Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Contents lists available at ScienceDirect

Comput. Methods Appl. Mech. Engrg.

journal homepage: www.elsevier.com/locate/cma

Concurrent topology optimization of shells with pattern-guided

infills for intuitive design and additive manufacturing
Yu Li a,b , Zeyu Zhang b,c , Jiaxiang Luo b,c , Wei Peng a,b , Weien Zhou a,b , Wen Yao a,b ,∗
a Defense Innovation Institute, Chinese Academy of Military Science, Beijing 100071, China
b Intelligent Game and Decision Laboratory, Beijing 100071, China
c
National University of Defense Technology, Changsha 410073, China

GRAPHICAL ABSTRACT

ARTICLE INFO ABSTRACT

Keywords: This paper introduces a systematic design optimization approach for shells with pattern-guided
Topology optimization infills, aiming to address both intuitive design requirements and considerations for additive
Shell-infill structure manufacturing. The approach utilizes a density-based topology optimization method and assigns
Pattern-guided design
two sets of design variables to distinctly define the base, coating, and infill structures. By
Coating technique
evolving the solid shells and enriched infills concurrently, the design optimization process
Visual appearance
ensures a cohesive integration of both components. To guide the infill design towards matching
the geometrical features of a predefined pattern, a geometrical constraint is imposed on the
infill field. Enhanced by an updated and approximate volume constraint, the infill materials
with prescribed appearance performances are accordingly distributed in the base region.
The simultaneous optimization of shell geometry and infill layout is then performed within
the framework of minimum compliance topology optimization. A range of numerical results
illustrate the effectiveness and robustness of the proposed approach. The topology optimized
results demonstrate that a given pattern can directly control the design features of the infill in
a visually explicit manner and the shell-infill composites ensure convenient manufacturability.

∗ Corresponding author at: Defense Innovation Institute, Chinese Academy of Military Science, Beijing 100071, China.
E-mail address: [email protected] (W. Yao).

https://doi.org/10.1016/j.cma.2023.116485
Received 19 June 2023; Received in revised form 22 September 2023; Accepted 22 September 2023
Available online 10 October 2023
0045-7825/© 2023 Elsevier B.V. All rights reserved.
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

1. Introduction

Topology optimization has developed from an academic discipline to an advanced and widely-used computer-aided approach
for designing lightweight structures in industries like aerospace and automotive [1]. With the rapid development of additive
manufacturing technologies, designers could directly fabricate such innovative and complex structures designed through topology
optimization. This shift has led to the consideration of practical requirements from both the design and manufacturing perspectives,
including factors such as stiffness, strength, buckling resistance, temperature effects, manufacturability-related geometrical features,
tooling accessibility, and so on.
Especially in the fields of architecture and consumer product design, aesthetics have gained increased attention and play an
important role in consumer acceptance. Ref. [2] confirmed these points and proposed applying a unity-in-variety principle to
improve aesthetics in topology optimization. Aesthetics primarily rely on the visual quality of a subject, and topology optimized
configurations often exhibit organic appearances.
However, topology optimized results are highly diversified and can take on arbitrary shapes within the design space. Achieving
visually appealing morphology becomes quite coincidental. Certain feature size, connectivity, or enclosed voids constraints have
already been proposed to control the topology in various studies (e.g., Refs. [3–6]). The desirable effects of such design methods are
incorporated into mathematical formulations and sophisticated algorithms, but the final outcomes may not be immediately visible
or intuitive to designers before conducting the topology optimization process.
Hence, the enhancement of visual geometrical features is frequently performed through manual trial and error or as a post-
processing measure subsequent to topology optimization. These manual interventions can significantly compromise the achieved
optimized mechanical performance. Consequently, it is advantageous to optimize both the mechanical performance and aesthetics of
the design concurrently. Furthermore, such concurrent design approach can facilitate the promotion of specific desirable geometrical
features for manufacturing or prevent the occurrence of detrimental flaws in the design.
In this paper, our aim is to tackle the aforementioned considerations by introducing an intuitive and manufacturable design
paradigm.

1.1. Intuitive design

In the topology optimization design considering both mechanical performance and aesthetics, a series of studies have been made.
Some studies focused on designing surface textures for 3D structures [7] and shells [8] using reference exemplars that indicated
solid and void regions. Ref. [9] explored the topology optimization of tessellations as a means to merge art and engineering. Ref. [10]
optimized shapes for both structural properties and appearance, where the latter was controlled by a user-provided pattern example.
In their modeling formulations, appearance was treated as an objective function and structural properties as the constraints. Hence,
the optimization process would likely induce troublesome problems in connectivity and convergence. Ref. [11] switched the above
settings with stiffness as the goal and appearance as the constraints. However, this approach required complex multi-resolution
optimization and similarity regulations. More recently, a more general and robust method called pattern-guided design was proposed
in Ref. [12]. It employed simple and effective appearance constraint calculations, allowing multiple 2D or 3D geometrical patterns
to directly steer topology optimization towards designs that were both stiffness-optimized and appearance-matched.
In addition, other attempts were made using the neural style transfer technique [13,14]. Ref. [15] simultaneously optimized
both structural performance for a given loading condition and geometric similarity for a reference design. However, the design
process relied on a single weighted objective function and the coupling between aesthetics and mechanical performance seemed
quite weak. Ref. [16] improved upon this by integrating the abstract structural style of a given image as a similarity constraint.
Architectures with an artistic flavor were generated and continually evaluated during the design optimization process. Refs. [17–19]
further developed interactive topology optimization frameworks considering the design details, the subjective preferences, and the
feature size controls of the designers, respectively.
However, the majority of the above researches rarely considers the manufacturability of the designs. Although the design
blueprints were obtained with fascinating appearance, they were often challenging to manufacture or required post-processing.
This limitation restricts their practical application in engineering.
The authors believe that such intuitive designs with artistic considerations should be accompanied by mechanical performance
and manufacturing requirements. Therefore, we further develop our design method for shell-infill structures suitable for additive
manufacturing.
It is worth highlighting that the notion of intuitive design, as explored in this paper, refers to a design approach where
structural configurations or styles are guided by a visible blueprint. This blueprint incorporates geometrical features that are readily
comprehensible to the designers, as opposed to being non-intuitively defined through complex mathematical formulations.

1.2. Additive manufacturing

In the realm of additive manufacturing, a shell-infill structure refers to a composite where a solid shell forms the exterior (also
called coating in the following) and a closed-walled porous infill occupies the interior. Such a structure exhibits superior mechanical
properties compared to solid counterparts, such as high strength-to-weight ratio, excellent buckling resistance, favorable energy
absorption characteristics, and strong robustness against material deficiency and force variation [20,21]. Additive manufacturing
provides the flexibility required to fabricate such coated structures, particularly in industrial metal printing processes like selective

2
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Fig. 1. A satellite bracket filled with lattice developed by Northwestern Polytechnical University [22].

laser melting (as shown in Fig. 1). This manufacturing technique allows for precise control and customization of the shell and infill,
enabling the widespread application of these structures in various disciplinary fields.
The conventional approach to designing shell-infill structures in engineering is usually undertaken by replacing the solid interior
of an optimized structure with predefined porous patterns, such as repetitive rectilinear lattices, concentric infills, or honeycomb
cells. Although this solution can effectively reduce the weight of the structure, it does not guarantee the desired overall mechanical
properties.
A more relevant approach involves employing the concurrent topology optimization design method, which optimizes both the
geometry of the shells and the layout of the infills simultaneously, based on the functional requirements. Naturally, there are two
critical aspects to consider in the optimization of shell-infill composites. Firstly, it is essential to identify the appropriate coated
shell, and secondly, the design of a suitable porous infill is crucial.
Several topology optimization approaches have been employed to address these two challenges. These research directions are
complementary as they optimize one component of the composite (either the shell or the infill) while assuming the other component
to be predetermined or fixed.
In addressing the interface-related challenges of the coated shell, Ref. [23] originally developed a two-step filtering and projection
strategy, which was further extended to 3D cases in the density-based method [24]. Ref. [25] simplified this approach and employed
an erosion-based interface identification method. The coated shell problem was also solved using the level-set method [26,27] in
Refs. [28–30], as well as in the MMC-MMV method in Ref. [31]. Ref. [32] investigated the effect of local material orientation on
shell-infill design, while Ref. [33] established a full-scale fiber model and focused on surface fiber reinforcement design for shells.
Regarding the design of porous infills in prescribed shapes, multi-scale topology optimization methods have been widely used
to determine optimal infill layouts [34–36]. Initially, repetitive infill patterns were adopted, where uniform microstructures were
distributed in the design domain [37]. However, fixed infill patterns may limit design flexibility and the performance of the optimized
structure. Additionally, high computational costs and disconnected graded cellular structures are issues to be handled.
To overcome these limitations, spatially-varying infill structures at only macroscopic scale have been developed. For example,
Ref. [21] studied the topology optimization of bone-like porous infill structures and Ref. [38] investigated the optimization design
of graded multi-phase infills. Ref. [39] further developed a design method to concurrently optimize the layout of the shells with
porous infills. Ref. [40] proposed to generate optimized structural typologies with homogenized properties. Ref. [41] considered
additive manufacturing overhang control in the design of shell-infill structures.
In summary, the shell-infill structure proves to be convenient for additive manufacturing. However, the determination of specific
infill shapes and geometrical features, including material porosity (local volume) and orientation, overhang angle, and minimum
size, remains non-intuitive and challenging for designers as discussed earlier.

