Structure and appearance optimization for controllable

shape design
Jonàs Martínez, Jérémie Dumas, Sylvain Lefebvre, Li-Yi Wei

To cite this version:

Jonàs Martínez, Jérémie Dumas, Sylvain Lefebvre, Li-Yi Wei. Structure and appearance opti-
mization for controllable shape design. ACM Transactions on Graphics, 2015, 34 (6), pp.12.
�10.1145/2816795.2818101�. �hal-01240642�

HAL Id: hal-01240642

https://inria.hal.science/hal-01240642v1
Submitted on 9 Dec 2015

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Structure and Appearance Optimization for Controllable Shape Design
Jonàs Martínez Jérémie Dumas Sylvain Lefebvre Li-Yi Wei
INRIA Université de Lorraine, INRIA INRIA University of Hong Kong

Problem

Symmetry

Exemplars

Figure 1: Our technique automatically generates rigid shapes answering a specific loading scenario and resembling an input
exemplar pattern, while using a user-specified quantity of material. Top left: Loading scenario; in this case the synthesized
shape is anchored to the ground by its bottom left/right corners, while supporting a road through four attachments. Each
attachment contains an empty region (white) surrounded by a solid boundary (blue), serving as a socket to plug in the road
plank. Bottom left: Two exemplars defining the desired appearance. Second column: Two synthesized bridges answering the
loading scenario but each using a different exemplar. Photograph: Fabricated objects using the synthesized shapes.

