
Figure 8.46. Finite element model for topology optimization of the A-pillar section
hT
A-pillar
Figure 8.47. Some results of topology optimization for the A-pillar
initial rib shape. The formulation of topology optimization is as follows:
Find the mass distribution (8.16a)

to minimize E bending E torsion (8.16b)
subject to mass k d 0.3mass initial

k
(8.16c)
where k means 130 design domains in both A-pillars. mass k and massinitial k
are
the mass of the kth design domain and the corresponding initial mass, respectively.
The mass constraint is defined for each design domain. Mass distributions are
output from the topology optimization, and some of them are selected as the
candidates for initial rib design. In Figure 8.47, some results for the 65 design
domains are illustrated. Among these results, six cases of the initial rib design are
selected and they are presented in Figure 8.48. The selection is carried out by the
designer’s judgment. Each rib design is assigned a name such as Section 2, 3, 4, 5,
6 or 7. The original A-pillar section without ribs is named Section 1.‫ٻ‬
As mentioned earlier, structural analysis is performed for every LSV model
with each rib candidate design. The stiffness increase of the six models relative to
the original model without ribs is illustrated in Figure 8.49. All sections have
better stiffness properties than Section 1. Section 5 has the best stiffness. The
bending and torsional stiffnesses of Section 5 are increased by 4.6% and 5.0%,
respectively. Therefore, Section 5 is selected for the A-pillar.
Size optimization is performed for the entire LSV model with Section 5 to
satisfy the stiffness target values. Size optimization is also performed for the LSV
with Section 1. The formulation of size optimization is as follows:
Find the thickness (8.17a)

(a) Topology optimization result (left) for determination of Sections 2 (center) and 3 (right)
(b) Topology optimization result (left) for determination of Sections 4 (center) and 5 (right)
(c) Topology optimization result (left) for determination of Section 6 (right)
(d) Topology optimization result (left) for determination of Section 7 (right)
Figure 8.48. Rib design candidates from some topology optimization results
to minimize mass (8.17b)

subject to G B d 2.255 mm (8.17c)
G T d 0.1415q (8.17d)
V j dV j
allowable (8.17e)
where G and T are the displacements of the bending and torsional loading
conditions, respectively. The limit values of G and T are taken from Table 8.19.
V j and V j allowable are the maximum and the allowable stresses for strength
analysis. Superscript j refers to each loading condition in the strength analysis.
Design variables are the thicknesses of the entire LSV members including the ribs.
The members in the front and the rear parts, which play a role in crashworthiness
performance, are not included in this optimization process.
In Figure 8.50, the history of the objective function is illustrated. Because the
current LSV model has a much lower stiffness than the target stiffness, the
ࣲ࣪࣬࣬࣡
ࣱ࣪࣬࣬࣡
एरथढढपडययࣜथऩबमफलडऩडपरय
ࣰ࣪࣬࣬࣡
࣯࣪࣬࣬࣡
࣮࣪࣬࣬࣡
࣭࣪࣬࣬࣡ ࣾडपठथपणࣜयरथढढपडयय
ऐफमयथफपࣜयरथढढपडयय
࣬࣪࣬࣬࣡
एडटरथफप࣭ࣜ एडटरथफप࣮ࣜ एडटरथफप࣯ࣜ एडटरथफपࣰࣜ एडटरथफपࣱࣜ एडटरथफपࣲࣜ एडटरथफपࣜࣳ
एडटरथफपࣜरवबडय
Figure 8.49. Stiffness comparison relative to Section 1
ͥ͢͟͡͡
ͤͦ͢͟͡ ΄ΖΔΥΚΠΟ͑ͦ
ͤ͢͟͡͡ ΄ΖΔΥΚΠΟ͑͢
;ΒΤΤ͙͑ΜΘ͚
ͣͦ͢͟͡
ͣ͢͟͡͡
ͦ͢͢͟͡
͢͢͟͡͡
ͦ͢͟͡͡
͢ ͣ ͤ ͥ ͦ ͧ ͨ ͩ ͪ ͢͡
͵Ζ Τ ΚΘΟ ͑Δ Ϊ Δ ΝΖ Τ
Figure 8.50. History of the objective function

constraint violation is very large. Therefore, the initial values of thickness are
increased by 1 mm compared to the current LSV. In the first three design cycles,
the mass of the LSV is increased to satisfy the stiffness constraints. Because the
rib is included in the A-pillars, the initial weight with Section 5 is a little heavier
by 0.9 kg than the original LSV without the rib. However, the situation is reversed
at the optimum. When the LSV with Section 5 is optimized, weight reduction is
improved by 1.9 kg compared to the case when the LSV with Section 1 is
optimized. Therefore, the rib design obtained from topology optimization
significantly reduces the weight.‫ٻ‬
8.5.5 Summary
(1) Structural analysis is performed for the low speed vehicle, which is
currently under development. Loading conditions of the general passenger
car are used for stiffness and strength analyses. The stiffness of the LSV is
compared with that of a general passenger car group. From the results of
stress, the LSV model is found to be safe when the safety factor is 1.0. A
high stress concentration occurs where the A-pillar and the mount meet.
(2) An approximation formula is proposed to find the stiffness with respect to
the vehicle weight and the wheelbase. Using this formula, the current LSV
with 600 kg for the weight and 1565 mm for the wheelbase should have
4345 N/mm and 6927 Nm/deg for bending and torsional stiffnesses,
respectively.
(3) The stiffness levels of the current LSV model are lower than the target
values defined by the formula. Structural optimization is carried out to
satisfy the stiffness target values. First, topology optimization is used to
find parts whose stiffness should be increased. After the A-pillar is
selected to be reinforced by the rib, topology optimization is performed
again to find six candidates for the rib shape. A rib design is selected from
the structural analysis. Finally, size optimization is performed to find the
optimum thickness of the entire LSV with the rib. Although the mass of
the LSV with the rib is initially heavier than that of the LSV without the rib,
the one with the rib shows better performance in optimization. Therefore,
the rib design process defined in this study is found to be efficient.
(4) To satisfy the stiffness target values, the weight of the vehicle should be
increased by 24% of the current model. This is reasonable because the
current model is designed by engineering intuition. Nevertheless, the rib
design process can be used for other parts such as the side sill and the rear
mount to further reduce the weight. Loading conditions for stiffness and
strength analyses were derived from requirements for general passenger
cars. Actually, the size and the operation range of the low speed vehicle
are different from general passenger cars and appropriate loading
conditions for the low speed vehicle should be investigated.
References
ABAQUS/Standard User’s Manual Version 6.4 (2003) Hibbitt, Karlsson, and
Sorensen Inc., Providence, RI
ANSYS User’s Manual Version 8.1 (2004) ANSYS Inc., Canonsburg, PA
GENESIS User's Manual Version 7.5 (2004) Vanderplaats Research &
Development, Colorado Springs, CO
Hwang KH, Lee KW, Park GJ (2001) Robust Optimization of an Automobile
Rearview Mirror for Vibration Reduction. Structural and Multidisciplinary
I-DEAS nx10 Help Library Bookshelf (2002) Unigraphics Solutions Inc., Cypress,
CA
National Highway Traffic Safety Administration (1998) U.S. Code of Federal
Regulations. Title 49, Part 571
Park KJ, Kang BS, Song KN, Park GJ (2003) Design of a Spacer Grid Using
Axiomatic Design. Journal of Nuclear Science and Technology 40(12):989í997
Park CM, Kim WS, Park GJ (2003) Thermal Analysis of the Roll in the Strip
Casting Process. Mechanics Research Communications 30(4):297í310
Shin MK, Shin JK, Hwang SJ, Park GJ (2006) Structural Analysis and
Optimization of a Low Speed Vehicle Body. Proceedings of the Institution of
Mechanical Engineers, Part D, Journal of Automobile Engineering, accepted
9 Application of Design Methods to Automobile
Safety Problems
9.1 Introduction
Various design methodologies are applied to problems involving design for
automobile safety. Generally, the safety of an automobile is measured by
experiments. Figure 9.1 shows a crash test and the test measures forces or
accelerations on the head, chest, legs, etc. The safety is evaluated by these
measured values. The crash test requires fairly expensive facilities and the test
itself is also expensive. Thus, sled tests are performed to simulate the crash test.
A sled test is presented in Figure 9.2. The sled is moving based on the acceleration
curve measured in the crash test and the behavior of parts is measured.
Sometimes, even the sled test can be quite costly. Therefore, computer
simulation is often utilized as a surrogate for real tests. Since computer simulation
is relatively inexpensive, it is frequently utilized when many experiments are
Figure 9.1. Crash test (courtesy of the Korea Automobile Test and Research Institute)
needed. The simulation for the crash test in Figure 9.1 generally uses nonlinear
finite element analysis and an example of a model is illustrated in Figure 9.3. The
simulation for the sled test in Figure 9.2 uses multibody dynamics and a model is
When a design is performed to improve safety, computational simulation is
usually employed. If nonlinear finite element analysis is used for computer
simulation, the computational cost is also high. Therefore, when a design
methodology such as optimization, which requires many function calculations, is
utilized, simulation using multibody dynamics is usually employed.
Figure 9.2. Sled test (courtesy of the Korea Automobile Test and Research Institute)
Figure 9.3. Finite element model of an automobile body for nonlinear analysis
(http://www.ncac.gwu.edu)
Application of Design Methods to Automobile Safety Problems 499
Figure 9.4. An occupant model for multibody dynamic analysis
9.2 Modelling for Safety Simulation
9.2.1 Modelling for Multibody Dynamic Analysis
When a safety problem is simulated by using multibody dynamic analysis, the

model illustrated in Figure 9.4 is required. It does not directly model parts, but the
body is modelled directly using nonlinear finite element analysis and the bodies are
modelled by concentrated masses or springs. The modelling technique is
important and the model is tuned by comparison with experimental results. In this
chapter, the software system called SAFE (safety analysis for occupant crash
Rim
Upper column
Shear capsule Plate bend
Lower column
Upper shaft
Hub Lower shaft
Upper shaft
Upper bracket
Lower bracket
Universal joint
(a) Steering system (b) Model of the steering system
Figure 9.5. Steering system

SD2 SD4
Force
Force
Displacement Displacement
SD6 SD7
Force
Force
Displacement Displacement
Figure 9.6. Force–displacement curves for the springs of the steering system
environment) is used (SAFE 1999). SAFE is a modified version of ATB

(articulated total body), which is in the public domain (ATB 1988). It is similar to
MADYMO (mathematical dynamical models), which is commercialized
(MADYMO 2003). The detailed theories involved in multibody dynamics are not
explained here; only the modelling process is stated.
The modelling of the human body is explained here. The HYBRID III dummy
is mainly used and it is presented in Figure 9.4 (Backaitis and Mertz 1994). Each
part of the body is modelled by an ellipsoid, and the parts are connected by springs
and dampers. Planar panels are utilized for the roof, the windshield, inner panels,
the knee bolster, bottom panels, the seat, etc. Figure 9.4 shows an example. A
steering system is illustrated in Figure 9.5a and it is modelled as shown in Figure
9.5b. The characteristics of the springs in Figure 9.5b are defined by the force–
displacement curves as illustrated in Figure 9.6. The springs and dampers for
connection are defined in the same way.
The size of the airbag in Figure 9.4 is determined by the mass flow rate, which
is the result of the inflow and outflow measured by an experiment. The surface of
the airbag can be modelled by a finite element model and the contact force with the
dummy and the steering system can be calculated. The seat belt is modelled by a
force–displacement curve similar to Figure 9.6 or a finite element model. In
simulation, the ellipsoids may contact other ellipsoids or panels. The force–
displacement curves for the contacts are obtained from experiments as illustrated
in Figure 9.7. The displacement in Figure 9.7 means the penetration between
ellipsoids or an ellipsoid and a panel. The deceleration curve of a car is obtained
from a crash test and used as an input in the simulation.
Acceleration of the head is the most important injury in occupant analysis. The
HIC (head injury criteria) is defined as follows:
Displacement
Figure 9.7. Force–displacement curve
⎛⎛ ⎞
2.5 ⎞
⎜ ⎜ 1 t2 ⎟
HIC = sup⎜ ∫ a (t )dt ⎟ (t 2 − t1 ) ⎟ (9.1)
⎜
t1 , t2 ⎜ ⎝ t 2 − t1 t1
⎟ ⎟
⎝ ⎠ ⎠
where a(t) is the acceleration of the center of the head, t1 and t 2 are the times
which maximize the HIC value during the impact. There is a chest injury index,
which is called CSI (chest severity index) and is defined as follows:
tend
CSI = ∫ a(t ) 2.5 dt (9.2)
tbegin
where a(t) is the acceleration of the center of the chest and t begin and t end are the
beginning and end times of the impact, respectively. There are injury indices for
the femur load and the maximum acceleration of the chest. Moreover, the
following index is used:
TTI (d ) 0.5 (GR + GLS )

where C is the damping matrix. The damping matrix is frequently replaced by the
linear combination of the mass matrix M and the stiffness matrix K.
There are geometric nonlinearity and material nonlinearity. M, C and K are
functions of z. The methods to solve the differential equation in Equation 9.4 are
classified into the explicit method and the implicit method. The explicit method
can solve the equation quickly; however, it does not obtain accurate solutions.
Since large-scale problems, as shown in Figure 9.3, require a large amount of
calculation, the explicit method is exploited. Details are explained in the reference
(Cook et al. 1989).
If we solve the nonlinear finite element equation for Figure 9.3, the crash test
in Figure 9.1 is reconstructed. Nonlinear finite element analysis can be applied to
parts as well and then the behaviors due to external forces can be calculated.
9.3 Design of an Airbag Using Orthogonal Arrays

An airbag is equipped at the hub of the steering system and is designed to protect a
driver at the front impact by using the method of Section 6.3.3. The following
multiobjective function is minimized:
minimize f 0.6 u HIC 0.4 u CSI (9.5)
where HIC and CSI are shown in Equations 9.1 and 9.2, respectively.
Design variables should be selected to solve Equation 9.5. Variables that
influence HIC or CSI are chosen as design variables. The design variables of the
airbag are the mass flow rate, the vent hole area, the firing time of the inflator, the
stiffness of the seat belt and the stiffness of the knee bolster. When a design
variable is selected for the tendency of a force–displacement curve, the scale factor
of the curve yields a design variable.
Table 9.1 shows the design variables and their levels for an airbag. Each
design variable has five levels and the initial designs are allocated to the third
levels. The initial design is adopted from the current existing design. The mass
Table 9.1. Design variables and their levels for an airbag
Level
1 2 3 4 5
Design variable
Scale factor of the mass flow rate 0.7 0.85 1.0 1.15 1.3
2
Vent hole area (cm ) 26 31 36 41 46
Firing time (ms) 23 26 29 32 35
Scale factor for the seat belt 0.7 0.85 1.0 1.15 1.3
Scale factor for the knee bolster 0.7 0.85 1.0 1.15 1.3
flow rate and the stiffnesses of the belt and knee bolster are defined by given
curves and the scale factors are increased or decreased from 15% to 30%. In Table
9.1, the value 1 for a scale factor means the initial design, 1.3 means an increase by
30% and 0.7 means a decrease by 30%. The vent hole area and the firing time are
changed in size. The orthogonal array L25 (5 6 ) is selected and is shown in Table
9.2. The simulation of each row is carried out by using the SAFE program
Table 9.2. Experiments using the L 25 (5 6 ) orthogonal array
Design Scale
variable Scale Objec-
Mass Vent hole Firing factor for
factor for HIC CSI tive
flow rate area time knee
seat belt function
Exp. no. bolster
1 1 1 1 1 1 312.0 335.0 321.2

2 1 2 2 2 2 447.4 342.8 405.6
3 1 3 3 3 3 464.2 319.8 406.4
4 1 4 4 4 4 350.8 323.2 339.8
5 1 5 5 5 5 434.0 314.3 386.1
6 2 1 2 3 4 503.0 340.2 437.9
7 2 2 3 4 5 478.5 350.1 427.1
8 2 3 4 5 1 542.0 341.9 462.0
9 2 4 5 1 2 438.8 346.2 401.8
10 2 5 1 2 3 184.7 302.5 231.8
11 3 1 3 5 2 526.9 406.6 478.8
12 3 2 4 1 3 509.1 415.9 471.8
13 3 3 5 2 4 491.0 352.0 435.4
14 3 4 1 3 5 232.1 299.3 259.0
15 3 5 2 4 1 562.2 331.8 470.0
16 4 1 4 2 5 755.2 437.5 628.1
17 4 2 5 3 1 670.7 495.1 600.5
18 4 3 1 4 2 328.6 306.3 319.7
19 4 4 2 5 3 352.6 315.2 337.7
20 4 5 3 1 4 237.6 329.2 274.2
21 5 1 5 4 3 643.7 406.8 548.9
22 5 2 1 5 4 320.2 303.2 313.4
23 5 3 2 1 5 259.9 310.9 280.3
24 5 4 3 2 1 633.2 399.5 539.7
25 5 5 4 3 2 330.5 350.1 338.3
Scale factor
Mass flow Vent hole Scale factor
Level Firing time for knee
rate area for seat belt
bolster
1 1859.1 2414.9 1445.1 1749.3 2393.4
2 1960.6 2218.4 1931.4 2240.6 1944.1
3 2115.0 1903.8 2126.3 2218.5 1996.7
4 2160.1 1877.9 2240.0 2105.6 1800.7
5 2020.7 1700.6 2372.7 1977.9 1980.7
Table 9.4. Initial and final solutions for the airbag
Injury Initial design Final solution

HIC 314.5 184.7
CSI 330.7 302.5
introduced in Section 9.2.

The one-way table is shown in Table 9.3 and a solution is obtained. The
solution is (level 1, level 5, level 1, level 1, level 4). The objective function is
239.1 from the confirmation experiment. The objective function of the tenth row
of Table 9.2 has a smaller value. Therefore, the tenth row is considered as the
final solution and the results are shown in Table 9.4. The injuries are decreased by
reduction of the force from the airbag to the occupant and reduction of the forward
movement of the occupant.
9.4 Design of an Energy Absorbing Steering System

Using Orthogonal Arrays

Although cars are equipped with airbags and seat belts, the steering system can
cause serious injuries of the occupant through contact. The steering system is
designed to absorb the impact energy. Here, the characteristics of an energy
absorbing steering system are investigated and the steering system is designed
using orthogonal arrays to improve the safety performance. The safety
performance is evaluated by the software system SAFE as introduced in Section
9.2. The structure of the steering system is modelled by rigid bodies and joints,
and the energy absorbing elements, and brackets are modelled by spring–damper
elements. The force–displacement curves of the spring–damper elements consider
Figure 9.8. Facility of the body block test (courtesy of the Korea Automobile Test and
Research Institute)
the unloading and reloading processes as illustrated in Figure 9.7. A model is

presented in Figure 9.5.
The analysis model is validated by comparison with the results of the body
block test. It is difficult to measure the behavior of parts by an experiment; they
are measured in detail by simulation. The facility of the body block test is
presented in Figure 9.8. A validated analysis model can show how the stiffness
changes of the energy absorbing elements affect the behavior of the steering
system and energy absorbing performance. A design idea for improving the safety
performance is proposed here. The maximum load on the body block is limited by
the FMVSS 203 regulation (FMVSS 2003). To reduce the load, an objective
function is defined using the squared value of the derivative of the load curve.
Design using DOE is efficient for highly nonlinear and noisy problems such as
occupant analysis. The method in Section 6.3.3 is employed to minimize the
objective function.
9.4.2 Modelling of the Energy Absorbing Steering System
The energy absorbing steering system consists of the steering wheel and the
column as illustrated in Figure 9.5. The rim and the hub are essential parts of the
wheel. The steering column consists of four components. The first is the upper
shaft, which is attached to the wheel. The second is the lower shaft, which
translates the torque from the upper shaft to the steering gear box through the
universal joint. The third is the upper column, which is connected to the upper
shaft through ball bearings and is attached to the mounting frame of the vehicle. It
supports the upper column. The fourth is the lower column, which fastens the
lower shaft on the mounting frame.
Six rigid segments are utilized to model the steering wheel and column, as
illustrated in Figure 9.5. Nine spring–damper elements are utilized to model the
energy absorbing parts and brackets. Using the general force–displacement
function to consider the unloading and the reloading processes, the spring–damper
can realistically estimate the behavior of the energy absorbing parts and brackets.
The force–displacement characteristics of the energy absorbing parts are obtained
from experiments with the components. The body block in Figure 9.8 is modelled
by one rigid segment with the mass and the moment of inertia.
The spring–damper SD1 in Figure 9.5b represents ball bearings which connect
the upper shaft to the upper column and transmit the rotation between them. Thus,
the force–displacement characteristics of the SD1 are modelled in order to prevent
the relative displacement between the upper shaft and the upper column. The
spring–damper SD2 represents the polyacetal molding that connects the upper
shaft to the lower shaft. SD2 absorbs the crash energy through compression. The
spring–damper SD4 represents the plate bend that absorbs the crash energy
through tension. The spring–damper SD6 represents the column fitting part that
connects the upper column to the lower column and absorbs the crash energy
through compression. The spring–damper SD7 represents the rubber flange that
connects the lower shaft to the steering gear box. The force–displacement
characteristics of the energy absorbing parts are represented in Figure 9.6. The
spring–dampers SD3, SD5, SD8 and SD9 represent brackets that connect the
steering system to the vehicle structure.
9.4.3 Validation of the Model Through the Body Block Test
The body block test is conducted in order to validate the simulation model of the
steering system. To evaluate the performance of a steering system, the test utilizes
the standard FMVSS 203 (FMVSS 2003). The standard indicates that the impact
force developed on the chest of a test dummy may not exceed 11,125 N (1,134
kgf). On impact, the dummy hits the steering system at a relative velocity of 24.14
km/h (15 mph). To correlate the analysis result to the test result, the scale factor of
the force–displacement curve is adjusted. The scale factor is the multiplier of the
1 .2
1 Test
Analysis
Normalized force
0 .8
0 .6
0 .4
0 .2
0
0 20 40 60 80
Tim e (m s)
Time (ms)
Figure 9.9. Comparison of test and analysis results

1.5
Normalizeddisplacement
displacement SD2
1
SD4
SD6
0.5
SD7
0
Normalized
0 10 20 30 40 50 60 70 80
-0.5
-1
-1.5
Time (ms)
Figure 9.10. Displacements of the energy absorbing parts
force term in the force–displacement curve. The shape (trend) of the force–
displacement curve for the energy absorbing parts is not modified from the
variation of the scale factor. As illustrated in Figure 9.9, the analysis results agree
very well with the test results.
The first peak, the second peak and the lowest concave value of the force curve
on the body block are 1.000, 0.826 and 0.315, respectively. The values for the
force curve are normalized and the forces are below 11125 N. Displacements and
forces of the energy absorbing parts are represented in Figures 9.10 and 9.11.
After approximately 22 ms, the maximum force on the body block occurs before
breakdown of the polyacetal molding (SD2) and the column fitting part (SD6).
After the breakdown, the upper column and the upper shaft move down to the
1.2
1
SD2
0.8 SD4
Normalized force
0.6 SD6
0.4 SD7
0.2
0
-0.2 0 10 20 30 40 50 60 70 80
-0.4
-0.6
-0.8
Time (ms)
Figure 9.11. Forces on the energy absorbing parts

0 ms 40 ms 80 ms
Figure 9.12. Behavior of the body block and the steering system
direction of the lower part of the steering system. The force on the body block
decreases to a lower value because of the breakdown of the molding and the
column fitting part. The force on the body block increases again according to the
metal-to-metal contact of both shafts. At a time between 22 ms and 45 ms, the
spring–damper SD4 absorbs the impact energy through the tensile deformation and
prevents the upper part of the steering system from excessive movement toward
the lower part. The SD4 controls the behavior of the upper part of the steering
system; therefore, the force on the body block does not rise too much at the time of
the metal-to-metal contact.
The behaviors of the body block and the steering system are represented in
Figure 9.12. Contact between the body block and the steering system occurs at
0 ms. The body block is separated from the steering system after 80 ms. In Figure
9.12, the steering wheel penetrates into the body block. The deformation is
evaluated by the penetration and the contact force is calculated from the
deformation. Thus, the penetration length is assumed to be the displacement
(deformation) and the impact force on the body block is evaluated from the force–
displacement curve. The same size of reaction force in the opposite direction acts
on the steering wheel. This force breaks the mold so that the upper column and the
upper shaft can have rigid body motions.
9.4.4 Parameter Study of the Energy Absorbing Parts
The parameter study of the characteristics of the steering system is conducted

through the stiffness variation of the energy absorbing parts. Parameters are the
scale factors of the force–displacement curves of the spring–dampers SD2, SD4,
SD6 and SD7. The force curves on the body block according to the stiffness
variation of the spring–dampers are represented in Figures 9.13í9.16. DN15 and
DN30 in the figures denote a 15% and a 30% decrease of the stiffness, respectively.
UP15 and UP30 are a 15% and a 30% increase of the stiffness, respectively. ORG
represents the initial value of the stiffness.
The increase of the SD2 stiffness results in the increase of the first peak and the
decrease of the second peak of the force curve on the body block, as illustrated in
Figure 9.13. The increase of the SD4 stiffness results in the increase of the first
peak and the lowest concave value, as illustrated in Figure 9.14. The second peak
is the increase after the initial decrease. This is the result due to the increase of the
lowest concave value. The trend of SD6 is similar to SD4, as illustrated in Figure
9.15. However, the increase or the decrease of the SD6 stiffness rarely alters the
trend of the force curve. The increase or decrease of the SD7 stiffness influences
the second peak of the force curve as illustrated in Figure 9.16. The trend of SD2
is opposite to that of SD7 and the softness of SD7 delays the destruction of SD2
(molding) according to the deformation increase of SD7. As represented in
Figures 9.13í9.16, the stiffness for SD2, SD4 and SD6 may be decreased and the
SD7 stiffness may be increased to reduce the first peak on the body block. In
1.2 1.2
1 1
Normalized force
Normalized force
0.8 DN30 0.8

DN15 1st Peak
0.6 ORG 0.6 Valley
UP15 2nd Peak
0.4 UP30 0.4
0.2 0.2
0 0
0 10 20 30 40 50 60 70 80 DN30 DN15 ORG UP15 UP30
Time (ms) Stiffness
Figure 9.13. Parameter study results for the spring–damper SD2
1.2 1.2
1 1
Normalized force
Normalized force
0.8 DN30 0.8

DN15 1st Peak
0.6 ORG 0.6 Valley
UP15 2nd Peak
0.4 UP30 0.4
0.2 0.2
0 0
0 10 20 30 40 50 60 70 80 DN30 DN15 ORG UP15 UP30
Time (ms) Stiffness
1.2 1.2
1 1
Normalized force
Normalized force
0.8 DN30 0.8

DN15 1st Peak
0.6 ORG 0.6 Valley
UP15 2nd Peak
0.4 UP30 0.4
0.2 0.2
0 0
0 10 20 30 40 50 60 70 80 DN30 DN15 ORG UP15 UP30
Time (ms) Stiffness

1.2 1.2
1 1
Normalized force
Normalized force
0.8 DN30 0.8

DN15 1st Peak
0.6 ORG 0.6 Valley
UP15 2nd Peak
0.4 UP30 0.4
0.2 0.2
0 0
0 10 20 30 40 50 60 70 80 DN30 DN15 ORG UP15 UP30
Time (ms) Stiffness
order to reduce the second peak, the SD2 stiffness may be increased and the SD7
stiffness may be decreased. According to the above characteristics for the spring–
dampers, it is not easy to reduce both peaks of the force curve.
9.4.5 Definition of the Objective Function
The objective function should be defined appropriately for an effective design of

energy absorbing parts. As mentioned in Section 9.4.3, the peak force is an
important measure in FMVSS 203; therefore, it should be minimized. To
minimize the peak force, the objective function utilizes the quantity related to the
force curve instead of the peak force itself. The reason for this is that both peaks
of the force curve move slightly from the initial position as illustrated in Figures
9.13í9.16.
Two idealized curves with identical impulse or momentum, which are
transmitted to the occupant in crash environment, are illustrated in Figure 9.17.
The areas between the curves and the horizontal axis are the same due to the
conservation of momentum. Suppose the curves are symmetric with respect to the
vertical line on t m . The curve C A in Figure 9.17 is more desirable than C B
because the maximum load is smaller. The curves in Figure 9.17 are differentiated
and the absolute values of the derivatives are integrated (summed). Then the
CB
f2
Force
f1 CA
Time
t1 t2 tm tf
Figure 9.17. The idealized force–time curves

