You are on page 1of 12

Composite Structures 321 (2023) 117305

Contents lists available at ScienceDirect

Composite Structures
journal homepage: www.elsevier.com/locate/compstruct

Designing 2D stochastic porous structures using topology optimisation


Hongjia Lu a , Ting-Uei Lee a , Jiaming Ma a , Da Chen b , Yi Min Xie a ,∗
a
Centre for Innovative Structures and Materials (CISM), School of Engineering, RMIT University, Melbourne 3001, Australia
b
Centre for Infrastructure Engineering and Safety (CIES), School of Civil and Environmental Engineering, The University of New South Wales, New South
Wales 2052, Australia

ARTICLE INFO ABSTRACT

Keywords: Functionally graded porous materials have superior properties such as high strength, low thermal conductivity
Functionally graded porous material and high energy absorption. More importantly, their density distribution can be customised to best suit different
Topology optimisation objectives, thus having a wide range of applications. With different manufacturing approaches, the geometry
Voronoi structure
of functionally graded porous materials can be deterministic (e.g., lattice structure) or stochastic (e.g., metal
foam). Whilst the design of the former has been extensively studied leveraging the long-established topology
optimisation approach, the latter is rarely investigated despite its widespread use in industrial fields, such
as aerospace and biomedical. In this study, we propose a novel two-step topology optimisation framework
to design 2D stochastic porous structures. In the first step, topology optimisation based on compliance
minimisation (or stiffness maximisation) is carried out to obtain the homogenised material density distribution.
In the second step, a new de-homogenisation approach is developed to transform the homogenised materials
into stochastic porous structures. Several numerical examples are presented to demonstrate the effectiveness
of the proposed approach. As a result of employing stochastic materials and the associated density constraint,
the optimised porous structures demonstrate higher compliance compared to the optimised structures made
of solid materials. Nevertheless, the results indicate that the difference in compliance values between the
homogenised structures and de-homogenised porous structures is minor (i.e., less than 6.99%). Furthermore, it
is observed that the structural stiffness is only minimally affected by the introduction of geometric stochasticity
during de-homogenisation, with variations of less than 1.94%. Therefore, the developed topology optimisation
is numerically robust. Various constraints have also been developed, enabling designers to choose structures
with the desired stiffness and geometric complexity from a diverse range of novel designs.

1. Introduction
applications in diverse fields, such as aerospace, biomedical, among
others [3].
In recent decades, due to a growing focus on carbon emissions,
Studies on the mechanical performance of FGPMs have been boom-
the demand for lightweight, high-performance materials has increased.
ing in recent years. For example, Magnucka-Blandzi [4,5] investigated
Among various modern materials, such as polymers, ceramics and
the dynamic stability and axisymmetrical deflection of FGPM plates
high-strength alloys, porous composites including metal foams (e.g.,
using an analytical approach. Gao et al. [6] studied the dynamic
Fig. 1) have attracted much interest in both research and industrial
buckling behaviour of laminated porous beam with different types of
area [1]. This composite material form has superior properties such as
porosity distributions. Chen et al. [7] investigated the elastic buckling
high strength, low thermal conductivity and high energy absorption.
properties of both regular and porous sandwich beams, with an em-
More importantly, they are programmable-material that can best meet
phasis on performance improvement through the use of non-uniform
different design objectives, attributed to their key feature of density-
porosity distributions. Moreover, in a recent study, Chen et al. [8]
dependent properties. By taking advantage of the programmability,
employed representative volume elements to model FGPM structures,
materials with different densities can be combined into new com-
examining Young’s modulus and the buckling resistance of beams with
posite materials, for example, functionally graded porous materials
varying relative densities. It is worth pointing out that, the investi-
(FGPMs) [2]. Remarkably, the behaviour of FGPMs can be manipulated
gated structures are usually designed using the concept of laminated
by creating a gradient in the porosity distribution within a continu-
composites. Specifically, designers need to first divide a beam into
ous design domain. Consequently, these materials have vast potential

∗ Corresponding author.
E-mail address: mike.xie@rmit.edu.au (Y.M. Xie).

https://doi.org/10.1016/j.compstruct.2023.117305
Received 22 November 2022; Received in revised form 16 June 2023; Accepted 22 June 2023
Available online 27 June 2023
0263-8223/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
H. Lu et al. Composite Structures 321 (2023) 117305

ratio [23]. Recently, a clustering approach has been incorporated into


Acronyms the multi-scale approach by Yan et al. [24] to enable the generation
FGPM Functionally graded porous material of structures with enhanced structural rigidity under mechanical and
SIMP Solid isotropic material with penalisation thermal loads, while limiting the number of module types. Li et al. [25]
applied the multi-scale concept with the level-set based optimisation
OC Optimality criteria
approach to design structures with smooth boundaries.
FEA Finite element analysis
While the multi-scale approaches mentioned above directly design
NURBS Non-uniform rational basis spline the material arrangements at multiple scales, another approach is
BESO Bi-directional evolutionary structural opti- focused on first obtaining a homogenised solution through topology
misation optimisation and then using a de-homogenisation step to generate
MBB Messerschmitt-Bölkow-Blohm structures with specific geometric features. For example, Bruggi et al.
[27] designed the density distribution of homogenised structures and
transformed them into porous structures by using circular pores with
different radii. Similarly, Do et al. [28] transformed optimised ho-
several layers and manually assign the relative density value of each mogenised structures into Voronoi diagram-based composites, whereby
layer. This conventional method can only compare limited cases of the density distribution was regulated by adjusting the number of
FGPMs, without comprehensive and in-depth insights into the optimum Voronoi cells. Groen and Sigmund [29] developed an approach to
cellular morphology construction. To fully unleash the potential of obtain manufacturable truss-like structures by generating the structure
porous structure, topology optimisation approach can be used to design based on the principal stress directions obtained with coarse-scale
the material density distribution of FGPM structures to best adapt to dif- solutions. These de-homogenisation studies focus on optimising struc-
ferent application cases, including infrastructure components, aircraft tures with deterministic geometries. With the development of stochastic
parts, and vehicle frames, etc. porous structures, researchers have started applying the multi-scale
Topology optimisation is a long-established approach which targets approaches to stochastic materials.
identifying structures that have optimised performance (e.g., maximum
stiffness, minimum volume) and meet certain conditions. Since the 1.3. De-homogenisation approaches with stochastic geometries
pioneering work of Michell [9], many different optimisation methods
have been developed: density-based approach with the Solid Isotropic When considering stochastic materials, the de-homogenisation step
Material with Penalisation scheme (SIMP) [10,11], Bi-directional Evo- incorporates parameters that involve randomness, allowing for the
lutionary Structural Optimisation (BESO) [12,13] and the level-set generation of various stochastic structures from a single homogenised
methods [14,15]. Porous structure optimisation has attracted signif- solution. In this area of research, Gómez et al. [30] designed stochastic
icant attentions in the field of topology optimisation. On this topic, bone structures based on the relative density distribution of natural
based on whether the structure/material is represented with a single bone tissues and analysed the properties of the resulting structures.
scale, the approaches can be classified as either full-scale or multi-scale. Similarly, Liu et al. [31] tailored the relative density field of a ho-
For the sake of completeness, a brief review of these approaches is mogenised model according to the relative elasticity field obtained from
provided in the following paragraphs. Note that a systematic review finite element analysis (FEA), and then generated Voronoi porous struc-
of this area is recently published in [16]. tures based on the optimisation results. In these studies, the relative
density distributions are deterministic, but the geometric details are
1.1. Full-scale approach stochastic. Consequently, the resulting designs can be manufactured
using metal foam, which has relatively low manufacturing costs in mass
Geometrical constraints are used in full-scale approaches to ensure production [32]. However, while functionally graded porous materials
that the optimised structures possess the desired porous geometrical (FGPMs) are often used in conjunction with solid materials for better
features. Two types of geometrical constraints are commonly used: stiffness, such as in sandwich structures, the preceding studies in this
pattern repetition and local volume constraints. Pattern repetition is section concentrated solely on utilising porous materials in the design
focused on designing structures formed by one or several types of process, potentially limiting the performance of porous structures.
modules [17,18], whereas local volume constraints are used to control In this study, we aim to develop an optimisation approach to con-
the local geometrical features of optimised structures [19,20]. Re- struct structures formed by porous and solid materials by designing the
cently, researchers [21] investigated the incorporation of a length scale overall density gradient, as well as local cellular morphologies (e.g., cell
constraint, which is a type of local volume constraint, in topology opti- size and cell wall thickness). For the sake of simplicity, our current
misation to generate high-stiffness, multi-porosity bone structures. Feng focus is solely on 2D problems. This approach contains two main steps:
et al. [22] has explored the optimisation of Voronoi-based structures, • In the first step (Section 2), we treat porous materials as homoge-
with parametrised and optimised geometrical features to develop high- neous materials with density-dependent properties and optimise
stiffness structures. Whilst full-scale approaches can efficiently produce the density distributions with a customised topology optimisation
optimised porous structures, the inclusion of additional geometrical approach.
constraints typically leads to a reduction in performance compared • In the second step (Section 3), a de-homogenisation approach
to traditional topology optimisation outcomes. Consequently, other is used to transform the homogenised structure into a physi-
studies have adopted the multi-scale approach. cal structure model consisting of cellular morphologies. In this
step, various stochastic structures can be generated with the
same homogenised structure (from the first step) with different
1.2. Multi-scale approach
randomness parameters.
The multi-scale topology optimisation approach focuses on design- This approach is able to provide detailed instructions on the porous
ing structures that incorporate both macro and micro scales, so that structural design to obtain the maximum stiffness under a restricted
properties of the local material and the global structure can both be material consumption. The rest of this article is arranged as follows:
customised. Customising microstructures for individual elements has Section 2 introduces the proposed topology optimisation approach;
the potential to produce materials with exceptional properties, such as Section 3 introduces the de-homogenisation approach used to generate
micro-modules with a Young’s modulus that approaches the theoretical the porous structure; Section 4 includes a few typical examples to verify
Hashin–Shtrikman limit [11], or materials with a negative Poisson’s the proposed approach, and the conclusion is in Section 6.

