Geometric constraints for shape and

topology optimization in architectural

design

Charles Dapogny, Alexis Faure, Georgios

Michailidis, Grégoire Allaire, Agnes
Couvelas & Rafael Estevez

Computational Mechanics
Solids, Fluids, Structures, Fluid-
Structure Interactions, Biomechanics,
Micromechanics, Multiscale Mechanics,
Materials, Constitutive Modeling,
Nonlinear Mechanics, Aerodynamics

ISSN 0178-7675

Comput Mech
DOI 10.1007/s00466-017-1383-6

1 23
Your article is protected by copyright and
all rights are held exclusively by Springer-
Verlag Berlin Heidelberg. This e-offprint is
for personal use only and shall not be self-
archived in electronic repositories. If you wish
to self-archive your article, please use the
accepted manuscript version for posting on
your own website. You may further deposit
the accepted manuscript version in any
repository, provided it is only made publicly
available 12 months after official publication
or later and provided acknowledgement is
given to the original source of publication
and a link is inserted to the published article
on Springer's website. The link must be
accompanied by the following text: "The final
publication is available at link.springer.com”.

1 23
 Author's personal copy
Comput Mech
DOI 10.1007/s00466-017-1383-6

ORIGINAL PAPER

Geometric constraints for shape and topology optimization in

architectural design
Charles Dapogny1 · Alexis Faure2 · Georgios Michailidis2 · Grégoire Allaire3 ·
Agnes Couvelas4 · Rafael Estevez2

Received: 24 August 2016 / Accepted: 30 January 2017

© Springer-Verlag Berlin Heidelberg 2017

Abstract This work proposes a shape and topology opti- Numerical examples are discussed in two and three space
mization framework oriented towards conceptual architec- dimensions.
tural design. A particular emphasis is put on the possibility
for the user to interfere on the optimization process by supply- Keywords Numerical methods · Optimization · Optimum
ing information about his personal taste. More precisely, we shape · Architecture
formulate three novel constraints on the geometry of shapes;
while the first two are mainly related to aesthetics, the third
one may also be used to handle several fabrication issues
that are of special interest in the device of civil structures. 1 Introduction
The common mathematical ingredient to all three models is
the signed distance function to a domain, and its sensitivity Designing shapes in an ‘optimal’ way is a task of utmost
analysis with respect to perturbations of this domain; in the interest in a wide variety of applications. Architects, design-
present work, this material is extended to the case where the ers, artists and engineers from various fields and with diverse
ambient space is equipped with an anisotropic metric tensor. sensitivities endeavour to create products and structures that
suit their purposes in the best possible way, while meeting
B Charles Dapogny some specifications at the same time. Consciously or not,
[email protected] they try to solve an ‘optimization problem’, relying on their
Alexis Faure knowledge, experience and intuition.
[email protected] Shape and topology optimization [2,14,22] is a discipline
Georgios Michailidis which combines mathematical and computational techniques
[email protected] in order to find the shape of a body or a structure as the
Grégoire Allaire solution to an optimization problem. It has aroused a consid-
[email protected] erable enthusiasm in engineering, where the design criteria
Agnes Couvelas usually lend themselves to mathematical formulations, and
[email protected] where engineers are already quite familiar with computa-
Rafael Estevez tional methods. The great headway made in computer science
[email protected] during the last decades and the subsequent soar in computa-
1
tional capacity made it possible to apply such techniques to
Laboratoire Jean Kuntzmann, CNRS, Université
Grenoble-Alpes, BP 53, 38041 Grenoble Cedex 9, France
the typical design problems in that field. Applications rang-
ing from nanoscience to airplanes [59] have been reported in
2 SIMaP, Université Grenoble-Alpes, BP 53, 38041 Grenoble the literature, where complex problems are tackled, in which
Cedex 9, France intuition and experience are limited. Nowadays, several
3 CMAP, Ecole polytechnique, CNRS, Université Paris-Saclay, dedicated commercial softwares exist (OptiStruct, TOSCA,
91128 Palaiseau, France GENESIS, etc.), which are extensively used in industry to
4 SHAPE IKE, Kritis 13, 14231 Nea Ionia, Attiki, Greece accelerate the design process.

123
 Author's personal copy
Comput Mech

On the contrary, shape and topology optimization tech- In this situation, we introduce three geometric constraints;
niques have hardly ever been applied in architecture. Exclud- two of them are aimed at alleviating the prejudicial lack of
ing the fact that scientific computing has only recently started information related to any user-defined notion of aesthet-
being acknowledged by architects as a valuable help in ics in shape and topology optimization problems, while the
design, there are mainly three sources of explanation for this third one allows among other things to control the visibility
fact: of structures, and to penalize the length of their bars. The
common point between these shape functionals is that they
bring into play the notions of signed distance function and
– Architects generally follow construction rules, based
maximum thickness, either in the usual Euclidean setting (as
on the long-standing intuition and mechanical under-
in, e.g. [31] of [6]), or in the anisotropic one, which is new to
standing of the discipline, to apprehend the physical
the best of our knowledge. The impact of these constraints
situation of a problem, and the multiple manufactur-
is discussed in the perspective of computational mechanics,
ing constraints at stake. Except in ‘exceptional cases’,
and especially architectural design, even though they could
where, for instance, the aesthetic quality of the designed
(and for some of them, already have) find an interest in totally
structure is the main priority, or where this structure is
different contexts.
‘unique’, by nature or by size (e.g. skyscrapers), chal-
The remainder of this article is organized as follows: in
lenging these construction rules proves costly, lengthy,
Sect. 2 we explain the sense in which we intend to apply shape
and eventually inefficient.
and topology optimization techniques to conceptual architec-
– Although shape and topology optimization results are
tural design. After discussing the limitations of the existing
often directly exploitable in the field of mechanical engi-
approaches, we present in Sect. 3 the shape and topology opti-
neering, this is untrue when it comes to architecture.
mization framework adopted in this work to deal with civil
Indeed, most of the available methods deal with continu-
structures: in particular, the mechanical model for structures,
ous media, while architects and civil engineers typically
the mathematical method for evaluating the shape sensitivity
simulate their structures using beams, plates, shells, etc.
of functions and the numerical representation of shapes are
– Perhaps the main point lies in that architectural design
detailed. In Sect. 4, we recall some fairly classical material
is mostly driven by aesthetics and is influenced by
about the signed distance function and the unit normal vector
the designer’s inspiration, which is not always easily
to a domain; these notions are then extended to the case where
amenable to a mathematical formulation. As a result,
the ambient space is equipped with a (constant), anisotropic
the architect’s intervention is limited to the mechanical
metric tensor. In Sect. 5, we formulate and analyze three
inputs of the optimization problem, which has a major
shape functionals of a geometric nature, which allow to add
drawback: the results can be reproduced by anyone mak-
information about the personal taste of the user regarding aes-
ing the same choices, i.e. the notion of personal creation
thetics into the optimization problem, and to model important
disappears!
constraints on civil structures, e.g. constraints over the vis-
ibility of shapes, or over the length of their bars in some
The main purpose of this article is to propose a shape and particular directions. In Sect. 6, numerical results are pre-
topology optimization framework which is oriented towards sented in two and three space dimensions to appraise the
applications in architecture. Depending on the relevance in main features of the proposed shape and topology optimiza-
the particular situation, the material of this work may serve tion framework. This article ends with a general conclusion
to improve the mechanical performance of the designed in Sect. 7, and with a technical appendix where all the math-
structure, or to influence the optimization process with an ematical details underlying the main issues of this article are
information about the personal taste of the architect. concentrated.
From the mathematical viewpoint, structures are described
as linearly elastic shapes, and their mechanical performance
is evaluated in terms of their compliance, which is to be min- 2 Shape and topology optimization for conceptual
imized under a volume constraint. This minimization uses a architectural design
numerical representation of shapes via the level set method
[7,8,49,57,64]; the clear description of shapes allowed by 2.1 Motivations of this work
this technique is pivotal in accounting for their geomet-
ric attributes (normal vector, principal curvatures, etc.), as Admittedly, one should not expect such a universal appli-
opposed to density-based methods [14] where the same can cability of shape and topology optimization in the field of
be done only at the price of an additional approximate recon- architectural design as in that of mechanical parts in high-
struction of the shape from the material density (see [25] and technology industry: as we have mentioned, the nature of
references therein). the mechanical simulations involved when dealing with civil

123
 Author's personal copy
Comput Mech

structures, and the generally overwhelming amount of man- All in all, our belief is that both engineers and archi-
ufacturing constraints limit the range of application of shape tects may benefit from such an approach, converging faster
and topology optimization to exceptional structures, where in in a proposal of high aesthetic quality and firm structural
particular, the total construction budget is not the main con- coherence: as hinted at in [33], an optimized shape can be
cern. However, there are certainly two purposes for which regarded as a convergence point between architecture and
shape and topology optimization techniques may be a valu- engineering, helping designers to shift from a typological to
able help during the stage of conceptual design: exploring a performance-driven design [16,35].
new aesthetic concepts which are mechanically viable, and
enforcing geometric constraints on the devised structures 2.2 Shape and topology optimization studies in
(thus accelerating the interactions between architects and architecture: pioneer and modern contributions
engineers).
Let us comment on the first point. Shape optimization Although modern techniques of optimal design were not
of structures may help in several directions to get inter- available at that time, perhaps the first famous example of
esting designs from an aesthetical point of view. Several shape and topology optimization in architecture dates back
architects indeed argue that an optimized form, e.g. a body to the French architect Gustave Eiffel [62]: indeed, the outline
bearing a load with the minimum possible material or max- of the Eiffel tower in Paris is similar to the optimal shape of
imizing some performance criterion for a certain material a simple cantilever when it is subjected to gravity and wind
volume, enhances its proper aesthetic qualities, as it reveals loads. Moreover, its topology, ordered over three different
the trajectory of loads in the structure, eliminating any under- levels, is strongly inspired by patterns commonly found in
performing part [46]. The visual impact of such a shape acts nature.
in an educational manner, helping the observer to understand The first applications of modern shape and topology opti-
how the structure works mechanically. mization techniques in architecture have been presented by
Shapes obtained as the result of a shape and topology Sasaki [55]. Sasaki used the so-called Bi-directional Evo-
optimization procedure are frequently characterized by a lutionary Structural Optimization (BESO) method in an
strong organic nature [33]. This is not a coincidence, since international competition for the design of the Florence sta-
such shapes exhibit an optimal mechanical response to the tion in 2002. Grossly speaking, he searched for the design
applied loads, in a similar way as those found in nature. which would achieve maximum mechanical efficiency and
One famous example of shape optimization in nature con- minimum use of material, imposing two non-optimizable
cerns the evolution of trees [13,43], which is guided by the regions for the slab and the entrance. The resulting opti-
necessity to withstand wind and to maximize their capacity mal design avoids bending forces and achieves an iso-stress
to collect nutrients. Another well-known example is bone distribution in the structure, where members work either in
remodelling [66], i.e. the process under which bones change tension or in compression [41]. The same method was used
their shape and constitution according to the loads they have in the design of the entrance of the Qatar National Con-
to bear. Although forces in nature act in a complex way, vention Centre, in Doha (see Fig. 1). The final design is
the principle remains the same [55] and boils down to the 250 m long by 30 m wide by 20 m high; it shows a dominant
minimization of some potential energy. Therefore, shape and tree-like structure constructed via double tubes supported by
topology optimization methods turn out to be a natural way
to mimic the rules of biological growth and produce non-
intuitive shapes. Let us mention that famous architects, such
as Antoni Gaudi and Frei Otto, have extensively used such
models in their works, either by directly taking inspiration
from natural shapes or by performing simplified experiments
under gravitational loads [41].
In addition to this ‘natural’ aesthetic behavior of mechani-
cally optimal shapes, shape and topology optimization allows
to devise optimization criteria to help in improving their
geometric features (for instance the curvature of their bound-
aries), see Sect. 5.2.
Beyond aesthetic motivations, shape and topology opti-
mization allows to model important geometric constraints,
for instance visibility constraints, which are pivotal in the
study of structures exhibiting a non-standard mechanical Fig. 1 Front facade of the Qatar National Convention Center, in Doha
behavior, such as skyscrapers [15]. (taken from [65])