Table 9.5. Summation values of the derivatives
Curve C A Curve C B
§ f1 · § f2 ·
¨¨ ¸¸ u t1 0 u (t m t1 ) f1 ¨¨ ¸¸ u t 2 0 u (t m t 2 ) f2
From 0 to t m
© t1 ¹ © t2 ¹
§ f1 · § f2 ·
From t m to t f
¨¨ ¸¸ u t1 0 u (t m t1 ) f1 ¨¨ ¸¸ u t 2 0 u (t m t 2 ) f2
© t1 ¹ © t2 ¹
Total 2 f1 2 f2
integrated value for C A is 2 f1 and that for C B is 2 f 2 . The calculation is shown

in Table 9.5. Since f1 f 2 , the integrated value for C A is smaller than that of
C B . Thus, the design variables are determined to minimize the integrated
(summation) value of the derivatives. Then we can find a curve with a low peak
value. To avoid the influence of the sign, an objective function is defined as
2
§ df ·
Fobjective (b) ³ ¨ ¸ dt (9.6)
© dt ¹
9.4.6 Design Using Orthogonal Arrays

Optimization using sensitivity is not appropriate, since occupant analysis is highly
nonlinear and noisy. Therefore, the method in Section 6.3.3 is employed to obtain
the design of the energy absorbing parts. An orthogonal array is used to minimize
Equation 9.6. As shown in Table 9.6, an orthogonal array is selected to consider
four design variables with five levels. Design variables are the scale factors of the
force–displacement curves for spring–dampers SD2, SD4, SD6 and SD7. In Table
9.6, the first and second levels are where the stiffnesses are reduced by 30% and
15%, respectively. The stiffness is unchanged for the third level. The stiffness is
increased by 15% and 30% for the fourth and fifth levels. The values for Equation
9.6 are shown in Table 9.6 and the one-way table for Table 9.6 is shown in Table
9.7.
A one-way table is constructed as shown in Table 9.7. The solution of each
design variable is the level with the minimum value. Thus, the solution is the first
level of SD2, the fifth level of SD4, the fifth level of SD6 and the first level of
SD7. Since the solution is better than any row in Table 9.6, it is considered as the
final solution. The summation of the derivatives, the first peak, the second peak
5
Table 9.6. Results using the L25 (5 ) orthogonal array
2
Exp. Design variable Normalized force § df ·
no.
³ ¨ ¸ dx
SD2 SD4 SD6 SD7 1st peak 2nd peak Valley © dx ¹
1 1 1 1 1 0.951 0.878 0.153 1.134
2 1 2 2 2 0.964 0.855 0.233 1.055
3 1 3 3 3 0.967 0.852 0.310 0.983
4 1 4 4 4 0.975 0.864 0.379 0.943
5 1 5 5 5 0.982 0.868 0.440 0.878
6 2 1 2 3 0.963 0.870 0.185 1.188
7 2 2 3 4 0.968 0.858 0.261 1.066
8 2 3 4 5 0.974 0.841 0.334 0.985
9 2 4 5 1 0.991 0.827 0.374 0.915
10 2 5 1 2 0.998 0.830 0.388 0.918
11 3 1 3 5 0.979 0.926 0.216 1.234
12 3 2 4 1 0.994 0.812 0.251 1.025
13 3 3 5 2 0.999 0.830 0.329 0.989
14 3 4 1 3 1.001 0.826 0.345 0.964
15 3 5 2 4 1.010 0.826 0.405 0.923
16 4 1 4 2 0.990 0.827 0.202 1.139
17 4 2 5 3 0.989 0.825 0.284 1.034
18 4 3 1 4 0.993 0.824 0.300 1.019
19 4 4 2 5 1.005 0.814 0.365 0.966
20 4 5 3 1 1.029 0.777 0.410 0.883
21 5 1 5 4 0.981 0.899 0.238 1.174
22 5 2 1 5 0.986 0.888 0.258 1.132
23 5 3 2 1 1.006 0.784 0.290 0.971
24 5 4 3 2 1.024 0.799 0.359 0.946
25 5 5 4 3 1.043 0.779 0.419 0.916
and the lowest concave value are 0.864, 0.987, 0.839 and 0.425, respectively. The
summation of the derivatives and the first peak are decreased by 13.6% and 1.3%,
respectively. The second peak and the lowest concave value are increased by
1.58% and 34.9%, respectively.
Generally, it is easy to obtain an improved solution for a problem with one
peak. However, a problem with two peaks is difficult to handle due to the valley.
In this case, the engineer can select an appropriate solution from the orthogonal
Level SD2 SD4 SD6 SD7
1 4.992 5.868 5.167 4.929

2 5.072 5.312 5.102 5.047
3 5.135 4.948 5.113 5.084
4 5.041 4.734 5.007 5.125
5 5.139 4.518 4.990 5.195
array. Suppose we select the third row in Table 9.6. The summation value is
reduced by 1.7% and the first peak is reduced by 3.3%. The lowest concave value
is reduced by 1.6% and the second peak is increased by 3.2%. This is the case
where the solution is picked from the rows of the orthogonal array by using
engineering sense.
9.4.7 Summary
Using the occupant analysis program SAFE, the characteristics of the energy
absorbing steering system are examined and improved. The analysis model is
established by comparing the analysis model and the experiments based on the
FMVSS 203 regulation. To reduce the impact load on the body block, an objective
function is proposed as the integrated value of the squared sum of the derivatives
of the load curve. An orthogonal array is employed for an improved design. A
design solution is obtained from the one-way table. Because there are two peaks, a
better solution is suggested by picking one from the rows of the orthogonal array.
In practical engineering, designers may generate multiple cases and determine the
final design from these cases. The solution here is the same as in the method
where the performances are evaluated using only orthogonal arrays.
9.5 Design of a Side Impact Beam of an Automobile

Door

There are regulations for side impact experiments. Here, the experiment for the
strength of the door is carried out by pushing a steel cylinder or semicylinder into
the door. The radius of the column is 305 mm. A load is imposed on the door by
the cylinder. The cylinder is pushed by a velocity of 12.7 mm/s until the total
displacement is 460 mm. The experiment should be finished within 120 seconds.
All dofs fixed by springs
Uniform velocity
800 mm/s
All dofs fixed by springs
Figure 9.18. Simulation of the impact beam
The side impact experiment is performed with an assembled door. In this

section, the design is carried out to enhance the resistance of the side impact beam.
Therefore, only the impact beam is considered. Computer simulation is utilized for
the evaluation. The simulation uses the nonlinear finite element software system
LS/DYNA3D (LS/DYNA 2003), which uses the explicit integration method. If
we use the velocity of the impact for regulation, the calculation requires an
extremely large CPU time. Therefore, the velocity is increased to 800 mm/s in the
simulation. Figure 9.18 represents the simulation process. The modelling and
boundary conditions are illustrated in Figure 9.19. The governing equation used in
this simulation is shown in Equation 9.4.
The results of the experiment are illustrated in Figure 9.20. The performance
of the door can be evaluated by the average penetration at 155 mm, 310 mm and
460 mm. Here, the resistance at 155 mm is considered as the performance for
Figure 9.19. Modelling and boundary conditions

Disjoint of the weld
Resistance in the early stage

(155 mm)
Figure 9.20. Experimental results of the impact beam
penetration. It is noted that the load is drastically changed around 120 mm of

penetration. This is because the welding parts are disjoined from the bracket.
Therefore, the value at 120 mm is used when the simulation model is tuned by
experiments.
9.5.2 Problem Formulation
In side impact, the distance between an occupant and the impact beam is short.
Thus, enhancement of the stiffness of the impact beam is more important than
energy dissipation. If the size of the impact beam is increased, then the stiffness is
increased. However, the weight, which should be minimized, is also increased.
Therefore, an optimization problem is formulated as follows:
Find a, b, t (9.7a)
to maximize f(a, b, t) (9.7b)
subject to w < w0 (9.7c)
The impact beam has an ellipsoidal shape. a is the major axis, b is the minor axis,
t is the thickness of the impact beam, f(a, b, t) is the contact force between the
impact beam and the impactor (cylinder), w0 is the initial weight of the impact
beam, and w is the weight of the impact beam.
Generally, the impact beam has a circular shape. However, here it is defined
by an ellipsoid. The impact beam is designed using the method in Section 6.5 and
the response surface method in Section 6.7. The contact force is the performance
index for the resistance to penetration.
9.5.3 Design of an Impact Beam Using Orthogonal Arrays
Using the penalty function P(b) in Equation 6.60, a characteristic function <(b)
to be maximized is defined as follows:
<(b) f (b) P(b) (9.8)
Equation 9.8 is slightly different from Equation 6.60 because f should be

maximized here. It may be difficult to satisfy all the constraints with Equation 9.8.
This difficulty can be overcome by using a large value for the scale factor.
However, the influence of the objective function can be ignored due to the large
scale factor. Therefore, the scale factor is defined to let the penalty function be
one order less than the objective function. The method in Section 6.5 is employed.
The numbers of factors and levels are 3. If we use the method to select an
orthogonal array in Section 6.3, the L9 (34 ) is selected. However, the L18 (21 u 37 )
orthogonal array is used to enhance accuracy. It is noted that the interactions are
evenly distributed in the L18 (21 u 37 ) orthogonal array. 18 experiments are
performed and the one-way table is used to determine the design solution.
The levels of design variables are shown in Table 9.8 and the results of
experiments are shown in Table 9.9. From the one-way table, the solution is
[a1 b3 t1 ]T [17 11 2.6]T (mm). The confirmation experiment is carried out and
the contact force is 3.16 u 10 7 kgmm/s. The constraints are violated. Therefore,
the solution, which has the maximum contact force while the constraints are
satisfied, is the 12th row of Table 9.9.
As explained in Section 6.5, the estimated values are lined up in ascending
order. The results are shown in Table 9.10. The constraints are satisfied in the
fourth row and this is the same as in the previous solution. Therefore, the final
solution is [a1 b3 t 2 ]T [17 11 2.4]T (mm). Table 9.11 shows the initial and final
designs.
9.5.4 Design Using the Response Surface Method
The response surface method is explained in Section 6.7. The objective function is
approximated by a quadratic function. The candidate points for the approximation
Table 9.8. Levels of design variables ( u 10 3 m)
Design Level
variable 1 2 3
a 17.0 15.5 14.0
b 14.0 12.5 11.0
t 2.6 2.4 2.2
Table 9.9. Results from the orthogonal array
Column no. and design variable (×10-3m)

Exp. < (b)
1 2 3 4 5 6 7 8
no.
Error a b t Error Error Error Error
1 0 17.0 14.0 2.60 0 0 0 0 3.51E+07
2 0 17.0 12.5 2.40 0 0 0 0 2.83E+07
3 0 17.0 11.0 2.20 0 0 0 0 2.69E+07
4 0 15.5 14.0 2.60 0 0 0 0 2.75E+07
5 0 15.5 12.5 2.40 0 0 0 0 2.71E+07
6 0 15.5 11.0 2.20 0 0 0 0 2.59E+07
7 0 14.0 14.0 2.40 0 0 0 0 2.50E+07
8 0 14.0 12.5 2.20 0 0 0 0 2.39E+07
9 0 14.0 11.0 2.60 0 0 0 0 2.72E+07
10 0 17.0 14.0 2.20 0 0 0 0 2.98E+07
11 0 17.0 12.5 2.60 0 0 0 0 3.07E+07
12 0 17.0 11.0 2.40 0 0 0 0 2.87E+07
13 0 15.5 14.0 2.40 0 0 0 0 2.48E+07
14 0 15.5 12.5 2.20 0 0 0 0 2.47E+07
15 0 15.5 11.0 2.60 0 0 0 0 2.87E+07
16 0 14.0 14.0 2.20 0 0 0 0 2.34E+07
17 0 14.0 12.5 2.60 0 0 0 0 2.82E+07
18 0 14.0 11.0 2.40 0 0 0 0 2.67E+07
are the points of L18 (21 u 37 ) in Table 9.9 and the results in Table 9.9 are utilized.
The approximated function is
f (a, b, c ) 421010 26773274a 16339739b 88044159t
6.25 u 108 a 2 1.89 u 108 b 2 1.44 u 1010 t 2 5.58 u 108 ab
1.27 u 109 bt 7.13 u 108 ta (9.9)
Optimization is performed by using Equation 9.9 as the objective function and

Equation 9.7c as a constraint. The optimum solution is [a b t ]T [17 11 2.403]T
(mm). a converges to the upper bound and b converges to the lower bound. The
solution is similar to the previous one.
Table 9.10. Results in ascending order
Exp. Design variable

Estimated value Constraint
no. a b t
1 30677777.8 1 3 1 Violated
2 29977777.8 1 2 1 Violated
3 29861111.1 1 1 1 Violated
4 28961111.1 1 3 2 Satisfied
5 28261111.1 1 2 2

25 24544444.4 2 1 3
26 24377777.8 3 2 3
27 24261111.1 3 1 3
Table 9.11. Initial and final designs
Result Improved rate

Design Contact force Contact
Weight (kg/kg) Weight
(N/N) force
Initial 1.000 1.000 - -
Final 1.196 0.896 19.6% í10.4%
9.5.5 Summary
The cross section of the impact beam is designed to maximize the contact force
and to reduce the weight. The design process is carried out by using an orthogonal
array and the response surface method.
(1) The design results from the two methods are compared and are found to be
similar.
(2) When an orthogonal array is used for the design, the contact force is
increased by 19.6% and the weight is reduced by 10.4%. The response
surface method increases the contact force by 19.44% and reduces the
weight by 10.39%.
9.6 Design of a Motor Driven Tilt/Telescopic Steering

System

A motor driven tilt/telescopic steering system is a steering system that has a few
more functions than the one introduced in Section 9.4. The added functions are
the control of the position to fit the size and the position of the driver and the
capability of memorizing the position. Thus, when a driver gets in or out of the car,
the steering system moves forward so that sufficient space is guaranteed. When
the driver drives, the position of the steering system returns to the original
memorized position. This is a tilt/telescopic function. For this function, the
system has additional parts such as one or two motors, gears, a gear box, a tilt tube,
a tile bracket, a tilt housing, a spline shaft, etc. The performance of the system
depends upon the selection of the parts.
The basis functions of the motor driven tilt/telescopic steering system are the
steering and tilt/telescopic functions. Moreover, the system should not vibrate
during driving. The steering function is determined by the steering shaft and the
universal joint that connects the shaft to the car body. The tilt/telescopic function
is operated by motors. The safety and vibration performances are dependent upon
the positions and sizes of various parts. Axiomatic design is utilized in the
conceptual design process, while optimization and orthogonal arrays are used in
the detailed design process.
9.6.2 Conceptual Design Using Axiomatic Design
The axiomatic approach is utilized to analyze the overall design process. The
functional requirements(FRs) for the product are as follows:
FR1 : Let the system have the capability of tilt/telescopic operation.
FR2 : Let the transmission of the torque be easy.
FR3 : Let the strength of the steering column be sufficient.
The design parameters (DPs) are defined as follows:

DP1 : Gears, motors and universal joints
DP2 : The positions and angles of universal joints
DP3 : The arrangement and sizes of components
DP1 for FR1 is defined because the tilt/telescopic function is enabled by the
rack and pinion gears, which are operated by a motor. The function becomes easy
to operate if we use the universal joints and splines. The positions and angles of
the two universal joints determine the steering function; therefore, they are
selected as DP2 . Since the strength of the system depends on the positions and
sizes of the parts, they are defined as DP3 .
The main interests are the safety and vibration performances. Thus, only FR3
and DP3 are explained. The lower level function requirements ( FR3 x ) for FR3
FR31 : Maintain the bending strength.
FR32 : Absorb impact energy from vehicle collision.
FR33 : Avoid the resonance caused by the combustion of the engine, the
irregularity of the road and the variation of the brake torque.
The bending strength should be maintained to keep the configuration of the
structure. If there is no airbag, the steering system directly contacts the driver.
Thus, the steering system should absorb the impact energy as much as possible.
The steering system should not vibrate excessively from the external forces during
operation. When the system is resonant, comfort is deteriorated. Therefore, DP3 x
DP31 : The positions of the support and bolt (tilt point)
DP32 : The mechanism for the energy absorption such as the bolt ejection,
tube-in/tube-out insertion, the main energy absorber and three-point
support
DP33 : The shapes, sizes and gaps of components
Because the positions of the support and the bolt for the tilt point are crucial for
FR31 , DP31 is defined as above. DP32 is selected with respect to FR32 as
follows: During impact, the bolt of the main bracket is ejected. The main bracket
is a part that is connected to the car body. The upper and lower columns are
compressed, forces are imposed on the energy absorbing parts and impact energy
is absorbed. Figure 9.21 shows the hierarchies for the functional requirements and
the design parameters.
Equation 9.10 is the overall design equation and Equation 9.11 is the design
equation for FR3 .
FR DP
FR1 FR2 FR3 DP1 DP2 DP3
FR31 FR32 FR33 DP31 DP32 DP33
(a) Functional domain (b) Physical domain
Figure 9.21. Hierarchies of the FR–DP mapping for the steering system
Figure 9.22. Prototype of a motor driven tilt/telescopic steering system
« FR » «X X 0 »» «« DP2 »» (9.10)
« 2» «
«¬ FR3 »¼ «¬ X X X »¼ «¬ DP3 »¼
« FR » «0 X 0 »» «« DP32 »» (9.11)
« 32 » «
«¬ FR33 »¼ «¬ X X X »¼ «¬ DP33 »¼
where 0 means there is no relation and X means there is a relation. Equations 9.10
and 9.11 represent decoupled designs.
Design for safety and vibration reduction should be performed according to the
sequence indicated by Equation 9.11. Then the Independence Axiom is satisfied.
The detailed design is carried out to find the detailed sizes of DP3 x . Figure 9.22
presents a prototype of the steering system.
9.6.3 Detailed Design for Safety
In this section, detailed design of DP32 is carried out to satisfy FR32 . The force
displacement curves for energy absorbing parts are determined. The steering
system consists of the steering wheel and the steering column. The rim and hub of
the steering wheel are modelled by a contact ellipsoid that does not have mass, as
explained in Section 9.4. Therefore, the corresponding mass and moment of
inertia are included in the steering shaft. The other parts are modelled by segments
that have the mass and the moment of inertia. The connections between parts are
modelled by slip joints or universal joints. Since the slip joint prevents the motion
in the axial direction, a spring–damper, which has a large stiffness coefficient, is
used. Figure 9.23 illustrates the model.
Figure 9.23. Model of a motor driven tilt/telescopic steering system
The force displacement curves for the spring–dampers are obtained from static
experiments and they are tuned in the modelling process. The parts in Figure 9.23
are as follows:
G S1: Ball bearing
G S2: Main bracket of the upper body
G S3: Main energy absorbing device
G S4: Spline
G S5: Screw of the telescopic motor
G S6: Tube-in and tube-out energy absorbing part
G S7: Bracket of the lower body
G S9: Universal joint
G S8, S10: Mounting parts
S11: Energy absorbing part by ejection of the bolt
G S12: Support for the telescopic housing
S1 is the spring–damper that models the ball bearing between the upper column
3 - p o in t su psupport
Three-point p o r t ( S 1 2(S12)
) F - D F-D
c u r vecurve
The energy
T h e en e rgy ababsorbing component
so rb in g co m p o n e n t (S 3 )(S3)
F -D F-D
cu rvecurve
value)
value)
͢ ͢ ͣ͟
Force/(nominal value)
͡ ͩ͟ ͢
Force/(nominal
Force/(nominal
͡ ͩ͟
͡ ͧ͟
͡ ͧ͟
͡ ͥ͟
͡ ͥ͟
͡ ͣ͟ ͡ ͣ͟
͡ ͡
͡ ͡ ͣ͟ ͡ ͥ͟ ͡ ͧ͟ ͡ ͩ͟ ͢ ͡ ͡ ͣ͟ ͡ ͥ͟ ͧ͟͡ ͡ ͩ͟ ͢
Displacement/(nominal
D eflec tio n /(n o m in al va lu value)
e) Displacement/(nominal
D e fle c tio n /( n o m in a l va luvalue)
e)
B olt
Bolt ejec tion(S11)
ejection (S 1 1 )F-D
F -D curve
cu rve ΆΡΡΖΣ͠ͽΠΨΖΣ͑ΔΠΝΦΞΟ͙͑΄͚ͧ͑ͷ͞͵͑ΔΦΣΧΖ
Upper/lower column (S6) F-D curve
value)
value)
͢ ͢
ͷΠΣΔΖ͙͠ͿΠΞΚΟΒΝ͑·ΒΝΦΖ͚
ͩ͟͡ ͩ͟͡
Force/(nominal
Force/(nominal
ͧ͟͡ ͧ͟͡
ͥ͟͡ ͥ͟͡
ͣ͟͡ ͣ͟͡
͡ ͡
͡ ͡ ͣ͟ ͡ ͥ͟ ͡ ͧ͟ ͩ͟͡ ͢ ͑͑͑͑͑͑͑͑͑͑͑͑͑͑͑ͣ͑͑͑͑͑͑͑͑͑͑͑͑͑ͥ͑͑͑͑͑͑͑͑͑͑͑͑͑ͧ͑͑͑͑͑͑͑͑͑͑͑͑͑͑ͩ͑͑͑͑͑͑͑͑͑͑͑͑͑͑͑͟͟͟͟͢͡͡͡͡͡
͟͡͡ ͟͡͡ ͤ͟͡ ͥ͟͡ ͩ͟͡ ͩ͟͡ ͪ͟͡ ͢͟͡
͵ΖΗΝΖΔΥΚΠΟ͙͠ͿΠΞΚΟΒΝ͑·ΒΝΦΖ͚
Displacement/(nominal value)
Displacement/(nominal
D efle ction /(n om in a l va luvalue)
e)
Figure 9.24. Force displacement curves for energy absorbing devices (S3, S6, S11, S12)
and the tilt-tube and its deformation is negligible. S2 connects the main bracket of
the steering system with the vehicle body. S3 is the main energy absorbing (EA)
device. When impact occurs, it absorbs the impact energy by elongation.
S4 connects male and female splines. S5 represents the screw for transmitting
power from the motor. It connects the telescopic housing with the motor mounted
on the tube-out and the deformation is negligible. S6 is an EA device to absorb the
Table 9.12. Design variables and levels
1 2 3 4 5
S3 0.85 0.93 1.0 1.07 1.15
S6 0.85 0.93 1.0 1.07 1.15
S11 0.85 0.93 1.0 1.07 1.15
S12 0.85 0.93 1.0 1.07 1.15
Table 9.13. The L25 (5 5 ) orthogonal array
Exp. no. S3 S6 S11 S12 Impact load

1 1 1 1 1 0.9227
2 1 2 2 2 0.9789
3 1 3 3 3 0.9948
4 1 4 4 4 1.0112
5 1 5 5 5 1.0210
6 2 1 2 3 0.9689
7 2 2 3 4 1.0048
8 2 3 4 5 1.0229
9 2 4 5 1 0.9592
10 2 5 1 2 0.9857
11 3 1 3 5 1.0282
12 3 2 4 1 0.9630
13 3 3 5 2 0.9988
14 3 4 1 3 0.9934
15 3 5 2 4 1.0104
16 4 1 4 2 1.0008
17 4 2 5 3 0.9891
18 4 3 1 4 1.0024
19 4 4 2 5 1.0169
20 4 5 3 1 0.9877
21 5 1 5 4 1.0248
22 5 2 1 5 1.0153
23 5 3 2 1 0.9859
24 5 4 3 2 1.0110
25 5 5 4 3 1.0140
1 2 3 4 5
S3 4.9286 4.9415 4.9938 4.9969 5.0510
S6 4.9454 4.9511 5.0048 4.9917 5.0188
S11 4.9195 4.9610 5.0265 5.0119 4.9929
S12 4.8185 4.9752 4.9602 5.0536 5.1043
Table 9.15. Expanded design variables and levels
1 2 3 4 5
S3 0.3 0.5 0.7 1.0 1.2
S6 0.3 0.5 0.7 1.0 1.2
S11 0.3 0.5 0.7 1.0 1.2
S12 0.3 0.5 0.7 1.0 1.2
impact energy by the relative movement between the tube-in and the tube-out. S7
and S9 are connections between the lower bracket of the steering system and the
vehicle body. S8 and S10 make the universal joint connected to front wheels. S11
is an EA device ejected from the main bracket to absorb the impact energy at the
beginning of impact. S12 is a three-point EA device connected between the main
bracket and the telescopic housing. It enhances the stability of the system.
The data for the analysis model has been tuned using the FMVSS 203 test,
which is standard for the steering system. The test is carried out at 24.14 km/h
Body block impact load

1.2
Initial
1
Impact load
Orthogonal
0.8
0.6
0.4
0.2
0
0 10 20 30 40 50 60 70 80
Time (ms)
Time (msec)
Figure 9.25. Comparison of the initial design and the solution

(15 mph) and the impact load acting on the body block must not exceed 11,125 N
(1,134 kgf). The analysis model is established by tuning the force-displacement
curves for the spring–dampers. The data are made from experiments and modified
to match the results between the FMVSS 203 test and the numerical analysis.
The objective of the design is to minimize the impact load on the body block.
The characteristics of the spring–dampers are the most important for the evaluation
of the impact load. Thus, the scale factors for the spring–dampers are selected as
design variables to minimize the impact load. Some important spring–dampers are
chosen for design variables through sensitivity analysis. They are the EA devices
such as S3, S6, S11 and S12 shown in Table 9.12. The number of levels for the
design variables is five where the stiffness is varied from í15% to 15%. The
stiffness is varied by the scale factor of the force displacement curves as illustrated
in Figure 9.24. Since we have four design variables with five levels, the L25 (5 6 )
orthogonal array in Table 9.13 is utilized. Then 25 analyses are conducted for the
design.
Table 9.16. Orthogonal array for Table 9.15
Exp. no. S3 S6 S11 S12 Impact load

1 1 1 1 1 0.9697
2 1 2 2 2 0.8968
3 1 3 3 3 0.8770
4 1 4 4 4 0.9543
5 1 5 5 5 0.9963
6 2 1 2 3 0.8669
7 2 2 3 4 0.9350
8 2 3 4 5 0.9884
9 2 4 5 1 0.9002
10 2 5 1 2 0.8499
11 3 1 3 5 0.9708
12 3 2 4 1 0.8983
13 3 3 5 2 0.9126
14 3 4 1 3 0.8862
15 3 5 2 4 0.9618
16 4 1 4 2 0.9074
17 4 2 5 3 0.9267
18 4 3 1 4 0.9539
19 4 4 2 5 1.0123
20 4 5 3 1 0.8929
21 5 1 5 4 1.0177
22 5 2 1 5 1.0079
23 5 3 2 1 0.8839
24 5 4 3 2 0.9056
25 5 5 4 3 0.9969
1 2 3 4 5
S3 4.6941 4.5404 4.6297 4.6932 4.8120
S6 4.7325 4.6647 4.6158 4.6586 4.6978
S11 4.6676 4.6217 4.5813 4.7453 4.7535
S12 4.5450 4.4723 4.5537 4.8227 4.9757
After the analysis model is established, the analysis for each row in Table 9.13
is carried out. The results in Table 9.13 are normalized. When the one-way table
in Table 9.14 is made, the solution is found to have the first levels. This is the first
row of Table 9.13 and also the best row in Table 9.13. The initial design is
compared with the solution in Figure 9.25. The peak is reduced by 7.7%.
The lowest value of the stiffness is selected. The ranges of the design variables
are expanded as shown in Table 9.15. New experiments are shown in Table 9.16
and the one-way table for Table 9.16 is shown in Table 9.17. The confirmation
experiment with the solution of Table 9.17 shows that the impact load is 0.8790
and this is compared with the results in Table 9.16. It is found that the tenth row
of Table 9.16 is the best and the impact load is 0.8499.
The values in Table 9.15 are arbitrarily expanded. It is found that the real parts
cannot be manufactured with the values in Table 9.15. Therefore, the usable
solution is the one from Table 9.13.
9.6.4 Detailed Design Considering Vibration
In the conceptual design process, the functional requirement FR33 is to prevent

resonance. Thus, DP33 is determined here. The detailed design for DP33 is
performed to improve the vibrational performance by using size and shape
Figure 9.26. Finite element model of the steering system

Table 9.18. Experimental and analysis results
Modal test Modal analysis

1st 25 25
2nd 90 110
optimizations. The objective is to increase the first natural frequency to avoid

resonance. The vibration of the steering system is caused by the vibration of the
engine, the irregularities of the road and the variation of the brake torque. In
general, the frequencies of the causes exist between 10 Hz and 30 Hz. Therefore,
the steering system is usually designed to keep the first natural frequency in the
range over 30 Hz to prevent resonance. The steering system is modelled for finite
element analysis and the model is validated by experiments. Then optimization is
conducted on the established model and a new design is proposed to increase the
first natural frequency.
The commercial software system GENESIS (GENESIS 2004) is employed for
structural analysis and optimization. Figure 9.26 depicts the finite element model.
Table 9.18 shows experimental and analysis results.
Part 4
Part 3
Part 1
Part 1
Part 2 Part 2 Part 3 Part 4
(a) Telescopic housing (b) Main bracket
Part 1
Part 2
Part 3
(c) Tilt shaft
Figure 9.27. Design variables for size optimization

