.

Structural Optimization

Gerald Kress, David Keller and Benjamin Schläpfer

January 22, 2015

Laboratory of Composite Materials and Adaptive Structures

Contents

1 Scope, Goals, and Sample Structural Optimization Problems 1

1.1 Introductory Remarks on Design Optimization . . . . . . . . . . . . . . . . 1
1.2 Overview of the Contents and Acknowledgment . . . . . . . . . . . . . . . . 2
1.3 Problems of Structural Optimization . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 Weight Minimization of a Motorcycle Tubular Frame . . . . . . . . . 4
1.3.2 Racing Car Rim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.3 Maximum-Strength Flywheel Design . . . . . . . . . . . . . . . . . . 6
1.3.4 Maximum Bond Strength Design . . . . . . . . . . . . . . . . . . . . 6
1.3.5 Composite Boat Hull . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.6 Minimum Weight Fuel Cell Stack End Plate . . . . . . . . . . . . . . 8

2 Treatment of a Structural Optimization Problem 9

2.1 Eschenauer’s Three-Columns Concept . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Structural Model or Structural Analysis . . . . . . . . . . . . . . . . 10
2.1.2 Optimization Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.3 Optimization Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Practical Optimization Problem Solution Setup . . . . . . . . . . . . . . . . 12
2.2.1 The DynOPS evaluator . . . . . . . . . . . . . . . . . . . . . . . . . 12

3 Design Evaluation 15
3.1 Local Optimality Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Global Objective Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Constraining Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4 Several Design Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4.1 Pareto Optimality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4.2 Substitute Problem and Preference Function or Scalarization . . . . 17
3.5 Transformation Methods and Pseudo Objectives . . . . . . . . . . . . . . . 18
3.5.1 Penalty Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.5.2 Method of Multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.5.3 Unconstrained Lagrange Problem Formulation . . . . . . . . . . . . 21
3.6 Fitness Function for Evolutionary Algorithms . . . . . . . . . . . . . . . . . 22
3.6.1 Mapping Functions for Objectives . . . . . . . . . . . . . . . . . . . 23
3.6.2 Constraint Mapping Functions . . . . . . . . . . . . . . . . . . . . . 24
3.7 Design Evaluation Exemplified on Selected Problems . . . . . . . . . . . . . 27
3.7.1 Weight Minimization of a Motorcycle Tubular Frame . . . . . . . . . 27
3.7.2 Racing Car Rim Design Evaluation . . . . . . . . . . . . . . . . . . . 28
3.7.3 Maximum-Strength Flywheel Design . . . . . . . . . . . . . . . . . . 30
3.7.4 Maximum Bond-Strength Design . . . . . . . . . . . . . . . . . . . . 32
3.7.5 Composite Boat Hull . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

c ETH Zürich IDMF-CMAS, March 1, 2018

 3.7.6 Minimum-Weight Fuel-Cell Stack End Plate . . . . . . . . . . . . . . 34

4 Parameterization and Variable Transformations 37

4.1 Classification of Design Variables and Structural Optimization Problems . . 37
4.2 Design Parameterization Sample Problems . . . . . . . . . . . . . . . . . . . 39
4.2.1 Motorcycle Frame and Sizing Parameters . . . . . . . . . . . . . . . 39
4.2.2 Shape Parameters of the Formula 1 Race Car Rim . . . . . . . . . . 40
4.2.3 Maximum-Strength Flywheel and Mesh-Dependent Shape Parameters 42
4.2.4 Onsert Design and Mesh-Independent Shape Parameters . . . . . . . 42
4.2.5 Composite Boat Hull and the Patch Idea . . . . . . . . . . . . . . . 44
4.2.6 Fuel-Cell-Stack End Plate . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3 The Parameterization Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.3.1 Influence of Mechanical Situation on Parameterization . . . . . . . . 47

5 Some Basic Concepts of Global Nonlinear Optimization 51

5.1 Nonlinear Optimization Task . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.2 Feasible Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.3 Convex and Non-Convex Functions . . . . . . . . . . . . . . . . . . . . . . . 52
5.4 Method of feasible directions . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.4.1 Inequality Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.4.2 Equality Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.4.3 Generalization of Lagrange factor calculation . . . . . . . . . . . . . 55
5.5 Lagrange Multiplier Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.5.1 Lagrange Multiplier Method Sample Problem . . . . . . . . . . . . . 56
5.6 Necessary and Sufficient Optimality Criteria . . . . . . . . . . . . . . . . . . 57
5.6.1 Unconstrained Objective Functions . . . . . . . . . . . . . . . . . . . 57
5.6.2 Kuhn-Tucker Optimality Conditions . . . . . . . . . . . . . . . . . . 57
5.7 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.7.1 The Max-Min Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.7.2 The Primal and Dual Problems . . . . . . . . . . . . . . . . . . . . . 60
5.7.3 Computational Considerations . . . . . . . . . . . . . . . . . . . . . 60
5.7.4 Use of Duality in Nonlinear Optimization . . . . . . . . . . . . . . . 61
5.8 Optimization Algorithms Overview . . . . . . . . . . . . . . . . . . . . . . . 63
5.8.1 An Argument for Mathematical Programming . . . . . . . . . . . . . 64
5.8.2 An Argument for Stochastic Methods . . . . . . . . . . . . . . . . . 65
5.8.3 Mathematical Programming and Stochastic Search Methods . . . . . 66
5.9 Practical Considerations for Numerical Optimization . . . . . . . . . . . . . 67
5.9.1 Advantages of Using Numerical Optimization . . . . . . . . . . . . . 67
5.9.2 Limitations of Numerical Optimization . . . . . . . . . . . . . . . . . 67

6 Search for Design Improvement: Mathematical Programming 69

6.1 Simplex Search Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.2 Method of Steepest Descent . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.3 Quadratic Objection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.4 Original and Modified Newton Methods . . . . . . . . . . . . . . . . . . . . 77
6.5 Nonlinear Conjugated Gradient Methods . . . . . . . . . . . . . . . . . . . . 79
6.6 Powell’s Conjugate Direction Method . . . . . . . . . . . . . . . . . . . . . . 83
6.7 Response-Surface Method Minimizing Algorithms . . . . . . . . . . . . . . . 87
6.7.1 Constructing a Response Surface Model from Supporting Points . . 88
6.7.2 Finding the Minimum Point of the Response Surface Model . . . . . 89
 6.7.3 The Relation between RSM and NM . . . . . . . . . . . . . . . . . . 89
6.7.4 Adaptive Response Surface Method . . . . . . . . . . . . . . . . . . 96
6.8 Line-Search Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.8.1 One-Dimensional Search in Multidimensional Variables Space . . . . 99
6.8.2 Interval Halving and Golden Section Methods . . . . . . . . . . . . . 99
6.8.3 Quadratic and Cubic Approximation Methods . . . . . . . . . . . . 101
6.8.4 Brent’s Routine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.9 Lagrange Multiplier Method Numerical Optimization . . . . . . . . . . . . . 111
6.9.1 Modified Lagrangian with Local Minima . . . . . . . . . . . . . . . . 111
6.9.2 Algorithm for Removing Constraint Violations . . . . . . . . . . . . 112
6.10 Objective Function Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.10.1 Differences Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.10.2 Sensitivity-Formula Gradient Calculation . . . . . . . . . . . . . . . 115
6.11 Global Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.11.1 Multi-Start Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.11.2 Tunneling Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

7 Stochastic Search 121

7.1 Introduction to stochastic search and optimization . . . . . . . . . . . . . . 121
7.1.1 Neighborhood concept . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.1.2 Strategies in stochastic search . . . . . . . . . . . . . . . . . . . . . . 122
7.1.3 A prototype of a stochastic search algorithm . . . . . . . . . . . . . 122
7.1.4 Performance of stochastic search . . . . . . . . . . . . . . . . . . . . 123
7.2 Stochastic Search Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.2.1 Random Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.2.2 Stochastic Descent . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.2.3 Metropolis Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.2.4 Simulated Annealing . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.2.5 Evolutionary Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.3 Representation Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.3.1 The universal genotype . . . . . . . . . . . . . . . . . . . . . . . . . 133

8 Composite Structures 135

8.1 Design of Fiber Reinforced Composites . . . . . . . . . . . . . . . . . . . . . 135
8.2 Laminated Composites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.2.2 Classical Laminate Theory . . . . . . . . . . . . . . . . . . . . . . . 138
8.3 Finite Element Representation . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.3.1 Layered Shell Elements . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.3.2 Laminate Sensitivities . . . . . . . . . . . . . . . . . . . . . . . . . . 144
8.4 Optimization with Lamination Parameter . . . . . . . . . . . . . . . . . . . 151
8.4.1 Basic Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
8.5 Optimization on Physical Design Variables . . . . . . . . . . . . . . . . . . . 155
8.5.1 Optimization of the Fiber Orientation . . . . . . . . . . . . . . . . . 155
8.5.2 Optimization of the Stacking Sequence . . . . . . . . . . . . . . . . . 157
8.5.3 Material Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.5.4 Optimization of the Laminate Thickness . . . . . . . . . . . . . . . . 158
8.5.5 Combined Laminate Optimizations . . . . . . . . . . . . . . . . . . . 160
8.6 Laminate Tailoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
 8.6.1 FEM-Based Parametrization of the Sub-Domains . . . . . . . . . . . 162
8.6.2 CAD-Based Parametrization of the Sub-Domains . . . . . . . . . . . 164

9 Selected Methods and Case Studies 167

9.1 Computer Aided Optimization after Mattheck . . . . . . . . . . . . . . . . . 167
9.2 Soft-Kill Option after Mattheck . . . . . . . . . . . . . . . . . . . . . . . . . 169
9.3 Flywheel Optimization and Inspired Mechanical Model . . . . . . . . . . . . 171
9.3.1 Sodola’s Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
9.3.2 Shape optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
9.3.3 Discussion of results . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
9.3.4 Simple Prediction of Optimum Shape Features . . . . . . . . . . . . 174
9.3.5 Conclusions on the Flywheel Optimization and Modeling . . . . . . 178
9.4 Topology Optimization after Bendsøe and Kikuchi . . . . . . . . . . . . . . 179
9.4.1 Topology optimization sample problem . . . . . . . . . . . . . . . . . 180
9.4.2 Objective Function and Design Evaluation . . . . . . . . . . . . . . . 181
9.4.3 Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
9.4.4 Optimization Problem Statement . . . . . . . . . . . . . . . . . . . . 183
9.4.5 Lagrange Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
9.4.6 Gradients with Respect to Densities . . . . . . . . . . . . . . . . . . 184
9.4.7 Calculation of Lagrange Factors . . . . . . . . . . . . . . . . . . . . 185
9.4.8 A Dual Algorithm for Topology Design . . . . . . . . . . . . . . . . 186
9.4.9 Sample Topology Design Problem . . . . . . . . . . . . . . . . . . . 188
9.5 Truss Optimization with Ground Structure Approach . . . . . . . . . . . . 189
9.5.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
9.5.2 Problem Statement Extension for Multiple Loads . . . . . . . . . . . 190
9.5.3 Problem Statement with Self-Weight Loading . . . . . . . . . . . . . 191
9.5.4 Fully Stressed Design and Optimality Criteria Methods . . . . . . . 191

10 Demonstration Programs 195

10.1 Program DEMO OPT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
10.2 Program TOP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

A Finite Element Method 213

A.1 Equation of Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
A.2 Formulation of a Layered Shell Element . . . . . . . . . . . . . . . . . . . . 217

B Material Invariants of Orthotropic Materials 221

Chapter 1

Scope, Goals, and Sample

Structural Optimization Problems

1.1 Introductory Remarks on Design Optimization

Optimization is inseparable from product development. Before the invention of structural

analysis methods, optimization was based on a combination of experience, or experimen-
tation, and intuition. The sensitivity of the performance properties of an existing product
to changes of its design became thus apparent only through some trial and error process.
The time needed for some design improvement was thus determined by building the new
design and testing it or finding out about its performance during service.
The invention of powerful structural analysis methods, such as the finite-element method
(FEM), reduced the time needed for one optimization cycle because a product does not
have to be built anymore to find out about its performance properties. The structural
analysis information may give more or less direct clues to the analyst, or the ”decision
maker” as to how the design of the structure may be changed to improve its properties
with respect to some requirement or objective. Still, the optimization process requires
intuition to identify the improving design changes and manual work to implement these
changes in the structural analysis models, evaluating the results of the new simulations,
and comparing these with the previous results. Apart from this, in complex situations
human intuition may fail at finding the improving design variations and the path from the
initial to a fully optimized design is often so winding that a large number of very small
straight steps is required to follow it up.
Fully automated optimization procedures let the analyst partake more efficiently at the
active and creative product development process. They reduce the process times greatly
and find the best design solutions systematically, letting the analyst focus on his human
role as planner and decision maker.
Structural optimization is quite well developed on academic level but is still not routinely
used in industrial environments. In the past, optimization procedures following closely the
path winding through design space from the initial to the optimum design have received
the highest attention. Such nonlinear search methods are summarized under the term
mathematical programming. They assume the optimization objective to be formulated in
terms of continuous and, often, at least twice differentiable functions that should also be
convex. Then, mathematical programming requires a numerical effort that can be orders
of magnitude less than with other methods. The practical difficulty with mathematical
programming lies in interfacing of the design, analysis, and evaluation models and the

c ETH Zürich IDMF-CMAS, March 1, 2018

2 Scope, Goals, and Sample Structural Optimization Problems

optimizers.
Generally, real optimization problems yield objective functions that can not be processed
with mathematical programming. Stochastic search methods do not suffer limitations like
mathematical programming and the interfacing problems are often less severe. They are
thus better suited for many practical problems but they require a much higher number of
function evaluations unless the optimum is hit early by chance. Consequently, the latest
research [1] focuses on rendering stochastic search methods more efficient by borrowing
concepts from mathematical programming. The more efficient stochastic search methods
are based on concepts inspired by the evolution of life or bacterial search mechanisms,
for instance, and involve the evaluation of whole populations of trial design solutions (or
their genotypes) within one generation of the evolution process. Since the individuals of
one population can easily be evaluated in-parallel, the use of modern massively parallel
computer architecture, such as Beowulf clusters with hundreds of processors, mitigates the
efficiency deficiency of the stochastic search methods.

1.2 Overview of the Contents and Acknowledgment

The students shall understand the essential aspects of structural optimization by studying
the various aspects given with this lecture class script. They begin with studying the
selected structural optimization problems listed in the following Section 1.3. The prob-
lems are used in the lecture to exemplify the various aspects of setting up an automated
optimization procedure. The task is supported by the schematic presented in Chapter 2.
Chapter 3 teaches the various components of design evaluation in terms of formulation of
the optimization objectives and the design constraints and the recovery of the necessary
data from the structural analysis results. The analysis models and the design models must
communicate through transformations between design and analysis variables which is the
subject of Chapter 4 on design parameterization. The parameterization is also linked
with the classification of structural optimization problems. Chapter 5 explains some basic
concepts related to minimization of constrained objective functions and the material has
been extended to include dual problem formulation in June 2007 although this material
will first be taught in class in 2008. Optimization algorithms of mathematical program-
ming and advanced stochastic search are explained in some detail in Chapters 6 and 7,
respectively. The topic of stochastic search, Chapter 7, has been provided by my colleague
Dr. David Keller in 2007. In 2013, Dr. Benjamin Schläpfer has contributed Chapter 8
on optimization of laminated composite structures; this topic must also address basics of
the mechanics of composite materials as well as of the finite-element method. Finally,
Chapter 9 presents special methods of importance, specifically the methods developed
by Claus Mattheck, the homogenization method pioneered by Bendsøe and Kikuchi, the
ground-structure approach for truss design. It also includes a mechanical model inspired
[2] by flywheel shape optimization to demonstrate that the interpretation of numerical
optimization results may enhance understanding of mechanical problems.
The sample problems of structural optimization presented in the following section have
been provided by my colleagues of the optimization group. The motorcycle-frame (Section
1.3.1) and the race-car rim (Section 1.3.2) design problems are contributed by M. Win-
termantel and O. König [3, 4, 5, 6]. The flywheel [2] and load-introduction for sandwich
problems [7, 8] are provided by G. Kress. The problem of global laminate optimization
(Section 1.3.5), with the composite boat hull sample problem, has been provided by N.
Zehnder [9, 10]. The problem of finding a minimum-weight design for the end plates of a
fuel-cell stack (Section 1.3.6) has been worked out by O. König [5, 11].

c ETH Zürich IDMF-CMAS, March 1, 2018

1.3 Problems of Structural Optimization 3

1.3 Problems of Structural Optimization

The structural optimization sample problems presented in the following subsections 1.3.1
through 1.3.6 have been studied at the Centre of Structure Technologies in the course of
building up know-how of optimization methods, research and development projects, Ph.D.
theses, and student graduate projects. Each of the problems is unique in terms of problem
statement and solution methods. They are here introduced to trigger the creativity of the
students who are invited to suggest ideas on how the various problems could be solved.
The problems are also used throughout the lecture class to illustrate the various solution
approaches, design parameterizations, and search techniques for finding the best design
solutions.
Different types of optimization problems are illustrated in Fig. 1.1. Often, design solu-
!  

  
2
G 
G
G !

G "

(a) (b) (c) (d)

Figure 1.1: Types of Structural Optimization

tions in terms of load-carrying frames, often used for bridges, motorcycle frames, or other
structures, have a fixed connectivity of the various members with each other, or topology.
If the topology is already specified, further possibility for improving the performance of
such structures lies in the adjustment of the various member’s size. The sizing changes
the area values, or moments of inertia, of members such as trusses or beams. The illus-
tration shown on Fig. 1.1(a) indicates not only the sizing of member properties but also
the change of the position of the nodes where members connect.
The position change of connectors indicated in Fig. 1.1(a) is very similar to the shape
optimization indicated in Fig. 1.1(b). The example shows an optimized design with a
shape that resulted, through a shape-optimization process, from a rectangular-shaped ini-
tial design.
Sizing and shape optimization always require some initial design where the topology is
already fixed. Less information, or pre-existing knowledge, requires the topology opti-
mization. It requires only a definition of the physical or geometric design space and the
specification of geometric boundary conditions and loads. By redistributing the material,
which is initially evenly distributed in the design space, some topology as indicated in Fig.
1.1(c) is automatically created. The method, along with a specific solution technique [12],
is explained in section 8.1. It was used to create the title page illustration. The shown
result is the stiffest structure with respect to the sketched boundary conditions and the
fineness of the finite-element mesh. The simulated best design hints at the optimal Michell
structures [13].
Advanced composite materials consist of fibers of high stiffness and strength embedded
in some matrix the material of which may be rather weak and compliant. The resulting
composite, for instance with unidirectional reinforcement, has mechanical properties that
are highly direction-dependent, or anisotropic. Such materials are most often used for thin
shell structures, where the walls consist of laminates of several layers of the composite ma-
terial. The orientation of the reinforcement of the individual layers, see Fig. 1.1(d), can be

c ETH Zürich IDMF-CMAS, March 1, 2018

4 Scope, Goals, and Sample Structural Optimization Problems

chosen to obtain some desired global structural behavior. Fiber orientation is one of the
internal parameters characterizing a laminate, and algorithms for the optimum adjustment
of these are called internal material parameter optimization.
Of course, different structural optimization types may be combined to solve one opti-
mization task. Topology optimization may be followed by shape optimization or shape
optimization may be coupled with internal material parameter optimization.

1.3.1 Weight Minimization of a Motorcycle Tubular Frame

The motor cycle manufacturer Ducati, active in racing, uses a tubular steel trellis frames
such as the one of the Ducati 996R Superbike shown in Fig. 1.2(a). A co-operation
with Ducati lead to a student graduate project and inspired the development of some
solution technique to be published soon [3]. The company is interested in making the frame

(a) (b)

Figure 1.2: Ducati 996R frame and geometry model [3]

structure as lightweight as possible whilst at the same time preserving prescribed structural
stiffness properties. The structural stiffness influences the racing performance of the bike
because it contributes to the suspension characteristics. The springs and shocks provide
the suspension characteristics of the bike in the straight position. The forced displacements
or dynamic loads, exerted by the road onto the bike at certain speeds, act then parallel
to the vertical axis of the bike. In curves, surrounded at high speeds, the forces act more
sideways and the suspension system takes only a small component in the vertical direction
of the bike. The lateral component must be damped by the structural compliance of the
bike in the lateral direction and with respect to the contact points between wheels and
road. A significant factor determining the lateral suspension properties to be provided by
the frame is its torsional stiffness. Therefore, the torsional stiffness of the frame, measured
in terms of twisting moment related to relative twist between its front and rear ends, is a
fixed value that must be kept constant when reducing the weight. Such a requirement is
called a constraint, specifically an equality constraint. Another constraint is derived from
the fact that the frame must withstand the loads to be transferred by it. This implies
that the stresses induced by external loading must not exceed the material strength at
any location of the frame. In order to obtain the necessary information on the structural
stiffness and the stressing of the frame, its behavior and performance under loads must
be simulated. The simulations are provided by the finite-element method FEM and Fig.
1.2(b) shows the frame geometry that was used for the FEM modelling.

c ETH Zürich IDMF-CMAS, March 1, 2018

1.3 Problems of Structural Optimization 5

1.3.2 Racing Car Rim

The problem was considered in the Ph.D. theses of O. König [5] and M. Wintermantel [6],
a number student projects. The present problem characterization is taken from [5].
The performance of a race car as shown in Fig. 1.3 depends on several factors: weight
should be as low as possible, the moment of inertia along the rotation axis has to be
small, and the stiffness should be high at the same time. The weight of the rim does

Figure 1.3: CAD-Model of the racing car rim (courtesy O. König [5])

not only influence the performance as a part of the cars overall weight. Since the wheels
belong to the so called unsprung mass, a low rim weight improves the mechanical grip of
the car especially on bumpy road surfaces. The moment of inertia along the wheels
rotational axis should be minimal for several reasons. Low moments of inertia allow faster
acceleration and deceleration of the wheels and therefore of the whole car. Furthermore,
the moment of inertia leads through the gyro effect to higher steering forces as well as a
higher inertia of the car with respect to direction changes. Finally, the stiffness of the
rim is of high importance in turns at high speed. Vertical loads of 5700N , resulting from
the car’s weight and aerodynamical descending forces, as well as maximum lateral forces
of 7000N , as a result of the centripetal forces, build up a bending moment on the rim.
Additionally, strength requirements must be fulfilled. Plastic yield must not occur in use,
whereas some parts reach temperatures well above 200◦ C. A special magnesium alloy
is used where maximum yield stress does hardly decrease with higher temperatures and
the mass-specific stiffness ratio is high. Manufacturing starts with a forging blank, the
rim’s bed is shaped by CNC-lathe, and the spokes form the interspace of CNC-milled
pockets. The forging blank’s shape is not to be changed, as this would exceed costs. FIA1
regulations affect the bead diameter as well as dimensions of the lower rim-bed.
For the optimization presented in here, maximum bending stiffness is defined as main
design objective from the rim manufacturer. Nevertheless, the desired properties

• Low mass

• Low rotational moment of inertia

• Sufficient margin of safety for mechanical stresses

of the existing design should also be matched or even surpassed by the optimized design.
Furthermore, the optimization must also take into account the functional, regulatory, and
manufacturing requirements discussed earlier in this section. Altogether, this constitutes
1
Fédération International d’Automobiles ([Link]

c ETH Zürich IDMF-CMAS, March 1, 2018

6 Scope, Goals, and Sample Structural Optimization Problems

a highly constrained optimization problem, which can not be tackled with classical math-
ematical optimization techniques. Since all original data is confidential, arbitrary new
geometries for the rim and the forging blank, as well as modified load cases are created
and optimized for the presentation in here.

1.3.3 Maximum-Strength Flywheel Design

Consider a flywheel of constant thickness with a central bore as illustrated in Fig. 1.4(a).
Because of the rotational symmetry, the structural behavior of the flywheel can be simu-
lated by a 2-dimensional FEM model where the finite elements are based on a rotational
symmetry assumption. Such a model is plotted in Fig. 1.4(b) and the radial and circum-
ferential stress distributions are also shown. The direct stress in radial direction falls down
to zero at the edge of the central bore and also at the outer rim because both are stress-free
boundaries. The direct stress in the circumferential direction, however, increases sharply
at the edge of the central bore. The material of the flywheel is unevenly stressed and,

(a) (b)

Figure 1.4: Flywheel with central bore, 2-dimensional FEM model and stresses [2]

therefore, not economically used.

A more even stress distribution can be found by varying the thickness of the flywheel along
the radius.

1.3.4 Maximum Bond Strength Design

The onsert is a joining element to achieve load introduction into lightweight structures
typical for the transportation sectors [7, 8]. In contrast to the insert, the onsert is simply
bonded to the surface of a plate that can be a sandwich structure as shown in Fig. 1.5(a).
The load that can be transferred by the system consisting of the parts onsert and sandwich
structure depends on the stress distribution in the bonding layer between the two parts.
Assuming that the onsert geometry is rotationally symmetric and a load is centrally applied
in the axial direction, a 2-dimensional model similar to that used for analyzing the flywheel
can be used. The 2-dimensional basic geometry model is shown in Fig. 1.5(b). The finite
element model shown in Fig. 1.5(c) obtains the stress distributions in the bonding layer
shown in Fig. 1.5(d). The stresses are very unevenly distributed with high concentrations
at the central bore. Failure of the bond will be initiated there.
The maximum-strength design of onserts and other bonded systems is one of the ongoing
research activities shared between Alcan and the chair Structure Technologies.

c ETH Zürich IDMF-CMAS, March 1, 2018

1.3 Problems of Structural Optimization 7

r s a n d w ic h

r o n s e rt

r b o re
O n s e rt h o n s e rt

B o n d L a y e r h g lu e
F a c e S h e e t

C o re h c o re

F a c e S h e e t h fa c e

(a) (b)

(c) (d)

Figure 1.5: Onsert design demonstrator (a) and geometry model (b)

1.3.5 Composite Boat Hull

The hull of a sail boat, made of composite materials, should be as stiff as possible under
typical service loads. The particular model shown in Fig. 1.6 should be low priced for
marketing reasons. The sample problem is thus useful to demonstrate a problem where a

Figure 1.6: ANSYS model of the sail boat hull with composite material patches (Courtesy
N. Zehnder [14])

mechanical and an economical objective are combined to give a multi-objective optimiza-

tion problem. Also, achieving desired mechanical properties with a composite material
design provides an interesting design parameterization problem. An elegant key idea to
solve this problem is studied and developed by N. Zehnder in the course of his Ph.D. thesis
work.

c ETH Zürich IDMF-CMAS, March 1, 2018

8 Scope, Goals, and Sample Structural Optimization Problems

1.3.6 Minimum Weight Fuel Cell Stack End Plate

The weight of an end plate of a fuel cell stack shall be minimized. The problem arose in
fuel cell research projects2 [15, 16] at the Swiss Federal Institute of Technology Zurich. The
end plates for the fuel cell stacks developed in these projects were designed by Tribecraft
AG3 [17], which also provided the detailed problem description for the optimization pre-
sented in this section. A fuel cell stack, as shown in Figure 1.7, consists of bipolar plates,
two collectors, electrical insulation, and the end plates that are connected with tie bolts.
The fuel and cooling supply line runs through the upper end plate. The weight of such

Figure 1.7: Conceptual model of a fuel cell stack.

an end plate is minimized subject to a stress constraint and manufacturing requirements,

whereas the structure is foreseen to be manufactured by extrusion molding. The objective
is to increase the power density, i.e. power per weight, of a fuel cell stack. In a second
optimization procedure, described in [17], the bottom surface of the end plate can be cam-
bered independently to guarantee a constant pressure distribution on the fuel cell stack
in built-in state. This sequential partitioning of the optimization problem allows to first
minimize the weight of the plate under a stress constraint without concern for the stiff-
ness of the plate. For the verification model mentioned above, the same CAD-model and
genotype is used to optimize a bridge-like structure. The objective of this optimization is
to minimize its compliance subject to a constant mass given in percent of the fully-filled
design domain.

2
[Link]
3
[Link]

c ETH Zürich IDMF-CMAS, March 1, 2018

Chapter 2

Treatment of a Structural
Optimization Problem

Optimizing a structure by some automated numerical procedure may seem very complex
and difficult to organize. The concept presented in the following sections was worked out
by Eschenauer [18] and decomposes the task into manageable subtasks so that it can be
solved in a straightforward manner. Although Eschenauer’s concept seems to have been
developed with regard to the optimization algorithms labelled by the term mathematical
programming, it is also valid when other solution techniques such as genetic algorithms
are used.

2.1 Eschenauer’s Three-Columns Concept

An optimization problem can generally be solved by applying the Three-Columns Concept
after Eschenauer [18]. The three columns are the structural model, the optimization model,
and the optimization algorithm. The concept is depicted in Fig. 2.1 and the following
sections follow to some extend the presentation in the textbook [18].

d e c is io n d a ta
m a k e r
in p u t
o p tim a l
d e s ig n
x * in itia l d e s ig n s tru c tu ra l
x 0 p a ra m e te rs

tra n s fo rm e d a n a ly s is
tr a n s fo r m a tio n d e s ig n m o d e l
v a r ia b le z ® x d e s ig n x ® y v a r ia b le
v a r ia b le
z x y
o p tim iz a tio n s tru c tu ra l m o d e l
a lg o r ith m o p tim iz a tio n m o d e l s tr u c tu r a l a n a ly s is
u = u (y )
F ,g ,h B ,f ,g ,h
o p tim iz a tio n e v a lu a tio n
s tr a te g ie s m o d e l u
¶ F ¶ g ¶ h
, ,
¶ x ¶ x ¶ x
s e n s itiv ity
a n a ly s is

Figure 2.1: Three-columns concept after Eschenauer [18]

c ETH Zürich IDMF-CMAS, March 1, 2018

10 Treatment of a Structural Optimization Problem

2.1.1 Structural Model or Structural Analysis

In order to be able to perform a computerized optimization process, the real-life structure
must be transferred into a structural model. The structural model describes mathemat-
ically, or numerically, the physical behavior of a structure. The physical behavior of a
mechanical system is its response to loads or structural properties such as eigenfrequen-
cies or weight. The optimization objective and constraints are formulated in terms of
state variables. For instance, if the structural model is formulated by using the finite-
element method (FEM), the state variables of a solid mechanics problem are the nodal-
point displacements u. Other interesting quantities such as stresses are calculated from
the displacements within the postprocessing step.

2.1.2 Optimization Algorithm

Real-life optimization problems generally lead to nonlinear and constrained optimization
problems. Optimization algorithms solve such problems. They are based on iteration
procedures that proceed from an initial design x0 and produce improved design variable
vectors xk . The procedure stops upon satisfaction of some predefined convergence crite-
rion. Experience teaches, and numerous studies have shown, that the choice of the most
appropriate optimization algorithm is problem-dependent.

2.1.3 Optimization Model

The optimization model builds a bridge between the structural model and the optimiza-
tion algorithms.
The evaluation model performs the design evaluation in terms of the optimization objec-
tive and the state (i.e. violated or not) of existing constraints from the values of the state
variables and other information from the structural model. The optimization objective is
often formulated as a scalar objective function f or, in case of multi-objective optimization,
a vector f . The constraints of a design are formulated in terms of constraining functions
vectors g (inequality constraints) and h (equality constraints). The evaluation model may
based on the state variables u (for calculating stresses, for instance) or some other vari-
ables influencing the design (for calculating weight, for instance).
The optimization model also contains variable definitions and transformations that can be
summarily called parameterization. The analysis variables y are chosen from the struc-
tural parameters. For example, nodal point positions at the boundary of a structural model
domain define the shape of the structural model and change during a shape optimization
process. The shape of a structure or design is explicitly defined in terms of design variables
x. The design model describes the mathematical relation between the analysis variables
y and the design variables x.
The design variables variables x may by additionally transformed into transformation
variables z in order to adapt the optimization problem to the special requirements of the
optimization algorithm.
The sensitivity analysis yields the sensitivity of the objective and constraints with re-
spect to small changes of the design variables. This information is used to control the
optimization algorithm and/or by the decision maker who judges the design.

Design Evaluation
Design evaluation and the setting up of an evaluation model is discussed in chapter 3
on the basis of the sample problems presented in chapter 1. In structural optimization,

c ETH Zürich IDMF-CMAS, March 1, 2018

2.1 Eschenauer’s Three-Columns Concept 11

one uses generally FEM to obtain the response of a structure to loads under specified
geometric boundary conditions. The solution of the numerical system of equations obtains
the primary solution in terms of the nodal-point degrees of freedom. In structural analysis,
the degrees of freedom are displacements. From the primary solution other results such
as stresses can be obtained. The stresses may be used to formulate an objective when
the strength of a part is to be maximized. Stresses may also be used to formulate a
constraint when, for instance, the weight of a load-carrying part is to be minimized but a
required strength must be preserved. The weight is then calculated from the integral of
the material densities over the volume of the considered part and does not depend on the
load response. Generally, however, the structural response is needed for the evaluation of
both, the objective and the constraints.

Design Parameterization

The parameterization of a design is crucial for achieving an efficient optimization process.

Parameterization concerns the design and the analysis models and the transformation be-
tween the design variables and the analysis variables.
One important aspect arises from using FEM structural models in terms of a finite-element
mesh. The mesh parameters comprise the element types with their physical properties,
such as thickness of plates or cross-sectional values of beams, and the fineness of the mesh
resp. number of elements or nodal points and the nodal point positions. Obviously, the
mesh parameters can be messy to handle because there are so many of them. Also, one
may wish to change the fineness of the mesh to specific needs, for instance to have a smaller
number of unknowns to reduce the solution time for one design evaluation or to increase
the accuracy of stress results. Such changes would entail changing the whole parameter-
ization. It is therefore desirable to use design parameters that are defined independent
of mesh size and to have an automated transformation between those design parameters
and the analysis parameters. The transformation can be in terms of a mesh generator
that uses, as input, some fixed and variable design parameters and some data controlling
the mesh fineness and generates the FEM mesh data establishing the analysis model. The
parameterization in terms of design variables is then mesh independent. Also, the number
of design variables is typically much smaller than that of the analysis variables so that the
optimization algorithm converges much faster, reducing the number of necessary design
evaluations and with it the solution time of the optimization process.
It is difficult to discuss more specific details of the parameterization issue in general be-
cause the kind and choice of design variables depend very much on the considered problem.
Therefore, the sample problems listed in chapter 1 are used to discuss their respective pa-
rameterization models in chapter 3.

Transformed Variables

The design variables may be transformed to meet certain requirements of the optimization
algorithm. For instance, genetic algorithms operate upon genotype variables that may be
defined in terms of bit strings whose encoded information must be transformed into the
phenotype design variables. The topic of transformed variables is therefore discussed in
context with genetic algorithms in section 6.3.

c ETH Zürich IDMF-CMAS, March 1, 2018

12 Treatment of a Structural Optimization Problem

2.2 Practical Optimization Problem Solution Setup

Some of the sample problems discussed throughout this lecture class (Flywheel, Onsert)
have been solved by writing dedicated computer codes where the design is controlled by
the variables read from a data input file. The variables determine details of a finite ele-
ment mesh which is built by a specially written mesh generator. The solution of the FEM
model, its evaluation with objective and constraints calculation, and the search algorithm
are also implemented in the one dedicated code. Although this concept may give perfect
programs for special problems, it is not very useful in practice where one desires to have a
general tool with which various and different problems can be used. A concept with more
practical significance has been developed by O. König [5], M. Wintermantel [6], and N.
Zehnder in the course of Ph.D. work. The concept has been developed and realized for
evolutionary algorithms as optimization engine but could also be applied with mathemati-
cal programming techniques. It is here described by using material from O. König’s Ph.D.
thesis [5].
To efficiently apply design optimization to engineering practice, it should be possible to
integrate external simulation software, only controllable through text files, with the op-
timizer. The in-house developed software DynOPS (Dynamic Optimization Parameter
Substitution) is made for that purpose: it allows to run evolutionary optimization using
arbitrary simulation software. The software is built on an Generic Evolutionary Algorithm,
it design parameters, and it also integrates the design evaluation. The main component
newly added in DynOPS is an interface transferring the optimization parameters from the
optimizer to the respective simulation program using file amendment as well as handling
of the results. Further more, a general concept for program handling is implemented,
allowing also sequences with different simulation programs to be used for evaluation. An
example for such an evaluation composed of two simulation programs is to handle a pa-
rameterized structure in a 3D CAD program, compute its mass and inertia, and export the
model to a FEM program where under given load cases deflections and maximum stresses
are evaluated.

2.2.1 The DynOPS evaluator

To explain the general functionality of DynOPS, the new evaluation module is first dis-
cussed as stand-alone tool. Given an arbitrary population of eoUniGene genotypes, the
DynOPS evaluator calculates fitness values for every individual as shown in Fig. 2.2. To
set up such an evaluation procedure, the following problem-dependent information must
be defined:

Sequence of simulation programs. The framework of the fitness evaluation is defined

through a sequence of involved simulation programs together with the needed in-
put/output files.

Input files for the simulation programs. The actual simulations to be carried out are
established with the input files, as well as the effective results to be calculated and
stored in output files.

Mapping from genotype to input files. Every optimization parameter in the input
files must be linked to the appropriate gene of the eoUniGene genotype. This is
done by storing the exact row and column, where every parameter must be inserted
in the input files.

c ETH Zürich IDMF-CMAS, March 1, 2018

2.2 Practical Optimization Problem Solution Setup 13

Offspring Dynops Evaluator

Fitness Value Optimization Parameter

Computation Transfer

Simulation
Objective & Constraint Program
Value Reading Manager File File File

File File

Simulation Simulation
Program B File Program A

Figure 2.2: DynOPS evaluator to calculate fitness values for a population of individuals
using external simulation software.

Objectives, constraints, and fitness function. How to read the objectives and con-
straints from the appropriate output files, as well as how the actual fitness should
be calculated from these values must be defined.

The simulation program manager starts the first simulation program together with the
appropriate input files, and waits for job completion. The results from the evaluation
stage stored in output files are either used as input for the next simulation program (e.g.
a geometry file), or are directly used for fitness calculation (e.g. a mass evaluated in the
CAD system). The next evaluation stage in the sequence is then started, and so forth.
After completion of the sequence of simulation programs, objective and constraint values
stored in result files are transferred back to the DynOPS evaluator. The evaluation of an
individual is completed by computing its fitness value. This evaluation loop is repeated
for the whole population.

c ETH Zürich IDMF-CMAS, March 1, 2018

14 Treatment of a Structural Optimization Problem

c ETH Zürich IDMF-CMAS, March 1, 2018

Chapter 3

Design Evaluation

3.1 Local Optimality Criteria

Material failure is a local affair. A material fails when its strength is exceeded by the
local stress state at a point. So, an equivalent stress can be a local optimality criterion.
The objective would be to reduce the highest equivalent stress, appearing at some spatial
position in the structure, to the smallest possible value or

min {max {σeqv (x, spatial position)}} (3.1)

by adjusting the design parameters x.

An added difficulty with such an objective is that the position where the highest stress
value appears is likely to change when the design changes. Evaluating the design then
implies calculating the stress distribution and selecting the highest value. The subsequent
systematic design change is likely to reduce the highest stress value at the selected position
but may also raise stress values at other positions. This holds the risk of a non-converging
optimization process or the continued switching between two or more design solutions.
Examples for local optimality criteria are provided by the growth-strategy approaches by
Mattheck, called computer aided design (CAO) and soft-kill option. The local optimality
criteria are implicitly able, under certain circumstances or restrictions, to improve or
optimize global design properties such as load transfer capacity.

3.2 Global Objective Functions

A global objective function defines how the global objective depends on the design pa-
rameters and how it changes with a change of the design variables. However, as global
property, it is independent of the spatial coordinates. It is customary to set up the global
objective function so that its absolute minimum value corresponds to the objective, or the
best design solution regarding the objective,

min {f (x)} , (3.2)

and if the objective is actually to maximize some property such as volume V , the objective
function can be set up so that minimizing it is equivalent to maximizing that property.
This is easily achieved by multiplying the property with a negative number, f (x) = −V (x)
1
or dividing a positive number by it, f (x) = V (x) . In context with evolutionary-algorithm
solution techniques, the objective function is also called a fitness function.

c ETH Zürich IDMF-CMAS, March 1, 2018

16 Design Evaluation

3.3 Constraining Functions

Generally a design is subject to one or several constraints such as discussed in section
3.2. The constraints are cast into constraining functions. Following the nature of the
considered constraints, one distinguishes between inequality constraining functions g and
equality constraining functions h:
gi (x) ≤ 0, i = 1, 2, ..., m
(3.3)
hj (x) = 0, j = 1, 2, ..., n
where l and m are the numbers of the respective constraining functions.
Constraining functions restrict the search space. Inequality constraining functions form
hyperplanes in search space dividing it into the feasible regions and infeasible regions,
where design constraints are violated. As long as the optimization variables vector points
into a feasible region, the inequality constraints are said to be inactive and then they do
not restrict the search at the momentary stage. Inequality constraints become active when
the respective constraint is violated or the limiting state is reached,
gi (x) ≥ 0. (3.4)
Equality constraints are always active.

3.4 Several Design Criteria

We have so far considered the optimization of a structure with respect to a single de-
sign criterion. If there are two or more criteria to be considered, the respective objective
functions can also be minimized simultaneously. Such optimization procedures are called
multicriteria optimization, vector optimization, multiobjective optimization. Problems of
this kind are of particular relevance to practice where, in general, several structural re-
sponse modes and failure criteria must be taken into account in the design process [18].
The form of a constrained vector optimization problem can be stated as
min {f (x)|h(x) = 0, g(x) ≤ 0} , (3.5)
X∈Rn

where f (x) is called a vector objective function of the design variables

 

 f1 (x) 

f (x) := .
. . (3.6)
 . 
fm (x)
 

At some stage during an multi-objective optimization process the situation appears that
a further minimization of one objective function goes on account of increasing some other
objective function value. Such a situation is called an objective conflict because none of
the possible solutions allows for simultaneous optimum fulfillment of all objectives.

3.4.1 Pareto Optimality

A vector is then - and only then - called Pareto-optimal if no vector xX exists for which
fj (x) ≤ fj (x∗ ) for all j ∈ {1, ...., m}
(3.7)
and fj (x) < fj (x∗ ) for at least one j ∈ {1, ...., m}

c ETH Zürich IDMF-CMAS, March 1, 2018

3.4 Several Design Criteria 17

Fig. 3.1 shows a projection from the two-dimensional design space X into the objective
function space Y . The Pareto optimal solutions lie on the lines AB. The designer may
choose from these solutions from assessment of the relative values of the two objective
functions.

Figure 3.1: Mapping of a feasible design space into the criteria space [18]

3.4.2 Substitute Problem and Preference Function or Scalarization

Multi-objective optimization problems can be reduced to optimization problems with a
scalar objective function by formulating a substitute problem with a preference function
p so that
min p [f (x)] , (3.8)
X∈Rn

such that
p [f (x̃)] = min p [f (x)] . (3.9)
X∈R\

Eschenauer [18] cites various formulations of the preference functions from which here only
the sum of weighted objectives is mentioned:

m
X
p [f (x)] := [wj fj (x)] , x ∈ Rn . (3.10)
j=1

The weighting factors wj are chosen by the designer

m
X
0 ≤ wj ≤ 1, wj = 1. (3.11)
j=1

If all objectives are convex, a full set of Pareto-optimal solutions can be generated by
running a sequence of substitute scalar problems where the preference functions cover an
appropriate range of values for the weighting factors wj . If one or several of the objectives
are not convex, the Pareto-optimal set is not so easily generated and Eckhart Zitzler
explains the topic in his lecture class Bio-Inspired Computation and Optimization at ETH
Zurich.

c ETH Zürich IDMF-CMAS, March 1, 2018

18 Design Evaluation

3.5 Transformation Methods and Pseudo Objectives

Searching for a minimum of a constrained objective function adds the burden of keeping
out of the infeasible region. Principally, when the algorithms for unconstrained search
exist, they can also be used for solving constrained problems if they are transformed into
unconstrained problems. Some of these transformation methods, namely two versions of
the penalty method and the multiplier method, are described in the following.

3.5.1 Penalty Methods

The penalty methods [19] transform the objective function f (x) and the constraining
functions h and g into a transformed objective function p without explicit constraints.
Such a transformed objective function is also called a pseudo objective and it is obtained
by adding to f a penalty function Ω that is composed of the constraining functions and
some penalty parameters R,

p(x, R) = f (x) + Ω (R, g(x), h(x)) . (3.12)

The function Ω can be defined so that either the exterior point method or the interior
point method results [20]. A disadvantage of both penalty methods shown in the follow-
ing subsections is that the respective pseudo-objective functions topologies make it more
difficult for deterministic optimization methods to perform well.

Exterior Point Penalty Method

An example for the exterior penalty method is provided by setting up the penalty function
by using the quadratic penalty term so that the constraint violations are penalized:
m
X l
X
Ω(x, R) = R {max [0, gj (x)]}2 + R {hk (x)}2 . (3.13)
j=1 k=1

Therefore, of the inequality constraining functions g(x), only the active ones may be consid-
ered in the penalty formulation (3.13). Inserting it into (3.12) results in an unconstrained
function the minimum point of which lies outside the feasible region, hence the name of
this method. As Fig. 3.2(a) shows, with increasing values of the penalty parameter R the
minimum moves closer to the feasible region but it can never quite reach it.

Interior Point Penalty Method

An interior point method is the result of selecting a form for Ω that will force stationary
points of P (x, R) to be feasible. Such methods are also called barrier methods, since
the penalty term forms a barrier of infinite P function values along the boundary of the
feasible region. Since the keeping of constraints may be crucial for safe design solutions,
it is generally to be preferred to use methods finding the improved design solutions within
the feasible region. The interior point method has that property because, for the inequality
constraints, it penalizes the closeness to the infeasible region, even before the constraints
are actually violated (barrier function). This is achieved, for instance, with the inverse
penalty term:
m
X −1
Ω(x, R0 ) = R0 . (3.14)
gj (x)
j=1

c ETH Zürich IDMF-CMAS, March 1, 2018

3.5 Transformation Methods and Pseudo Objectives 19

Note that the interior-point penalty method does not work with equality constraining
functions h as violations of these may lead to either positive or negative values. Fig.
3.2(b) illustrates that with decreasing value of the penalty parameter R0 the minimum
point of the transformed objective function moves closer to the infeasible region or the
constrained minimum point.

Illustrative Sample Problem

The objective function
1
f (x) = (x + 2)2 (3.15)
20
depends on only one variable x and has its unconstrained minimum point at x = −2. The

4 4

3 3

2 2

1 1

0 0
-2 -1 0 1 2 3 -2 -1 0 1 2 3

Figure 3.2: Examples for the exterior (a) and interior (b) penalty functions

variable x is subject to the side constraints

1<x ∧ x < 2. (3.16)

Thus, the feasible region of the optimization variable is 1 < x < 2. The two side constraints
define the inequality constraining functions

g1 (x) : 1 − x ≤ 0, g2 (x) : x − 2 ≤ 0. (3.17)

It can be seen that the minimum point of the constrained original problem is at x = 1.
The substitute problem resulting from the transformation with the outer penalty method
has its minimum point between that of the unconstrained and the constrained original
objective functions, −2 < x∗ < 1. In Fig. 3.2(a) the unconstrained function f and three
transformed functions p are plotted for the penalty parameter values 1, 10, and 100. It
can be seen how the minimum of the transformed function moves closer to x = 1 as the
penalty parameter increases from 1 to 100.
The inner penalty method obtains an unconstrained substitute function whose minimum
point lies within the feasible region. As the penalty parameter R0 decreases from 100 to 1
the minimum point moves closer to the edge of the infeasible region. It can also be seen
from Fig. 3.2(b) that it will take very small values of R0 to move the minimum of p close
to x∗ = 1.
One might conclude that solutions close to the constrained minimum point can be obtained
simply by using very high or small values for the penalty parameters R or R0 , respectively,
and just minimizing the pseudo objective for these values. However, it can be seen from
Fig. 3.2 that the unconstrained pseudo objectives are more distorted if compared to the

c ETH Zürich IDMF-CMAS, March 1, 2018

20 Design Evaluation

original objectives. The search methods of mathematical programming (section 6.3) are
designed to work best on objective functions that behave almost like quadratic functions
and might fail at the highly distorted pseudo-objectives resulting from choosing the penalty
parameters so that the minimum point of the unconstrained pseudo objective is very close
to the true constrained minimum point.
It is then inevitable that the constrained optimization problem is solved by a sequence
of unconstrained subproblems, where the penalty parameters are updated at each step.
Considering the exterior point method, the parameter R is chosen small, for instance zero,
at the first stage, and gradually increased with the subsequent stages. For the interior
point method, one starts with a high value of R0 and decreases it from stage to stage. The
minimum point of each subproblem is then used as a starting point for solving the next
subproblem. So the considered regions in search space become smaller as the pseudo objec-
tives become more distorted. Nevertheless it is inevitable that the generated subproblems
become progressively ill-conditioned so that, at one point, the sequence terminates not
because of finding a very close approximation to the true constrained minimum point but
because of failure of the search algorithms.

3.5.2 Method of Multipliers

The method of multipliers (MOM) [21, 22, 23] uses the penalty function
J n
X o K n
X o
P (x, σ, τ ) = f (x) + R hgj (x) + σj i2 − σj2 + R [hk (x) + τk ]2 − τk2 (3.18)
j=1 k=1

where R is a constant scale factor (R may vary from constraint to constraint but remains
constant from stage to stage), and the bracket operator is defined as

α if α > 0
hαi = . (3.19)
0 if α ≤ 0
The σj and τk parameters are constant during each unconstrained minimization but are
updated from stage to stage. It is not necessary for the starting vector x0 to be feasible,
and the parameters can be conveniently chosen for the first stage as σ = τ = 0. Thus the
first minimization stage is identical to the first unconstrained minimization using standard
exterior point method penalty terms.