1.3. Our bridging method

Our current research is motivated by the above two perspectives. We present a complete solution to the concurrent topology
optimization of shells with pattern-guided infills, which is a bridging way between intuitive design and additive manufacturing. The
proposed design method emphasizes intuitive design principles, allowing designers to have more control and understanding of the
final results, providing them with a tangible starting point for the optimization process.
To achieve intuitive design objectives, we utilize and adapt the pattern-guided design method [12], as it provides an accurate
appearance descriptor and an easy-to-extend design framework. Regarding manufacturing properties, the shell-infill modeling
present in this article is based on the previous work of Ref. [40], enabling the creation of a clear and well-defined shell-infill
layout. Since each of them partially addresses the concerning problem from one side, the combination is non-trivial and significant.
The remainder of this paper is organized as follows. Section 2 presents the modeling method of shell-infill structure and
the pattern-guided design method. Section 3 describes the concurrent topology optimization problem along with the associated
sensitivity analyses and the control strategies of the optimization parameters. Section 4 elaborates numerical results with detailed
discussions. Conclusions are drawn in Section 5.

3
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Fig. 2. Design Variable 1, denoted as 𝜇, is used to identify the base region 𝜑 and the shell 𝜏. Design Variable 2, denoted as 𝜐, is used to design the infill’s
feature𝜓. A honeycomb picture is selected as the guided geometrical pattern 𝐼. The intermediate fields 𝜑, 𝜏, 𝜓 are interpolated to obtain the final physical
density field 𝜌. The color legend for density visualization is also depicted.

2. Shell-infill modeling and pattern-guided design methods

Fig. 2 illustrates an overview of the conceptual design methods and detailed technical roadmap proposed in this work. For the
concurrent design of shells and pattern-guided infills, we make use of two scalar fields of design variables (i.e. DV1 and DV2) defined
on the design domain as shown in Fig. 2(a). From the first field of design variables 𝜇, a Double Smoothing and Projection (DSP)
approach and coating step are applied to ensure a clear distinction between coating 𝜏, base 𝜑 and void (see Section 2.1). From the
second field of design variables 𝜐, an enrichment field 𝜓 is designed with a pre-defined pattern I by imposing constraints on an
appearance distance function (see Section 2.2). The integration of these intermediate fields into the infill 𝜓𝜑 and the final physical
density field 𝜌 is realized by a material interpolation model (see Section 2.3 as shown in Fig. 2(b)).

2.1. Coated-base structure with filtering technique

A coated structure consists of two material phases, a coating, and an infill. And these two sub-structures are related by a base
region. To effectively separate the base structure and the coating structure, Ref. [23] introduced an elegant method to obtain a
coated structure using only a single field as design variable. Ref. [40] further improved the two-step filtering process with a Double
Smoothing and Projection, DSP approach. In this paper, we adopt the improved procedure for a clear distinction between coating,
infill, and void. Related filtering techniques are presented here.

2.1.1. Smoothing
The first method in the procedure is a smoothing operation using the Helmholtz-type PDE-based density filter [42], i.e.,
( )2
𝑅 { }
− √ ∇2 𝑥̂ + 𝑥̂ = 𝑥 , 𝑥 ∈ 𝜇 , 𝜇̄̂ , 𝜑 , 𝜐 , (1)
2 3

4
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

where the scalar-valued 𝑅 corresponds to the filter radius in the standard filtering technique by a convolution operator and indirectly
controls the length-scale imposed by the filter operation. In the current design framework, there are four smoothing steps with
generally four different values of radii. We consistently use the PDE-filter for the four smoothing steps.

2.1.2. Projection
The second one is a projection step to force the smoothed values on the interval [0, 1] towards either 0 or 1. We use the
formulation for the smoothed Heaviside projection based on the hyperbolic tangent function [43,44], stated as
( )
tanh(𝛽𝜂) + tanh 𝛽(𝑥̂ − 𝜂) { }
𝑥̂̄ = ( ) , 𝑥̂ ∈ 𝜇̂ , 𝜇̂̄̂ , ‖∇𝜑‖
̂ 𝛼 , (2)
tanh(𝛽𝜂) + tanh 𝛽(1 − 𝜂)
where 𝛽 controls the sharpness of the projection and 𝜂 is the threshold value between 0 to 1. When 𝛽 → +∞, designs (𝑥) ̄̂ are forced
towards pure 0–1 solution. A parameter continuation process starting from a small 𝛽 value is applied to improve convergence
behavior, in contrast to directly starting with a very large 𝛽 value. There are three projection steps with generally three different 𝛽
values in the current design procedure.

2.1.3. Double smoothing and projection

The Double Smoothing and Projection, DSP approach [40,45] is to use the smoothing and projection steps twice in a row with
filter radius of 𝑅1 , and projection parameters 𝛽1 , 𝜂1 in the design field. The motivation for using the DSP approach is two-fold
compared to the original Single Smoothing and Projection (SSP) approach proposed in Ref. [23].
One is that the DSP approach could indirectly control the length scale of the base region. As shown in Fig. 2(b), the middle
tiny gap is eliminated in the base region after the DSP filtering process. As reported in Ref. [39], when the width of the base
region was small, shell–infill would locally degenerate to a shell enclosing no infill. To prevent this, they additionally adapted the
robust formulation [43] to ensure a minimum length scale of the base. A combined objective function between the blue-printed and
eroded designs was further minimized with a gradually increasing weighting factor. The DSP approach is easy to be implemented
and simplifies the design method to produce a well-defined coating.
The other reason is that the DSP approach could greatly reduce the possibility of the base density being non-binary. As observed
in the result of the SSP approach [23,39], the base structure and coating did not converge exactly to 0 or 1. To circumvent such
undesired effects, Ref. [40] successfully used the DSP approach and compared the effects.

2.1.4. Coating
The third one is a spatial gradient-based shell modeling where the coating layer is derived from the interface between the base
region and void. The binary base structure is smoothed again using 𝑅2 < 𝑅1 , such that the gradient across the sharp edges is
well-defined. The shell is modeled by a gradient norm of the smoothed base field and subsequently projected using 𝛽2 and 𝜂2 , i.e.,

𝜏 = ‖∇𝜑‖
̂ 𝛼, (3)
√
where the term ‖∙‖𝛼 is represented as the Euclidean norm divided by scalar 𝛼 = 𝑅2 ∕ 3, rather than the usual meaning of 𝐿𝛼 -
norm. An analytical relation is shown in Ref. [23] between 𝑅2 and the maximum coating thickness. This is used to select 𝑅2 for a
user-specified coating thickness 𝑡ref as
√
3
𝑅2 = 𝑡 ≈ 2.5𝑡ref . (4)
ln (2) ref
To circumvent an undesired boundary effect during the smoothing operation of Eq. (1) and help calculate the spatial gradient of
Eq. (3), we make use of the domain extension approach [46]. As shown in the top right of Fig. 2(b), half of the coating 𝑡ref ∕2 near
the boundary would exceed the design domain. To effectively restrict the optimized structure to the design domain, a parameter
𝑞 < 1 is penalized on the elastic property only in the padded domain, as implemented in Ref. [40] and will be seen in the stiffness
interpolation in the following subsection.

2.2. Appearance constrained infill based on a geometrical pattern

In this section, we creatively integrate an appearance constraint into the design of infills’ geometrical features. For the concept
of pattern-guided design, Ref. [12] proposed a simple and effective appearance constraint in structural topology optimization. With
a geometrical pattern predefined, the optimized structure will share features from the patterns. Rather than a fixed infill without
any specific detail and a conceptual design of pure solid without coating, the pattern-guided infill promotes design flexibility and
structural manufacturability. We will briefly introduce the pattern-guided design method originally adopted from Ref. [12]. And
some modifications for further implementation in the shell-infill design framework are also described.