Abstract CR Categories: I.3.5 [Computer Graphics]: Computa-

tional Geometry and Object Modeling—[Physically based
The field of topology optimization seeks to optimize shapes modeling]; I.3.7 [Computer Graphics]: Three-Dimensional
under structural objectives, such as achieving the most rigid Graphics and Realism—[Color, shading, shadowing, and
shape using a given quantity of material. Besides optimal texture]; J.6 [Computer Applications]: Computer-aided
shape design, these methods are increasingly popular as de- Engineering—[Computer-aided manufacturing (CAM)];
sign tools, since they automatically produce structures hav-
ing desirable physical properties, a task hard to perform by Keywords: manufacturing, modeling, texture synthesis,
hand even for skilled designers. However, there is no simple topology optimization
way to control the appearance of the generated objects.
1 Introduction
In this paper, we propose to optimize shapes for both their
structural properties and their appearance, the latter being Recent years have witnessed a significant spread of rapid
controlled by a user-provided pattern example. These two manufacturing technologies, such as 3D printing and laser
objectives are challenging to combine, as optimal structural cutting. In principle, these techniques empower casual users
properties fully define the shape, leaving no degrees of free- with the ability to create tangible objects from their virtual
dom for appearance. We propose a new formulation where ap- counterparts. In practice, it remains extremely difficult to
pearance is optimized as an objective while structural proper- design objects which are aesthetically pleasing and at the
ties serve as constraints. This produces shapes with sufficient same time structurally sound for real world constraints, such
rigidity while allowing enough freedom for the appearance of as being rigid enough to perform their intended function.
the final structure to resemble the input exemplar.
An important effort towards simplifying the creation of com-
Our approach generates rigid shapes using a specified quan- plex yet functional objects emerged from the field of topology
tity of material while observing optional constraints such optimization [Bendsøe 1989; Sigmund 2009; Brackett et al.
as voids, fills, attachment points, and external forces. The 2011]. In this field, the primary consideration is to design
appearance is defined by examples, making our technique lightweight structures that are as rigid as possible. That is,
accessible to casual users. We demonstrate its use in the con- optimizing for the most rigid shape using a prescribed amount
text of fabrication using a laser cutter to manufacture real of material. This is a key engineering problem as material use
objects from optimized shapes. and weight are directly related to cost and efficiency. These
techniques are a perfect match to additive manufacturing
technologies as they typically produce complex geometries
impossible to manufacture otherwise.
However, these approaches only consider rigidity as an op-
timization objective, and the appearance of the final ob-
ject cannot be controlled besides explicit constraints such
as avoiding regions of space or enforcing symmetries [Kosaka
and Swan 1999]. In this work we propose to jointly opti-
mize for the rigidity and the appearance of the structure,
as defined by a user-specified exemplar pattern. This is
different from after-the-facts reinforcement of the final re-
sult [Stava et al. 2012], and from synthesizing uniform, man-
ufacturable patterns [Dumas et al. 2015] (Figure 2): the op-
timized shape is obtained as the result of a single optimization
problem integrating both appearance and rigidity, and oper-
ates under a constrained material budget. It is easy to use
via by-example specification and simple constraints such as
solid and void regions to enforce, as well as external forces.
These constraints are general and allow us to optimize for
appearance, mechanical strength, and material cost.
Dumas [2015] Our result
A natural intuition is to combine topology optimization [Sig- Exemplar Loading Scenario Volume: 50% Volume: 34%
mund 2009] and by-example texture synthesis [Wei et al. No Control Controllable
2009] to satisfy both structural and appearance objectives.
However, these are challenging to combine, preventing the Figure 2: Comparison with [Dumas et al. 2015]. From left
algorithm to properly converge. This is confirmed experi- to right: input exemplar; loading scenario, attachment points
mentally (see Figure 5) as finding good compromises with a and optimization domain (gray square); result of [Dumas et al.
simple combination of these two objectives requires tedious 2015]; our result. Regardless of the loading scenario Dumas
parameter tuning, if possible at all. et al. always seek to produce a structure that fills a given
region, while we generate a rigid shape using the user-specified
Contributions. Our contributions are: quantity of material and resembling the input exemplar.
• A new formulation in which appearance is optimized as
an objective while rigidity is understood as a constraint.
of our knowledge, these structures are however not designed
• Controls that are powerful yet easy to understand: single from examples nor optimized for a specific loading scenario.
parameter for the appearance-rigidity tradeoff (α), vol- In concurrent work, Dumas et. al. [2015] propose a technique
ume usage bounds (vmin , vmax ), and appearance speci- synthesizing a uniform stochastic pattern covering a surface,
fied by example. while ensuring that it is printable. However, the application is
different: our approach generates a shape under a prescribed
• First order derivatives for the appearance objective, en-
material budget and will generally not fill a domain. Instead,
abling gradient descent under non-linear constraints.
it seeks for the optimal compromise between appearance and
rigidity while distributing material in space. Figure 2 high-
Scope. In this paper, we target the synthesis of flat (and lights the differences: while [Dumas et al. 2015] targets the
possibly curved) shapes that are manufactured with laser equivalent of uniform texturing, we target controllable shape
cutting. Our formulation directly translates to volumes (see design.
Section 5.3), but results in large problems which are too slow
to solve via our current implementation. Scalability to higher
dimensions is an important direction of future work. Style transfer. Style transfer is a variant of by-example syn-
thesis where the process is guided to enrich existing content
with details. Hertzmann et. al. [2001] pioneered this idea by
2 Related Work proposing to transfer details specified by a pair of images A:A’
to produce an image B’ from a different source image B. Re-
Our work is at the crossing of two fields: by-example texture cent approaches have explored how to exchange styles within
synthesis and topology optimization. and across collections of shapes [Xu et al. 2010; Xu et al.
2012; Li et al. 2013; Han et al. 2015; Ma et al. 2014]. These
By-example texture synthesis. The problem of synthesizing approaches typically require a collection or a pair defining
a new image resembling an exemplar image is a long-standing style by analogy, which are not available in our context.
problem in Computer Graphics [Wei et al. 2009]. We focus
here only on works most related to ours. Our work jointly optimizes appearance and structural objec-
tives, instead of transferring style after the facts. The global
The initial methods for by-example texture synthesis are structure therefore emerges from the details of the pattern,
based on Markov Random Fields [Efros and Leung 1999; Wei which becomes an intrinsic part of the final shape.
and Levoy 2000]: a probabilistic model is defined by sampling
neighborhoods from the exemplar. These approaches are not
trivially amenable to our context due to the stochastic nature Topology optimization. Topology optimization is the pro-
of the optimization process. Kwatra et. al. [2005] proposed cess of determining the optimal distribution of material inside
a different point of view of the problem which is based on a domain [Bendsøe and Kikuchi 1988; Deaton and Grandhi
formulating an energy — the pixels of the output image 2014], so as to optimize various objectives such as rigidity.
being the variables. By optimizing this energy a new image We discuss these approaches in more details later in the text,
is synthesized which resembles the example. We adopt this as they are an important component of our work.
point of view to define the appearance energy relating the
produced shape to the exemplar. Different manufacturing constraints for topology optimiza-
tion have been investigated [Sigmund 2009], however, there
Few works have considered both texture synthesis and fabrica- exists little work regarding aesthetics. Paulino et. al. [2015]
tion. Zhou et al. [2014] synthesize patterns along curves while recently proposed to perform topology optimization over a 2D
precisely controlling their topology. The results can then be polygonal tessellated domain, for artistic purposes. Kosaka
laser-cut or 3D printed as they form singly-connected ob- et. al. [1999] constrain the process to enforce symmetries,
jects. The software Magics by Materialize contains libraries which can be used for aesthetic considerations. Christiansen
of structures that can be tiled inside a 3D object. To the best et. al. [2015] demonstrate the use of topology optimization
for modeling purposes. However, none of these methods can work exerted by the forces on the structure, i.e. the sum of
handle general exemplar patterns. the dot product of forces and displacements, as illustrated
in Figure 4. A low compliance implies that forces produce
only small displacements, which characterizes a high rigidity.
Fabrication. Our work targets modeling for fabrication. Sev-
eral recent approaches considering structural properties have u1 u2
been proposed in this context, for instance, to reinforce frag-
ile 3D models [Stava et al. 2012], to analyze rigidity prior to
f f
fabrication [Zhou et al. 2013; Umetani and Schmidt 2013], to
achieve strong but lightweight parts [Wang et al. 2013; Lu Low compliance C1 = f · u1 High compliance C2 = f · u2
et al. 2014], or to fabricate objects with prescribed mechan-
ical behavior [Panetta et al. 2015; Schumacher et al. 2015]. Figure 4: Two 1D linear elements (springs) with a single
However, these methods consider mechanical properties only, degree of freedom. A same force f generates a displacement
but not appearance as our method does. of respectively u1 and u2 . The compliances reflect that the
stronger spring (leftmost) is more rigid: C1 < C2 . The same
3 Our Formulation concept translates to elements of higher dimensions.

