Journal of Computational Design and Engineering, 2023, 10,

1736–1766
DOI: 10.1093/jcde/qwad072
Advance access publication date: 13 July 2023
Review article

Topology optimization via machine learning and deep

learning: a review
Seungyeon Shin1 ,† , Dongju Shin1,2 ,† and Namwoo Kang1,2 , *
1
Cho Chun Shik Graduate School of Mobility, Korea Advanced Institute of Science and Technology , 34051, Daejeon, Republic of Korea
2
Narnia Labs , 34051, Daejeon, Republic of Korea
∗
Correspondence: [email protected]
†Contributed equally to this work.

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

Abstract
Topology optimization (TO) is a method of deriving an optimal design that satisfies a given load and boundary conditions within a
design domain. This method enables effective design without initial design, but has been limited in use due to high computational
costs. At the same time, machine learning (ML) methodology including deep learning has made great progress in the 21st century,
and accordingly, many studies have been conducted to enable effective and rapid optimization by applying ML to TO. Therefore, this
study reviews and analyzes previous research on ML-based TO (MLTO). Two different perspectives of MLTO are used to review studies:
(i) TO and (ii) ML perspectives. The TO perspective addresses “why” to use ML for TO, while the ML perspective addresses “how” to
apply ML to TO. In addition, the limitations of current MLTO research and future research directions are examined.

1. Introduction the parametric LSM for non-uniform mesh of fluid TO problems

(Li et al., 2022).
Topology optimization (TO) is a field of design optimization that
However, although the aforementioned TO methodologies can
determines the optimal material layout under certain constraints
produce good conceptual designs, one of the main challenges in
on loads and boundaries within a given design space. This method
performing TO is its high computational cost. The overall cost of
allows the optimal distribution of materials with desired perfor-
the computational scheme is dominated by finite element anal-
mance to be determined while meeting the design constraints of
ysis (FEA), which computes the sensitivity for each iteration of
the structure (Bendsøe, 1989). TO is meaningful in that, compared
the optimization process. The required FEA time increases as the
with conventional optimization approaches, designing is possible
mesh size increases (e.g., when the mesh size is increased by 125
without meaningful initial design. Due to these advantages, var-
times, the required time increases by 4137 times; Liu & Tovar,
ious TO methodologies have been studied to date (Allaire et al.,
2014). Amid this computational challenge, performing TO for a
2002; Bendsøe, 1989; Mlejnek, 1992; Rozvany et al., 1992; Wang et
fine (high-resolution) topological mesh can take a few hours to
al., 2003; Xie & Steven, 1993). The following four TO methodologies
days (Rade et al., 2020). Furthermore, the 3D TO process requires
are described in detail in Appendix 1: density-based method [i.e.,
much higher computational costs with increasing demands in the
the solid isotropic material with penalization (SIMP) method], evo-
order SIMP, bidirectional ESO (BESO), and LSM in terms of the
lutionary structural optimization (ESO) method, level-set method
number of iterations (Yago et al., 2022). Therefore, various studies
(LSM), and moving morphable component (MMC) method.
have aimed to reduce the computation of solving these analysis
Recent TO methods aim to solve various industrial applica-
equations in TO (Amir et al., 2009, 2010). For instance, by using an
tions. Examples include TO for patient-specific osteosynthesis
approximate approach to solve the nested analysis equation in
plates (Park et al., 2021), microscale lattice parameter (i.e., the strut
the minimum compliance TO problem, Amir and Sigmund (2011)
diameter) optimization for TO (Cheng et al., 2019), homogeniza-
reduced the computational cost by one order of magnitude for an
tion of three-dimensional (3D) TO with microscale lattices (Zhang
FE mesh with 40 500 elements.
et al., 2021b), and multiscale TO for additive manufacturing (AM;
Similarly, aiming to improve this computational challenge, var-
Kim et al., 2022). Other notable works for more complex TO prob-
ious methods have been developed to accelerate TO (e.g., Limkilde
lems include a multilevel approach to large-scale TO accounting
et al., 2018) discussed the computational complexity of TO, while
for linearized buckling criteria (Ferrari & Sigmund, 2020), the lo-
Ferrari and Sigmund (2020) conducted large-scale TO with re-
calized parametric LSM applying a B-spline interpolation method
duced computational cost. Martínez-Frutos et al. (2017) performed
(Wu et al., 2020), the systematic TO approach for simultaneously
efficient computation using GPU, while Borrvall and Petersson
designing morphing functionality and actuation in 3D wing struc-
(2001) and Aage et al. (2015) attempted to accelerate TO by par-
tures (Jensen et al., 2021), the parametrized LSM combined with
allel computing. Despite the aforementioned efforts to reduce
the MMA algorithm to solve non-linear heat conduction problems
TO computing time, the computational costs remain high. This
with regional temperature constraints (Zhuang et al., 2021b), and

Received: October 16, 2022. Revised: June 30, 2023. Accepted: July 1, 2023
© The Author(s) 2023. Published by Oxford University Press on behalf of the Society for Computational Design and Engineering. This is an Open Access article
distributed under the terms of the Creative Commons Attribution-NonCommercial License (https://creativecommons.org/licenses/by-nc/4.0/), which permits
non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact
[email protected]
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1737

2.1. Acceleration of iteration

The optimization process of iterative TO can be divided into two
stages: the early stage, in which the design change is significant,
and the latter stage, in which the design changes slowly until it
converges. Studies applying ML techniques supplement the lat-
ter step of the optimization process, which is especially time-
consuming. Thus, studies that aim at accelerating iterative TO
consist of two stages as follows:

(i) Until the intermediate stage of the iteration, the iterative

TO is performed.
(ii) In the intermediate stage, the optimal topology can be di-
rectly predicted by ML by using the intermediate topology
obtained from (i).

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

This process can accelerate the time-consuming and costly lat-
ter stage in the conventional optimization process.
Figure 1: Overview of AI, ML, and DL.
As a representative study, Sosnovik and Oseledets (2019) pro-
posed an optimization method that predicts the final optimized
challenge has encouraged several researchers to develop ML- image through a fully connected network by using the density
based TO (MLTO) to accelerate TO. gradient between iterations obtained through the SIMP method to
Artificial intelligence (AI) includes any technology that allows reduce the total time consumption through the method of early
machines to emulate human behavior, and machine learning (ML) stopping the SIMP method. Lin et al. (2018) also predicted the op-
is a subset of AI, aimed at learning meaningful patterns from data timal shape of the conductive heat transfer by using two figures
by using statistical methods (Fig. 1). Deep learning (DL) is a subset obtained by the SIMP method: the distribution of the design pa-
of ML, inspired by the neuron structure of the human brain, seeks rameter and the gradient distribution of the design parameter be-
to improve learning ability through methods of training hierarchi- tween iterations. Extending this method to 3D, Banga et al. (2018)
cal neural network (NN) structures consisting of multiple layers proposed the 3D MLTO by using three types of input, the 3D den-
from the data itself (Goodfellow et al., 2016; LeCun et al., 2015). sity distribution of voxels at iteration, the gradient of voxel den-
Data-based TO methodologies complement the problem of sities between iterations, loading and boundary conditions along
conventional TO methodologies with computational cost prob- the x, y, and z directions, into a 3D convolutional NN (CNN) to
lems due to design iterations over thousands of times, and many predict optimized voxel topology. This study is similar to the work
studies have been conducted to improve TO algorithms through of Sosnovik and Oseledets (2019), except for adding loading and
AI (Banga et al., 2018; Cang et al., 2019; Chandrasekhar & Suresh, boundary conditions as input and applying them to the 3D do-
2021a; Guo et al., 2018; Oh et al., 2019; Sosnovik & Oseledets, 2019). main.
TO methods applying AI can quickly and effectively optimize the In addition, to significantly reduce the computational cost of
initial topology. In this study, MLTO is expressed throughout stud- fine-discretized meshes, Kallioras et al. (2020) utilized deep be-
ies that incorporate ML or DL. In particular, we focus more on var- lief networks (DBNs) with the SIMP method to perform optimiza-
ious methodologies by using DL. tion. The intermediate optimization results from the SIMP method
This material intends to review various MLTO studies cur- are input into the DBN to predict the near-optimal element den-
rently developed and analyze each characteristic. Sections 2 and sity and then the preliminary results were refined using the SIMP
3 present the existing studies from a TO point of view and an method. Similarly, Kallioras and Lagaros (2021a) utilized an iter-
ML point of view, respectively, and analyzed the characteristics ative re-meshing and convolution process for the computational
of the study. Section 2 groups and introduces studies focusing on advantage of the fine-mesh domain. Similarly, Bi et al. (2020) used
the purpose, i.e., “why” ML is applied to TO from a TO perspec- the gradient predicted by deep NN (DNN) instead of the gradi-
tive. Section 3 groups and introduces studies focusing on the ML ent of every iteration during the optimization to accelerate the
method, i.e., “how” TO can be converted to ML problems from TO. They presented a method of predicting a given design and its
an ML perspective. In Section 4, the limitations of MLTO studies corresponding gradient based on iteration history data through
and the desirable future directions of the field are presented. Ap- online updates by training DNN with gradients obtained through
pendices 1 and 2 introduce the fundamental background of TO local sampling. With this approach, designs can be quickly up-
and basic DL methodologies required to understand this study. dated using the predicted gradient with notably few training
Figure 2 presents various analytical perspectives on MLTO to be steps.
described in Sections 2 and 3.
2.2. Non-iterative optimization
Studies on non-iterative TO immediately derive an optimal shape
2. TO Perspective: Why Use ML that satisfies design conditions, such as loading and bound-
From the perspective of TO, studies that focus on ML/DL to ary conditions, without the iterative process. However, since
improve the existing TO techniques have seven main pur- this method can yield topology that falls short of the objective
poses: acceleration of iteration, non-iterative optimization, meta- performance or topology with lower resolution than the opti-
modeling, dimensionality reduction of design space, improvement mized shape derived from iterative TO, researchers have proposed
of optimizer, generative design (design exploration), and postpro- frameworks consisting of two or more networks to obtain high-
cessing. Each purpose is shown in Fig. 3, and the studies corre- resolution results or optimal designs with satisfying engineering
sponding to each purpose are organized in Table 1.
1738 | Topology optimization via artificial intelligence: a review

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

Figure 2: MLTO review framework.

performance. Single-stage MLTO consists of one network, whereas volume fraction and loading and boundary conditions for the in-
multistage MLTO consists of two or more networks. put of deep CNN but also the strain field information through FEA
to generalize this method to various loading and boundary con-
2.2.1. Single-stage process ditions and various initial designs. Similarly, Nie et al. (2021) pro-
First, as a single-stage method to predict optimization results, posed cGAN-type TopologyGAN, in which the strain energy den-
Abueidda et al. (2020) presented a CNN (ResUnet) that di- sity and von Mises stress values obtained from FEA are used as
rectly derives an optimized shape with 2D loading and bound- the condition, along with the volume fraction, displacement, and
ary conditions, and volume fraction as the input for three loading and boundary conditions.
material properties, such as linear elasticity with small de- Cang et al. (2019) introduced a “theory-driven” mechanism
formation (i.e., without non-linear constraints), non-linear hy- based on adaptive sampling rather than a batch-mode training
perelasticity with geometric non-linearity, and linear elasticity method to reduce computational costs in the optimization pro-
with stress constraint (i.e., non-linear constraint). Sharpe and cess. The deviation of the topology predicted by the NN from the
Seepersad (2019) derived the optimized topology in real-time optimality conditions for TO is calculated and used to select a new
by inputting conditions, such as volume fraction, loading lo- data point (optimization problem set) for the next training. This
cation, and material type, by using a conditional generative method allows for the better prediction of designs with optimal
adversarial network (cGAN). Kollmann et al. (2020) predicted structural compliance for unseen loadings than using static data
the optimal topology of 2D meta-material via encoder and collection under the same computational budget.
decoder from ResUNet by using volume fraction, filter ra-
dius, and design objective (e.g., maximum bulk modulus, max- 2.2.2. Multistage process
imum shear modulus, or minimum Poisson’s ratio) as the in- Scholars have also proposed a multistage non-iterative process
put. to predict the optimization result with better engineering perfor-
Similarly, by using 2D optimization results obtained from the mance by integrating two or more networks.
SIMP method under specific loading and boundary conditions as First, Yu et al. (2019) performed TO in a non-iterative man-
the reference data, Rawat and Shen (2018) used Wasserstein gen- ner and then upscaled the predicted low-resolution image into
erative adversarial network (WGAN) to predict the optimal shape high-resolution. This study was the first to propose a near-
in real-time and then additionally predicted the optimization con- optimal optimization process without repeated iterations. The
dition of the generation result. This approach allows the gener- low-resolution optimal images are first predicted by inputting x-,
ated shape to be mapped to the corresponding optimization con- and y-directional loading and boundary conditions as the input,
ditions, such as the volume fraction, penalty, and radius of the and then the final optimal topology in high-resolution can be pre-
smoothing filter through CNN. Extending this scheme to the 3D dicted with the low-resolution image. Li et al. (2019a) also applied
domain, Rawat and Shen (2019a) predicted the 3D optimal topol- a non-iterative method similar to that of Yu et al. (2019) to a con-
ogy through WGAN, and the optimization conditions (e.g., volume ductive heat transfer problem. Low-dimensional topology images
fraction, penalty, and radius of smoothing filter) are predicted by are generated through the generative adversarial network (GAN)
CNN. Rawat and Shen (2019b) attempted to predict quasi-optimal based on the heat sink position, the heat source position, and the
topology by using volume fraction as a condition to conditional mass fraction information, and upscale the result through super-
WGAN. resolution GAN (SRGAN). Keshavarzzadeh et al. (2021) constructed
Furthermore, researchers have attempted to predict the opti- a prediction model that approximates a high-resolution optimal
mized shape instantly aided by physics information from FEA for design, given training data of multiple resolutions. To this end,
higher accuracy. For example, Zhang et al. (2019) used not only they proposed a model integrating the disjunctive normal shape
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1739

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

Figure 3: Visualization of the methodology by MLTO purpose.

model (DNSM) at the end of a fully connected DNN to construct sity sequence (DS) prediction network that predicts the initial den-
a multiresolution network with both low-resolution and high- sity by using the initial compliance and the volume fraction ini-
resolution designs. After mapping low-dimensional parameters tialization to predict the final optimal density. The second method
such as loading, boundary conditions, and volume fraction to the is coupled density and compliance sequence (CDCS) prediction
DNSM coefficient by using the DNN, the DNSM model is used to network, which uses the intermediate density and the intermedi-
predict the optimal topologies of various resolutions. ate compliance information together to predict the optimal den-
Rade et al. (2020) proposed two methods to explore high- sity though five repetitive processes. By using the first 30 iteration
resolution TO (HRTO) for 2D and 3D features. The first is the den- results from ESO, Qiu et al. (2021a) trained the U-net CNN model to
1740 | Topology optimization via artificial intelligence: a review

Table 1: Classification of relevant studies by MLTO purpose.

