Download as pdf or txt
Download as pdf or txt
You are on page 1of 90

Form-finding and patterning of fabric structures using

shape optimization techniques

Thomas Linthout

Supervisors: Prof. dr. ir. Wim Van Paepegem, Dr. Ali Rezaei
Counsellor: Tien Dung Dinh

Master's dissertation submitted in order to obtain the academic degree of


Master of Science in Civil Engineering

Department of Materials Science and Engineering


Chair: Prof. dr. ir. Joris Degrieck
Faculty of Engineering and Architecture
Academic year 2015-2016
Form-finding and patterning of fabric structures using
shape optimization techniques

Thomas Linthout

Supervisors: Prof. dr. ir. Wim Van Paepegem, Dr. Ali Rezaei
Counsellor: Tien Dung Dinh

Master's dissertation submitted in order to obtain the academic degree of


Master of Science in Civil Engineering

Department of Materials Science and Engineering


Chair: Prof. dr. ir. Joris Degrieck
Faculty of Engineering and Architecture
Academic year 2015-2016
Acknowledgements

“Rome wasn’t built in a day”, and similarly, this dissertation is the result of a year of
hard work and research. However, the completion of this dissertation would not have been
possible without the advice and support of some people who I would like to thank.

I would like to thank professor Van Paepegem for his efforts as supervisor, and for intro-
ducing me to the wonderful world of optimization and tensile architecture.

I am grateful for the advice of my counselor Tien Dung Dinh. Whenever I had a question
about membrane materials, about optimization, or when a software bug slipped past my
eye, he was always ready to assist.

Special thanks to Ali Rezaei for helping me out numerous times.

I am thankful for my friends, who provided distraction whenever I couldn’t see the forest
for the trees.

And finally, I would like to thank my parents, for supporting me throughout my studies.

Thomas Linthout, June 2016


Permission

“The author gives permission to make this master dissertation available for consultation
and to copy parts of this master dissertation for personal use. In the case of any other use,
the limitations of the copyright law have to be respected, in particular with regard to the
obligation to state expressly the source when quoting results from this master dissertation”

Thomas Linthout, June 2016


Form-finding and patterning of
fabric structures using shape
optimization techniques
by
Thomas Linthout
Master’s dissertation submitted in order to obtain the academic degree of
Master of Science in Civil Engineering
Academic year 2015–2016
Supervisors: Prof. Dr. Ir. W. Van Paepegem, Dr. A. Rezaei
Counsellor: M. Sc. T.D. Dinh
Faculty of Engineering and Architecture
Ghent University
Department of Materials Science and Engineering

Abstract
Fabric materials used for tent structures are form-active, behave non-linearly and anisotrop-
ically, and can only transfer loads through tension. Due to this, designing and modelling
such structures is quite complicated and different from traditional structures. The so-
called conventional method is the most frequently used design method for designing tensile
structures. In this method, a form-finding is carried out, followed by the development of a
cutting pattern on the surface. Another design method is the integrated method, in which
flat panels are iteratively generated, assembled and subjected to tension, until a structure
with a permissible stress distribution is reached.
This thesis is a continuation on the method proposed by Dinh et al., based on the integrated
approach. In this method, the shape of the assembled panels is iteratively adjusted until
the final structure has a permissible stress distribution. It is assumed that the shape of
the assembled panels, also called the intermediate configuration, is stress-free. Expressing
the shape of the assembled panels by parameters, a SolidWorks model is created, which
is then imported in Abaqus. After calculation, the stresses in the membrane, along with
the forces in the boundary cables is compared to a specified stress and cable force, and the
parameters are adjusted. This process is repeated until the stress and force deviations are
minimal.

Keywords
Tensile architecture, shape optimization, fabric materials, Bayesian optimization, optimiza-
tion
Form-finding and patterning of fabric structures
using shape optimization techniques
Thomas Linthout

Supervisors: Prof. Dr. Ir. Wim Van Paepegem, Dr. Ali Rezaei, M. Sc. Tien Dung Dinh

Abstract Multiple methods exist to design fabric does not correctly represent the material. A material
structures. The most commonly used method is the so- model was proposed by Dinh et al.[1], taking into
called conventional method, which has some deficiencies. account these characteristics. However, time-
However, an accurate method, taking into accounts the dependent behavior is not included in this model. This
material characteristics of the membrane is non-existent.
model is further referred to as the fabric plasticity
In this thesis, a new method is proposed, taking into
account the correct loading path the material follows. model.
The proposed method assumes that the assembled panels
form a similar shape as the final structure, albeit with a III. DESIGN METHODS
smaller curvature. Using shape optimization, the shape
of this intermediate shape is determined in such a way A. Conventional method
that the stress distribution in the final structure is
permissible.
The most frequently used method in the industry is
the so-called conventional method. The method
Keywords Tensile architecture, shape optimization,
consists of a form-finding step, in which a fictitious
fabric materials, Bayesian optimization, optimization
membrane material is subject to a prestress, resulting
in an equilibrium shape. This shape is then subjected
I. INTRODUCTION
to a structural analysis, using the actual material
For over 50 years, coated woven fabrics have been characteristics, in order to assess the impact of
used in modern lightweight structures. However, clear environmental loads such as wind and snow loads.
guidelines for such architecture do not exist, and their Finally, a cutting pattern is generated on the
design is mainly reliant on experience. Fabric prestressed shape. This cutting pattern is then
materials used for tent structures are form-active, flattened, resulting in flat panels. In this step, the
behave non-linearly and anisotropically, and can only deformations caused by this flattening operation are
transfer loads through tension. neglected. The shape of the stress-free panels is then
The aim of this thesis is to propose a new method to determined by unloading these panels. Due to the non-
design tensile structures, taking into account the linear behavior of the membrane and the difference
characteristics of the material and the loading path it between the stress-strain curves for loading and
follows. To this extent, shape optimization methods unloading, this is not straightforward. As such, a
may provide an answer. linear orthotropic material model is often used in this
step, and the shape of the panels is adjusted to
II. MEMBRANE MATERIALS compensate for the inaccuracy of the model.
Textile fabrics used for tensile structures are a Due to this compensation, and because deformations
combination of a system of orthogonally woven yarns occurring during flattening are ignored, this method is
and a coating applied to withstand weathering and less than optimal.
ageing. The coating is applied continuously, by
applying the liquid coating while the fabric is moving B. Integrated approach
under a knife. During this process, the fabric moved While the conventional method starts with a
through drums by pulling the warp fibres. Due to this, prestressed shape, the integrated approach includes the
the warp fibres are more stretched than the fill fibres, influence of the cutting pattern in the form-finding and
which will have an impact on the behavior of the analysis. Cutting patterns are iteratively adjusted such
material. that, after applying the deformations resulting in the
As a result of the production process and the final shape, the structure has a permissible stress-
composite action, the fabric has a quite complex distribution, minimizing the difference between the
material behaviour. The material behaves design prestress and the actual stress in the membrane.
anisotropically, inelastically and nonlinearly. In a study conducted by Dinh et al.[2], a new design
Additionally, the material behavior is dependent on method was proposed. This method is a variation of
the load ratio between warp and fill threads. As a the integrated approach. In the method, the shape of
result, a simple orthotropic elasticity material model the assembled panels is assumed to be flat, and three
parameters describing the intermediate shape are The result of this optimization is a certain
adjusted until the stress distribution in the final intermediate shape. As this shape consists of fabric
construction is permissible. The method proposed in panels assembled together, this shape should be stress-
this study is a continuation on this method. free. This assumption can be validated by cutting the
three-dimensional intermediate shape into panels,
IV. PROPOSED METHOD computationally flatten these panels using Rhino3D’s
‘Squish’ algorithm, and check the deformations. If
A. Proposed design method these deformations are small enough to neglect, the
In this case study, the proposed method is tested on method is deemed valid, and the panels can correctly
a 4 m x 4 m hypar, with a height of 80 cm once assemble the hyperstructure.
assembled and tensioned. First, the shape of the
adjoined panels is considered. It is assumed that this B. Objective function
intermediate stress-free shape will have a similar The objective function is used to convert the results
shape of that of the final hypar, albeit with a smaller from the FEM analysis to one value, which is to be
height. To describe this intermediate shape, a set of minimized. In order to do this, the average squared
five state variables P is used, as seen in Figure 1. deviation from a predefined membrane stress and
cable force is calculated. This results in two values, of
which a weighted average is taken resulting in one
value, which is the value of the objective function, as
seen in the formulae below.

𝑊 𝑃 = 𝑤! ⋅ 𝐷! + 𝑤! ⋅ 𝐷!

1 !"!#$%&"
𝐷! = 𝜎 !"!#$%&" − 𝜎!"#
𝑛!"!#$%&"
!!"!#$%&"

Figure 1 Five state variables describing the intermediate 1 !"#$%


shape of the membrane 𝐷! = 𝐹 !"#$% − 𝐹!"#
𝑛!"#$%
!!"#$%
In Abaqus, the membrane material characteristics
are assigned to the shape, and a cable element is
modeled in sleeves at the edges of the membrane. The
cable has an effective stiffness of 2000 kN. In a first C. Optimization algorithm
step, an orthotropic material model is used for the Many capable algorithms exist to perform the
membrane, and next the fabric plasticity model will be optimization of the objective value, however it is not
used. The characteristics of the orthotropic material clear which algorithm is the most suited for this
model can be found in table 1. Next, the purpose. To know the most appropriate algorithm,
displacements are described, resulting in a stress three algorithms are compared to each other: a genetic
distribution in the membrane and a force in the cable algorithm, a Bayesian optimization algorithm adopted
element. from a study conducted by Snoek et al.[3] and a
Pattern Search algorithm. All the algorithms are
Table 1: Parameters of the orthotropic elasticity model
applied to the objective function described above, and
Ew Ef 𝜈!" 𝜈!" G the algorithms are compared based on their number of
975.6 MPa 716.1 MPa 0.23 0.17 87.7 MPa function evaluations and the resulting value of the
objective function. This comparison is done for five
Using an Abaqus Python script, the average values of the weight factors.
absolute deviation from a predefined membrane stress In this comparison, the pattern search algorithm
and cable force is calculated. This results in two returns the fastest results. However, the Bayesian
values, of which a weighted average is taken resulting optimization always returns a better value, albeit with
in one value. The goal of this design method is to more function evaluations. The genetic algorithm uses
minimize this value, and as such bringing the stress in the most function evaluations, and never returns a
the membrane as close to the design stress as possible. better value than the Bayesian optimization. As a
The target stress in the membrane is 5 MPa, while the result, the Bayesian optimization is deemed the most
target force in the cable is 20 kN. suitable for this project.
To minimize this value, an optimization algorithm is
used. The goal of optimization is to find the variables D. Verification of stress-free hypothesis
at which the objective function reaches a minimum. In Finally, it has to be verified that the resulting
this case, the objective function is the result from the intermediate shape can be assembled from flat panels
post processing, which is a function of the state without inducing stresses that are too high. In
variables. Rhino3D, the intermediate shape is divided into six
panels, along geodesic lines. Using the ‘Squish’
feature, these panels are then flattened, while
returning maximum and average deformations caused
by the flattening. A clear overview of the process can
be seen in figure 2.

Figure 3 Warp stresses in the optimal configuration, using


the orthotropic elasticity model (w1=0.7, w2=0.3)

Figure 2 Flowchart of the proposed design method

E. Benchmark
As the method proposed in this thesis is a
continuation of the method proposed by Dinh et al.[2],
results from this study can be compared to results Figure 4 Fill stresses in the optimal configuration, using the
from the earlier work. As the intermediate shape is orthotropic elasticity model (w1=0.7, w2=0.3)
controlled by five parameters instead of three, better Table 2: Parameters of the optimal configuration, using the
stress distributions and cable forces are expected. orthotropic elasticity model (w1=0.7, w2=0.3)

V. RESULTS AND CONCLUSION p1 (m) p2 (m) p3 (m) p4 (m) p5 (m) Fcable (kN)
-0.011 0.367 0.310 0.000 0.242 19.998
Varying the weight factors, intermediate shapes
causing a lower stress and force deviation were indeed
found, both using the orthotropic elasticity model and
the fabric plasticity model. After dividing the found
intermediate configurations into six panels of equal
width and flattening these panels, deformations up to
0.15% occur. These deformations correspond with
negligible stresses, so the intermediate configuration
can indeed be assembled without inducing stresses.
Results of the shape optimization can be seen in
figures 3-4 and table 2 for the orthotropic elasticity
material model, and in figures 5-6 and table 3 for the
fabric plasticity material model.

Figure 5 Warp stresses in the optimal configuration, using


the fabric plasticity model (w1=0.7, w2=0.3)
ACKNOWLEDGEMENTS
The author would like to thank the suggestions of
his supervisors and counsellors.

REFERENCES
[1] T.D. Dinh, A. Rezaei, L. De Laet, M. Mollaert, D. Van
Hemelrijck, and W. Van Paepegem. A new elasto-plastic
material model for coated fabric. Engineering Structures,
71:222–233, 2014.
[2] T.D. Dinh, A. Rezaei, L. De Laet, M. Mollaert, D. Van
Helmelrijck, and W. Van Paepegem. A shape optimization
approach to integrated design and nonlinear analysis of cable
reinforced fabric membrane structures.
[3] J. Snoek, H. Larochelle, and R.P. Adams. Practical Bayesian
optimization of machine learning algorithms.

Figure 6 Fill stresses in the optimal configuration, using the


fabric plasticity model (w1=0.7, w2=0.3)
Table 3: Parameters of the optimal configuration, using the
fabric plasticity model (w1=0.7, w2=0.3)

p1 (m) p2 (m) p3 (m) p4 (m) p5 (m) Fcable (kN)


-0.033 0.314 0.350 0.000 0.349 19.993

It can be concluded that the method can indeed


provide form-finding and patterning of fabric
structures, taking into account their material
characteristics. The difference between the used
material model manifests itself in the resulting shape
of the panels. As seen in figure 7, the resulting panels
when using the orthotropic elasticity model (blue) are
bigger than those resulting when using the fabric
plasticity model (red).