2.2.1. Guided design infill

Inspired by the work of Ref. [39], an enrichment infill is evolved from a second design field and further superimposed on the
plain base region to create pattern-guided structures. The second design field, 𝜐, is smoothed (giving 𝜐̂ ,) and projected, resulting in
a mesh-independent infill field, 𝜓 = 𝜐̄̂ . The smoothing radius 𝑅3 and projection parameters 𝜂3 , 𝛽3 reduce erroneous checkerboard
patterns and introduce a minimum length scale in the optimization.
Upon the enrichment field 𝜓, a further smoothing is applied, giving 𝜓.
̂ This smoothing is meant to quantify the local appearance
distance. By imposing an upper bound on this measure based on a user-defined pattern, the infill’s material distribution is effectively
formed into guided patterns. The constraint will be explained in later sections.

5
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

2.2.2. Geometrical pattern and patch discretization

In this paper, only 2D patterns are considered (e.g., given in *.PNG form files), which prescribe certain geometrical features or
design styles. For the density-based topology optimization design method, a normalization of pixel RGB values and proper image
resizing are possibly needed in the pre-processing.
As shown in the second row of Fig. 2(b), two pairs of patches are assigned to the design domain 𝛺 and pattern 𝐼. Patches are
rectangular windows of size |𝜔| = (2𝑙x + 1) × (2𝑙y + 1), where 𝑙x , 𝑙y are integers defining the half-sizes of the patch along each axis.
𝜔𝑒 is the neighborhood of central element 𝑒 specified by
{ }
𝜔𝑒 = 𝑖| ‖𝑖𝑘 − 𝑒𝑘 ‖ ⩽ 𝑙𝑘 , 𝑘 ∈ {x, y} . (5)

These local square patches are further used to describe a global appearance function field.
As discussed in Ref. [12], a large pattern size with relatively small patch sizes will increase computational cost and lose detailed
geometrical features. In our numerical results, the effects of different sizes are demonstrated.

2.2.3. Distance metric

An elemental distance metric is defined to evaluate the appearance constraint. Firstly, suppose a mapping function  ∶ 𝛺(𝜓)
̂ ⇒
𝐼(𝛹 ) between all possible patches from 𝛺 and the matching patch from the pattern 𝐼 according to the chosen distance metric. For
the pattern-guided design, we adopt the mathematically simple and intuitive formulation proposed in Ref. [12]. The distance metric
is expressed as
( )2
( ) ∑ 𝜓̂ 𝑒, 𝑗 − 𝛹(𝑒), 𝑗
𝐷𝑖𝑠𝑡 𝜓̂ 𝑒 , 𝛹(𝑒) = 𝜓̂ 𝑒 ∑ ( )2 , (6)
𝑗∈𝜔𝑒 2 𝑘∈𝜔 0.5 − 𝛹(𝑒), 𝑘 +𝜀
𝑒

where the notation 𝜓̂ 𝑒, 𝑗 is used to refer to the filtered density of an element 𝑒 which is in a neighborhood 𝑗 of 𝑒.
A low appearance value indicates a small difference between a patch centered on 𝜓̂ 𝑒 in the design domain 𝛺 and the
corresponding matched patch 𝛹(𝑒) in the pattern domain 𝐼. In an extreme case, where 𝜓̂ 𝑒 = 1, 𝛹(𝑒) = 0, 𝐷𝑖𝑠𝑡max could be up to
2, which rarely happens in practical mapping.

2.2.4. Fast-matching algorithm

To efficiently search a possibly best matching with the given distance metric of Eq. (6), a fast randomized matching algorithm
PatchMatch [47] and Random walk [48] are applied. Ref. [12] further coupled and extended the above methods into a fast-matching
algorithm to handle multiple patterns with a Switch search phase. For computing simplicity, we adopt the fast-matching algorithm
and only consider a single geometrical pattern during the matching process and corresponding design problem.
The fast-matching algorithm is only suitable for patterns appearing without any rotation or scaling. While Ref. [49] proposed a
PatchMatch 2.0, the generalized algorithm could efficiently find global, dense matches across all scales and rotations. Thus, future
work could be made based on PatchMatch 2.0 for much more design freedom along with an inevitable increase in the computational
cost.
We run one execution of the matching process at the start of each topology optimization iteration for evaluating the appearance
constraint. And the propagation for a coherent spreading of the connectivity information of the finite element mesh is performed
by turns in one pass per optimization loop. The propagation algorithm is for matching assignments. It sweeps through the design
domain and visits each element sequentially. The propagation in the original method [12,49] is performed by 2 passes (both forward
and backward) in one time or one optimization loop, while in our implementations we alternately perform 1 pass (either forward
or backward) in one loop for better efficiency with the same matching effect.
After the fast-matching process, every element in the design domain1 with its corresponding patch would be matched to a locally
nearest patch lying inside the pattern according to the distance metric. Then an elemental appearance distribution is founded and
continually improved along with the topology optimization iteration. Readers are referred to the relevant literature for more details
about the matching procedure.

2.2.5. Appearance constraint

After calculating all the elemental distances of Eq. (6) in the design domain as well as the patch match, a simple sum aggregation
is defined and further constrained to reflect and design the structural appearance performance, i.e.,
1 ∑ ( )
𝐴(𝜓)̂ = 𝐷𝑖𝑠𝑡 𝜓̂ 𝑒 , 𝛹(𝑒) ⩽ 𝐴∗ , 𝐴∗ ∈ (0, 1) . (7)
|𝛺| 𝑒
The appearance constraint 𝐴(𝜓) ̂ expects a global average similar pattern feature between 𝜓̂ and 𝛹 , which is quantified as 𝐴∗ .
As discussed in Ref. [12], the topology optimization process is quite sensitive to the choice of 𝐴∗ . A poor choice of the constraint
value would lead to designs exhibiting non-converged densities or unrecognizable geometrical patterns. In our shell-infill design
framework, a proper threshold value of 𝐴∗ is suggested to choose from 0.1 ∼ 0.2 according to different patterns and design
requirements.

1 For the elements near the boundary where corresponding patches are partially outside the design domain, a zero-padding strategy is applied.

6
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Although the appearance function is non-smooth and piece-wise differentiable due to the discontinuous operation of fast-
matching (involving multiple phases: the initialization, switch search, and walk search) [12], the numerical results exhibit a fairly
stable optimization process. This can be attributed to the fast convergence of the matching algorithm. To further improve the
optimization performance and eliminate any numerical instability, a continuation strategy is also employed for the appearance
constraint. The control expression is as follows and updated every 10 iterations,

𝐴∗iter = max(0.5𝐴iter , 𝐴∗ ) , (8)

where 𝐴∗iter and 𝐴iter are the appearance constrain value and appearance function value at the current iteration, respectively.

2.3. Material interpolation

After getting the intermediate density fields (𝜑, 𝜏, 𝜓) as earlier mentioned, the final physical density 𝜌 and corresponding stiffness
𝐸 are defined as interpolations:
[ ]
𝜌(𝜑, 𝜏, 𝜓) = 𝑚0 𝜆𝑚 𝜑𝜓 + (1 − 𝜆𝑚 𝜑𝜓)𝜏 , (9)
[ ]
𝑝 𝑝 𝑝 𝑝
𝐸(𝜑, 𝜏, 𝜓) = 𝐸0 𝜆𝐸min + 𝑞(𝜆𝐸 − 𝜆𝐸𝑚𝑖𝑛 )𝜑 𝜓 + 𝑞(1 − 𝜆𝐸 𝜑 𝜓 )𝜏 , (10)

where a scaling factor 𝜆𝑚 for mass density and a factor 𝜆𝐸 for stiffness between two materials of coating and infill are applied to
the material field. We assume that 𝜆𝐸 satisfies the upper bound of the Hashin–Shtrikman bounds in 2D [50], i.e.,
𝜆𝑚
𝜆𝐸 = . (11)
3 − 2𝜆𝑚
With 𝜆𝑚 contained in the interval [0, 1], the infill is (optionally) made of a lighter, and softer material than the coating material.
To prevent the global stiffness matrix from becoming singular, a small minimum stiffness factor 𝜆𝐸min is also used. 𝑞 is the
penalty parameter only active on the padded domain. 𝑝 is the penalty parameter to penalize intermediate values of 𝜑 and 𝜓. No
penalization for the coating is in order not to penalize low gradients at early design iterations which would easily end up in worse
local minima [24,40].
The interpolations of Eqs. (9) and (10) are further examined here by three (𝜑, 𝜏, 𝜓) cases one of which gives out three material
phases of coating, infill, and void, respectively:
{
𝜌(𝜑, 1, 𝜓) = 1
Coating material (12a)
𝐸(𝜑, 1, 𝜓) = 𝐸0
{
𝜌(1, 0, 𝜓) = 𝜆𝑚 𝜓
Infill material (12b)
𝐸(1, 0, 𝜓) = 𝜆𝐸 𝜓 𝑝 𝐸0
{
𝜌(0, 0, 𝜓) = 0
Void material (12c)
𝐸(0, 0, 𝜓) = 𝐸min

For the numerical examples throughout this paper, the material parameters are set as 𝑚0 = 1, 𝐸0 = 1, 𝐸min = 𝜆𝐸min 𝐸0 = 10−9
with Poisson’s ratio of 0.3. The penalties are set as 𝑞 = 0.2 and 𝑝 = 3.