2
H. Lu et al. Composite Structures 321 (2023) 117305

Fig. 1. Aluminium foam structures comprised of solid materials, porous materials and void regions (image usage permission granted by the open-source license of [26]).

2. Topology optimisation of homogenised structure It is worth noting that in this study we focus on the classical
topology optimisation problem that applies boundary conditions on
2.1. Density-based topology optimisation with solid isotropic material pe- generalised forces. If non-zero prescribed displacements are required
nalisation scheme please refer to the approaches in [36,37].

The topology optimisation approach developed in this study is 2.2. Modifications on the traditional topology optimisation approach
based on the traditional density-based approach with the SIMP scheme.
Therefore, it is first reviewed in this section for clarity. This approach While the traditional SIMP-based approach designs structures com-
prised of solid materials and voids (i.e., 𝜌 ∈ {𝜌min , 𝜌max } with 𝜌min
uses finite element discretisation approach to divide the domain into
donating voids and 𝜌max donating solid materials), in the current study,
numerous elements, and each element’s existence is determined by
we include porous materials in between the voids and solids, that is:
the corresponding element density 𝜌. The traditional 2D optimisation
formulation under a conventional Cartesian coordinate system (x, y) is: 𝜌 ∈ {𝜌min , [𝜌p,min , 𝜌p,max ], 𝜌max }, (4)
T
min 𝐶(𝝆) = 𝐮 𝐊𝐮 (1a) where 𝜌p,min and 𝜌p,max are the minimum and maximum relative den-
𝝆
sities for the porous materials. Note that most previous studies focus
𝑉 (𝝆)
s.t. =f (1b) on high-porosity FGPMs due the fact that they possess innovative
V0
properties and high application potential in various fields (e.g., 𝜌 ∈
𝐊𝐮 = 𝐅 (1c) [0.04, 0.13] in [8]; 𝜌 ∈ [0.16, 0.52] in [38]). Therefore, in the current
𝟎 < 𝝆min ≤ 𝝆 ≤ 𝟏, (1d) study, we assume that 𝜌p,min = 0.04 and 𝜌p,max = 0.5 following existing
works. Different values of 𝜌p,min and 𝜌p,max can be used for various
where 𝐶 is the structural compliance (i.e., the inverse of structural porosity ranges based on the developed approach.
stiffness); 𝝆 = [𝜌1 , 𝜌2 , … , 𝜌𝑛 ]T is the elemental density vector with 𝑛 The traditional SIMP-based topology optimisation approach is mod-
donating the number of elements; 𝐮 = [𝑢1,x , 𝑢1,y , … , 𝑢𝑠,x , 𝑢𝑠,y ]T is the ified in two aspects to design structures that satisfy Constraint (4), as
displacement vector with 𝑠 donating the number of the nodes; 𝐊 rep- introduced in Sections 2.3 and 2.4. Furthermore, two optional con-
resents the stiffness matrix; 𝑉 (𝝆), V0 and f are volume of the optimised straints controlling the optimised structural shape are elaborated in
structure, volume of the design domain and the pre-defined volume Sections 2.5 and 2.6.
fraction, respectively; 𝐅 = [F1,x , F1,y , … , F𝑠,x , F𝑠,y ]T is the external load
vector; 𝝆min = [𝜌min , 𝜌min , 𝜌min , …]T and 𝜌min is a parameter close to 2.3. Material interpolation
zero and used to avoid the numerical singularity problem [10].
In the FEA process, the material interpolation scheme is written as: The effective mechanical properties of stochastic porous materials
have been extensively studied in the literature, and various regression
𝐸𝑒 = E0 𝜌𝑝𝑒 , (2) models of fitted scaling laws have been developed. One frequently used
model is [39]:
of which 𝐸𝑒 is Young’s modulus for the 𝑒th element; E0 is the Young’s
modulus for the base material; 𝑝 is a penalty factor for eliminating 𝐸𝑒 = 𝛼E0 𝜌𝛽𝑒 , (5)
grey elements with densities between 𝜌min and 𝜌max ; usually 𝑝 = 3 is
where the constants 𝛼 and 𝛽 change for different cellular morphologies.
adopted.
In this study, we employ the values 𝛼 = 0.535 and 𝛽 = 1.81, which
With Eq. (2), the constitute matrix can be obtained as:
is demonstrated by [40] to be a reliable estimation for the Young’s
modulus of porous materials with a relative density between 0.026 and
E0 𝜌𝑝𝑒 ⎡⎢
1 𝜈 0 ⎤
𝐃= 𝜈 1 0 ⎥, (3) 0.35.
1 − 𝜈2 ⎢ 1−𝜈 ⎥ Therefore, the elemental Young’s modulus can be obtained by:
⎣0 0 ⎦
2 {
where 𝜈 represents the Poisson’s ratio. E0 𝜌e if 𝜌𝑒 = 𝜌max or 𝜌𝑒 = 𝜌min
𝐸𝑒 = (6)
To solve the optimisation problem, the optimality criteria (OC) 0.535E0 𝜌1.81
𝑒 if 𝜌p,min ≤ 𝜌𝑒 ≤ 𝜌p,max
based approach [33] or the Method of Moving Asymptotes approach
We employ the Poisson’s ratio relation also suggested by [40],
[34] can be adopted. In addition, mesh-independency filters and the
which was derived by fitting data from open-cell foam materials with
Heaviside projection approach are frequently used to improve the
relative densities ranging between 0 and 1:
quality of the optimised results. More details about these approaches
can be found in [10,35]. 𝜈(𝜌𝑒 ) = 0.24 − 0.04𝜌𝑒 (7)

3
H. Lu et al. Composite Structures 321 (2023) 117305

Note that although the data ranges of Eqs. (6) and (7) are different
from the current study (i.e, [0.04, 0.5]), we show that the caused error
is minor in Section 5.