Purpose of MLTO in TO perspective Corresponding research

Acceleration of iteration Banga et al., 2018; Bi et al., 2020; Kallioras & Lagaros, 2021a; Kallioras et al., 2020; Lin et al., 2018;
Sosnovik & Oseledets, 2019
Non-iterative optimization Abueidda et al., 2020; Almasri et al., 2020; Cang et al., 2019; Keshavarzzadeh et al., 2021; Kollmann et al.,
2020; Lew & Buehler, 2021; Li et al., 2019a; Nie et al., 2021; Qiu et al., 2021a; Rade et al., 2020; Rawat &
Shen, 2018, 2019a, b; Sharpe & Seepersad, 2019; Yu et al., 2019; Zhang et al., 2019
Meta-modeling Asanuma et al., 2020; Aulig & Olhofer, 2014; Chi et al., 2021; Deng et al., 2022; Doi et al., 2019;
Keshavarzzadeh et al., 2020; Kim et al., 2021; Lee et al., 2020; Li et al., 2019b; Patel & Choi, 2012; Qian &
Ye, 2021; Sasaki & Igarashi, 2019a, b; Takahashi et al., 2019; White et al., 2019; Xia et al., 2017; Zheng et
al., 2021; Zhou & Saitou, 2017
Dimensionality reduction Guo et al., 2018; Li et al., 2019c; Ulu et al., 2016
of design space
Improvement of optimizer Bujny et al., 2018; Chandrasekhar & Suresh, 2021a, b, c; Deng & To, 2020; Halle et al., 2021; Hoyer et al.,

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

2019; Jiang et al., 2020; Lei et al., 2019; Lynch et al., 2019; Zehnder et al., 2021; Zhang et al., 2021a
Generative design Blanchard-Dionne & Martin, 2021; Gillhofer et al., 2019; Greminger, 2020; Ha et al., 2021; Jang et al.,
2022; Jiang & Fan, 2019a, b; Jiang et al., 2019; Kallioras & Lagaros, 2020; Kallioras & Lagaros, 2021b;
Kudyshev et al., 2021; Li et al., 2021; Malviya, 2020; Oh et al., 2018, 2019; Shen & Chen, 2019; Sim et al.,
2021; Sun & Ma, 2020; Wang et al., 2020; Wen et al., 2019, 2020; Yamasaki et al., 2021
Postprocessing Chu et al., 2016; Karlsson et al., 2021; Napier et al., 2020; Strömberg, 2019, 2020; Wang et al., 2021; Xue et
al., 2021; Yildiz et al., 2003
Others Brown et al., 2022; Gaymann & Montomoli, 2019; Hayashi & Ohsaki, 2020; Kumar & Suresh, 2020; Liu et
al., 2015; Qiu et al., 2021b

predict the intermediate topologies and then used the evolution of ods to proceed immediately without undertaking FEA. Deng et al.
the predicted topologies to learn the subsequent structure trans- (2022) also used DNN to approximate the objective function.
formations using long short-term memory (LSTM). This method Kim et al. (2021) used the representative volume element (RVE)
required less training data but could offer much broader appli- method for TO of functionally graded composite structure design,
cability. Lew and Buehler (2021) utilized representations in latent constructing a DNN-based continuous model to predict the effec-
space to explore the latent design space via the variational au- tive elasticity tensor (a material property) for given design param-
toencoder (VAE)-LSTM model. The LSTM was trained to learn the eters to replace the FEA process in the iterative TO process. This
trajectories in a latent space that could compress the optimiza- methodology is applicable to both 2D and 3D problems.
tion routes, allowing the system to learn the valid optimization In the work of White et al. (2019), which deals with the prob-
progression. lem of multiscale TO, the computational process for effective elas-
In a study that additionally considered the manufacturing con- tic stiffness of meta-material is replaced by the NN surrogate
ditions, Almasri et al. (2020) used a conditional convolutional dual- model. Through this NN, microscale geometric parameters can
discriminator GAN to accelerate TO, considering both a single gen- be mapped to macroscale stiffness. Li et al. (2019b) used feed-
eral manufacturing condition and a mechanical condition. In the forward NNs (FFNN) to map macroscale strain components and
prediction of the 2D optimized shape, both geometrical (e.g., the macroscale stress components. Using the series of stress values
total number of bars) and mechanical (e.g., boundary conditions, predicted by FFNN, they predicted whether the microstructure is
loading configuration, and volume fraction) constraints are used damaged or not through CNN to provide a criterion for sensitivity
as the input to the generator, and two discriminators ensure that analysis.
both mechanical and geometrical constraints are both satisfied.
Furthermore, to evaluate the performance of the generated topol- 2.3.2. Sensitivity
ogy, a CNN-based network is used to predict the compliance of the
As a study to predict sensitivity, Aulig and Olhofer (2014) replaced
generated structure.
sensitivity analysis using a regression model . The material den-
sity of an element, which is a local state feature, and the displace-
2.3. Meta-modeling ment values of the element nodes obtained through FEA are used
In the TO process, the calculation of objective function and sen- to predict the sensitivity. Xia et al. (2017) proposed a TO method
sitivity analysis process through FEA occupies a high computa- applied with a constrained maximum-weight connected graph
tional cost. Therefore, some studies that accelerate the TO process (CMWG) to remove the checkerboard pattern from the optimiza-
through ML are conducted to accelerate the optimization process tion results and used kriging models based on the support vector
by replacing the solvers, such as FEA. machine (SVM) to predict the sensitivity. Through the CMWG al-
gorithm, the result of the previous ten iterations and the data of
the surrounding elements are combined and used as training data
2.3.1. Objective function to calculate the sensitivity value of the element.
As a representative study to predict the objective function Takahashi et al. (2019) attempted to find the global optimum
value, Lee et al. (2020) attempted to accelerate the optimiza- by calculating the sensitivity through CNN in the TO process. Ad-
tion process by replacing the objective function calculation pro- ditionally, filtering was also performed at the end to enable the
cess through FEA with CNN . The compliance and volume optimization results to be more meaningful from an engineering
fraction are predicted by inserting the intermediate material perspective. Chi et al. (2021) also trained the ML model to directly
layout as an input, allowing for the optimality criteria (OC) meth- predict sensitivity without solving the state equations. However,
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1741

in contrast to the work of Takahashi et al. (2019), they proposed and Igarashi (2019a) predicted the average torque and torque rip-
an online training concept in the optimization process and used ple value of the interior permanent magnet (IPM) motor through
the history data of the optimization process as a training sample. CNN to calculate the objective function. According to the pre-
In providing sufficient information for training while avoiding the dicted objective function value, probability P of FEA is calculated
most time-consuming step of solving the state equation, TO for- to determine whether to execute FEA or not. The probability P
mulas with two discretization levels, fine-scale mesh and coarse- helps to skip unnecessary FE processes because FE calculations
scale mesh, are used. Design variables are used from the fine- are only performed when P is large. In the existing genetic al-
scale mesh, and the strain vectors are used from the coarse-scale gorithm (GA), many candidates are randomly generated. How-
mesh. ever, as these initial shapes do not all satisfy the torque perfor-
mance, CNN predicts the torque performance to reduce the num-
2.3.3. Objective function and sensitivity ber of samples to be subjected to FEM to save time and comput-
As a study for predicting both objective function calculation and ing cost. Sasaki and Igarashi (2019b) also reduced the FE com-
sensitivity analysis, Qian and Ye (2021) used dual model NN for putation through a classification model that predicts the per-
more accurate sensitivity analysis. The initial design obtained formance of IPM motors. In contrast to the work of Sasaki and

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

through SIMP is used to predict compliance through the forward Igarashi (2019a), this study used various input values and con-
model of dual model CNN, and the sensitivity of each element ducted various experiments based on multi-objective optimiza-
is calculated through the adjoint model. These NNs are used tion. Doi et al. (2019) also accelerated multi-objective optimiza-
to perform forward calculations and sensitivity analysis, which tion (e.g., maximizing average torque and minimizing torque rip-
can reduce time because they are much faster than higher-order ple) for IPM motors. Similar to the previous studies, the prob-
simulations. Zheng et al. (2021) proposed a multiscale TO frame- ability P to determine whether or not FEM is performed was
work based on spinodoid TO for cellular structures. By simulta- calculated based on the average torque and torque ripple pre-
neously optimizing macroscale material distribution, microstruc- dicted by CNN. In addition, when the predicted average torque
tural design, and orientation, FE calculations can be replaced with and torque ripple values exceed a critical value, the iron loss is
a data-based surrogate model on a computationally expensive mi- calculated. Similarly, Asanuma et al. (2020) aimed to predict mo-
croscale to predict the stiffness tensor quickly and easily obtain tor performance with a small amount of training data by apply-
the sensitivity of the microscale through backpropagation of gra- ing transfer learning (classifier VGG16) to reduce the FEM process
dients. Keshavarzzadeh et al. (2020) adopt the Gaussian process during GA.
(GP) with a covariance kernel constructed with low-resolution fi-
nite element (FE) simulations to predict the maximum von Mises
stress value and sensitivity for unseen features. GP was trained
using a small amount of simulation data by treating the variabil- 2.4. Dimensionality reduction of design space
ity of material properties, geometry, and loadings as parameters To efficiently explore the design space in the iterative TO process,
in the parameter space. researchers have used the dimensionally reduced latent space
through ML.
2.3.4. Constraints Guo et al. (2018) proposed an optimization process in the la-
Scholars have attempted to replace constraints used in TO with tent space for acceleration and efficiency and performed a multi-
ML. Reliability-based TO, which is used for highly non-linear or objective problem optimization for heat conduction. Using the
disjoint failure domain problems, require extremely high compu- optimized image through SIMP as training data, latent space is
tational cost because the finite element method (FEM) is used to learned through the VAE, and style transfer is also used to allow
evaluate reliability constraints. Therefore, Patel and Choi (2012) the generated images to resemble the training data. Several op-
used the classification-based probabilistic NNs (PNNs) to pre- timization methods, such as GA, gradient-based algorithm, and
dict whether a random variable is a safe or unsafe region with hybrid algorithm, were attempted by extracting the initial popu-
a limit state function in the iterative TO process. The probabil- lation from the trained latent space, of which MOGA performed
ity of failure is calculated using the predicted value, which is the best.
used to check the convergence of the solution in the optimization Ulu et al. (2016) reduced the dimensionality of the existing
algorithm. optimized image in high-dimensional space by using principal
Zhou and Saitou (2017) replaced the manufacturing constraint components analysis (PCA) to generate new optimized topologies.
with ML. They presented a method for designing a minimum The optimal topologies (e.g., reduced PCA weight) and the corre-
compliance composite structure while reducing resin filling time sponding loading configuration are mapped using FFNN. With the
in a transfer molding process. Random forest was used to pre- trained model, PCA components (weight) for a new loading config-
dicting the resin filling time, and the kriging-interpolated level- uration can be predicted, and by synthesizing this with an eigen-
set (KLS) approach and multi-objective genetic algorithm (MOGA) image, new optimal topologies can be obtained. Li et al. (2019c)
were used for optimization methods. proposed efficient dimensionality reduction and surrogate-based
TO to reduce the computational cost of TO of periodic structures.
2.3.5. Hybrid approach This study used logistic PCA to find the lower dimensional rep-
In the case of the studies above, the FEA calculation process was resentation of the topologies. To be used as training data, the
completely replaced. In this section, studies first predict the ob- design variables selected from the latent design space are con-
jective function value by replacing the FEA process with ML and verted into topologies through inverse transformation of logistic
then perform actual FEA to the predicted results with good en- PCA, and the bandgap characteristic corresponding to each topol-
gineering performance to use accurate objective function values ogy is obtained through a numerical model. With this, a surrogate
in the optimization process. As a study that greatly reduced the model (kriging-based efficient global optimization) for the numer-
computational cost of TO based on a stochastic algorithm, Sasaki ical model is constructed and used to predict the bandgap. By
1742 | Topology optimization via artificial intelligence: a review

using this method, the optimal topology can be efficiently found uated the displacement of the topology via physics simulation
within the global optimization. and updated the network parameters based on the physics-based
loss function by using the standard gradient-based algorithm. In
2.5. Improvement of the optimizer this manner, the optimal topology can be obtained with the de-
sired compliance. This process highlights the structural optimiza-
2.5.1. Optimization parameter tuning
tion method, which can be generalized to various optimization
Scholars have also supplemented conventional TO methods by problems. Similarly, Zhang et al. (2021a) proposed TO via a neu-
tuning the optimization parameters based on ML to quickly and ral reparameterization framework that would not need to con-
efficiently search for parameters used in the optimization process. struct a dataset in advance and would not suffer from structural
The parameters applied to the MMC are tuned by experience disconnection for various problems, including non-linear TO
and can obtain an infeasible solution. To solve this problem, Lei problems. Physics information was introduced into the loss func-
et al. (2019) combined ML to optimize parameters used in the MMC tion and the design constraint processing method, which was
method. After obtaining the eigenvalue and eigenvector of the pa- based on the volume-preserving Heaviside function with sigmoid
rameter vector through PCA, non-linear regression was performed transformation.
using support vector regression (SVR) and K-nearest neighbor

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

Halle et al. (2021) proposed a two-stage scheme named the
(KNN), respectively. After that, the design variables for character- predictor–evaluator network to predict optimized images without
izing the optimal shape and the corresponding loading may be the use of a pre-optimized dataset. First, a predictor with an artifi-
mapped. This approach is meaningful in that the design parame- cial NN (ANN) architecture was used to predict optimized images,
ters of the MMC method, which have fewer parameters than the with the input of static and kinematic boundary conditions and
pixel-level SIMP method, are sampled using ML. Similarly, Jiang the target degree of filling both considered. Then, the generated
et al. (2020) used extra-tree (ET)-based image classification for par- image was processed by an evaluator that was used to calculate
ticle swarm optimization algorithm to tune the parameters of the the objective function, and the network was trained using the gra-
method of moving asymptote, which is used as an optimizer for dient. Among the various evaluators used, the compliance evalu-
MMC. The ET classifier extracts image features from the image ator was determined to be an FEM-based algorithm, and it can be
dataset, then normalizes and quantifies them to perform k-means used to predict the compliance of images generated by the predic-
clustering. With the grouped clusters, users can judge whether tor.
the input image is feasible or infeasible. In this manner, users can Similarly, scholars have applied the process mentioned above
quickly determine whether the optimized solution is feasible dur- to the 3D domain by using an implicit neural representation of
ing the optimization process. the density field with spatial coordinates utilized as the input.
Bujny et al. (2018) proposed the A-EA-LSM, an advanced version Deng and To (2020) used a new geometric representation method
of the evolutionary level-set method (EA-LSM), and complements for deep representation learning (DRL) that approximates density
it, to converge faster and find good designs in low-dimensional fields in the design domain by using DNN to reduce design vari-
space. On this basis, they used a new graph-based parameteriza- ables in a computationally expensive 3D TO problem and updates
tion instead of the MMC parameterization originally used in EA- the network parameters in a direction to minimize the compli-
LSM to increase the convergence speed by reducing the dimen- ance obtained through FEM. The density function is parametrized
sion. Furthermore, after generating various designs, the perfor- by parameters of a NN by predicting the corresponding density
mance of each design was evaluated and trained for the NN via value for the input coordinates based on the FEM results. Chan-
the ranking of the designs. Through this trained NN, various topol- drasekhar and Suresh (2021a) also obtained the density value of
ogy variations for EA-LSM can be predicted. elements through a NN, calculated the objective function value
Many design optimization algorithms have a problem in through FEA, and backpropagated the loss function. The method
that a wide range of tuning parameters for controlling the has been extended to 3D and 2D, and projection operators can
algorithmic function and convergence must be manually set. be used to satisfy design and manufacturing constraints. Chan-
Accordingly, Lynch et al. (2019) proposed a ML-based framework drasekhar and Suresh (2021b) proposed a Fourier-TO using neu-
to recommend tuning parameters to users to reduce the trial ral network (TOuNN) that enhances the performance of the ex-
and error and computational cost associated with manual tun- isting TOuNN to generate manufacturable designs. In this study,
ing. To this end, the framework consisted of two stages: a meta- a length scale control strategy was proposed to control the mini-
learning stage that derives parameter recommendations from mum and maximum length scale by adding Fourier space projec-
similar problems and a meta-modeling stage that uses Bayesian tion to TOuNN. No additional constraint is required because the
optimization to optimize parameters for a specific TO problem ef- maximum and minimum dimension range is determined using
ficiently. This scheme allows tuning parameters with similar prob- Fourier projection. Chandrasekhar and Suresh (2021c) extended
lems from previous repositories, which is determined using dis- this study and applied it to multimaterial TO (MMTO) also to pre-
similarity metrics based on the problem’s meta-data, to the cur- dict the material types of elements. The proposed method calcu-
rent problem by using a Bayesian optimization approach. lates the loss by evaluating the volume fraction of each material
per element for a given material type through the NN and by ob-
2.5.2. Reparameterization taining Young’s modulus through FEA.
To achieve efficient optimization, researchers have reparameter- Zehnder et al. (2021) parameterized not only density but also
ized the density fields using NN. In this method, TO is conducted displacement fields and used implicit neural representation to
by backpropagating objective function values obtained through perform TO in continuous solution spaces in a mesh-free man-
solvers such as FEA in the training process. This method works ner. First, the total gradient of the compliance loss with respect
similarly to the process of updating the sensitivity and can obtain to the NN parameters was computed to obtain the target density
the optimized image by optimizing the parameters of the NN. fields. Then, the density network was updated to minimize the
Hoyer et al. (2019) obtained the optimal topology for the input mean squared error (MSE) between the current and target densi-
noise by adopting the deep image prior of CNN and then eval- ties. In this manner, convergence to the bad local minima could be
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1743

avoided. Although the proposed method requires more computa- mal topology), to learn the corresponding failure distribution of
tional time than the conventional FEM methods, this mechanism the optimal topology. To generate only high-performance shapes,
offers a fully mesh-free approach to TO for 2D and 3D problems. training was conducted by updating the dataset with samples
Additionally, researchers have suggested integrating the meth- with a good performance through FEM among those generated in
ods above with generative models to generate various optimal the previous iteration. Through a data-driven approach, Yamasaki
topologies with high engineering performance. This topic is elab- et al. (2021) produced new material distribution with good perfor-
orated on in Section 2.6. mance by constructing VAE, a deep generative model, based on
current high-rank data, which is the best design evaluated by FEM
2.6. Generative design among initial material distribution data from the SIMP method.
There are studies for generative design that quickly and diversely By integrating this into the current high-rank data, the pro-
generate new optimal designs. cess of screening new high-rank data is iteratively repeated, and
when the elite data converges, the final optimal topology can be
2.6.1. Naïve approach derived.
Engineering performance aided by NN
The most naïve method is to train a generative model using the ex-