Figure 7 Panels from the orthotropic elasticity result (blue)


compared to panels from the fabric plasticity result (red)
However, some further research could be done on
this method. First, only a handful of optimization
algorithms were tested, so it is possible that a better
algorithm exists for this purpose. Only parametric
optimization was considered, and non-parametric
optimization might provide even better results. Also,
due to SolidWorks API restrictions, parallelization
was not possible, which could have provided faster
solutions. Finally, the method was only tested on a
single type of structure, a hyperstructure. It remains to
be confirmed whether the method can also be applied
to other shapes, such as conical membranes or more
complex structures.
CONTENTS xi

Contents

Acknowledgements iv

Permission v

Abstract vi

Extended abstract vii

Contents xi

List of Figures xiv

List of Tables xvii

Used abbreviations xviii

Used symbols xix

1 Introduction 1

2 Fabric architecture 3
2.1 History of fabric architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Analysis of fabric tension structures . . . . . . . . . . . . . . . . . . . . . . 6
2.2.1 Conventional method . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.2 Integrated design method . . . . . . . . . . . . . . . . . . . . . . . 12

3 Membrane materials 14
3.1 Textile fabrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Coating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 Mechanical Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.1 Tensile strength . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3.2 Anisotropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3.3 Nonlinear behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . 18
CONTENTS xii

3.3.4 Inelastic behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . 18


3.3.5 Shear stiffness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.3.6 Kink resistance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3.7 Material model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.4 Structural behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4 Optimization 23
4.1 Optimization problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1.1 Linear vs. Nonlinear . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1.2 Constrained vs. Unconstrained . . . . . . . . . . . . . . . . . . . . 24
4.1.3 Discrete vs. Continuous . . . . . . . . . . . . . . . . . . . . . . . . 25
4.1.4 Single vs. Multiobjective . . . . . . . . . . . . . . . . . . . . . . . . 25
4.1.5 Single vs. Multiple Minima . . . . . . . . . . . . . . . . . . . . . . 26
4.1.6 Deterministic vs. Nondeterministic . . . . . . . . . . . . . . . . . . 27
4.1.7 Simple vs. Complex Problem . . . . . . . . . . . . . . . . . . . . . 27
4.2 Solution approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.1 Experimental Optimization . . . . . . . . . . . . . . . . . . . . . . 27
4.2.2 Analytical Optimization . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.3 Numerical Optimization . . . . . . . . . . . . . . . . . . . . . . . . 28
4.3 Summary of optimization algorithms . . . . . . . . . . . . . . . . . . . . . 28
4.3.1 Gradient-based methods . . . . . . . . . . . . . . . . . . . . . . . . 28
4.3.2 Genetic algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3.3 Pattern Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.3.4 Simulated annealing . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3.5 Bayesian optimization with Gaussian process priors . . . . . . . . . 31
4.4 Comparison of optimization algorithms . . . . . . . . . . . . . . . . . . . . 32

5 Shape optimization 34
5.1 Parametric shape optimization . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.1.1 Design model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.1.2 Analysis model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.1.3 Optimization algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.2 Non-parametric shape optimization . . . . . . . . . . . . . . . . . . . . . . 36

6 Shape optimization project 39


6.1 Proposed design method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.2 Design model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.3 Analysis model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.3.1 Creating the model . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.3.2 Computation and post-processing . . . . . . . . . . . . . . . . . . . 43
CONTENTS xiii

6.4 Optimization algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44


6.5 Patterning and verifying the stress-free hypothesis . . . . . . . . . . . . . . 45
6.6 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.7 Benchmark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.7.1 Orthotropic elasticity material model . . . . . . . . . . . . . . . . . 47
6.7.2 Fabric plasticity material model . . . . . . . . . . . . . . . . . . . . 47

7 Results 49
7.1 Comparing optimization algorithms using the orthotropic elasticity model . 49
7.2 Shape optimization using the orthotropic elasticity material model . . . . . 52
7.2.1 w1 = 0.7 and w2 = 0.3 . . . . . . . . . . . . . . . . . . . . . . . . . 52
7.2.2 w1 = 0.9 and w2 = 0.1 . . . . . . . . . . . . . . . . . . . . . . . . . 54
7.2.3 w1 = 0.95 and w2 = 0.05 . . . . . . . . . . . . . . . . . . . . . . . . 55
7.2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
7.3 Shape optimization using the fabric plasticity material model . . . . . . . . 58
7.3.1 w1 = 0.7 and w2 = 0.3 . . . . . . . . . . . . . . . . . . . . . . . . . 58
7.3.2 w1 = 0.9 and w2 = 0.1 . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.3.3 w1 = 0.95 and w2 = 0.05 . . . . . . . . . . . . . . . . . . . . . . . . 62
7.3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

8 Conclusion and recommendations for future work 65


8.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
8.2 Recommendations for future work . . . . . . . . . . . . . . . . . . . . . . . 65

Bibliography 67
LIST OF FIGURES xiv

List of Figures

2.1 A reconstruction of a 10,000-year-old tent from remains found at Pincevent,


Northern France, adapted from [2]. . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Artist’s render of the Colosseum’s velarium, while deploying, adapted from [3]. 5
2.3 The Olympia Stadium in Munich, Germany, designed by Frei Otto, adapted
from [4]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.4 A membrane structure from the ’10 Expo in Shanghai, China, adapted from [5]. 6
2.5 The air-supported roof of the Tokyo Dome, adapted from [6]. . . . . . . . . 7
2.6 The different steps in creating a membrane structure, adapted from [7]. . 8
2.7 Tangential surface stress field, adapted from [9]. . . . . . . . . . . . . . . 9
2.8 Example of cloth unfolding, adapted from [10]. . . . . . . . . . . . . . . . . 11
2.9 An illustration of the patterning in the conventional method, adapted from
[7]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3.1 Diagram of different weaving methods, adapted from [16]. . . . . . . . . . . 15


3.2 Typical stress-strain curves from a uni-axial test in different orientations,
adapted from [17]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3 Stress strain curves for uniaxial test in warp and fill directions in case of
cyclic load, adapted from [15]. . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.4 A proposed elasto-plastic material model for coated fabrics, adapted from [19]. 20

4.1 Traditional vs. Optimal Design Process, adapted from [20]. . . . . . . . . . 23


4.2 Multiobjective optimization, adapted from [20]. . . . . . . . . . . . . . . . 25
4.3 Pareto frontier, adapted from [20]. . . . . . . . . . . . . . . . . . . . . . . . 26
4.4 Procedure for a Genetic Algorithm, adapted from [20]. . . . . . . . . . . . 30
4.5 An example of a Bayesian optimization on a 1D optimization problem,
adapter from [25]. The figure also contains the acquisition function, which
is highest in the points where the Gaussian process predicts a high objective
along with a high uncertainty. . . . . . . . . . . . . . . . . . . . . . . . . 32
LIST OF FIGURES xv

5.1 General shape optimization process, adapted from [28]. . . . . . . . . . . . 35


5.2 Example of cloth unfolding, adapted from [10]. . . . . . . . . . . . . . . . . 37

6.1 The different steps in creating a membrane structure, adapted from [7]. . 40
6.2 Final configuration of the hypar, adapted from [19]. . . . . . . . . . . . . . 40
6.3 Intermediate configuration of the hypar, including its state variables . . . . 41
6.4 Flowchart depicting the proposed design method . . . . . . . . . . . . . . . 46
6.5 Membrane stresses in the optimized shape found by Dinh. et al. [15] using
the orthotropic elasticity model . . . . . . . . . . . . . . . . . . . . . . . . 47
6.6 Membrane stresses in the optimized shape found by Dinh. et al. [15] using
the fabric plasticity model . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

7.1 Membrane stresses in the optimized shape . . . . . . . . . . . . . . . . . . 51


7.2 Membrane stresses in the optimized shape using the orthotropic elasticity
model (w1 = 0.7 and w2 = 0.3) . . . . . . . . . . . . . . . . . . . . . . . . . 52
7.3 Optimization process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.4 Optimized shape divided in panels . . . . . . . . . . . . . . . . . . . . . . 53
7.5 Membrane stresses in the optimized shape using the orthotropic elasticity
model (w1 = 0.9 and w2 = 0.1) . . . . . . . . . . . . . . . . . . . . . . . . . 54
7.6 Optimization process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
7.7 Optimized shape divided in panels . . . . . . . . . . . . . . . . . . . . . . 55
7.8 Membrane stresses in the optimized shape using the orthotropic elasticity
model (w1 = 0.95 and w2 = 0.05) . . . . . . . . . . . . . . . . . . . . . . . 56
7.9 Optimization process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
7.10 Optimized shape divided in panels . . . . . . . . . . . . . . . . . . . . . . 57
7.11 Membrane stresses in the optimized shape using the fabric plasticity model
(w1 = 0.7 and w2 = 0.3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
7.12 Optimization process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.13 Optimized shape divided in panels . . . . . . . . . . . . . . . . . . . . . . 59
7.14 Membrane stresses in the optimized shape using the fabric plasticity model
(w1 = 0.9 and w2 = 0.1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.15 Optimization process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.16 Optimized shape divided in panels . . . . . . . . . . . . . . . . . . . . . . 61
7.17 Membrane stresses in the optimized shape using the fabric plasticity model
(w1 = 0.95 and w2 = 0.05) . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.18 Optimization process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
LIST OF FIGURES xvi

7.19 Optimized shape divided in panels . . . . . . . . . . . . . . . . . . . . . . 63


7.20 Difference between the panels when using the orthotropic elasticity model
(blue) or the fabric plasticity model (red), when using weight factors w1 =
0.7 and w2 = 0.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
LIST OF TABLES xvii

List of Tables

6.1 Boundaries of the state variables . . . . . . . . . . . . . . . . . . . . . . . 42


6.2 Parameters of the orthotropic elasticity material model . . . . . . . . . . . 43
6.3 Genetic algorithm parameters . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.4 Parameters and results (w1 = 0.7 and w2 = 0.3) . . . . . . . . . . . . . . . 47
6.5 Parameters and results (w1 = 0.7 and w2 = 0.3) . . . . . . . . . . . . . . . 48

7.1 Different weight factor combinations . . . . . . . . . . . . . . . . . . . . . 49


7.2 Algorithm comparison for the case w1 = 0.1 and w2 = 0.9 . . . . . . . . . . 50
7.3 Algorithm comparison for the case w1 = 0.3 and w2 = 0.7 . . . . . . . . . . 50
7.4 Algorithm comparison for the case w1 = 0.5 and w2 = 0.5 . . . . . . . . . . 50
7.5 Algorithm comparison for the case w1 = 0.7 and w2 = 0.3 . . . . . . . . . . 50
7.6 Algorithm comparison for the case w1 = 0.9 and w2 = 0.1 . . . . . . . . . . 50
7.7 Results for each weight factor combination . . . . . . . . . . . . . . . . . . 51
7.8 Parameters and results (w1 = 0.7 and w2 = 0.3) . . . . . . . . . . . . . . . 52
7.9 Equivalent strains caused by assembly, panels numbered from left to right . 53
7.10 Parameters and results (w1 = 0.9 and w2 = 0.1) . . . . . . . . . . . . . . . 54
7.11 Equivalent strains caused by assembly, panels numbered from left to right . 54
7.12 Parameters and results (w1 = 0.95 and w2 = 0.05) . . . . . . . . . . . . . . 55
7.13 Equivalent strains caused by assembly, panels numbered from left to right . 56
7.14 Parameters and results (w1 = 0.7 and w2 = 0.3) . . . . . . . . . . . . . . . 59
7.15 Equivalent strains caused by assembly, panels numbered from left to right . 59
7.16 Parameters and results (w1 = 0.9 and w2 = 0.1) . . . . . . . . . . . . . . . 60
7.17 Equivalent strains caused by assembly, panels numbered from left to right . 61
7.18 Parameters and results (w1 = 0.95 and w2 = 0.05) . . . . . . . . . . . . . . 62
7.19 Equivalent strains caused by assembly, panels numbered from left to right . 63
xviii

Used abbreviations

2D Two-dimensional
3D Three-dimensional
ALLIE Total strain energy
ALLSD Viscous damping energy
CAD Computer-aided design
FEM Finite element method
GA Genetic algorithm
MSAJ Membrane Structure Association of Japan
PTFE Polytetrafluoroethylene
PVC Polyvinyl chloride
SLS Serviceability limit state
UV Ultraviolet
VBA Visual Basic for Applications
xix

Used symbols

 Strain
ν Poisson’s ratio
F Force
σ Stress
σ11 or S11 Stress in the warp direction
σ22 or S22 Stress in the fill direction
c Damping factor
nmembrane Number of finite element nodes in the membrane
ncable Number of finite element nodes in the cable
D1 Averaged absolute stress deviation in the membrane
D2 Averaged absolute force deviation in the cable
w1 Weight factor regarding to D1
w2 Weight factor regarding to D2
1

Chapter 1

Introduction

For over 50 years, coated woven fabrics have been used in modern lightweight structures.
The membrane materials have a strong architectural value, can be installed quickly and
are environmentally friendly. In comparison with traditional construction materials such
as concrete and steel, architectural fabrics are lighter and cheaper, making them an ideal
choice when a large area needs to be covered.

Fabric materials used for tent structures are form-active, behave non-linearly and anisotrop-
ically, and lack bending or compressive stiffness. Because of this, they can only transfer
loads through tension, while compression stresses will lead to wrinkling, which can eventu-
ally cause the failure of the structure. These material characteristics make the design and
modelling of tensile structure difficult.

A fabric tensile structure is constructed in three steps: first, panels are cut out from the
rolls which the fabric is supplied from. These panels are then welded together, and the
assembled membrane is tensioned. Due to the complex nature of such structures, along
with the material characteristics of the membrane, designing tensile structures is more
difficult than designing traditional structures.

Usually, fabric tensile structures are designed using the conventional method. This method
consists of finding a shape of the tensioned structure, followed by developing a pattern on
the surface. Next, the flat shapes of the 2D panels are to be determined. Due to the
non-linear and inelastic behaviour of the membrane material, this is hard to do, and in
practice, an orthotropic material law is used during the flattening, and the resulting panels
are ‘compensated’ to account for this simplification. The flaws of this method are apparent,
as the correct material characteristics are not used throughout the process, and because
2

the compensation is poorly documented.

