See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/261173987

Design of Material Structures Using Topology Optimization

Thesis · January 1994

CITATIONS READS

402 13,226

1 author:

Ole Sigmund
Technical University of Denmark
469 PUBLICATIONS 48,217 CITATIONS

SEE PROFILE

All content following this page was uploaded by Ole Sigmund on 26 August 2015.

The user has requested enhancement of the downloaded file.

 Design of Material Structures
using Topology Optimization

Ole Sigmund
December, 1994

Department of Solid Mechanics

(now: Department of Mechanical Engineering/Solid Mechanics)
Technical University of Denmark
Lyngby, Denmark

DCAMM, Special report, S69

Note added December 7th, 2005:

Please note that almost all the contents of this thesis have been published in various journal pa-
pers, thus reference to the work should preferably be given to the journal papers and not to this
thesis.
 Preface
The work presented in this thesis for the Ph.D. degree has been carried out at the Department
of Solid Mechanics, Technical University of Denmark (DTU) and has been supervised by Docent
Pauli Pedersen (Dept. of Solid Mech., DTU), and co–supervised by Docent Martin P. Bendsøe
(Institute of Mathematics, DTU) and Lektor Jon J. Thomsen (Dept. of Solid Mech., DTU). I
would like to express my sincere thanks to my supervisors for inspiring me to do this work, and
for their helpful advice, suggestions, support, encouragement and patience when I needed it.

I would also like to thank Robert B. Zetterlund and Kurt Fries Weihe (both Dept. of Solid Mech.,
DTU) for their valuable ideas and help in producing working models of the material microstruc-
tures.

The work was supported by Denmarks Technical Research Council through the Program on
Computer Aided Design. This support is gratefully acknowledged.

The introductory part of this Ph.D. work (Oktober 1991 to December 1992) was performed
under supervision of Professor George I.N. Rozvany, University of Essen, Germany. The stay
in Essen was sponsored by the Deutsche Forschungsgemeinshaft whose support is gratefully
acknowledged. I would like to express my sincere thanks to G.I.N. Rozvany for giving me a valu-
able introduction to optimality criteria methods and to him and his staff for a professionally and
personally instructive and memorable period.

Part of this work (Appendix 1), was done together with Professor Alejandro R. Díaz, Michigan
State University (MSU). I would like to thank A. R. Díaz for our co–operation and also Profes-
sors Noboru Kikuchi and John E. Taylor, University of Michigan (UofM) for many inspiring
research discussions during my stay in Michigan in the spring of 1994.

Last but not least, I would like to thank my fellow students at the Department of Solid Mechanics,
DTU for great support and many social activities through the years.

Lyngby, December 1994

Ole Sigmund
 Abstrakt (in Danish)
En numerisk metode er udviklet til design af materialer med foreskrevne termoelastiske parame-
tre. Materiale design problemet er formuleret som en optimering af topologien for en periodisk
mikrostruktur repræsenteret ved en enhedscelle. Materialets vægt er valgt som objektfunktion,
men funktionen er modificeret for at kunne tage hensyn til produktionsbegrænsninger. Mikro-
strukturen er diskretiseret med stang, ramme eller kontinuum finite elementer, hvor respektivt
tværsnitsareal, bredde og densitet af de individuelle elementer indgår som design variable. De
effektive elastiske parametre for den diskretiserede mikrostruktur er fundet ved hjælp af en
numerisk homogeniseringsmetode og materiale design problemet kan derved defineres som et
inverst homogeniseringsproblem.

Numeriske eksperimenter, i hvilke den foreslåede metode er anvendt til design af gitter– (i to og
tre dimensioner) og rammeagtige mikrostrukturer, indikerer, at det er muligt at designe nye
mikrostrukturer med alle termodynamisk tilladelige elastiske parametre og som kun har en læng-
deskala.

Anvendelse af metoden på kontinuumagtige mikrostrukturer var vanskelig på grund af proble-

mer med skakternede områder i de resulterende topologier. Skakternede områder er et velkendt
problem indenfor generel topologi optimering af strukturer, og problemet er derfor undersøgt
i detaljer og er forklaret ved, at lavere ordens finite elementer modellerer stivheden af en skakter-
net mikrostruktur dårligt. Der foreslåes en procedure til at eliminere problemet uden at forøge
beregningstiden. Den foreslåede procedure kan samtidig mindske net–afhængighed af de opti-
male løsninger. Selvom metoden er heuristisk baseret, opnåes lovende resultater ved dens
anvendelse i topologi optimering af strukturer og materialer.

Brug af den foreslåede net–uafhængigheds procedure gør design af kontinuumagtige mikro-

strukturer simplere, og eksempler viser, at materialer med ekstreme termoelastiske parametre
også kan designes i kontinuum formuleringen.

Af produktionstekniske årsager foretrækkes de kontinuumagtige mikrostrukturer fremfor de git-

ter– og rammeagtige materialer, men det kan konkluderes, at enhver topologisk begrænsning
indskrænker de opnåelige termoelastiske parametre.
 Abstract

A numerical method for the design of linear elastic materials with prescribed thermoelastic prop-
erties is proposed. The material design problem is formulated as a topology optimization prob-
lem of a periodic material microstructure represented by a base cell. The objective function is
a measure of weight, modified to take manufacturability constraints into account. The micro-
structure is discretized by truss–, frame– or continuum–type finite elements where cross–sec-
tional areas, widths and densities of individual elements respectively are taken as design vari-
ables. Effective properties of the discretized microstructure are found by a numerical
homogenization method which leads to the definition of the material design problem as an
inverse homogenization problem.

Numerical experiments with the proposed method applied to the design of truss– (2–d and 3–d)
and frame–like microstructures indicate that it possible to design novel microstructures on one
length scale that attain any elastic properties admissible by thermodynamics.

Applying the method to continuum–like microstructures was difficult due to occurrence of

checkerboard patterns in the solutions. The reason for the occurrence of checkerboard patterns
which is a well known problem in general topology optimization of continuum structures is stu-
died in detail and explained by poor modelling of the stiffness of checkerboards by lower order
finite elements. A procedure is proposed to eliminate the problem without increasing computa-
tional complexity and at the same time preventing the mesh–dependency problem which refers
to the non–convergence of solutions with mesh refinement. Although based on heuristics, the
method shows promising results applied to topology optimization of structures and materials.

Taking advantage of the proposed mesh–independency algorithm, design of continuum–like

materials was facilitated and examples show that materials with extreme thermoelastic proper-
ties can be designed by the continuum formulation as well.

Due to manufacturability constraints, the continuum type microstructures are preferred to the
truss– and frame–like materials but it is concluded that any topological constraint on the material
microstructure damps out attainable elastic properties.
 Table of Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I

Abstrakt (in Danish) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . II

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . III

Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . IV

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Hierarchical materials and structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Negative Poisson’s ratio materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
Approaches to material design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Topology optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
About this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1. Numerical homogenization procedures . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.1 Homogenization of elastic properties . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2 Homogenization of thermal properties . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Finite element discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 Solution procedures for the finite element problem . . . . . . . . . . . . . . 15
1.5 Discretizations of microstructure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2. Design of materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.1 The inverse homogenization problem . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 Solving the optimization problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3 Truss–like materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4 Frame–like materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.5 Continuum–like materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.6 Continuum–like materials with prescribed thermoelastic properties . 53
2.7 Practical considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.8 Summary and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

3. The checkerboard and mesh–dependency problems

in layout optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.1 Design of structures by compliance optimization . . . . . . . . . . . . . . . . 62
3.2 Checkerboard prevention scheme using image processing . . . . . . . . . 70
3.3 Implementation of the checkerboard prevention scheme . . . . . . . . . . 72
3.4 Elimination of mesh–dependency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

List of publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
 Table of contents V

Appendices

Appendix 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A1
Checkerboard patterns in layout optimization . . . . . . . . . . . . . . . . . . . . . . A1

Appendix 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A2
Notations for the constitutive tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A2

Appendix 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A4
The Element–by–Element Preconditioned Conjugate Gradient Method . A4

Appendix 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A6
Continuum equivalence of bar elements and equivalent nodal loads . . . . A6
 Introduction
Optimization of material properties play an important role in the design of structures. Using the
right material for a specific application can give great payoffs such as lower weight, higher stiff-
ness or strength or improvement of thermal or dynamic response.

Looking at any material through a microscope, a microstructure is seen if the magnification is

strong enough. If the microstructure itself has a microstructure, the material can be called a hier-
archical material. The topologies and compositions of the individual microstructural levels are
crucial for the effective properties of the material.

A microstructure can be regarded as a small structure and therefore, techniques developed for
the optimization of general structures can be used in the optimization of materials. Using this
idea, the present thesis suggests that a numerical method based on topology optimization of
structures can be used in the design of new material microstructures with prescribed thermoelas-
tic properties.

In the following sections, an introduction to matters important for microstructural design is given
together with references to literature. First the benefits obtained from designing hierarchical
materials and structures are summarized. Then examples of approaches to material design are
given together with a description of the important subproblem of designing materials with nega-
tive Poisson’s ratios. Finally, approaches to topology optimization of structures are listed
together with their possible applications to material design.

Hierarchical materials and structures

The usual engineering practice in the design of elastic structures has been to take an existing
material and try to utilize it in the best possible way by varying structural dimensions or shape.
After the introduction of fiber based composites a few decades ago, ways were opened for the
design of more efficient structures taking advantage of the high stiffness to weight ratio and ani-
sotropy of such materials. Recently, a new approach to the design of structures has been sug-
gested by different authors. Instead of trying to utilize an existing material the best possible way,
the approach suggests that the material microstructure itself should be tailored to the specific
application. Ideally, the material at each point of a structure should be specifically designed for
the local loading condition. The idea of tailoring material microstructures for specific needs has
been presented by Ashby, (1991), who suggests that efficient structures can be constructed by
combining the design of structural shape with the design of material microstructure; by Lakes,
(1993)2, who discusses the benefits of using hierarchical materials; by Ashley, (1994)1, who pre-
dicts big property payoffs from design of materials on the atomic (nano–) scale and finally by
2 Introduction

Bendsøe et al., (1993)1,2 and Ringertz, (1993), who suggest that weight optimal structures sub-
ject to simple and multiple loading cases are obtained by a free parametrization of the elasticity
tensor and the former conclude, that a lower bound on compliance can be reached under the
condition that materials with properties that are only restricted to be thermodynamically admissi-
ble can be made.

As a result of evolution, nature has tailored many hierarchical and highly specialized materials.
One of the most structurally efficient materials that exists is wood, as pointed out by Ashby,
(1991). Wood has a porous microstructure where the solid part is distributed as prismatic cells.
In this way, material is moved away from the bending axis, making the tree stronger in bending
per unit weight than the solid it is made of. Human bone structure is another example of naturally
occurring hierarchical materials. Bone material has two hierarchical levels [e.g. Hollister and
Riemer, (1993)], the first is the architectural level on a scale ranging from 100 to 500 microns,
the second is the tissue microstructural level which contains structure ranging in size from 3 to
50 microns. During life time the microstructure of bone constantly adapts to outer stimuli. As
pointed out in a review paper by Huiskes and Hollister, (1993), it is not yet fully understood how
bone structure adapts to mechanical stimuli but it is generally believed that the microstructure
remodels to support given loads in some optimal way.

Apart from the direct structural benefits discussed above, designing materials with specific prop-
erties has many other applications. For example, it is rewarding to design materials that mimic
naturally occurring materials as seen in the next two examples. Human bone structure is known
to exhibit extreme elastic properties, and Evans, (1990) points out that it is important to mimic
replaced tissue as closely as possible, in order to make the best artificial replacements of human
bone structure in surgery. The other example is the interesting task of constructing an ”artificial”
Stradivarius. In Ashby and Jones, (1986) a desire is expressed of being able to design a synthetic
material that reproduces the acoustic properties of wood as closely as possible, in order to
construct great sounding violins.

Negative Poisson’s ratio materials

The design of a material with a very special form of elastic properties has been the goal of many
researchers for the last few years. The existence of materials with negative Poisson’s ratios, or
in other words, the existence of materials that expand transversely subject to an applied tensile
load has been questioned by researchers and engineers for a long time and many basic books on
elasticity limit the possible value of the Poisson’s ratio to the interval [0, 0.5[. However, thermo-
dynamics or the requirement that the constitutive tensor be positive semi–definite allow the Pois-
son’s ratio of an isotropic material to approach –1. Particular microstructures of materials with
negative Poisson’s ratios were reported by Almgren, (1985) who suggested a mechanism like
 Introduction 3

structure with hinges, springs and sliding collars and Kolpakov, (1985) and Gibson and Ashby,
(1988) who suggested an inverted honeycomb. Later, Evans, (1989) proposed microstructures
composed of rotating disks connected by inextensible strings and Rothenburg, Berlin and
Bathurst, (1991) proposed a triangular truss structure with telescopic bars. Many of these solu-
tions are not isotropic or can not be characterized as elastic continua because they have internal
sliding surfaces. This lead Milton, (1991) to ask the question: is it possible to produce isotropic
materials with negative Poisson’s ratios within the framework of linear continuum elasticity?
Milton answered the question himself in the affirmative, by introducing a two–phase composite
based on layerings at different length scales and made a mathematically rigorous proof that the
composite has Poisson’s ratio close to –1 and has no sliding surfaces.

Two approaches have been made to produce negative Poisson’s ratio materials in practice. The
first is by Lakes (1987) (and several following papers by the same author) who modifies open
walled foam materials by subjecting them to heat and compression, to get Poisson’s ratios as low
as –0.7. The second is by Caddock and Evans, (1989) who obtained an expanded polytetrafluo-
roethylene microporous material with negative Poisson’s ratio. However, the latter is strongly
anisotropic and therefore many of the positive features reported for isotropic negative Poisson’s
ratio materials might not exist for this material.

Applications and benefits of negative Poisson’s ratio materials are numerous. Lakes, (1993)1 has
investigated several applications and reports that negative Poisson’s ratio materials are good as
core materials for convex sandwich plates because they can be easily bent into a convex shape
as opposed to a core material with positive Poisson’s ratio. Lakes also investigates the implica-
tions of having a negative Poisson’s ratio on stress concentration at holes or cracks and con-
cludes, that stress concentration factors are reduced in some cases and unchanged or increased
in others. Lakes and Elms, (1993) have investigated indentability of negative Poisson’s ratio
foams and calculated energy absorption rates that were considerably higher than for conven-
tional foams. Another application of negative Poisson’s ratio materials are as fasteners, investi-
gated in Choi and Lakes, (1991), where insertion of the fastener is facilitated due to the transverse
contraction of the fastener when pressed into the hole while removal is resisted due to the trans-
verse expansion of the fastener when pulled. The fastener can therefore be seen as a sort of ”raw-
plug”. An open but interesting question is whether fibre pull–out would be resisted if the fiber
is made of a negative Poisson’s ratio material? Finally, an application due to the change in bulk
modulus of a negative Poisson’s ratio material shall be mentioned. The bulk modulus of a mate-
rial is defined as K=E/(3(1–n)), where E is Young’s modulus and n is Poisson’s ratio. A low bulk
modulus makes a material more sensitive to hydrostatic pressure. The sensitivity to hydrostatic
pressure is increased with a factor of three for a Poisson’s ratio –1 material compared to one with
Poisson’s ratio 0.33 and it is expected that the high sensitivity of Poisson’s ratio –1 materials can
4 Introduction

be utilized in the design of micro–hydrophones or in sensors and actuators for use in smart mate-
rials and structures.

Approaches to material design

The problem of designing materials with specific properties is being attacked by researchers
from many different areas, ranging from theoretically developed composite materials by applied
mathematicians over lamination and braiding of existing composite materials to the manipula-
tion of individual atoms by nano–technologies.

In Milton and Cherkaev, (1993) materials with elastic parameters ranging over the entire scale
compatible with thermodynamics are constructed by a multi scale layering of isotropic soft and
strong material. From bounds obtained by Cherkaev and Gibiansky, (1993), it is known that the
stiffness ratio between the strong and the soft material should be large in order to be able to design
materials with extreme elastic properties. A layered material with Poisson’s ratio close to –1 was
constructed from rigid and infinitely weak material by Milton, (1992). These materials are usu-
ally regarded as mathematical tools rather than practical composites due to their widely differing
length scales [Lakes, (1993)2].

Stiffness, strength and other material properties can be tailored to a certain extent by varying
layer directions, layer thicknesses, types, etc. in fibre/matrix materials [e.g. Autio, Laitinen and
Pramilla, (1993)] and by varying yarn orientations and fiber densities in braided composites [e.g.
Kostar and Chou, (1994)]. Dvorak, (1993) tailored conduction properties of a periodic 2–phase
material by varying the shape of the inclusion phase. All three papers produce solid materials
which are more or less readily manufacturable, but the attainable material properties are limited
by the choice of fibre/matrix material.

Considering design of materials at a smaller scale, Moore, (1993) lists different attempts to
chemically synthesize hinged molecular crystals which apart from exhibiting negative Poisson’s
ratio behaviour can be tailored to have special mechano–coupled optical and electrical proper-
ties.

Designing materials on the atomic scale is the subject of the rapidly expanding research field of
nanostructures where techniques are developed to manipulate microstructures atom by atom.
Great efforts are invested on producing materials with a variety of properties such as strength,
hardness, toughness, thermal expansion, conductivity, corrosion–resistance, optical and mag-
netic properties and many others as listed in Ashley, (1994)1.
 Introduction 5

Material design can roughly be divided into three different length scales, namely nano–, meso–
and micro–scale. The nano–scale includes design at atomic and molecular levels, the meso– or
submicron–scale goes up to approximately one micron and micro–scale is here defined as every-
thing above one micrometer. Usual continuum elasticity laws hold for the micro–scale and with
some caution for the meso–scale [e.g. Nkansah, Evans and Hutchinson, (1994)] but the nano–
scale requires other non–linear analysis tools.

Topology optimization
Topology optimization of structures is a rapidly growing field of particular interest to the auto-
motive and aerospace industries. Topology optimization, as opposed to shape optimization,
allows the introduction of holes or cavities in structures which usually results in great savings
in weight or improvement of structural behaviour such as stiffness, strength or dynamic
response.

Topology optimization has received revived interest since the introduction of the ”homogeniza-
tion approach to topology optimization” by Bendsøe and Kikuchi, (1988) but has its origins back
to the minimum weight structures of Michell, (1904).

Topology optimization has in the last decades been applied to discrete structures such as trusses
[e.g. Pedersen, (1970), Ringertz (1988) or Sankaranaryanan, Haftka and Kapania, (1992)] or
frames [e.g. Zhou and Rozvany, (1993)], to continuum structures such as plates [e.g. Cheng and
Olhoff, (1981) and Díaz, Lipton and Soto (1993)] or to membranes [e.g. Bendsøe and Kikuchi,
(1988)]. Objective functions are most often compliance or weight but topology optimization has
also been performed with other performance criteria such as dynamic behaviour [e.g. Díaz and
Kikuchi, (1992) and Rozvany, Zhou and Sigmund, (1994)], thermoelastic behaviour [e.g. Rodri-
ques and Fernandes, (1994)]. There are several examples of topology optimization used in or
inspired by medical engineering reviewed in Huiskes and Hollister, (1993). The author has, as
a part of his Ph.D.–work, published analytical and numerical work on topology optimization of
trusses, grillages and continua with single and multiple static and dynamic constraints in collabo-
ration with G.I.N. Rozvany and his research group at University of Essen, Germany. References
are given in ”List of publications” at the end of this report.

Above mentioned references to applications of topology optimization methods are not complete
but are selected due to their relevance to the present work. For a more exhaustive overview of
the field, the reader is referred to Bendsøe (1994) and the references herein.

The material design problem can formulated as a topology optimization problem. Depending on
density and topology one can interpret a microstructure as a micro truss–, frame– or continuum–
6 Introduction

like structure. Furthermore, desirable properties for general structures such as low weight, high
stiffness or specific thermal responses are also desirable properties in material design. Therefore,
assuming that material microstructure can be designed freely, the tools developed for topology
optimization of general structures described above can be directly applied to the design of mate-
rial microstructure.

Problem formulation
Having realized that the design of materials ranging from atomic to millimeter scale with specific
elastic, strength, optic, magnetic, thermal, electric and many other properties opens exciting per-
spectives, the subject of this work shall be limited to the design of materials with prescribed ther-
moelastic properties on a scale which makes it possible to use the rules of continuum elasticity.

Until now, many approaches to design of materials with specific properties have been based on
a trial and error basis. In theoretical approaches, many solutions have been based on intuitive
ideas of structures which could exhibit the wanted behaviour. Practical approaches often consist
of numerous experiments trying out mixtures of different constituents and processing types and
often, materials with new properties are discovered by coincidence.

As examples of systematical approaches to material design were mentioned the mathematical

approach with multi–scale composites by Milton and Cherkaev, (1993) and the approach of lam-
inating fiber/matrix materials by Autio, Laitinen and Pramilla, (1993). However, the former
approach is not regarded as practical because of the differing length scales and the latter can only
produce materials with a limited range of properties.

Consequently, a major goal of this thesis is to propose a unified approach to the design of effi-
cient linear elastic materials, with any possible prescribed elastic properties and with vari-
ation of microstructure limited to a single length scale

To facilitate the work,

S the material microstructures will be assumed to be periodically repeated so that

material properties are uniquely determined by analysis of a single representative
base cell.

S the choice of the base material of which the microstructure is built is considered
unimportant for the design problem. Whether the actual microstructure is built in
metal or polymer is just a matter of production method as we do not consider fatigue
life, plasticity or other properties strongly related to a specific base material.
 Introduction 7

S manufacturability constraints such as exclusion of sliding surfaces should be taken

into account in the definition of the design problem but apart from this, no further
restrictions such as choice of base material, overall density or microstructural fea-
tures of specific materials shall be imposed.

S actual production of the resulting microstructures is outside the scope of this thesis.
The resulting microstructures can be used as inspiration for people actually involved
in manufacturing of materials with novel microstructures. However, simple test
models and ideas to build more practical prototypes will be discussed.

The effective properties of a periodically repeated microstructure represented by a base cell can
be computed by a homogenization method. Homogenized or effective properties are found by
subjecting the discretized base cell to different test fields and calculating the response by the
finite element method. The material design problem can be seen as an inverse homogenization
problem. In a usual homogenization problem we seek the effective properties of a given material.
In the inverse problem, we seek the material that has given effective properties.