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

To reflect the physics information in the training process, schol-
isting optimized shape as reference data. Malviya (2020) compared
ars have also suggested a method of evaluating the objective
the performance of CNN, UNet, and GAN to explore the effects of
function value of design generated through a DL-based predic-
different types of deep generative models in a wide range of TO
tion model instead of a solver. Ha et al. (2021) proposed Fit2Form
problems, along with various design constraints, and loading and
to automate the design of 3D task-specific robot gripper fingers,
boundary conditions, and consequently evaluated TOPCNN and
which consists of 3D CNNs and 3D autoencoders (AEs). A 3D AE
TOPGAN as the best.
that takes the truncated signed distance function volume of the
target object as input and outputs the appropriate left and right
2.6.2. Engineering performance-aided approach
finger shapes. After that, for the predicted left and right finger
Among the studies that aimed at generative design, many stud-
shapes, objective values, such as grasp success, force stability, and
ies have suggested a generative design method that guarantees
robustness, were predicted through a fitness network composed of
good engineering performance by integrating additional meth-
3D CNN to be reflected in the training process. Blanchard-Dionne
ods. To this end, there are studies that utilize physics informa-
and Martin (2021) also generated an optimal cloak with an opti-
tion with the help of additional methods, such as solvers, NNs,
mal configuration using DCGAN and forward network based on
dimensionality reduction, and clustering, for the designs gener-
the existing optimized shapes as the reference. For the designs
ated by generative models.
generated by the generator of DCGAN, the scattering coefficient
Engineering performance aided by simulation
is also predicted through the forward model. Therefore, the per-
First, scholars have reflected the objective function value ob-
formance of the image generated by the generator is reflected in
tained through the solver to the loss function in the generation
the loss function to minimize the scattering coefficient to gener-
process. This is a method of backpropagating the objective func-
ate designs with high performance and, at the same time, deceive
tion value obtained through simulation, in which a generative de-
the discriminator. In increasing the accuracy of the DCGAN and
sign framework reflects the physics information. Most of these
forward model, FEA was performed on data predicted to have high
methods have the advantage of not requiring training data.
performance among the generated shapes, and DCGAN and for-
For example, Jiang and Fan (2019a) proposed a new type of
ward model were re-trained by adding the data as an additional
global optimization algorithm, generative NN, which can produce
dataset.
high-performance meta-surface designs without training data.
A study was also conducted to generate a high-performance
For the input noise, the efficiency of the generated structure is
design by slightly modifying the discriminator in GAN. Gillhofer
evaluated through an EM solver. By backpropagating the efficiency
et al. (2019) proposed a derivative-free optimization method. Fur-
gradient, the generator can generate a high-performance device,
thermore, in contrast to the existing GAN, the discriminator is
allowing the generation of designs that reaches the desired per-
trained to classify whether the image generated by the genera-
formance. Extending this, Jiang and Fan (2019b) proposed global
tor corresponds to the region with higher responses or the region
TO networks to design electromagnetic devices with the target
with lower responses. In this manner, the generator is trained to
wavelength and deflection angle as the additional condition. Fur-
generate samples with a lower response, i.e., better performance.
thermore, Jiang et al. (2019) extended this into an iterative opti-
Engineering performance aided by dimensionality reduction
mization method. Using the same input as the previous research,
Researchers have also used dimensionality reduction to re-
which is the meta-grating deflection angle, operating wavelength,
flect engineering performance in the generative design pro-
and random noise, devices generated by cGAN undergo a rigorous
cess can also be observed. Wang et al. (2020) constructed a
coupled-wave analysis (RCWA) solver to calculate the efficiency.
model that combines VAE and regression to solve the prob-
Among them, devices with good efficiency are used as additional
lems of high-dimensional topological design space, multiple
GAN training data, and devices with good efficiency can be gener-
local optima, and high computational cost. When VAE takes
ated during iterations.
meta-material as input to reconstruct the meta-material, latent
Although with a similar purpose to the previous studies, schol-
space that compresses the features of the meta-material is con-
ars have used the physics information from FEM to update the
structed. Simultaneously, a regressor that predicts a stiffness
training data instead of reflecting it into the loss function. Li
matrix with the latent variable is attached to the VAE archi-
et al. (2021) proposed subset simulation (SS)-based TO that in-
tecture and trained together with the VAE model. In this man-
tegrates SS and GAN to generate samples from the failure dis-
ner, the latent space can contain the mechanical features of the
tribution of each SS level to generate periodic structures with
microstructure.
the desired bandgap properties efficiently. GAN is trained based
Kudyshev et al. (2021) applied the adversarial AE (AAE) for pho-
on failure samples, which are high-performance features (opti-
tonic meta-device design by presenting three new global opti-
1744 | Topology optimization via artificial intelligence: a review

mization frameworks. First, a model consisting of conditional AAE this method, not only the training sample and condition but also
(c-AAE) and regression network is presented, secondly, c-AAE + DE a random condition was additionally inputted to the discrimina-
framework, which uses differential evolution (DE) global opti- tor, and the minibatch discrimination technique was applied to al-
mizer for c-AAE, and finally, c-AAE + rDE framework that also uses low more stable convergence and prevent the mode collapse phe-
important physics-driven regularization. Among them, the third nomenon in which the generator continues to generate similar
model allows the physical regularization of the compressed de- shapes.
sign space to adjust the design space configuration to perform GO
exploration better. This method utilizes the physics information 2.6.5. Manufacturing constraints
inherent in the optical performance of meta-devices with com- In general, many cases claim that manufacturing is impossible
plex material composition. by simply referring to the topological optimization results. To ap-
Engineering performance aided by clustering ply manufacturing constraints to TO, Greminger (2020) used syn-
In the previous section, the studies used engineering perfor- thetic 3D voxel training set to represent the distribution of a set
mance information during the training process. Next, a study con- of geometries that can be manufactured by a specific manufac-
sidered engineering performance by applying clustering to the turing technique to generate specific manufacturable 3D designs

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

generated designs. Sim et al. (2021) attempted to collect optimized through multiscale gradient GAN.
shapes with meaningful performance among the designs gener-
ated through clustering analysis (CA) with GAN. Data were clas- 2.6.6. Aesthetics-based generative design
sified through k-means clustering based on the performance of Moreover, Oh et al. (2018) proposed a generative design method
topologies generated through GAN and DCGAN. They also pro- considering aesthetics. Their study presents a method to generate
posed TO validation curve to collect optimized valid data among new designs by training BEGAN with authentic wheel images and
the generated results. postprocessing the generated image through the SIMP as an initial
Engineering performance aided by TO design to derive final wheel designs. Although data through GAN
For more accurate optimization results, scholars have also con- does not guarantee stiffness, the proposed method can consider
ducted generative design studies that consider engineering perfor- both stiffness and aesthetics by integrating SIMP method. Further-
mance by utilizing the conventional iterative TO method with DL. more, Oh et al. (2019) performed optimization with the same pur-
Kallioras and Lagaros (2020) utilized the optimization results from pose by using SIMP to resemble the reference data, filtered similar
SIMP to diversify into several different results through convolu- designs, and then trained BEGAN with the data to generate various
tional layers. After extracting intermediate topology up to a spe- wheel designs. These data are filtered and entered into SIMP again
cific iteration through SIMP, 24 different topologies were created as reference data and repeated until various designs are gener-
through DBN and four different convolutional layers with the in- ated beyond the desired goal, which is called the deep generative
termediate domain with reduced mesh size through re-meshing. framework . This work has modified the SIMP method to resemble
The 24 different topologies created using this method are re- the reference design and considered both stiffness and aesthetics
entered into SIMP, and the final 24 new optimized structures are by generating new designs while filtering similar designs.
created. In a similar manner, Kallioras and Lagaros (2021b) uti-
lized SIMP with LSTM to estimate the final density distribution 2.6.7. Reinforcement learning
and performed generative design by using different image filters.
Finally, scholars have studied reinforcement learning (RL) algo-
This method showed great generalizability to various loading con-
rithms to generate a vast amount of generative design in a new
ditions, mesh sizes, and support conditions, as well as 3D prob-
approach. Sun and Ma (2020) extended SIMP and BESO, which is
lems.
the main density-based structural TO methods, based on RL ex-
2.6.3. Detailed features ploration strategies. Similarly, Jang et al. (2022) proposed a gen-
erative design process based on RL to maximize design diversity.
While reflecting on the physics information, scholars have at-
The study has the distinction of guaranteeing both maximum di-
tempted to learn detailed features of complex designs. Wen et al.
versity and aesthetics and proposed a RL-based generative design
(2019) used the progressive growth of GAN (PGGAN) to extract
process that applies two networks by defining the maximization of
detailed features through high-resolution training data . In the
design diversity as a reward function. They used TopOpNet to ap-
initial training stage, starting from a low-resolution image and
proximate the TO process and found the design parameter com-
gradually training with a high-resolution image, large features
binations with maximum diversity through GDNet.
with low-resolution data may be learned in the beginning, and
then detailed features from high-resolution data may be learned
as the training progresses. While growing the network architec- 2.7. Postprocessing
ture progressively, they could also augment the training dataset. The purpose of the studies pertaining to the postprocessing of the
For this purpose, high-performance devices are selected through optimized image by applying ML is largely divided into two cat-
the RCWA EM solver and added as the training data, while low- egories. First, as the generated optimized images are often low-
efficiency data are deleted. To supplement the previous study to resolution images, the studies convert them into high-resolution
learn global geometric trends, Wen et al. (2020) conducted the images . Second, other studies have proposed a structural bound-
training by adding a self-attention layer after the deconvolu- ary processing method through ML to postprocess the generated
tion and convolution layers in the generator and discriminator, images with gray scale or unclear boundary .
respectively.
2.7.1. High resolution
2.6.4. Improved GANs A large computational cost is required to obtain a high-resolution
Shen and Chen (2019) introduced a new type of GAN to gener- optimal image instead of a low-resolution image. If high-
ate various topologies by proposing a convex relaxing CGAN that resolution images can be predicted from low-resolution struc-
converges more efficiently and stably than the existing CGAN. In tures, then a large amount of computational cost can be reduced.
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1745

Therefore, by studying ML to obtain high-resolution optimized im- sists of three steps: clustering, optimization, and connectiv-
ages, Napier et al. (2020) used ANN to upscale a low-dimensional ity. This study proposed a density–strain-based method with
input to a high-dimensional one. In this study, the FFNN is trained a fixed clustering pattern according to the static density and
to upscale the low-resolution data to high-resolution data by us- stress state . However, fixed clustering patterns can limit the
ing the preprocessed data of high-resolution beam optimization performance. Consequently, Qiu et al. (2021b) excluded den-
data to low resolution through averaging. Similarly, Wang et al. sity from CA to optimize cellular structures with uniform
(2021) used CNN to map the relationship between low-resolution porosity and used the direction and ratio of principal stress
and high-resolution structures in TO problems. This approach is for clustering. By introducing a stress ratio into clustering,
helpful for large-scale structural design. the stiffness performance and connectivity of the microstruc-
In addition, Xue et al. (2021) proposed a super-resolution CNN ture can be improved. The clustering pattern should be up-
(SRCNN) framework with superior applicability to improve the dated by the current stress tensor in grouping the microstruc-
resolution of TO designs and the HRTO method established by im- ture by using the dynamic clustering method. When clus-
plementing SRCNN and pooling strategy. SRCNN is a network that tering, the stress direction-based cluster analysis is first per-
maps low-resolution images to high resolution, and is available formed, and then the elements of each direction cluster are

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

for 2D and 3D problems with arbitrary boundary conditions, all further clustered into several groups according to the stress
design domain shapes, and various loadings, demonstrating supe- ratio.
rior applicability and high efficiency. In addition to this, this study
was able to solve the 3D HRTO problem by extending the 2D SR-
2.8.2. Multimaterial process
CNN.
By conducting a study on clustering techniques to reduce de-
sign variables of multimaterial density-based optimization, Liu
2.7.2. Boundary processing
et al. (2015) demonstrated the optimization process by dividing the
Chu et al. (2016) used the SVM-based boundary processing method steps into continuous density distribution, clustering, and meta-
as an effective 2D structural boundary processing to obtain clear model-based optimization steps. First, with an intermediate den-
and smooth boundaries. Extending this into the 3D domain, sity distribution obtained through iterative iteration, the continu-
Strömberg (2020) also used SVM for postprocessing of the opti- ous density distribution is classified into a finite number of clus-
mized 3D design to smooth the boundary by classifying the out- ters based on similarity. In this case, the number of clusters means
ermost part of the optimized structure into parts with or with- the number of materials used. Finally, the optimal material dis-
out material. Furthermore, for design optimization for AM, Ström- tribution can be obtained by constructing a meta-model (kriging
berg (2019) classified the benchmark optimized structure from the interpolation) and updating it iteratively according to the global
SIMP method into parts with or without elements by using the 1- optimization algorithm (i.e., GA).
norm SVM to automate the postprocessing and then added the
resulting structure lattice structures. Karlsson et al. (2021) used
SVM to obtain an implicit surface representation of the TO result. 2.8.3. Others
Boundaries were processed like those in the previous works, but Gaymann and Montomoli (2019) proposed the fluid-structure TO
this study also conducted an internal design through a Boolean (FSTO) problem and the structural TO as gamification by optimiz-
operation. ing the fluid structure that minimizes the pressure difference us-
To automate the combination of TO and shape optimization for ing DNN and Monte Carlo tree search. As an optimization method
boundary processing, Yildiz et al. (2003) used MLP NN. After per- that proceeds in a manner in which the winner among the players’
forming TO and changing the grayscale image output as a mesh shapes can be determined, the DNN receives the player’s shape
image, the shape of each hole is labeled with shapes, such as tri- as input to determine who the winner is and whether to end the
angles, squares, trapezoids, and circles. After predicting the shape game and predicts the pressure difference . Based on these re-
of the hole obtained from the result of TO, shape optimization is sults, the process of determining the player’s future movements
performed based on the predicted shape to select a feature-based through a Monte Carlo search is repeated, and the final optimiza-
model. tion shape is obtained at the end of the game.
Hayashi and Ohsaki (2020) demonstrated the TO of truss struc-
ture using graph embedding through RL. To consider the connec-
2.8. Multiscale, multimaterial, and other
tivity between the members, the truss structure is regarded as a
processes
graph, and an optimization topology can be obtained by perform-
Other MLTO studies have a purpose other than the previously de-
ing RL. In addition, Brown et al. (2022) utilized RL to optimize el-
fined ones to solve optimization problems, such as microstruc-
ementally discretized 2D topologies with a multistep progressive
tural TO (MTO), multimaterial density-based optimization, and
refinement approach to improve the computational cost. The pro-
other structural design problems.
posed method can be tested on unseen load cases during training.