Multiplier Update Rule

Suppose that the vector xt minimizes the tth -stage penalty function:
D 
(t) 2
E
(t) (t)
PJ (t) 2
P (x, σ , τ ) = f (x) +R j=1 gj (x) + σj − [σj ]
h  (3.20)
(t) 2
i
PK (t) 2
+R k=1 hk (x) + τj − [τk ]

Multiplier estimates for the (t + 1)st stage are formed according to the following rules:
D E
(t+1) (t)
σj = gj (x(t) ) + σj j = 1, 2, 3, ...J
(3.21)
(t+1) (t) (t)
τj = hk (x ) + τk k = 1, 2, 3, ...K
Because of the bracket operator, σ has no negative elements, whereas the elements of τ
can take either sign.

c ETH Zürich IDMF-CMAS, March 1, 2018

3.5 Transformation Methods and Pseudo Objectives 21

3.5.3 Unconstrained Lagrange Problem Formulation

An optimization problem with linear equality constraints can be transformed into an
unconstrained substitute problem by introducing the Lagrange function L:
min {f (x)|h(x) = 0} → min {L(x)} , L = f + λT h (3.22)
X∈Rn X∈Rn

Suppose that improvement is sought by moving from a reference point, which must be
feasible, along a search direction s. The search direction is the linear combination of

N
Ñ D
I ×Ñ D = 
I
Ñ B

D =  I = - Ñ B - l Ñ D

N 

Figure 3.3: Illustration to the unconstrained Lagrange problem formulation

a usable direction, along which smaller values of f are found and an obvious choice of
which is the direction of steepest descent, and the gradients of the equality constraining
functions:
s = −∇f − λT ∇h (3.23)
The search direction is feasible, or will not violate the linear constraint, if it is orthogonal
to the constraining function gradient:
sT ∇h = 0 (3.24)
Inserting the definition (3.23) of the search direction into the orthogonality condition (3.24)
yields the Lagrange factors λi :
∇f T ∇hi
λi = − (3.25)
∇hTi ∇hi
The Lagrange function, or Lagrangian, is the function whose negative gradient provides a
feasible search direction obeying all constraints.
m
X l
X
L(x, λ) = f (x) + λj gj (x) + λm+k hk (x) (3.26)
j=1 k=1

More information on the method of feasible directions, the Lagrangian, and the Kuhn-
Tucker conditions for constrained optimization problems is given in Sections 5.4 through
5.6. Section 5.5 on page 55 explains why the stationary point of the Lagrangian is a
saddle point and Section 5.7 on page 59 introduces the concept of duality, where the dual
problem is that of solving the constrained optimization problem in terms of the Lagrange
multipliers. Section 6.9 considers numerical solution of the Lagrangian by searching for a
minimum. If the constraining function g is not linear, the found search direction is feasible
only at the reference point and moving along it will eventually violate the inequality
constraints. It will then be necessary to remove the violations, or find a feasible point
close to the infeasible one just obtained. Such an algorithm is explained in Section 6.9.2.

c ETH Zürich IDMF-CMAS, March 1, 2018

22 Design Evaluation

3.6 Fitness Function for Evolutionary Algorithms

Evolutionary Algorithms evaluate a given design by a fitness function. The term is in
context with Genetic Algorithms, which are inspired by the understanding of the mecha-
nisms of biological development under environmental pressure. The analogy lead also to
distinguishing between phenotype and genotype. The phenotype refers to the parameters
of the real design while the genotype results from transforming the phenotype parameters
into a form upon which the genetic algorithms can work efficiently. In mimicry of nature’s
working, genotype representation used to be in terms of binary bit strings. For this reason,
the symbol x, familiar to those working with mathematical programming, is replaced by
p or g, to point out that phenotype or genotype variables, respectively, are meant. The
fitness function is a sum of products of weights w and demands D.
X
F (p) = wi Di (p) (3.27)
i

The demands represent ratings for one or several objectives and constraints so that the
fitness function F appears on first sight analogous to the pseudo-objective function P
explained in section 3.5.1.
Recent research [6, 5] has elaborated schemes for defining the ratings in such a way that
the problem of finding appropriate weight values dissolves. The following is direct citation,
or uses material, from Oliver König’s Ph.D. thesis [5].

In order to avoid that one of these terms becomes much larger than the other
ones and therefore dominant, only bounded functions scaled to the interval
[0, 1] are used. Moreover, this facilitates the adjustment of the weight coeffi-
cients wi . Further, to enhance general usability of the fitness formulations, the
mapping functions Di (~ p) are defined range-independent. This means that a
certain mapping function does not change its behavior if applied to objectives
or constraints operating in different number ranges.
Based on these requirements, functions Di for the different possible types of
optimization objectives and constraints are presented. The focus for the formu-
lation of these mapping functions is put on good practical usability. The user
of an evolutionary design optimization program shall be able to define good
mapping functions by only bringing in know-how about the problem he wants
to solve. Thus three types of general mapping functions are defined with their
defining parameters as listed in Table 3.1. The defining parameters are chosen
so that they relate directly to engineering practice. For a problem at hand,

Table 3.1: General types of fitness functions with defining parameters.

Type Parameters
Required Optional
Design objective Oinit , Oestim αpost
Upper/lower limit constr. Climit , Cfeas tol -
Target constraint Ctarget , Cadm tol , Cfeas tol -

the parameters Oinit and Oestim refer to the initial value and the estimated
best-possible value of the design objective respectively. The design objective
function can optionally be modified using an amplification factor αpost . A

c ETH Zürich IDMF-CMAS, March 1, 2018

3.6 Fitness Function for Evolutionary Algorithms 23

limit-value constraint is defined through Climit for the given limit value, and
through a parameter Cfeas tol specifying a tolerance range for constraint values
still considered feasible for the problem at hand. Finally there are target-value
constraints defined through three parameters. Ctarget defines the target value
to be achieved. It is useful to specify an admissible tolerance Cadm tol , defining
an interval for the target value to be reached at the end of the optimization.
Additionally a feasible value tolerance Cfeas tol should be specified, defining
which constraint values should still be considered during optimization.
In the following, fitness functions for the different types and parameters are
presented.

3.6.1 Mapping Functions for Objectives

Be O (~
p) a measure of the objective of a design optimization problem, as for
instance the mass or the compliance of the considered structure. Then, one
can define (
0
O (~
p) : if O (~
p) to be minimized
O (~
p) = (3.28)
−O (~p) : if O (~
p) to be maximized
and therefore O0 (~
p) is always representing a value to be minimized. The func-
tional mapping Di (O0 ) should satisfy the following requirements:

1. The resulting fitness values must fit into the interval [0, 1].
2. Relevant design improvements should be reflected in distinct decreases of
Di (O0 ).
3. Selection pressure is initially strong and slows down close at the optimum.

To meet these requirements, the mapping function is defined as

α
Di O0 = aO0 + b


(3.29)

where the choice of the exponential factor α = 5 is based on experience, and a

and b are scaling factors defined by the conditions

Di (O = Oinit ) = 1
(3.30)
Di (O = Oestim ) = 0.1

Oinit represents an initial value of the design objective, which shall result in
the maximum fitness value 1. Oestim is the estimated goal value that can be
achieved in the optimization. The fitness value 0.1 for Oestim was adjusted to-
gether with an exponential factor α = 5 in order to fulfill the third requirement
defined above. Furthermore, a small fitness value for the estimated objective
also ensures conformance with the second requirement defined. The scaling
factors a and b can be computed as:
√
1 − α 0.1
a = (3.31)
Oinit − Oestim
b = 1 − aOinit

The user has only to specify Oinit and Oestim to define the fitness function for
a design objective of a problem at hand. The bold line in Fig. 3.4 pictures an

c ETH Zürich IDMF-CMAS, March 1, 2018

24 Design Evaluation

O b je c tiv e V a lu e
5 0 6 0 7 0 8 0 9 0 1 0 0 1 1 0
1 .2

1 .0

0 .8

0 .6
F itn e s s

0 .4 3
4
0 .2 5
6
0 .0 8
1 0
-0 .2

Figure 3.4: Fitness function for a design objective defined through Oinit and Oestim .

example of the fitness function computed for Oinit = 100 and Oestim = 60. The
graph demonstrates that the three given requirements for the fitness function
are met for arbitrary ranges of objective values. As a last tuning parameter
αpost is introduced. Leaving the scaling parameters a and b constant, the
exponential factor α can be varied subsequently to adapt for the problem at
hand as presented in Fig. 3.4. Usually these variations will be of negligible
influence on the overall performance of the algorithm.

3.6.2 Constraint Mapping Functions

The community of evolutionary algorithms use for constraints as well a different
vocabulary than the community of mathematical programming. The inequality
constraints become upper and lower limit constraints and equality constraints
are called target value constraints.

Upper and Lower Limit Constraint Mapping Functions

For constraint values C (~p) that are not allowed to exceed/fall below a certain
value Climit , the most straight-forward approach would be to define a simple
step function of the form
(
0 : C (~p) ≤ Climit
Di (C) = (3.32)
1 : else

With this penalty function a strong selection pressure is exerted towards so-
lutions that meet the constraint. On the other hand, the constraint function
makes no distinction between values that hurt the constraint only marginally
and values that clearly violate the restriction. However the former values can
contain valuable information for the problem and should therefore not be ex-
cluded strictly. In order to take this fact into account, a smoothed step function
is defined as
1
Di (C) = −λ(C(~p)−Climit −∆)
(3.33)
1+e

c ETH Zürich IDMF-CMAS, March 1, 2018

3.6 Fitness Function for Evolutionary Algorithms 25

allowing the algorithm to also extract valuable information from solutions with
marginally violated constraints. The step function is controlled with two pa-
rameters: ∆ adjusts the horizontal positioning of the function, and λ deter-
mines the steepness of the step. However, the parameters ∆ and λ are difficult
to adjust correctly, since they depend on the range of occurring constraint val-
ues. Furthermore it is difficult to estimate their quantitative effect on the step
function.
Based on these findings, another definition of the step function has been de-
veloped. For practical purpose it would be much more comfortable to define
the step function only with its limit value Climit , and an additional tolerance
Cfeas tol giving an upper limit of feasible constraint values to be taken into ac-
count by the EA. A method to define the step functions that way is to specify
two conditions

Di (C = Climit ) = Dlimit (3.34)

Di (C = Climit + Cfeas tol ) = Dfeas

where Dlimit is the penalty value typically reached at the end of an optimiza-
tion, where the EA has found an equilibrium between the different demands
of the fitness function. Dfeas corresponds to the penalty value which is typi-
cally still taken into account by the EA. For the normalized fitness formulation
used within this thesis, Dlimit = 0.01 and Dfeas = 0.5 proved to be reasonable.
With these two conditions given, the original parameters of Equation 3.33 can
be computed as
    
1 1 1
λ = ln − 1 − ln −1 (3.35)
Cfeas tol Dlimit Dfeas

 
1 1
∆ = ln −1
λ Dlimit
Fig. 3.5 presents the mapping functions for a critical value given as Climit = 60,
and for tolerance values Cfeas tol = 0.6...6. This corresponds to a 1 − 10%
tolerance of the limit value Climit . For this formulation, it has to be Cfeas tol ∈
R+ for upper limits and Cfeas tol ∈ R− for lower limits, respectively.

Target Value Constraint Mapping Functions

A formulation for constraints where a parameter has to achieve a given target
value Ctarget , as e.g. a defined stiffness of a structure, is discussed. For real-
valued parameters it makes sense to define a tolerance Cadm tol leading to an
acceptance interval where the resulting constraint values should fit in. Further
smooth transitions are defined, similar to the formulations for limit constraint
values. Therefore, the function is defined as

 0 : |C(~
p) − Ctarget | < Cadm tol
Di (C) = p)−Ctarget |−Cadm tol )
(|C(~ 2

1 − e− 2σ 2 : |C(~
p) − Ctarget | ≥ Cadm tol

(3.36)
where C(~ p) refers to the actual constraint value. The parameter σ determines
how fast the penalty value increases when the acceptance interval is left. In this

c ETH Zürich IDMF-CMAS, March 1, 2018

26 Design Evaluation

C o n s tr a in t V a lu e
4 0 4 5 5 0 5 5 6 0 6 5 7 0 7 5 8 0
1 .2

1 .0

0 .8

0 .6
F itn e s s
0 .4 0 .6
1 .0
0 .2 2 .0
c to l 3 .0
0 .0 4 .0
6 .0
-0 .2

Figure 3.5: Upper limit constraint penalty functions defined through Climit and Cfeas tol .

form, the penalty function can be used to solve practical problems. However,
the parameter σ depends on the number range of the constraint considered,
and is therefore very difficult to adjust for a problem at hand. To paraphrase
a constraint function defined through the parameters given in Table 3.1, an
additional condition is introduced.

Di (C = Ctarget + Cfeas tol ) = Dfeas (3.37)

As introduced before, the feasible value tolerance Cfeas tol determines which
constraint values should still be taken into consideration during optimization.
Dfeas corresponds to the penalty value which is typically still taken into account
by the EA. For practical purpose, Dfeas = 0.5 proved to be reasonable. The
initial parameter σ can now be determined as

(Cfeas tol − Cadm tol )2

σ2 = (3.38)
−2 ln (1 − Dfeas )
Fig. 3.6 shows examples for this penalty function where a target value of 60±5
shall be achieved. Therefore Ctarget = 60 and Cadm tol = 5 are set, and the
feasible tolerance is varied with Cfeas tol = 6 − 15.

c fe a s _ to l

c a d m in _ to l = 5

Figure 3.6: Penalty functions for a target constraint

c ETH Zürich IDMF-CMAS, March 1, 2018

3.7 Design Evaluation Exemplified on Selected Problems 27

3.7 Design Evaluation Exemplified on Selected Problems

Chapter 1 presents selected structural optimization problems and specifies the design ob-
jectives and constraints. This section discusses the objectives and constraints. The form
of objective and constraining functions depends on the optimization algorithms which are
chosen for problem. The methods of Mathematical Programming function with the formu-
lations given in Sections 3.1 through 3.5 while Evolutionary Algorithms draw advantage
from those provided in section 3.7. Table 3.2 tells whether the respective sample problem
is solved with mathematical programming or evolutionary algorithm optimization engines.
The explanations to the problems solved with evolutionary algorithms are taken from the
recently finished Ph.D. work of O. König [5] and M. Wintermantel [6] and the ongoing
research of N. Zehnder. The constraining function formulations depend sometimes on the

Table 3.2: Problems and solution methods.

Problem Solution Method

Motorcycle Frame Evolutionary Algorithms

Racing Car Rim Evolutionary Algorithms
Flywheel Mathematical Programming
Onsert Mathematical Programming
CFRP Boat Hull Evolutionary Algorithms
End Plate Evolutionary Algorithms

parameterization and cannot be considered separately. In these cases, reference is made

to the respective sections in chapter 4. Side constraints giving lower and upper bounds
on analysis or design variables are directly discussed in the respective sections of chapter
4 on parameterization and variable transformations.

3.7.1 Weight Minimization of a Motorcycle Tubular Frame

The motorcycle frame problem is presented in section 1.3.1 and the evaluation model pre-
sented here is being published [3]. The objective is minimum weight. As a constraint,
the initial structural torsional stiffness of the frame must be the same as in the optimized
design. Another constraint is the strength of the frame under various selected severe load
cases. Further constraints come from the manufacturing considerations regarding the de-
sign parameters: it is intended to make the frame from a selection of standard size tubes
or a limited number of customized tubes. The problem is thus used to demonstrate a
search method based on Evolutionary Algorithms to solve the discrete problem.
The objective of minimum weight and the constraints are transformed into a fitness func-
tion of the form (3.27). Normalizing the weight W (g) with respect to the initial weight
Winitial gives
W (g)
Dweight (g) = . (3.39)
Winitial
The chassis is to keep the torsional stiffness value of the existing initial design, which is
included in the fitness function by the target-value constraint mapping function 3.36. The
frame must withstand the applied loads so that it must be demanded that the maximum
equivalent stress value σeqvmax (g) is less than the critical value σcrit . This is achieved by
the upper-limit-constraint mapping function (3.32). The chassis is a space truss design

c ETH Zürich IDMF-CMAS, March 1, 2018

28 Design Evaluation

F B

M T

F B

Figure 3.7: FEM model of the frame and load cases, courtesy =. König[5]

and the tubes are suffering mainly loads in the axial directions. However, since they
are welded together at the joints, bending will also contribute slightly to the structural
stiffness. Thus, the fitness of the design is more sensible to changes of the cross-sectional
areas of the tubes but depends also slightly on their second moments of inertia. The area
A, depending on the geometrical parameters diameter D and wall thickness t, influences
both the weight and the extensional stiffness linearly. Given a certain area value, the
bending stiffness increases quadratically with increasing diameter:

1
I = AD2 . (3.40)
8
Due to the influence of bending stiffness on the fitness tubes will tend to grow to large
diameters and small wall thicknesses. It is more difficult to weld tubes with small than
those with large wall thickness values. Tubes with smaller diameter and larger wall thick-
ness are therefore preferable. A further term is introduced to the fitness function, giving
tubes with thin walls a slight penalty,
PNtubes 1
i=1 (αi (ti (g)−tmin )+1)βi
i
Dthick (g) = , (3.41)
Ntubes

where ti (g) is the actual thickness and tmin

i the minimum allowed thickness of the ith tube.
The factor αi is used for scaling and βi adjusts the severity of the penalty.
Thanks to the normalized transformation functions developed by König [5], the weight
factors wi of each term in (3.27) can be chosen unity.

3.7.2 Racing Car Rim Design Evaluation

The optimization task is to improve the bending stiffness of an existing design. The
bending stiffness is measured by applying loads and relating them to their conjugate
displacements at the points indicated in Fig. 3.8. The rim design is subject to several
types of constraints:

• mass and inertia constraints

c ETH Zürich IDMF-CMAS, March 1, 2018

3.7 Design Evaluation Exemplified on Selected Problems 29

B2
B1

Figure 3.8: Rim compliance measure (courtesy O. König [5])

• strength constraints

• geometric constraints

The mass and the moment of inertia around the wheel’s axis of the existing design are low
and the same values should be attained by any new design. This is achieved by using the
target-value constraint mapping function 3.36. The complex geometry makes it necessary
to use CAD as well as FEM to simulate and evaluate the design. Therefore, the mass and
moment of inertia can be evaluated by the CAD system but the bending-stiffness objective
and the strength constraints require costly FEM analysis.
For evaluating the strength constraints, the load case combined from the loads indicated in
Fig. 3.9 is considered. The loads are applied to the rim’s shoulders – since this is the only
area of contact with the tire – and include horizontal (car’s weight plus aerodynamic force),
vertical (centripetal), and rotational forces (braking). Essential boundary conditions of

Figure 3.9: Loading of the Rim (courtesy O. König [5])

the analysis model simulate clamping at the contact area with the car’s suspension. The
margin of safety for mechanical stresses is defined as
σ
msaf ety = − 1, (3.42)
σyield

where σ is the acting stress and σyield is the yield stress of the material at a given tem-
perature. Values less than zero indicate plastic yield to occur. The margin of safety is
calculated for every node, a counter S(p) records the number of nodes with msaf ety < 0.

c ETH Zürich IDMF-CMAS, March 1, 2018

30 Design Evaluation

The geometric constraints stem from

• manufacturing constraints

• assembly constraints

• regulatory constraints (FIA1 )

and are kept by the parameterization of the design model, which will be explained in
Section 4.2.2.
The manufacturing constraints include the given shape of the forging blank. A 1mm
distance to its outer contour must be kept to allow for properly machined surfaces. In
addition, manufacturing techniques must not be changed. Finishing is restricted to a
CNC-lathe and a CNC-mill. Also, the rim’s bed wall thickness must not be thinner than
2mm.
The assembly constraints include contact areas to the car’s suspension as well as to the
nut holding the wheel and the tire. An additional constraint is a 3mm distance to the
brake assembly positioned on the inside of the rim.
The FIA regulations applying to the rim include that the maximum bed diameter is limited
to 330mm. The minimum depth of the lower bed is 13.57mm and the maximum distance
from the outside surface is 43.3mm.

3.7.3 Maximum-Strength Flywheel Design

Kinetic Energy and Stressing of a Flywheel

A flywheel of constant mass density ρ and a radius-dependent thickness distribution t(r)

stores at an angular speed ω the kinetic energy
Z
U = ω 2 πρ r3 t(r)dr. (3.43)
r

The inertia effects induce a body force distribution

f = trω 2 πρ (3.44)

so that the equilibrium of forces in the radial direction requires

t
(σr t) ,r + (σr − σθ ) + trω 2 ρ = 0. (3.45)
r

The equilibrium equation 3.45 shows that the radial and the circumferential stresses, σr
and σθ , increase quadratically with increasing rotational speed ω. Thus, the rotational
speed, and with it the kinetic energy, of the flywheel are limited by the maximum stress
that the material can bear.
As outlined in section 1.3.3 and shown in Fig. 1.4, the distribution of stresses in a flywheel
of constant thickness and with a central bore is uneven. Particularly, the circumferential
stress increases steeply towards the edge of the bore. Failure will initiate at the bore
although other regions are not critically stressed.
1
Fédération International d’Automobiles ([Link]

c ETH Zürich IDMF-CMAS, March 1, 2018

3.7 Design Evaluation Exemplified on Selected Problems 31

Objective and Constraint Formulations

Assume the objective of maximizing the kinetic energy storage capability of a flywheel.
Let it also be assumed that both, the total mass and the rotational mass inertia, of the
initial design with constant thickness t0 must be kept constant. Then, the kinetic energy
capability can only be maximized by increasing the failure rotational speed of the flywheel.
This, in turn, can only be done when the maximum stress can somehow be minimized:

minimize max {σ(r)} . (3.46)

Let the symbol σ stand for a failure criterion such as maximum principal stress σI or von
Mises equivalent stress σeqv . The problem with evaluating the objective (3.46) is that not
only the stress values at all points along r must be calculated but that the location at
which the maximum stress appears may change when the design changes. This causes the
first and second derivatives of the objective function to be discontinuous. Consequently,
one would be forced to use genetic algorithms instead of a mathematical programming
technique which would increase the numerical solution effort considerably.
A continuous objective function with continuous derivatives can be constructed by the
following argument. The maximum stress at some point in the flywheel is minimized when
its value equals the minimum stress at some other point. Then, the stress distribution
must be a constant. This idea was expanded by Stodola [24] who derived analytically
the shape of a turbine disk of constant strength. The constant stress distribution implies
that the local stress σ(r) is everywhere equal to the mean stress σ̄ and that therefore its
variance (average quadratic deviation) is zero:
Z
σ(r) = σ̄ → s = r (σ(r) − σ̄)2 dr = 0. (3.47)
r

In reality a perfectly spatially constant stress state can not be reached for physical reasons:
the surfaces perpendicular to the radial direction at the bore and at the outside are stress-
free and the radial direct stress must drop off to zero. This causes the stress distribution
inevitably not be a constant. However, a global objective function f can be based on
(3.47) so that the objective is reached by minimizing
Z
f = r (σ(r) − σ̄)2 dr. (3.48)
r

Now let us consider again the two assumptions of the mass and the rotational moment of
inertia remaining constant during the optimization process. They constitute two equality
constraints which can be written in terms of constraining functions h1 and h2 ,
Z
h1 = 2πρ r (tr − t0 ) dr = 0 (3.49)
r

and Z
2
h2 = ω πρ r3 (tr − t0 ) dr = 0. (3.50)
r

c ETH Zürich IDMF-CMAS, March 1, 2018

32 Design Evaluation

Flywheel Evaluation Model

The objective function (3.48) and the constraining functions (3.49) and (3.50) must be
evaluated from the analysis model. The analysis model is based on the finite-element
method and delivers a numerical system of equations the solution of which gives the nodal-
point displacements. This primary solution is used on element-level to calculate stresses
from which some equivalent stress is formed. Since the stresses are numerically evaluated
at the discrete stress points, the integrals in the objective and constraining functions must
be replaced by sums. When the finite element mesh is equidistantly spaced in the radial
direction as indicated in Fig. 3.10, it is not necessary to attach individual weights to the
stress values.

H  E

Figure 3.10: Region out of the Flywheel Analysis Model with Nodal and Optimum Stress
Points of Quadratic Serendipity Type Finite Elements

3.7.4 Maximum Bond-Strength Design

Objective and Constraint Formulations
The objective is to maximize the load that can be transferred by the bonding layer from
the onsert into the substrate. Failure within the bond layer is predicted by evaluating
equivalent stress criteria for thin-layer adhesives.
Material failure is a local affair and therefore its prediction requires the evaluation of the
stress states at all points within the adhesive domain. The location where the highest
failure probability occurs may change when changing the onsert’s shape. In order to
have a smooth objective function that can be very efficiently minimized by using gradient
methods, the original objective of minimizing the maximum value of a failure criterion
at any point in the bonding layer is recast into a global objective functional [2]. The
functional penalizes the variance of some equivalent stress σeqv along the radial direction,
Z r2
1 p
f= (σeqv − σ eqv )2 dr, (3.51)
r2 − r1 r1

from the mean stress σ eqv . For p = 1, the objective function becomes the variance (square
of the standard deviation), of the stress distribution. Higher values of p further penalize
stress peaks so that the absolute values of minimum and maximum stress deviate less from
the mean stress. A value of p = 4 has been used for the sample calculations. The stresses
are evaluated at the optimum stress points of each finite element in the bonding layer.
The parameterization of the onsert problem will be described in section 4.2.4. However,
the shape optimization, to maximize the bond strength, regards the thickness distribution
of the onsert. There it is assumed that the interface between the onsert and the bonding

c ETH Zürich IDMF-CMAS, March 1, 2018

3.7 Design Evaluation Exemplified on Selected Problems 33

layer remains straight, preserving the initial constant thickness distribution of the bonding
layer. Thus, the thickness t(x) = p2 (x) − p1 depends only on the vertical position of the
points p2 (x) on the upper onsert surface. The thickness must be positive and greater than
or at least equal to some predefined small minimum-thickness value t which gives the
constraining function
g(x) = t − t(x) = t + p1 − p2 (x) ≤ 0 (3.52)

Onsert Problem Evaluation Model

It is quite obvious that the analysis model of the onsert problem has much in common
with that of the flywheel problem: both are based on the rotational-symmetry assumption
and they use the same finite-element formulations.
The stresses threatening the cohesion of the bonding layer material are evaluated at the
optimum stress within each finite element of the bonding layer. They are then inserted
into a failure criterion for thin layers of bonding materials.
As explained in section 4.2.4, considering the nature of the stress distributions indicated in
Fig. 1.5(d), it is a good idea to have the mesh density increase at the inner and the outer
onsert edge as indicated in Fig. 1.5(c). In order to approximate the integration required
by (3.51), it is therefore necessary to weigh the stresses at the stress points with the width
of the element wherein the stress points are located.

3.7.5 Composite Boat Hull

Evaluation of objective and constraints is again based on a combination of CAD and FEM.
The objective is to maximize stiffness which is the same as minimizing the displacements
conjugate to the loads indicated in Fig. 3.11. It is mapped onto a demand D by using the

Figure 3.11: ANSYS model of the sail boat hull with composite material patches

mapping function 3.28. The constraints include

• upper limit on mass

• upper limit on cost

• sufficient strength
The parameterization of the hull is explained in Chapter 4 and allows the laminate con-
struction to change over the hull’s shell area. It is foreseen that the hull is made from
glass-fiber as well as carbon-fiber reinforced prepreg material. Carbon fibers are much

c ETH Zürich IDMF-CMAS, March 1, 2018

34 Design Evaluation

more expensive than glass fibers and so the material cost responds to the total amount
as well as the amount ratio of the two types of material. However, under the given mass
constraint the carbon fiber may be more effective, if used in certain regions, because of its
low mass and high stiffness properties. The upper limits on mass and cost are expressed
as demands D by using the mapping function 3.32. Their evaluations do not require FEM
analysis.
The finite element model is needed for the stiffness and strength evaluations. It is assem-
bled from shell elements. Their extensional and bending stiffness depends on the local
laminate construction which is given by the number of layers, their respective thicknesses
and the materials and orientation of principal material axes used for each layer, and is
evaluated by the theory of laminates plates [25]. Once the primary unknowns of the finite-
element model are known, the stresses and failure probabilities of the laminate can be
evaluated.

3.7.6 Minimum-Weight Fuel-Cell Stack End Plate

Commercial versions of fuel cell sources of electric power for cars should be light and
inexpensive. An idea for reaching the objective of low manufacturing costs is to use an
extruded-aluminum-profile design [17]. A quarter model of it is rendered in Fig. 3.12. The

sym
sym

Fb

ps

Figure 3.12: Quarter end-plate model in the initial design (courtesy [5])

other objective, minimum weight, is reached through an automated optimization process

where the parameters of the cross-sectional design are adjusted. The necessity to have an
almost constant pressure distribution over the cross-sectional area of the bipolar plates
is considered another objective. The other constraint follows from demanding that the
maximum von Mises stress anywhere in the plate domain not exceed an allowable value.
The objective and constraining function formulations are closely connected with the chosen
design variables so that the design model is presented here.

Weight and Gas-Tightness Objectives

Since the plate is homogeneous, the total weight is given by the sum of the weights of all
finite elements,
Nel
X
W =ρ Vel . (3.53)
k=1

The other objective of having an even pressure distribution on the bipolar plate cross-
section areas is reached independently of, or after completing, the optimization procedure.
In the middle of the stack, the cross section will be plane because of symmetry. When
the other cross sections also remain plane, the interface between the fuel-cell stack and
the end plates should be so as well. Then, the stack is under a state of plane strain

c ETH Zürich IDMF-CMAS, March 1, 2018

3.7 Design Evaluation Exemplified on Selected Problems 35

but the pressure will not necessarily by constant everywhere because of Poisson’s ratio
effects. However, the animus behind the constant pressure idea is that the bipolar plates
be gas tight and one may expect that that goal is equally well reached by demanding the
longitudinal strain to be constant. A structural analysis modelling the optimized design
will reveal the out-of-plane displacement field in the interface between the end plate and
the stack. The displacement field is interpreted as a shape and the end plate’s bottom
surface will then be given the negative of that shape. The displacements under load will
then cancel out that shape so that the bottom becomes plane, consistent with a plane
state of the stack. The surface shape variations are expected to be small enough so that
there effect on the bending stiffness is negligible, decoupling the objective of gas tightness
from the minimum weight objective and stress constraints.

Stress Constraining Equations

The stress constraint gk = σeqv(k) − σmax ≤ 0 is evaluated for each finite element, giving a
number of constraints that equals the number of finite elements contained in the model.

Optimization Problem Statement

It remains the single-objective constrained optimization problem

min {W (x)|g(x) ≤ 0} . (3.54)

x∈Rn

c ETH Zürich IDMF-CMAS, March 1, 2018

36 Design Evaluation

c ETH Zürich IDMF-CMAS, March 1, 2018

Chapter 4

Parameterization and Variable

Transformations

4.1 Classification of Design Variables and Structural Opti-

mization Problems
Design parameterization defines the fixed design parameters and the design variables. The
design variables may describe the configuration of a structure, element quantities such as
cross sections, wall thicknesses, shapes, and physical properties of the material. Eschenauer
[18] classifies structural optimization problems in terms of their design variables. Consid-
ering a truss structure, and following [26] and [27], possible design variables can be divided
into the different classes indicated in Fig.

a ) C o n s tr u c tio n

b ) T o p o lo g y

M a te r ia l
c ) S te e l A lu m in u m C o m p o s ite
P r o p e r tie s

G e o m e try
d )
o r S h a p e

e ) S u p p o rt
o r lo a d in g

f) C r o s s s e c tio n
o r s iz in g

Figure 4.1: Classification of design optimization problems for truss-like structures in terms
of different types of design variables, after Eschenauer [18]

c ETH Zürich IDMF-CMAS, March 1, 2018

38 Parameterization and Variable Transformations

a) Constructive layout
The determination of the best suited layout requires optimizing each layout coming
into consideration and comparing the calculated optimum solutions.

b) Topology
The topology or arrangement of the elements in a structure is often described by
parameters that can be modified in discrete steps only. Different topologies can
also be obtained by eliminating nodes and linking elements. Note also the topology
optimization method introduced by Bendsøe and Kikuchi [12] explained in section
9.1.

c) Material properties
The material properties of isotropic building materials such as steel or aluminum
describe the stiffness in terms of Young’s modulus or Poisson’s ratio, the strength
in terms of the yield stress or other strength limit, or the weight in terms of specific
weight or mass density. The designer can often select the most suitable material
from a selection of alloys and sometimes he may decide whether steel, aluminum
or any other type of metal alloy would best meet the requirements of the design
objective. All of these choices are discrete in nature, leading to a discrete design
variable set in terms of the materials contained in a data base. Only laminates made
from anisotropic composite materials have some continuous design variables in terms
of fiber orientation.

d) Geometry and Shape

The geometry of trusses or frames is described by the coordinates of the nodes. The
shape of solid bodies is determined by their bounding surfaces.

e) Supports and loading

Often a design may be enhanced by changing the geometric and the natural boundary
conditions. Process optimization, for instance in the case of injection molding, is
usually dealing with optimum adjustment of boundary conditions such as injection
flow speed or injection point location.

f) Sizing
Structures in terms of members such as bars, trusses, beams, plates, or shells and also
their FEM models offer properties such as thickness, cross-sectional area, moment
of inertia, or thickness to be used as design variables. It is important to distinguish
between independent and dependent design variable. When a cross-section geometry
is determined by some variable, the geometrical properties listed above are depen-
dent on that variable. Sizing usually leads to a discrete optimization problem as
commercially available members with I-sections or channel sections come in different
discrete sizes.

The foregoing classification is exemplified on truss-like structures but remains valid for
shells or solid body structures. A number of design parameterization models and trans-
formations between design and analysis variables are discussed on the basis of the sample
structural optimization problems used for the lecture class.

c ETH Zürich IDMF-CMAS, March 1, 2018

4.2 Design Parameterization Sample Problems 39

4.2 Design Parameterization Sample Problems

4.2.1 Motorcycle Frame and Sizing Parameters
The design freedom is limited by the consideration that some key measures such as the
spatial arrangement of the nodes should not be changed. The remaining parameters regard
the sizing of the tubes. Deciding that tubes have a circular cross-section geometry, the
inner and the outer radii are used to determine the properties area and second moments of
inertia which are needed to simulate the structural stiffness behavior of the frame. Tubes
are available in standard sizes or they can be custom made. Considering standard tubes,
t= 0 .5 m m t= 0 .8 m m t= 1 .0 m m t= 1 .2 5 m m t= 1 .5 m m t= 1 .7 5 m m t= 2 .0 m m t= 2 .5 m m

ra= 1 5 m m

ra= 2 0 m m

ra= 2 5 m m

Figure 4.2: Standard Tube Sizes after DIN 2394

one can choose from a catalogue or a set of existing discrete design variable values.
Considering custom made tubes, one can freely adjust the inner and outer diameters as
continuous variables but the number of the customized tube sizes should be small, say
three. All tubes of the frame should then be chosen from the small number of custom
made tubes and when it is not predefined which of the tubes are to be of the same size,
a choice must be made. That choice again renders the optimization problem discrete.
The structural model must be so that all finite elements along one of the tubes must

1 1
2 3

2 15 9
1 1 2
2 1 2 14
2 1 2 1
2 2 8
1 2 12 13
2 4 7
3 1
1 2 2
1 3 6
2 11
1 10
2 1 1 3
2 5
1

2 2
Tube types: 1 : ri= 12.5mm, t=1.5mm
1
2 : ri= 9.5mm, t=1.5mm
3 : ri= 6.75mm, t=1.5mm

Figure 4.3: FEM-Model and Parameterization of the Motorcycle Frame, after [5, 6]

have the same properties determined by the respective design variable. Consequently, the
independent design variables must be transformed into the dependent geometric element
cross-section property value variables, and these must be assigned to all the elements
constituting the respective tube. Also, the frame must be symmetric with respect to the

c ETH Zürich IDMF-CMAS, March 1, 2018

40 Parameterization and Variable Transformations

midplane spanned by vertical and longitudinal directions so that each tube on one side of
the midplane must have the same dimensions as the respective tube on the other side of
it.

4.2.2 Shape Parameters of the Formula 1 Race Car Rim

This sub section is taken from [5]. Four substructures (features) build up the basic geom-
etry of the CAD model as shown in Fig. 4.4:

• A rotational body for the rim’s bed.

• A second rotational body for the spokes.

• Two pockets that remove the spokes’ interspace.

Half of the rim bed. Spoke body with the two pocket features subtracted

Figure 4.4: Structure of the CAD model of the rim.

All these features are built on fully parametric two-dimensional sketches. In addition to
these basic features, various chamfers are applied, chosen completely parametric as well.
The basic design is taken from an existing racing car rim.

Parameterization
For performance reasons as outlined before, a parameterization including implicitly as
many of the mentioned constraints as possible has to be found, without excluding relevant
feasible solutions. The bed is parameterized by 9 wall thicknesses including the inner and
the outer bead as shown in Figure 4.5. With this approach the manufacturing constraint
of a minimum wall thickness of 2 mm can directly be included. The bed’s outer contour
can not be altered in some sections. The outer bead diameter and some shoulder diameters
are restricted by FIA regulations. The inner bed’s maximum diameter is limited to the
shoulder’s diameter to allow the assembly of the tire. The lower bed’s minimum depth
and the maximum distance from the rim’s outer face are also subject to FIA regulations.
Therefore the coordinates of the point in question also form parameters allowing to comply
with those restrictions. For the remaining degrees of freedom of the bed, contour lines of
the forging blank and the brake assembly limit the geometric design space. To make sure

c ETH Zürich IDMF-CMAS, March 1, 2018

4.2 Design Parameterization Sample Problems 41

wall thicknesses

distances

point
coordinates

radii

Figure 4.5: Optimization variables for spoke-body and bed contour

the minimum distances of 1 mm to the forging blank and 3 mm to the braking contour
can be approached as closely as possible without crossing them, the contour lines form
construction elements in the sketch and the bed’s contour is directly dimensioned to those
contours, setting the distances as parameters. The rotational body for the spokes is
parameterized in a way similar to the rim’s bed. Front and rear contour are dimensioned
to the blank’s contours. The parts of the contour-forming interfaces to the suspension
and the nut are non-parametric. The interface between spokes and bed depends on the
shape of both features. The sketch for the bed contains the spokes’ rear and front contour
lines as construction elements. The second sketch for the rotational body of the spokes is
referenced to those construction elements and the interface line with reference dimensions.
This way, the parameterization for both features is done in only one sketch and update
loops are avoided. At the same time, two separate features are needed, because the base
part between the spokes is not defined by prismatic pockets but the contour of the bed.
The pockets are removed from the spoke body going beyond the outer diameter. The
remaining spokes are then added to the rim bed. The sharp-angled base of the spokes is
then smoothed out by two parametric fillet features. The two pockets are parameterized in
a third sketch (see Fig. 4.6). A base line with a constant radius is defined. Parameters for

Poc
ket2
Pocket 1

Figure 4.6: Optimization variables for the pockets

the larger pocket include height and width at the base as well as at a continuous transition
point, and the radius of the upper rounding. The gap between the two pockets, forming

c ETH Zürich IDMF-CMAS, March 1, 2018

42 Parameterization and Variable Transformations

the spoke, is parameterized by two widths, one at the base and one at the same transition
point. This leaves only the upper rounding as a free parameter of the small pocket.

4.2.3 Maximum-Strength Flywheel and Mesh-Dependent Shape Param-

eters
The thickness values at the element interfaces in the radial direction take the role of shape
parameters as outlined in Fig. 4.7. Thus, the number of optimization variables depends
linearly on the number of elements in the radial direction but is independent of the number
of elements in axial direction. The position of the nodes on the element sides is always
on the straight line connecting the respective corner nodes. The flywheel problem uses

t0+ , t t0

Figure 4.7: Relation between nodal point coordinates and shape parameter t
[2]

very little pre-existing knowledge, allowing a great design freedom: assuming an analysis
model with 100 element columns in the radial direction, there are 101 element interface
lines associated with independent thickness values to define a design with individual fea-
tures. Thus, it became possible that the numerical shape optimization results suggested
a simplified mechanical model for finding the optimum flywheel design by exact formulae
[2].

4.2.4 Onsert Design and Mesh-Independent Shape Parameters

The onsert shape is defined by the curves of its surfaces along the radial coordinate x. A
change of thickness implies a change of the y-coordinate values of the finite-element nodal
points. The surface nodal point positions p of the initial design are partitioned into the
axial nodal point positions p1 at the bottom and p2 at the top surfaces, see Fig. 4.8. The
reference plane of the nodal-point positions is the surface of the sandwich. The values of
p1 gives thus the thickness of the bonding layer. In this study the bonding layer thickness
is taken as constant and all the entries of p1 have the same value, p1i = pi . Changing
F in itia l- s h a p e p o s itio n s

p 1, p 2 i

F o p tim u m - s h a p e p o s itio n s (p 2 i - p 1) m in

Figure 4.8: Definition of shape parameters p2 .

the values of p2 changes the thickness distribution of the onsert. This allows to study the
influence of the onsert structural stiffness properties on the bonding strength. In the initial

c ETH Zürich IDMF-CMAS, March 1, 2018

4.2 Design Parameterization Sample Problems 43

configuration the onsert has a rectangular cross section with a regular mesh consisting of
rectangular elements evenly arranged in rows and columns. Mesh distortion is minimum
when the positions of nodes between the unmovable bottom node and the top node are
adjusted so that a constant spacing in the axial direction is maintained as can be seen
from the mesh plot shown in Fig.

Figure 4.9: Final Design and Mesh Distortion

The number of entries of p increases with increasing fineness of the finite-element mesh in
the radial direction and may be quite high when an accurate stress analysis is required.
The number of optimization variables becomes decoupled from mesh size when the nodal
point positions p are made to depend on a different set of design parameters x. By
choosing that the onsert surface shape is composed of simple polynomial functions P, the
optimization variables x control the original shape parameters p as weight coefficients of
the polynomials P:
p = P n xn , 0 ≤ n ≤ N. (4.1)
Einstein’s summation convention is implied in (4.1) and the set of polynomials P is com-
plete, ranging in degree from zero to a specified maximum value n.
The shape depicted in Fig. 4.8 corresponds with a polynomial including the constant,
linear, and quadratic terms.

Transformation of the Constraining Equation

The transformation between the design variables and the analysis variables is given by
(4.1). The constraining function (3.52) is defined in terms of the nodal-point positions
p2 and must be evaluated at all element interfaces. An analysis model with NEL element
columns in the onsert region must therefore fulfill the constraint (3.52) at all discrete
NEL + 1 interface positions,
gi = t − ti = t + p1 − p2i ≤ 0 (4.2)
Corresponding with the discrete analysis model, the continuous constraining function has
thus been replaced by a number of NEL + 1 discrete constraining equations that can be
written as one vector equation:
g =  − t =  + p1 − p2 ≤ 0 (4.3)
The constraining equations are written in terms of the analysis variables p2 . Next we
wish to express them in terms of the design variables x which is achieved by use of the
transformation (4.1):
g =  − t =  + p1 − Px ≤ 0 (4.4)
The transformation does not reduce the number of constraining equations. Generally, the
design variables are subject to a larger number of constraining equations, say one-hundred.

c ETH Zürich IDMF-CMAS, March 1, 2018

44 Parameterization and Variable Transformations

Mathematical method for observing the design variables constraints

Only for cubic polynomials whose complexity is not higher than cubic, the constraints
can be mathematically expressed in terms of the design variables. For this purpose it is
convenient to normalize the radius range with −1 ≤ ξ ≤ 1 and then choose for the shape
parameterization the form:

t(ξ) = t0 + t1 ξ + t2 ξ 2 + t3 ξ 3 (4.5)

A first step is to require that the minimum thickness  is kept at the points ξ = −1, ξ = 0,
and ξ = 1:

g1 = g(−1) =  − t0 + t1 − t2 + t3 ≤ 0
g2 = g( 0) =  − t0 ≤ 0 (4.6)
g3 = g( 1) =  − t0 − t1 − t2 − t3 ≤ 0

If these conditions are satisfied, violations in the interior can only exist together with real
polynomial extrema. These are found with the condition:
 
1
q
2 2
g(ξ),xi = 0 → t1 + 2t2 ξ + 3t3 ξ = 0 → ξ1,2 = − t2 ∓ t2 − 3t1 t3 (4.7)
3t3
Real extrema do not exist if the radikant of the root is negative, or if the discriminat D is
positive:
D = 3t1 t3 − t22 ≥ 0 . (4.8)
This could be used complementary with (4.6) to form a set of conditions for keeping the
minimum thickness everwere within the considered domain. However, the search space
would be unnecessarily reduced. The maximum possible constrained search space is main-
tained by first checking with the discriminat whether real extremas exist. Then, the
position of the maximum
 
1
q
ξmax = − 2
t2 − t2 − 3t1 t3 (4.9)
3t3
is substituted in (4.5). This result is then used as complementary condition (4.6)

g4 =  − tξmax ≤ 0 . (4.10)

4.2.5 Composite Boat Hull and the Patch Idea

The idea of patches, illustrated in Fig. 4.10, relates the parameterization of a laminated
structure to its manufacturing process. A structure made from prepregs consists of an

Figure 4.10: Patch Pattern and Laminate (Courtesy N. Zehnder [14])

assembly of prepreg sheets each of which is characterized by its instance in the lay-up

c ETH Zürich IDMF-CMAS, March 1, 2018

4.2 Design Parameterization Sample Problems 45

sequence, position in space, shape, and size. These parameters are geometrical patch
parameters. Each prepreg sheet constitutes a patch and is in addition assigned a choice of
material and orientation of the principal material axes, which are internal material patch
parameters. For each point on the shell-type structure, the local laminate construction is
determined by the geometrical and internal material patch parameters. Of the complete
set of parameters, the lay-up sequence and the choice of material are discrete, rendering
the parameterization suitable for Evolutionary Algorithms rather than for Mathematical
Programming optimization engines.
The composite boat hull problem is multi-objective where stiffness and prize are in conflict
with each other. This assigns a crucial role to the choice and placement of materials which
can be chosen from a range of inexpensive low-stiffness glass-fiber and expensive high-
stiffness carbon-fiber reinforced plastics. A patch pattern on the boat hull is indicated in
Fig. 1.6.

4.2.6 Fuel-Cell-Stack End Plate

The CAD model is assembled from three types of CAD features: the lower plate, the
upper plate, and four ribs. Fig. 4.11 details how these entities are defined. The lower
hu1 tu1 tb1

hu2 tu2 tb2

xli

xui

hu3 tu3 tb3

hu4 tu4 tb4

tr
i

hu5 tu5 tb5

hu6 tu6 tb6

Figure 4.11: CAD features: lower plate, upper plate and a rib

plate is bounded through a planar functional face at the bottom and through assembled
face segments at the top. These segments are defined through equidistant sampling points
defining six optimization variables tl1 ...tl6 of the lower plate. Additionally, the edges of the
top faces are chamfered. The upper plate is defined through two sets of sampling points,
i.e. six equidistant height parameters hu1 ...hu6 defining the bottom face and six thickness
variables tu1 ...tu6 defining the top face of the upper plate. Again, the edges of these faces
are chamfered. Finally, each rib is defined through three optimization variables: a lower
position xli , an upper position xui , and a thickness tri . For all these optimization parame-
ters a range and a step size σ for Gaussian mutation or initialization are assigned as listed
in Table 4.1. The genotype for the Evolutionary Algorithm is defined as follows. The ribs
are chosen as CAD features to be optimized; one rib is defined by three parameters that
are represented in one gene. For the lower and upper plate sampling positions are chosen
as genes. That means for the lower plate each thickness represents a gene, and for the
upper plate the height and the thickness at a sampling position form a gene. This leads
to the following genotype, where each gene is marked through accolades:
{tl1 } {hu1 , tu1 } {tl2 } {hu2 , tu2 } {tl3 } {hu3 , tu3 } {tl4 } {hu4 , tu4 } {tl5 } {hu5 , tu5 } {tl6 }
{hu6 , tu6 } {xl1 , xu1 , tr1 } {xl2 , xu2 , tr2 } {xl3 , xu3 , tr3 } {xl4 , xu4 , tr4 }.
This representation implicitly fulfills the manufacturing requirement, i.e. extrusion mold-

c ETH Zürich IDMF-CMAS, March 1, 2018

46 Parameterization and Variable Transformations

Table 4.1: Ranges and step sizes for the optimization variables.
Parameters Range Step
[mm] [mm]
tl1 ...tl6 [1, 7] 1
hu1 ...hu6 [1, 33] 4
tu1 ...tu6 [1, 7] 1
xl1 ...xl4 [0.1, 73.3] 8
xu1 ...xu4 [0.1, 73.3] 8
tr1 ...tr4 [1, 7] 1

ing. For the end plate, the holes for medium flow and the tension bolts are made in a
subsequent machining process, and the global vertical edges are rounded. Since material
can only be removed from the extruded base block, planar horizontal faces have to be
machined into the upper plate for a well defined load introduction from the bolts to the
end plate. The position of these contact faces must be adapted to the varying slope and
curvature of the top face of the upper plate, whereas for some solutions this face even can
be split in two subregions.

4.3 The Parameterization Spectrum

The various parameterization models presented in the previous sections can be character-
ized by their respective levels of design-parameter and constraint densities. This idea of a
parameterization spectrum is due to M. Wintermantel and O. König [4]. Fig. 4.12 gives
a two-dimensional representation of it. The vertical axis stands for the number of design
D e s ig n P a r a m e te r d e n s ity

K n o w -H o w / C o n s tr a in t D e n s ity

Figure 4.12: Sample problems and parameterization spectrum

variables. The horizontal axis indicates the measure of constraint density, invested pre-
existing know-how, or parameterization sophistication. The indicated example problems
are placed in the scheme according to the summarizing description below.

c ETH Zürich IDMF-CMAS, March 1, 2018

4.3 The Parameterization Spectrum 47

Motorcycle Frame The problem was solved by assigning tube dimensions to each of the
15 independent tubes. An additionally posed, and interesting, problem was that only
three different dimensions where to be used but these had to be adjusted optimally.
On the other hand, the design freedom is reduced by prescribing fixed values for the
node positions where the tubes are connected with each other. This can be regarded
as pre-existing knowledge.

Race Car Rim The problem was solved by adjusting about 30 parameter values of CAD
entities. The problem is highly constrained as explained in Section 3.7.2.

Flywheel The problem was solved by using a mesh-dependent shape parameterization

where, depending on the fineness of the mesh, the number of variables can be a hun-
dred or more. This allows the forming of optimum design solutions with unexpected
shapes. Also, only two equality constraints, namely preservation of the mass and
torsional inertia of the initial design, where prescribed.

Onsert The problem was solved by using a mesh-independent shape parameterization

where the number of variables depends on a desired shape complexity in terms
of the maximum considered polynomial power. The calculated examples include
powers from zero to three, which implies that the small number of four weighting
coefficients act as optimization variables. This reduces the numerical effort of finding
the optimum design which is, on the other hand, limited to a shape whose complexity
is limited to a polynomial of the third degree. In order to guarantee meaningful
solutions, the onsert thickness must explicitly constrained to a minimum positive
value.

FRP-Boat Hull The problem is solved by patch parameterization including geometry

and internal material parameters whose number is a multiple of the number of
patches, here prepreg cuttings, used for making the structure. On the other hand,
there is much freedom in adjusting these parameters and the only restrictions used
in the example problem are the given shape of the boat hull and that the mass of
the initial design must not be exceeded.

Fuel-Cell-Stack End Plate Since four ribs where foreseen for the extruded-profile de-
sign, there are 27 free CAD parameters used in the optimization. Thickness values
must be larger than the minimum value that can be realized by the extrusion mold-
ing technique. The upper plate must be in a position above the lower plate. The four
ribs must be positioned within the geometric design space of the structure. Apart
from these obvious constraints that must be imposed to guarantee meaningful de-
sign solutions, there is much freedom to arrive at an optimum design solution whose
shape was not intuitively expected.