3. Concurrent topology optimization problem

3.1. Problem model

We consider a compliance (𝑐) minimization problem under the assumption of linear elasticity and plane stress. The design domain
𝛺 is discretized by bi-linear finite elements, and the material properties are assumed to be element-wise constant. The topology
optimization problem is solved in nested form. The objective function  is calculated for each design iteration with the equilibrium
equations satisfied by FE-analysis. Two types of design vectors 𝝁, 𝝊 are updated based on their gradients. A global volume constraint
is added and an appearance constraint imposed on the second design field is introduced to fulfill the pattern-guided design mode.
The discretized optimization problem can thus be written as

min ∶ (𝝁, 𝝊, 𝑼 ) = 𝑐 = 𝑼 𝖳 𝑲𝑼 , (13a)

𝝁, 𝝊

s.t. ∶ 𝑲(𝝁, 𝝊)𝑼 = 𝑭 , (13b)

∗
∶ 𝐴(𝝊) ⩽ 𝐴 , (13c)
∗
∶ 𝐺(𝝁, 𝛾) ⩽ 𝑉 , (13d)
∶ 𝜇𝑒 , 𝜐𝑒 ∈ [0, 1] , ∀𝑒 ∈ 𝛺 . (13e)

7
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Here, 𝑼 , 𝑲, and 𝑭 are displacement vector, stiffness matrix, and force vector, respectively. The functions 𝐴 and 𝐺 represent
appearance and volume constraint, respectively. 𝐴∗ is the maximum appearance constraint value controlling the geometrical
configurations guided by a pre-defined pattern. 𝑉 ∗ is the maximum allowed global volume fraction of the material in the design
domain. The global stiffness matrix 𝑲 is assembled from element stiffness matrix 𝒌𝑒 , which is defined as

𝒌𝑒 = 𝐸𝑒 (𝜑𝑒 , 𝜏𝑒 , 𝜓𝑒 )𝒌0 , (14)

where 𝒌0 is the elemental stiffness matrix for an element with unit Young’s modulus.
For the volume constraint, Ref. [39] employed an approximate one to avoid undesirable large void patches among the infills. As
later shown in the numerical results, we also utilize a global volume constraint based on an approximate elemental volume value 𝛾
to control the overall structural weight as well as the allowed volume for the infill. The global volume constraint used in this paper
is expressed as
1 ∑
𝐺(𝝁, 𝛾) = 𝜌́ 𝑣 ⩽ 𝑉 ∗ , (15)
|𝛺| 𝑒 𝑒 𝑒

where 𝑣𝑒 is the (constant) element volume and 𝜌́𝑒 is an approximate physical density of element 𝑒. The approximate physical density
is calculated as

𝜌́𝑒 (𝜑, 𝜏, 𝛾) = 𝜆𝑚 𝜑𝛾 + (1 − 𝜆𝑚 𝜑𝛾)𝜏 . (16)

And the approximate value 𝛾 in the second design variable field is computed as
𝑛𝑠 𝑛𝑠 ( √ )
1 ∑ 1 ∑ ́ | |
𝛾= 𝜓́𝑠 = 𝛹(𝑠) + 2𝐴∗ |0.5 − 𝛹́ (𝑠) | . (17)
𝑛𝑠 1 𝑛𝑠 1 | |

where 𝜓́𝑠 is the average density value of 𝜓̂ 𝑠 within surrounding neighbors of element 𝑠, and further approximated by the average
one of corresponding matched patch 𝛹́ (𝑠) in the pattern. As the average density values of different patches in the pattern may vary
spatially, uniformly random sampling is employed to determine the final 𝛾. The sampling number is set as 𝑛𝑠 = 0.01|𝛺|.
During the optimization process, the matching procedure proceeds iteratively until a converged and well-matched mapping
between patches of the design structure and corresponding ones in the pattern. Therefore, we will update 𝛾 every 10 iterations until
𝛽1, 2 = 64 to achieve a fine value of approximate volume. Numerical experiments (e.g., see Fig. 8) show that a good match is already
found and the shell-infill structure is stable and ready to converge when 𝛽1, 2 are increased to 64.
Detailed derivations of the approximate volume in Eq. (17) are elaborated in Appendix A. The major idea is that the density
distribution of 𝜓̂ would be exactly the same as the matched patch 𝛹 in the pattern when the appearance value is equal to 0,
i.e. 𝛾(𝐴∗ = 0) = 𝜓̂ = 𝛹 .

3.2. Sensitivity analysis

To solve the optimization problem, the gradient-based Method of Moving Asymptotes (MMA) [51] is used. The derivatives of the
objective function and constraints with respective to elemental design variables 𝜇𝑒 , 𝜐𝑒 are needed. All these sensitivities are given
in the following paragraphs.

3.2.1. Sensitivity analysis of compliance

The sensitivity of the self-adjoint compliance function can be derived as
d𝑐 𝜕𝑲 ∑ 𝜕𝐸𝑖
= −𝑼 𝖳 𝑼 =− (𝒖𝖳 𝒌 𝒖 ) , 𝑥 ∈ {𝜇, 𝜐} . (18)
d𝑥𝑒 𝜕𝑥𝑒 𝑖
𝜕𝑥𝑒 𝑖 0 𝑖
𝜕𝐸𝑖
Considering two sets of design variables, the sensitivity of stiffness interpolation 𝜕𝑥𝑒
is further separated into two parts.

(1) 𝜕𝐸𝑖 ∕𝜕𝜇𝑒 .

Following the chain rule and product rule, we arrive at
( )
𝜕𝐸𝑖 (𝜑, 𝜏, 𝜓) 𝜕𝐸𝑖 𝜕𝐸𝑖 𝜕𝜏𝑖 𝜕𝜑𝑖
= + , (19)
𝜕𝜇𝑒 𝜕𝜑𝑖 𝜕𝜏𝑖 𝜕𝜑𝑖 𝜕𝜇𝑒
with
𝜕𝐸𝑖 [ ] 𝜕𝐸𝑖 [ ]
= 𝑞𝐸0 (𝜆𝐸 − 𝜆𝐸min − 𝜆𝐸 𝜏𝑖 )𝜓𝑖𝑝 𝑝𝜑𝑝−1
𝑖 , = 𝑞𝐸0 1 − 𝜆𝐸 𝜑𝑝𝑖 𝜓𝑖𝑝 . (20)
𝜕𝜑𝑖 𝜕𝜏𝑖
𝜕𝜏 𝜕𝜑𝑖
The sensitivity analysis of 𝜕𝜑𝑖 is elaborated in Appendix B. The derivation of 𝜕𝜇𝑒
is a standard filter modification factor for the two
𝑖
smoothing and two projection steps, which is well-known and omitted here.

8
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

(2) 𝜕𝐸𝑖 ∕𝜕𝜐𝑒 .

Similarly, we have
𝜕𝐸𝑖 (𝜑, 𝜏, 𝜓) 𝜕𝐸𝑖 𝜕𝜓𝑖
= , (21)
𝜕𝜐𝑒 𝜕𝜓𝑖 𝜕𝜐𝑒
with
𝜕𝐸𝑖 [ ]
= 𝑞𝐸0 (𝜆𝐸 − 𝜆𝐸min − 𝜆𝐸 𝜏𝑖 )𝜑𝑝𝑖 𝑝𝜓𝑖𝑝−1 . (22)
𝜕𝜓𝑖
𝜕𝜓𝑖
The derivation of 𝜕𝜐𝑒
is also a standard filter modification factor for the one smoothing and one projection steps.