Table 9.19. Design variables for optimization
Design variable
Tilt-tube, tilt-bracket, main-bracket
Size variables Base-bracket, housing, motor-bracket,
Column-out
Shape variables Housing, main-bracket, spline shaft male
Using the analysis model in Figure 9.26, optimization is conducted to increase

the first natural frequency. Design variables are shown in Table 9.19 and some of
these are indicated in Figure 9.27. The optimization formulation is as follows:
Find design variables (b) (9.12a)

to minimize the weight of the steering column (9.12b)
subject to the first natural frequency > 30 Hz (9.12c)
V d V all (9.12d)
bL d b d bU (9.12e)
Size and shape optimizations are carried out simultaneously. The first natural
frequency is increased to 27.1 Hz from the optimization process and the constraint
in Equation 9.12c is violated. In other words, a feasible region is not found from
the current configuration.
3
Initial
Magnitude (displacement)
Optimum
2
0
0 20 40 60 80 100 120 140
Frequency (Hz)
Figure 9.28. Harmonic response analysis results for the initial and optimum designs
Table 9.20. Optimization results
Optimum value
Weight (kg) 11.1
Frequency (Hz) 30
Table 9.21. Optimum design variables (unit: mm)
Design variable Initial Final

Tilt-tube 2 4
Tilt-bracket 5.5 7.5
Main-bracket
(Part 1) 10 12.5
(Part 2) 6 8
(Part 3) 7.5 7
(Part 4) 6 9.5
Base-bracket 4.5 4.5
Housing
(Part 1) 6 6.5
(Part 2) 9 10
(Part 3) 6 10
(Part 4) 8 16.1
Column-out 2.25 2.5
Motor-bracket 8 10
Tilt shaft
(Part 1) 2 2.5
(Part 2) 3.5 4
(Part 3) 5.15 5
To find a feasible region, the diameter of the tilt/telescopic screw is increased

by 50%. Actually, the diameter is not included in the design variable set of Table
9.19. When optimization is performed, the first natural frequency is increased to
30 Hz as shown in Table 9.20. It is noted that the weight is increased to satisfy the
constraint in Equation 9.12c. A steering system is manufactured based on the
optimization results, and the natural frequency is measured as 31 Hz from the
experiment. The optimum design variables are shown in Table 9.21. The initial
and optimum designs are compared in Figure 9.28 for harmonic response analysis.
It is noted that the peaks are considerably reduced although they are not included
in the optimization process.
9.6.5 Summary
Axiomatic design is applied to the conceptual design process of a motor driven

tilt/telescopic steering system. Funtional requirements and design parameters are
defined and the hierarchy of the design process is established. The design is found
to be decoupled and the detailed design is carried out based on the sequence
indicated by the axiomatic design. Orthogonal arrays are utilized for the design
that considers crash performance. Structural optimization is employed to improve
the vibration performance.
9.7 Crash Analysis and Design of a Belt Integrated Seat

for Occupant Safety
The structure of a belt integrated seat (BIS) is shown in Figure 9.29. The
conventional seat is typically composed of the seat back, the headrest, the seat
cushion and the recliner. Traditionally, the seat back is a one piece circular frame
and is connected to the seat cushion frame by connecting screws. The recliner on
the left side of the seat cushion frame allows the seat back to rotate back and forth.
(a) Front view (b) Side view
Figure 9.29. Belt integrated seat

In a BIS, the seat belt is attached to the seat, while in a general passenger car it
is attached to the vehicle body. In a crash, the seat belt is located close to the
driver regardless of the occupant size and the injury level is lowered. On the other
hand, a large load is imposed on the seat by the seat belt. Figure 9.29 shows a belt
integrated seat. A belt integrated seat is composed of seven parts: the seat back
frame, the seat back plastic injection mold, the headrest, the headrest adjustment,
the recliner, the seat cushion and the seat sliding frame. The BIS has a different
structure in the seat back frame compared to the conventional seat of a general
passenger car.
The seat back frame has two hollowed rectangular frames on both sides. The
headrest frame, which supports the headrest, is inserted in the right hollowed
rectangular seat back frame allowing for easy height adjustment. Both seat back
frames are connected by a plastic injection mold to support the lower back of the
Table 9.22. Specifications of a conventional seat and a belt integrated seat
Conventional seat Belt integrated seat

Height
690 840
(mm)
Width
480 510
(mm)
Length
510 735
(mm)
Thickness
45 54
(mm)
Seat back structure
Seat back structure 2.7 (including headrest and 9.1
height adjuster)
Seat cushion structure 9.6 Seat cushion structure 18.4
Seat back cover 1.5 Seat back cover 3.8
Weight Seat cushion cover 1.6 Seat cushion cover 5.4

(kgf)
Headrest 0.7 Driving motor 2.9
Motor controller 2.7
Weight without motor part 36.7
Total weight 16.1 Total weight 42.3

occupant. The seat cushion is constructed with a thicker panel than the
conventional one and is connected to the seat sliding frame. The lateral movement
of the seat is accomplished with electric motors located on either side of the seat
sliding frame. The back and forth rotation of the seat back is accomplished with
another set of electric motors driving the screw type gears, which provide
additional structural strength.
The specifications of both a conventional seat and a BIS are described in Table
9.22. The numbers in the table are taken from an actual product. The
specifications of height, width and thickness are based on the skeleton frame. The
weight of the BIS is approximately three times greater than that of the
conventional seat. Therefore, a method to reduce the weight of the seat should be
considered during the design phase.
9.7.2 Simulation Modelling
According to FMVSS208, the sled test with an existing BIS is performed by the
manufacturers (FMVSS 2003). In the sled test, the Hybrid III 50% dummy is used
and the crash velocity is 27.8 mph (44.7 km/h) (Backaitis and Mertz 1994). The
deceleration pulse of the sled from the sled test is illustrated in Figure 9.30.
The multibody dynamic simulation program SAFE is used to model the sled
and the seat (SAFE 1999). In order to model the accurate behavior of the seat
during the crash, a segment element that has the mass and the moment of inertia as
illustrated in Figure 9.31 is used instead of the simple panel element. The panel
element has been used for the seat. This may be sufficient to evaluate the overall
behavior. The panel element can make a resistant force only on the orthogonal
direction of the panel. Thus, it is not accurate for seat behavior. The segment
element can generate resistant forces in any direction. Hyper-ellipsoid segments
are employed to describe the geometry of the seat more accurately. Figure 9.31
35
30
25
(G)
Deceleration(G)
20
Deceleration
15
10
0
0 10 20 30 40 50 60 70 80
-5
Time((ms)
Time ms )
Figure 9.30. The deceleration curve from a frontal crash

Segment
Hyper-ellipsoid
(a) Seat back (b) Seat cushion
Figure 9.31. The modelling of a seat back and a seat cushion
shows a seat modelled by hyper-ellipsoids. The seat model is composed of two

parts: the seat back and the seat cushion. The seat back consists of nine segments.
Segments are connected to each other by pin joints and spring–damper systems.
The headrest is supported by the seat back through a pin joint. The seat cushion
has three segments attached by pin joints and is connected to the car floor via
sliding joints. The seat back and cushion are connected by pin joints.
9.7.3 Simulation of Frontal and Rear Impacts
The dummy and the seat are modelled by segment elements that have the mass and
ͩ͡͡
΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ͽΖΗΥ͚
Shoulder belt (left) ͩ͡͡ ΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ͽΖΗΥ͚
Shoulder belt (left)
ͨ͡͡ ΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙΃ΚΘΙΥ͚
Shoulder belt (right) ΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙΃ΚΘΙΥ͚
Shoulder belt (right)
ͨ͡͡
ͧ͡͡ ͽΒΡ͑ͳΖΝΥ͙ͽΖΗΥ͚
Lap belt (left)
ͧ͡͡ ͽΒΡ͑ͳΖΝΥ͙ͽΖΗΥ͚
Lap belt (left)
ͽΒΡ͑ͳΖΝΥ͙΃ΚΘΙΥ͚
Lap belt (right)
(kgf)
(kgf)
ͷΠΣΔΖ͙ΜΘ
ͦ͡͡ ͽΒΡ͑ͳΖΝΥ͙΃ΚΘΙΥ͚
Lap belt (right)
ͦ͡͡
ͷΠΣΔΖ͙ΜΘΗ͚
ͥ͡͡
Force
Force
ͥ͡͡
ͤ͡͡
ͤ͡͡
ͣ͡͡
ͣ͡͡
͢͡͡
͢͡͡
͡
͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪ͡ ͢͡͡ ͢͢͡ ͣ͢͡ ͤ͢͡ ͥ͢͡ ͦ͢͡ ͧ͢͡ ͡
͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪ͡ ͢͡͡ ͢͢͡ ͣ͢͡ ͤ͢͡ ͥ͢͡ ͦ͢͡ ͧ͢͡
΅ΚΞΖ͙͑ΞΤ͚͑
Time (ms) ΅ΚΞΖ͙͑ΞΤ͚͑
Time (ms)
(a) Conventional seat (b) BIS
Figure 9.32. Comparisons of belt force curves on anchor points

Table 9.23. Maximum belt forces and times
Left shoulder Right shoulder Left lap Right lap
Conventional Force (kgf) 695.2 226.48 735.06 746.95

seat Time (ms) 62 66 63 63
Belt Force (kgf) 541.13 375.32 702.76 680.27
integrated
seat Time (ms) 90 53 53 53
the moment of inertia while the car body is modelled by simple panel elements.
Force–displacement functions are used for both elements to describe the
deformation characteristics in a crash environment. A crash pulse of the car is
obtained from a crash test and is used in the simulation. After validating the
dummy behavior with high speed photos from the sled test, the simulation model is
established.
The force curves on the anchor points for both the conventional seat and the
BIS are illustrated in Figure 9.32. The maximum forces with respect to times are
described in Table 9.23. The times to reach the maximum forces on each anchor
point are similar to those of the conventional seat. However, the time to reach the
maximum force on the left anchor point of the shoulder belt is slower by
approximately 30 ms, compared to other anchor points. This delay occurs because
the left anchor point of the shoulder belt follows the movement of the seat back
during the crash. Although the summation of forces from both anchor points is
similar in the conventional seat and the BIS, the forces are more evenly distributed
to both anchor points in the case of the BIS. Therefore, the energy is scattered in
the body of the dummy.
The acceleration curves on the head, the chest and the pelvis of the dummy for
both the conventional seat and the BIS are illustrated in Figure 9.33. The
ͦ͡
ͦ͡ ͹ΖΒΕ͑ͲΔΔ͟
Head acc. ͹ΖΒΕ͑ͲΔΔ͟
Head acc.
ͥͦ ͥͦ
ʹΙΖΤΥ͑ͲΔΔ͟
Chest acc. ʹΙΖΤΥ͑ͲΔΔ͟
Chest acc.
ͲΔΔΖΝΖΣΒΥΚΠΟ͙͑͸͚͑(G)
ͥ͡ ͥ͡
(G)
ͲΔΔΖΝΖΣΒΥΚΠΟ͙͑͸͚͑
ͤͦ ΁ΖΝΧΚΤ͑ͲΔΔ͟
Pelvis acc. ͤͦ ΁ΖΝΧΚΤ͑ͲΔΔ͟
Pelvis acc.
Acceleration
ͤ͡
Acceleration
ͤ͡
ͣͦ ͣͦ
ͣ͡ ͣ͡
ͦ͢ ͦ͢
͢͡ ͢͡
ͦ ͦ
͡ ͡
͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪ͡ ͣͤͥͦͧ͢͢͢͢͢͢͢͢͡͡͡͡͡͡͡͡ ͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪͣͤͥͦͧ͢͢͢͢͢͢͢͢͡͡͡͡͡͡͡͡͡
Time (ms) ΅ΚΞΖ͙͑Ξ(ms)
Time Τ͚͑
Figure 9.33. Comparison of accelerations

Table 9.24. Maximum accelerations and times
Head Chest Pelvis
Acceleration (G) 34.640 36.722 41.835

Conventional seat
Time (ms) 72 55 58
Acceleration (G) 37.075 27.818 46.811

Belt integrated seat
Time (ms) 94 54 53
maximum forces with respect to the times are described in Table 9.24. The
maximum head acceleration is similar to that of the conventional seat. However,
the time to reach the maximum is delayed by approximately 20 ms due to the
movement of the left anchor point of the shoulder belt. As a result, the maximum
chest acceleration is reduced by approximately 24% compared to that of the
conventional seat. In addition, the times of maximum accelerations for the head
and chest are more evenly distributed. Since the distance between the occupant's
chest and the steering wheel or the instrument panel is narrow, the reduction of the
maximum chest acceleration is quite important.
Figure 9.34 shows the relation between the forces and accelerations for both
the conventional seat and the BIS. The head acceleration and the force on the left
anchor points of the shoulder belt are similar. The chest acceleration and the force
on the right anchor point of the shoulder belt curves are also similar. These
similarities show that the seat belt plays an important role in the safety of the
occupant head and chest. It is important to select the proper type of the seat belt
for the BIS. An additional safety system, such as an airbag, would be more
efficient in protecting the occupants from potential injuries.
The seat belt is categorized into two types: (1) the emergency locking retractor
(ELR), which controls the web length in the belt spool only by the retractor, and
(2) the web locking retractor (WLR), which controls the web length in the belt
ͥ͡ ͨ͡͡
ͥ͡ ͹ΖΒΕ͑ͲΔΔ͟
Head ͩ͡͡
Chest acc.
ʹΙΖΤΥ͑ͲΔΔ͟ ͤͦ ͧ͡͡
ͤͦ ͨ͡͡
Shoulder belt (left)
΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ͽΖΗΥ͚
ͲΔΔΖΝΖΣΒΥΚΠΟ͙͑͸͚͑(G)
ͤ͡
ͤ͡ ͧ͡͡ ͦ͡͡
(G)
΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙΃ΚΘΙΥ͚
ͲΔΔΖΝΖΣΒΥΚΠΟ͙͑͸͚͑
Acceleration
ͣͦ ͹ΖΒΕ͑ͲΔΔ͟
Head
ͷΠΣΔΖ͙͑ΜΘΗ͚͑
ͣͦ ͦ͡͡
(kgf)
ͷΠΣΔΖ͙͑ΜΘΗ͚͑
(kgf)
ͥ͡͡
Acceleration
Chest acc.
ʹΙΖΤΥ͑ͲΔΔ͟
ͣ͡
ͣ͡ ͥ͡͡ Shoulder belt (left)
΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ͽΖΗΥ͚ ͤ͡͡
Force
Force
ͦ͢ ͤ͡͡ ͦ͢ ΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙΃ΚΘΙΥ͚
ͣ͡͡
͢͡ ͣ͡͡ ͢͡
ͦ ͢͡͡
ͦ ͢͡͡
͡ ͡ ͡ ͡
͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪ͡ ͣͤͥͦͧ͢͢͢͢͢͢͢͢͡͡͡͡͡͡͡͡ ͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪ͡ ͢͡͡ ͣͤͥͦ͢͢͢͢͢͢͡͡͡͡͡ ͧ͢͡
Time (ms) ΅ΚΞΖ͙͑ΞΤ͚͑
Time (ms)
Figure 9.34. Comparison of accelerations and shoulder belt forces

spool by the retractor and the lift bar. The modelling of a seat belt is dependent on
the web length in the belt spool: the slack of ELR between the chest and the belt is
initially 0 mm and the slack of WLR is initially í20 mm. The seat belt stiffness is
also categorized into two types: (1) low stiffness, and (2) high stiffness. The belt
stiffness is defined by the force–displacement function curve.
The types of seat belts are also divided into three types based on the slack
length: (1) í40 mm, (2) 0 mm and (3) 40 mm. The case of í40 mm slack length
has a pretensioner. A pretensioner reduces the slack at the initial stage of impact
by winding the seat belt. The fuse type belt is also considered and categorized.
The fuse type belt is constructed by sewing together certain lengths of the belt web.
As the certain stitched lengths are torn, the absorption of the load is controlled
during the crash. Four kinds of stitched length are considered. The fuse type belts
are defined by the force–displacement function curves as illustrated in Figure 9.35.
The severity of occupant injury for both the conventional seat and the BIS is
defined by the head injury criteria (HIC) and the chest severity index (CSI) for the
various seat belts. The simulations are performed for the different seat belts and
the results for occupant injuries such as HIC and CSI are shown in Tables 9.25 and
9.26. The results show a reduced injury to the head with 0 mm belt slack length
with the BIS compared to the results of the minimum (í40 mm) belt slack length.
Because the belt slack in the BIS is significantly less than that of the conventional
seat during the initial stage of the crash, the structure of the BIS is expected to
function similarly to that of the conventional seat belt with a pretensioner.
Reduced injury to the chest with minimum belt slack length occurs with the BIS.
As the stitched length in the fuse type seat belt is increased, HIC and CSI are
reduced in both cases for the conventional seat and the BIS. Therefore, it is
recommended that the travel distance of the dummy be maximized within the limit
1130 1130
Force (kgf)
Force (kgf)
272.35 272.35
0 2.2 4.3 11.0 0 1.2 2.4 6.2

Displacement (mm) Displacement (mm)
(a) Belt with low stiffness (b) Belt with high stiffness
Figure 9.35. Modified force–displacement curves of a fuse type belt

Table 9.25. ELR and WLR types of the BIS
Slack Stitched ELR WLR

Stiffness
(mm) length (mm) HIC CSI HIC CSI
0 191.87 119.26 199.50 119.19
44.5 133.27 95.66 143.11 95.86

í40
88.9 86.71 87.72 91.83 86.37
133.4 78.25 84.49 76.36 81.60
0 179.88 123.31 185.05 120.68
44.5 121.21 101.76 126.04 97.53

Low 0
88.9 82.27 95.51 84.22 91.00
133.4 80.72 93.86 80.36 88.37
0 176.12 131.25 177.01 127.07
44.5 121.01 110.80 119.83 106.50

40
88.9 84.46 102.71 83.22 99.61
133.4 83.40 100.78 82.08 97.95
0 229.88 134.05 233.93 143.42
44.5 209.98 110.36 216.99 120.80

í40
88.9 159.07 99.47 172.78 107.16
133.4 118.30 92.97 135.82 100.51
0 215.78 129.17 222.81 128.82
44.5 187.01 107.13 198.58 106.59

High 0
88.9 131.37 98.27 144.33 97.23
133.4 88.08 92.80 100.75 90.89
0 198.73 139.34 207.10 133.94
44.5 174.70 114.61 178.82 109.78

40
88.9 119.89 106.38 122.04 101.62
133.4 84.20 101.92 82.74 96.99
to avoid contact with the interior of the automobile. Lower injury of the occupant
occurs in the case of the low stiffness seat belt. In the case of the BIS, the ELR,
low stiffness and 0 mm slack are excellent. The WLR, low stiffness and í40 mm
slack are superior for the conventional seat.
In a rear crash, a stationary vehicle is hit by a moving vehicle from behind.
Injuries to the occupant’s spine and neck typically occur. To minimize these
injuries, the seat back frame should be stiff and easily collapse as it absorbs the
crash energy to reduce the applied load on the occupant.
A deceleration pulse curve as illustrated in Figure 9.36 is obtained from a test
and is used to establish the simulation model of a rear crash with the BIS. The test
Table 9.26. ELR and WLR of the conventional seat
Slack Stitched ELR WLR

Stiffness
(mm) length (mm) HIC CSI HIC CSI
0 151.12 187.03 146.25 179.77
44.5 156.71 167.88 154.38 162.50
í40
88.9 143.18 130.02 144.11 127.81
133.4 117.81 98.78 120.86 98.50
0 156.90 205.45 152.11 198.99
44.5 161.19 181.22 157.22 174.44
Low 0
88.9 147.34 139.08 142.78 133.86
133.4 117.58 102.41 117.40 100.09
0 160.51 211.05 157.74 209.57
88.9 163.37 183.92 161.75 183.67
40
88.9 146.40 138.95 147.17 139.65
133.4 111.52 100.67 113.70 102.31
0 145.49 203.30 138.07 189.07
44.5 154.67 203.17 149.53 191.56
í40
88.9 161.46 189.42 158.69 180.57
133.4 162.87 165.62 160.09 161.16
0 166.29 244.61 156.45 222.10
88.9 171.99 239.46 161.77 219.81
High 0
88.9 173.93 221.10 165.86 203.76
133.4 170.37 192.95 163.70 177.34
0 169.29 263.19 170.74 257.27
44.5 173.35 254.07 175.57 249.49
40
88.9 179.59 232.88 175.52 228.89
133.4 168.31 201.23 170.55 198.88
was carried out by the BIS manufacturer. The test impact velocity is 18.4 mph
(29.6 km/h). Test regulations are not yet defined for BIS. The utilized impact
velocity is arbitrarily defined by the company for the preliminary design of the
product. The anchor points of the seat belt and stiffness are the same as the ones
used in the case of the frontal crash analysis. Figure 9.37 illustrates the behavior
of the dummy in the rear crash for both the conventional seat and the BIS. The
movement of the dummy in the conventional seat and the conventional seat itself
are similar to that of the BIS during the early stage of the collision. However, in
the conventional seat, the head displacement is greater than that of the chest during
80í160 ms. At approximately 160 ms, the chest is pushed forward as a result of
the elastic rebound of the seat back while the head continues to travel backward.
This causes the relative velocities between the head and the chest to increase,
thereby increasing the occupant’s chances of neck injury. However, in the BIS,
12
10
Deceleration( G )(G)
8
Deceleration
0
0 20 40 60 80 100 120 140 160
-2
Time
Time((ms)
ms )
Figure 9.36. The deceleration curve of a rear crash
0 ms 80 ms 160 ms
(a) Conventional seat
0 ms 80 ms 160 ms
(b) Belt integrated seat
Figure 9.37. Comparison of dummy behaviors in a rear end impact
the seat belt continuously restrains the dummy to the seat back and absorbs the
crash energy from the initial stage of the collision. The relative velocity between
the head and the chest is reduced due to this restraint. Thus, neck injury is
expected to be reduced.
Generally, the injury caused by a rear crash is influenced by the torque that
occurs in the neck while the upper body of the dummy is rotated backward during
the collision. The torque that occurs in the neck of the dummy during the
simulation is illustrated in Figure 9.38. The torques are similar from 0í80 ms in
both the conventional seat and the BIS. In the case of the conventional seat, the
torque occurs after 120 ms and reaches its maximum of approximately 310 kgfcm
Figure 9.38. Comparison of torque values of necks
at around 170 ms. In the case of the BIS, the torque is maintained from 80 ms and
reaches its maximum of 70 kgfcm at around 120 ms. Unlike the conventional seat,
the BIS restrains the dummy with the seat belt throughout the entire region during
rear crash. Thus, it has an important role in the safety of the occupant.
9.7.4 Structural Optimization of the BIS
The finite element model to obtain the preliminary structural design trend is
established based on the conventional seat as illustrated in Figure 9.39 with the
commercial structural optimization program GENESIS (GENESIS 2004). The
z
y
Figure 9.39. The BIS model for structural optimization

frame of the conventional seat is used as the initial model of structural

optimization. The existing BIS as illustrated in Figure 9.29 is heavy and
complicated due to the additional parts, motors and etc. Components used in the
conventional seat structure are considered to be inadequate to support the load
from the seat belt during a crash. A reinforced structure as shown in Figure 9.39 is
defined to compensate for the inadequacy. The reinforced structure is developed
by applying the results of the joint’s load from the dynamic simulation and
considering the shape of the existing BIS product. The reinforced components are
as follows:
(1) An extended member from the left seat back frame to attach the left anchor
point of the shoulder belt.
(2) A crossed member between the left and the right seat back frames.
(3) Coil and linear spring elements attached on the recliner of the right seat
back frame.
The type of elements and the properties of materials used are shown in Tables
9.27 and 9.28, respectively. All of the mounting brackets of the seat are fixed as
the boundary condition. During an actual collision, plastic deformation of the seat
occurs in a dynamic environment. Nonlinear dynamic analysis should be included
in the optimization process. However, optimization with nonlinear analysis has
not yet been developed for general purposes. This is because the sensitivity
analysis and optimization process are extremely complicated and costly. Moreover,
it is quite difficult to optimize a structure based on dynamic analysis. The
difficulty is that dynamic behavior is defined in a time domain. Therefore,
mathematical optimization with those analyses is almost impossible for a large-
scale structure.
Optimization is carried out for a BIS. That is, optimization with static loads is
performed. Design specification is not well defined for the BIS design. For
structural design, the loading conditions and the allowable design region are not
yet well established. The design environment is defined based on the dynamic
simulation, the experience and specifications of the conventional seat design.
Therefore, the final design of this research only shows the design trend for the
preliminary design, not for the final design.
The points where the load is imposed are the left shoulder belt anchor point,
Table 9.27. Element types and number used for the BIS model
Type Quad. shell Triang. shell Rigid Spring Total

Number 235 1922 52 3 2212
Table 9.28. Properties used for the BIS model
Young’s modulus Density Poisson’s ratio

8 2 3 3 0.29
2.068 u 10 kN / m 7.82 u 10 kg / m
Table 9.29. Loading conditions
Crash Anchor point Ave. belt load (kN) Max. belt load (kN)
direction location
x y z x y z
Left of shoulder belt 2.5 í0.9 í0.5 5.1 í1.9 í1.0
Right of shoulder belt 1.1 0.4 1.1 2.5 0.9 2.7

Front
Left of lap belt 1.2 í0.4 2.2 3.4 í1.2 6.1
Right of lap belt 1.2 0.4 2.3 3.2 1.2 5.9
Left of shoulder belt 2.4 1.1 í0.5 4.9 2.3 í1.0
45q Right of shoulder belt 0.5 1.0 1.1 1.1 2.5 2.7
left Left of lap belt 1.2 0.6 2.3 3.3 1.5 6.1
Right of lap belt 0.6 1.2 2.3 1.4 3.1 5.9
Left of shoulder belt 1.1 í2.4 í0.5 2.3 í4.9 í1.0
45q Right of shoulder belt 1.0 í0.5 1.1 2.5 í1.1 2.7
right Left of lap belt 0.6 í1.2 2.3 1.5 í3.2 6.1
Right of lap belt 1.1 í0.6 2.3 3.1 í1.4 5.9
the right shoulder belt anchor point, the left lap belt anchor point and the right lap
belt anchor point. To consider the oblique crash of 45 degrees, the load of the
frontal crash is divided into the components in 45 degrees. Table 9.29 shows the
components of the average and maximum values. The displacement and stress
constraints for optimization are limited to 45 mm and 4.5 u 10 5 kN/m 2 .
Structural optimization as a preliminary study is performed for the four cases as
follows:
Case 1. Average belt loads for frontal impact.
Case 2. Average belt loads for multidirection impact.
Case 3. Maximum belt loads for frontal impact.
Case 4. Maximum belt loads for multidirection impact.
The optimization results are shown in Table 9.30. The design variables of the
BIS are slightly scaled to protect the manufacturer’s proprietary security and due
to the preliminary development stage of the product. However, the optimum
values are utilized to obtain the overall design trend, so the result values in Table
9.30 are sufficient. An extended member from the left seat back frame is
functionally linked with the left seat back frame using the design variable linking
Table 9.30. Results from structural optimization of the BIS
Design variable
Initial Case 1 Case 2 Case 3 Case 4
(thickness /spring coeff.)
1 Left seat frame 2.52 5.62 7.74 14.2 14.2
2 Right seat frame 2.52 1.4 1.4 7.64 7.93

Lower and upper seat
3 2.52 1.4 1.4 2.98 5.77
frame
Crossed member between
4 1.4 1.4 1.4 1.4 4.87
seat back frames
5 Left/upper recliner 7.72 18.45 17.64 23.64 23.07
6 Left/lower recliner 8.96 20.12 19.68 38.05 35.95
7 Right/upper recliner 5.6 2.18 8.89 28.88 17.21
8 Right/lower recliner 5.6 4.58 3.71 14.94 9.52
9 Left coil spring coeff. 140000 184800 139860 122920 54600
10 Left liner spring coeff. 140000 94600 100000 59080 210000
11 Right coil spring coeff. 0.14 0.39 0.43 1.59 2.35
12 Right liner spring coeff. 0.14 1.58 3.72 2436 1397