4.3.1 Influence of Mechanical Situation on Parameterization

The two problems Maximum-Strength Flywheel Design and Maximum Strength Onsert De-
sign use different levels of parameterization. The flywheel problem uses design variables
that transform very directly to the mesh-dependent analysis variables, resulting in a high
number of optimization variables, say 200. The onsert problem, on the other hand, uses
design variables controlling the global onsert shape via a linear combination low-degree
polynomials, resulting in a much lower number of optimization variables, say 4.
Apart from pure arbitrariness or the curiosity to try out things - what could one motivate

c ETH Zürich IDMF-CMAS, March 1, 2018

48 Parameterization and Variable Transformations

to use such different parameterizations? For both problems seem to be very similar re-
garding the structural models, the objectives, the formulation of the objective functions,
and the side constraints.
Both mechanical problems are rotationally symmetric and use in fact the same differen-
tial equations and the same quadratic eight-node Serendipity finite-elements to create the
structural stiffness matrix.
The objectives are to maximize strength and the basic idea behind reaching these ob-
jectives is to make the distribution of stresses, within the regions of interest, as even as
possible. So both objective functions globally penalize the sum of the quadratic (if n = 0)
deviations of the local stresses from the mean stress.
The two problems have also in common that thickness values must be positive, yielding
basically identical side constraints.
Only the loading sets the two problems apart. The flywheel is loaded by inertia body forces
in the radial direction due to the rotation while the onsert is loaded by some external force
in the axial direction.
Anticipating the study of the behavior of the two optimization programs, the flywheel
optimization procedure yields, without any problems, shape results such as shown in Fig.
1.1(b). On the other hand, an early version of the onsert program, using the same parame-
terization as the flywheel program, failed to produce a smooth shape such as shown in Fig.
4.9 [28, 29]. Rather, jagged shapes such as shown in Fig. 4.13 were obtained [30]. The

Figure 4.13: Jagged onsert shape obtained with mesh-dependent analysis variables [30]

parameterization based on global polynomial shape representation makes sure that the ob-
tained shapes are nice in appearance and easy to manufacture although the jagged shape
results are quite correct regarding the successful minimization of the objective function.
Another positive effect of the global shape representation is that the reduced number of
independent optimization variables tends to speed up convergence and reduce the number
of necessary design evaluations.
The question remains as to why the flywheel problem is better-natured than the onsert
problem. The answer lies in its loading situation and how it affects the stress distribution,
forming an objective function topology favoring smooth shape results. The loading of the
flywheel is in the radial direction and, assuming a mesh with only one element row, the
finite elements are loaded in-series. Consequently, when one of these elements becomes
thinner, the radial stress must increase and as it becomes thicker, the stress decreases. But

c ETH Zürich IDMF-CMAS, March 1, 2018

4.3 The Parameterization Spectrum 49

these stress deviations are immediately penalized by the objective function. Particularly,
the objective function tends to prevent exceedingly small thickness values, having the same
effect as if the side constraint were implemented via a penalty method transformation. On
the other hand, the loading of the onsert with respect to the stressing of the bonding layer
is partially in-parallel and does not produce the same stabilizing effect occurring in the
flywheel problem.
The different design variable models place the two problems at different positions on the
parameterization spectrum. Also, one could say that the program for finding maximum-
strength onsert shapes is useful as a preliminary design tool while the program for finding
maximum-strength flywheel shapes gave some deeper understanding of the mechanical
problem and inspired a new simplified mechanical model [2].

c ETH Zürich IDMF-CMAS, March 1, 2018

50 Parameterization and Variable Transformations

c ETH Zürich IDMF-CMAS, March 1, 2018

Chapter 5

Some Basic Concepts of Global

Nonlinear Optimization

The nonlinear constrained optimization problem can be written as follows [20]:

Minimize: f (x) objective function

Subject to:
gi (x) ≤ 0 j = 1, m inequality constraints
hk (x) = 0 k = 1, l equality constraints
xli ≤ xi ≤ xui i = 1, n side constraints
 

 x1  
x
 
2

 

 
where x = x 3 design variables
 .. 
. 

 

 

xn
 

The vector x is called vector of design variables. The objective function as well as the
constraining functions may be linear or nonlinear functions of x. They may be explicit or
implicit in x and may be evaluated by any analytical or numerical techniques we have at
our disposal. However, if mathematical programming is used, it is important that these
functions be continuous and have continuous first derivatives in x. If these conditions
are not satisfied, for instance when discrete-valued variables appear, one must either in-
vent homogenization techniques such as [12] or resort to other methods such as genetic
algorithms.

5.1 Nonlinear Optimization Task

The nonlinear optimization tasks require iterative solution processes some of which are de-
scribed in chapter 6. The nonlinearity makes iterative solution process, such as illustrated
in Fig. 6.3, necessary

c ETH Zürich IDMF-CMAS, March 1, 2018

52 Some Basic Concepts of Global Nonlinear Optimization

5.2 Feasible Region

The inequality constraints divide the search space into feasible and infeasible regions.
In the infeasible regions are all the designs violating one or several constraints. The
boundaries gi (x) = 0 between feasible and infeasible regions define hyperplanes in the
multi-dimensional search space. The equality constraints hi (x) = 0 define also hyperplanes
but they do not separate feasible from infeasible regions - they are feasible regions. Any
feasible design must lie on those equality-constraint hyperplanes.

5.3 Convex and Non-Convex Functions

Fig. 5.1 illustrates convex functions, concave functions, and functions that are neither
convex nor concave. We share Vanderplaats [20] mental picture that a convex function
looks like a bowl that will be able to hold water. Thus, a concave function is the negative
of a convex function. The function Fig. 5.1(c) is neither convex nor concave. A function

=  >  ? 

Figure 5.1: Convex (a), concave (b), and neither convex nor concave function (c)

f (x) is convex if for any two points x1 and x2 contained in the set it holds that

f θx1 + (1 − θ) x2 ≤ θf (x1 ) + (1 − θ) f (x2 )

 
0≤θ≤1 (5.1)

For further illustration we consider the sample optimization problem:

Minimize: f (x) = (x1 − 0.5)2 + (x2 − 0.5)2 (5.2)

1 1
Subject to: g(x) = + −1≤0 (5.3)
x1 x2
The objective and constraining functions are plotted in Fig. 5.2. The objective function

4
1 6
3
8

2 4
1 2
1
1
2
0 0 .5
0
0 1 2 3 4

Figure 5.2: Convex design space. Source: Vanderplaats [20]

is convex and its feasible region is bounded by a convex constraining function. Then, the
design space is a convex set.

c ETH Zürich IDMF-CMAS, March 1, 2018

5.4 Method of feasible directions 53

5.4 Method of feasible directions

As the unconstrained Lagrangian, for taking constraints in account, is so closely connected
with the method of feasible directions, a first first glimpse on the latter is given with Section
3.5.3. Here, more information is provided as the equality and the inequality constraints
may be treated differently.

5.4.1 Inequality Constraints

The method of feasible directions seeks to minimize the given original, or untransformed,
objective function and to preserve a feasible design at each iteration step. As with the
unconstrained optimization task, we search for an improved point of the optimization
variables x by determining a useful search direction s and adjusting a factor α so that the
new point,
x(k+1) = x(k) + αs, (5.4)
further minimizes the objective f . The requirement that the new design must not violate
the constraints restricts the search direction to the space of the feasible directions. The
direction-finding algorithm is now provided with the double task of producing search direc-
tions that are not only useful (having the descent property) but also feasible (not violating
constraints). Consider a design with an active constraint g(x( 0)) as illustrated in Fig.
5.3. The gradients of the objective as well as of the constraining functions are indicated

x 2 g ra d f(x (0 ) )
fe a s ib le s e c to r

x (0 )

s
g ra d g (x (0 )
) F (x )= c o n s t.
u s a b le a n d
fe a s ib le s e c to r
u s a b le s e c to r

g (x (0 )
)= 0

x 1

Figure 5.3: Usable and feasible sectors of a search space

in the figure. The line passing through the reference point x( 0) and being vertical to the
object-function gradient ∇f divides the two-dimensional search space into the usable and
unusable sectors. The line passing through the reference point x( 0) and being vertical to
the constraining-function gradient ∇g divides the feasible and the unfeasible sectors. For
a usable search direction it must be that
∇f (x(0) )s ≤ 0 (5.5)
and for a feasible search direction
∇g(x(0) )s ≤ 0 (5.6)

c ETH Zürich IDMF-CMAS, March 1, 2018

54 Some Basic Concepts of Global Nonlinear Optimization

must hold. The feasible search direction with steepest descent is tangential to the contour
line of the constraining function, g = const. However, if the constraining function is
nonlinear, the feasible search direction of steepest descent may stray into the infeasible
region which is unwanted. Therefore, one can introduce into (5.6) a positive parameter
θ (push-off factor), pushing the search direction away from the tangent at the infeasible
region:
∇g(x(0) )s + θ ≤ 0 (5.7)
The concept of including push-off factors and the resulting algorithms are well described
in Vanderplaat’s textbook [20].

5.4.2 Equality Constraints

Equality constraints are always active. If they are nonlinear, the line search necessarily
causes constraint violations, demanding their subsequent removal. At the beginning of the
line search, the search direction is perpendicular to the constraining function gradients.
Since the search direction must also be usable, it is composed of the direction of steepest
descent and the necessary corrections determined by the gradients of the constraining
functions.
Fig. 3.3 illustrates the finding of a feasible search direction with an objective function
that depends on two design variables subject to one equality constrained. The following
equations, however, refer to a situation where the number of design variables is not specified
and two equality constraints are assumed. The situation with two constraints is readily
generalized to one with m constraints. One proposes a search direction composed of the
direction of steepest descent and the gradients of the constraining functions:

s = −∇f − λ1 ∇h1 − λ2 ∇h2 . (5.8)

The weight factors λi must now be adjusted so that the search direction s is orthogonal
to the gradients of the constraining functions:

sT ∇h1 = −(∇f + β1 ∇h1 + β1 ∇h1 )T ∇h1 = 0

. (5.9)
sT ∇h2 = −(∇f + β1 ∇h1 + β1 ∇h1 )T ∇h2 = 0

After executing the multiplications and re-ordering the terms one obtains the system of
equations for the weight factors λi

∇h1 T ∇h1 ∇h2 T ∇h1 ∇f T ∇h1

" #( ) ( )
λ1
=− . (5.10)
∇h1 T ∇h2 ∇h2 T ∇h2 λ2 ∇f T ∇h2

The system of equations can be resolved when the determinant D of the coefficient matrix
is not equal to zero. The determinant D

D = (∇h1 T ∇h1 )(∇h2 T ∇h2 ) − (∇h1 T ∇h2 )(∇h2 T ∇h1 ) ≥ 0 (5.11)

vanishes only if the two equality constraints are identical.

c ETH Zürich IDMF-CMAS, March 1, 2018

5.5 Lagrange Multiplier Method 55

5.4.3 Generalization of Lagrange factor calculation

The instantaneous feasible search direction s considering the equality and all active in-
equality constraints is:
m
X l
X
s = −∇f (x) − λj ∇gj (x) − λm+k ∇hk (x) (5.12)
j=1 k=1

It is written in the more concise form:

s = −∇f − Bλ (5.13)
where all constraining function gradients and the Lagrange factors are collected in arrays
abbreviated with:  
λg
B = [∇g∇h] λ= (5.14)
λh
The feasible search direction s must be orthogonal to all active constraining function
gradients Bso that the inner product vanishes:
BT s = −BT ∇f − BT Bλ = 0. (5.15)
Resolving for the Lagrange factors gives:
−1 T
λ = − BT B

B ∇f (5.16)

5.5 Lagrange Multiplier Method

After having developed the background in terms of the method of feasible directions, one
finds a Lagrange function simply to be the function whose negative gradient falls together
with the feasible search direction:
m
X l
X
L(x, λ) = f (x) + λj gj (x) + λm+k hk (x) (5.17)
j=1 k=1

Solving a constrained optimization problem by the method of Lagrange multipliers

amounts to finding a stationary point of the Lagrangian L in terms of the optimiza-
tion variables x and the Lagrange multipliers λ. The gradient of L with respect to the
optimization variables, ∇x L, obtains the negative of the feasible search direction s (5.12):
∇x L(x, λ) = −s (5.18)
At the stationary point ∇x L = −s = 0 the objective f possesses its smallest value within
the feasible region, so that the stationary point is a minimum with respect to the opti-
mizations variables x:
L(x∗ , λ∗ ) ≤ L(x, λ∗ ) ; (5.19)
The gradient of L with respect to the optimization variables, ∇λ L, obtains the constraining
functions g and h:  
g
∇λ L(x, λ) = . (5.20)
h
At the stationary point ∇λ L = 0 all constraining equations are satisfied. The inequality
constraining equations g have negative values within the feasible region and the respective
Lagrange factors are all positive; for the equality constraints the opposite might hold.
Therefore the stationary point is a maximum with respect to the Lagrange factors λ:
L(x∗ , λ∗ ) ≥ L(x∗ , λ) . (5.21)

c ETH Zürich IDMF-CMAS, March 1, 2018

56 Some Basic Concepts of Global Nonlinear Optimization

5.5.1 Lagrange Multiplier Method Sample Problem

For an illustrative example the sample problem for penalty methods in Section 3.5.1 on
page 18 is reconsidered. Objective and constraining functions are provided with (3.15)
and (3.17), respectively. The one constraint that will be activated when searching for the
best solution is g1 . The Lagrangian formulation of the optimization problem is:
1
L= (x + 2)2 + λ (1 − x) (5.22)
20
The gradients of the objective and constraining functions are:
1
∇f = (x + 2) ∇g1 = −1 (5.23)
10
At the known constrained minimum x = 1 the Lagrange multiplier λ is calculated with
(3.25) on page 21:
∇f T ∇g1 3
λ=− T
= (5.24)
∇g1 ∇g1 10
With λ thus specified, Fig. 5.4(a) shows that the Lagrangian, as a function of the op-
timization variable x, is smooth with a minimum at the smallest possible value of the
objective f . To demonstrate that L as a function of λ has a maximum at the constrained
minimum point, the gradient ∇x L is set equal to zero and then resolved for x:
x+2
∇x L = −λ=0 → x = 10λ − 2 (5.25)
10
The result allows to eliminate x from (5.22) to express the Lagrangian as a function of the
Lagrange multiplier λ:
L = λ (3 − 5λ) (5.26)
The plot in Fig. 5.4(b) shows that the Lagrangian, as a function of the Lagrange multiplier

1.5 0.5
f g L
0.4
1
Lagrangian

0.3
0.5
0.2

0 0.1
‐2 ‐1 0 1 2 3
0
‐0.5 x 0 0.2  0.4 0.6

a b

Figure 5.4: Sample problem Lagrangian versus x and versus λ

λ, is smooth with a maximum at the smallest possible value of the objective f .

c ETH Zürich IDMF-CMAS, March 1, 2018

5.6 Necessary and Sufficient Optimality Criteria 57

5.6 Necessary and Sufficient Optimality Criteria

5.6.1 Unconstrained Objective Functions
The minimum point has a smaller function value than all other points. Probing the
points in the vicinity of a minimum point must always yield higher function values. When
the objective function is smooth with existing first and second derivatives, the necessary
optimality conditions can be expressed in terms of its first and second derivatives. The
first derivative with respect to the optimization variables is a vector called gradient ∇f (x)
that must vanish at the optimum:
 ∂f (x) 
∂x1 

 
 ∂f (x) 

 

∂x1
∇f (x) = 0, ∇f (x) = ..  . (5.27)


 .  
 ∂f (x) 
 
∂xn

The second derivative is a matrix called Hessian H and, at the minimum point, it must
be positive definite:
 2 
∂ f (x) ∂ 2 f (x) ∂ 2 f (x)
. . .
 ∂x21 ∂x1 ∂x2 ∂x1 ∂xn 
 
 ∂ 2 f (x) ∂ 2 f (x) ∂ 2 f (x) 
 ∂x ∂x ∂x22
. . . ∂x2 ∂xn 
 1 2 
|H|> 0,  . (5.28)
 .. .. .. .. 

 . . . . 

 
 2 
∂ f (x) ∂ 2 f (x) ∂ 2 f (x)
∂x1 ∂xn ∂x2 ∂xn . . . ∂x2 n

Positive definiteness means that this matrix has all positive eigenvalues. If the gradient is
zero and the Hessian matrix is positive definite for a given x, this insures that the design
is at least a relative minimum, but does not insure that the design is a global minimum.
Only when the objective is known to be convex will (5.27) and (5.28) suffice to identify
the global minimum.

5.6.2 Kuhn-Tucker Optimality Conditions

Constraints restrict the search space to the feasible region, and the optimal solution is the
one indicated in Fig. 3.3 and Fig. 5.5with a minimum objective value within the feasible
region. The necessary Kuhn-Tucker conditions [31] for x∗ being a constrained minimum
are:

1. x is feasible
gj (x∗ ) ≤ 0 j = 1, m
∗
hk (x ) = 0 k = 1, l
2. λ∗j gj (x∗ ) =0 j = 1, m λ∗j ≥0
m
X l
X
3. ∇x f (x∗ ) + λ∗j ∇x gj (x) + λ∗m+k ∇x hk (x) = 0
j=1 k=1
λ∗j ≥ 0
λ∗m+k unrestricted in sign

c ETH Zürich IDMF-CMAS, March 1, 2018

58 Some Basic Concepts of Global Nonlinear Optimization

The first of the Kuhn-Tucker conditions requires that the minimum of the constrained
function lie within the feasible region. The second condition implies that inactive con-
straints, as the one inactive constraint shown in Fig. 5.5, must not be considered. The
third condition can be understood by the method of feasible directions: At the constrained
minimum point, the length of the feasible search direction becomes zero. This appears
with one constraint, see Fig. 3.3, or several active constraints, see Fig. 5.5.

Ñ B x * 
N B x  = c o n s t l Ñ C x * 

l Ñ C  x * 

Ñ C x *  Ñ B x * 

Ñ C x * 
C x  = 

C x  = 

N 

Figure 5.5: Illustrations to Kuhn-Tucker conditions for constrained optima

c ETH Zürich IDMF-CMAS, March 1, 2018

5.7 Duality 59

5.7 Duality
Section 5.5 explains that the critical, or stationary, points of the Lagrangian are saddle
points. The saddle-point nature is connected with the max-min problem considered on
Section 5.7.1 and the duality concept introduced with 5.7.2; the presentation of both of
these topics rely on textbook material [20, 32, 33, 34].

5.7.1 The Max-Min Problem

The results of the previous sections allow defining the Lagrangian in terms of λ alone as

L(λ) = min L(λ, x) (5.29)

x

For finding a solution to the optimization problem the Lagrangian must be maximized
with respect to λ, or
max L(λ) = max min L(λ, x) (5.30)
λ λ x

The same solution is obtained by stating this as an equivalent min-max problem as

min L(x) = min max L(λ, x) (5.31)

x x λ

Vanderplaats [20] presents the example of a simple max-min problem:

1
F = ; g = x − 1 ≤ 0; x≥0 (5.32)
x
The Lagrangian is
1
L(x, λ) = + λ(x − 1) (5.33)
x
From (5.29) the gradient of L with respect to x must vanish for any value of λ, so
1 1
λx L(x, λ) = − +λ=0 → x = ±√ (5.34)
x2 λ
where only the positive root is meaningful since x is required to be positive. Now for any
value of λ the value of x which will minimize L is known. The Lagrangian can now be
written in terms of λ only as
√ √
 
1
L(λ) = λ + λ √ − 1 = 2 λ − λ (5.35)
λ
The constrained minimum is found by maximizing L(λ) and the gradient with respect to
λ must therefore vanish. This gives:
1
∇λ L(λ) = −1 → λ∗ = 1 (5.36)
λ
and then from ((5.34))
x∗ = 1 (5.37)
The original inequality-constrained problem has been solved by conversion to a max-
min problem. Next, the primal and equivalent dual problems are defined and finally
the problem types, where the presented concepts give a computational advantage, will be
identified.

c ETH Zürich IDMF-CMAS, March 1, 2018

60 Some Basic Concepts of Global Nonlinear Optimization

5.7.2 The Primal and Dual Problems

We are already familiar with the primal problem:
n o
min f (x) g ≤ 0, h = 0 (5.38)
x
and it can be solved without concern for the values of the Lagrangian multipliers λ. At
the optimum x∗ the values of λ∗ can be calculated with the third Kuhn-Tucker condition
(5.29), which is actually n equations, n being the number of variables x. Usually the
Lagrange multiplier values are not calculated because they are of no particular use. If,
however, the saddle-point values λ were known in advance, the constrained problems could
be solved with only one unconstrained minimization.
The dual optimization problem is stated as
n o
max f (λ) λj ≥ 0, j = 1, m (5.39)
λ
The Lagrange multipliers are now called dual variables. Before we have found out were it
is useful we may just play with the idea to solve the dual problem, or finding the Lagrange
multipliers, first and then retrieve the optimum primal variables x∗ . Actually in many
design problems only a few of many constraints are critical at the optimum so that only
the few corresponding Lagrange multipliers are nonzero. The maximization problem (or
minimization problem for −L) has only simple non-negativity constraints on the Lagrange
multipliers corresponding with the inequality constraints on x.
A further attractive aspect of duality is that the primal and dual objective values become
the same at the saddle-point. The dual always provides a lower bound on the primal
problem.

5.7.3 Computational Considerations

The usefulness of using dual methods depends on the finding of a way of conveniently
handling the primal variables within the dual problem. This is why dual methods become
attractive when the primal problem is mathematically separable. Separability exists when
the objective and constraint functions are each calculated as the sum of functions of the
individual design variables. Thus, a separable function shows the following characteristic
f (x) = f1 (x1 ) + f1 (x1 ) + · · · + fn (xn ) (5.40)
The primal problem we complete by writing the inequality constraining equations
gj (x) = gj1 (x1 ) + gj2 (x2 ) + · · · + gjn (xn ) ≤ 0, j = 1, m (5.41)
and equality and side constraints are omitted here for brevity. The Lagrangian is then
written as
Xn m
X n
X
L(x, λ) = fi (xi ) + λj gji (xi ) (5.42)
i=1 j=1 i=1
The dual problem, depending on the Lagrange multipliers only, is prepared by using (5.29)
and the property that the minimum of the separable function is the sum of the minima of
the individual parts:
 
Xn m
X
Maximize: L(λ) = min fi (xi ) + λj gji (xi ) (5.43)
xi
i=1 j=1

subject to: λj ≥ 0 j = 1, m (5.44)

c ETH Zürich IDMF-CMAS, March 1, 2018

5.7 Duality 61

5.7.4 Use of Duality in Nonlinear Optimization

Vanderplaats [20] discusses three cases where the dual form is used for computational
advantage, namely:

- Separable objective with linear and quadratic terms and linear constraints,

- Separable objective and constraints both with linear and quadratic terms, and

- nonseparable objective and separable linear constraints.

Separable Objective with Linear and Quadratic Terms and Linear Constraints
The primal problem is
n  
X 1
f (x) = ai xi + bi x2i (5.45)
2
i=1
n
X
g(x) = cji xi j = 1, m (5.46)
i=1

Referring to the result (5.43) the following function is minimized with respect to x:
 
m
1 X
min ai xi + bi x2i + λj cji xi  (5.47)
xi 2
j=1

At the minimum the derivative with respect xi must vanish which condition yields
m
X
ai + bi xi + λj cji = 0 (5.48)
j=1

Solving for xi in terms of λj gives

Pm
−ai − j=1 λj cji
xi = (5.49)
bi
Since some of the bi may be zero, it is important to foresee upper and lower limits for the
primal variables to be treated explicitly:

xli ≤ xi ≤ xui i = 1, n (5.50)

In case of bi being equal to zero, xi is set to its lower bound if the numerator in (5.49)
is positive and vice versa. The result is to be inserted into (5.43) which is then to be
maximized with respect to the dual variables λj .

Separable Objective and Constraints Both with Linear and Quadratic Terms
n  
X 1 2
f (x) = ai xi + bi xi (5.51)
2
i=1
n  
X 1
gj (x) = cji xi + dji x2i j = 1, m (5.52)
2
i=1

c ETH Zürich IDMF-CMAS, March 1, 2018

62 Some Basic Concepts of Global Nonlinear Optimization

Inserting the objective and constraints into (5.43) gives the sub problem
 
m  
1 X 1
min ai xi + bi x2i + λj cji xi + dji x2i  (5.53)
xi 2 2
j=1

Differentiating with respect to xi and equating to zero gives

m
X
ai + bi xi + λj (cji + dji xi ) = 0 (5.54)
j=1

the resolution of which gives the primal variable as

−ai − m
P
λj cji
xi = Pmj=1 (5.55)
bi + j=1 λj dji
Again, if the denominator becomes zero the xi is set to its upper or lower bound if the
numerator is negative or positive, respectively.

Nonseparable Objective and Separable Linear Constraints

Here the primal quadratic programming problem is written in matrix notation:
1
Minimize: f (x) = xT Ax + bT x (5.56)
2

subject to: λj ≥ 0 Bx ≤ 0 (5.57)

Nonnegativity requirements on xi can for convenience be included in the set of linear
inequality constraints. The matrix A is the Hessian matrix and is required to be positive
definite for convexity and because it must be invertible. Here the Lagrangian becomes
1
L(x, λ) = xT Ax + bT x + λT (Bx − c) (5.58)
2
Solving for the stationary condition with respect to x gives
∇x L(x, λ) = Ax + b + BT λ = 0 (5.59)
from which the primal variables x
x = −A−1 b + BT


(5.60)
are obtained. Substituting this into (5.58) and simplifying gives the Lagrangian in terms
of λ alone:
1 1
L(λ) = − λT Dλ − dT λ − bT A−1 b (5.61)
2 2
where the abbreviations
D = BA−1 BT ; d = BA−1 b (5.62)
have been used.
The special cases presented by Vanderplaats permit writing explicit expressions for the
primal variables in terms of the dual variables. often this may not be possible and then
the minimization subproblem can be solved as an optimization problem itself. When this
can be done with great efficiency, dual methods can provide efficient solutions techniques
for nonlinear optimization. Examples are the topology optimization problems explained
in Sections 9.4 and 9.5.

c ETH Zürich IDMF-CMAS, March 1, 2018

5.8 Optimization Algorithms Overview 63

5.8 Optimization Algorithms Overview

The nonlinear optimization tasks, stated in section 5.1, require iterative solution methods
for finding the minimum of a global objective function that may be constrained or un-
constrained. The iteration scheme illustrated in Fig. 5.6 applies largely to all structural
optimization processes.

Figure 5.6: Design Optimization Iteration Basic Scheme

There are many different methods from which optimization algorithms are derived. Müller
provides in her recent dissertation [1] a concise scheme for dividing the methods. The direct
methods require only the values of the objective function itself and the indirect methods
need additional information in terms of first or higher derivatives. Complete evaluation is
an extremely simple and inefficient direct method devoid of any strategy to utilize topol-
ogy information from previous function evaluations for improving convergence. The more
sophisticated direct methods are further subdivided into stochastic and deterministic ones.
Stochastic methods use random numbers to identify new design solutions. The stochastic
methods include, for instance, evolutionary algorithms some of which are explained in
chapter 7.
The deterministic methods rely on some information on the local objective function topol-
ogy to identify more systematically promising points in variable space. They include the
simplex search method or Powell’s conjugate direction method, explained in sections 6.1
and 6.6, respectively.
The indirect methods are all deterministic. One distinguishes there between first-order
methods, needing first-derivative information, and second-order methods, needing in ad-
dition second-derivative information. The first-order methods include the methods of
steepest descent (section 6.2) or the nonlinear conjugated gradient method of Fletcher
and Reeves (section 6.5). The second-order methods include the Newton’s method (sec-
tion 6.4) or surface-response methods (section 6.7).
In the literature [18], the deterministic methods, or the algorithms based on them, are cus-
tomarily called mathematical programming and the direct deterministic methods go under
the label zeroth-order methods of mathematical programming. The illustration in Fig. 5.7
adopts parts from [1] but also indicates the realm of mathematical programming. Apart
from the method order indicated in Fig. 5.7, search methods are characterized by the
model order which is is the order of the Taylor series expansion that approximates the ob-
jective function. Fig. 5.8 arranges the methods in a matrix whose rows and columns stand
for the model and method orders, respectively. Search methods based on quadratic models
of the objective function converge very quickly on quadratic functionals. For instance, the
Newton and the surface-response methods converge in one step on quadratic functionals

c ETH Zürich IDMF-CMAS, March 1, 2018

64 Some Basic Concepts of Global Nonlinear Optimization

P a r a m e te r O p tim iz a tio n T e c h n iq u e s

D ir e c t In d ir e c t

S to c h a s tic D e te r m in is tic U s e f, f U s e f, f, 2
f

- S im p le x M e th o d
- E v o lu tio n a r y - C a u c h y 's M e t h o d
- P o w e ll's M e t h o d - N e w t o n 's M e t h o d
C o m p u ta tio n - F le tc h e r R e e v e s M .
- R e s p o n s e -S u rfa c e

0 th
o rd e r 1 s t o rd e r 2 n d
o rd e r

M a th e m a tic a l P r o g r a m m in g

Figure 5.7: Overview of Optimization Algorithms. Source: Müller [1]

- R e s p o n s e -S u r fa c e M . - F le tc h e r -R e e v e s C G M . - N e w to n M e th o d
2
- P o w e ll's C D M e t h o d - V a r . M e tr ic M e th o d - Q u a s i-N e w to n M .
r d e r

1 - C a u c h y M e th o d
O
o d e l
M

0 - S im p le x S e a r c h M e th o d

0 1 2

M e t h o d O r d e r

Figure 5.8: Search methods arranged after model and method orders

although the methods are of order two and zero, respectively. The lower triangle of that
matrix is empty because model order limits method order.

5.8.1 An Argument for Mathematical Programming

The simplest search method is complete evaluation of the search space. Value ranges
are assigned to the respective variables so that a region of the search space is covered
where the optimum is suspected. Along each value range several coordinate values are
selected, for instance equidistantly. These coordinate values define a hypermesh of points
in n-dimensional space of the optimization variables. Fig. 5.9 illustrates the situation for
n = 2. The objective function is then evaluated at each point and the point having the
lowest function value is taken to be optimum. The numerical effort of the method explodes
with increasing number n of search space dimensions. The situation is vividly depicted by
Vanderplaats [20]:

”Consider, for example, a design problem described by three variables. Assume

we wish to investigate the designs for 10 values of each variable. Assume also

c ETH Zürich IDMF-CMAS, March 1, 2018

5.8 Optimization Algorithms Overview 65

x * P o in t M e s h
x 2

x 2
o p t
f ( x ,y ) = c o n s t.

x 1
o p t x 1

Figure 5.9: Objective function and hypermesh in two dimensions

that any proposed design can be analyzed in one-tenth of a central processor

unit (CPU) second on a digital computer. There are then 103 combinations of
design variables to be investigated, each requiring one-tenth second for a total
of 100 CPU seconds to obtain the desired optimum design. This would probably
be considered an economical solution in most design situations. However, now
consider a more realistic design problem where 10 variables describe the design.
Again, we wish to investigate 10 values of each variable. Also now assume that
the analysis of a proposed design requires 10 CPU seconds on the computer.
The total CPU time now required to obtain the optimum design is 1011 seconds,
or roughly 3200 years of computer time! Clearly, for most practical design
problems, a more rational approach to design automation is needed.

He uses this to underline the importance of the much more efficient mathematical pro-
gramming algorithms.

5.8.2 An Argument for Stochastic Methods

Most of the mathematical programming methods work local. Usually they proceed from
one starting point and estimate from the function value at that point and from derivative
information at the same point, or from information of previously obtained function values
at other points, a new point with a function value lower than those at the other points.
Thus, all points with higher function values than at the starting point, or any point reached
during the search, are systematically not evaluated which explains the incredible efficiency
gain if compared with complete evaluation.
The local nature of mathematical programming leads thus to great efficiency but, on the
other hand, lies at the heart of a great disadvantage: the search tends to get trapped at
local minima of non-convex objective functions. Practical ways and conceptual approaches
to mitigate this problem are addressed in section 6.11. However, the multi-start method
encounters a similar efficiency problem as complete evaluation and the tunneling method
is not yet fully developed.
In addition, the first- and second-order methods require the objective function to be con-
tinuous with continuous first and second derivatives, respectively. But generally one has
to expect that an engineering design problem may also be of a discrete nature, and then

c ETH Zürich IDMF-CMAS, March 1, 2018

66 Some Basic Concepts of Global Nonlinear Optimization

mathematical programming is not applicable.

All of this poses no problem for the stochastic methods which gain much higher efficiency
than complete evaluation by using random numbers to define design changes.

5.8.3 Mathematical Programming and Stochastic Search Methods

Mathematical programming can be so much more efficient than other methods that suc-
cessful methods have been derived where problems of a discrete nature are transformed so
that minimization of continuous objective functions solves the problem. Other aspects are
addressed in the introductory remarks on design optimization in section 1.1. Chapter 6
explains some selected algorithms of mathematical programming and chapter 7 deals with
stochastic search methods.

c ETH Zürich IDMF-CMAS, March 1, 2018

5.9 Practical Considerations for Numerical Optimization 67

5.9 Practical Considerations for Numerical Optimization

From Vanderplaats [20] we quote advantages and limitations of numerical optimization.

5.9.1 Advantages of Using Numerical Optimization

• A major advantage is the reduction in design time—this is especially true
when the same computer program can be applied to many design projects.
• Optimization provides a systematized logical design procedure.
• We can deal with a wide variety of design variables and constraints which
are difficult to visualize using graphical or tabular methods.
• Optimization virtually always yields a design improvement.
• It is not biased by intuition or experience in engineering. Therefore, the
possibility of obtaining improved, nontraditional designs is enhanced.
• Optimization requires a minimal amount of human-machine interaction.

5.9.2 Limitations of Numerical Optimization

• Computational time increases as the number of design variables increases.
If one wishes to consider all possible design variables, the cost of auto-
mated design is often prohibitive. Also, as the number of design variables
increases, these methods tend to become numerically ill-conditioned.
• Optimization techniques have no stored experience or intuition on which
to draw. They are limited to the range of applicability of the analysis
program.
• If the analysis program is not theoretically precise, the results of opti-
mization may be misleading, and therefore the results should always be
checked very carefully. Optimization will invariably take advantage of
analysis errors in order to provide mathematical design improvements.
• Most optimization algorithms have difficulty in dealing with discontinuous
functions. Also, highly nonlinear problems may converge slowly or not
at all. This requires that we be particularly careful in formulating the
automated design problem.
• It can seldom be guaranteed that the optimization algorithm will obtain
the global design optimum. Therefore, it may be necessary to restart the
optimization process from several different points to provide reasonable
assurance of obtaining the global optimum.
• Because many analysis programs were not written with automated design
in mind, adaption of these programs to an optimization code may require
significant reprogramming of the analysis routines.

c ETH Zürich IDMF-CMAS, March 1, 2018

68 Some Basic Concepts of Global Nonlinear Optimization

c ETH Zürich IDMF-CMAS, March 1, 2018

Chapter 6

Search for Design Improvement:

Mathematical Programming

Mathematical Programming (MP) is a rather generic term and the algorithms embraced by
it assume that in general the objective and the constraining functions are continuous and
at least twice differentiable [18]. Moreover, MP assumes convex objective functions. Given
a non-convex objective and a starting point, MP will generally obtain a local minimum
and an additional effort, for instance by employing a multi-start technique, is needed to
find other local minima with smaller function values, and, hopefully, the global minimum
too. Some mathematical programming methods are regarded as highly efficient but it is
also important to note that the efficiency of these methods depends very much on the
actual objective function. Therefore, general statements, claiming one certain method to
be always better than another certain method, are usually untenable.
The terms Sequential Linear Programming and Sequential Quadratic Programming are
used to characterize solution methods for nonlinear constrained optimization problems
which use linear and quadratic approximations, respectively, to the problems. The follow-
ing definitions have been taken from Vanderplaats’ textbook [20].

Sequential Linear Programming The nonlinear optimization problem

stated at the beginning of Chapter 5 is linearized via a first-order Taylor series
expansion:

Minimize: f (x) ≈ f (x0 ) + ∇f (x0 )∆x

Subject to: g(x) ≈ g(x0 ) + ∇g(x)∆x ≤ 0
(6.1)
h(x) ≈ h(x0 ) + ∇h(x)∆x = 0
where ∆x = x − x0

Whereas the unconstrained problem requires iterative line searches along a

current search direction, SLP can replace these line searches with a direct step
to activate the closest and as of yet inactive constraining equation.

Sequential Quadratic Programming Newton’s method for solving non-

linear minimization problems uses a quadratic approximation to the uncon-
strained objective by Taylor-series approximation up to the quadratic term, see
Section 6.4. SQP methods use the same quadratic Taylor-series approximation
of the Lagrangian, see Section 3.5.3, of the nonlinear constrained minimization

c ETH Zürich IDMF-CMAS, March 1, 2018

70 Search for Design Improvement: Mathematical Programming

problem.

L(x, λ) ≈ L(x0 , λ0 ) + ∇x f (x0 , λ0 )∆x + 21 ∆xT ∇2x f (x0 , λ0 )∆x (6.2)

Because of the approximation both methods require iterations and each ap-
proximation is called a quadratic sub problem. If all constraints are equality
constraints, the search method remains identical to that of the Newton method.
If the problem includes inequality conditions, the search becomes similar to the
modified Newton method but a nonlinear line search along may be replaced
with a direct step to activate the closest and as of yet inactive constraining
equation.

Fig. 5.7 illustrates how the methods of mathematical programming are divided into the
categories direct methods, gradient-based methods, and second-order methods. Zeroth-
order, or direct, methods require only objective-function values. First-order, or gradient-
based, methods require in addition the gradient (first derivative). Second-order methods
require also the Hesse matrix (second derivative).
Some methods are derived upon the model of a quadratic function, and thus have a the-
oretical basis. The textbook by Reklaitis, Ravindran, and Ragsdell [23] gives two reasons
for choosing a quadratic model:

1. It is the simplest type of nonlinear function to minimize, and hence any general
technique must work well on a quadratic if it is to have any success with a general
function

2. Near the optimum, all nonlinear functions can be approximated by a quadratic

(since in the Taylor expansion, the linear part must vanish). Hence, the behavior of
the algorithm on the quadratic will give some indication of how the algorithm will
converge for general functions.

The direct method due to Powell [35] is also based on a quadratic model. Cauchy’s method
of steepest descent [36] is a gradient method which is based on a linear model: improved
points are suspected in directions along the most local descent. Since no assumption is
made how that descent property might change at some distance away from the reference
point, the method is often less efficient than those based on quadratic models. Its one
advantage over more sophisticated strategies lies in its descent property.

c ETH Zürich IDMF-CMAS, March 1, 2018

6.1 Simplex Search Method 71

6.1 Simplex Search Method

The rendering of the simplex search method of Spendley, Hext, and Himsworth [37] follows
the textbook by Reklaitis, Ravindran, and Ragsdell [23]. Consider the problem of finding
some first-order information on direction-depending function-value distribution of a func-
tion in N dimensions by function values at points. For N = 1, two points can give the
information whether the function is increasing or decreasing with increasing variable value.
For N = 2 three points are necessary to give an equivalent information. In N dimensions
there are always N + 1 points necessary to obtain the directional trend of a function in
some region. When this smallest number of points are arranged equidistantly, they define
a regular simplex. For example, the equilateral triangle is a simplex in two dimensions; a
tetrahedron is a simplex in three dimensions. The main property of a simplex employed
by the simplex search method is that a new simplex can be generated on any face of the
old one by projecting any chosen vertex a suitable distance through the centroid of the
remaining vertices of the old simplex. The new simplex is then formed by replacing the old
vertex by the newly generated projected point. In this way each new simplex is generated
with a single evaluation of the objective. This process is demonstrated for two dimensions
in Fig. 6.1.
x (2 )
x (2 )

C e n tr o id o f x (2 )
a n d x (3 )

x (1 )
x (1 ) x (4 )

( a ) o ld s im p le x (a ) n e w s im p le x
x (3 )
x (3 )

x (1 )
, x (2 )
, x (3 )
x (2 )
, x (3 )
, x (4 )

Figure 6.1: Construction of new simplex (after [23])

The method begins by setting up a regular simplex in the space of the independent vari-
ables and evaluating the function at each vertex. The vertex with highest functional value
is located. This vertex is then reflected through the centroid to generate a new point,
which is used to complete the next simplex. As long as the function values obtained at
each new point decrease monotonously, the iterations move along until either the min-
imum point is straddled or the iterations begin to cycle between two or more vertices.
The minimum point is straddled when always the same vertex is reflected back and forth
through the iterations. These situations are resolved using the following three rules:
1. Minimum ”Straddled”
If the vertex with the highest function value was generated in the previous iteration,
then choose instead the vertex with the next highest value for reflection.
2. Cycling
If a given vertex remains unchanged for a M iterations, reduce the size of the simplex
by some factor. Set up a new simplex with the currently lowest point as the base
point. Spendley et al. suggest that M be predicted via M = 1.65N + 0.05N 2 , where
N is the problem dimension and M is rounded to the nearest integer. This rule
requires the specification of a reduction factor.

c ETH Zürich IDMF-CMAS, March 1, 2018

72 Search for Design Improvement: Mathematical Programming

3. Termination Criterion
The search is terminated when the simplex gets small enough or else if the standard
deviation of the function values at the vertices gets small enough. This rule requires
the specification of a termination parameter.

Apart from the objective evaluations, there are only two types of calculation required for
the simplex search algorithm: (1) generation of simplex at a given base point x(0) and
scale factor α in N -dimensional space and (2) calculation of the reflected point. For a
given base point the other vertices of the simplex are calculated from
 (0)
 xj + δ 1 if j = i
(i)
xj = (6.3)
 (0)
xj + δ 2 if j 6= i

for i and j = 1, 2, ..., N . The increments δ1 and δ2 , which depend only on N and the
selected scale factor α, are calculated from
h√ i
δ1 = √ −1 α
N +1+N
N 2
h√ i . (6.4)
N +1−1
δ2 = √
N 2
α

The design of a simplex in two dimensions is illustrated in Fig. 6.2(a). Suppose x(j) is the

d 2 d 1- d 2 N

x 4

x 3

a d 1- d 2

d 1 a x c

x c-x 1
x 2

d x
a 2 1

d 1 N 

(a) (b)

Figure 6.2: Design Principle (a) and Reflection (b) of a Simplex in Two Dimensions

point to be reflected as illustrated in Fig. 6.2(b). Then the centroid of the remaining N
points is
N
1 X (i)
xc = x , i 6= j. (6.5)
N
i=0

All points on the line from x(j) through xc are given by

 
x = x(j) + λ xc − x(j) . (6.6)

Choosing λ = 2 will yield the new vertex point so that the regularity of the simplex is
retained. Thus,
(j)
x(j)
new = 2xc − xold . (6.7)

c ETH Zürich IDMF-CMAS, March 1, 2018

6.2 Method of Steepest Descent 73

6.2 Method of Steepest Descent

Consider a reference point x0 and the function value f at that point. Let the gradient
∇f also be known. The gradient gives the direction of steepest ascent of the objective
function f in the space of the optimization variables x. Therefore, in the neighborhood
of the reference point, points with lower function values exist in the negative gradient
direction, or the direction of steepest descent, s:

s = −∇f (6.8)

Thus, the direction of steepest descent is a useful search direction along which to locate
points with lower function values. Moving along the search direction will initially obtain
ever smaller function values but after reaching a minimum the values will, from there on,
again increase. That minimum point along the search direction can be used as a new
reference point. This generates the sequence

xk+1 = xk + αk sk (6.9)

where the αk are obtained by line searches (see section 6.8). The slope f 0 is the component
of the gradient in the normalized search direction,

sk
f 0 = (∇f (xk+1 ))T (6.10)
|sk |

It vanishes at the minimum point along the search direction because there the search
direction and the gradient are perpendicular to each other. Therefore, any two consecutive
line searches are perpendicular to each other as Fig. 6.3 illustrates. However, exact

Figure 6.3: Cauchy-method zigzag path through two-dimensional search space

values of the step-length αk can be calculated only when the objective is a quadratic
functional. Even then the sequence 6.9 is not guaranteed to yield the absolute minimum

c ETH Zürich IDMF-CMAS, March 1, 2018

74 Search for Design Improvement: Mathematical Programming

of a convex function in a finite number of iteration steps - the zigzagging course indicated
in Fig. 6.3 could be continued forever. Moreover, the step-length can only be numerically
estimated and very accurate estimations may cost many function evaluations, further
reducing the efficiency of the sequence 6.9. Therefore, using in-exact line searches with
reduced numerical effort have the effect that two consecutive search directions are no
longer perpendicular to each other but the overall efficiency may increase as long as any
new point generated by the line search has a lower function value than the reference point.

c ETH Zürich IDMF-CMAS, March 1, 2018

6.3 Quadratic Objection 75

6.3 Quadratic Objection

First we recall a quadratic function depending on a single variable x only,

f (x) = a0 + a1 x + a2 x2 , (6.11)

and its first and second derivatives,

f 0 (x) = a1 + 2a2 x
. (6.12)
f 00 (x) = + 2a2

Setting the first derivative equal to zero determines the point xe where the function value
is an extremum:
a1
xe = − . (6.13)
2a2
If the value of the second derivative, or a2 , is positive, the extremum is a minimum. When
the minimum point is considered an optimum, the two conditions are called optimality
conditions.
Next we generalize the concept to quadratic functions depending on several variables xi
that are arranged in the variables vector x,

f (x) = p + xT p + xT Px, (6.14)

where the function f and the constant coefficient p are scalars, the variables vector x and
the coefficient p of the linear term are vectors, and the coefficient P is a matrix. The
extreme point xe in variables space is identified by setting the first derivative of (6.14)
equal to zero:
∂f
∇f (x) = = p + 2Px = 0. (6.15)
∂x
The first derivative of a function of several variables, or variables vector, is a vector called
gradient ∇f . From (6.15) follows the extremum point xe
1
xe = − P−1 p. (6.16)
2
The extreme point is a minimum if for any arbitrary vector v with non-zero length it
holds:
vT Pv > 0. (6.17)

c ETH Zürich IDMF-CMAS, March 1, 2018

76 Search for Design Improvement: Mathematical Programming

c ETH Zürich IDMF-CMAS, March 1, 2018

6.4 Original and Modified Newton Methods 77

6.4 Original and Modified Newton Methods

We realize that the quadratic polynomial is the simplest form of a function with just
one minimum that can be determined by the optimality criteria. Next we consider the
expansion of a smooth function at a reference point x0 by a Taylor series up to the quadratic
terms. The Taylor series approximates the function the better the smaller the considered
region is. The quadratic Tailor series approximation of a function f depending on a single
variable x is
1
fˆ(x0 + 4x) = f (x0 ) + 4xf 0 (x0 ) + (4x)2 f 00 (x0 ), (6.18)
2
where 4x is the distance from the reference point x0 . When f depends on a variables
vector x, the gradient ∇f is the first derivative and the Hesse matrix H = ∇∇f is the
second derivative of f (x):
1
fˆ(x0 + 4x) = f (x0 ) + 4x∇f (x0 ) + 4xT H4x, (6.19)
2
The first and second derivatives of the Tailor-series expansions are

∇fˆ(x0 + 4x) = ∇f (x0 ) + H4x

. (6.20)
∇∇fˆ(x0 + 4x) = H

Setting the gradient of the Tailor-series approximation fˆ(x0 + 4x) equal to zero yields the
distance 4x from the reference point x0 to the extremum point xapp e :

4x = −H−1 ∇f (x0 ). (6.21)

The extremum point itself follows from

xapp
e = x0 + 4x = x0 − H−1 ∇f (x0 ). (6.22)

The extremum is a minimum if, for any arbitrary vector v, it holds that vT Hv > 0.
The extremum point of the Tailor-series approximation is generally not identical with the
extremum of the objective function itself. Under certain conditions it will be closer to
the true extremum than the reference point. It can then replace the reference point and
applying (6.22) again yields an improved estimate so that the continued sequence,

xk+1 = xk − H−1 ∇f (6.23)

converges to the true extremum. It is understood that the gradient and the Hesse matrix
are evaluated at the point xk . This procedure is called the Newton Method.
The original objective function may not be well approximated by its quadratic Taylor-
Series expansion. This is likely to occur when the reference point is too far away from
the minimum point of the original objective. Then, the calculated extremum point of the
approximating function may be farther away from the true minimum than the reference
point so that the sequence does not converge. Introducing a variable step length αk leads
to the so-called Modified Newton Method,

xk+1 = xk − αk H−1 ∇f. (6.24)

The step length is not beforehand known and must be determined by a so-called line
search. Line search methods are explained in Section 6.7.

c ETH Zürich IDMF-CMAS, March 1, 2018

78 Search for Design Improvement: Mathematical Programming

In cases the Newton and Modified Newton Methods may be very efficient. However, in
structural optimization they enjoy only limited popularity. This is because the derivatives
can only be extracted numerically by using the methods explained in Section 6.8. The
number of function evaluations to calculate the Hesse matrix scales quadratically with the
number of design variables so that for a large number of design variables the numerical
effort becomes too high. The methods presented in the following sections avoid the need
for explicit calculation of the Hesse matrix.

c ETH Zürich IDMF-CMAS, March 1, 2018

6.5 Nonlinear Conjugated Gradient Methods 79

6.5 Nonlinear Conjugated Gradient Methods

The disadvantage of Cauchy’s Method is that the number of line search searches to obtain
the exact minimum point of even a convex objective function may be infinite. Although
this does not necessarily stand in the way of finding significant improvements over an
initial design after a small number of steps, one is interested in a more efficient sequence
for very close approximations of a minimum point. This is achieved by the Method of
Conjugated Gradients (CG) that is also used to efficiently solve large systems of linear
equations in iterations. The systems of equations then define a quadratic functional (Total
Potential Energy) of the unknown solution parameters the minimum point of which yields
the desired solution. The solution of those quadratic functional is guaranteed to be found
after as many steps as there are unknowns but may be found earlier depending on the
condition of the coefficient matrix. The method for such linear problems is well derived
and explained in [38] and it is also considered in the lecture class Structural Analysis with
FEM. One important feature of CG is that, instead of the orthogonality property of two
consecutive search directions, they fulfill the condition of C-conjugacy, or

s0 T Cs1 = 0, (6.25)

if the objective function is a quadratic functional, for instance

1
f = xT Cx. (6.26)
2
The point of C-conjugacy is illustrated in Fig. 6.4. However, a structural optimization task

s (0 )

r (1 ) = s (1 )
x (1 )
s (0 )

x x
s (1 )
x (1 )

r (1 )

Figure 6.4: Orthogonal and Conjugate Search Directions

usually leads to nonlinear objective functions so that the algorithm that is well-developed
for quadratic functionals cannot be applied. Then, the Method of Conjugated Directions
developed by Fletcher and Reeves [39] can be used. It generates new search directions
that are approximately C-conjugated to the previous one from the information of previous
line searches after the formula

|∇f k |2 k−1
sk = −∇f k + s . (6.27)
|∇f k−1 |2

The sequence requires memorizing the previous search-direction and gradient vectors. At
the first step of the sequence, these quantities are not known so that the first step is the
same as for the Method of Steepest Descent.

c ETH Zürich IDMF-CMAS, March 1, 2018

80 Search for Design Improvement: Mathematical Programming

In deriving the sequence (6.27) one assumes a quadratic functional (6.26) so that the
gradient at a given point x is given by

∇f = Cx. (6.28)

The search starts at a point x0 and the initially chosen search directions is that of steepest
descent at this point,

s(0) = −∇f (x(0) ) = −Cx(0) → x(0) = C−1 ∇f (0) . (6.29)

The vector of optimization variables along the search direction depends on a factor λ,

x = x(0) + λs(0) . (6.30)

That factor is chosen to minimize the value of f along the search direction. At that
minimum point the derivative of f with respect to λ must vanish. This condition is used
along with (6.29) and (6.30) to derive the minimizing value of λ:

f,λ = xT Cx,λ = 0

= (x(0) + λs(0) )T Cs(0)

= (C−1 ∇f (0) − λ∇f (0) )T C∇f (0)