2.4. New OC update scheme

To obtain optimised solutions that satisfy (4), a new variable update


scheme is proposed:
𝜌new =
⎧ 𝜂
⎪𝜌max if 𝜌𝑒 𝐵𝑒 ≥ 𝜌𝑡,𝛼
⎪ 𝜂
⎪min(𝜌p,max , 𝜌𝑒 + 𝑚) if 𝜌p,max < 𝜌𝑒 𝐵𝑒 ≤ 𝜌𝑡,𝛼
⎪ 𝜂 𝜂
(8)
⎨𝜌𝑒 𝐵𝑒 if 𝜌p,min < 𝜌𝑒 𝐵𝑒 ≤ 𝜌p,max

⎪max(𝜌p,min , 𝜌𝑒 − 𝑚) if 𝜌𝑡,𝛽 < 𝜌𝑒 𝐵𝑒 ≤ 𝜌p,min

⎪𝜌min if 𝜌𝑒 𝐵𝑒𝜂 ≤ 𝜌𝑡,𝛽 ,

where 𝜌𝑒 and 𝜌new are the element densities of the current and the
subsequent iterations, respectively; 𝑚 and 𝜂 are step size and damping
factors that are used for convergence control; 𝜌𝑡,𝛼 and 𝜌𝑡,𝛽 are pre-
defined numerical thresholds for modifying the density values located
within the unwanted ranges (i.e., (𝜌p,max , 𝜌max ) or (𝜌min , 𝜌p,min )); for
𝜌 +𝜌 𝜌 +𝜌
simplicity, it is assumed that 𝜌𝑡,𝛼 = p,max2 max and 𝜌𝑡,𝛽 = p,min2 min ;
𝐵𝑒 is found from the optimality condition as:
𝜕𝐶
− 𝜕𝜌
𝑒
𝐵𝑒 = 𝜕𝑉
, (9)
𝜆 𝜕𝜌
𝑒

where 𝜆 is a Lagrangian multiplier that can be determined by a bi-


sectional method (available in a educational MATLAB code of [10]);
the sensitivity of the objective function is found as:
{
𝜕𝐶 −𝐮𝑒 T 𝐤0 𝐮𝑒 if 𝜌𝑒 = 𝜌max or 𝜌𝑒 = 𝜌min
= 𝜕𝐤 (10)
𝜕𝜌𝑒 −𝐮𝑒 T 𝜕𝜌𝑒 𝐮𝑒 if 𝜌p,min ≤ 𝜌𝑒 ≤ 𝜌p,max ,
𝑒

where 𝐮𝑒 is the displacement vector of the 𝑒th element, 𝐤0 denotes the


Fig. 2. Flowchart of the customised SIMP-based approach, where FEA represents finite
local stiffness matrix computed with the constant Young’s modulus E0 element analysis; 𝝆new represents density values used in the subsequent iteration; 𝑉 (𝝆),
and Poisson’s ratio 𝜈0 of the base material; 𝐤𝑒 is the local stiffness matrix V0 , f, ftol are the optimised structure volume, design domain volume, target volume
computed using Eqs. (6) and (7). Note that in the current study, for fraction and numerical volume fraction tolerance, respectively; 𝛥𝐶 and Ct are the
simplicity, 2D square elements with unit side length are used in the change in compliance between iterations and the convergence tolerance, respectively.

optimisation stage. The local 8 × 8 stiffness matrix is available in the


𝜕𝐤
widely used code of [10]. Therefore, the analytical expression of 𝜕𝜌𝑒
𝑒
can be obtained by first substituting Eqs. (6) and (7) into the local
stiffness matrix, and then using symbolic computation to obtain the the bottom-ranked elements (i.e., elements with low 𝜌𝑒 𝜕𝜌 𝜕𝐶
values) are
first-order derivative. 𝑒
assigned with 𝜌min . In addition, during the iterative process the black
and white volume fractions are gradually moved towards the target
2.5. Volume constraints for black and white elements values using the following equations:
{
Whilst Eqs. (6) and (7) are used for the homogenisation of the min(fk,𝑡 (1 + 𝑒𝑟), fk ) if fk,𝑡 ≤ fk
fk,𝑡+1 = (12)
porous materials, numerically, 𝛽 = 1.81 in Eq. (6) tends to push the max(fk,𝑡 (1 − 𝑒𝑟), fk ) if fk,𝑡 > fk ,
{
optimised solution towards black-and-white. Therefore, to control the min(fw,𝑡 (1 + 𝑒𝑟), fw ) if fw,𝑡 ≤ fw
proportion of the porous materials, volume constraints for black and fw,𝑡+1 = (13)
max(fw,𝑡 (1 − 𝑒𝑟), fw ) if fw,𝑡 > fw ,
white elements are introduced:

∀𝜌𝑒 =𝜌max v𝑒 where fk,𝑡 , fw,𝑡 represent the black and white volume fractions for the
= fk (11a) structure at the current iteration; fk,𝑡+1 and fw,𝑡+1 represent the target
V0
∑ black and white volume fraction for the next iteration; 𝑒𝑟 is an evolution
∀𝜌𝑒 =𝜌min v𝑒
= fw , (11b) rate and 𝑒𝑟 = 1% is used in the current study.
V0
where v𝑒 is the volume of the 𝑒th element; fk and fw are the given 2.6. Layer constraint
volume fractions for black and white elements. Note that fk ≤ f and
fk + fw ≤ 1.0. Laminated FGPM beam (i.e., beams with uniform density across
A heuristic approach that inspired by the BESO concept [12] is used each horizontal layer) are frequently considered in previous studies [8,
to apply Constraints (11). In this approach, the elements are ranked 41]. To obtain similar structures in the optimisation stage for the
𝜕𝐶
based on the values of 𝜌𝑒 𝜕𝜌 . Subsequently, the top-ranked elements purpose of comparison, layer constraints are used:
𝑒
𝜕𝐶
(i.e., elements with high 𝜌𝑒 𝜕𝜌 values) are assigned with 𝜌max and 𝜌𝑒 = 𝜌layer,𝑘 , ∀𝑒 ∈ E𝑘 , (14)
𝑒

4
H. Lu et al. Composite Structures 321 (2023) 117305

Fig. 3. Graphic abstract of the de-homogenisation approach: (a) a homogenised FGPM beam; (b) extract black regions; (c) generate Voronoi seeds with a coarse-and-fine subdivision
approach; (d) an iterative process to generate porous structures with the target volume fractions; (e) assemble porous and solid structural parts; 𝑉p , V0 , f, ftol are the porous structure
volume, design domain volume, target volume fraction and numerical volume fraction tolerance.