A second method is the integrated method. In this method, flat panels are iteratively
generated, assembled, and subjected to tensioning. This is repeated until a permissible
stress distribution in the membrane is obtained. Throughout the design, the correct
material characteristics are used.

The aim of this master’s dissertation is to propose a new design method for tensile architec-
ture, continuing on the research by Dinh et al. and on the integrated approach in general.
The purpose of this method is to find a cutting pattern of a fabric in such a way that the
difference between the final stress distribution in the membrane is negligible compared to
the design. A combination of finite element analysis and shape optimization is used to
iteratively find the optimal shape of the structure, performing a nonlinear analysis in each
iteration.
3

Chapter 2

Fabric architecture

When thinking about fabric structures, tents directly come to mind, being a small and
portable solution. However, due to modern construction technology and materials, con-
temporary tents can be enormous and can provide shelter for huge crowds. Two types of
tensile architecture are commonly used: air-supported fabric structures, and fabric tension
structures. These structures are incredibly light when compared to structures made of
traditional construction materials, such as steel and concrete. Fabric tension structures
have a low maintenance cost, no need for air-tightness, and a large variety of double-curved
surfaces [1]. In this master’s dissertation, the focus lies on fabric tension structures.

2.1 History of fabric architecture


The earliest structures made by mankind were, like tents, both functional and transient.
These humans were hunters, and had the ability to use animal by-products to make shelters,
as well as weapons, tools and clothing. In Northern France, archaeologists have discovered
4.5 m wide tents, with wooden poles covered by animal skin, as seen in figure 2.1. Some of
these prehistoric structures are still recognizable today, such as the Lapp Keti, a conical
structure clad in hides, used by Siberian nomads and hunters. The skins used have a
low tensile strength, and are relatively heavy, so when available, they were replaced by
fabrics [2].

Small tents were also used by more developed nations in commercial, civic and military
settings. The Roman military used ’papilio’, or butterfly tents that were constructed
from a calf leather membrane. These tents were 2 m high at their central ridge, with
0.3 m walls that could be lifted for ventilation. Sometimes, tents had greater importance.
2.1. HISTORY OF FABRIC ARCHITECTURE 4

Figure 2.1: A reconstruction of a 10,000-year-old tent from remains found at Pincevent, North-
ern France, adapted from [2].

For example, Ottoman rulers considered their tents as main residences. Their tents were
arranged in complex structures including many different buildings and were larger than
Roman tents [2].

To increase in size, membrane materials had to be tested in another area of design, such
as ships. From 2000 BC, animal skins were used as sails by the Minoans to propel their
ships using wind force. The curved sail can be compared by a pneumatic structure, which
got optimized over the ages. This technology reached a peak in the 1869 tea clipper Cutty
Sark, which could sail at 17 knots with its 3000 m2 of sail [2].

This technology was transferred back to buildings throughout history. Roman architec-
ture used featured temporary and retractable shading called vela (Latin for sail). The
Colosseum in Rome featured a tensile canvas roof of 189 m by 156 m, which was probably
suspended from 7.5 m high masts around the building, as seen in figure 2.1. At a political
meeting in 1520 called ‘Camp du Drap d’Or’ in France, a circular, 16-sided structure with
a 40 m high central mast was used as main building [2].

At the beginning of the twentieth century, temporary fabric structures were used for other
purposes, such as performance venues. This was pioneered by the company Stromeyer,
which worked in close collaboration with the great innovator in fabric architecture design,
2.1. HISTORY OF FABRIC ARCHITECTURE 5

Figure 2.2: Artist’s render of the Colosseum’s velarium, while deploying, adapted from [3].

Frei Otto. From the 1950’s, Otto designed membrane structures with the combination
of engineering, research, and sensitivity to the beauty and sustainability attributes of
lightweight structures. Frei Otto invented the weighted cable and soap bubble method to
produce natural physical models resulting in an optimal form [2].

Fabric architecture at this time was suited to big events and was therefore temporary. A
precedent for permanent fabric structures was set by Otto’s German Pavilion on the Expo
’67 in Montreal, Canada. This however was a building with a steel cable-net mesh, from
which polyester fabric was hung. Nowadays, vinyl-coated polyesters (PVCs) are the main
material for fabric architecture, due to their low cost and 15-20-year lifespan [2].

Figure 2.3: The Olympia Stadium in Munich, Germany, designed by Frei Otto, adapted from [4].
2.2. ANALYSIS OF FABRIC TENSION STRUCTURES 6

Figure 2.4: A membrane structure from the ’10 Expo in Shanghai, China, adapted from [5].

While the majority of fabric structures are fabric tension structures, air-supported struc-
tures are another trend in fabric architecture. These fabrics achieve their tensioning by a
difference in air pressure at both sides of the membrane [2]. The average air pressure at
kN
sea level is 1.013 25 bar, or about 100 m 2 . As a comparison, this is fifty times the usual live

load on a floor. Because of this value, small variations in air pressure can easily lift and
tension a lightweight membrane [6].

Air-supported structures are commonly shaped as a dome, or as a half cylinder. By realizing


an air pressure difference of about 0.001 MPa, or 10 mbar compared to the outside air
pressure, the roof of the structure is held in place. As this is only a 1 % difference compared
to the average air pressure, it is hardly noticeable by people inside of the building. However,
such a structure is never completely air-tight, due to entrances, exits, and gaps in the
membrane. Because of this, fans are used to guarantee a constant airflow. An example of
this type of structure is the Tokyo Dome, seen in figure 2.1.

2.2 Analysis of fabric tension structures


Modern fabric tension structures are constructed from numerous flat panels that are welded
together and tensioned. The panels are cut from rolls of fabric, and the shape of the
assembled stress-free panels is called the intermediate configuration, as seen in figure 2.6.
By tensioning this intermediate configuration, the final configuration is reached. Due to the
complex assembly of membrane structures, along with the complex behaviour of membrane
materials, designing these structures is not straightforward.
2.2. ANALYSIS OF FABRIC TENSION STRUCTURES 7

Figure 2.5: The air-supported roof of the Tokyo Dome, adapted from [6].

The most frequently used method to design fabric tension structures is the so-called con-
ventional method. An important property of the conventional method is that the form-
finding and the patterning are decoupled. By contrast, the patterning is incorporated in
the form-finding in the so-called integrated approach. Both methods will be discussed in
what follows.

2.2.1 Conventional method


The conventional method is generally used when designing and patterning fabric struc-
tures. This method consists of four steps: the form-finding analysis, the load analysis,
the patterning analysis and the fabrication analysis [1]. A commonly used software pro-
gram following the conventional method is EASY, which incorporates every step from the
form-finding up to the patterning analysis.

Form-finding analysis

Traditional structures, often made of steel and concrete, are designed to guarantee ser-
viceability. In limit state design, a structure that satisfies the Serviceability Limit State
(SLS) does not exceed a defined limit for deflections or local deformations [8]. As such,
the shape of the structure under loads, assuming it has not yet reached failure, is not
that different from the initial shape. By contrast, membranes tend to deform more than
traditional structures, as they lack bending stiffness. As such, the final shape is different
from the initial configuration, and should be determined beforehand.
2.2. ANALYSIS OF FABRIC TENSION STRUCTURES 8

Figure 2.6: The different steps in creating a membrane structure, adapted from [7].

When no loads are acting on the structure, except for the prestress, a force equilibrium
can be expressed. As seen in figure 2.7, stresses σx and σy act on the membrane. The force
equilibrium in the z-direction is given by:

σx σy
+ =0 (2.1)
Rx Ry

As the stresses σx and σy are both positive, the curvatures R1x and R1y should have an
opposite sign, resulting in a negative Gaussian curvature K = R1x R1y , which is a typical
property of a so-called anticlastic surface. If the curvatures R1x and R1y are equal but
opposite in every point of the surface, the membrane is subjected to a uniform stress in
both directions. Additionally, the membrane describes a so-called minimum surface.

Minimum surfaces are frequently used in the design of tension structures. The advantages
of using minimum surfaces are their aesthetically pleasing shapes, and the uniform stress
in the membrane. However, minimum surfaces tend to be quite flat due to the mean
curvature being zero. Because the load-bearing capacities of the membrane are dependent
on its curvature, this is not desirable [10].

To create such a surface, a constant prestress is maintained, independent of changes in


strain, as seen in figure 2.7. This corresponds in a Young’s modulus of zero. However,
in calculations, a very small Young’s modulus is used, to avoid numerical singularities.
This process is an iterative process, and mostly Newton-Raphson iterations are used to
2.2. ANALYSIS OF FABRIC TENSION STRUCTURES 9

Figure 2.7: Tangential surface stress field, adapted from [9].

determine an equilibrium shape in which the specified prestresses are achieved [10]. Ex-
perimentally, minimal surfaces can be found using soap bubbles (as discussed in 2.1).

One of the most frequently used shapes in fabric tension architecture is the hyperbolic
paraboloid, also known as a hypar or a saddle shape. This shape can be constructed by
lifting two opposite corners of a square membrane, while lowering the other two corners.
Another standard shape is a conical shape, which can be realized by lifting an internal
point of a membrane while keeping the edges in place.

However, form-finding ignores any external effects, such as the dead load of the membrane,
wind loads, crimp, creep and temperature effects, and material anisotropy. To counteract
deformation resulting from external loads, the prestress in the membrane has to be high
enough to guarantee a sufficient out-of-plane stiffness. This is where the load analysis
comes into play.

Load analysis

After the equilibrium state of the membrane is found, a structural analysis can be con-
ducted to study the behaviour under environmental loads. In this analysis, the prestress
is taken into account, along with the external loads such as loads from wind and snow.
2.2. ANALYSIS OF FABRIC TENSION STRUCTURES 10

Due to the form-active behaviour of fabric tension structures, changes in the geometry
can occur that cannot be neglected. Therefore, a nonlinear analysis should be conducted.
Additionally, the membrane will buckle or wrinkle under compressive stresses, so the correct
material characteristics should be included. Again, this step can be performed in EASY,
but also in general FEM software such as Abaqus or ADINA [1].

Cutting pattern analysis

Finally, a cutting pattern has to be generated on the prestressed membrane. Because archi-
tectural fabrics are made in rolls of 2 m-3 m wide, the 3D structure has to be constructed
out of 2D panels of this width or smaller. Cutting patterns are usually created along
geodesic lines, both for aesthetic reasons and to minimize fabric usage [11]. Tension fabric
structures are highly varied in size, material characteristics and curvature, and the cutting
pattern depends on all these factors. In plane cloth geometry determination, there is al-
ways some approximation, as double-curved surfaces are by definition undevelopable [7].
However, in order to produce reliable data, this approximation should be minimized. In
other words, the flat panels resulting from the flattening should resemble the curved strips
as good as possible [10].

The patterning process consists of two steps. First, a global layout of individual cloths
has to be developed. The mesh used for the shape-finding can be used for this purpose.
Afterwards, 3D-data for each cloth has to be converted in usable plane form. If a fabric
cloth consists of only one layout of elements, the cloth may be developed to a distortion-
free plane form. This method is called cloth unfolding and is the simplest one. It is a
geometric technique and is therefore reliable. An example is shown in figure 2.8 [10].

However cloth unfolding is a major approximation, as every cloth is only one element wide.
In order to accurately simulate double-curved membranes, the cloths have to be flattened.
To minimize stresses caused by the flattening, the change in link length is minimized, using
a least-squares approach. In this approach, the sum of squared differences between the link
lengths in the flat situation and the link lengths in the curved situation is minimized, as
seen in equation 2.2-2.4 [10].
m
X
S(x) = φ2i (x) (2.2)
i=1
q
φi = (xi,1 − xi,2 )2 + (yi,1 − yi,2 )2 − di (2.3)
2.2. ANALYSIS OF FABRIC TENSION STRUCTURES 11

(a) Double curved membrane structure (b) Cloth unfolding of (a)

Figure 2.8: Example of cloth unfolding, adapted from [10].

q
di = (Xi,1 − Xi,2 )2 + (Yi,1 − Yi,2 )2 + (Zi,1 − Zi,2 )2 (2.4)

To ensure that the cloth strips remain compatible after flattening, the seam boundaries
have to remain unchanged. To achieve this, weights ωi are are assigned to links lying on the
edge of a cloth strip. Instead of minimizing equation 2.2, equation 2.5 is minimized. [10]
m
X
S(x) = ωi φ2i (x) (2.5)
i=1

The choice in location of the panel’s edges, or seams, requires artistic judgement, so that the
seaming does not visually corrupt the structure. It should also be economic, to minimize
waste of material, and it should guarantee safe transfer of stresses between panels. For
these reasons, it is optimal that the seams follow geodesic lines on the surface. [7]

The traditional patterning method described above is a geometrical method, and does not
take into account the prestress of the membrane, nor its material characteristics. To take
this into account, the panels are ‘compensated’ afterwards. This is an ad hoc method,
so there is ample room for improvement. Figure 2.9 describes the traditional patterning
process as a whole [7].

After the creation of the pattern, the shape of the stress-free panels are determined by
unloading the prestress. However, these materials have a strongly non-linear behaviour,
and the stress-strain curves for loading and unloading are different, resulting in multiple
strain levels for a certain stress level. As a result, a correct unloading is impossible.
2.2. ANALYSIS OF FABRIC TENSION STRUCTURES 12

Figure 2.9: An illustration of the patterning in the conventional method, adapted from [7].

In practice, a linear orthotropic elasticity model is used, and the cutting patterns are
compensated afterwards.

The conventional method obviously has its disadvantages. The compensation of the panels
is badly documented and the stresses caused by deforming the flat panels are not taken
into account, although they may be significant. As a consequence, wrinkles can occur in
the structure, and the final shape of the membrane may differ from the shape found by
the form-finding [7].