3.2.2. Sensitivity analysis of appearance

The sensitivity analysis of appearance is given as
( )
d𝐴 1 ∑ 𝜕𝐷𝑖𝑠𝑡 𝑥𝑘 , (𝑘)
= , 𝑥 ∈ {𝜇, 𝜐} , (23)
d𝑥𝑒 |𝛺| 𝑘|𝑒∈𝜔 𝜕𝑥𝑒
𝑘

where element 𝑘 lies in the neighborhood of element 𝑒. Thus a summation result of all the contributions from corresponding patches
𝜔𝑘 is calculated.
As the appearance constraint is only imposed on the second design variable field, we have
( )
𝜕𝐷𝑖𝑠𝑡𝑘 𝜕𝐷𝑖𝑠𝑡𝑘 𝜕𝐷𝑖𝑠𝑡 𝜓̂ 𝑘 , 𝛹(𝑘) 𝜕 𝜓̂ 𝑘
= 0, = , (24)
𝜕𝜇𝑒 𝜕𝜐𝑒 𝜕 𝜓̂ 𝑘 𝜕𝜐𝑒
with
( )
( ) ⎧ 𝐷𝑖𝑠𝑡 𝜓̂ 𝑘 , 𝛹(𝑘) + 𝜓̂ 𝜓̂ 𝑒 −𝛹(𝑘), 𝑒
( )2 , for 𝑘 = 𝑒
𝜕𝐷𝑖𝑠𝑡 𝜓̂ 𝑘 , 𝛹(𝑘) ⎪ 𝜓̂ 𝑘 𝑘∑
𝑗∈𝜔𝑘 0.5−𝛹(𝑘), 𝑗 +𝜀
=⎨ 𝜓̂ 𝑒 −𝛹(𝑘), 𝑒 . (25)
𝜕 𝜓̂ 𝑘 ⎪𝜓̂ 𝑘 ∑ ( )2 , for 𝑘 ≠ 𝑒
⎩ 𝑗∈𝜔𝑘 0.5−𝛹(𝑘), 𝑗 +𝜀

𝜕 𝜓̂ 𝑘
The derivation of 𝜕𝜐𝑒
is still a standard filter modification factor for the two smoothing and one projection steps.

3.2.3. Sensitivity analysis of volume

The sensitivity of the global volume constraint has the form
d𝐺 1 ∑ ∑ 𝜕 𝜌́𝑖
= 𝜌́ , 𝑥 ∈ {𝜇, 𝜐} . (26)
d𝑥𝑒 |𝛺| 𝑗 𝑗 𝑖 𝜕𝑥𝑒

For the two design variables, we similarly have

𝜕 𝜌́𝑖 𝜕𝜑 𝜕𝜏 𝜕𝜑𝑖 𝜕 𝜌́𝑖
= 𝜆𝑚 𝛾(1 − 𝜏) 𝑖 + (1 − 𝜆𝑚 𝛾𝜑𝑖 ) 𝑖 , = 0. (27)
𝜕𝜇𝑒 𝜕𝜇𝑒 𝜕𝜑𝑖 𝜕𝜇𝑒 𝜕𝜐𝑒

3.3. Optimization control parameters

The control parameters used during the optimization are listed as follows unless specified otherwise. The starting design guess
is a uniform density distribution of 𝜇 start = min(𝑉 ∗ ∕𝜆𝑚 , 1), 𝜐start = 1. Parameter continuation is applied for the sharpness parameters
of the projection operations. The parameters of 𝛽1,2 used in the projections for the base and the shell are started at the value of
start = 2, and doubled every 50 iterations until the ending value of 𝛽 end = 128 (or at convergence). Meanwhile, the sharpness of
𝛽1,2 1,2
infill starts from 𝛽3start = 1, and double every 50 iterations until 𝛽3end = 16. The projection thresholds are set as 𝜂1, 2, 3 = 0.5. The
optimization loop is terminated when the maximum change between two consecutive designs (in terms of design variables) is less
than 0.01 or the iteration step reaches a maximum number of 500.

4. Numerical results

For the numerical experiments in this paper, we use three different types of examples for which the loads and boundary
conditions, including padded domains, are shown in Fig. 3. Solid blocks of material are used at the boundaries which guarantee a
minimum feature size of twice the coating thickness, e.g., 2𝑡ref , and reinforce the structure thus preventing load concentrations. 𝐹
is the external loading force, 𝐿 is the length of height, and 𝑑ext is the extension distance.
The numerical implementation is conducted in Python 3.9 running on a laptop equipped with an Intel(R) Core(TM) i9-12900H
@2.50-GHz CPU and 16 GB of RAM. A CPU parallelization utilizing 16 threads is applied in the whole optimization procedures for
enhanced computational performances. Further benefits could come from several more efficient ways, such as using platforms in
C/C++ [52], sub-domain strategy [53], and other acceleration techniques [54].

9
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Fig. 3. Design domains and boundary conditions for different problems considered in this paper, where the extended filter domains are bounded by the dotted
lines.

Fig. 4. Design result of the half-MBB beam guided by a honeycomb pattern.

4.1. Half-MBB beam

For the experiments on the half-MBB beam example, we use a discretization of 400 × 200 finite elements. A global volume
constraint of 𝑉 ∗ = 0.4 and an appearance constraint of 𝐴∗ = 0.1 are used. Other design parameters are 𝑅1 = 0.20𝐿, 𝑡ref = 0.04𝐿, 𝑅3, 4 =
0.01𝐿, 𝜆𝑚 = 1.

4.1.1. Guided by a pattern of hexagons

Firstly, we consider a repeated hexagon pattern 𝛹 with a size of 80 × 50 (in elements) and a patch size of 31 × 31 for the
matching. The optimized design result is given in Fig. 4. Corresponding integrated design results of the shell-infill structure are
shown in Fig. 5. The intermediate density distributions as results of successful filter operations are also illustrated.
In the first row of Fig. 5, a near binary density distribution of 𝜑 after DSP operation is obtained as the base region. And a shell
structure 𝜏 with a uniform thickness is derived from the coating technique. In the second row of Fig. 5, a clear density field 𝜓
is formed after SSP operation and further composed into an enrichment infill with the base region. In the third row of Fig. 5, a
smoothed field 𝜓̂ is designed to exhibit the guided geometrical shape. In the elemental appearance distance result, the average value
is 𝐷𝑖𝑠𝑡mean = 0.100 to satisfy the appearance constraint, while locally larger distance values occur near the loading boundaries for
better load-carrying capabilities. At last, an optimized shell-infill structure 𝜌 filled with recognizable patterns is achieved by the
proposed concurrent design method.
To further verify the approximate volume effect of the proposed Eq. (15), a design result with a global volume constraint based
on the real physical density is compared in Fig. 6. As shown in Fig. 6(a), multiple large void parts appear in the base region,
counteracting the intent to have distributed infill. While in Fig. 6(b) based on the approximate physical density, the infill spreads
over the entire base, making a distinguishable shell–infill composite.
The convergence history of the half-MBB beam design is shown in Fig. 7 for the objective function, global volume constraint, and
appearance constraint. After the volume constraint becomes active, the compliance history shows a smooth decrease along with the
gradually introduced appearance constraint across all optimization iterations, except for an temporary increase when the appearance
constraint value is relatively strict (from below 0.4). After 100 iterations, both the appearance constraint and the volume constraint
are satisfied and stable.
Several intermediate design results during the optimization process are presented in Fig. 8. The infill accompanied by pattern-
guided features comes out at first and the non-penalized coating gradually becomes clear. The continuation strategy on the sharpness
parameter 𝛽 described in Section 3.3 plays an effective role in a stable optimization process.

(1) Effect of shell thickness.

In Fig. 9, the same design problem as above is solved with varying coating thickness. The modeled coating thickness is controlled
by modifying the filter radius 𝑅2 in Eq. (4). The compliance improves when a thicker coating takes up more material 𝑣shell and
thereby the amount of infill material 𝑣infill is reduced.

10
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Fig. 5. The intermediate density distributions, elemental appearance distance, and corresponding pattern involved in the optimized shell-infill structure.

Fig. 6. Shell-infill design results obtained with different global volume calculations based on (a) the real physical density of Eq. (9) and (b) the approximate
physical density of Eq. (16).

Fig. 7. Design history of the half-MBB beam optimization problem.

(2) Effect of infill stiffness.

In Fig. 10(a), the half-MBB beam is optimized for a infill with a relative density of 𝜆𝑚 = 0.7. Both physical density and stiffness
distributions are displayed. As the infill stiffness becomes lower, more amount of material is needed and the base region is thus
getting larger. While for meeting the appearance constraint, the volume constraint of 𝑉 ∗ = 0.4 is inactive. When the volume

11
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Fig. 8. The shell-infill structure results of the half-MBB beam problem during the iterative (iter) optimization process.

Fig. 9. Design results of the half-MBB beam with varying coating thickness.

constraint value is further restricted to 𝑉 ∗ = 0.3, the corresponding result is listed in Fig. 10(b) with an active approximate
volume fraction of 𝑣́ = 0.300. Comparing the compliance results with different upper bounds of volume constraint, it seems that
the shell-infill structure with less real physical material (the actual and accurate material volume of 𝑣) has a lower compliance.
It should be noted that in the less material design result of Fig. 10(b), the ratio of the coating material volume to the composite
one, 𝑟𝑎𝑡𝑖𝑜shell = 𝑣shell ∕𝑣 × 100%, is relatively larger. Hence, the overall stiffness could be better with less amount of weak infill
material.