123
 Author's personal copy
Comput Mech

steel bars, which is strongly reminiscent of typical outlines cant progress has been made in this direction, via techniques
obtained in shape and topology optimization. Sasaki further such as CNC-milling, HWC, or even additive manufacturing
explored geometric optimization techniques in his work, i.e. methods, they are still far from being practical for common
techniques for optimizing the shape of a structure, starting applications. On the contrary, when shape and topology opti-
from an initial design, without affecting its topology. Two mization is applied for conceptual design, i.e. when its results
famous examples presented in [55] are the Island City Cen- serve as a guide for novel shapes or efficient structural sys-
tral Park GrinGrin and the Crematorium in Kakamigahara, tems, this limitation is of minor concern.
which achieve optimal mechanical efficiency under stress Concerning the second class of limitations, the use of
and displacement constraints. a simplified mechanical description of the real mechanical
Since then, several studies have relied on shape and problem is legitimated by several arguments. Firstly, when
topology optimization techniques for the design of struc- shape and topology optimization is intended for conceptual
tures, most of them employing density-based methods. In design, the obtained results are likely to undergo a seri-
[33], the authors use a commercial software to optimize ous post-treatment. Thus, considering more sophisticated
pre-fabricated, post-tensioned fiber-reinforced concrete ele- mechanical descriptions, as in [9,10] (for different appli-
ments. In that work, the optimization problem only considers cations however), would only complicate the optimization
the mechanical behavior of the structure—the deformation problem unreasonably. Furthermore, when aesthetics is the
energy is minimized under manufacturing constraints linked principal interest, it is quite reasonable to slip in the intuitive
to the CNC-milling process—and aesthetic aspects are con- framework where simple gravity loads occur and to consider
sidered implicitly as a post-processing issue, by imposing the compliance as the objective function. Admittedly when
a subsequent constraint on the already optimized design, so the performance of the structural system is at the heart of the
as to drive it towards an organic shape. Similar techniques problem, a more complete model including e.g. a possible
are used in [53] to optimize the topology of concrete shel- material anisotropy, i.e. different mechanical behaviors under
ters, in [46] for concrete shell structures, as well as in [35] tension and compression, or even multiple phases could be
where the so-called Robotic Hotwire Cutting (HWC) tech- beneficial [1,33].
nique is selected for processing the structure. Finally, in [36] Last but not least, let us comment on the shape and topol-
three examples of topology optimization using a commercial ogy optimization methodologies used so far in the field of
software are presented. The optimization employs a density- architecture. They either feature a density-based descrip-
based method and the various possible interpretations of the tion of structures, like the SIMP method [14], or rely on
optimized shapes are of particular interest. an evolutionary approach, like the BESO method [54], in
Eventually, let us mention the works [15] and [61], which elements of the Finite Element mesh are removed or
where examples using topology optimization in the ini- added according to a (usually heuristic) mechanical crite-
tial conceptual phase are presented and the importance of rion.
these techniques in obtaining knowledge about the structural In our opinion, a significant flaw of density-based meth-
behaviour of optimized systems is highlighted. ods as regards architectural applications is the loss of almost
any description of the shape, and handling geometric cri-
2.3 Limitations teria is then far from obvious. Recently, some works have
been presented in this direction, such as the consideration of
Let us now discuss the potential limitations of the use of shape flexible void areas in [24] and the application of projection
and topology optimization in architecture, some of which filters in [25] to simulate an interface. However, such meth-
have already arised from the above bibliography. These can ods are expected to be less specific and thus less efficient
be roughly classified into three categories with limitations than traditional geometric methods for shape optimization,
related to: in which an explicit description of structures is brought into
play. Another significant drawback of density-based methods
1. The practical realization of the obtained results. is the well-known difficulty of interpreting the final result:
2. The simplified mechanical framework considered. a classical shape is revealed only in case the resulting den-
3. The shape and topology optimization method. sity field from the optimization process takes values close
to 0 or 1. Although there is usually an inherent mecha-
The first class of limitations mainly comes up when the nism (as in the SIMP formulation) that favors this tendency,
result of the optimization process is intended to be produced numerical results very often contain a lot of intermediate
and used as such, or at least with minimal modifications. densities, as shown in [35,46]. Hence, an interpretation step
Then, the utility of shape and topology optimization meth- is necessary, and beyond the difficulty of such a step, it
ods is mostly restricted to art works, since it requires the use can cause in a significant loss of optimality for the final
of costly advanced fabrication techniques. Although signifi- shape.

123
 Author's personal copy
Comput Mech

Finally, in BESO methods, the heuristic criterion used for The considered shapes Ω are clamped on a region Γ D
the shape change can be proved efficient only for very sim- of their boundary; they are subjected to traction loads g ∈
ple criteria. In case of complicated functionals, where the L 2 (Γ N )d , exerted on a disjoint subset Γ N ⊂ ∂Ω, and body
gradient does not have a constant sign (as in the case of the forces f ∈ L 2 (Ω)d (accounting for instance for gravity).
compliance or the volume), such methods may not provide In this situation, the displacement u Ω of Ω is the unique
satisfying results. solution in the space
In a nutshell, we believe that there is room for improving  
the existing structural optimization methods in the perspec- HΓ1D (Ω)d := u ∈ H 1 (Ω)d , u = 0 on Γ D
tive of applications in architecture, notably by bringing into
play a precise geometric description of shapes. The frame- to the linear elasticity system:
work and geometric constraints introduced in this work are
proposals in this direction. ⎧
⎪
⎪ −div(Ae(u)) = f in Ω,
⎨
u=0 on Γ D ,
(1)
⎪ Ae(u)n = g
⎪ on Γ N ,
⎩
3 An architecture-oriented approach Ae(u)n = 0 on Γ,

In this section, we detail the basic theoretical and numerical where e(u) = 21 (∇u + ∇u T ) is the strain tensor associated
ingredients of the proposed shape and topology optimization to u.
method for architecture: in Sect. 3.1, we present the mathe- In this situation, we are interested in minimizing a criterion
matical model of linearized elasticity for the description of J (Ω) of the domain. One such criterion is the compliance
the physical behavior of our shapes; then, in Sect. 3.2, we C(Ω) of shapes, which is equivalently defined as the elastic
summarize Hadamard’s method for describing variations of energy stored in Ω, or as the work of external loads:
shape, and the inferred notion of shape derivative. Eventu-

ally, in Sect. 3.3, we recall the main features of the level set
method, used in the numerical representation of shapes and C(Ω) = Ae(u Ω ) : e(u Ω ) d x
Ω
their motions.  
Let us emphasize once more that the selection of such = f · uΩ d x + g · u Ω ds. (2)
Ω ΓN
‘geometric’ (theoretical and numerical) representations of
shapes—as opposed to the so-called density-based methods The optimization is carried out on the set Uad of admissible
for topology optimization—is motivated by our intention to shapes defined by
incorporate their geometric features (normal vector, etc.) into

the optimization problem: see the constraint functionals pro-
Uad = Ω ⊂ Rd is a smooth bounded domain,
posed in Sect. 5.
Γ D ∪ Γ N ⊂ ∂Ω} . (3)
3.1 Mechanical framework: compliance minimization in
linear elasticity In most applications, some constraints, modelled by a
functional P(Ω), should be imposed on shapes; there are
In this work, we limit ourselves with the model problem classically two ways for incorporating them in the optimiza-
of compliance minimization in the stationary context of lin- tion problem:
earized elasticity, which already reflects a great deal of the
complex physics involved. – One solution consists in penalizing the performance cri-
From the mathematical viewpoint, the considered struc- terion, in which case the optimization problem rewrites:
tures (or shapes) in the present work are bounded domains
Ω ⊂ Rd enjoying at least Lipschitz regularity (in our appli- min L(Ω), where L(Ω) := J (Ω) + m P(Ω),
Ω
cations, d = 2, 3). These shapes are filled with an isotropic
linear elastic material with Hooke’s tensor A given by: where m > 0 is a given weighting coefficient.
– They could alternatively be enforced as hard constraints,
∀e ∈ Sd (R), Ae = 2μe + λtr(e)I, so that the optimization problem reads:

min J (Ω),
where Sd (R) denotes the set of symmetric d × d matrices, Ω s.t. P(Ω)≤α
and λ, μ are the Lamé coefficients of the material, satisfying
μ > 0, and λ + 2μ/d > 0. where α is a maximum threshold for the values of P(Ω).

123
 Author's personal copy
Comput Mech

6
1
z

x y

6
1
z

x y

6
1
z

x y

Fig. 2 Three optimized designs of a three-dimensional bridge associated to different sets of fixations (see [12] for details)

The perhaps most obvious example of a constraint is associ- from [12]1 . In Fig. 2, we see the optimized shapes for three
ated to the volume function P(Ω) = Vol(Ω) := Ω d x; the optimization problems, where a uniform force is distributed
main purpose of this paper is to devise other penalty func- on the upper side of the design domain, and which differ only
tionals associated to geometric constraints which could find by the nature and locations of the imposed fixations. For the
applications in architectural design. upper case, the displacement is fixed at the two right corners
of the lower side, while the corners on the left side are free to
move in the x-direction. The optimized structural system has
Remark 1 In the above framework, we have considered that a clear physical interpretation: in the middle of the structure,
boundary conditions (that is, the fixations and loaded regions) straight bars in compression transfer the loads to an inverted
are predetermined, as is usually the case in applications of arc in tension. Then, two inclined bars in compression, as
mechanical engineering [see the definition (3) of admissible well as some horizontal bars in tension, transfer the loads to
shapes]. However, observing the change in the optimal shape
when various sets of fixations are used is of great interest for 1 The unpublished work carried out in this reference was obtained with
architects, since there is a certain latitude as for the choice of the collaboration of some of the authors, using the exact same algorithms
this set. Let us explain this point with an example extracted as those detailed in this article

123
 Author's personal copy
Comput Mech

Ωθ have to be restricted to a set Θad of admissible perturbations

defined by:
 
Θad = θ ∈ W 1,∞ (Rd , Rd ) smooth, θ = 0 on Γ D ∪ Γ N .

Recall that, according to the so-called Structure theorem

(see e.g. [40], §5.9, or [31], Chap. 9, Th. 3.6), the shape
derivative F  (Ω)(θ ) of a ‘large’ class of functions F(Ω)
only depends on the values of the normal component θ · n of
the considered perturbation on ∂Ω. Actually, as we shall see,
Ω
the shape derivatives of the considered integral functions of
θ the domain will enjoy the more precise structure

Fig. 3 Variations Ωθ of a shape Ω using Hadamard’s method 
F (Ω)(θ ) = vΩ θ · n ds, (5)
∂Ω
the fixations. Setting the same type of fixations all along the
where vΩ is a scalar field depending on F. For instance, it is
lower left and right-hand sides, the material in the optimized
well-known (see e.g. [8]) that when F(Ω) is the compliance
shape is mainly concentrated in the centre of the structure.
(2), one has:
Finally, imposing null displacement on both sides, the tractor
in the inferior part of the structure, accounting for the hori-
vΩ = −Ae(u Ω ) : e(u Ω ). (6)
zontal displacements is removed and the solution resembles
an arc in compression, with some vertical bars attached.
In the general case, from the expression (5), a descent direc-
tion for F(Ω) is revealed under the choice
3.2 Mathematical optimization framework:
perturbations of shapes and differentiation with
θ = −tvΩ n, (7)
respect to the domain
for a small, positive pseudo time-step t > 0.
Most optimization algorithms rely on the knowledge of
derivatives of the involved objective function. In the present
context, we aim at optimizing functionals of the domain. 3.3 Numerical framework: encoding shapes via the level
In this situation, several notions of differentiation are avail- set method
able and we rely on Hadamard’s boundary variation method
[2,40,45] (see [47] about the notion of topological deriva- Among the existing ‘geometric’ techniques for the numerical
tive): variations of a shape Ω are considered under the form representation of shapes in the perspective of their optimiza-
(see Fig. 3 for an illustration): tion, we have selected the ‘classical’ level set method on a
fixed mesh, due to its relative simplicity and robustness. To
Ωθ := (Id + θ )(Ω), θ ∈ W 1,∞ (Rd , Rd ), mention a few, alternative ‘geometric’ techniques for shape
and topology optimization represent designs e.g. by means of
||θ ||W 1,∞ (Rd ,Rd ) < 1.
a computational mesh, as in [5,23,51], or by a set of simple
primitives [38,69].
Definition 1 A function F(Ω) is shape differentiable at a The level set method was pioneered in [50], then intro-
particular shape Ω if the underlying mapping θ → F(Ωθ ), duced in the context of shape optimization in [7,8,49,57,64];
from a neighborhood of 0 in W 1,∞ (Rd , Rd ) into R, is see the monographs [48,56] for extensive discussions around
Fréchet differentiable at 0. The corresponding derivative the features of this method. The main idea is to describe a
θ → F  (Ω)(θ ) is the shape derivative of F at Ω, and the shape Ω ⊂ Rd via a scalar ‘level set’ function φ : Rd → R
following expansion holds: with the properties (see Fig. 4):
⎧
F(Ωθ ) = F(Ω) + F  (Ω)(θ ) + o(θ ), ⎨ φ(x) < 0 if x ∈ Ω,
|o(θ )| φ(x) = 0 if x ∈ ∂Ω, (8)
where lim = 0. (4) ⎩
θ→0 ||θ ||W 1,∞ (Rd ,Rd ) φ(x) > 0 if x ∈ c Ω.

Notice that, in general, so that perturbations Ωθ of admis- During the optimization process the shape Ω ≡ Ω(t)
sible shapes Ω ∈ Uad stay admissible, the vector fields θ evolves in (pseudo-)time with a motion driven by a normal

123
 Author's personal copy
Comput Mech

solved with the Finite Element method, which requires a dis-

cretization of Ω by means of a mesh. In our context, Ω is not
discretized exactly, and is only known via an associated level
set function φ, defined at the vertices of a (fixed) mesh of the
computational domain D. To get past this difficulty, we rely
on an approximation which is quite common in the treatment
of shape and topology optimization problems on a fixed mesh
(see e.g. [8,14]), namely the Ersatz material approximation:
the void region D \ Ω is filled with a soft material, i.e. (1)
may be accurately approximated by the same linear elastic-
ity system posed on the whole domain D, where the Hooke’s
tensor A is replaced by that Aε defined by:

A if x ∈ Ω,
Aε (x) =
ε A if x ∈ D \ Ω,

for a small value ε 1. This last system may be solved by

a standard Finite Element method.