2.2.2 Integrated design method


While the conventional method can be considered a shape optimization problem, in which
the patterning is done independently of the form-finding, the integrated design method,
proposed by Abel et al. [12], tries to optimize the shape of the flat panels, in such a way
that the shape of the structure and the stresses in the membrane are permissible.

The approach begins with the shape found by the conventional method, followed by a
initial cutting pattern generation using the technique used in the conventional method.
2.2. ANALYSIS OF FABRIC TENSION STRUCTURES 13

Next, the flat panels are deformed to form the 3D shape from the beginning, and the
stresses in the membrane are compared to the desired prestress, and a new cutting pattern
is generated [13]. Iteratively, cutting patterns are adjusted such that the final structure
has a stress distribution that has a minimal difference from the designed prestress [14].

In the integrated design method, the influence of patterning is incorporated in the form-
finding, and the correct non-linear mechanics of the membrane are used. Naturally, this
method is more accurate than the conventional method, as there are no approximations,
and the correct material model is used throughout the process. However, as it is an iterative
process, it takes longer to reach a solution.

Recently, a new design method based on the integrated approach was proposed by Dinh et
al. [15], based on shape optimization techniques. The method assumes that the intermedi-
ate shape of the assembled panels is flat, and uses shape optimization to find the optimal
intermediate shape, such that the final structure has a permissible stress distribution. Due
to this simplification, a solution could be found faster.

The method proposed in this thesis will continue on the work of Dinh et al. [15]. However,
the intermediate shape is not assumed to be flat anymore, but it is assumed that the inter-
mediate shape has a similar shape of the final structure, albeit with a smaller curvature.
The method will be discussed in detail in Chapter 6.
14

Chapter 3

Membrane materials

Membrane materials are very thin and lightweight, and therefore suited to applications
where large areas are to be covered. They are mostly classified in technical textiles and
technical plastics. Technical textiles consist out of large membrane parts, joined together
at the seam, and can only transfer loads by tension. Architectural fabrics are composites,
consisting of woven yarns coated to withstand weathering, loads and temperature effects.
Non-coated fabrics also exist for other purposes. Another option for tensile architecture is
to use technical plastics. These films are made of fluorothermoplastics and are transparent.
Because of this, the material is highly popular among architects. However, the strength is
far lower than that of coated fabrics, and frequently multiple layers are used. This has the
added benefit of providing isolation. Still, the lower strength of the material inhibits its
use in larger spans [16]. In what follows, coated textile fabrics will be discussed.

Coated textile fabrics are a combination of multiple materials, and are therefore composite
materials. They form a system of woven yarns, arranged orthogonally when unstressed,
which consist of single threads, parallel or twisted together. Two types of materials are
mostly used as architectural membranes: PVC-coated polyester, and PTFE-coated high-
strength glass fiber. Less frequently used are silicone-coated glass fiber fabrics [16].

3.1 Textile fabrics


A fabric is made by interlacing warp and fill threads. Multiple methods exist to weave
these fibers, namely the plain weave, basket weave and modified plain weave. Usually in
polymers, warp and fill threads are equally curved, as is the case in figure 3.1. However,
if large rolls of the fabric have to be manufactured, the coating process requires that the
3.1. TEXTILE FABRICS 15

fabric is pulled through drums along its warp direction. Because of this, the warp threads
are more stretched than the fill threads. This waviness is called crimp [16].

Figure 3.1: Diagram of different weaving methods, adapted from [16].

Polyester fabrics are cheap, and offer great resistance during production and erection.
Before yielding, the fabric can deform significantly, enabling timely corrections. However,
the material properties decrease under UV light, and it is subject to ageing. Polyester
fabrics are often combined with a PVC coating, and is mostly used for temporary structures.
The PVC coating also gives it fire-resistant properties.

Fibreglass fabrics have a higher tensile strength, but fail in a brittle way, with low elastic
straining. Because of this, the manufacturing process needs to be very accurate, and the
material should be handled with proper care. On the other hand, the material is almost
insusceptible to the effects of ageing, making it a preferred choice for long-time solutions. In
combination with a PTFE coating, the fabric is non-combustible, and relatively translucent
[17].

Finally, aramid fibers, most notably Kevlar, are used in tensile structures. Kevlar is used
for its fire-resistive properties, however it is far more expensive than its competitors. Again,
UV light can diminish its properties, requiring a coating [6].
3.2. COATING 16

3.2 Coating
Textile fabrics used in architecture have to be resistant against alternating loads, creep,
ageing, climatic and atmospheric effects, and should be non-flammable. In order to guar-
antee these characteristics, a suitable coating has to be applied to the fabric. The applied
coating has no load bearing function, only a protective function. However, applying the
coating gives the fabric a certain shear stiffness [16]. Additionally, the coating influences
the flexibility, the weldability, and the price of the fabric [6].

A commonly used coating material is polyvinyl chloride, or PVC. Usually, a PVC layer
alone is not sufficient, and an extra protection layer has to be applied to prevent ageing
under UV light. Mostly, a polyvinyl fluoride film is used for this purpose. A second coating
material is polytetrafluorethylene or PTFE, commercially available under the name Teflon.
PTFE is resistant against most chemicals and solvents, humidity, and UV light. It is dirt
repellent, and can face high temperature fluctuations. However, it has a lower flexibility
than PVC [6].

Recently, synthetic rubber coatings have been developed, such as Hypalon. These materials
have better sustainability and strength properties, along with a lower manufacturing price.
Finally, also polyurethane is used as a coating material [6].

Different methods exist to apply a coating. A pre-heated film can be pressed on the fabric,
or the coating can be applied in liquid form, after which the fabric has to be heated in an
oven [6]. However, in order to produce large quantities, a continuous method is required. In
the industry process, the coating material is heated so it starts flowing. Using a knife, the
coating is spread on the fabric while it is rolling under the knife. Afterwards, the coating
is treated, to give additional protection and prevent moisture penetration. To guarantee
that the coating provides sufficient protection, it should be thick enough. Generally, the
thicker the coating is, the better the protection of the thread. Usually, the thickness is
between 0.08 mm and 0.25 mm [16].

3.3 Mechanical Properties


The weaving process described above results in a specific material behaviour. Traditional
materials used in construction show linear elastic and isotropic behaviour, while membrane
materials exhibit non-linear, anisotropic and non-elastic characteristics [17].
3.3. MECHANICAL PROPERTIES 17

3.3.1 Tensile strength


The tensile strength of a fabric material is dependent on the tensile strength of its fibres. If
a single thread in the fabric fails, the load it initially carried is transferred to its neighbours.
If this happens to multiple threads, threads in the neighbourhood will also fail due to the
force distribution, and a rupture will occur [6].

3.3.2 Anisotropy
When a uniaxial tension test is conducted for strips that differ from each other in yarn
orientation, results such as in figure 3.2 are obtained. It is apparent that yarn orientation
has a noticeable influence on the material characteristics. When the yarns are oriented
perpendicular to eachother, the material behaviour is described as orthogonally anisotropic
or orthotropic.

Figure 3.2: Typical stress-strain curves from a uni-axial test in different orientations, adapted
from [17].

This anisotropy can be explained by the woven base of the fabric, and the way it is
3.3. MECHANICAL PROPERTIES 18

produced. As mentioned above, the warp threads are tensioned during production, while
the fill threads are woven in between. As a result, when the membrane is subjected to
tension in the fill direction, the fill fibres first straighten before they elongate. In contrast,
when the warp fibres are tensioned, they will deform less as they are already straight [17].

Additionally, the warp and fill yarns interact with eachother. When the membrane is
tensioned in its fill direction, the curvature of the warp threads changes at the intersection
points, causing a shortening in the warp direction. Similarly, when both directions are
tensioned, the tensioned warp threads exert a resistance on the fill threads. In other words,
the crimp in the fill direction is complemented by the crimp in the warp direction. This
behaviour is called crimp interchange. As a result, fabrics behave differently under varying
load ratios. The ratio of the strain 1 in the direction of the applied load to the resulting
crimp in the transverse direction 2 is Poisson’s ratio, or µ. In traditional materials,
Poisson’s ratio is constant and related to the modulus of elasticity of the material. In
contrast, the Poisson’s ratio of a fabric is dependant on the force applied [16].

3.3.3 Nonlinear behaviour


Unlike traditional materials used in construction, no linear relation can be found between
the stresses and strains in the material, as seen in figure 3.2, and as a result Hooke’s law
of elasticity cannot be used to describe material behaviour.

3.3.4 Inelastic behaviour


When fabric materials are subjected to repetitive, cyclic loads, some hysteresis is shown, as
seen in figure 3.3. It is clear that all loading and unloading paths are different, indicating
that the textile undergoes permanent strains, and that its material characteristics are
dependent on its loading history. Additionally, the extent, speed, duration and number of
load cycles have an influence on the material behaviour.

3.3.5 Shear stiffness


The shear stiffness of architectural fabrics is typically quite low. As a rule of thumb, 1/20th
of the tensile elastic modulus is commonly used as an estimation [18]. Due to this, the
detailed shear behaviour is often neglected in calculations. The shear stiffness is low at low
shear angles, since the yarns can rotate freely and only the coating provides some shear
resistance. However, starting from a certain angle, called the limiting angle, warp and fill
3.3. MECHANICAL PROPERTIES 19

Figure 3.3: Stress strain curves for uniaxial test in warp and fill directions in case of cyclic load,
adapted from [15].

yarns lock eachother, increasing the resistance against further shear displacement. [16].

3.3.6 Kink resistance


During erection, fabric materials are often subjected to significant mechanical effects. Local
damage to the coating of the fabric caused by folding or lifting should be prevented. Dam-
age to the coating can locally expose the fabric threads, which are subsequently subjected
to environmental influences. This can lead to premature failure of the construction [16].

3.3.7 Material model


Because of the previous material characteristics, it is a difficult task to model the material
behaviour. Despite the fact that the yarns in the fabric are oriented perpendicularly, the
material doesn’t follow an orthotropic elasticity material law. A full description of the load
response of these fabrics for biaxial and shear loading has not been carried out yet, and
even if it were, few techniques exist to interpret this data [18].

In the Commentary to the Membrane Structures Association of Japan (MSAJ) Testing


Method for Elastic Constants of Membrane Materials, a standardized method is described
for testing fabrics. For plain weave fabrics, a method consisting of biaxial and uniaxial
tests is proposed to define mechanical constants. An orthotropic elasticity model is often
assumed, since calculation software is often based on elastic parameters.

An elasto-plastic material model for coated fabrics was proposed by Dinh et al. in [19]. In
this paper, uniaxial tension, biaxial tension and uniaxial bias extension tests were carried
3.3. MECHANICAL PROPERTIES 20

out. The model handles interaction between warp and fill yarns, as well as the influence
of the coating. Different load ratios were considered, in order to capture the material
characteristics as correctly as possible. A stress-strain diagram of this model is shown in
figure 3.4.

Figure 3.4: A proposed elasto-plastic material model for coated fabrics, adapted from [19].

It is apparent that this material model is more complex than a simple orthotropic elasticity
law. The material follows a different stress-strain behaviour in the warp and fill direction,
as is the case in real life. In the warp direction, a linear correlation exists between the
stress and the strain, with a modulus Ew1 as long as the stress is lower than the yield stress
σwy . After this, linear hardening will occur, modeled with the plastic modulus Hw . When
unloading the textile, the relationship between stress and strain is again linear, albeit with
a different modulus Ew3 . Reloading the textile again follows a different material law, with
a modulus Ew4 .

The fabric acts similar in the fill direction, although the relationship between stress and
strain is nonlinear in the elastic region. Additionally, linear hardening happens in two
stages, with two different plastic moduli Hf 1 and Hf 2 . For the membrane used in this
thesis, the following moduli are used:

In the warp direction:


3.4. STRUCTURAL BEHAVIOUR 21

Ew1 = 1364 MPa Hw = 448 MPa

Ew3 Hw
Ew2 = Hw +Ew3
σwy = 15 MPa

Ew3 = 1130 MPa

In the fill direction:

Ef 1 = 10 000 MPa(14.1309(f )2 + 0.04f + 0.0115) Hf 1 = 593 MPa

Hf 1 Ef 4
Ef 2 = Hf 1 +Ef 4
Hf 2 = 236.82 MPa

Hf 2 Ef 4
Ef 3 = Hf 2 +Ef 4
σfy1 = 5 MPa

Ef 4 = 825 MPa σfy2 = 15 MPa

The warp and fill threads interact with each other according to two Poisson’s ratios ν12 =
0.09076 and ν21 = 0.46923. To take into account the load ratio dependency, two moduli
Ew4 = 852.241 MPa and Ef 5 = 618.402 MPa are used. Using a user material subroutine
(UMAT), this material model can be implemented in Abaqus.

This material model accurately represents the nonlinear behaviour in both warp and fill
directions, along with the permanent strains and dependency on load ratio. However,
the model is only verified for load ratios 1:1, 2:1, 1:2, 1:0 and 0:1. Also, time dependent
behaviour and resistance to shear are not considered.

3.4 Structural behaviour


As membrane materials are thin and flexible, they have low, negligible shear, flexural and
compressive stiffnesses. Due to prestress and geometry, some out-of-plane stiffness can be
achieved. However, if this prestress is too low, this out-of-plane stiffness is too low, and
wrinkling can occur [16].

Clearly, the prestress of the membrane is of the utmost importance for the structural
integrity of the structure. However, fabric materials exhibit creep behaviour, of which very
3.4. STRUCTURAL BEHAVIOUR 22

little is documented. As a result, prestress may vanish over time in some locations, along
with the out-of-plane stiffness. One can imagine that if rainfall occurs, out-of-plane stiffness
guarantees that water simply flows of the structure. However, over time, it can happen
that a certain portion of the structure forms a pond, in which water can accumulate. This
ponding phenomenon is to be avoided, as it can cause ruptures in the fabric. As the creep
behaviour is tough to predict, creep is usually compensated by re-tensioning the boundary
elements [6] [18].

To conceive the prestress in the membrane, boundary elements such as cables are used.
These boundary elements are commonly installed in sleeves at the edges of the membrane,
and by tensioning the elements, a prestress is introduced in the membrane. These flexible
boundary elements usually are made out of spiral steel wire ropes, or band-shaped textile
belts. Additionally, rigid boundary elements exist, which can be implemented to reinforce
the edge of the surface.