. (6.31)
T T
= ∇f (0) C−T C∇f (0) − λ∇f (0) C∇f (0)
T
∇f (0) ∇f (0)
→ λ = T
∇f (0) C∇f (0)

We obtain the minimum point along the search direction s(0) by inserting the just derived
minimizing value of λ in (6.30) and call it x(1) . The gradient at this point,
T
(1) (0) (0) (0) ∇f (0) ∇f (0)
∇f = C(x + λs ) = ∇f − T
C∇f (0) , (6.32)
∇f (0) C∇f (0)

is naturally orthogonal to ∇f (0) . Guided by the pre-existing knowledge, that the mini-
mum of quadratic functionals is obtained after a finite number of iteration steps if two
consecutive search directions fulfill C-conjugacy, we require that the form

s(1) = −∇f (1) + βs(0) (6.33)

satisfy (6.25). That obtains

s(0)T Cs(1) = 0

= −s(0)T C∇f (1) + βs(0)T Cs(0)

= ∇f (0)T C∇f (1) + β∇f (0)T C∇f (0) . (6.34)

∇f (0)T C∇f (1)

→ β = −
∇f (0)T C∇f (0)
The problem with the result (6.34) is that the coefficients C appear explicitly. But these
coefficients are generally not known. They can, however, be eliminated by an identity that

c ETH Zürich IDMF-CMAS, March 1, 2018

6.5 Nonlinear Conjugated Gradient Methods 81

is derived by analyzing the scalar product of the gradient ∇f (1) with itself:
T
∇f (0) ∇f (0)
∇f (1)T ∇f (1) = ∇f (1)T ∇f (0) − T
∇f (1)T C∇f (0)
∇f (0) C∇f (0)

T
∇f (0) ∇f (0)
= − T
∇f (1)T C∇f (0) . (6.35)
∇f (0) C∇f (0)

T
∇f (1) C∇f (0) ∇f (1)T ∇f (1)
→ T
= −
∇f (0) C∇f (0) ∇f (0)T ∇f (0)

With the identity (6.35) we find for the factor β:

∇f (1)T ∇f (1)
β= (6.36)
∇f (0)T ∇f (0)

and for the search direction s(1)

∇f (1)T ∇f (1) (0)

s(1) = −∇f (1) + s . (6.37)
∇f (0)T ∇f (0)

Generalizing the indices 0 and 1 to k − 1 and k, respectively, gives the formula (6.27)
found by Fletcher and Reeves [39]. The derivation is based on a quadratic functional and
the formula will result in a better choice of search directions than Cauchy’s Method if the
objective function is dominated by the quadratic terms of its Taylor series representation.
The quadratic terms tend to overwhelm the other terms with decreasing distance to the
minimum point.
If many line searches must be performed within a region of the objective where it deviates
much from a quadratic functional, the search directions may lose their descent property
so that it becomes useful to occasionally restart the method by setting β = 0. The Polak-
Ribière method [40],
∇f (i)T ∇f (i) − ∇f (i−1)


PR
βi = , (6.38)
∇f (i−1)T ∇f (i−1)
allows for re-starting the search through periodically setting βi to zero, thereby offering
potential performance enhancement for non-quadratic functions compared to Fletcher-
Reeves. The underlying mechanism can be explained as follows [41]. Given a quadratic
objective function, both strategies coincide (∇f (i)T ∇f (i−1) = 0). The presence of non-
quadratic terms, potentially combined with insufficiently accurate line searches, is likely
to induce conjugacy loss as the search progresses. This manifests itself in the determined
search direction being nearly orthogonal to ∇f (i) , which is accompanied by:

∇f (i) ≈ ∇f (i−1) . (6.39)

Hence βi becomes almost zero, directing the search towards the steepest descent:

s(i) ≈ −∇f (i) . (6.40)

c ETH Zürich IDMF-CMAS, March 1, 2018

82 Search for Design Improvement: Mathematical Programming

c ETH Zürich IDMF-CMAS, March 1, 2018

6.6 Powell’s Conjugate Direction Method 83

6.6 Powell’s Conjugate Direction Method

According to a textbook [23] the most successful of the direct search methods is the method
due to Powell [35] which has further improved by modifications by Zangwill [42] and Brent
[43]. This method is motivated by the observation that a quadratic function of N variables
in the form of a sum of perfect squares can be minimized in N steps, one with respect to
each of the variables. Also, general quadratics can be transformed into a sum of perfect
squares in terms of the transformed variables. The quadratic form

Q(x) = xT Cx (6.41)

is generally not a sum of perfect squares unless all off-diagonal elements are zero. The
process of transforming it into a sum of perfect squares is equivalent to finding a transfor-
mation matrix T,
x = Tz, (6.42)
so that the functional becomes a sum of perfect squares in terms of the transformed
variables z and the diagonal matrix D,

Q(x) = zT TT CTz = zT Dz. (6.43)

We realize that the columns tj of the transformation matrix T give a new set of coordinates
that, because they diagonalize the quadratic C, correspond to its principal axes:

x = Tz = t1 z1 + t2 z2 + ... + tN zN . (6.44)

The new coordinates tj are called conjugate directions and the remaining problem is how
to calculate such a set of conjugate directions.
The basic approach to finding the conjugate directions is based on the parallel subspace
property of a quadratic function. The parallel subspace property is explained by using
the two-dimensional example illustrated in Fig. 6.5. Consider two arbitrary but distinct
points x(1) and x(2) and a direction vector d. Let y(1) be the point corresponding to the
minimum of Q(x(1) + λd) and (y)2 be the solution to Q(x(2) + λd). Then the direction
(y(2) − y(1) ) is C conjugate to d. From Fig. 6.5 it is apparent that two line line searches
determine the points y(1) and y(2) , establishing the set of C conjugate directions d and
(y(2) −y(1) ), and a third line search with reference point y(1) or y(2) and along the direction
(y(2) −y(1) ) finds the minimum point of the quadratic. All this is achieved without gradient
calculation.
The foregoing can be extended to give a elucidating definition of conjugacy that is taken
from the textbook [23]. Given an N × N symmetric matrix C, the directions s(1) , s(2) , ...
, s(r) , r ≤ N are called C conjugate if the directions are linearly independent, and

s(i)T Cs(j) = 0 for all i 6= j. (6.45)

Now consider the general quadratic function

1
q(x) = a + bT x + xT Cx. (6.46)
2

The points along the direction d from x(1) depend on the single variable λ,

x = x1 + λd. (6.47)

c ETH Zürich IDMF-CMAS, March 1, 2018

84 Search for Design Improvement: Mathematical Programming

N  
O  

N  
O  

Figure 6.5: Conjugacy in two dimensions. Source: [23]

The minimum of q along the line defined by (6.47) is obtained by finding λ∗ such that the
derivative of q with respect to λ is zero:

∂q ∂q ∂x
= = (bT + xT C)d. (6.48)
∂λ ∂x ∂λ

We call the minimum y(1) so that

h  i
y(1)T C + bT d = 0. (6.49)

Similarly, by using the same arguments, we have

h  i
y(2)T C + bT d = 0. (6.50)

Subtracting (6.49) from (6.50) gives us

 
y(2)T − y(1)T Cd = 0. (6.51)

Accordingly, the directions (y(2) − y(1) ) and d are C conjugate, and the parallel subspace
property of quadratic functions has been verified.
It remains to develop the actual minimization algorithm. The parallel subspace property
has been explained by using two starting points and a direction. Having to generate a
number of starting points is, however, not elegant from a computational point of view.
Therefore we construct a way to find C conjugate directions and the minimum by using
only one starting point. To this end we employ the coordinate unit vectors e(1) , e(2) ,
..., e(N ) . We consider, as before, the two-dimensional case and let e(1) = [1, 0]T and
e(2) = [0, 1]T . We use a starting point x(0) and calculate λ(0) so that f (x(0) + λ(0) e(1) ) is
minimized. This gives us the new point x(1)

x1 = x0 + λ0 e1 . (6.52)

From the new point we start a line search in the direction of e(2) so that f (x(1) + λ(1) e(2) )
is minimized which obtains a second new point

x2 = x1 + λ1 e2 . (6.53)

c ETH Zürich IDMF-CMAS, March 1, 2018

6.6 Powell’s Conjugate Direction Method 85

(0 ) ( 2 ) (1 ) (1 ) ( 2 ) ( 2 ) (3 ) (3 ) (1 )
I = A I = A I = A I = N - N

( 4 ) *
N (3 ) N = N

(1 ) ( 2 )
N N

( 0 )
N

Figure 6.6: Powell’s Method. Source: [23]

Next, we use the first coordinate unit vector again for a line search to calculate λ(2) so
that f (x(2) + λ(2) e(1) ) is minimized and let

x3 = x2 + λ2 e1 . (6.54)

Then, the directions (x(3) − x(1) ) and e(1) will be conjugate as Fig. 6.6 illustrates. The
construction can be extended from 2 to N dimensions and yields Powell’s Conjugate
Direction Method :

1. Define the starting point x(0) and a set of N linearly independent directions, possibly
s(i) = e(i) , i = 1, 2, 3, ..., N .

2. Minimize along the N + 1 directions, using the previous minimum to begin the next
search and letting s(N ) be the first and last searched

3. Form the new conjugate direction using the extended parallel subspace property

4. Delete s(1) and replace it with s(2) , and so on. Place the new conjugate direction in
s(N ) . Goto step 2.

Figure 6.7 illustrates the path taken by Powell’s method through the variables space of a
non-quadratic function.

c ETH Zürich IDMF-CMAS, March 1, 2018

86 Search for Design Improvement: Mathematical Programming

Figure 6.7: Path Taken by Powell’s Method Algorithm Through Non-Quadratic Function
Space. Source: [20]

c ETH Zürich IDMF-CMAS, March 1, 2018

6.7 Response-Surface Method Minimizing Algorithms 87

6.7 Response-Surface Method Minimizing Algorithms

A standard non-linear optimization problem is usually formulated as at the beginning of
Chapter 5 which can also be written more compact:

min {f (x)|gi ≤ 0, hk = 0} , j = 1, ..., J, k = 1, ..., K, xL,i ≤ xi ≤ xU,i (6.55)

x∈Rn

The lower and upper bounds xL,i and xU,i are initially imposed by the designer. To reduce
the number of costly function evaluations, a response surface model, also called a surrogate,
is built for the objective and possibly also for the constraining functions. The search for
the optimum design can then be partly make use of the surrogate functions and then the
original optimization problem is replaced by
n o
minn f˜(x)|g̃i ≤ 0, h̃k = 0 , j = 1, ..., J, k = 1, ..., K, xL,i ≤ xi ≤ xU,i (6.56)
x∈R

where the tilde symbol indicates that the respective surrogates are meant instead of the
original functions.
The response surface methodology was originally developed for constructing empirical
response functions based on physical experiments. When the physical experiments are
replaced by numerical function evaluations, the methodology can be used to find minimum
points of the function. For that purpose quadratic polynomials are obviously suited.
They are the simplest functions with a minimum point and can easily be constructed and
evaluated. The methodology is based on the elements:

1. Construct the response surface model for the given number of variables

2. Apply a scheme for placing the supporting points

3. Find response surface model parameters from the supporting point evaluations

4. Find minimum point of the response surface model, or surrogate objective function

5. Update response surface model in iterations

6. Apply some termination criteria

Response surface approximations shift the computational burden from the optimization
problem to the problem of constructing the approximations, and accommodate the use of
detailed analysis techniques without the need of derivative information, [44]. Additionally,
response surface approximations filter out numerical noise inherent to most numerical
analysis procedures, by providing a smooth approximate response function, and simplify
the integration of the analysis and the optimization codes.

c ETH Zürich IDMF-CMAS, March 1, 2018

88 Search for Design Improvement: Mathematical Programming

6.7.1 Constructing a Response Surface Model from Supporting Points

The objective function, actually a functional, is alternatively called a response surface.
It is approximated by a response surface model constructed from quadratic polynomials
Π2 in terms of the optimization variables. The following example is written out for two
variables:
Π2 = a1 + a2 x1 + a3 x2 + a4 x21 + a5 x1 x2 + a6 x22 . (6.57)
The polynomial Π2 (6.57) can in general also be written as inner product of the coefficients
and parameters vectors, c and a:
 T 
c (p1 )
 .. 
 . 
T T
 T 
Π2 = C a ; C =  c (pk ) 

 (6.58)
 .. 
 . 
T
c (pns )

where c contains the polynomial powers and pk is the k th supporting point. The parameters
ai of Π2 can be adjusted to fit the objective function values fk at the supporting points
pk . The number ns of supporting points spanning the search space must at least equal the
number nc of parameters or coefficients in Π2 . That number depends on the number nx
of optimization variables:
1
nc = 1 + nx + [nx (nx + 1)] . (6.59)
2
If the number ns of supporting points equals the number nc of the parameters of the
quadratic approximation, the parameters a are obtained by inverting (6.58):

a = C−T f . (6.60)

It is also possible to use a higher number of supporting points for constructing the quadratic
model. In general the approximation cannot match the function values for all of the
supporting points if their number is higher than the number of terms in the polynomial,
ns > nc , so that a residual r remains:

CT a − f = r (6.61)

A best-fit approximation follows from minimizing the inner norm rT r of the residual r
with respect to the polynomial parameters a:

∂ rT r



−1
=0 ⇒ a = CCT C f. (6.62)
∂a
If there are more supporting points ns than model parameters nc , the coefficient matrix
C has dimensions ns × nc .

c ETH Zürich IDMF-CMAS, March 1, 2018

6.7 Response-Surface Method Minimizing Algorithms 89

6.7.2 Finding the Minimum Point of the Response Surface Model

For estimating the minimum point of the quadratic approximation it is more convenient
to recast (6.57) into the form

Π2 = p + xT p + xT Px, (6.63)

where the parameters a are contained in the scalar p, the vector p, and the matrix P. As
a necessary condition, the gradient must vanish at an extremum,

4Π2 = p + 2Px = 0, (6.64)

yielding the extreme point

1
xE = − P−1 p. (6.65)
2
The extreme point is a minimum when for any vector v with non-zero length it holds:

vT Pv > 0, |v| =
6 0. (6.66)

6.7.3 The Relation between RSM and NM

The Response Surface Method is a zero-order methods since they do not require function
derivatives. The Newton Method, on the other hand, is second-order method because it
requires first and second derivatives. Therefore, the two methods may seem to be very
different from each other. A closer look reveals that the methods have more in common
than is often realized, and they can in fact be unified by using the supporting placement
scheme [45] explained in the following Paragraph.

The Unifying Supporting Point Placement Scheme

Fig. 6.8 shows the placement of the supporting points that can be used for a three-
dimensional response-surface approximation. The point x(1) at the center of the point set
 

& '

# $

 !

"

Figure 6.8: Placement of supporting points x(1) through x(10) in 3-dimensional variables
space

is the reference for constructing the other supporting points. Assuming the placement
of the points on a regular lattice, indicated in Fig. 6.8, with a spacing D along the
respective coordinate directions, the coordinate variations of the individual supporting

c ETH Zürich IDMF-CMAS, March 1, 2018

90 Search for Design Improvement: Mathematical Programming

points with respect to the reference point x(1) are given in the matrix presentation of
Table 6.1. The rows correspond to the optimization variables and the columns correspond
to the supporting points x(2) through x(10) . The coordinates of the points can generally

2 3 4 5 6 7 8 9 10

1 −D D 0 0 D 0 0 D 0
2 0 0 −D D D 0 0 0 D
3 0 0 0 0 0 −D D D D

Table 6.1: Supporting points in terms of changes with respect to x(1)

be identified for nx variables by the scheme developed in the following. The points are
identified successively from variable 1 through nx . For the ith variable, a number of i + 1
supporting points must be added to the set of nc = nc (i − 1) (6.59) existing points. The
first two of the new points span the direction of the ith variable:

(nc +1) 1 0, k 6= nx + 1
xk = xk − (6.67)
D, k = nx + 1
and 
(n +2) 0, k 6= nc + 1
xk c = x1k + . (6.68)
D, k = nc + 1
The remaining i − 1 points are variations of the new point x(nc +2) (6.68):

(nc +2+l) (nc +2) 0, k 6= l
xk = xk + , l = 1, i − 1 (6.69)
D, k = l
The supporting points span some appropriate region of the search space. Along with the
function values at each point, they are used to construct the response-surface approxima-
tion.
It is interesting to examine the presented supporting-point arrangement for the NM. The
NM uses a set of supporting points in the vicinity of a reference point to calculate the
gradient and Hessian matrix at that point. Point x(1) then becomes the reference point
at which the gradient and Hessian matrix must be obtained by the method of differences.
The other supporting points must then move very close to the reference point which is
achieved by choosing a very small value for the distance D:

D =  << 1. (6.70)

It is interesting to note that the point arrangement allows exact calculation of both first
and second derivatives of quadratic polynomials independent of the value of . This is
illustrated on the three dimensional case, where the points 1 through 10 are arranged as
shown in Fig. 6.8. Let the gradient and Hesse matrix be calculated from
 
f3 − f2 
1 
∇f = f5 − f4 (6.71)
2 
f8 − f7


and
1  a b c

∇∇f = Hij + H ij + Hij , (6.72)
2

c ETH Zürich IDMF-CMAS, March 1, 2018

6.7 Response-Surface Method Minimizing Algorithms 91

where Ha is formed in terms of the reference point,

 
2f1 −f1 −f1
Hija = − 
 
 −f1 2f1 −f1 
, (6.73)
−f1 −f1 2f1

Hb is formed in terms of the points on the coordinate axes,

 
(f2 + f3 ) −(f3 + f5 ) −(f3 + f8 )
Hijb = 
 
 −(f3 + f5 ) (f4 + f5 ) −(f5 + f8 ) , (6.74)

−(f3 + f8 ) −(f5 + f8 ) (f7 + f8 )

and Hc is formed in terms of the points shifted in away from the coordinate axes in other
directions:  
0 f6 f9
Hijc = 
 
 f6 0 f10 
. (6.75)
f9 f10 0
Obviously, the gradient of quadratic functions is calculated exactly because (6.71) is the
central differences method. It can be seen from (6.73), (6.74), and (6.75) that (6.72)
produces a symmetric Hessian matrix. Moreover, it can be shown that the evaluations
reproduce the symmetric version of the matrix P, multiplied by two:

∇∇f = (Pij + Pji ) (6.76)

The presented point arrangement and evaluation scheme calculate for quadratic functions
f := Rn → R and regardless of the value of the epsilon parameter not only the Hessian
but also the gradient exactly.
The RSM evaluation scheme obtaining the approximation parameters p, p, and P, ex-
plained in Subsections 6.1 and 6.2, can therefore be replaced by p = f (x(1) , p = ∇f , and
P = 12 ∇∇f where ∇f , and ∇∇f are obtained by (6.71) and (6.72), respectively for arbi-
trary values of D. For quadratic functions, both methods obtain the same approximation
parameters. Therefore, the evaluation scheme of the Newton method becomes identical
with that of the RSM method at the limit of D approaching very small values D → .
The entries of the gradient vector, where (6.71) represents the example for three variables,
are calculated by the general scheme
fnc (k−1)+2 − fnc (k−1)+1
∇fk = . (6.77)
2
The entries of the matrices Ha , Hb , and Hc follow from
(
2f1 , j = i
Hija = − , (6.78)
−f1 , j 6= i
(
(fnc (i−1) + fnc (j−1) ) , j = i
Hijb = , (6.79)
−(fnc (i−1)+2 + fnc (j−1)+2 ) , j =
6 i
and (
0 , j=i
Hijc = . (6.80)
fnc (j)−i+1 , j 6= i

c ETH Zürich IDMF-CMAS, March 1, 2018

92 Search for Design Improvement: Mathematical Programming

Rigid Lattice Minimization strategy

If the point defined by (6.65) has a smaller function value than any of the supporting
points, it is used as the reference point for the next set of supporting points. The distance
value D of the next supporting point set depends on the distance between the current and
next reference points:
D = |xk+1 − xk |. (6.81)
This is illustrated in Fig. 6.9. A current point set in 2-dimensional variables space (solid
circles) defines a response surface approximation. Its minimum point xE is the encircled
square. It becomes the reference x(1) for the new set of supporting points that is marked
with squares. The minimum point estimate of this set is indicated by the encircled triangle,
defining the next point set. As can be seen from Fig. 6.9, the distance values D of new
point sets are always chosen so that the region covered by the new set extends to the
previous reference point. Thus, if the successful minimum point estimate is outside of the
region of the current set, the new set will span a larger region. If the minimum point
is found inside of the region of the current point set, the new region will automatically
contract. Expanding the point set region makes it more likely to find the minimum point
of the objective function inside that region when future minimum point estimates will
fail, and contracting the region helps accelerating convergence once it seems likely that
the objective minimum point is being closed in. If, on the other hand, the point defined

Figure 6.9: Updated supporting point sets around successful minimum point estimates

by (6.65) does not yield a smaller function value than any of the supporting points, it is
discarded and the strategy considers two cases.
In the one case, the current reference point has a smaller function value than any other
point of the set, thus remaining the best estimate of the objective minimum point. Then,
the reference point is kept and convergence is facilitated by shrinking the√set of supporting
points about the reference point. The shrinking factor can be chosen 1/ 2, for instance.
In the other case, a point of the set other than the referenced point appears as best
estimate of the objective minimum point. The whole point set is then shifted by moving
the reference point to the point with smallest function value. Instead of using this mental
picture, one can also say that say that the point with smallest function value becomes the
new reference point and the distance value defining the extension of the region covered by
the point set is kept constant. This keeping constant of the distance value is motivated
by the opportunity to reduce the number of function evaluations of the new point set
as some points of the new set will then have the same coordinates as other points of the

c ETH Zürich IDMF-CMAS, March 1, 2018

6.7 Response-Surface Method Minimizing Algorithms 93

Figure
√ 6.10: Shrinking of supporting point region around the reference point by a factor
of 1/ 2

previous set whose function values are already known. Fig. 6.11 illustrates the shifts to the
points 2 through 6, respectively. The figure obtains that in each case three points of the
shifted sets coincide with previous points so that three function evaluations can be saved,
making the optimization process more time-efficient. In general, however, when moving

5 6

2 1 3
c u r r e n t p o in t s e t s h ifte d to p o in t 6
4

s h ifte d to p o in t 5 s h ifte d to p o in t 3

s h ifte d to p o in t 2 s h ifte d to p o in t 4

Figure 6.11: Supporting point set shifted to center around the respective points with
smallest function value

the supporting point set so that the reference point of the shifted set coincides with one of
the other points in the original position, the total number of coinciding points depends on

c ETH Zürich IDMF-CMAS, March 1, 2018

94 Search for Design Improvement: Mathematical Programming

the point to which the reference point is shifted. Table 6.2 shows the number of coinciding

2 3 4 5 6 7 8 9 10 11 12 13 14 15
2 2
3 3 3 3 3
4 4 4 4 3 4 4 3 4
5 5 5 5 3 5 5 3 3 5 5 3 3 3

Table 6.2: Number of coinciding points for the different reference point positions of the
shifted sets corresponding to 1-, 2-, 3-, and 4-dimensional variable spaces.

points when the reference point of the new supporting point set coincides with the other
points of the previous set. In the 1-dimensional variables space there are two other points
where the reference point can be shifted and in each case there are two coinciding points.
The situation in the 2-dimensional variables space is illustrated in Fig. 6.11. In the 3-
dimensional variables space with ten supporting points the number of coincidences depends
on which point of the previous point set the reference point of the new point set is located.
The number of coincidences is either 3 or 4. In the 4-dimensional space the number of
coincidences is either 3 or 5. For a set of supporting points corresponding to nx variables,
the average number nc of coinciding points can be shown to be
nx
nc = [2 (nx + 1) + 3 (nx − 1)] (6.82)
np − 1

The minimum, maximum, and average numbers of coinciding points, depending on the
number of variables, are shown in Fig. 6.12. The maximum number increases linearly

Figure 6.12: Maximum, average, and minimum number of coinciding points versus number
of variables

and the minimum number remains constant at three for variables spaces higher than one-
dimensional. The average number shown in the figure is calculated with (6.82). The points
of each shifted set coinciding with previously evaluated points need not be evaluated again.
Fig. 6.13 shows the point set size and its reduction due to coinciding points depending on
the number of variables. The potential computing time savings are rather large when the
number of variables is small but tend to become insignificant at large numbers of variables
as Fig. 6.14 illustrates. The figure plots the point set size over the point set size reduced

c ETH Zürich IDMF-CMAS, March 1, 2018

6.7 Response-Surface Method Minimizing Algorithms 95

Figure 6.13: Points set size and reduction due to coinciding points versus number of
variables

by the average number of coinciding points,

np
e= , (6.83)
np − p c

which is a measure of the potential efficiency gain. As the figure shows, a program may
run up to three times faster when searching for the minimum point of only one variable, or
twice as fast when there are two variables, but only insignificantly faster when the numbers
of variables is large.

Figure 6.14: Relative computing time savings due to coinciding points versus number of
variables

c ETH Zürich IDMF-CMAS, March 1, 2018

96 Search for Design Improvement: Mathematical Programming

6.7.4 Adaptive Response Surface Method

The Adaptive Response Surface Method (ARSM) reduces systematically the design space.
Ideally, the reduced design space contains only points with function values smaller than a
certain threshold value. Some take this threshold value from the second highest function
value of a current supporting point set. The reduction of the design space can be mentally
compared with the shrinking water surface area of an evaporating puddle. For such a
minimization strategy, a more flexible support placement scheme, than the one presented
above, is useful, and explanations in the following Subsection are taken from Wang [46].

Latin Hypercube Sampling Points Scheme

Latin Hypercube Sampling, or Latin Hypercube Design (LHD), was first introduced by
McKay et al. [47]. In practice, Latin Hypercube Design samples can be obtained as follows.
The range of each design input variable is divided into n intervals, and one observation
on the input variable is made in each interval using random sampling with each interval.
Thus, there are n observations on each of the d input variables. One of the observations on
x1 is randomly selected (each observation is equally likely to be selected), matched with a
randomly selected observation on x2 , and so on through xd . These collectively constitute a
design alternative x1 . One of the remaining observations on x1 is then matched at random
with one of the remaining observations on x2 , and so on, to get x2 . A similar procedure is
followed for x3 , , xn , which exhausts all observations and results in n LHD sample points
[15]. The sampling method is indicated in Fig. 6.15. LHD sampling is well suited for

$
#
"
!


  ! " # $

Figure 6.15: Latin Hypercube Sampling in a Two-Dimensional Design Space

ARSM because the points are easy to construct and they fill the complete reducing design
space.
The quadratic response surface model, or surrogate, is fitted to the data using the least
square method. When constraints are considered, a global optimization algorithm is used
to find the optimum. Following this step, the value of the actual objective function at the
optimum of the surrogate is calculated through an evaluation of the computation-intensive
objective function. If the value of the actual objective function at the surrogate optimum
is better than the values at all other experimental designs, the point is added to the set of
experimental designs for the following iteration because the point represents a promising
search direction.

c ETH Zürich IDMF-CMAS, March 1, 2018

6.7 Response-Surface Method Minimizing Algorithms 97

Search Space Reduction

All experimental designs and the accepted model optimum are recorded in a design library.
Then a threshold value of the objective function is chosen. The design space that leads to
objective function values larger than the threshold is then discarded. In ARSM, the second
highest value of the objective function in the set of experimental designs is chosen as the
threshold. If this second highest value cannot help to reduce the design space, the next
highest value of the design function will be used, and so on. The optimization process will
terminate if a satisfactory design emerges in the design library or the difference between
the lower limit and the upper limit of each design variable is smaller than a given small
number.

Using Inherited Points

The new reduced design space is again subdivided into square intervals (hyper cubes).
Then, some of the sampling points of the previous, larger space will fall within the new
space, providing an incomplete sampling. These points are inherited and complemented
with new points that are placed by using the LHD scheme and the details of the method
suggested by Wang can be found in his recent publication [46].

c ETH Zürich IDMF-CMAS, March 1, 2018

98 Search for Design Improvement: Mathematical Programming

c ETH Zürich IDMF-CMAS, March 1, 2018

6.8 Line-Search Methods 99

6.8 Line-Search Methods

The iteration rule
xk+1 = xk + αk sk (6.84)
has first been stated in the context of the method of steepest descent but, in its general
form, stands also for all the other methods where the minimum is sought from a reference
point along a useful search direction s. It has been mentioned before that the distance
between the reference point x0 and the minimum point along the search direction s must
be calculated for each iterate. Finding that distance is left to the so-called line-search
methods. Since we are dealing with nonlinear optimization, the line-search methods are
iterative. So, the minimization process is a nestled loop where the outer loop determines
a sequence of search directions while the inner loop finds the minimum point along each
search direction.

6.8.1 One-Dimensional Search in Multidimensional Variables Space

The iteration rule (6.84) finds an improved point xk+1 in multi-dimensional variables space
x by going from the reference point xk a certain way αk along the useful search direction
sk . The known search direction reduces the problem of obtaining the new point xk+1 to
the problem of determining the scalar αk . When the search direction is normalized,
s
s̄ = (6.85)
|s|

the distance α becomes a direct measure of distance. The a change of position 4x is

related to the absolute value of it, 4x, by the normalized search direction s̄:

4x = 4xs̄. (6.86)

At the minimum point the slope f 0 , or the derivative of the objective function f with
respect to the distance variable x, along the search direction s̄ must vanish. The slope can
be calculated as scalar product of the gradient vector and the normalized search direction:

f 0 = s̄T ∇f. (6.87)

6.8.2 Interval Halving and Golden Section Methods

The scalar-valued reference point x = x0 corresponds with the point in variables space
xk . The initial value of the increment 4x, obtaining x1 = x0 + 4x, must be chosen. If
the function value f (x1 ) is less than the value f (x0 ), the new point x1 is a success and
the search continuous by adding another increment to obtain the point x2 = x0 + 24x.
The variable x increases incrementally until a newly obtained point xm+1 has a higher
function value than the previous point xm . Then the minimum point xmin is assumed
to be bounded by xm−1 < xmin < xm+1 . The orientation of the search is reversed and
the increment is divided by two, 4x = − 21 4x, so that the new point lies in the middle
between xm and xm+1 as Fig. 6.16 indicates. In the figure, the open circles indicate
points whose function values have been obtained by a previous line search. The necessary
computational effort, up to a certain level of accuracy, thus follows from summing up the
number of solid circles. The increment is from now on modified after no more than two
steps, respectively, until the process is terminated by some criteria such as

c ETH Zürich IDMF-CMAS, March 1, 2018

100 Search for Design Improvement: Mathematical Programming

y m + 1

y m

y m + 1

x m -1 x m + 3 x m + 4 x m x m + 5 x m + 2 x m + 1

Figure 6.16: Sketch of the interval-halving algorithm

• The relative change between two consecutive function values is smaller than a preset
value

• The increment is smaller than a preset value

The golden-section method uses an optimized division of the intervals to obtain the desired
containment of the minimum point with the least possible number of iteration steps. For
deriving the method, consider a function with one single minimum within the normalized
interval xL = 0 and xU = 1, see Fig. Next, two more points x1 and x2 are chosen inside
of the initial interval. The example illustrated in Fig. 6.17 indicates a higher function
value at x1 so that this point yields an improved lower bound. The coordinates of the new
points are chosen such that the interval is reduced by the same fraction. Since generally

y o

y u

y 1
y 2

x u x 2x 1 x o

Figure 6.17: Sketch of the golden-section algorithm

c ETH Zürich IDMF-CMAS, March 1, 2018

6.8 Line-Search Methods 101

it is not known which of the points x1 and x2 will become the new bound, the new points
must be chosen symmetrical to the midpoint of the interval:

xU − x2 = x1 − xL . (6.88)

The ratio between the old and the new intervals remains constant at each iterate, if the
points abide to
x1 − xL x2 − x1
= . (6.89)
xU − xL xL − x1
From Fig. 6.17 we have that
x2 = 1 − x1 (6.90)

and with this we obtain from (6.89) that

1 − 2x1
x1 = → x21 − 3x1 + 1 = 0 → x1 = 0.38197. (6.91)
1 − x1

Thus, we obtain for the ration between x2 and x1 ,

x2
= 1.61803 (6.92)
x1

This ratio has some significance in fine arts, architecture, and philosophy: proportions in
the golden-section ratio are sensed esthetic or harmonious. In the more general case, when
xL 6= 0 and xU 6= 1, new bounds are obtained by the formulae

x1 = xL + τ (xU − xL )
. (6.93)
x2 = xU − τ (xU − xL )

6.8.3 Quadratic and Cubic Approximation Methods

The approximation methods approximate the objective function along the search direction
by a simple polynomial the minimum point of which is calculated analytically. Whether
or not the approximation methods are far better or far worse than the interval-halving or
golden-section methods, depends on how well the objective can approximated. As a rule,
the objective can be better approximated by quadratic or cubic polynomials the closer one
gets to its minimum point.
The quadratic approximation methods model the original objective function by

f = a0 + a1 x + a2 x2 . (6.94)

One version uses three supporting points (x1 , f1 ), (x2 , f2 ), and (x3 , f3 ) to determine the
coefficients ai of (6.94) fitting through these points. An illustrative example of this is given
in Fig. 6.18 The coefficients a0 , a1 , and a2 follow from the system of equations
    
1 x1 x21 
 a0 
 
 f1 

 2
   
 1 x2 x2  a = f (6.95)
  1   2 
2
 
  
1 x3 x3 a2 f3
  

c ETH Zürich IDMF-CMAS, March 1, 2018

102 Search for Design Improvement: Mathematical Programming

1 .8
1 .6
1 .4
1 .2
1

f(x )
0 .8
0 .6
0 .4
0 .2
0
-1 -0 .8 -0 .6 -0 .4 -0 .2 0 0 .2 0 .4 0 .6 0 .8 1
x

Figure 6.18: Approximation of the objective f = x4 + 13 x3 + 21 x2 −0.4x+1 with a quadratic

polynomial and three supporting points at x1 = 0, x2 = 0.5, and x2 = 1.0.

The equations 6.94 can be resolved after Cramer’s Rule:

     
f1 x1 x21 1 f1 x21 1 x1 f1
1  1  1
     
a0 = D  f2 x2 x22 
 , a 1 = D  1 f2 x22  , a2 = D
 1 x2 f2 
 
f3 x3 x23 1 f3 x3 2 1 x3 f3
  (6.96)
1 x1 x21
 
2 
 1 x2 x2 
D=
1 x3 x23

If one places value upon a translucent symbolic resolution, the form [23]

f = b0 + b1 (x − x1 ) + b2 (x − x2 )(x − x3 )
, (6.97)
f0 = b1 + b2 (2x − x2 − x3 )

after substituting the supporting points for the variable x, leads to the less strongly coupled
system of equations

f1 = f (x1 ) = b0 + b2 (x1 − x2 )(x1 − x3 )

f2 = f (x2 ) = b0 + b1 (x2 − x1 ) . (6.98)
f3 = f (x3 ) = b0 + b1 (x3 − x1 )

It can be resolved more concisely and yields the result

f2 −f1
f3 − f2 x2 −x1 − b1
b1 = , b0 = f2 − b1 (x2 − x1 ), b2 = (6.99)
x3 − x2 x1 − x3
The coefficients ai of the polynomial (6.94) are related with the bi by

a0 = b0 − b1 x1 + b2 x2 x3
a1 = b1 − b2 (x2 + x3 ) . (6.100)
a2 = b2

c ETH Zürich IDMF-CMAS, March 1, 2018

6.8 Line-Search Methods 103

The extreme xe is given by (6.13). The iteration requires replacement of one of the three
supporting points by the newly found point xe . The function value fe at the new point
follows from evaluating the original objective function. For instance, one can discard the
old supporting point with the highest function value. The algorithm is relatively easy to
implement and works without additional gradient information.
The cubic approximation method uses the model

f = a0 + a1 x + a2 x2 + a3 x3
. (6.101)
f0 = a1 + 2a2 x + 3a3 x2

It needs only two supporting points but at each of these the function value f as well as the
slope f 0 must be calculated. An illustrative example is given in Fig. 6.19. The coefficients

Figure 6.19: Approximation of the objective f = x4 + 31 x3 + 12 x2 − 0.4x + 1 with a cubic

polynomial and two supporting points at x1 = 0 and x2 = 1.

follow from the system of equations

    
1 x1 x21 x31  a0
 
 
 f1 

  
 
 

 0 1 2x1 3x2    a1 
  f2 
 
1 
= . (6.102)


 1 x2 x22 x32  

  a2 




 f3 


   
0 1 2x2 3x22  a3
 
 
 f 
4

Again, a suitable form allows a more concise resolution by hand:

f = b0 + b1 (x − x1 ) + b2 (x − x1 )(x − x2 ) + b3 (x − x1 )2 (x − x2 )
. (6.103)
f 0 = b1 + b2 [(x − x1 ) + (x − x2 )] + b3 [(x − x1 )2 + 2(x − x1 )(x − x2 )]

By using the abbreviations 4x = (x2 − x1 ), 4f = (f2 − f1 ), and 4f 0 = (f20 − f10 ) we write

the system of equation the solution of which determines the coefficients bi :

f1 = f (x1 ) = b0
f2 = f (x2 ) = b0 + b1 4x
. (6.104)
f10 = f 0 (x 1) = b1 − b2 4x
f20 = f 0 (x2 ) = b1 + b2 4x + b3 4x2

c ETH Zürich IDMF-CMAS, March 1, 2018

104 Search for Design Improvement: Mathematical Programming

The result is
4f 4f − 4x4f 0 4xf10 + 4xf20 − 24f
b0 = f1 , b1 = , b2 = , b 3 = . (6.105)
4x 4x2 4x3
The coefficients ai of the polynomial (6.101) are related with the bi by

a0 = b0 − b1 x1 + b2 x1 x2 − b3 x21 x2
a1 = b1 − b2 (x1 + x2 ) + b3 (2x1 x2 + x21 )
. (6.106)
a2 = b2 − b3 (2x1 + x2 )
a3 = b3

The extreme points then follow from setting the derivative (6.101) equal to zero:
a2 a1
x2e + 2pxe + q = 0, p= , q= (6.107)
3a3 3a3
If the original objective is a quadratic, the approximation can only be quadratic as well so
that the coefficient a3 vanishes. Then xe follows from (6.13). Else the solution of (6.107)
is given by p
−a2 ± a22 − 3a1 a3
xe = (6.108)
3a3
The sign of the root must be chosen correctly. The cubic approximation offers a maximum
of two extreme points the one of which is the desired minimum and the other one is a
maximum. At the minimum point the second derivative is positive:

f100 = 2a2 + 6a3 xe ≥ 0. (6.109)

Combining (6.109) with (6.108) yields after some simplifications the result
q
± a22 − 3a1 a3 ≥ 0 (6.110)

and reveals that by choosing the positive sign of the root in (6.108) we always obtain
the minimum point xe = xmin . Extreme values exist only if the argument of the root
is positive. An unfortunate choice of the supporting points may yield an approximating
polynomial without extreme points. This can be avoided by checking the condition

x ≤ xe → f0 ≤ 0 ∧ x ≥ xe → f 0 ≥ 0. (6.111)

is fulfilled or that the extreme point is bounded between the two supporting points.
Often, the numerical effort, or the programming difficulty, of obtaining derivative infor-
mation lets the cubic approximation method based on four supporting points appear more
attractive. Here, the parameters ai can easily be obtained analytically by using the form

f = b0 + b1 (x − x1 ) + b2 (x − x2 )(x − x3 ) + b3 (x − x2 )(x − x3 )(x − x4 ). (6.112)

Inserting the four supporting points yields the system of equations:

f1 = f (x1 ) = b0 + b2 4x12 4x13 + b3 4x12 4x13 4x14

f2 = f (x2 ) = b0 + b1 4x21
, (6.113)
f3 = f (x3 ) = b0 + b1 4x31
f4 = f (x4 ) = b0 + b1 4x41 + b2 4x42 4x43

c ETH Zürich IDMF-CMAS, March 1, 2018

6.8 Line-Search Methods 105

where 4xij = xi − xj . The equations (6.113) are resolved to give the parameters bi :

4f32
b1 = ;
4x32
b0 = f3 − b1 4x31 ,
. (6.114)
f4 − b0 − b1 4x41
b2 = ,
4x42 4x43
f1 − b0 − b2 4x12 4x13
b3 =
4x12 4x13 4x14

Ordering the terms of (6.112) after powers of the variable x yields the form

f = x0 [b0 − x1 b1 + (x2 x3 )b2 − (x2 x3 x4 )b3 ]

+ x1 [ b1 − (x2 + x3 )b2 + (x2 x3 + x3 x4 + x4 x2 )b3 ]
. (6.115)
+ x2 [ b2 − (x2 + x3 + x4 )b3 ]
+ x3 [ b3 ]

From (6.115) the parameters ai , in terms of bi , can be immediately obtained.

6.8.4 Brent’s Routine

The success, or efficiency, of the various line search methods depends very much on the
objective function topology. For instance, if a gradient calculation is costly compared to
a function evaluation, the line search methods depending on gradient information can be
less efficient in terms of computing time although the number of iterations for one line
search may be significantly smaller. Principally, we desire the routine to work reliable on
a wide range of function topologies and to consume as little computing time as possible.
In structural optimization we generally do not know much in advance about the topology
of a given problem. Therefore, we do not wish the performance to depend on topology-
dependent tuning parameters.
The combination of several methods in one routine often gives the best results in terms
of reliability as well as efficiency for a range of function topologies. In the following, a
collection of methods and algorithms is presented that is combined to give a reliable and
efficient minimization routine. All of this follows closely the idea’s underlying Brent’s
routine by [43] available in the Numerical Recipes. The method presented here shares
much of the basic approach with Brent’s Routine available from the Numerical Recipes
Library but has been coded with no direct reference to that routine. Both routines perform
similar.

c ETH Zürich IDMF-CMAS, March 1, 2018

106 Search for Design Improvement: Mathematical Programming

Basic Ideas and Algorithms

The routine starts with bracketing the minimum point along the search direction. The
reliability of the further approximation iterations is based on the knowledge that the
minimum point is between those brackets. The further steps try to reduce the bracketed
range. At the beginning, the reduction is based on the slower but more robust golden-
section bracketing method and at a certain point the more efficient but less robust method
of three-point quadratic approximation is used instead.
Upon calling of the search routine the following information is relevant input:
• Search direction vector s

• Maximum step length xmax

• Starting point in optimization variables space x0

• Object function value at starting point f

• The integer argument value of minimization loop

• Small epsilon value 

The step-length value is initially set equal to . The starting point in variables space refers
to a zero distance along the search direction. Points at other distances are calculated
according to. The initially given point, transferred in the array OP V AR, must be stored
in another array that is called OP ST ORE because OP V AR is updated every time a new
point at the distance x from the starting point is evaluated. After its initiation and during
one line search OP ST ORE is not changed but rather used as reference.
The routine first checks the descent property. It ensures that the line search makes any
sense at all because one may not hope to find a minimum in directions of increasing
function values. When the inner product of the gradient and search direction vectors is
non-positive the descend property is given. Otherwise, the routine returns.
Next, the minimum point is bracketed, i.e. an upper and a lower bound of the minimum
point are established. The first bracket spans the distance from the starting point x1 = 0
to the point x4 where an increase of the initially decreasing objective function value is
observed. The value of the point x4 is obtained by multiplying the value STEP with
increasing powers of ten until the newly calculated objective function value FUN exceeds
the previously calculated value fmin by at least ten percent. However, x4 may not be
greater then the maximum step length xmax that prevents the line search to lead into the
infeasible region.
An initial set of four supporting points is created by the points x1 and x4 that mark the
range of the bracket, and two additional points x2 and x3 that lie in the interior of the
bracket and divide the range of the bracket according to the golden-section ratio, see Fig.
6.20. The variables xL and xU carry the information on the upper and lower limits of the
range. The variable iQ is used to switch from the golden section method the quadratic
approximation method depending on a criterion explained below.

Loop on Refining the Brackets on the Minimum Point

First, from the range defined by the four initially found points a new and smaller range
containing the minimum point is selected. The smaller range is bracketed either by the
points x1 and x3 or by the points x2 and x4 . Referring to the situation sketched in Fig.
6.20, the smaller range is defined by x2 and x4 because of the four points x4 is the one

c ETH Zürich IDMF-CMAS, March 1, 2018

6.8 Line-Search Methods 107

N  N N ! N "

Figure 6.20: Golden section method supporting point on a range bracketing the minimum
point

with the smallest objective function value. A set of three points is now defined that
consequently contains x2 , x3 , and X4 of the four-point set. The lower limit xL of the
narrowed range is set equal to x1 of the three-point set and the upper limit is set to x3
of the three-point set as shown in Fig. 6.21. In the Situation depicted in Fig. 6.21 the

X 4 1 X 3 1= X 4 2 X 3 2= X 4 3 X 3 3= X 4 4

Figure 6.21: Narrowing the bracket by reduction of four supporting points to three

minimum point, in terms of the original four supporting points, is point x4 . Using the
quadratic approximation method for estimating a new minimum point implies a risk that
the new minimum point is estimated outside of the current bracketed range as indicated
in Fig. 6.22. This is not wanted because it has been established before that the minimum
must be inside the bracketed range. Therefore, the integer switch iQ is set equal to zero
so that the golden section method instead of the quadratic approximation methods is
selected. The golden section method uses the reduced range defined by the set of three

X 4 1 X 3 1= X 4 2 X 3 2= X 4 3 X 3 3= X 4 4

Figure 6.22: Quadratic approximation incorrectly indicated minimum outside of bracketed

range

supporting points to replace the previous set of four points by a new set of four points.
Based on the situation depicted so far, the result is shown in Fig. 6.23. Characteristic for
the golden section method is that the new four points now divide the new and narrower
range with the same ratios as the previous set divided the previous range. This is achieved
by using the selected three points of the previous set and adding just one new point. The

c ETH Zürich IDMF-CMAS, March 1, 2018

108 Search for Design Improvement: Mathematical Programming

new point is point x3 of the new set in Figure 6.23. The objective function is evaluated

: "  : " : " ! : " "

Figure 6.23: Updated four-point set

at the new point and it is determined which point of the set has the minimum function
value found so far. As Fig. 6.23 indicates, point x4 remains the minimum point for the
considered example. The minimum point and function value found by the golden section
method are called xGS and fGS . One could continue iterations on the golden section
method by restarting the narrowing of the bracketed range as explained above. However,
the routine CBS EARCH performs a quadratic approximation as well because the case
of the estimated minimum point lying outside of the bracketed range is only a possibility
but, generally, not a certainty.
The quadratic approximation method with three supporting points is explained in section
6.8.3. Basically, a quadratic polynomial is fitted to the three supporting points of the
current three-point set. As a condition for an extreme point, the spatial derivative of the
polynomial is set equal to zero. The resulting equation is resolved for x. The second
spatial derivative must be positive if x is a minimum point. If it is not positive or if the
minimum point x is calculated outside the bracketed range, the result is disregarded and
c is placed in the middle between the two brackets. In the program, x is named XQA and
the objective function is evaluated and the value is called F QA. Even if the quadratic
approximation as such failed, it is possible that the newly found point in the middle of the
range might be a valid minimum point. Therefore, the results of the golden section and
the quadratic approximation codes are compared with each other.
The step of comparing the golden-section and the quadratic-approximation methods is
skipped when the golden section method has not been used. A criterion is introduced
that characterizes the shape of the objective function within the bracketed range. If
the function is extremely steep, numerical round-off errors may impair the quality of the
quadratic-approximation results. If the function is not terribly steep and the function
value found by the quadratic approximation method is smaller than that found by the
golden section method, the result of the quadratic approximation is chosen for the new
minimum point and the switch IQ is set equal to 1 so the golden section method is not
used anymore in order to save computing time. Referring to the situation depicted in Figs.
6.20 through 6.23 one can expect one or two more iterations on the golden section method
before the program switches to the then more effective quadratic approximation method.
After it has been decided upon which one of the points XGS and XQA determines the
new minimum-point step length XN EW , the new point x0 in variable space is calculated.
If F N EW is smaller than the previous found minimum value F M IN , the minimum point
is update by replacing XM IN with XN EW and F M IN with F N EW . The newly found
step-length value XN EW must now replace one of the points of the existing four-point
sets according to its value with respect to the other points.
The newly found step-length value XN EW and the current three-point produce the new

c ETH Zürich IDMF-CMAS, March 1, 2018

6.8 Line-Search Methods 109

four-point set. In case the new point is found by quadratic approximation the golden-
section ratio is generally lost. Fig. 6.24 sketches that a newly found point is located
between points two and three of the current three-point set so that it becomes point three
of the new four-point set. Its position deviates from that corresponding with the golden
section ratio, which does not really matter because when the quadratic search starts being
preferred over the golden section method, the latter one is not used anymore during the
remainder of the current line search. The point with the smallest function value has the

: !  : ! : ! !

: "  : " : " ! : " "

Figure 6.24: Minimum point successfully estimated by quadratic approximation

step length XM IN and the function value F M IN . The point with the second smallest
function value has the step length XSEC and the function value F SEC. The points are
needed for the termination or convergence criteria, respectively. Referring to Fig. 6.24,
the third point would be XM IN and the fourth point would be XSEC.
There are two criteria that must both be fulfilled simultaneously in order to terminate the
line search.

1. Close at the true minimum point, the actual function value FUN must be very close
to the function value F N EW A as estimated by the quadratic approximation curve.
The variable DF carries a relative error that is calculated as the absolute value of
the difference divided by the sum of the two values

2. Close at the true minimum point, the relative change of the step length with respect
to the absolute value is small. The criterion is set up so that smaller values result
as the minimum point moves closer to the center of the bracketed range.

A third criterion terminates the search even if the two described above do not. A perfectly
solved DoD problem corresponds to a zero objective function value. It is therefore safe
to terminate the search when the value of the objective function itself is smaller than the
chosen epsilon value.
If the criteria do not terminate the search, the iterations continue with extracting a new
three-point set from the four-point set shown in Fig. 6.24.

c ETH Zürich IDMF-CMAS, March 1, 2018

110 Search for Design Improvement: Mathematical Programming

c ETH Zürich IDMF-CMAS, March 1, 2018

6.9 Lagrange Multiplier Method Numerical Optimization 111

6.9 Lagrange Multiplier Method Numerical Optimization

The Lagrange multiplier method requires knowledge of the Lagrange multipliers, whose
values are often not accessible through analytic formulas. The problem can be mitigated
by expanding the set of optimization variables with the set of Lagrange multipliers so
that numerical optimization methods search within both sets simultaneously for the best
numerical approximation of constrained minima. At constrained minima the Lagrangian
formulation seen in Sections 3.5.1 and 5.5 is stationary; the stationary points are minimum
with respect to the optimization variables and maximum with respect to the Lagrange-
multipliers. Within the unified space of optimization variables and Lagrange multipliers
the Lagrangian form has saddle points which cannot be found with minimum search meth-
ods. Recasting the Lagrangian formulation into an alternative formulation transforms all
saddle points to minima by the method explained in the following Section 6.9.1.