where 𝜌layer,𝑘 corresponds to the density of the 𝑘th layer; E𝑘 is an 3.1. Voronoi diagram
elemental index set which contains the indices for the elements in the
𝑘th layer. For the de-homogenisation approach, it is vital to establish an
In the optimisation process, Constraint (14) is considered by assign- effective connection between the local density and the porous structure
ing the same value of 𝜌layer,𝑘 for all the elements in the 𝑘th layer. The generated via Voronoi diagram (Voronoi structure). In view of the fact
sensitivity of the compliance with respective to the layer density reads: that its volume is determined by the total cell wall length (Lw ) and the
𝜕𝐶 ∑ 𝜕𝐶 cell wall thickness (𝑡c ), the following relation stands when fixing 𝑡c is
= , (15)
𝜕𝜌layer,𝑘 𝑒∈E 𝜕𝜌𝑒 used for all the Voronoi cells:
𝑘

𝑉local ⋅ 𝜌local = Lw, local ⋅ 𝑡c , (16)


2.7. Summary of the customised topology optimisation approach
where 𝑉local , 𝜌local and Lw, local are the volume, relative density and
With the modifications in Sections 2.3 to 2.6, this new approach is total cell wall length of the target local region. Here, the volume
able to obtain structures that satisfy Constraint (4). It is worth noting intersections at joint positions are ignored for simplicity.
that the discontinuity in Constraint (4) and Eq. (6) cause numeri- In addition, Chen et al. [8] suggests the following relation for a
cal difficulties such as convergence problem or sub-optimal solutions. Voronoi structure within a square region:
Therefore, to ensure a stable optimisation process, firstly we start the √
optimisation process with the traditional SIMP-based method with 𝑝 = Lw, local ∝ 𝑁c , (17)
3 for 10 iterations to obtain a reliable starting point; secondly, we where 𝑁c is the Voronoi cell number in the concerned square region.
employed 𝜂 = 0.1 in Eq. (8) for better convergence control. Steps in By combining (16) and (17), one may obtain:
Sections 2.3 to 2.6 are carried out within the inner loops of the OC √
approach. For clarity, a flow chart of the new approach is shown in 𝜌local ∝ 𝑁c , (18)
Fig. 2.
Therefore, according to (18), the density of a local square region can
be controlled by adjusting the corresponding Voronoi cell numbers.
3. A de-homogenisation approach to generating porous structure
Following this finding, we propose a coarse-and-fine domain sub-
division process (i.e., Fig. 4) to generate the graded Voronoi diagram
With the density distribution optimised by the approach in Sec-
according to the optimised density distribution. In the coarse subdivi-
tion 2, we now use a de-homogenisation approach to transform the
sion step (Fig. 4(b)), the overall design domain is divided into a number
homogenised structures into porous structures. For clarity, the steps
of square regions with side length Lc ; and the local mean density 𝜌 of
of the de-homogenisation approach are demonstrated with a simple
each region is calculated by averaging the densities of the optimisation
FGPM beam shown in Fig. 3(a). In the first step, considering that the
geometry stochasticity only exists in the porous structures, we separate elements located within the region of interest. In the fine subdivision
the grey (porous) regions from the black (solid) and white (void) step (Fig. 4(c)), each square region is further divided into a number of
regions (Fig. 3(b)). Note that to avoid zigzag boundaries, a simple equal-sized square cells; and following (18), the cell number in each
moving average filter (i.e, Chapter 15 of [42]) is used to smooth the square region (i.e., 𝑁c ) is determined by:
black and white regions in topology-optimised structures. Although this 𝑁c = max(1, 𝜌 ⋅ NL,max )2 , (19)
simple technique may introduce errors, we demonstrate in Section 5
that these errors are negligible. For applications that require higher where NL,max is a pre-defined parameter used to control the complexity
accuracy, alternative approaches such as the one proposed in [43] may of the porous structure. As shown in Fig. 4(c), larger NL,max leads more
be considered. Voronoi cells, which further leads to more complex porous structure.
After the region separation, we follow the approach in [44,45] and Here, each square region contains at least one square cell to avoid
use Voronoi diagrams to generate the stochastic porous geometries. the formation of overly large internal pores, differentiating the porous
In this process, the cellular morphology is generated according to material part from the preset void part.
the density distribution from the corresponding homogenised structure After the coarse-and-fine subdivision step, the standard uniform
(Fig. 3(c)). Then an iterative approach is used to construct the porous distribution is adopted to place one Voronoi seed in each square cell.
structure with the target volume (Fig. 3(d)). These two steps are Then the Voronoi diagram can be constructed with the Voronoi seeds.
comprehensively described in Sections 3.1 and 3.2. It is worth noting that to obtain a smooth transition between different

5
H. Lu et al. Composite Structures 321 (2023) 117305

Fig. 5. Generate internal pores using the Voronoi cells, where 𝑡c is the fixed offset
distance for all cell walls.

Fig. 6. Flowchart for generating porous structures with target volumes, where 𝑡c is the
fixed offset distance for all the Voronoi cells; 𝑉p , V0 , f, ftol are the porous structure
volume, design domain volume, target volume fraction and numerical volume fraction
tolerance.

Fig. 4. Generation of Voronoi seeds: (a) a FGPM beam with six layers (a selected part
for illustration); (b) domain coarse subdivision; (c) domain fine subdivision for Voronoi basis spline of each offsetted cell boundary. With this approach, the
seeds; Lc is the given length of the square sub-regions; NL,max is a given parameter to volume of the optimised structure is related to a single parameter 𝑡c .
control the number of Voronoi seeds; 𝜌 is the average density of the concerned square Therefore, by calculating the 𝑡c value using a bi-sectional approach, a
region. porous structure with a target volume fraction can be obtained. For
the sake of clarity, a flow chart of this approach is shown in Fig. 6. It
is worth noting that the volume of the porous structure is calculated
regions (i.e., the transition between black and grey regions, and the by subtracting the volume of all voids (i.e., including those for porous
transition between white and grey regions), although the Voronoi holes in Fig. 5 and the white regions obtained in the optimisation stage)
diagram is built up for the whole design domain, only the cellular from the volume of the design domain.
morphologies in the grey regions are kept.
4. Numerical examples
3.2. Generate porous structure
Several examples are presented in this section to demonstrate the
With the Voronoi diagram created, the next step is to establish the effectiveness of the proposed approach and to examine the performance
actual geometry of the porous structure. Here we use an approach of generated FGPM structures. In these examples, 𝜌max = 1.0, 𝜌min =
similar to [44–46]. In this approach (i.e., Fig. 5), firstly, each Voronoi 0.001; E0 = 1 is used for the base material of both the optimisation
cell is offsetted towards inside by a uniform distance 𝑡c ; secondly, and the FEA; Eq. (7) is used for the Poisson’s ratio in the optimisation
the internal pore shapes are constructed by interpolating a third-order and 𝜈 = 0.3 is used in the FEA; unit load P is used in all cases; and

6
H. Lu et al. Composite Structures 321 (2023) 117305

Fig. 7. MBB beam example: (a) case description and optimised homogenised structure; (b) smoothed black, white and grey regions of the optimised structure; (c) corresponding
Voronoi diagram; (d) generated porous structure; (e) triangular mesh used in the finite element analysis process.