The deformation capability of flexible membranes is an important property of tensile


structure. While traditional structures mostly act by bending, requiring a large cross-
sectional area, membrane materials only transfer loads by tensile forces, resulting in a
lightweight design.
23

Chapter 4

Optimization

In [20], optimization is defined as “the art of making things the best”, or more correctly
“the process of maximizing and/or minimizing one or more objectives without violating
specified design constraints, by regulating a set of variable parameters that influence both
the objectives and the design constraints”. Without optimization, ideal solutions are
searched for using experience, intuition and plain luck. It goes without saying that this
method is suboptimal, ineffective, and dated. Using optimization, these solutions are found
in a systematic way, using the power and speed of a computer. Additionally, optimization
algorithms make sure this search is done in an efficient fashion.

Figure 4.1: Traditional vs. Optimal Design Process, adapted from [20].

Figure 4.1 illustrates the difference between traditional design approaches and optimal
design approaches. Box A represents the input, consisting of an initial design, a desired
performance, for example a minimal cost or material usage, and any constraints that may
4.1. OPTIMIZATION PROBLEMS 24

apply. Box B includes the analysis, or post-processing, which returns the output results
for any given set of variables. The optimization cycle begins at box C, which evaluates
the analysis results, and assesses whether the performance is acceptable. If this is not the
case, a design change will occur using optimization (Box D), or via intuition, experience,
... (Box E).

4.1 Optimization problems


Optimization problems can be classified by seven major characteristics. These character-
istics are important, as they can indicate which algorithms are applicable to the problem
at hand, and its complexity. These characteristics are as follows [20]:

• Linear vs. Nonlinear

• Constrained vs. Unconstrained

• Discrete vs. Continuous

• Single vs. Multiobjective

• Single vs. Multiple Minima

• Deterministic vs. Nondeterministic

• Simple vs. Complex

4.1.1 Linear vs. Nonlinear


A linear optimization or linear programming problem is a problem in which both the
objective function and the constraints are linear. When either the objective function or
any of the constraints is nonlinear, the problem is called a nonlinear problem. In general,
linear problems are more easy to solve. Solving a large nonlinear problem is also less
reliable than solving a large linear problem, and this difference is more pronounced when
the number of variables increases.

4.1.2 Constrained vs. Unconstrained


A problem that has constraints, such as most practical problems, is called a constrained
problem. In linear programming, constraints are required, as the solution would otherwise
not be finite.
4.1. OPTIMIZATION PROBLEMS 25

4.1.3 Discrete vs. Continuous


When any of the design variables is discrete, the optimization problem is no longer contin-
uous. Several cases exist. If the design variables can only take on the values 0 or 1, this
is called 0-1 programming or binary programming. If the design variables can only take
on integer values, this is called integer programming. When the design variables can only
take on a given set of values, this is called discrete optimization. Often, a problem consists
of a mixture of these cases. Usually, continuous optimization problems are easier to deal
with.

4.1.4 Single vs. Multiobjective


In some cases, more than one objective is to be optimized. However, improving one
objective often involves compromising one or more of the other objectives. In figure 4.2,
an optimization problem with 2 objectives is shown.

Figure 4.2: Multiobjective optimization, adapted from [20].

If only Objective 1 is to be optimized, a minimum is found in point M1, and similar with
Objective 2. However, if a better value for Objective 2 is to be found than the one in M1,
we have to move to the right, and compromise the value of Objective 1. This is true for
every point between M1 and M2, which are called the Pareto optimal solutions, or non-
dominated solutions. Points outside this region are called dominated solutions, as there
4.1. OPTIMIZATION PROBLEMS 26

are always points in which both objectives perform better; for example, point A can be
optimized by moving to the right. If the Pareto optimal solutions are plotted with a design
objective on each axis, a line known as the Pareto frontier appears.

Figure 4.3: Pareto frontier, adapted from [20].

It is clear that to solve a multiobjective problem, a Pareto optimal point is needed. In


two-objective problems, this is done with the weighted sum method, as seen in equation
4.1. With this method, a weighted average is taken from the objectives, which is the new
value to be minimized. Minimizing this value yields a Pareto point, and by changing the
weights, a series of Pareto points can be found.

J(x) = w1 µ1 (x) + w2 µ2 (x) (4.1)

4.1.5 Single vs. Multiple Minima


Some optimization problems have only one minimum or maximum (optimum), while others
have have to find the best solution among multiple local optimum values. These problems
are called unimodal and multimodal, respectively. Solving multimodal problems is referred
to as global optimization, which is a lot more difficult than single optimum optimization.
4.2. SOLUTION APPROACHES 27

4.1.6 Deterministic vs. Nondeterministic


In practical problems, for example when manufacturing a part, tolerances are very impor-
tant. Low tolerances often go hand in hand with a high cost. Because of this, information
is almost never exact. Similarly, some variables are simply unknown, and can only be
estimated. These variables are called nondeterministic.

4.1.7 Simple vs. Complex Problem


Assessing whether a problem is simple or complex involves all of the above characteristics.
A problem is simple if it involves only continuous variables, if it is linear, when local
optimization will suffice, ... In practice, each of above items only gives an indication of the
complexity of the problem.

4.2 Solution approaches


Three main solution approaches exist: experimental, analytical, and numerical. These are
briefly discussed in the following section.

4.2.1 Experimental Optimization


As mentioned before, without optimization, ideal solutions are searched for using experi-
ence, intuition and plain luck. Experimental optimization involves building a version of the
system, evaluating its performance, and determining whether the performance is accept-
able. If this is not the case, changes are made, and the process is repeated (see Figure 4.1).
It is very clear that this approach is obsolete, as it can be very costly and time-consuming,
with no guarantee to convergence at all.

4.2.2 Analytical Optimization


If a mathematical function exists that represents the performance of the design, an ana-
lytical approach is the most commonly used. From calculus, if a point is found where the
derivative is zero, this point is a local optimum. The second derivative of the function
can be used to find out whether the optimum is a minimum, a maximum or a saddle
point. If the problem contains constraints, Lagrange multipliers are used, which will not
be explained in further detail.
4.3. SUMMARY OF OPTIMIZATION ALGORITHMS 28

Before the use of computers, an analytical approach was the most viable. However, for
most practical issues, this approach is way too complicated, especially when compared with
a numerical approach. Furthermore, the majority of problems cannot be described by a
mathematical function.

4.2.3 Numerical Optimization


Indeed, a numerical optimization approach is the most modern way to optimize a design.
Using a computer, it can explore far more possibilities than any manual way. A numerical
approach starts with one (or more) design(s), evaluates the objective function, and esti-
mates a next design, based on an algorithm. This process is repeated until the stopping
criteria are met, hopefully reaching an optimal design.

This type of optimization uses iterations to reach an optimal solution. Compared to the
experimental approach, the next iteration is based on an algorithm and not on sheer luck,
and the use of a computer can result in a fast result.

A large assembly of algorithms and software exists to this purpose. In this thesis, the
MATLAB Toolbox is used, along with third-party algorithms implemented in MATLAB.

4.3 Summary of optimization algorithms


In this section, some frequently used global optimization algorithms implementable in
MATLAB are reviewed.

4.3.1 Gradient-based methods


Gradient-based methods find local extrema using first and second order derivatives, compa-
rable to analytical optimization as described above. In MATLAB, two gradient-based solvers
are incorporated: MultiStart and GlobalSearch. Both generate a series of starting points,
from which a MATLAB function called fmincon is used to find the local minima. The global
minimum is simply the local minimum with the lowest objective value.

At default, MultiStart uses uniformly distributed starting points, within the specified
bounds. User-supplied starting points are also supported. In contrast, GlobalSearch
generates the starting points by a scatter-search mechanism. GlobalSearch then analyzes
4.3. SUMMARY OF OPTIMIZATION ALGORITHMS 29

these points, and dismisses of those that are unlikely to generate a better minimum than
the best already found [21].

The fmincon algorithm evaluates the first- and second-order derivatives of the objective
function at a certain point, and uses this information to obtain the direction in which
to evaluate a new point. However, once a local optimum is found, the algorithm stops,
regardless of the existence of superior optima.

4.3.2 Genetic algorithms


Genetic algorithms, or GA, are inspired by Darwin’s principle of survival of the fittest, using
the objective function as a measure of fitness, and the so-called genes of an individual to
represent the solutions. Usually, the parameters of each solution are encoded, resulting in
the genes. For example, the number 21 can be represented in binary as 10101. The five
digits of this binary number can be treated as genes [20].

During each iteration, a new generation is created, consisting of individuals created from
the genes of so-called ‘parents’, which are randomly selected from the previous iteration.
Children are created using crossover, mutation or elitism. Elite children are unchanged
across generations, mutation involves a random change in the genes of a parent, and
crossover randomly combines the genes of two parents. Using the objective function,
the function values of the individuals are calculated. After each iteration, the individ-
uals of that iteration are assigned a fitness value, which is based on the objective value.
Afterwards, a new generation is created, and the process is repeated.

In GA, it is important that the population has a large genetic diversity. For example, if
only crossover is considered, new children can only exist out of genes from the previous
iteration. The only way to introduces new genes is through mutation.

A genetic algorithm does not require information about the objective function, in contrast
to gradient-based methods, so the objective function is treated as a black-box function. As
all individuals of a generations are known when the generation is created, multiple objective
values can be calculated at once, increasing the performance of the algorithm. This
technique is called parallelization. Genetic algorithms are powerful and computationally
simple, making it a useful algorithm for a myriad of optimization problems [21] [22].
4.3. SUMMARY OF OPTIMIZATION ALGORITHMS 30

Figure 4.4: Procedure for a Genetic Algorithm, adapted from [20].

4.3.3 Pattern Search


The term pattern search refers to a family of direct search algorithms. These types of
algorithms do not require information about the gradient or objective function, and are
therefore more suited to non-smooth or discontinuous objective functions. The pattern
search algorithm constructs a pattern of multiple points around a starting point, using a
specified poll method. Next, the function values at each of these points are calculated.. A
new pattern is then constructed around the point with the best function value, and the
process is repeated.

However, as the searching direction is unknown, these methods can be computationally


expensive [21]. The global convergence of pattern search algorithms was proven by V.
Torczon [23]. However, if the objective function to be minimized has many local minima,
the algorithms might get stuck at one of the local minima, instead of continuing to find
the global minimum.
4.3. SUMMARY OF OPTIMIZATION ALGORITHMS 31

4.3.4 Simulated annealing


Simulated annealing (SA) is an optimization process inspired by the physical process of
annealing, or reducing defects in a material by heating and cooling it. It is a direct search
function, meaning it does not require information about the gradient or the objective
function. As a result, this process is more suited to non-smooth or discontinuous objective
functions. However, problesimms occur when applying this function to a problem with
constraints. Combining SA with another solver can alleviate these problems. This is
called hybrid simulated annealing (HSA) [21].

4.3.5 Bayesian optimization with Gaussian process priors


In computer science, machine learning algorithms are often dependent on parameters
that significantly impact the performance of the algorithm. To enhance the quality of
the algorithm, these parameters have to be optimized. In this, the performance of the
algorithm is a so-called black-box function, meaning only the input and output of the
function are known, without any knowledge of its internal workings. In this optimization,
the performance evaluation is usually computationally expensive, so the number of function
evaluations should be minimal [24].

When function evaluations are expensive, more computational time should be spent on
determining the next point where the function should be evaluated. For this kind of prob-
lem, Bayesian optimization provides an elegant approach, and outperforms some other
state of the art global optimization algorithms. For continuous functions, Bayesian opti-
mization assumes the unknown black-box function was sampled from a Gaussian process,
and maintains a posterior distribution for this function after observations [24].

Bayesian optimization constructs a probabilistic model for f (x), and uses this model to
choose the location of the next function evaluation, while integrating out uncertainty.
In contrast to traditional optimization algorithms, Bayesian optimization uses information
from previous evaluations of f (x), and it does not rely solely on a local gradient or Hessian.
In this procedure, a minimum of difficult functions can be found using a relatively small
amount of function evaluations, which is especially interesting when a function evaluation
is expensive to perform [24]. An illustration of the optimization algorithm can be found in
figure 4.5.

In Bayesian optimization a prior has to be chosen, which has to give assumptions about
the function that is to be optimized. In this case, a Gaussian process prior is chosen. This
4.4. COMPARISON OF OPTIMIZATION ALGORITHMS 32

Figure 4.5: An example of a Bayesian optimization on a 1D optimization problem, adapter


from [25]. The figure also contains the acquisition function, which is highest in
the points where the Gaussian process predicts a high objective along with a high
uncertainty.

prior is responsible for the blue zone in figure 4.5. Additionally, an acquisition function
has to be chosen, which has to allows us to determining the ideal location for the next
function evaluation. This function determines the green graph in figure 4.5.

4.4 Comparison of optimization algorithms


The genetic algorithm, pattern search and simulated annealing algorithms were compared
in [21], optimizing a hot rolling process. The study found that Pattern Search and Multi-
Start provide on average the fastest result, but that MultiStart performs more consistently.
It is also noted that the GA is too slow and not accurate.

In MATLAB documentation, several solvers are used to optimize Rastrigin’s function, a


4.4. COMPARISON OF OPTIMIZATION ALGORITHMS 33

function with a lot of local minima [26]. Among the solvers used are the GA, GlobalSearch
and Pattern Search algorithms. The Pattern Search algorithm returned a value within 174
function evaluations, compared to 5400 evaluations for the GA and 2178 evaluations for
the GlobalSearch. However, only GlobalSearch and GA reached the global minimum of
the function, while the Pattern Search algorithm got stuck in one of the local minima.