Our method combines two fundamental ingredients. The first We model the shape in an n-dimensional grid of square el-
is the notion of appearance as defined by neighborhood sim- ements denoted by x, each having 2n corner nodes shared
ilarities between a synthesized structure and an example with their neighbors. Each element e in x receives a density
pattern. The second comes from mechanical engineering and xe which through optimization has to converge towards void
is the notion of compliance. Figure 3 provides an overview of (= 0) or solid (= 1), thus defining an interior and exterior. In
our method. practice, intermediate values remain after optimization, and
we apply a thresholding after convergence.
The basic problem in which compliance appears is the pre-
diction of the mechanical behavior of a structure when it is We formulate our goal as a multi-objective optimization prob-
subjected to precise boundary conditions — that is, a set of lem that minimizes both an appearance energy AI (x) and
attachment points and external loads applied to the struc- the structural compliance C (x):
ture. In this work we consider small deformations for which x = arg min (AI (x) , C (x))
the behavior of the structure can be characterized by linear x
elasticity. We consider isotropic materials described by their , where I is the input exemplar — a black and white pattern
Young’s modulus and Poisson’s ratio. The compliance is the defining void (pixel = 0) and solid (pixel = 1) regions; x
is the outcome — densities defining a shape in the grid —
computed through our optimization procedure. The user has
Boundary conditions Exemplar to specify at least one attachment point for the problem to be
well-posed. She can optionally impose additional conditions,
such as regions of void or fill, symmetry, and external forces;
see results in Figure 1 and Section 6.
Previous works exist to optimize each of these energies in
isolation (Section 2). Therefore, a straightforward approach
Compliance Compliance and would optimize for a linear blend of both energies, that is:
Level
optimization appearance optimization
AI (x) + λC (x) (1)
with λ > 0 allowing to explore the tradeoff. Unfortunately,
0
such a simple scheme does not produce reliable results: the
values of λ that can produce a reasonable output differ widely
between exemplars, boundary conditions, and domain size,
when they exist at all. Figure 5 illustrates this issue and
compares to our formulation.
1
We therefore propose to modify the formulation of the prob-
lem. We note that the goal is not necessarily to obtain the
most rigid structure, but rather a structure with sufficient
rigidity, i.e. which does not yield under the given loads. Thus,
2 our insight is that rigidity should be considered as a con-
straint, which can be relaxed to allow more freedom for the
appearance objective. Thus, our goal is now to minimize
AI (x) such that the structural compliance is below a thresh-
Figure 3: Overview of our multi-resolution optimization old Cmax and the volume is bounded:
approach. Given an input exemplar (upper right) and out-
put boundary conditions (upper left), our method optimizes arg min : AI (x)
x
the corresponding output in a multi-resolution fashion (lower subject to : C (x) 6 Cmax
rows). The boundary conditions can include support, solid, X
and void. In this example, the volume is constrained to 35% vmin 6 xe 6 vmax
of the overall output domain, and the relaxation factor of the e
compliance with respect to the optimal is set to α = 1.2.
∀e 0 6 xe 6 1
 Exemplar Weighted sum optimization via eq. (1) Our method
λ=1 λ = 50 λ = 300 α = 1.2

×15 ×1.48 ×1.35 ×1.2

×42 ×3.5 ×1.68 ×1.2

Figure 5: Comparison of a straightforward weighted sum approach and our formulation. All results use the same parameters
and a volume constrained to 30% of the entire domain. We give below each result the ratio between its compliance and the
compliance of the shape optimized without appearance objective (Copt ) ; e.g. ×1.2 implies the result is within 20% of the
computed optimum. λ weights the importance of rigidity versus appearance. On the left hand side, a low λ gives results with
good appearance but mediocre compliance. On the right hand side, a large λ produces more rigid results but degraded appearance.
We show in orange the values of λ producing reasonable compromises, and in green and red the best and worst compliance ratios,
respectively. Note that these differ significantly between both exemplars. Our method (rightmost column) does not need any
specific setting besides the threshold from the computed optimum (20%). The boundary conditions are the same as in Figure 3.

The volume bounding constraint is important. The weight of Each square element e receives a continuous scalar density
the structure is often negligible compared to external forces, 0 6 xe 6 1, and its Young’s modulus is defined as:
in which case the most rigid shape would tend towards a full
block of material. vmax prevents this naive solution to exist. Ee = Emin + (xe )p (E0 − Emin )
On the contrary, when only considering the weight of the
, where E0 is the material Young’s modulus, Emin > 0 is
structure — i.e. solving a self-weight problem — the naive
a small value to prevent numerical instabilities, and p = 3
solution is an empty shape. vmin prevents it. The volume
is the standard SIMP penalization factor which penalizes
constraint is also a natural control for the user. Combined to
intermediate values in the solution.
appearance it allows changing the overall size of the structure
(see Figure 6). For the sake of clarity we express volume The compliance of the output is measured by summing the
constraints as a percentage of the design domain. We often compliance of each individual element. Let us denote:
use only external forces or only self-weight in which case we
respectively set vmin = 0% or vmax = 100%. In such cases u Global displacement vector.
we report only the non-trivial bound. ue Element displacement vector.
f Global force vector.
K Global stiffness matrix.
A meaningful value of the Cmax constraint is crucial to ensure K0 Element stiffness matrix with unit Young’s modulus.
a feasible solution. We describe how the threshold is com- t Element thickness.
puted in Section 3.1, and describe the appearance objective
in Section 3.2. We discuss our solver and numerical scheme Compliance minimization is formulated as follows:
in Section 4 and extensions in Section 5.
Copt = min : C (x)
x
X
3.1 Compliance Constraint subject to : vmin 6 xe 6 vmax
e