2.8.1. Multiscale process

MTO, a simultaneous optimization of macroscale topologies 2.9. Relations between MLTO purposes
and microscale structures, has a problem of high compu- Previously, ML was applied in TO to describe the process from a
tational cost and microstructural connectivity loss. To re- TO perspective. The main purpose of applying ML was classified
duce the computational cost of such MTO, scholars have into seven categories. For each purpose, the advantages and dis-
conducted iterative TO using K-clustering. Among them, Ku- advantages of accuracy and efficiency inevitably exist due to the
mar and Suresh (2020) reduced computational cost and guar- tradeoff relationship. Although MLTO has been divided into seven
anteed connectivity between microstructures in MTO and purposes, the boundaries cannot be strictly distinguished. The re-
performed TO by K-clustering elements based on the ba- lations between the varying purposes of MLTO can be described
sic strain tensor and SIMP density. This methodology con- as follows.
1746 | Topology optimization via artificial intelligence: a review

Various studies have replaced the inefficient process in itera- variables through NN. The design parameters predicted by
tive TO with ML. For example, the studies presented in Section Keshavarzzadeh et al. (2021) correspond to the DNSM co-
2.1 were aimed at accelerating TO by combining ML at the end efficient and are predicted to map DNSM’s coefficient for
of the iterative TO, thereby ensuring accuracy and improving ef- design parameters (load and boundary conditions and vol-
ficiency. The studies in Section 2.3 replaced the time-consuming ume fraction) via NN through a DNSM metric model. In
simulation stage in the iterative TO with an ML-based meta-model addition, in the work of Yildiz et al. (2003), the design pa-
to reduce the computational cost of the conventional TO. Some rameters correspond to the class for the shape of the hole.
studies in Section 2.4 utilized the latent design space to efficiently Yildiz et al. (2003) predicted which class the hole of the
seek the optimal solution. The studies attempted to accelerate shape corresponds to for the optimized feature and per-
the TO while maintaining the iterative process to ensure accu- formed TO based on the results. Second, in studies pre-
racy. Meanwhile, the studies in Section 2.2 were aimed at gen- dicting the performance of a shape (Guo et al., 2021), NN is
erating the optimal topology in one shot by using ML without used to replace some calculations in the TO process. Deng
the need for implementing a time-consuming iterative process. et al. (2022) used DNN to replace the calculation of the ob-
This method can instantly provide optimal topologies once the jective function during the optimization process. Chi et al.

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

datasets are trained. However, the accuracy of this model may (2021) and Aulig and Olhofer (2014) replaced the sensitivity
be low due to the pixelwise training of the reference data. In calculation in the optimization process, while Zheng et al.
addition, the computational cost, including the data collection (2021) and White et al. (2019) allowed the mapping of ob-
cost, can be considered a burden as evidenced by the compu- jective function to parameters defining microscale meta-
tational complexity of the supervised MLTO. Aimed at address- materials. Bujny et al. (2018) predicted the performance for
ing these issues, the studies in Section 2.5.2 would reparame- topology change by mapping graph (structural informa-
terize the density field by using NNs that utilize physics infor- tion) to performance through NN, and Patel and Choi (2012)
mation, such as structural compliance to the loss function. This used PNN to predict the feasibility in terms of safety or risk
method can replace the optimization process by updating the in the TO process to determine the convergence of opti-
parameters of the NN. In addition, much fewer datasets are re- mization.
quired, and structural connectivity can be ensured. However, as 2. Optimization: Several studies also train NN for design op-
most studies that use this method iteratively evaluate objec- timization purposes. Halle et al. (2021) predicted the opti-
tive function values in FEA, the computational cost problem still mized design by using design parameters as inputs, and
exists. The limitation of the MLTO will be further explained in Kim et al. (2021) optimized the periodic material structure
Section 4. by replacing the process of obtaining the elasticity ten-
sor for the design parameter with DNN in the FEA pro-
cess to effectively utilize RVE. Chandrasekhar and Suresh
3. ML Perspective: How to Use ML (2021a, b, c) and Zehnder et al. (2021) optimized the topol-
ML techniques used to improve TO can be classified from various ogy density of the corresponding coordinates. In this case,
perspectives based on the different usages of learning algorithms topology prediction becomes possible by reflecting FEA
(Section 3.1), input data form (Section 3.2), ML loss (or objective) in the loss function. Napier et al. (2020) approximated
function (Section 3.3), usage of physical information (Section 3.4), the high-dimensional (fine) shape with a low-dimensional
and various applications and generalizability (Section 3.5). (coarse) shape, and Bi et al. (2020) predicted and up-
dated the gradient of the shape through the optimization
3.1. Learning algorithm history.
This section analyzes the types of ML applied to TO. Types of ML CNN
are classified and compared by supervised learning which has la- When CNN is applied to TO, it is advanced in terms of the pur-
beled outputs, unsupervised learning which has unlabeled out- pose of prediction for utilizing images rather than NN. Similar to
puts, semi-supervised learning, an algorithm conceptually situ- NN, CNN is mainly used for the purpose of prediction of the per-
ated between supervised learning and unsupervised learning, and formance of shape and design optimization:
RL, consisting of action and reward algorithms (Table 2). An expla-
nation of the DL methodology is given in more detail in Appendix 1. Prediction: For the purposes of predicting the performance,
2. Doi et al. (2019), Sasaki & Igarashi (2019a, b), Asanuma
et al. (2020), Lee et al. (2020), and Li et al. (2019b) predicted
3.1.1. Supervised learning the value of the objective of constraint function by CNN.
For supervised learning, NNs and CNNs were used the most, and Among them, Li et al. (2019b) provided a criterion for sensi-
various other methodologies, such as LSTMs, SVMs, SVRs, KNNs, tivity analysis by predicting whether the shape is damaged
random forest regression, and GP, were applied. or not. Takahashi et al. (2019) also replaced sensitivity cal-
NN culations via CNN. Gaymann and Montomoli (2019) deter-
The application of ANN or DNN (both referred to as NN) to TO, mined a topology with better performance among a given
is distinguished for the purpose of prediction of features or per- topology (player) as a winner.
formance and design optimization: 2. Optimization: For the design optimization purposes, Zhang
et al. (2019), Abueidda et al. (2020), and Kollmann et al. (2020)
1. Prediction: First, researchers have conducted optimization derived the optimal design by inputting the design condi-
by predicting design parameters constituting the shape by tions (load and boundary conditions, etc.). Sosnovik and
using NN. The design parameters predicted by Deng and Oseledets (2019) and Lin et al. (2018) predicted the final op-
To (2020) correspond to DRL design variables that repre- timized design by using the gradient between the nth it-
sent density. The number of design parameters was re- eration shape and the gradient between nth and (n−1)th
duced by converting the existing density field to DRL design iteration. Banga et al. (2018) derived the final optimized de-
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1747

Table 2: Classification of studies according to the learning algorithm.

Category Methodology Use Studies

Supervised learning Neural network Prediction Aulig & Olhofer, 2014; Bujny et al., 2018; Chi et al., 2021;
Deng & To, 2020; Deng et al., 2022; Keshavarzzadeh et al.,
2021; Patel & Choi, 2012; White et al., 2019; Yildiz et al.,
2003; Zheng et al., 2021
Optimization Bi et al., 2020; Chandrasekhar & Suresh, 2021a, b, c; Halle
et al., 2021; Kim et al., 2021; Napier et al., 2020; Zehnder et
al., 2021
Convolutional neural Prediction Asanuma et al., 2020; Doi et al., 2019; Gaymann &
network Montomoli, 2019; Lee et al., 2020; Li et al., 2019b; Sasaki &
Igarashi, 2019a, b; Takahashi et al., 2019
Optimization Abueidda et al., 2020; Banga et al., 2018; Hoyer et al., 2019;
Kollmann et al., 2020; Lin et al., 2018; Rade et al., 2020;

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

Sosnovik & Oseledets, 2019; Wang et al., 2021; Xue et al.,
2021; Zhang et al., 2019, 2021a
Long short-term Optimization Kallioras & Lagaros, 2021b; Lew & Buehler, 2021; Qiu et al.,
memory 2021a
Support vector Postprocessing Chu et al., 2016; Karlsson et al., 2021; Strömberg, 2019, 2020
machine
Prediction Xia et al., 2017
Others Prediction Keshavarzzadeh et al., 2020; Lynch et al., 2019; Zhou &
Saitou, 2017
Unsupervised Generative adversarial Exploration Almasri et al., 2020; Greminger, 2020; Jiang et al., 2019; Li et
learning network al., 2019a, 2021; Malviya, 2020; Nie et al., 2021; Oh et al.,
2018, 2019; Rawat & Shen, 2018, 2019a, b; Sharpe &
Seepersad, 2019; Shen & Chen, 2019; Sim et al., 2021; Wen
et al., 2019, 2020; Yu et al., 2019
Autoencoder Exploration Guo et al., 2018; Ha et al., 2021; Yamasaki et al., 2021
Principal component Exploration Lei et al., 2019; Ulu et al., 2016
analysis
Prediction Li et al., 2019c
Clustering Optimization Jiang et al., 2020; Kumar & Suresh, 2020; Liu et al., 2015;
Qiu et al., 2021b
Semi-supervised Active learning Optimization Cang et al., 2019
learning
Modified networks Optimization Blanchard-Dionne & Martin, 2021; Gillhofer et al., 2019;
Jiang & Fan, 2019a, b; Kudyshev et al., 2021; Wang et al.,
2020
Deep belief network Optimization Kallioras & Lagaros, 2020, 2021a; Kallioras et al., 2020
Reinforcement Value-based Optimization Brown et al., 2022; Hayashi & Ohsaki, 2020
learning
Policy-based Exploration Jang et al., 2022; Sun & Ma, 2020

sign density by combining design conditions and FEA infor- reflect compliance in the loss function. Finally, Xue et al.
mation. Rade et al. (2020) used the CDCS method to train (2021) and Wang et al. (2021) used CNN to upscale optimal
three encoder–decoder-shaped CNN models. The compli- designs.
ance prediction model could predict the next compliance
LSTM
by using the inputs of initial compliance and current den-
LSTM is a useful method for predicting a series of processes
sity. The density prediction model could also predict the
(e.g., the TO process). It is often used in conjunction with other
next density by using the inputs of current compliance and
methodologies for TO optimization:
current density. During the five loops, the two models uti-
lized each other’s outputs as inputs. Finally, with the out- 1. Optimization: Qiu et al. (2021a) utilized the LSTM model
put density determined after five loops, final density pre- to predict the next sequence of TO and then performed
diction model predicts optimal density. In addition, Hoyer DLTO integrated with ESO. An initial 1st–10th sequence im-
et al. (2019) updated the objective function for the pre- age was predicted via CNN(U-net). From the 11th sequence,
dicted design via the deep image prior to the implemen- optimization was performed by predicting the shape of the
tation of CNNS, reflected this information using gradient- next sequence, in which the shapes of the previous 1st to
methods, and allowed the design to be generated to sat- 10th sequences were inputted. Similarly, Rade et al. (2020)
isfy the objective function. Similarly, Zhang et al. (2021a) used DS method with CNN-LSTM and SIMP method to pre-
proposed the TONR method and introduced physics infor- dict the density of subsequent sequences with the input
mation into the loss function. With the noise input, the shapes of the initial two iterations. Lew and Buehler (2021)
model could predict the quasi-density and conduct FEA to trained a 2D convolutional VAE with the shapes of TO se-
1748 | Topology optimization via artificial intelligence: a review

quences to encode an image into a 2D latent vector. With 3.1.2. Unsupervised learning
two sets of initial latent vectors, the LSTM model could pre- The use of unsupervised learning for TO usually heavily uti-
dict the time series latent vectors for the subsequent 147 lizes the generative model and generally has a design exploration
sequences. Kallioras and Lagaros (2021b) used LSTM for the purpose similar to the original purpose of the generative model
time-series classification for generative design. With mul- methodology.GAN
tiple mesh domain, after 20 iterations of SIMP, the density
values of each element were input into the LSTM model 1. Exploration: Greminger (2020) generated feasible data by
to derive the final density class of every element of the FE using a 3D GAN. Li et al. (2021) evaluated the performance
discretization. Then, multiple image filters were applied to of designs generated by GAN through FEM, and Sim et al.
the output shape to generate various optimized shapes. Fi- (2021) clustered the designs generated by GAN model into
nally, 20 SIMP iterations were performed in the final desired k-means clustering. Besides simply generating shapes, Oh
mesh size. et al. (2018, 2019) refined the design through the process of
re-optimizing the GAN-generated structures as an initial
SVM design. Rawat and Shen (2018, 2019a) used GAN to gen-
SVMs are also often used for postprocessing and prediction in erate structures and then replace FEA with meta-models

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

TO: to predict design characteristics over CNN. As a manner of
reflecting the design features or performance during train-
1. Postprocessing: Chu et al. (2016), Strömberg (2019, 2020),
ing, Sharpe and Seepersad (2019), Shen and Chen (2019),
and Karlsson et al. (2021) discretized the boundary sur-
Rawat and Shen (2019b), Malviya (2020), Nie et al. (2021),
faces of designs by using SVMs. The design domains could
and Almasri et al. (2020) added features, such as design
be classified into two sets. Separating hypersurfaces were
constraints and load and boundary conditions, as the con-
used to maximize the distance from the boundary to the
ditions to GAN to generate designs under the influence of
closest point of each set. This approach enables the de-
features (conditions) during training. Jiang et al. (2019), and
signs to converge with smooth surface boundaries and it
Wen et al. (2019, 2020) utilized design constraints as the
can be manufactured. Furthermore, Strömberg (2019, 2020)
input conditions in the GAN, while simultaneously eval-
and Karlsson et al. (2021) performed FEA and meta-model-
uating performance on generated designs and reflecting
based design optimization (MBDO). For the MBDO, they
performance in training to generate performance-good de-
used an ensemble of meta-models, including quadratic re-
signs. In this case, Wen et al. (2019, 2020) derived high-
gression, kriging, radial basis function networks, polyno-
resolution designs by using PGGAN. Yu et al. (2019) and Li
mial chaos expansion, polynomial regression models, and
et al. (2019a) also used GAN for upscaling purposes for low-
SVR models. Their proposed schemes were using GAs fol-
resolution designs predicted by CNN.
lowed by sequential quadratic programming. In particular,
Strömberg (2019) and Karlsson et al. (2021) considered the AE
use of lattice structures for AM.
2. Prediction: Xia et al. (2017) used SVM to predict sensitiv- 1. Exploration: There are also quite a few methodologies us-
ity for the purpose of removing the checkerboard pattern ing AEs, another representative methodology of unsuper-
while optimizing the process. vised learning. In the work of Ha et al. (2021), AE generates
a left, right robot gripper that fits the optimized designs
Others and 3D CNN model predicts its performance. Yamasaki et
al. (2021) implemented the re-training process using high-
1. Prediction: Among the other researchers that use super-
performance designs among the data generated using VAE,
vised learning methodologies, the work of Keshavarzzadeh
and Guo et al. (2018) combined style transfer with the VAE
et al. (2020) applied the GP to uncertainty-based stress-
so that the designs generated in latent space resemble the
based TO. On the basis of the low-resolution FEA data, they
original design and are feasible.
constructed a GP kernel and trained a GP emulator to pre-
dict maximum stress values and sensitivity for data points PCA
in design domains through high-resolution FEA data. At As another unsupervised learning methodology, the method
this time, high-resolution and low-resolution data points using PCA is distinguished for the purpose of design exploration
are extracted through Monte Carlo search. Lynch et al. and prediction:
(2019) recommended the tuning of parameters by measur-
ing the similarity between the current optimization prob- 1. Exploration: Lei et al. (2019) used PCA for parameter opti-
lem and the optimization problem of the previous database mization used in MMC TO. Through PCA dimension reduc-
and the meta-learning approach (potential tuning parame- tion, they get principal components of MMC variables and
ters for the cantilevered beam SIMP TO problem: weight of conduct SVR or KNN regression to obtain a new design. Ulu
objective functions, mesh size, constraint scaling, density et al. (2016) trained a meta-model that maps the principal
initial values, SIMP exponent penalty, objective tolerance, component of the PCA dimensionally reduced space for the
maximum number of forward problem evaluations, min- optimal shape and the load configuration corresponding to
imum allowable stiffness, and meshing method/element the design parameters of the shape. Accordingly, it is possi-
type). By using parameters defined through meta-learning, ble to predict the principal components of PCA for the new
the parameters are tuned through Bayesian optimization design for the new loading configuration and derive a new
in the optimization progression (meta-modeling). Zhou design.
and Saitou (2017) approximated the objective function for 2. Prediction: Li et al. (2019c) sampled data in a low-
the optimal design through an ensembled random forest dimensional space gained through PCA, obtained the per-
regression model. formance of the sample data numerically, and built a meta-
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1749