4 Preliminaries about the signed distance function

In this section, we collect some material about the notions

of signed distance function and unit normal vector to a
domain Ω ⊂ Rd ; we also discuss their extensions to the
context where the ambient space is equipped with a Rieman-
nian structure, induced by an anisotropic metric tensor. This
material plays a fundamental role in the expression of many
geometric features of shapes (such as minimum and maxi-
mum length scales, see for instance [3,6,39,67]), in particular
in the definitions of our geometric constraints in Sect. 5,
Fig. 4 (Top) One shape Ω ⊂ R2 , and (bottom) the graph of an associ- The present section purposely stays at an elementary level,
ated level set function and all the underlying technical details (including mathemat-
ical proofs) are postponed to “Appendix”.
velocity field V (t, x), which stems from the shape deriva-
tive of the objective criterion under consideration, in the 4.1 The signed distance function
line of Sect. 3.2 (see (7). The motion translates in terms of
an associated level set function φ(t, x) as the well-known Let us start with the definition of the main notion of this
Hamilton–Jacobi equation: section.

Definition 2 Let Ω ⊂ Rd be a bounded, Lipschitz domain.

∂φ
+ V |∇φ| = 0, (9) The signed distance function dΩ : Rd → R to Ω is defined
∂t
as:
which can be solved, e.g., on a Cartesian grid of a large ⎧
computational box D, with an explicit second order upwind ⎨ −d(x, ∂Ω) if x ∈ Ω,
∀x ∈ Rd , dΩ (x) = 0 if x ∈ ∂Ω,
scheme [56], or on a simplicial mesh of D (i.e. composed of ⎩
triangles in 2d and tetrahedra in 3d) with a semi-Lagrangian d(x, ∂Ω) if x ∈ c Ω,
scheme [60].
where d(x, ∂Ω) = inf |x − y| is the usual Euclidean dis-
y∈∂Ω
Remark 2 In the shape and topology optimization context of tance function to ∂Ω.
interest in this article, the practical calculation of the veloc-
ity field V (t, x) cannot be performed exactly: its expression We now state several definitions and properties in connec-
(6–7) involves the elastic displacement u Ω , the solution of tion with the signed distance function. Even though they will
(1). This system is posed on the shape Ω, and is typically not be used as such in the sequel, they are very similar (yet

123
 Author's personal copy
Comput Mech

much simpler to understand) to their anisotropic counterparts n(p∂Ω (y))

addressed in Sect. 4.3, and they are key in the proofs of the
latter. Notice that we do not attempt to provide statements p∂Ω (y)
under minimal assumptions. Π∂Ω (x)

Definition 3 Let Ω ⊂ Rd be a bounded, Lipschitz domain. y

– For x ∈ Rd , the set of projections Π∂Ω (x) of x onto ∂Ω is x

defined by: Π∂Ω (x) = {y ∈ ∂Ω, |x − y|= d(x, ∂Ω)}; Σ
it is a non empty, closed subset of ∂Ω. When Π∂Ω (x) is
a singleton, its unique element is denoted as p∂Ω (x) and
is called the projection of x onto ∂Ω. z
Ω
– The skeleton Σ of Ω is the set of points x ∈ Rd where
dΩ2 is not differentiable; note in passing that Σ ∩∂Ω = ∅, Σ
and that Σ has null Lebesgue measure as a consequence ray∂Ω (z)
of Rademacher’s theorem.
– The ray of ∂Ω emerging from a point y ∈ ∂Ω is the set:
 
−1 Fig. 5 Illustration of Definition 3 and Proposition 1: projection set of
ray∂Ω (y) = p∂Ω (y) = x ∈ Rd , Π∂Ω (x) = {y} . a point x ∈ Σ, projection point p∂Ω (y) of y ∈ / Σ, and ray emerging
from z ∈ ∂Ω
The following proposition is a pot-pourri of several results
in [20,31,40] and [42], which are illustrated on Fig. 5.
fundamental form of ∂Ω at y, and κ1 (y), . . . , κd−1 (y) (resp.
Proposition 1 Let Ω ⊂ Rd be a bounded, Lipschitz domain. (e1 (y), . . . , ed−1 (y))) are the associated principal curva-
tures (res.p principal directions), obtained as the eigenvalues
(i) A point x ∈
/ ∂Ω has a unique projection over ∂Ω if and (resp. eigenvectors) of I I y .
only if x ∈
/ Σ. Then, dΩ is differentiable at x and its
gradient reads: (i) For any point x ∈ Rd , and p ∈ Π∂Ω (x), one has:

x − p∂Ω (x) −κi ( p)dΩ (x) ≤ 1, i = 1, . . . , d − 1. (12)

∇dΩ (x) = . (10)
dΩ (x)
(ii) The closure Σ of the skeleton Σ reads as the disjoint
In particular, |∇dΩ (x)|= 1 wherever it makes sense. union: Σ = Σ ∪ Γ , where Γ is the set of points x ∈ Rd
(ii) If Ω is of class C 1 , then for x ∈
/ ∂Ω, x ∈
/ Σ, where one of the inequalities in (12) is an equality.
(iii) The projection mapping p∂Ω is differentiable at any
∇dΩ (x) = n( p∂Ω (x)),
x ∈ Rd \ Σ and its derivative reads:
and x = p∂Ω (x) + dΩ (x)n( p∂Ω (x)). (11) ⎛ dΩ (x)κ1 ( p) ⎞
1− 1+dΩ (x)κ1 ( p) 0 ... 0
⎜ .. .. .. ⎟
(iii) If Ω is of class C 2 , the closure Σ also has null Lebesgue ⎜ . . ⎟
⎜ 0 .⎟
∇ p∂Ω (x) = ⎜ ⎟,
measure. Moreover, there exists a tubular neighborhood ⎜ .. .. dΩ (x)κd−1 ( p)
⎟
⎝ . . 1− 0⎠
U of ∂Ω such that Σ ∩ U = ∅. 1+dΩ (x)κd−1 ( p)
0 ... 0 0
(iv) Still in the case where Ω is of class C 2 , dΩ is of class
where p ≡ p∂Ω (x), (13)
C 2 on Rd \ Σ. Actually, dΩ is twice differentiable on
Rd \Σ, and (11) actually holds on Rd \Σ (in particular,
it holds in the vicinity of ∂Ω). where the above matrix is written in the orthonormal
basis (e1 ( p), . . . , ed−1 ( p), n( p)) of Rd .
The foregoing properties are completed with the following
proposition which describes the second-order derivatives of 4.2 The normal vector to a domain
dΩ in terms of the second-order character of Ω (see again
[20] and [3]). In the following, we equivalently denote by n Ω or n the unit
normal vector to ∂Ω, pointing outward Ω. When Ω is of
Proposition 2 Let Ω ⊂ Rd be a bounded domain of class class C 2 , Proposition 1 indicates that n Ω : ∂Ω → Sd−1 may
C 2 ; for y ∈ ∂Ω, we denote by Ty ∂Ω be the tangent plane to be extended to an open neighborhood U of ∂Ω (and actually
∂Ω at y; I I y : Ty ∂Ω × Ty ∂Ω → R stands for the second to Rd \ Σ) via the formula:

123
 Author's personal copy
Comput Mech

n Ω (x) = ∇dΩ (x) = n( p∂Ω (x)), x ∈ U.

This choice of an extension for the normal vector enjoys the

interesting symmetry property:

∇n Ω = ∇ 2 dΩ = ∇n Ω
T
. (14)

In the following, we will need to evaluate the sensitivity of

the normal vector n Ω with respect to the shape Ω; the main
result in this direction is the following (see for instance [45]):

Proposition 3 The normal vector Ω → n Ω has a material

derivative in the sense that the mapping θ → n Ωθ ◦ (Id + θ ),
from C 1,∞ (Rd , Rd ) into C(∂Ω) is Fréchet differentiable at
0. Its derivative n˙Ω (θ ) reads:

n˙Ω (θ ) = ∇θ T n Ω , n Ω n Ω − ∇θ T n Ω .

4.3 The anisotropic, signed distance function

The material of Sect. 4.1 may be extended to the anisotropic

context. Let M be a symmetric, positive definite d × d
matrix (we do not discuss the case of space-dependent met-
ric tensors, which is much more involved). In the anisotropic
context, the space Rd is equipped with the inner product
·, · M and associated norm |·| M defined as:

∀x, y ∈ Rd , x, y M = M x, y , and |x| M = x, x M ,

where ·, · is the usual Euclidean inner product.

Definition 4 Let Ω ⊂ Rd be a bounded, Lipschitz domain.

M to Ω reads:
The signed, anisotropic distance function dΩ
⎧
⎨ −d M (x, ∂Ω) if x ∈ Ω,
dΩ (x) = 0
M
if x ∈ ∂Ω,
⎩ M
d (x, ∂Ω) if x ∈c Ω̄,

where d M (x, ∂Ω) := inf y∈∂Ω |x − y| M .

Before going any further, let us consider an intuitive example

in the case of the two-dimensional shape depicted in Fig. 6
(top), and of the metric tensor, written in the canonical basis:
 
0.1 0
M= . (15)
0 1

Here, imposing a metric with a much lower coefficient

in the horizontal direction implies that, from a point x ∈
R2 , travelling until hitting ∂Ω ‘costs’ almost nothing (that
is, does not increase much the travelled distance), whereas Fig. 6 (Top) One shape Ω (in black); (middle) isolines of the isotropic
it does cost very much when the travel has some vertical distance function dΩ ; (bottom) isolines of the anisotropic distance func-
M associated to the tensor (15)
tion dΩ
component. Hence, in the vertical section of the cross, where

123
 Author's personal copy
Comput Mech

any point can be connected to a point of the boundary by an (iii) The projection p∂ΩM is differentiable on Rd \ Σ M and

horizontal travel, the anisotropic signed distance function is its derivative reads:
very low, whereas it is larger on the horizontal bar.
M
∇ p∂Ω (x)
4.3.1 Properties of the signed, anisotropic distance function ⎛ −1 ⎞
M − 21 − 21
⎜ I + |M − 21 n ( p)| dΩ (x)M I I p M
1
0⎟
In this section, we discuss the generalization of the properties =⎜ ⎝
Ω ⎟,
⎠
of Sect. 4.1 in the anisotropic case. This material is new to
0 0
the best of our knowledge, and we provide the proofs in
(18)
“Appendix”.
First, note that Definition 3 extends verbatim to the
anisotropic context. Hence, we shall speak about: where p ≡ p∂Ω
M (x), the latter matrix being written in

an orthonormal basis of Rd of the form

– The anisotropic projection set Π∂Ω M (x) (or point p M (x)
∂Ω  1 1 1

when it is a singleton) of a point x ∈ Rd , M 2 τ1 ( p) M 2 τd−1 ( p) M − 2 nΩ
,..., , ,
– The anisotropic skeleton Σ M of Ω, 1
|M 2 τ1 ( p)|
1 1
|M 2 τd−1 ( p)| |M − 2 n Ω ( p)|
M (x) of a point x ∈ Rd .
– The anistropic ray ray∂Ω
for any set of tangent vectors τi ( p) ∈ T p ∂Ω, i =
Proposition 4 Let Ω ⊂ Rd be a bounded, Lipschitz domain, 1, . . . , d − 1 satisfying the identity Mτi ( p), τ j ( p) =
and M be a symmetric, positive definite matrix. δi j .

(i) For x ∈/ ∂Ω, the set Π∂ΩM (x) is a singleton if and only if
4.3.2 Shape differentiation of the anisotropic, signed
x∈/ Σ M . Then, dΩM is differentiable at x and its gradient
distance function
reads:
  We now proceed to the shape differentiation properties of the
M x − p∂Ω M (x)
anisotropic signed distance function. Our first result is about
∇dΩ (x) =
M
M (x)
. (16) M:
dΩ the pointwise shape differentiability of dΩ

In particular, M −1 ∇dΩ M , ∇d M  = 1 holds wherever it Proposition 5 Let Ω ⊂ Rd be a bounded and Lipschitz

Ω
makes sense. domain, and let x ∈ Rd \ ∂Ω. The mapping (0, ∞)  t →
dΩM (x) ∈ R is differentiable at t = 0+ , and its derivative
(ii) If Ω is of class C 1 , then for x ∈
/ ∂Ω, x ∈ / ΣM, tθ
reads:
∇dΩ
M
(x) = n M ( p),
d  M  y − x, θ (y) M
and x = p + dΩ
M
(x)n M ( p), p ≡ p∂Ω (x), (17) dΩtθ (x) (0+ ) = inf M (x)
.
dt M (x)
y∈Π∂Ω dΩ

where the distorted normal n M (y) reads n M (y) :=

In particular, if Ω is of class C 1 , and x ∈
/ Σ M , the above
1
M −1 n(y),n(y)
M −1 n(y) (note that |n M (y)| M = 1.
formula rewrites:

We now assume Ω to be at least of class C 2 . d  M 

dΩtθ (x) (0+ ) = −n M ( p), θ ( p) M ,
dt
(i) The set Σ M has zero Lebesgue measure. Moreover, where p ≡ p∂Ω
M
(x).
there exists an open neighborhood U of ∂Ω such that
U ∩ Σ M = ∅, and (17) actually holds on Rd \ Σ M . We now come to the differentiation of integral functionals
(ii) For any point x ∈ Rd , the matrix M.
involving dΩ
 