6.9.1 Modified Lagrangian with Local Minima

A modified form of the Lagrangian (3.26) on page 21 is built with its gradients with respect
to the optimization variables x and its Lagrange multipliers λ:

∇x L = ∇x f + λg ∇x g + λh ∇x h ; ∇λ L = g + h (6.116)

If no other means are available, the partial derivatives can be approximated with:

L(x + i , λ) − L(x, λ) L(x, λ + i ) − L(x, λ)

(∇x L)i ≈ ; (∇λ L)i ≈ (6.117)
 
The squared norms of both gradients are non-negative and vanish at critical points. There-
fore, the objective L̄
L̄ = ∇x LT · ∇x L + ∇λ LT · ∇λ L (6.118)
possesses minima at the original Lagrangian saddle points and these can be found with
any minimum-searching optimization routine simultaneously for both x and λ.
The sample problem addressed in Section 3.5.1 and revisited in Section 5.5 gives the
modified Lagrangian:
 2
x+2
L̄ = − λ + (1 − x)2 (6.119)
10
Fig. 6.25 shows the contour lines of the modified Lagrangian, which is a quadratic poly-

Figure 6.25: Two-dimensional variable space with two linear constraining functions

nomial, and that the method of conjugate gradients by Fletcher and Reeves finds the
constrained minimum exactly within two line searches.

c ETH Zürich IDMF-CMAS, March 1, 2018

112 Search for Design Improvement: Mathematical Programming

6.9.2 Algorithm for Removing Constraint Violations

The following algorithm removes several constraint violations in one step if the constraining
functions are linear. The derivations are illustrated for a 2-dimensional problem with two
constraining functions by Fig. 6.26. The figure also illustrates that each constraining
equation reduces the search space dimensionality by one so that the feasible region is
reduced to one feasible point if the number of constraints equals the number of problem
variables.

N 

D =  D = 

N 

? N
N L 
L

u @
u @ 
N L

Ñ D
Ñ D 

Figure 6.26: Two-dimensional variable space with two linear constraining functions

Let xv be a violating point in an infeasible region of the search space. It is then desired
to find a new point x0 where all the violated constraints remain active with zero values.
The vector pointing from xv to x0 is called c and is a linear combination of the gradients
∇gi of the constraining functions at the point xv :

c = µi ∇gi . (6.120)

It is assumed that xv is close enough to the feasible region so that the constraining func-
tions gi can be linearly approximated along c:

∇giT ∇gi
gi = βi gi0 = βi = βi |∇gi |. (6.121)
|∇gi |

In the case of side constraints, the assumption made above is fully justified because the
constraining functions are linear anyway and (6.121) holds exactly. The distances of the
functions (or hyper planes) gi = 0 to the point xv are
gi
βi = . (6.122)
|∇gi |

The constraining function gradients ∇gi will generally not be linearly independent. There-
fore one may not confuse the linear-combination weight factors µi with the respective dis-
tances βi . The following method for constructing vectors di that have components in the
direction of the gradients ∇gi only, i.e. that are completely decoupled from all the other
gradients ∇gj
di T ∇gj = 0, i 6= j, (6.123)

c ETH Zürich IDMF-CMAS, March 1, 2018

6.9 Lagrange Multiplier Method Numerical Optimization 113

requires that not two or several of the gradients ∇gi are equal to each other. The vectors
di are then constructed from all remaining violated constraining function gradients dj :

di = ∇gi − γij ∇gj (6.124)

After substituting (6.124) in the condition (6.123) and re-ordering one obtains the factors
γij : 
 i=j → γij = 0
T (6.125)
 i 6= j → γij = ∇giT ∇gj
∇g ∇gj j

Of course the correction vector c (6.120) can as well be written in terms of the vectors di :

c = νi di . (6.126)

In contrast to the gradients ∇gi the vectors di are linearly independent from each other
and so the weight factors νi can be calculated from the respective single constraining
functions. The component νi di of the vector di in the direction of the gradient ∇gi must
be equal to the distance βi :
∇gi
νi di = βi . (6.127)
|∇gi |
Solving (6.127) for νi and substituting (6.122) for βi yields
gi
νi = T
. (6.128)
di ∇gi
In a program the first step is to identify the violated constraints and to obtain the respective
gradients. The factors γij are calculated after (6.125) and the linearly independent set of
vectors di is computed after (6.124). Next the factors νi are calculated after (6.128) and
the correction vector c follows from (6.126). In cases where the assumption (6.121), the
constraining functions being linear along c, is only an approximation, the new point must
be checked for remaining constraint violations. If any of the constraining functions have
values greater than zero the procedure must be repeated until all constraints are satisfied.
An illustration to the here presented concept is given in Fig. 6.26. If multiple vectors
appear, they must first be removed from the set. From a set of multiple gradient vectors
∇gi , the one associated with the constraining function with the greatest value should be
kept.

c ETH Zürich IDMF-CMAS, March 1, 2018

114 Search for Design Improvement: Mathematical Programming

c ETH Zürich IDMF-CMAS, March 1, 2018

6.10 Objective Function Derivatives 115

6.10 Objective Function Derivatives

The most effective optimization algorithms, at least in terms of small numbers of itera-
tions, can be constructed when gradient information is available for the calculating the
search directions as well as performing the line searches. In terms of computing time, effi-
ciency depends highly on the numerical effort for calculating the gradients. If the gradient
calculation is very expensive, recourse to gradient-free minimization methods will yield a
better time efficiency.
How expensive is the gradient calculation when the objective function evaluation involves
solving a FEM model? First of all, that depends on the computational effort for solving
the system of equations. The computational effort is mainly dependent on the number of
unknowns (degrees-of-freedom) as well as the structure of the stiffness matrix. When it
comes to gradient calculation, the numerical effort increases linearly with the number of
independent optimization variables.

6.10.1 Differences Method

The differences method calculates the gradient at a reference point (x0 , f0 ) by performing a
loop i = 1, N on all N optimization variables. Each variable xi is increased by an increment
4xi and the objective is then evaluated. The difference between the objective value of
the modified design and f0 , divided by the increment, approximates the sensitivity of the
objective with respect to the respective optimization variable and, thus, the respective
entry of the gradient vector.

f (x0 + 4xi ) − f (x0 )

∇fi = . (6.129)
4xi

The differences method implies that the structural model must be evaluated N times for
each gradient calculation and is therefore also referred to as a brute-force method. The
expense for one gradient calculation equals roughly N function evaluations. For problems
with a high number of optimization variables the expense for one gradient calculation may
become much greater than one function evaluation, abolishing the efficiency of gradient-
based minimization methods.

6.10.2 Sensitivity-Formula Gradient Calculation

Gradient calculation after the sensitivity formula for finite-element models requires no ad-
ditional function evaluations. The finite-element method for linear static elasticity prob-
lems assembles a system of linear equations

Kũ = r (6.130)

that must be solved for the unknown nodal-point displacements, or degrees of freedom, ũ.
The stiffness matrix K is a numerical representation of the behavior of the structure and
is assembled from the stiffness matrices of the individual finite elements into which the
structure domain is divided. The right-hand-side vector r contains the nodal forces derived
from the loads acting upon the structure. Once (6.130) is resolved, or the primary solution
ũ is obtained, any other quantity is calculated as a function of the primary solution, and
possibly model parameters, by the postprocessing step. Having gradient calculation in
mind, we assume that the objective function depends explicitly and implicitly, through
the primary solution, on the optimization variables x, f = f (x, ũ(x)). Therefore, the

c ETH Zürich IDMF-CMAS, March 1, 2018

116 Search for Design Improvement: Mathematical Programming

sensitivity of the objective with respect to a change of the ith optimization variable is
obtained by the rule of chains:

∂f ∂f ∂ũ
∇fi = + (6.131)
∂xi ∂ũ ∂xi
The postprocessing is most often numerically insignificant if compared with the effort to
∂f ∂f
solve (6.130). Therefore, the quantities ∂x i
and ∂ũ are obtained at low numerical cost.
The cost of calculating the sensitivity of the primary solution ũ upon the optimization
variables x is much reduced by the formula derived in the following.
The first step in deriving the sensitivity formula is to form the partial derivative of (6.130)
with respect to the ith optimization variable xi ,

∂K ∂ũ ∂r
ũ + K = . (6.132)
∂xi ∂xi ∂xi

Resolving (6.132) for the sensitivity of the primary solution upon xi yields the desired
sensitivity formula  
∂ũ ∂r ∂K
= K−1 − ũ . (6.133)
∂xi ∂xi ∂xi
On first sight one could conclude that nothing has been gained since, after all, the inversion
of the stiffness matrix is at least numerically equivalent to resolving (6.130). In fact,
however, (6.132) allows to calculate the gradient very efficiently, especially for large models.
Let us realize that the inverse at the reference point, at least in terms of a triangulated
matrix, exists at the point (x0 , f0 ) because the objective has just been evaluated there.
It remains to calculate the terms within the braces. The sensitivity of the nodal forces
upon the optimization variables, which are design parameters in structural optimization,
is easily calculated at practically no cost. Often, if the design is subject to fixed and
concentrated loads, the nodal forces are independent of the design parameters so that the
first term in braces becomes zero. The second term in braces is a product of the sensitivity
of the stiffness matrix with the existing primary solution vector. Often, the change of a
design parameter affects only part of the structural model, or only a small subset of all
the finite elements making up the whole model. It is therefore not necessary to assembly
the complete stiffness matrix. Instead, one may only consider the elements whose shape or
other properties are affected by a change of xi . Then, the numerical effort may be much
less than for a complete stiffness matrix assembly, which is again much less expensive than
a solution of the system equations. Fig. 6.27 illustrates the point. The example refers

Figure 6.27: Shape optimization and sensitivity formula for gradient calculation

to shape optimization where the positions of boundary nodes is variable in the direction
perpendicular to the boundary. The variation of the position of one node affects the shape

c ETH Zürich IDMF-CMAS, March 1, 2018

6.11 Global Optimization 117

of the two adjacent finite elements while all other elements remain unchanged. The stiffness
matrices of the two elements with the reference shapes are subtracted from those with the
modified shapes. The difference is multiplied with the existing displacement vector, where
only the displacements on the nodes of the affected elements need to be considered.
Although the sensitivity formula allows gradient calculation at much lower cost than the
differences method, it may be difficult to implement in existing general-purpose FEM
programs.

6.11 Global Optimization

Mathematical programming may obtain a local minimum and terminate there because
the necessary optimality conditions are satisfied. Thus, the local minima of a non-convex
objective function may stand in the way of obtaining its absolute, or global, minimum.
This is a disadvantage and practical weakness of mathematical programming that may
otherwise be so much more efficient than genetic algorithms, for instance. Attempts to
mitigate this weakness of mathematical programming lead to methods such as

• Multi-start method

• Tunneling method

6.11.1 Multi-Start Method

The multi-start method assigns reasonable ranges of values to the individual optimization
variables. The ranges should be chosen inside of the feasible region. From these ranges a
number of values, maybe in equidistant spacing, are selected. The values define a hyper
mesh of points in the N dimensional space of the optimization variables. One after another,

x * p o in t m e s h
x 2

x o p t
2
f(x 1,x 2)= c o n s t.

x 1
o p t x 1

Figure 6.28: Function with minima in considered region and starting-point mesh.

these points are used as starting points for a minimum search by using mathematical
programming. If the search from each starting points leads to the same minimum point,
one may be dealing with a convex function. Operating upon a non-convex function, the
multi-start method will generally obtain several minimum points (if there are several in
the feasible region). The one with the lowest function value may be taken for the global
minimum although there is no proof that the absolute minimum may not have escaped the
search. The probability of finding the global minimum increases with increasing density
of the starting-point mesh. However, a high mesh implies an immense number of starting
points, or optimization processes, when the number of dimensions is high. Covering the

c ETH Zürich IDMF-CMAS, March 1, 2018

118 Search for Design Improvement: Mathematical Programming

range of each of the N variables with M points yields M N starting points altogether. It
goes without saying that the method is forbidding when it comes to large optimization
problems.

6.11.2 Tunneling Method

Assume non-convex function and a local minimum point having just been obtained by
mathematical programming. Starting from this local minimum point, the tunneling method
invented by Levy and Gomez [48] finds systematically a new minimum point with lower
function value. The iterative process ends with finding the global minimum point. The
method has the property that the other local minimum points, having higher function
values than the most recently found minimum point, are not considered, or tunneled,
hence the name of the method. Each iteration of the tunneling method consists of two
phases:

1. Minimization phase: starting from the point x1 , a new local minimum point x∗1
is obtained by mathematical programming

2. Tunneling phase: starting from the local minimum point x∗1 , a different point x2 ,
having the same function value than x∗1 or fulfilling f (x2 ) = f (x∗1 ), is searched

After the tunneling phase follows the return to the minimization phase. With given min-
imum point x∗i a tunneling function t,

F (x) − f (x∗i )
t(x, x∗i , si ) = , (6.134)
[(x − x∗i )T (x − x∗i )]si

is defined. The parameter si must be chosen so that the denominator tends stronger
against zero than the numerator when x → x∗i so that the tunneling function value is
different from zero:
t(x → x∗i , x∗i , si ) 6= 0. (6.135)
On the other hand, the value si should also be such that the tunneling function decreases
with increasing distance from the latest minimum point x∗1 :

t(x∗i + 4x, x∗i , si ) ≤ t(x∗i , x∗i , si ). (6.136)

If the conditions (6.135) and (6.136) are satisfied, any null of the tunneling function t is a
point where the objective function f has the same value as the reference minimum point.
This new point is a suitable starting point for a new minimum search and guarantees that
the next obtained minimum point has a lower function value than the previous one. The
tunneling function corresponding to the global minimum has no null.
The tunneling method is demonstrated on a polynomial of degree six that depends on only
one variable. The objective function shown in Fig. 6.29(a),

f (x) = x6 − 16x5 + 100x4 − 310x3 + 499x2 − 394x + 140, (6.137)

has a minimum point at x∗1 = 1 with a value f (x∗1 ) = 20. For purpose of demonstrating
an unusable tunneling function we choose a value si = 0.5 for the root parameter. Then,
the tunneling function shown in Fig. 6.29(b) is obtained by subtracting 20 from the
polynomial (6.137) and dividing the result by x − 1. Obviously, that tunneling function
does not satisfy the condition (6.135) because it has a null at x∗1 or t(x∗1 ) = 0. However,
choosing si = 0.5 obtains the tunneling function shown in Fig. 6.29(c) by subtracting 20

c ETH Zürich IDMF-CMAS, March 1, 2018

6.11 Global Optimization 119

(a) (b) (c)

Figure 6.29: Function with objective (a) and unusable (b) and usable (c) tunneling func-
tions

from the polynomial (6.137) and dividing the result by (x − 1)2 . This tunneling function
is usable because its value at x∗1 is different from zero and it decreases monotonously
from there to the next null at x2 = 2. The null is quickly approximated by employing
the Raphson method. From there, mathematical programming will obtain the other local
minimum point x∗2 ≈ 2.517.

c ETH Zürich IDMF-CMAS, March 1, 2018

120 Search for Design Improvement: Mathematical Programming

c ETH Zürich IDMF-CMAS, March 1, 2018

Chapter 7

Stochastic Search

A general optimization problem can be characterized by the structure of its search space, its
objective space and its objective function. The optimization variables are parameterized
within the search space (e.g. Rn in a real-valued optimization problem). Candidate
solutions are rated in the objective space (e.g. R for single-objective problems or Rn for
multi-objective problems). The methods from the field of Mathematical Programming are
well suited for continuous, homogeneous search spaces, continuous, scalar objective spaces
and smooth, convex objective functions. These qualities can not always be expected to be
present in an engineering problem at hand, indicating that deterministic algorithms are
inappropriate. Therefore, a lot of research has focused on the development and applications
of stochastic search methods to overcome this limitation. This chapter gives an overview
over the field of stochastic optimization. Some of the concepts presented earlier in this
lecture are directly transferable to stochastic search. However, it should be pointed out
that the separation of the parameterization concept (or in general the representation) and
the search method does no longer hold for all search methods presented in the following.
Some of the methods evolved closely affiliated representation concepts dedicated at a
specific problem type.

7.1 Introduction to stochastic search and optimization

According to [49] a stochastic search method is characterized by the presence of one or
both of the following properties:

• There is random noise in the measurements of the criterion to be optimized and/or

related information.

• There is a random choice made in the search direction as the algorithm iterates
toward a solution.

These properties contrast with one of the basic hypothesis of Mathematical Programming,
where it is assumed that one has perfect knowledge about the objective and depending
on the method on its derivatives and that this information is used to determine a search
direction through to a deterministic rule. In many applications, this information is not
available. This is usually referred to as black-box optimization.
A large group of stochastic search methods is inspired by biology or physics, e.g.
Simulated Annealing, Evolutionary Algorithms, Ant Colony Optimization and Swarm Op-
timization. Others imitate learning mechanisms, e.g. Tabu Search or Neural Networks.

c ETH Zürich IDMF-CMAS, March 1, 2018

122 Stochastic Search

The methods presented in the following sections rely on two model assumptions con-
cerning the objective function:

1. Pointwise sampling of the search space allows to get a kind of a problem landscape
at least locally.

2. Better solutions can be found close to already visited good solutions.

These assumptions are less restrictive than the typical polynomial models employed in
Mathematical Programming. However, there are problem types which are excluded by
this model. A typical optimization problem which is not a model of the above concept is :

0 if x = 0
minimize F (x) = (7.1)
1 otherwise

7.1.1 Neighborhood concept

From the second model assumption stated above the requirement to define a measure for
closeness arises. This is usually done by the definition of a problem specific neighborhood -
function. The neighborhood -function neighborhood(x) returns a subset Xn of the search
space X which is close to the solution x. In a real-valued search space X = Rn the
Euclidian distance is a common choice to define a neighborhood function, i.e.

neighborhood(x, r) = x0 ∈ X||x0 − x| ≤ r ⊆ X

(7.2)

In a binary search space the Hamming distance H(x, x0 ) may serve as distance measure.
It denotes the number of bits which need to be flipped to change x into x0 .

7.1.2 Strategies in stochastic search

Stochastic global search involves two opponent search strategies named exploration and
exploitation. Both of them characterize the way how new solutions are created within
the iterations of the algorithm. Explorative search addresses unacquainted regions of
the search space, while exploitation inquires the regions around the best solutions known
so far. Exploitation drives an algorithm toward an optimum (which could be a local
or the global one), thus it is usually referred to as convergent component. Exploration
should avoid a premature (i.e. local) convergence and is therefore also called divergent
component. In typical implementations of stochastic search both strategies occur and
the balance between them is either determined from a process parameter to be set by
the user or adapted by the algorithm according to some rules. It should be pointed out
that the chance of finding a global optimal solution depends crucially on the weighing of
exploration versus exploitation. An optimal setup is not known a priori but it depends on
the objective function at hand.

7.1.3 A prototype of a stochastic search algorithm

The search algorithms presented in the following sections fit all in this general
scheme:
1: Create a set of start solutions {xi } ∈ X
2: Calculate F (xi )
3: Initialize the memory M ← {(xi , F (xi ))}
4: Initialize the iteration counter t ← 0

c ETH Zürich IDMF-CMAS, March 1, 2018

7.2 Stochastic Search Algorithms 123

5: while continue(M, t) = 1 do
6: t←t+1
7: Create a set of alternative solutions {x̂i } ∈ X
8: Calculate F (x̂i )
9: Update the memory M ← update (M, {(x̂i , F (x̂i ))})
10: end while
11: Return the best solution x∗ in M
The steps 1, 7 and 9 may contain a random component. The function continue(M, t)
stands for some kind of termination criterion. It returns 1 as long as the current state
of the search (i.e. the iteration counter t and the solutions in the memory M ) does not
indicate a termination.

7.1.4 Performance of stochastic search

The performance of optimization algorithms is usually measured by the number of objec-
tive function evaluations required to find an optimal solution. Since the way a stochastic
algorithm iterates toward a solution is influenced by random decisions this number may
change between different runs of the same algorithm on one problem. Therefore, an empir-
ical performance assessment has to consider multiple runs of an optimization algorithm.

No Free Lunch Theorems for Optimization

D.H. Wolpert, W.G. Macready [50] present a pure theoretical investigation on the perfor-
mance of optimization methods. All optimization problems are considered. The average
performance of an arbitrary pair of optimization algorithms over all optimization problems
is investigated. The so called No Free Lunch Theorems show that the average performance
of these two algorithms has to be the same. The No Free Lunch Theorems are based on
the assumptions that all optimization problems are equally likely and that each algorithm
knows when an optimal solution is found. These assumptions are not met by typical opti-
mization applications. However, the No Free Lunch Theorems point out that a comparison
reporting the performance of a particular algorithm on a few sample problems can only
indicate the behavior on this range of problems considered. A generalization to other
problems may not be possible. Moreover, the optimization algorithm should be chosen
to match the problem type at hand as far as possible. A general optimization algorithm
capable of solving a wide range of problems may loose performance when compared with
a problem-specific algorithm.

7.2 Stochastic Search Algorithms

7.2.1 Random Search
Random search is a global stochastic search algorithm:
1: Create a random start solution x ∈ X
2: Calculate F (x)
3: Initialize the memory M ← {(x, F (xi ))}
4: Initialize the iteration counter t ← 0
5: while continue(M, t) = 1 do
6: t←t+1
7: Create one alternative solutions x̂ ∈ X
8: Calculate F (x̂)

c ETH Zürich IDMF-CMAS, March 1, 2018

124 Stochastic Search

U
9: Add the new solution to the memory M ← {(x, F (xi ))}
10: end while
11: Return the best solution x∗ in M
Random search applies a pure explorative search by just randomly sampling the search
space. The creation of new solutions is not based on solutions stored in the memory.

7.2.2 Stochastic Descent

Stochastic descent is a local search algorithm:
1: Create a random start solution x ∈ X
2: Calculate F (x)
3: Initialize the iteration counter t ← 0
4: while continue(x, t) = 1 do
5: t←t+1
6: Create a random solution x̂ ∈ neighborhood(x)
7: Calculate F (x̂)
8: if F (x̂) ≺ F (x) then
9: x ← x̂
10: end if
11: end while
12: Return the best solution x∗ in M
It transfers the concept of gradient-based descent methods to stochastic methods what
makes it applicable to discrete problems. Stochastic descent is based on a pure exploitation
strategy.

7.2.3 Metropolis Algorithm

The Metropolis Algorithm is a generalization of the stochastic descent algorithm. The
method accepts uphill steps with a certain probability. Therefore, the objective is inter-
preted as energy. The energy difference between two states ∆E = F (x̂) − F (x) determines
the probability for a state transition from x to x̂. The probability is given by
∆E
p(∆E) = e− T (7.3)

whereas T denotes a constant referred to as temperature of the algorithm.

1: Create a random start solution x ∈ X
2: Calculate F (x)
3: Initialize the iteration counter t ← 0
4: while continue(x, t) = 1 do
5: t←t+1
6: Create a random solution x̂ ∈ neighborhood(x)
7: Calculate ∆E = F (x̂) − F (x)
8: Chose a random number k in [0, 1]
9: if ∆E > 0 then
10: x ← x̂
11: else if k ≤ p(∆E) then
12: x ← x̂
13: end if
14: end while
15: Return the best solution x∗ in M

c ETH Zürich IDMF-CMAS, March 1, 2018

7.2 Stochastic Search Algorithms 125

Figure 7.1: Metropolis Algorithm: Probability p for the acceptance of a new solution in
dependence of the energy difference ∆E and the temperature T .

The probability for uphill steps increases with increasing temperature, thus exploration is
emphasized. For low temperatures the algorithm becomes similar to the stochastic descent
algorithm.

7.2.4 Simulated Annealing

In many applications an exploration strategy seems to be useful at the very beginning of

the optimization run. Exploitative search is used in the regions where good solutions have
been found. This requires an adaption of the temperature T on the run of a Metropo-
lis algorithm. Simulated annealing is inspired by the problem of finding an equilibrium
state for frozen n-body systems. There, the decreasing temperature reduces the proba-
bility for state transitions with increasing time. The basic scheme is the same as for the
Metropolis algorithm but with an additional update step for the temperature within the
iteration:
1: Create a random start solution x ∈ X
2: Calculate F (x)
3: Initialize the iteration counter t ← 0
4: while continue(x, t) = 1 do
5: t←t+1
6: T ← update(T, t)
7: Create a random solution x̂ ∈ neighborhood(x)
8: Calculate ∆E = F (x̂) − F (x)
9: Chose a random number k in [0, 1]
10: if ∆E > 0 then
11: x ← x̂
12: else if k ≤ p(∆E) then
13: x ← x̂
14: end if
15: end while
16: Return the best solution x∗ in M

c ETH Zürich IDMF-CMAS, March 1, 2018

126 Stochastic Search

There are several alternatives introduced concerning the temperature update. A common
choice is a linear rule:
Tt+1 = update(Tt , t) = αTt (7.4)
The constant α (0 < α < 1) is called annealing constant. Simulated annealing is a global
search algorithm. The probability Pt that simulated annealing finds a global optimal
solution in t iterations converges to one:

lim Pt = 1 (7.5)
t→∞

This is a special property of the Simulated Annealing algorithm within the field of stochas-
tic search. However, it is usually not relevant for application since the value of t has to
take huge values in order to achieve a high probability Pt .

7.2.5 Evolutionary Algorithms

Müller [1] summarizes concisely the field of evolutionary computation [51]. Fig. 7.2 visu-
alizes that the field divides into the three branches evolutionary programming [52], genetic
algorithms [53, 54], and evolution strategies. Each branch contributes to the class of evo-

E v o l u t i o n a r y C o m p u t a t i o n
E v o lu tio n a r y G e n e tic E v o lu tio n
P r o g r a m m in g A lg o r ith m s S tr a te g ie s

E v o l u t i o n a r y A l g o r i t h m s

Figure 7.2: Selected mathematical programming methods and ordering scheme

lutionary algorithms. The well-written characterization of the evolutionary computation

algorithms, mirrored against mathematical programming, [1], is quoted here:

Evolutionary computation algorithms have in common that they are based

on biological principles such as reproduction, mutation, isolation, recombina-
tion, and selection, applied to individuals in a population. The members of a
population are capable of evolving over time, that is, of adapting to their en-
vironment. In each generation, a surplus of individuals is created by means of
mutation and/or recombination, out of which the most promising individuals
are selected as members of the next generation. In contrast to many determin-
istic algorithms, evolutionary algorithms require only the value of the objective
function for a given point in the parameter space, but do not need gradient
information. Like other stochastic algorithms, evolutionary algorithms employ
randomness which makes them less sensitive to noise, discontinuities, and the
danger of getting trapped in a local optimum. While evolutionary algorithms
do not possess highly efficient convergence properties like for example the con-
jugate gradient method, they have the important feature of being inherently
parallel.

Evolutionary algorithms are inspired by and based upon evaluation in nature. They typi-
cally use an analogy to natural evolution to perform search by evolving solutions to prob-
lems, usually working with large collection or populations of solutions at a time. From the

c ETH Zürich IDMF-CMAS, March 1, 2018

7.2 Stochastic Search Algorithms 127

point of view of classical optimization, EAs represent a powerful stochastic zeroth order
method which can find the global optimum of very rough functions.
In Structural Optimization, evolutionary algorithms are used in many different forms.
Commonly they are divided into four categories with respect to their historical background.
They are:

• Genetic Algorithms (GAs) developed by John Holland [55]. The basic terminology
of genetic search and its principal components are discussed by Goldberg [56]. An
introduction to the application of Genetic Algorithms to Structural Optimization
using traditional binary string coding is given by Hajela [57].

• Evolutionary Programming (EP) created by Lawrence Fogel [58] and developed fur-
ther by his son David Fogel [59].

• Evolution Strategies (ES) created by Ingo Rechenberg [60] and today strongly pro-
moted by Thomas Bäck [61].

• Genetic Programming (GP) is the most recent development in the field by John
Koza [62]

However, they are all based on the same evolutionary principles. Therefore we will use
a more modern terminology also used by Bentley [63] and Schoenauer [64] and generally
speak just of Evolutionary Algorithms (EAs). All the above listed strategies can be seen
as specializations of general EAs which we will describe below.

Architecture of Evolutionary Algorithms

Based on biology and the theory of Universal Darwinism, an object to evolve must meet
several criteria which are:

• Reproduction

• Inheritance

• Variation

• Selection

They perform the reproduction of individuals, either directly cloning parents or by using
recombination and mutation operators to allow inheritance with variation. These opera-
tors may perform many different tasks, from a simple random mutation to a complete local
search algorithm. All EAs also use some form of selection to determine which solutions will
have the opportunity to reproduce, and which will not. The key thing to remember about
selection is that it exerts selection pressure, or evolutionary pressure to guide the evolu-
tionary process towards specific areas of the search space. To do this, certain individuals
must be allocated a greater probability of having offspring compared to other individuals.
As will be shown below, selection does not only mean parent selection - it can also be
performed using fertility, replacement, or even death operators. It is also quite common
for multiple evolutionary pressures to be exerted towards more than one objective in a
single EA.
But unlike natural evolution, EAs further require three other important features:

• Initialization

c ETH Zürich IDMF-CMAS, March 1, 2018

128 Stochastic Search

• Evaluation

• Termination

Because we are not prepared to wait for the computer to evolve for several million of
generations, EAs are typically given a head start by initializing (or seeding) them with
solutions that have fixed structures and meanings, but random values. This means we
feed a certain amount of knowledge already at the beginning of the evolution into the
algorithm. Evaluation in EAs is responsible for guiding evolution towards better solutions.
Unlike natural evolution, evolutionary algorithms do not have a real environment in which
the survivability or goodness of its solutions can be tested, they must instead rely on
simulation, analysis and calculation to evaluate solutions.
Extinction is the only guaranteed way to terminate natural evolution. This is obviously
a highly unsuitable way to halt EAs, for all the evolved solutions will be lost. Instead,
explicit termination criteria are employed to halt evolution, typically when a good solution
has evolved or when a predefined number of generations have passed. In practice even more
often it happens that just the programmers patience is exceeded and he manually stops
the evolution.
There is one more important processes, which, although not necessary to trigger or
control evolution, will improve the capabilities of evolution: Mapping. Even though not
always necessary we typically separate between the search space (genotype space) and
the phenotype space. The search space is a space of coded solutions (genotypes) to the
problem, and the solution space is the space of actual solutions (phenotype). The genotype
can be understood as a recipe of how to build an actual solution. E.g., it encodes attributes
of a geometry model. Coded solutions must be mapped onto actual solutions, before the
fitness of each solution can be evaluated. The process of decoding the recipe and building
the actual solution is called mapping.
Figure 7.3 shows the general architecture of EAs. This architecture should be regarded
as a general framework for EAs, not an algorithm itself. Indeed, most EAs use only a subset
of the stages listed. In the following we will now briefly examine each of these possible
stages of EAs.

Initialization EAs typically seed the initial population with entirely random values.
Evolution is then used to discover which of the randomly sampled areas of the search
space contain better solutions, and then to converge upon that area. Sometimes the entire
population is constructed from random mutants of a single user-supplied solution. Often
random values are generated inside specified ranges (a form of constraint handling). It is
not uncommon for explicit constraint handling to be performed during initialization, by
deleting any solutions which do not satisfy the constraints and subsequently creating new
ones. More complex problems often demand alternative methods of initialization. Some
researches provide the EA with embryos - simplified non-random solutions which are then
used as starting points for the evolution of more complex solutions. Some algorithms
actually attempt to evolve representations or low-level building blocks first, then use the
results to initialize another EA which will evolve complex designs using these representa-
tions or building blocks. Although most algorithms do use solutions with fixed structures
(i.e. a fixed number of decision variables), some allow the evolution of the number and
organization of parameters in addition to parameter values. In other words, some evolve
structure as well as details. For such algorithms, initialization will typically involve the
seeding of solutions with both random values and random structures. paragraphMap Since

c ETH Zürich IDMF-CMAS, March 1, 2018

7.2 Stochastic Search Algorithms 129

Figure 7.3: General architecture of an evolutionary algorithm.

typically EAs distinguish between search space (genotype) and solution space (phenotype)
they require a mapping stage to convert genotypes into phenotypes. This process, known
by biologists as embryogeny, is highly important for the following reasons:

• Reduction of search space. Embryogeny permits highly compact genotypes to de-

fine phenotypes. This reduction (often recursive, hierarchical and multi-functional)
results in genotypes with fewer parameters than their corresponding phenotypes,
causing a reduction in the dimensionality of the search space, hence a smaller search
space for the EA.

• Better enumeration of search space. Mapping permits two very differently

organized spaces to coexist, i.e. a search space designed to be easily searched allows
the EA to locate corresponding solutions within a hard-to-search solution space.

• More complex solutions in solution space. By using growing instructions

within genotypes to define how phenotypes should be generated, a genotype can
define highly complex phenotypes.

• Improved constraint handling. Mapping can ensure that all phenotypes always
satisfy all constraints, without reducing the effectiveness of the search process in any
way, by mapping every genotype onto a legal phenotype.

Especially in the field of Structural Optimization this mapping is a very crucial point for
the performance of the whole optimization. Therefore we will have a closer look at it in
Section 7.3.

c ETH Zürich IDMF-CMAS, March 1, 2018

130 Stochastic Search

Evaluation Every new phenotype must be evaluated to provide a level of goodness for
each solution. Often a single run of an EA will involve thousands of evaluations, which
means that almost all computation time is spent performing the evaluation process. In
Structural Optimization, evaluation is often performed by dedicated analysis software
(CAD-software, FEM-tools etc.) which can take minutes or even hours to evaluate a
single solution. Therefore often a strong emphasis exists toward reducing the number
of evaluations during evolution. Sometimes one even knows at the beginning how many
evaluations can be afforded and therefore the EA should be designed to make the best out
of it.
Evaluation involves the use of fitness functions to assign fitness scores to solutions.
These fitness functions can have single or multiple objectives, they can be unimodal or
multi-modal, continuous or discontinuous, smooth or noisy, static or continuously chang-
ing. EAs are known to be proficient at finding good solutions for all these types of fitness
functions. Nevertheless, the implementation of the fitness function often has tremendous
influence on the performance of the EA.

Mating Selection Parent solutions are always required in an EA, otherwise no child
solutions can be generated. However, their preferential selection of some parents instead
of others is not essential for the EA to work. Evolution will still occur without it, as long
as evolutionary pressure is exerted by one of the the other selection methods: fertility,
replacement and death. Nevertheless, most forms of EAs perform parent selection.
Choosing the fitter solutions to be parents of the next generation is the most common
and direct way of inducing a selective pressure towards the evolution of fitter solutions.
Typically, one of three selection methods are utilized: fitness ranking, tournament selection
or fitness proportionate selection. Fitness ranking sorts the population in order of the
fitness values and bases the probability of a solution being selected for parenthood on its
position in the ranking. Tournament selection bases the probability of a solution being
selected on how many other randomly picked individuals it can beat. Fitness proportionate
selection (or roulette wheel selection) bases the probability of a parent being selected on
the relative fitness score of each individual, e.g. a solution ten times as fit as another is ten
times more likely to be picked as parent (Goldberg [56]). This method also incorporates
fertility selection (see below).
Although normally fitter parents are selected, this does not have to be the case. It is
possible to select parents based on how many constraints they satisfy, or how well they
fulfill other criteria, as long as a fitness-based selecting pressure is introduced elsewhere in
the algorithm. In algorithms that record the age of individuals, parent selection may be
limited to individuals that are mature or individuals which are below their maximum life
spans.

Reproduction Reproduction is the cornerstone of every evolutionary algorithm - it is

the stage responsible for the generation of child solutions from parent solutions. Crucially,
child solutions must inherit characteristics from their parents, and there must be some
variability between the child and parent. This is achieved by the use of the genetic oper-
ators: recombination and mutation.
Recombination operators require two or more parent solutions. The solutions (or geno-
types) are shuffled together to generate child solutions. EAs normally use recombination
to generate most or all offspring. Recombination is normally performed by crossover op-
erators in EAs.
Mutation operators modify a single solution at a time. Some EAs mutate a copy of a

c ETH Zürich IDMF-CMAS, March 1, 2018

7.2 Stochastic Search Algorithms 131

parent solution in order to generate the child, some mutate the solution during the appli-
cation of the recombination operators, others use recombination to generate children and
then mutate these children. In addition, the probability of mutation varies depending on
the EA.
There are huge numbers of different mutation operators in use today. Examples include:
bit-mutation, translocation, segregation, inversion, structure mutation, permutation, edit-
ing, encapsulation, mutation directed by strategy parameters, and even mutation using
local search algorithms (see [56, 62, 65, 61]).
An important feature of both recombination and mutation is non-disruption. Although
variation between parent and child is essential, this variation should not produce excessive
changes to phenotypes. In other words, child solutions should always be near to their
parent solutions in the solution space. If this is not the case, i.e. if huge changes are per-
mitted, then the semblance of inheritance from parent to child solutions will be reduced,
and their position in the solution space will become excessively randomized. Evolution
relies on inheritance to ensure the preservation of useful characteristics of parent solution
in child solutions. When disruption is too high, evolution becomes no more than a random
search algorithm.

Environmental Selection Once offspring have been created, they must be inserted
into the population. EAs usually maintain populations of fixed sizes, hence for every new
individual that is inserted into the population, an existing individual must be deleted.
Therefore, this stage is also called replacement. The simpler EAs just delete every individ-
ual and replace them with new offspring. However, some EAs use an explicit replacement
operator to determine which solutions are replaced by the offspring. Replacement is often
fitness-based, i.e. children always replace solutions less fit than themselves, or the weakest
in the population are replaced by fitter offspring.
Replacement is clearly a third method of introducing evolutionary pressure to EAs, but
instead of being a selection method, it is a negative selection method. In other words,
instead of choosing which individuals should reproduce or how many offspring they should
have, replacement chooses which individuals will die.
Replacement needs not be fitness-based, it can be based on constraint satisfaction, the
similarity of genotypes, the age of solutions, or any other criterion, as long as a fitness-
based evolutionary pressure is exerted elsewhere in the EA. Replacement is also limited by
speciation within EAs: a child from two parents of one species/population/island should
not replace an individual in a different species/population/island.

Termination Evolution by an EA is halted by termination criteria, which are normally

based on solution quality and time. Most EAs use quality-driven termination as the
primary halting mechanism: they simply continue evolving until an individual which is
considered sufficiently fit has been evolved. Some EAs will also re-initialize and restart
evolution if no solutions have attained a specific level or fitness after a certain number of
generations.
For algorithms which use computationally heavy fitness functions, or for algorithms which
must generate solutions quickly, the primary termination criterion is based on time. Nor-
mally evolution is terminated after a specific number of generations, evaluations, or sec-
onds. In order to reduce the number of unnecessary generations, some algorithms measure
the convergence rates during evolution, and terminate when convergence has occurred (i.e.
when the genotypes, phenotypes or fitness values of all individuals are static for a number
of generations). Many EAs also permit the user to halt evolution.

c ETH Zürich IDMF-CMAS, March 1, 2018

132 Stochastic Search

c ETH Zürich IDMF-CMAS, March 1, 2018

7.3 Representation Concepts 133

7.3 Representation Concepts

7.3.1 The universal genotype
The universal genotype consists of a collection of different gene types, where each type
represents a common parameter type.
Therefore any genotype that is representable by an arbitrary collection of the available
gene types can be realized by just composing a heterogeneous list of the appropriate genes.
Up to date, the following gene types are available:

Float-gene represents an arbitrary floating-point parameter.

Upper and/or lower limits can be provided. If the parameter is unbounded, an
additional step size  for uniform mutation must be specified.

Integer-gene represents an integer parameter.

Again upper and/or lower limits, as well as a step size, can be provided.

Bool-gene represents a binary parameter that can be true or false.

Float-list-gene is a list of arbitrary floating point values. The parameter always has to
represent one of the values specified in the definition of the gene.

Const-float-list-gene are equally distributed floating-point values, i.e. they have a con-
stant distance between two neighboring values.
This gene is quite similar to the integer gene. Optionally, limits or a mutation step
size can be provided.

String-list-gene is a list of arbitrary discrete values upon which no norm or ordering can
be applied.
This is in contrast to the integer-gene, the float-list-gene, and the const-float-list-
gene.

For the definition of the listed gene types, the standard deviation σ used for Gaussian
mutation must be specified additionally for every gene.
Float-gene, integer-gene, float-list-gene, and const-float-list-gene can further be pro-
vided with so called cyclic properties.
This is suitable when between two possible gene values a distance but no absolute
order can be defined, as e.g. for angle values.

Examples of different gene types

To clarify how an universal genotype for a given optimization problem must be defined,
examples for the different gene types are explained in the following, as they are read from
file during initialization of an actual EA:

• [float gene 1.2 true 0.5 false 0.0 3.0 5.0 false] denotes a float gene with a default
value of 1.2, with a lower limit 0.5, and with no upper limit defined (the booleans
before the limit values specify, whether a limit exists or not).
Further the mutation parameters  = 3.0 and σ = 5.0 are specified.
 denotes the range for uniform mutation of an unbounded gene, and σ defines the
standard deviation to be used for Gaussian mutation.

c ETH Zürich IDMF-CMAS, March 1, 2018

134 Stochastic Search

The last boolean value indicates that this example gene does not have cyclic prop-
erties.

• [int gene 3 true 0 true 360 0 10.0 true] denotes an int-gene with a default value
of 3, with a lower limit 0, and an upper limit of 360.
Further, mutation parameters  = 0 (useless for this fully bounded gene), and σ =
10.0 are specified.
In addition, this gene has cyclic properties, indicated by the last boolean value.

• [bool gene true 2.0] denotes a bool-gene with default value true.
A standard deviation σ = 2.0 for some sort of Gaussian mutation is also defined.

• [float list gene 4 3 0.0 1.0 false 1.2 2.5 4.5 4.9] denotes a float-list-gene with 4
values, where the value with index 3 (starting from 0) is the actual default value.
Mutation parameters  = 0.0 (useless for this gene) and σ = 1.0 follow.
Then it is defined that this gene has no cyclic properties, and the four possible float
values are appended.

• [const float list gene 1.75 1.0 true 3.0 false 8 0.5 1.2 false] denotes a const-
float-list-gene with a default value of 1.75.
The gene has an active lower limit of 1 and an inactive (indicated through the boolean
after the limit value) upper limit of 3.
By specifying the number of intervals to be 8, the possible values for this genes are
[1.0, 1.25, 1.5, ...].
Finally,  = 0.5, σ = 1.2 and no cyclic properties are given.

• [string gene 3 1 1.0 blue green red ] denotes a string-gene with 3 possible values
and where the value with index 1 (again starting from 0) is chosen to be the actual
default value. Further a σ = 1.0 value is specified, and the different possible values
are appended.

During the initialization of the Evolutionary Algorithm, a list containing such genes is
read from file.
This list defines the genotype structure for a problem at hand.
Further, the given default values can also be used to incorporate existing solutions for
knowledge-based initialization.

c ETH Zürich IDMF-CMAS, March 1, 2018

Chapter 8

Composite Structures

This chapter discusses the structural optimization of laminated composite materials with
focus on fiber reinforced materials. The thickness of these structures is usually thin com-
pared to the other dimensions which is why they are typically represented in FEM simula-
tions with layered shell finite elements. The first section gives a short introduction on the
design of laminated composites including a selection of applications. Consequently, the
Classical Lamination Theory (CLT) is introduced which is the standard calculus for lam-
inated composites. Afterward, the basics of a shell finite element are explained. Together
with the CLT, the element is enhanced to a layered shell element which can be used for
the computational representation of laminates. The majority of these fundamentals are
based on the doctoral thesis of B. Schläpfer [66]. The second part of this chapter discusses
different disciplines in laminate optimization including fiber orientation, laminate thick-
ness, stacking sequence and material optimization. Finally, the theories are specialized to
optimization techniques for finding locally varying laminates which is also called laminate
tailoring. Additionally, a selection of laminate tailoring methods is presented.

8.1 Design of Fiber Reinforced Composites

Fiber reinforced composite materials have superior mass-specific mechanical properties
which is why they are primarily used for light-weight applications. In contrast to metallic
materials such as aluminum or steel, they are usually strongly anisotropic. Together with
the layer-wise building-technique, the connection between the design parameters and the
resulting structural behavior becomes very complex and can barely be understood intu-
itively. Consequently, employing computer-based engineering methods becomes important.
Common optimization techniques are used for laminate design even if the objective is not
to find the globally best solution but designs with significantly increased properties. This
is why this process can basically be understood as automated design process. The thick-
ness of light-weight structures is typically thin compared to the other dimensions which
is why they have shell characteristics. From the structural engineering point of view, a
shell is a plane structure with curved inner and outer surfaces which are separated by the
thickness t which is smaller compared to the other dimensions of the shell. The thickness
can be constant or may vary and the mid-surface is defined to be the distance t/2 from
both surfaces. Since the thickness of shell structures is parameterized with a scalar value
t, it can be adapted without changing the representation of the spatial model. This is an
important benefit in terms of computational cost which is a decisive factor in preliminary
design. The optimization and the automated design of fiber reinforced plastics needs a
deeper understanding of the lamination theory and layered shell elements which are pre-

c ETH Zürich IDMF-CMAS, March 1, 2018

136 Composite Structures

sented within the next sections. A selection of applications with shell characteristics are

(a) Fuselage section of the Air- (b) Payload fairing of the (c) CFRP-monocoque of a
bus A350 Ariane 5 launcher Lamborghini Aventador

Figure 8.1: Examples of applications with shell structures

presented in Figure 8.1. Figure 8.1(a) shows a fuselage section of the Airbus 350 which
is entirely made of Carbon Fiber Reinforced Plastics (CFRP). Considering light-weight
structures or especially aircraft structures, the thickness of almost any components is
usually much smaller compared to the other dimensions wherefore they are prevailingly
modeled with shells. Another typical shell application is the payload fairing of the Ariane 5
launcher shown in Figure 8.1(b). Its mission is to protect the payload during the launch
and it should of course be as light as possible in order to maximize the payload capacity of
the launcher. Additionally, it must fulfill acoustic and dynamic requirements. The third
example shown in Figure 8.1(c) is a CFRP-monocoque of a Lamborghini. With increasing
fuel costs, weight saving has become an important issue in automotive engineering and
components are increasingly often made of fiber reinforced plastics.

c ETH Zürich IDMF-CMAS, March 1, 2018

8.2 Laminated Composites 137

8.2 Laminated Composites

8.2.1 Introduction
Composite materials are a combination of at least two different sub-materials. The com-
bination aims for a resulting material having superior properties by taking advantage of
the properties of its sub-materials. On a macroscopic scale, the composite material can
be considered as homogeneous which simplifies the design and analysis significantly. The
composite may have material properties that cannot be achieved by the single compo-
nents. Since not all properties can be improved simultaneously, the composites have to be
designed considering the specific requirements of an application.
The most important class of composites are the Fiber Reinforced Plastics (FRP) which
consist of fiber material embedded in matrix material. The superior material properties of
FRP are mainly arising from the fiber material. Typical fiber materials are carbon, glass,
boron, or aramid. The benefit of fiber materials is the low number of voids due to their
small diameters which restrains crack growing and therefore increases the strength [67].
Moreover, the mentioned fiber materials have low mass densities compared to common
metallic materials (e.g. carbon ∼1.8 g/cm3 ) which results in outstanding mass-specific
mechanical properties. However, the fibers themselves are not able to build a continuous
material wherefore they are embedded in a matrix system (see Figure 8.2(a)) which are
preferably duroplastic or thermoplastic materials. Widely spread matrix materials are
epoxy resins which are easy to process and have acceptable mechanical properties with
low specific mass ( ∼1.2 g/cm3 ). In order to simplify the handling in the manufacturing
process, FRP are usually prefabricated. Fiber rovings are joined to laminar plies, the so-
called laminae. The fibers are processed to textiles with two or multiple fiber directions or
to unidirectional plies with a single principal fiber orientation. In so-called PREPREGs,
the fiber plies are already pre-impregnated with the polymer resin. This ensures an ideal
mixing ratio of the components and a simple handling of the fabrics which results in a
high quality of the final laminated composite. The mechanical properties of such FRP plies

(a) Matrix-embedded unidirectional (b) Laminated composite

fibers

Figure 8.2: Laminated composites

are highly anisotropic. While the mass-specific properties in fiber direction may be one
magnitude higher than for metallic materials, the material properties perpendicular to the
fiber orientation, which are mainly governed by the properties of the matrix material, are
comparatively low. Therefore, structures are built by stacking several plies with different
orientation angles which are then called laminated composites (see Figure 8.2(b)). Basi-
cally, the orientation angles can be chosen in a way so that the overall mechanical behavior
becomes nearly isotropic (also called quasi-isotropic). This is achieved by distributing the
fiber orientations regularly in all directions. A big advantage of laminated composites

c ETH Zürich IDMF-CMAS, March 1, 2018

138 Composite Structures

is the possibility to specifically design their structural response to different loadings and
requirements. In contrast to isotropic materials for which the amount of material, in
particular the thickness, is the only design parameter, the behavior of composites can be
designed by varying the orientation angles, the number of plies, the stacking sequence or
in special cases the fiber volume content. Material can be added easily to highly loaded
regions and the fibers can be oriented dependent on the principal directions of the loads.
Furthermore, the layered design enables to build laminates of different ply materials. Due
to the behavior of anisotropic materials, the layered design method and the resulting large
number of design variables, the design process of laminated composites becomes complex
and time consuming. The experience and the intuition of structural engineers may be
pushed to the limit and, excepting structures with a low degree of complexity, e.g. plates
or cylinders, finding solutions with good light-weight properties is hard to be accomplished
manually. The application of computer-based design methods is indispensable aiming for
high quality designs in a reasonable time frame.

8.2.2 Classical Laminate Theory

This section gives a short introduction to the Classical Laminate Theory (CLT) which is
the standard theory for the analysis and the design of thin laminated composites. There
exist different conventions in literature which may vary slightly. The chosen conventions
for this thesis are based on the textbook of Jones [25]. The CLT bases on the assumptions
of the Kirchhoff-Love theory, namely that:

• the structure be thin compared to the other dimensions and a constant thickness.

• the Bernoulli-theory be valid, namely plane sections and no transverse shear strains.

• the stress state be plane (σz = τxz = τyz = 0).

• material behavior be linear-elastic.

• deformations be small.

Supplementary to the assumptions for homogeneous materials, it is assumed that the single
layers are bonded perfectly. Due to the stacking of several layers, the material stiffness
properties of a laminate through the thickness are inhomogeneous. In order to provide
a linear relation between plate deformations to plate line loads of the global laminate,
the CLT performs a stiffness homogenization. This requires that the layer stiffnesses,
which are usually given in material principal coordinates 1,2, be transformed to the global
coordinates x, y of the problem.