model for the performance of the sample data to proceed For design optimization, value-based RL is used. Hayashi and
with optimization. Ohsaki (2020) derived designs corresponding to design parame-
ters on a graph basis by using Q-learning (Watkins & Dayan, 1992).
Clustering
Brown et al. (2022) reformulated TO problems to a sequential RL
1. Optimization: As a methodology of using clustering al- task, a Markov decision process (MDP; Bellman, 1957), and opti-
gorithm, scholars have added clustering for shapes, de- mized elementally discretized 2D topologies by using the double-
sign characteristics, or FEM results to an iterative TO step deep Q-learning algorithm (Van Hasselt et al., 2016).
to reduce the calculation costs (Jiang et al., 2020; Kumar For design exploration, Sun and Ma (2020) used the e-greedy
& Suresh, 2020; Liu et al., 2015; Qiu et al., 2021b). Among policy to disturb the search direction, UCB to add a sensitivity fea-
them, Jiang et al. (2020) trained an ensembled ETs surro- ture, TS and IDS to direct the search. The optimal shape changes
gate model that determines whether it is feasible through were compared on the basis of the parameter changes. Jang et al.
the results of clustering. (2022) conducted a generative design method on designs to ex-
plore and eventually maximize design diversity by using the prox-
3.1.3. Semi-supervised learning imal policy optimization (PPO) algorithm (Schulman et al., 2017).

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

Traditional semi-supervised learning is the branch of ML con-
cerned with the use of labeled data as well as unlabeled data
3.2. Data
to perform certain learning tasks; it is conceptually situated
In TO, ML is used for various purposes, as mentioned in Section 2.
between supervised and unsupervised learning (Van Engelen &
The input type of data also varies depending on the purpose of the
Hoos, 2020). This method includes the field of active learning, in
ML model. The form of the input may be characterized according
which the learning algorithm can ask queries for the labels of pre-
to the form of use of the ML model.
viously unlabeled data points to be labeled by an oracle (e.g., a
Design parameters input to derive features that meet the de-
human annotator). Learning algorithms can achieve much greater
sign constraints, intermediate optimization design input to ac-
accuracy with fewer labeled training data (Settles, 2012). However,
celerate TO or FEA meta-modeling, final optimization, or struc-
in some cases, some algorithms are difficult to define as super-
tural analysis input to predict the performance, and FEA meta-
vised or unsupervised learning. These configurations are concep-
modeling, design generation, and coordinate information input
tually situated between supervised and unsupervised learning be-
to derive the optimal density value as a means of coordinating
cause they combine the concepts of both algorithms, using them
the various combined information input. This section reviews the
in parallel (e.g., combining GANs or AEs with NNs; McDermott
DLTO frameworks focusing on DL input.
et al., 2018) or step by step (e.g., modifying GANs and DBNs). Ac-
cordingly, we consider both traditional semi-supervised learning
and models that combine the concepts of supervised and unsu- 3.2.1. Design parameter (condition)
pervised learning as semi-supervised models. From this point of The design parameter corresponds to condition information for
view, semi-supervised learning is a useful methodology for design designing. For example, the value of the load and boundary con-
generation while ensuring good performance. All of the studies in- ditions, the value set as constraints in the TO process, such as
troduced in this section were aimed at design optimization. Some volume fraction and mass, or the coordinate information, are in-
methods of training were achieved by active learning, modifying cluded, and the design parameter is variously defined according to
traditional networks (e.g., combining GANs or AEs with NNs or the applications. When design parameters are used as input, the
modifying existing GANs) and DBNs (combination of supervised model is trained to the optimal design shape to satisfy the design
and unsupervised learning steps). parameter settings, and it can be divided into three methods. First,
Cang et al. (2019) conducted active learning with adaptive sam- a prediction method was used to predict the performance corre-
pling. By providing design parameters to NN, they could train the sponding to the design parameter. Second, a design optimization
model, obtain near-optimal solutions and find new data points method was implemented to map the information in the design
to compute corresponding optimal designs. In this manner, the parameters to the optimal geometry. Finally, a design exploration
model could compute the design with a less expensive computa- method was designed to create new geometry under the influence
tional budget. of design parameters.
Wang et al. (2020) used a combined model of VAE and regression Prediction
NN to allow latent space to contain physical properties. Similarly, As a study to predict the performance corresponding to the de-
Gillhofer et al. (2019) and Blanchard-Dionne and Martin (2021) sign parameters, White et al. (2019) predicted the performance of
proposed a methodology for training to generate performance- parameters that present the meta-material to the microscale as
satisfactory designs by changing the label of the discriminator an input. On this basis, sensitivity analysis of the optimization of
to reflect it on loss or by attaching a forward network. Kudyshev the macroscale is performed. Kim et al. (2021) predicted effective
et al. (2021) used an AAE, which combines conditional-AAE with elasticity tensor using the characteristic parameter of the mate-
NN to create a design containing physical information. Jiang and rial by using RVE methodology as an input. Thereafter, the FEA
Fan (2019a, b) trained the model to approximate the optimal de- process may be replaced, and the structure that considers the pe-
sign by reflecting the performance of design generated using the riodic structure may be optimized.
generative NN in loss function to reflect the physical informa- OptimizationIn the studies for design optimization, Cang et al.
tion. Further that, Kallioras and Lagaros (2020, 2021a) and Kallio- (2019), Abueidda et al. (2020), Kollmann et al. (2020), Halle et al.
ras et al. (2020) proposed methods that combined DBNs and the (2021), and Malviya (2020) predicted the design with design pa-
SIMP method to refine on similar predicted features. rameters as inputs for non-iterative optimization. Keshavarz-
zadeh et al. (2021) generates design through the DNSM parameters
3.1.4. Reinforcement learning after training a model that maps design parameters to DNSM pa-
RL has a high potential for generating optimal designs, but the re- rameters for non-iterative optimization. To carry out sequential
search on the application of RL to TO remains to be insufficient. optimization process, Lew and Buehler (2021) used latent vectors
1750 | Topology optimization via artificial intelligence: a review

as design parameters. Lynch et al. (2019) recommended the tuning discretization provided by the initial steps of the SIMP method.
of parameters for the corresponding design problem during opti- Although Hoyer et al. (2019) and Zhang et al. (2021a) used noise
mization and uses Bayesian optimization to tune the parameters as an input, not the intermediate optimization results, it deter-
and proceed with optimization. Lei et al. (2019) used PCA dimen- mines the performance of the shape generated from the noise and
sion reduction for design parameters to obtain eigenvector and optimizes the design by allowing the performance to affect net-
eigenvalue. Using this information, the design is derived through work parameter updates with the backwards pass. Liu et al. (2015)
non-linear regression. Ulu et al. (2016) predicted the PCA dimen- classified the shape of intermediate iteration into k number of ma-
sionally reduced value of the topology by using the design param- terials through K-means clustering. On this basis, optimization is
eter (i.e., loading configuration) as an input. When a new design re-progressed and obtained optimal material distribution.
parameter is introduced, the shape corresponding to the predicted
PCA value is derived. 3.2.3. Final optimization result
Exploration Studies with the final shape as an input can be largely divided into
As for the studies on design exploration, Jiang et al. (2019), Shen three purposes: the purpose of prediction, the purpose of design
and Chen (2019), Wen et al. (2019), Jiang and Fan (2019b), Rawat exploration, and the purpose of postprocessing.

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

and Shen (2019b), Almasri et al. (2020), and Sharpe and Seepersad Prediction
(2019) trained generative models that generate a new design by In studies for performance prediction purposes, Rawat and
entering design parameters. Yu et al. (2019) and Li et al. (2019a) Shen (2018) and Blanchard-Dionne and Martin (2021) predict the
generated a low-resolution image by inputting a design parame- parameters or performance of the shape based on that generated
ter and upscale it into a GAN. Hayashi and Ohsaki (2020) selected with the generative model trained with the optimal designs. Bujny
design parameters with initial environment and derived the graph et al. (2018) predicted the performance ranking of the node–edge
features through Q-Learning. pair of the graph representing the shape as an input graph. This
method accelerates optimization by creating a predictor for shape
3.2.2. Intermediate optimization result changes.
When the intermediate optimization result is used as the input, Li et al. (2019c) dimensionally reduced the dimension of the op-
it is divided into a method for prediction and a method for design timal shape by using logistic PCA and sampled data in the latent
optimization from that point. space. Then, the shape is obtained through inverse transforma-
Prediction tion, and then a surrogate model is trained to predict the perfor-
The purpose of prediction is primarily meta-modeling. DL mod- mance of the shape. Zhou and Saitou (2017) predicted the perfor-
els replace calculation of TO process (calculation of objective mance (resin filling time) corresponding to the input of the resin
function or sensitivity, etc.) to accelerate TO. Doi et al. (2019), filling data generated using the KLS.
Sasaki and Igarashi (2019a, b), Asanuma et al. (2020), Lee et al. Exploration
(2020), Deng et al. (2022), and Zheng et al. (2021) proposed surro- In studies aimed at design exploration, Guo et al. (2018) trained
gate models that replace objective functions, such as performance VAE with optimization features as inputs and to find new designs
with DL models, by inputting intermediate optimization results in latent space. To create a better latent space, Wang et al. (2020)
for each application. Takahashi et al. (2019) utilized intermediate determined the optimization features by adding a regressor to the
optimization results as the input to replace sensitivity calcula- VAE, and predicted the performance of the design when latent
tions with CNNs, while Xia et al. (2017) removed checkerboard pat- variables are input through the regression model. Through this
terns by replacing the process of calculating sensitivity with in- approach, latent space can be formed to generate designs with
putting information from specific data points at n-th iteration to good performance. Yamasaki et al. (2021) generated designs by us-
SVM. Gaymann and Montomoli (2019) focused on a game format ing VAE, to generate high-performance shape, re-training is con-
in which the design means player, and is used as input. They pre- ducted using high-performance generated data. In addition, stud-
dicted the winner and the end of the game, repeated the process of ies that generate new designs with GANs are trained using shapes
determining the next movement through Monte Carlo search, and generated from noise and final optimization results. Oh et al. (2018,
optimized the fluid structure that could minimize the pressure 2019), Jiang and Fan (2019a), Greminger (2020), Li et al. (2021), Sim
difference. Keshavarzzadeh et al. (2020) predicted the maximum et al. (2021), and Gillhofer et al. (2019) performed studies in which
stress value and sensitivity for data points of unknown designs a discriminator is trained to distinguish optimal shapes and gen-
through GPemulator. At this time, a covariance kernel affecting erated shapes by the generator without any condition. Oh et al.
the GPemulator is constructed through low-resolution FEA data. (2019) further trained AE to evaluate design originality through
Patel and Choi (2012) performed a process of determining reconstruction error, and Gillhofer et al. (2019) trained the discrim-
whether the region exceeds the safety criterion by inputting each inator to classify performance criteria rather than simply focus-
data point when the iteration is repeated. Jiang et al. (2020) prepro- ing on real/fake classification, allowing the generator to generate
cessed the shape with representing features (called visual word better performance features.
vectors in the study) during the optimization process and quickly Jang et al. (2022) conducted PPO by using AE as an environment
determined whether the shape is feasible through the correspond- pre-trained by final optimization results. At this time, the agent
ing vectors. network generates a wheel and gives its parameters as action, and
Optimization the environment outputs reward and state. Furthermore, Ha et al.
In a study for design optimization, Sosnovik and Oseledets (2021) generated left and right robot grippers suitable for the op-
(2019) and Lin et al. (2018) predicted the final design by inputting timized 3D shape by using the optimized 3D shape as input and
the intermediate density distribution of the design in the mid- predicted their performance.
dle nth iteration and the gradients between the nth and n−1th Postprocessing
iteration to a model for acceleration of TO. Kallioras and Lagaros In the postprocessing fields, Chu et al. (2016), Strömberg (2019,
(2020, 2021a, b) and Kallioras et al. (2020) predicted their respec- 2020), and Karlsson et al. (2021) processed the boundaries through
tive designs by using the density fluctuation pattern of the FE SVM when optimization designs are input. Moreover, Wang et al.
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1751

(2021), Xue et al. (2021), and Napier et al. (2020) predicted high- of deriving optimal designs. Qiu et al. (2021a) also used the design
resolution shapes when optimal low-resolution shapes are input. parameter (load condition) and the shapes of previous iterations
In addition, Yildiz et al. (2003) predicted the shape of the hole (tri- as inputs. Rade et al. (2020) implemented two methodologies, DS
angle, trapezoid, rectangle, etc.) of the design by using topology and CDCS prediction, as mentioned in Section 2.2, in which the
optimized shapes as input, thereby enabling appropriate shape DS methodology was used to predict optimal designs by using ini-
optimization to proceed. tial design parameters and FEA information. The CDCS methodol-
ogy can predict optimal designs through intermediate optimiza-
3.2.4. Structural information (FEA information) tion results and intermediate FEA information. Qian and Ye (2021)
The two main purposes in which FEA information is used as input replaced compliance and sensitivity calculations by inputting de-
are prediction and design optimization. sign parameters and intermediate optimization results for meta-
Prediction modeling purposes.
As a study aimed at FEA meta-modeling, Aulig and Olhofer
(2014) replaced sensitivity calculations with predictive models by 3.3. Loss function
entering density FEA information (displacement) of shapes, and The form of loss function of DL model in TO is defined as follows