13 Seat cushion frame 1.4 1.4 1.4 1.4 2.56
14 Seat mounting bracket 3.5 3.5 5.81 6.10 11.98
15 Seat rail 3.5 2.8 2.8 2.8 2.8
Objective function (kg) 9.704 10.75 11.35 16.44 20.33
No. of iterations 7 11 22 16
method. All the mounting brackets for the seat are linked into one design variable.
The results of the structural optimization are summarized as follows:
(1) The thickness of the left seat frame is increased more than the others to
bear the force from the left shoulder belt.
(2) The thickness of the right recliner and the crossed member between the left
and right seat back frames are slightly increased in the case of a frontal
crash. However, these are increased remarkably in the case of a
multidirection crash. The right recliner and crossed member are expected
to be important in a practical situation such as the oblique crash.
(3) The spring coefficients of the left recliner are significantly increased. A
particular energy absorbing device would be useful for a practical design.
(4) The thickness of mounting brackets is increased under the maximum load
cases. It is insignificantly influenced by the direction of the crash because
the brackets are located on the car floor.
(5) The seat rail and the seat cushion frame have lower bounds. Therefore, if
the seat is appropriately attached to the bottom, a new design can be made
without increasing the thickness.
(6) The weight of the BIS is increased from 9.7 kg to 20.33 kg. This is
because the initial design is made from an existing conventional design.
9.7.5 Summary
(1) The structural configuration of the existing belt integrated seat (BIS) is
analyzed and the simulation models for the frontal and rear end collisions
are established through the occupant analysis program SAFE. The hyper-
ellipsoid element is used for the seat to represent the geometry and forces
more accurately.
(2) The movement of the dummy’s upper body occurs simultaneously with
that of the seat back of the BIS during both frontal and rear end crashes.
The seat belt slack is significantly small. The maximum values of the seat
belt force and the dummy accelerations with respect to times are
distributed. The right shoulder load is increased compared to that of a
conventional seat. This is because the left and right loads are evenly
distributed. The pelvis acceleration is increased. However, the chest
acceleration, which is more important, is considerably decreased.
Therefore, the safety performance of a BIS is generally excellent. There
are not any good performance indices for the BIS. In the future, research is
required for this aspect.
(3) The parametric study is performed on various seat belts to obtain the
proper type for the BIS. From the numerical results, the belt type with ELR,
high elongation and 0 mm slack is proposed.
(4) The finite element model of the BIS is established by the commercial
structural analysis and optimization software GENESIS. Structural
optimization is performed considering all of the seat frames and brackets as
the design variables including multidirection collision with the oblique
crash. The thicknesses of various frames including the left seat and the
recliner are significantly increased compared to those of the conventional
seat.
(5) Design specification is not well defined for the design of BIS. The design
environment is defined for this research. Therefore, the proposed design
example can only be used for a preliminary design of the BIS. The
industrial sponsor of this research is making a proto type product according
to the results of this research. In the future, establishment of design
specifications is required for the design of belt integrated seats.
9.8 Determination of a Crash Pulse and Optimization

of Crash Components Using Response Surface
Approximate Optimization

An automobile is sequentially designed to improve safety. First, the body structure
is designed to satisfy various conditions including crashworthiness performance.
Once the body structure is determined, a crash pulse is determined from a real test
or nonlinear finite element analysis of crashworthiness. Second, occupant analysis
is performed for occupant safety. The crash pulse is utilized as input for the
occupant analysis. Finally, crash components are designed based on the results of
the occupant analysis. The set of crash components includes seat belts, airbags,
seats and interior panels. Therefore, the crash component design is passively
carried out and the designer does not have much freedom due to the pre-
determined crash pulse. However, if an appropriate crash pulse is proposed
beforehand, a better crash component can be designed. Once the pulse has been
determined, the car body structure can be designed to achieve it, and crash
components can be designed to minimize potential injury to the occupants. In this
section, we describe how a good crash pulse can be estimated and discuss the
optimal design of cabin crash components.
The deceleration of a car during a crash determines the movement of occupants
and injuries are induced by the contact between an occupant and the interior parts.
An optimization problem is formulated with the shapes of the deceleration as
design variables and the injury as the objective function. After the deceleration
curve is optimized, the safety devices are optimized based on the curve.
As mentioned earlier, optimization of the safety devices mainly uses a DOE
with orthogonal arrays. It was mentioned that the gradient-based method is not
appropriate. The response surface method is also usable in that it uses only
function values without gradient information. A response surface method in
VisualDOC (VisualDOC 2004) is utilized here and interfaced with an external
program. SAFE (SAFE 1999) is used for occupant analysis.
9.8.2 Response Surface Approximate Optimization

An algorithm, which is modified from the response surface method, is used for the
optimization. It is included in VisualDOC and is called the response surface
approximate optimization (RSAO) or the second-order approximation method (Lee
et al. 1993). The response surface method approximates a function by a quadratic
function based on the information at the candidate points. It is explained in
Section 6.7.
The RSAO algorithm is explained here. From a few candidate points, the
objective function and constraints are approximated. The least squares method in
Guess initial designs
Approximate
objective function and constraints
Const. ĺ Linear ĺ Quadratic
Find optimum using

approximated functions Update designs
Satisfy No
termination criterion?
Yes
Stop
Figure 9.40. Flow of RSAO
Section 6.7 is used for the approximation process. At the beginning, only a few
coefficients are obtained. For example, only the coefficients for linear terms can
be calculated. Optimization is performed with the approximated functions and the
optimum becomes a new candidate point. In other words, the number of candidate
points increases by one.
The process proceeds in an iterative manner until the new candidate point is
unchanged. As the iteration proceeds, constant, linear and quadratic terms are
sequentially produced for the approximated functions. This process is represented
in Figure 9.40. The interface between SAFE and VisualDOC is depicted in Figure
9.41.
VisualDOC Third party program
GUI Input
Input (Graphic User Interface)
Define a design problem
Interface routine Analysis

Optimizer program
Control third party
program
Post-processor
Output
Figure 9.41. Interface between VisualDOC and a third party program

9.8.3 Determination of a Crash Pulse to Minimize Occupant

Injury
Currently, the design for safety is conducted as illustrated in Figure 9.42a. It is ass
umed that the simulation model is verified by real tests. Crashworthiness analysis
using the nonlinear finite element method is carried out to verify energy absorbing
characteristics and the occupant survival space in the development process. A cras
h pulse is extracted from the analysis or a real impact test. This crash pulse is one
of the input data for occupant analysis. Crash components can be designed using t
he occupant analysis results with this crash pulse. In Figure 9.42a, the crash pulse
is predetermined for the design of the crash components. Therefore, this method h
as limitations in that the design does not have much freedom.
A new method is proposed as illustrated in Figure 9.42b. First, an excellent
crash curve is determined based on the validity of physics. The curve is obtained
through an optimization process. The crash components are designed by using an
occupant analysis that uses the generated crash pulse. Optimization is also used in
this process. It is proposed that the automobile body structure be designed to
generate the crash pulse.
Design car body Find crash pulse

and layout that minimizes
occupant injury
Crashworthiness
analysis
or Crash pulse
test
Crashworthiness analysis
Crash pulse Design car body and layout
Occupant analysis Occupant analysis

Design Design
crash components crash components
to minimize to minimize
occupant injury occupant injury
(a) Conventional method (b) Proposed method
Figure 9.42. Conventional and proposed methods

9.8.4 Optimization Formulation for the Determination of a Crash

Pulse
An optimization problem is formulated to determine an excellent crash pulse that
minimizes the injuries of the occupants. The form of the crash pulse is selected as
a design variable. Some points on the curve are selected and linear interpolations
are used to represent the crash pulse. This is illustrated in Figure 9.43. Some
important points on the curve are selected and are used as design variables in the
optimization process.
In optimization, some physical phenomena in the crash must be kept valid.
That is, an automobile must satisfy Newton’s second law in a crash environment.
Using Newton’s law, the motion of the automobile is expressed as:
dv
Fdt d(mv) m dt (9.13)
dt
where F is the external force, m is the mass, t is the time, and v is the velocity.
Because the mass is unchanged during a crash, the momentum of the body is as
follows:
dv dv
³ F dt ³m dt m³ dt m ³ a dt (9.14)
dt dt
Real pulse
1.2 Fitted pulse
1.0 4 Design point

6
0.8
Deceleration (G/G)
0.6 5
2
0.4
3
0.2
7
1 8
0.0
0.0 0.2 0.4 0.6 0.8 1.0
–0.2
Time (ms/ms)
Figure 9.43. Definition of design variables by linear interpolation

dv
where a.
dt
Suppose there are two different automobile body structures, model 1 and model
2. These represent two different body designs. In the same crash situation, each
model has to satisfy the law of conservation of momentum as follows:
T T
For model 1 ³ F model 1dt m model 1 ³ a model 1dt (9.15a)
0 0
T T
model 2
For model 2 ³F dt m model 2 ³ a model 2 dt (9.15b)
0 0
where T is the length of the time range.

It is assumed that the two designs have the same amount of mass. The
properties in Equation 9.15 must be identical because the crash situations for the
two designs are identical. Therefore, the following equation must be satisfied:
T T
model 1 model 2
³a dt ³a dt (9.16)
0 0
Although the design is changed, Equation 9.16 must be physically satisfied in a

crash environment. Equation 9.16 means that the integration of deceleration in the
time domain must be kept constant in each case. That is, the area of the
deceleration curve has to be kept constant. Therefore, this condition should be
considered as a constraint in the optimization process.
The velocity of a car is drastically reduced in a frontal crash. The car is
accelerated forward in a rear end impact. When we determine the deceleration
curve, the shape should be constrained to have these typical behaviors. The
bounds on the design variables should be defined to account for the reasonable
shape.
An optimization problem is formulated here. The objective function is the
injury of the occupant. An appropriate injury index can be used depending on the
characteristics of the impact. The shape of the crash pulse (deceleration curve) is
defined as design variables. Constraints are defined to keep the physics valid. The
optimization problem is formulated as follows:
Find crash pulse (9.17a)

to minimize injury of the occupant (9.17b)
subject to side constraints (9.17c)
conservation of linear momentum (9.17d)
The crash pulse is the input for occupant analysis and is the discrete data for
the deceleration curve of the vehicle body with respect to time. The number of the
data points is 90 to 900. Therefore, all the data points cannot be defined for the
design variable set. The linear interpolation can be utilized as illustrated in Figure
9.43. Then a few data points can be used as design variables. The optimization
problem is formulated as follows:
Find t i , ai , i 1, ..., n (9.18a)
to minimize injury of the occupant (9.18b)

subject to t iL d t i d t iU , aiL d ai aiU (9.18c)
T
³ a dt const. (9.18d)
0
where ti is a point for the linear interpolation, ai is the acceleration at ti , n is the

number of ti and ai , L represents the lower bound and U represents the upper
bound.
9.8.5 Frontal Crash with an Airbag
The deceleration curve for a frontal crash is determined for a car with an airbag.
Eight data points are connected by linear functions as illustrated in Figure 9.43.
The times and accelerations of the eight points are selected as design variables.
The injury indices for the frontal crash are HIC, CSI and the femur load (FMVSS
2003). Here HIC is chosen for the objective function. The optimization
formulation is the same as Equation 9.18 except for replacing Equation 9.18b by
HIC. Table 9.31 shows the lower and upper bounds of the design variables.
RSAO is applied to find the optimum value of the above problem. The
optimum solution is shown in Tables 9.32 and 9.33. The optimum curve is
Table 9.31. Lower and upper bounds on design variables
Time point Deceleration

Point
Lower bound Upper bound Lower bound Upper bound
1 0. 000 0. 000 í0. 003 í0. 003
2 0. 047 0. 170 0. 366 0. 494
3 0. 179 0. 330 0. 258 0. 385
4 0. 340 0. 415 0. 873 1. 000
5 0. 425 0. 566 0. 557 0. 685
6 0. 575 0. 755 0. 732 0. 860
7 0. 858 1. 000 0. 000 0. 112
8 1. 000 1. 000 í0. 048 í0. 048
illustrated in Figure 9.44. HIC is improved by 6.15%. In Figure 9.44, the peaks
are reduced and the distance between the peaks is enlarged.
Table 9.32. Optimum solution for the frontal crash curve
Initial Optimum
Point
Time Deceleration Time Deceleration
1 0.000 −0.003 0.000 −0.003
2 0.071 0.459 0.071 0.527
3 0.143 0.344 0.143 0.332
4 0.286 1.000 0.286 0.932
5 0.357 0.663 0.357 0.595
6 0.500 0.850 0.500 0.782
7 0.714 0.051 0.786 0.043
8 1.000 −0.048 1.000 −0.048
Table 9.33. Initial and optimum values of HIC
HIC Remarks
Initial pulse 1.000 1
Optimum 0.938 −6.15 %
Figure 9.44. Initial and optimum curve shapes

9.8.6 Rear End Impact
The deceleration curve for rear end impact is determined. When a rear end impact
occurs, the occupant body and the seat move forward together. The head is
supported by the body and moves more slowly than the body. Therefore, the
relative rotational movement of the head causes neck injuries. The objective
function is defined by the extension moment of the neck, which has the limit value
in the safety regulation.
The deceleration curve of the rear end impact is presented in Figure 9.45 and
11 data points are selected for linear interpolation. Constraints are defined from
the conservation of momentum. The bounds of the design variables are shown in
Table 9.34. The optimization formulation is the same as in Equation 9.18 with n =
11 except for the objective function in Equation 9.18b. The objective function is
defined as follows:
Table 9.34. Bounds on design variables for rear end impact
Time point Deceleration

Point
Lower bound Upper bound Lower bound Upper bound
1 0.000 0.000 −0.061 −0.061

2 0.001 0.015 −0.700 −0.500
3 0.018 0.030 −13.500 −10.500
4 0.032 0.045 −11.500 −8.500
5 0.050 0.065 −11.500 −8.500
6 0.078 0.088 −7.500 −5.000
7 0.091 0.101 −7.500 −5.000
8 0.108 0.118 −5.500 −3.000
9 0.121 0.129 −5.500 −3.000
10 0.166 0.166 0.594 0.594
Figure 9.45. Definition of design variables for rear end impact

Table 9.35. Initial and optimum values for rear end impact
Extension moment (Nm) Remarks
Initial pulse 13.426 /
Optimum 11.672 −13.07 %
Figure 9.46. Initial and optimum curves for rear end impact
Minimize the extension moment of the neck (9.19)

RSAO is used to solve the optimization problem. The results are shown in
Figure 9.46 and Table 9.35. The objective function is reduced by 13.07% and the
first peak is reduced.
Crash pulses are obtained to reduce the injury indices by using optimization.
The crash pulse is used for definition of the design variables. The injury indices
are used for the objective function and the constraints are defined by the
conservation of momentum. The safety devices will be designed based on the
optimum crash pulses.
9.8.7 Design of Safety Devices for a Frontal Crash
An airbag system is designed using the crash pulse for a frontal crash, which was
determined in the previous section. The design variables are defined as in Section
9.3. The design variables are the mass flow rate, the vent hole area, the firing time
of the inflator and the stiffnesses of the seat belt and the knee bolster. They are
shown in Table 9.36. There are no constraints and the objective function is HIC.
Table 9.36. Design variables for an airbag
Design variable number Description of design variables

1 (A) Mass flow of the airbag
2 (B) Vent hole area of the airbag
3 (C) Firing time of the inflator in the airbag system
4 (D) Stiffness of the seat belt
5 (E) Stiffness of the knee bolster
Table 9.37. Orthogonal array for an airbag design
Column number and assigned factor

Exp.
1 2 3 4 5 6 HIC
no.
1 (A) 2 (B) 3 (C) 4 (D) 5 (E) error
1 1 1 1 1 1 1 0.720
2 1 2 2 2 2 2 0.785
3 1 3 3 3 3 3 0.734
4 1 4 4 4 4 4 0.702
5 1 5 5 5 5 5 0.781
6 2 1 2 3 4 5 0.668
7 2 2 3 4 5 1 0.678
8 2 3 4 5 1 2 0.772
9 2 4 5 1 2 3 0.721
10 2 5 1 2 3 4 0.643
11 3 1 3 5 2 4 0.748
12 3 2 4 1 3 5 0.763
13 3 3 5 2 4 1 0.609
14 3 4 1 3 5 2 0.708
15 3 5 2 4 1 3 0.636
16 4 1 4 2 5 3 0.816
17 4 2 5 3 1 4 0.856
18 4 3 1 4 2 5 0.720
19 4 4 2 5 3 1 0.934
20 4 5 3 1 4 2 0.577
21 5 1 5 4 3 2 1.000
22 5 2 1 5 4 3 0.780
23 5 3 2 1 5 4 0.639
24 5 4 3 2 1 5 0.674
25 5 5 4 3 2 1 0.760
An orthogonal array is used for the design. The L25 (5 6 ) orthogonal array is
selected as shown in Table 9.37. The levels are normalized as in Section 9.3. The
first level is 0.7, the second level is 0.85, the third level is 1.0, which is the current
Level mean
Factor Mean
1 2 3 4 5
1 (A) 0.745 0.696 0.693 0.781 0.771 0.737
2 (B) 0.790 0.772 0.695 0.748 0.679 0.737
3 (C) 0.714 0.732 0.682 0.763 0.793 0.737
4 (D) 0.684 0.705 0.745 0.747 0.803 0.737
5 (E) 0.732 0.747 0.815 0.667 0.724 0.737
value, the fourth level is 1.15 and the fifth level is 1.3. The one-way table is made
as shown in Table 9.38 and the solution is A3 B5 C 3 D1 E 4 . A confirmation
experiment is carried out and HIC 0.69. The 20th row of Table 9.37 is better
with HIC 0.577. Thus, it is considered as the final solution. This phenomenon
appears because the interactions are ignored.
Now an optimization problem is defined. The design variables are the same as
the above DOE. The first and fifth levels are the upper and lower bounds,
respectively. A gradient-based method and RSAO are applied. Figure 9.47
presents the history of the objective function from each method. The values are
normalized by the initial objective functions. The gradient-based method obtains
an inferior solution even though it is mathematically rigorous. The reason is that
the functions in occupant analysis are quite noisy and the local sensitivities
frequently differ from the overall trend.
The results are compared in Table 9.39. All the methods can obtain an
improved solution. The RSAO method obtains the best solution. It is noted that
1.0
0.8
Optimum
0.6
0.4
0.2 Gradient-based optimization

Response surface method
0.0
0 2 4 6 8 10
Iteration number
Figure 9.47. Optimization history of the airbag system

Table 9.39. Comparison of airbag designs by various methods
Design Design variable

HIC
method 1 (A) 2 (B) 3 (C) 4 (D) 5 (E)
Orthogonal
1.150 1.000 1.000 0.700 1.150 1.000
array
Gradient-
1.000 0.868 0.966 1.049 0.900 0.991
based method
RSAO 0.700 0.962 0.966 0.700 1.142 0.687
the solution from the orthogonal array is obtained from a discrete space. Thus, it
may be more reasonable because an available design can be used.
9.8.8 Seat Design in Rear End Impact
A seat is designed based on the previously determined crash pulse. In rear end
impact the neck injuries depend on the safety performance of the seat because the
seat is the only safety device. Rear end impact generally occurs at a low speed.
However, it generates a relative motion between the neck and the body, and the
relative motion causes neck injuries.
Neck injuries result from the sudden extension moment, shear force, tension
and compression force at the occipital condyles. Especially, the limit value of the
extension moment is low and the extension moment is the main cause of neck
Table 9.40. Design variables of a seat in rear end impact
Design variable number Description of design variables

1 (A) Clearance between the occupant’s head and headrest
2 (B) Stiffness of the seatback upholstery
3 (C) Friction coefficient of the seat back
4 (D) Torsional stiffness of the seat back
5 (E) Headrest stiffness
Table 9.41. Design variables and levels for seat design
Design variable
Level
A B C D E
1 5 cm í50% 0. 30 10¶ í50%
2 7 cm 0 0. 45 20¶ 0
3 9 cm 50% 0. 60 30¶ 50%
Table 9.42. Orthogonal array for seat design
Column number and design variable Extension

Exp.
1 2 3 4 5 6 7 8 moment
no.
(Nm)
error A error B C D E error
1 1 1 1 1 1 1 1 1 13.357
2 1 1 2 2 2 2 2 2 15.278
3 1 1 3 3 3 3 3 3 14.034
4 1 2 1 1 2 2 3 3 18.434
5 1 2 2 2 3 3 1 1 15.788
6 1 2 3 3 1 1 2 2 11.829
7 1 3 1 2 1 3 2 3 16.964
8 1 3 2 3 2 1 3 1 11.731
9 1 3 3 1 3 2 1 2 18.218
10 2 1 1 3 3 2 2 1 13.220
11 2 1 2 1 1 3 3 2 20.188
12 2 1 3 2 2 1 1 3 12.054
13 2 2 1 2 3 1 3 2 12.103
14 2 2 2 3 1 2 1 3 14.308
15 2 2 3 1 2 3 2 1 19.767
16 2 3 1 3 2 3 1 2 14.700
17 2 3 2 1 3 1 2 3 12.916
18 2 3 3 2 1 2 3 1 16.043
injuries. Thus, a design should be carried out to decrease the extension moment.
Design variables are shown in Table 9.40. An optimization problem is formulated
here. The objective function is the extension moment and there are no constraints.
An orthogonal array is utilized for the seat design. The design variables and
Design Level mean

Mean
variable 1 2 3
1 (A) 14.689 15.371 15.095 15.052
2 (B) 17.147 14.705 13.304 15.052
3 (C) 15.448 15.327 14.380 15.052
4 (D) 12.332 15.917 16.907 15.052
5 (E) 14.738 14.996 15.422 15.052
1.0
function
0.8
Optimum
0.6
Objective
0.4
0.2 Gradient-based optimization

Response surface method
0.0
0 2 4 6 8 10 12
Iteration number
Iteration number
Figure 9.48. Optimization history for rear end impact
levels are defined in Table 9.41 and the utilized orthogonal array is shown in Table
9.42. The one-way table for Table 9.42 is shown in Table 9.43. The solution is
A1 B3C3 D1 E1 and the objective function is 19.335 from the confirmation
experiment. It is compared with all the rows of the orthogonal array. It is found
that the eighth row of the orthogonal array is the best with 11.731. Thus, the
eighth row is selected as the final solution. The phenomenon occurs because the
interactions are ignored. The optimization results are shown in Figure 9.48 and
Table 9.44. The trend of the results is similar to that of the airbag example.
9.8.9 Summary
The crash pulse for occupant analysis is obtained first and this differs from the
conventional process. An optimization problem is defined to obtain the crash
pulse and it is applied to frontal crash and rear end impact. The safety devices are
Table 9.44. Optimum values for the seat from various methods
Design Design variable Extension

method A B C D E moment (Nm)
Orthogonal
í5.0 1.5 0.300 70.0 1.5 11.731
array
Gradient
í5.0 0.5 0.6 70.0 1.25 11.446
method
RSAO í2.396 0.5 0.6 70.0 0.591 11.250

Table 9.45. Design results according to different crash pulses
Original pulse Optimum pulse
Initial design Optimum Initial design Optimum

for components design for components design
HIC 1.000 0.613 0.938 0.425
from frontal
impact í í38.75% í6.15% í57.53%
Extension 1.000 0.930 0.869 0.838

moment
from rear end í í7.01% í13.07% í16.21%
impact
designed based on the crash curves, and the crash performances are considerably
improved. The overall results are shown in Table 9.45. It is noted that the safety
performances are considerably improved by the optimized crash pulses.
References
ATB User’s Guide (1988) Wright Patterson Air Force Base, Dayton, OH
Backaitis SH, Mertz HJ (eds.) (1994) Hybrid III: The First Human-like Crash Test
Dummy. Society of Automotive Engineers, Warrendale, PA
Cook RD, Malkus PS, Pleska ME (1989) Concepts and Application of Finite
Element Analysis. Wiley, New York
Federal Motor Vehicle Safety Standards (FMVSS) (2003) Code of Federal
Regulations: Title 49–Transportation. Part 571, National Archives and Records
Administration, Washington D.C.
GENESIS User's Manual Version 7.5 (2004) Vanderplaats Research &
Lee WI, Park GJ, Park YS (1993) A Study on the Second-order Approximation
Method for Engineering Optimization. Proceedings of the IFIP WG7.5 Fifth
Working Conference, pp 213í220
Lee JW, Yoon KH, Park GJ (2003) A Study on Occupant Neck Injury in Rear End
Collisions. Proceedings of the Institution of Mechanical Engineers, Part D, Journal
of Automobile Engineering 217(1):23í29
Lim JM, Choi JW, Park GJ (1997) Automobile Side Impact Modeling Using ATB
Software. International Journal of Crashworthiness 2(3):287í297
Lim JM, Wu DH, Park GJ (2000) Analysis and Design Consideration of an Energy
Absorbing Steering System Using Orthogonal Arrays. International Journal of
Crashworthiness 5(3):271í278
LS/DYNA User’s Manual (2003) Livermore Software Technology, CO, USA
MADYMO User’s Manual Version 6.1 (2003) TNO Automotive, The Netherlands
Park YS, Lee JY, Lim JM, Park GJ (1996) Optimum Design of a Steering Column
to Minimize the Injury of a Passenger. International Journal of Vehicle Design
17(4):398í414
Park YS, Park GJ (2001) Crash Analysis and Design of a Belt Integrated Seat for
Occupant Safety. Proceedings of the Institution of Mechanical Engineers, Part D,
Journal of Automobile Engineering 215(8):875í889
SAFE User’s Manual (1999) Hanyang University, Korea
Shin MK, Hong SW, Park GJ (2001) Axiomatic Design of the Motor Driven
Tilt/Telescopic Steering System for Safety and Vibration. Proceedings of the
Institution of Mechanical Engineers, Part D, Journal of Automobile Engineering
215(2):179í187
Shin MK, Park KJ, Park GJ (2003) Occupant Analysis and Seat Design to Reduce
Neck Injury from Rear End Impact. International Journal of Crashworthiness
8(6):573í581
VisualDOC User’s Manual Version 4.0 (2004) Vanderplaats Research &
10 Multidisciplinary Design Optimization
10.1 Introduction
At present, engineering systems are fairly large and complicated. Design
requirements are rigorous and stringent for such systems. Accordingly, design
engineers are seeking new methods and one of them is multidisciplinary design
optimization (MDO). MDO is a design optimization method. Generally,
optimization is applied by considering only one discipline. However, it is fairly
difficult to solve modern engineering problems with a single discipline. Therefore,
we need design methodologies that can cover multidisciplines. Diverse MDO
methods have been proposed to overcome these difficulties.
Structural optimization is a sophisticated method that considers a single
discipline of structural analysis. On the other hand, multiple disciplines are
included in MDO. MDO usually has many design variables, objective functions
and constraints. Analysis can be independently performed for a discipline, and the
disciplines can be linked through analyses. That is, output from a discipline can be
input to other disciplines. This phenomenon is called the coupled process. In
MDO, it is important to efficiently manage the coupled process and to optimize
large-scale problems.
The need for MDO has emerged in the design of airplane wings. Pressures are
imposed on airplane wings by the flow of the surrounding air. The pressures act
on the wings as the lift force and the drag force. The pressures are external forces
on the wing structure and the shape of the wing is changed due to these external
forces. Then the air flow changes and the pressures change accordingly.
Equilibrium is obtained when the coupled process converges. This phenomenon is
investigated in aeroelasticity and computer simulation is adopted to solve the
problem. The pressure distribution is solved by computational fluid dynamics and
the deformation of the wing is analyzed by the finite element method. The
analysis method considering the coupled process is called multidisciplinary
analysis (MDA).
When we design airplane wings, the above two disciplines should be
considered simultaneously. A design problem is separately defined for each
discipline. For example, the design to minimize the drag force is performed by the
output from computational fluid dynamics and the weight of the wing structure is
minimized by using the output of the finite element method. When optimization
technology is employed in this case, it is called MDO. The two disciplines have
common design variables, objective functions and constraints. They can be
independently defined in each discipline.
Coupling is predominant in the analysis and design of airplane wings. The
problem has not yet been thoroughly solved from an MDO viewpoint, so designers
are examining various methods. Sometimes, a discipline of the control theory is
involved in the design of airplane wings and there are many references for the
design of wings. Therefore, the details of airplane wings are not explained in this
chapter. Instead, general methods for MDO are discussed and their application is
investigated (Sobieszczanksi-Sobieski and Haftka 1996).
The optimization formulation for a target problem requires the use of MDO.
Based on the formulation, the overall flow and utilized optimization methods are
determined and can be modified accordingly. Gradient-based methods are usually
used for the optimization process. Recently, some MDO research has adopted
nongradient-based methods such as genetic algorithms, neural networks, simulated
annealing, etc. Meanwhile, some software companies have announced various
MDO frameworks such as ModelCenter (ModelCenter 2004), iSIGHT (iSIGHT
2004), Optimus (Optimus 2003), MIDAS (Team Midas 1995), VisualDOC
(VisualDOC 2004), etc. In this chapter, the formulation of MDO is explained and
the methods to solve it are described.
Overall, there are seven methods for MDO. The formulation of the
multidisciplinary feasible (MDF) method is basic. It is easy to use even though
complicated system analysis should be carried out for each step. The coupled
relationship is solved in the system analysis. The formulation of the individual
discipline feasible (IDF) method has been developed to eliminate the system
analysis, and each discipline is independently solved. In the all-at-once (AAO)
method, even the analysis process is eliminated in the disciplines. None of the
disciplines determine the design in the above methods. Analysis is only carried
out in the disciplines (Cramer et al. 1993).
Various methods have been developed to allocate the design process to
disciplines. They are concurrent subspace optimization (CSSO) (Sobieszczanski-
Sobieski 1988), collaborative optimization (CO) (Braun 1996), bilevel integrated
system synthesis (BLISS) (Sobieszczanski-Sobieski et al. 1998) and
multidisciplinary optimization based on independent subspaces (MDOIS) (Shin
and Park 2005).
In CSSO, sensitivity information is obtained by solving the global sensitivity
equation (GSE), which is a derivative of the system analysis. Each discipline
considers the design conditions of the other disciplines indirectly. The overall
process is controlled by the coordination optimization problem (COP) in the upper
level. BLISS also uses GSE. Each discipline of the lower level has its own local
design variables and the common design variables are determined in the upper
level. MDOIS proposes an MDO method for a problem that has an independent
design problem for each discipline. These three methods conduct the system
analysis at the beginning of each cycle. On the other hand, CO does not have the
Multidisciplinary Design Optimization 563
system analysis. The upper level determines target values for disciplines, and each
discipline solves its own design problem while a given target value is satisfied.
Here, the characteristics of MDO are investigated and the above seven methods
are compared. The methods are constantly evolving and being improved upon. In
this chapter, they are explained based on the initial proposition of the methods.
The methods are derived mathematically. They can still solve small-scale
problems or problems with weak couplings. Approximation methods are
frequently utilized for large-scale problems or problems with strong couplings.
Currently, it seems that no method can solve all the MDO problems. For large-
scale problems, an improved solution is sought with specific approximation
methods instead of a mathematically optimum solution. Therefore, none of the
introduced methods is said to be the ultimate one.
10.2 Multidisciplinary Design Optimization