Fig. 8. Convergence history of the MBB beam example, where the traditional SIMP-based approach is carried out between iterations 1 to 10 and the customised approach starts
at the 11th iteration.

ftol = 1 × 10−3 and Ct = 1 × 10−4 for the approaches in Figs. 2 and 6. The and the associated cell offset length 𝑡c is 0.112mm. In the final step,
compliance of the optimised porous structure is evaluated using FEA this porous structure is analysed using FEA, where 759,706 quadratic
via Abaqus 2021 [47]. triangular shell elements (STRI65 element) are used (i.e., Fig. 7(e)). The
compliance obtained through FEA is 65.48 N⋅mm. In comparison to the
4.1. A simple example to demonstrate the process compliance presented in Fig. 7(a), the disparity is merely 6.99%.
A convergence history graph is shown in Fig. 8. This example takes
A simple Messerschmitt-Bölkow-Blohm (MBB) beam is used in this 22 iterations to converge. In the first 10 iterations, since the traditional
section to demonstrate the proposed approach. The case description is SIMP-based approach is used, there exist no clear boundary between
shown in Fig. 7. In the topology optimisation stage, 180 × 30 elements solid and grey regions. The modified update scheme starts from the
are applied and the volume fraction f = 0.5. The optimised structure 11th iteration, which introduce clear boundary between the solid and
is shown in Fig. 7(a), which has a compliance equal to 61.2 N⋅mm. grey regions. The iterative process then takes another 11 iterations to
Since it is assumed that 𝜌 ∈ {0.001, [0.04, 0.5], 1.0}, clear boundaries reach the convergence tolerance Ct = 1 × 10−4 .
exist between black (i.e., 𝜌 = 1.0), white (i.e., 𝜌 = 0.001) and grey
regions (i.e., 𝜌 ∈ [0.04, 0.5]), leading to an easy region separation.
In Fig. 7(b), a smoothing step (Section 3) is conducted to eliminate 4.1.1. Stiffness comparison
zigzag boundaries between the black and white regions. Note that To test the stiffness of the optimised porous structure, the MBB
the optimised density distribution within the grey region is preserved, beam example is designed with different approaches and various porous
although not depicted in Fig. 7(b) for the sake of clarity. In the coarse- geometries in this section. To achieve a fair comparison, all the struc-
and-fine subdivision step, Lc = 3 mm and NL, max = 6 are used, and tures in this section possess the same volume (i.e., 2700 mm2 ), and
the generated Voronoi diagram is shown in Fig. 7(c). Subsequently, the their normalised stiffnesses are revealed in Fig. 9. The lowest stiffness
process in Fig. 6 is employed to construct the porous structure with f = is linked with Fig. 9(e), a six-layer FGPM beam. The layer densi-
0.5 (Fig. 7(d)). The constructed structure contains 1493 porous holes, ties are denoted by 1.0, 0.4, 0.1, 0.1, 0.4, 1.0 from top to bottom

7
H. Lu et al. Composite Structures 321 (2023) 117305

(i.e., Fig. 3). The stiffness of this laminated beam provides the normal-
isation benchmark (i.e., 𝑆0 ). Whilst Fig. 9(e) represents a conventional
design with manually preset layer thickness and densities, the op-
timisation approach (with layer constraint described in Section 2.6)
is introduced to optimise the density of each layer as depicted in
Fig. 9(d). It can be observed that from Fig. 9(e) to (d), the top and
bottom solid regions are reduced, while more materials are added
into the middle region, leading to a 0.94𝑆0 increase in stiffness. From
Fig. 9(d) to (c), the layer constraint is released, and hence a free-form
structure is gained with another 0.44𝑆0 stiffness increase. The volume
constraints for black and white elements (i.e., Section 2.5) are then
introduced in Fig. 9(b), which bring another 0.36𝑆0 stiffness increase.
Finally, we include a (near-) optimal benchmark solution computed
with the BESO approach (Fig. 9(a)) using 180 × 30 elements. The
stiffness of the benchmark solution is 19.6% higher than the structure
in Fig. 9(b). The notable distinction between Figs. 9(a) and (b) is in
line with previous research [11], which demonstrates that materials
possessing intermediate densities exhibit less efficiency with respect to
Young’s modulus when compared to solid materials, in accordance to
the Hashin–Shtrikman bound. Furthermore, since this study focuses on
stochastic porous materials, the Young’s modulus is even lower than
the Hashin–Shtrikman bound.
Nevertheless, Fig. 9 show that the proposed approach can lead to
diversed designs with different geometrical complexities. Although the
designs in Figs. 9(b-c) have relatively low stiffness compared with
Fig. 9. Performance comparison of conventional and optimised designs: (a) an opti-
the design in Fig. 9(a), they present the potential for achieving low
mised beam using the traditional BESO approach; (b) an optimised porous beam with
manufacturing expenses. Particularly, for the design in Fig. 9(c), the fk = 0.35 and fw = 0.35; (c) an optimised porous beam without volume constraints
infill with graded densities may be manufactured using the metal foam for black and white elements activated; (d) an optimised porous beam with layer
fabrication technique [48]. In addition, although the porous material constraints; (e) a conventional 6-layer functional graded porous beam (i.e., Fig. 3);
components in Fig. 9(b) have free-form shapes, by setting fk = fw , all the structural stiffness are normalised against the stiffness of (e) (i.e., 𝑆0 ); fk and
fw represent the volume fractions for solid and void elements used in the topology
all the porous material will have the same density of 0.5 (i.e., see
optimisation stage.
detailed reason in Section 4.4). Consequently, the porous and solid
components can be generated using subtractive techniques, and sub-
sequently integrated via the approach outlined in [49]. Future studies
varying complexities, during the homogenisation stage, NL,max values
may consider varies geometry constraints and clustering approaches to
reduce the manufacturing costs. However, these are beyond the scope of 4, 6, 8 and 10 are utilised. The results are disclosed in Figs. 12(b–e);
of the present study. which have 1631, 3304, 5465 and 8306 porous holes, respectively. As
mentioned in Section 3.2, the cell offset distance 𝑡c is identified with a
4.2. Influence of stochastic parameters bi-sectional method for achieving porous structures with target volume.
As the structural volumes in Figs. 12(b–e) remain constant, a reduced 𝑡c
A cantilever example is investigated in this section to quantify the value is employed for the configuration with a greater number of cells
influence of geometrical stochasticity on the structural compliance. The (i.e., 𝑡c equals to 0.092 mm, 0.063 mm, 0.054 mm and 0.041 mm for
case description and the homogenised optimised solution (i.e., obtained Figs. 12(b–e), respectively). The calculated compliances for Figs. 12(b–
with 80 × 128 elements and f = 0.5) are shown in Fig. 10(a). It takes e) are 8.39 N⋅mm, 8.31 N⋅mm, 8.25 N⋅mm and 8.23 N⋅mm accordingly.
34 iterations for the algorithm to reach a convergence. Note that the This example shows that once the homogenised structure is determined,
volume constraints for black and white elements are not used in this the compliance variation caused by the porous structure complexity
example. In the de-homogenisation step, Lc = 4 mm and NL,max = 6 can also be negligible (in this example, compliance variation is smaller
in the coarse-and-fine subdivision step. By using randomly positioned than 1.94%). Therefore, the proposed approach is applicable to various
Voronoi seeds after the fine subdivision step, 20 different porous struc- porous materials with different ranges of pore sizes.
tures are generated (i.e., with the same optimised homogenised solution
in Fig. 10(a)). For the sake of clarity, we show five of them in Figs. 10(b- 4.4. Influence of black and white volume
f). The compliances of all 20 porous structures are illustrated in Fig. 11.
It can be seen that the variation of the compliances is small (i.e., within In this section, a cantilever example is used to investigate the influ-
[−0.68%, 0.69%]). This suggests that the influence of the geometrical ence of fk (i.e., the black volume fraction in Eq. (11a)) and fw (i.e., the
stochasticity on compliance appears to be negligible, highlighting the white volume fraction in Eq. (11b)) on the structural compliance. The
numerical robustness of the proposed approach. case description is indicated in Fig. 13. In the topology optimisation
stage, the design domain is meshed into 140 × 70 elements and f =
4.3. Influence of Voronoi cell number 0.5. The cases take between 80–120 iterations to converge because a
relatively small evolution rate of 𝑒𝑟 = 1% is used in Eqs. (12) and
To access the effect of the Voronoi cell number on the structural (13). In the de-homogenisation stage, Lc = 7 mm and NL,max = 7. The
compliance, a ‘half wheel’ example is investigated in this section. The optimised results with various values for fk and fw are shown in Fig. 14.
case description is displayed in Fig. 12(a). In the topology optimisation Note that when f = 0.5 and 𝜌p,max = 0.5, feasible structures only exist
stage, the design domain is discretised into 120 × 60 elements and for fk ≥ fw (see Condition (22)).
f = 0.5 is used. It takes 25 iterations for the algorithm to achieve It can be observed that the structural compliance tends to decrease
convergence. The obtained solution is shown in Fig. 12(a), which has as fk and fw increase, suggesting a higher stiffness with less porous
a compliance equal to 8.81 N⋅mm. To generate porous structures with materials while more solid components (similar to the findings from