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

Bi et al. (2020) used high-fidelity FEM simulation data to predict depending on the objectives to achieve through DL models.
the gradient of shape in each iteration, replacing gradient calcu-
lation, accelerating TO. Chi et al. (2021) conducted an online train- 1. Difference between generated and ground truth shape
ing of TO and predicted the sensitivity information by using the 2. Difference between predicted and ground truth perfor-
FEA value of the coarse data and the filtered design variables of mance
fine data of the same shape as inputs for TO. Furthermore, Li et al. 3. Add performance value/gradient obtained from FEA
(2019b) used the FEA value as an input to determine whether the 4. Add additional constraints
shape is damaged and to determine whether the sensitivity anal- 5. Add two types of loss function for multiple models
ysis proceeds.
Optimization 3.3.1. Difference between generated and ground truth
As design optimization studies, Zhang et al. (2019) and Nie et al. shapes
(2021) predicted and generated optimal shapes accordingly by in- To minimize the difference between generated and ground truth
putting both design parameter information and FEA information. shapes, Abueidda et al. (2020) and Kollmann et al. (2020) calcu-
Qiu et al. (2021b) utilized FEA information as the input, and this late the difference between the shape predicted using the NN
FEA information was clustered to determine the internal shape and the ground truth shape in quadratic norm form. Xue et al.
structure, and Kumar and Suresh (2020) performed clustering (2021) uses MSE loss to calculate the difference between high-
with the density of the shape and the FEA information input to resolution generated shape and real shape to map low-resolution
improve the performance. shape to high-resolution shape. Furthermore, Keshavarzzadeh
et al. (2021) predicted multiresolution TO results, including low-
3.2.5. Coordinates in design space resolution predictive shape and combining MSE loss function for
The two purposes related to the inputting of the coordinate value low-resolution shapes and high-resolution shapes. Lastly, because
of the design domain are the design optimization purpose and Qiu et al. (2021a) used binary data (i.e., ESO method), they chose
the meta-modeling purpose. Chandrasekhar and Suresh (2021a, binary cross-entropy (BCE) loss.
b, c) predicted the density value of the coordinates, and the op-
timal shape is derived through minimization of loss function. 3.3.2. Difference between predicted and ground truth per-
At this time, the loss function contains FEA to satisfy the TO formance
settings. Zehnder et al. (2021) also used data points as input To minimize the difference between predicted and ground truth
and subsequently trained the displacement network by mini- performance, Lee et al. (2020) used MSE and MAPE as loss func-
mizing the total potential energy of the system and updating tion to calculate the difference between the predictive and real
the displacement. Then, sensitivity analysis was performed to values of objective function and constraints in the optimiza-
compute the density–space gradients, which, after applying our tion. White et al. (2019) defined the loss function in combina-
sensitivity filtering, resulted in target density fields. Finally, the tion with the quadratic loss form of the difference of the ob-
density was updated by minimizing the loss function between jective function and the derivative component of the objective
the current and target densities. Deng and To (2020) predicted function. Gaymann and Montomoli (2019) also used the quadratic
the DRL design variables that represent voxel shapes. Through loss form to approximate the objective function of the opti-
this approach, the design variable was reduced to accelerate mization process. Patel and Choi (2012) used a quadratic loss
optimization. function to determine the feasibility of the shape in the TO
process.
3.2.6. Multimodal input Gillhofer et al. (2019) used BCE to determine the low or high
The study of multimodal input data refers to studies that use response region. The weight was adjusted for shapes with a high
various input data for a model. The works of Aulig and Olhofer response higher than the threshold value. In this manner, the gen-
(2014), Zhang et al. (2019), Nie et al. (2021), and Chi et al. (2021) erator can generate a shape of the low response region.
mentioned in Section 3.2.4 are exactly included in this section,
but they are not elaborated in the previous section to highlight 3.3.3. Adding performance value/gradient obtained from
examples with FEA information. Brown et al. (2022) provided the FEA
design parameters (load and boundary conditions) and the FEA The studies aimed at adding performance value/gradient objec-
information (stress) to the RL agent to take a new action. Banga tive from FEA defer the use of performance data as the label, but
et al. (2018) predicted the optimal designs by inputting design pa- they add a physics term to the loss function itself to satisfy the
rameters and intermediate optimization results for the purpose objective function, as shown in Method 2 in Fig. 4. Chandrasekhar
1752 | Topology optimization via artificial intelligence: a review

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

Figure 4: Methods of using FEA information for training models.

and Suresh (2021a, b, c) defined the loss function to reflect the 3.3.5. Adding two types of loss functions for multiple mod-
objective function and constraint values calculated through FEA els
of shape in each iteration to optimize the structure. Hoyer et al. In studies involving two or more networks, the use of more than
(2019), Jiang and Fan (2019a, b), Qian and Ye (2021), and Zhang two loss functions for each network is reasonable. Rawat and Shen
et al. (2021a) also incorporated the performance values obtained (2018, 2019a) generated shapes by WGAN by using Wasserstein
through FEA into loss function and repeat the process of updating distance (Section 3.3.1) and predicted the parameters of the gen-
the variables by calculating gradients through backpropagation to erated shapes by using the MSE loss (Section 3.3.2). Similarly, Ha
optimize the shape. et al. (2021) used MSE loss for a generative network to generate
shapes and used BCE to determine the feasibility of generated
shapes through 3D CNN. Li et al. (2019a) used standard GAN loss
to generate low-resolution shapes over GAN and added the MSE
loss between real, predicted shapes in SRGAN for upscaling (Sec-
tion 3.3.1). Yu et al. (2019) selected mean absolute error (MAE)
3.3.4. Integrating additional constraints
loss of real and predicted low-resolution shapes on CNN network
The loss function to add additional constraints is defined in the (Section 3.3.1) and added MAE loss term on the standard GAN
form of adding a specific constraint loss term to the existing loss. loss to resemble the real shape on the cGAN network (Section
Zhang et al. (2019) used Kullback–Leible divergence (Kingma & 3.3.1). Li et al. (2019b) used MSE loss to map macroscale strain and
Welling, 2013) to obtain the difference between the prediction and macroscale stress (FFNN) and predict the feasibility of microstruc-
the real shapes corresponding to Section 3.3.1, and further defines ture with a stress value (Section 3.3.2). Rade et al. (2020) utilized a
additional loss in the form of L2 regulation term to prevent overfit- different type of loss for two methodologies. The study proposed a
ting. Guo et al. (2018) added a reconstruction error corresponding DS method with MSE loss to calculate the differences between real
to standard VAE loss, a style loss (Gatys et al., 2015), which eval- and predicted shapes (Section 3.3.1) and a CDCS prediction net-
uates how well the generated shape match the style of the real work with loss that combines MAE loss that calculates the differ-
shape, and a mode-collapse loss (Zhao et al., 2016) to prevent only ence between real and predicted shapes in the early stages (Sec-
similar samples from being generated. Nie et al. (2021) defined BCE tion 3.3.1) while the BCE loss used to predict compliance in terms
of standard GAN loss, L2 loss term between real and fake shapes of optimal design convergence (Section 3.3.2).
corresponding to Section 3.3.1 and further added an absolute er-
ror term between the target volume and the volume of predicted
shape. Sosnovik and Osledets (2019), Lin et al. (2018), and Banga 3.4. Physics
et al. (2018) defined the BCE loss to determine the difference be- FEA is used in DLTO to ensure a much more efficient convergence
tween predicted and real shapes (Section 3.3.1) and add MSE loss of optimal designs and produce meaningful results. This concept
for the volume constraint term to the loss. is similar to a physics-guided NN and a physics-informed NN
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1753

(PINN) methodology, which have recently emerged in engineering 3.5.1. Application

fields, and represent a method for allowing DL models to learn the Studies on DLTO is a representative application such as beam,
physical meaning of the data (Zhuang et al., 2021a). For DLTO prob- bridge, truss, nanophotonics, meta-surface, wheel, motor, heat
lems, FEA information is reflected largely in three ways (Fig. 4): conduction system, and fluid structure, etc. Until now, studies on
2D shapes have been dominated. It is gradually being used in var-
ious fields, and 3D shape optimization research is also being con-
ducted (Table 3).
1. Method 1. FEA as a method of evaluating performance dur-
ing the training process, 3.5.2. Generalization
2. Method 2. FEA as a term of loss function during the training
The generalizability of the methodology can be classified into
process, and
three possibilities:
3. Method 3. FEA as an information to use as input when
training. 1. Can be applied to various load and boundary conditions,
2. Can be applied to 3D design problem, and
3.4.1. FEA as a method of evaluating performance in the

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

3. Can be applied to any design domain.
training process
Method 1 uses FEA as a performance evaluation tool for the result This section explored each topic in more detail.
shapes derived through the trained model. Then, the dataset is re- Application to various load and boundary conditions
constructed with the generated data that is determined to be good The generalization for various load and boundary conditions
through FEA and re-training is conducted on the reconstructed has the advantage of being able to cover various structures. Stud-
dataset, allowing the model to predict the shape with good perfor- ies applicable to various load and boundary conditions within the
mance (Halle et al., 2021; Jiang et al., 2019; Wen et al., 2019, 2020; fixed design domain may be considered in this section.
Yamasaki et al., 2021). Thus, FEA is used as a performance evalu-
ation tool to reinforce the performance of generated data. 1. Diversification of load conditions: Abueidda et al. (2020),
Qiu et al. (2021a), and Ulu et al. (2016) changed the position
3.4.2. FEA as a loss function term in the training process and direction of the load while the boundary conditions
Method 2 reflects FEA information into a loss function during were fixed for 2D brackets, so allowed load condition gen-
training. The loss function reflecting FEA information converges eralization. Similarly, Sharpe and Seppersad (2019) gener-
in the direction of minimization. In other words, if the FEA infor- ated optimal shapes by diversifying loading conditions for
mation is used as a loss function, then the model is trained in the the 2D truss structure. Cang et al. (2019) also fed diversi-
direction of increasing performance through backpropagation for fied loading conditions to a model to adaptively learns from
the corresponding FEA loss term (Chandrasekhar & Suresh, 2021a, true solutions from a distribution of problems. Thereafter,
b; Hoyer et al., 2019; Jiang & Fan, 2019a, b; Zehnder et al., 2021; solutions to other unobservable problems from the same
Zhang et al., 2021a, b). distribution could be predicted. Bujny et al. (2018) selected
one of the three cases of boundary conditions and the load
conditions at random locations within the design domain.
3.4.3. FEA as an information for use as input during train-
Oh et al. (2018, 2019) and Jang et al. (2022) changed the di-
ing
rection and ratio of load conditions with fixed boundary
Method 3 allows the model to be trained with the FEA infor-
conditions in the automobile wheel design. Li et al. (2019a)
mation itself by inputting FEA information as an input of the
proposed a methodology to allow the heat sink position,
model. Nie et al. (2021) and Zhang et al. (2019) predicted the
the heat source position to be changed for the conductive
features by putting basic design conditions and FEA informa-
heat transfer system. Qian and Ye (2021) diversified the
tion together as inputs. Bi et al. (2020) used only the data ob-
load conditions for 2D bracket designs and used Poisson’s
tained through FEA as an input to predict the gradient of the
ratio for meta-material design. Furthermore, Li et al. (2019c,
shapes during TO, and Chi et al. (2021) predicted the sensitiv-
2021) reconstructed or generated periodic structures from
ity by using FEA information of low-resolution shapes to op-
latent space trained with optimal shapes at various load
timize high-resolution shapes. Rade et al. (2020) presented two
conditions, respectively, allowing for the use of the general
methodologies by using (design parameter + FEA information),
load conditions.
(intermediate optimization result + FEA information) to pre-
2. Diversification of load and boundary conditions: Yu et al.
dict optimal shapes and the methodology presented by Kudy-
(2019), Zhang et al. (2019), Nie et al. (2021), and Malviya
shev et al. (2021) also used design parameters and physical in-
(2020) utilized load and boundary conditions for design as
formation through FEA to generate data. As a study consider-
inputs to model, enabling optimal shape prediction for var-
ing microscale, Qiu et al. (2021b) performed clustering with only
ious load and boundary conditions. Halle et al. (2021) also
FEA information, and Kumar and Suresh (2020) used density
predicted the optimal shapes by inputting load and bound-
and FEA information of shape to determine microstructure sim-
ary conditions for the 2D bracket and repeated re-meshing
ilar to that in the work of Li et al. (2019b) that considered mi-
and predicting until the objective function is converged.
cro/macroscales, predicts damage by using FEA information on
Hoyer et al. (2019) and Zhang et al. (2021a) updated the loss
macro/microstructures.
function containing load and boundary condition informa-
tion for the 2D shape predicted by NN through a gradient
3.5. Applications and generalization based algorithm to enable generalization. Wang et al. (2021)
The application fields of DLTO are diverse and are gradually ex- trained a model that upscale low-resolution 2D bracket
panding. This section reviews application examples and general- shapes to high resolution with diversified load and bound-
ization cases of this methodology. ary conditions.
1754 | Topology optimization via artificial intelligence: a review

Table 3: DLTO applications.

Scale Applications Studies Examples

2D Abueidda et al., 2020; Almasri et al., 2020; Aulig & Olhofer, 2014; Bi
et al., 2020; Brown et al., 2022; Bujny et al., 2018; Cang et al., 2019;
Chu et al., 2016; Deng et al., 2022; Gillhofer et al., 2019; Halle et al.,
2021; Hayashi & Ohsaki, 2020; Hoyer et al., 2019; Jiang et al., 2020;
Kallioras & Lagaros, 2020, 2021a, b; Kallioras et al., 2020;
2D design domain Keshavarzzadeh et al., 2020; Lee et al., 2020; Lei et al., 2019; Lew &
(beam, bridge, truss, Buehler, 2021; Li et al., 2019a; Liu et al., 2015; Lynch et al., 2019;
bracket, etc.) Malviya, 2020; Napier et al., 2020; Nie et al., 2021; Patel & Choi, 2012;
Qian & Ye, 2021; Qiu et al., 2021a; Rawat & Shen, 2018, 2019b;
Sharpe & Seepersad, 2019; Shen & Chen, 2019; Sim et al., 2021;
Sosnovik & Oseledets, 2019; Takahashi et al., 2019; Ulu et al., 2016;
Wang et al., 2021; Yamasaki et al., 2021; Yu et al., 2019; Zehnder et

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

al., 2021; Zhang et al., 2019, 2021a; Zhou & Saitou, 2017

Nanophotonics, Blanchard-Dionne & Martin, 2021; Jiang & Fan, 2019a, b; Jiang et al.,
spinodoid, 2019; Kim et al., 2021; Kollmann et al., 2020; Kudyshev et al., 2021;
meta-surface, periodic Kumar & Suresh, 2020; Li et al., 2019b, c, 2021; Qian & Ye, 2021; Qiu
structure et al., 2021b; Wang et al., 2020; Wen et al., 2019, 2020

Wheel Jang et al., 2022; Oh et al., 2018, 2019

Motor Asanuma et al., 2020; Doi et al., 2019; Sasaki & Igarashi, 2019a, b

Deng et al., 2022; Guo et al., 2018; Keshavarzzadeh et al., 2021; Li et

Heat conduction
al., 2019a; Lin et al., 2018; Zhang et al., 2021a

Fluid structure Deng et al., 2022; Gaymann & Montomoli, 2019

Banga et al., 2018; Chandrasekhar & Suresh, 2021a, b, c; Chi et al.,

2021; Deng & To, 2020; Deng et al., 2022; Greminger, 2020;
3D design domain
3D Keshavarzzadeh et al., 2021; Rade et al., 2020; Rawat & Shen, 2019a;
(beam, bracket, etc.)
Strömberg, 2019, 2020; Sun & Ma, 2020; Xue et al., 2021; Yildiz et al.,
2003; Zehnder et al., 2021