Clearly, no optimization algorithm stands out as ‘the best’. Instead, some algorithms are
more suited to certain problems. For the project discussed in this thesis, an optimization
algorithm is required. As it is not known beforehand which optimization technique is the
most suited for this issue, the different algorithms have to be compared against eachother.
In this comparison, the Pattern Search algorithm will be considered along with the GA
and the Bayesian algorithm.
34

Chapter 5

Shape optimization

Shape optimization is a mathematical approach that optimizes a material layout within a


given design space, for a given set of loads and boundary conditions, such that the resulting
layout meets a prescribed set of performance targets. Using shape optimization, engineers
can find the best concept design that meets the design requirements. The method is of
great importance in aircraft design, bridge construction, and many other industries. Shape
optimization problems can also be described as the minimization of a certain objective
function within certain boundaries, where the objective function depicts the quality of
the design [27]. Two main types of shape optimization are parametric and nonparametric
shape optimization.

5.1 Parametric shape optimization


In general, a parametric shape optimization process consists of three steps. A parametrized
geometry, also known as the design model, the analysis of the structural response, also
known as the analysis model, and finally the optimization algorithm. Some drawbacks
herein are the exchange between the CAD model and the FEM solver, and limited vari-
ability of the possible shapes, as they are described by a finite number of parameters [28].

5.1.1 Design model


In the design model, the potential variations of the considered shape are governed by
certain parameters, or design variables. This design model is often described by a CAD
model. Several CAD applications exist to create a design model, such as SolidWorks or
AutoCAD.
5.1. PARAMETRIC SHAPE OPTIMIZATION 35

Figure 5.1: General shape optimization process, adapted from [28].

Many methods exist to represent a shape by parameters. For example, Bezier curves or
B-splines can be used to define the edges of a certain shape. These curves are variation
diminishing, meaning that they lie within the convex hull of the control points that define
it. They are independent of the coordinate system used to measure the location of the
control points, and they can be used to create lots of shapes. If the edges of a shape that is
to be optimized are represented by these curves, smoothness is guaranteed, resolving some
problems of the methods above [29]. When considering 3D shells, or membrane structures,
splines are often used [19] [30].

5.1.2 Analysis model


In the analysis model, the considered shape is subjected to its boundary conditions and
external loads. Afterwards, the structural response of said shape is computed using nu-
merical methods such as FEM. If the analysis model is not linked with the design model,
as is the case when using Bezier curves or B-splines, the approach requires an automated
meshing procedure [31].

Numerous FEM applications exist for this purpose, such as Ansys and Abaqus. However,
not all FEM software is compatible with all CAD software, which is one of the negative
aspects of this method.
5.2. NON-PARAMETRIC SHAPE OPTIMIZATION 36

5.1.3 Optimization algorithm


Lastly, the optimization algorithms handles the iterative process of reaching an optimum.
As discussed before, a large variety of optimization algorithms exists. One can make the
difference between gradient algorithms, in which the gradient information can be accessed
by the optimization software, or response-surface-based algorithms, in which this is not
the case [31]. When using a commercial CAD system, it is often not straightforward to
calculate the gradient.

In practice, genetic algorithms are frequently used in shape optimization problems in which
the gradient cannot be calculated. For example, in the study conducted by Woon et al.,
a genetic algorithm was used to optimize the shape of a simple spanner head. The shape
of the spanner head was described by 20 parameters, and a GA with a population of 18
was used for 807 generations [32]. Similarly, in the study conducted by Gilbert et al., a
shape optimization for thin-walled steel cold-formed steel profiles was conducted using a
genetic algorithm. The algorithm reached an optimal solution after 50 generations with a
population size of 50 [33]. Finally, in a study conducted by Annicchiarico et al, the shape
of a perforated cantilever is optimized using GA. The shape is described by 5 parameters,
and 50 generations were used with a population of 100 [34].

However, this does not take away that other algorithms are not suited for shape optimiza-
tion. In contrary, similar to any optimization problem, some algorithms perform better
than others, depending on the nature of the problem.

5.2 Non-parametric shape optimization


The main disadvantage of the parametric approach are that only shapes within the so-
lution space are considered, meaning that limiting the number of parameters will lead to
suboptimal results. Additionally, a CAD geometry is required, along with automatic mesh
generation, limiting the approach to simple triangle and tetrahedron meshes.

If these limitations are considered a problem, a non-parametric shape optimization can be


carried out instead. In this method, the design space is built up by implicit parameters
from a chosen set of surface nodes. These implicit parameters are the displacements of
the design nodes, along a local optimization displacement vector. As each design node
can be modified, all possible solutions lie in the design space. In non-parametric shape
optimization, there is no real difference between the design model and the analysis model.
5.2. NON-PARAMETRIC SHAPE OPTIMIZATION 37

If the edges of the considered shape are to be optimized, one could use the finite element
method to mesh the geometry, and vary the coordinates of nodes on the design boundary.
This method is called independent node movement. As proven by Neuber [35], the stress
in an edge is only influenced by the geometry in the vicinity of the stress peak. Non-
parametric shape optimization can change the displacements of nodes in this vicinity in
order to eliminate these stress peaks.

However, because the stresses are sensitive to geometric distortion of the elements, nu-
merical instabilities can occur when using this method. For example, the boundary could
become jagged after optimization, which is of course undesirable. Additionally, this method
uses a large number of design variables, which inhibits fast results [36].

To overcome these numerical instabilities, design elements could be used. These design
elements consist out of multiple finite elements.. The shape of these design elements is also
controlled by its nodes, however the number of nodes is less than the sum of the number
of nodes of its components. As shown in figure 5.2, these design elements exist in both
two-dimensional [37] and three-dimensional cases [38].

(a) Some two-dimensional design elements (b) 20-noded three-


[37]. dimensional design element,
adapted from [38].

Figure 5.2: Example of cloth unfolding, adapted from [10].

To further improve non-parametric shape optimization, optimality criteria methods should


be implemented. This means that an optimization displacement for each node is calculated
5.2. NON-PARAMETRIC SHAPE OPTIMIZATION 38

as a result of the stress in the FEM analysis. This method can solve problems with a large
number of design variables in a timely fashion, as the convergence speed is independent on
the number of variables [31].

One of the advantages of this method is that a bigger number of solutions can be evaluated
when compared to parametric shape optimization.
39

Chapter 6

Shape optimization project

As discussed earlier, the conventional method of designing and patterning has its limita-
tions. The method neglects certain aspects of the used fabric material, such as its non-linear
and non-elastic behaviour. An alternative is to use the so-called integrated approach. In
this thesis, a variation on the integrated method for designing and patterning hyperbolic
paraboloid structures, or hypars, is proposed. The method, which is based on shape opti-
mization, is a continuation of the work done by Dinh et al. [15], and takes into account the
material characteristics of the membrane material by using the material model proposed
by Dinh et al. [19].

6.1 Proposed design method


As described earlier, the integrated method consists of iteratively generating cutting pat-
terns, which are then assembled to achieve a so-called intermediate configuration, as seen
in figure 6.1. This intermediate configuration is then subjected to tensioning using FEM,
and after computation the stresses are studied. However, this method can be quite com-
putationally intensive, as the whole assembly process is simulated at every iteration.

In the method proposed in this thesis, the integrated method will be simplified by postpon-
ing the first step in this process. Instead, an intermediate configuration will be iteratively
generated and subjected to tensioning. It is assumed that the intermediate configuration
of the structure has a similar shape, albeit with a smaller curvature.

In this case, the method will be tested using a hyperstructure or hypar, a commonly used
shape in tensile architecture. The hypar considered in this thesis has the same dimensions
as the hypar used in the study conducted by Dinh et al. [19], and is a square hypar with a
6.1. PROPOSED DESIGN METHOD 40

Figure 6.1: The different steps in creating a membrane structure, adapted from [7].

side of 4 m and a height of 0.8 m. Figure 6.2 is a depiction of the final shape of the hypar.
The hypar will be reinforced in its edges with cable elements, installed through sleeves in
the membrane. The target stress in the membrane is 5 MPa, while the target force in the
boundary elements is 20 kN.

Figure 6.2: Final configuration of the hypar, adapted from [19].

First, an intermediate configuration is described by a set of parameters, which is then


subjected to deformations leading to the final shape. Next, analysis is done using FEM,
leading to stresses in the membrane and forces in the edge cables. After this, these
stresses and forces are compared to a certain reference stress and force. This process is
repeated until the difference between the reference stress/force and the actual stress/force
6.2. DESIGN MODEL 41

is minimal. In this way, an optimal intermediate shape is found using an optimization


algorithm. As the shape of the intermediate configuration is prescribed by parameters,
this is a case of parametric optimization. It is assumed that this intermediate shape is
stress-free. After the optimal shape is found, the patterning can be performed. Finally,
the stress-free assumption is controlled by flattening the patterned shape, and looking at
the deformations. If these are sufficiently low, the stress-free assumption is verified. The
influence of the welds is neglected, and environmental loads are not taken into account.

6.2 Design model


To build the model used in this study, an intermediate shape has to be prescribed by
parameters. In figure 6.3, five state variables are shown which determine the initial shape.
p2 and p5 indicate the vertical coordinates of the corners relative to a flat plane, while
p3 indicates the curvature of the edge of the membrane. The middle point of the edge
vertically lies in the middle the two corners. p1 and p4 indicate the extra lengthening or
shortening of the membrane; this might be necessary to prevent negative stresses in the
membrane.

Figure 6.3: Intermediate configuration of the hypar, including its state variables

These parameters determine the location of eight points: four points at the corners of the
membrane, and four points at the middle of the edges. Using SolidWorks, the edges are
drawn as splines through these points. With the boundary surface feature, a minimum
6.3. ANALYSIS MODEL 42

surface is created using these splines as edges. The shape is then exported as a STEP-file.
This step of the process is done using a SolidWorks VBA script. To prevent issues in
creating the CAD model, or singularities in the FEM model, the parameters are limited
between certain boundaries. These boundaries can be found in table 6.1.

Table 6.1: Boundaries of the state variables


p1 p2 p3 p4 p5
Lower boundary (m) -0.05 0 0 -0.05 0
Upper boundary (m) 0.05 0.40 0.75 0.05 0.40

6.3 Analysis model

6.3.1 Creating the model


In order to calculate the stresses in the membrane after erection, a FEM model has to
be created. This can be done by importing the previously created STEP-file as a part in
Abaqus. In order to make computation possible, boundary conditions have to be created,
along with material descriptions.

Cable elements are installed in a sleeve along the edge of the membrane. To control the
displacements, a tie constraint is put in place, linking the displacements of the cable to those
of the membrane. To represent the cable-sleeve interaction, SLIPRING connectors are
created along the cable. These connectors represent a frictionless pulley system, allowing
the cable to slip through the sleeve without resistance. The cable elements have an effective
stiffness of 2000 kN. In order to prevent any singularities during computation, the corners
of the membrane are chamfered using a sphere with a 5 cm radius.

Next, displacements are applied at the corners of the membrane, in order to transform
the intermediate configuration to the final configuration. The deformations are vertical,
to reach the final height of the structure, and horizontal, to compensate for the length
difference of the diagonals (p1 and p4 in figure 6.3). Consecutively, the job is submitted.
This step of the process is done using an Abaqus Python script.

Finally, the material characteristics of the membrane are assigned. In a first step, an
orthotropic elasticity material model is used for computations. The material characteristics
6.3. ANALYSIS MODEL 43

of this model can be found in table 6.2. Later, the model proposed by Dinh et al. [19] will
be used. The characteristics of this model are described in 3.3.7.

Table 6.2: Parameters of the orthotropic elasticity material model


Ew Ef ν12 ν21 G
975.6 MPa 716.1 MPa 0.23 0.17 87.7 MPa

6.3.2 Computation and post-processing


As the problem studied here is a nonlinear one, Abaqus uses a volume-proportional damp-
ing to prevent instabilities. This implies adding viscous forces Fv to the global equilibrium
equations.

P − I − Fv = 0 (6.1)
Fv = c · M ∗ · v (6.2)

In these formulas, M ∗ is an artificial mass matrix, c is a certain damping factor, v = ∆u


∆t
is the nodal velocity vector and ∆t is the time increment. As a rule of thumb, the viscous
damping energy (ALLSD) should be smaller than about 5% or 10% of the total strain
energy (ALLIE), in order to guarantee a correct solution. After experimenting with some
values of the damping factor, a value of c = 10−8 returned good results for the orthotropic
elasticity material model, while a value of c = 2.5 · 10−7 is used when using the fabric
plasticity model.

After completion of the job, the results are available for post-processing. In this case, the
stress distribution in warp and fill direction in the membrane is used, together with the
force in the cable. In order to optimize the shape of the membrane, these values have to
be converted into a single value, also known as the objective. This step of the process is
done using an Abaqus Python script.

To do this, an average of absolute deviations over all the nodes of the model is calculated,
compared to a certain reference stress/force. For the membrane, this reference stress is
5 MPa, while for the cable, the reference force is 20 kN. The formula used is formula 6.3. As
shown, the formula consists out of two parts: a segment for the membrane, and a segment
for the cable. Using weight factors these two segments are combined into a single value.
6.4. OPTIMIZATION ALGORITHM 44

W (P ) = w1 · D1 + w2 · D2 (6.3)
where

1 X
D1 = |σ membrane − σref
membrane
| (6.4)
nmembrane nmembrane

and

1 X
D2 = |F cable − Fref
cable
| (6.5)
ncable ncable

Of course, these weight factors are not known beforehand, and can be arbitrarily chosen.
However, different choices for the weight factors will, after optimization, result in different
solutions. To resolve this, different combinations of weight factors will be used.

Additionally, to enhance the quality of the final solution, some penalty factors are intro-
duced. If the viscous damping energy is more than 10% of the total strain energy, the
result is not accurate as describe above. Therefore, if this occurs, or when the calculations
do not converge, a high value is returned. This way, the optimization algorithm will steer
away from points in which this is the case.

6.4 Optimization algorithm


Finally, an optimization algorithms is necessary in order to find the configuration in which
the objective function has the lowest value. As mentioned before, some good candidates
for this purpose are the genetic algorithm, and the pattern search algorithm, both available
using MATLAB. Additionally, a Bayesian optimization algorithm with 1000 iterations based
on the study conducted by Kawaguchi et al. [39] was used. For the genetic algorithm, the
same parameters as in [34] are used, as that study also considers a 5-parameter problem.
These parameters can be found in table 6.3.