We determine the compliance threshold Cmax by first com- ∀e 0 6 xe 6 1

puting a solution Copt to the problem considering compliance
, where the compliance term is given by the equation:
alone. We adopt the well-established Solid Isotropic Material
Penalization (SIMP) topology optimization method [Bendsøe X
and Kikuchi 1988]. Implementation details of this method C (x) = uT f = t Ee ueT K0 ue
can be found in [Sigmund 2001; Andreassen et al. 2011]. e
subject to : Ku = f
3.1.1 Topology Optimization for Determining Copt We use the Method of Moving Asymptotes (MMA) [Svanberg
1987] to minimize C (x) and obtain a solution Copt .
The SIMP method seeks to minimize compliance by assign- Note that when the boundary conditions include large ex-
ing densities to each element
P in x given a constrained total ternal forces, the weight of the structure tends to have a
material budget vmin 6 e xe 6 vmax . The elastic structure negligible influence, and can be ignored. We discuss in Sec-
is simulated with the Finite Element Method (FEM). tion 5.1 how self-weight can be taken into account.
3.1.2 Setting the Threshold Cmax The first term compares pairwise densities throughout neigh-
borhoods and corresponds to the appearance energy (eq. (2)).
We first perform standard topology optimization to compute The second term is used to increase the spatial uniformity of
a lower bound on the achievable minimal compliance Copt . the appearance energy [Kopf et al. 2007; Kaspar et al. 2015].
We then set Cmax = α Copt , where α > 1. As we increase O is an occurrence map storing how many times each exem-
the value of α, we obtain results having a higher compliance, plar pixel is used in the different closest neighborhoods. It is
but that are more similar to the exemplar. This is a simple computed from the previous iteration. λocc > 0 controls the
and predictable parameter controlling how much freedom is |I|
amount of enforced spatial uniformity. |x|N 2 is a normalizing
allowed to appearance, as can be seen in Figure 6. factor making λocc independent of synthesis resolution. In
our implementation λocc = 20.
3.2 Appearance Objective
Optimizing densities. Optimizing the appearance objective
Our method can be used with any appearance energy that
(eq. (2)) given the best matching neighborhoods is more chal-
has first order derivatives available, as required by the solver
lenging. We are facing a non-linear, non-convex optimiza-
(Section 4). Our formulation builds upon the work of Kwatra
tion problem for both objective and constraints. In addition,
et. al. [2005] but adapts it to facilitate the computation of
evaluating the compliance constraint is computationally ex-
first order derivatives with respect to x.
pensive as it requires solving for the FEM equation. For
In the following, when using the notation xe , e is meant these reasons we rely on the Globally Convergent Method of
as a coordinate within the grid of elements. We denote by Moving Asymptotes (GCMMA) [Svanberg 1995]. GCMMA
N = 2δ +1 the texture synthesis neighborhood size (typically converges in fewer iterations than augmented Lagrangian
15 × 15, i.e. δ = 7). We denote by m(e) the coordinate of methods, reducing the number of required FEM solutions.
the matching pixel in I whose neighborhood ze is the most It iteratively solves subproblems that are convex approxima-
similar to the current neighborhood of xe in the result. The tions of the original problem, and rely on the gradient of
matching process is described in Section 4. We denote by both objective and constraints functions to do so. GCMMA
zce = I[m(c) + (e − c)] the value at coordinate e from the therefore requires the first order derivatives of the objective
best matching neighborhood of c, where I[z] accesses the and constraint functions.
pixel value (0 or 1) at coordinate z in I. The appearance
energy contributed by an element xe ∈ x is: Derivatives. The derivative of the volume constraint is 1.
X The compliance derivative is [Bendsøe and Sigmund 2003]:
AI (xe ) = |xe − zce |r (2)
∂C ∂f ∂K
c , kc−ek∞ 6δ = 2ueT − ueT ue
∂xe ∂xe ∂xe
P
The total appearance energy is AI (x) = e AI (xe ). We use
∂f
r = 1.2, set experimentally as in [Kwatra et al. 2005]. The term ∂x e
is null when the weight of the structure is
neglected. For self-weight problems (Section 5.1) this term
will however influence the result.
4 Solver
The derivatives are typically filtered to prevent numerical is-
We are now ready to solve for the global optimization prob- sues, and to control the optimization quality. We consider the
lem, that is to minimize appearance under the compliance extensively used smoothing operator described by [Sigmund
constraint (Section 3). This is a challenging optimization as 1997] to filter the derivative of the function C:
all terms are non-linear and are subject to inequality con-
straints. In addition, the appearance objective contains com- ∂C
d 1 X ∂C
binatorial terms: the best matching neighborhoods. = P we,i xi
∂xe max(, xe ) i we,i ∂xi
i
We thus optimize for the density x in an iterative block coor-
dinate descent scheme alternating between finding the best , where we,i = max (0, γ − dist(e, i)) is a weight factor (con-
matching neighborhoods coordinates m(e) and the densities volution), controlled by a parameter γ (filtering radius), and
xe for each element e.  (10−3 ) is a small coefficient avoiding division by zero.
At this step, all zce are assumed to be constants. Therefore,
Initialization. We set xe = min( vmax , 1)
for all elements the derivative of AI (x) can be expressed as:
|x|
and select random coordinates for best matches m(e). One ∂AI X |xe − zce |r
can choose a different initial guess, but a better convergence = r
∂xe (xe − zce )
is typically observed when starting from a uniform gray; see c , kc−ek∞ 6δ
Figure 2 in [Sigmund and Maute 2013].
The derivative is not defined when xe = zce . Thus, using a
Neighborhood matching. Given the densities, the coor- small ε > 0, we regularize AI (x) as follows:
dinates of the most similar neighborhoods are computed X
r/2
with PatchMatch [Barnes et al. 2009] and the random walk A
eI (xe ) = (xe − zce )2 + ε
of [Busto et al. 2010]. For a neighborhood at coordinate e c , kc−ek∞ 6δ
in x and a neighborhood at coordinate e0 in I the similarity
used for matching is defined as: , and it follows that the derivative is:
X O [e0 ] |I| ∂A
eI X
r/2−1
d(e, e0 ) = |xc − I[e0 + (c − e)])|r + λocc · = r (xe − zce ) (xe − zce )2 + ε
N2 |x| N 2 ∂xe
c , kc−ek∞ 6δ c , kc−ek∞ 6δ
 α = 1.4, vmax = 30% α = 1.6, vmax = 30%

α = 1.2, vmax = 30%

α = 1.2, vmax = 35% α = 1.2, vmax = 40%

Figure 6: Progressive relaxation of the maximum compliance (top row) and maximum volume (bottom row) constraints.