Design is the process to make the responses of a target object satisfy the designer,
while analysis is the process to evaluate the responses. Recently, the analysis
process has mostly been carried out by computers. In an MDO process, analysis of
a discipline is the basis of a design. Various analyses are required in MDO. The
analyses of the disciplines are related and this phenomenon is called coupling.
Coupling is the relationship between disciplines. It can be direct or indirect. A
change in a discipline may or may not return to the discipline again. Between
independent disciplines, a change in one discipline does not affect the other one.
Coupling is quite important in MDO. Attention is given to the solving process of
the coupling in analysis while coupling in the design process is also considered.
The couplings in analysis or in design can be explained as follows:
Coupling in analysis: When results (responses) of an analysis are input for the
analysis of other disciplines, coupling in analysis exists. We can use the
terminologies of Chapter 2. Coupling in one direction is a decoupled relationship
while coupling in both directions is a coupled relationship. An uncoupled
relationship means that there is no coupling between disciplines. Generally,
coupling in MDO refers to the coupling in analysis.
Coupling in design: When design variables are shared by multiple disciplines, a
coupled relationship exists. An uncoupled relationship means that there are no
common variables.
The relationship between analysis and design: Analysis is the process to find an
equilibrium condition for given design variables. Design is the process where the
analysis results are changed as the designer drives the design process. The process
that uses the analysis results in design is called nested analysis and design (NAND).
Meanwhile, analysis may not be performed and an evaluator is used instead. This
process is called simultaneous analysis and design (SAND) (Balling and
Sobieszczanski-Sobieski 1996).
In this chapter, methods are explained based on the assumption that there are
two disciplines. Although we have more than two disciplines, the derivation is still
valid.
10.2.1 Coupling in Analysis
In MDO, there are multiple disciplines that can have independent analysis and
design procedures. Suppose we have two disciplines. The coupling for analysis
can be expressed as
z1 h1 (b1 , b c , z c2 ) (10.1a)
z2 h 2 (b 2 , b c , z1c ) (10.1b)
where h1 and h 2 are the analyzers for disciplines 1 and 2, respectively. Each
analyzer can have its own design variables and coupling variables. b1 R nd 1 and
b 2 R nd 2 are local design variable vectors for disciplines 1 and 2, respectively.
b c R ndc is the common design variable vector shared by the two disciplines.
The state variable vectors z1 R ns1 and z 2 R ns 2 are the results (responses) from
the two disciplines, respectively.
When we have couplings between disciplines, some or all of the state variables
of a discipline are used as input to the other discipline. Such state variables are
called coupling variable vectors z1c R nc1 and z c2 R nc 2 . The state variables that
are not coupling variables are the noncoupling variables z1nc R nnc1 and
nnc 2 T T
z nc
2 R . The state variable vectors are expressed by z1 [z1nc z1c ]T and
T T
z2 [z nc
2 z c2 ]T .
Equation 10.1 can be expressed as
z1nc h11 (b1 , b c , z c2 ) (10.2a)
z1c h12 (b1 , b c , z c2 ) (10.2b)
z nc
2 h 21 (b 2 , b c , z1c ) (10.2c)
z c2 h 22 (b 2 , b c , z1c ) (10.2d)
Equations 10.1 and 10.2 have the same meaning and they represent analyzers.
If we have couplings in analysis, the coupling variables should be evaluated. A
representative method is the fixed point iteration method. It is difficult to obtain
coupling variables in practice. For example, aeroelasticity is utilized in the wing
design. There are still many difficulties in the analysis of a wing with MDO
application. Approximation methods are frequently adopted for calculation of the
coupling variables. The state where all the coupled relationships are satisfied is
called multidisciplinary feasibility (MF). That is, the equilibrium between
disciplines is satisfied. The state where the equilibrium of a discipline is satisfied
is called individual feasibility (IF). In IF, the coupled relationship may not be
satisfied. Achieving MF or IF is a fairly important problem in MDO.
The above three coupled relationships of analysis involved in MDO are
represented in Figure 10.1. In Figure 10.1, the design variable vectors b1 , b 2 and
b c are fixed and the state variable vectors z1 and z 2 are obtained from h1 and
h2 .
In Figure 10.1a, analysis is independently performed for each discipline. In
b1 , b 2 , b c
z1 h1 (b1 , b c ) z 2 h 2 (b 2 , b c )
z1 z2
(a) Uncoupled analysis
b1 , b 2 , b c
z1 h1 (b1 , b c )
z1c
z2 h 2 (b 2 , b c , z1c )
z1 z2
(b) Decoupled analysis
b1 , b 2 , b c
z c2
z1 h1 (b1 , b c , z c2 ) z2 h 2 (b 2 , b c , z1c )
z1c
z1 , z 2
(c) Coupled analysis
Figure 10.1. Patterns of analysis in MDO

Figure 10.1b, z 1 is obtained from h1 and z 1 c is a subset of z 1 . z 1 c becomes

input to h 2 . One analysis for each discipline is required in these two cases. On
the other hand, results of a discipline are used as input for the other discipline. The
analysis process terminates when z1c and z c2 are not changed. MF is the state
when the coupling variables are not changed. This is the state when analysis of a
discipline is performed regardless of the coupling variables. The process in Figure
10.1c is called a system analysis in MDO. MDO generally includes the process in
Figure 10.1c, and the relationship in Figure 10.1c is the most important in MDO.
10.2.2 Formulation of MDO
MDO formulation is defined based on the coupled aspects of the design variables,
the objective function and constraints. An MDO problem is formulated as follows:
Find b1, b 2 , b c (10.3a)
to minimize f (b1 , b 2 , b c , z1 , z 2 ) (10.3b)
subject to g (b1 , b 2 , b c , z1 , z 2 ) d 0 (10.3c)
z1 h1 (b1 , b c , z c2 ) (10.3d)
z2 h 2 (b 2 , b c , z1c ) (10.3e)
where b1 , b 2 and b c are local design variable vectors of disciplines 1 and 2, and
the common design variable vector shared by the two disciplines, respectively.
f [ f1 f 2 f c ] consists of the local objective functions for disciplines 1 and 2,
and the common objective function. g R m is the constraint vector.
The formulation in Equation 10.3 can be rewritten as follows:
Find b1 , b 2 , b c (10.4a)
ª f1 (b1 , b c , z1 ) º
« f (b , b , z ) »
to minimize « 2 2 c 2 » (10.4b)
«¬ f c (b1 , b 2 , b c , z1 , z 2 )»¼
subject to g1 (b1 , b c , z1 ) d 0 (10.4c)
g 2 (b 2 , b c , z 2 ) d 0 (10.4d)
g c (b1 , b 2 , b c , z1 , z 2 ) d 0 (10.4e)
z1nc h11 (b1 , b c , z c2 ) (10.4f)

z1c h12 (b1 , b c , z c2 ) (10.4g)
z 2nc h 21 (b 2 , b c , z1c ) (10.4h)
z c2 h 22 (b 2 , b c , z1c ) (10.4i)
where g i R mi is the constraint vector of discipline i and g c R mc is the vector

of common constraints. m1 , m2 and mc are the numbers of constraints for
disciplines 1, 2 and of common constraints, respectively. Generally, f c is not
used. h11 , h12 , h 21 and h 22 are the same as in Equation 10.2. The formulation
process is the same when there are three or more disciplines.
10.2.3 Classification of MDO Methods
Multiple disciplines are involved in MDO. The relationships between disciplines

are represented in Figure 10.2a. MDO methods solve the problem having the
structure of Figure 10.2a. The methods are classified into single level methods and
multilevel methods. The single level methods generally have a single optimizer
and directly use the structure in Figure 10.2a. On the other hand, multilevel
methods modify the relationship in Figure 10.2a into the hierarchical structure
illustrated in Figure 10.2b. Each level has an optimizer. Two-level methods are
Discipline 4
Discipline 1 Discipline 3
Discipline 2
(a) Nonhierarchical structure
Upper level
Lower level
Discipline 1 Discipline 2 Discipline 3 Discipline 4
(b) Hierarchical structure
Figure 10.2. Structures of the MDO processes

Design optimization
b1 , b 2 , b c z1 , z 2
z1c
z1 h1 (b1 , b c , z c2 ) z2 h 2 (b 2 , b c , z1c )
z c2
Figure 10.3. Flow of the MDO methods Using MF
generally used for multilevel MDO methods.

The single level method is classified into three groups. The first is the method
that directly uses the nonhierarchical relationship in analysis. It is the same as
general optimization methods as shown in Figure 10.3. MF (multidisciplinary
feasibility) in analysis is always satisfied. The MDF method has this approach.
In the second single level method, the disciplines are decomposed to avoid the
expensive system analysis as illustrated in Figure 10.4. There is only one
optimizer. The process to obtain the coupling variables is included in the
optimization cycle. The second method is classified into two methods. The first
one defines the complimentary vectors s1c and s c2 . Coupling vectors z1c and
z c2 are replaced by the complimentary vectors in each discipline. Each discipline
conducts analysis with the complimentary vectors. The complimentary vectors are
included in the optimization process and updated during the design cycles. When
the final solution is found, the coupling is solved. Instead of performing the
system analysis, the design problem becomes complicated. The overall flow is
illustrated in Figure 10.4. At the analysis stage of each discipline, IF is satisfied.
However, MF is satisfied when the design process ends. The IDF method has
these characteristics.
The second method uses the aforementioned concept of SAND (simultaneous
analysis and design). It does not even have an analysis process. The overall flow
is illustrated in Figure 10.5. It decomposes the disciplines and has one optimizer.
Analysis is not performed for each discipline. The lower level in Figure 10.5 only
checks the validity of s1 and s 2 , which are the output of the optimization process.
b1 , b c , s c2 Design optimization b 2 , b c , s1c
(z1c s1c ) 0, (z c2 s c2 ) 0
z1c z c2
z1c h1 (b1 , b c , s c2 ) z c2 h 2 (b 2 , b c , s1c )
Figure 10.4. Flow of the MDO methods Using IF

b1 , b c , s1 , s c2 Design optimization b 2 , b c , s1c , s 2

e1 0, e 2 0
e1 e 2
e1 h1' (b1 , b c , s c2 ) s1 e2 h '2 (b 2 , b c , s1c ) s 2
Figure 10.5. Flow of the MDO methods Using SAND
This is validated by checking e1 0 and e 2 0 . Therefore, IF and MF are

satisfied simultaneously at the end. When evaluators that calculate e1 and e 2 are
available, the method can be exploited. However, if they are not available, the
existing analyzers should be transformed to the evaluators. Also, since MF and IF
should be satisfied in the optimization process, the design problem is fairly
complicated. The AAO method is representative of this method.
The third single level method performs the system analysis and decomposes the
disciplines. It has a hierarchical structure as shown in Figure 10.2b. MDOIS is
included in this method. The system analysis is performed at each cycle. The
lower part of Figure 10.3 is carried out for the system analysis at the beginning of
each optimization cycle. The coupling variables are conveyed to the disciplines
and each discipline performs optimization with the conveyed variables. It is noted
that each discipline has an optimizer.
All the multilevel methods decompose disciplines. They have the structure
shown in Figure 10.2b and have two levels, such as the upper level and the lower
level. Optimizers reside in both levels. In other words, optimization is performed
for the entire system as well as for each discipline. The method is classified into
two groups according to the way MF is satisfied.
The first one uses a concept similar to that shown in Figure 10.3. CSSO and
BLISS are included. The system analysis is performed at the beginning. The
upper level determines some parameters and transmits them to the lower level by
using optimization. Each discipline of the lower level conducts an optimization.
The parameters are considered as constants or are linearly approximated. Some
values for the upper level are determined and transmitted to the upper level in the
optimization cycle.
The second one has a flow similar to Figure 10.4. For instance, CO belongs to
this method. It does not conduct the system analysis. IF is satisfied during the
optimization process and MF is satisfied at the end. Each discipline of the lower
level has an optimizer and determines the design variables. In the lower level, IF
is satisfied while the target values for the design and the coupling variables are
accomplished. The design variables determined in the lower level are transmitted
to the upper level. The upper level determines the target values for the design and
the coupling variables to satisfy MF by using an optimizer. IF is always satisfied
during the optimization process of the upper level. The target values are conveyed
to the lower level.
The classification is made based on the existing MDO methods. For example,
there is no multilevel method that uses evaluators. Thus, it is not included. Each
method will be explained later in detail.
Example 10.1 [Finding MF and IF]

In the following equations:
z1c 2b1 3b32 0.5 z 2c
z 2c 5b22 2b3 z1c
when [b1 b2 b3 ]T [1 1 1]T
(1) Using the fixed point iteration method, find a solution that satisfies MF.
(2) Use [ s1c s 2c ] [1 1] of Figure 10.4 and find a solution that satisfies IF.
(3) Using the condition of (2), find e1 and e 2 of Figure 10.5.
Solution
(1) The initial values of z1c and z 2c are defined by 1’s
z1c 2b1 3b32 0.5 z 2c ( z 2c 1) (10.5a)
z 2c 5b22 2b3 z1c ( z1c 1) (10.5b)
The result of Equation 10.5a is substituted into Equation 10.5b and the
result of Equation 10.5b is substituted into Equation 10.5a. The process is
repeated until it converges. Then a solution satisfying MF is obtained.
The process is shown in Table 10.1. The final solutions are z1c 3.0 and
z 2c 4.0.
(2) The solutions satisfying IF are
z1c h1 (b1 , b3 , s 2c )
2b1 3b32 0.5s 2c
(2)(1) (3)(1) 2 0.5(1) 4.5 (10.6)
z 2c h2 (b2 , b3 , s1c )
5b2 2b3 s1c

Table 10.1. The process of the system analysis in Example 10.1
No. z1c z 2c
1 1.00000 1.00000
2 4.50000 6.00000
3 2.00000 2.50000

39 3.00000 3.99999
40 3.00000 4.00000
(5)(1) (2)(1) 1 6.0 (10.7)
(3) The solutions for e1 and e 2 are
e1 h1 ' (b1 , b3 , s 2c ) s1c
2b1 3b32 0.5s 2c s1c
(2)(1) (3)(1) 2 0.5(1) 1 3.5 (10.8)
e2 h2 ' (b2 , b3 , s1c ) s 2c
5b2 2b3 s1c s 2c
(5)(1) (2)(1) 1 1 5.0 (10.9)
10.3 Basic Theories for MDO Methods

The basic theories for the MDO methods are introduced here.
10.3.1 Linear Decomposition and Global Sensitivity Equation
The optimization problem in Equation 10.3 may be difficult to solve for large-scale
problems. A large-scale problem can be decomposed with respect to each
discipline. The decomposition is usually carried out based on the type of analysis.
Using linear addition for objective functions in Equation 10.3 ( f f1 f 2 f c ) ,
Equation 10.3 can be modified as follows:
Find b1 , b 2 , b c (10.10a)
to minimize f (b1, b 2 , b c , z1, z 2 ) (10.10b)
subject to g (b 1 , b 2 , b c , z 1 , z 2 ) d 0 (10.10c)
In Equation 10.10, the objective function and constraints are coupled with
respect to the design variables. The objective function and constraints can be
linearly expanded as follows:
df df df
f # f0 'b1 'b 2 'b c (10.11a)
db1 db 2 db c
dg j dg j dg j
g j # g j0 'b1 'b 2 'b c (j 1, 2,..., m) (10.11b)
db1 db 2 db c
d ª d d d º T
where « », 'b l ['bl1 'bl 2 'blr ] l {1, 2,
db l db
¬« l1 d b l2 db »
lr ¼
c} and r {nd1, nd 2, ndc}. The subscript 0 denotes the values of the current
design and ' is perturbation from the current design.
Using Equation 10.11, an optimization problem for each discipline can be
formulated by its own design variables. Suppose b1 and b c are allocated to
discipline 1 and b 2 is allocated to discipline 2. Then the optimization problem for
each discipline is defined as follows:
(1) Discipline 1
Find b1 , b c (10.12a)
df df
to minimize f 0 'b1 'b c (10.12b)
db1 db c
dg j dg j
subject to g j0 'b1 'bc d 0 ( j 1, 2, ..., m) (10.12c)
db1 db c
(2) Discipline 2
Find b2 (10.12d)
df
to minimize f 0 'b 2 (10.12e)
db 2
dg j
subject to g j0 'b 2 d 0 ( j 1,2,..., m) (10.12f)
db 2
In Equation 10.12, the optimization problem in Equation 10.3 is linearized and

decomposed. This process is called linear decomposition. Derivatives (sensitivity
information) of the objective function and constraints are required for the
decomposition. The derivative terms in Equation 10.11 are
df wf wf dz nc wf dz c wf dz nc wf dz c
nc 1 c 1 nc 2 c 2 (10.13a)
db1 wb1 wz1 db1 wz1 db1 wz 2 db1 wz 2 db1
df wf wf dz nc wf dz c wf dz nc wf dz c
nc 1 c 1 nc 2 c 2 (10.13b)
db 2 wb 2 wz1 db 2 wz1 db 2 wz 2 db 2 wz 2 db 2
df wf wf dz nc wf dz c wf dz nc wf dz c2
nc 1 c 1 nc 2 c (10.13c)
db c wb c wz1 db c wz1 db c wz 2 db c wz 2 db c
dg j wg j wg j dz1nc wg j dz1c wg j dz nc
2
wg j dz c2
(10.13d)
db1 wb1 wz1nc db1 wz1c db1 wz nc
2 db1 wz c2 db1
dg j wg j wg j dz1nc wg j dz1c wg j dz nc wg j dz c2
nc
c nc 2 c (10.13e)
db 2 wb 2 wz1 db 2 wz1 db 2 wz 2 db 2 wz 2 db 2
dg j wg j wg j dz1nc wg j dz1c wg j dz 2nc wg j dz c2

(10.13f)
db c wb c wz1nc db c wz1c db c wz 2nc db c wz c2 db c
dz qk ª dz qk dz qk dz qk º
where « » (l {1, 2, c}, k {1, 2}, q {nc, c}) and
db l «¬ dbl1 dbl 2 dblr »¼
r {nd1, nd 2, ndc}, j = 1, 2,…,m. The partial derivatives on the right hand side
of Equation 10.13 are explicit ones and the ordinary derivatives are obtained from
the system analysis.
Equation 10.2 can be derived as follows:
wh11 wh wh
dz1nc db1 11 db c 11c
dz c2 (10.14a)
wb1 wb c wz 2
wh12 wh wh
dz1c db1 12 db c 12c
dz c2 (10.14b)
wb1 wb c wz 2
wh 21 wh wh
dz nc
2 db 2 21 db c 21c
dz1c (10.14c)
wb 2 wb c wz1
wh 22 wh wh
dz c2 db 2 22 db c 22 dz1c (10.14d)
wb 2 wb c wz1c
wh pq ª wh pq wh pq wh pq º wh pq
where « » (l {1,2, c}, p, q {1,2}),
wb l «¬ wbl1 wbl 2 wblr »¼ wz lc
ª wh pq wh pq wh pq º
« c » (l , p, q {1,2}), s {nc1, nc 2} and r {nd1, nd 2,
«¬ wzl1 wzlc2 wzlcs »¼
ndc}. From Equation 10.14, derivatives of the state variables are obtained. For
instance, derivatives of the state variables with respect to b1 are
dz1nc wh11 wh11 dz c2

c (10.15a)
db1 wb1 wz 2 db1
dz1c wh12 wh12 dz c2

c (10.15b)
db1 wb1 wz 2 db1
dz 2nc wh 21 dz1c
(10.15c)
db1 wz1c db1
dz c2 wh 22 dz1c
(10.15d)
db1 wz1c db1
Using the matrix form, Equation 10.15 is modified to
ª wh11 º ª dz1nc º
«I 0 0 »« »
« wz 2 c » « db1 » ª wh11 º
wh « c » « wb »
«0 I 0 12c » « dz1 » « 1»
« wz 2 » « db1 »
« » « wh12 » (10.16)
«0 wh 21
nc « wb1 »
I 0 » «« dz 2 »» « 0 »
« wz1c » db
« »
« wh » « c1 »
«¬ 0 »¼
«0 22 0 I » « dz 2 »
«¬ wz1c »¼ « db1 »
¬ ¼
The terms related to the coupling variables in Equation 10.16 are expressed as
ª wh12 º ª dz1c º
« I « » ª wh12 º
« wz c2 »» « db1 » « wb » (10.17)
« wh 22 «
I » « dz 2 »
c » « 1»
« wz c » db ¬ 0 ¼
¬ 1 ¼¬ 1 ¼
Equations 10.16 and 10.17 are called global sensitivity equations (GSE). Their
solution is the derivative of coupling variables with respect to design variables. It
can be used in Equation 10.13 and various MDO methods.
10.3.2 Optimum Sensitivity
When a parameter, which is considered as a constant, is changed, the optimum

sensitivity is the ratio between the change of optimum values and the change of the
parameter. For example, when the external load, allowable stress, material
properties, etc. are changed in structural optimization, the optimum sensitivity is
the sensitivity of optimum values with respect to the parameters. In MDO, if a
variable is determined in the upper level and considered as a constant in the lower
level, the variable is a parameter in the lower level. The equation for the optimum
sensitivity is derived from the KKT necessary condition.
Suppose we have only inequality constraints in Equation 3.1. The KKT
necessary condition for such a problem is
f (b) ¦ O j g j (b) 0 (10.18a)

jJ
g j (b) 0, jJ (10.18b)
Oj !0 (10.18c)
where J represents the active set. Since equality constraints are always active, the
equality constraints can be considered in the same manner except for the sign of
the Lagrange multiplier vector Ȝ in Equation 10.18c. Therefore, the derivation in
this section is valid for problems with equality constraints.
When the parameter p is changed, Equation 10.18 should not be changed at the
optimum point with respect to p. In other words, the derivatives of Equation
10.18a with respect to p are zeroes. It is expressed as
n ª w 2 f (b ) w 2 g j (b) º wbk wg j (b) wO j w 2 f (b)

¦« ¦ Oj » ¦
k 1« wbi wb k jJ wbi wbk » wp jJ wbi wp wbi wp
¬ ¼
w 2 g j (b)
¦ Oj 0, i 1,..., n (10.19a)
j J wbi wp
n wg j (b) wbi wg j (b)

¦ 0, jJ (10.19b)
i 1 wbi wp wp
Using a matrix form, Equation 10.19 is
ª A nun B nu J º ªGb º ª c nu1 º

«B 0 (10.20a)
¬ J un 0 J u J »¼ «¬GȜ »¼ «¬d J u1 »¼
where
w2 f w2g j
Aik ¦ Oj (10.20b)
wbi wbk jJ wbi wbk
wg j
Bij , jJ (10.20c)
wbi
w2 f w2 g j
ci ¦ Oj (10.20d)
wbi wp jJ wbi wp
wg j
dj , jJ (10.20e)
wp
T
ª wb1 wb2 wbn º
Gb « ... » (10.20f)
¬ wp wp wp ¼
T
ª wO1 wO2 wO J º
GȜ « ... » (10.20g)
¬ wp wp wp ¼
Gb and GȜ are obtained from Equation 10.20a. Thus, the optimum sensitivity
of the objective function is
df (b) wf (b)
T f (b)Gb (10.21a)
dp wp
or
df (b) wf (b) wg(b)

ȜT (10.21b)
dp wp wp
Equation 10.21b is derived from Equations 10.18a and 10.19b. Generally, the
df (b)
optimum sensitivity means Gb or .
dp
Example 10.2 [Global Sensitivity for a Coupling Variable]

Find b1 , b2 , bc (10.22a)
to minimize f f1 f 2 ( z1nc 0.5) 2 ( z 2nc 0.5) 2 (10.22b)
subject to g1 1.0 z1nc d 0 (10.22c)
g2 1.0 z 2nc d 0 (10.22d)
z1nc (b1 2.5) (bc 2.0) 0.5 u z 2c (10.22e)
z1c (b1 2.5) (bc 2.0) 0.4 u z 2c (10.22f)
z 2nc (b2 3.0) (bc 2.0) 0.7 u z1c (10.22g)
z 2c (b2 3.0) (bc 2.0) 0.6 u z1c (10.22h)
0.0 d b1 , b2 , bc ( 5.0) d 10.0 (10.22i)
ª dz c dz 2c º ª dz1c dz 2c º
Evaluate the global sensitivities « 1 » and « ».
¬« db1 db1 ¼» ¬« db2 db2 ¼»
Solution
ª dz c dz 2c º ª dz1c dz 2c º
Using Equation 10.15 « 1 » and « » are included in the
«¬ db1 db1 »¼ «¬ db2 db2 »¼
following simultaneous equations:
dz1c wz1c wz1c dz 2c dz 2c

1 ( 0.4) u (10.23a)
db1 wb1 wz 2c db1 db1
dz 2c wz 2c wz 2c dz1c dz1c
0 ( 0.6) u (10.23b)
db1 wb1 wz1c db1 db1

0 ( 0.4) u (10.23c)
1 ( 0.6) u (10.23d)
ª dz c dz 2c º
The solutions of Equation 10.23 are « 1 » [1.31579 0.78947] and
«¬ db1 db1 »¼
ª dz1c dz 2c º
« » [ 0.52632 1.31579] .
«¬ db2 db2 »¼
Example 10.3 [Calculation of the Optimum Sensitivity]

We have the following optimization problem:
Find b1 , b2 (10.24a)
to minimize f (b) (b1 p) 3 b2 (10.24b)
subject to g1 b22 2b2 (b1 p ) 3 d 0 (10.24c)
0 d b1 , b2 d 5 (10.24d)
df (b)
When p 2, evaluate optimum sensitivity .
dp
Solution
The optimum point is b1 0.0 and b2 0.64575 . Each component of the
optimum sensitivity is obtained from Equation 10.21b as follows:
wf (b)
3(b1 p) 2 3(0 2) 2 12 (10.25a)
wp
O 0.18892 (10.25b)
wg
2b2 2 u 0.64575 1.2915 (10.25c)
wp
df (b)
Therefore, the optimum sensitivity is
dp
df (b)
12 (0.18892)(1.2915) 11.75601 (10.26)
dp
10.4 Multidisciplinary Design Optimization Methods

The aforementioned seven MDO methods are explained here. It is assumed that
there are two disciplines.
10.4.1 Multidisciplinary Feasible (MDF)

The MDF method is a single level method which has the system analysis. It is also
called the all-in-one (AiO) method. In the MDF method, the system analysis is
treated as the analyzer of the single optimization.
Figure 10.6 presents the general structure of the MDF method. The lower part
of the figure is the system analysis. The system analysis usually uses the fixed
point iteration method. Meanwhile, sensitivity information is needed in the
optimization process. The sensitivity information is obtained from the
aforementioned global sensitivity equations or the finite difference method on the
system analysis. When the process converges, the MDF method can find a
mathematical optimum. Therefore, the solution from the MDF method is
considered as the standard solution when MDO methods are compared.
The MDF method directly solves the problem in Equation 10.3. Multiple
objective functions are modified into a single objective function. Equations 10.3d
and 10.3e are handled in the system analysis. The MDF method does not
decompose the disciplines. The number of optimizers is one and the number of
design variables is the sum of the number of local variables (nd1 nd 2) and the
Find b1 , b 2 , b c
to minimize f (b1 , b 2 , b c , z1 , z 2 )
subject to g1 (b1 , b c , z1 ) d 0
g 2 (b 2 , b c , z 2 ) d 0
g c (b1 , b 2 , b c , z1 , z 2 ) d 0
b1 , b 2 , b c z1 , z 2
z c2
z1 h1 (b1 , b c , z c2 ) z2 h 2 (b 2 , b c , z1c )
z1c
Figure 10.6. Structure of the MDF method

number of common variables (ndc).
Example 10.4 [Application of the MDF Method to Example 10.2]