Table 6.3: Genetic algorithm parameters


Population size 100
Number of generations 50
Selection scheme RSSWR (Remainder stochastic sampling without replacement)
Crossover probability 0.8
6.5. PATTERNING AND VERIFYING THE STRESS-FREE HYPOTHESIS 45

6.5 Patterning and verifying the stress-free hypothe-


sis
As discussed, the intermediate shape of the membrane is assumed to be stress free. In
reality, this shape consists out of several flat panels welded together. However, the inter-
mediate shape of the membrane is a double-curved surface, and is therefore undevelopable
to a flat surface. Because of this, some deformation will always occur when assembling the
panels. The goal is to minimize this deformation, and as a result to minimize the stress in
the intermediate shape. In order to validate this stress-free hypothesis, the intermediate
configuration found in the shape optimization will be divided into six pieces of equal width,
with the edges of the panels laying on geodesic lines, as discussed earlier. These panels
should be able to be flattened without significant strains.

To computationally flatten these panels, Rhino3D’s Squish algorithm is used. This algo-
rithm first meshes the surface, and then flattens the mesh by minimizing the changes in
facet area and changes in facet edge lengths. The panels are flattened in such a way that
they only stretch during assembly, to prevent wrinkling. The algorithm returns the average
and maximal equivalent strain caused by the flattening. If this maximal equivalent strain
is small enough, it can be assumed that the intermediate shape can be assembled without
introducing significant stresses.

6.6 Overview
In a first step, a shape optimization will be carried out, using the orthotropic elasticity
material model. The GA, Pattern Search and Bayesian algorithms will be used for this.
From these optimization processes, it will be clear which algorithm is the most suited to
this type of problem. Subsequently, a shape optimization will be done using the algorithm
determined in the first step, using both the orthotropic elasticity and the fabric plasticity
model. In these shape optimizations, the stress-free hypothesis has to be confirmed on the
found shapes. A clear overview of the method is given in figure 6.4.
6.7. BENCHMARK 46

Figure 6.4: Flowchart depicting the proposed design method

6.7 Benchmark
As this study is a continuation on the study conducted by Dinh et al. [15], the results
obtained in that study can be used as a benchmark for the results in this study. As the
intermediate shape used here has more parameters than the shape used by Dinh et al.,
lower stress and force deviations are expected. Therefore, obtained results will be compared
with the results from said research. A weight factor combination of w1 = 0.7 and w2 = 0.3
was used to obtain the benchmark results.
6.7. BENCHMARK 47

6.7.1 Orthotropic elasticity material model


As shown in figure 6.5 and table 6.4, a reasonable stress distribution was found in the
membrane along with a permissible cable force. However, there are some zones where the
stress in one of the directions is lower than 3 MPa, indicated in blue, which might inhibit
the out-of-plane stiffness of the membrane.

(a) S11 (b) S22

Figure 6.5: Membrane stresses in the optimized shape found by Dinh. et al. [15] using the
orthotropic elasticity model

Table 6.4: Parameters and results (w1 = 0.7 and w2 = 0.3)


p1 (m) p2 (m) p3 (m) p4 (m) p5 (m) Fcable (kN) D1 D2 W
0.020 0.000 0.350 0.021 0.000 20.247 1.13079 0.12361 0.82863

6.7.2 Fabric plasticity material model


As shown in figure 6.6 and table 6.5, a reasonable stress distribution was found in the
membrane along with a permissible cable force. However, the stresses are more varied in
this case, and the force in the cable is also further apart from the target value. Again,
zones are visible where the stress in one of the directions is lower than 3 MPa. These zones
are larger than in the previous case.
6.7. BENCHMARK 48

(a) S11 (b) S22

Figure 6.6: Membrane stresses in the optimized shape found by Dinh. et al. [15] using the fabric
plasticity model

Table 6.5: Parameters and results (w1 = 0.7 and w2 = 0.3)


p1 (m) p2 (m) p3 (m) p4 (m) p5 (m) Fcable (kN) D1 D2 W
-0.011 0.000 0.482 0.032 0.000 19.575 1.81705 0.21243 1.33566
49

Chapter 7

Results

7.1 Comparing optimization algorithms using the or-


thotropic elasticity model
First, the algorithm used for the shape optimization has to be determined. To do this, the
shape optimization is carried out using three different algorithms. The results obtained
from these optimizations are then compared based on their speed and quality. Simulations
are carried out for the five combinations of weight factors listed in table 7.1.

Table 7.1: Different weight factor combinations


Combination w1 w2
1 0.1 0.9
2 0.3 0.7
3 0.5 0.5
4 0.7 0.3
5 0.9 0.1

From the results it is visible that Pattern Search is by far the fastest algorithm. However,
its results are inferior to those of the GA and the Bayesian algorithm. From those two
algorithms, the Bayesian algorithm is both faster and more reliable than the genetic
algorithm. While the Pattern Search algorithm is faster than the Bayesian algorithms,
the Bayesian algorithm is chosen for further optimization procedures as it returned the
best result in a reasonable number of function evaluations. In some cases the genetic
7.1. COMPARING OPTIMIZATION ALGORITHMS USING THE ORTHOTROPIC
ELASTICITY MODEL 50
algorithm did not provide results at all due to a software error preventing the optimization
from completing.

Table 7.2: Algorithm comparison for the case w1 = 0.1 and w2 = 0.9
Function evaluations D1 D2 W
Bayesian 1000 5.4230 0.0126 0.5597
GA NA NA NA NA
Pattern Search 378 12.5622 0.0004 1.2565

Table 7.3: Algorithm comparison for the case w1 = 0.3 and w2 = 0.7
Function evaluations D1 D2 W
Bayesian 1000 5.6420 0.0180 1.7527
GA NA NA NA NA
Pattern Search 289 9.3203 0.0033 2.7771

Table 7.4: Algorithm comparison for the case w1 = 0.5 and w2 = 0.5
Function evaluations D1 D2 W
Bayesian 1000 5.6568 0.6396 2.7611
GA 4300 6.2179 0.0005 3.0000
Pattern Search 272 6.8295 0.2210 3.5293

Table 7.5: Algorithm comparison for the case w1 = 0.7 and w2 = 0.3
Function evaluations D1 D2 W
Bayesian 1000 5.4137 0.0038 3.8372
GA 4600 5.7680 0.0678 3.8995
Pattern Search 280 5.6954 0.0479 3.9725

Table 7.6: Algorithm comparison for the case w1 = 0.9 and w2 = 0.1
Function evaluations D1 D2 W
Bayesian 1000 5.6711 0.0796 4.8988
GA NA NA NA NA
Pattern Search 291 4.1361 11.1634 5.1495
7.1. COMPARING OPTIMIZATION ALGORITHMS USING THE ORTHOTROPIC
ELASTICITY MODEL 51
The five obtained results should lie on Pareto points or non-dominated points. However,
this is not the case here. In this case, as seen in table 7.7, one result clearly stands out
from the rest in both objectives - result 4. This might be explained by the difference in
magnitude of the objectives. In the solutions, D1 is two orders of magnitude bigger than
D2 . As a result, the optimizations converge to similar results, regardless of the weight
factors.

Table 7.7: Results for each weight factor combination


Result w1 w2 p1 p2 p3 p4 p5 D1 D2
1 0.1 0.9 0.0067 0.3778 0.4154 0.0000 0.3333 5.4230 0.0126
2 0.3 0.7 0.0082 0.3333 0.3648 0.0000 0.3778 5.6420 0.0180
3 0.5 0.5 -0.0030 0.3778 0.4500 0.0000 0.3778 5.6568 0.6396
4 0.7 0.3 0.0073 0.3752 0.4130 0.0000 0.3333 5.4137 0.0038
5 0.9 0.1 -0.0030 0.3926 0.4167 0.0067 0.3333 5.6711 0.0796

(a) S11 (b) S22

Figure 7.1: Membrane stresses in the optimized shape

Fcable =20.008 kN

However, after determining the ideal algorithm, a bug was discovered in the code creating
the model. The result obtained above will therefore be disregarded. Still, the algorithm
comparison is useful, and it is assumed that the Bayesian algorithm still provides the best
solution out of the tested algorithms. In what follows, the Bayesian algorithm is used for
shape optimization.
7.2. SHAPE OPTIMIZATION USING THE ORTHOTROPIC ELASTICITY MATERIAL
MODEL 52

7.2 Shape optimization using the orthotropic elastic-


ity material model

7.2.1 w1 = 0.7 and w2 = 0.3


First, a shape optimization is carried out using the orthotropic elasticity material model.
Initially, the weight factors will be w1 = 0.7 and w2 = 0.3, so a good comparison can be
made with the benchmark result. Clearly, this result is better than the benchmark result,
as both D1 and D2 are lower. A better stress distribution is found, and there are no large
zones with small stresses. The stress-free hypothesis has to be confirmed, so the membrane
is split symmetrically into six panels along geodesic lines, which are then flattened. The
deformations resulting from this flattening can be seen in table 7.9. These deformations
are low enough to assume a stress-free state after assembly.

(a) S11 (b) S22

Figure 7.2: Membrane stresses in the optimized shape using the orthotropic elasticity model
(w1 = 0.7 and w2 = 0.3)

Table 7.8: Parameters and results (w1 = 0.7 and w2 = 0.3)


p1 (m) p2 (m) p3 (m) p4 (m) p5 (m) Fcable (kN) D1 D2 W
-0.011 0.367 0.310 0.000 0.242 19.998 0.77947 0.00100 0.54953

While D2 is already very close to zero, and with it the cable force close to 20 kN, the stress
deviation in the membrane D1 can still be improved. Better results for D1 can be found
7.2. SHAPE OPTIMIZATION USING THE ORTHOTROPIC ELASTICITY MATERIAL
MODEL 53
10 2

Function evaluations
9
Best found value 1.8

8 1.6

7 1.4
Function value

Function value
6 1.2

5 1

4 0.8

3 0.6

2 0.4

1 0.2

0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Function evaluation Function evaluation

Figure 7.3: Optimization process

Figure 7.4: Optimized shape divided in panels

Table 7.9: Equivalent strains caused by assembly, panels numbered from left to right
Panel 1 Panel 2 Panel 3
Maximum 0.27% 0.11% 0.02%
Average 0.06% 0.03% 0.01%

by shifting the weight towards it. Therefore, the optimization will be repeated with weight
factor combinations w1 = 0.9 and w2 = 0.1, and w1 = 0.95 and w2 = 0.05.
7.2. SHAPE OPTIMIZATION USING THE ORTHOTROPIC ELASTICITY MATERIAL
MODEL 54

7.2.2 w1 = 0.9 and w2 = 0.1


Indeed, by moving the weight more towards D1 , a lower value for D1 is reached, while
compromising on D2 . The force in the cable is still very close to the target value, and
the stress distribution in the membrane is allowable. Once more, no large zones with low
stresses are observed. To verify the stress-free hypothesis, the membrane is again split into
panels and flattened. The maximum deformation is low enough to neglect and thus to
verify the stress-free state of the intermediate configuration.

(a) S11 (b) S22

Figure 7.5: Membrane stresses in the optimized shape using the orthotropic elasticity model
(w1 = 0.9 and w2 = 0.1)

Table 7.10: Parameters and results (w1 = 0.9 and w2 = 0.1)


p1 (m) p2 (m) p3 (m) p4 (m) p5 (m) Fcable (kN) D1 D2 W
-0.012 0.379 0.312 -0.004 0.251 19.983 0.76071 0.00841 0.68548

Table 7.11: Equivalent strains caused by assembly, panels numbered from left to right
Panel 1 Panel 2 Panel 3
Maximum 0.19% 0.12% 0.01%
Average 0.05% 0.03% 0.01%
7.2. SHAPE OPTIMIZATION USING THE ORTHOTROPIC ELASTICITY MATERIAL
MODEL 55
10 2

Function evaluations
9
Best found value 1.8

8 1.6

7 1.4
Function value

Function value
6 1.2

5 1

4 0.8

3 0.6

2 0.4

1 0.2

0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Function evaluation Function evaluation

Figure 7.6: Optimization process

Figure 7.7: Optimized shape divided in panels

7.2.3 w1 = 0.95 and w2 = 0.05

While the obtained cable force deviation is equal to the previous result, the membrane
stress deviation is actually higher. As such, this result is dominated by the previous result
and is therefore not useful.

Table 7.12: Parameters and results (w1 = 0.95 and w2 = 0.05)


p1 (m) p2 (m) p3 (m) p4 (m) p5 (m) Fcable (kN) D1 D2 W
-0.011 0.249 0.315 -0.004 0.378 20.017 0.77789 0.00841 0.73657
7.2. SHAPE OPTIMIZATION USING THE ORTHOTROPIC ELASTICITY MATERIAL
MODEL 56

(a) S11 (b) S22

Figure 7.8: Membrane stresses in the optimized shape using the orthotropic elasticity model
(w1 = 0.95 and w2 = 0.05)

10 2

Function evaluations
9
Best found value 1.8

8 1.6

7 1.4
Function value

Function value

6 1.2

5 1

4 0.8

3 0.6

2 0.4

1 0.2

0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Function evaluation Function evaluation

Figure 7.9: Optimization process

Table 7.13: Equivalent strains caused by assembly, panels numbered from left to right
Panel 1 Panel 2 Panel 3
Maximum 0.16% 0.11% 0.02%
Average 0.04% 0.03% 0.01%
7.2. SHAPE OPTIMIZATION USING THE ORTHOTROPIC ELASTICITY MATERIAL
MODEL 57

Figure 7.10: Optimized shape divided in panels


7.3. SHAPE OPTIMIZATION USING THE FABRIC PLASTICITY MATERIAL MODEL 58

7.2.4 Conclusion
Using the orthotropic elasticity model, the method provides good results, not only for the
stress distribution in the membrane, but also for the force in the cable and the deformation
due to flattening. As the first two obtained configurations are Pareto solutions, they are
equal solutions to the design problem. However, the second result showed the smallest
deformation during flattening, and is therefore deemed the optimal solution.