Meta-material Karlsson et al., 2021; White et al., 2019; Zheng et al., 2021

Robot gripper Ha et al., 2021

Application to 3D design problems bility of DLTO. Therefore, we explore the studies that enable the
DLTO still lacks an expansion mechanism for the 3D design do- change in design domains. Hayashi and Ohsaki (2020) expressed
main. However, as expansion to 3D is essential for future usability, the 2D truss structure with nodes and edges, allowing the size of
studies with 3D applications were reviewed. the design domain to be adjusted. The RL agent learned general-
Rawat and Shen (2019a) optimized for 3D beams but did not izations for the design domain and could discover optimal truss
consider various load and boundary conditions. Meanwhile, Banga topologies for various load cases. Brown et al. (2022) trained the
et al. (2018), Rade et al. (2020), and Keshavarzzadeh et al. (2021) gen- RL agent to design optimal topologies (i.e., 6 × 6 in size) with di-
eralized for various load and boundary conditions and proceeded verse load and boundary conditions. Then, they tested a multistep
with extensions to 3D objects. Chi et al. (2021) updated the strain progressive refinement approach to prove that the RL agent could
information about coarse mesh online and optimizes fine mesh. adopt a generalized design strategy.
The study calculated strain for various load and boundary con- Bi et al. (2020) applied the approach to general problems by up-
ditions and applied it to 3D. Kim et al. (2021) performed the op- dating the corresponding gradient online for a given 2D shape
timization of structures considering materials in the problem of rather than using the trained model. Kumar and Suresh (2020) and
2D and 3D structure optimization through FEA meta-modeling of Qiu et al. (2021b) optimized the 2D design domain by clustering mi-
periodic structure. Extensions to 3D are also applied in various ar- crostructure with strain and stress information, respectively. The
eas. Greminger (2020) conducted a study of generating synthetic clustering process may be applied to various design domains.
3D voxel dataset. White et al. (2019) replaced the process of calcu- While some studies have expanded their focus to various
lating effective elasticity of 3D metal with NN, and Ha et al. (2021) 2D design domains, other studies have expanded to 3D. Chan-
optimized 3D robot grippers for 3D optimized shapes. drasekhar and Suresh (2021a, b, c), Zehnder et al. (2021), and Deng
Application to any design domain and To (2020) used the network itself as an optimization calcula-
The most important point of expanding the application area is tion tool, thereby allowing the network to be used for any design
whether the design domain can be set flexibly beyond (1) and (2) domain and any load and boundary condition, and it can even
in Section 3.5.2, which is the most important issue in the applica- be expanded to 3D. The DLTOP methodology (Kallioras et al., 2020)
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1755

and other applied versions (Kallioras & Lagaros, 2020, 2021a, b) are of high computational cost and time consumption (Blanchard-
also applicable to any design domain and various load and bound- Dionne & Martin, 2021; Doi et al., 2019; Gaymann & Montomoli,
ary conditions when optimization progression is combined with 2019; Sasaki & Igarashi, 2019b; Zhang et al., 2021a). For example,
SIMP and DL. Zheng et al. (2021) matched the stiffness tensors of Chandrasekhar and Suresh (2021a) demonstrated that their FEA
3D macroscale structures with the predicted stiffness tensors of consumed approximately 50% of the computational cost in the
microscale spinodoid meta-materials to optimize the spinodoid case of a 60 × 30 mesh size, and the percentage would likely in-
topologies of macrostructures with various load and boundary crease with increasing mesh size. The design generated in this
conditions. Xue et al. (2021) generalized the scheme to all differ- method guaranteed engineering performance and increase the
ent load and boundary conditions, and even 3D design domains, accuracy of the network, but it imposed a burden on the computa-
for the upscaling at high resolution. tional cost; this scenario implies a more significant disadvantage
when used in large-scale problems, such as 3D design problems.
Consequently, some studies have attempted to introduce meth-
4. Limitations and Research Opportunities ods such as parallel FEM computation (Deng et al., 2022), although
The previous sections have verified that studies are being con- more research is needed to utilize FEA while attempting to effec-

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

ducted with various approaches to accelerate TO and improve tively minimize the computational cost.
the accuracy. However, limitations exist in relation to the usability
and necessity of MLTO. In this section, the limitations commonly 4.1.3. Physics versus data
found in existing MLTO studies are summarized. On the basis, fu- The computational complexity of MLTO shows the need for PINN
ture research opportunities are examined. that can learn the underlying physics instead of the use of a large
pre-optimized dataset as the label. Similarly, recent studies have
4.1. Computational complexity shown the possibility of solving the partial differential equation
(PDE) through DL by capturing the underlying physics patterns
As shown in the previous sections, the aim of MLTO is to reduce
(Anitescu et al., 2019; Samaniego et al., 2020). PINN is a method of
the computational cost of TO problems. For example, as shown
reflecting additional information obtained by applying the laws
in Table 4, MLTO can instantly infer the optimal topology once
of physics into network training and utilizing PDEs that describe
trained compared with the conventional TO.
physics problems into the loss functions (Raissi et al., 2019).
According to Zhang et al. (2019), by reflecting the physics in-
4.1.1. Data collection burden
formation in the loss function, a structural disconnection of the
An important consideration in the MLTO field is the advantageous
generated optimization result may be prevented by generating the
feature of MLTO, which is to focus on the short inference time af-
design based on engineering knowledge rather than simply pre-
ter completing the training. As shown in Table 4, the data collec-
dicting features based on the similarity between pixels. In addi-
tion cost to complete the MLTO and the efficiency of front load-
tion, this method can train even with a small amount of training
ing, such as the training time of the model, should also be consid-
data, and it can reduce the burden and cost of data collection. In
ered. Many studies on supervised MLTO require a large amount of
other words, an advantageous method may be provided for 3D TO.
TO to collect optimized training data (Kallioras & Lagaros, 2020,
Consequently, scholars have planned to apply PINN to MLTO as a
2021b; Rade et al., 2020; Sharpe & Seepersad, 2019; Wang et al.,
future work (Chandrasekhar & Suresh, 2021c; Ha et al., 2021; Ke-
2021; Wen et al., 2019; Wen et al., 2020; Yu et al., 2019). In partic-
shavarzzadeh et al., 2021; Rade et al., 2020; Zhang et al., 2019). In
ular, more data are required in 3D MLTO, but sufficient amounts
addition, as it is not a method of simply predicting a shape based
of data are not easy to collect, and the initial computational cost
on the reference data, extrapolation may help to generate a new
for data generation is much larger. For instance, Zhang et al., 2019
result that has not been seen before (Sim et al., 2021). However,
generated a dataset of 80 000 with 88 lines of code (Andreassen
PINN research showing the product-level output that can be used
et al., 2011) to train a non-iterative MLTO NN. The input of the
in the actual industry in MLTO is not yet shown, and it remains
41 × 81 × 6 tensor, which included the boundary conditions, load-
only a benchmark problem.
ings, initial displacement field, strain field, and volume fraction,
was used. The output of a 40 × 80 matrix to represent the den-
sity of the elements was used for the NN, replacing the classical 4.2. Scalability
compliance minimization problem solved by SIMP. Similarly, Yu et One of the most important issues of MLTO studies is the scala-
al., 2019 used 100 000 optimized structures to train non-iterative bility of applicability to actual product development and various
MLTO networks, with a solution domain of a 32 × 32 grid at low domains. With a single methodology, we should consider whether
resolution and a 128 × 128 grid at high resolution. To solve a 3D TO it can be extended to various TO or real-world problems and not
problem in a supervised manner, Rade et al., 2020 used a 3D mesh simply to one design problem or condition. However, as reviewed
cube as an initial design domain, which entailed 31 093 nodes in Section 3.5.2, most studies show limitations that these exten-
and 154 677 elements. A total of 13 500 samples were generated sions and applications are impossible. The problem of the compu-
by the iterative TO, which took an average of 13 iterations each tational complexity of MLTO is linked to the scalability of MLTO. If
time. MLTO does not have enough scalability, then it raises the question
of whether MLTO is effective enough to risk the computational
4.1.2. Computational cost of FEA cost of data collection willingly. Therefore, the problem definition
Computational cost and accuracy generally have a tradeoff re- and usage scenario should be studied to emphasize and prove the
lationship. Although MLTO aims for acceleration, the accuracy need for MLTO to overcome the aforementioned shortcomings.
is still inferior to that of the conventional iterative TO. However,
studies attempted to use solvers, such as FEA in the intermedi- 4.2.1. 3D design
ate process or the iterative TO in the middle part of the MLTO Most of the current studies have a limitation in that they cannot
process, to increase the accuracy would also raise the problems be applied to 3D problems (Abueidda et al., 2020; Almasri et al.,
1756 | Topology optimization via artificial intelligence: a review

Table 4: Computational cost of TO and MLTO.

Improvement
Research Element size TO time Training Data generation MLTO (times faster) Computational resources

Almasri et al., 2020 101×101 0.3244 s - - 0.0046 ms 70 521× GPU

Bi et al., 2020 480×160 600 s - - ∼150 s 8.60× ORNL OLCF Summit supercomputer
960×320 1400 s - - ∼200 s -
Cang et al., 2019 - 100∼1000 s - - 0.01 s 10 000∼ Intel Xeon CPU E5-1620 @ 3.50 GHz
100 000×
Chi et al., 2021 85 750 1346 s - - 924 s 1.46× two Titan Xp GPUs
250 000 3665 s - - 1515 s 2.42×
1 458 000 25 455 s - - 6069 s 4.19×
Halle et al., 2021 64×64 1.9 s - - 7.3 ms 259× Nvidia Titan RTX
Hayashi & Ohsaki, 2020 3×2 1.6 s - - 0.7 s 2.29× Intel® Core(TM) i9-7900X @ 3.30GHz
10×10 137.7 s - - 7s 19.67×
Jang et al., 2022 128×128 600 s - - 0.01 s 60,000× NVidia GeForce 1080 Ti GPU
Kallioras et al., 2020 83 200 54 287 s - - 20 910 s 2.60× Intel Xeon E5-1620 @ 3.70 GHz

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

quad-core (with eight threads) with 16
GB RAM & NVIDIA GeForce 640 with
384 cores and 2 GB RAM
Keshavarzzadeh et al., 128×128 2s 1639.3 s 20 203.8 s 0.002 s 1000× Intel 12 Core i7−5930K @ 3.5 GHz CPU
2021 with 32 GB RAM & 2x Intel 8 Core
E5−2660 @ 2.20 GHz processor with 64
GB of RAM & 2x Nvidia K20 GPUs
40×20×10 33 s 8836.4 s 331 213.1 s 0.002 s 16 500×
Li et al., 2019a 160×160 23 s 14 400 s 18 s 0.006 s 3833.33× Intel i7-8750H CPU with 8 GB RAM &
NVidia GeForce GTX 1060 GPU
Lin et al., 2018 80×80 64.635 s - - 18.297 s 3.5× -
Napier et al., 2020 160×100 591.7 s - 259 200 s 22.6 s 26.18× Intel i7-5930k CPU at 3.50 GHz and
16GB of ram
320×120 3912.62 s - - 554.9 s 7.3×
Nie et al., 2021 64×128 119.83 s - - 0.036 s 3328.61× TO : Intel i7-6500U central processing
unit (software ToPy) MLTO : NVIDIA
GeForce GTX 2080Ti GPU
Qian & Ye, 2021 36×36 0.42 s 917.4 s 973.8 s 0.0127 s 33.1× Intel(R) Xeon(R) CPU E5-2687W v2
(3.40 GHz)
64×64 0.97 s 4545.6 s 1897.80 0.0101 s 96.04× -
Qiu et al., 2021a 20×60 486 s 1710 s 7560 s 0.6 s 810× Intel(R) Core(TM) i9-9900 3.10 GHz
processors
20×60×10 2250 s 2946 s 33 840 s 2.9 s 775.86× -
Rawat & Shen, 2019b - ∼114.635 s - - ∼0.353 s 324.75× -
Sosnovik & Oseledets, 40×40 - 4800 ∼ 5400 s - - 20× NVIDIA Tesla K80
2019
Ulu et al., 2016 20×20 - ∼625.19 s - ∼0.0004 s - 2.4 GHz Core CPU and 8 GB RAM
Wang et al., 2021 128×128 24.13 s - - 0.19 s 127× Intel CoreTM i5-7500 processor
Yu et al., 2019 128×128 22.7 s - 2377 900 s 0.014 s 1621.4× Intel Xeon (8 cores and 2.5 GHz CPU
clock) & NVIDIA Tesla P100 (3584
CUDA cores of 1.19 GHz GPU clock and
16 GB memory)
Zhang et al., 2019 40×80 3.913 s - - 0.001 s 3913× Intel(R)Core(TM) i7-8700k CPU NVIDIA
GeForce GTX1080 Ti GPU
Zheng et al., 2021 - 300 s 600 s - 0.001 s 300 000× Two 18-core 2.7 GHz Intel Xeon Gold
6150 processors and 192 GB of DDR4
memory at 2666 MHz

2020; Chandrasekhar & Suresh, 2021b, c; Halle et al., 2021; Jiang 4.2.2. Boundary conditions
& Fan, 2019a; Kumar & Suresh, 2020; Liu et al., 2015; Napier et al., Many studies are applicable to only one design problem (Asanuma
2020; Nie et al., 2021; Oh et al., 2019; Qian & Ye, 2021; Qiu et al., et al., 2020; Ha et al., 2021; Hayashi & Ohsaki, 2020; Jang et al., 2022;
2021b; Sharpe & Seepersad, 2019; Ulu et al., 2016; Wang et al., 2020, Jiang & Fan, 2019a, b; Jiang et al., 2019; Wang et al., 2020, 2021; Wen
2021; Yildiz et al., 2003; Yu et al., 2019; Zhou & Saitou, 2017). In other et al., 2019; Zhang et al., 2019; Zheng et al., 2021; Zhou & Saitou,
words, the works mentioned above are not yet applicable in the ac- 2017), and in most cases, the initial design domain has to be fixed
tual product development stage (Bujny et al., 2018; Oh et al., 2019; (Nie et al., 2021; Sharpe & Seepersad, 2019) and cannot be applied
Wang et al., 2021). In reality, the 3D design domain must be applied, to new boundary conditions or loading conditions (Bujny et al.,
especially in large-scale problem. Furthermore, the TO process is 2018; Halle et al., 2021; Hayashi & Ohsaki, 2020; Lew & Buehler,
more time-consuming, and the computational cost is expensive, 2021; Malviya, 2020; Rawat & Shen, 2018; Zheng et al., 2021). Fur-
so the application of MLTO is an urgent research topic. However, as thermore, additional studies on complex problems (Kallioras &
the 3D MLTO problem entails numerous design parameters, more Lagaros, 2021a; Li et al., 2019b, 2021; Lynch et al., 2019), such as
training data are required, which causes the problem of computa- multi-objective problems or various objective functions (Bujny et
tional cost attributable to data collection and simulation analysis. al., 2018; Chandrasekhar & Suresh, 2021a, c; Jiang & Fan, 2019a;
Therefore, MLTO can be used while supplementing this problem. Kumar & Suresh, 2020; Qiu et al., 2021b; Sun & Ma, 2020), robust
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1757

TO (Chi et al., 2021), non-linear TO (Halle et al., 2021; Qian & Ye, fore, more studies are needed to predict manufacturable opti-
2021), and MMTO (Kim et al., 2021; Li et al., 2019c; Napier et al., mization results considering complex manufacturing constraints
2020), are needed. (Karlsson et al., 2021; Wang et al., 2020; Zhang et al., 2019), such
as AM conditions (Almasri et al., 2020) and injection molding
(Greminger, 2020).
4.2.3. Industrial applicability
Scalability is a crucial factor directly related to the usability of
4.3.3. Robust learning performance
MLTO in industrial applications. From an industrial perspective,
an adaptive tool with excellent generalizability to solve various Whether the network used for MLTO has robust quality should
problems is urgently needed. Most MLTO methods require a large be verified and improved. This issue is also linked to the cherry-
amount of pre-optimized data, such as those in the studies pre- picking problem, showing only one-time good results frequently
sented in Section 2.2. However, collecting sufficient data, espe- occurring in ML/DL research. As some of the networks are unsta-
cially those used in the industry, takes time. Furthermore, the ble, additional research on the convergence of the network (Gill-
computational complexity of supervised MLTO limits its use in hofer et al., 2019; Li et al., 2021; Takahashi et al., 2019) is needed. Fur-
the industry, as explained in Section 4.1. Considering that these thermore, scholars have attempted to additionally apply various

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

studies were conducted on the basis of benchmark data, the com- DL models, such as state-of-the-art GAN models (Li et al., 2019a)
putational complexity of supervised MLTO for industrial appli- and cGAN that inputs the desired target volume or manufactur-
cations will be more complicated. Therefore, recent studies have ing process as a condition (Greminger, 2020), to improve further
suggested the use of adaptive MLTO that can reparameterize the the accuracy of the network (Jiang & Fan, 2019b; Kallioras & La-
density field by using a NN as shown in Section 2.5.2. This method garos, 2021a; Malviya, 2020; Wen et al., 2020). Other scholars have
can be generalized to various design domains and boundary con- focused on generative design to solve the mode collapse problem,
ditions and extended to 3D problems without the need to pre- which lacks diversity (Rawat & Shen, 2019b). Furthermore, studies
pare data in advance. Even though the problem of computational that use GAN for generative design lack an appropriate evaluation
cost still needs to be supplemented due to the FEA process, this method, and they need to consider the evaluation metric for the
method can be advantageous for the industry. generated designs (Rawat & Shen, 2019a, b).