1 M − 21 − 21 Proposition 6 Let Ω ⊂ Rd be a bounded domain of class
I+ dΩ (x)M I Ip M ,
|M − 21
n Ω ( p)| C 1 , contained in a larger domain D, and let j : R → R
be a function of class C 1 . Let J (Ω) be the shape functional
where p ≡ p∂Ω
M
(x),
defined by:
  
is positive semi-definite. Moreover, if x ∈
/ Σ M , this
matrix is positive definite. J (Ω) = M
j dΩ (x) d x.
D

123
 Author's personal copy
Comput Mech

Then J (Ω) is shape differentiable at Ω in the sense of

Gâteaux, and its shape derivative reads:

∀θ ∈ W 1,∞ (Rd , Rd ),
     1
J  (Ω)(θ ) = − j  dΩ
M
(x) n M p∂Ω
M
(x) ,
0.5
  D 1
θ p∂ΩM
(x)  M d x. 6
1
z

x y
5 Three geometric criteria for architectural
applications

In the shape and topology optimization literature, several

geometric constraints have been proposed, which may find
a great interest in architectural design (even though it seems
that such applications have not yet been investigated):

– Constraints on the maximum and minimum feature size

of shapes - hereafter referred to as ‘maximum thickness’
and ‘minimum thickness’—have notably been dealt with
Fig. 7 (Top) Reproduction of the definition of the test-case of Fig. 2
in [6,58]. Let us mention that whereas restricting the max-
modified by the addition of an ‘empty’ non-optimizable zone; (bottom)
imum thickness of shapes is of particular interest, mainly resulting optimal shape
due to fabrication limitations, controlling the minimum
feature size is rarely necessary for applications in archi-
tecture since the scale of civil structures is quite large. to a user-defined metric tensor M. This comes in handy
Penalizing small features could be used mainly to sim- when trying to improve the visibility of shapes, or in
plify the optimized design. an attempt to prevent the emergence of too elongated
– Another way to control the complexity of the shape is members in some direction.
to add a penalization over the total perimeter (see for
instance [8,68] on this topic), but the amount of person-
alized information offered by this ingredient is limited to These criteria are intended to be inserted either as penal-
the extent of the penalization. ization in the minimized function, or as constraints in the
– The possibility of imposing non-optimizable areas [24] optimization problem (see the discussion in the end of Sect.
leaves larger room for a personalized shape control. 3.1, and the examples of Sect. 6).
For instance, imposing fictitious holes in the working
domain (that is, areas where the optimized structure
cannot infringe) is quite simple and allows to explore 5.1 Requiring that shapes stay close to a specific design
different designs, using different strategies to bear the
imposed loads (see the example of Fig. 7, excerpted from The first functional of interest for our purpose aims at keeping
[12]). the proposed design Ω close to a reference shape ΩT . In fact,
if one tries to optimize the shape and topology of a structure
In the present work, we propose three novel geometric cri- starting from a reference design, nothing guarantees that the
teria that are very much suited for architectural applications: final optimized shape with ressemble the initial proposal.
Thus, one shall try to formulate a functional penalizing some
– In Sect. 5.1, we propose a shape functional Pm (Ω) sort of deviation from the reference domain.
which constrains shapes to stay ‘close’ to a user-specified Suppose we want Ω to resemble some specific design ΩT ,
design. supplied via its signed distance function dΩT . Morphing the
– In Sect. 5.2, we introduce a shape functional Pp (Ω) design domain Ω to the target domain ΩT can be achieved
which constrains Ω to conform to a user-specified pattern by minimizing the functional:
(or general trend) rather than strictly a design.

– In Sect. 5.3, a function PaM (Ω) is devised with the goal
to reduce the anisotropic thickness of shapes with respect Pm (Ω) = dΩT d x. (19)
Ω

123
 Author's personal copy
Comput Mech

The intuition behind (19) is that every region of Ω comprised ∇g

ng = .
inside ΩT (resp. inside c ΩT ) has a negative (resp. positive) |∇g|
contribution in Pm (Ω). Hence, the minimization of Pm (Ω)
As a result of the optimization process, we are interested in
drives the regions of Ω lying inside ΩT to expand, and those
that the normal vector n Ω of Ω be as close as possible to
lying outside ΩT to retract. In particular,
that of the pattern; this may be of interest either for aesthetic
purposes, or to drive the light reflection properties of Ω,
– If ΩT is connected, the unique local minimizer of Pm (Ω)
which are directly tied to the orientation of its boundary.
is ΩT itself,
The function Pp (Ω) of interest is then:
– In the general case where ΩT has several connected com-
ponents, say O1 , . . . , O N , the local minimizers of Pm (Ω) 

are exactly the sets i∈I Oi , where I ⊂ {1, . . . , N }, and Pp (Ω) = |n Ω − n g |2 ds. (21)
∂Ω
the unique global minimizer of Pm (Ω) is again ΩT .
The shape differentiability of Pp (Ω) is now studied in a
The shape derivative of (19) is easily calculated as (see slightly larger context—see “Details of the calculations of
e.g. [40]): the shape derivatives of Sect. 5” of Appendix section for a
proof of the forthcoming results.
Proposition 7 The functional Pm (Ω) defined by (19) is
shape differentiable at any shape Ω ∈ Uad , and its shape Proposition 8 Let P(Ω) be the shape functional defined by:
derivative reads: 
 P(Ω) = ϕ(n Ω ) ds, (22)
∂Ω
∀θ ∈ Θad , Pm (Ω)(θ ) = dΩT θ · n ds. (20)
∂Ω
where ϕ : Rd → R is a function of class C 1 . Then P(Ω)
Remark 3 is shape differentiable in the sense of Fréchet at any shape
Ω ∈ Uad , and its derivative reads:
– Functionals of the form (19), based on the signed distance
function have already been used in [29] (or prior in [11]) ∀θ ∈ Θad ,

in the contexts of shape morphing and shape registration.
P  (Ω)(θ ) = κ ϕ(n) θ · n ds
– Different functionals P(Ω) from (19) could be devised Γ

for the same shape matching purpose; for instance, one
− ∇ϕ(n) · ∇∂Ω (θ · n) ds,
may imagine using a least-square functional of the form: Γ

where ∇∂Ω f := ∇ f − ∇ f, n Ω n Ω is the tangential gradi-
P(Ω) = (dΩ − dΩT )2 d x,
D ent of a smooth enough function f : ∂Ω → R.

where D is a fixed computational domain. However, in In particular, we infer the following corollary:
our opinion, (19) has the nice feature to enjoy a very Corollary 1 The functional Pp (Ω) defined by (21) is shape
simple shape derivative; besides, least-square functionals differentiable at any shape Ω ∈ Uad , and its shape derivative
often cause conditionning issues in optimization prob- reads:
lems (see for instance [26]), and we prefer to avoid the 
latter insofar as possible. ∀θ ∈ Θad , Pp (Ω)(θ ) = 2 (κ − div∂Ω (n g ))θ · n ds,
∂Ω
5.2 Imposing a resemblance with a user-defined pattern
where div∂Ω V := div(V )−∇V n, n is the tangential diver-
genceof a smooth enough vector field V : ∂Ω → Rd .
We are now interested in imposing that the considered shapes
fit a user-defined pattern, defined throughout the ambient The result of Corollary 1 is quite intuitive: in order to min-
space Rd , rather than sticking to a shape, strictly speaking. As imize Pp (Ω), the mean curvature of ∂Ω has to be adjusted
an example, one may look for an optimized structure which to that of the pattern.
enjoys a wavy boundary (see the numerical illustration in
Sect. 6.2). 5.3 Penalization of the directional thickness of shapes
In our applications, the pattern is supplied as the 0 level
set of a given smooth function g : Rd → R. We assume In this section, we discuss a third class of constraints which
that the gradient ∇g does not vanish, so that the unit normal have interesting applications in architectural design, namely
vector n g to the level sets of g reads: constraints on the anisotropic maximum thickness of shapes.

123
 Author's personal copy
Comput Mech
  2
5.3.1 Mathematical formulation of anisotropic maximum P(Ω) = 0, where P(Ω) = M
dΩ − δ/2 d x, (23)
thickness constraints Ω −

where we denote by t− = max(0, −t) the negative part of a

Let us start with a rigorous account of the notion of (isotropic
real number t ∈ R.
or anisotropic) maximum thickness of shapes.
For reasons that are extensively detailed in [6,44], impos-
Definition 5 Let Ω ⊂ Rd be a bounded domain, M be a ing maximum thickness constraints with this kind of func-
symmetric, positive definite d × d matrix. Then, tional turns out to be too strict, and produces undesirable
numerical artifacts. The numerical experience reported in
– Ω has maximum thickness smaller than δ > 0 if: there suggests to impose this maximum anisotropic thick-
ness as a constraint in the optimization algorithm (see the
∀x ∈ Ω, dΩ (x) ≥ −δ/2. discussion in the end of Sect. 3.1) by using the following
relaxed version of (23):
– Ω has anisotropic maximum thickness smaller than δ > δ
0 if: PaM (Ω) ≤ ,
2
where
∀x ∈ Ω, dΩ
M
(x) ≥ −δ/2. ⎛ ⎞1
2
    2 ⎟
⎜ 1
Imposing constraints on the maximum thickness of shapes PaM (Ω) = ⎜
⎝
M
h dΩ M
dΩ dx⎟ .
⎠ (24)
has countless applications in mechanical engineering. It is, M
h(dΩ )d x Ω

for instance, a means to prevent the appearance of too thick Ω

regions, which could slow down the cooling process dur- In the above formula, h : R → R is a regularized character-
ing manufacturing; see [6] and references therein. However, istic function, defined as:
to our knowledge, constraints on the anisotropic maximum
  
thickness of shapes have not yet been considered, and could 1 |s| − δ/2
help in testing (then constraining) the thickness of shapes h(s) = 1 + tanh ,
2 α δ/2
in only one (or several) selected direction—for instance the
length of bars in certain directions, or the area of surfaces α > 0 being a parameter that controls the regularization of
with a prescribed orientation. the constraint. The interpretation of h and α is the following:
To appraise the difference between the isotropic and when α goes to 0, h converges to the function
anisotropic versions of the notion of maximum thickness,

consider again the example of the cross Ω, depicted in Fig. 1 if |s|> δ/2,
h ∗ (s) = ,
6, (left), when the anisotropic metric tensor M is defined 0 otherwise
by (15). In this case, the anisotropic signed distance func-
tion dΩM (x) measures approximately the closest horizontal a function that penalizes only regions where the maximum
distance from a point x to a region of the boundary ∂Ω. thickness constraint is violated. Of course, h ∗ is way too steep
Hence, if one aims at decreasing (in terms of absolute value) to be used in practice, and we use a ‘small’, positive value for
this value at any point x ∈ Ω, the solution (reflected by the α; our experience suggests the heuristic choice α = 25 Δx,
shape derivative of Proposition 5) consists in reducing the where Δx stands for the constant mesh size.
‘horizontal thickness’ of Ω. Hence, imposing a constraint on Remark 4 This idea of using a Riemannian, anisotropic
the anisotropic maximum thickness of Ω with respect to M structure to measure lengths as a means of constraining the
is a way to constrain the length of its horizontal bars. Notice features of geometric objects in different ways depending
that relying on the notion of isotropic maximum thickness in on their orientation in space has already been used in other
such a situation is useless to this purpose, since the isotropic contexts, notably in meshing; see e.g. [37,63].
maximum thickness of Ω is already small in this situation.
Let us then construct a shape functional for imposing a 5.3.2 Shape derivative of the constraint functional PaM (Ω)
bound δ > 0 on the anisotropic thickness of a shape Ω.
Ideally, the desired constraint reads: Let us now provide the main result concerning the shape
differentiability of PaM (Ω); its proof is postponed to Sect. 1.
∀x ∈ Ω, dΩ
M
(x) ≥ −δ/2,
Theorem 1 The constraint functional PaM (Ω) defined by
which can be equivalently formulated using an integral (24) is shape differentiable at any admissible shape Uad , and
penalty function: its shape derivative in a direction θ ∈ Θad reads:

123
 Author's personal copy
Comput Mech

PaM (Ω)(θ ) = DΩ θ · n ds, where violated, and almost zero values elsewhere. In this perspec-
Γ tive, (25) expresses a means to convey this information about
 
d−1 constraint violation from internal regions of Ω to the bound-
DΩ (x) = G(z) ary ∂Ω; in other terms, (25) uses the volumetric information
M (x)∩Ω
ray∂Ω

i=1
 encoded in G to predict how the boundary ∂Ω should be
dΩM (z) deformed to guarantee a better fulfillment of the constraint
1+ λi (x) d(z), (25) (24).
− 21
|M n(x)|
Another means to achieve this information transfer is
where d(z) stands for the one-dimensional Hausdorff mea- accounted for by (27), which may be understood as a
sure on the rays ray∂ΩM (x), the λ (x), i = 1, . . . , d − 1, are
i
preconditionner for the shape optimization algorithm. The
the eigenvalues of the (d − 1) × (d − 1) symmetric matrix information contained in G undergoes a diffusion process
1 1
M − 2 II∂Ω (x)M − 2 , and the integrand factor G(z) is given from the regions where it takes large values to the boundary
by the expression: of Ω. This process is chosen to be anisotropic so as to favor
   diffusion in the direction where the maximum thickness con-
− M )d x
h(dΩ h  (dΩ
M (z))d M (z)2 + h(d M (z))2d M (z)
Ω Ω Ω straint is prevailing. The coefficient αr is chosen so that the
Ω
G(z) =  2 diffusion rate is large enough for the information to hit the
2PaM (Ω) M )d x
h(dΩ boundary.
Ω
 