Solve Example 10.2 by using the MDF method.
Solution
Equation 10.22 is the optimization formulation of Example 10.2 to use in the MDF
method. The values of b1 , b2 and bc are initialized by 5.0. At each cycle, the
system analysis is performed until the state variables z1nc , z1c , z 2nc and z 2c converge.
This process is the same as the process of Example 10.1(1). When a system
analysis is finished, the results are incorporated into the optimization process. The
final solutions are b1 3.44059, b2 4.16281, bc 2.61497 and f 0.5.
10.4.2 Individual Discipline Feasible (IDF)
The IDF method is a single level method that decomposes the disciplines. It does
not have system analysis. Complimentary vectors ( s1c and s c2 ) are adopted for the
coupling variable vectors ( z1c and z c2 ). Instead of MF, compatibility conditions
are utilized and they are included as equality constraints in the optimization
process. The formulation for the IDF method is as follows:
Find b1 , b 2 , b c , s1c , s c2 (10.27a)
to minimize f (b1 , b 2 , b c , z1 , z 2 ) (10.27b)
subject to g1 (b1 , b c , z1 ) d 0 (10.27c)
g 2 (b 2 , b c , z 2 ) d 0 (10.27d)
g c (b1 , b 2 , b c , z1 , z 2 ) d 0 (10.27e)
c1 s1c z1c 0 (10.27f)
c2 s c2 z c2 0 (10.27g)
z1 h1 (b1 , b c , s c2 ) (10.27h)
z2 h 2 (b 2 , b c , s1c ) (10.27i)
An analyzer of a discipline does not wait for the output from the other
analyzers. Instead, it performs the analysis (IF) by using the complimentary
vectors (s1c and s c2 ). Eventually, the coupling variables (z1c and z c2 ) should be
the same (MF) as the complimentary vectors (s1c and s c2 ). Thus, compatibility
Find b1 , b 2 , b c , s1c , s c2
to minimize f (b1 , b 2 , b c , z1 , z 2 )
subject to g1 (b1 , b c , z1 ) d 0
b1 , b c , s c2 b 2 , b c , s1c
g 2 (b 2 , b c , z 2 ) d 0
g c (b1 , b 2 , b c , z1 , z 2 ) d 0
MF c1 s1c z1c 0
c2 s c2 z c2 0
z1 z2
z1 h1 (b1 , b c , s c2 ) z2 h 2 (b 2 , b c , s1c )
IF IF
Figure 10.7. Structure of the IDF method
conditions (c1 and c 2 ) are defined as in Equations 10.27f and 10.27g. The
overall flow of the IDF method is presented in Figure 10.7.
When the IDF method is used, the number of design variables is increased by
the number of complimentary variables. The number of design variables is
nd1 nd 2 ndc plus N (s1c ) N (s c2 ) . N (s1c ) means the number of the elements
in vector s1c . The number of equality constraints is the number of coupling
variables. IF is always satisfied during the optimization process and MF is
satisfied when the optimum design is obtained.
Example 10.5 [Application of the IDF Method to Example 10.2]

Solve Example 10.2 by using the IDF method.
Solution
Complimentary variables s1c and s 2c are defined for state variables z1c and z 2c .
Find b1 , b2 , bc , s1c , s 2c (10.28a)
to minimize f ( z1nc 0.5) 2 ( z 2nc 0.5) 2 (10.28b)
subject to g1 1.0 z1nc d 0 (10.28c)
g2 1.0 z 2nc d 0 (10.28d)
c1 s1c z1c 0 (10.28e)

c2 s 2c z 2c 0 (10.28f)
z1nc (b1 2.5) (bc 2.0) 0.5 u s 2c (10.28g)
z1c (b1 2.5) (bc 2.0) 0.4 u s 2c (10.28h)
z 2nc (b2 3.0) (bc 2.0) 0.7 u s1c (10.28i)
z 2c (b2 3.0) (bc 2.0) 0.6 u s1c (10.28j)
The solutions of the above problem are b1 3.44476, b2 4.16617,

bc 2.61121, s1c 1.11054, s 2c 1.11196 and f 0.49999 .
10.4.3 All-at-once (AAO)
The AAO method is a single level method that does not perform either the system
analysis or individual analysis for each discipline. Analysis is generally carried
out many times for a design. The analysis is quite expensive and the analysis
results during the design process are meaningless when the final solution is
obtained. The AAO method was developed to abolish the expensive analysis
process. Evaluators are utilized instead of analyzers in the AAO method.
State variables are calculated with respect to design variables in analysis. The
analysis process can be viewed as illustrated in Figure 10.8. Figure 10.8 represents
the analysis process of discipline 1. State variable vector z1 is obtained from
design variable vectors b1 and b c and coupling variable vector z c2 .
The analysis process can be viewed from a different perspective. Figure 10.9
presents an evaluator. The design and state variable vectors are inputs. When the
results of the evaluator are the zero vector, the analysis is regarded as finished.
The concept in Figure 10.9 is expanded to MDO by using the SAND concept.
Input Output
Analyzer
b1 , b c , z c2 z1
z1 h1 (b1 , b c , z c2 )
Figure 10.8. The analysis process
Input Output
Evaluator
b1 , b c , z1 , z c2 0
e1 e1 (b1 , b c , z1 , z c2 )
h1' (b1 , b c , z c2 ) z1
Figure 10.9. The role of the evaluator

Complimentary variable vectors (s1c and s c2 ) are adopted for state variable vectors
(z1c and z c2 ). The process of the evaluator is considered in the same way as
equality constraints. This is similar to the IDF method. In the IDF method, the
complimentary variables for coupling variables are used as design variables in the
optimization process. On the other hand, the complimentary variables for coupling
and noncoupling variables are included in the design variable set. That is, all the
state variables are regarded as design variables.
The optimization formulation of the AAO method is as follows:
Find b1 , b 2 , b c , s1nc , s 2nc , s1c , s c2 (10.29a)
to minimize f (b1 , b 2 , b c , s1nc , s1c , s nc c

2 ,s2 ) (10.29b)
subject to g1 (b1 , b c , s1nc , s1c ) d 0 (10.29c)
g 2 (b 2 , b c , s nc c
2 ,s2 ) d 0 (10.29d)
g c (b1 , b 2 , b c , s1nc , s1c , s 2nc , s c2 ) d 0 (10.29e)
e1 0 (10.29f)
e2 0 (10.29g)
e1 h1c (b1 , b c , s 2 c ) s1 (10.29h)
e2 h c2 (b 2 , b c , s1c ) s 2 (10.29i)
Find b1 , b 2 , b c , s1c , s1nc , s c2 , s nc

2
to minimize f (b1 , b 2 , b c , s1nc , s1c , s nc c

2 ,s2 )
b1 , b c , s1, s c2 subject to g1 (b1 , b c , s1nc , s1c ) d 0 b 2 , b c , s 2, s1c

g 2 (b 2 , b c , s 2nc , s c2 ) d 0
g c (b1 , b 2 , b c , s1nc , s1c , s nc c
2 ,s2 ) d0
e1 0
e2 0 MF, IF
e1 e1 (b1 , b c , s1 , s c2 ) e1 e2 e2 e 2 (b 2 , b c , s 2 , s1c )
h1' (b1 , b c , s c2 ) s1 h '2 (b 2 , b c , s1c ) s 2
Figure 10.10. Structure of the AAO method

Equations 10.29h and 10.29i represent the evaluator.

The overall flow of the AAO method is illustrated in Figure 10.10. Each
discipline evaluates the validity of the given input. When the optimization process
ends, IF and MF are simultaneously satisfied. The number of equality constraints
is the same as the number of state variables. The number of design variables is
nd1 nd 2 ndc plus N (s1nc ) N (s 2nc ) N (s1c ) N (s c2 ) .
The above three methods have the following characteristics:
(1) They are single level methods.
(2) An analyzer or an evaluator is used in each discipline.
(3) The formulations are simple and easy to understand.
(4) The common objective function and constraints are handled in the same
manner as the independent ones.
Some MDO methods have the design process in each discipline. These
methods are introduced in the subsequent sections.
Example 10.6 [Application of the AAO Method to Example 10.2]

Solve Example 10.2 by using the AAO method.
Solution
In the IDF method, complimentary variables s1nc , s1c , s2nc and s 2c are added to the
design variable set. The formulation is as follows:
Find b1 , b2 , bc , s1nc , s1c , s 2nc , s 2c (10.30a)
to minimize f ( s1nc 0.5) 2 ( s 2nc 0.5) 2 (10.30b)
subject to g1 1.0 s1nc d 0 (10.30c)
g2 1.0 s 2nc d 0 (10.30d)
e11 (b1 2.5) (bc 2.0) 0.5 u s 2c s1nc 0 (10.30e)
e12 (b1 2.5) (bc 2.0) 0.4 u s 2c s1c 0 (10.30f)
e21 (b2 3.0) (bc 2.0) 0.7 u s1c s 2nc 0 (10.30g)
e 22 (b2 3.0) (bc 2.0) 0.6 u s1c s 2c 0 (10.30h)
The solutions of Equation 10.30 are b1 6.05743, b2 6.77966, bc 0.0, s1nc

1.00001, s1c 1.11354, s2nc 0.99951, s 2c 1.11111 and f 0.49957.
10.4.4 Concurrent Subspace Optimization (CSSO)
CSSO is a multilevel method, which has a system analysis and decomposes

disciplines. The design process is used in the disciplines. CSSO divides an MDO
problem into upper and lower levels. In the upper level, some parameters are
determined by an optimization process for the design process of the lower level.
The parameters enable a discipline to consider the objective function and
constraints of other disciplines. The optimization problem of the upper level is
called the coordinate optimization problem (COP). In the lower level (this is also
called a subspace) of each discipline, optimization is carried out with the
parameters from the upper level. The functions in a discipline are defined by using
the linear decomposition introduced in Section 10.3.1. The lower level provides
the optimization results and the optimum sensitivity. The optimization results are
used in the system analysis and the optimum sensitivity is used in the optimization
of the upper level. The process proceeds in an iterative fashion.
CSSO first performs the system analysis in Equation 10.3. Equations 10.3d
and 10.3e are utilized for the system analysis. A method such as the fixed point
iteration method is employed for the system analysis until the coupling variable
vectors z1c and z c2 are not changed. Details are given in Equation 10.2.
Sensitivity information is needed to decompose the entire problem into
optimization problems for the disciplines. Global sensitivity information in
Equations 10.15í10.17 is needed for the coupling variables and is used for linear
approximation of the coupling variables. When the number of constraints in a
discipline is large, all the constraints can be transformed to a cumulative constraint
using the Kreisselmeier–Steinhauser (KS) function. It is
1 ªm º
C ln « ¦ expUg i » (10.31)
U ¬i 1 ¼
where U is the adjustment coefficient, m is the number of constraints and g i is

the ith constraint. If any one of the constraints is violated, the value of Equation
10.31 is positive. When all the constraints are satisfied, it is zero or negative.
Thus, if the value of 10.31 is zero or negative, all the constraints are satisfied.
The flow of CSSO is illustrated in Figure 10.11. At the first cycle, the system
analysis is performed and the global sensitivities are calculated with respect to
b1 , b 2 and b c . With the results of the upper level, optimization is carried out for
each discipline in the lower level and the optimum sensitivity is evaluated. In the
upper level, COP is defined by the optimum sensitivity information. Then the
coefficients for the next cycle are defined for the lower level. The convergence
criteria from the second cycle are checked.
The problem in the lower level is investigated. In CSSO, the common variable
b c should be considered in one discipline. The optimization problems of the two
disciplines are defined as follows:
(1) Discipline 1
Definition of design variables
System analysis
Converge? Yes
Stop
No
Sensitivity from global sensitivity equations
Update
design
Optimization for each discipline variables
Optimum sensitivity analysis
Coordination optimization problem
Figure 10.11. Flow of CSSO
Given r11 , r12 , t11 , t12 , s1 , s 2 (10.32a)
find b1 , b c (10.32b)
df df
to minimize f # f 0 'b1 'b c (10.32c)
db1 db c
subject to C1 d C10 [ s1 (1 r11 ) (1 s1 )t11 ] (10.32d)
~ dC 2 dC 2
C 2 # C 20 'b 1 'b c d C 20 [ s 2 (1 r12 ) (1 s 2 )t12 ]
db 1 db c
(10.32e)
z1nc h11 (b1 , b c , ~z 2c ) (10.32f)
z1c h12 (b1 , b c , ~z 2 c ) (10.32g)
~z nc nc dz 2nc dz nc
2 z10 'b1 2 'b c (10.32h)
db1 db c
~z c dz c2 dz c
2 z c20 'b1 2 'b c (10.32i)
db1 db c
(2) Discipline 2
Given r21 , r22 , t 12 , t 22 , s1 , s 2 (10.33a)
find b2 (10.33b)
df
to minimize f # f 0 'b 2 (10.33c)
db 2
~ dC
subject to C1 # C10 1 'b 2 d C10 [ s1 (1 r21 ) (1 s1 )t12 ] (10.33d)
db 2
C 2 d C 20 [ s 2 (1 r22 ) (1 s 2 )t 22 ] (10.33e)
~z nc nc dz1nc
1 z10 'b 2 (10.33f)
db 2
~z c c dz1c
1 z10 'b 2 (10.33g)
db 2
z 2nc h 21 (b 2 , b c , ~z1c ) (10.33h)
z c2 h 22 (b 2 , b c , ~z1c ) (10.33i)
s p , rkp and t kp are constants in the lower level and are determined in COP of
the upper level. k and p represent the discipline number, and s p , rkp and t kp are
given by the user at the first cycle. When the objective function of a discipline is a
function of the local variables of the corresponding discipline, the above
approximation is not needed. Otherwise, the objective function is linearly
approximated. A cumulative constraint is employed to include all the constraints.
The cumulative constraint of a discipline is considered through approximation by
another discipline. This process enables on to consider the constraints of the other
discipline because a change of the design in a discipline may affect the other
discipline through coupling. This is shown in Equations 10.32e and 10.33d.
dC
(i = 1, 2, c) is the information obtained from the global sensitivity equation.
db i
The constraints of the other discipline are considered by s p , rkp and t kp .
The coefficients are illustrated in Figure 10.12, which explains rkp . Suppose
the constraint of discipline 1, C1 is violated. It is corrected by r11 in discipline 1
and by r21 in discipline 2. Thus, r21 is the amount of responsibility of discipline 2
for the constraint violation in discipline 1. This is called the responsibility
coefficient and has the following relationship:
C1
r11
r11 r21 1
r21
(a) Responsibility coefficient r
C1
t11 t 12 0
t 12
t11
(b) Trade-off coefficient t
Figure 10.12. Responsibility and trade-off coefficients
NSS
p
¦ rk 1 (p 1, 2,..., NSS ) (10.34)
k 1
where NSS is the number of disciplines. The meaning of Equation 10.34 is that
“if the constraint violation of the pth discipline C p is 1, rkp is the amount of
responsibility for the kth discipline.”
The trade-off coefficient t is schematically drawn in Figure 10.12b. Suppose
constraint C1 in discipline 1 is satisfied. t11 is negative, thus, discipline 1 tries to
satisfy C1 that is oversatisfied by t11 . t 12 is positive and discipline 2 tries to
~
further reduce the objective function although C1 is violated by t12 . t kp is the
trade-off coefficient of constraint C p in the kth discipline. The trade-off
coefficients have the following relationship:
NSS
p
¦ tk 0 (p 1, 2,..., NSS ) (10.35)
k 1
The meaning of Equation 10.35 is that “if constraint C p of the pth discipline is
satisfied, the constraint is oversatisfied or violated by t kp in the kth discipline.”
s p is a switch coefficient for the responsibility and trade-off coefficients.
When C p is violated, s p 1 and the trade-off coefficient is not activated. When
p
C p is satisfied, s 0 and the responsibility coefficient is not activated. The
switch coefficient is automatically determined according to the status of the
constraint violation. On the other hand, the responsibility and trade-off
coefficients can be defined by the user. A method exists to initialize the
responsibility and trade-off coefficients (Sobieszczanski-Sobieski 1988). From the
second cycle, they are determined from the COP.
The COP is formulated as follows:
Find rkp , t kp (k , p 1, 2) (10.36a)

2 2 df 2 2 df
to minimize f 0 ¦ ¦ p
'rkp ¦ ¦ p 't kp (10.36b)
p 1k 1drk p 1 k 1dt k
2
p
subject to ¦ rk 1 (0 d rkp d 1) ( p 1, 2) (10.36c)
k 1
2
p
¦ tk 0 (p 1, 2) (10.36d)
k 1
Equation 10.36b needs the optimum sensitivity of the objective function with
respect to the responsibility and trade-off coefficients. Since the coefficients are
treated as constants in each discipline, the optimum sensitivity is obtained at each
discipline.
CSSO is a two-level optimization method. Each discipline of the lower level
(discipline) needs an optimizer and the upper level also needs an optimizer. IF and
MF are satisfied by the system analysis. Constraints are indirectly shared by all
the disciplines and common design variables are specifically handled by one
discipline.
Example 10.7 [Application of CSSO]

Solve Example 10.2 by using CSSO.
Solution
From the system analysis, the global sensitivities are obtained as follows:
dz1nc wz1nc wz1nc dz 2c dz 2c

c 1 (0.5) u (10.37a)
db1 wb1 wz 2 db1 db1

1 (0.4) u (10.37b)
dz 2nc wz 2nc dz1c dz1c

(0.7) u (10.37c)
db1 wz1c db1 db1
dz 2c wz 2c dz1c dz1c
(0.6) u (10.37d)
db1 wz1c db1 db1
dz1nc wz1nc dz 2c dz 2c
(0.5) u (10.37e)
db2 wz 2c db2 db2
dz1c wz1c dz 2c dz 2c
(0.4) u (10.37f)
db2 wz 2c db2 db2
dz 2nc wz 2nc wz 2nc dz1c dz1c

c 1 (0.7) u (10.37g)
db2 wb2 wz1 db2 db2
1 (0.6) u (10.37h)
dz1nc wz1nc wz1nc dz 2c dz 2c

c 1 (0.5) u (10.37i)
dbc wbc wz 2 dbc dbc

1 (0.4) u (10.37j)
dbc wbc wz 2c dbc dbc
dz 2nc wz 2nc wz 2nc dz1c dz1c

c 1 (0.7) u (10.37k)
dbc wbc wz1 dbc dbc
1 (0.6) u (10.37l)
dbc wbc wz1c dbc dbc
dz1nc dz1c dz 2nc

Solving Equation 10.37, 1.39474, 1.31579, 0.92105,
db1 db1 db1
dz 2c dz1nc dz1c dz 2nc
0.78947, 0.65790, 0.52632, 1.36842,
db1 db2 db2 db2
dz 2c dz1nc dz1c dz 2nc
1.31579, 0.73684, 0.78947, 0.44737 and
db2 dbc dbc dbc
dz 2c
0.52632. These are utilized for the formulation of each discipline. The
dbc
common variable is handled within discipline 1. The formulations are as follows:
(1) Discipline 1
Given r11 , r12 , t11 , t12 , s1 , s 2 (10.38a)
find b1 , bc (10.38b)
df df
to minimize f1 f0 u 'b1 u 'bc
db1 dbc
§ wf dz nc wf dz nc ·
f 0 ¨¨ nc 1 nc 2 ¸¸ u 'b1
© wz1 db1 wz 2 db1 ¹
§ wf dz nc wf dz nc ·
¨¨ nc 1 nc 2 ¸ u 'bc
¸
© wz1 dbc wz 2 dbc ¹
f 0 [2( z1nc 0.5)(1.39474) 2( z 2nc 0.5)(0.92105)] u 'b1
[2( z1nc 0.5)(0.73684) 2( z 2nc 0.5)(0.44737)] u 'bc

(10.38c)
subject to C1 d C10 [ s1 (1 r11 ) (1 s1 )t11 ] (10.38d)
~ dC2 dC
C2 # C20 ǻb1 2 ǻbc d C20 [ s 2 (1 r12 ) (1 s 2 )t12 ]
db1 dbc
(10.38e)
1
C1 ln[exp( U1 g1 )] g1 (10.38f)
ȡ1
z1nc (b1 2.5) (bc 2.0) 0.5 u ~z 2c (10.38g)
z1c (b1 2.5) (bc 2.0) 0.4 u ~z 2c (10.38h)
~z c c dz 2c dz c
2 z 20 u 'b1 2 u 'bc
db1 dbc
c
z 20 (0.78947) u 'b1 (0.52632) u 'bc (10.38i)
~ dz 2nc dz nc
z 2nc nc
z 20 u 'b1 2 u 'bc
db1 dbc
nc
z 20 (0.92105) u 'b1 (0.44737) u 'bc (10.38j)
where subscript 0 represents the initial value of each cycle.

(2) Discipline 2
Given r21 , r22 , t 12 , t 22 , s1 , s 2 (10.39a)
find b2 (10.39b)
wf § wf dz nc wf dz nc ·
to minimize f 2 f0 u 'b2 f 0 ¨¨ nc 1 nc 2 ¸¸ u 'b2
wb2 © wz1 db2 wz 2 db2 ¹
(10.39c)
f 0 [2( z1nc 0.5)(0.65790) 2( z 2nc 0.5) u 1.36842]'b2

(10.39d)
~ dC
C1 # C10 1 ǻb2 d C10 [ s1 (1 r21 ) (1 s1 )t12 ] (10.39e)
db2
C 2 d C 20 [ s 2 (1 r22 ) (1 s 2 )t 22 ] (10.39f)
1
C2 ln[exp( U 2 g 2 )] g2 (10.39g)
U2
z 2nc (b2 3.0) (bc 2.0) 0.7 u ~

z1c (10.39h)
z 2c (b2 3.0) (bc 2.0) 0.6 u ~z1c (10.39i)
~z c c dz1c c
1 z10 u 'b2 z10 (0.52632) u 'b2 (10.39j)
db2
~z nc nc dz1nc nc
1 z10 u 'b2 z10 (0.65789) u 'b2 (10.39k)
db2
(3) Upper level (COP)
Find rkp , t kp (k , p 1,2) (10.40a)

2 2 df 2 2 df
to minimize f 0 ¦ ¦ p
'rkp ¦ ¦ p 't kp (10.40b)
p 1k 1drk p 1 k 1dt k
2
p
subject to ¦ rk 1 (0 d rkp d 1) ( p 1, 2) (10.40c)
k 1
2
p
¦ tk 0 (p 1, 2) (10.40d)
k 1
df df
where p
and p are the optimum sensitivity of each discipline in the
drk dt k
lower level. The solutions are b1 3.08234, b2 3.08524, bc 2.97466
and f 0.50243 .
10.4.5 Bilevel Integrated System Synthesis (BLISS)
BLISS is a multilevel method, which has a system analysis and decomposes

disciplines. It divides an MDO problem into an upper level and a lower level.
Each discipline in the lower level has local design variables, while the common
variables are considered as constants. On the other hand, the upper level has
common variables as design variables while the local variables of the disciplines
are regarded as constants.
BLISS can handle an MDO problem without common constraints. Thus, it
solves Equation 10.4 without Equation 10.4e. Using the linear decomposition, the
problem in Equation 10.4 can be reformulated as follows:
(1) Discipline 1
Given b1 , b c , z c2 (10.41a)
find 'b1 (10.41b)
df
to minimize 'b1 (10.41c)
db1
subject to g1 (b1 , b c , z1 ) d 0 (10.41d)
z1nc h11 (b1 , b c , z c2 ) (10.41e)
z1c h12 (b1 , b c , z c2 ) (10.41f)
(2) Discipline 2
Given b 2 , b c , z1c (10.42a)
find 'b2 (10.42b)
df
to minimize 'b 2 (10.42c)
db 2
subject to g 2 (b 2 , b c , z 2 ) d 0 (10.42d)
z 2nc h 21 (b 2 , b c , z1c ) (10.42e)
z c2 h 22 (b 2 , b c , z1c ) (10.42f)
(3) Upper level
Given bc (10.43a)
find 'b c (10.43b)
df
to minimize 'b c (10.43c)
db c
The objective functions in the lower level are linearized with respect to the
local variables as shown in Equations 10.41c and 10.42c. The global sensitivity is
utilized for the linearization. For the upper level, the optimum sensitivity of the
objective function with respect to common variables is required as shown in
Equation 10.43c. The optimum sensitivity is obtained from the lower level. Since
there are two disciplines in the lower level, different optimum sensitivities can be
obtained from different disciplines. Therefore, the optimum sensitivity is
calculated based on satisfaction of the system analysis. BLISS uses GSE/OS
(GSE/optimized subsystems) for this. GSE/OS is a union of the global sensitivity
equation and the optimum sensitivity equations in Equations 10.41 and 10.42. It is
as follows:
nc
ª wh11 wh11 º ª dz1 º ª wh11 º
«I 0 0 0 » « db »
wz c2 wb1 « wb »
« » « cc » « c »
wh wh « » « wh12 »
«0 I 0 12 12 0 » « d z1 »
« wz c2 wb1 » db « wb c »
« »« c » « wh »
«0 wh 21 I 0 0
wh 21 » « dz nc »
2 « 21 »
« wz1c wb 2 » « db » « wb c »
« « c» (10.44)
wh wh » c « wh 22 »
«0 22 0 I 0 22 » « dz 2 » « »
« wz1c wb 2 » « db » « wb c »
« wb1 »« c » « wb1 »
«0 0 0 I 0 » « db1 » « wb »
« wz c2 » « db c » « wb c »
« wb 2 »« » « 2 »
«0 wz c 0 0 0 I » « db 2 »
«¬ wb c »¼
¬ 1 ¼ ¬« db c ¼»
wb1 wb1 wb 2
where and are the optimum sensitivities of discipline 1, and and
wz c2 wb c wz1c
wb 2
are those of discipline 2.
wb c
The optimum sensitivities can be obtained by an optimization problem derived
by using the global sensitivity of the objective function and constraints as follows:
Given 'p one of ('b c , 'z ic ) (10.45a)
find 'b i (10.45b)
df
to minimize f 0 'b i (10.45c)
db i
dg i
subject to gi0 ǻb i d 0 (10.45d)
db i
where i is the discipline number. The results of Equation 10.45 are utilized to
calculate the approximate optimum sensitivity as follows:
wb i 'b i
| (10.46)
wp 'p
The optimum sensitivity can also be obtained from the other methods in Section
10.3.2. The results of Equation 10.46 are substituted into Equation 10.44 and the
solution of Equation 10.44 is evaluated. The other partial derivative terms in
Equation 10.44 are the same as those of Equation 10.16. The solution of Equation
10.44 is adopted for the linearization of Equation 10.43c.
The steps of BLISS are as follows:
Step 1. Design variable vectors b1 ,b 2 and b c are initialized.
Step 2. The system analysis is performed to calculate z1 and z 2 .
Step 3. Convergence is checked from the second cycle.
Step 4. Global sensitivities are evaluated.
Step 5. Problems in Equations 10.41 and 10.42 are solved in the lower level.
Step 6. Equation 10.44 for GSE/OS is established and solved.
Step 7. In the upper level, the problem in Equation 10.43 is solved.
Step 8. The design variables are updated and Step 2 is redone.
In BLISS, IF and MF are satisfied during optimization because the system analysis
is performed first.
Example 10.8 [Application of BLISS]

Solve Example 10.2 by using BLISS.
Solution
After the system analysis is performed with the current design, global sensitivities
are obtained for the objective functions in the lower level. This is the same as
Example 10.7, which uses CSSO. Each discipline and the upper level solve the
following optimization problems:
(1) Discipline 1
Given b1 , bc , z 2c (10.47a)
find 'b1 (10.47b)
df § wf dz1nc wf dz 2nc ·¸
to minimize u 'b1 ¨ u 'b1
db1 ¨ wz nc db wz 2nc db1 ¸¹
© 1 1
[2( z1nc 0.5)(1.39474) 2( z 2nc 0.5)(0.92105)] u 'b1

(10.47c)
subject to g1 1.0 z1nc d 0 (10.47d)
z1nc (b1 2.5) (bc 2.0) 0.5 u z 2c (10.47e)
z1c (b1 2.5) (bc 2.0) 0.4 u z 2c (10.47f)
(2) Discipline 2
Given b2 , bc , z1c (10.48a)
find 'b2 (10.48b)
to minimize u 'b2 ¨ u 'b2 (10.48c)
db2 ¨ wz nc db wz 2nc db2 ¸¹
© 1 2
[2( z1nc 0.5)(0.65790) 2( z 2nc 0.5) u 1.36842] u 'b2