The goal of designing efficient materials with prescribed elastic properties, here termed the
inverse homogenization problem, can be defined as an optimization problem of minimization of
material weight, subject to equality constraints on the prescribed elastic properties. Minimizing
weight is considered a reasonable design criteria because high stiffness to weight ratio is desir-
able in most cases – especially in automotive and aerospace applications. Choosing the pre-
scribed elastic properties as equality constraints ensures that the specified properties are obtained
as demanded. The optimization problem can be seen as a topology optimization problem, where
the design domain is the material microstructure represented by the base cell, and the design vari-
ables are the dimensional properties of finite elements used to discretize the base cell.

About this thesis

This thesis is divided into three self contained chapters dealing with: general numerical homoge-
nization procedures, design of materials with prescribed thermoelastic properties and elimina-
tion of checkerboard and mesh dependency problems in topology optimization.

In chapter one, the standard numerical homogenization procedure is described. The procedure,
is reformulated and extended in a way that makes it well suited for the topology optimization
of material microstructure discussed in chapter two. Material microstructure is represented by
groundstructures composed of truss– (two and three dimensional), frame– and continuum–like
elements and it is suggested that any specific microstructure can be modelled by one of these
groundstructures. The use of the groundstructure approach in homogenization is illustrated with
8 Introduction

examples. Readers not interested in numerical homogenization procedures may skip chapter one
and proceed directly to chapter two on material design.

An optimization algorithm is introduced in chapter two that can solve the inverse homogeniza-
tion problem of finding a material with prescribed thermoelastic properties. The optimization
problem is formulated in a general way, independent of the considered microstructure and the
basic objective function, minimization of weight, is modified to account for manufacturability
constraints as well. The optimization problem is solved by a multiple constraint optimality crite-
ria method, based on procedures known from general topology optimization. The chapter shows
numerous examples of truss–, frame– and continuum–like microstructures with a wide range of
thermoelastic properties. It is shown that the proposed procedure indeed can design one–length–
scale materials with any possible elastic properties by truss–like microstructures but the
introduction of manufacturability constraints damps out the attainable properties as demon-
strated by frame– and continuum–like microstructures.

Chapters one and two represent a compilation and extension of work published in Sigmund,
(1994)1,2. Especially the part on design of continuum–like materials has been revised, and all
examples have been recomputed with many more design degrees of freedom. This was partly
made possible by the implementation of an iterative equation solver (appendix 3) but not least
by the implementation of the mesh–independency algorithm developed in chapter three.

In chapter three a procedure to eliminate two well known problems occurring in topology opti-
mization of continuum structures is proposed, namely the checkerboard problem and the mesh–
dependency problem. The checkerboard problem refers to the formation of regions in optimized
topologies with alternating solid and void elements ordered in a checkerboard like fashion. The
problem is seen in many works on topology optimization of continuum structures and was one
of the reasons why early results on the material design problem published in Sigmund, (1994)1
only considered coarse discretizations where checkerboards were less prone to appear. The rea-
son for the occurrence of checkerboards is discussed in Díaz and Sigmund, (1994), supplied in
appendix 1, and can simply be explained by poor numerical modelling of the stiffness of checker-
boards by lower order finite elements. The mesh dependency problem refers to the non–conver-
gence of solutions with mesh refinement. Chapter three suggests a procedure that eliminates both
problems and is based on ideas borrowed from image processing. Applying the procedure can
be seen as applying a low–pass filter to eliminate structural variation underneath a certain length
scale. Although based on heuristics, the proposed procedure shows promising results when
applied to topology optimization of continuum structures.
 1. Numerical homogenization procedures
In this chapter a numerical procedure to determine effective thermoelastic properties of nonho-
mogeneous periodic materials will be described. The material microstructure represented by a
base cell is discretized in different ways, namely by micro truss–, frame– and continuum–like
discretizations. The advantages and disadvantages of the different discretizations will be dis-
cussed and illustrated by examples.

Section 1.1 describes the general formulas used in homogenization of linear elastic materials and
section 1.2 discusses the extension to homogenization of thermoelastic properties. In section 1.3
the homogenization problem is discretized using the finite element method. Section 1.4
describes details related to the actual implementation and advantages gained by using sparse
methods to solve the finite elements problem. Characteristics and implementation issues of the
different discretizations are discussed in section 1.5 and illustrated by examples in section 1.6.

Considering linear elasticity, material behaviour is governed by the generalized Hooke’s law
given as

s ij + E ijkl åkl – DT E ijkl a kl (1.1)

where s ij and å kl are the stress and strain tensors, E ijkl is the positive semi–definite elasticity ten-
sor, a kl is the coefficient tensor of thermal expansion and DT is the temperature change from
a reference temperature.

The generalized Hooke’s law (1.1) can also be written as

s ij + E ijkl åkl – DT b ij (1.2)

where b ij is a thermal coefficient related through the elasticity tensor to the coefficient of thermal
expansion a kl.

If we have a homogeneous material that is, a material with no microstructural variation, the mate-
rial behaves according to (1.2) subjected to some strain field or temperature change. For an inho-
mogeneous material the elasticity and thermal coefficient tensors E ijkl and b ij will fluctuate on
the microscopic scale and therefore make an accurate analysis of a structure composed of the
inhomogeneous material an overwhelming computational task. We are therefore interested in
determining an ”average” or homogenized measure of the thermoelastic properties. The homog-
enized coefficients can be used for the analysis of the global structure which means, that we can
ignore the microstructure in the global analysis and thereby simplify analysis considerably.
10 1. Numerical homogenization procedures

ijkl and b ij and can be found

The ”averaged” or homogenized material properties are denoted E H H

by several different methods.

Considering simple microstructures such as honeycombs or circular inclusions in matrix materi-

als etc., homogenized properties or good bounds can be determined analytically. The theory is
developed in Bensoussan, Lions and Papanicolau, (1978), Sanchez–Palencia, (1980) and Ciora-
nescu and Saint Jean Paulin, (1986), and examples of analytical determination of effective
parameters are seen in Bakhalov and Panasenko, (1989), Aboudi, (1991) or Nemat–Nasser and
Hori, (1993). Considering more complicated microstructures, analytical determination of
homogenized coefficients becomes an impossible task and numerically based methods must be
used. Examples of numerically based homogenization methods are the Finite Element (FE)
based numerical methods described in Bourgat, (1977) and Guedes and Kikuchi, (1990), Four-
ier–series approaches described in Walker, Freed and Jordan (1991) and Moulinec and Suquet
(1994) and boundary integral methods described in Greenbaum, Greengard and McFadden
(1993).

1.1 Homogenization of elastic properties

The numerical homogenization method is based on analysis of a base cell together with an
asymptotic expansion of the displacement field of the inhomogeneous material. A base cell is
the smallest repetitive unit of a periodic material as sketched in figure 1.1.

Figure 1.1 (left) shows a periodic material in two dimensions where the microstructure consists
of a matrix material with circular inclusions of another material and a three dimensional material
(right) consisting of fibres embedded in a matrix material. Two coordinate systems are defined
to describe the material behaviour on the macroscopic scale X and on the microscopic scale Y.

X X Y

Figure 1.1 Periodic materials in two and three dimensions represented by their base cells.
 1. Numerical homogenization procedures 11

The microstructure of a two or three dimensional material is thus represented by rectangular base
cells in 9 2 or 9 3 defined as

Y + ] 0, y 01 [ ] 0, y 0 [ (2 * d)
2 (1.3)
Y + ] 0, y 01 [ ] 0, y 0 [
2
] 0, y 0 [ (3 * d)
3

where y 01, y 02 and y 03 is the horizontal length, vertical length and depth of the base cells respec-
tively.

The microscopic displacement u of an inhomogeneous material can be represented by an asymp-

totic expansion with respect to the size of the base cell h and is written as

u + u 0(x, y) ) h u 1(x, y) ) h 2 u 2(x, y) ) AAA , y + xńh (1.4)

Only considering the first order terms of the asymptotic expansion and using the notation
å 0ij + 12 (ēu 0i ńēx j ) ēu 0j ńēx i) and å *ij + 12 (ēu 1i ńēy j ) ēu 1j ńēy i) we can write the strain å ij of an
inhomogeneous material as

å ij + å ij(u) [ å 0ij ) å *ij (1.5)

The microscopic strain field å ij is thus expressed as the sum of the strain field due to the average
displacement over the base cell å 0ij and the fluctuation strain å *ij due to the first order variation
or inhomogeneity of the material.

By letting the size of the base cell h go to zero and by choosing four linearly independent test
strain fields å 0ij defined as the four unit tensors å 0(11)
ij
+ (1, 0, 0, 0), å 0(22)
ij
+ (0, 1, 0, 0),
å ij + (0, 0, 1, 0) and å ij + (0, 0, 0, 1) in two dimensions and similarly with nine unit ten-
0(12) 0(21)

sors in three dimensions, it can be shown that the homogenized constitutive tensor can be written
as

EH 1
ijkl + Y ŕE ijpq
ǒå0(kl) *(kl)Ǔ
pq * å pq dY (1.6)
Y

where the fluctuation strain å *ij is the Y–periodic solution to the variational type problem

ŕE ijpq å ij(v) å pq
*(kl)
dY + ŕE ijpq å ij(v) å pq
0(kl)
dY ôvŮ V
(1.7)
Y Y

V + { v : v is Y * periodic }
12 1. Numerical homogenization procedures

Finding all the coefficients in the homogenized constitutive tensor E H ijkl, requires (1.7) to be
solved for the four linearly independent test strain fields å ij and the solutions å *ij to be substituted
0

into (1.6). Using the symmetries of the elasticity tensor, namely Eijkl = Ejikl = Eijlk = Eklij , the
number of test fields can be reduced to three in two dimensions (å 0(12)
ij
+ å 0(21)
ij
) and six in three
dimensions.

As mentioned in the introduction, construction of materials with prescribed constitutive parame-

ters can conveniently be treated as a topology optimization problem. The optimality criteria
methods used in topology optimization are based on energy expressions such as compliance or
eigenfrequencies, so expressing the constitutive parameters in energy terms, makes it possible
to use effective existing algorithms used in the topology optimization of trusses, plates, grillages,
etc.

Using that (1.7) is satisfied, the homogenized constitutive tensor can be expressed in the energy
bilinear form

1
ijkl + Y
EH ŕ ǒE pqrs ǒå0pq(kl) * å *pq(kl)Ǔ ǒå0rs(ij) * å*rs(ij)ǓǓ dY (1.8)
Y

The homogenization formulas (1.6) and (1.7) correspond to the standard numerical homogeniza-
tion formulas [e.g. Bourgat, (1977) or Guedes and Kikuchi, (1990)]. The standard equations are
derived for continuum structures under the assumption that no holes intersect the boundaries of
the base cell.

1.2 Homogenization of thermal properties

Parallel to the derivations of the homogenized elasticity tensor, it can be shown that the homoge-
nized thermal coefficient tensor b H ij can be written as

bH 1
ij + Y ŕE ijkl
ǒa kl – å CklǓ dY (1.9)
Y

where å C
kl is the Y–periodic solution to the variational type problem

ŕE ijkl å ij(v) å kl
C
dY + ŕE ijkl å ij(v) a kl dY ôvŮ V
(1.10)
Y Y

V + { v : v is Y * periodic }
 1. Numerical homogenization procedures 13

Equation (1.10) is very similar to the elasticity problem (1.7) – only difference is that the pres-
train in (1.10) is not a unit prestrain but a prestrain due to the variation in the thermal expansion
coefficient a kl. Solving equation (1.10) requires an extra analysis of the base cell. Using the Max-
well–Betti reciprocal theorem as described in Rodrigues and Fernandes, (1994), we can save this
extra analysis and reuse the results from equation (1.7).

Using the Maxwell–Betti reciprocal theorem we get that

ŕE ijkl å ij
C
dY + ŕbå *(kl)
ij ij dY (1.11)
Y Y

Using (1.11) the homogenized thermal coefficient tensor b H

ij can be written as

1
ij + Y
bH ŕb kl
ǒå0(ij)
kl
– å *(ij)
kl
Ǔ dY + Y1 ŕ apq Epqkl ǒå0(ij)
kl
– å *(ij)
kl
Ǔ dY (1.12)
Y Y

thus saving the computation of (1.10).

1.3 Finite element discretization

In the following, the homogenization procedure described in the last sections will be discretized
by the standard finite element method. The notations used will follow standard finite element
notation as defined in Cook, Malkus and Plesha, (1989). The derivations in this section are valid
for any type of element discretization. Details related to the specific finite element discretizations
will be discussed in section 1.5.

The base cell is discretized by NE finite elements of various types and geometries as will be dis-
cussed later. The equation for the homogenized constitutive tensor (1.8) can then be written in
discretized form as

ȍ ŕ ǒ{åe0}ij – {åe}ijǓT[E ]ǒ{åe0}kl – {åe}klǓ dYe

NE
1
ijkl + Y
EH (1.13)
e+1 Ye

where Ye is the area of the e’th element. Performing the integration over each element we get

ȍ ǒ{de0}ij – {de}ijǓT[se]ǒ{de0}kl – {de}klǓ

NE
EH
ijkl +1 (1.14)
Y
e+1

where the element stiffness matrix [s e] is defined as [s e] +

e
ŕ T
[B e] [E e][B e]dY e and {d e} is the
Y
local displacement vector related to the strain field {e} of element e by the strain–displacement
matrix [B e].
14 1. Numerical homogenization procedures

Solving the variational type problem (1.7) corresponds to solving a finite element problem of
the base cell subject to each of the test strain fields {å 0} ij . In FE–notation this problem can be
written as

[ S ]{D } ij + {R} ij (1.15)

where [S] is the global stiffness matrix assembled from the local stiffness matrices [s e], {D}ij is
the global displacement vector, and the load vector associated with test strain field å ij0 is defined
ŕ
as {R} ij + ȍ [Be] [E e]{å 0} ij dY e where the summation implicates the usual FE–assembly
NE
T
e+1 Y e
procedure.

The solution of the finite element problem (1.15), the displacement vector {D}ij , is subject to
periodic boundary conditions. The boundary conditions can be imposed in different ways as will
be discussed in section 1.4.

For reasons explained in the section on defining the inverse homogenization problem it is an
advantage to define the homogenized properties (1.14) in terms of element mutual energy densi-
ties

ȍ Qeijkl
NE
EH
ijkl + (1.16)
e+1

where the element mutual energy densities Q eijkl are defined as

Q eijkl + 1 ǒ{d e0} ij – {d e} ijǓ [s e]ǒ{d e0} kl – {de} klǓ

T
(1.17)
Y

Design of materials in chapter 2 only considers microstructures where the elasticity tensor is lin-
early dependent on the design variables x e. This feature makes the design problem well posed
for the optimality criteria procedures used in topology optimization and we can conveniently
write the homogenized elastic properties as

ȍ qeijkl xe
NE

ijkl +
EH (1.18)
e+1

where

q eijkl + 1 ǒ{d e0} ij – {d e} ijǓ [s e0]ǒ{d e0} kl – {de} klǓ

T
(1.19)
Y
Here it was assumed that the element stiffness matrices are linearly dependent of the design vari-
ables such that [s e0] + [s e]ńxe where x e is the e’th design variable.
 1. Numerical homogenization procedures 15

Similar to the discretization of the computation of the homogenized elastic tensor, we can write
the homogenized thermal coefficients as

+ ȍ q eij x e
NE
bH
ij (1.20)
e+1

where

q eij + Y NJb e0Nj [B ]ǒ{d e0} ij – {d e} ijǓ + Y NJa e0Nj ƪE 0ƫ[ B ]ǒ{d e0} ij – {de} ijǓ ,
e T e T
(1.21)
Y Y

{b 0} + {b e}ńxe and [E 0] + [E e]ńx e and it was used that we are considering bilinear elements.

1.4 Solution procedures for the finite element problem

This section discusses ways to impose the periodic boundary conditions on the cell problem and
shows how computer time requirements can be reduced by using an iterative element–by–ele-
ment finite element solver.

The periodic boundary conditions in the FE–problem (1.15) can be imposed in different ways.
The simplest way is to use a penalty method described in Cook, Malkus and Plesha, (1989). The
basic idea in the penalty method is to add a very stiff (2–node) element connecting the linked
degrees of freedom thus forcing the relative displacements of the two nodes to be equal. Another
method is to give opposing nodes the same node number. This method has the advantage of redu-
cing the number of degrees of freedom of the system but it also makes the housekeeping with
the node numbering scheme more complicated.

Both above mentioned methods have the big disadvantage that the bandwidth of the global stiff-
ness matrix [S] becomes almost equal to the number of Degrees Of Freedom (DOF) of the linear
system. Apart from making computations time expensive and frustrating, the large bandwidth
makes storage requirements huge. As an example, consider a base cell modelled by 30x30=900
4–node quadrilateral finite elements. The number of DOF for this system is 1800 making the size
of the stiffness matrix approximately equal to 26 Mb. The huge storage requirement for such a
relatively small linear system makes the use of sparse methods promising.

One way to ”overcome” the storage and computer time problem is to use a so-called skyline
solver. In this approach only non–zero elements of the global stiffness matrix are stored. For a
linear system with large variations in the bandwidth (as for the base cell), the skyline solver can
be very efficient.
16 1. Numerical homogenization procedures

Iterative linear equation solvers belong to another class of sparse methods. Iterative methods are
especially attractive for use in optimization problems. In the beginning steps of an optimization
procedure we do not need an exact solution for the displacement vector. The convergence
requirements for the iterative solver can therefore be relaxed. In fact, this idea was fully taken
advantage of by Sankaranarayan, Haftka and Kapania, (1992) or Ringertz, (1988), where the
exact solution of the FE problem and the optimal values of the design variables only are found
in the very last iteration.

In this work we will use the so-called Element–by–Element Preconditioned Conjugate Gradient
solver (EBE–PCG). The general PCG method is described in Numerical Recipes, Press et al.,
(1992). The EBE–PCG method is discussed and applied to FE analysis in Ferencz, (1989) and
applied to the analysis of human bone structure discretized by up to one million finite elements
in Hollister and Riemer, (1993).

The iterative algorithm of the EBE–PCG solver is described in appendix 3. Shortly summarized
it solves the NxN linear system

[ S ]{D } + {R} (1.22)

by minimizing the function

f ({ D }) + 1 {D } [ S ]{ D } – {R} T{D }
T
(1.23)
2

which means that we are physically back in the minimization of potential energy in the linear
elastic body.

The function f (potential energy) has a stationary point when the gradient

ʼnf + [S ]{ D } – {R} (1.24)

is zero which corresponds to (1.22).

The minimization is performed as an iterative procedure computing the search direction {p}k and
improved minimizers {D}k . At each step k in the iteration a factor  is determined such that
f ({ D } k ) a {p} k) is minimized. The new minimizer {D}k+ 1 is then set equal to the new point
{ D } k ) a {p} k .
 1. Numerical homogenization procedures 17

The computationally heavy part of the EBE–PCG algorithm is the multiplication of the global
stiffness matrix [S] and the search direction vector {p}. However, this multiplication can be done
on the element level as

ȍ [se]{pe}k
NE
{v} k + (1.25)
e+1

using the appropriate assembly procedure from local to global coordinates. As seen in equation
(1.25), the global stiffness matrix never has to be assembled – this feature makes the requirement
to storage space very small in the EBE–PCG solver. Furthermore, the procedure is independent
on the bandwidth of the FE–problem again because the global stiffness matrix is never
assembled.

In practice, we do not require an exact solution to the FE–problem and we can therefore save
considerable computational time by stopping the iterative solver when the norm of the residual
vector gets below a stopping criteria d. In fact d can be chosen very gross when the EBE–PCG
algorithm is used as a part of an iterative optimization procedure. Only in the final iterations of
the optimization procedure we require an exact solution and then d can be made smaller for final
convergence. Furthermore old solutions of the FE–problem can be used as starting guesses for
the new iteration, making further savings in computational time possible.

Numerical experience shows that the EBE–PCG solves the FE–problems considered in this work
very efficiently. Using regular continuum element discretizations, the storage space for a FE–
problem with 10.000 bilinear elements is less than 1 Mb (vector of element densities, one element
stiffness matrix and some work–area for the solver). In the beginning iterations, computational
time is comparable to time requirements of conventional solvers (for problems without periodic
boundary conditions) but as the optimization procedure converges, the computational time using
the iterative FE–solver is reduced by up to a factor of 20. Further details of the computational
procedure are described in appendix 3.

1.5 Discretizations of microstructure

Microstructures of materials can vary from very thin or porous microstructures such as sponges
or foams over solid metals with no porosity, to mixtures of materials such as carbon fibers
embedded in a matrix. To analyze various microstructures properly we therefore have to use dif-
ferent discretizations to model different material types. In this section three, different discretiza-
tions, namely truss–, frame– and continuum–like microstructures are described.

All the considered microstructures in this section are based on the ”ground structure” idea. This
18 1. Numerical homogenization procedures

means that the number of elements to discretize the base cell has a fixed large number. Specific
microstructures can then be built by assigning different properties such as area, Young’s modu-
lus, thickness or material type to each element in the base cell. The properties of elements which
are non–important for the specific microstructures are set to small values such that they do not
have significant influence on the computed properties. This procedure will be demonstrated by
a simple example later in this section. Using the ground structure approach we get the following
advantages

S The program can be standardized independent on the considered discretization

and material topology

S Existing methods from topology optimization of structures can be applied to the

material design problem

S The requirement that no holes can intersect the boundaries of the base cell will
always be fulfilled because no elements are removed from the initial structure

S Regular meshes for continuum modelled microstructures allow the use of image
processing techniques for pre– and post–processing of the resulting designs

The disadvantage is, that storage and computer time requirements are increased because of the
large number of elements which are passively present in the structure.

Truss–like microstructures
As one of the aims of this work is to design materials with extreme elastic properties and such
materials are expected to show mechanism type behaviour (see chapter two), we will, as one of
the discretizations of base cells, use a truss like–microstructure which can model such behaviour.
A truss structure is defined as a structure composed of bars which can only transmit longitudinal
forces. Examples of truss–like microstructures in two and three dimensions are shown in figure
1.2 and 1.3 respectively.

Figure 1.2 (right) shows a two dimensional base cell with 16 nodal points connected by bar ele-
ments – outside the nodal points, bars are allowed to cross each other without interaction.

The standard homogenization formulas discussed in section 1.1 were derived for continuum
structures. Therefore, it is necessary to consider the truss structure as a continuum structure with
holes which is done by treating the individual bar members as 2–node continuum elements as
described in appendix 4.