M M 2  M
h(dΩ )dΩ d x h (dΩ (z))
Ω
+  2 . (26)
M )d x
6 Numerical examples
2PaM (Ω) h(dΩ
Ω
In this last section, we present several numerical simulations
5.3.3 Practical implementation of the maximum thickness to appraise the applicability of the shape optimization frame-
constraint work introduced in Sect. 3 and the geometric functionals of
Sect. 5 in the field of architecture.
Unfortunately, the expression (25) is difficult to evaluate Our implementation differs slightly between the two- and
in numerical practice; taking inspiration from the ‘approxi- three-dimensional cases. In both situations, a large com-
mate shape derivative’ proposed in [6] to handle the isotropic putational box D is equipped with a fixed quadrilateral
maximum thickness constraint, we rely on a quite heuristic (Cartesian) mesh G, and each considered shape Ω is rep-
simplification. This simplification, which proves very effi- resented by means of a level set function φ (i.e. (8) holds)
cient in practice, builds upon the classical idea of changing which is discretized at the vertices of G. The (constrained
inner products to infer a descent direction θ from the knowl- or unconstrained) optimization process is driven by an SLP-
edge of the shape derivative of a functional (see e.g. [18,30]). type algorithm similar to that presented in [34], implemented
At first, an anisotropic diffusion equation is solved on in Scilab [19].
the whole computational domain D; i.e. one calculates the When it comes to the Finite Element analyses involved
solution Q ∈ H 1 (D) to the following variational problem: in the resolution of the linearized elasticity system (1), our
two-dimensional computations are performed using Scilab:
∀v ∈ H 1 (D), the Cartesian mesh G of D also serves as a computational
   
αr2 M −1 ∇ Q · ∇v + Qv d x = Gv d x, (27) support for the execution of the Q1 Finite Element method.
D Ω In the three-dimensional case, we use a tetrahedral mesh T
of D, together with P1 Finite Elements, in the FreeFem++
where G is given in (26), and the coefficient ar is set as environment [52]; interpolation procedures are used to trans-
  fer data about the level set function φ from G to T , and
maxΩ |dΩ (x)| conversely, to transfer data about the calculated elastic dis-
αr2 = 2 (Δx)2 .
Δx maxi (Mii−1 ) placement u Ω from T to G.
In both situations, the anisotropic signed distance function
A descent direction θ for the functional PaM (Ω) is then dΩM is calculated on G by solving an anisotropic counterpart

revealed as θ = −Qn. to the so-called redistancing equation (see e.g. [48] about
Let us comment shortly on the rationale behind this pro- the redistancing equation and [28] and references therein
cedure (see [6], section 7.3 for further details). The function about its anisotropic version), with a second-order numer-
G carries all the needed information about the violation of ical scheme [56].
the maximum thickness constraint (24); in particular, it takes Let us emphasize that the fixed Cartesian mesh G (and
significant values in the regions of Ω where this constraint is tetrahedral mesh T in the 3d setting) could be easily replaced

123
 Author's personal copy
Comput Mech

In our applications, the maximum authorized volume is

VT = 0.163|D|. Starting from the rather arbitrary initial-
20
ization shown in Fig. 9 (top), the optimized shape of Fig. 9
(bottom) is obtained, whose compliance equals 9347.
250 Let us now assume that, out of aesthetic purposes, we
would like that the optimal design resemble the target shape
Fig. 8 Description of the two-dimensional bridge test case of Sect. 6.1
ΩT represented on Fig. 10 (top), which uses the same
amount VT of material. The compliance of ΩT is very high
(C(ΩT ) = 48693), and we would like to improve its struc-
by a simplicial mesh of D, possibly changing in the course of tural performance, that is to optimize the shape Ω in view
the optimization process, so that it is consistently adapted to of its compliance so that it has the same material volume
the shape; see for instance [37] for related mesh adaptation |Ω|= VT , while staying ‘close’ to ΩT .
techniques. Unfortunately, solving (28) with ΩT as an initial guess
For all results, the considered elastic material is assumed gives no guarantee that the resulting optimized shape will
to be isotropic, with normalized Young modulus E = 1 and retain any resemblance with ΩT , as evidenced in Fig. 10 (bot-
Poisson’s ratio ν = 0.33. The ersatz material (simulating the tom). Although this final shape has a significantly improved
void) has Hooke’s tensor ε A, where ε = 1.10−3 . performance over ΩT (its compliance equals 12817), it is not
satisfactory from an aesthetics point of view.
In order to incorporate the desired aesthetic information,
6.1 Imposing shapes to stick to a user-defined design we trade (28) for the new optimization problem (29),

Our first example deals with the geometric constraint intro-

min L(Ω)
duced in Sect. 5.1, on the example of the optimal design of Ω
a two-dimensional bridge depicted on Fig. 8. The working s.t. Vol(Ω) ≤ VT ,
domain D has size 250 × 20; the considered structures are C(Ω) Pm (Ω)
where L(Ω) := t + (1 − t) (29)
clamped on two small regions on their lower side, no body C(ΩT ) Pm (ΩT )
forces apply (i.e. f = 0 in (1)), and a uniform, unit vertical
load g = (0, −1) is distributed on their upper side. Taking involving the penalization functional Pm (Ω) defined by (19),
advantage of the symmetry of the test-case, the calculation and a weighting coefficient t ∈ [0, 1]. Using the target shape
is performed on only one half of D. ΩT as initialization for the problem (29) and different val-
In a first experiment, we seek to minimize the compli- ues of t, the optimized shapes of Fig. 11 are obtained. As
ance of the considered structures, under a volume constraint, expected, for low values of t, the topology of ΩT is retained
namely: by the optimized shape, while increasing t accounts for a
smooth transition of optimal shapes from ΩT to the one of
Fig. 10 (bottom). The values of the compliance C(Ω) and of
min C(Ω)
Ω , where C(Ω) is the compliance (2). the penalization Pm (Ω) for the resulting optimal designs are
s.t. Vol(Ω) ≤ VT reported on Table 1, and the associated convergence histories
(28) are represented on Fig. 12.

Fig. 9 (Top) Initial shape, and

(bottom) resulting optimized
shape for the problem (28)

Fig. 10 (Top) ‘target’ domain

ΩT provided by the designer;
(bottom) optimized shape for the
problem (28) starting from ΩT

123
 Author's personal copy
Comput Mech

Fig. 11 Optimized shapes for

the problem (29) setting: a
t = 0.35, b t = 0.45, c
t = 0.55, d t = 0.70

Table 1 Values of the compliance C(Ω) and of the penalization Pm (Ω) In a first example, we simply aim at evaluating the penalty
for the optimal designs obtained in the 2d bridge example of Sect. 6.1 functional Pp (Ω) defined by (21): starting from the initializa-
Weight t 0.7 0.55 0.45 0.35 tion of Fig. 13 (middle), we solve the optimization problem
Compliance C(Ω) 12927 13108 11558 12864
min Pp (Ω). (31)
Value of Pm (Ω) 81.48 60.76 17.96 −39.68 Ω
The algorithm converges to a local minimum, displayed on
Fig. 13 (top-right). As indicated by the shape derivative of
t=0.70
7
t=0.55 Corollary 1, the mean curvature of the boundary at the opti-
t=0.45
t=0.35 mal shape equals that of the motif at each point. This means
6
that in regions where n · n g > 0, the algorithm tries to align
the exterior normal vector with n g , while if n ·n g < 0, a local
Objective function

5
minimum is achieved when n · n g = 0.
4 Let us now incorporate this aesthetic constraint in a struc-
tural optimization problem: the situation is that depicted on
3
Fig. 14. The shape Ω is enclosed in a computational box D
2 of dimensions 1 × 2, and is clamped at some part of its lower
side; a unit vertical load g = (0, −1) is distributed on its
1 upper side and no body forces occur, i.e. f = 0.
0 10 20 30 40 50 60 70 80 90 We aim at minimizing the compliance under a volume con-
Iterations
straint, and additionnally want to influence the shape with the
Fig. 12 Convergence histories for the optimization processes leading sinusoidal pattern (30) of Fig. 13 (left). To this end, we solve
to the shapes in Fig. 11: values of the objective function L(Ω) versus once more the optimization problem (29), replacing Pm (Ω)
number of iterations with Pp (Ω). The results for different values of the weighting
factor t are shown in Fig. 15, where the influence of the pat-
6.2 Constraining shapes to fit a user-defined pattern tern is obvious for decreasing values of t. The corresponding
values for the compliance C(Ω) and the penalty Pp (Ω) of
Let us now apply the material of Sect. 5.2 to influence struc- the optimized shapes are reported in Table 2.
tures with a user-specified pattern. The pattern used in this
example is given by the function g : D → R defined by (see 6.3 Illustrations of the constraint over the anisotropic
Fig. 13, top-left): maximum thickness of shapes

∀x = (x1 , x2 ) ∈ D, In this section, we now exemplify how the notion of

 
L1 2π x2 anisotropic maximum thickness introduced in Sect. 5.3 may
g(x) = x1 − sin , L 1 = 1, L 2 = 2. (30) be used in architectural applications. In all the examples
10 3L 2

123
 Author's personal copy
Comput Mech

Fig. 14 Boundary conditions in the mechanical test case of Sect. 6.2

Fig. 13 (Top-left) Isovalues of the ‘pattern’ function g defined in (30),

(top-right) initial shape, and (bottom) optimized shape for Problem (31)

below, the considered optimization problem features an addi-

tional constraint corresponding to a maximum authorized
value over the function PaM (Ω) defined in (24) (for a metric (a) (b) (c)
tensor M to be specified). Hence, the considered optimization
problem reads:

min C(Ω)
Ω

Vol(Ω) ≤ VT (32)
s.t.
PaM (Ω) ≤ δ/2,

where δ > 0 is a threshold depending on the particular situ-

ation.

6.3.1 Simplification of the shape of a three-dimensional (d) (e) (f)

bridge
Fig. 15 Optimized shapes for problem (29) setting: a t = 1.0,
b t = 0.90, c t = 0.60, d t = 0.45, e t = 0.35, f t = 0.15
We first show how structures may be simplified by enforc-
ing a control on their anisotropic maximum thickness. The

123
 Author's personal copy
Comput Mech

Table 2 Values of the compliance C(Ω) and of the penalization Pp (Ω) for the optimized designs obtained in the 2d pattern-fitting example of
Sect. 6.2

Weight t 0 0.1 0.4 0.55 0.65 0.85

Compliance C(Ω) 11.66 11.95 15.13 13.20 10.75 13.22
Value of Pp (Ω) 27.59 26.77 13.47 14.27 10.35 8.22

6
1
z

x y
(a) (b)

(c) (d)
Fig. 16 a Description of the optimal bridge test case, b optimized shape without adding any thickness constraints; (bottom) optimized shapes for
problem (32) setting: c m 1 = 0.1, d m 1 = 0.01

considered test case is that of the optimal bridge depicted on more classical truss-like shape composed of bars in tension
Fig. 2, which is reproduced on Fig. 16. or compression.
At first, a simple minimization of the compliance C(Ω) To help this purpose, we now solve Problem (32) with a
of the bridge under the volume constraint Vol(Ω) = VT is metric tensor M of the form
performed—i.e. (28) is solved—where the target volume is ⎛ ⎞
the fraction VT = 0.1|D| of that of the working domain D, m1 0 0
and D acts as an initial shape. One can see that the resulting M = ⎝ 0 m2 0 ⎠ , (33)
optimized shape Ω ∗ contains regions with large thickness 0 0 m3
in the (yz)-plane, although its isotropic maximum thickness
is low. The mechanical relevance of these parts is not clear where the coefficients m 1 , m 2 , m 3 are yet to be chosen. The
and shall be further interpreted by the designer, to result in a threshold parameter δ is taken as

123
 Author's personal copy
Comput Mech

Table 3 Values of the compliance of the optimized shapes of Fig. 16 9. 106

no thickness constraint
Test case Figure 16b Figure 16c Figure 16d 8. 106 m1=0.10
m1=0.01

Compliance 751 766 753 7. 106

Volume 0.1|D| 0.1|D| 0.1|D|

Compliance
6. 106

5. 106

4. 106

δ = 0.3 min {m 1 , m 2 , m 3 }, 3. 106

2. 106
and the optimal shapes associated to the values m 1 = 0.1,
then m 1 = 0.01, and m 2 = m 3 = 1 are displayed on Fig. 16. 0 10 20 30 40 50 60
The associated values of the compliance of the optimized Iterations
shapes are reported in Table 3, and the related convergence
no thickness constraint
histories are those of Fig. 17. m1=0.10
m1=0.01
One observes that the performance of all three results are
quite similar. The thick part has disappeared and the material 1

Volume violation
has been redistributed in the rest of the domain to retain
rigidity.
0.5
6.3.2 Improvement of the visibility of a three-dimensional
short cantilever

Let us now show how a constraint on the anisotropic maxi- 0

mum thickness may improve the visibility of shapes. 0 10 20 30 40 50 60
Consider the ‘short cantilever’ test case of Fig. 18 (left): Iterations
in a working domain D of dimensions 1 × 2 × 1, the shape is
clamped at its left side and two different loads g1 , g2 , defined 0.05 m1=0.10
m1=0.01
by: 0.04

0.03
g1 = (−1, 0, 0), and g2 = (0, −1, 0)
Thickness violation

0.02

are applied on the centre of its right-hand side. Again, no 0.01

body forces apply: f = 0. The mechanical criterion whose 0
minimization is sought is the total compliance S(Ω)
-0.01

S(Ω) = C 1 (Ω) + C 2 (Ω), -0.02


-0.03
where C (Ω) =
i
Ae(u iΩ ) : e(u iΩ ) d x, 0 10 20 30 40 50 60
Ω Iterations

and u iΩ is the elastic displacement of the shape when sub- Fig. 17 Evolution of (top) the compliance C(Ω), (middle) the viola-
tion Vol(Ω) − VT of the volume constraint, and (bottom) the violation
jected to the surface loads gi (i.e. the solution of (1) with g PaM (Ω) − δ/2 of the thickness constraint during the optimization of the
replaced by gi ). three-dimensional bridge examples of Sect. 6.3.1
First, we perform the minimization of S(Ω) under the
volume constraint Vol(Ω) = VT , VT = 0.2|D|, without
adding any constraint on the anisotropic maximum thickness
of shapes (i.e. (28) is solved by replacing C(Ω) with S(Ω)).
Doing so results in the optimized shape of Fig. 18 (right). maximum thickness and solve the optimization problem (32)
One may argue that this shape is not particularly interest- for δ = 0.024, by using different metric tensors of the form
ing, mainly due to the large bulk of material concentrated in (33). The optimized shapes appear in Fig. 19 and the corre-
the middle of the shape. Moreover, it is possible that the large sponding results in Table 4. Once more, the performance of
surface in the (x, y) plane hinders the light entering from the all structures is quite similar, although the shapes are quite
z-direction. Therefore, we add a constraint on the anisotropic different.