4.1.2. Guided by a pattern of triangles

Secondly, the half-MBB beam is designed with a pattern of repeated triangles. Different sizes of triangles are considered, and
corresponding optimized results are shown in Fig. 11. Both pattern sizes are 60 × 60 but with different sizes of triangles inside the
patches of 31 × 31. Expected design effects are generated with coated infills which are mostly alike to the given patterns.
From the optimized results, it is demonstrated that the triangle member size in the pattern could directly control the infill’s
feature under the appearance constraint. In traditional topology optimization methods considering feature size control, a designer
could use robust formulation [55], implicit geometric constraint [56], or other complex approaches [57,58]. Although the minimum
size or other geometrical features could be mathematically analyzed [59], it would be more convenient for the designers to guide
the optimization process using a visible blueprint with predetermined geometrical features in the final structural configurations.
In contrast, the pattern-guided design favors a visual and explicit way where designers could intuitively determine the final
design factors from the outset. Furthermore, prior experiences and desired physical performance characteristics could be potentially
incorporated in the pre-defined geometrical patterns. For instance, honeycomb substructures are widely applied in engineering for
thermal protection and vibration isolation [60], while triangular frames are buckling-superior [61,62].
Another advantage of the pattern-guided infill structure enclosed with shells is in manufacturing. These customized shell-infill
composites can be easily realized through additive manufacturing, providing convenience in fabrication and assembly processes. In
Fig. 12, two specimens of the design results are showcased. The printing process is carried out using an Fused Deposition Modeling
(FDM) 3D printer capable of utilizing multiple colors for both the coating and infill components.

12
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Fig. 10. Design results of the cantilever beam with 𝜆𝑚 = 0.7, 𝜆𝐸 = 0.4375, 𝐴∗ = 0.1.

Fig. 11. Design results of the half-MBB beam guided by repeated patterns of triangles with different member sizes.

Fig. 12. 3D printed architectures of the optimized MMB beam, where the coating is orange and infill white.

13
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Fig. 13. Design result of the cantilever beam guided by a pattern of gears [63].

Fig. 14. Design result of the cantilever beam guided by a pattern of waves [64].

4.2. Cantilever beam

In this section, several complex patterns are further tested on a cantilever beam problem. For the experiments on the cantilever
beam example, we use a discretization of 300 × 200 finite elements. A global volume constraint of 𝐺∗ = 0.4, an appearance constraint
of 𝐴∗ = 0.15 and design parameters of 𝑅1 = 0.10𝐿, 𝑡ref = 0.02𝐿, 𝑅3, 4 = 0.01𝐿 are used.

4.2.1. Guided by a pattern of gears

Fig. 13 gives a pattern filled with gears and its corresponding design result. It can be observed that when using a small patch
size, the design infills inherit partial textures from the reference pattern. However, due to the inefficient load-bearing characteristics
of the gear pattern, only a few complete gears are present in the structure, mainly located away from the primary load-carrying
path.

4.2.2. Guided by a pattern of waves

A wave-like pattern and corresponding design result are displayed in Fig. 14. The structural performance and infill appearance
are simultaneously optimized. The infill material is arranged in the design area in the form of wavy patches, where uniform coating
layers are surrounded.

4.3. Multiple loading bridge

A pattern of spider webs in Fig. 15 is used to guide the design of the multiple loading bridge in Fig. 3(c), where five vertical loads
∑
are individually applied. The objective function is modified as the average of five compliances (𝑐mean = 15 5𝑖=1 𝑐𝑖 ), each of which
corresponds to an individual load. We use a discretization of 400 × 200 finite elements. A global volume constraint of 𝑉 ∗ = 0.3 and
an appearance constraint of 𝐴∗ = 0.15 are imposed at first. Other design parameters are 𝑅1 = 0.15𝐿, 𝑡ref = 0.03𝐿, 𝑅3, 4 = 0.01𝐿 and
a relatively weak infill material is assumed with 𝜆𝑚 = 0.8. Corresponding design results with different sizes of matched patches are
illustrated in Fig. 16.
For a relatively small patch size (21 × 21) with respect to the pattern size, only a few geometrical features of webs are presented
in the resulting structure. When the patch size is increased to include more pattern’s elements (31 × 31), the style of the design

14
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Fig. 15. A pattern of spider webs with a size of 90 × 90 and three patches with sizes of 11 × 11, 21 × 21, 31 × 31, respectively.

Fig. 16. Design results of the multiple loading bridge for different patch sizes with 𝐴∗ = 0.15, 𝜆𝑚 = 0.8, 𝑉 ∗ = 0.3.

Table 1
CPU time (unit: second) per optimization iteration in Fig. 16.
Patch size Iteration runtime Matching process Sensitivity analysis
– 23 ∼ 24 – –
21 × 21 60 ∼ 61 5∼6 36 ∼ 37
31 × 31 66 ∼ 67 6∼7 42 ∼ 43

Fig. 17. Design results of the multiple loading bridge with 𝐴∗ = 0.10, 𝜆𝑚 = 1, 𝑉 ∗ = 0.4.

result is quite similar to the given pattern. Such patch size effects are consistent with the ones illustrated in Fig. 15. Corresponding
elemental appearance distribution clearly shows such similarity level over the whole infill region. The average value of similarities
is measured to be 0.15, precisely obeys the appearance design requirement.
A comparison of the computational performances for various patch sizes is provided in Table 1. The matching process is really
fast and not expensive compared to a costly brute-force one (since the searching domain is 400 × 200). The most time-consuming
part is the sensitivity analysis of the appearance constraint. As the neighboring elements in the patch region are involved, the total
sensitivity calculation quantities are 𝐷𝑒𝑠𝑖𝑔𝑛 𝑠𝑖𝑧𝑒 × 𝑃 𝑎𝑡𝑐ℎ 𝑠𝑖𝑧𝑒 × 𝐿𝑜𝑎𝑑 𝑐𝑎𝑠𝑒, e.g., 400 × 200 × 31 × 31 × 5 = 3.844 × 108 !!! As multiple
threads used, the total execution time per optimization iteration is around 1 minute. Hence the computational burden is considered
acceptable and the proposed method is practical for further applications.
To achieve better appearance performance, we further tighten the constraint on the appearance function by setting a lower value
of 𝐴∗ = 0.10. Additionally, we set the infill material property to be the same as the shell, with 𝜆𝑚 = 1, and allow for more volume
with 𝑉 ∗ = 0.4. The patch size is selected as 31 × 31. As shown in Fig. 17, although the maximum elemental distance value increases
a bit, the appearance colormap is more evenly distributed with the mean value decreased. It is worth noting that as the matching
algorithm involves randomness, the optimized result of the pattern-guided infill may converge to a slightly asymmetrical one for
the symmetrical bridge problem. One may just optimize one-half of the design domain or add symmetry constraints on the design
variables to ensure perfect symmetry if required.

15
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

5. Conclusion

We have presented a topology optimization method for designing a shell-infill composite structure, where the exterior coating
layout and the interior infill configuration are concurrently optimized. Two scalar fields of design variables are employed to
simultaneously evolve the solid shells and the enriched infills. In the first field, successive density filtering operations of the double
smoothing and projection approach make a clear distinction between coating, infill, and void. In the second field, a simple constraint
on the appearance of the geometrical features is applied and integrated into the infill region. Based on a compliance minimization
framework with an approximate volume constraint and a predefined pattern, the overall mechanical properties of the composites
could be efficiently optimized with controllable visual performances of the infill material distributions. Numerical results and design
influences of different geometrical patterns, feature sizes, material stiffness ratios, and load conditions are fully illustrated and
comprehensively studied. An intuitive design manner is demonstrated and its capability of manufacturing is also ensured. The
promising idea allows for extensions of the proposed method to 3D or more complex problems, such as strength and buckling design,
thermal-mechanical design, etc. Further application to the topology optimization for bio-inspired design could also be explored in
the future.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared
to influence the work reported in this paper.

Data availability

Data will be made available on request.

Acknowledgments

The authors would like to thank the TopOpt group at the Technical University of Denmark for constructive discussions and also
thank anonymous reviewers for their valuable comments.
This work is supported by National Natural Science Foundation of China under Grant No. 51675525, 11725211, 52005505,
and 12372123. Zeyu Zhang receives financial support from Post-graduate Scientific Research Innovation Project of Hunan Province
under Grant No. CX20220059.
Finally, the authors wish to thank Krister Svanberg for the MMA code.