Multi-resolution. For improved performance and quality we Forces due to the structure weight are modeled by a vertical
optimize through a multi-resolution scheme. The process force acting on each grid node q as follows:
starts from downsampled versions of the grid x and exem- X me
plar I. The resolution is iteratively doubled, using the pre- qx = 0, qy = −g xe
|qe |
vious result to initialize the next finer resolution by bilinear e , q∈qe
up-sampling. The process is illustrated in Figure 3.
, where qe is the set of nodes belonging to element e (as
defined in Section 3), me is the element mass, and g is the
Our algorithm optimizes three resolution levels. The compli-
absolute value of the gravitational acceleration. Let us em-
ance relaxation parameter α remains constant throughout
phasize that this force depends on the current densities of
the process. The exemplar is downscaled to match the resolu-
the elements xe .
tion of each level. The strain-displacement and constitutive
material matrices are also changed according the resolution. When using the SIMP formulation on a problem taking into
account the weight of the structure, the displacements might
We use the same multi-resolution scheme to compute the become unbounded for low density regions, resulting in nu-
compliance solution (Section 3.1) and obtain Cmax for each merical issues [Bruyneel and Duysinx 2005]. We therefore
resolution level. In practice, we observe that the value of use a modified formulation of material stiffness as suggested
Cmax is remarkably stable across resolutions. by [Pedersen 2000] to overcome this problem:

Emin + (xe )p (E0 − Emin ) µ < xe 6 1
Ee = (3)
Convergence. The optimization process ends when k∇Ak
is Emin + xe µp−1 (E0 − Emin ) 0 < xe 6 µ
kAk
below a small threshold (we use 0.001), or when a maximum In our experiments we set µ = 0.25. This switches to a linear
number of iterations is reached (in our implementation we stiffness model in regions of low densities. The derivatives are
use 40, 20, 10 on the three successive resolution levels). updated accordingly. The non-differentiable point where the
switch between models occurs does not have a detrimental
impact in practice [Bruyneel and Duysinx 2005].
5 Extensions
We observe that on self-weight problems the volume bounds
vmin , vmax have to allow for some freedom to achieve con-
Optional constraints can be added to our basic formulation. vergence. Indeed, the optimized shape is a subtle tradeoff:
We describe two important ones. The first is to consider adding matter makes some regions more rigid but also adds
the weight of the structure itself during optimization (Sec- stress to others through gravity. Figure 7 illustrates results
tion 5.1). This is useful when there is no external force besides obtained on self-weight problems.
gravity applied to the structure. The second is to consider
symmetry constraints, which are useful for aesthetics pur-
poses but also to reduce computation time when the solution
5.2 Symmetry
is known to have symmetries (Section 5.2). We also describe
Symmetry plays an important role in aesthetics for
how to optimize for 3D outputs even though our method is
shape design. We adopt the symmetry reduction approach
dimension agnostic (Section 5.3).
of [Kosaka and Swan 1999] for topology optimization. The
design domain x is partitioned into S > 1 subdomains xi .
5.1 Self-weight We define a mapping between xi and an imaginary domain
x∗ , and only optimize for x∗ . The derivatives are:
We optionally take into account the weight of the structure ∂C 1 X ∂C
∗
=
and the forces it generates under gravity. Note that on self- ∂xe S xie
i
weight problems — i.e. no external forces — the complete
void is a trivial optimal. We therefore impose vmin > 0 in That is, we compute the derivatives for the design domain x,
such cases. and optimize x∗ according the averaged derivatives given by
 Compliance optimization 6 Results
Boundary conditions Standard Using eq. (3)
Most of our results are obtained by laser cutting from the syn-
thesized shapes. We then assemble objects by gluing several
planks together. In most cases, we compute and assemble
independent 2D results, under the assumption that forces
remain in a plane — which works well in practice for most
32% volume 27% volume scenarios. We however also investigated full 3D solutions, as
Compliance and appearance optimization described in Section 5.3.
Before presenting our results in more details Section 6.2, we
describe Section 6.1 how we obtain the final curves for laser
cutting. We discuss performance in Section 6.3 and validation
tests in Section 6.4.

6.1 Contour Extraction

41% volume 43% volume 34% volume Contours for laser cutting are extracted in a few simple steps.
The optimized shape x is first thresholded (0.5) to snap values
Figure 7: Optimizing self-weight problems (no external which are between 0–1 to void or solid. In rare cases, this
forces). Top: A different material stiffness model is required results in the creation of small disconnected components. We
to avoid degeneracies in low density regions. Bottom: All filter these by keeping the connected components anchored
results use the same parameters α = 2.2, vmin = 20%, to attachment points. Disconnected components are further
vmax = 100%. Self-weight problems are more challenging discussed in Section 6.5. Finally, the filtered grid is upsampled
to optimize (see text), and therefore the bounds are relaxed by bilinear filtering (x2 in our implementation), and paths
to let the optimizer converge. Note how different volumes are for laser cutting are extracted along the isovalue 0.5.
obtained depending on the exemplar.
6.2 Fabricated Objects
Compliance optimization only We created several objects using our approach. In all cases,
the user only specified the attachment points, external forces,
x1 x2 target volume and example pattern. The algorithm automat-
Symmetry x∗ ically synthesizes the structure. Thus, many results can be
easily produced using a variety of patterns: the algorithm
Compliance and appearance deals with the complex task of generating the intricate de-
tails of the final structure. While we only laser cut miniatures,
industrial cutters could be employed to fabricate large-scale
objects in a variety of materials.
Besides attachment points and external forces, the
user may also rely on passive elements, which can
represent non-designable parts with a fixed density.
Figure 8: Constrained symmetry, with boundary conditions The inset figure illustrates the use of
defined on x∗ . Maximum volume is constrained to 45%. passive elements to optimize a struc-
ture around the SIGGRAPH logo.
This is obtained by constraining the
the mapping. Note that even though all xie variables map value of xe , that is xmin 6 xe 6 xmax .
to the same x∗e , their individual gradients on the right-hand If desired, other passive element prop-
side may differ due to asymmetric loading scenarios. erties can be predefined, such as the material volumetric mass
density and stiffness.