As an example of how the groundstructure approach is used in the truss–like microstructure, we

will show how to analyze a simple octagonal microstructure using the base cell in figure 1.2 as
 1. Numerical homogenization procedures 19

X
Figure 1.2 Composite material composed of a truss–like microstructure in two dimen-
sions. Nodal points are illustrated by small circles.

X Y

Figure 1.3 Composite material composed of a truss–like microstructure in three dimensions.

Nodal points are not shown

groundstructure. Figure 1.4 illustrates how the important members for the octagonal cell are
given an area Abar (thick bars) and the unimportant bar members are given a small area Amin (thin
bars). The minimum area Amin is chosen such that the homogenized coefficients are unaffected
by their presence (usually Amin /Abar = 10–7).

The truss structure has two DOF per node in two dimensions and three DOF per node in three
dimensions (two and three translational DOF respectively) thus the element stiffness matrix [s e]
has dimension 4x4 in two dimensions and 6x6 in three dimensions.
20 1. Numerical homogenization procedures

Abar

Amin

Figure 1.4 Modelling of a simple

octagonal microstructure using the
groundstructure approach.

Frame–like microstructures
Modelling porous materials as micro truss structures might be criticized in a couple of ways.
First, no known naturally occurring material has moment free links in the microstructure and
second, for the design of material purpose, it will be very difficult to build links at the microlevel.
For these reasons we will also consider base cells discretized by frame elements. A frame struc-
ture is defined as a structure composed of beam elements transferring longitudinal forces and
moments and connected by stiff joints. An example of a frame–like microstructure is shown in
figure 1.5. The frame structure has three DOF per node (two translational and one rotational) thus
the element stiffness matrix [s e] has dimension 6x6 in two dimensions.

Y
X

Figure 1.5 Composite material composed of a frame–like microstructure in two dimensions.

 1. Numerical homogenization procedures 21

Continuum–like microstructures
Truss and frame–like microstructures can only be used for the modeling of porous materials.
Considering materials with small voids or mixtures of different materials, a continuum–like
discretization of the base cell must be used. In the continuum–like discretization the base cell
is discretized by four node bilinear finite elements. The discretization can be characterized as
a pixel–like microstructure where each pixel represents one finite element. Choosing a regular
discretization as shown in figure 1.6, all pixels or elements will have the same geometrical prop-
erties. This feature makes the whole problem very cheap in storage space because we only have
to calculate and store the element stiffness matrix [s e] once for each material type used in model-
ling the microstructure. This advantage was fully utilized by Hollister and Riemer, (1993) where
they calculated three dimensional human bone structure discretized by over one million 8–node
brick elements.

Figure 1.6 shows a microstructure consisting of a (strong) fiber embedded in a (soft) matrix and
discretized by square finite elements. The enlargement (figure 1.6 right) shows the resulting
step–like boundaries between the two materials which is due to the pixel–like discretization. It
might be objected that this ”jagged” discretization will lead to erroneous results but examples
in the next section will show that the effective properties are in fact quite insensitive to the
”jagged” boundaries.

Y
X
Figure 1.6 Composite material composed of a continuum–like microstructure in two dimensions.
22 1. Numerical homogenization procedures

1.6 Examples
To illustrate the difference in modelling a porous microstructure by a frame–like or a continuum–
like discretization and to discuss the consequence of modelling smooth edges by step–like edges
in the continuum discretization, we will use two ”famous” examples of porous microstructures,
namely the honeycomb and the inverted honeycomb.

The Perfect Honeycomb

Figure 1.7 shows the base cell and the dimensions of the perfect honeycomb. Choosing a cell with
a=3–3/4, the area of the base cell becomes unity. In this example the ratio of wall thickness to cell
size is t/a= Ǹ3 ń12, the Youngs modulus of the isotropic base material is 0.91 and the Poisson’s
ratio  is 0.3.

Figure 1.8 shows three different continuum discretizations used to model the base cell. Areas
with material are shown in black and white areas should be void. However, due to the ground
structure approach, white areas actually are modelled by weak material with a very low stiffness
(Evoid /Esolid =10–4). The resulting homogenized elastic properties from the three discretizations
are shown in table 1.1.

3a
a

2t
Ǹ3 a

60o

t
Figure 1.7 Dimensions of perfect honeycomb.

a) b) c)

Figure 1.8 Three different discretizations of the perfect honeycomb: a) 21x12=252 elements, b) 63x36=2268
and c) 126x72=9072 elements.
 1. Numerical homogenization procedures 23

Size CPU–time  E1111 E2222 E1212 E1122 

(elements) (seconds)
21x12=252 2 0.310 0.0912 0.0877 0.0121 0.0691 0.77
63x36=2268 58 0.305 0.0934 0.0940 0.0122 0.0703 0.75
126x72=9072 309 0.306 0.0958 0.0950 0.0125 0.0713 0.75
PREMAT2D, 9520 – 0.306 0.0968 0.0968 0.0124 0.0720 0.74
Frame model, 8 – 0.334 0.0961 0.0962 0.0129 0.0703 0.73

Table 1.1. Resulting homogenized properties of the perfect honeycomb calculated by continuum and
frame–like microstructures and compared with results from the program PREMAT2D supplied by J.M.
Guedes.

The computed effective properties for the three discretizations are compared with values
obtained by running the program PREMAT2D developed and kindly supplied by José M.
Guedes, Institute Superior Technico, Portugal [see Guedes and Kikuchi, (1990) and Guedes,
Neves and Orlando, (1994)]. The program PREMAT2D uses the same numerical homogeniza-
tion procedure as described in this chapter and uses a mesh generator which can model bound-
aries between solid and void (or two materials) smoothly, furthermore the program uses no ele-
ments in the void regions. PREMAT2D was used to calculate the effective properties of the
perfect honeycomb discretized by 9520 4 node linear displacement finite elements.

Comparing the effective properties obtained by the pixel approach in this report with the effec-
tive properties from PREMAT2D, we notice a reasonable agreement even for the coarse mesh
shown in figure 1.8a which consists of only 252 elements. The error relative to the PREMAT2D
results is less than 10%. From this example we can conclude that effective properties are pre-
dicted well even for very coarse meshes with badly ”jagged” edges by the proposed pixel–like
discretization.

The last row in table 1.1 shows the effective properties obtained by discretizing the perfect hon-
eycomb with frame elements. Fine agreement with the results from PREMAT2D is seen by
comparing the effective stiffnesses. However, there there is an error in the prediction of the den-
sity r for the frame case This error can be explained by the fact that the frame modelling allows
overlapping and therefore areas where two frame elements overlap are counted twice when the
total material density is calculated.

It can be observed that the perfect honeycomb is an isotropic material because E1212 =
(E1111+E1122)/2 and E1111 = E2222 but it also follows from the fact that the microstructure has
600 symmetry.
24 1. Numerical homogenization procedures

The inverted honeycomb

In this example we will look at the influence of the volume fraction on the effective properties.
The inverted honeycomb which has a negative Poisson’s ratio was first reported by Almgren,
(1985), Kolpakov, (1985) and Gibson and Ashby, (1988). Figure 1.9 shows the inverted honey
comb with three different volume fractions and using 4032 square elements for the discretiza-
tions. The microstructure was also modelled by frame elements. The resulting effective proper-
ties are shown in table 1.2.

A FE–mesh for the computation of the homogenized properties of the inverted honeycomb is
not directly available in PREMAT2D (although customized FE meshes can be generated) but as
in the last example, good agreement (especially for low densities) between continuum and
frame–like discretization is seen in table 1.2.

Some observations can be made on the effective properties of the inverted honeycomb. The
material does not fulfill the requirements for isotropy because neither is E1212 = (E1111+E1122)/2
nor is E1111 = E2222. In fact the shear stiffness E1212 is close to zero and the inverted honeycomb
is therefore a highly anisotropic material. Most of the advantages of materials with negative Pois-
son’s ratios discussed in the introduction are lost because of the anisotropy of the inverted honey-
comb.

t/a  E1111 E2222 E1212 E1122 n n

Ǹ3 ń12 (continuum) 0.397 0.0692 0.1011 0.0033 –0.0457 –0.66 –0.45
Ǹ3 ń12 (frame) 0.445 0.0694 0.0817 0.0021 –0.0507 –0.73 –0.62
Ǹ3 ń24 (continuum) 0.210 0.0299 0.0342 0.0004 –0.0274 –0.92 –0.80
Ǹ3 ń24 (frame) 0.222 0.0322 0.0338 0.0003 –0.0296 –0.92 –0.88
Ǹ3 ń48 (continuum) 0.109 0.0139 0.0144 0.0001 –0.0135 –0.97 –0.94
Ǹ3 ń48 (frame) 0.111 0.0157 0.0160 0.0000 –0.0154 –0.98 –0.96

Table 1.2. Resulting homogenized properties of the inverted honeycomb calculated by continuum and
frame–like microstructures for different volume fractions of material to void.

a) b) c)

Figure 1.9 Three different volume fractions of the inverted honeycomb. Groundstructure is 84x48 elements.
 1. Numerical homogenization procedures 25

1.7 Summary
A numerical homogenization procedure is described and applied to truss–, frame– and continu-
um–like discretizations of microstructures. The advantages gained by using sparse methods in
solving the associated finite element problem is discussed and the EBE–PCG solver is chosen
due to its iterative nature which can be taken fully advantage of when it is applied as solver in
an iterative optimization procedure. The homogenized coefficients are expressed in a form that
make them well suited for topology optimization methods. The proposed procedures and discre-
tizations are applied to the homogenization of two practical microstructures and the results are
compared with those obtained by the software PREMAT2D and show good agreement. Bounda-
ries in the continuum formulation appear as jagged edges due to the pixel–like discretization and
this could be expected to give erroneous results. However, numerical evidence shows only small
errors in the homogenized properties even for very coarse meshes which is of importance when
the discretization is used as groundstructure in the design of materials in chapter two.
 2. Design of materials
In this chapter the problem of how to design a material with prescribed thermoelastic properties
is considered. The design problem is formulated as an inverse homogenization problem and is
solved by an optimality criteria method. The design domain is the microstructure of a periodic
material represented by base cells and discretized by various types of finite elements.

The problem of how to design a material with prescribed thermoelastic properties can be
approached in different ways as discussed in the introduction. We can try to utilize existing
(fiber) materials the best possible way by varying lamination parameters, such as angles or thick-
nesses [e.g. Autio, Laitinen and Pramila, (1993)], or we can take the mathematical approach and
design ranked materials consisting of laminations at several length scales [e.g. Milton and Cher-
kaev, (1993)]. By the former approach, we can not design materials with extreme thermoelastic
properties but on the other hand the results are directly manufacturable. By the latter approach,
we can design materials with any thermodynamically admissible properties. However, they will
be difficult if not impossible to manufacture, because of the assumption of several length scales
in the microstructure. In this chapter, we will try to design materials with extreme elastic proper-
ties but restricting the design space to microstructures with only one length scale.

The material design problem can be formulated as an inverse homogenization problem; figure
2.1 illustrates the problem.

As described in chapter one, the usual homogenization problem is the following: given a material
with periodic inhomogeneous microstructure (figure 2.1 right), we seek the effective homoge-
nized properties E H ijkl (figure 2.1 left). These are found by analysis of the representative base cell
(figure 2.1 center). The inverse homogenization problem is just the opposite. Given the effective
properties E H
ijkl (figure 2.1 left), we seek a material (figure 2.1 right) which has these effective
properties. The problem can be solved by designing the microstructural topology (figure 2.1 cen-
ter).

EH
ijkl
?
Figure 2.1 Illustration of the inverse homogenization problem
 2. Design of materials 27

Generally many different material microstructures can have the same constitutive properties.
Consequently, we will expect the inverse homogenization problem to have several different solu-
tions. To reduce the number of possible solutions, the inverse homogenization problem can be
recast as an optimization problem, where the goal is to find the lightest material which has the
prescribed elastic properties. The goal of minimizing weight is chosen because we are generally
interested in designing materials with high stiffness–to–weight ratios as discussed in the
introduction. The design domain of the optimization problem is the periodic material microstruc-
ture represented by the base cell.

The optimization problem can be seen as a topology optimization problem, where the design
domain is the interior of the base cell, and the design variables are the dimensional properties
of the finite elements used to discretize the base cell. By formulating the inverse homogenization
problem as a topology optimization problem, we can use some of the well proven algorithms used
in topology optimization of general structures; the reader is referred to Bendsøe, (1994), for an
overview of methods.

Basically topology optimization consists in finding the optimal topology of a structure supported
at its boundaries and subjected to some load case. The optimal topology is found within the
design space defined by the groundstructure. The groundstructure can consist of different types
of finite element discretizations, and should be chosen with so many design degrees of freedom
as needed to ensure a good solution. During an iterative procedure, unimportant elements are
removed from the groundstructure and the procedure is iterated until an optimum (usually con-
sisting of only a small fraction of the elements of the original groundstructure) has been reached.
For computational simplicity of this work, unimportant elements are not removed but their
dimensional properties are made small such that they have no structural significance. This
means, that we are actually considering a sizing problem because the topology is fixed, defined
by the groundstructure. However, due to structural insignificance of the unimportant elements,
the approach is stated as a topology optimization problem.

The best discretization type used in the groundstructure is dependent on the type of problem con-
sidered. Here, we will use truss, frame and continuum elements to model material microstruc-
tures for the following reasons:

Materials with extreme elastic behaviour are expected to show mechanism type behaviour. An
example of an extreme material could be a two dimensional material with Poisson’s ratio equal
to one. The constitutive tensor for this material has only one non–zero eigenvalue and therefore,
the material will have ”zero–stress modes” where it can collapse with no structural resistance.
Groundstructures consisting of truss elements in two or three dimensions (figures 2.2a and 2.2b
28 2. Design of materials

respectively) can model mechanism type behaviour and are therefore good choices when we
want to design extreme materials.

For the design of more realistic (stable) materials, groundstructures made up of frame elements
(figure 2.2c) can be used. Frame elements can take up bending moments and in the limit, when
the thicknesses approach zero, they approach truss behaviour.

For the design of materials with high densities and only small holes, the frame element discretiza-
tion is inadequate and should be replaced by a continuum type element discretization (figure
2.2c).

In the truss formulation the individual bar cross–sectional areas are the design variables, in the
frame formulation element widths are chosen as the design variables and in the continuum for-
mulation, the density of material in each element represents the design variables.

a) b)

c) d)
Figure 2.2 The three different groundstructures used to model the microstruc-
tures. a) truss–like (two dimensions), b) truss–like (three dimensions), c)
frame–like and d) continuum–like microstructure.
 2. Design of materials 29

Common for the three different types of groundstructure formulations is, that the homogenized
elastic properties are assumed linearly dependent on the design variables (see chapter one). The
design problem can therefore be formulated in a standard way independent on the choice of
microstructure.

This chapter is structured as follows. In section 2.1, the inverse homogenization problem is for-
mulated as an optimization problem, which is common for the four different discretizations
shown in figure 2.2. In section 2.2, we describe the optimality criteria method used to solve the
optimization problem and discuss details related to the computational procedure. In sections 2.3,
2.4 and 2.5 a number of examples using the three different types of groundstructures (truss– 2d
and 3d, frame– and continuum–like microstructures, respectively) are shown and discussed. In
section 2.6 we design continuum–like materials with prescribed thermoelastic properties and
finally, in section 2.7, we verify the numerical results by simple macro size working models.

2.1 The inverse homogenization problem

In this section, the inverse homogenization problem is defined as an optimization problem of
minimizing the weight of the base cell subject to equality constraints on the elastic properties.

As described in chapter one, the homogenized elastic properties E Hijkl of a periodic material can
be found by finite element analysis of the base cell. Under the assumption that the base cell is
discretized by NE finite elements and that the homogenized properties are linearly dependent on
the design variables x e, the homogenized elastic tensor was written as

ȍ qeijkl xe
NE
EH
ijkl + (1.26)
e+1

where q eijkl are the relative element mutual strain energies defined in (1.19).

Denoting the prescribed elastic tensor E *ijkl, we can write the equality constraint for the optimiza-
tion problem as

ȍ qeijkl xe + 0
NE
E *ijkl – (2.1)
e+1

Equation (2.1) represents six linearly independent equality constraints in two dimensions and
21 in three dimensions. In most of the examples in this chapter we are considering symmetric
(implying orthotropicity) microstructures. Implying the symmetry conditions by element link-
ing, the number of equality constraints in (2.1) can be cut down to four and nine in two and three
dimensions respectively. Definitions of notations used for the elasticity tensor and its values for
specific material behaviour (isotropic, orthotropic, plane stress etc.) are given in appendix 2.
30 2. Design of materials

Making use of the abbreviation ijkl ³ I defined by

1111 ³ 1, 2222 ³ 2, 1122 ³ 3, 1212 ³ 4 (2–d)

1111 ³ 1, 2222 ³ 2, 3333 ³ 3ȣ (2.2)

1122 ³ 4, 1133 ³ 5, 2233 ³ 6Ȧ (3–d)
1212 ³ 7, 1313 ³ 8, 2323 ³ 9Ȥ

equation (2.1) can be written in a simpler form as

ȍ qeI xe + 0
NE
E *I – , I + 1,..., NC (2.3)
e+1

where NC is the number of equality constraints (NC=4 in 2–d and NC=9 in 3–d).

Considering porous microstructures, the material density  of the homogenized material can be
defined as the ratio between the volume of solid material in the base cell and the total volume
of the base cell. The density  is therefore bounded to the domain ]0, 1], where =1 represents
solid material and =0 means no material at all.

Assuming that the total volume of the base cell is unity, the density of a specific microstructure
can be calculated as

ȍ e xe
NE
+ (2.4)
e+1

where  e is a ”volume” factor which multiplied by the design variable x e gives the volume of
material in element e. The actual weight of a base cell can then be found by multiplying the den-
sity  by the weight of a solid base cell.

As previously defined, the goal of the optimization problem is to minimize the weight of the base
cell (with constraints on the elastic properties). A minimization problem with (2.4) as the objec-
tive function could now be defined. However, before we do this, we will try to incorporate some
manufacturability constraints into the objective function.

It is a well known fact in the field of topology optimization, that we get complicated solutions
or solutions with many intermediate values of the design variables, if we do not use some sort
of penalty formulation. Choosing proper penalization factors, we can get simple and manufactur-
able solutions and we have the possibility to choose between solutions with different topological
features as will be explained in the following.
 2. Design of materials 31

A penalization of specific design variables can generally be implemented by modifying the den-
sity function (2.4) to

ȍ e e (xe)1ńp
NE
+ (2.5)
e+1

where the physical meaning of the penalty factors  e and p as well as the meaning of the ”vol-
ume” factor  e is related to the considered microstructure as follows:

Truss–like materials (figure 2.2a and 2.2b):

The design variable x e is the cross sectional area of element e and therefore  e is defined as the
length l e of the element. For manufacturing reasons we will introduce the penalizing term

ǒl prefńl eǓ , which makes it possible to penalize solutions with long or short elements. l pref is a
preferred length of the elements, and elements longer or shorter than l pref will be penalized by
choosing  positive or negative. The density function for two dimensional truss–like materials
can be written as

 + ȍ l eǒl prefńl eǓ x e
NE 
(2.5)truss2d
e+1

For truss–like materials in three dimensions we will multiply by another penalizing term  e,
which makes it possible to penalize elements on the surface of the 3–d base cell. Choosing  e
smaller or greater than one for elements on the surfaces makes it possible to prefer one solution
from another. By choosing  e smaller than one, we get solutions with only interior elements and
vice versa.

ȍ leǒlprefńleǓ e xe
NE
+ (2.5)truss3d
e+1

In the example section 2.3, different choices of penalty factors and their influence on the solu-
tions will be discussed in more detail.

Frame–like materials (figure 2.2c):

The design variable x e represents the width of element e and therefore  e for this case is defined
as the height h e of element e times the length l e. We will use the same penalization of long or short
elements as for the truss–like materials such that the density function can be written as

 + ȍ h e l e ǒl prefńl eǓ x e
NE 
(2.5)frame
e+1
32 2. Design of materials

Continuum–like materials (figure 2.2d):

In this case, the design variable x e represents the density of material in element e, where x e is
restricted to the interval ]0, 1] and  e is the volume Ye of element e. The design problem is likely
to have solutions with areas of intermediate values of the design variable x e. Such ”grey” areas
are undesirable because we want to make porous microstructures. As discussed in chapter three
of this report (on general layout optimization of structures), solid/void solutions are obtained by
introducing a penalty factor p, such that the modified material density can be written as

ȍ Ye (xe)1ńp
NE
+ (2.5)continuum
e+1

By choosing p>1 we make intermediate values of x e artificially ”heavy” and thus unpreferable
in the optimal solution.

All material models

The four different forms of the modified density functions for the different microstructure discre-
tizations are all contained in the general density expression (2.5).

Using the objective function (2.5) and the equality constraints (2.3) we can now state the general
optimization problem as

Minimize :  + ȍ  e  e (x e)1ńp
NE

e+1

ȍ qeI xe + 0
NE
Subject to : E *I – , I + 1,..., NC (2.6)
e+1
: 0 t x min v x e v xmax , e + 1,..., NE
and : equilibrium equations

where x min and x max are lower and upper limits on the design variables respectively.
 2. Design of materials 33

2.2 Solving the optimization problem

The optimization problem (2.6) can be solved by many different numerical optimization algo-
rithms. However, the problem has many design variables but few global constraints and is there-
fore well suited to be solved by some of the efficient optimality criteria algorithms used in topo-
logy optimization. In defining the optimality criteria algorithm, the mutual influences from
element to element is ignored which means that the design variables are updated independently
in each iteration step. The method is widely used for large scale structural optimization problems,
(more because of its computational efficiency than its mathematical rigour).

The optimality criteria algorithm described here includes multiple equality constraints such that
the procedure must be divided in to an inner and an outer problem. The outer problem consists
in the update of the design variables using a fixed–point type algorithm and the inner problem
consists in determining the values of the multiple Lagrangian multipliers. The inner problem is
solved by a standard Newton–Raphson procedure as proposed in Zhou and Rozvany, (1993).

The Lagrangian function for the optimization problem (2.6) can be written as

ȍ e e (xe)1ńp ) ȍ IȡȧE*I – ȍ qeI xeȣȧ)

NE NC NE
L+
e+1 I+1 Ȣ e+1 Ȥ
(2.7)
ȍ ȍ
NE NE
e ǒ– xe ) x min
Ǔ) e ( xe – xmax)
e+1 e+1

where  I are the NC Lagrangian multipliers for the NC equality constraints and e and e are
the NE Lagrangian multipliers for the NE lower and upper side constraints respectively. In (2.7)
we have omitted the equilibrium equations as they do not have any influence on the updating
scheme for the design variables. For the correct handling of equilibrium equations in the Lagran-
gian function, see chapter three. Here we will assume that they are fulfilled as the solutions of
the finite element problem.