8
H. Lu et al. Composite Structures 321 (2023) 117305

Fig. 10. Cantilever example: (a) case description and optimised homogenised structure, 𝐶 = 22.26 N⋅mm; (b) stochastic design 1, 𝐶=23.41 N⋅mm; (c) stochastic design 2,
𝐶=23.62 N⋅mm; (d) stochastic design 3, 𝐶=23.55 N⋅mm; (e) stochastic design 4, 𝐶=23.45 N⋅mm; (f) stochastic design 5, 𝐶=23.56 N⋅mm; 𝐶 represents the structural compliance
(i.e., 1/stiffness).

Fig. 11. Compliances for the porous cantilevers generated from the same homogenised structure (i.e., Fig. 10(a)) and different Voronoi seed locations.

Fig. 12. Half wheel example: (a) case description and optimised homogenised structure, 𝐶 = 8.81 N⋅mm; (b) porous structure obtained with NL,max = 4, 𝐶 = 8.39 N⋅mm; (c) porous
structure obtained with NL,max = 6, 𝐶 = 8.31 N⋅mm; (d) porous structure obtained with NL,max = 8, 𝐶 = 8.25 N⋅mm; (e) porous structure obtained with NL,max = 10, 𝐶 = 8.23 N⋅mm;
a higher NL,max value leads more porous holes; 𝐶 represents the structural compliance (i.e., 1/stiffness).

9
H. Lu et al. Composite Structures 321 (2023) 117305

Table 1
Compliances for the homogenised structures (𝐶H ) and de-homogenised porous structures (𝐶P ).
Fig. 7 Fig. 9(b) Fig. 9(c) Fig. 9(d) Fig. 9(e) Fig. 10(b) Fig. 10(c)
𝐶H (N⋅mm) 61.20 56.81 61.20 79.97 153.63 22.26 22.26
𝐶P (N⋅mm) 65.48 55.37 65.48 80.37 156.67 23.41 23.62
Difference 6.99% 2.53% 6.99% 0.59% 1.98% 5.16% 6.11%
Fig. 10(d) Fig. 10(e) Fig. 10(f) Fig. 11(b) Fig. 11(c) Fig. 11(d) Fig. 11(e)
𝐶H (N⋅mm) 22.26 22.26 22.26 8.81 8.81 8.81 8.81
𝐶P (N⋅mm) 23.55 23.45 23.56 8.39 8.31 8.25 8.23
Difference 5.80% 5.35% 5.84% 4.77% 5.67% 6.36% 6.58%
Fig. 14(1) Fig. 14(2) Fig. 14(3) Fig. 14(4) Fig. 14(5) Fig. 14(6) Fig. 14(7)
𝐶H (N⋅mm) 39.81 40.96 39.75 42.44 40.96 39.87 44.42
𝐶P (N⋅mm) 37.80 38.83 38.56 40.33 39.58 39.30 42.03
Difference 5.35% 5.46% 3.06% 5.24 % 3.48 % 1.47 % 5.70 %
Fig. 14(8) Fig. 14(9) Fig. 14(10) Fig. 14(11) Fig. 14(12) Fig. 14(13) Fig. 14(14)
𝐶H (N⋅mm) 42.55 41.09 39.99 46.59 44.50 42.58 41.15
𝐶P (N⋅mm) 40.97 40.41 40.10 43.56 43.38 42.54 41.34
Difference 3.86 % 1.67% 0.27% 6.96 % 2.61 % 0.10% 0.44%
Fig. 14(15) Fig. 14(16) Fig. 14(17) Fig. 14(18) Fig. 14(19) Fig. 14(20) Fig. 14(21)
𝐶H (N⋅mm) 40.06 48.97 46.34 44.48 42.65 41.21 40.12
𝐶P (N⋅mm) 41.25 46.18 44.41 44.10 43.65 42.08 42.13
Difference 2.90 % 6.04 % 4.36 % 0.88 % 2.30 % 2.06% 4.77%

Fig. 9). It is also worth noting that the range of the density variation
within the grey regions can be reduced by using a certain combination
of fk and fw . This is illustrated in Fig. 13, where varied grey densities
(i.e., 0.04 ≤ 𝜌𝑝 ≤ 0.5) are used in Fig. 13(a) and uniform grey density
(i.e., 𝜌𝑝 = 0.5) is given in Fig. 13(b). To be more specific, we analyse
this change analytically in the following process. Firstly, the domain
volume is composed of three parts:

fk + fk + fp = 1, (20)

where fp is the volume fraction of porous structure.


In addition, the structure volume (apart from the void part) consists
of those from solid and porous materials:

fk + 𝜌p fp = f, (21)

where 𝜌p is the average density for the grey regions.


Since 𝜌p ≤ 𝜌p,max , the following condition must be satisfied (i.e., ob-
tained via combining Eqs. (20) and (21)):
f − fk
𝜌p = ≤ 𝜌p,max (22)
1 − fk − fw
In the current case, substitute f = 0.5 and 𝜌p,max = 0.5 into (22)
leads to: fk ≥ fw and 𝜌p = 0.5 when fk = fw . Condition (22) can be
used to build up an optimised structure with uniform porous density,
potentially reducing manufacturing costs.