5.3 Optimizing 3D Structures In Figure 10 different bridge sides supporting a road are
obtained by changing the loading scenario. A symmetry con-
straint is also used, but this is optional. Results using different
Our formulation is amenable to 3D, adding a third dimension patterns on the same set of conditions are shown Figure 1.
and using a grid of cubic (hexahedral) elements. While this In Figure 11 a set of shelves is produced. They are designed
would provide a full volume synthesis, such an approach is to be fixed on the ground and to support several shelves offset
computationally expensive and requires a 3D exemplar as from the attachment point: the weight is entirely supported
input (schemes using several 2D exemplars to define a volume by the sides. Yet, the structure remains visually similar to
could be adapted [Wei et al. 2009]). the exemplar pattern. In Figure 12 we apply the same prin-
ciple to produce phone stands. Figure 13 shows a variety of
Instead, we propose to optimize structures along several in- tables obtained by interleaving three planks. Using different
terleaved planes in 3D, as illustrated in Figure 9. This is patterns immediately changes the look and feel of the results.
different from independently optimizing 2D shapes: the 3D The tables are very strong and can support large weights.
hexahedral elements at the crossing of several planes are
shared, and stresses and appearance propagate across the Finally, we show in Figure 9 3D results where the structures
different planes. The results in Figure 9 show how pattern along each plane are optimized jointly in an interleaved 3D
features are able to flow from one plane to another. problem. This allows the pattern to flow from one plane
Figure 9: 3D synthesis. From left to right: Boundary conditions, result optimized without appearance, two fabricated chairs
from our optimized results using different patterns.
Symmetry
Symmetry

Figure 11: A set of shelves, meant to be fixed on the ground.

Two exemplars are used on the same problem, producing re-
sults of very different styles but the same purpose.
Symmetry

Figure 10: Bridges obtained from the same exemplar us-

ing different loading scenarios. Symmetry is used, which is
why the boundary conditions are shown for only half of the
problem. Top: Forces are applied at the top, supporting only
the road. Middle: Forces are applied below the top for the
Figure 12: Phone stands fabricated from optimized struc-
road, and at the top to create a handrail. Bottom: Forces are
tures using two different exemplar patterns.
applied in the middle. Combined with the passive elements
this produces an arch.

summarizes performance for the main results. On average,

to another, while in previous results the pattern features in 2D, 75% of the time is spent on the FEM computation,
could be interrupted between different planks. Also, note 20% on the appearance gradient computation, and 5% on
how the result optimized without appearance looks much less the GCMMA optimization.
appealing in 3D, while our results produce intricate, visually
interesting details. As performance was not our focus, our reference implementa-
tion uses a single thread and takes in the order of minutes to
6.3 Performance converge. Nevertheless, the multi-resolution approach allows
the user to preview the result being computed.
We implemented our approach with Python and use the
GCMMA implementation of the NLopt [Johnson 2007] op- As can be seen in Table 1, the 3D result of Figure 9 takes
timization library. We measured the execution time on an roughly three times longer to compute than a 2D problem
Intel R CoreTM i7-4770K @ 3.50GHz, 16 GB RAM. Table 1 with the same number of elements.
 volume = 45%

Figure 14: Two structures in MDF wood optimized for

different forces but with the same volume constraint. The
ratio between the two forces is 10 : 1. The arrow indicates
where the top structure starts to rupture when loaded with
Figure 13: Three tables produced with our system, using the heavier weight intended for the bottom structure (please
three planks and structures optimized independently. The two refer to the accompanying video).
tables on the left use the same external conditions but different
patterns. The right table weights 76g and is supporting a filled Boundary cond. Compliance optimization
cup of 762g (10× heavier).

# elements exemplar time # iter

Figure 1 202K Web 240 × 240 80 3300 120
Figure 2 62K Sponge 200 × 200 1 5600
0
77
Figure 5 150K Cells 210 × 210 40 2000 112
Figure 7 80K Cells 200 × 200 10 0400 56
Figure 8 120K Grid 160 × 160 50 2400 118
Figure 9 114K Sponge 104 × 104 140 2000 113
Figure 10∗ 162K Flower 176 × 169 50 4500 106 Exemplar Compliance and appearance
Figure 12 230K Grid 330 × 330 70 3800 89
Figure 14 90K Cells 105 × 105 30 3400 119

Table 1: Performance on our main results. The exemplar

resolutions are shown as width×height. Times are formated
as min0 sec00 . ∗ Figure 10 middle result.

6.4 Structural Properties

Figure 15: Comparison of Von Mises stresses. vmax = 40%,
α = 1.3. We use the classical L-Beam boundary conditions
We first verify that the structural properties optimized by our for stress analysis [Duysinx and Bendsøe 1998] (see top left
system can be observed after fabrication. We show in Fig- image). Stress colors use the same normalization of both
ure 14 two bridge structures using the same volume but rows. Our structure (bottom row) has a higher compliance
optimized with a different force in the center. As can be seen, (α = 1.3), but local stresses are comparable to the ones found
for the case where the force is small the optimizer grows in the result without appearance (top row).
features on the top of the bridge, as this does not violate
the compliance constraint. On the contrary, when the force
is large the optimizer concentrates material below the arch,
for reinforcement. When a heavy load is applied to both, mization. Optimizing local stresses is still an active research
the bridge optimized for a small load collapses, whereas the topic [París et al. 2005; Lee et al. 2012; Holmberg et al. 2013].
bridge optimized with the correct force withstands it easily.
6.5 Limitations, Future Work
Compliance reflects the global rigidity of a shape but does
not consider local stresses. Therefore, there is a concern that Connectivity and convergence. While the SIMP method
shapes of low compliance but high local stresses could be does not explicitly prevent disconnected components from
produced, resulting in local failures under loads. In practice appearing, they are typically not present in an optimized de-
the results usually exhibit low local stresses, but for a few sign. Whenever the compliance constraint is violated, matter
specific places such as sharp corners. This is illustrated is redistributed in weaker regions to reduce the compliance.
in Figure 15 for the classical L-Beam test. Our approach Combined with the volume constraint this discourages the
inherits this limitation from compliance-based methods. It existence of disconnected components. In all our results, the
is however worth noting that our method produces results disconnected elements represent less than 3% of the total vol-
having comparable local stresses to those optimized without ume. However, in cases where the exemplar has many small
appearance, as shown in Figure 15. disconnected components and where the user allows for a
high compliance threshold, the appearance objective is free
Note that in the field of topology optimization compliance to generate disconnected components (Figure 16, col.1–2).
is widely used due to its smoother behavior (see e.g. §2.4
in [Deaton and Grandhi 2014], §6.10 in [Sigmund and Maute Self-weight (Section 5.1) explicitly penalizes disconnected
2013]) making it amenable to efficient gradient descent mini- components: unsupported matter produces high compliance.
No self-weight, α = 1.15 No self-weight, α = 1.20 With self-weight, α = 1.20
Disconnected exemplar Non-stochastic exemplar Strongly oriented exemplar

Figure 16: Some challenging exemplars.