123
 Author's personal copy
Comput Mech

Fig. 18 a Description of the 3d short cantilever test case, b optimized shape without adding any thickness constraint

Fig. 19 Optimized shapes for the short cantilever test case with different metric tensors M in the definition of the anisotropic maximum thickness
a m 1 = 0.01, b m 1 = 0.01, m 2 = 0.01, c one constraint with m 1 = 0.01, m 2 = 1.0 and a second constraint with m 1 = 1.0, m 2 = 0.01,
d m 3 = 0.01

123
 Author's personal copy
Comput Mech

Table 4 Results for the

Test case Figure 18 (right) Figure 19a Figure 19b Figure 19c Figure 19d
optimized shapes of Figs. 18
and 19 Compliance 49 51 50 52 50
Volume 0.2|D| 0.2|D| 0.2|D| 0.2|D| 0.2|D|

g = 0.5

Fig. 22 Example of a junction between two HEA beams

3
Fig. 20 Setting of the MBB beam example of Sect. 6.3.3

Fig. 23 Optimized shapes in the MBB beam example with a constraint

on the anisotropic maximum thickness, and parameters (top) m 1 = 0.1;
(bottom) m 1 = 0.01

Fig. 21 (Top) Initial and (bottom) optimized shapes obtained in the

MBB beam example without adding any thickness constraint In a first experiment, we minimize the volume Vol(Ω) of
shapes, under a compliance constraint, that is:

min Vol(Ω)
6.3.3 Constraining the length of horizontal bars in a Ω ,
two-dimensional MBB beam s.t. C(Ω) ≤ C T

In our last example, we show how a constraint over the where the maximum authorized compliance C T equals 150.
anisotropic maximum thickness may serve as a manufac- The result is displayed on Fig. 21.
turing constraint for civil structures. The resulting, optimized shape shows a very long, hori-
To this end, let us consider the two-dimensional bench- zontal bar in the upper side. Such members could be unde-
mark MBB beam example, of Fig. 20, where only one half sirable: because of their length, it is usually not possible to
of the structure is considered, taking advantage of the sym- produce them in industry. In this kind of situation, engineers
metry of the problem (so that all the designs represented in assemble such long bars by joining two (or more) smaller
Figs. 20, 21 and 23 are to be understood after symmetry with bars and constructing very strong connections between them
respect to their right-hand side): the structures are enclosed (see Fig. 22). Besides the increased cost of such connections,
in a working domain D with dimensions 3 × 1; they are the final beam may not behave mechanically in the predicted
clamped on their lower left-hand corner and a vertical load way, and so structural engineers generally wish to avoid them.
g = (0, −0.5) is applied at their upper right-hand corner. To penalize the length of such horizontal bars, we add
Body forces are omitted. a constraint on the anisotropic maximum thickness to the

123
 Author's personal copy
Comput Mech

45 •
in blue. Six load cases gi , i = 1, . . . , 6 are considered, as
D represented in Fig. 25: the first two correspond to extreme
wind scenarii, while the last four correspond to situations
41 •
where some wires have broken. We additionnally impose
31 • the symmetry of the structure with respect to the two planes
containing the point (7.5, 7.5, 0) and parallel to the (x z)- and
23 • (yz)-planes respectively.
As in the example of Sect. 6.3.2, the objective function of
interest in this multiple load setting is the sum S(Ω) of the
z 0 individual compliances C i (Ω) associated to each load case
ΓD
y x gi :
15
15 6
Fig. 24 Setting and boundary conditions of the pylon design example S(Ω) = C i (Ω),
of Sect. 6.4 i=1

optimization problem, and now solve where C i (Ω) = Ae(u iΩ ) : e(u iΩ ) d x,
Ω
min Vol(Ω)
Ω
and u iΩ is the elastic displacement of the shape when the
C(Ω) ≤ C T
s.t. , loads gi are applied (i.e. the solution of (1) with g replaced
PaM (Ω) ≤ δ/2.
by gi ).
with a threshold value δ = 1.0, and where the anisotropy is Several numerical experiments are conducted in this con-
dictated by the metric tensor: text, and the corresponding details are reported in Table 5.

 
m1 0 6.4.1 Simplification of the optimal design and improvement
M= . of its visibility
0 m2

The optimized shapes corresponding to the values m 1 = 0.1, At first, we merely aim at minimizing S(Ω) under a vol-
m 1 = 0.01, and m 2 = 1 are presented on Fig. 23: so that ume constraint, i.e. we solve the optimization problem (28)
the distance in the horizontal direction is reduced, the upper with C(Ω) replaced by S(Ω), and with target volume VT =
feature is rotated and broken in two bars, while a third element 0.1|D|. The resulting optimized shape is displayed in Fig. 26.
is added at their joint. Recall that while the appearance of Although this shape looks structurally coherent, it contains
thinner, smaller bars could be a hindrance in other contexts, several extended surfaces; as we have already mentioned, this
the typical scale involved in architectural is so large that even is an undesirable feature, since it undermines the visibility
small members (on the reduced scale) are manufacturable. of the shape, and makes it difficult to interprete the shape
in terms of bars, which is a key step in the civil engineering
6.4 A large-scale example design process.
One first idea towards circumventing this drawback sim-
In this last section, we intend to deal with a more realistic ply consists in reducing the maximum authorized volume VT ;
example of how the proposed material in this article may help it is then expected that mechanically unnecessary extended
in the design of realistic civil structures. surfaces will disappear. Acting this way on the target volume
The proposed test-case is inspired by the Pylon Design VT , which is given as a strict requirement in most indus-
Competition organized by the Royal Institute of British trial design situations, is perfectly legitimate in the present
Architects (RIBA) in 2014 [17]. This competition invited context: indeed, let us bear in mind that the designs resulting
experts from various fields to search for innovating designs from the numerical optimization process are only conceptual
of electric pylons, with minimal impact on the surrounding outlines aimed at guiding the architectural design process. In
environment, which are structurally coherent in the sense that this direction, we again solve Problem (28) (still with S(Ω)
they safely carry the loads transmitted by the cables. replacing C(Ω)) with target volumes VT = 0.075|D| and
Taking into account the specifications imposed by the VT = 0.05|D|; the resulting shapes are displayed in Fig. 27.
organizers of the competition, and after several simplifica- Evidently, the structure gets simplified and the visual result
tions, the setting is that presented in Fig. 24: the working is better.
domain has size 15 × 15 × 45; the structure is clamped on its An alternative way to achieve this simplification (thus
lower side, and loads are applied at the seven points depicted allowing to explore different structural systems) consists in

123
 Author's personal copy
Comput Mech

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
20 −20 0
⎝ 0 ⎠ ⎝ 0 ⎠ ⎝ −25 ⎠
−15 ⎛ ⎞ −15 ⎛ ⎞ −5 ⎛ ⎞
70 −70 0
⎝ 0 ⎠ ⎝ 0 ⎠ ⎝
⎛ ⎞ ⎛ ⎞ ⎛ ⎞ 0 ⎠
70 −85 −70 −85 0
⎝ −50
⎝ 0 ⎠ 0 ⎠ ⎛ ⎞ ⎝ −35 ⎠
⎛ ⎞ ⎛ ⎞
−85 70 −85 −70 −100 0
⎝ 0 ⎠ ⎝ 0 ⎠ ⎝
⎛ ⎞ ⎛ ⎞ ⎛ ⎞ 0 ⎠
70 −85 −70 −85 0 −50
⎝ 0 ⎠ ⎝ 0 ⎠ ⎛ ⎞ ⎝ −35 ⎠
⎛ ⎞ −70 ⎛ ⎞
−85 70 −85 −100 0
⎝ 0 ⎠ ⎝
⎝ 0 ⎠ ⎛ ⎞ ⎛ ⎞ 0 ⎠
⎛ ⎞ −70 −85 0
70 −85 −50
⎝ 0 ⎠ ⎝ −35 ⎠
⎝ 0 ⎠
−85 −100
−85
z z z
x x x
y y y

(1) (2) (3)

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 0 0
⎝ 25 ⎠ ⎝ −25 ⎠ ⎝ 25 ⎠ ⎛ ⎞
⎛ ⎞ 0
−5 0 −5 ⎛ ⎞ −5
⎝ 0 ⎝ 100 ⎠
0 ⎠
⎛ ⎞ ⎝ −100 ⎠ ⎛ ⎞ −35
0 −50 ⎛ ⎞ 0
⎝ 35 ⎠ 0 −35 ⎝ 0 ⎠ ⎛ ⎞
⎛ ⎞ ⎝ 0 ⎠ 0
−100 0 ⎛ ⎞ −50 ⎝ 100 ⎠
⎛ ⎞ ⎝ 0 ⎠ −50 0
0 ⎝ −100 ⎠ ⎛ ⎞ −35
−50 ⎛ ⎞ 0
⎝ 35 ⎠ ⎛ ⎞ 0 −35 ⎝ 0 ⎠ ⎛ ⎞
0 ⎝ 0 ⎠ 0
−100 −50 ⎝ 100 ⎠
⎝ 0 ⎠ −50 ⎛ ⎞
⎛ ⎞ 0 ⎛ ⎞ −35
0 −50 ⎛ ⎞ ⎝ −100 ⎠ 0
⎝ 35 ⎠ 0 ⎝ 0 ⎠
⎝ 0 ⎠ −35
−100 −50
−50

z z z
x x x
y y y

(4) (5) (6)

Fig. 25 Definition of the six load cases considered in the pylon design example of Sect. 6.4

Table 5 Values of the

Design Ω S(Ω) Vol(Ω) Pm (Ω)/PaM (Ω) δ
compliance S(Ω), the volume
Vol(Ω), and of the constraint Figure 26 0.137 1012 – –
function Pm (Ω)/PaM (Ω) for the
optimized designs of Sect. 6.4 Figure 27a 0.224 758 – –
Figure 27b 0.369 502 – –
√
Figure 28a 0.269 1009 – 4 min {m 1 , m 2 , m 3 }Δx
√
Figure 28b 0.117 1011 – 3 min {m 1 , m 2 , m 3 }Δx
Figure 30a 0.706 1012 −344 –
Figure 30b 0.690 1012 −317 –
Figure 30c 0.667 1012 −284 –
Figure 30d 0.625 1012 −182 –
Figure 31a 0.400 1012 – –
Figure 31b 0.600 758 – –
Figure 31c 1.050 503 – –
√
Figure 32a 0.320 1012 – 4 min {m 1 , m 2 , m 3 }Δx
√
Figure 32b 0.357 999 – 3 min {m 1 , m 2 , m 3 }Δx

using a constraint on the anisotropic maximum thickness of PaM (Ω) ≤ δ/2, where PaM (Ω) is defined by (24),
shapes, in the spirit of Sect. 6.3, while keeping the initial
target volume VT = 0.1|D|. More precisely, we now solve a associated to the two metric tensors of the form (33) charac-
version of Problem (32) featuring S(Ω) instead of C(Ω), as terized by
well as two constraints of the form

123
 Author's personal copy
Comput Mech

Fig. 26 Resulting pylon from the minimization of S(Ω) with target volume VT = 0.1|D|

Fig. 27 Optimized designs for the minimization of S(Ω) with target volumes a VT = 0.075|D| and b VT = 0.05|D|

123
 Author's personal copy
Comput Mech

Fig. 28 Optimized design in the minimization

√ of S(Ω) under a volume√ constraint VT = 0.1|D|, and anisotropic maximum thickness constraints
associated to the parameter a δ = 4 min {m 1 , m 2 , m 3 }Δx, b δ = 3 min {m 1 , m 2 , m 3 }Δx (Sect. 6.4.1)

m 1 = m 3 = 0.01, m 2 = 1 6.4.2 Imposing a resemblance between the optimal design

(constraint on the thickness in the (x z)-plane), and a target shape
then, (34)
m 2 = m 3 = 0.01, m 1 = 1 Let us now imagine that one desires a pylon design in which
(constraint on the thickness in the (yz)-plane). the material is concentrated around a central axis, as in the
shape ΩT of Fig. 29.
√ optimized shapes associated to the threshold values δ =
The One possibility to achieve this purpose is to constrain
4 min√ {m 1 , m 2 , m 3 }Δx and shapes to stay close to ΩT , in the line of Sect. 5.1. More
δ = 3 min {m 1 , m 2 , m 3 }Δx (recall that Δx is the mesh precisely, we solve Problem (29) with target volume VT =
size) are represented in Fig. 28. Notice that these shapes 0.1|D|, and for different values of the weight t. The resulting
still present extended surfaces, which are not parallel to optimized designs are displayed in Fig. 30. Understandably
the two√penalized planes. Interestingly, in the case where enough, as the value of t increases, the optimized shape gets
δ = 4 min {m 1 , m 2 , m 3 }Δx, a large quantity of material more and more influenced by the structural criterion S(Ω):
is concentrated at the bottom of the structure and the visual its bending and torsional stiffness dramatically increases at
result is quite satisfying. its bottom side.