Fig. 13. Optimised homogenised cantilevers with different volume fractions for solid
5. Discussion
elements (i.e., fk ) and void elements (i.e., fw ): (a) solution with fk = 0.32 and fw = 0.2;
(b) solution with fk = 0.32 and fw = 0.32.
In this study, a homogenisation-based approach is used in the
optimisation process. Since the adopted Eqs. (6) and (7) are both
obtained through data-fitting approach, there is a possibility of er-
6. Conclusion
rors arising in the homogenisation step. Therefore, the compliances
of the homogenised structures and the generated porous structures
are summarised and compared in Table 1. It can be seen that for all In this study, a novel topology optimisation-based approach is pro-
the cases investigated, the maximum compliance difference is 6.99%, posed to design the density distributions of 2D functional graded porous
suggesting that the errors introduced in the homogenisation and de- material (FGPM) structures. This approach contains two steps. In the
homogenisation steps are negligible. first step, the traditional SIMP-based topology optimisation approach
In addition, the results in Sections 4.2 and 4.3 show that once the is customised so that an optimised homogenised structure comprised of
homogenised solution is determined, the stochasticity introduced in the solid, void and porous materials can be obtained. The second step uses
de-homogenised step affects little on the global structural performance a Voronoi structure-based technique to generate a stochastic porous
(i.e., maximum influence less than 1.94%), highlighting the robustness structure based on the optimised density distribution from the first
of the adopted stochastic model. step. Note that in the second step, the volume of the porous struc-
While this study focuses on 2D problems, the developed optimisa- ture is determined by a single parameter (i.e., the cell offset distance
tion scheme can be extended to 3D problems in future research. During 𝑡c ). Therefore, the volume of the porous structure can be precisely
the optimisation stage, the values of 𝜌p,min and 𝜌p,max and Eqs. (6) controlled by applying a bi-sectional method on 𝑡c .
and (7) can be adjusted to achieve the desired cellular morphology. The effectiveness of the proposed approach is demonstrated through
Additionally, the Voronoi diagram-based de-homogenisation scheme several numerical examples, and the following conclusions can be
can be customised to apply to either open or closed-cellular structures. drawn:

10
H. Lu et al. Composite Structures 321 (2023) 117305

Fig. 14. Compliances for the porous cantilever structure designed with different volume fractions for solid elements (fk ) and void elements (fw ); 𝐶 represents the structural
compliance (i.e., 1/stiffness).

• The compliance differences between the homogenised and gener- Data availability
ated porous structures are negligible in all investigated examples,
with a maximum difference of only 6.99%. This highlights the high Data will be made available on request.
accuracy of the homogenisation approach.
• Once the optimised homogenised structure is determined, different Acknowledgements
parameters in the de-homogenisation step may lead to porous
structures with various geometrical shapes. It is found that in the This research is financially supported by Australian Research Coun-
cases investigated, the stochastic parameters impose a minor influ- cil (FL190100014, DE220100876).
ence on the structural stiffness (i.e., less than 1.94%), highlighting
the robustness of the proposed de-homogenisation approach. References
• The considered stochastic porous material displays a lower Young’s
modulus than the theoretical Hashin–Shtrikman bound, resulting [1] Tan WC, Saw LH, San Thiam H, Xuan J, Cai Z, Yew MC. Overview of porous
in decreased stiffness values (19.6% in the investigated MBB beam media/metal foam application in fuel cells and solar power systems. Renew
example) for the optimised porous structures compared to those Sustain Energy Rev 2018;96:181–97. http://dx.doi.org/10.1016/j.rser.2018.07.
made from solid materials and created using the traditional topol- 032.
[2] García-Moreno F. Commercial applications of metal foams: Their properties and
ogy optimisation approach. Nevertheless, compared to traditional
production. Materials 2016;9(2):85. http://dx.doi.org/10.3390/ma9020085.
FGPM structures, by optimising the density distribution using the [3] Shi H, Zhou P, Li J, Liu C, Wang L. Functional gradient metallic biomaterials:
proposed approach, the mechanical stiffness can be increased by Techniques, current scenery, and future prospects in the biomedical field.
175% (i.e., in the investigated MBB beam example). In addition, Front Bioeng Biotechnol 2021;8:616845. http://dx.doi.org/10.3389/fbioe.2020.
various constraints can be used to generate diverse designs with 616845.
[4] Magnucka-Blandzi E. Axi-symmetrical deflection and buckling of circular porous-
different stiffness and geometrical complexities.
cellular plate. Thin-Walled Struct 2008;46(3):333–7. http://dx.doi.org/10.1016/
j.tws.2007.06.006.
CRediT authorship contribution statement [5] Magnucka-Blandzi E. Dynamic stability of a metal foam circular plate. J Theoret
Appl Mech 2009;47(2):421–33.
Hongjia Lu: Methodology, Software, Investigation, Writing – orig- [6] Gao K, Huang Q, Kitipornchai S, Yang J. Nonlinear dynamic buckling of
inal draft. Ting-Uei Lee: Methodology, Structure analysis, Writing functionally graded porous beams. Mech Adv Mater Struct 2021;28(4):418–29.
http://dx.doi.org/10.1080/15376494.2019.1567888.
– review & editing. Jiaming Ma: Methodology, Structure analysis,
[7] Chen D, Yang J, Kitipornchai S. Elastic buckling and static bending of shear
Writing – review & editing. Da Chen: Conceptualisation, Methodol- deformable functionally graded porous beam. Compos Struct 2015;133:54–61.
ogy, Writing – review & editing. Yi Min Xie: Supervision, Project http://dx.doi.org/10.1016/j.compstruct.2015.07.052.
administration, Conceptualisation, Writing – review & editing. [8] Chen D, Rezaei S, Rosendahl PL, Xu B-X, Schneider J. Multiscale modelling of
functionally graded porous beams: Buckling and vibration analyses. Eng Struct
Declaration of competing interest 2022;266:114568. http://dx.doi.org/10.1016/j.engstruct.2022.114568.
[9] Michell AGM. The limits of economy of material in frame-structures. Phil Mag
1904;8(47):589–97. http://dx.doi.org/10.1080/14786440409463229.
The authors declare that they have no known competing finan- [10] Sigmund O. A 99 line topology optimization code written in mat-
cial interests or personal relationships that could have appeared to lab. Struct Multidiscipl Optim 2001;21(2):120–7. http://dx.doi.org/10.1007/
influence the work reported in this paper. s001580050176.