Stationarity of the Lagrangian function with respect to the design variables gives us the optima-
lity criteria

ȍ I qeI –
NC
ēL + 1  e  e (x e) (1ńp–1) – e
) e
+ 0 , e + 1, ... , NE (2.8)
ēx e p
I+1
34 2. Design of materials

On the basis of (2.8), we can formulate an updating scheme for the design variable x e. Assuming
that the lower and upper side constraints are not active (i.e. e= e= 0), equation (2.8) can be
written as

ȍ I qeI
NC

I+1
B ek() + + 1 , e + 1, ... , NE (2.9)
1
p  e  e (x e) (1ńp–1)

Following Bendsøe, (1994), we can define the fixed point type updating scheme

x ek)1 + x ek ǒ B ek() Ǔ (2.10)

where is a damping factor and subscript k denotes the iteration number.

The updating scheme for the fixed point type algorithm is defined as

ȡmax {(1– ) xek, xmin } if xek ǒ B ek() Ǔ v max {(1– ) xek, xmin }
ȧe e
x k)1 + ȥx k ǒ B k() Ǔ if max{(1– ) x ek, x min } v x ek ǒ B ek() Ǔ v min{(1– ) x ek, x max }
e
(2.11)
ȧmin{(1 ) ) xe, x } if min {(1 ) ) xe, x } v xe ǒ B e() Ǔ
Ȣ k max k max k k

where is a move limit.

Determination of the Lagrangian multipliers  I has to be done iteratively in an inner loop. In

order to get an implicit set of equations for the Lagrangian multipliers, we use that the equality
constraints (2.3) should be satisfied for the new values of the design variables x ek)1

ȍ qeI xek)1 + 0
NE
E *I – , I + 1,..., NC (2.12)
e+1

Substituting the updating formula (2.11) into the equality constraints (2.12), we get

I + E *I – ȍ q eI x ek ǒ B ek() Ǔ – ȍ q eI x ek)1 + 0 , I + 1,..., NC (2.13)

eŮRact eŮRpas

where I denotes the residual in obtaining the I’th equality constraint, Ract denotes the group of
elements governed by the center updating rule in (2.11) and Rpas denotes the group of elements
 2. Design of materials 35

which are governed by the lower or upper side constraints xmin or xmax . To find the set of Lagran-
gian multipliers that minimize the residuals I, we will use the standard Newton–Raphson itera-
tive procedure. The recurrence formula is

–1
 n)1 +  n – ǒʼn Ǔn n (2.14)

where n is the iteration number of the Newton–Raphson algorithm and ʼn is the NCxNC Jaco-
bian matrix. The individual elements of ʼn are given by

ēB ek()
ē I
ē J
+– ȍ q eI x ek ǒ B ek( n) Ǔ
ǒ –1 Ǔ

ē J
(2.15)
eŮRact

and

ēB ek() q eJ
+1 (2.16)
ē J p   (x )
e e e (1ńp–1)

The Newton–Raphson procedure is iterated until satisfactorily accuracy in obtaining the equality
constraints (2.12) is obtained, i.e., the inner loop is stopped when Ť IŤ t  for I=1,..., NC, where
 is a small number.

The overall computational procedure is summarized in figure 2.3.

The above described computational procedure is also applicable to the design of materials with
prescribed thermoelastic properties. The only change is, that the three extra equality constraints

ȍ qeij xe + 0
NE
*
ij – (2.17)
e+1

are added to the optimization problem (2.6). As the equality constraints (2.17) are in the same
form as those for the elastic properties (2.1), the computational procedure is again given by figure
2.3, with the addition of three extra equality constraints such that NC=7 in the two dimensional
case.

Convergence of the optimization procedure is generally quite stable. For the two dimensional
truss problems, 20–30 iterations are needed to converge (with six digits accuracy in the change
in design variables). For the three dimensional truss and the frame–like microstructures, conver-
gence is generally a little slower and for the continuum–like microstructures up to 100 iterations
are needed. The actual convergence rate is dependent on the choice of move limit and damping
factor in the updating scheme (2.11). For the truss and frame models, 0.02 and 0.8 respectively,
36 2. Design of materials

are good choices and for the continuum model 0.05 and 0.4 give stable convergence. Prescribing
impossible elastic properties (e.g. elastic tensor non–positive definite) or prescribing elastic
properties which are non–attainable by the chosen groundstructure, results in ”jumping” of the
design variables and lack of convergence, indicating that the operator should ”relax” the pre-
scribed elastic properties.

In the following sections, we will show how materials with a wide range of thermodynamically
admissible properties can be designed by the proposed procedure. In the two dimensional exam-
ples we will assume plane stress state when we refer to the isotropic materials. In the figures, the
elasticity tensor is written in a matrix form which is defined in appendix 2 together with defini-
tions for different ”families” of materials.
Initialize

Factorize stiffness matrix for

homogenization
discretized model of base cell

Solve for 3/6 prestrain cases given by (1.15)

Calculate element mutual energies by (1.19)

Update design variables by (2.11)

Update Lagrangian multipliers  by (2.14)

no  converged ?

yes

no xe converged ?

yes
plot microstructure

stop
Figure 2.3 Flowchart of the optimization algorithm
 2. Design of materials 37

2.3 Truss–like materials

This section presents a number of examples obtained from groundstructures discretized by truss
elements. It is shown that materials with a variety of properties, including materials with Pois-
son’s ratios close to –1 and 0.5 (–1 and 1 in 2–d), can be constructed as truss modelled microstruc-
tures in two and three dimensions.

The presented results are limited to orthotropic materials with axis of symmetry coalligned with
the coordinate system, although experiments show that all kinds of anisotropic and orthotropic
materials with principal axes oriented in different directions can be constructed with the pro-
posed algorithm.

In the 2–d examples all microstructures are constructed from a 4 by 4 node full ground structure
(known from topology optimization), with all nodes connected by bar elements, i.e. there are 120
potential bar members when overlapping is allowed (figure 2.2a). The groundstructure size of
4 by 4 nodes seems to be the ideal size of ground structure. Smaller ground structures do not have
enough degrees of freedom to model complicated micro mechanical behaviour, and larger
ground structures tend to result in more complicated topologies and are not so easily interpret-
able. In the 3–d examples a 4 by 4 by 4 ground structure (figure 2.2b) is chosen for the same
reasons as in 2–d. Connecting all nodes by bars gives 2016 potential bar members.

The aim of the optimization procedure is to find out which of the potential members of the ground
structure should vanish and to find the areas of the non–vanishing members which gives the
desired elastic properties. As described in chapter one, vanishing members are not removed from
the base cell but their areas are set to the minimum value xmin . By choosing xmin approximately
107 times smaller than the maximum member area, elements governed by the minimum
constraint in (2.6) will have no structural significance, and they will not be shown in the examples
in order to ensure simplicity of the graphics.

In all examples, Youngs modulus of the base material (the material the individual bars are made
of) is normalized to unity. In order to ensure that the resulting truss–like microstructures are thin
(have low density), the prescribed elastic properties E *ijkl are chosen so small that the resulting
density is smaller than one. Using realistic elastic properties for the base material just requires
linear scaling of the prescribed elastic properties due to the linearity of the design problem.
38 2. Design of materials

Examples in two dimensions

In the first example we seek the material topology of a material with Poisson’s ratio equal to 1.
Optimizing a truss modelled base cell for the constitutive parameters obtained by inserting 1
in the elasticity tensor for an isotropic material in plane stress [see appendix 2, equation (A2.2)]
and prescribing E *1111 to be 0.05, results in several different solutions with the same overall den-
sity – four of them are shown in figure 2.4a–2.4d. The four different topologies are obtained by
varying the penalty terms in (2.5)truss2d. For instance, 2.4a is obtained by penalizing long bars,
and topology 2.4c by penalizing short bars. The two others were obtained by experimenting with
different sizes of the move limits, but common to all results was a ”switching” between the
”basic” topologies such that move limits had to be made quite small to ensure final convergence.
It is seen that the topologies are mechanisms as expected. In figure 2.5, periodic materials are
constructed from the base cells in figure 2.4 and here it is easier to imagine the actual topologies
of the materials.

a) b) c) d)

ǒ Ǔ
1 1 0
ǒE *Ǔ + 0.05 1 1 0 , ò + 0.2
0 0 0

Figure 2.4 Microstructures of materials with Poisson’s ratio close to 1.0.

Figure 2.5 Periodic materials with Poisson’s ratio 1.0 constructed from the first three microstructures in
figure 2.4.
 2. Design of materials 39

The microstructure in figure 2.4a can be interpreted as an octagonal honeycomb cell and can be
compared with the original hexagonal honeycomb discussed in chapter one.

Figure 2.6 shows a honeycomb with less stiffness in the vertical direction and figure 2.7 shows
a material where the only non–zero elastic parameter is E *1111. It is obvious, that the latter mate-
rial only can take up loads in the horizontal direction and has zero shear stiffness.

Choosing an elastic tensor where E *1122 t E *1212, the microstructures tend to be somewhat more
complex than those previously shown. Figure 2.8 shows two topologies of truss–like anisotropic
negative Poisson’s ratio materials with E *1111 + –E *1122 + 0.05 and E *1212 + 0. The topologies
have overlapping bars, so their shear displacement modes (å *(12)
pq from equation (1.7)) are shown
to illustrate the overlapping. The materials compress vertically subject to horizontal compres-
sion, but in shear, they will collapse in the mechanism–type motion shown in figure 2.8 (right).

The previous example showed anisotropic materials with negative Poisson’s ratio, where the
shear stiffness E *1212 was zero. The next examples will show ”real” isotropic negative Poisson’s
ratio materials, i.e. negative Poisson’s ratio materials with shear stiffness.

Requesting a material with Poisson’s ratio close to –1 the optimization algorithm returns only
one simple solution (figure 2.9 left). As shown in the physical interpretation (figure 2.9 right)
the base cell consists of two ”stiff” quadratic frames (shown in black and grey) rotating about
the center, such that expansion in one direction rotates the frames in opposite directions and
thereby causes expansion perpendicular to the applied expansion. Subjected to shear, the frames
will ”lock” because of the periodicity conditions.

ǒ 1 0.5 0
Ǔ
ǒE *Ǔ + 0.05 0.5 0.25 0 , ò + 0.113
0 0 0
ǒ Ǔ
1 0 0
ǒE *Ǔ + 0.05 0 0 0 , ò + 0.05
0 0 0

Figure 2.6 Honeycomb weakened in Figure 2.7 Rank 1 like material.

the vertical direction.
40 2. Design of materials

ǒE *Ǔ + ǒ 1 –1 0
Ǔ
–1 1 0 , ò + 0.36 (top) and 0.45 (bottom)
0 0 0

Figure 2.8 Anisotropic negative Poisson’s ratio materials.

ǒ1 –1 0
Ǔ
ǒE *Ǔ + 0.05 –1 1 0 , ò + 0.44
0 0 1

Figure 2.9 2–d base cell of an isotropic material with Poisson’s ratio close to –1.0 (left). The physical interpreta-
tion (right) shows that the material is composed of two stiff frames (shown in black and grey) free to rotate about
their centers. When pulled horizontally (big arrows) the frames will rotate in the directions of the small arrows
and thus expand vertically.

It was noted earlier, that a more complicated starting guess than the 4 by 4 node groundstructure
does not give rise to simpler microstructures. Two examples of this are shown in figure 2.10,
where the resulting microstructures for a Poisson’s ratio 1 (figure 2.10 left) and –1 (figure 2.10
right) material based on a 6 by 6 node groundstructure with 630 potential bar members are
shown. The Poisson’s ratio 1 microstructure (figure 2.10 left) can be compared with figure 2.4(c)
– only differences are the coordinates of the nodal points. The Poisson’s ratio –1 microstructure
 2. Design of materials 41

Figure 2.10 Effect on choosing more complicated starting guesses. Pois-

son’s ratio 1 (left) and –1 (right) microstructures are based on 6 by 6 node
ground structures with 630 potential bar members.

(figure 2.10 right) can be compared with figure 2.9, in this case we see, that starting with a more
complicated groundstructure results in a more complicated though functionally similar solution.

Examples in three dimensions

This section will show examples of 3–d microstructures with various material parameters, and
their similarities with 2–d microstructures will be discussed.

Asking for a 3–d isotropic material with Poisson’s ratio close to 0.5, the optimization algorithm
returns several different topologies depending on choice of penalty factors, symmetry require-
ments and starting guesses. Four of the resulting microstructures which have equal structural
weight are shown in figure 2.11a–d.

Figures 2.11a and 2.11b show Poisson’s ratio 0.5 materials with full cubic symmetry. It is inter-
esting to see that the microstructure in figure 2.11a corresponds to a 3–d version of the 2–d micro-
structure shown in figure 2.4b. Although figure 2.11b has an apparently more complicated
microstructure this material too is composed of 2–d microstructures, namely a combination of
those from figure 2.4a and b. Both figure 2.11a and b have bars in the interior of the base cells.
It was tried to find a cubic symmetric Poisson’s ratio 0.5 material with surface bars only by choos-
ing the penalty factor me in (2.4) equal to 0.5, i.e. bars on the surfaces are made ”cheaper” than
interior bars. This approach does not result in any solution which was seen in lack of convergence
of the optimization algorithm. In order to obtain a ”surface” solution, requirements on symmetry
are lessened, i.e. only symmetry about the three coordinate planes individually are required
instead of the full cubic symmetry. This lessening of symmetry requirement results in the micro-
structure shown in figure 2.11c. This microstructure too is seen to be composed of 2–d micro-
structures, namely figure 2.4c in the xy–plane and figure 2.4a in the yz– and zx–planes. Lessen-
ing the symmetry requirements also results in another ”interior” solution shown in figure 2.11d
which actually is a simple version of figure 2.11b. We note that it is possible to construct isotropic
materials which are not cubic symmetric in the microstructure.
42 2. Design of materials

(b)
(a)

y y

z x x
z

(c) (d)

y y

z x x
z

ȡ11 11 11 00 00 00ȣ
ǒE Ǔ + .05ȧ
*
ȧ10 10 10 00 00 00ȧ
ȧ, ò + 0.45
ȧ0 0 0 0 0 0ȧ
Ȣ0 0 0 0 0 0Ȥ
Figure 2.11 3–d base cells of isotropic materials with Poisson’s ratio close to 0.5.

A base cell of a material with Poisson’s ratio equal to zero is shown in figure 2.12. The solution
is obtained by penalizing interior bars, thus the microstructure has only bars on the surface, but
still it has a very complex microstructure.

Another interesting example in 3–d, is the tailoring of an isotropic Poisson’s ratio –1.0 material.
The first solution for this problem which is obtained without any penalty factors, results in a very
complicated microstructure with a lot of crossing and overlying bars. The solution is structurally
quite complicated and will not be shown here. Experiments with penalty factors (me = 0.5, c =
–1.0 and l pref = 0.5) lead to a much nicer solution which is shown in figure 2.13. The microstruc-
 2. Design of materials 43

y
z x y
1
0
ȡ 0
1
0
0
0
0
0
0
ȣ
0
0 x
ǒE *Ǔ + 0.05
0 ȧ
ȧ 0 1 0 0 ȧ
ȧ
0 z
0 , ò + 0.38
0
0 ȧ 0
0
0
0
.5
0
0
.5 ȧ
0
0 Ȣ 0 0 0 0 Ȥ
.5

Figure 2.12 3–d base cell of an isotropic material with Poisson’s ratio 0.

y
z x y
ȡ 1 –.5 –.5 0 0 0
–.5 1 –.5 0 0 0
ȣ x
ǒE *Ǔ + 0.02ȧ
ȧ–.5 –.5 1 0 0 0 ȧ
ȧ
z
0 0 0 .75 0 0 , ò + 0.29
ȧ 0 0 0 0 .75 0 ȧ
Ȣ 0 0 0 0 0 .75 Ȥ
Figure 2.13 3–d base cell of an isotropic material with Poisson’s ratio close to –1.0. The base cell is a cube
where the six sides are built up by rotatable frames like in the two dimensional case (figure 2.9).

ture in figure 2.13 has only exterior bars, and is seen to be built up by rotatable frames like in
the two dimensional case (figure 2.9).

The last example will show that it is also possible to tailor 3–d materials with different behaviour
in different directions. Asking for Poisson’s ratio –1.0 behaviour in the xy–plane and Poisson’s
ratio 0.5 behaviour in the yz– and zx–planes we get the 3–d microstructure shown in figure 2.14.
44 2. Design of materials

y
z x

z x

ȡ–.51 –.51 .5.5 00 0

0
ȣ
0
0
ǒE Ǔ + 0.04ȧ
ȧ .50 .50 10 .750 0 ȧ
ȧ
0
0 , ò + 0.37
*

ȧ0 0 0 0 0
0 ȧ
0
Ȣ0 0 0 0 0 Ȥ
0

y
x z
Figure 2.14 3–d base cell of an isotropic material with Poisson’s ratio –1.0 in the xy–plane and
Poisson’s ratio 1.0 behaviour in the yz– and zx–planes.

Examining figure 2.14, we notice that the microstructure is composed of the 2–d microstructures
in figure 2.4(b) and figure 2.9. It can furthermore be seen, that the xy–plane with Poisson’s ratio
–1 behaviour is not connected to the two other planes.

The 3–d microstructures in this section were obtained independently of the 2–d results and they
all came from the 2016 bar ground structure (figure 2.2b). The results in this section can indicate
that 3–d materials with arbitrary structural behaviour can be composed by ”braiding” of 2–d
microstructures into 3–d microstructures with small modifications. However, further investiga-
tion of this hypothesis is required.

2.4 Frame–like materials

This section shows that the topologies obtained by discretizing the groundstructure by frame ele-
ments resembles those obtained from the truss discretization, only difference is that the attainable
elastic properties are ”damped”.
 2. Design of materials 45

For the design of frame like materials, the design variables are given by the individual element
widths and the height of each element is kept constant for all elements . The groundstructure with
120 possible beam elements, shown in figure 2.2c, is used as starting guess.

Design of the Poisson’s ratio plus and minus 1 materials is used to illustrate the difference
between discretizing the groundstructure by truss or frame elements. Table 2.1 shows attainable
values of Poisson’s ratios for varying beam thicknesses. The first row in the table shows the
attainable Poisson’s ratio for a truss modelled base cell. It is seen that the extreme values ( =
1 and    can not be obtained exactly, even for truss modelled materials. This lies in the
fact that the elements governed by the minimum thickness constraint xmin prevent the base cell
in becoming a true mechanism. The next rows in the table show the attainable values of Poisson’s
ratio for frame modelled microstructures of increasing thickness. It is seen that modelling the
base cell from thick frame elements ”damps” out extreme material behaviour as expected. Maxi-
mum beam thickness in the table is restricted to 20% of the base cell size. Allowing for thicker
beams the microstructure should be thought of as a continuum with small holes.

Although not shown graphically, it should be noted that the material topologies for frame model-
led base cells are apparently the same as for truss modelled base cells. This means that it is not
so important to optimize the base cells using frame elements which, especially in 3–d, is a com-
putationally more complicated task than the truss case (the number of degrees of freedom would
be twice that of the truss case). Having optimized a truss modelled base cell, we will assume that
the frame modelled microstructure can be constructed from the same topology, but it must be
taken into account that the material parameters will be ”damped” according to table 2.1.

=1  = –1
Beam thickness in % Attainable Attainable
of base cell dimension Poisson’s ratio Poisson’s ratio
0 (truss) 0.99996 –0.9998
1 0.9993 –0.997
2 0.997 –0.990
5 0.984 –0.94
10 0.94 –0.80
20 0.79 –0.63
Table 2.1. Attainable values of Poisson’s ratio, from base cells modelled by
frame elements with different thicknesses.
46 2. Design of materials

2.5 Continuum–like materials

This section demonstrates how a wide range of elastic properties can be obtained from ground-
structures discretized by continuum type elements.

The microstructures shown in the following are all obtained from rectangular base cells discre-
tized by four node bilinear elements. The design variables x e, are the material densities in the
individual elements with upper and lower bounds xmax = 1 and xmin = 10–4. The base material
for all examples has Youngs modulus 0.91 and Poisson’s ratio 0.3 such that a solid base cell will
have E1111=1.0. To obtain solid/void solutions, the penalty factor p in (2.5)continuum was varied
between 2 and 5 depending on the problem. Very bad ”checkerboard” patterns were seen in
results from the preliminary tests of the algorithm. Consequently, early results published in Sig-
mund, (1994)1, only considered coarse design domains where the checkerboard problem was
less prone to appear. In this chapter the checkerboard prevention scheme based on image proces-
sing techniques described in chapter 3 of this report, is used to obtain nice solutions for fine
discretizations.

In usual topology optimization one takes a uniform distribution of material as a starting guess.
This approach can not be taken here, because a uniform material distribution does not cause any
microstructural variation and thus the overall strain fields å ij are equal to the test strain fields å 0ij
in (1.5). Consequently, the values of the initial energies q eijkl are constant. Therefore, we will use
two different inhomogeneous starting guesses, for the first, the value of the design variable is
proportional to its distance from the center of the base cell, and for the second, the values are
chosen by a random process.

Generally many different solutions are found for the same prescribed elastic properties depend-
ing on the choice of starting guess, penalty factor and move limits. As for the truss– and frame–
like discretizations many local minima exists, so the examples in this section are selected
amongst many others mostly because they were the lightest found but also because they were
the topologically simplest.

The EBE–PCG equation solver described in appendix 3 works very efficiently for the problems
considered here. As an example, one FE–solution of a 60x60 element base cell might take one
minute on an HP 735 workstation during the first design steps but later, during final convergence,
the solution time is less than five seconds.

To test that the optimization algorithm gives reasonable results, we will try to ”reinvent” the per-
fect honeycomb analyzed in chapter 1. Giving the design domain the same rectangular dimen-
sions as specified in figure 1.7, and prescribing the elastic properties to the ones obtained from
 2. Design of materials 47

the discretizations with 21x12 and 63x36 elements in table 1.1, we get the ”optimal” microstruc-
tures shown in figure 2.15a and 2.15b. We see, that the optimized microstructures are very simi-
lar to the perfect honeycomb. Changing the design domain to a square cell discretized by 15x15
elements results in the topology shown in figure 2.15c. Now we get an ”octagonal honeycomb”
like the one obtained from the truss discretization in figure 2.4a. The density of the octagonal
honeycomb is approximately the same as for the original honeycomb and therefore, manufactur-
ability considerations must decide which of the topologies to choose.