7.3 Shape optimization using the fabric plasticity ma-


terial model

7.3.1 w1 = 0.7 and w2 = 0.3


The same optimizations are repeated using the fabric plasticity material model. Again, the
obtained result is better than the benchmark result, as both the stress deviation and the
force deviation are smaller. Compared to the orthotropic elasticity solutions, a larger zone
with low fill stresses is apparent. Similar to the orthotropic elasticity optimization, there is
room to improve D1 , while compromising on D2 . When flattening the panels, deformation
up to 0.57% occurs. The panels resulting from this shape optimization are smaller than
those resulting from the orthotropic elasticity optimization.

(a) S11 (b) S22

Figure 7.11: Membrane stresses in the optimized shape using the fabric plasticity model (w1 =
0.7 and w2 = 0.3)
7.3. SHAPE OPTIMIZATION USING THE FABRIC PLASTICITY MATERIAL MODEL 59

Table 7.14: Parameters and results (w1 = 0.7 and w2 = 0.3)


p1 (m) p2 (m) p3 (m) p4 (m) p5 (m) Fcable (kN) D1 D2 W
-0.033 0.314 0.35 0.000 0.349 19.993 1.26782 0.00355 0.88854

10 2

Function evaluations
9
Best found value 1.8

8 1.6

7 1.4
Function value

Function value
6 1.2

5 1

4 0.8

3 0.6

2 0.4

1 0.2

0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Function evaluation Function evaluation

Figure 7.12: Optimization process

Figure 7.13: Optimized shape divided in panels

Table 7.15: Equivalent strains caused by assembly, panels numbered from left to right
Panel 1 Panel 2 Panel 3
Maximum 0.57% 0.08% 0.02%
Average 0.10% 0.02% 0.01%
7.3. SHAPE OPTIMIZATION USING THE FABRIC PLASTICITY MATERIAL MODEL 60

7.3.2 w1 = 0.9 and w2 = 0.1

(a) S11 (b) S22

Figure 7.14: Membrane stresses in the optimized shape using the fabric plasticity model (w1 =
0.9 and w2 = 0.1)

Table 7.16: Parameters and results (w1 = 0.9 and w2 = 0.1)


p1 (m) p2 (m) p3 (m) p4 (m) p5 (m) Fcable (kN) D1 D2 W
-0.0163 0.253 0.338 0.005 0.289 20.127 1.34759 0.06371 1.21920

10 2

Function evaluations
9
Best found value 1.8

8 1.6

7 1.4
Function value

Function value

6 1.2

5 1

4 0.8

3 0.6

2 0.4

1 0.2

0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Function evaluation Function evaluation

Figure 7.15: Optimization process

Using weight factors w1 = 0.9 and w2 = 0.1, a solution is obtained that is dominated by the
previous solution. While this configuration is inferior to the previous solution regarding
7.3. SHAPE OPTIMIZATION USING THE FABRIC PLASTICITY MATERIAL MODEL 61

Figure 7.16: Optimized shape divided in panels

Table 7.17: Equivalent strains caused by assembly, panels numbered from left to right
Panel 1 Panel 2 Panel 3
Maximum 0.19% 0.15% 0.03%
Average 0.04% 0.03% 0.02%

stress and force deviations, deformations caused by assembling the panels are lower, causing
the result to be more accurate.
7.3. SHAPE OPTIMIZATION USING THE FABRIC PLASTICITY MATERIAL MODEL 62

7.3.3 w1 = 0.95 and w2 = 0.05


The solution obtained by using weight factors w1 = 0.95 and w2 = 0.05 is a Pareto solution,
as D1 is lower compared to the first solution, while D2 is higher. Again, the deformations
caused by assembling the panels are negligible.

(a) S11 (b) S22

Figure 7.17: Membrane stresses in the optimized shape using the fabric plasticity model (w1 =
0.95 and w2 = 0.05)

Table 7.18: Parameters and results (w1 = 0.95 and w2 = 0.05)


p1 (m) p2 (m) p3 (m) p4 (m) p5 (m) Fcable (kN) D1 D2 W
-0.035 0.353 0.324 0.006 0.264 21.949 1.22046 0.97395 1.20813

10 2

Function evaluations
9
Best found value 1.8

8 1.6

7 1.4
Function value

Function value

6 1.2

5 1

4 0.8

3 0.6

2 0.4

1 0.2

0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Function evaluation Function evaluation

Figure 7.18: Optimization process


7.3. SHAPE OPTIMIZATION USING THE FABRIC PLASTICITY MATERIAL MODEL 63

Figure 7.19: Optimized shape divided in panels

Table 7.19: Equivalent strains caused by assembly, panels numbered from left to right
Panel 1 Panel 2 Panel 3
Maximum 0.29% 0.18% 0.02%
Average 0.06% 0.04% 0.01%

7.3.4 Conclusion
Using the fabric plasticity material model, the method again provides good results, not
only for the stress distribution in the membrane, but also for the force in the cable and
the deformation due to flattening. While both the first and the third solution are Pareto
solution, the latter is more accurate as the panels deform the least during assembly. The
difference in material model manifests itself in the shape of the patterns, as seen in figure
7.20.

However, the results are inferior to those of the orthotropic elasticity model, as the so-
lutions contain large zones where S22 is quite low. With the chosen design model, all
possible configurations are relatively rotationally symmetric: when turning the shape 90
degrees about its Z-axis, a similar shape appears. As the fabric plasticity model is highly
anisotropic, this explains the regions where S22 is low. Further improvements can be made
by preventing this rotational symmetry, by introducing more shape parameters.
7.3. SHAPE OPTIMIZATION USING THE FABRIC PLASTICITY MATERIAL MODEL 64

Figure 7.20: Difference between the panels when using the orthotropic elasticity model (blue) or
the fabric plasticity model (red), when using weight factors w1 = 0.7 and w2 = 0.3
65

Chapter 8

Conclusion and recommendations for


future work

8.1 Conclusion
Compared to previous work, some improvements are made regarding the design method.
As more parameters were used to describe the intermediate configuration, better stress
distributions can be found. The Bayesian algorithm also provides faster and better results
than the genetic algorithm used in previous work. The obtained results are able to be split
into panels and flattened without substantial deformations, and are therefore considered
reliable.

8.2 Recommendations for future work


Some recommendations can be made for future work regarding this design method. For
starters, only parametric optimization was considered, and non-parametric optimization
might also provide a solution. Regarding parametric optimization, one could use more than
the current number (5) of parameters to describe the intermediate shape of the membrane.
Due to restrictions regarding the SolidWorks API, parallellization was not possible. If this
can be resolved, or if other CAD software is used to create the design model, a faster
conclusion can be obtained.

A myriad of optimization algorithms exists, and only a few of those have been considered
for this specific shape optimization. Therefore it is reasonable to believe that a better
optimization algorithm exists for this purpose. While a new algorithm might provide a
8.2. RECOMMENDATIONS FOR FUTURE WORK 66

faster solution, it might also return solutions that perform better than the results from
this study.

A next step in developing the design method can be to integrate the structural analysis.
Similarly, an intermediate shape can be prescribed by parameters, followed by its erec-
tion. If then the environmental forces such as wind and snow are considered, the shape
optimization could prevent rupturing of the fabric due to these loads.
BIBLIOGRAPHY 67

Bibliography

[1] Fujikake Masahisa, Kojima Osamu, and Fukushima Seiichiro. Analysis of fabric ten-
sion structures. Computers & Structures, 32(3-4):537–547, 1989.

[2] J. Llorens. Fabric Structures in Architecture. Elsevier Science, 1st edition, 2015.

[3] Greg Amato. History of the Roller Shade. http://www.bayscreens.net/blog/bid/


87389/History-of-the-Roller-Shade, 2013.

[4] Association Athlétique Allaudienne 3A Olympiastadion München. http:


//www.as-allaudienne.com/compte-rendu-de-course/marathon-de-munich/
attachment/img{_}0972-copier/.

[5] Snipview. Expo Axis. http://www.snipview.com/q/Expo{_}Axis.

[6] Jan Belis. Spatial structures - Ruimtelijke constructies. Department of Structural


Engineering-Ghent University, 2014.

[7] John S. Brew and Wanda J. Lewis. Spline-based and stress-monitored patterning of
fabric structures. Computers and Structures, 119:203–214, 2013.

[8] D. Vandepitte. Berekening van Constructies - Bouwkunde en Civiele Techniek. Weten-


schappelijke uitgeverij E.Story-Scientia, 1982.

[9] Kai-Uwe Bletzinger, Roland Wüchner, Fernaß Daoud, and Natalia Camprubı́. Com-
putational methods for form finding and optimization of shells and membranes. Com-
puter Methods in Applied Mechanics and Engineering, 194(30-33):3438–3452, 2005.

[10] B Tabarrok and Z. Qin. Form Finding and Cutting Pattern Generation for Fabric
Tension Structures. Microcomputers in Civil Engineering, 8:377–384, 1993.

[11] Shon Su-deok, L E E Seung-jae, and L E E Kang-guk. Smooth cutting pattern


generation technique for membrane structures using geodesic line on subplane and
spline interpolation. pages 3131–3141, 2013.
BIBLIOGRAPHY 68

[12] J.F. Abel, R.B. Haber, D.C. Salmon, and M.L. Brown. Development of computer
aided-design of cable-reinforced membrane structures in the U.S.A. Shells, Membrane
and Space Frames, 2:111–118, 1986.

[13] Jae-yeol Kim and Jang-bog Lee. A new technique for optimum cutting pattern gen-
eration of membrane structures. Engineering Structures, 24:745–756, 2002.

[14] J. B. Pargana, D. Lloyd-Smith, and B. a. Izzuddin. Fully integrated design and


analysis of Tensioned Fabric Structures: Finite elements and case studies. Engineering
Structures, 32(4):1054–1068, 2010.

[15] T. D. Dinh, A. Rezaei, L. De Laet, M. Mollaert, D. Van Helmelrijck, and W. Van


Paepegem. A shape optimization approach to integrated design and nonlinear analysis
of cable reinforced fabric membrane structures. International Journal of Solids and
Structures, 83:114–125, 2016.

[16] Michael Seidel. Tensile Surface Structures: A Practical Guide to Cable and Membrane
Construction. 2009.

[17] R Houtman and M Orpana. Materials for Membrane Structures. Bauen mit Textilen,
4, 2000.

[18] B Bridgens and M Birchall. Form and function : The significance of material properties
in the design of tensile fabric structures. Engineering Structures, 44:1–12, 2012.

[19] T.D. Dinh, A. Rezaei, L. De Laet, M. Mollaert, D. Van Hemelrijck, and W. Van
Paepegem. A new elasto-plastic material model for coated fabric. Engineering Struc-
tures, 71:222–233, 2014.

[20] Achille Messac. Optimization in Practice with MATLAB: For Engineering Students
and Professionals. Cambridge University Press.

[21] Joakim Agnarsson, Mikael Sunde, and Inna Ermilova. Parallel Optimization in Mat-
lab. (January), 2013.

[22] S. Rajeev and C. S. Krishnamoorthy. Discrete optimization of structures using genetic


algorithms. 118(5):1233–1250, 1992.

[23] Virginia Torczon. On the convergence of pattern search algorithms. 7(1):1–25, 1997.
BIBLIOGRAPHY 69

[24] Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical Bayesian optimization
of machine learning algorithms. pages 1–12, 2001.

[25] Eric Brochu, Vlad M Cora, and Nando De Freitas. A Tutorial on Bayesian Optimiza-
tion of Expensive Cost Functions , with Application to Active User Modeling and
Hierarchical Reinforcement Learning. 2010.

[26] Comparison of five solvers. http://nl.mathworks.com/help/gads/


example-comparing-several-solvers.html. Accessed: 2016-04-17.

[27] Helmut Harbrecht. Analytical and Numerical Methods in Shape Optimization. (305),
2006.

[28] Wolfgang A Wall, Moritz A Frenzel, and Christian Cyron. Isogeometric struc-
tural shape optimization. Computer Methods in Applied Mechanics and Engineering,
197(40):2976–2988, 2008.

[29] V. Braibant and C. Fleury. Shape optimal design using B-splines. 1984.

[30] K Bletzinger and E Ramm. Form Finding of Shells by Structural Optimization. pages
27–35, 1993.

[31] R Meske, J Sauter, and E Schnack. Nonparametric gradient-less shape optimization


for real-world applications. Structural and Multidisciplinary Optimization, 30(3):201–
218, 2005.

[32] S Y Woon, O M Querin, and G P Steven. Structural application of a shape optimiza-


tion method based on a genetic algorithm. (1975):57–64, 2001.

[33] Benoit P Gilbert, Lip H Teh, and Hong Guan. Thin-Walled Structures Self-shape
optimisation principles : Optimisation of section capacity for thin-walled profiles.
Thin Walled Structures, 60:194–204, 2012.

[34] W Annicchiarico and M Cerrolaza. Structural shape optimization 3D finite-element


models based on genetic algorithms and geometric modeling. 37:403–415, 2001.

[35] Heinz Neuber. Kerbspannungslehre, Grundlagen für genaue Festigkeitsberechnung


mit Berücksichtigung von Konstruktionsform und Werkstoff. Springer-Verlag Berlin
Heidelberg GmbH, 1958.
BIBLIOGRAPHY 70

[36] OC Zienkiewicz and JS Campbell. Shape optimization and sequential linear program-
ming. Optimum structural design, pages 109–126, 1973.

[37] J A Bennett and M E Botkin. Structural Shape Optimization with Geometric De-
scription and Adaptive Mesh Refinement. 23(3):458–464, 1984.

[38] M. Hasan Imam. Three-dimensional shape optimization. 18(December 1980):661–673,


1982.

[39] Kenji Kawaguchi, Leslie Pack Kaelbling, and Tomás Lozano-Pérez. Bayesian Op-
timization with Exponential Convergence. In C Cortes, N D Lawrence, D D Lee,
M Sugiyama, and R Garnett, editors, Advances in Neural Information Processing
Systems 28, pages 2809–2817. Curran Associates, Inc., 2015.

You might also like