11
H. Lu et al. Composite Structures 321 (2023) 117305

[11] Bendsøe MP, Sigmund O. Material interpolation schemes in topology op- [31] Liu B, Cao W, Zhang L, Jiang K, Lu P. A design method of Voronoi porous
timization. Arch Appl Mech 1999;69(9):635–54. http://dx.doi.org/10.1007/ structures with graded relative elasticity distribution for functionally gradient
s004190050248. porous materials. Int J Mech Mater Des 2021;17(4):863–83. http://dx.doi.org/
[12] Huang X, Xie YM. Evolutionary topology optimization of continuum structures: 10.1007/s10999-021-09558-6.
Methods and applications. John Wiley & Sons; 2010, http://dx.doi.org/10.1002/ [32] Kulshreshtha A, Dhakad S. Preparation of metal foam by different methods: A
9780470689486. review. Mater Today: Proc 2020;26:1784–90. http://dx.doi.org/10.1016/j.matpr.
[13] Querin O, Young V, Steven G, Xie YM. Computational efficiency and validation of 2020.02.375.
bi-directional evolutionary structural optimisation. Comput Methods Appl Mech [33] Bendsøe MP. Optimization of structural topology, shape, and material, Vol. 414.
Eng 2000;189(2):559–73. http://dx.doi.org/10.1016/S0045-7825(99)00309-6. Springer; 1995, http://dx.doi.org/10.1007/978-3-662-03115-5.
[14] Wang MY, Wang X, Guo D. A level set method for structural topology optimiza- [34] Svanberg K. The method of moving asymptotes—A new method for structural
tion. Comput Methods Appl Mech Eng 2003;192(1–2):227–46. http://dx.doi.org/ optimization. Internat J Numer Methods Engrg 1987;24(2):359–73. http://dx.
10.1016/S0045-7825(02)00559-5. doi.org/10.1002/nme.1620240207.
[15] Allaire G, Jouve F, Toader A-M. Structural optimization using sensitivity analysis [35] Guest JK, Prévost JH, Belytschko T. Achieving minimum length scale in topology
and a level-set method. J Comput Phys 2004;194(1):363–93. http://dx.doi.org/ optimization using nodal design variables and projection functions. Internat
10.1016/j.jcp.2003.09.032. J Numer Methods Engrg 2004;61(2):238–54. http://dx.doi.org/10.1002/nme.
[16] Wu J, Sigmund O, Groen JP. Topology optimization of multi-scale structures: A 1620240207.
review. Struct Multidiscipl Optim 2021;63:1455–80. http://dx.doi.org/10.1007/ [36] Montemurro M. On the structural stiffness maximisation of anisotropic continua
s00158-021-02881-8. under inhomogeneous Neumann–Dirichlet boundary conditions. Compos Struct
[17] Wu J, Dick C, Westermann R. A system for high-resolution topology optimization. 2022;287:115289. http://dx.doi.org/10.1016/j.compstruct.2022.115289.
IEEE Trans Vis Comput Graphics 2015;22(3):1195–208. http://dx.doi.org/10. [37] Montemurro M, Rodriguez T, Pailhès J, Le Texier P. On multi-material topol-
1109/TVCG.2015.2502588. ogy optimisation problems under inhomogeneous Neumann–Dirichlet boundary
[18] Wu J, Wang W, Gao X. Design and optimization of conforming lattice structures. conditions. Finite Elem Anal Des 2023;214:103867. http://dx.doi.org/10.1016/
IEEE Trans Vis Comput Graphics 2019;27(1):43–56. http://dx.doi.org/10.1109/ j.finel.2022.103867.
TVCG.2019.2938946. [38] Kováčik J, Marsavina L, Linul E. Poisson’s ratio of closed-cell aluminium foams.
[19] Guest JK. Imposing maximum length scale in topology optimization. Struct Mul- Materials 2018;11(10):1904. http://dx.doi.org/10.3390/ma11101904.
tidiscipl Optim 2009;37:463–73. http://dx.doi.org/10.1007/s00158-008-0250- [39] Gibson LJ. Cellular solids. Cambridge University Press; 2003, http://dx.doi.org/
7. 10.1017/CBO9781139878326.
[20] Dou S. A projection approach for topology optimization of porous struc- [40] Roberts AP, Garboczi EJ. Elastic properties of model random three-dimensional
tures through implicit local volume control. Struct Multidiscipl Optim open-cell solids. J Mech Phys Solids 2002;50(1):33–55. http://dx.doi.org/10.
2020;62(2):835–50. http://dx.doi.org/10.1007/s00158-020-02539-x. 1016/S0022-5096(01)00056-4.
[21] Zhao Z, Zhang XS. Design of graded porous bone-like structures via a [41] Kitipornchai S, Chen D, Yang J. Free vibration and elastic buckling of func-
multi-material topology optimization approach. Struct Multidiscipl Optim tionally graded porous beams reinforced by graphene platelets. Mater Des
2021;64(2):677–98. http://dx.doi.org/10.1007/s00158-021-02870-x. 2017;116:656–65. http://dx.doi.org/10.1016/j.matdes.2016.12.061.
[22] Feng F, Xiong S, Liu Z, Xian Z, Zhou Y, Kobayashi H, et al. Cellular topology [42] Roberts RA, Mullis CT. Digital signal processing. Addison-Wesley Longman Pub-
optimization on differentiable voronoi diagrams. 2022, http://dx.doi.org/10. lishing Co., Inc.; 1987, http://dx.doi.org/10.1016/B978-0-7506-7444-7.X5036-
48550/arXiv.2204.10313, arXiv Preprint. 5.
[23] Clausen A, Wang F, Jensen JS, Sigmund O, Lewis JA. Topology optimized [43] Li Z, Lee T-U, Yao Y, Xie YM. Smoothing topology optimization results using
architectures with programmable Poisson’s ratio over large deformations. Adv pre-built lookup tables. Adv Eng Softw 2022;173:103204. http://dx.doi.org/10.
Mater 2015;27(37):5523–7. http://dx.doi.org/10.1002/adma.201502485. 1016/j.advengsoft.2022.103204.
[24] Yan J, Sui Q, Fan Z, Duan Z, Yu T. Clustering-based multiscale topology opti- [44] Xiao F, Yin X. Geometry models of porous media based on Voronoi tessellations
mization of thermo-elastic lattice structures. Comput Mech 2020;66(4):979–1002. and their porosity–permeability relations. Comput Math Appl 2016;72(2):328–48.
http://dx.doi.org/10.1007/s00466-020-01892-4. http://dx.doi.org/10.1016/j.camwa.2015.09.009.
[25] Li H, Luo Z, Zhang N, Gao L, Brown T. Integrated design of cellular composites [45] Martínez J, Dumas J, Lefebvre S. Procedural Voronoi foams for additive manufac-
using a level-set topology optimization method. Comput Methods Appl Mech Eng turing. ACM Trans Graph 2016;35(4):1–12. http://dx.doi.org/10.1145/2897824.
2016;309:453–75. http://dx.doi.org/10.1016/j.cma.2016.06.012. 2925922.
[26] Spimhero. Photograph by spimhero, distributed under a CC BY-SA 4.0 license. [46] Kou X, Tan S. A simple and effective geometric representation for irregular
2016, URL https://commons.wikimedia.org/wiki/File:Aluminium_foam.png. porous structure modeling. Comput Aided Des 2010;42(10):930–41. http://dx.
[27] Bruggi M, Ismail H, Lógó J, Paoletti I. Lightweight design with displacement con- doi.org/10.1016/j.cad.2010.06.006.
straints using graded porous microstructures. Comput Struct 2022;272:106873. [47] Smith M. ABAQUS/standard user’s manual, version 6.9. United States: Dassault
http://dx.doi.org/10.1016/j.compstruc.2022.106873. Systèmes Simulia Corp; 2009.
[28] Do QT, Nguyen CHP, Choi Y. Homogenization-based optimum design of addi- [48] Yi Z, Xiao-Yun Z, Ke W, Si-Yuan H, Jia-Gui L, Wei Z, et al. Fabrication of
tively manufactured Voronoi cellular structures. Addit Manuf 2021;45:102057. functionally radial graded metallic foam. Mater Lett 2020;264:127292. http:
http://dx.doi.org/10.1016/j.addma.2021.102057. //dx.doi.org/10.1016/j.matlet.2019.127292.
[29] Groen JP, Sigmund O. Homogenization-based topology optimization for high- [49] Leitlmeier D, Degischer HP, Flankl HJ. Development of a foaming
resolution manufacturable microstructures. Internat J Numer Methods Engrg process for particulate reinforced aluminum melts. Adv Energy Mater
2018;113(8):1148–63. http://dx.doi.org/10.1002/nme.5575. 2002;4(10):735–40. http://dx.doi.org/10.1002/1527-2648(20021014)4:
[30] Gómez S, Vlad M, López J, Fernández E. Design and properties of 3D scaffolds 10<735::AID-ADEM735>3.0.CO;2-Y.
for bone tissue engineering. Acta Biomater 2016;42:341–50. http://dx.doi.org/
10.1016/j.actbio.2016.06.032.

12

You might also like