Therefore, using self-weight encourages well connected shapes Acknowledgements

(Figure 16, col.3). However, combining self-weighting with
external forces generally leads to very challenging optimiza- We would like to thank anonymous reviewers for their valu-
tion problems and the resulting shapes are not perfectly con- able comments. This work was supported by ERC grant
verged. In addition, self-weight require relaxing the compli- ShapeForge (StG-2012-307877) and general research fund
ance bound (see Figure 7). This typically requires the user Dynamic Element Textures (HKU 717112E).
to test a few different parameters. Strictly enforcing connec-
tivity thus remains an open direction of future work.
References
Note that there are assumptions about the exemplars that
are inherited from texture synthesis. They should exhibit an Andreassen, E., Clausen, A., Schevenels, M., Lazarov,
overall stochastic, homogeneous appearance, and have fea- B., and Sigmund, O. 2011. Efficient topology optimiza-
tures roughly the size of the selected neighbourhood size. tion in Matlab using 88 lines of code. Struct. Multidiscip.
The appearance of organized patterns is not properly repro- Optim. 43, 1, 1–16.
duced (Figure 16, col.4). Strongly oriented patterns make it
challenging to create a rigid structure when the features do Barnes, C., Shechtman, E., Finkelstein, A., and Gold-
not align with stress directions (Figure 16, col.5). man, D. B. 2009. PatchMatch: A randomized correspon-
dence algorithm for structural image editing. ACM Trans.
Graph. 28, 3, 24:1–24:11.
Thickness constraints. While we did not impose a mini-
mum length-scale on our designs, previous works exist to Bendsøe, M. P., and Kikuchi, N. 1988. Generating opti-
control the minimum thickness and hole size in the SIMP mal topologies in structural design using a homogenization
framework [Sigmund 2009; Zhou et al. 2015]. Combining method. Computer Methods in Applied Mechanics and En-
these approaches to ours is left as future work. gineering 71, 2, 197–224.