Another testcase is the ”reinvention” of the inverted honey comb from figure 1.9 and table 1.2.
Using a 42x24 groundstructure as starting guess, the optimized topology is shown in figure 2.16.
Again, the topology resembles the original in figure 1.9. Only difference is a ”shifting” of the
microstructure of half a period.

Changing the design domain to a square domain with 30x30 elements and imposing horizontal
and vertical symmetry but prescribing the same elastic properties as in figure 2.16 resulted in
the microstructure shown in figure 2.17a. The density of the quadratic groundstructure based
material is 33% lower than the original inverted honeycomb but again, one might choose the

a)

ǒ 1 .75
ǒE *Ǔ + 0.09 .75 1
0
0
0 0 .125
Ǔ
ò + 0.29

b)

ǒ 1 .75
ǒE *Ǔ + 0.094 .75 1
0
0
0 0 .125
Ǔ
ò + 0.30

c)

ǒ 1 .75
ǒE *Ǔ + 0.09 .75 1
0
0
0 0 .125
Ǔ
ò + 0.28

Figure 2.15 ”Reinvention” of the perfect honeycomb. Microstruc-

tures with Poisson’s ratio equal to 0.75. Three different groundstruc-
tures: a) 21x12, b) 63x36 elements and c) 15x15 elements.
48 2. Design of materials

ǒ 1 –.5
ǒE *Ǔ + 0.08 –.5 1
0
0
0 0 0.06
Ǔ
ò + 0.407

Figure 2.16 ”Reinvention” of the inverted honeycomb. Anisotropic

microstructure obtained from groundstructures with 42x24 elements.

inverted honeycomb anyway for manufacturability reasons. The ”optimal anisotropic negative
Poisson’s ratio material” in figure 2.17a can be extremized with Poisson’s ratio equal to –0.7 and
–0.85 as seen in figures 2.17b and c.

For frame like materials, the attainable properties were dependent on the density of material in
the base cell; the same is the case for continuum like materials. The dependency is illustrated in
figure 2.18. Prescribing E *1111 to be 0.021, the highest attainable Poisson’s ratio is 0.94, obtained
with the microstructure shown in figure 2.18a. Again, we notice the similarity to the truss solu-
tion. Increasing the desired value of E *1111 (figure 2.18b–2.18e) decreases the attainable Pois-
son’s ratio until finally, we get to the solid base cell with E *1111 + 1 and n + 0.3. From this and
other examples it appears that the attainable stiffness ratio between the effective properties and
the properties of the base material defined as max(Eijkl / E) is low when we prescribe extreme
elastics properties and we may conclude, that we must accept low overall stiffness if we want
materials with extreme values of Poisson’s ratio.

The anisotropic negative Poisson’s ratio materials shown in figure 2.17 and the inverted honey-
comb have very low shear stiffnesses. Almgren, (1985), suggests to modify the inverted honey-
comb with struts and sliding hinges such that it gets a high shear stiffness. An interesting possibil-
ity to make a simple isotropic negative Poisson’s ratio material by the present optimization
algorithm, will therefore be to take the inverted honeycomb, say figure 1.9b as a starting guess,
specify isotropic properties, and see if small structural modifications of the inverted honeycomb
can lead to an isotropic solution. This showed out to be impossible; there was no convergence
of the optimization algorithm.

Instead of starting out from the inverted honeycomb, it was tried to make an isotropic material
with n + ć0.7 using a quadratic and fully symmetric 60x60 element groundstructure. The
resulting topology is shown in figure 2.19 and looks very beautiful, but the micro–mechanical
interpretation is difficult. The microstructure can be compared with the reentrant foams
described in Lakes (1987). When stretched in one direction, the small bars and plates unfold and
results in expansion in the opposite direction.
 2. Design of materials 49

ǒ Ǔ
a)
1 –.5 0
ǒE *Ǔ + 0.08 –.5 1 0
0 0 .06
ò + 0.27

ǒ Ǔ
b)
1 –.7 0
ǒE *Ǔ + 0.1 –.7 1 0
0 0 .1
ò + 0.44

ǒ Ǔ
c)
1 –.85 0
ǒE *Ǔ + 0.1 –.85 1 0
0 0 .2
ò + 0.64

Figure 2.17 Anisotropic ”negative Poisson’s ratio” material (low shear

stiffness). Solutions obtained from groundstructures with 30x30 elements
and horizontal and vertical symmetry enforcement.
50 2. Design of materials

a)
ǒ 1 .94 0
ǒE Ǔ + 0.02 .94 1 0
*

0 0 .03
Ǔ b)
ǒ 1 .9 0
ǒE *Ǔ + 0.05 .9 1 0
0 0 .05
Ǔ c)
*
ǒ
1 .7 0
ǒE Ǔ + 0.1 .7 1 0
0 0 .15
Ǔ
ò + 0.078 ò + 0.19 ò + 0.31

d)
ǒ 1 .55 0
ǒE *Ǔ + 0.2 .55 1 0
0 0 .23
Ǔ e)
ǒ 1 .44 0
ǒE *Ǔ + 0.3 .44 1 0
0 0 .28
Ǔ f)
ǒ 1 .3 0
ǒE *Ǔ + 1.0 .3 1 0
0 0 .35
Ǔ
ò + 0.47 ò + 0.59 ò + 1.0

Figure 2.18 Example of the dependence of the obtainable Poisson’s ratio for various densities.
Groundstructures are a) 40x40 elements, b) 20x20 elements and c) to f) 16x16 elements.

Figure 2.19 Isotropic and symmetric microstruc-

ture with Poisson’s ratio equal to –0.7. Ground-
structure is 60x60 elements.
 2. Design of materials 51

To get a simpler solution, the symmetry requirements were dropped and replaced by two extra
equality constrains on the prescribed properties. Trying out different starting guesses and
groundstructure sizes result in different more or less complicated solution. The simplest topo-
logy obtained is shown in figure 2.20a (left). By looking at the base cell alone it is difficult to
see what is going on, therefore the base cell was repeated periodically in figure 2.20a (right).
Now, the ”mechanism” is seen clearly. Horizontal compression causes the triangles to ”close”
and thereby contract vertically.

a)

ǒ 1 –.8 0
ǒE *Ǔ + 0.02 –.8 1 0
0 0 .9
Ǔ
ò + 0.25

b)

ǒ 1 –.6 0
ǒE *Ǔ + 0.04 –.6 1 0
0 0 .8
Ǔ
ò + 0.38

c)

ǒ 1 –.5 0
ǒE *Ǔ + 0.065 –.5 1 0
0 0 .75
Ǔ
ò + 0.56
Figure 2.20 Negative Poisson’s ratio materials obtained from
groundstructures with 30x30 elements and no symmetry enforce-
ment.
52 2. Design of materials

The advantage of the continuum modelled negative Poisson’s ratio material in figure 2.20
compared to the truss modelled ”rotating frames” version in figure 2.9 is, that the continuum ver-
sion has no sliding surfaces. The values of the stiffness and Poisson’s ratio can be varied to a cer-
tain extent as seen in figure 2.20b–2.20c. It is interesting to note, that a microstructure neither
has to be horizontally and vertically symmetric to be orthotropic, nor does it have to show 60
degrees symmetry to be isotropic.

The continuum modelled microstructure with Poisson’s ratio –.8 in figure 2.20a can be inter-
preted as a rod and hinge model shown in figure 2.21 (left). The rod–and–hinge interpretation
can be compared with the fish–bone like microstructures possessing Poisson’s ratio –1 proposed
shown in figure 2.21 (right) by Milton, (1992). Milton assumed that his fish bone structure is
built up from laminates of three different length scales and thereby proved mathematically that
isotropic materials with negative Poisson’s ratios exist within the framework of continuum elas-
ticity. The fish bone structure of Milton belongs to the so-called ranked materials, which usually
are considered as mathematical tools rather than being practical composites because of their
widely differing length scales. The example in figure 2.20a and its possible interpretation in fig-
ure 2.21 indicates that the ”mechanisms” of the extreme ranked material is reproduced in the one
length scale microstructure proposed here.

fig. 2.20a Milton, (1992)

Figure 2.21 Rod–and–hinge interpretation (left) of negative Pois-

son’s ratio material from figure 2.20a compared with Rod–and–
hinge interpretation of Milton’s fish bone microstructure (right)
with Poisson’s ratio –1 [Milton, (1992)].
 2. Design of materials 53

For practical reasons, we might be interested in designing solid instead of porous microstruc-
tures. As an example, this could be done by embedding the considered microstructure in a soft
matrix material. The matrix material has to be very soft compared with the base material in order
not to damp out attainable properties to much. In fact, Cherkaev and Gibiansky, (1991), provide
lower and upper bounds for the attainable values of Poisson’s ratio for mixtures of two well
ordered materials. They conclude that a stiffness ratio between the strong and the soft phase of
more than 25 is necessary to achieve negative Poisson’s ratio at all.

To study the influence on embedding one of the numerically developed microstructures in a soft
matrix material, we will consider the Poisson’s ratio –.8 microstructure in figure 2.20a. Simply
embedding the microstructure in a matrix material with E M1111 + x m and Poisson’s ratio 0.3, can
be done by specifying the lower bound on the design variables to be xm . By numerical experi-
ments, we found that the highest value of xm which still gives a material with negative Poisson’s
ratio is 0.014 which means that the ratio between the strong and the soft phase should be more
than 71 to get a negative Poisson’s ratio for this particular numerical example. Examples of mate-
rial combinations with high ratios could be steel/polystyrene (70).

2.6 Continuum–like materials with prescribed thermoelastic properties

In this section, a simple example demonstrating the design of a material with prescribed thermo-
elastic properties is given.

The design of materials with prescribed thermal coefficients has large potentials in many engi-
neering applications where temperature changes are large. Two well known examples are civil
structures where temperature changes between summer and winter cause large structural expan-
sion and space structures, where the temperature difference between the ”sunny” and the ”shady”
side can be large and cause big structural distortions.

The design of a material with a prescribed thermal expansion tensor a *ij can not be done alone
by designing the microstructure based on one material – we have to consider a mixture of at least
two materials with different thermal expansion coefficients a 1ij and a 2ij. The explanation of this
comes from studying equation (1.12). We note that if we only have one material (or one material
and void or one material with cracks), a kl is constant throughout the microstructure such that a kl
can be put outside the integral in (1.12) and we get that b H ij + E ijkl a kl .
H

To make a well posed example for design of materials with prescribed thermoelastic properties,
we will consider the design problem sketched in figure 2.22. We are considering a mixture of
two materials with different thermal expansion coefficients. The matrix phase has low thermal
expansion coefficient a 111 + a 122 + 1 and the square inclusion phase has high expansion coeffi-
54 2. Design of materials

a 1ij + ǒǓ
1
1
0

a 2ij + ǒǓ
10
10
0

Figure 2.22. Initial base cell.

cient a 211 + a 222 + 10. The goal is to design elastically isotropic materials with Poisson’s ratio
0.5 and various prescribed thermal expansion coefficients a *ij. Having specified a *ij, the pre-
scribed thermal coefficient tensor b *ij used as constraint in the optimization problem is calculated
as b *ij + E *ijkl a*kl.

Prescribing a low overall thermal expansion coefficient a *11 + a *22 + 1.33 we get the solution
shown in figure 2.23a. It is seen that the solution resembles the octagonal honeycomb and that
most of the material concentrates in the outer matrix phase with low expansion coefficient. Spec-
ifying a higher overall expansion coefficient results in material forming in the inclusion phase
as seen in figure 2.23b and finally specifying a high overall expansion coefficient
a *11 + a *22 + 3.66 results in a ”shifted” version of figure 2.23a as seen in figure 2.23c. We see
that the microstructure tries to utilize the high expansion material in the square inclusion phase
to make a high overall expansion coefficient. Finally, we we will show that an elastically iso-
tropic material can be designed with different thermal expansion coefficient in the horizontal and
the vertical direction. Figure 2.23d shows a microstructure where a *11 and a *22 were prescribed
to 5.33 and –0.66 respectively. It is seen, that the two horizontal bars in the high expansion inclu-
sion phase will cause the material to expand heavily in the horizontal direction when heated,
whereas only low expansion phase matrix material contributes to the vertical expansion.

The above example was quite basic and shall just be seen as a demonstration of which material
properties it is possible to control. The high expansion inclusion phase can also be seen as a
build–in actuator which could be controlled by an electric field (piezo electric actuator) instead
of by temperature change, and the design goal could, for example, be to make a material with
the largest possible expansion in a specified direction due to an electric input signal. Conse-
quently, the author expects the method to be applicable (with minor modifications) to the design
of smart or intelligent materials which can change their properties actively as a response to exter-
nal stimuli.
 2. Design of materials 55

a) a *11 + a*22 + 1.33 b) a * 11 + a *22 + 2.66 c) a * 11 + a *22 + 3.66

ǒ1 .5 0
ǒE *Ǔ + 0.2 .5 1 0
0 0 .25
Ǔ
ò + 0.45

d) a *11 + 5.33, a*22 + –0.66

Figure 2.23 Design of materials with prescribed thermoelastic properties. Groundstructure has
30x30 elements.

2.7 Practical considerations

In this section, we will discuss simple working models which were made to test that the numeri-
cally obtained results are practically realizable.

To test the function of the truss modelled two dimensional Poisson’s ratio –1 material from figure
2.9, a lab model of the material consisting of an array of 4 by 4 connected base cells was
constructed (figure 2.24). The model consists of two thin layers of celluloid, put on top of each
other, and placed between two stiff transparent plates. Figure 2.25 shows the mechanical beha-
viour of the working model. Comparing the picture of the unstressed material (figure 2.25 left)
with the stressed version (figure 2.25 right) the deformation mechanism is clearly seen.

The three dimensional truss modelled Poisson’s ratio –1 material from figure 2.13 was also tested
by a working model. A 15 by 15 by 15 cm. model of this microstructure was built in the lab,
consisting of 72 plastic tubes connected by bent pieces of thin wire (figure 2.26). The test cube
works nicely, which means that if compressed vertically, it will also compress in the two other
directions. The negative Poisson’s ratio behaviour of the base cell is, of course, limited to the
point where the rotating square frames become parallel.
56 2. Design of materials

Figure 2.25 Working model of the proposed negative Poisson’s ratio material from figure 2.9. The undeformed
material consisting of an array of 4 by 4 base cells is shown left. Expanded vertically we get the deformation pat-
tern shown right.

Figure 2.24 Working model consisting of an array of 4 by 4 base cells of the pro-
posed negative Poisson’s ratio material from figure 2.9.
 2. Design of materials 57

Figure 2.26 Working model of the Poisson’s ratio –1 material from figure 2.13.

One might ask whether having two layers rotating on top of each other is essential for obtaining
the isotropic negative Poisson’s ratio material shown in figure 2.9 and ask if the much simpler
one layer version in figure 2.27 could do the job. Comparing the modified microstructure in fig-
ure 2.27 with the continuum modelled microstructure from figure 2.17c, we get the answer. The
two microstructures essentially have the same topology and because the continuum version (fig-
ure 2.17c) was computed to have low shear stiffness, we will expect the truss version to be aniso-
tropic as well. We can therefore conclude that the two interconnected sliding layers in the original
microstructure in figure 2.9 are essential for the isotropy of the truss like material.

Figure 2.27 One layer version of the

Poisson’s ratio –1 microstructure
from figure 2.9.
58 2. Design of materials

This conclusion is a serious drawback for the practical application of the truss modelled Pois-
son’s ratio –1 material. The sliding surfaces and the linking of the frames will be difficult to
manufacture and may cause large microstructural stress concentrations.

The continuum modelled microstructures do not have the disadvantages of sliding surfaces and
microscale links. Taking the Poisson’s ratio –0.6 material from figure 2.20b and post–processing
it by taking the contour–line of density 0.5 as the border between solid and void, we get the
microstructure seen in figure 2.28. Although it still has jagged edges due to the coarse discretiza-
tion, we can clearly interpret the microstructure as a micro–frame like material with no sliding
surfaces or micro–scale links. To remove the jagged edges, either more elements could be used
in the groundstructure or, the microstructure could be taken into a shape optimization software
to smooth boundaries and minimize stress concentrations.

Figure 2.28 Post–processed picture of microstructure with Poisson’s ratio –0.6 from fig-
ure 2.20b.
 2. Design of materials 59

2.8 Summary and discussion

In this chapter we have proposed a numerical approach to the design of materials with prescribed
thermoelastic properties. The method is based on algorithms and discretizations used in topology
optimization of general structures. The material design problem is formulated as an optimization
problem of minimizing weight of a base cell subject to equality constraints on the prescribed
thermoelastic properties. The design domain is the topology of the representative base cell which
has been discretized by truss, frame and continuum type finite elements to account for different
material types. The groundstructure approach is used as a starting guess to the optimization pro-
cedure, which means that the best microstructure is sought within a fixed but very detailed discre-
tization.

As demonstrated by many examples, the proposed numerically based design procedure can be
used to tailor materials with a wide range of thermoelastic properties.

Using truss–like discretizations in two and three dimensions makes it possible to design materials
with properties very close to the theoretical bounds (positive semi–definiteness of the elastic ten-
sor) whereas using more practical frame elements damps out attainable properties. For the solu-
tions in both the truss and the frame formulation, bars and beams are allowed to cross without
(theoretical) interaction. This means that in practical realization of the microstructures, there will
be parts sliding on top of each other which can cause difficulties in manufacturing or cause large
stress concentrations. Both these problems are difficult to prevent in the truss and frame formula-
tions.

More realistic microstructures are obtained by the continuum formulation because the discretiza-
tion only has one layer and consequently does not allow sliding surfaces. Examples show that
extreme elastic parameters can be obtained by the continuum formulation as well, however,
experiments show that only low overall stiffness can be obtained if extreme elastic properties
are prescribed. Further manufaturability constraints on the continuum like microstructures such
as minimum size of holes or structural members can be imposed by using the mesh independency
procedure developed in chapter 3 however, it must be taken into account that any topological
constraint put on the microstructural geometry damps out the attainable elastic properties. As
demonstrated in an example, the attainable properties are also damped when a porous micro-
structure is imbedded in a matrix material.

The possibility of designing materials with prescribed thermoelastic properties was demon-
strated by a simple example using the continuum formulation. The obtained microstructure
topologies showed ”actuator” like elements which suggests that the present approach to material
design can be used for design problems like optimal placement of actuators or sensors in design
of smart materials or structures.
60 2. Design of materials

To test if the numerically obtained microstructures really work the predicted way, two very
coarse models were built and both showed the expected (negative Poisson’s ratio) behaviour.
However, due to coarseness of the models, none of them allowed actual measurements of their
elastic properties. Therefore, it would be interesting to make more realistic prototypes consisting
of several base cells that could be tested by standard material testing methods. This is outside
the scope of this thesis but in continuation of the work, prototypes of the proposed microstruc-
tures can be produced by the rapid prototyping or stereolitography technique. Briefly described,
the method can produce prototypes of complicated three dimensional structures by layerwise
hardening of liquid polymers or other base materials [e.g. overview in Ashley, (1994)2]. Com-
mercially available rapid prototyping machinery can produce parts with resolutions down to a
tenth of a millimeter but methods are developed which can produce geometries with resolution
of a few micrometers [e.g. Ikuta et al. (1994)].

The microstructures obtained by the numerical method developed in this thesis shall not be seen
as readily manufacturable solutions but rather as sources of inspiration to people involved in
practical synthesis of novel microstructures.
 3. The checkerboard and mesh–dependency problems
in layout optimization
This chapter will present a method to prevent the formation of checkerboard patterns occurring
in layout optimization of linear elastic structures and at the same time eliminate the mesh–depen-
dency problem of optimal layouts.

In section 3.1, the layout problem expressed as a compliance minimization problem will be
described, and the occurrence of checkerboard patterns for certain discretizations will be demon-
strated. A checkerboard prevention scheme is proposed based on ideas borrowed from image
processing techniques and therefore a brief introduction to image processing is given in section
3.2 and the application to structural layout problems is treated in section 3.3. Applying the pro-
posed checkerboard prevention scheme to typical design examples a well known problem turns
up, which is that optimal layouts are dependent on the finite element discretization. The checker-
board prevention scheme is therefore modified to obtain an algorithm that, although based on
heuristics, produces efficient mesh–independent designs that satisfy manufacturability
constraints such as exclusion of geometry change below a prescribed scale (section 3.4).

The general layout optimization problem is sketched in figure 3.1 and can be defined as follows.
Find the distribution of material for a structure supported on its boundaries and subjected to a
given loading condition, such that an objective function is optimized. The amount of material
is constrained and the spatial distribution of material is limited to the design domain .. The
design domain can have holes (void regions) and regions with fixed solid material. As discussed
in the introduction a structure can be discretized in a number of different ways and it can be mod-
elled by full three dimensional analysis, plates with inplane or out of plane forces etc. In this
chapter we will consider the layout optimization of continuous linear thin elastic plates (mem-
branes) subject to inplane forces (plane stress) and discretized by four node quadrilateral finite
elements.

void
W load

support
solid

Figure 3.1 The generalized layout optimization problem.

 62 3. The checkerboard and mesh dependency problems in layout optimization

Practical layout optimization of continuum modelled plates can basically be divided into two dif-
ferent approaches. The first is the ”Homogenization approach to topology optimization”
introduced by Bendsøe and Kikuchi, (1988) and the second is the ”Solid Isotropic Microstruc-
tures with Penalty” (SIMP) approach proposed in Bendsøe, (1989) and used in Rozvany et al.,
(1992). The penalization approach is also extensively used in biomechanics [see Huiskes and
Hollister, (1993) and references herein]. Here we shall use a ”mixture” of the two methods but
the proposed prevention algorithm should be readily applicable to other formulations.

3.1 Design of structures by compliance optimization

Consider a linear elastic structure bounded on the domain , subject to surface tractions {R} and
proper supports. Using standard finite element notation from Cook, Malkus and Plesha, (1988)
the structure is discretized by NE finite elements. The compliance of the structure can be written
as

NE
T
W + {D } [ S ]Ă{ D } + ȍ {de}T[se]Ă{de} (3.1)
e+1

where the global displacement vector {D} is the solution to the equilibrium problem

[ S ]Ă{ D } + {R } (3.2)

and [ S ] is the global stiffness matrix of the discretized design domain . The global stiffness
matrix is assembled from the local element stiffness matrices [s e] , defined as
ŕ
[s e] + [B e] T[E e](ò e) [B e] dv e where [B e] is the strain–displacement matrix, [E e](ò e) is the
ve
constitutive matrix as a function of the density re and v e is the volume of element e. Similarly
the local displacement vectors {d e} are the elements of the global displacement vector associated
with element e.