123
 Author's personal copy
Comput Mech

6.4.3 Exploring innovating designs by changing the

boundary conditions

Eventually, as exemplified in Remark 1, significantly differ-

ent designs may be explored by acting on the imposed set of
fixations. In this section, we slightly modify the definition of
the test-case supplied in Fig. 24 by shrinking the clamping
surface Γ D into a 5 × 5 square centered at the middle of the
bottom side of the working domain D. Grossly speaking, it
is expected that doing so will entail optimized designs show-
ing curved members, benefitting from the surrounding space
of the working domain to connect Γ D to the supports of the
loads.
In a first step, we simply minimize the mechanical crite-
Fig. 29 Target shape ΩT in the example of Sect. 6.4.2 rion S(Ω) for several values of the target volume VT (that is,
we solve Problem (28) with C(Ω) replaced by S(Ω)), with
this new definition of the set Γ D of fixations for shapes. The

Fig. 30 Optimized designs under the constraint that the shape resembles the design ΩT of Fig. 29 (Sect. 6.4.2); values of the weighting coefficient
a t = 0.2, b t = 0.3, c t = 0.4, d t = 0.6

123
 Author's personal copy
Comput Mech

Fig. 31 Optimized designs in the minimization of S(Ω) performed in Sect. 6.4.3 with target volumes a VT = 0.1|D|, b VT = 0.075|D|,
c VT = 0.05|D|

associated optimized designs, represented in Fig. 31, obvi- and are associated to the metric tensors defined in (34) (so
ously show the aforementioned trend. that we solve exactly the same problem as in Sect. 6.4.1, up
Then, we incorporate two constraints on the anisotropic to the aforementioned change of boundary conditions). The
maximum thickness of shapes to the optimization problem: resulting shapes are presented on Fig. 32.
these constraints are identical to those detailed in Sect. 6.4.1,

123
 Author's personal copy
Comput Mech

Fig. 32 Optimized design in the minimization

√ of S(Ω) under a volume√ constraint VT = 0.1|D|, and anisotropic maximum thickness constraints
associated to the parameter a δ = 4 min {m 1 , m 2 , m 3 }Δx, b δ = 3 min {m 1 , m 2 , m 3 }Δx (Sect. 6.4.3)

7 Concluding remarks – Constrain shapes to resemble a specific design,

– Constrain shapes to fit a user-defined pattern,
In this article, we have proposed a shape and topology – Penalize the lengths of bars or the areas of extended sur-
optimization framework oriented towards applications in faces of shapes, thus allowing to simplify their design, or
architecture. This framework emphasizes on a precise, geo- to improve their visibility.
metric description of shapes: from the theoretical point of
view, Hadamard’s method is used to evaluate the sensitivity Realistic numerical simulations in two and three space
of optimization criteria with respect to variations of shapes, dimensions have been presented to confirm the efficiency
while their numerical representation is achieved via the level of the presented framework.
set method. Not only does this setting allow for an accurate
account of the mechanical behavior of shapes, but it naturally Acknowledgements Part of this work has been supported by the
OptArch Project (H2020-MSCA-RISE-2015). The authors are indebted
lends itself to an efficient modelling of constraints of a geo- to A. Beghini, from Skidmore Owings & Merrill, for his help in the def-
metric nature. In this perspective, we have devised three new inition of the examples discussed in Remark 1 and Sect. 5. They also
shape functionals, which are either related to the mechanical ackowledge fruitful discussions with N. Aage (Technical University
performance of structures, or allow to add personal, aesthetic of Denmark) and O. Amir (Technion, Israel Institute of Technology).
G. Allaire is a member of the DEFI Project at INRIA Saclay Ile-de-
requirements into the shape optimization problem. In partic- France. A. Faure and R. Estevez also acknowledge support from the
ular, our shape functionals make it possible to Labex CEMAM at Grenoble.

123
 Author's personal copy
Comput Mech

1
Appendix: Mathematical details M
p∂Ω (x) = M − 2 p∂Ω M (x M ). (37)

In this appendix, we present the mathematical proofs of the

Hence, for x ∈/ ∂Ω, x ∈/ Σ M , (16) follows and (i) is proved.
results stated in Sects. 4.3.1 and 5.
To obtain (ii), we simply apply (11) to the domain Ω M ,
and rely on Lemma 1; for x ∈ / ∂Ω, x ∈ / Σ M , the sequence
Proofs of the properties of the anisotropic, signed
of equalities
distance function
1 1
Proof of Proposition 4 The proof can be achieved in two ∇dΩ
M
(x) = M 2 ∇dΩ M (x M ) = M 2 n Ω M ( p∂Ω M (x M ))
ways: one could directly adapt the proofs of the results of 1
Propositions 1 and 2, or observe that the anisotropic context = n Ω ( p∂Ω
M
(x))
− 21
|M n Ω ( p∂ΩM (x))|
can be derived from the isotropic one via a suitable transfor-
mation. We rely on this second point of view.
More precisely, we consider the mapping Rd  x → proves the first part of (17). As for the second part, one has:
1 1
M 2 x ∈ Rd , and set x M := M 2 x for short. As a straightfor-
ward consequence of definitions, for any two points x, y ∈ x M = p∂Ω M (x M ) + dΩ M (x M )n Ω M ( p∂Ω M (x M )).
Rd , on has |x − y| M = |x M − y M |. It follows that the signed
distance functions dΩ and dΩ M (resp. sets of projections Π
∂Ω
and Π∂Ω ) are related as:
M Then, using (35), (37) and Lemma 1, it follows:
1 1
x = M − 2 p∂Ω M (x M ) + dΩ M (x M )M − 2 n Ω M ( p∂Ω M (x M )),
∀x ∈ Rd , dΩ
M
(x) = dΩ M (x M ), dΩM (x) 1 1
= p∂Ω
M (x) +
1 M − 2 com(M 2 )n Ω ( p∂Ω
M (x)),
− 21 |com(M 2 )n Ω ( p∂Ω
M (x))|
and Π∂Ω
M
(x) = M Π∂Ω M (x M ), (35) dΩM (x)
M (x) +
= p∂Ω M −1 n Ω ( p∂Ω
M (x)),
1
|M − 2 n Ω ( p∂Ω
M (x))|

where we have introduced the transformed domain Ω M :=

1
M 2 Ω.
which is the desired result.
The following lemma will be used repeatedly in the fol-
Point (iii) is a straightforward consequence of the first
lowing. Its proof is postponed to the end of this section. 1
two points (and notably the fact that Σ M = M − 2 Σ M ), and
Lemma 1 Let Ω be a bounded domain of class C 2 and M Proposition 1.
be a symmetric, positive definite matrix. Then, Likewise, point (iv) follows from Proposition 2, and the
transformation rule of Lemma 1 for the second fundamental
1. The unit normal vector n Ω M to ∂Ω M , pointing outward form of Ω M .
Ω M reads: Eventually, point (v) follows easily from a combination of
(37), Lemma 1, and the expression (13) of the derivative of
1
com(M 2 )n Ω (x) the usual projection mapping. 

∀x ∈ ∂Ω, n Ω M (x M ) = 1 We now provide the missing ingredient in the proof of
|com(M 2 )n Ω (x)| Proposition 4.
1
M − 2 n Ω (x)
= 1
, (36)
|M − 2 n Ω (x)| Proof of Lemma 1 The first point is classical (and actually
holds in a much larger context than that of linear mappings),
where com(B) stands for the cofactor matrix of a d × d and we concentrate on that of the second one. To this end,
matrix B. we use the following elementary characterization: for a given
1
2. The second fundamental form I I M of the boundary tangent vector v M = M 2 v to ∂Ω M (i.e. v M ∈ Tx M ∂Ω M ),
∂Ω M satisfies the identity: where v ∈ Tx ∂Ω,

1 1 1
∀x ∈ ∂Ω, M 2 I I xMM M 2 = I Ix . I I xMM (v M , v M ) = γ  (0), n Ω M (x M ), (38)
− 21
|M n Ω (x)|

It follows immediately from (35) that the skeleton Σ M is for any curve γ : (−t0 , t0 ) → ∂Ω M of class C 2 such that
1 1
obtained from that Σ M of Ω M via Σ M = M − 2 Σ M . Hence, γ (0) = x M , γ  (0) = v M . Let us choose γ (t) = M 2 ζ (t),
for x ∈/ ∂Ω, dΩM is differentiable at x if and only if x ∈
/ ΣM; where ζ : (−t0 , t0 ) → ∂Ω is a curve of class C 2 such that
using again (35), one has then: ζ (0) = x, ζ  (0) = v. Then, (38) together with (36) yield:

123
 Author's personal copy
Comput Mech

1 1
I I xMM (v M , v M ) = 1
ζ  (0), M 2 com(M 2 )n Ω (x), we obtain:
1
|com(M )n Ω (x)|
2
1  
det(M 2 )
= 1 ζ  (0), n Ω (x), 
P (Ω)(θ ) = κϕ(n)θ · n ds − ∇ϕ(n) · ∇∂Ω (θ · n) ds
|com(M )n Ω (x)|
2
Γ Γ
= 1
1
 ζ (0), n Ω (x), 
|M − 2 n Ω (x)|
+ (∇ϕ(n) − ∇∂Ω (ϕ(n))) · θ ds.
Γ
and the desired result follows. 

Eventually, using the symmetry property (14), the last inte-
Proofs of Propositions 5 and 6 Again, both results may be gral in the right-hand side of the above equality vanishes, and
equivalently inferred from their isotropic counterparts, by the expected result follows. 

1
using the isomorphism x → M 2 x in the same way as in the
proof of Proposition 4, or by adapting their original proofs Proof of Corollary 1 Using ϕ(n) = |n − n g |2 , ∇ϕ(n) =
(see [32] and [27]) to the present anisotropic context, and we 2(n − n g ) in Proposition 8 yields:
omit further details about these points. 


Pp (Ω)(θ ) = κ|n Ω − n g |2 θ · n ds
Details of the calculations of the shape derivatives of Γ

Sect. 5
− (n Ω − n g ) · ∇∂Ω (θ · n) ds.
Γ
Proof of Proposition 8 The proof is carried out in the work
[4], and is reproduced in here for convenience. Using again integration by parts on the boundary on the sec-
For θ ∈ Θad , a change of variables in the boundary inte- ond term in the right-hand side of the above equality yields:
gral (22) yields (see e.g. [40] Prop. 5.4.3):
 Pp (Ω)(θ )
P(Ωθ ) = ϕ(n Ωθ ◦ (Id + θ )) |com(Id + θ )| ds.  
Γ = κ|n Ω −n g | θ · n ds − (n Ω −n g )·∇∂Ω (θ · n) ds,
2

Γ  Γ
Using the well-known identity over matrices: = κ|n Ω − n g | −2κ(n Ω − n g ) · n Ω
2
Γ

com(M) = det(M)M −T , + 2(κ − div∂Ω (n g )) θ · n ds,

we obtain that θ → |com(I + ∇θ )| is Fréchet-differentiable which leads to the desired result. 


at θ = 0, and
Proof of Theorem 1 It follows from Proposition 6 that
|com(I + ∇θ )|= 1 + div∂Ω (θ ) PaM (Ω) is shape differentiable at an arbitrary shape Ω ∈
|o(θ )| θ→0 Uad . An easy calculation then yields:
+ o(θ ), where −→ 0.
||θ ||C 1,∞ (Rd ,Rd )

∀θ ∈ Θad , PaM (Ω)(θ )

In this formula, we recall that div∂Ω V := div(V )−∇V n, n
= G(dΩM
(x))Mn M ( p∂Ω
M
(x)), θ ( p∂Ω
M
(x)) d x.
stands for the tangential divergence of a (smooth enough) Ω
vector field V : ∂Ω → Rd . Hence, using Lebesgue domi-
nated convergence theorem and Proposition 3, it follows that We now proceed to achieve an expression of the above deriva-
θ → P(Ωθ ) is Fréchet-differentiable at 0, and that the cor- tive with the convenient structure (5). To this end, recalling
responding derivative reads: that Σ M has zero Lebesgue measure (see Proposition 4), we
apply the coarea formula to the mapping p∂Ω M : Ω \ ΣM →
∀θ ∈ Θad , P  (Ω)(θ ) =
       ∂Ω, as a curvilinear version of the classical Fubini theorem
∇ϕ(n) · ∇θ T n · n n −∇θ T n + ϕ(n)div∂Ω (θ ) ds. (see [21], or [3] in the same context):
Γ