Plane-Stress State and Stiffness Transformation

The stiffness matrix of an orthotropic layer is defined as
 
C11 C12 C13 0 0 0
C12 C22 C23 0 0 0 
 
C13 C23 C33 0 0 0 
Cortho =
  (8.1)
 0 0 0 C44 0 0 
 0 0 0 0 C55 0 
0 0 0 0 0 C66

c ETH Zürich IDMF-CMAS, March 1, 2018

8.2 Laminated Composites 139

In order to express the stiffness entries as functions of the engineering constants such as
the Young’s moduli, the Poisson ratios and the shear moduli, it is easier to formulate
the so-called compliance matrix S which is the inverse of the stiffness matrix C. The
compliance matrix of an orthotropic material is defined as

1
− Eν21 − Eν31
 
E11 22 33
0 0 0
− ν12 1
− Eν32 0 0 0 
 E11 E22 33 
− ν13 − Eν23 1
0 0 0 
 E11 E33
Sortho =  22 (8.2)

1
 0 0 0 0 0

G23 

 0 1 
0 0 0 G31 0 
1
0 0 0 0 0 G12

Measured engineering constants can be inserted directly and the stiffness matrix results
of the inversion. Considering a unidirectional reinforced composite material, the material
properties are transversely-isotropic. Such a material is defined with 5 elastic coefficients
which are usually the Young’s modulus in fiber direction E11 , the Young’s modulus per-
pendicular to the fibers E22 , the in-plane shear modulus G12 , the in-plane Poissons ratio
ν12 and the transverse shear modulus G23 . The transversal shear Poisson ratio is related
to the variables above with
E22
ν23 = −1 (8.3)
2G23
Due to the assumption of a plane stress state, the stresses with out-of-plane components
are zero and the compliance matrix S can be reduced to a 3×3-matrix. Consequently, the
stiffness matrix C is reduced to the so-called reduced stiffness matrix Q and the Hooke’s
law is simplified to
    
σ1 Q11 Q12 0 ε1
 σ2  = Q12 Q22 0   ε2  (8.4)
τ12 0 0 Q66 γ12

Keep in mind that the strains having out-of-plane components, namely ε3 , γ13 and γ23 ,
are not zero, they are just not considered. Strains and stresses are expressed in material
principal coordinates 1,2. However, the orientation angle of a unidirectional laminate
layer can be chosen arbitrarily wherefore the material principal directions do not coincide
with the global coordinates of the problem formulation. Consider a laminate illustrated
in Figure 8.3 where the material coordinates 1,2 are rotated with respect to the global
coordinates x,y by a given angle ϕ. A relation which transforms the stresses and the

2 y

φ
x

Figure 8.3: Material principal and global coordinate systems

strains from the material coordinate system to the global coordinate system is needed.

c ETH Zürich IDMF-CMAS, March 1, 2018

140 Composite Structures

This is achieved by applying the rotation matrix T.

cos2 ϕ sin2 ϕ
 
2 sin ϕ cos ϕ
T =  sin2 ϕ cos2 ϕ −2 sin ϕ cos ϕ  (8.5)
− sin ϕ cos ϕ sin ϕ cos ϕ cos2 ϕ − sin2 ϕ
This rotation matrix is different from the common transformation matrix for a vector
rotation by a coordinate axis. This is because the stresses and the strains are 2nd -order
tensors (not vectors) and they are only arranged in arrays. While the material stresses are
mapped directly using the matrix T according to equation (8.6),
   
σ1 σx
 σ2  = T  σy  (8.6)
τ12 τxy
it has to be taken into account that the engineering shear strains γij are 2 times the
tensorial shear strains
γij = 2εij (8.7)
which leads to the following relation.
   
ε1 εx
 ε2  = T  εy  (8.8)
1 1
2 γ12 2 γxy

In order to be able to work directly with the engineering strains without a pre-factor,
Reuter [68] introduced the simple matrix R.
 
1 0 0
R = 0 1 0 (8.9)
0 0 2
An application of this Reuter matrix reduces the potential for mistakes to a minimum
since it is clear that only engineering stains are utilized. The engineering strains are then
transformed with equation (8.10).

εxy = RTR−1 ε12 (8.10)

Contrariwise, the global strains and stresses can be mapped to the material principal
strains and stresses performing a multiplication with the inverse rotation matrix T−1 .
The connection between the strains and stresses in global coordinates is based on the
transformed reduced stiffness matrix Q.
   
σx εx
 σy  = Q  εy  (8.11)
τxy γxy

The transformed reduced stiffness matrix Q can be derived by using equations (8.4), (8.6)
and (8.10) and is therefore a function of Q, R and T only.

Q = T−1 QRTR−1 = T−1 QT−T (8.12)

The transformed reduced stiffness matrix Q has to be evaluated for every laminate layer
in order to be able to homogenize the material data of the entire laminate stack. However,
the evaluation is straight forward and the computational costs are low.

c ETH Zürich IDMF-CMAS, March 1, 2018

8.2 Laminated Composites 141

Stiffness Homogenization
In contrast to a homogeneous material, the material properties of a laminate are alter-
nating through the thickness. This complicates the analysis but also the formulation of
a finite element, since the integration through the thickness becomes more complex. The
stiffness discontinuity due to the layer-wise design technique results in a discontinuous
stress distribution. To quantify a stress state in a laminate, the components have to be
evaluated for each layer particularly. In order to have a load unit that includes all layers,
the line loads are introduced, namely the force per unit length N and the moment per
unit length M. The force per unit length N is the integration of the stress components
over the laminate thickness t.

t
zk
zk-1 hk
hk-1 t
zj 2
hj
z1 h2
z0 h1

Figure 8.4: General layup of a laminated composite material

   
Nx Z t σx
2
N =  Ny  =  σy  dz (8.13)
t
Nxy − 2 σxy
Analogously, the moment per unit length M is given by the integration of the stress
components multiplied with the stacking position z.
   
Mx Z t σx
2
M =  My  =  σy  zdz (8.14)
t
Mxy − 2 σxy
Since the stiffness is constant within a single layer, the integration can be replaced with
a summation of the integrals over each layer taking advantage of the transformed reduced
stiffness matrix Q derived above. Using the kinematic relations saying that the strains are
composed of the membrane strains and a part arising from the curvature (see equation
(A.27))
ε = ε0 + zκ (8.15)
equations (8.13) and (8.14) yield
Z t n Z zj n Z zj
2 X X
ε0 + zκ dz


N= σdz = Qj εdz = Qj (8.16)
− 2t j=1 zj−1 j=1 zj−1
n  
X 1 2
0 2


= Qj (zj − zj−1 ) ε + zj − zj−1 κ (8.17)
2
j=1

Z t n Z zj n Z zj
2 X X
ε0 + zκ zdz


M= σzdz = Qj εzdz = Qj (8.18)
− 2t j=1 zj−1 j=1 zj−1
n  
X 1 2 2

0 1 3 3


= Qj zj − zj−1 ε + zj − zj−1 κ (8.19)
2 3
j=1

c ETH Zürich IDMF-CMAS, March 1, 2018

142 Composite Structures

Considering equations (8.17) and (8.19), the matrices A, B and D can be extracted which
are all of dimension 3×3. The A-matrix connects the membrane strains ε0 with the force
per unit length N.
Xn
A= Qj (zj − zj−1 ) (8.20)
j=1

and the D-matrix connects the plate curvatures κ with the moments per unit length M.
n
1X
Qj zj3 − zj−1
3


D= (8.21)
3
j=1

The B-matrix is responsible for the coupling of membrane and bending components.
n
1X
Qj zj2 − zj−1
2


B= (8.22)
2
j=1

Dependent on the laminate layup, some entries may become zero. In case of a symmet-
ric laminate, there is no coupling between bending and membrane effects wherefore the
B-matrix vanishes. These matrices build the so-called ABD-matrix which is the main
achievement of the homogenization process.

A B ε0
    
N
= (8.23)
M B D κ

ε0x
    
Nx A11 A12 A16 B11 B12 B16
 Ny   A12 A22 A26 B12 B22 B26  ε0y 
ε0xy
    
 Nxy  
= A16 A26 A66 B16 B26 B66  
   (8.24)

 Mx  
  B11 B12 B16 D11 D12 D16 
 κx 

 My   B12 B22 B26 D12 D22 D26  κy 
Mxy B16 B26 B66 D16 D26 D66 κxy

c ETH Zürich IDMF-CMAS, March 1, 2018

8.3 Finite Element Representation 143

8.3 Finite Element Representation

The application of shell elements requires that the thickness of the represented structures
is much smaller compared to its other dimensions. Thin structures are usually not modeled
with 3D elements, especially if analyzing plate bending. If 3D elements are modeled thin
in only the thickness direction according to Figure 8.5(a), there may be the problem
of shear locking and ill-conditioning [69]. This problem can be avoided by using many
elements according to Figure 8.5(b). However, the number of degrees-of-freedom (d.o.f.) is
increased drastically and the models become large wherefore solving the problems becomes
computationally expensive. Using shell elements and the underlying plate theory, the
problem of having a large number of d.o.f. is mitigated (see Figure 8.5(c)).

(a) 3D-modeling with large aspect ratios may lead to shear lock-
ing and ill-conditioning

(b) 3D modeling with smaller mesh leads to an extensive number

if d.o.f.

(c) 2D-modeling with plate elements mitigates both problems

Figure 8.5: Front view of different modeling techniques for bending

8.3.1 Layered Shell Elements

Stiffness Matrix

A detailed derivation of a general shell element is given in Appendix A.2. Based on

these results and the Classical Lamination Theory introduced above, a layered shell el-
ement can be formulated. In contrast to the general formulation, the material here is
not homogeneous through the thickness which is illustrated in Figure 8.6. Consequently,
the integration of the stiffness parts through the thickness (see equations (A.34), (A.37)
and (A.39)) have to be replaced with a summation according to the formulation of the
ABD-matrix. The stiffness matrix of a layered shell element is given by

k = km + k c + kb + ks (8.25)

c ETH Zürich IDMF-CMAS, March 1, 2018

144 Composite Structures

with
n Z
X 
km = BTm Qj Bm dA (zj − zj−1 ) (8.26)
j=1 A
n Z 
X 1 2
BTm Qj Bb BTb Qj Bm 2



kc = + dA zj − zj−1 (8.27)
A 2
j=1
n Z 
X 1 3
BTb Qj Bb dA 3


kb = z − zj−1 (8.28)
A 3 j
j=1
n Z 
X S
ks = BTs κQj Bs dA (zj − zj−1 ) (8.29)
j=1 A

Consider that the matrices Bm , Bb and Bs are strain-displacement matrices (see Appendix
A.2) and should not be confused with the B-matrix of the lamination theory. In contrast to
homogeneous materials, an additional term from the coupling stiffness matrix kc appears
which becomes zero again for symmetric laminates. However, no additional components
are need since it is a combination of membrane and bending parts.

Mass Matrix
The mass matrix formulation for a layered shell element is simple when using a lumped
mass model where rotary inertia is neglected. It is feasible for small deformations which is
usually true for harmonic vibration problems. The lumped mass matrix is expressed with
equation
Xn  Z 
T
m= ρj tj φ φ dA (8.30)
j=1 A

The shape functions φ map the mass element to its nodes.

8.3.2 Laminate Sensitivities

The expression sensitivity origins from the so-called sensitivity analysis which is a common
methodology in the field of optimization with mathematical programming. The sensitiv-
ities express the influence of a change of the design variables x to the objective function
f . From the mathematical point of view, the sensitivities are the gradients of an objective
function. The general expression is therefore

df f (x + ∆x) − f (x)
∇f = = lim (8.31)
dx ∆x→0 ∆x
A simple method for the numerical determination of the sensitivities is the finite difference
approximation. The forward finite difference approximation is defined as

df f (x1 , ..., xi + ∆xi , ..., xn ) − f (x1 , ..., xi , ..., xn )

= (8.32)
dxi ∆xi
For an objective function f with n design variables, n + 1 function evaluations are needed
to fully evaluate the sensitivities. In almost any problem of structural optimization, a
finite element run is included in the evaluation of the function. If the size of the finite
element model is large or if many design variables are used, the numerical determination

c ETH Zürich IDMF-CMAS, March 1, 2018

8.3 Finite Element Representation 145

of the sensitivities may become too expensive. Additionally, the numerical determination
includes a numerical error which is caused by the finite value of ∆xi . However, a numerical
determination of the sensitivities is the only possibility if the objective function is not
known explicitly, for example in black box optimizations.
Considering for instance the common sensitivity equations for minimal compliance
∂W
= −uT ∂K T ∂r W = uT r

∂x ∂x u + 2u ∂x (8.33)

for displacements which has already been derived in Section 6.10.2 (see equation (6.133))
∂u
= K−1
 ∂r ∂K

∂x ∂x − ∂x {u = Kr} (8.34)

or eigenfrequencies
∂K ∂M
∂λn ΦT
n ( ∂x −λn ∂x )Φn
∂x = T
Φn MΦn
{(K − λn M) Φn = 0}
(8.35)
it can be noticed that they both require the sensitivities of the stiffness matrices. Since the
formulation of the finite elements, and therefore the connections between design variables
and the stiffness matrix are known, the sensitivities can be derived analytically which is
done here for the thickness of the layers and the orientation angles.

Stiffness Matrix Sensitivities With Respect To Layer Thicknesses

The general formulation of the stiffness matrix of a layered shell element is given by
equations (8.26) to (8.25). The formulation is dependent on the positions of the laminate
interfaces zj . To differentiate the stiffness matrix with respect to the layer thicknesses hj ,
they have to be reformulated. To keep things simple, the sensitivities of the stiffness matrix
with respect to layer thickness is here only derived for symmetric laminates. However,
general formulations are given in [66]. For symmetric laminates, the integration may
be done for one half of the laminate only if the value of the stiffness matrix is doubled
simultaneously. Figure 8.7 illustrates one half of the symmetric laminate where the layers
are numbered here from the laminate mid-plane. Consequently, the interface coordinates
zj may be replaced by the thickness values with

j
X
zj = hk (8.36)
k=1

The finite element stiffness matrices for membrane, bending and transverse shear conse-
quently yield

n/2 Z j j−1
" !#
X X X
km = 2 BTm Qj Bm dA hk − hk (8.37)
j=1 A k=1 k=1

n/2 Z

j
!3 j−1
!3  
X 1
 BTb Qj Bb dA 
X X
kb = 2 hk − hk   (8.38)
A 3
j=1 k=1 k=1
n/2 j j−1
"Z !#
X S X X
ks = 2 BTs κQj Bs dA hk − hk (8.39)
j=1 A k=1 k=1

c ETH Zürich IDMF-CMAS, March 1, 2018

146 Composite Structures

Due to the symmetry of the laminate, the coupling stiffness matrix kc is zero. Taking
advantage of relation
Xj j−1
X
hk − hk = hj (8.40)
k=1 k=1

the stiffness for the membrane and shear parts can be simplified to

n/2 Z 
X
km = 2 BTm Qj Bm dA hj (8.41)
j=1 A

n/2 Z 
X S
ks = 2 BTs κQj Bs dA hj (8.42)
j=1 A

Consequently, the sensitivities of these parts are only dependent on the material properties
of the respective layer and yield
Z
dkm
=2 BTm Ql Bm dA (8.43)
dhl A
Z
dks S
=2 BTs κQl Bs dA (8.44)
dhl A

For simplicity, it is assumed that the stiffness correction factor κ is independent on the
thickness hl . The derivative of bending part is more complex because the thickness appears
with a higher order. Considering equation (8.38), it must be distinguished whether the
layer with respect to which the derivative is taken is part of the summation or not. This
can be expressed with
j
! 
∂ X 1 for j ≥ l
hk = (8.45)
∂hl 0 for j < l
k=1

Consequently, the bending stiffness sensitivities yield

Z l
!2
dkb X
=2 BTb Ql Bb dA hk
dhl A k=1

n/2

j
!2 j−1
!2 
X Z X X
+2  BTb Qj Bb dA  hk − hk  (8.46)
j=l+1 A k=1 k=1

Here it becomes obvious that a change of the bending stiffness is caused by two different
effects. The first term expresses the stiffness change due to the thickness change of the
layer itself. The terms within the summation consider the change of the stiffness caused by
pushing outward the overlaying layers which results in a higher area moment of inertia (see
Figure 8.8). This relation can be written alternatively with expressions (8.47) and (8.48).
The top layer (l = n/2) is independent on the subjacent layers wherefore its sensitivity is
expressed with
 2
Z n/2
dkb X
=2 BTb Ql Bb dA  hk  (8.47)
dhl A k=1

c ETH Zürich IDMF-CMAS, March 1, 2018

8.3 Finite Element Representation 147

The lower layers (l = 1, ..., n/2 − 1) take into account the stiffness contribution of the
overlaying layers wherefore
Z Z  X l
!2
dkb dkb
=2 BTb Ql Bb dA − BTb Ql+1 Bb dA hk + (8.48)
dhl A A dhj+1
k=1

The total sensitivities of the stiffness matrix are built with an addressed summation of the
several parts corresponding to equation (8.25).
dk dkm dkb dks
= + + (8.49)
dhl dhl dhl dhl
All the derivations above have been performed on the element level for the element stiffness
k. However, the sensitivity equations (8.33) through (8.35) require the sensitivities of the
global stiffness matrix K. They can however be derived directly from the element stiffness
parts. The global stiffness matrix is an addressed summation of all the element matrices
which is denoted schematically in equation (8.50).
 
 
 

 k1 0 
 
 

 k2 

K= (8.50)
 


 k3 

 
 
 
 
 .. 
 0 . 

The element stiffness matrix derivatives are dependent only on the thickness of the respec-
tive element. The derivatives with respect to the thicknesses of other elements are all zero.
dK
Thus, the global stiffness matrix derivative dh l
contains only zeros except of the entries
corresponding to the considered element l which is shown schematically in equation (8.51).
 
 
 

 0 0 

 
 
 
dK  dkl

= (8.51)
 
dhl  hl 

 
 
 
 
 
 
 0 0 

The sensitivity of the mass matrix with respect to a layer thickness change is simply
Z
dm
= ρl φT φ dA (8.52)
dhl A

c ETH Zürich IDMF-CMAS, March 1, 2018

148 Composite Structures

z,w

A
y,v
t
x,u

Figure 8.6: Layered shell element

zn/2 t
zj 2
z1

Figure 8.7: Schematic illustration of one half of a symmetric laminate

∆t z3
k3(z2,z3) 3
z z2 k3(z2+∆h,z3+∆h)
z
k2(z1,z2) 2 z1 k2(z1+∆h,z2+∆h)
z k1(z0+∆h,z1)
k1(z0,z1) z1 z0
0

Figure 8.8: Effect of a thickness change to the overlaying layers

c ETH Zürich IDMF-CMAS, March 1, 2018

8.3 Finite Element Representation 149

Stiffness Matrix Sensitivities With Respect To Fiber Orientation

The calculation of the sensitivities with respect to the fiber orientation is much simpler
since the stiffness of the surrounding layers is not influenced by changing the fiber orienta-
tion of a particular layer. Taking the derivative with respect of the fiber orientation angles
of the membrane stiffness part given in equation (8.26) simply yields

∂Qj
Z
∂km
= BTm Bm dA (zj − zj−1 ) (8.53)
∂ϕj A ∂ϕj

since only the transformed reduced material stiffness matrix Qj is dependent on the fiber
orientations This is the same for all other stiffness parts given in equations (8.27) through
(8.29). The derivative if the reduced material stiffness matrix starts from equation (8.12)
and yields
∂Q ∂T−1 ∂T −1
= QRTR−1 + T−1 QR R (8.54)
∂ϕj ∂ϕj ∂ϕj
The derivative of the transformation matrix introduced in equation (8.5) is

2 cos2 ϕ − 2 sin2 ϕ
 
−2 sin ϕ cos ϕ 2 sin ϕ cos ϕ
∂T
=  2 sin ϕ cos ϕ −2 sin ϕ cos ϕ −2 cos2 ϕ + 2 sin2 ϕ (8.55)
∂ϕj
− cos2 ϕ + sin2 ϕ cos2 ϕ − sin2 ϕ −4 cos ϕ sin ϕ

The derivatives of the inverse can simply be determined by considering that

T−1 (ϕ) = T (−ϕ) (8.56)

The mass matrix is not sensitive to a change of the fiber orientations wherefore their
sensitivities become zero.

c ETH Zürich IDMF-CMAS, March 1, 2018

150 Composite Structures

c ETH Zürich IDMF-CMAS, March 1, 2018

8.4 Optimization with Lamination Parameter 151

8.4 Optimization with Lamination Parameter

The theory of lamination parameters is a common approach in laminate optimization and
widely spread in literature (e.g. [70, 71, 72, 73, 74]). It basically operates on the ABD-
matrix which is given in equation (8.23). The idea is to find an optimal set of ABD entries
which represent an optimal laminate regarding the considered objective function. Assum-
ing that all layers are of the same material and same thickness and the laminate
thickness is defined to a given value t, the ABD-matrix may be formulated as a sum
of the matrices Γi containing the invariants of a orthotropic material and the so called
lamination parameters ViA,B,D according to equations (8.57) through (8.59).

A = t Γ0 + Γ1 V1A + Γ2 V2A + Γ3 V3A + Γ4 V4A

(8.57)
t2
Γ1 V1B + Γ2 V2B + Γ3 V3B + Γ4 V4B


B= (8.58)
4
t3
Γ0 + Γ1 V1D + Γ2 V2D + Γ3 V3D + Γ4 V4D


D= (8.59)
12
The invariants of an orthotropic material contained in Γi are dependent on the entries of
the stiffness matrix and are explicitly listed in Appendix B. The lamination parameters
are defined as non-dimensional integrals of the orientation angles through the thickness
according to equations (8.60), (8.61) and (8.62).
Z 1
2
V1A , V2A , V3A , V4A

= Φdz (8.60)
− 12
Z 1
 B B B B 2
V1 , V2 , V3 , V4 = 4 zΦdz (8.61)
− 12
Z 1
2
V1D , V2D , V3D , V4D z 2 Φdz

= 12 (8.62)
− 12

with
Φ = {cos 2ϕ, sin 2ϕ, cos 4ϕ, sin 4ϕ} (8.63)
Assuming that the thickness of all layers is unique and taking into account the fact that
the orientation angles ϕ are constant through the thickness of one layer, the integrals can
be replaced with summations according to equations (8.64), (8.65) and (8.66).
n
 A A A A 1X
V1 , V2 , V3 , V4 = [Φi ] (8.64)
n
i=1
n  2  
 B B B B 2 X n 2  n
V1 , V2 , V3 , V4 = 2 −i+1 − −i Φi (8.65)
n 2 2
i=1
n  3  
 D D D D 4 X n 3  n
V1 , V2 , V3 , V4 = 3 −i+1 − −i Φi (8.66)
n 2 2
i=1

with
Φi = {cos 2ϕi , sin 2ϕi , cos 4ϕi , sin 4ϕi } (8.67)
Laminate optimization by means of the lamination parameter theory operates directly on
the lamination parameters ViA,B,D and not on the physical design variables such as for

c ETH Zürich IDMF-CMAS, March 1, 2018

152 Composite Structures

example the orientation angle or layer thickness. Consequently, the number of linearly in-
dependent parameters is 12, for laminates which are symmetric and balanced even only 4.
Naturally, the lamination parameters cannot be chosen arbitrarily since they are bounded
to a feasible domain through the underlying trigonometric functions. A few publications
[75, 76, 77] focus on the description of the feasible lamination parameter domains, but so
far, there exists no generally valid analytical approach. However, if assuming a balanced
and symmetric laminate, the feasible domain of the in-plane lamination parameters can
be found rather easily. The contours of the feasible domain for the in-plane lamination
parameters can be determined, by evaluating equation (8.64) for the border case. An
illustration of the feasible domain for the lamination parameters V1A and V3A is given
in Figure 8.9. The lower boundary is defined by (±ϕ)S -laminates. The upper straight

VA3
(0/903)S 1 (03/90)S
(904)S (04)S
(02/902)S

(0/45/902)S (02/45/90)S
(±752)S (±152)S
(45/903)S (03/45)S
(0/452/90)S V A1
−1 (452/902)S (02/452)S 1

(±302)S
(±602)S (453/90)S (0/453)S

−1 (±452)S

Figure 8.9: Feasible domain of the in-plane lamination parameters V1A, , V3A,

boundary is given by (0j /90n−j )S -laminates. For symmetric and balanced laminates, the
in-plane lamination parameters V2A and V4A are always zero due to the integration over
the thickness. Additionally, all coupling lamination parameters ViB are equal to zero for
the symmetric case. For simplicity, only the +45◦ -layers are labelled in Figure 8.9. Since
the lamination parameters V1A and V3A are based on cosine-functions, each +45◦ -layer can
directly be replaced with a −45◦ -layer
It can be shown that the feasible domain is convex wherefore an optimal solution of the
ABD-matrix can be found efficiently by using algorithms of mathematical programming.
However, the major drawback of using lamination parameters is the fact that the infor-
mation of the ABD-matrix does not explicitly include the information of the physical
laminate. The back-substitution from the ABD-matrix or from the lamination parame-
ters, respectively, to the physical laminate with a known stacking sequence and orientation
angles is not unique and includes a second optimization problem. Considering again the
feasible domain in Figure 8.9, the lamination parameters can take every value in the do-
main. However, the feasible physical laminates are distributed discretely in the design
domain. Having a given number of layers, which is a basic assumption of the theory, the
optimal lamination parameters can only be represented approximately by a physical lam-
inate. In case of a symmetric and balanced laminate as shown above, the complexity of
the back-substitution is acceptable. The problem becomes much more complex for general
laminates.

c ETH Zürich IDMF-CMAS, March 1, 2018

8.4 Optimization with Lamination Parameter 153

8.4.1 Basic Examples

A simple application of lamination parameters is demonstrated on the example of an in-
plane stiffness design of balanced and symmetric laminates. According to the derivations
above, V2A and V4A are zero and V1A and V3A must lie in the feasible domain given in
Figure 8.9. The A-matrix is a linear combination of the lamination parameters and the
material invariants.
       
A11 A12 A16 U1 U4 0 U2 0 0 U3 −U3 0
A12 A22 A26  = U4 U1 0  +V1A  0 −U2 0 +V3A −U3 U3 0 
A16 A26 A66 0 0 U5 0 0 0 0 0 −U3
| {z } | {z } | {z } | {z }
A Γ0 Γ1 Γ3
(8.68)
In a first example, we assume that the in-plane stiffness in x-direction has to be maximized.
Consequently, the entry A11 must be maximal. The material invariants U2 and U3 are
strictly positive (see Appendix B). Using the plot of the feasible domain in Figure 8.9 and
the matrices in equation (8.68) we can determine that the entry A11 is maximal for being
both, V1A and V3A , equal to one. This corresponds to a laminate consisting of 0◦ -layers
only.

max A11 = U1 + V1A · U2 + V3A · U3 → V1A = 1, V3A = 1 → (04 )S

The same calculations can be done for a laminate with maximal in-plane shear stiffness:

max A66 = U5 + V1A · 0 + V3A · (−U3 ) → V3A = −1, (V1A = 0) → (±452 )S

Alternatively, we may seek for a laminate which has equal stiffness in all directions, which
is also known as quasi-isotropic laminate. The equations characterizing a quasi-isotropic
laminate are
A11 = A22 → U1 + V1A · U2 + V3A · U3 = U1 − V1A · U2 + V3A · U3
1
A66 = 2 (A11 − A12 ) → U5 − V3A · U3 = 12 U1 + V1A · U2 + V3A · U3 − U4 + V3A · U3

The first equation directly implies

V1A = 0
Using the explicit formulation of the invariants given in Appendix B, it can be shown that

V3A = 0

Since we assume the lamination parameters V2A and V4A to be zero, the laminate must
be balanced wherefore the solution corresponds to (0/45/-45/90)S -laminate. However,
this is not the only solution for a quasi-isotropic laminate. Also a laminate with layup
(0/60/-60)S is quasi-isotropic, symmetric and balanced and results in the same lamina-
tion parameters (V1A = V3A = 0). This demonstrates that the back-substitution for the
lamination parameters to the physical laminate is not is not unique. Of course, the two
laminates (0/45/-45/90)S and (0/60/-60)S produce different entries ViD . However, the
difference becomes small if having a large number of layers.

c ETH Zürich IDMF-CMAS, March 1, 2018

154 Composite Structures

c ETH Zürich IDMF-CMAS, March 1, 2018

8.5 Optimization on Physical Design Variables 155

8.5 Optimization on Physical Design Variables

The parametrization schemes presented in this section all base on the physical design
variables such as fiber orientation angles, layer thickness, stacking sequence, number of
layers of material. In contrast to the parametrization with lamination parameter, which
directly operates on the entries of the ABD-matrix, the ABD-matrix is here a result of
the physical design variables which are processed with the homogenization routine of the
CLT introduced in section 8.2.2. As long as the geometry of the considered structures are
simple (e.g. plates or cylinders) and the loads are distributed homogeneously, a solution
can directly be evaluated with the CLT. However, if the geometries become more complex,
the application of a FEM model becomes necessary. A flowchart of a typical optimization
process including a Finite Element evaluation is shown in Figure 8.10. It is separated into

x0

Optimizer
x Optimization
Environment

Read Read Write

FEM FEM
FEM FEM
Output FEM-Solver Input
Model Environment
File File

Figure 8.10: Flowchart of a typical optimization with an FEM-Solver

two different environments, the optimization environment which is usually implemented

in a programming language (e.g. MATLAB, Python, C), and a FEM environment which
is represented by a FEM solver. Typically, the process is starting with an initial finite
element model whose laminate layup has to be optimized. An input-interface is reading all
the relevant data for the optimization. Naturally, the optimization can be combined with
a simultaneous shape optimization. The optimization algorithm then generates an initial
design variable vector x0 which is processed to the new model. Based on that, the output-
interface generates a input-file for the finite element framework and the finite element
solver is activated by the optimizer. The resulting output-file is read again to calculate
the objective value. The evaluation of the objective value can either be carried out in the
finite element framework or the optimization environment. Based on the objective value,
the optimizer creates a new set of design variables. Many programming environments
provide a set of pre-implemented optimization algorithms wherefore the programming
effort can be reduced. However, the implementation of the input/output-interfaces for an
efficient communication between the two parts is more programming expensive.

8.5.1 Optimization of the Fiber Orientation

The optimization of the fiber orientation is a very common discipline in the field of laminate
optimization. The adaption of the fiber orientation allows a specific design of material
anisotropy which is one of the major benefits of laminates compared to isotropic materials.
The fiber orientations of each layer are adapted with respect to a specific design criterion.
Taking a very simple approach, it is assumed that the number of layers and the material

c ETH Zürich IDMF-CMAS, March 1, 2018

156 Composite Structures

is predefined and fixed during the optimization. The orientation angles are taken as
design variables wherefore the optimization problem becomes continuous. Figure 8.11(a)
schematically illustrates a laminate consisting of four layers whose orientation angles are
optimized. The design variables can simply be arranged in an array which is illustrated

φ1 = x1,1 φ1 = x1,n
φ2 = x2,1 φ2 = x2,n
φ3 = x3,1 φ3 = x3,n
φ4 = x4,1 φ4 = x4,n
(a) Fiber orientation optimization scheme

x1,1 x2,1 x3,1 x4,1 x1,n x2,n x3,n x4,n

(b) Fiber orientation optimization parametrization

Figure 8.11: Fiber orientation optimization

in Figure 8.11(b). In general, the fiber orientation values are real numbers (ϕj ∈ R).
Due to the periodicity of the orientation angles, their are often bounded to an interval of
180◦ , such as [0◦ , 180◦ ] or [−90◦ , 90◦ ]. Consequently, the search space is restricted without
reducing the potential solution quality. However, some optimization algorithms cannot
handle constrained problems. Of course, the search space is not restricted using this type
of algorithms. The continuity of the design variables allows to solve fiber orientation
optimization with algorithms of the class of mathematical programming. However, it
must be considered that the search space is multi-modal which may be problematic for
this class of algorithms since they risk to get stuck in a local optima. The non-convexity
of a search space with two orientation angles has been visualized by Keller [78]. Figure

-90 -75 -60 -45 -30 -15 0 15 30 45 60 75

0.18
75 0.18 0.19 0.16 75

0.17 0.2
60 0.18 0.18 60
0.21
0.2 0.18 0.21 0.19
45 0.18 45

30 30

15 15
θ2

0 0.16 0.17 0.18 0.17 0.16 0.18 0.17 0.16 0

y -15 -15

-30 -30
0.18
-45 -45
0.21 0.18 0.21 0.19
θ
-60 0.18 -60
x 0. 0.17 .2
-75 0.19 0.16 -75
0.18 0.18 0.18
-90 0.18 0.18 -90
-90 -75 -60 -45 -30 -15 0 15 30 45 60 75
θ1

(a) The structure is clamped at the (b) Contour plot of the objective for the tensile
left side and a uniform line-load of specimen experiment as a function of the ply an-
0.01 N/mm is applied to right side. gles φ1 and φ2 (in degrees).

Figure 8.12: Geometry and objective of the tensile specimen experiment. Source: Keller
[78]

8.12(a) illustrates a simple plate which is clamped at the left side and loaded with a

c ETH Zürich IDMF-CMAS, March 1, 2018

8.5 Optimization on Physical Design Variables 157

uniform line load. The plate consists of two layers with orientation angles x1 and x2 .
Figure 8.12(b) shows the corresponding contour plot of the objective function which is
obviously multi-modal. There will even be more local optima when using more than two
layers.
Allowing the orientation angles to be real may lead to solutions with many positions
after decimal points which is inappropriate for manufacturing since processes are often re-
stricted to discrete steps, for instance to 5◦ -steps or higher. This renders the problem to be
discrete wherefore the application of algorithms of the category of mathematical program-
ming becomes impossible. Stochastic algorithms may mitigate that problem. Alterna-
tively, continuous orientation angles may be rounded to the next feasible angle. However,
if the discrete steps are too large, the solution may change its behavior significantly.

8.5.2 Optimization of the Stacking Sequence

Another optimization discipline, which has already been performed in the early stages of
computer-aided laminate design, is the optimization of the laminate stacking sequence. It
basically modifies the order of the laminate stack by exchanging the layers or its position
in the laminate, respectively. Stacking sequence optimization is important if the load cases
include bending. For such a case, it is absolutely essential which layers are located in the
outer positions since these layers have a greater impact on the area moment of inertia.
Moreover, the stacking sequence is important for the mechanical coupling between bend-
ing and twisting which is mathematically represented by the B-matrix. If the layers are
distributed appropriately, an in-plane load may induce out-of-plane deformations and vice
versa which may be interesting for specific applications. Additionally, the interlaminar
stress distribution is strongly dependent on the distribution of the single layers. Figure
8.13(a) schematically illustrates a stacking sequence optimization. The parametrization of

A B
B D
C A
D C
(a) Stacking sequence optimization scheme

A B C D B D A C
(b) Stacking sequence optimization parametrization

Figure 8.13: Stacking sequence optimization

a stacking sequence is usually solved by arranging the numbered layers in an array accord-
ing to Figure 8.13(b). An exchange of the array sequence is equivalent to an exchange of
the corresponding layers. When exchanging two layers in the laminate stack, the objec-
tive function changes by leaps and bounds wherefore the optimization problem becomes
discrete. Consequently, these category of problems can only be solved with stochastic
algorithms. A type of algorithms which has been used very often to optimize the stacking
sequence are Genetic Algorithms (GA) [79, 80, 81, 82, 83, 84, 85, 86, 87]. Basically, a
stacking sequence problem can be transformed in to a problem of fiber orientation op-
timization. Instead of exchanging the layers, their fiber orientations may be varied and
potentially, the same solution can be found.

c ETH Zürich IDMF-CMAS, March 1, 2018

158 Composite Structures

8.5.3 Material Optimization

Another parameter which can be optimized is the material of the single layers. Of course, it
does not make sense to vary the engineering parameters of the sub-materials. However, the
applied materials can be selected from a list of predefined materials which are compatible
with each other. Usually, a set of candidate materials is predefined in a database. Each
material has a unique labeling number which is used as design variable. Consequently, the
design variables are discrete and the problem is solved with stochastic algorithms. Pure
material optimizations are rather rare since the material which may be used for the design
is often predefined. Moreover, it is intuitively clear to the engineer which material best fits
the needs. However, material optimization is sometimes used in combination with other
laminate optimization disciplines.
The Discrete Material Optimization (DMO) method which has been introduced by
Stegmann and Lund [88, 89] renders the discrete problem a continuous formulation. The
constitutive matrix C is expressed as a weighted sum of candidate materials which are
characterized by their own constitutive matrices C i .
n
X
C= wi C i , 0 ≤ wi ≤ 1 (8.69)
i=1

The weighting factors wi , which are bounded by 0 and 1, are used as design variables where-
fore the problem becomes continuous. According to the topology optimization method of
Bendsøe and Kikuchi (see Section 9.4), the design variables are pushed towards their
boundaries and the final weighting factor array of each layer must have one value of 1
and the rest must be zero. If defining several constitutive matrices made of one material
but different orientation angles, the method can be used for fiber orientation or stacking
sequence optimizations.

8.5.4 Optimization of the Laminate Thickness

In contrast to stacking sequence or fiber orientations, the laminate thickness has an influ-
ence on the structural weight. Consequently, the thickness must be adapted at a certain
stage of the design process if aiming for structures with minimal weight. There are two
options for varying the laminate thickness, whether a variation of the thickness of the
particular layers or a variation of the number of layers.

Variable Layer Thickness

Analogously to the fiber orientation optimization, the thickness of each layer can be varied.
Basically, the layer thickness are continuous wherefore the algorithms of mathematical
programming can be used. It must be ensured, that the thickness values do not become
negative since this would lead to physically meaningless solutions. Some structural models
can handle negative thickness values (e.g. the CLT) which leads to negative stiffness and
negative mass. If allowing the thickness values to become zero, layers can vanish which
results in a topology optimization. A layer having zero thickness is non-existent wherefore,
practically, the number of layers is changed. If using the Finite Element Method to
represent the structural model, it must be taken care that the thickness may not become
zero for all layers since this would result in numerical problems.
From the practical point of view, the variation of the layer thickness with continuous
variables is limited. If aiming for manufacturing-friendly solutions, the layer thickness
is restricted to values given by the semi-finished products which are provided by the

c ETH Zürich IDMF-CMAS, March 1, 2018

8.5 Optimization on Physical Design Variables 159

t1 = x1,1 t1 = x1,n
t2 = x2,1 t2 = x2,n
t3 = x3,1 t3 = x3,n
t4 = x4,1 t4 = x4,n

(a) Layer thickness optimization scheme

x1,1 x2,1 x3,1 x4,1 x1,n x2,n x3,n x4,n

(b) Layer thickness optimization parametrization

Figure 8.14: Layer thickness optimization

manufacturers. Practically, this can be done differently. A pragmatic approach is to let

the thickness variables to be continuous during the optimization process. The thickness
values of the final solution are then rounded to the next value predefined by the semi-
finished product. It is important to assure that the new solution is feasible and that its
objective value is not differing too much from the initially found solution. It might be that
the rounded solution behaves completely different. Alternatively, the thickness variables
can be set to the discrete values given by the semi-finished product. This requires the
application of stochastic algorithms.
If using the layer thicknesses as design variables, it is important to include the structural
weight in the objective function or in the side constraints. In case that the optimization
objective is to maximize for instance the stiffness or the strength, the weight must be
bounded with a side constraint. Alternatively, the thicknesses will increase to very high
values and the structural weight tends to infinity. If aiming for solutions with minimal
weight, stiffness or strength must be bounded. Otherwise, the thickness tends to become
zero.

Variable Number of Layers

An alternative approach to vary the thickness of the laminate is to let the number of lay-
ers be variable. The handling of these problems becomes slightly more difficult since the
number of design variables is not constant anymore. Figure 8.15 schematically shows an

φ1 = x1,n
φ1 = x1,1
φ2 = x2,n
φ2 = x2,1
φ3 = x3,n
φ3 = x3,1
φ4 = x4,n
φ4 = x4,1
φ5 = x5,n
(a) Scheme of a problem with variable number of layers

x1,1 x2,1 x3,1 x4,1 x1,n x2,n x3,n x4,n x5,n

(b) Parametrization of a problem with variable number of layers

Figure 8.15: Optimization with variable number of layers

optimization problem with a variable number of layers. For each new appearing layer, the
fiber orientation, the thickness, the material and the position in the laminate stack has to
be determined (even if in Figure 8.15 only the orientation angles are indicated), which is

c ETH Zürich IDMF-CMAS, March 1, 2018

160 Composite Structures

usually done randomly or by choosing them randomly from a predefined database, respec-
tively. Due to the variable length of the design variable array, these kinds of problems can
not be solved with algorithms from Mathematical Programming. Usually, Genetic Algo-
rithms with variable genotype length are applied. It is important to implement appropriate
reproduction and mutation routines in order to come close to optimal solutions.

8.5.5 Combined Laminate Optimizations

Often, the above introduced disciplines are combined. For instance, both, the fiber ori-
entation angles and the layer thickness are optimized simultaneously. On one hand, it
increases the potential for finding optimal solutions. On the other hand, the search space
is becoming more complex wherefore finding the optimal solution becomes more difficult.
Basically, the parametrization must be chosen in a way that the solutions are feasible
for production. Even if mathematically optimal solutions might be interesting from the
mechanical point of view, they do not really make sense if they cannot be realized. An
appropriate parametrization scheme helps to reduce the search space in order to obtain
manufacturing-friendly solutions.

c ETH Zürich IDMF-CMAS, March 1, 2018

8.6 Laminate Tailoring 161

8.6 Laminate Tailoring

The methods above are addressing problems for which an optimal unique laminate has
to be [Link]
If the design domain is geometrically simple (e.g. rectangular plates or
cylinders), and the loads are applied homogeneously as illustrated in Figure 8.16, the
 In contrast
internal load distribution is tohomogeneous
isotropic materials, laminates
as well. can loadings,
For such be designed
there exists a unique
specifically to the occurring load conditions

 For homogeneous loading states, finding an optimal laminate layup

isFigure 8.16: Homogeneous loads: Tension and shear
rather easy

 Homogeneous
optimal laminate which can be load
found rather easily.
distributions occurHowever, in general,
only in academic the internal loads
examples
are inhomogeneous and the stress state at each point of the laminate is different. Figure
Motivation
8.17 shows the strain field of a plate with a centered hole which is loaded uniaxially. Even
18. December 2012 Doctoral Examination – Benjamin Schläpfer 7

Find global
Figure laminate layup strain
8.17: Inhomogeneous which field
is theofbest solution
a plate with for the hole
centered
structural behavior → local loadings are not respected
Laminatoptimierung
if this problem (Laminate-Tailoring)
is quite simple, the internal load distribution is highly inhomogeneous.
 Tailor
In order to the local
fully exploit thelaminate properties
potential to the the
of composites, locallaminate
needs by taking have to be
properties
advantage of locally varying thickness or fiber orientations
tailored to the local load conditions. This can be achieved by splitting the design domain
locally
Ω into sub-domains Ω varying
ƒ Die grundlegenden anisotropy
each havingParametrisierungen
a unique layup, und Optimierungsmethoden
which is schematically illustrated in
bleibeni dieselben
Figure 8.18. Consequently, the objective here is to find an optimal
ƒ Die Design-Domain wird in sogenannte Sub-Domains aufgeteilt laminate for each sub-

18. December 2012 Doctoral Examination – Benjamin Schläpfer

ȍ 8

ȍ3

ȍ1 ȍ2

ƒ Jede Sub-Domain besitzt einen eigenen Laminataufbau

ƒ Die Optimierung
Figure wird aufwendiger
8.18: Global domain Ωdasplit
die Anzahl Optimierungs-Ω
with subdomains i
variablen ansteigt
domain. The optimization of the local laminate layup is also called laminate tailoring.
Even if the optimization techniques are the same as presented above, laminate tailoring is
Montag, 13. Mai 2013 Optimierung von Schalentragwerken aus Faserverbundwerkstoffen 45

a relatively new discipline since the computer system requirements are rather high. The
number of design variables increases linearly with the number of sub-domains which makes
the optimization more computational expensive.
Performing laminate tailoring, there might be the risk of finding solutions which can-
not be physically realized since they are not manufacturing friendly. Consequently, man-
ufacturing aspects should be regarded in the parametrization concept. For instance, the

c ETH Zürich IDMF-CMAS, March 1, 2018

162 Composite Structures

connectivity of the sub-domain must be guaranteed with a minimal number of layers

covering multiple sub-domains. Figure 8.19 shows a laminate section consisting of three

Elemente

Layer 1 Global Ply A

25

Layer 2 Global Ply B

12

Layer 3 Global Ply C

31

Layer 4 Global Ply D

9

Layer 5 Global Ply E

16

Sub-domain
Region 11 Sub-domain
Region 2 2 Sub-domain
Region 33

Figure 8.19: Schematic illustration of layers covering multiple sub-domains

sub-domains. Layers A, C and E are covering the entire domain and a global connectiv-
ity is given. A simple method to guarantee connectivity is to define a number of plies
which cover the entire domain and may not be adapted during the optimization. This
may restrict again the search space.
Basically, the geometric representation of the sub-domains can be predefined and as-
sumed to be fixed during the optimization. This results in a constant number of design
variables which simplifies the optimization, but also restricts the search space and the so-
lution quality, respectively. More freedom in terms of design is achieved if the parametriza-
tion additionally includes the geometrical representation of the sub-domains. A variable
number of sub-domains leads to a variable number of design variables which can only
be solved by means of stochastic algorithms. Two parametrization approaches for the
sub-domains are presented in the next two sections.

8.6.1 FEM-Based Parametrization of the Sub-Domains

A simple approach, which is often used to find locally varying laminates (e.g. [89, 90,
91, 92, 93]), is the direct application of the finite elements as sub-domains. Consequently,
the goal is to find a unique laminate within each finite element wherefore the number of
design variables increases linearly with finite element mesh size. A schematic illustration
of such a parametrization is shown in Figure 8.20(a). The contours of the sub-domains are

Ω1 Ω4 Ω1

Ω2 Ω…

Ω3 Ωn Ω2 Ωn
(a) Parametrization with single elements (b) Parametrization with multiple ele-
representing a sub-domain ments pooled into a sub-domain

Figure 8.20: FEM-based Parametrization of Sub-Domains

dependent on the finite element mesh which may restrict the design freedom, but only if a
rough mesh is chosen. However, no additional parametrization of the geometry is needed.

c ETH Zürich IDMF-CMAS, March 1, 2018

8.6 Laminate Tailoring 163

Alternatively, the number of design variables can be reduced by pooling a given number
of finite elements into groups which then represent the design domains. A schematic
sketch is shown in Figure 8.20(b). One possibility is to pre-define the elements belonging
to one sub-domain and keep it fixed during the optimization. This of course restricts
the freedom of design but it might be helpful to make the solutions more manufacturing
friendly. Alternatively, the elements which belonging to a sub-domain may be chosen to
be variable. The realization and the optimization become more complex but the potential
space for improved solutions becomes bigger. Consider that for both parameterizations,
the solution quality, but also the computational costs are dependent on the finite element
mesh.
Giger et al.[94] proposed a graph-based parametrization scheme which allows to make
the elements which are pooled in a sub-domain to be variable. This parametrization

Figure 8.21: Graph-based parametrized vertex patches. Source: Giger [94]

scheme allows maximal freedom in terms of design since regions having a unique laminate
are morphing. The material data and the orientation angle of a so called patch, which
can be understood as a layer that partially covers the design domain, is stored in a patch
vertex as shown in Figure 8.21. The elements belonging to a patch, but also the material
and the orientation angle can be modified by means of a genetic algorithm. In order to
guarantee the connectivity of a patch, only boundary elements may be added or removed
which are located automatically by a computer routine. The total laminate layup in each
finite element is set up by summarizing the patch layers the respective element belongs
to in the right order (according to Figure 8.21 for instance from left to right). This
parametrization scheme is inspired by the manufacturing process and the solutions are
manufacturing-friendly.
An alternative approach for laminate tailoring based on a finite element based
parametrization is proposed by Schläpfer and Kress [93, 95, 66]. The method basically
locally reinforces a predefined laminate structure with laminate patches regarding a spe-
cific design criterion such as stiffness, strength or dynamic behavior by simultaneously
aiming for minimal mass. Layers with given material and fiber orientation are predefined
in the design domain. Some of these layers have a thickness of zero wherefore they do not
really contribute to the finite element model. These zero-thickness layers act as potential
reinforcements. The local reinforcements are generated by partially setting the thickness
of these layers to values of the semi-finished material. The decision, whether a layer in
one finite element is set to a real value or not is made on the sensitivities which have been

c ETH Zürich IDMF-CMAS, March 1, 2018

164 Composite Structures

introduced in section 8.3.2. In particular, the sensitivities of a particular design criterion

with respect to the layer thickness in each finite element are calculated. Since a change of
the thickness has always an influence on the stiffness, even if the layer has a thickness of
zero (consider that in equation (8.43) and (8.44), the thickness is not existent anymore),
they are non-zero for the defined zero-thickness layers. Consequently, the sensitivities
provide information how big the impact of a thickness change of a layer on the objective
function is. Figure 8.22 shows a simple optimization of a vibrating plate, whose second and
third eigenfrequency are separated. The sensitivities in Figure 8.22(a) are a superposition
of the sensitivities of the second and the third vibrating mode. The resulting layers are
represented by the black areas in Figure 8.22(b). The present approach is inspired by the
manufacturing process. The global connectivity is guaranteed by basic layers which are
predefined and not participating in the optimization process. The obtained solutions can
basically all be manufactured.