The optimization problem is to minimize the compliance for a fixed amount of available material
in the design domain. Assuming that the density of material in element e is re, the total volume
of the structure can be written as

ȍ òe ve
NE
V+ (3.1)
e+1

In order to get a well posed problem, we will assume that the density in an element is bounded
by 0 < rmin vre v , where rmin %  such that low–density regions have little influence on the
overall structural behaviour.
 3. The checkerboard and mesh dependency problems in layout optimization 63

We can now define the compliance optimization problem as

ȍ {de}T[se] {de}
NE
T
minimize : W + { D } [ S ]{D } +
e+1

ȍ òe ve – V v 0
NE
subject to : (3.2)
e+1
[S ] {D } – { R } + {0}
0 t ò min v ò e v 1 , e + 1,..., NE

Remaining now is the problem of choosing a suitable material model for the density to stiffness
relation at a point (element) in the structure.

The goal of the optimization algorithm is to find the optimal structure composed of solid and void
regions. This problem is a 0–1 type optimization problem which is too large to be solved by any
integer programming method even for a low number of elements. Furthermore the solution of
the 0–1 problem does not converge towards a macroscopic solid–void layout when the mesh is
refined. Instead, refining the mesh leads to chattering designs or solutions with an increasing
number of small holes. This problem was first described in Cheng and Olhoff, (1981) for plates.
To prevent these problems, a ”relaxed” formulation is necessary [see Allaire and Kohn, (1993)1
and references therein]. The relaxed formulation introduces a material microstructure consisting
of so-called ranked materials made of microscopically oscillating material on differing length
scales. Using homogenization methods to determine the effective properties of ranked materials
and substituting the homogenized parameters into the ”macroscopic” topology optimization
problem, leads to the ”homogenization approach to topology optimization” method introduced
as a computational tool in the work of Bendsøe and Kikuchi, (1988). It can be shown [e.g. Avella-
nada, (1987)], that solutions using ranked microstructures attain the lower bound on compliance
for mixtures of two isotropic materials. Although proven optimal, the solutions obtained by
using ranked materials have large ”grey areas” of intermediate densities consisting of micropo-
rous regions. Aiming at producing structures with macroscopic voids more applicable to practi-
cal purposes, we have to consider ”suboptimal” microstructures [e.g. Bendsøe and Kikuchi
1988], penalizations of intermediate densities [e.g. Allaire and Kohn, (1993)2, Allaire and
Francfort, (1993), Jog, (1994)] or artificial microstructures [e.g. Rozvany et al., (1992),
Bendsøe, Díaz and Kikuchi, (1993) and from biomechanics, Mullender, Huiskes and Weihnans,
(1994)].
 64 3. The checkerboard and mesh dependency problems in layout optimization

Here we will use an artificial model for the stiffness to density relation, namely writing the
constitutive matrix of element e as

ȱ1n 1n 00 ȳ
[E e](ò e) + (òe) p ƪE 0ƫ + (ò e) p E
ȧ ȧ (3.3)
Ȳ
(1–n) 2 0 0 (1–n)ń2
ȴ

where E is the Young’s modulus, n is the Poisson’s ratio and [E0] is the constitutive matrix for
solid isotropic material. By choosing different values of the penalty factor p, ”grey” regions or
regions with intermediate values of re can be penalized. Figure 3.2 shows E1111 plotted versus
re for different values of p compared with E1111 for a rank–2 material. The stiffness of a rank–2
material is not uniquely defined because it is optimized with respect to the given stress or strain
field. The possible range of optimal stiffnesses for a rank–2 laminate with lamination directions
aligned with the principal stress or strain directions are marked as a grey region in figure 3.2.
Loosely interpreted the figure shows that the artificial material law (3.3) with penalty factor up
to 1.5 should give solutions similar to solutions obtained by using rank 2 materials. Choosing
a higher value for the penalty factor p will penalize intermediate densities and results in more
distinct solutions.

E1111

p=1

rank–2

p=2
p=3

Figure 3.2 Stiffness versus density for different penalty factors p compared with stiffness
range of optimized rank–2 material (grey region).
 3. The checkerboard and mesh dependency problems in layout optimization 65

Using the stiffness to density relation (3.3), the compliance optimization problem (3.2) can be
written as

ȍ {d ȍ (òe)p{de}T[se0]{de}
NE NE
} T[ e} T[ e]
minimize : W + { D S ]{D } + s {d e}
+
e+1 e+1

ȍ òe ve–V v 0
NE
subject to :
e+1 (3.4)
[S ] {D } – { R } + {0}
0 t ò min v ò e v 1 , e + 1,..., NE
p
where : [s e] + (ò e) [s e0] , [s e0] + ŕ [B ] [E ][B ] dv , p w 1
e T
0
e e

ve

T
For convenience, we will define the element relative strain energies q e as q e + {d e} [s e0] {d e}
such that the compliance can be written in the simple form

ȍ (òe)p qe
NE
W+ (3.5)
e+1

To solve the optimization problem (3.4), we will use an optimality criteria based algorithm.
Using the standard optimality criteria method for structural optimization [Bendsøe, (1994)], the
Lagrangian function for (3.4) can be written

L + {D } [ S ]{ D } ) lȧ
T ȡȍNE
ȣ
ò e A e – Vȧ) { m } ([ S ]{D } – {R }) )
T
Ȣe+1 Ȥ
(3.6)
ȍ ȍ
NE NE
a e( – òe ) ò min) ) be (ò e – 1)
e+1 e+1

T
where l, a e b e and { m } are the Lagrangian multipliers for the volume constraint, the lower and
upper side constraints and the equilibrium equation respectively.

The necessary condition for optimality is stationarity of the Lagrangian function (3.6) with
respect to the displacement vector and the design variables.

Stationarity with respect to the displacement vector gives

ēL + 2{ D } T[S ] ) {m } T[S ] + {0} å {m } + –2{ D } (3.7)

ē{ D }
 66 3. The checkerboard and mesh dependency problems in layout optimization

Stationarity with respect to design variable ò e gives

ēL + 2{D } T[ S ] ē{ D } ) { D } T ē[ S ] { D } )
ēò e ēò e ēò e

ēò
ǒ
[ ] { }
ēò
Ǔ
{m } T ē Se {D } ) [ S ] ē De ) l v e – a e ) b e + 0 , e + 1,..., NE
(3.8)

Substituting (3.7) into (3.8) we get

ēL + – {D } T ē[S ] {D } ) l v e – a e ) b e +
ēò e ēò e
(3.9)
T ē[s e] { e}
– {d e} d ) l v e – a e ) b e + 0 , e + 1,..., NE
ēò e

p–1
and using that ē[s e]ńēò e + p (òe) [s e0] we get

ēL + – p (ò e)p–1 q e ) l v e– ae ) b e + 0 , e + 1,..., NE (3.10)

ēò e

For intermediate densities ( ò min t ò e t 1), the conditions (3.10) can be written

p (ò e)p–1 q e
B0 + + 1 , e + 1,..., NE (3.11)
l ve
The interpretation of (3.11) is that the sensitivities of a design change should be constant for all
intermediate densities in the optimal solution.

Following Bendsøe, (1994), we can define the fix–point type update algorithm

h
ò e + ò e0 ǒBe0Ǔ , e + 1,..., NE (3.12)

where h is a damping factor and ò e0 is the value of the design variable from the previous iteration.

The updating scheme for the fixed point algorithm is defined as

ȡmax {(1–j) òe0, òmin } if òe0 B h0 v max {(1–j) òe0, òmin }

ȧe h e e h
ò e + ȥò 0 B0 if max {(1–j) ò 0, òmin } v ò 0 B 0 v min {(1–j) ò 0, 1}
e
(3.13)
ȧmin {(1 ) j) òe0, 1} if min {(1 ) j) òe0, 1} v òe0 B h
Ȣ 0

where j is a move limit.

 3. The checkerboard and mesh dependency problems in layout optimization 67

As the new values of the design variables, defined by (3.12), are dependent on the Lagrangian
multiplier l, we have to perform an inner loop in the optimization algorithm to update the multi-
plier. The volume of the structure is a decreasing function of the multiplier and therefore we can
use a bisection method for the inner loop.

A few words shall be devoted to the choice of penalty factor p in the stiffness to density relation
(3.3). Choosing p=1 we are back to the ”thickness of a sheet” design problem which is a well
posed problem. However, this problem generally results in solutions with large ”grey” areas
which we are trying to avoid. The higher we chose the value of p, the more distinct black and
white (solid and void) structures we get. For p3 we get solutions consisting almost entirely of
solid and void regions.

A problem is, that different choices of p can lead to widely different topologies. In order to get
solutions seemingly independent on the choice of penalty factor, we will make use of the so–
called continuation approach used by Jog, (1994).

The idea of the continuation approach is to start with a low value p=p0 ; letting the subproblem
converge; increase p by np; converge, and continue this, till the desired value p=pmax has been
reached. Although increasing the total number of required iterations considerably (up to 300),
we do get more consistent and detailed solutions. Experiment with the values of p0 , np and pmax
have shown that 1.4, 0.1 and 3.0 respectively are good choices. The overall computational proce-
dure is summarized in figure 3.3.

The optimization procedure was used to optimize the standard MBB–beam–example treated in
Olhoff, Bendsøe and Rasmussen, (1992). The MBB–beam is a simply supported beam subject
to a point load at the mid of the top edge as sketched in figure 3.4. Minimizing the compliance
of the MBB–beam discretized by 72x24 4–node quadrilateral elements and constraining the
amount of available material to 50% of the design domain resulted in the ”optimal” layout shown
in figure 3.5.

The resulting structure in figure 3.5 is seen to contain large regions with checkerboard patterns,
where checkerboard patterns are defined as regions with alternating void and solid elements
ordered in a checkerboard like fashion. The appearance of checkerboard patterns is a numerical
problem due to poor modelling of the stiffness of a checkerboard by lower order finite elements
and can not be interpreted as a kind of optimal porous microstructure. The actual stiffness of a
checkerboard consisting of solid and void elements is zero due to the stress singularities at the
corners of the solid regions as discussed in Berlyand and Kozlov, (1992).
68 3. The checkerboard and mesh dependency problems in layout optimization

Initialize
p = p0
p = p + np

Solve the FE problem (3.2)

Update design variables by (3.13)

Update Lagrangian multiplier 

by bisectioning

 converged ?
no
yes

no re converged ?

yes

pwpmax ?
no
yes
stop

Figure 3.3 Flowchart of the optimization algorithm

1/3

2
Figure 3.4 Design domain of the MBB–beam.

Figure 3.5 Checkerboard patterns in layout optimization.

 3. The checkerboard and mesh dependency problems in layout optimization 69

The checkerboard problem is discussed and the reason for its occurrence is explained extensively
in Jog and Haber, (1994) and Díaz and Sigmund, (1994), the latter is supplied in appendix 1 of
this report. The conclusion of the above mentioned papers is, that some precautions should be
and can be taken to avoid checkerboards.

Different approaches to overcome the checkerboard problem has been described in literature. A
very simple approach has been to ignore the presence of checkerboards during the design itera-
tion procedure and then to interpret checkerboard regions as regions with intermediate uniform
density during post–processing. Although this approach is simple to implement, it makes no
sense to use it because the correct stiffness of the checkerboard pattern is zero [Berlyand and
Kozlov, (1992)] and it can therefore not be interpreted as ”grey”. Another approach has been to
discretize the design domain by 8– or 9–node elements [Rodriques and Fernandes, (1994) and
Jog, Haber and Bendsøe (1994)1]. Although no checkerboards have been reported in practical
problems using 8– or 9–node elements so far, Jog and Haber, (1994) and Díaz and Sigmund,
(1994) show that checkerboards might appear in special cases, furthermore the use of 8– or
9–node elements suffers from the major drawback that computing time is increased by up to a
factor of 16 and storage space by a factor of 4 (or more) compared to the same discretization using
4–node elements. In a recent paper by Mullender, Huiskes and Weihnans, (1994) a bone remodel-
ling algorithm based on the assumption that bone growth at some point is dependent on loads
at all other point in the bone structure was proposed. The algorithm is reported to prevent forma-
tion of checkerboards patterns and at the same time producing mesh–independent solutions. The
algorithm proposed in this chapter can be compared to the algorithm proposed by Mullender,
Huiskes and Weihnans, (1994) although the origins of the algorithms are widely different.

In order to maintain the use of lower order finite elements and at the same time prevent the forma-
tion of checkerboards, a prevention scheme was proposed by Bendsøe, Díaz and Kikuchi,
(1993). The prevention scheme, which is applied after each design update, divides the design
domain into patches of four elements (a kind of super elements) and modifies the cell parameters
if checkerboards are detected in the patch. The prevention scheme does well in removing check-
erboards but still, a considerable amount of checkerboards regions are observed in the resulting
topologies. A reason for the methods lack in obtaining absolute checkerboard free solutions
might lie in the fact that each patch of 4 elements does not overlap with the neighboring patches
and thus checkerboards occurring ”between” patches are not ”detected” by the prevention
scheme. An improved checkerboard prevention scheme is therefore needed.
 70 3. The checkerboard and mesh dependency problems in layout optimization

Requirements to a good checkerboard prevention algorithm are the following:

S Checkerboard patterns must be removed entirely

S Computer time requirement must not be increased

S Must be simple to implement

S Must not destroy the stability of the optimization algorithm

S Must be applicable to any design domain

In the next sections, we will suggest a prevention scheme that tries to fulfill all the above men-
tioned criteria.

3.2 Checkerboard prevention scheme using image processing

This section will give a brief introduction to noise–reduction techniques in digital image proces-
sing and their possible applications to the prevention of checkerboard patterns in layout opti-
mization.

Digital image processing techniques are commonly used to process graphical data: good descrip-
tions and many examples of applications are given in Pratt, (1991). A graphical image is digitized
in to a finite number of pixels. The resolution, i.e. the number of pixels in the horizontal and verti-
cal directions, is mostly determined by the quality of the recording device (camera) or the dis-
playing device (monitor). Each pixel in a digital image represents a gray value and it is common
to distinguish 256 levels or gray, thus the color of one pixel can be stored in eight bits or one byte.
Here we will assume that the pixel values are continuous variables.

A common problem in image processing is noise coming from electrical sensor noise, transmis-
sion errors etc. Noise mostly appears as discrete isolated pixel variations, where pixels in error
visually appear markedly different from their neighbors.

The discretized design domain in layout optimization can be seen as a digital image. Each ele-
ment represents one pixel and the density in each element is represented by the gray scale: white
is void and black is solid material. Checkerboard patterns can then be interpreted as unwanted
noise in the layout optimized structure and thus one of the many powerful techniques developed
for noise cleaning can be used as a checkerboard prevention scheme.

Pratt, (1991) describes linear and non–linear noise cleaning techniques. Non–linear techniques
are reported to be superior to linear techniques but they can not be applied to layout optimization
 3. The checkerboard and mesh dependency problems in layout optimization 71

because the smoothness of the design problem would be destroyed. We will therefore settle for
the linear techniques.

Linear noise cleaning techniques can again be divided into two approaches: Fourier transforma-
tion based techniques and convolution based techniques.

Application of Fourier transformation based techniques immediately sound promising. The

Fourier transform of an image provides insight into the different frequency components of an
image. Cutting of high frequencies (low–pass filtering) and performing the inverse Fourier trans-
formation will result in an image where high frequency components (checkerboards) are effi-
ciently removed. The only, however grave, disadvantage of this method is that it only can be
applied to regular rectangular meshes. Of course the design domain could be divided into regular
subsections but this would require a complicated algorithm.

Convolution based techniques can also be used for linear noise cleaning. Properly executed, the
convolution technique corresponds to the Fourier based techniques however, the former does not
give the same intuitive insight into the nature of the noise cleaning filtering.

Following Pratt, (1991), the color (or density) of a pixel in the i’th row and the j’th column of
an image is denoted F(i, j). A spatially filtered output image G(i, j) is formed by a discrete con-
volution of the input image F(i, j) with an LxL impulse response array H(i, j). The convolution
process is formulated as

ȍ ȍ F(m, n) H(i–m ) C, j–n ) C)

L L
G(i, j) + (3.14)
m+1 n+1

where C=(L+1)/2 and the limits of i and j are max(1, i–C ) 1) v m v min(i ) C–1, M) and
max(1, j–C ) 1) v n v min(j ) C–1, N), where M and N are the numbers of pixels in the
vertical and horizontal directions of the picture respectively.

Choosing different impulse response matrices H, makes the convolution filter good for different
purposes, two examples are edge crispening and image restoration. For noise cleaning purposes,
H should be of low–pass form with only positive elements. Widely used is the 3x3 parametric
low–pass filter whose impulse response matrix is defined as

ȱ1 b 1ȳ
ǒ Ǔ ȧb
2
H+ 1 b 2 bȧ (3.15)
b)2
Ȳ1 b 1ȴ

where bŮ[1, R[.

 72 3. The checkerboard and mesh dependency problems in layout optimization

j
1 0 1 0 1 1/2 1/2 1/2 1/2 1/2

Ǔ ȱȧ2
0 1 0 1 0 1 2 1ȳ 1/2 1/2 1/2 1/2 1/2
ǒ
2
H+ 1 4 2ȧ
i 1 0 1 0 1 4 Ȳ1 2 1ȴ
1/2 1/2 1/2 1/2 1/2
0 1 0 1 0 1/2 1/2 1/2 1/2 1/2
1 0 1 0 1 1/2 1/2 1/2 1/2 1/2

F
ȍ ȍ F(m, n) H(m ) i–C, n ) j–C) +
3 3 G
G(i, j) +
m+1 n+1
2
ǒ1ń4Ǔ (1 1)0 2)1 1) 0 2)
1 4)0 2 ) 1 1)0 2)1 1) + 1ń2

Figure 3.6 Convolution of input image F with impulse response matrix H gives G. The filter parameter is
set to b=2.

For b approaching infinity, G(i, j) approaches F(i, j) and thus no filtering is performed. For
smaller values of b the output pixel value G(i, j) is a weighted average over the neighboring input
values.

The convolution procedure is demonstrated by a small example sketched in figure 3.6. Consider
the checkerboard pattern figure 3.6(left), where the numbers in each pixel denotes the density
of the pixel. Performing the convolution (3.14) with b=2 in (3.15), gives a uniform output pat-
tern. Note that the average density of the image remains constant under the convolution.

The linear low–pass convolution procedure reduces checkerboard noise efficiently, is easy to
implement, is applicable to any design domain and is therefore ideally suited for our purposes.

3.3 Implementation of the checkerboard prevention scheme

In this section, the linear convolution noise cleaning method will be applied to layout optimiza-
tion of structures.

The simplest way to perform the noise cleaning procedure is to take the resulting layout from
the optimization algorithm (i.e. figure 3.5) and postprocess it by the convolution procedure. This
approach would convert checkerboard regions into ”grey” regions which also are unpreferable
in the final design. Therefore the noise cleaning must be performed as a part of the optimization
algorithm.
 3. The checkerboard and mesh dependency problems in layout optimization 73

Applying the convolution procedure after each iteration in the optimization algorithm was tested
and resulted in checkerboard free final designs however, applying the noise cleaning after each
optimization step is a rather empirical approach and certainly destroys the optimality conditions
in (3.9). The conclusion is that the noise cleaning process must be integrated in the optimization
problem – how to do this is described in the following.

The element relative strain energies q e are modified as

ȍ Hf qf ǒòfǓp
9
q + 1e p
e
(3.16)
(ò )
f+1

where H f is the impulse response matrix defined in (3.15) written in vector form and the summa-
tion is taken over the nine elements in the neighborhood of element e. The modified element rela-
tive strain energies q e can be interpreted as the weighted average of strain energies in element
e and its eight direct neighbors.

A problem in the summation of (3.16) arises when element e is situated on the boundary of the
design domain and therefore has less than eight neighbors. One possibility to avoid this problem
is to assume that the non–existing elements have q f=0 but this tends to suppress formation of
material at the boundaries.

Instead we will only perform the summation over the nH existing neighboring elements of e. We
will therefore introduce a new filter vector H f* which is equal to H f defined in (3.15) without the
scaling factor 1/(b+2)2 . The modified relative strain energies (3.16) can then be written as

nH
q +
e 1
nH
ȍ Hf* qf ǒòfǓp (3.17)
(ò e)p ȍ Hi* f+1

i+1

where nH v 9. For nH = 9, (3.17) corresponds to (3.16).

The optimality criteria for the modified problem is defined as

ēL + – p (ò e)p–1 q e ) l v e– ae ) b e + 0 , e + 1,..., NE (3.18)

ēò e

and it is seen that the only change in (3.18) compared to the original optimality criteria (3.10),
is the exchange of the relative strain energies q e with the modified or averaged ones q e. The mod-
ification is rather heuristic and the actual objective function or Lagrangian with minimum
defined by (3.18) is not known for the time being however, numerical evidence shows that imple-
 74 3. The checkerboard and mesh dependency problems in layout optimization

mentation of the modified stationarity condition (3.18) in an iterative process, generates struc-
tures with decreasing compliance where checkerboard patterns are removed. For the filter factor
b approaching infinity the modified stationarity condition (3.18) approaches the original one
(3.10). Further explanation of the modified objective function will be subject to future research.

The overall computational procedure corresponds to the original procedure (figure 3.3), except
for the substitution of the element relative strain energies q e in (3.10) with the weighted averages
q e computed by (3.17).

Physically an infinite value of b corresponds to specifying an infinite ”cut–off” frequency of the

low–pass filter. Decreasing b corresponds to lowering the cut–off frequency and thereby filtering
out rapidly varying densities. When the filter factor is chosen below a certain value bcheck , all
checkerboard patterns are filtered out. To find a critical value of bcheck , we will apply the con-
volution filter to the previously considered MBB–beam design problem.

Figure 3.7 shows the optimal layouts of the MBB beam discretized by 120x40=4800 elements,
where different filter parameter values b were used for the convolution filter. The filtered designs
can be compared with the unfiltered design at the top of the figure. It is seen that the filter value
b=100 hardly removes any checkerboards. For decreasing values of b, the checkerboard areas
gets smaller and for b=15, they are entirely removed. Decreasing b further results in thicker bars
of the structure. It can be concluded that the value of the filter parameter b in (3.16) should be
15 (or 12 to be on the safe side), to get the most efficient solutions without checkerboards. The
critical value of bcheck might change for other formulations (other elements or material models)
and should be determined by experiments as done here.