Using integration by parts on the boundary Γ (see [40], Prop. PaM (Ω)(θ ) = Mn M (x), θ (x)
5.4.9), together with the identity: ∂Ω
 
G(dΩM (z))
d(z) ds(x),
∇∂Ω (θ · n) = ∇θ T n + ∇n T θ − (∇θ T n · n)n, M (x)∩Ω
ray∂Ω
M (z))
Jac( p∂Ω

123
 Author's personal copy
Comput Mech

where we recall that ray∂ΩM (x) = ( p M )−1 (x), and the Jaco- 15. Besserud K, Katz N, Beghini A (2013) Structural emergence: archi-
∂Ω
bian Jac( p∂Ω (x)) is defined as:
M tectural and structural design collaboration at som. Architect Des
83(2):48–55
! 16. Bradner E, Iorio F, Davis M (2014) Parameters tell the design story:
ideation and abstraction in design optimization. In: Proceedings of
M
Jac( p∂Ω (z)) = ∇ p∂Ω
M (z)∇ p M (z)T , z ∈ Rd \ Σ M .
∂Ω the symposium on simulation for architecture and urban design.
Society for Computer Simulation International, p 26
17. Royal Institute of British Architects, P.D.C.: http://www.
The only remaining task is then to calculate this Jacobian, ribapylondesign.com/home
which we do by using the representation (18) of ∇ p∂Ω
M in an
18. Burger M (2003) A framework for the construction of level set
orthonormal basis of R :
d methods for shape optimization and reconstruction. Interfaces Free
Bound 5:301–329
 −1 19. Campbell SL, Chancelier JP, Nikoukhah R (2006) Modeling and

d−1 M (z)
dΩ simulation in scilab/scicos. Springer, New York
M
Jac( p∂Ω (z)) = 1+ 1
λi ( p∂Ω M (z)) . 20. Cannarsa P, Cardaliaguet P (2004) Representation of equilibrium
i=1 |M − 2 n( p∂Ω M (z))| solutions to the table problem of growing sandpiles. Interfaces Free
Bound 6(4):435
21. Chavel I (2006) Riemannian geometry: a modern introduction.
This ends the proof. 
 Cambridge University Press, Cambridge
22. Christensen P, Klarbring A (2009) An introduction to structural
optimization, vol 153. Springer, Berlin
23. Christiansen AN, Nobel-Jørgensen M, Aage N, Sigmund O,
References Bærentzen JA (2014) Topology optimization using an explicit inter-
face representation. Struct Multidiscip Optim 49(3):387–399
1. Aage N, Amir O, Clausen A, Hadar L, Maier D, Søndergaard A 24. Clausen A, Aage N, Sigmund O (2014) Topology optimization with
(2015) Advanced topology optimization methods for conceptual flexible void area. Struct Multidiscip Optim 50(6):927–943
architectural design. In: Advances in architectural geometry 2014. 25. Clausen A, Aage N, Sigmund O (2015) Topology optimization of
Springer, pp 159–179 coated structures and material interface problems. Comput Meth-
2. Allaire G (2007) Conception optimale de structures, Mathéma- ods Appl Mech Eng 290:524–541
tiques & Applications (Berlin), vol 58. Springer, Berlin 26. Dambrine M, Kateb D (2010) On the ersatz material approximation
3. Allaire G, Dapogny C, Delgado G, Michailidis G (2014) Mutli- in level-set methods. ESAIM Control Optim Calc Var 16(3):618–
phase structural optimization via a level-set method. In: ESAIM: 634
control, optimisation and calculus of variations, 20, pp. 576-611. 27. Dapogny C (2013) Optimisation de formes, méthode des lignes
doi:10.1051/cocv/2013076 de niveaux sur maillages non structurés et évolution de maillages.
4. Allaire G, Dapogny C, Faure A, Michailidis G (2017) Shape opti- Ph.D. thesis, Université Pierre et Marie Curie-Paris VI. http://tel.
mization of a layer by layer mechanical constraint for additive archives-ouvertes.fr/tel-00916224
manufacturing (submitted) Preprint: https://hal.archives-ouvertes. 28. Dapogny C, Frey P (2012) Computation of the signed distance
fr/hal-01398877v2 function to a discrete contour on adapted triangulation. Calcolo
5. Allaire G, Dapogny C, Frey P (2014) Shape optimization with a 49(3):193–219
level set based mesh evolution method. Comput Methods Appl 29. De Buhan M, Dapogny C, Frey P, Nardoni C (2016) An opti-
Mech Eng 282:22–53 mization method for elastic shape matching. C R Acad Sci Paris
6. Allaire G, Jouve F, Michailidis G (2014) Thickness control in struc- 354:783–787
tural optimization via a level set method. (to appear in SMO), HAL 30. De Gournay F (2006) Velocity extension for the level-set method
preprint: hal-00985000v1 and multiple eigenvalues in shape optimization. SIAM J Control
7. Allaire G, Jouve F, Toader AM (2002) A level-set method for shape Optim 45(1):343–367
optimization. C. R. Acad. Sci. Paris Série I 334:1125–1130 31. Delfour M, Zolésio JP (2001) Shapes and geometries, Advances
8. Allaire G, Jouve F, Toader AM (2004) Structural optimization in design and control, vol 4. Society for Industrial and Applied
using sensitivity analysis and a level-set method. J Comput Phys Mathematics (SIAM), Philadelphia
194(1):363–393 32. Delfour M, Zolésio JP (2005) Shape identification via metrics
9. Amir O, Bogomolny M (2011) Topology optimization for concep- constructed from the oriented distance function. Control Cybern
tual design of reinforced concrete structures. In: Proceeding of the 34(1):137
9th world congress on structural and multidisciplinary optimiza- 33. Dombernowsky P, Søndergaard A (2009) Three-dimensional topol-
tion, June 13–17 2010, Shizuoka, Japan ogy optimisation in architectural and structural design of concrete
10. Amir O, Sigmund O (2013) Reinforcement layout design for con- structures. In: Proceedings of the international association for shell
crete structures based on continuum damage and truss topology and spatial structures (IASS) symposium, Valencia, Spain
optimization. Struct Multidiscip Optim 47(2):157–174 34. Dunning P, Kim H (2015) Introducing the sequential linear pro-
11. Bajcsy R, Kovacic S (1989) Multiresolution elastic matching. Com- gramming level-set method for topology optimization. Struct
put Vis Gr Image Process 46:188–198 Multidiscip Optim 51(3):631–643
12. Bando K, Din R, Fouquerand M, Gilbert L, Moissenot A, Nicolas M 35. Feringa J, Søndergaard A (2012) An integral approach to structural
(2016) Optimisation d’une structure et application architecturale. optimization and fabrication. In: ACADIA 12: synthetic digital
PSC MEC07, Ecole Polytechnique (X) ecologies [proceedings of the 32nd annual conference of the asso-
13. Baumgartner A, Harzheim L, Mattheck C (1992) Sko (soft kill ciation for computer aided design in architecture (ACADIA) ISBN
option): the biological way to find an optimum structure topology. 978-1-62407-267-3], pp 491–497
Int J Fatigue 14(6):387–393 36. Frattari L (2011) The structural form: Topology optimiza-
14. Bendsoe M, Sigmund O (2004) Topology optimization: theory, tion in architecture and industrial design. Ph.D. thesis, School
methods and applications. Springer, Berlin of Advanced Studies. Dottorato di ricerca in“ Architettura e

123
 Author's personal copy
Comput Mech

Design”: curriculum Conoscenza e progetto delle forme di inse- 54. Querin O, Steven G, Xie Y (1998) Evolutionary structural optimisa-
diamento”(XXIII cycle) tion (eso) using a bidirectional algorithm. Eng Comput 15(8):1031–
37. Frey P, George PL (2010) Mesh generation. Wiley-ISTE, New York 1048
38. Guo X, Zhang W, Zhong W (2014) Doing topology optimization 55. Sasaki M, Itō T, Isozaki A (2007) Morphogenesis of flux structure.
explicitly and geometrically a new moving morphable components AA Publications, London
based framework. J Appl Mech 81(8):081,009 56. Sethian J (1999) Level set methods and fast marching methods:
39. Guo X, Zhang W, Zhong W (2014) Explicit feature control in struc- evolving interfaces in computational geometry, fluid mechanics,
tural topology optimization via level set method. Comput Methods computer vision, and materials science. Cambridge University
Appl Mech Eng 272:354–378. doi:10.1016/j.cma.2014.01.010 Press, Cambridge
40. Henrot A, Pierre M (2005) Variation et optimisation de formes: une 57. Sethian J, Wiegmann A (2000) Structural boundary design via level
analyse géométrique, vol 48. Springer, Berlin set and immersed interface methods. J Computat Phys 163(2):489–
41. Larena, A.B.: Shape design methods based on the optimisation of 528
the structure. historical background and application to contempo- 58. Sigmund O (2007) Morphology-based black and white filters for
rary architecture. In: Proceedings of the third international congress topology optimization. Struct Multidiscip Optim 33(4–5):401–424
on construction history (2009) 59. Sigmund O, Bendsøe M (2004) Topology optimization: from air-
42. Li Y, Nirenberg L (2005) The distance function to the boundary, planes to nano-optics. In: Bridging from technology to society.
finsler geometry. Commun. Pure Appl. Math. 58(1):85-146-1209 Technical University of Denmark, Lyngby, Denmark
43. Mattheck C (1990) Design and growth rules for biological struc- 60. Strain J (1999) Semi-lagrangian methods for level set equations. J
tures and their application to engineering. Fatigue Fract Eng Mater Comput Phys 151(2):498–533
Struct 13(5):535–550 61. Stromberg LL, Beghini A, Baker WF, Paulino GH (2012) Topol-
44. Michailidis G (2014) Manufacturing constraints and multi-phase ogy optimization for braced frames: combining continuum and
shape and topology optimization via a level-set method. Ph.D. the- beam/column elements. Eng Struct 37:106–124
sis, Ecole Polytechnique X 2014. http://pastel.archives-ouvertes. 62. Sundaram MM, Ananthasuresh G (2009) Gustave eiffel and his
fr/pastel-00937306 optimal structures. Resonance 14(9):849–865
45. Murat F, Simon J (1976) Etude de problèmes d’optimal design. 63. Vallet MG, Hecht F, Mantel B (1991) Anisotropic control of mesh
Optim Tech Model Optim Serv Man Part 2:54–62 generation based upon a Voronoi type method. Arcilla AS, Häuser
46. Nahmad Vazquez A, Bhooshan S, Inamura C, Sondergaard A J, Eiseman PR, Thompson JF (eds) Third international conference
(2014) Design, analysis and fabrication of expressive, efficient shell on numerical grid generation in computational fluid dynamics and
structures: a prototype exploring synergy between architecture, related fields. North-Holland, pp 93–103
engineering and manufacture. In: Proceedings of the IASS-SLTE 64. Wang M, Wang X, Guo D (2003) A level set method for struc-
2014 symposium “Shells, Membranes and Spatial Structures: Foot- tural topology optimization. Comput Methods Appl Mech Eng
prints”. IASS SLTE 192(1):227–246
47. Novotny A, Sokołowski J (2013) Topological derivatives in shape 65. Wikipedia: https://commons.wikimedia.org/wiki/file:the_qatar_
optimization. Springer, Berlin national_convention_center_qncc_(6953154154).jpg
48. Osher S, Fedkiw R (2003) Level set methods and dynamic implicit 66. Wolff J (2012) The law of bone remodelling. Springer, Berlin
surfaces. In Applied mathematical sciences, vol 153. Springer, New 67. Xia Q, Shi T (2015) Constraints of distance from boundary to
York skeleton: for the control of length scale in level set based structural
49. Osher S, Santosa F (2001) Level set methods for optimization topology optimization. Comput Methods Appl Mech Eng 295:525–
problems involving geometry and constraints: I. frequencies of a 542
two-density inhomogeneous drum. J Comput Phys 171(1):272–288 68. Yamada T, Izui K, Nishiwaki S, Takezawa A (2010) A topology
50. Osher S, Sethian J (1988) Fronts propagating with curvature- optimization method based on the level set method incorporat-
dependent speed: algorithms based on Hamilton–Jacobi formula- ing a fictitious interface energy. Comput Methods Appl Mech Eng
tions. J Comput Phys 79(1):12–49 199(45):2876–2891
51. Persson PO (2004) Mesh generation for implicit geometries. Ph.D. 69. Zhang W, Yuan J, Zhang J, Guo X (2016) A new topology optimiza-
thesis, Citeseer tion approach based on moving morphable components (mmc) and
52. Pironneau O, Hecht F, Le Hyaric A. Freefem++ version 2.15-1 the ersatz material model. Struct Multidiscip Optim 53(6):1243–
53. Ponginan R. Danish team uses hyperworks to prove the value 1260
of topology optimization for concrete architectural structures.
http://www.altairuniversity.com/2015/04/14/danish-team-uses-
hyperworks-to-prove-the-value-of-topology-optimization-for-
concrete-architectural-structures/

123