8.6.2 CAD-Based Parametrization of the Sub-Domains

Alternatively, the laminate patches can be parametrized by taking advantage of a CAD-
environment as for example made by Zehnder et al.[96]. There, the geometry of the patches
with given material and fiber orientations are parametrized by means of a CAD-software.
The shape of the patches is only restricted by the parametrization capacities of the software
itself. If using splines, the shapes can basically be chosen arbitrarily. As illustrated in
Figure 8.23, the sub-domains having a unique laminate layup are defined by summarizing
the patches in the right order. Also here, the structural model is evaluated with help of
the finite element method. However, here finite element mesh is created dependent on the
parametrization of the parametrization of the sub-domain. Consequently, the sub-domains
are not dependent on the finite element mesh as it was the case above. An evaluation of
a solution always includes the generation and solving of the finite element model which is

0° 0°

45° 45°

-45° -45°

90° 90°

low high

(a) Red areas indicate high sensitivities which (b) The resulting local reinforcements are rep-
implies that the thickness of the corresponding resented by the black areas
layer has to be increased in order to improve the
design value

Figure 8.22: Sensitivities and local reinforcements of a vibration plate for which the second
and the third eigenfrequency are separated. Source: Schläpfer [66]

c ETH Zürich IDMF-CMAS, March 1, 2018

8.6 Laminate Tailoring 165

Figure 8.23: Connection between global layers and laminate regions. Source: Zehnder [96]

rather computational expensive. Also this method works with a genetic algorithm.

c ETH Zürich IDMF-CMAS, March 1, 2018

166 Composite Structures

c ETH Zürich IDMF-CMAS, March 1, 2018

Chapter 9

Selected Methods and Case

Studies

9.1 Computer Aided Optimization after Mattheck

The growth-strategy method by Claus Mattheck [97] can, under certain circumstances
more efficiently than other methods, reduces geometry-caused stress concentrations in de-
signs. His algorithm, called computer-aided optimization (CAO) is modelled to be an
analog to the growth behavior of trees which he derived from nature observations. He
identified the axiom of constant stresses as the basic principle of the growth and the heal-
ing mechanisms of trees. Trees are annually forming a region where growth takes place,
called cambium, located right under the bark. The thickness growth ensues from the
cambium that is growing faster at more highly stressed locations than at others. So, by
time, a strength optimization results in the sense of reducing local stress concentrations
by increasing the areas of the more highly stressed cross sections.
Mattheck transforms this observed mechanism to an automated optimization method.
The structural model of the considered part, suffering reduced strength from stress con-
centrations, is modeled with a layer of constant thickness underneath its surface along the
highly stressed region, as shown in Fig. 9.1. Each iteration of the CAO process consists of

r e g io n to b e o p tim iz e d

c o n s ta n t- th ic k n e s s s u r fa c e la y e r o r c a m b iu m

Figure 9.1: Part with stress raising geometry and modelled cambium

two phases. The first phase determines the stress distribution within the cambium, due to
the specified geometric boundary conditions and applied forces, by a structural analysis.
The stresses are reduced to some equivalent stress distribution. During the second phase,
a temperature distribution is derived from the equivalent stress distribution. A tempera-

c ETH Zürich IDMF-CMAS, March 1, 2018

168 Selected Methods and Case Studies

ture expansion coefficient is appropriately chosen and Young’s modulus of the cambium is
reduced by some factor, for instance 400. The structural analysis of the second phase then
simulates a growth of the cambium effected by the temperature strains. The strains tend
to be higher where higher equivalent stress values were calculated by the preceding first
phase. The strains imply a growth of the structural model. The growth is numerically
recorded by adding the nodal-point displacements to the nodal point reference positions.
The so obtained geometry changes can have the effect that the stress concentrations due
to the mechanical loading are reduced. The process is illustrated in Fig. 9.2 During the

Figure 9.2: Two-phase CAO process after Mattheck

iterations the geometry changes automatically so that the stress concentrations are sys-
tematically reduced.
The advantage of the method with its ingenious simplicity lies in the relatively low
numerical effort to obtain a significantly improved design. The method was used with
much success to reduce stress concentrations, and dramatically increase lifetime, in load-
carrying automotive parts that could be re-designed to achieve lower weight.
The disadvantage of it lies in the very nature of is its local approach: more sophis-
ticated solutions other than increasing the thickness of highly stressed parts can not be
found by it. For example, the maximum-strength flywheel design, see section 9.3, could
not be solved by CAO.

c ETH Zürich IDMF-CMAS, March 1, 2018

9.2 Soft-Kill Option after Mattheck 169

9.2 Soft-Kill Option after Mattheck

The topology optimization method after Bendsœand Kikuchi presented in the preceding
section uses as a global objective some specified compliance property resulting from the
behavior of the whole structure. Claus Mattheck [97] replaces the minimization of the
global compliance objective function with an algorithm, sketched in Fig. 9.3, based on a
local optimality criterion. Again, the geometric design space is specified, spanned with

Figure 9.3: Soft-Kill Option process after Mattheck

a finite-element mesh, and geometric boundary conditions as well as forces specified. An

average Young’s modulus is initially assigned to each of the finite elements in the design
space. A structural analysis obtains a primary displacement solution the element deriva-
tives of which are used to obtain a stress distribution over the domain. The stresses are
combined to establish the distribution of an equivalent stress. The local optimality cri-
terion used by Mattheck assumes that the stiffness of the design will globally increase
when Young’s modulus is increased in regions with higher stresses and reduced where the
stresses are lower. When the stresses fall below a certain threshold, Young’s modulus is
reduced to zero, or with respect to numerical stability, small values.

c ETH Zürich IDMF-CMAS, March 1, 2018

170 Selected Methods and Case Studies

c ETH Zürich IDMF-CMAS, March 1, 2018

9.3 Flywheel Optimization and Inspired Mechanical Model 171

9.3 Flywheel Optimization and Inspired Mechanical Model

This section brings together the results of automated shape optimization of a flywheel and
Stodola’s solution on high-strength turbine disk design. The example exposes that oppor-
tunities exist that computer aided optimization may inspire understanding and modeling
of complex mechanical situations.
Within a flywheel of constant thickness having a central bore stresses are distributed as
plotted in Fig. 9.4. The normal stress in radial direction σr must meet the natural bound-

Figure 9.4: Initial flywheel shape and radial and circumferential stress distributions

ary condition at the bore and at the rim. The circumferential stress σθ is higher than the
radial stress everywhere and has a stress peak at the bore. Because of the inhomogeneous
distribution of both the normal and the circumferential stresses, the material of the disk
of constant thickness is not economically used.

9.3.1 Sodola’s Solution

Stodola [24] invented an optimum-strength design for the steam-turbine center disk. The
center disk connects the turbine’s blades to its drive shaft. Stodola’s design essentially
avoids a central hole, Fig. 9.5. The inertia forces of the blades acting on the rim of

Figure 9.5: Blueprint of a turbine disk design solution after Stodola

the central disk are taken to be equivalent to an average radial stress which enters the
mathematical model as a natural boundary condition. By choosing the more general
formulation - allowing for a changing thickness t - of the equilibrium of forces in the radial
direction,
t
(σrt) ,r + (σr − σθ ) + trω 2 ρ = 0 , (9.1)
r

c ETH Zürich IDMF-CMAS, March 1, 2018

172 Selected Methods and Case Studies

and requiring both stresses to be equal and also constant everywhere,

σr = σθ = σ , (9.2)

the general equilibrium equation is reduced to a differential equation where only the thick-
ness remains as a dependent variable

ρω 2
t,r +t0 r=0, (9.3)
σ
and the shape of the evenly stressed flywheel is defined by the solution
ρω 2 2
t + t0 e− 2σ
r
(9.4)

It must, however, be kept in mind that the model is based on radial equilibrium (9.1)
only, wherefore it is accurate only if the reference thickness t0 is small compared to the
diameter of the disk. Then, the normal stress in the axial direction σz and the shear stress
τrz can be neglected. As many flywheels possess a central bore for the purpose of fixing
them on a rotating shaft, the study of Stodola’s problem provokes the question whether
a shape can be found for which an almost even stress distribution exists in the presence
of a central bore and in the absence of external radial stress. An answer to this question
is found by using a numerical shape optimization scheme where a FEM-based structural
model provides the system equations.

9.3.2 Shape optimization

The objective is to find a shape of the flywheel resulting in an even stress distribution
within it. Using a flywheel of constant thickness t0 as the initial configuration its mass
and rotational energy are also to be kept constant. Mathematically, this is expressed
by using the standard deviation of the radial stress distribution σ as the objective and
by introducing the equality constraints h1 (constant mass) and h2 (constant rotational
energy), Z
n
O= (σ − σ̄)2 dr = min, σ̄ : average stress (9.5)
r
Z
n
h1 = 2πρ r (t(r) − t0 )2 dr = 0 (9.6)
r
Z
n
h2 = 2πρ r3 (t(r) − t0 )2 dr = 0 (9.7)
r
The thickness values at the element interfaces in the radial direction take the role of shape
parameters as outlined in Fig. 9.6. Thus, the number of optimization variables depends

t0+ , t t0

Figure 9.6: Parameterization of thickness and mesh generator

c ETH Zürich IDMF-CMAS, March 1, 2018

9.3 Flywheel Optimization and Inspired Mechanical Model 173

linearly on the number of elements in the radial direction but is independent of the number
of elements in the axial direction. The position of the nodes on the element sides is always
on the straight line connecting the respective corner nodes. The objective function is min-
imized using the method of feasible directions according to the textbook of Vanderplaats
[20] as well as exact gradient information from the objective and the constraint functions.
Additional corrections to ensure that the equality constraints remain satisfied are made
after each line search. Quadratic approximations of the objective function are used for
quickly finding the minimum value along the current search directions as described in the
textbooks of Reklaitis et al. [33] and and Baier et al. [98].

9.3.3 Discussion of results

Fig. 9.4 shows the initial configuration, the finite element mesh, and the initial stress
distributions. These stress distributions indicate that the region adjacent to the central
hole is significantly more stressed than other parts of the flywheel.
The shape optimization procedure yields the results plotted in Figs. 9.7 and 9.8. Both the
stress distributions and the shape suggest a division of the flywheel into three regions for
a characterization of the optimized geometry. Region I adjacent to the hole increases pro-
gressively in thickness with decreasing distance to the hole. The circumferential stress also
increases with decreasing distance to the hole but its maximum value is much lower than
in the initial configuration. The most striking feature of region II is the near constancy
of both the radial and the circumferential stresses. Moreover, the values of both stresses
are almost equal. The thickness of region II decreases progressively but at a slow rate
with increasing distance to the hole. Within region III, the shape of which is somewhat
reminiscent of a bell, the radial stress drops to zero.
The difference between the results shown in Fig. 9.7 and in Fig. 9.8 is in the choice of
the stress used for the objective. Fig. 9.7 shows the shape corresponding to the most

Figure 9.7: Shape optimized after maximum stress criterion and stress distributions

evenly distributed circumferential stress. Since the circumferential stress has higher val-
ues everywhere in the initial configuration, it is identical to the first principal stress and
minimizing the objective represents the engineering aspect of seeking a spatially constant
probability of brittle failure in the flywheel. Fig. 9.8 shows the shape corresponding to
the most evenly distributed von Mises stress. The stress distribution obtained here comes
closest to a spatially constant probability of yield failure in the flywheel. Thus, it may
be concluded that local details of the shape depend on the choice of stress for the objec-
tive function while global features such as the existence of regions I, II, and III remain

c ETH Zürich IDMF-CMAS, March 1, 2018

174 Selected Methods and Case Studies

Figure 9.8: Shape optimized after yield stress criterion and stress distributions

the same. Both the shape of region II and the almost even distribution of the stresses
within it bring to mind Stodola’s evenly stressed turbine disk. The deviations from per-
fect constancy of the stresses are explained by the effects of steep thickness changes on the
two-dimensional stress equilibrium. The hypothesis is proposed that region I alleviates
the stress perturbations of the central hole and that region III introduces radial stresses
on the outer edge of region II which are used in Stodola’s model to simulate the forces
exerted by the blades on the rotating turbine disk.

9.3.4 Simple Prediction of Optimum Shape Features

A mechanical system is proposed, consisting of a disk modelling region II and two discrete
rings located at the edges of the disk, modeling regions I and III (Fig. 9.9). The disk has

d is c r e te r in g w ith d is c r e te r in g w ith
a r e a A 1 fo r r e g io n I a r e a A 2 fo r r e g io n II
a x is o f r o ta tio n

S t o d o la 's d is k o f v a r y in g
th ic k n e s s fo r r e g io n II

Figure 9.9: Mechanical model consisting of Stodola’s disk and two discrete rings repre-
senting regions I and III

a thickness distribution as given by Stodola [24]. The inner and outer rings at positions
r1 and r2 have the cross-sectional areas A1 and A2 , respectively.
The line load Nr resulting from the integration of the radial stress σr over the thickness
is assumed to be positive, and the circumferential stresses σθ in the rings are
r1
σθ1 = ρω 2 r12 + Nr1
A1
r2
σθ2 = ρω 2 r22 − Nr2
A2
where the first terms on the right-hand side give the circumferential stress in the inde-
pendently rotating ring and the second terms make corrections for the tensile radial line
load which tends to pull the inner ring apart and the outer ring together. The rotational

c ETH Zürich IDMF-CMAS, March 1, 2018

9.3 Flywheel Optimization and Inspired Mechanical Model 175

symmetry of the problem automatically satisfies the condition of force equilibrium in the
circumferential direction. The condition of force equilibrium in the radial direction re-
quires that the radial line loads be continuous at the interfaces between the disk and the
rings. If the material of the flywheel is homogeneous, the kinematics are satisfied if the
circumferential stress is also continuous. The particular shape of the disk as defined by
Stodola’s equation guarantees the constant stress distribution indicated in (9.2) as long as
the conditions at the discrete rings
Nri = σt|ri , i = 1, 2 (9.8)
apply. This determines the cross-sectional areas of the rings, A1 and A2 ,
σr1 ρω 2 r 2
− 2σ 1
A1 = t 0 e
σ − ρω 2 r12
σr1 ρω 2 r 2
− 2σ 2
A1 = t 0 e
σ − ρω 2 r22
The cross-sectional areas of the rings are positive as long as σ is higher than the circum-
ferential stress due to the inertia body forces that would arise in the rotating inner ring
as an isolated system, and also lower than the circumferential stress that would arise in
the rotating outer ring:
ρω 2 r12 ≤ σ ≤ ρω 2 r22 . (9.9)
Then the stress σ, acting in the disk, tends to pull the inner ring apart in addition to
its own inertia effect and to restrain the outer ring against the inertia effect acting on it.
Thus, the stress σ can be chosen freely as long as it lies between the bounds defined by
(9.9). In accordance with the finite-element shape optimization described above it was
decided to fix the mass M as well as the rotational energy U of the disk to the respective
values of the initial shape (IS),
 Z r2 2

2 2 − ρω r 2


M = tIS πρ r2 − r1 = 2πρ r1 A1 + r2 A2 + t0 re 2σ , (9.10)
r1
 Z r2 
1 3 − ρω
2
r2
U = tIS πρω 2 4 4 2 3 3


r2 − r1 = 2πρω r1 A1 + r2 A2 + t0 r e 2σ . (9.11)
4 r1
This can be achieved by adjusting the reference thickness t0 and the constant stress value
σ. This stress value will always be lower than the maximum circumferential stress at the
edge of the hole in the initial design. The practical significance of the optimized shape
lies in the wider choice of less expensive materials that can be used for the flywheel. The
integrals in (9.10) and (9.11) can not be solved algebraically. They can, however, be
eliminated by using the identity
Z r2 Z r2
ar2 r2 ar2 r2 2 ρω 2
re dr ≡ e −a r3 ear dr , a=− . (9.12)
r1 2 r1 r1 2σ
Combining (9.10), (9.11), and (9.12) leads to the stress σ,
U 1
= ρω 2 r22 + r22 .


σ=ρ (9.13)
M 4
The maximum circumferential stress at the edge of the hole of the initial configuration
depends on the material’s Poisson’s ratio ν,
1
σθIS = ρω 2 r22 (3 + ν) + r12 (1 − ν) .
 
(9.14)
4

c ETH Zürich IDMF-CMAS, March 1, 2018

176 Selected Methods and Case Studies

Its value, even when neglecting the influence of Poisson’s ratio, can be up to three times
higher than that of the constant stress of the optimized flywheel.
Substituting the result given in (9.13) into the exponent of the thickness function of the
disk,
2r 2
ρω 2 2 M ω2 2 −
tr = t0 e− = t0 e−
r r 2 +r 2
r2
2σ 2U = t0 e 1 , (9.15)
reveals that the diameters of the rim and the central hole uniquely determine the shape
of the optimized disk as well as the ring areas,
2r 2
r2 + r12 − 2 12
A1 = t0 22 r1 e r2 +r1
r2 − 3r12
2r 2
r2 + r2 − 2 22
A2 = t0 22 12 r2 e r2 +r1 . (9.16)
3r2 − r1

The condition stated by (9.9) can now be written as

√
r2 ≥ 3r1 . (9.17)

Under the constraints of having the same mass and rotational energy as a reference flywheel
of constant thickness, a shape for constant stressing can be found only when the outer
diameter is at least 1.73205 times larger than the diameter of a central hole. Introducing
the ratio α of the two radii
r1 1
α= , 0≤α≤ √ . (9.18)
r2 3
yields more transparent equations for the ring areas and the thickness function of the disk,

1 + α2 2
− 2α 2
A1 = t 0 r1 e 1+α ,
1 − 3α2
1 + α2 − 2
A2 = t 0 2
r2 e 1+α2 ,
3−α
2r/r2
−
t(r) = t0 e 1+α2 . (9.19)

The ring areas depend on the diameter ratio as well as linearly on the absolute size of the
flywheel. The thickness function is independent of the diameter of the flywheel. Thus,
the ratio α uniquely determines the overall shape of the optimized flywheel. In order to
determine the remaining model parameter t0 it is necessary to evaluate one of the integrals
in (9.10) or (9.11), for instance, by using a numerical integration scheme.
Figure 9.10 gives the areas of the two rings and the disk over the full permitted range
of α as (9.18) indicates. The considered initial-shape flywheels have a unit radius and a
thickness of one hundredth.
In the case of very small values of α, the disk and the outer ring A2 dominate the area
of the optimized flywheel. With increasing α, the areas of the inner ring A1 and the disk
quickly increase and decrease, respectively. In the case of values of α higher than 0.3 the
inner ring tends to take up the major portion of the total area. The area of the outer ring
decreases slowly with increasing α. As α approaches the upper limit, the whole design
degenerates in the sense that all the available area moves into the inner ring.
The shapes according to the simplified model are visualized on the left-hand side of Fig.

c ETH Zürich IDMF-CMAS, March 1, 2018

9.3 Flywheel Optimization and Inspired Mechanical Model 177

Figure 9.10: Relative Values of the areas of the inner and outer rings and the disk

Figure 9.11: Optimum shapes for different values of the diameter ratio. Source: [2]

9.11 where ring areas are represented by solid circles. On the right-hand side of Fig.
9.11 the results of the numerical two-dimensional FEM analysis, based on the same input
data, are shown. It appears that the shape results agree well for lower values of α up
to 0.3, which seems remarkable in view of the crudeness of the simplified model. The
main difference in shape is that the extra areas at the inner and the outer rim are more
smoothly distributed according to the FEM model than according to the simplified model.
The two models agree in the extra area at the outer rim remaining fairly constant, the
extra area at the inner rim increasing quickly, and the thickness of the disk decreasing, all
with increasing values of α.
The reason for the degeneration of the agreement between the two models with increasing
values of α becomes obvious when examining the stress plot in Fig. 9.7. The radial stress
has to vanish at the boundaries causing a perturbation to the even stress distribution. The
perturbation is modeled well by the FE method but is not considered by the simplified one-
dimensional model. Thus, the shape plotted in Fig. 9.7 is optimal although the stresses
are not even throughout the domain. Equation (9.13) also implies that the capacity of
an evenly stressed flywheel for storing kinematic energy per unit mass equals the specific
strength of the material,
U σ
= σ∗ = . (9.20)
M ρ

This provides an immediate estimate of the energy storage capacity per unit mass at
optimum shape for any given material.

c ETH Zürich IDMF-CMAS, March 1, 2018

178 Selected Methods and Case Studies

9.3.5 Conclusions on the Flywheel Optimization and Modeling

A procedure based on FEM-analysis for finding a maximum strength design for flywheels
with central bore has been coded in FORTRAN and explained. The program yields satis-
factory results which approximate the objective of even stressing. A simplified analytical
model yields rough shape features of the maximum strength design and elucidates the
underlying essential mechanisms. According to the simplified model, an even-stress shape
design does not exist for bore radii greater than the square root of one-third of the radius
of the flywheel. The optimum-shape predictions of both models agree well if the radius of
the bore does not exceed approximately one third of the radius of the disk. The stress level
in the optimized design is approximately only one third of the maximum circumferential
stress at edge of the bore in a flywheel of constant thickness. The specific energy-storing
capacity of the evenly stressed flywheel according to the closed-form solution of the sim-
plified model equals the specific strength of the material being used.
.

c ETH Zürich IDMF-CMAS, March 1, 2018

9.4 Topology Optimization after Bendsøe and Kikuchi 179

9.4 Topology Optimization after Bendsøe and Kikuchi

Topology optimization of solid structures involves the determination of features such as
the number and location of holes and the connectivity of the domain. The only known
quantities in the problem are the applied loads, the possible support conditions, the vol-
ume or domain shape of structure and some prescribed topological items such as size
and location of prescribed holes. Topology optimization is well established for minimum
compliance design, which method obtains a design proposal for highest structural stiff-
ness under the specified geometric and natural boundary conditions and a specified total
amount of material to be distributed in the physical design space. It is modelled by a
finite-element mesh and the deformation of the given volume is numerically evaluated be
the finite-element method for any kind of geometric boundary conditions and loads. The
distribution of the material is initially constant and then changed so that a topology of
geometrically favorable regions with maximum density and void regions is created. Any
element in the often uniform finite-element mesh maps either material or a void, rendering
the topology optimization problem discrete.
Finding a solution by checking all possible combinations appears impractical since the
number of topologies nT increases exponentially with the number of finite elements n,

nT = 2n , (9.21)

and the sketch in Fig. 9.12 illustrates the equation with n = 4. The moderate-sized sample

H = 0 .0 0 H = 0 .2 5 H = 0 .5 0 H = 0 .7 5 H = 1 .0 0

Figure 9.12: Illustration of topologies for a mesh of 1 by 4 elements

problem given in Fig. 9.14 uses n = 300 finite elements and the number of unconstrained
topologies is nT = 2.037 × 1090 . Of course, a mass, or volume, constraint reduces the
number of feasible topologies drastically. Let n continue to give the mesh size and introduce
m to give the number of elements that must be filled with material due to a specified
average density in the design space. Then the number of combinations is only
n!
nT = (9.22)
m! (n − m)!

and Fig. 9.12 illustrates how, for instance, a mean density ρ = 0.25 reduces the number of
solutions from 16 to 6. The number of combinations for the sample problem with n = 300
and specified material density , giving m = 75, is nT = 9.796 × 1072 . A further constraint
arises from the consideration that the topology must at least connect the boundaries
where displacements are prescribed with those where non-vanishing stresses or forces are
applied. Design solutions violating that requirement are called illegal. A simple formula
for obtaining the legal solutions for any mesh does of course not exist but the illustrative
sample shown in Fig. 9.13 obtains that, under the mass constraint ρ̄ = 0.5, only four of the
924 design solutions after (9.22) are legal. Here it would appear attractive to investigate

c ETH Zürich IDMF-CMAS, March 1, 2018

180 Selected Methods and Case Studies

Figure 9.13: The legal (top) and some illegal (bottom) topologies with 4 by 3 elements

the possibilities of stochastic or even exhaustive search provided that illegal topologies
could be systematically identified and kept from being numerically evaluated.
However, the minimum compliance problem has been proven to be convex and the well-
known homogenization approach of Bendsøe and Kikuchi [12] uses a distributed-function
parameterization rendering the objective function continuous. Each element is equipped
with a variable density function whose minimum value is zero, representing a void, and
whose maximum value represents the density of the of the massive material. Between those
bounds are intermediate states of more or less thinned out material. These intermediate
states are not wanted in the resulting topology, or layout, but accepted in order to enable a
continuous optimization process. The final design should have only material-filled or void
elements and the filled elements present the generated shape of the part to be designed.
Then the best design solution can be found quickly by using some method provided by
mathematical programming.

9.4.1 Topology optimization sample problem

These points are illustrated by the example shown in Fig. 9.14. The rectangular design
space is meshed with 10 times 30 finite elements. The boundary conditions are not shown
in the figure: all nodes on the left vertical part of the boundary are fixed (homogeneous
geometric boundary condition) and a downward oriented concentrated force acts upon
the nodal point at the lower right corner. The compliance of the initial configuration,

(a) (b) (c)

(d) (e) (f)

Figure 9.14: Topology optimization example: initial (a), intermediate (b through e), and
final (f) states for an average density ρ̄ = 0.252

Fig. 9.14(a), is high enough to reveal the deformation of the finite-element mesh with the
highest displacement at the node where the force attacks. The displacement scale factor is

c ETH Zürich IDMF-CMAS, March 1, 2018

9.4 Topology Optimization after Bendsøe and Kikuchi 181

the same for all plots and it can seen that the mesh deformation reduces quickly through
the optimization process. The mass density is presented by gray level. Unfortunately, there
are not enough different gray levels to portray the density differences between elements
in more detail. Nevertheless, it can be seen that already after the first iteration step a
differentiated mass density distribution with concentrations at the corners of the clamped
boundary emerges, Fig. 9.14(b). The state after 20 iterations, Fig. 9.14(b), reminds of a
sandwich design where less strained regions are already thinned out. After 40 iterations
emerges some differentiation in the interior in terms of diagonally arranged members,
Fig. 9.14(d). The intermediate crossing-members design, emerging after 60 iterations and
shown in Fig. 9.14(e), dissolves again in favor of the simple diagonal member pattern of
the final design obtained after 83 iterations and shown in Fig. 9.14(f).

9.4.2 Objective Function and Design Evaluation

The objective of the highest possible structural stiffness is elegantly cast into a global
objective function by minimizing the work W done by the external forces r along the
conjugated displacements ũ caused by the forces:

W = ũT r = min. (9.23)

The load vector r remains constant. The displacements ũ depend on the state of the
design parameters and require the solution of system of equations

Kũ = r (9.24)

assembled by the FEM model. The tilde reminds of the fact that the displacements are
discrete values defined on the nodal points of the finite-element mesh. Then, the objective
function calculation itself requires only to perform the scalar product (9.23).

9.4.3 Parameterization
The structural stiffness depends on the stiffness contributions of the finite elements. The
finite elements do not change their size or shape but their stiffness is controlled by Young’s
modulus E. It depends, see Fig. 9.15 in turn on the density distribution function ρ so

1 .0
N o r m a liz e d Y o u n g 's m o d u lu s

0 .8

0 .6
F
- = -  r
0 .4

0 .2

0 .0
0 .0 0 .2 0 .4 0 .6 0 .8 1 .0
D e n s ity

Figure 9.15: Young’s modulus and density

that
E = E0 ρp , → 0 ≤  ≤ 1, (9.25)

c ETH Zürich IDMF-CMAS, March 1, 2018

182 Selected Methods and Case Studies

where E0 is a reference stiffness value and  is a small but finite number so that ρ is
prevented to become exactly zero for numerical reasons. The exponent in (9.25), usually
chosen from the range 3 ≤ p ≤ 4, is justified by some material modelling considerations
but also effects Young’s modulus moving against its bounds more quickly. In the original
approach, the values of the density functions ρ are directly related to the each finite
element and used as variable design variables. The lower bound signifies a void and the
upper bound signifies a region of massive material extending over the sub-domain of just
one finite element. Thus, there are as many design variables as finite elements. If the
vector x denotes the sought topology result, the density distribution function ρ(x) is

 if x ∈ Ωv
ρ(x) = (9.26)
1 if x ∈ Ωs
where Ωv and Ωs denote the void and solid regions in the design space Ω, respectively. So
we have a typical case of a mesh-dependent parameterization and the number of variable
design parameters, or optimization variables, may become quite high: the relatively simple
example with its course mesh presented in section 9.4.1 already features 300 independent
optimization variables. An evenly spaced mesh, where all finite elements have the same
shape and size, reduces the numerical effort because all element matrices can be traced
back to a reference matrix K0 so that
Kk = Ek K0 (9.27)
and the reference element stiffness matrix needs to be calculated only once. The stiffness
matrix K of the whole structure, appearing in the system of equations (9.24), is thus
assembled from the element matrices by
NEL

ρpk .
X
K = E0 K0 (9.28)
k=1

Fig. illustrates how the nodal point addresses are used to connect the element stiffness
matrix entries with those of the system matrix. Finally it is important to note that the

5 ,6 1 1 ,1 2 1 7 ,1 8 2 3 ,2 4

3 ,4 9 ,1 0 1 5 ,1 6 2 1 ,2 2

1 ,2 7 ,8 1 3 ,1 4 1 9 ,2 0

F in ite E le m e n t M o d e l S y s te m S tiffn e s s M a tr ix

Figure 9.16: Illustration to the assembly of the global stiffness matrix

total mass must be constrained to some initially specified value M0 , or mean density ρ̄ in
the volume V ,
N
X EL
ρi Vi = M0 = ρ̄V , (9.29)
i=1

c ETH Zürich IDMF-CMAS, March 1, 2018

9.4 Topology Optimization after Bendsøe and Kikuchi 183

where Vi is the volume of one finite element, because otherwise the objective of maximum
structural stiffness would be reached by the trivial solution that the design space Ω is
completely filled with material.

9.4.4 Optimization Problem Statement

The topology optimization problem is summarized as

minn W (ρ) hM (ρ) = 0, g+ ≤ 0, g− ≤ 0


(9.30)
ρ∈R

The equality constraining function for constant mass, hM (ρ), is

N
X EL
M
h (ρ) = ρk Vk − M0 = 0 <; . (9.31)
k=1

The side constraints on the design variables ρi are cast into the inequality constraining
functions
gi+ = ρmin − ρi ≤ 0 (9.32)
and
gi− = ρi − 1 ≤ 0 (9.33)

9.4.5 Lagrange Function

The optimization problem with objective and constraining functions can be expressed by
a Lagrange function L(ρ, Λ, λ+ , λ− ):
N
XEL N
X EL
+ −
L(ρ, Λ, λ , λ ) = W (ρ) + Λh (ρ) + M
λ+ +
i gi + λ− −
i gi (9.34)
i=1 i=1

Optimal topology is obtained at the saddle point of the Lagrangian and the derivative of
it with respect to the optimization variables must vanish:

∂L(ρ, Λ, λ+ , λ− )
=0 (9.35)
∂ρi
The next step is to derive the gradients of all parts of the Lagrange function.

c ETH Zürich IDMF-CMAS, March 1, 2018

184 Selected Methods and Case Studies

9.4.6 Gradients with Respect to Densities

The large number of optimization variables of this type of topology optimization prob-
lem calls for a very efficient optimization procedure. The ingenious step of Bendsœ’s and
Kikuchi’s method, the introduction of smooth density functions as optimization variables
into the otherwise discrete-natured problem, allows using one of the methods of mathe-
matical programming.

Objective Function Gradient

Specifically, gradient calculation after the sensitivity formula, described in section 6.10,
is extremely simple and numerically effective as the following derivation [98] shows. The
derivative of the objective function (9.23) with respect to the ith is

∂ũ (T )
 
∂W ∂r
= r + ũ(T ) (9.36)
∂ρi ∂ρi ∂ρi
The sensitivity of the displacement solution ũ to the optimization variable ρi is described
by the sensitivity formula (6.133). Substituting it into (9.36) yields

∂K (T ) (−T )
 
∂W ∂r ∂r
= − ũ K r + ũ(T ) (9.37)
∂ρi ∂ρi ∂ρi ∂ρi
Because of the symmetry of the global stiffness matrix K we have that K(−T ) r = ũ and
therefore (9.36) simplifies to
∂W ∂r ∂K
= 2ũ(T ) − ũ(T ) ũ (9.38)
∂ρi ∂ρi ∂ρi
A change of the density value ρi of the ith finite element influences only the stiffness
of it and not the stiffness of any other element. If it influences, in the case of density-
dependent weight, only the load of the ith finite element and not the load of other elements,
the sensitive of the objective with respect to the variable ρi can be calculated on element
level,
∂W (T ) ∂ri (T ) ∂Ki
= 2ũi − ũi ũi , (9.39)
∂ρi ∂ρi ∂ρi
where ũi and Ki are the displacement vector and stiffness matrix of the ith finite element.
Fixed loads do not depend on the state of ρi and vanish from (9.39). With a reference
stiffness matrix K0 for a unit Young’s modulus E = 1 the stiffness matrix of the ith finite
element is obtained by
Ki = ρ4i K0 . (9.40)
The reference stiffness matrix must be calculated only once if the mesh is a rectangular
array or otherwise regular in a way so that all elements have the same geometry. This
reduces the numerical effort for the assembly of the global stiffness matrix drastically.
Returning to our gradient-calculation problem, we discover that the derivative of the
stiffness matrix can now be obtained analytically,
∂Ki
= 4ρ3i K0 . (9.41)
∂ρi
Last we assume fixed loads and insert (9.41) in the element-level sensitivity equation (9.39):
∂W (T )
= −4ρ3i ũi K0 ũi . (9.42)
∂ρi

c ETH Zürich IDMF-CMAS, March 1, 2018

9.4 Topology Optimization after Bendsøe and Kikuchi 185

Constraining Functions Gradients

From (9.31), (9.32), and (9.33) it follows by derivation that

∂hM (ρ)
= Vi (9.43)
∂ρi

∂gi+ (ρ)
= −1 (9.44)
∂ρi
∂gi− (ρ)
=1 (9.45)
∂ρi

9.4.7 Calculation of Lagrange Factors

With the just calculated gradient information the optimality condition (9.35) can be writ-
ten in detail,
∂L
= −pE0 ρip−1 ũTi K0 ũi + ΛVi − λ+ −
i + λi = 0 , (9.46)
∂ρi
but it remains to calculate the hitherto unknown Lagrange multipliers. From Section 3.5.3
it is known that
(∇W )T ∇h
Λ= (9.47)
(∇h)T ∇h
if all inequality conditions are inactive. The inner product of the gradient of the equality
constraining function is:
V2
(∇h)T ∇h = NEL Vk2 = (9.48)
NEL
The inner product of the gradients of the objective and the equality constraining functions
is:
NEL
V
ρp−1
X
(∇W )T ∇h = −p E0 T
k ũk K0 ũk (9.49)
NEL
k=1

Inserting the results into (9.47) gives, after simplification,

N EL
p
ρp−1
X
T
Λ=− E0 k ũk K0 ũk (9.50)
V
k=1

Substituting the result into (9.46) and taking the negative of the gradient gives a search
direction along which the mass remains the same:
N EL
p
pE0 ρp−1 ρp−1
X
si = i ũTi K0 ũi + E0 T
k ũk K0 ũk (9.51)
V
k=1

Since it has been assumed that the inequality constraints are inactive, it must be discussed
what happens if they become active. Then, since they apply individually to the density
functions, the respective entries in the gradient vectors can be considered constants and
the respective entries in all gradients can be set equal to zero:

ρi ≤ ρmin ∧ ∇Wi > 0 → ρi = ρmin ; ∇hi = 0 . (9.52)

ρi ≥ 1 ∧ ∇Wi < 0 → ρi = 1; ∇hi = 0 . (9.53)

c ETH Zürich IDMF-CMAS, March 1, 2018

186 Selected Methods and Case Studies

9.4.8 A Dual Algorithm for Topology Design

The same topology optimization problem has been visited by Jog [99] and he uses a
dual formulation of the problem for finding a different solution method. We present his
derivations and results but continue to use, where possible, the notation we have become
familiar with. He uses a more general expression for the compliance, or external work:
Z Z Z
W = σ̄ T udΓ − σ T ūdΓ + bT udΩ (9.54)
Γσ Γu ΓΩ

where σ̄ and ū are prescribed natural and geometric boundary conditions, respectively,
and b is distributed body force. The first term on the right-hand-side carries the same
meaning as (9.23), with the only difference being that the latter is expressed with the
discrete forces and displacements of a finite element model. The second term gives the
compliance due to prescribed displacement rather than traction and one wishes, for a stiff
structure, the reaction stress as high as possible which explains the negative sign. The
third term gives the work done by distributed body force such as self-weight (sometimes
also called dead weight). The total potential energy Π of an elastic system is the difference
between the stored elastic deformation energy U and the external work W or

Π=U −W . (9.55)

The principle of the minimum of the total potential energy requires that Π be a minimum
with respect to the displacement. Then, at the equilibrium point, Clapeyron’s theorem
states that Π = −W/2 or
Z Z Z
T
Π(u, ρ) = udΩ − σ̄ udΓ − bT udΩ (9.56)
Ω Γσ Ω

where u denotes the strain energy density function (please distinguish this from the bold
printed displacement) which for a linear elastic material is given by

u = T C (9.57)

where  denotes the linearized strain tensor in vector notation and C the materials law in
matrix notation.
Jog defines the dual problem by the task of finding the Lagrange multiplier Λ associated
with the volume constraint that solves

min L(Λ), Λ>0 (9.58)

Λ

where Z 
L(Λ) = max min Π(u, ρ) − Λ ρdΩ − M0 (9.59)
ρ u Ω
From Section 5.7.4 we recall that dual formulations require the primal problem consist of
separable functions. Jog proceeds with establishing a separable approximation for L(Λ)
and approximates the total potential energy by using the reciprocal variables κ to denote
the field 1/ρ. Inventing the symbols u0 and ρ0 to denote the equilibrium displacement field
corresponding to the design ρ0 , respectively, a first-order approximation for Π by Taylor
series is given by Z  
0 0 0 ∂u 1 1
Π(u , ρ) ≈ Π(u , ρ ) + − dΩ (9.60)
Ω ∂κ κ
0 ρ ρ0

c ETH Zürich IDMF-CMAS, March 1, 2018

9.4 Topology Optimization after Bendsøe and Kikuchi 187

The total potential energy can now be maximized without considering the constant term.
Jog next introduces the discrete representation corresponding to the finite element method.
Then, the Lagrangian can be written as a sum over the separate contributions of each finite
element, or
!
X Z ∂u  1 1
 XZ
L(Λ) = max − 0 dΩ − Λ ρdΩ − M 0 (9.61)
ρ Ωi ∂κ κ0 ρ ρ Ωi
i i

where Omegai is the domain occupied by the ith finite element. Since the density values
are assumed to be constant over each element, the densities can be removed from within
the integrals, or integrals can be replaced by summations. Thus we get
X 2  1 1
Z
∂u

0
L(Λ) = max ρi − dΩ − Λρi Vi + ΛM 0 (9.62)
ρ
i
ρ0i ρi Ωi ∂ρ ρ
i i
0

The maximization is carried out by maximizing each term under the summation sign, or
  Z 
X
02 1 1 ∂u
L(Λ) = max ρi − dΩ − Λρi Vi + ΛM 0 (9.63)
i
ρ ρ0i ρi Ωi ∂ρi ρ 0
i

The density ρi can take on either of the two values (ρmin , ρmax ). If the maximum of the
term in square brackets occurs at the value of (ρmax ), it holds that
 Z  Z
02 1 1 ∂u 02 1 1 ∂u
ρi 0 − dΩ − Λρmax Vi > ρi 0 − dΩ − Λρmin Vi
ρi ρmax Ωi ∂ρi ρi
0 ρi ρmin Ωi ∂ρi ρi
0

(9.64)
and simplification leads to the condition
2
ρ0i
Z
∂u
ρi = ρmax if dΩ > Λρmin Vi (9.65)
ρmax ρmin Ωi ∂ρi ρ0i

The other case gives the condition

2
ρ0i
Z
∂u
ρi = ρmin if dΩ < Λρmin Vi (9.66)
ρmax ρmin Ωi ∂ρi ρ0i

In fact the dual problem has only the one single dual variable Λ which needs to be found for
minimizing the Lagrangian, all other variables follow from the above conditions. The dual
optimization problem requires an iterative procedure. In alternations the displacement
field is solved for given density distribution and the Lagrangian function is then minimized
with respect to the Lagrangian multiplier Λ. A line search is conducted for the Lagrange
multiplier only. Its value is increased if the current mass is higher than the specified mass,
and vice versa. Jog reports difficulty with obtaining topologies close to the optimum
arising from the fact that the linear approximations of the Lagrangian close to a reference
point permit only small changes in the design variables. As a remedy he uses a filtering
technique.

c ETH Zürich IDMF-CMAS, March 1, 2018

188 Selected Methods and Case Studies

9.4.9 Sample Topology Design Problem

One of the sample problems used by Jog [99] is explained with Fig. 9.17. Jog uses Young’s

1 6 c m

5 c m
P = 4 8 0 0 0 N

Figure 9.17: Topology design problem considered by Jog. Source: Jog [99]

modulus and Poisson’s ratio E = 2.1 × 107 and 0.25, the force attacks at the domain
center, and the average mass density is 30 per cent. He utilizes the existing symmetry and
uses for the half-model a mesh of 40 × 25 finite elements with quadratic shape functions
(Q9). His solution, see Fig. 9.18, is obtained with the dual problem and here compared

Figure 9.18: Topology design solution after Jog. Source: Jog [99]

with a solution, see Fig. 9.19 obtained with the demonstration program TOP which is
based on the primal problem. The performance data, as reported by Jog and obtained

Figure 9.19: Topology design solutions with half (left) and full (right) models

with our program TOP, are summarized in Table 9.1. Using a half model with TOP yields

Table 9.1: Topology optimization process performance data

dual primal half primal full

iterations ≈ 200 142 272

compliance 2087 2088 2117
element type Q9 L4 L4

a design with almost the same compliance as the one obtained by Jog; it exhibits more of
the expected symmetries and the number of iterations is even smaller.

c ETH Zürich IDMF-CMAS, March 1, 2018

9.5 Truss Optimization with Ground Structure Approach 189

9.5 Truss Optimization with Ground Structure Approach

The ground structure approach is well established for the optimization of the geometry
and topology of trusses. A number of mathematical formulations for solving the problem
of finding minimum compliance is explained in the review paper [100] and the material
presented in here is selected from it. The layout of a truss structure is found by allowing a
certain set of connections between a fixed set of nodal points as active structural members
or vanishing members. Fig. 9.20 exemplifies a ground structure in two dimensions with

= >

? @

Figure 9.20: Various ground structures

fifteen points and four pre-defined sets of connections with increasing complexity. In case
of the so-called complete ground structure, where each nodal points is connected with all
others and which is illustrated by Fig. 9.20(d), the number of bars, m, increases with the
square of the number of points, n, by

1
m = n (n − 1) . (9.67)
2

Then the number of bars is much higher than the number of degrees-of-freedom of the truss-
structure finite-element model. Please note that with the topology optimization method,
commented in Section 9.4, the numbers of density variables and degrees-of-freedom grow
only linearly with the mesh density. Also, the full ground structure produces a structural
stiffness matrix lacking any sparseness and bandedness. It should therefore not surprise
that, for the two approaches, the respective most efficient problem formulations and solu-
tion methods are not the same.
Obviously, the ground-structure approach is only a sizing problem, where the optimum
combination of the member cross-sectional areas must be found, if the areas must not
be smaller than a lower limit. Permitting zero area values combines the sizing with the
topology optimization problem. If, moreover, the positions of the nodal points of the
structure are allowed to move, shape optimization joins in and the combination of all
three disciplines is called a layout problem.

c ETH Zürich IDMF-CMAS, March 1, 2018

190 Selected Methods and Case Studies

9.5.1 Problem Statement

The problem of finding the minimum compliance truss for a given amount of material is
expressed as

m
X
T
min r u, subject to: ti Ki u = r,
u,t
i=1

m
X
ti = V, ti ≥ 0, i = 1, . . . , m (9.68)
i=1

Here, ti symbolizes the volume of the ith bar, and ti = ai li is introduced to achieve a more
compact notation. In local coordinates, the stiffness matrix of a bar reads
" #
EA 1 −1
K= (9.69)
L −1 1

and consequently the Ki appearing in (9.68) must be normalized with respect to bar
volume: " #
E 1 −1
Ki = 2 (9.70)
li −1 1

If a non-negative lower bound is imposed on the volumes ti , the stiffness matrix remains
positive definite for all ti > 0 and the remaining problem of just adjusting the bar volumes
for maximum stiffness with respect to the defined forces has been shown by Svanberg [101]
to be convex with assured existing solutions.
The zero lower bound on the variables ti implies that bars of the ground structure can be
removed and the problem statement thus covers topology design. It also implies that the
stiffness matrix is not necessarily positive definite and that the displacement solution u can
not simply be removed from the problem by solving the finite-element-method equations.

9.5.2 Problem Statement Extension for Multiple Loads

The problem of finding the minimum compliance design for several load cases can be solved
by treating the problem of minimizing a weighted average of the compliances of each load
case [102]. For a set of M different load cases pk , k = 1, . . . , M , the multiple load
problem can be written as

M m
k kT
X X
k
min w r u , subject to: ti Ki uk = rk ,
u,t
k=1 i=1

m
X
k = 1, . . . , M, ti = V, ti ≥ 0, i = 1, . . . , m . (9.71)
i=1

For convenience the extended displacement vector û = (u1 , . . . , uM ), an extended force

vector r̂ = (w1 r1 , . . . , wM rM ) of the weighted force vectors, and the extended element

c ETH Zürich IDMF-CMAS, March 1, 2018

9.5 Truss Optimization with Ground Structure Approach 191

stiffness matrices as the block diagonal matrices

 1
w Ki


w2 Ki
 
 
K̂i =  (9.72)
 
.. 

 . 

w M Ki
are introduced. This allows writing problem (9.71) as
m
T
X
min r̂k ûk , subject to: ti K̂i ûk = r̂k ,
u,t
i=1

m
X
ti = V, ti ≥ 0, i = 1, . . . , m . (9.73)
i=1

which is of the same form as (9.68).

9.5.3 Problem Statement with Self-Weight Loading

For a truss it can be assumed that the weight of a bar is carried equally by the joints at its
ends. With gi denoting the specific nodal gravitational force vector due to the self-weight
of bar i, the problem of finding the optimal topology takes the form
 !T 
Xm
min rT u + ti gi u ,
u,t
i=1

m
X m
X m
X
subject to: ti Ki u = r + ti gi , ti = V,
i=1 i=1 i=1

ti ≥ 0, i = 1, . . . , m . (9.74)

9.5.4 Fully Stressed Design and Optimality Criteria Methods

This section will cite the derivation of the optimality conditions for the general minimum
compliance problem (9.74) with self-weight and explain how these conditions constitute
the basis for the optimality criteria method for the numerical solution of the general layout
and topology design problem.
In order to obtain the necessary conditions for optimality for problem (9.74) the Lagrange
multipliers ǔ, Λ, and λ for the equilibrium constraint, the volume constraint and the zero
lower bound constraint; respectively, must be introduced. The necessary conditions are
then found as the conditions of stationarity of the Lagrangian L
m
!T m m
!
X X X
L= r+ ti gi u − ǔT ti Ki u − r − ti gi
i=1 i=1 i=1

m m
!
X X
+Λ ti − V + λi (−ti ) . (9.75)
i=1 i=1

c ETH Zürich IDMF-CMAS, March 1, 2018

192 Selected Methods and Case Studies

The conditions of vanishing derivatives,

∂L ∂L
= 0; → =0, (9.76)
∂u ∂t
obtain the necessary conditions
m
X m
X
ti Ki ǔ = r + ti gi
i=1 i=1

m
!
X
ǔT Ki u − 2gi = Λ − λi
i=1

λi ≤ 0; λi ti = 0; i = 1, . . . , m; Λ≥0. (9.77)

Let µ(u) denote the maximal mutual energy with self-weight uT (Ki u − 2gi ) of the indi-
vidual bars, i.e.
µ(u) = max uT (Ki u − 2gi ) |i = 1, . . . , m ,

(9.78)
and let J(u) denote the set of bars for which the mutual energy attains this maximum
level,
J(u) = i|uT (Ki u − 2gi ) = µ(u) .

(9.79)
After defining the non-dimensional element volumes t̃i = ti /V the necessary conditions
are satisfied with

ǔ = u; ti = t̃i V, i ∈ Ju; ti = 0 i ∈
/ J(u); Λ = U (u)

λi = 0, i ∈ J(u); λi = µ(u) − uT (Ki u − 2gi ) , i ∈

/ J(u) , (9.80)

provided that there exists a displacement field u with corresponding set J(u) and non-
dimensional element volumes t̃i , i ∈ J(u), such that
X X X
V t̃i Ki u = r + V t̃i gi ; t̃i = 1 (9.81)
i∈J(u) i∈J(u) i∈J(u)

The reduced optimality conditions (9.81) state that a convex combination of the gradients
of the quadratic functions
 
1 T T
V u Ki u − gi u , i ∈ J(u) , (9.82)
2

equals the load vector r.

The following derivations, copied from [103], show that a pair (u, t) exists which is a
solution to (9.81) which implies that there exists an optimal truss having bars with constant
mutual energies and the set J(u) is the set of these active bars. Because of these properties
the solution is labelled fully stressed design. The proof shows that the assumed existing
optimum design (u, t) has smaller compliance uT r than any other design (v, s). The proof
also utilizes that the total potential energy of an elastic system,

Π=U −W , (9.83)

c ETH Zürich IDMF-CMAS, March 1, 2018

9.5 Truss Optimization with Ground Structure Approach 193

7 2 9

K~ 0 K~

Figure 9.21: Total potential energy visualization for only one dependent variable

at equilibrium equals one half of the negative of the external work or

W = −2Π (9.84)

which is visualized in Fig. 9.21. Here, we identify the external work W and the total
potential energy Π with
m
!T m
!
X 1 T X
W = r+ ti gi u, Π= u ti Ki u − 2r − 2 ti gi (9.85)
2
i=1 i=1

Substituting the expressions (9.85) into the equilibrium condition (9.84) gives
m
!T m
!T m
X X X
r+ ti gi u=2 r+ ti gi u− ti uT Ki u (9.86)
i=1 i=1 i=1

Re-ordering of the terms appearing on the right-hand-side of (9.84) produces an expression

that is identified with the maximal mutual energy (9.78) found in the active bars of the
fully stressed design:
m
X m
X
T T T
2r u − ti u (Ki u − 2gi ) = 2r u − ti µ(u) (9.87)
i=1 i=1

Since the maximal mutual energy µ(u) is a constant, the sum over the volumes of all bars
can be replaced with the total volume V of the truss. Changing the volumes (which is
traced back to changing the cross-sectional areas) gives a different truss design si but let
the maximum mutual energy µ(u) continue to correspond to the fully-stressed design ti .
These thoughts give
Xm
2rT u − V µ(u) = 2rT u − si µ(u) (9.88)
i=1
However, the other design si may have different bar volumes only for the active bars but
generally it must be assumed that volumes are assigned to bars which are inactive in the
fully stressed design ti . The mutual energy of the inactive bars is smaller than µ(u) and
therefore it holds that
m
X m
X
2rT u − si µ(u) ≤ 2rT u − si uT (Ki u − 2gi ) (9.89)
i=1 i=1

Next observe that the design si can not be at equilibrium with the here assumed dis-
placements corresponding with the fully stressed design ti . The extremum principle for

c ETH Zürich IDMF-CMAS, March 1, 2018

194 Selected Methods and Case Studies

equilibrium requires that Π be a minimum or, by (9.84), W be maximum. Let the maxi-
mum of W for design si be found by adjusting a variable displacement vector w:
 !T 
m m m
X X 1 X
2rT u − si uT (Ki u − 2gi ) ≤ 2 max  r + si gi w− si wT Ki w (9.90)
w 2
i=1 i=1 i=1

Practically the equilibrium adjustment of w is found by simply solving the elasticity prob-
lem for given design si for the displacements which are here here labelled v. Using (9.84)
again identifies

m
!T m m
!T
X X X
2 r+ s i gi v− si vT Ki v = r+ s i gi v (9.91)
i=1 i=1 i=1

The sequence (9.87) through (9.91) of equalities and inequalities proves that

m
!T m
!T
X X
r+ ti gi u≤ r+ si gi v (9.92)
i=1 i=1

which says that the best design solution is the fully-stressed design.
The optimality criterion (9.81) with its implicated fully stressed design solution allows
devising a very simple and effective search method which is known in the literature [104,
105] as optimality criterion method. The iterative method assign in each step volumes to
the bars proportionally to the respective mutual energies in order to the state of constant
mutual energy in the active bars. At iteration step k the following computations are
performed:

- the current intermediate design tk−1

i is given

- compute displacements uk−1 by solving the truss system equations

 
- assign preliminary bar volumes ζik = max tk−1 uk−1T K uk−1 , t
i i min

Pm
- find actual volumes satisfying volume constraint V k = k tki = ζik V /V k


i=1 ζi ,

c ETH Zürich IDMF-CMAS, March 1, 2018

Chapter 10

Demonstration Programs

Two demonstration programs are used for exercises and are available to the students. The
data input files of both programs are explained in the following sections.