A few words shall be devoted to the practical application of the checkerboard prevention scheme
and its applicability to other design formulations or problems.

Finding the neighbors defined by the filter window is very simple in the case where the design
domain is discretized as a regular mesh in a rectangular domain. In the case of more complicated
geometries and irregular meshes another approach must be taken. In preprocessing the design
problem, we will make a call to a subroutine which produces a list of neighbors of each element.
The subroutine must consist of a search loop where the (eight) neighboring elements to the con-
sidered element must be found. The search might be computationally expensive but as the sub-
routine only has to be called once in the beginning of procedure, the problem is not considered
of major concern.

The method is directly applicable to the ”homogenization approach to layout optimization”,

where the design variables are densities and orientation of the ranked material in each element.
 3. The checkerboard and mesh dependency problems in layout optimization 75

no control
W=199.6

b=100
W=202.0

b=30
W=204.6

b=20
W=205.7

b=15
W=207.0

b=10
W=208.2

b=5
W=211.2

b=2
W=215.0

Figure 3.7 Influence of filter factor b on the optimal layout of the MBB–beam defined in figure 3.4. The ground
structure (half of the beam) consists of 120x40=4800 4–node elements and the volume is restricted to 50% of
the design domain.
76 3. The checkerboard and mesh dependency problems in layout optimization

The element relative strain energies are modified as described above and the lamination angles
can be determined the usual way.

Discretizing the structure by triangular elements would require a modification of the impulse
response matrix H in (3.15), which would have to be reduced to four elements (only direct neigh-
bors). Experiment would have to be performed to find the right weighting factors.

The described procedure can easily be extended to three dimensions by adding a third dimension
to the impulse response matrix. Assuming a discretization by 8–node brick elements, the impulse
response matrix H will contain 27 factors.

3.4 Elimination of mesh–dependency

Applying the checkerboard prevention scheme proposed in the last section to the MBB–beam
example discretized by 210x70=14700 elements, we get the optimal topology shown in figure
3.8.

Comparing figure 3.8 with the optimal topology using 4800 elements and filter factor 15 (figure
3.7), we notice that the solution has become more detailed. In other words, the solution becomes
more chattering when the mesh is refined. The mesh–dependency is explained by the ill–posed-
ness of the penalized problem discussed earlier. For practical reasons we are interested in
”coarse” solutions with a few large holes and we must therefore find a method to ”fix” the topol-
ogy no matter how fine the mesh–discretization is made. Eventually mesh–refinement shall only
result in smoother and better descriptions of the boundaries between solid and void regions in
the optimal layout.

An approach to eliminate the problem of mesh dependency and at the same time ensuring a
unique solution is the perimeter control method proposed by Jog, Haber and Bendsøe, (1994)2.
The formulation of the perimeter control method ensures a well–posed problem by combining
a constraint on the global perimeter of the structure with penalization on intermediate density
values. Implementation of the method involves adding a (non–linear) constraint in the optimality

Figure 3.8 MBB–beam example discretized by 210x70=19200 elements and filter factor b=15.
W=203.4.
 3. The checkerboard and mesh dependency problems in layout optimization 77

criteria and a critical choice of penalty factors and therefore, convergence is reported to be rather
unstable.

To get a robust procedure which requires no interaction of the engineer and ensures mesh–inde-
pendency of the solution, a modification of the checkerboard prevention algorithm described in
the last section is proposed in the following.

From figure 3.7 can be observed that low values of the filter factor b give more manufacturable
designs. Mesh–independent designs could therefore be obtained by specifying lower values of
b for finer meshes. However, this approach is rather empirical and it is difficult to assign a filter
factor to a given mesh. We will therefore modify the checkerboard convolution filter such that
the filter area is independent on the mesh size.

For the checkerboard prevention algorithm, the size of the convolution filter H i* was fixed to 3x3
elements. Extending the filter size to the elements within a radius rmin of element e, we can define
a convolution filter

^
H f + v f [rmin – dist(e, f)] , {f Ů NE | dist(e, f) v r min} (3.19)

where the operator dist(e,f) is defined as the distance between the center of element e and the
center of element f, and v f is the volume of element f.

If n r denotes the number of elements which have centers within the radius rmin of element e, the
modified relative strain energies can be written in parallel to (3.17) as

ȍ Hf qf ǒòfǓp
nr
^
^e 1
q + (3.20)
ȍ
nr
^
(ò e)p H i f+1
i+1

The weights of the filter can be seen as a cone with bottom radius rmin and volume one.

To clarify what happens to the strain energies when the filter is applied, we will consider two
simple (one dimensional examples). Without any filtering a bar in a layout can have the thickness
of only one element as sketched in figure 3.9a. The coordinate system shows the unfiltered strain
energies q e(ò e) p as a function of the elements in a column of a layout. Applying the filter with
e
rmin equal to three times the element size results in the modified strain energies q^ (ò e) p shown
in figure 3.9b. The sums of the unfiltered and filtered strain energies are equal but the strain
energy distribution is changed to a cone like shape clearly seen for the fine mesh in figure 3.9c.
The base radius of the cone is rmin .
 78 3. The checkerboard and mesh dependency problems in layout optimization

Figure 3.10 shows the filter effect on a double impulse which physically can represent two thin
bars. In figure 3.10a, the distance between the two bars is rmin . The effect of filtering with filter
size rmin is shown in figure 3.10b and we notice, that the two bars have been merged to one but
the sum of the strain energies (=compliance) is preserved. It can be concluded that applying the
filter introduces a length scale rmin underneath which structural variation is not allowed.

q e (re) p q e (re) p q e (re) p

filter

e rmin e rmin e

a) No filter, coarse mesh b) Filtered, coarse mesh c) Filtered, fine mesh

Figure 3.9 Change of impulse after filtering with the mesh–independency filter.

q e (re) p q e (re) p

filter

rmin e e

a) No filter b) Filtered

Figure 3.10 Change of double impulse after filtering with the mesh–indepen-
dency filter.
 3. The checkerboard and mesh dependency problems in layout optimization 79

The modified optimality criteria for the mesh–independency algorithm is written as

ēL + – p (ò e)p–1 q^ e ) l v e– ae ) b e + 0 , e + 1,..., NE (3.21)

ēò e

As for the checkerboard scheme defined earlier it must be emphasized that the modification is
based on heuristics but numerical evidence shows that implementation of the procedure in an
iterative process generates structures with decreasing compliance.

The present mesh–independency algorithm can be seen as a ”mix” of compliance optimization

and image processing which gives good results. By filtering the strain energies with the extended
^
impulse response matrix H f, we have introduced a length scale rmin , underneath which structural
variation is not allowed.

We can also say, that we have separated the design variables from the finite element discretiza-
tion. Keeping the value of rmin fixed, independent on mesh refinements, means that the allowed
variation of the density is mesh–independent and we will therefore expect to get mesh–indepen-
dent optimal layouts. The boundaries between solid and void regions will be monotonously vary-
ing grey scales of width approximately equal to rmin .

Choosing rmin to be 1.1 times the element size makes the modified filter very similar to the check-
erboard filter (3.15) with filter factor b=15. Therefore, rmin should always be chosen greater or
equal to 1.1 times the maximum element size to prevent formation of checkerboards.

The modified impulse response filter (3.19) was applied to the MBB–beam example. Figure 3.11
shows the optimal topologies obtained for rmin =0.04 and varying numbers of elements. It is seen
in the figure that topologically similar structures are obtained independent on the grid refine-
ment.

Figure 3.12 shows three different topologies obtained from a fixed discretization of 120x40 ele-
ments and varying values of rmin . It is seen from the figure, that the complexity of the topologies
can be controlled by the choice of rmin and that the smallest bars in the topologies are of approxi-
mately the same size as rmin .
80 3. The checkerboard and mesh dependency problems in layout optimization

30x10
W=229.9

60x20
W=237.7

90x30
W=239.5

120x40
W=240.7

150x50
W=241.5

Size of rmin:

Figure 3.11 Elimination of the mesh–dependency problem by application of the modified convolution filter with
rmin =0.04. Mesh discretizations ranging from 30x10 to 150x50. The design problem is defined in figure 3.4.
 3. The checkerboard and mesh dependency problems in layout optimization 81

a) rmin
W=208.8

b) rmin
W=220.9

c) rmin
W=240.7

Figure 3.12 Dependency of optimal topology on filter size. Groundstructure has 120x40 elements and
a) rmin =0.01, b) rmin =0.02 and c) rmin =0.04.
 82 3. The checkerboard and mesh dependency problems in layout optimization

Another example is shown in figure 3.13. In this example the topology of a ”bicycle wheel” is
optimized. The figure shows the optimal layouts for different mesh sizes obtained with and with-
out the modified convolution filter.

Design problem .625

.125

16x20 16x20
W=35.5 W=40.2

48x60 48x60
W=31.2 W=40.6

64x80 48x60
W=31.0 W=40.8

Figure 3.13 The Bicycle Wheel example with volume fraction 20%. The three solutions left shows mesh–depen-
dency of optimal layout (original checkerboard filter, b=15). The three solutions right are generated using the
mesh–independency algorithm with rmin =0.04 and shows topologically similar solution for different meshes.
 3. The checkerboard and mesh dependency problems in layout optimization 83

Many test examples used in literature (and here) are very idealized in the sense that groundstruc-
tures are symmetric, rectangular, discretized by quadratic elements and loads are applied either
horizontally or vertically. ”Real life” problems have complex geometries and non–symmetric
loadings and therefore algorithms which work well on ”laboratory problems” might fail to work
on more complicated problems. Consequently, A. R. Díaz (in private communication) has sug-
gested a standard ”worst case” test example shown in figure 3.14(upper right). The ”worst case”
design problem is a variation of the well known short cantilever problem in figure 3.14(upper
left). The challenge in the ”worst case” problem has three sources. First, the original problem
has been rotated, second, the individual elements are rectangular not quadratic such that ele-
ment–to–element symmetry of the solution is hindered and third, the starting guess is non–uni-
form varying from zero density at the left edge to full density at the right edge. Ideally, solving
the original and the modified problem using a topology optimization algorithm should result in
two similar (symmetric) topologies because we are seeking mesh–independent solutions.

The ”worst case” example was used to test the proposed checkerboard and mesh–independency
algorithms. As a reference topology we have figure 3.14a which was produced by the mesh–in-
dependency algorithm with rmin =0.035 (1.2 times maximum element size). The solution to the
worst case problem using the checkerboard prevention scheme is seen in figure 3.14b. It is seen
that there are no checkerboards but the topology is unsymmetric which could be expected
because the checkerboard filter has become unsymmetric due to the rectangular elements. Solv-
ing the problem with the mesh–independency algorithm with rmin =0.035 results in the topology
in figure 3.14c which is seen to be symmetric and similar to the topology of the original problem
in figure 3.14a. To demonstrate the importance of using the continuation method (gradual
increase of penalty factor p), the worst case problem was solved by the mesh–independency algo-
rithm but fixing the penalty factor p to 3 from the first iteration. The resulting topology (figure
3.14d) is seen to be very unsymmetric.

From this example, it can be concluded that the proposed mesh–independency algorithm must
be preferred to the checkerboard prevention scheme because the former produces mesh–inde-
pendent and checkerboard free designs.
84 3. The checkerboard and mesh dependency problems in layout optimization

Original design problem ”Worst case” design problem

1 1

40 elements

40 elements
5/8

5/8

60 elements 60 elements

a) Solution to original problem. c) Solution to ”Worst case” problem.

W=40.7,rmin =0.035. W=40.0, rmin =0.035.

b) Solution to ”Worst case” problem. d) Solution to ”Worst case” problem.

W=38.7, b=15. W=40.2, rmin =0.035.
Without continuation approach.

Figure 3.14 ”Realistic” design problem, shows difference in optimal topologies obtained by: b) checker-
board control, c) mesh–independency control and d) mesh–independency control without continuation
approach. All groundstructures have 60x40=2400 elements and volume fraction are 25%.
 3. The checkerboard and mesh dependency problems in layout optimization 85

Figure 3.15 Contour plots of the above 0.5 density level of the the MBB–example in figure 3.12a–c.

The optimal topologies obtained by the mesh–independency filter do not need much post–pro-
cessing. A contour plotting program (here we have used the software MATHEMATICA) can
threshold the density values 0.5 as shown in figure 3.15. The resulting boundaries between
solid and void are seen to be smooth but should still be analyzed for stress concentrations by some
general FE–program or shape optimization software.

A fundamental difference exists between the mesh–independency algorithm proposed here and
the perimeter control algorithm proposed in Jog, Haber and Bendsøe, (1994)2. In the latter
approach, a constraint on the overall perimeter is imposed which means that we do get mesh–in-
dependent solutions but in principle nothing prevents a particular bar in becoming thinner and
thinner with mesh refinement. The approach presented here prevents bars or holes that are
smaller than rmin in being formed, and therefore the method both eliminates mesh–dependency
and keeps characteristic sizes constant with mesh refinement. This does not mean that one
method should be preferred from the other but it means, that layouts obtained from the two meth-
ods can not be directly compared.
86 3. The checkerboard and mesh dependency problems in layout optimization

3.5 Summary
In this chapter, we have proposed a computationally simple procedure to prevent two well known
problems in layout optimization, namely, the problem of checkerboard–like patterns occurring
in optimal designs due to bad finite element discretizations and the problem of mesh–depen-
dency of optimal solutions due to non–existence of solutions to the penalized problem.

The prevention scheme is based on noise cleaning methods from the field of image processing.
The scheme can be seen as a low–pass filter which prevents rapid oscillation of the densities of
a structure. The cut–off frequency is independent on the discretization and we thereby obtain an
algorithm where the operator can control characteristic sizes such as minimum bar or hole sizes.

Several numerical examples show that the proposed method gives efficient checkerboard–free
and mesh–independent designs, readily applicable to manufacturing and with the possibility of
controlling characteristic sizes such as minimum diameter of holes and minimum thickness of
bars.

Although based on heuristics, the numerical experiments are encouraging and current effort
therefore tries to put the procedure into a mathematical framework.
 Conclusions
In the present thesis, a numerical method for design of linear elastic materials with prescribed
thermoelastic properties has been proposed. The design problem is formulated as a topology
optimization problem where the objective function is to minimize weight pr. unit volume and
the design degrees of freedoms are geometrical properties of the periodic material microstruc-
tures represented by base cells.

To analyze effective properties of a base cell discretized by finite elements, a numerical homoge-
nization procedure is developed and applied to truss–, frame– and continuum–like discretiza-
tions of microstructures. Great savings in computer time and storage requirements can be
obtained by taking advantage of sparse and iterative methods in solving the associated finite ele-
ment problem.

The truss–, frame– and continuum–like discretizations are used as groundstructures or starting
guesses for the optimization problem, which means that the optimal microstructure for pre-
scribed elastic properties is sought within a fixed but very detailed discretization.

As demonstrated by many examples, the proposed numerically based design procedure can be
used to tailor materials with a wide range of thermoelastic properties.

Using truss–like discretizations in two and three dimensions makes it possible to design materials
with properties very close to the theoretical bounds (positive semi–definiteness of the elastic ten-
sor) whereas using more practical frame elements damps out attainable properties. Both the truss
and the frame formulations allow bars and beams to cross without interaction. This means that
practical realization of the microstructures will be difficult due to crossing members and sliding
surfaces.

More realistic microstructures are obtained by the continuum formulation. Examples show that
extreme elastic parameters can be obtained by the continuum formulation as well, however,
experiments show that only low overall stiffness can be obtained if extreme elastic properties
are prescribed. Generally it can be said that any topological constraint imposed on the micro-
structural geometry damps out the attainable elastic properties. As demonstrated in an example,
the attainable properties are also damped when one of the proposed porous microstructures is
imbedded in a softer matrix material.

Isotropic materials with negative Poisson’s ratios are obtained from all three discretization types
but the continuum topology is preferred due to the above mentioned manufacturability
constraints.
88 Conclusions

The possibility of designing materials with prescribed thermoelastic properties is demonstrated

by a simple example using the continuum discretization.

The numerically predicted material properties are tested by two working models of truss like
microstructures with negative Poisson’s ratios and showed the expected behaviour. Unfortu-
nately, due to simplicity and coarseness of the models, none of them allowed actual measure-
ments of their elastic properties.

The microstructures obtained by the numerical method developed in this thesis shall not be seen
as readily manufacturable solutions but rather as sources of inspiration to people involved in
practical synthesis of novel microstructures.

Two well known problems in general topology optimization of structures, namely the checker-
board problem and the mesh–dependency problem were encountered in design of the continu-
um–like microstructures. The checkerboard problem refers to the formation of regions in opti-
mized topologies with alternating solid and void elements ordered in a checkerboard like
fashion. The reason for the occurrence of checkerboards is discussed in Díaz and Sigmund,
(1994), supplied in appendix 1, and can simply be explained by poor numerical modelling of the
stiffness of checkerboards by lower order finite elements. The mesh dependency problem refers
to the non–convergence of solutions with mesh refinement. A procedure to eliminate both prob-
lems based on ideas borrowed from image processing has been developed. Although based on
heuristics, the proposed procedure shows promising results when applied to topology optimiza-
tion of continuum structures.

As a continuation of the work presented in this thesis, following subjects are suggested:

S More realistic prototypes of the proposed materials can be produced by the rapid
prototyping technique. Actual properties of the prototype materials can be tested
by standard mechanical test procedures and can be compared with the theoreti-
cally predicted values.

S Contact to researchers involved in practical synthesis of novel microstructures

can be established. In connection with this, the idealized modelling of micro-
structures used in this work might have to be extended to handle non–linearities
or other problems encountered in the practical implementation.

S The optimal microstructural topologies for the example with prescribed thermo-
elastic properties showed ”actuator” like elements which suggests that the present
approach to material design can be used for design problems like optimal place-
ment or configuration of actuators or sensors in the design of smart materials.
 Conclusions 89

S Although the suggested procedure to prevent the checkerboard and mesh–depen-

dency problems is based on heuristics, the numerical results are encouraging and
efforts invested in formulating the procedure in a mathematical framework seem
rewarding.
 List of publications
Lyngby, January 1993 – December 1994.
Díaz, A.R. and Sigmund, O. (1995). Checkerboard patterns in layout optimization, Structural
Optimization, 10(1), pp. 40–45.
Sigmund, O. (1994)1. Materials with prescribed constitutive parameters: an inverse homogeniza-
tion problem, Int. J. Solids Struct., 31(17), pp. 2313–2329.
Sigmund, O. (1995)2. Tailoring materials with prescribed elastic properties”. Mech. Materials.
20, pp. 351–368.
Sigmund, O. (1995)3. Tailoring materials for specific needs. J. of Intelligent Structures and
Materials). 5, pp. 736–742.
Work published in Essen, October 1991 – December 1992.
Rozvany, G.I.N., Lewinski, T, Sigmund, O., Gerdes, D. and Birker, T. (1993). Optimal topology
of trusses or perforated deep beams with rotational restraints at both ends. J. Struct. Optim., 5,
pp. 268–270.
Rozvany, G.I.N., Sigmund, O., and Birker, T. (1993). Optimal design of composite and fibre–re-
inforced plates. In Optimization with Advanced Materials, (ed. Pedersen, P.), Elsevier, Amster-
dam, pp. 293–309.
Rozvany, G.I.N., Sigmund, O., Lewinski, T, Gerdes, D., and Birker, T. (1993). Exact optimal
layouts for non–self–adjoint problems. J. Struct. Optim., 5, pp. 204–206.
Rozvany, G.I.N., Sigmund, O., Zhou, M. and Birker, T. (1993). Iterative discretized methods for
layout optimization and generalized shape optimization. In Structural Optimization 1993, Pro-
ceedings (ed. Herskovits, J.), Rio de Janeiro, Brazil, pp. 139–151.
Rozvany, G.I.N., Zhou, M., Birker, T. and Sigmund, O. (1993). Topology optimization using
iterative Continuum–type Optimality Criteria (COC) methods for discretized systems. In Topol-
ogy optimization of structures, (eds. Bendsøe, M.P.; Mota Soares, C.A.), Kluwer, Dordrecht, pp.
273–286.
Rozvany, G.I.N., Zhou, M., Lewinski, T. and Sigmund, O. (1993). Optimal layout theory: new
classes of exact solutions for trusses including several load conditions and non–self–adjoint
problems. In Structural Optimization 1993, Proceedings (ed. Herskovits, J.), Rio de Janeiro,
Brazil, pp. 37–46.
Rozvany, G.I.N., Zhou, M. and Sigmund, O. (1994). Topology optimization in structural design.
In Advances in Design Optimization, (ed. Adali, H.), Chapman and Hall, London, UK, pp.
340–399.
Sigmund, O., Zhou, M. and Rozvany, G.I.N. (1993). Layout optimization of large FE systems
by new optimality criteria methods: applications to beam systems. In Concurrent Engineering
Tools and Technologies for Mechanical Systems Design, (ed. Haug, E.J.), Springer Verlag, New
York, pp. 803–819.
 References
Aboudi, J. (1991). Mechanics of Composite Materials. Elsevier Science Publishers B.V.

Allaire, G. and Francfort, G.A. (1993). A numerical algorithm for topology shape optimization.
In Topology optimization of structures, (eds. Bendsøe, M.P.; Mota Soares, C.A.), Kluwer, Dor-
drecht, pp. 239–248.

Allaire, G. and Kohn, R.V. (1993)1. Optimal design for minimum weight and compliance in
plane stress using extremal microstructures. European J. Mech. A, 12, pp. 839–878

Allaire, G. and Kohn, R.V. (1993)2. Topology optimization and optimal shape design using
homogenization. In Topology optimization of structures, (eds. Bendsøe, M.P.; Mota Soares,
C.A.), Kluwer, Dordrecht, pp. 207–218.

Almgren, R.F. (1985). An isotropic three–dimensional structure with Poisson’s ratio = –1. J.
Elasticity, 15, pp. 427–430.

Ashby, M.F (1991). Materials and shape, Acta. Metall. mater., 39(6), pp. 1025–1039.

Ashby, M.F and Jones, D.R.H. (1986). Engineering Materials 2, Pergamemnon Press.

Ashley, S. (1994)1. Small–scale structure yields big property payoffs. Mech. Engng., February,
pp. 52–57.