(10.48d)
subject to g2 1.0 z 2nc d 0 (10.48e)
z 2nc (b2 3.0) (bc 2.0) 0.7 u z1c (10.48f)
z 2c (b2 3.0) (bc 2.0) 0.6 u z1c (10.48g)

Coupling variables z1c and z 2c are considered as constants. Using

Equation 10.44, the equation for GSE/OS is made for the optimization
problem in the upper level as follows:
ª dz1nc º
« »
« dbc »
ª1 0 0 0.5 1 0 º « dz1c » ª 1 º
«0 1 0 0.4 1 0 »» «« dbc »» « 1 »
« « »
«0 0.7 1 0 0 1» « dz 2nc » « 1 »
« »« » « »
« 0 0.6 0 1 0 1» « dbc » « 1 » (10.49)
«0 wb1 c
« wb1 »
0 0 c 1 0 » « dz 2 »
« wz 2 »« » « wbc »
« wb » « dbc » « wb »
«0 2 0 0 0 1 » « db1 » « 2»
«¬ wz1c »¼ « dbc » «¬ wbc »¼
« db »
« 2»
«¬ dbc »¼
(3) Upper level
Given bc (10.50a)
find 'bc (10.50b)
to minimize u 'bc ¨ u 'b c (10.50c)
dbc ¨ wz nc dbc wz nc dbc ¸
© 1 2 ¹
The local variables are obtained from Equations 10.47 and 10.48, and the
common variables are determined from Equation 10.50. The solution converges to
b1 3.02503, b2 3.75753, bc 3.02503 and f 0.501046.
10.4.6 Collaborative Optimization (CO)
Collaborative optimization (CO) is a multilevel method, which does not perform

the system analysis and decomposes disciplines. It has two levels: the upper level
that manages the overall process and the lower level that manipulates the design
for each discipline. The system analysis is replaced by adopting complimentary
variables and compatibility conditions. The upper level minimizes the objective
function with the results from the lower level and determines the target values for
disciplines in the lower level. The lower level tries to find a design to satisfy the
constraints and the target values from the upper level.
The first step of CO is to define the complimentary variables. This can vary
according to the user. Here, they are defined for local variables (b1 and b 2 ),
common design variables (b c ) and coupling variables as follows:
b1 s1 , b 2 s 2 , bc s c , z1c s1c , z c2 s c2 (10.51)
The left terms of Equation 10.51 are design variable vectors of the lower level or
the coupling variable vectors. The right terms are vectors determined in the upper
level and used as target values in the lower level.
The problem in Equation 10.4 is modified for each discipline as follows:
(1) Discipline 1
Given s1 , s c , s1c , s c2 (10.52a)
find b1 , b c , z c2 (10.52b)
to minimize c1 || b1 s1 || 2 || b c s c || 2 || z 1c s1c || 2 || z c2 s c2 ||
(10.52c)
subject to g1 (b1 , b c , z1 ) d 0 (10.52d)
z1nc h11 (b1 , b c , z c2 ) (10.52e)
z1c h12 (b1 , b c , z c2 ) (10.52f)
(2) Discipline 2
Given s 2 , s c , s1c , s c2 (10.53a)
find b 2 , b c , z1c (10.53b)
to minimize c 2 || b 2 s 2 || 2 || b c s c || 2 || z 1c s1c || 2 || z c2 s c2 ||
(10.53c)
subject to g 2 (b 2 , b c , z 2 ) d 0 (10.53d)
z 2nc h 21 (b 2 , b c , z1c ) (10.53e)
z c2 h 22 (b 2 , b c , z1c ) (10.53f)
where || || means the l 2 norm, which generally represents the magnitude of a

vector.
Each discipline receives target values from the upper level. In the first cycle,
the targets are arbitrarily defined and the optimization for each discipline of the
lower level is initiated. The objective function of a discipline is the difference

between the target values and the local design variables (state variables and design
variables). CO can only consider local constraints. It is noted that the common
and coupling variables are considered as design variables in all the disciplines.
The following formulation is utilized in the upper level:
Find s1 , s 2 , s c , s1c , s c2 (10.54a)
to minimize f (s1 , s 2 , s c , z1nc , z1c , z nc c

2 ,z2) (10.54b)
dc1
subject to c1 c10 's 0 (10.54c)
ds
dc2
c2 c20 's 0 (10.54d)
ds
z1nc h11 (s1 , s c , s c2 ) (10.54e)
z1c h12 (s1 , s c , s c2 ) (10.54f)
z 2nc h 21 (s 2 , s c , s1c ) (10.54g)
z c2 h 22 (s 2 , s c , s1c ) (10.54h)
The sensitivity of the objective function in the upper level can be obtained
analytically or by the finite difference method as follows:
df 'f f (s 's) f (s 's)

# , s {s1 , s 2 , s c , s1c , s c2 } (10.55)
d s 's 2's
The equality constraints in Equations 10.54c and 10.54d are objective functions
of the disciplines. The derivative in Equations 10.54c and 10.54d is the optimum
sensitivity. Using Equation 10.21b, the optimum sensitivity is
dc j wc j wg j
Ȝ jT ( j 1,2, s {s1 , s 2 , s c , s1c , s c2 }) (10.56)
ds ws ws
Equation 10.56 is investigated. The local constraints are implicit functions

with respect to the target values, therefore, the second term of the right hand side is
zero. Since the first term is a partial derivative with respect to the target values,
dc1
2(d1 s) (d1 {b1 , b c , z1c , z c2 }, s {s1 , s c , s1c , s c2 }) (10.57a)
ds
Find s1 , s 2 , s c , s1c , s c2
s1 , s c , s1c , s c2 s 2 , s c , s1c , s c2
to minimize f (s1 , s 2 , s c , z1c , z1nc , z c2 , z nc
2 )
subject to ci 0 i 1,2
Find b1 , b c , z c2 Find b 2 , b c , z1c

to minimize to minimize
c1 || b1 s1 || 2 || b c s c || 2 c 2 || b 2 s 2 || 2 || b c s c || 2
|| z 1c s1c || 2 || z c2 s c2 || 2 || z 1c s1c || 2 || z c2 s c2 || 2
subject to g1 (b1 , b c , z1 ) d 0 subject to g 2 (b 2 , b c , z 2 ) d 0
z1nc h11 (s1 , s c , s c2 )

z1nc h11 (b1 , b c , z c2 ) z1c h12 (s1 , s c , s c2 ) z nc
2 h 21 (b 2 , b c , z1c )
z1c h12 (b1 , b c , z c2 ) z nc
2 h 21 (s 2 , s c , s1c ) z c2 h 22 (b 2 , b c , z1c )
z c2 h 22 (s 2 , s c , s1c )
Figure 10.13. Flow of CO
dc2
2(d 2 s) (d 2 {b 2 , b c , z1c , z c2 }, s {s 2 , s c , s1c , s c2 }) (10.57b)
ds
As shown in Equation 10.57, the optimum sensitivity is easily obtained. The

overall flow of CO is illustrated in Figure 10.13.
CO is basically an expansion of the IDF method and is a two-level optimization
method. It does not have the system analysis. The objective functions of the
disciplines are closely related to the constraints of the upper level in order to
satisfy MF. Common constraints cannot be handled. CO can be easily applied to
problems where disciplines are obviously distinguishable. Analysis and design are
independently performed in a discipline.
Example 10.9 [Application of CO]

Solve Example 10.2 by using CO.
Solution
In CO, various formulations can be made. The objective functions in disciplines
(c1 and c2 ) are varied according to the selection of complimentary variables. In
this example, the complimentary variables are defined for common design
variables and common coupling variables.
The optimization formulation for each discipline is as follows:
(1) Discipline 1
Given sc , s1c , s 2c (10.58a)
find b1 , bc , z 2c (10.58b)
to minimize c1 (bc sc ) 2 ( z1c s1c ) 2 ( z 2c s 2c ) 2 (10.58c)
z1nc (b1 2.5) (bc 2.0) 0.5 u z 2c (10.58e)
z1c (b1 2.5) (bc 2.0) 0.4 u z 2c (10.58f)
(2) Discipline 2
Given sc , s1c , s 2c (10.59a)
find b2 , bc , z1c (10.59b)
to minimize c 2 (bc s c ) 2 ( z1c s1c ) 2 ( z 2c s 2c ) 2 (10.59c)
subject to g1 1.0 z 2c d 0 (10.59d)
z 2nc (b2 3.0) (bc 2.0) 0.7 u z1c (10.59e)
z 2c (b2 3.0) (bc 2.0) 0.6 u z1c (10.59f)
Each discipline transmits the optimum sensitivities to the upper level.

dc dc1 dc1
They are 1 2 u (bc sc ), 2 u ( z1c s1c ), 2 u ( z 2c s2c ),
dsc ds1c ds2c
dc2 dc2 dc
2 u (bc sc ),
c
2 u ( z1c s1c ) and 2c 2 u ( z 2c s2c ).
dsc ds1 ds2
The optimization formulation for the upper level is as follows:
(3) Upper level
dc1 dc1 dc1 dc dc dc

Given b1, , c , c ,b2 , 2 , 2c , 2c (10.60a)
dsc ds1 ds2 dsc ds1 ds2
find sc , s1c , s 2c (10.60b)
to minimize ( z1nc 0.5) 2 ( z 2nc 0.5) 2 (10.60c)
dc1 dc dc
subject to c1 c10 u 'sc 1c u 's1c 1c u 's2c 0 (10.60d)
dsc ds1 ds 2
dc2 dc dc
c2 c20 u 'sc 2c u 's1c 2c u 's2c 0 (10.60e)
dsc ds1 ds 2
z1nc (b1 2.5) ( sc 2.0) 0.5 u s 2c (10.60f)
z 2nc (b2 3.0) ( s c 2.0) 0.7 u s1c (10.60g)
New target values are generated from Equation 10.60 and transmitted to
the lower levels. The final solutions are b1 6.04083, b2 6.76783,
bc 0.0, s1c 1.10539, s 2c 1.0854 and f 0.49223.
10.4.7 Multidisciplinary Design Optimization Based on

Independent Subspaces (MDOIS)
MDOIS is a single level method, which has a system analysis and decomposes
disciplines. In many MDO problems, the disciplines are physically separated and a
separate analyzer is used for a discipline. Disciplines are coupled only in the
analysis process. An independent design problem is defined for each discipline
(subspace) of such problems. These are relatively simpler problems than a general
MDO problem. MDOIS has been developed to solve such problems efficiently.
The fact that the disciplines are separated means that the following two
conditions are met:
(1) Only local design variables exist for disciplines.
(2) Each discipline has its own objective function and constraints.
Generally, many mathematical problems and the airplane wing design problem do
not satisfy the above conditions. However, many design problems do satisfy the
conditions. A practical example is introduced in the Appendix. To satisfy the
conditions, a problem should not have common design variables, common
objective functions and common constraints. The problem is formulated as
follows:
Find b1, b 2 (10.61a)
to minimize f1 (b1 , z1 ) f 2 (b 2 , z 2 ) (10.61b)
subject to g1 (b1 , z1 ) d 0 (10.61c)

g 2 (b 2 , z 2 ) d 0 (10.61d)
z1 h1 (b1 , z c2 ) (10.61e)
z2 h 2 (b 2 , z1c ) (10.61f)
The system analysis should be performed to solve Equation 10.61. The global
sensitivity information may or may not be needed. The design problem for each
discipline is formulated as follows:
(1) Discipline 1
Given z c2 (10.62a)
find b1 (10.62b)
to minimize f1 (b1 , z1 ) (10.62c)
subject to g1 (b1 , z1 ) d 0 (10.62d)
z1 h1 (b1 , z c2 ) (10.62e)
(2) Discipline 2
Given z1c (10.63a)
find b2 (10.63b)
to minimize f 2 (b 2 , z 2 ) (10.63c)
subject to g 2 (b 2 , z 2 ) d 0 (10.63d)
z2 h 2 (b 2 , z1c ) (10.63e)
Equations 10.62a and 10.63a represent the coupling variables obtained from the
system analysis. If the global sensitivity is available, the coupling variables in
Equations 10.62 and 10.63 can be replaced to consider the changes of the coupling
variables with respect to the changes of the design variables by the following
equations:
dz1c
z1c c
z10 'b 2 (10.64a)
db 2
dz c2
z c2 z c20 'b1 (10.64b)
db1
Initial design variables
System analysis
Sensitivity from the global sensitivity equation
Optimization for discipline 1 Optimization for discipline 2
Design variables
Figure 10.14. Flow of MDOIS
c
where z10 and z c20 are the outputs of the system analysis for z1c and z c2 .
Equation 10.64 may not be used. That is, the coupling variables are considered as
constants in the optimization process of a discipline.
The flow of MDOIS is illustrated in Figure 10.14, which represents the case
where global sensitivities are utilized. MDOIS is a practical method in design
practice. The design proceeds until IF and MF are satisfied. Design is carried out
for each discipline.
Example 10.10 [Application of MDOIS]

Solve Example 10.2 by using MDOIS.
Solution
From the system analysis, the coupling variables ( z1c and z 2c ) are found. The
global sensitivities are not used in this example. The common design variable
vector is considered in discipline 1.
The disciplines have the following formulations:
(1) Discipline 1
Given z 2c (10.65a)
find b1 , bc (10.65b)
to minimize f1 ( z1nc 0.5) 2 (10.65c)
z1nc (b1 2.5) (bc 2.0) 0.5 u z 2c (10.65e)

z1c (b1 2.5) (bc 2.0) 0.4 u z 2c (10.65f)
(2) Discipline 2
Given bc , z1c (10.66a)
find b2 (10.66b)
to minimize f 2 ( z 2nc 0.5) 2 (10.66c)
g2 1.0 z 2nc d 0 (10.66d)
z 2c (b2 3.0) (bc 2.0) 0.6 u z1c (10.66e)
z 2nc (b2 3.0) (bc 2.0) 0.7 u z1c (10.66f)
The results of the disciplines are used by the system analysis. The final solution is
b1 3.02838, b2 3.75068, bc 3.02838 and f 0.50149.
10.4.8 Comparison of MDO Methods
In the previous sections, various MDO methods were introduced. MDO methods
basically solve the nonhierarchical system in Figure 10.2a directly or by
transforming it to the hierarchical system in Figure 10.2b. The problem
formulation and the solution process can be newly defined according to the
characteristics of the problem. Therefore, the methods introduced are several
methods selected out of many methods. There are many other MDO methods.
MDO methods are classified into single level methods and multilevel methods.
MDF, IDF, AAO and MDOIS are single level methods, while CSSO, BLISS and
CO are the multilevel methods. The single level methods are simple. However,
the characteristics of multidisciplines are not exploited in single level methods
such as MDF, IDF and AAO methods. The multilevel methods are complicated to
use.
MDO methods are also classified according to the existence of the system
analysis. The system analysis finds the coupling variables between analyses.
MDF, CSSO, BLISS and MDOIS have the system analysis. They have the
advantage that MF (multidisciplinary feasibility) is always satisfied during the
design process. However, many system analyses are required in MDF. In some
multilevel methods, the results of the system analysis and the global sensitivity in
the upper level are transmitted to disciplines in the lower level. The global
sensitivity is generally fairly expensive to calculate.
IDF and CO have the advantage that they do not need the system analysis. IF
(individual feasibility) is satisfied, but MF is not satisfied during the design
process. Moreover, the analysis cost in each discipline can be considerably high.
To overcome this difficulty, AAO uses evaluators instead of analyzers. This is the
simplest method to use. In the application of AAO, if we have only analyzers, we

have the disadvantage that an analyzer must be transformed to the evaluator.
Neither MF nor IF is satisfied during the design process.
MDO methods handle common constraints and common design variables in
different ways. Single level methods such as MDF, IDF and AAO can handle
them in unified ways. MDOIS cannot have common variables or functions. CSSO
cannot have common constraints and common design variables. Constraints of
other disciplines are considered by the global sensitivity. BLISS can have
common design variables; however, it cannot handle common constraints.
Common design variables are considered in the upper level. The effects of
common design variables in the lower level are transmitted to the upper level via
the optimum sensitivity. Calculation of the optimum sensitivity is quite expensive
except in CO. The characteristics of the methods are compared in Table 10.2. O
means “yes” and u means “no.” In this chapter, all the methods are explained in
order to obtain an exact solution. In design practice, it may not be easy to obtain
an exact optimum solution. In many cases, approximation is appropriately adopted
in the processes of the MDO methods.
Table 10.2. Comparison of MDO methods
Common Common System Global Optimum Satisfaction Satisfaction

Method Level constraints design analysis sensitivity sensitivity of IF of MF
variables
u
MDF single O O O always always
u
u
u
IDF single O O always At optimum
u
u
u
AAO single O O At optimum At optimum
u
u
CSSO multi O O O always always
u
BLISS multi O O O O always always
u
u
u
CO multi O O always At optimum

u
u
u
u
MDOIS single O O always always
Multidisciplinary Design Optimization
607
10.5 Discussion
Multidisciplinary design optimization (MDO) was created to manage design
problems defined throughout multidisciplines. The formulation of MDO can be
relatively easily defined. However, it is very difficult to find a universal method to
solve MDO problems. Efforts are still needed to solve such problems.
The first difficulty in MDO is the coupling aspect between analyses. For
example, fluid mechanics and solid mechanics are coupled in aeroelasticity
analysis for airplane wings. Solving this coupling may not be impossible, but
difficult. Moreover, it is extremely difficult to include it in optimization since it
requires a large number of analyses. Some MDO methods need a complicated
mathematical process such as global sensitivity or optimum sensitivity. Therefore,
MDO methods are rarely applied to aeroelasticity problems.
There are engineering problems that have weak couplings between disciplines.
It is relatively easy to apply an MDO method to such problems. Some problems
need large-scale analysis. Even single disciplinary optimization is not easy to use
in such problems. As analysis tools are developing significantly, many problems
are enormous in size. Therefore, the exact process of the introduced MDO
methods may not be easy to use in practical design.
Although there are no coupled aspects between analyses, multiple analyzers
can be utilized in some problems. Recently, such problems have been considered
as MDO problems. Also a problem that has common design variables or common
constraints can be regarded as an MDO problem.
Due to the above difficulties, a different MDO method may be utilized for
different applications. Sometimes, a combination of methods is used or a new
method is defined using the ideas of existing methods. Approximation is
frequently adopted to solve complicated problems. The response surface method
introduced in Chapter 6 is frequently employed especially when analysis is
complicated. An example is introduced in the Appendix.
MDO is used to solve large-scale design problems. There is still no universal
method and, therefore, more research is anticipated in this area.
10.A Application of an MDO Algorithm to the Design of

a Belt Integrated Seat Taking Crashworthiness into
Consideration
10.A.1 Problem Description

An MDO problem is defined for the design of a belt integrated seat (BIS) taking
crashworthiness into consideration. The crash model consists of an airbag, a BIS,
an energy absorbing steering system and a safety belt. As illustrated in Figure
10.A.1, a safety belt is attached to the frame of the BIS. In an ordinary safety belt
system, one end is attached to the vehicle structure. Out of position (OOP)
situations occur according to the occupant position. Thus, an ordinary seat belt
may not contact the occupant body in OOP. However, because the safety belt is
attached to the frame of the BIS, safety performance of the belt does not change
according to the position of the occupant. The BIS can reduce the injury of the
occupant by keeping the occupant in position. Moreover, the structure of the BIS
absorbs the impact energy during a vehicle crash. In an ordinary safety belt system,
the safety belt load is transferred to the vehicle. However, in the BIS system, the
safety belt load is transferred to the seat frame. Thus, the BIS is normally heavier
than ordinary seats for structural stability and needs to be lightened while
maintaining a good safety performance.
This example was introduced in Section 9.7 where the seat structure was
optimized based on the external forces. The structure had a linear behavior. This
section accounts for nonlinearity of the structure and the design of the seat is
carried out by using MDO.
Figure 10.A.1. Belt integrated seat with an attached belt

10.A.2 The Interdisciplinary Relationship Between the

Disciplines Involved
In order to design a BIS, we need two types of analyses: nonlinear structural
analysis and occupant analysis. In the nonlinear structural analysis discipline, the
weight of the BIS is minimized while the constraints of the structural performance
are satisfied. In the occupant analysis discipline, the HIC value is reduced. The
belt load is the input for the structural analysis and the stiffness of the seat back
frame is the input for the occupant analysis.
Meanwhile, a design change in the structural analysis discipline causes a
change in the stiffness of the seat back frame. Again, the stiffness of the seat back
frame is the input for the occupant analysis. Thus, a design change in the
structural analysis discipline results in a change of the input for the occupant
analysis discipline. Moreover, a design change in the occupant analysis discipline
causes a change in the belt load, which is the input for the structural analysis
discipline. Thus, a design change in occupant analysis results in a change in the
input for the structural analysis discipline. Therefore, there is a definite
interdisciplinary relationship between the two disciplines, and the relationship can
be identified by interdisciplinary variables: the stiffness of the seat back frame and
belt load as illustrated in Figure 10.A.2. This interdisciplinary relationship makes
the design process complicated.
10.A.3 Formulation for the Design of a BIS Taking

Crashworthiness into Consideration
The formulation of the design of a BIS taking crashworthiness into consideration is

shown in Equation 10.A.1. The formulation has adopted a multiobjective function
since both the weight of the BIS and the HIC value are to be minimized at the
same time.
Find b0 ,bs (10.A.1a)
to minimize F D 1W D 2 HIC (10.A.1b)
Stiffness
Structure Occupant
Belt load
Weight HIC
Displacement Acceleration of the chest
Figure 10.A.2. Dependency diagram of a BIS design taking crashworthiness into

consideration
x
S3 S4
S1
S5 S6
y
z S7
S2
S9
S8
S10
Figure 10.A.3. Energy absorbing steering system and its model for occupant analysis
subject to g1 'd anchor d 200 mm (10.A.1c)
g2 'd d 120 mm (10.A.1d)
g3 HIC d HIC 0 (10.A.1e)

0
g4 Gchest d Gchest (10.A.1f)
where b 0 is a design variable vector in the occupant analysis discipline whose

elements are the scale factors of a predefined force–displacement curve profile for
a seat belt, an airbag mass flow, and an energy absorbing steering column. b s is a
design variable vector in the structural analysis discipline whose elements are the
thicknesses of various parts of the seat. 'd anchor is the displacement at the belt
anchor point. 'd is the relative displacement between the column tube-in and
tube-out of the energy absorbing steering system as shown Figure 10.A.3. Gchest
is the resultant acceleration of the dummy chest. HIC is an injury to the head and
is introduced in Equation 9.1. Equation 10.A.1 does not show the analysis part
that needs the system analysis.
In Equation 10.A.1, the design variable vectors, the constraints and the
objective functions are clearly divided into two disciplines. However, it is not
easy to solve the optimization problem in Equation 10.A.1 with a single
optimization method. This is because the weighting factors for the objective
functions are not logically determined and a method to solve the coupling process
should be used. MDOIS in Section 10.4.7 is utilized to solve the problem.
10.A.4 Application of MDOIS to the Design of a BIS
Remember that MDOIS is only applicable if the objective function, design

variables, and constraints of one discipline are independent of the other disciplines.
The optimization problem formulated in Equation 10.A.1 meets this condition.
Thus, MDOIS is applicable to the optimization of the BIS.
System Analysis for Determination of Interdisciplinary Variables in a BIS

Design
The system analysis is the process of solving a system of nonlinear equations for
interdisciplinary (coupled) variables. In many MDO algorithms that have the
system analysis step, the system analysis adopts a fixed-point iteration method or
the Newton–Raphson method to solve a system of nonlinear equations. However,
in many practical engineering problems, this approach is almost unusable because
most problems cannot even be expressed in the rigorous form of a system of
equations.
In the present example, the interdisciplinary variables between the two
disciplines are the belt load and the stiffness (or force–displacement curve profile)
of the seat back frame. These interdisciplinary variables cannot be determined by
a fixed-point iteration method or by the Newton–Raphson method because the
interdisciplinary relationship in the present example cannot be expressed in an
explicit form of a system of equations. Thus, we need another approach to find the
solution of the system analysis. In this example, since structural analysis is more
costly than occupant analysis, a smaller number of structural analyses is desired.
Figure 10.A.4 illustrates the system analysis algorithm for determination of the belt
load and the force–displacement curve profile of the seat back frame.
Start
Perform occupant analysis
Perform structural analysis
Given GS
find force–displacement (f–d) curve profile
Update of the seat back frame
f–d curve T
S 0 2
to minimize E ³ [G (t ) G (t )] dt
0
subject to the results of occupant analysis
No Converged?
(f–d curve)
Yes
Stop
Figure 10.A.4. Flow chart for the system analysis in MDOIS applied to a BIS design
taking crashworthiness into consideration
In Figure 10.A.4, a target G S is given, which is the displacement of the seat

belt anchor point on the seat by using structural analysis. The force–displacement
curve is determined for the anchor point to be the target. The design variable is the
scale factor of the curve. This has been explained in Section 9.3. The steps are as
follows:
Step 1. Initialize the stiffness of the seat back frame at the initial design point.
Step 2. Find the belt load to be applied to the belt anchor point for structural
analysis by using occupant analysis.
Step 3. Find the displacement of the belt anchor point (G S (t )) by structural
analysis as illustrated in Figure 10.A.5b. The load condition for structural
analysis is obtained in Step 2.
Step 4. Determine the force–displacement curve profile of the seat back
frame by matching the displacement of the belt anchor point from occupant
(a) Model of the occupant analysis based on multibody dynamics
(b) Model of the structural analysis based on nonlinear structural analysis
Figure 10.A.5. Models of the BIS for system analysis and discipline optimization
analysis (G 0 (t )) to the displacement of the belt anchor point (G S (t ))

calculated in Step 3 as in Equation 10.A.2.
Find scale factors of the predefined force–displacement curves

(10.A.2a)
T
S 0 2
to minimize E ³ [G (t ) G (t )] dt (10.A.2b)
0
subject to results of occupant analysis (10.A.2c)
Step 5. If the termination criterion is satisfied, the process is terminated.