7 Conclusion Bendsøe, M. P., and Sigmund, O. 2003. Topology Opti-

mization: Theory, Methods and Applications.
Our work enables a novel way to design shapes that are rigid Bendsøe, M. P. 1989. Optimal shape design as a material
under a set of external conditions. It offers an unprecedented distribution problem. Structural Optimization 1 , 192–202.
control over appearance through the specification of an ex-
emplar. Rigidity is understood as a constraint affording for Brackett, D., Ashcroft, I., and Hague, R. 2011. Topol-
a simple and predictable control on the tradeoff between ogy optimization for additive manufacturing. In Proc. of
appearance and structural properties. the 24th Solid Freeform Fabrication Symposium, 6–8.
We envision that expert users will use our technique to quickly Bruyneel, M., and Duysinx, P. 2005. Note on topology
produce initial designs serving as a starting point, while non- optimization of continuum structures including self-weight.
expert users will explore a large variety of appearances for Struct. Multidiscip. Optim. 29, 4, 245–256.
objects having the same mechanical purpose.
Busto, P. P., Eisenacher, C., Lefebvre, S., and Stam-
There are several avenues of future work. First, our technique minger, M. 2010. Instant texture synthesis by numbers.
does not scale well to dense 3D problems which are almost In Proc. of the VMV Workshop, 81–85.
impractical due to their large computational cost. Second,
some patterns have appearance that complicates the task of Christiansen, A. N., Bærentzen, J. A., Nobel-
achieving a rigid structure, as illustrated in Figure 16. While Jørgensen, M., Aage, N., and Sigmund, O. 2015.
this is currently a limitation we believe that integrating the Combined shape and topology optimization of 3D
pattern orientation as an optimization variable will help on structures. Computers & Graphics 46 , 25 – 35.
the rightmost exemplar, allowing features to align with the
local directions of stress. Finally, we are looking forward to Deaton, J., and Grandhi, R. 2014. A survey of structural
explore design tools exploiting our technique. and multidisciplinary continuum topology optimization:
post 2000. Struct. Multidiscip. Optim. 49, 1, 1–38.
We hope to bring a novel modeling tool that will empower
users — experts or otherwise — to produce shapes that are Dumas, J., Lu, A., Lefebvre, S., Wu, J., and Dick, C.
unique, visually appealing, and yet structurally sound for a 2015. By-Example Synthesis of Structurally Sound Pat-
given usage scenario. terns. ACM Trans. Graph. 34, 4, 137:1–137:12.
Duysinx, P., and Bendsøe, M. P. 1998. Topology optimiza- Schumacher, C., Bickel, B., Rys, J., Marschner, S.,
tion of continuum structures with local stress constraints. Daraio, C., and Gross, M. 2015. Microstructures to
Int. J. Numer. Methods. Eng. 43, 8, 1453–1478. control elasticity in 3D printing. ACM Trans. Graph. 34,
4, 136:1–136:13.
Efros, A. A., and Leung, T. K. 1999. Texture synthesis
by non-parametric sampling. In Proceedings of the Inter- Sigmund, O., and Maute, K. 2013. Topology optimization
national Conference on Computer Vision, 1033–1038. approaches: A comparative review. Struct. Multidiscip.
Optim. 48, 6, 1031–1055.
Han, Z., Liu, Z., Han, J., and Bu, S. 2015. 3D shape
creation by style transfer. Vis. Comput. 31, 9, 1147–1161. Sigmund, O. 1997. On the design of compliant mechanisms
using topology optimization. Mechanics of Structures and
Hertzmann, A., Jacobs, C. E., Oliver, N., Curless, B., Machines 25, 4, 493–524.
and Salesin, D. H. 2001. Image analogies. In Proc. of
SIGGRAPH 2001, 327–340. Sigmund, O. 2001. A 99 line topology optimization code
written in Matlab. Struct. Mult. Optim. 21, 2, 120–127.
Holmberg, E., Torstenfelt, B., and Klarbring, A.
2013. Stress constrained topology optimization. Struct. Sigmund, O. 2009. Manufacturing tolerant topology opti-
Multidiscip. Optim. 48, 1, 33–47. mization. Acta Mechanica Sinica 25, 2, 227–239.
Stava, O., Vanek, J., Benes, B., Carr, N., and Měch,
Johnson, S. G., 2007. The NLopt nonlinear-optimization
R. 2012. Stress relief: Improving structural strength of 3D
package.
printable objects. ACM Trans. Graph. 31, 4, 48:1–48:11.
Kaspar, A., Neubert, B., Lischinski, D., Pauly, M., and Svanberg, K. 1987. The method of moving asymptotes—a
Kopf, J. 2015. Self tuning texture optimization. Com- new method for structural optimization. Int. J. Numer.
puter Graphics Forum 34, 2, 349–359. Methods. Eng. 24, 2, 359–373.
Kopf, J., Fu, C.-W., Cohen-Or, D., Deussen, O., Lischin- Svanberg, K. 1995. A globally convergent version of MMA
ski, D., and Wong, T.-T. 2007. Solid texture synthesis without linesearch. In Proc. of the first world congress of
from 2D exemplars. ACM Trans. Graph. 26, 3. structural and multidisciplinary optimization, 9–16.
Kosaka, I., and Swan, C. C. 1999. A symmetry reduction Umetani, N., and Schmidt, R. 2013. Cross-sectional struc-
method for continuum structural topology optimization. tural analysis for 3D printing optimization. In SIGGRAPH
Computers & Structures 70, 1, 47 – 61. Asia 2013 Technical Briefs, 5:1–5:4.
Kwatra, V., Essa, I., Bobick, A., and Kwatra, N. 2005. Wang, W., Wang, T. Y., Yang, Z., Liu, L., Tong, X.,
Texture optimization for example-based synthesis. ACM Tong, W., Deng, J., Chen, F., and Liu, X. 2013. Cost-
Trans. Graph. 24, 3, 795–802. effective printing of 3D objects with skin-frame structures.
ACM Trans. Graph. 32, 6, 177:1–177:10.
Lee, E., James, K. A., and Martins, J. R. 2012. Stress-
constrained topology optimization with design-dependent Wei, L.-Y., and Levoy, M. 2000. Fast texture synthe-
loading. Struct. Multidiscip. Optim. 46, 5, 647–661. sis using tree-structured vector quantization. In Proc. of
SIGGRAPH 2000, 479–488.
Li, H., Zhang, H., Wang, Y., Cao, J., Shamir, A., and
Cohen-Or, D. 2013. Curve style analysis in a set of Wei, L.-Y., Lefebvre, S., Kwatra, V., and Turk, G.
shapes. Computer Graphics Forum 32, 6, 77–88. 2009. State of the art in example-based texture synthesis.
In Eurographics 2009, State of the Art Report.
Lu, L., Sharf, A., Zhao, H., Wei, Y., Fan, Q., Chen, X.,
Savoye, Y., Tu, C., Cohen-Or, D., and Chen, B. 2014. Xu, K., Li, H., Zhang, H., Cohen-Or, D., Xiong, Y., and
Build-to-last: Strength to weight 3D printed objects. ACM Cheng, Z. 2010. Style-content separation by anisotropic
Trans. Graph. 33, 4, 97:1–97:10. part scales. ACM Trans. Graph. 29, 6, 184:1–184:10.

Ma, C., Huang, H., Sheffer, A., Kalogerakis, E., and Xu, K., Zhang, H., Cohen-Or, D., and Chen, B. 2012. Fit
Wang, R. 2014. Analogy-driven 3D style transfer. Com- and diverse: Set evolution for inspiring 3D shape galleries.
puter Graphics Forum 33, 2, 175–184. ACM Trans. Graph. 31, 4, 57:1–57:10.

Panetta, J., Zhou, Q., Malomo, L., Pietroni, N., Zhou, Q., Panetta, J., and Zorin, D. 2013. Worst-case
Cignoni, P., and Zorin, D. 2015. Elastic textures for structural analysis. ACM Trans. Graph. 32, 4, 137:1–
additive fabrication. ACM Trans. Graph. 34, 4, 135:1– 137:12.
135:12. Zhou, S., Jiang, C., and Lefebvre, S. 2014. Topology-
constrained synthesis of vector patterns. ACM Trans.
París, J., Muínos, I., Navarrina, F., Colominas, I., and
Graph. 33, 6, 215:1–215:11.
Casteleiro, M. 2005. A minimum weight FEM formu-
lation for structural topological optimization with local Zhou, M., Lazarov, B. S., Wang, F., and Sigmund, O.
stress constraints. In VI World Congress on Structural 2015. Minimum length scale in topology optimization
and Multidisciplinary Optimization. by geometric constraints. Computer Methods in Applied
Mechanics and Engineering 293 , 266 – 282.
Paulino, G. H., and Gain, A. L. 2015. Bridging art and
engineering using Escher-based virtual elements. Struct.
Multidiscip. Optim. 51, 4, 867–883.
Pedersen, N. 2000. Maximization of eigenvalues using
topology optimization. Struct. Mult. Optim. 20, 1, 2–11.