Appendix A. Approximate volume

Regarding the weighting factor 𝜓̂ 𝑒 in the distance metric of Eq. (6), we have the relationship in each patch as below
( )2 ( )2
∑ 𝜓̂ 𝑒, 𝑗 − 𝛹(𝑒), 𝑗 ∑ 𝜓̂ 𝑒, 𝑗 − 𝛹(𝑒), 𝑗
𝜓̂ 𝑒 ∑ ( )2 ⩽ ∑ ( )2 . (A.1)
𝑗 2 𝑘 0.5 − 𝛹(𝑒), 𝑘 +𝜀 𝑗 2 𝑘 0.5 − 𝛹(𝑒), 𝑘

Assume that a uniformly distributed appearance is realized and a well-matched mapping between the smoothed design field of
𝜓̂ and corresponding pattern field of 𝛹 is founded, thus every patch lying in the design is equivalent to the mapped patch in the
pattern under the constraint value of 𝐴∗ . Then the right term in Eq. (A.1) becomes into
∑( )2 ∑( )2
𝜓̂ 𝑒, 𝑗 − 𝛹(𝑒), 𝑗 = 2𝐴∗ 0.5 − 𝛹(𝑒), 𝑘 . (A.2)
𝑗 𝑘

To further approximate the elemental material volume, we have

| | √ | |
|𝜓́𝑒 − 𝛹́ (𝑒) | = 2𝐴∗ |0.5 − 𝛹́ (𝑒) | , (A.3)
| | | |
where a homogenized density value in the neighborhood is used as
1 ∑ 1 ∑
𝜓́𝑒 = 𝜓̂ , 𝛹́ (𝑒) = 𝛹 . (A.4)
|𝜔| 𝑗 𝑒, 𝑗 |𝜔| 𝑘 (𝑒), 𝑘

For a simple compliance minimized problem, the design tends to occupy more material for better stiffness. Hence Eq. (A.3) is
further derived as
√ | |
𝜓́𝑒 = 𝛹́ (𝑒) + 2𝐴∗ |0.5 − 𝛹́ (𝑒) | . (A.5)
| |
The approximate volume used in Eq. (17) is finally obtained.

16
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

Appendix B. Sensitivity analysis

𝜕𝜏𝑖
Following the chain rule, the sensitivity of 𝜕𝜑𝑖
is written as

𝜕𝜏𝑖 𝜕‖∇𝜑̂𝑖 ‖𝛼 𝜕‖∇𝜑̂𝑖 ‖𝛼 𝜕‖∇𝜑̂𝑖 ‖

= . (B.1)
𝜕𝜑𝑖 𝜕‖∇𝜑̂𝑖 ‖𝛼 𝜕‖∇𝜑̂𝑖 ‖ 𝜕𝜑𝑖
Here the gradient operator is defined as
( ) (𝜕)
∇x 𝜑̂𝑖
∇𝜑̂𝑖 = = 𝜕x 𝑵 𝖳 𝝃̂ 𝑖 = 𝑩 𝝃̂ 𝑖 , (B.2)
∇y 𝜑̂𝑖 𝜕
𝜕y

where 𝑵 is a vector of shape functions and 𝝃̂ 𝑖 is a vector of nodal densities of the second smoothed field (𝝋
̂ N ) corresponding to
element 𝑖. The calculation is further vectorized as

∇x 𝝋
̂ = 𝑮x 𝝋
̂N, ∇y 𝝋
̂ = 𝑮y 𝝋
̂N, (B.3)

where each row 𝑖 of matrix 𝑮 contains the components of 𝑩 at the dof of element 𝑖. The normalized norm of the density gradient
within each element is computed by
√
‖∇𝜑̂𝑖 ‖𝛼 = 𝛼 (∇x 𝜑̂𝑖 )𝖳 (∇y 𝜑̂𝑖 ) . (B.4)
𝜕‖∇𝜑̂𝑖 ‖𝛼
Thus, we have 𝜕‖∇𝜑̂𝑖 ‖
= 𝛼 and
[ ( ) ( )]
𝜕‖∇𝜑̂𝑖 ‖ 1 𝜕∇x 𝜑̂𝑖 𝜕∇y 𝜑̂𝑖
= ∇x 𝜑̂𝑖 + ∇y 𝜑̂𝑖 , (B.5)
𝜕𝜑𝑖 ‖∇𝜑̂𝑖 ‖ 𝜕𝜑𝑖 𝜕𝜑𝑖
where
𝜕∇x 𝜑̂𝑖 𝜕 𝜑̂ N, 𝑖 𝜕∇y 𝜑̂𝑖 𝜕 𝜑̂ N, 𝑖
= 𝑮x , = 𝑮y . (B.6)
𝜕𝜑𝑖 𝜕𝜑𝑖 𝜕𝜑𝑖 𝜕𝜑𝑖
For more details, readers are suggested to Refs. [23,32,33].

References

[1] O. Sigmund, EML webinar overview: Topology optimization—Status and perspectives, Extreme Mech. Lett. 39 (2020) 100855.
[2] S. Loos, S.v.d. Wolk, N.d. Graaf, P. Hekkert, J. Wu, Towards intentional aesthetics within topology optimization by applying the principle of unity-in-variety,
Struct. Multidiscip. Optim. 65 (7) (2022) 185.
[3] C. Wang, B. Xu, Z. Duan, J. Rong, Structural topology optimization considering both performance and manufacturability: strength, stiffness, and connectivity,
Struct. Multidiscip. Optim. 63 (2021) 1427–1453.
[4] L. Zhou, W. Zhang, Topology optimization method with elimination of enclosed voids, Struct. Multidiscip. Optim. 60 (2019) 117–136.
[5] J.L. Barrera, M.J. Geiss, K. Maute, Minimum feature size control in level set topology optimization via density fields, Struct. Multidiscip. Optim. 65 (3)
(2022) 94.
[6] Q. Wang, H. Han, C. Wang, Z. Liu, Topological control for 2D minimum compliance topology optimization using SIMP method, Struct. Multidiscip. Optim.
65 (1) (2022) 38.
[7] J. Dumas, A. Lu, S. Lefebvre, J. Wu, C. Dick, By-example synthesis of structurally sound patterns, ACM Trans. Graph. 34 (4) (2015) 1–12.
[8] C. Schumacher, B. Thomaszewski, M. Gross, Stenciling: Designing structurally-sound surfaces with decorative patterns, in: Computer Graphics Forum. Vol.
35. No. 5, Wiley Online Library, 2016, pp. 101–110.
[9] G.H. Paulino, A.L. Gain, Bridging art and engineering using Escher-based virtual elements, Struct. Multidiscip. Optim. 51 (2015) 867–883.
[10] J. Martínez, J. Dumas, S. Lefebvre, L.-Y. Wei, Structure and appearance optimization for controllable shape design, ACM Trans. Graph. 34 (6) (2015)
1–11.
[11] J. Hu, M. Li, S. Gao, Texture-guided generative structural designs under local control, Comput. Aided Des. 108 (2019) 1–11.
[12] T. Navez, M.-P. Schmidt, O. Sigmund, C.B. Pedersen, Topology optimization guided by a geometrical pattern library, Struct. Multidiscip. Optim. 65 (4)
(2022) 108.
[13] L.A. Gatys, A.S. Ecker, M. Bethge, Image style transfer using convolutional neural networks, in: Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, 2016, pp. 2414–2423.
[14] Y. Jing, Y. Yang, Z. Feng, J. Ye, Y. Yu, M. Song, Neural style transfer: A review, IEEE Trans. Vis. Comput. Graph. 26 (11) (2019) 3365–3385.
[15] P.S. Vulimiri, H. Deng, F. Dugast, X. Zhang, A.C. To, Integrating geometric data into topology optimization via neural style transfer, Materials 14 (16)
(2021) 4551.
[16] W. Zhang, Y. Wang, Z. Du, C. Liu, S.-K. Youn, X. Guo, Machine-learning assisted topology optimization for architectural design with artistic flavor, Comput.
Methods Appl. Mech. Engrg. 413 (2023) 116041.
[17] X. Yan, D. Bao, Y. Zhou, Y. Xie, T. Cui, Detail control strategies for topology optimization in architectural design and development, Front. Archit. Res. 11
(2) (2022) 340–356.
[18] Z. Li, T.-U. Lee, Y.M. Xie, Interactive structural topology optimization with subjective scoring and drawing systems, Comput. Aided Des. 160 (2023)
103532.
[19] D.Q. Ha, J.V. Carstensen, Human-Informed Topology Optimization: Interactive application of feature size controls, Struct. Multidiscip. Optim. 66 (3) (2023)
59.
[20] L.J. Gibson, Cellular solids, Mrs Bull. 28 (4) (2003) 270–274.
[21] J. Wu, N. Aage, R. Westermann, O. Sigmund, Infill optimization for additive manufacturing—approaching bone-like porous structures, IEEE Trans. Vis.
Comput. Graph. 24 (2) (2017) 1127–1140.