Otherwise, update the design variables, and go to Step 2.
Final outputs of the above steps are the belt load calculated in Step 2 and the
force–displacement curve profile obtained in Step 4 at the last iteration. Note that
costly structural analysis is performed just once during one iteration of Step 2–Step
5 while inexpensive occupant analysis is performed many times during one
iteration of Step 2–Step 5. Thus, overall cost for the system analysis is reduced.
Here is an example of the system analysis that is performed in the first cycle.
The optimization process of Step 4 is carried out by using the RSAO method
introduced in Section 9.8. The system analysis in the first cycle is converged in
five iterations. Accordingly, the determined force–displacement curve profile and
the belt load are illustrated in Figures 10.A.6 and 10A.7. The determined force–
displacement curve profile of Figure 10.A.6 is transferred to the occupant analysis
discipline and is utilized as the stiffness of the seat back frame. Moreover, the belt
load illustrated in Figure 10.A.7 is transferred to the structural analysis discipline
and is applied to the belt anchor point. With the transferred interdisciplinary
variables, each discipline independently determines its own design variables. Note
again that the interdisciplinary variables are treated as constant parameters in
subsequent disciplinary designs.
Independent Optimizations of Structural and Occupant Disciplines

Each discipline receives the interdisciplinary variables from the system analysis
and performs design in its own way with the interdisciplinary variables held
constant. Structural design can be formulated as follows:
Find b S for a given belt load (10.A.3a)
subject to g 'd anchor d 200 mm (10.A.3c)
where b S is a design variable vector whose elements are the thicknesses of

various parts of the seat and 'd anchor is the displacement of the anchor point.
ͤͦ
ͤ͡
ͣͦ
Force (kN)
ͣ͡
GZ
Iteration 3
G[
Iteration 4
ͦ͢
͢͡
͡
͡ ͦ ͢͡ ͦ͢ ͣ͡ ͣͦ ͤ͡
Displacement (mm)
Figure 10.A.6. Determined force–displacement curve in the system analysis of MDOIS
6
5
4
Force (kN)
3
2
1
0
0 50 100 150
Time (ms)
Figure 10.A.7. Determined belt load in the system analysis at the first cycle of MDOIS
The occupant discipline design can be formulated as follows:
Find b 0 for a predefined force–displacement curve for the seat

back frame (10.A.4a)
to minimize HIC (10.A.4b)
subject to g1 'd d 120 mm (10.A.4c)
g2 HIC d HIC 0 (10.A.4d)

0
g3 Gchest d Gchest (10.A.4e)
Objective function
1.2 Structure
1 Occup ant
function
0.8
Objective
0.6
Objective
0.4
0.2
0
0 1 2 3 4 5 6
Iteration
Figure 10.A.8. History of each objective function
where b 0 is a design variable vector whose elements are the scale factors of the
predefined force–displacement curve profiles for a seat belt, an airbag mass flow,
and an energy absorbing steering column.
Because the results of the involved analyses are very noisy, local sensitivity
information is not as meaningful as that of the smooth response problem. Thus, it
is quite difficult to apply a gradient-based optimization technique to raw responses
of nonlinear structural analysis and occupant analysis. DOE or some nongradient-
8000
7000
6000
5000
No. 1
Load
4000
No. 2
3000 No. 3
2000 No. 4
1000 No. 5
0
0 0.5 1 1.5 2
Displacement
Figure 10.A.9. Force–displacement curve profile history of the seat back through the entire
cycles of MDOIS
Design variable value

6 DV1 DV2 DV3
DV4 DV5 DV6

5
2
0 1 2 3 4 5 6
Iteration
Figure 10.A.10. Design variable history of the structural discipline
based optimization techniques could be candidates for the design tool for the
present example. We chose the RSAO algorithm in Section 9.8.2 to apply the
gradient-based optimization technique to the present example. The reason is that
gradient-based optimization is very fast once the problem is well formulated and
the responses are smooth enough to be differentiated.
Each discipline is independently designed after the interdisciplinary
relationship is solved by the system analysis. The solution is converged in six
cycles. Figure 10.A.8 is a design history of the objective functions for each
discipline. They are normalized by the initial values because the orders of the
1.2
1
Design variable value
0.8
0.6 DV1 DV2

0.4 DV3 DV4
0.2 DV5
0
0 1 2 3 4 5 6
Iteration
Figure 10.A.11. Design variable history of the occupant discipline

Table 10.A.1. Details of the design variables in the structural discipline and the occupant
discipline
Structural discipline Occupant discipline

Thickness of the lower right Scale factor for the force-displacement
DV1
recliner curve profile of the shear capsule
Thickness of the upper right Scale factor for the force-displacement
DV2
recliner curve profile of the ball sleeve
Scale factor for the force-displacement
Thickness of the upper left
DV3 curve profile of the energy absorbing
recliner
steering column
Thickness of the lower left Scale factor for the mass-flow rate of
DV4
recliner the airbag
Scale factor for the force-displacement
DV5 Thickness of the left back frame
curve profile of the seat belt
DV6 Thickness of the right back frame
objective functions in the two disciplines are quite different. As illustrated in

Figure 10.A.8, the objective functions are decreased. Figure 10.A.9 shows design
histories of the force–displacement profiles of the seat back frame. They are
determined in the system analysis, and transferred to the occupant discipline as
design parameters. Moreover, they are used for the convergence criterion. When
they are not changed, the process is regarded as converging. Figures 10.A.10 and
10A.11 are the design histories of the designs. Details of the design variables are
shown in Table 10.A.1. In Figure 10.A.11 all the design variables in the occupant
discipline are converged to the lower bounds, where the initial values are scaled to
1. In other words, the initial design of the occupant safety system has a fairly high
stiffness. From the results, both the weight and HIC of the BIS are reduced.
Summary
One of the MDO methods, MDOIS is applied to the design of the belt integrated
seat. In the original MDOIS, all the processes are mathematically defined by using
the system analysis, sensitivity analysis, etc. However, appropriate approximation
methods can be used in design practice.
References
ANSYS User’s Manual Version 8.1 (2004) ANSYS Inc., Canonsburg, PA
Balling RJ, Sobieszczanski-Sobieski J (1996) Optimization of Coupled Systems: A
Critical Overview of Approach. AIAA Journal 34(1):6í17
Barthelemy JF, Sobieszczanski-Sobieski J (1983) Optimum Sensitivity Derivatives
of Objective Function in Nonlinear Programming. AIAA Journal 21:913í915
Braun RD (1996) Collaborative Optimization: An Architecture for Large-scale

Distributed Design. Ph.D. Dissertation, Stanford University
Cramer EJ, Dennis JE Jr., Frank PD, Lewis RM, Shubin GR (1993) Problem
Formulation for Multidisciplinary Optimization. Center for Research on Parallel
Computation, Rice Univ. Report CRPC-TR93334
Goldberg DE (1989) Genetic Algorithm in Search, Optimization and Machine
Learning. Addison-Wesley, New York
Hajela P, Bloebaum CL, Sobieszczanski-Sobieski J (1990) Application of Global
Sensitivity Equations in Multidisciplinary Aircraft Synthesis. Journal of
Aircraft 27(12):1002–1010
iSIGHT User’s Guide Version 7.1 (2004) Engineous Software Inc., Cary, NC
ModelCenter User’s Manual Version 8.1 (2004) Phoenix-Int. Inc., Blacksburg,
VA
Optimus User’s Manual Version 4.0 (2003) LMS Int., Leuven, Belgium
Pham DT, Karaboga D (2000) Intelligent Optimization Techniques. Springer,
Renaud JE, Gabriele GA (1994) Approximation in Nonhierarchic System
Optimization. AIAA Journal 32(1):198í205
Sellar RS, Renaud JE, Batille SM (1994) Optimization of Mixed
Discrete/Continuous Design Variables System Using Neural Networks. 5th
AIAA/NASA/USAF/ISSMO Symposium on Multidisciplinary Analysis and
Optimization, Panama City, FL
Shin MK, Park GJ (2005) Multidisciplinary Design Optimization Based on
Independent Subspaces. International Journal for Numerical Methods in
Sobieski IP, Kroo I (1996) Collaborative Optimization Applied to an Aircraft
Design Problem. Proceedings of AIAA Aerospace Sciences Meeting, AIAA-
96-0715
Sobieszczanski-Sobieski J (1982) A Linear Decomposition Method for Large
Optimization Problems–Blueprint for Development. NASA TM 83248
Sobieszczanski-Sobieski J (1988) Optimization by Decomposition: A Step from
Hierarchic to Nonhierarchic Systems. Recent Advances in Multidisciplinary
Analysis and Optimization. NASA TM 101494
Sobieszczanski-Sobieski J (1990a) On the Sensitivity of Complex, Internally
Coupled Systems. AIAA Journal 28(1):153í160
Sobieszczanski-Sobieski J (1990b) Sensitivity Analysis and Multidisciplinary
Design Optimization for Aircraft Design: Recent Advances and Results. Journal
of Aircraft 27(12):993í1001
Sobieszczanski-Sobieski J, Agte J, Sandusky R Jr. (1998) Bilevel Integrated
System Synthesis. Proceedings of AIAA/USAF/NASA/ISSMO Symposium on
Multidisciplinary Analysis and Optimization, AIAA-98-4916
Sobieszczanski-Sobieski J, Barthelemy JF, Riley KM (1982) Sensitivity of
Optimum Solutions to Problem Parameters. AIAA Journal 20(9):1291í1299
Sobieszczanski-Sobieski J, Haftka RT (1996) Multidisciplinary Aerospace Design
Optimization: Survey of Recent Developments. AIAA Paper 96í0711
Team Midas (1995) Multidisciplinary Integrated Design Assistant for Spacecraft

(MIDAS). Jet Propulsion Lab., Pasadena, CA
Townsend JC, Weston RP, Eidson TM (1993) A Programming Environment for
Distributed Complex Computing: An Overview of the Framework for
Interdisciplinary Design Optimization (FIDO) Project. NASA TM 109058
Design. McGraw-Hill, New York
VisualDOC User’s Manual Version 4.0 (2004) Vanderplaats Research &
Index
A
ABAQUS .............................................458 axiomatic design ................................7, 17
acceleration constraint..........................285 axiomatic design (related to robust design)
active constraint ...........................124, 148 .........................................................428
active set strategy .................................147
additive statistical model..............317, 331 B
additivity backward difference method ................201
- axiomatic design..............................33 barrier function method........................153
- design of experiments (DOE) ........329 basic feasible solution ..........................131
adjoint variable method................206, 218 basic solution .......................................131
admissible function ..............................275 beam adjuster .........................................82
airbag ...................................................502 beam element .......................................176
all-at-once (AAO) ................................582 belt integrated seat (application of MDO)..
all-in-one ..............................................579 .........................................................609
aluminum space frame (ASF) ..............474 belt integrated seat (BIS)......................530
analysis of mean (ANOM) ...........329, 405 bilevel integrated system synthesis
analysis of variance (ANOVA) ............312 (BLISS)............................................593
ANSYS ..................................................97 Brachistochrone problem .....................107
antisynergistic interaction ............320, 340 Broyden–Fletcher–Goldfarb–Shanno
artificial variable .................. 227, 256, 302 (BFGS) Method ...............................143
approximation
- force approximation ......................224 C
- global approximation in dynamic canonical form .....................................132
response.......................................276 central difference method.....................201
- Hessian matrix.......................144, 152 characteristic function
- inverse Hessian matrix ..................143 - design of experiments (DOE) .......311,
- linear approximation..... 146, 166, 188, 312, 331, 351
223, 585 - robust design .........................407, 413
- local approximation in dynamic chest injury...........................................501
response.......................................277 chest severity index (CSI) ............501, 536
- quadratic approximation................166 collaborative optimization (CO)...........597
- response surface method................360 common constraint ...............................567
- structural optimization...................223 common design variable ......................566
augmented Lagrangian method ...154, 263, common objective function..................566
274 common range........................................36
automobile body...........................498, 549 compatibility condition ................580, 597
automobile door ...................................513 complexity .............................................33
axiom .....................................................17 complimentary variable................581, 597
622 Index
conceptual design.....................4, 108, 519 - multidisciplinary design optimization

concurrent subspace optimization (CSSO) (MDO).........................................563
.........................................................585 - variable..........................................564
conditional probability ...........................35 crash .....................................................497
configuration optimization .....................10 crashworthiness............................545, 609
confirmation experiment ..............331, 341 creative design ...................................5, 52
conjugate gradient method ...................141 cumulative constraint ...........................587
constrained optimization customer attribute(s) (CA(s)) ...........17, 65
- augmented Lagrangian method ....154,
263, 274 D
- barrier function method .................153 damping matrix ....................................502
- direct method.........................145, 222 Davidon–Fletcher–Powell (DFP) method ..
- exterior penalty function method..153, .........................................................142
263 decision-making process ..........................2
- indirect method..............................145 decomposition
- interior penalty function method ...153 - axiomatic design..............................23
- optimality criteria method .............145 - multidisciplinary design optimization
- penalty function method................153 (MDO).................................567, 571
- primal method .......................145, 159 decoupled
- recursive quadratic programming - axiomatic design......................19, 521
(RQP) ..................................145, 150 - multidisciplinary design optimization
- sequential linear programming (SLP).. (MDO).........................................565
.............................................145, 149 decoupling........................................76, 77
- sequential quadratic programming degenerate basic feasible solution ........131
(SQP)...................................145, 150 degenerate basic solution .....................131
- sequential unconstrained minimization design case studies
technique (SUMT)................145, 153 - application of an MDO algorithm to
- transformation method...................153 the design of a belt integrated seat
constrained quasi-Newton method .......152 taking crashworthiness into
constraint..............................................107 consideration ..............................609
constraint function................................110 - automotive door design with the
constraints (Cs) ......................................19 ULSAB concept using structural
continuous design.................................172 optimization.................................244
control factor................................395, 406 - axiomatic design of a beam adjuster
convergence criterion for a laser marker...........................82
- a method using equivalent loads....281 - beverage can design ........................33
- optimization...................................109 - bottle–can opener ............................32
- sequential algorithm with orthogonal - crash analysis and design of a belt
arrays (SOA) ...............................358 integrated seat for occupant safety
convex .....................................................530
- function ......................... 120, 166, 167 - design of a motor driven tilt/telescopic
- hull ................................................131 steering system ............................519
- sufficient condition for a convex - design of a refrigerator door ............20
problem .......................................120 - design of a side impact beam of an
coordination optimization problem (COP) automobile door...........................513
.........................................................585 - design of a spacer grid spring........455
coupled analysis ...................................565 - design of a strip casting roll ..........444
coupled design .......................................25 - design of a water faucet...................22
coupling - design of an automobile rearview
- axiomatic design..............................25 mirror ..........................................464
- in analysis......................................564 - design of an airbag using orthogonal
Index 623
arrays...........................................502 dynamic load..................................11, 255

- design of an energy absorbing steering dynamic response optimization......11, 255
system using orthogonal arrays ...504
- determination of a crash pulse and E
optimization of crash components eigenvalue ............................................179
using response surface approximate eigenvector...................................179, 268
optimization.................................545 element design .........................................4
- development of a design system for a emergency locking retractor (ELR)......535
TV glass bulb ................................90 energy absorbing steering system ........504
- development of a design system for engineering design ...........................1, 108
the EPS cushioning package of a equal interval search.............................134
monitor .........................................99 equality constraint............ 8, 108, 580, 599
- example of analysis of existing equivalent
designs: liquid crystal display holder - functional.......................................257
.......................................................56 - inertia load.....................................280
- example of a creative design: - static load ......................................279
refrigerator design .........................53 estimated value.....................................331
- example of design improvement: evaluator...............................................582
parking mode of an automatic expectation value .................................397
transmission...................................60 explicit function ................... 277, 310, 419
- optimization of cross sections of a exterior penalty function method .153, 263
robot arm .....................................290 extreme point ...............................117, 131
- structural analysis and optimization of İ-active set strategy ..............................147
a low speed vehicle body ............474
density method .....................................237 F
descent condition..........................124, 139 factor ....................................................311
descent direction ..........................124, 139 factorial design...............................12, 312
design domain ..............................229, 248 factors in experiments ....................12, 311
design equation ......................................19 F-distribution .......................................384
design matrix..........................................19 feasibility .............................................147
design of experiments (DOE).........12, 309 feasible region
design optimization ..........................8, 107 - axiomatic design..............................39
design parameter (DP) domain.........39, 41 - optimization...................................116
design parameter (DP) range..................40 feasible set............................................120
design parameter(s) (DP(s)) ...................17 feasible solution ...................................131
design range .....................................30, 36 FEM software
design using orthogonal arrays (DOA) 336 - ABAQUS ......................................458
design variable linking .........................226 - ANSYS............................................97
detailed design ................. 4, 108, 443, 519 - I-DEAS..........................................458
deviation function ................................426 - LS/DYNA3D...........................99, 514
Dirac delta function......................273, 431 Fibonacci sequence ..............................135
direct differentiation method........205, 265 finite difference method .......................200
direct method ...............................145, 221 finite element analysis..................174, 498
direction vector ............................123, 171 finite element method (FEM)...9, 172, 274
discipline ..............................................561 561
discrete design..............................172, 336 fixed point iteration method .................564
domain .................................................172 flexible multibody dynamic system .....286
DOT .....................................................427 flexural rigidity ....................................485
dummy .................................................500 FMVSS 203 .................................505, 524
dynamic analysis..........................276, 499 FMVSS 208 .........................................532
dynamic factor .....................................255 force approximation .............................224
624 Index
forward difference method ...................201 independence..........................................25

fractional design...................................326 Independence Axiom ..................8, 17 , 19
free variable .........................................127 independent module ...............................64
frequency dependent constraint............293 indirect method ....................................145
frequency domain.................................292 individual discipline feasible (IDF)......580
frequency response optimization..........292 individual feasibility (IF) .....................565
frontal crash analysis............................538 inequality constraint.................8, 108, 188
functional domain ..........................18, 520 inertia force ..........................................280
functional requirement (FR) domain......41 inflator..................................................502
functional requirement (FR) range .........42 information.............................................33
functional requirement(s) (FR(s)) .....7, 17, Information Axiom ......................8, 18, 33
519 information content ..........................18, 33
inner array ....................................406, 474
G insensitive design ...........................13, 418
GENESIS ............................. 244, 467, 527 interaction ............................ 312, 320, 340
genetic algorithm....................15, 337, 562 interdisciplinary relationship................610
glass bulb ...............................................90 interior penalty function method..........153
global intermediate variable............................278
- approximation in dynamic response inverse barrier function ........................153
.....................................................276 iSIGHT.................................................562
- convergence...................................124 iso-objective contour............................116
- minimum .......................................167
- sensitivity ......................................585 K
- sensitivity equation (GSE).............571 Karush–Kuhn–Tucker (KKT) condition ...
global sensitivity equation/optimized .................................................119, 575
subsystem (GSE/OS) .......................594 Kreisselmeier–Steinhauser (KS) function .
golden section ratio ..............................135 .........................................................585
golden section search ...........................135 Ȥ2-distribution.......................................383
gradient ........................................117, 164
gradient vector ..................... 117, 165, 265 L
graphical method....................................38 Lagrange multiplier.............. 107, 119, 575
graphical optimization..........................116 Lagrangian function .....................119, 152
grid dimple ...........................................456 Lagrangian method ..............................154
grid spring ............................................456 laser marker............................................82
least squares method .............. 12, 364, 545
H line search ............................................150
harmonic (response) analysis .......293, 528 linear approximation .... 166, 188, 223, 585
head injury ...........................................276 linear decomposition ............................571
head injury criteria (HIC).............500, 610 linear programming (LP) method.........125
headrest ................................................530 linear static problem.................................9
Hessian matrix..............................118, 165 local
homogenization method .......................237 - approximation in dynamic response
.....................................................277
I - constraint .......................................599
I-DEAS ................................................458 - design variable...............................564
imitative design ........................................5 - minimum ...............................117, 164
impact...................................................501 - objective function..........................566
impact beam .........................................513 log barrier function...............................153
impact energy.......................................504 loss function ........................... 13, 401, 429
implicit function...................................599 low speed vehicle (LSV)......................474
implicit method ....................................502 lower bound .........................................108
Index 625
LS/DYNA3D .................................99, 514 Administration (NHTSA) ................475

LS/INGRID............................................99 neighborhood electric vehicle (NEV) ..476
l2 norm..................................................598 nested analysis and design (NAND) ....563
Newmark method.................................266
M noise factor...................................395, 401
MADYMO ...........................................500 nominal design .......................................13
mass matrix .................... 11, 179, 256, 502 nominal value...............................399, 418
material derivative................................232 nonhierarchical system.........................605
matrix experiment ........................358, 359 nonlinear analysis.........................498, 541
mean..................................... 312, 382, 394 nonlinear dynamic analysis..................541
mean square .........................................314 nonlinear finite element analysis..........498
mean standard deviation (MSD) ..........413 nonlinear programming (NLP) method 122
metamodel ...................................311, 361 normal distribution............... 312, 383, 422
modal analysis..............................289, 478 nuclear fuel rod ....................................456
modal matrix ................................267, 296
mode superposition method .........267, 295 O
ModelCenter ........................................562 objective function.........................107, 171
monocoque ...........................................475 occupant analysis .........................505, 612
motor driven tilt/telescopic steering system occupant injury
.........................................................519 - chest severity index (CSI)......501, 536
move limit ....................................150, 277 - head injury criteria (HIC) ......500, 610
multidisciplinary analysis (MDA)........561 occupant safety.....................................530
multidisciplinary design optimization one-way table ...............................329, 405
(MDO) .......................................13, 561 one-dimensional minimization
multidisciplinary design optimization - equal interval search......................134
(MDO) formulation .........................566 - golden section search.....................135
multidisciplinary design optimization optimality criteria method ....................145
(MDO) method optimization ..................... 8, 107, 171, 561
- all-at-once (AAO)..........................582 optimization formulation......................109
- bilevel integrated system synthesis optimization software
(BLISS) .......................................593 - DOT ..............................................427
- collaborative optimization (CO)....597 - GENESIS ...................... 244, 462, 527
- concurrent subspace optimization - iSIGHT..........................................562
(CSSO) ........................................585 - ModelCenter..................................562
- individual discipline feasible (IDF)..... - Optimus .........................................562
.....................................................580 - VisualDOC ............................545, 562
- multidisciplinary feasible (MDF) ..579 optimum sensitivity..............................575
- multidisciplinary design optimization Optimus................................................562
based on independent subspaces orthogonal array ...........................311, 386
(MDOIS) .....................................602 outer array ....................................406, 473
multidisciplinary design optimization
based on independent subspaces P
(MDOIS)..........................................602 parameter design ..................................404
multidisciplinary feasibility (MF) ........564 penalty function ................... 153, 351, 407
multidisciplinary feasible (MDF).........579 penalty function method.......................153
multiobjective function ................426, 502 penalty parameter.........................153, 263
multiple loading conditions..........226, 301 physical domain .....................................18
multiplier method.................................154 physical integration................................32
pointwise constraint .............................258
N population mean................... 315, 385, 395
National Highway Traffic Safety positive definite............................118, 166
626 Index
positive semidefinite ....................118, 166 equations .....................................205

preliminary design....................................4 - analytic sensitivity analysis using
pressurized light water reactor (PWR) .455 continuum equations ...................216
primal method ..............................149, 145 - direct differentiation method .........265
probability density function ...........35, 422 - direct integration method...............265
probability of success.............................33 - dynamic response optimization .....265
production variable(s) (PV(s)) ...............17 - finite difference method ................200
Pshenichny's method ............................151 - mode superposition method...........267
- numerical method ..........................125
Q - semianalytic method......................214
quadratic approximation.......................166 sensitivity index (SI) ............................419
quadratic loss function .................153, 401 sequential algorithm with orthogonal
quadratic programming (QP) ...............150 arrays (SOA)....................................357
quadratic weighting function................431 sequential linear programming (SLP) ..145
quasi-Newton method ..........................142 149
sequential quadratic programming (SQP) ..
R .................................................145, 150
random variable ...................................382 sequential unconstrained minimization
reangularity ............................................25 technique (SUMT) ...................145, 153
rear end impact.............................552, 556 shape design parameterization .............231
rearview mirror ....................................464 shape optimization ................. 10, 231, 528
recursive quadratic programming (RQP) ... shape variable.......................................231
.................................................145, 150 signal-to-noise (S/N) ratio....................401
redundant design ..............................71, 77 simplex method ....................................131
response surface approximate optimization simulated annealing (SA)...............15, 562
(RSAO)....................................545, 614 simultaneous analysis and design (SAND)
response surface method (RSM) ..276, 360 .................................................563, 568
419, 448, 516 size optimization .................... 10, 229, 468
responsibility coefficient......................587 slack variable .......................................126
robust design .................... 13, 18, 393, 472 sled test ................................................497
robustness space frame structure ...........................475
- axiomatic design............................428 spacer grid spring .................................455
- index......................................394, 429 standard deviation ................ 382, 395, 474
- Taguchi method.............................399 state variable ........................ 188, 564, 574
robust optimization ..............................416 static response optimization ... 11, 187, 279
stationary point.....................................117
S statistical interval estimation................385
safety analysis for occupant crash statistical point estimation....................384
environment (SAFE)........................499 statistics................................................382
scale factor steepest descent direction.............124, 139
- design of experiments (DOE) ........351 steepest descent method .......................138
- Taguchi method.....................402, 407 stepsize.........................................123, 260
seat belt stiffness matrix.........................9, 175, 279
- emergency locking retractor (ELR)535 stress constraint ............................111, 257
- web locking retractor (WLR) ........535 strip casting roll....................................444
semangularity .........................................27 structural analysis.........................474, 527
semianalytic method ....................200, 214 structural design ................... 110, 187, 540
sensitivity analysis structural optimization ...... 9, 11, 171, 187,
- adjoint variable method ........200, 206, 489, 540, 561
265, 270, 296 subproblem...................................145, 188
- analytic method with discrete sufficient condition ..............................120
Index 627
sum of squares......................................312 U
surplus variable ....................................126 ultra light steel auto body (ULSAB) ....244
switch coefficient .................................589 unconstrained optimization
synergistic interaction ..........................320 - Broyden–Fletcher–Goldfarb–Shanno
system analysis.....................................566 (BFGS) method ...........................143
system range...........................................36 - conjugate gradient method ............141
- Davidon–Fletcher–Powell (DFP)
T method.........................................142
Taguchi method....................................399 - quasi-Newton method....................142
tailor welded blank (TWB) ..................244 - steepest descent method ........124, 138
target value uncontrollable factor ....................395, 402
- collaborative optimization (CO)....599 uncoupled design........................18, 22, 25
- Taguchi method.............................400 uniform distribution .............................430
Taylor expansion..................................146 Unigraphics ............................................97
Taylor series.........................................165 unit step function....................................43
t-distribution.........................................384 UNIX .....................................................95
time domain .................................256, 281 upper bound .........................................108
time finite element method...................274 upper level
time grid point..............................257, 260 - axiomatic design........................23, 24
time-dependent constraint ............257, 263 - multidisciplinary design optimization
tolerance (MDO).........................................567
- axiomatic design........................31, 35
- robust design .................................401 V
tolerance design ...........................401, 404 variance........................................382, 394
tolerance range ...............................38, 421 variance of the error .............................312
top down approach .................................70 variational method................................217
topology optimization ............10, 237, 467 vertex point ..........................................131
total potential energy............................107 virtual variable .....................................127
trade-off coefficient..............................588 VisualDOC...................................545, 562
transformation matrix...........................180 V-model .................................................68
transformation method .........................153
transient analysis..........................276, 281 W
transient load........................................256 web locking retractor (WLR) ...............535
TRIZ...................................................6, 15 weighting factor ...................................407
two-step optimization...................404, 413 worst case approach .............................258
two-way factorial design ......................317
Y
YAG laser ..............................................82
Z
zigzagging process ...........................23, 66

Gyung-Jin Park PHD (Auth.) - Analytic Methods For Design Practice-Springer-Verlag London (2007)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Gyung-Jin Park PHD (Auth.) - Analytic Methods For Design Practice-Springer-Verlag London (2007)

Uploaded by

Copyright:

Available Formats

Analytic Methods for Design Practice

British Library Cataloguing in Publication Data

ISBN 978-1-84628-472-4 e-ISBN 978-1-84628-473-1 Printed on acid-free paper

mathematics. Also, practical engineers can use it in their design practice. If an

2 Axiomatic Design ...........................................................................................17

2.B.2 Axiomatic Analysis of an Existing Design.................................84

3 Design Optimization ....................................................................................107

4 Structural Optimization ..............................................................................171

5 Dynamic Response Optimization................................................................255

5.7 Structural Optimization in the Frequency Domain...............................292

6 Design of Experiments .................................................................................309

7.5 Robust Design by Axiomatic Design ...................................................428

8 Case Studies with Analytic Design Methods..............................................443

9 Application of Design Methods to Automobile Safety Problems.............497

9.4.5 Definition of the Objective Function........................................510

10 Multidisciplinary Design Optimization......................................................561

1.1 What is Engineering Design?

Figure 1.1. Flow of product design and manufacturing

Figure 1.2. An example of a drawing

Figure 1.3. A car with an aesthetic shape

rules or general principles for an excellent design? Engineering designers

Figure 1.4. Inside design of a car

1.2 Role and Classification of Design

Figure 1.5. Sequence of design

Analysis and experiment

Figure 1.6. Detailed design process

number of parts involved. In system design, many design factors are

appropriate method should be chosen according to the design processes or the

1.3 Analytic Design Methodology

should be mathematically defined. Many researches have proposed other methods

1.4 Axiomatic Design

Customer domain G Functional domain Physical domain Production domain

What? How? How? How?

Figure 1.7. Relationship between domains, mapping and spaces

the two domains.

1.5 Design Optimization

subject to hi (b) 0 , i 1, ,l (1.1c)

the lower and upper bounds of b, respectively, l is the number of equality

1.6 Structural Optimization

Table 1.1. Classification of structural optimization

Design variables are thickness or section properties. The

Design variables are the elastic modulus of the elements for

In structural optimization, the governing equation is computationally solved

subject to K (b)z f (1.3c)

g j (b, z ) d 0 , j 1,, m (1.3d)

1.7 Structural Optimization Under Dynamic Loads

subject to M (b)z  K (b) z f (t ) (1.5c)

g j (b, z, t ) d 0 , j 1, ,m (1.5d)

Equation 1.4 is imposed as equality constraints as shown in Equation 1.5c.

is discretized and transformed to multiple static loads and a structure is optimized

1.8 Design with DOE

1.9 Robust Design

1.10 Multidisciplinary Design Optimization

Structural analysis Control

Figure 1.8. Typical coupling relationship between various analysis methods

Analysis of multiple disciplines No Integration

Sensitivity analysis Simultaneous optimization

Figure 1.9. An example of a definition for MDO

relationship is illustrated in Figure 1.8.

Hurst K (1999) Engineering Design Principles. Wiley, New York

Axiom 1: The Independence Axiom

Customer domain Functional domain Physical domain Process domain

What? How? How? How?

subject to hi (b) 0 , i 1, ,l (1.1c)

g j (b, z ) d 0 , j 1,, m (1.3d)

subject to M (b)z K (b) z f (t ) (1.5c)

g j (b, z, t ) d 0 , j 1, ,m (1.5d)

FR1 A11 u DP1 A12 u DP2

R sin T (1 cos2 T )1 / 2 (2.11)