10.1 Program DEMO OPT

The various methods of mathematical programming are implemented in DEMO OPT to
solve a few typical test functions in two-dimensional variables space. The students can ob-
serve the search paths taken and the number of iterations and function evaluations needed
to reach the respective optimum and thus obtain some feeling for the advantages and lim-
itations of the various methods. The most important educational goal is to understand
that the performance of any method depends very much on the nature of the objective
function.
The input data file is shown in Fig. 10.1. It can be seen from it that there are five dif-
C H O O S E T E S T F U N C T I O N : 5

1 : Q U A D R A T I C P O L Y N O M I A L
2 : H I M M E L B L A U F U N C T I O N
3 : R O S E N B R O C K F U N C T I O N
4 : F E N T O N E A S O N F U N C T I O N
5 : C A N T I L E V E R B E A M P R O B L E M

N O I S E : F R E Q U E N C Y A N D A M P L I T U D E 0 . 0 0 E + 0 0 0 . 0 0 E + 0 0

D E F I N E V I E W I N G R E G I O N : X M I N X M A X Y M I N Y M A X
2 . 0 0 3 5 . 0 0 2 . 0 0 2 5 . 0 0

S T A R T V A L U E S : X 0 Y 0 S I Z E
5 . 0 0 0 1 0 . 0 0 0 0 . 1 0 E + 0 0

S E L E C T O N E O F T H E M E T H O D S L I S T E D B E L O W : 4

1 S I M P L E X S E A R C H
2 R E S P O N S E S U R F A C E : S P E C I F I Y V E R S I O N 0 - 1 1
3 P O W E L L S M E T H O D
4 S T E E P E S T D E S C E N T ( E X A C T L I N E S E A R C H E S ) 1
5 F L E T C H E R - R E E V E S
6 N E W T O N M E T H O D
7 M O D I F I E D N E W T O N M E T H O D

F O R W A R D ( 0 ) O R C E N T R A L ( 1 ) D I F F E R E N C E S M E T H O D 0

N U M B E R A N D D I S T R I B U T I O N O F C O N T O U R L I N E S : 1 0 0 8

D E M O N S T R A T I O N T I M E S P A N : 1

Figure 10.1: DEMO OPT input data file

c ETH Zürich IDMF-CMAS, March 1, 2018

196 Demonstration Programs

ferent test functions, where the functions 2 through 4 often appear in the literature. The
user defines a viewing region and choices of the corner coordinates are suggested in table
10.1. The table suggests also starting point coordinates and the number and distribution

Table 10.1: Topology optimization process performance data

xmin xmax ymin ymax number exp xS ys

quadratic polynomial
Himmelblau −6.0 6.0 −6.0 6.0 100 4
Rosenbrock −3.0 3.0 −4.0 2.0 100 4 −1.2 1.0
Fenton-Eason 0.1 3.0 0.1 3.0 100 16 0.5 0.5
Cantilver Beam 1.0 40.0 1.0 30.0 100 16 2.0 2.0

of contour lines. According to the suggested values DEMO OPT produces a series of plots
showing each iteration of the optimization process and Fig. 10.2 shows the respective last

 #

! "

Figure 10.2: DEMO OPT input data file

images. The parameter size has an effect on the initial size of the simplex, the initial lat-
tice spacing of the supporting point set of the response-surface method, and takes on the
meaning of a fixed step length when these are combined with choosing steepest-descend
search directions. Choosing the central differences method over the forward differences
method tends to increase computation time. The parameter demonstration time span can
be adjusted to view the optimization process as quickly or slowly as desired.

c ETH Zürich IDMF-CMAS, March 1, 2018

10.2 Program TOP 197

10.2 Program TOP

Topology optimization after the homogenization method pioneered by Bendsøe and
Kikuchi [12] is implemented in TOP. The making of the new version of the program
has been motivated by the ETH project NOVA ([Link] where the
sequence of improving design solutions, Ein perfektes Tragwerk can be seen on a three-
dimensional display. The program supports the student to creatively set up their own
topology design problems and find solutions.
Fig. 10.3 shows a design solution and corresponding data file the meaning of which is illus-
trated by the sketches. The numbers within the rectangle signify the number of elements

L E N G T H A N D H E I G H T O F D E S I G N S P A C E : X L E N G T H Y H E I G H T Y H E IG H T

1 0 . 0 1 0 . 0
X L E N G T H
N O . O F E L E M E N T S I N X A N D Y D I R E C T I O N S : N E L X N E L Y

1 5 0 1 5 0

S U P P . E D G E S ( 0 = F R E E , 1 = X , 2 = Y , 3 = B O T H ) : 0 0 0 0 N E L Y

S U P P . C O R N E R S ( 0 = F R E E , 1 = X , 2 = Y , 3 = B O T H ) : 3 3 3 3

E D G E T R A C T I O N S : 0 . 0 0 0 0 . 0 0 0 1
N E L X
E D G E T R A C T I O N S : 0 . 0 0 0 0 . 0 0 0 1
E D G E T R A C T I O N S : 0 . 0 0 0 0 . 0 0 0 1 E D G E 3
E D G E T R A C T I O N S : 0 . 0 0 0 0 . 0 0 0 1
E D G E 4 E D G E 2
C O R N E R F O R C E S : 1 . 0 0 0 - 1 . 0 0 0 7 2
C O R N E R F O R C E S : 1 . 0 0 0 1 . 0 0 0 7 2
C O R N E R F O R C E S : - 1 . 0 0 0 1 . 0 0 0 7 2 E D G E 1
C O R N E R F O R C E S : - 1 . 0 0 0 - 1 . 0 0 0 7 2

M I D S I D E F O R C E S : 0 . 0 0 0 0 . 0 0 0 1 C O R N E R 4 C O R N E R 3
M I D S I D E F O R C E S : 0 . 0 0 0 0 . 0 0 0 1
M I D S I D E F O R C E S : 0 . 0 0 0 0 . 0 0 0 1
M I D S I D E F O R C E S : 0 . 0 0 0 0 . 0 0 0 1

G R A V I T Y 0 . 0 0 0 C O R N E R 1 C O R N E R 2

A V E R A G E D E N S I T Y 0 . 3 2 0
E X P O N E N T P 4 . M ID S ID E 3
T O L E R A N C E R A N G E [ % ] 2 . 0
F I L T E R O F F O R O N ? 0 O R 1 1
N U M B E R O F I T E R A T I O N S 1 0 0 0 M ID S ID E 4 M ID S ID E 2

R E A D M O V I E F I L E ? 0 O R 1 0 
M ID S ID E 1

Figure 10.3: TOP input data file and explaining sketches

by which the respective load introduction points are moved away from the boundary into
the domain. Rather time consuming optimization processes, for which the presented data
set is an example, can be viewed at much higher speed after the end of the computations.
Then, the encircled number must be set to 1 and the intermediate topology information
is read from a data file and plotted.

c ETH Zürich IDMF-CMAS, March 1, 2018

198 Demonstration Programs

c ETH Zürich IDMF-CMAS, March 1, 2018

List of Figures

1.1 Types of Structural Optimization . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Ducati 996R frame and geometry model . . . . . . . . . . . . . . . . . . . . 4
1.3 CAD-Model of the racing car rim . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Flywheel with central bore, 2-dimensional FEM model and stresses . . . . . 6
1.5 Onsert design demonstrator and geometry model . . . . . . . . . . . . . . . 7
1.6 ANSYS model of the sail boat hull with composite material patches . . . . 7
1.7 Conceptual model of a fuel cell stack. . . . . . . . . . . . . . . . . . . . . . 8

2.1 Three-columns concept after Eschenauer . . . . . . . . . . . . . . . . . . . . 9

2.2 DynOPS evaluator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3.1 Mapping of a feasible design space into the criteria space . . . . . . . . . . . 17

3.2 Examples for the exterior and interior penalty functions . . . . . . . . . . . 19
3.3 Illustration to the unconstrained Lagrange problem formulation . . . . . . . 21
3.4 Fitness function for a design objective . . . . . . . . . . . . . . . . . . . . . 24
3.5 Upper limit constraint penalty functions . . . . . . . . . . . . . . . . . . . . 26
3.6 Penalty functions for a target constraint . . . . . . . . . . . . . . . . . . . . 26
3.7 FEM model of the frame and load cases . . . . . . . . . . . . . . . . . . . . 28
3.8 Rim compliance measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.9 Loading of the Rim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.10 Region out of the Flywheel Analysis Model . . . . . . . . . . . . . . . . . . 32
3.11 ANSYS model of the sail boat hull with composite material patches . . . . 33
3.12 Quarter end-plate model in the initial design . . . . . . . . . . . . . . . . . 34

4.1 Classification of design optimization problems for truss-like structures . . . 37

4.2 Standard Tube Sizes after DIN 2394 . . . . . . . . . . . . . . . . . . . . . . 39
4.3 FEM-Model and Parameterization of the Motorcycle Frame . . . . . . . . . 39
4.4 Structure of the CAD model of the rim. . . . . . . . . . . . . . . . . . . . . 40
4.5 Optimization variables for spoke-body and bed contour . . . . . . . . . . . 41
4.6 Optimization variables for the pockets . . . . . . . . . . . . . . . . . . . . . 41
4.7 Relation between nodal point coordinates and shape parameter . . . . . . . 42
4.8 Definition of shape parameters . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.9 Final Design and Mesh Distortion . . . . . . . . . . . . . . . . . . . . . . . 43
4.10 Patch Pattern and Laminate . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.11 CAD features: lower plate, upper plate and a rib . . . . . . . . . . . . . . . 45
4.12 Sample problems and parameterization spectrum . . . . . . . . . . . . . . . 46
4.13 Jagged onsert shape obtained with mesh-dependent analysis variables . . . 48

5.1 Convex and concave functions . . . . . . . . . . . . . . . . . . . . . . . . . . 52

c ETH Zürich IDMF-CMAS, March 1, 2018

200 LIST OF FIGURES

5.2 Convex design space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

5.3 Usable and feasible sectors of a search space . . . . . . . . . . . . . . . . . . 53
5.4 Sample problem Lagrangian versus x and versus λ . . . . . . . . . . . . . . 56
5.5 Illustrations to Kuhn-Tucker conditions for constrained optima . . . . . . . 58
5.6 Design Optimization Iteration Basic Scheme . . . . . . . . . . . . . . . . . . 63
5.7 Overview of Optimization Algorithms . . . . . . . . . . . . . . . . . . . . . 64
5.8 Search methods arranged after model and method orders . . . . . . . . . . 64
5.9 Objective function and hypermesh in two dimensions . . . . . . . . . . . . . 65

6.1 Construction of new simplex . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

6.2 Design Principle and Reflection of a Simplex in Two Dimensions . . . . . . 72
6.3 Cauchy-method zigzag path through two-dimensional search space . . . . . 73
6.4 Orthogonal and Conjugate Search Directions . . . . . . . . . . . . . . . . . 79
6.5 Conjugacy in two dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.6 Powell’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.7 Path Taken by Powell’s Method . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.8 Placement of supporting points in 3-dimensional variables space . . . . . . . 89
6.9 Updated supporting point sets around successful minimum point estimates 92
6.10 Shrinking of supporting point region around the reference point . . . . . . . 93
6.11 Supporting point set shifted . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.12 Number of coinciding points versus number of variables . . . . . . . . . . . 94
6.13 Points set size and reduction due to coinciding points . . . . . . . . . . . . 95
6.14 Relative computing time savings due to coinciding points . . . . . . . . . . 95
6.15 Latin Hypercube Sampling in a Two-Dimensional Design Space . . . . . . . 96
6.16 Sketch of the interval-halving algorithm . . . . . . . . . . . . . . . . . . . . 100
6.17 Sketch of the golden-section algorithm . . . . . . . . . . . . . . . . . . . . . 100
6.18 Approximation with a quadratic polynomial and three supporting points . . 102
6.19 Approximation with a cubic polynomial and two supporting points . . . . . 103
6.20 Golden section method supporting points . . . . . . . . . . . . . . . . . . . 107
6.21 Narrowing the bracket by reduction of four supporting points to three . . . 107
6.22 Quadratic approximation incorrectly indicated minimum . . . . . . . . . . . 107
6.23 Updated four-point set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.24 Minimum point successfully estimated by quadratic approximation . . . . . 109
6.25 Two-dimensional variable space with two linear constraining functions . . . 111
6.26 Two-dimensional variable space with two linear constraining functions . . . 112
6.27 Shape optimization and sensitivity formula for gradient calculation . . . . . 116
6.28 Function with minima in considered region and starting-point mesh. . . . . 117
6.29 Function with objective and unusable and usable tunneling functions . . . . 119

7.1 Metropolis Algorithm: Probability for the acceptance of a new solution . . 125
7.2 Selected mathematical programming methods and ordering scheme . . . . . 126
7.3 General architecture of an evolutionary algorithm. . . . . . . . . . . . . . . 129

8.1 Examples of applications with shell structures . . . . . . . . . . . . . . . . . 136

8.2 Laminated composites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.3 Material principal and global coordinate systems . . . . . . . . . . . . . . . 139
8.4 General layup of a laminated composite material . . . . . . . . . . . . . . . 141
8.5 Front view of different modeling techniques for bending . . . . . . . . . . . 143
8.6 Layered shell element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
8.7 Schematic illustration of one half of a symmetric laminate . . . . . . . . . . 148

c ETH Zürich IDMF-CMAS, March 1, 2018

LIST OF FIGURES 201

8.8 Effect of a thickness change to the overlaying layers . . . . . . . . . . . . . . 148

8.9 Feasible domain of the in-plane lamination parameters . . . . . . . . . . . . 152
8.10 Flowchart of a typical optimization with an FEM-Solver . . . . . . . . . . . 155
8.11 Fiber orientation optimization . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.12 Geometry and objective of the tensile specimen experiment . . . . . . . . . 156
8.13 Stacking sequence optimization . . . . . . . . . . . . . . . . . . . . . . . . . 157
8.14 Layer thickness optimization . . . . . . . . . . . . . . . . . . . . . . . . . . 159
8.15 Optimization with variable number of layers . . . . . . . . . . . . . . . . . . 159
8.16 Homogeneous loads: Tension and shear . . . . . . . . . . . . . . . . . . . . . 161
8.17 Inhomogeneous strain field of a plate with centered hole . . . . . . . . . . . 161
8.18 Global domain Ω split with subdomains Ωi . . . . . . . . . . . . . . . . . . 161
8.19 Schematic illustration of layers covering multiple sub-domains . . . . . . . . 162
8.20 FEM-based Parametrization of Sub-Domains . . . . . . . . . . . . . . . . . 162
8.21 Graph-based parametrized vertex patches . . . . . . . . . . . . . . . . . . . 163
8.22 Sensitivities and local reinforcements of a plate with eigenfrequency band . 164
8.23 Connection between global layers and laminate regions . . . . . . . . . . . . 165

9.1 Part with stress raising geometry and modelled cambium . . . . . . . . . . 167
9.2 Two-phase CAO process after Mattheck . . . . . . . . . . . . . . . . . . . . 168
9.3 Soft-Kill Option process after Mattheck . . . . . . . . . . . . . . . . . . . . 169
9.4 Initial flywheel shape and radial and circumferential stress distributions . . 171
9.5 Blueprint of a turbine disk design solution after Stodola . . . . . . . . . . . 171
9.6 Parameterization of thickness and mesh generator . . . . . . . . . . . . . . 172
9.7 Shape optimized after maximum stress criterion and stress distributions . . 173
9.8 Shape optimized after yield stress criterion and stress distributions . . . . . 174
9.9 Mechanical model consisting of Stodola’s disk and two discrete rings . . . . 174
9.10 Relative Values of the areas of the inner and outer rings and the disk . . . . 177
9.11 Optimum shapes for different values of the diameter ratio . . . . . . . . . . 177
9.12 Illustration of topologies for a mesh of 1 by 4 elements . . . . . . . . . . . . 179
9.13 The legal and some illegal topologies with 4 by 3 elements . . . . . . . . . . 180
9.14 Topology optimization example . . . . . . . . . . . . . . . . . . . . . . . . . 180
9.15 Young’s modulus and density . . . . . . . . . . . . . . . . . . . . . . . . . . 181
9.16 Illustration to the assembly of the global stiffness matrix . . . . . . . . . . . 182
9.17 Topology design problem considered by Jog . . . . . . . . . . . . . . . . . . 188
9.18 Topology design solution after Jog . . . . . . . . . . . . . . . . . . . . . . . 188
9.19 Topology design solutions with half and full models . . . . . . . . . . . . . . 188
9.20 Various ground structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
9.21 Total potential energy visualization for only one dependent variable . . . . 193

10.1 DEMO OPT input data file . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

10.2 DEMO OPT input data file . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
10.3 TOP input data file and explaining sketches . . . . . . . . . . . . . . . . . . 197

A.1 Loads and kinematics of a Kirchhoff-plate . . . . . . . . . . . . . . . . . . . 217

c ETH Zürich IDMF-CMAS, March 1, 2018

202 LIST OF FIGURES

c ETH Zürich IDMF-CMAS, March 1, 2018

List of Tables

3.1 General types of fitness functions with defining parameters. . . . . . . . . . 22

3.2 Problems and solution methods. . . . . . . . . . . . . . . . . . . . . . . . . 27

4.1 Ranges and step sizes for the optimization variables. . . . . . . . . . . . . . 46

6.1 Supporting points in terms of changes with respect to x(1) . . . . . . . . . . 90

6.2 Number of coinciding points for the different reference point positions . . . 94

9.1 Topology optimization process performance data . . . . . . . . . . . . . . . 188

10.1 Topology optimization process performance data . . . . . . . . . . . . . . . 196

c ETH Zürich IDMF-CMAS, March 1, 2018

204 LIST OF TABLES

c ETH Zürich IDMF-CMAS, March 1, 2018

Bibliography

[1] Müller, S.D. Bio-inspired optimization algorithms for engineering applications. Dis-
sertation ETH No. 14719, Swiss Federal Institute of Technology Zurich, 2002.

[2] Kress, G.R. Shape optimization of a flyweel. Struct. Multidisc. Optim., 19:74–81,
2000.

[3] König, O. M. Wintermantel, U. Fasel, and P. Ermanni. Using evolutionary methods

for design optimization of a tubular steel trellis motorbike frame. Journal article
under preparation.

[4] Wintermantel, M. and O. König. Design-entity based structural optimization with

evolutionary algorithms. Project performed at ETH, Structure Technologies, sup-
ported by Swiss National Science Foundation grant no. 21-66879.01, 4/1/2002 -
9/30/2004.

[5] König, O. Evolutionary design optimization: Tools and applications. Dissertation

ETH No. 15???, Swiss Federal Institute of Technology Zurich, 2004.

[6] Wintermantfel, M. Design-encoding for evolutionary algorithms in the field of struc-

tural optimization. Dissertation ETH No. 15323, Swiss Federal Institute of Technol-
ogy Zurich, 2004.

[7] Kress, G., P. Naeff, M. Niedermeier, and P. Ermanni. Onsert strength design. Int.
J. of Adhesion & Adhesives, 24:201–209, 2004.

[8] Kress, G. and P. Ermanni. The onsert: A new joining technology for sandwich
structures. In Proc. International Conference on Buckling and Postbuckling Behavior
of Composite Laminated Shell Structures, Eilat, Israel, 2004.

[9] Zehnder, B. and P. Ermanni. A methodology for the global optimization of laminated
composite structures. Composite Structures, 72(3):311–320, 2006.

[10] Zehnder, B. and P. Ermanni. Optimizing the shape and placement of patches of
reinforcement fibers. Composite Structures, 77(1):1–9, 2007.

[11] Gätzi, R., M. Uebersax, O. König. Structural optimization tool using genetic al-
gorithms and ansys. Proc. 18. CAD-FEM User’s Meeting, Internationale FEM-
Technologietage, Graf-Zeppelin-Haus, Friedrichshafen, 2000.

[12] Bendsœ, M.P., N. Kikuchi. Generating optimal topologies in structural design using a
homogenization method. Computer Methods in Applied Mechanics and Engineering,
71:197–224, 1988.

c ETH Zürich IDMF-CMAS, March 1, 2018

206 BIBLIOGRAPHY

[13] Michell, A.G.M. The limits of economy of material in frame-structures. Phil. Mag.
(Series 6), 8:589–597, 1904.
[14] N. Zehnder. Global Optimization of Laminated Structures. PhD thesis, ETH Zurich,
2008. DISS. ETH. NO. 17573.
[15] Ruge, M. Entwicklung eines flüssigkeitsgekühlten Polymer-Elektrolyt-Membran-
Brennstoffzellenstapels mit einer Leistung von 6.5 kW. Ph.D thesis, Swiss Federal
Institute of Technology, VDI Verlag GmbH, Reihe 6, Nr. 494, Fortschrittberichte
VDI, 2003.
[16] Schmid, D. Entwicklung eines Brennstoffzellenstapels für portable Aggregate unter-
schiedlicher Leistungsbereiche. Ph.D thesis, Swiss Federal Institute of Technology,
VDI Verlag GmbH, Reihe 6, Nr. 500, Fortschrittberichte VDI, 2003.
[17] Evertz, J. and M. Günthart. Structural concepts for lightweight and cost-effective
end plates for fuel cell stacks. In In 2nd European PEFC Forum, Lucerne, Switzer-
land, 2003.
[18] H. Eschenauer, H., N. Olhoff, W. Schnell. Applied Structural Mechanics. Springer,
1997.
[19] Courant, R. Variational methods for the solution of problems of equilibrium and
vibrations. Bull. Am. Math. Soc., pages 1–23, 1943.
[20] Vanderplaats, G.N. Numerical Optimization Techniques for Engineering Design:
with Applications. McGraw-Hill: Series in Mechanical Engineering, 1984.
[21] Schuldt, S.B. A method of multipliers for mathematical programming methods with
equality and inequality constraints. JOTA, 17:155–162, 1975.
[22] Schuldt, S.B., G.A. Gabriele, R.R. Root, E. Sandgren, and K.M. Ragsdell. Applica-
tion of a new penalty function method to design optimization. ASME J. Eng. Ind.,
99:31–36, 1977.
[23] Reklaitis, G.V., A. Ravindran, K.M. Ragsdell. Engineering Optimization - Methods
and Applications. John Woley and Sons, 1983.
[24] Stodola, A. Dampf- und Gasturbinen. Springer, 1924.
[25] Jones, R.M. Mechanics of Composite Materials. Hemisphere Publishing Corporation,
1975.
[26] Schmit, L.A., R.H. Mallet. Structural synthesis and design parameters. Hierarchy
Journal of the Struct. Division, Proceedings of the ASCE, 89:269–299, 1963.
[27] Olhoff, N., J.E. Taylor. On structural optimization. J. Applied Mechanics, 50:1139–
1151, 1983.
[28] Fritsche, D. Lasteinleitungselemente maximaler verbindungsfestigkeit f/”ur sand-
wichbauteile. Student project thesis no. winter semester 2000/2001, Structure Tech-
nologies, Institute of Mechanical Systems, ETH Zürich.
[29] Naeff, P. Experimentelle verifikation strukturell geklebter lasteinleitungselemente.
Student project thesis no. 02-115 summer semester 2002, Structure Technologies,
Institute of Mechanical Systems, ETH Zürich.

c ETH Zürich IDMF-CMAS, March 1, 2018

BIBLIOGRAPHY 207

[30] Heer, A. Gestaltung von lasteinleitungslementen f/”ur sandwichbauteile. Student

project thesis summer semester 2000, Structure Technologies, Institute of Mechanical
Systems, ETH Zürich.

[31] Kuhn, H.W., A.W. Tucker. Nonlinear programming. Proc. Second Berkeley Symp.
on Math. Statist. and Probability, pages 481–492, 1951. J. Neyman, Ed.

[32] Schumacher, A. Optimierung mechanischer Strukturen. Springer, 2005.

[33] Reklaidis, G.V., A. Ravindran, K.M. Ragsdell, A. Engineering Optimization: Meth-

ods and Applications. Wiley Interscience, 1983.

[34] Spellucci, P. Numerische Verfahren der nichtlinearen Optimierung. Birkhä user,

1993.

[35] Powell, M.J.D. An efficient method for finding the minimum of a function of several
variables without calculating derivatives. Computer J., 7:155–162, 1964.

[36] Cauchy, A. Method generale pour la resolution des systemes d’equations simultanees.
Compt. Rend. Acad. Sci., 25:536–538, 1847.

[37] Spendley, W., G.R. Hext, F.R. Himsworth. Sequential application of of simplex
designs in optimization and evolutionary operation. Technometrics, 4:441–461, 1962.

[38] Shewchuck, J.R. An introduction to the conjugate gradient method without the
agonizing pain. School of Computer Science, Carnegie Mellon University, Pittsburgh,
PA, 15213, Aug. 4, 1994.

[39] Fletcher, R., C.M. Reeves. Function minimization by conjugate gradients. Computer
J., 7(5):149–154, 1964.

[40] Polak, E., G. Ribière. Note sur la convergence de méthodes de directions conjuguées.
Revue française d’informatique opérationelle, série rouge, 3(1):35–43, 1969.

[41] Navon, I.M., D.M. Legler. Conjugate-gradient methods for large-scale minimization
in meteorology. Mon. Wea. Rev., 115:1479–1502, 1987.

[42] Zangwill, W.I. Minimizing a function without calculating derivatives. Computer J.,
10:293–296, 1967.

[43] Brent, R.P. Algorithms for Minimization without Derivatives. Prentice-Hall, Engle-
wood Cliffs, NJ, 1973.

[44] Venter, G. Non-dimensional response surfaces for structural optimization with un-
certainty. Dissertation, University of Florida, 1998.

[45] Kress, G. and P. Ermanni. Comparison between newton and response surface meth-
ods. Struct. Multidisc. Optim., accepted for publication, 2004.

[46] Wang, G.G. Adaptive response surface method using inherited latin hypercube
design points. Journal of Mechanical Design, 125:210–220, 2004.

[47] McKay, M.D., R.J. Bechmann, and W.J. Conover. A comparison of three methods
for selecting values of input variables in the analysis of output from a computer code.
Technometrics, 21(2):239–245, 1997.

c ETH Zürich IDMF-CMAS, March 1, 2018

208 BIBLIOGRAPHY

[48] Levy, A.V., S. Gomez. The tunneling method applied to global optimization. Proc.
SIAM Conf. on Num. Optimization, pages 213–244, 1984.

[49] James C. Spall. Introduction to stochastic search and optimization : estimation,

simulation, and control. Wiley-Interscience series in discrete mathematics and opti-
mization, 2003.

[50] D.H. Wolpert and W.G. Macready. No free lunch theorems for optimization,. IEEE
Transactions on Evolutionary Computation, 1:67–82, 1997.

[51] Bäck, Th., U. Hammel, H.-P. Schwefel. Evolutionary computation: Comments on

the history and current state. IEEE Transactions on Evolutionary Computation,
1(1), 1997.

[52] Fogel, L.J., A.J. Owens, A.J. Walsh. Artificial Intelligence Through Simulated Evo-
lution. Wiley, New York, 1966.

[53] Holland, J.H. Nonlinear environments permitting efficient adaptation. Computer

and Information Sciences II, Acadameic Press, New York, 1967.

[54] Holland, J.H. Adaptation in Natural and Artificial Systems. Michigan Press, Ann
Arbor, MI, 1975.

[55] J.H. Holland. Adaptation in Natural and Artificial Systems. University Michigen
Press, Ann Arbor, MI, 1975.

[56] D.E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning.
Addison-Wesley Publishing Company, Inc., 1989.

[57] P. Hajela. Stochastic search in structural optimization: Genetic algorithms and

simulated annealing. In Structural Optimization: Status and Promise, volume 150,
pages 611–637. Progress in Astronautics and Aeronautics, 1992.

[58] L. J. Fogel. Biotechnology: Concepts and Applications. Prentice Hall, Englewood

Cliffs, NJ, 1963.

[59] D. Fogel. Evolving Artificial Intelligence. PhD thesis, University of California, San
Diego, CA, 1992.

[60] I. Rechenberg. Evolutionsstrategie: Optimierung Technischer Systeme nach Prinzip-

ien der Biologischen Evolution. Frommann-Holzboog, Stuttgart, 1973.

[61] T. Bäck. Evolutionary Algorithms in Theory and Practice. Oxford Univerity Press,
New York, 1996.

[62] J. Koza. Genetic Programming: On the Programming of Computers by Means of

Natural Selection. MIT Press, Cambridge, MA, 1992.

[63] Peter J. Bentley, editor. Evolutionary Design by Computers. Morgan Kaufmann

Publishers, Inc, San Francisco California, 1999.

[64] M. Schoenauer and Z. Michalewicz. Evolutionary computation: An introduction.

Control and Cybernetics Special Issue on Evolutionary Computation, 26(3):307–338,
1997.

c ETH Zürich IDMF-CMAS, March 1, 2018

BIBLIOGRAPHY 209

[65] N.J. Radcliffe and P.D. Surry. Formal memetic algorithms. Technical report, Edin-
burgh Parallel Computing Centre, 1994.

[66] B. Schläpfer. Optimal Design of Laminated Structures with Local Reinforcements.

PhD thesis, ETH Zurich, 2013. DISS. ETH. NO. 20894.

[67] J. E. Gordon. The new science of strong materials: or, Why you don’t fall through
the floor. Pelican books. Penguin, 1991.

[68] R. C. Reuter. Concise property transformation relations for an anisotropic lamina.

Journal of Composite Materials, 5(2):270–272, 1971.

[69] R. D. Cook, D. S. Malkus, M. E. Plesha, and R. J. Witt. Concepts and applications

of finite element analysis. John Wiley & Sons Inc, 4 edition, 2001.

[70] S. W Tsai and N. J. Pagano. Invariant properties of composite materials. Technical

report, Air Force Materials Laboratory, Wright-Patterson Air Force Base, Ohio,
AFB, 1968.

[71] H. Fukunaga. On laminate configurations for simultaneous failure. Journal of Com-

posite Materials, 22(3):271–286, 1988.

[72] J. L. Grenestedt. Layup optimization and sensitivity analysis of the fundamental

eigenfrequency of composite plates. Composite Structures, 12(3):193–209, 1989.

[73] J. L. Grenestedt. Layup optimization against buckling of shear panels. Structural

Optimization, 3:115–120, 1991.

[74] J. Foldager, J. S. Hansen, and N. Olhoff. A general approach forcing convexity of

ply angle optimization in composite laminates. Structural Optimization, 16:201–211,
1998.

[75] H. Fukunaga and H. Sekine. Stiffness design method of symmetric laminates using
lamination parameters. AIAA Journal, 30:2791–2793, 1992.

[76] C.G. Diaconu, M. Sato, and H. Sekine. Layup optimization of symmetrically lami-
nated thick plates for fundamental frequencies using lamination parameters. Struc-
tural and Multidisciplinary Optimization, 24(4):302–311, 2002.

[77] J. M. J. F. van Campen, C. Kassapoglou, and Z. Grdal. Generating realistic laminate

fiber angle distributions for optimal variable stiffness laminates. Composites Part B:
Engineering, 43(2):354–360, 2012.

[78] D. Keller. Optimization of ply angles in laminated composite structures by a hybrid,

asynchronous, parallel evolutionary algorithm. Composite Structures, 92(11):2781 –
2790, 2010.

[79] K. J. Callahan and G. E. Weeks. Optimum design of composite laminates using

genetic algorithms. Composites Engineering, 2(3):149–160, 1992.

[80] N. Kogiso, L. T. Watson, Z. Gürdal, and R. T. Haftka. Genetic algorithms with

local improvement for composite laminate design. Structural and Multidisciplinary
Optimization, 7:207–218, 1994.

c ETH Zürich IDMF-CMAS, March 1, 2018

210 BIBLIOGRAPHY

[81] R. Le Riche and R.T. Haftka. Improved genetic algorithm for minimum thickness
composite laminate design. Composites Engineering, 5(2):143 – 161, 1995.

[82] A. Todoroki and R. T. Haftka. Stacking sequence optimization by a genetic algorithm

with a new recessive gene like repair strategy. Composites Part B: Engineering,
29(3):277 – 285, 1998.

[83] Z. Gürdal, R. Haftka, and P. Hajela. Design and Optimization of Laminated Com-
posite Materials. John, 1999.

[84] J. H. Park, J. H. Hwang, C. S. Lee, and W. Hwang. Stacking sequence design of

composite laminates for maximum strength using genetic algorithms. Composite
Structures, 52(2):217 – 231, 2001.

[85] G. Soremekun, Z. Grdal, R. T. Haftka, and L. T. Watson. Composite laminate design

optimization by genetic algorithm with generalized elitist selection. Computers &
Structures, 79(2):131 – 143, 2001.

[86] A. Rama Mohan Rao and N. Arvind. A scatter search algorithm for stacking sequence
optimisation of laminate composites. Composite Structures, 70(4):383–402, 2005.

[87] J.-S. Kim. Development of a user-friendly expert system for composite laminate
design. Composite Structures, 79(1):76 – 83, 2007.

[88] J. Stegmann. Analysis and Optimization of Laminated Composite Shell Structures.

PhD thesis, Institute of Mechanical Engineering, Aalborg University, 2005.

[89] J. Stegmann and E. Lund. Discrete material optimization of general composite shell
structures. International Journal for Numerical Methods in Engineering, 62(14):2009
– 2027, 2005.

[90] E. Lund and J. Stegmann. On structural optimization of composite shell structures

using a discrete constitutive parametrization. Wind Energy, 8(1):109–124, 2005.

[91] N. Pedersen. On design of fiber-nets and orientation for eigenfrequency optimization

of plates. Computational Mechanics, 39:1–13, 2005.

[92] Shahriar Setoodeh, Mostafa M. Abdalla, and Zafer Grdal. Design of variablestiffness
laminates using lamination parameters. Composites Part B: Engineering, 37(45):301
– 309, 2006.

[93] B. Schläpfer and G. Kress. A sensitivity-based parameterization concept for the

automated design and placement of reinforcement doublers. Composite Structures,
94(3):896–903, 2012.

[94] M. Giger, D. Keller, and P. Ermanni. A graph-based parameterization concept

for global laminate optimization. Structural and Multidisciplinary Optimization,
36(3):289–305, 2008.

[95] B. Schläpfer, B. Rentsch, and G. Kress. Specific design of laminated composites

regarding dynamic behavior by the application of local reinforcements. submitted to
Composite Structures, 99:433–442, 2013.

[96] N. Zehnder and P. Ermanni. A methodology for the global optimization of laminated
composite structures. Composite Structures, 72(3):311–320, 2006.

c ETH Zürich IDMF-CMAS, March 1, 2018

BIBLIOGRAPHY 211

[97] Mattheck, C. Design in der Natur - der Baum als Lehrmeister Bd. 1. Rombacher
Ökologie, 1993.

[98] Baier, H., Ch. Seeßelberg, B. Specht. Optimierung in der Struktrumechanik. Vieweg,
1994.

[99] Jog, C.S. A robust dual algorithm for topology design of structures in discrete
variables. Int. J. Numer. Meth. Engng, 50:1607–1618, 2001.

[100] Bendsøe, M.P., A. Ben-Tal, J. Zowe. Optimization methods for truss geometry and
topology design. Structural Optimization, 7:141–159, 1994.

[101] Svanberg, K. On local and global minima in structural optimization. In Atrek,

A.; Gallager, R.H.; Ragsdell, K.M.; Zienkiewicz, O.C. , editor, New Directions in
Optimum Structural Design, pages 327–341. Wiley, New York, 1984.

[102] Diaz, A., Bendsøe, M.P. Shape optimization of structures for multiple loading con-
ditions using a homogenization method. Structural Optimization, 4:17–22, 1992.

[103] Taylor, J.E. Maximum strength elastic structural design. Proc. ASCE, 95:653–663,
1969.

[104] Olhoff, N., Taylor, J.E. On structural optimization. J. Appl. Mech., 50:1134–1151,
1983.

[105] Rozvani,. Structural Design via Optimality Criteria. Dordrecht: Kluwer, 1989.

[106] J. N. Reddy. An introduction to the finite element method. McGraw-Hill series in

mechanical engineering. McGraw-Hill Higher Education, 2006.

[107] H.R. Schwarz. Methode der finiten Elemente: eine Einführung unter besonderer
Berücksichtigung der Rechenpraxis. Leitfäden der angewandten Mathematik und
Mechanik. Teubner, 1991.

[108] H. Goldstein. Classical Mechanics. Addison-Wesley, 2001.

[109] A. E. H. Love. The small free vibrations and deformation of a thin elastic shell.
Philosophical Transactions of the Royal Society of London. A, 179:pp. 491–546, 1888.

[110] M. Iura and S. N. Atluri. Formulation of a membrane finite element with drilling
degrees of freedom. Computational Mechanics, 9(6):417–428, 1992.

[111] R. D. Cook. Four-node flat shell element: Drilling degrees of freedom, membrane-
bending coupling, warped geometry, and behavior. Computers & Structures,
50(4):549–555, 1994.

c ETH Zürich IDMF-CMAS, March 1, 2018

212 BIBLIOGRAPHY

c ETH Zürich IDMF-CMAS, March 1, 2018

Appendix A

Finite Element Method

Within this section, a short overview of the Finite Element Method (FEM) is given. The
content is mainly based on the textbooks of Cook et al. [69], Reddy [106] and Schwarz
[107]. The FEM is a numerical method for solving partial differential equations. It can
be applied to a wide field of physical problems such as structural analysis, heat transfer,
magnetic fields, flow processes and many more. The geometry of the physical problem is
discretized by dividing the domain into smaller parts called the finite elements. A finite
element has a domain, a boundary and so-called nodes on which the degrees of freedom
are defined. While the partial differential equations can only be solved analytically for
simple geometries, there is no geometric restriction using the FEM. Moreover, there is
no restriction to boundary conditions or material properties which makes the method ap-
plicable to any continuum mechanical problems. There exist numerous commercial finite
element codes which are still enhanced and refined today. While a deeper understanding of
the method requires some effort, the usage of FEM-codes is possible with little knowledge
of the method and the underlying problem. However, the consequences of an incorrect
application ”may range from embarrassing to disastrous” [69].
Within this thesis, the further explanations are focused on the FEM for structural anal-
ysis problems. The solution of the elasticity problem demands for the fulfillment of the
fundamental equation
LT CLu + f = ρü (A.1)

within a given domain Ω taking into consideration the prescribed surface stresses σ̂ and
displacements û on the surface Γ as well as the internally acting body forces f . As men-
tioned, an exact solution of the problem can only be found for simple geometries and
simple boundary conditions, e.g. a bar under uniaxial load. Generally, the FEM provides
only approximate solutions. Accepting a sufficient numerical effort, the obtained solution
may come close to the true solution.

A.1 Equation of Motion

There are several ways to derive the linear equation system of the FEM which can be solved
for the unknown displacements u. A general approach is based on Lagrangian mechanics
[108]. It is a formalism based on Hamilton’s Principle which describes the dynamics of
a system with a scalar function called the Lagrangian L and can be understood as a
generalization of the principle of virtual displacements to dynamics of solids [106]. The
equation of motion in Lagrangian mechanics, also known as the Euler-Lagrange equation,

c ETH Zürich IDMF-CMAS, March 1, 2018

214 Finite Element Method

is defined as  
d ∂L ∂L
− =0 (A.2)
dt ∂ ẋ ∂x
where L denotes the Lagrangian and x the generalized coordinates. The Lagrangian L is
defined as
L = T − Π = T − (U − W ) (A.3)
where T denotes the kinetic energy and Π the potential energy. The Hamilton’s principle
for an elastic body is represented with equation
Z t2 Z t2
δ [T − (U − W )] dt = δ Ldt = 0 (A.4)
t1 t1

which calls the time integral of the Lagrangian L to be stationary. The fundamental lemma
of the calculus of variation shows that solving the Lagrange equations is equivalent to find-
ing a solution of the Hamilton’s principle. However, taking advantage of the Lagrangian
mechanics transforms the equation of motion to its weak form which is needed for finite
element formulation.
The potential energy Π is the sum of the deformation energy U and the negative of the
potential of the external forces W . If the displacements u are small, which is required for
the linear elasticity problem, the velocities can be approximated with displacement time
derivatives u̇. The kinetic energy can thus be formulated as the integral over the domain Ω
of the density ρ and the scalar product or the velocities u̇.
Z
1
T = ρu̇T u̇dΩ (A.5)
2 Ω
The deformation energy U is defined as the domain integral of the scalar product of
stresses σ and strains ε. Z
1
U= εT σdΩ (A.6)
2 Ω
and the potential of the external forces W is dependent on the body forces f and the
surface stresses σ̂. Consider that the surface stresses are integrated over the surface Γ.
Z Z
W = f udΩ + σ̂ T udΓ
T
(A.7)
Ω Γ

The Lagrangian L is formulated for the entire domain Ω. In order to get the FEM for-
mulation, the domain must be discretized into smaller sub-domains Ωe , namely the finite
elements. The integration over the domain Ω is replaced with a summation of the integrals
over the sub-domains Ωe . Additionally, local approximation functions φ, which are also
called shape functions, are defined. They map the finite element nodal displacements ũ,
which represent also the degrees of freedom, to the continuous displacements u. The nodal
velocities ũ˙ and the accelerations ũ
¨ are mapped analogously.

u ≈ φT ũ , u̇ ≈ φT ũ˙ , ü ≈ φT ũ
¨ (A.8)

Considering the kinematic equation

ε = Lu, (A.9)
the strains ε can be expressed as a function for the nodal point displacements ũ.

ε = Lu = LφT ũ = Bũ (A.10)

c ETH Zürich IDMF-CMAS, March 1, 2018

A.1 Equation of Motion 215

A strain-displacement matrix B, which contains the spacial derivatives of the local shape
functions, is built by applying the differential matrix operator L to the shape functions φ.
The total Lagrangian for the discretized system then takes the discrete form
X 1 Z  X Z
1

L= ˙ T T ˙
ρũ φφ ũdΩe − T T
ũ B CBũdΩe
nelm
2 Ωe nelm
2 Ωe
X Z  X Z 
+ f T φT ũdΩe + σ T φT ũdΓe (A.11)
nelm Ωe nelm Γe

Finally, the unknown nodal displacements ũ are taken as generalized coordinates wherefore
the Lagrange-Euler equation (A.2) is modified to
 
d ∂L ∂L
− =0 (A.12)
˙
dt ∂ ũ ∂ ũ
The evaluation of the Lagrangian formalism leads to the equation of motion for the dis-
cretized linear system.
X Z  X Z 
T¨ T
ρφφ ũdΩe + B CBũdΩe
nelm Ωe nelm Ωe
X Z  X Z 
− φf dΩe − φσ̂dΓe = 0 (A.13)
nelm Ωe nelm Γe

The equation is then rearranged so that terms with no dependence on the nodal
displacements ũ are brought to the right hand side.
X Z  X Z 
T
B CBdΩe ũ + T ¨
ρφφ dΩe ũ
nelm Ωe nelm Ωe
X Z  X Z 
= φf dΩe + φσ̂dΓe (A.14)
nelm Ωe nelm Γe

The sums on the left side represent the global stiffness matrix K and the global mass
matrix M while the right hand side represents the load vector r. With these symbols,the
basic problem of the FEM for the linear elastic case is written as
¨=r
Kũ + Mũ (A.15)
¨ vanish so that the equation simplifies to
Assuming a static problem, the accelerations ũ

Kũ = r (A.16)

which is a simple linear equation system for the unknown displacements ũ. The equation
of motion can also be used for the determination of the harmonic eigenfrequencies of the
structural system. There, it is assumed that the load vector r is zero. Taking advantage
of the harmonic approach, the equation of motion can be transferred into an eigenvalue
problem with the unknown eigenvalues λ and eigenvectors Φ, whereas λ is equal to the
square of the angular frequency ω.

ũ = Φ sin (ωt) (A.17)

¨ = −ω 2 Φ sin (ωt)
ũ (A.18)

c ETH Zürich IDMF-CMAS, March 1, 2018

216 Finite Element Method

The combination of the equation of motion (A.15) and the harmonic approach (A.17, A.18)
yields to the eigenvalue problem for the harmonic vibration.

K − ω2M Φ = 0


(A.19)

Considering equation (A.14), the global system matrices can be extracted directly. The
global stiffness matrix K is defined as the sum of the element stiffness matrices
X Z 
T
K= B CBdΩe (A.20)
nelm Ωe

whereas the element stiffness matrices k are given as

Z
k= BT CBdΩe (A.21)
Ωe

Analogously, the global mass matrix M and the element mass matrices m are defined as
X Z 
T
M= ρφφ dΩe (A.22)
nelm Ωe

and Z
m= ρφφT dΩe (A.23)
Ωe

The load vector r contains the internal body forces and the prescribed forces on the surface
of the structure. The shape functions φ distribute the loads to the nodes.
X Z  X Z 
r= φf dΩe + φσ̂dΓe (A.24)
nelm Ωe nelm Γe

The summations must consider the connectivity of the respective element with the globally
numbered mesh nodes.

c ETH Zürich IDMF-CMAS, March 1, 2018

A.2 Formulation of a Layered Shell Element 217

A.2 Formulation of a Layered Shell Element

A general shell element contains both, membrane and flexural stress theory. The Kirchhoff-
Love theory [109], which has already been postulated for the derivation of the CLT, is
used here for the formulation of the bending behavior. The displacements of a point
are composed of the mid-plane displacements u0 , v0 and a term arising from the plate
curvature. The curvature term is dependent on the position through the thickness z and
the rotations of the normal to the middle plane θx and θx according to equations (A.25)
and (A.26).

u = u0 + z θy (A.25)
v = v0 + z (−θx ) (A.26)

The chosen conventions for the derivation of the finite shell element are shown in
Figure A.1. However, an alternative convention is feasible as well. Taking advantage

z,w
z,w
zθy
θx y,v
θy Ny x,u
Mxy θy P z w
My Nxy
x,u Mx
Nxy
Nx Mxy

Figure A.1: Loads and kinematics of a Kirchhoff-plate

of the kinematic relations, the strains yield

       
εx u,x u0,x θy,x
 εy  =  v,y  =  v0,y  +z  −θx,y  (A.27)
γxy u,y + v,x u0,y + v0,x θy,y − θx,x
| {z } | {z } | {z } | {z }
ε ε ε0 κ

where ε0 denotes the mid-plane or membrane strains and κ the plate curvatures. Thin
plate theory assumes the strain distribution to be linear through the thickness. The out-
of-plane deflections w are connected to the in-plane displacement u, v with the transverse
shear which is  
∂w ∂u
 
γxz + 
∂x ∂z 

=  ∂v ∂w  (A.28)
γyz
+
∂z ∂y
Alternatively it can be expressed in terms of θx and θy taking advantage of the derivatives
of the kinematic relations (A.25) and (A.26). The derivatives yield

∂u
= θy (A.29)
∂z
∂v
= −θx (A.30)
∂z

c ETH Zürich IDMF-CMAS, March 1, 2018

218 Finite Element Method

which leads to  
∂w
 
γxz + θy 
∂x

=
 ∂w
 (A.31)
γyz 
− θx
∂y
Taking advantage of the Finite Element Method (FEM) (see Appendix A), the strains
are composed of the matrix multiplication of the differential operator L, the shape func-
tions φ and the nodal displacements ũ (see equation (A.10)).

ε = Lu = LφT ũ = Bũ (A.32)

According to equation (A.21), the stiffness matrix is defined as

Z
k= BT CBdΩe (A.33)
Ωe

Since membrane (m), bending (b) and transverse shear (s) parts are differently dependent
on the out-of-plane-coordinates z (see equation (A.27)) and on different material laws,
respectively, the integration over the shell thickness is separated. The membrane stiffness
matrix assuming a homogeneous material yields
Z Z t Z
2
km = BTm Cf Bm dzdA = t BTm Cf Bm dA (A.34)
A − 2t A

with
 u v 
∂φ
 ∂x 0 
 
 ∂φ  (A.35)
Bm = 
 0

 ∂y 

 ∂φ ∂φ 
∂y ∂x
and  
1 ν 0
E ν 1
Cf = 0  (A.36)
(1 − ν)2 1−ν
0 0 2

The bending stiffness is given by

t
t3
Z Z Z
2
kb = BTb zCf zBb dzdA = BTb Cf Bb dA (A.37)
A − 2t 12 A

with
 θx θy 
∂φ
0

 ∂x 

 ∂φ  (A.38)
Bb =  − 0 

 ∂y 

 ∂φ ∂φ 
−
∂x ∂y

c ETH Zürich IDMF-CMAS, March 1, 2018

A.2 Formulation of a Layered Shell Element 219

Also the transverse shear stiffness part

Z Z t Z
2
ks = BTs κCc Bs dzdA =t BTs κCc Bs dA (A.39)
A − 2t A

is linearly dependent on the stiffness. The stress-strain relation in transverse shear is

defined as  
G 0
Cc = (A.40)
0 G
and a shear correction factor κ is introduced to mitigate the shear locking problem. The
corresponding strain-displacement matrix is given by

 w θ x θy 
∂φ
0 φ (A.41)
∂x
 
Bs = 
 ∂φ


−φ 0
∂y
The total element stiffness matrix k is the addressed summation of the single parts which
means, that the degrees-of-freedom indicated at the top of equations (A.35), (A.38) and
(A.41) must be respected.
k = k m + kb + ks (A.42)
A general shell element must have 6 degrees-of-freedom in order to be feasible for spatial
3D-modeling. However, the formulation above covers only the 5 d.o.f. u, v, w, θx and θy
but not θz The so called drilling degree-of-freedom θz must be introduced artificially (see
[110, 111]) in order to make the equation system solvable. Considering equations (A.34),
(A.37) and (A.39), it becomes obvious that no re-evaluation of the integrals is needed if the
thickness t is changed since it is contained explicitly. The thickness can be adapted with
very low additional computational cost which makes the application of the shell elements
very efficient and adequate for preliminarily design.

c ETH Zürich IDMF-CMAS, March 1, 2018

220 Finite Element Method

c ETH Zürich IDMF-CMAS, March 1, 2018

Appendix B

Material Invariants of Orthotropic

Materials

The theory of lamination parameters introduced in Section 8.4 require the matrices Γi
containing the material invariants of orthotropic materials. Tsai and Pagano [70] accom-
plished to formulate the reduced stiffness transformation equations by a combination of
trigonometric identities and 5 material invariants Ui (which are invariant under rotations
about the z-axis). After the textbook of Jones [25] the relations are

Q11 = U1 + U2 cos 2ϕ + U3 cos 4ϕ (B.1)

Q12 = U4 − U3 cos 4ϕ (B.2)
Q22 = U1 − U3 cos 2ϕ + U3 cos 4ϕ (B.3)
1
Q16 = − U2 sin 2ϕ − U3 sin 4ϕ (B.4)
2
1
Q26 = − U2 sin 2ϕ + U3 sin 4ϕ (B.5)
2
Q66 = U5 − U3 cos 4ϕ (B.6)

in which
1
U1 = (3Q11 + 3Q22 + 2Q12 + 4Q66 ) (B.7)
8
1
U2 = (Q11 − Q22 ) (B.8)
2
1
U3 = (Q11 + Q22 − 2Q12 − 4Q66 ) (B.9)
8
1
U4 = (Q11 + Q22 + 6Q12 − 4Q66 ) (B.10)
8
1
U5 = (Q11 + Q22 − 2Q12 + 4Q66 ) (B.11)
8
whereas Qii are the entries of the material stiffness matrix.

c ETH Zürich IDMF-CMAS, March 1, 2018

222 Material Invariants of Orthotropic Materials

These equation can be transformed to a matrix notation

Q = Γ0 + Γ1 cos 2ϕ + Γ2 sin 2ϕ + Γ3 cos 4ϕ + Γ4 sin 4ϕ (B.12)

with
 
U1 U4 0
Γ0 = U4 U1 0  (B.13)
0 0 U5
 
U2 0 0
Γ1 =  0 −U2 0 (B.14)
0 0 0
 
0 0 U2
1
Γ2 = 0 0 U2  (B.15)
2
U2 U2 0
 
U3 −U3 0
Γ3 = −U3 U3 0  (B.16)
0 0 −U3
 
0 0 U3
Γ4 =0 0 −U3  (B.17)
U3 −U3 0

These are also the matrices Γi which are employed in the lamination parameter theory in
Section 8.4.

c ETH Zürich IDMF-CMAS, March 1, 2018