17
Y. Li et al. Computer Methods in Applied Mechanics and Engineering 418 (2024) 116485

[22] Z. Jihong, Z. Han, W. Chuang, Z. Lu, Y. Shangqin, W. Zhang, A review of topology optimization for additive manufacturing: Status and challenges, Chin.
J. Aeronaut. 34 (1) (2021) 91–110.
[23] A. Clausen, N. Aage, O. Sigmund, Topology optimization of coated structures and material interface problems, Comput. Methods Appl. Mech. Engrg. 290
(2015) 524–541.
[24] A. Clausen, E. Andreassen, O. Sigmund, Topology optimization of 3D shell structures with porous infill, Acta Mech. Sinica 33 (2017) 778–791.
[25] Y. Luo, Q. Li, S. Liu, Topology optimization of shell–infill structures using an erosion-based interface identification method, Comput. Methods Appl. Mech.
Engrg. 355 (2019) 94–112.
[26] M.Y. Wang, X. Wang, D. Guo, A level set method for structural topology optimization, Comput. Methods Appl. Mech. Engrg. 192 (1–2) (2003) 227–246.
[27] V.J. Challis, A discrete level-set topology optimization code written in Matlab, Struct. Multidiscip. Optim. 41 (2010) 453–464.
[28] Y. Wang, Z. Kang, A level set method for shape and topology optimization of coated structures, Comput. Methods Appl. Mech. Engrg. 329 (2018) 553–574.
[29] J. Fu, H. Li, M. Xiao, L. Gao, S. Chu, Topology optimization of shell-infill structures using a distance regularized parametric level-set method, Struct.
Multidiscip. Optim. 59 (2019) 249–262.
[30] L. Jiang, Y. Guo, S. Chen, P. Wei, N. Lei, X.D. Gu, Concurrent optimization of structural topology and infill properties with a CBF-based level set method,
Front. Mech. Eng. 14 (2019) 171–189.
[31] C. Liu, Z. Du, Y. Zhu, W. Zhang, X. Zhang, X. Guo, Optimal design of shell-graded-infill structures by a hybrid MMC-MMV approach, Comput. Methods
Appl. Mech. Engrg. 369 (2020) 113187.
[32] D. Harvey, P. Hubert, Extensions of the coating approach for topology optimization of composite sandwich structures, Compos. Struct. 252 (2020) 112682.
[33] S. Xu, J. Liu, X. Li, Y. Ma, A full-scale topology optimization method for surface fiber reinforced additive manufacturing parts, Comput. Methods Appl.
Mech. Engrg. 401 (2022) 115632.
[34] O. Sigmund, Materials with prescribed constitutive parameters: An inverse homogenization problem, Int. J. Solids Struct. 31 (17) (1994) 2313–2329.
[35] F. Wang, O. Sigmund, Numerical investigation of stiffness and buckling response of simple and optimized infill structures, Struct. Multidiscip. Optim. 61
(6) (2020) 2629–2639.
[36] C. Wang, X. Gu, J. Zhu, H. Zhou, S. Li, W. Zhang, Concurrent design of hierarchical structures with three-dimensional parameterized lattice microstructures
for additive manufacturing, Struct. Multidiscip. Optim. 61 (2020) 869–894.
[37] J. Fu, H. Li, L. Gao, M. Xiao, Design of shell-infill structures by a multiscale level set topology optimization method, Comput. Struct. 212 (2019) 162–172.
[38] H. Li, L. Gao, H. Li, H. Tong, Spatial-varying multi-phase infill design using density-based topology optimization, Comput. Methods Appl. Mech. Engrg.
372 (2020) 113354.
[39] J. Wu, A. Clausen, O. Sigmund, Minimum compliance topology optimization of shell–infill composites for additive manufacturing, Comput. Methods Appl.
Mech. Engrg. 326 (2017) 358–375.
[40] J.P. Groen, J. Wu, O. Sigmund, Homogenization-based stiffness optimization and projection of 2D coated structures with orthotropic infill, Comput. Methods
Appl. Mech. Engrg. 349 (2019) 722–742.
[41] M. Zhou, Y. Lu, Y. Liu, Z. Lin, Concurrent topology optimization of shells with self-supporting infills for additive manufacturing, Comput. Methods Appl.
Mech. Engrg. 390 (2022) 114430.
[42] B.S. Lazarov, O. Sigmund, Filters in topology optimization based on Helmholtz-type differential equations, Internat. J. Numer. Methods Engrg. 86 (6)
(2011) 765–781.
[43] F. Wang, B.S. Lazarov, O. Sigmund, On projection methods, convergence and robust formulations in topology optimization, Struct. Multidiscip. Optim. 43
(6) (2011) 767–784.
[44] E. Andreassen, A. Clausen, M. Schevenels, B.S. Lazarov, O. Sigmund, Efficient topology optimization in MATLAB using 88 lines of code, Struct. Multidiscip.
Optim. 43 (1) (2011) 1–16.
[45] R.E. Christiansen, B.S. Lazarov, J.S. Jensen, O. Sigmund, Creating geometrically robust designs for highly sensitive problems using topology optimization:
Acoustic cavity design, Struct. Multidiscip. Optim. 52 (2015) 737–754.
[46] A. Clausen, E. Andreassen, On filter boundary conditions in topology optimization, Struct. Multidiscip. Optim. 56 (2017) 1147–1155.
[47] C. Barnes, E. Shechtman, A. Finkelstein, D.B. Goldman, PatchMatch: A randomized correspondence algorithm for structural image editing, ACM Trans.
Graph. 28 (3) (2009) 24.
[48] P.P. Busto, C. Eisenacher, S. Lefebvre, M. Stamminger, et al., Instant texture synthesis by numbers, in: VMV, Citeseer, 2010, pp. 81–85.
[49] C. Barnes, E. Shechtman, D.B. Goldman, A. Finkelstein, The generalized patchmatch correspondence algorithm, in: Computer Vision–ECCV 2010: 11th
European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part III. Vol. 11, Springer, 2010, pp. 29–43.
[50] Z. Hashin, S. Shtrikman, A variational approach to the theory of the elastic behaviour of multiphase materials, J. Mech. Phys. Solids 11 (2) (1963)
127–140.
[51] K. Svanberg, The method of moving asymptotes—a new method for structural optimization, Internat. J. Numer. Methods Engrg. 24 (2) (1987) 359–373.
[52] N. Aage, E. Andreassen, B.S. Lazarov, Topology optimization using PETSc: An easy-to-use, fully parallel, open source topology optimization framework,
Struct. Multidiscip. Optim. 51 (2015) 565–572.
[53] Z.-L. Zhao, Y. Rong, Y. Yan, X.-Q. Feng, Y.M. Xie, A subdomain-based parallel strategy for structural topology optimization, Acta Mech. Sinica 39 (9)
(2023) 422357.
[54] E.A. Träff, A. Rydahl, S. Karlsson, O. Sigmund, N. Aage, Simple and efficient GPU accelerated topology optimisation: Codes and applications, Comput.
Methods Appl. Mech. Engrg. 410 (2023) 116043.
[55] O. Sigmund, Morphology-based black and white filters for topology optimization, Struct. Multidiscip. Optim. 33 (2007) 401–424.
[56] M. Zhou, B.S. Lazarov, F. Wang, O. Sigmund, Minimum length scale in topology optimization by geometric constraints, Comput. Methods Appl. Mech.
Engrg. 293 (2015) 266–282.
[57] E. Fernández, M. Collet, P. Alarcón, S. Bauduin, P. Duysinx, An aggregation strategy of maximum size constraints in density-based topology optimization,
Struct. Multidiscip. Optim. 60 (2019) 2113–2130.
[58] E. Fernández, K.-k. Yang, S. Koppen, P. Alarcón, S. Bauduin, P. Duysinx, Imposing minimum and maximum member size, minimum cavity size, and
minimum separation distance between solid members in topology optimization, Comput. Methods Appl. Mech. Engrg. 368 (2020) 113157.
[59] D. Trillet, P. Duysinx, E. Fernández, Analytical relationships for imposing minimum length scale in the robust topology optimization formulation, Struct.
Multidiscip. Optim. 64 (2021) 2429–2448.
[60] D. Gu, X. Shi, R. Poprawe, D.L. Bourell, R. Setchi, J. Zhu, Material-structure-performance integrated laser-metal additive manufacturing, Science 372 (6545)
(2021) eabg1487.
[61] A. Clausen, N. Aage, O. Sigmund, Exploiting additive manufacturing infill in topology optimization for improved buckling load, Engineering 2 (2) (2016)
250–257.
[62] C.F. Christensen, F. Wang, O. Sigmund, Topology optimization of multiscale structures considering local and global buckling response, Comput. Methods
Appl. Mech. Engrg. 408 (2023) 115969.
[63] Seamless gear pattern in blue colors. https://www.shutterstock.com/zh/image-vector/seamless-gear-pattern-blue-colors-vector-103849379.
[64] Free vector graphic on Pixabay. https://pixabay.com/vectors/background-wallpaper-waves-ocean-2025984/.

18