Ashley, S. (1994)2. Prototyping with advanced tools. Mech. Engng., June, pp. 48–55.

Autio, M., Laitinen, M. and Pramila, A. (1993). Systematic creation of composite structures with
prescribed thermomechanical properties. Comp. Eng. 3(3), pp. 249–259.

Avellaneda, M. (1987). Optimal bound and microgeometries for elastic two–phase composites.
SIAM J. App. Math., 47, pp. 1216–1228.

Bakhvalov, N. and Panasenko, G. (1989). Homogenization: averaging processes in periodic

media. Kluwer Academic Publishers, Dordrecht.

Bendsøe, M.P. (1989). Optimal shape design as a material distribution problem. Struct. Optim.,
1, pp. 193–202.

Bendsøe, M.P. (1994). Methods for the Optimization of Structural Topology, Shape and Mate-
rial. Springer.

Bendsøe, M.P., Díaz, A.R. and Kikuchi, N. (1993). Topology and generalized layout optimiza-
tion of elastic structures. In Topology optimization of structures, (eds. Bendsøe, M.P.; Mota
Soares, C.A.), Kluwer, Dordrecht, pp. 159–206.

Bendsøe, M.P., Díaz, A.R., Lipton, R. and Taylor, J.E. (1993)2. Optimal design of material prop-
erties and material distribution for multiple loading conditions. To appear in Int. J. Num. Meth.
Engng. 38(7), pp. 1149–1170.
 References 3

Bendsøe, M.P., Guedes, J.M., Haber, R.B., Pedersen, P. and Taylor, J.E. (1994)1. An analytical
model to predict optimal material properties in the context of optimal structural design. J. Appl.
Mech., 61(4), pp. 930–937.

Bendsøe, M.P. and Kikuchi, N. (1988). Generating optimal topologies in structural design using
a homogenization method. Comp. Meth. Appl. Mechs. Engng., 71, pp. 197–224.

Benssousan, A., Lions, J.L. and Papanicolau, G. (1978). Asymptotic analysis for periodic struc-
tures. North Holland, Amsterdam.

Berlyand, L.V. and Kozlov, S.M. (1992). Asymptotics of homogenized moduli for the elastic
chess–board composite. Arch. Rational Mech. Anal., 118, pp. 95–112.

Bourgat, J.F. (1977). Numerical experiments of the homogenization method for operators with
periodic coefficients. Lecture Notes in Mathematics 704, Springer Verlag, Berlin, pp. 330–356.

Caddock, B.D. and Evans, K.E. (1989). Microporous materials with negative Poisson’s ratios:
I. Microstructure and mechanical properties. J. Phys. D: Appl. Phys., 22, pp. 1877–1882.

Cheng, G. and Olhoff, N. (1981). An investigation concerning optimal design of solid elastic
plates. Int. J. Solids Struct., 16, pp. 305–323.

Cherkaev, A.V. and Gibiansky, L.V. (1993). Coupled estimates for the bulk and shear moduli of
a two–dimensional isotropic elastic composite. J. Mech. Phys. Solids 41(5), pp. 937–980.

Choi, J.B. and Lakes, R.S. (1991). Design of fastener based on negative Poisson’s ratio foam.
Cellular Polymers, 10(3), pp. 205–212.

Cioranescu, D. and Saint Jean Paulin, J. (1986). Reinforced and honey–comb structures, J. Math.
Pures et Appl. 65, pp. 403–422.

Cook, R.D., Malkus, D.S., Plesha, M.E. (1989). Concepts and Applications of Finite Element
Analysis. John Wiley & Sons.

Díaz, A.R. and Kikuchi, N. (1992). Solutions to shape and topology eigenvalue optimization
problems using a homogenization method. Int. J. Num. Meth. Engng., 35, pp. 1487–1502.

Díaz, A.R., Lipton, R. and Soto, C.A. (1994). A new formulation of the problem of optimum
reinforcement of Reissner–Mindlin Plates, Preprint, Dept. of Mech. Engng., Michigan State
University, East Lansing, USA.

Díaz, A.R. and Sigmund, O. (1994). Checkerboard patterns in layout optimization. (Submitted).

Dvorak, J. (1993). Optimization of composite materials. Master’s Thesis, Faculty of Mathema-

tics and Physics, Charles University, Prague, Czech Republic.

Evans, K. E. (1989). Tensile network microstructures exhibiting negative Poisson’s ratios. J.

Phys. D: Appl. Phys., 22, pp. 1870–1876.

Evans, K. (1990). Tailoring the negative Poisson’s ratio. Chemistry & Industri, 15, October .
4 References

Ferencz, R.M. (1989). Element–by–element preconditioning techniques for large–scale, vecto-

rized finite element analysis in nonlinear solid and structural mechanics. Ph.D.–thesis, Stanford
University.

Gibson, C.S. and Ashby, M.F. (1988). Cellular Solids. Pergamemnon, Oxford.

Greenbaum, A., Greengard, L. and McFadden, G.B. (1993). Laplace’s equation and the dirich-
let–neumann map in multiply connected domains. J. Comp. Phys., 105, pp. 267–278.

Guedes, J.M. and Kikuchi, N. (1991). Preprocessing and postprocessing for materials based on
the homogenization method with adaptive finite element methods. Comp. Meth. Appl. Mech.
Eng. 83, pp. 143–198.

Guedes, J.M., Neves, M.M. and Orlando, J. (1994). Users manual for the program PREMAT2D,
Vers. 2.1. Report Instituto Superior Technico, Cemul, Portugal.

Hollister, S.R. and Riemer, B.A. (1993). Digital image based finite element analysis for bone
microstructure using conjugate gradient and Gaussian filter techniques. SPIE Vol. 2035 Math.
Meth. in Medical Imaging II, pp. 95–106.

Huiskes, R. and Hollister, S.J. (1993). From structure to process, from organ to cell: recent deve-
lopments of FE–analysis in orthopaedic biomechanics. J. Biomech. Engng., 115, pp. 520–527.

Ikuta, K., Hirowatari, K. and Ogata, T. (1994). Ultra high resolution stereo lithography for three
dimensional micro fabrication. In Proc. 5’th Int. Conf. on Rapid Prototyping, June 12–15, Day-
ton, Ohio.

Jog, C. (1994). Variable–topology shape optimization of linear elastic structures. Ph.D.–thesis,

Dept. of Theoretical and Applied Mechanics, Univ. of Illinois at Urbana–Champaign, USA.

Jog, C. and Haber, R.B. (1994). Stability of finite element models for distributed–parameter opti-
mization and topology design. TAM–report No. 758, Dept. of Theoretical and Applied Mechan-
ics, Univ. of Illinois at Urbana–Champaign, USA.

Jog, C., Haber, R.B. and Bendsøe, M.P. (1994)1. Topology optimization with optimized, self–
adaptive materials. Int. J. Num. Meth. Engng., 37, pp. 1323–1350.

Jog, C., Haber, R.B. and Bendsøe, M.P. (1994)2. Variable–topology shape optimization with a
constraint on the perimeter. In Proc. 20’th ASME Design Automation Conference, Minneapolis,
Mn., Sept. 11–14, ASME publ. DE–Vol. 69–2, 261–272.

Kohn, R.V. and Strang, G. (1986). Optimal design and relaxation of variational problems.
Comm. Pure Appl. Math., 39, pp. 1–25, 139–182, 353–377 (Part I–III).

Kolpakov, A.G. (1985). Determination of the avarage characteristics of elastic frameworks.

PMM J. appl. Math. Mech. U.S.S.R. 49, pp. 739–745.

Lakes, R. (1987). Foam Structures with Negative Poisson’s Ratio. Science 235(Feb.), p. 1038.
 References 5

Lakes, R. (1993)1. Design considerations for materials with negative Poisson’s ratios. J. Mech.
Design, 115, pp. 696–700.

Lakes, R. (1993)2. Materials with structural hierarchy, review article. Nature 361(Feb.), pp.
511–515.

Lakes, R. and Elms, K. (1993). Indentability of conventional and negative Poisson’s ratio foams.
J. Comp. Mat., 27(12), pp. 1193–1202.

Kostar, T.D. and Chou, T.–W. (1994). Microstructural design of advanced multi–step three–di-
mensional braided preforms. J. Comp. Mater., 28(13), pp. 1180–1201.

Michell, A.G.M. (1904). The limit of economy of material in frame structures. Philosophical
Magazine, Series 6, 8, 589–597.

Milton, G.W. (1992). Composite materials with Poisson’s ratios close to –1. J. Mech. Phys. So-
lids, 40(5), pp. 1105–1137.

Milton, G.W. and Cherkaev, A.V. (1993). Materials with elastic tensors that range over the entire
set compatible with thermodynamics. Presented at the Joint ASCE–ASME–SES Mett’N’, June
1993, University of Virginia, Charlottesville, Virginia, USA.

Moore, J.S. (1993). Odd solid predictions. Nature, 365, p. 690.

Moulinec, H. and Suquet, P. (1994). A fast numerical method to compute the linear and nonlinear
mechanical properties of composites. C. R. Acad. Sci. Paris, t. 318, Serie II, pp. 1417–1423.

Mullender, M.G., Huiskes, R. and Weihnans, H. (1994). A Physiological approach to the simula-
tion of bone remodelling as a self–organizational control process. J. Biomech., 11, pp.
1389–1394.

Nemat–Nasser, S. and Hori M. (1993). Micromechanics. Elsevier, Amsterdam.

Nkansah, M.A., Evans, K.E. and Hutchinson, I.J. (1994). Modelling the mechanical properties
of an auxetic molecular network. Modelling Simul. Mater. Sci. Eng., 2, pp. 337–352.

Olhoff, N. Bendsøe, M.P. and Rasmussen, J. (1992). On CAD–integrated structural topology and
design optimization. Comp. Meth. Appl. Mechs. Engng., 89, pp. 259–279.

Sanchez–Palencia, E. (1980). Non–homogeneous Media and Vibration Theory. Lecture Notes

in Physics 127, Springer Verlag, Berlin.

Pedersen, P. (1970). On the minimum mass layout of trusses. AGARD conf. proc. No. 36, Sympo-
sium on Structural Optimization, AGARD–CP–36–70.

Pratt, W.K. (1991). Digital Image Processing. John Wiley and Sons, New York.

Press, W.H., Teukolsky, S.A., Vetterling, W.T. and Flannery, B.P. (1992). Numerical Recipes in
Fortran, Cambridge University Press.
6 References

Ringertz, U. (1988). A mathematical programming approach to structural optimization.

Research Report No. 88–24, Dept. of Leightweight Structures, The Royal Institute of Technol-
ogy, Stockholm, Sweeden.

Ringertz, U. (1993). On finding the optimal distribution of material properties. Struct. Optim.,
5, pp. 265–267.

Rodriques, H.C. and Fernandes, P. (1995). A material based model for topology optimization
of thermoelastic structures. Int. J. Num. Meth. Engng., 38(12), pp. 1951–1965.

Rothenburg, L., Berlin, A.A. and Bathurst, J. (1991). Microstructures of isotropic materials with
negative Poisson’s ratio. Nature, 354, pp. 470–472.

Rozvany, G.I.N., Zhou, M., Birker, T. and Sigmund, O. (1992). Topology optimization using
iterative continuum type Optimality Criteria (COC) methods for discretized systems. In Topol-
ogy Design of Structures, (ed. Bendsøe, M.P. and Soares, C.A.), Springer, pp. 273–286.

Rozvany, G.I.N., Zhou, M. and Sigmund, O. (1994). Topology optimization in structural design.
In Advances in Design Optimization, (ed. Adali, H.), Chapman and Hall, London, UK, pp.
340–399.

Sanchez–Palencia, E. (1980). Non–homogeneous media and vibration theory. Lecture Notes in

Physics, 127, Springer Verlag, Berlin.

Sankaranarayanan, S., Haftka, R.T. and Kapania, R.K. (1993). Truss topology optimization with
stress and displacement constraints. In Topology Design of Structures, (ed. Bendsøe, M.P. and
Soares, C.A.), Springer, pp. 71–78.

Sigmund, O. (1994)1. Materials with prescribed constitutive parameters: an inverse homogeni-

zation problem, Int. J. Solids Structures, 31, pp. 2313–2329.

Sigmund, O. (1994)2. Tailoring Materials with Prescribed Elastic Properties, to appear in Mech.
Materials. 20, pp. 351–368.

Walker, K.P., Freed, A.D. and Jordan, E.H. (1991). Microstress analysis of periodic composites.
Comp. Engng., 1(1), pp.29–40.

Weihnans, H., Huiskes, R. and Grootenboer, H.J. (1992). The behaviour of adaptive bone–re-
modeling simulation models. J. Biomech., 25(12), pp. 1425–1441.

Zhou, M. and Rozvany, G.I.N. (1993). DCOC: an optimality criteria method for large systems,
Part I: Theory, Struct. Optim., 5, pp. 12–25.
 Appendix 1

Checkerboard patterns in layout optimization

Alejandro R. Díaz
Department of Mechanical Engineering
Michigan State University
East Lansing, MI 48824
USA

Ole Sigmund
Department of Solid Mechanics
Technical University of Denmark
DK–2800 Lyngby
Denmark

Abstract
Effective properties of arrangements of strong and weak materials in a checkerboard fashion are
computed. Kinematic constraints are imposed so that the microscale deformation is consistent
with typical finite element approximations. It is shown that when four–node quadrilateral ele-
ments are involved, these constraints result in a numerically induced, artificially high stiffness.
This can account for the formation of checkerboard patterns in continuous layout optimization
problems of compliance minimization.

Comment to pdf–version, July, 1998: This paper is left out but has appered as:
Díaz, A.R. and Sigmund, O. (1994). Checkerboard patterns in layout optimization, Structural
Optimization, 10, 40–45
 Appendix 2
Notations for the constitutive tensor
Reference to the constitutive tensor will be given in the standard engineering notation.

For two dimensional orthotropic materials the constitutive tensor is written in matrix form as

ȡE 1111 E 1122 0 ȣ
( E ) 2–d
ort +ȧE 1122 E 2222 0 ȧ (A2.1)
Ȣ 0 0 E 1212Ȥ

and for isotropic materials assuming plane stress state as

ȡ 1n n 0ȣ
+ E 2ȧ ȧ
( E ) 2–d 1 0
iso, pl. stress (A2.2)
1–n 0 1–n
Ȣ0 2Ȥ

For three dimensional orthotropic materials, the constitutive tensor is written in matrix form as

ȡE 1111 E 1122 E 1133 0 0 0 ȣ

ȧE 1122 E 2222 E 2233 0 0 0
ȧ
ȧE 1133 ȧ
( E ) 3–d + ȧ 0 E 2233 E 3333 0 0 0
ȧ
ort ȧ 0 0 E 1212 0 0 ȧ (A2.3)
ȧ 0 ȧ
ȧ 0 0 0 E 1313 0 ȧ
Ȣ 0 0 0 0 0 E 2323Ȥ

Limiting the properties of the three dimensional materials further by assuming isotropy in each
plane of symmetry, and assuming that the moduli in the three principal directions are equal
(E1111= E2222= E3333), we get

ȱ 1 – n 223 n 12 ) n 13n 23 n13 ) n 12n 23

0 0 0 ȳ
D D D
ȧn n ȧ
ȧ ) n n 1 – n ) n 12n 13
ȧ
2
12 13 23 13 23

ȧ D D D
1 0 0
ȧ
ȧn 1 – n212 ȧ
ȧ 13 ) n 12n 23 n 23 ) n 12n 13
ȧ
ȧ ȧ
0 0 0
( E ) 3–d D D D
ort,iso + Eȧ 1 ȧ (A2.4)
ȧ 0 0 0
2(1 ) n 12)
0 0 ȧ
ȧ ȧ
ȧ 1 ȧ
ȧ 0 0 0 0
2(1 ) n13)
0
ȧ
ȧ 1 ȧ
Ȳ 0 0 0 0 0
2(1 ) n 23 )ȴ
 Appendix 2, Notations for the constitutive tensor A3

where

D + 1 * n 212 * n 223 * n 213 * 2 n 12 n 23 n 13

and n ij are the Poisson’s ratios in the ij–plane.

Assuming full isotropy (n 12 + n 23 + n 13 + n) we have

n n
ȡ1 1–n 1–n
0 0 0 ȣ
ȧn 1 n 0 ȧ
ȧ1–n 0 0 ȧ
ȧn n 1
1–n
ȧ
ȧ1–n 0 0 0 ȧ
ȧ 1–n ȧ
( E ) 3–d 1–n ȧ
iso + E (1 ) n)(1–2n)ȧ 0 0 0 1–2n 0 0 ȧ
ȧ (A2.5)
ȧ 2(1–n) ȧ
ȧ0 ȧ
ȧ 0 0 0 1–2n 0 ȧ
ȧ 2(1–n) ȧ
ȧ0 0 0 0 0 1–2n ȧ
Ȣ 2(1–n)Ȥ
 Appendix 3
The Element–by–Element Preconditioned Conjugate Gradient Method
This appendix is more or less a repetition of the description of the EBE–PCG method introduced
in section 1.4 but some details on the computational implementation are emphasized.
The Conjugate Gradient method solves the NxN linear system

[ S ]{D } + {R} (A1.1)

by minimizing the function

f ({ D }) + 1 {D } [ S ]{ D } – {R} T{D }
T
(A1.2)
2

which corresponds to the minimization of potential energy in the linear elastic elastic body. The
function f (potential energy) has a stationary point when the gradient

ʼnf + [S ]{ D } – {R} (A1.3)

is zero which correspond to (1.22).

The minimization is performed as an iterative procedure computing the search direction {p}k and
improved minimizer {D}k . At each step k in the iteration a factor a is determined such that
f ({ D } k ) a {p} k) is minimized; the new minimizer {D}k+ 1 is then set equal to the new point
{ D } k ) a {p} k.

Choosing the search direction {p}k as the direction of greatest descent, namely he negative of
the current residual vector {r} k + [S ]{ D } k–{R} k we have defined the Method of Steepest
Descent. The convergence of this method can be very slow and therefore the search direction
should be made [S]–conjugate, i.e.

{p} Ti[S]{p} j + 0, j t i (A1.4)

By making the search vector [S]–conjugate, the procedure is guarantied to converge to the exact
solution in at most N iterations.

The convergence of the conjugate gradient algorithm is dependent of the square of the condition
number (ratio between highest and lowest eigenvalue of the stiffness matrix [S]). Convergence
can therefore be speeded up by preconditioning the system. Many different choices of precondi-
tioners have been discussed in literature but Hollister and Riemer, (1993) report good results by
simply taking the diagonal of the stiffness matrix as the preconditioner [L]=diag([S]). The com-
 Appendix 3, The Element–by–Element Preconditioned Conjugate Gradient Method A5

putational procedure of the PCG–method is summarized in figure 1.1.

1) {r } 0 + {R }–[ S ]{ D } 0
2) {z } 0 + {r } T0 [L ] –1
3) {p }0 + {z } 0
T
4) g + {r } 0 {z } 0
5) k+0
6) {v } k + [S ]{ p } k
T
7) a + gń({p } k {v } k)
8) {D }k)1 + { D } k ) a{p } k
9) {r } k)1 + {r } k – a{v } k
10) if Ť{r } k)1Ť t d then STOP
T
11) {z } k)1 + {r } k)1[L ] –1
12) gȀ + g
T
13) g + {r } k)1{z } k)1
14) b + gńgȀ
15) {p }k)1 + {z } k)1 ) b { p } k)1
16) k+k)1
17) goto 6

Figure 1.1. Computational procedure for the

Element–By–Element Preconditioned Con-
jugate Gradient method.

The computationally most time consuming step of this procedure is step 6, the multiplication of
the global stiffness matrix [S] and the search direction vector {p}k . However, this multiplication
can be done on the element level as

{v} k + ȍ [s e]{pe} k
NE
(A1.5)
e+1

using the appropriate assembly procedure from local to global coordinates. As seen in equation
(A1.5), the global stiffness matrix never has to be assembled – this feature makes the requirement
to storage space very small in the EBE–PCG procedure.

In practice we do not require the procedure to converge exactly and we can therefore stop the
algorithm when the norm of the residual vector gets below a stopping criteria d. In fact d can be
chosen very gross when the EBE–PCG algorithm is used as a part of an iterative optimization
procedure. Only in the final iterations of the optimization procedure we require an exact solution
and then d can be made smaller for final convergence.
 Appendix 4
Continuum equivalence of bar elements and equivalent nodal loads

Modelling a truss member as a 2–node continuum element is done the following way. The shape
functions Ni for a bar element with a linear axial displacement field are given as

N 1 + L * t and N 2 + t A(3.1)
L L
where L is bar length and 0 v t v L. The bar is rotated the angle a with respect to the global
coordinate system. The shape functions Ni are therefore functions of a and the derivatives of Ni
with respect to horizontal and vertical coordinates x1 and x2 are

ēN 1 ēN 1 ēN 2 ēN 2
+*c , +*s , + c and +s A(3.2)
ēx 1 L ēx 2 L ēx 1 L ēx 2 L

where c and s is cos(a) and sin(a) respectively.

Following the matrix notation of Cook et al., (1989), the strain–displacement matrix [B] can be
written as

ȱ* c 0
[ B ] + 1ȧ 0 * s
c
0
0ȳ
sȧ A(3.3)
LȲ* s * c s cȴ

The constitutive matrix for a bar modelled as a 2–node continuum in a local coordinate system
aligned with the bar is

ȱE 0 0ȳ
[ EȀ ] +ȧ0 0 0ȧ A(3.4)
Ȳ0 0 0ȴ
which of course corresponds to a material with only stiffness in one direction and no shear stiff-
ness. Now the 2–node continuum element can be treated the usual way known from plane FE–
problems [e.g. Cook et al., (1989)] and we get the stress–strain relation

{s } + [ T ] T[EȀ ][T ]{ å } A(3.5)

where {s } and {å } are the 3x1 stress and strain vectors respectively and [T] is a transformation
matrix given as

ȱ c2 s2 cs ȳ
[ T ] +ȧ s 2 c2 * cs ȧ A(3.6)
Ȳ* 2cs 2cs c 2 * s 2ȴ
 Appendix 4, Continuum equivalence of bar elements and equivalent nodal loads A7

To calculate the equivalent nodal load vector {re } for elements subject to the prestrainfield NJå 0Nj
we have the following relation

{r e} +
4x1
ŕ [B] [T] [EȀ][T]NJå Nj dY
4x3
T T
0
3x1
e
A(3.7)
Ye

View publication stats