4.3. Connectivity, feasibility, and robustness 5. Conclusions

4.3.1. Topology resolution and connectivity Various studies on data-based TO have been recently conducted,
Compared with the optimization result through the conventional and the aim is to effectively derive optimal topologies by apply-
iterative TO, the result of MLTO has a lower resolution, and a ing various ML techniques to supplement the problems of con-
separate refinement process is often required. In particular, be- ventional methodologies. Each study was conducted using various
cause physics information is not properly reflected and the opti- ML methods, such as supervised learning, unsupervised learning,
mization results are predicted with the similarity between pixels, and RL, for the purpose of acceleration of iteration, non-iterative
connectivity may not be guaranteed, resulting in TO results that optimization, meta-modeling, dimensionality reduction of design
the user cannot trust and cannot be manufactured (Banga et al., space, improvement of the optimizer, generative design, and post-
2018; Sharpe & Seepersad, 2019; Zhang et al., 2019). In addition, processing to TO field. These approaches allowed the studies to
the checkerboard pattern phenomenon of the TO result is a sig- speed up the optimization process and reduce the computational
nificant problem to be solved (Chandrasekhar & Suresh, 2021a), cost.
and some studies have shown that MLTO’s results do not produce Although various studies in this field have been conducted,
detailed features or have blurry results because detailed features many studies still have problems, such as the scalability of ap-
cannot be learned during the training process (Gaymann & Mon- plicability to various design domains (e.g., 3D domain), the cost
tomoli, 2019; Guo et al., 2018; Halle et al., 2021; Keshavarzzadeh et burden of data collection, low resolution of the resulting topology,
al., 2021; Li et al., 2019a). the infeasibility of manufacturing, weak performance of the ML
method, and expensive FEA and computational cost. We expect
4.3.2. Manufacturing feasibility effective MLTO studies applicable to 3D TO to be conducted in the
future to obtain optimal results in a manner that can explain en-
For the optimized result to be used in actual product develop-
gineering phenomena based on physics information.
ment, it is necessary to examine the manufacturing possibility.
Given that AM processing is often impossible for many TO results,
it is important to consider the manufacturing constraints in the
Acknowledgments
optimization process. Recent TO studies that consider manufac-
turability are being conducted. For instance, Cheng et al. (2019) This work was supported by theNational Research Foundation of
proposed a functional gradient lattice structure TO for AM. Liu Korea (grant no. 2018R1A5A7025409) and the Ministry of Science
et al. (2021) proposed a stress-constrained TO method for fused and ICT of Korea (grant numbers 2022–0-00969 and 2022–0-00986).
deposition modeling in AM. Liu et al. (2022) constructed a simul- A previous version of this manuscript was presented at the 2021
taneous product design workflow while considering the materials World Congress on Advances in Structural Engineering and Me-
and the manufacturing process for AM. Kim et al. (2022) imple- chanics (ASEM21; Seoul, Korea, August 23–26, 2021) (Shin et al.,
mented computational homogenization to increase the compu- 2021).
tational efficiency and reliability of FEA results in microscale and
multiscale TOs. Wu and Xiao (2022) incorporated self-supporting
factors to generate manufacturable designs. However, manufac-
Conflict of interest statement
turing constraints are rarely considered in the MLTO field. There- Authors declare no conflict of interest associated with this
manuscript.
1758 | Topology optimization via artificial intelligence: a review

References Bendsøe, M. P., & Sigmund, O. (1999). Material interpolation schemes

in topology optimization. Archive of Applied Mechanics, 69, 635–654.
Aage, N., Andreassen, E., & Lazarov, B. S. (2015). Topology optimization https://doi.org/10.1007/s004190050248.
using PETSc: An easy-to-use, fully parallel, open source topology Bendsoe, M. P., & Sigmund, O. (2003). Topology optimization: Theory,
optimization framework. Structural and Multidisciplinary Optimiza- methods, and applications. Springer Science & Business Media.
tion, 51, 565–572. https://doi.org/10.1007/s00158- 014- 1157- 0. Bi, S., Zhang, J., & Zhang, G. (2020). Scalable deep-learning-
Abueidda, D. W., Koric, S., & Sobh, N. A. (2020). Topology optimization accelerated topology optimization for additively manufactured
of 2D structures with nonlinearities using deep learning. Comput- materials. arXiv preprint arXiv:2011.14177.
ers & Structures, 237, 106283. https://doi.org/10.1016/j.compstruc. Blanchard-Dionne, A. P., & Martin, O. J. (2021). Successive training of a
2020.106283. generative adversarial network for the design of an optical cloak.
Albawi, S., Mohammed, T. A., & Al-Zawi, S. (2017). Understanding of Osa Continuum, 4(1), 87–95. https://doi.org/10.1364/OSAC.413394.
a convolutional neural network. In Proceedings of the 2017 Interna- Borrvall, T., & Petersson, J. (2001). Large-scale topology optimization
tional Conference on Engineering and Technology (ICET)(pp. 1–6). IEEE. in 3D using parallel computing. Computer Methods in Applied Me-
https://doi.org/10.1109/ICEngTechnol.2017.8308186. chanics and Engineering, 190(46-47), 6201–6229. https://doi.org/10
Allaire, G., Jouve, F., & Toader, A. M. (2002). A level-set method for .1016/S0045- 7825(01)00216- X.

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

shape optimization. Comptes Rendus Mathematique, 334(12), 1125– Bourdin, B. (2001). Filters in topology optimization. International Jour-
1130. https://doi.org/10.1016/S1631- 073X(02)02412- 3. nal for Numerical Methods in Engineering, 50(9), 2143–2158. https:
Almasri, W., Bettebghor, D., Ababsa, F., & Danglade, F. (2020). //doi.org/10.1002/nme.116.
Shape related constraints aware generation of mechanical Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32. https:
designs through deep convolutional GAN. arXiv preprint //doi.org/10.1023/A:1010933404324.
arXiv:2010.11833. Brown, N. K., Garland, A. P., Fadel, G. M., & Li, G. (2022). Deep reinforce-
Amir, O., & Sigmund, O. (2011). On reducing computational effort in ment learning for engineering design through topology optimiza-
topology optimization: How far can we go?. Structural and Multi- tion of elementally discretized design domains. Materials & Design,
disciplinary Optimization, 44, 25–29. https://doi.org/10.1007/s00158 218, 110672. https://doi.org/10.1016/j.matdes.2022.110672.
- 010- 0586- 7. Bruns, T. E., & Tortorelli, D. A. (2001). Topology optimization of non-
Amir, O., Bendsøe, M. P., & Sigmund, O. (2009). Approximate reanalysis linear elastic structures and compliant mechanisms. Computer
in topology optimization. International Journal for Numerical Methods Methods in Applied Mechanics and Engineering, 190(26-27), 3443–
in Engineering, 78(12), 1474–1491. https://doi.org/10.1002/nme.25 3459. https://doi.org/10.1016/S0045- 7825(00)00278- 4.
36. Bujny, M., Aulig, N., Olhofer, M., & Duddeck, F. (2018). Learning-based
Amir, O., Stolpe, M., & Sigmund, O. (2010). Efficient use of iterative topology variation in evolutionary level set topology optimiza-
solvers in nested topology optimization. Structural and Multidisci- tion. In Proceedings of the Genetic and Evolutionary Computation Con-
plinary Optimization, 42, 55–72. https://doi.org/10.1007/s00158-009 ference(pp. 825–832). https://doi.org/10.1145/3205455.3205528.
- 0463- 4. Burges, C. J. (1998). A tutorial on support vector machines for pattern
Andreassen, E., Clausen, A., Schevenels, M., Lazarov, B. S., & Sigmund, recognition. Data Mining and Knowledge Discovery, 2(2), 121–167. ht
O. (2011). Efficient topology optimization in MATLAB using 88 tps://doi.org/10.1023/A:1009715923555.
lines of code. Structural and Multidisciplinary Optimization, 43, 1–16. Cang, R., Yao, H., & Ren, Y. (2019). One-shot generation of
https://doi.org/10.1007/s00158- 010- 0594- 7. near-optimal topology through theory-driven machine learning.
Anitescu, C., Atroshchenko, E., Alajlan, N., & Rabczuk, T. (2019). Ar- Computer-Aided Design, 109, 12–21. https://doi.org/10.1016/j.cad.
tificial neural network methods for the solution of second order 2018.12.008.
boundary value problems. Computers, Materials & Continua, 59(1), Chandrasekhar, A., & Suresh, K. (2021a). TOuNN: Topology optimiza-
345–359. https://doi.org/10.32604/cmc.2019.06641. tion using neural networks. Structural and Multidisciplinary Opti-
Asanuma, J., Doi, S., & Igarashi, H. (2020). Transfer learning through mization, 63, 1135–1149. https://doi.org/10.1007/s00158- 020- 027
deep learning: Application to topology optimization of electric 48-4.
motor. IEEE Transactions on Magnetics, 56(3), 1–4. https://doi.org/10 Chandrasekhar, A., & Suresh, K. (2021b). Multi-material topology
.1109/TMAG.2019.2956849. optimization using neural networks. Computer-Aided Design, 136,
Aulig, N., & Olhofer, M. (2014). Topology optimization by predicting 103017. https://doi.org/10.1016/j.cad.2021.103017.
sensitivities based on local state features. In Proceedings of the 5th Chandrasekhar, A., & Suresh, K. (2021c). Length scale control in
European Conference on Computational Mechanics (ECCM V)(pp. 3578– topology optimization using Fourier enhanced neural networks.
3589). arXiv preprint arXiv:2109.01861.
Banga, S., Gehani, H., Bhilare, S., Patel, S., & Kara, L. (2018). 3D topology Cheng, L., Bai, J., & To, A. C. (2019). Functionally graded lattice struc-
optimization using convolutional neural networks. arXiv preprint ture topology optimization for the design of additive manufac-
arXiv:1808.07440. tured components with stress constraints. Computer Methods in
Bellman, R. (1957). A Markovian decision process. Indiana University Applied Mechanics and Engineering, 344, 334–359. https://doi.org/10
Mathematics Journal, 6, 679–684. https://doi.org/10.1512/iumj.195 .1016/j.cma.2018.10.010.
7.6.56038. Chi, H., Zhang, Y., Tang, T. L. E., Mirabella, L., Dalloro, L., Song, L., &
Bendsøe, M. P. (1989). Optimal shape design as a material distribution Paulino, G. H. (2021). Universal machine learning for topology op-
problem. Structural Optimization, 1, 193–202. https://doi.org/10.100 timization. Computer Methods in Applied Mechanics and Engineering,
7/BF01650949. 375, 112739. https://doi.org/10.1016/j.cma.2019.112739.
Bendsøe, M. P., & Kikuchi, N. (1988). Generating optimal topologies Chu, S., Xiao, M., Gao, L., Gui, L., & Li, H. (2016). An effective structural
in structural design using a homogenization method. Computer boundary processing method based on support vector machine
Methods in Applied Mechanics and Engineering, 71(2), 197–224. https: for discrete topology optimization. In Proceedings of the 2016 20th
//doi.org/10.1016/0045- 7825(88)90086- 2. International Conference on Computer Supported Cooperative Work in
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1759

Design (CSCWD)(pp. 10–15). IEEE. https://doi.org/10.1109/CSCWD. Ha, H., Agrawal, S., & Song, S. (2021). Fit2Form: 3D generative model
2016.7565954. for robot gripper form design. In Proceedings of the Conference on
Cover, T., & Hart, P. (1967). Nearest neighbor pattern classification. Robot Learning(pp. 176–187). PMLR.
IEEE Transactions on Information Theory, 13(1), 21–27. https://doi.or Halle, A., Campanile, L. F., & Hasse, A. (2021). An artificial
g/10.1109/TIT.1967.1053964. intelligence–Assisted design method for topology optimization
Deng, H., & To, A. C. (2020). Topology optimization based on deep rep- without pre-optimized training data. Applied Sciences, 11(19), 9041.
resentation learning (DRL) for compliance and stress-constrained https://doi.org/10.3390/app11199041.
design. Computational Mechanics, 66(2), 449–469. https://doi.org/10 Hayashi, K., & Ohsaki, M. (2020). Reinforcement learning and
.1007/s00466- 020- 01859- 5. graph embedding for binary truss topology optimization
Deng, C., Wang, Y., Qin, C., Fu, Y., & Lu, W. (2022). Self-directed online under stress and displacement constraints. Frontiers in
machine learning for topology optimization. Nature Communica- Built Environment, 6, 59. https://doi.org/10.3389/fbuil.2020.00059.
tions, 13(1), 388. https://doi.org/10.1038/s41467- 021- 27713- 7. Hoyer, S., Sohl-Dickstein, J., & Greydanus, S. (2019). Neural repa-
Doi, S., Sasaki, H., & Igarashi, H. (2019). Multi-objective topology op- rameterization improves structural optimization. arXiv preprint
timization of rotating machines using deep learning. IEEE Trans- arXiv:1909.04240.
actions on Magnetics, 55(6), 1–5. https://doi.org/10.1109/TMAG.201 Huang, X., & Xie, Y. M. (2007). Convergent and mesh-independent

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

9.2899934. solutions for the bi-directional evolutionary structural optimiza-
Ferrari, F., & Sigmund, O. (2020). Towards solving large-scale topol- tion method. Finite Elements in Analysis and Design, 43(14), 1039–
ogy optimization problems with buckling constraints at the cost 1049. https://doi.org/10.1016/j.finel.2007.06.006.
of linear analyses. Computer Methods in Applied Mechanics and Engi- Jang, S., Yoo, S., & Kang, N. (2022). Generative design by reinforce-
neering, 363, 112911. https://doi.org/10.1016/j.cma.2020.112911. ment learning: Enhancing the diversity of topology optimization
Gatys, L. A., Ecker, A. S., & Bethge, M. (2015). A neural algorithm of designs. Computer-Aided Design, 146, 103225. https://doi.org/10.1
artistic style. Journal of Vision, 16(12), 326. https://doi.org/10.1167/ 016/j.cad.2022.103225.
16.12.326. Jensen, P. D. L., Wang, F., Dimino, I., & Sigmund, O. (2021). Topology op-
Gaymann, A., & Montomoli, F. (2019). Deep neural network and timization of large-scale 3D morphing wing structures. Actuators,
Monte Carlo tree search applied to fluid-structure topology op- 10(9), 217. https://doi.org/10.3390/act10090217.
timization. Scientific Reports, 9(1), 1–16. https://doi.org/10.1038/s4 Jiang, J., & Fan, J. A. (2019a). Dataless training of generative models
1598- 019- 51111- 1. for the inverse design of metasurfaces. 401, 402. arXiv preprint
Gillhofer, M., Ramsauer, H., Brandstetter, J., Schäfl, B., & Hochreiter, arXiv:1906.07843.
S. (2019). A GAN based solver of black-box inverse problems. In Jiang, J., & Fan, J. A. (2019b). Global optimization of dielectric metasur-
Proceedings of the NeurIPS 2019 Workshop on Solving Inverse Problems faces using a physics-driven neural network. Nano Letters, 19(8),
with Deep Networks. 5366–5372. https://doi.org/10.1021/acs.nanolett.9b01857.
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT Jiang, J., Sell, D., Hoyer, S., Hickey, J., Yang, J., & Fan, J. A. (2019). Free-
Press. form diffractive metagrating design based on generative adver-
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., sarial networks. ACS Nano, 13(8), 8872–8878. https://doi.org/10.1
Ozair, S., Courville, A., & Bengio, Y. (2020). Generative adversarial 021/acsnano.9b02371.
networks. Communications of the ACM, 63(11), 139–144. https://do Jiang, X., Wang, H., Li, Y., & Mo, K. (2020). Machine learning based
i.org/10.1145/3422622. parameter tuning strategy for MMC based topology optimization.
Greminger, M. (2020). Generative adversarial networks with syn- Advances in Engineering Software, 149, 102841. https://doi.org/10.1
thetic training data for enforcing manufacturing constraints on 016/j.advengsoft.2020.102841.
topology optimization. In Proceedings of the ASME 2020 International Jolliffe, I. T. (2002). Principal component analysis for special types of
Design Engineering Technical Conferences and Computers and Informa- data(pp. 338–372). Springer.
tion in Engineering Conference(Vol. 84003, p. V11AT11A005). Amer- Kallioras, N. A., & Lagaros, N. D. (2020). DzAIN: Deep learning based
ican Society of Mechanical Engineers. https://doi.org/10.1115/DE generative design. Procedia Manufacturing, 44, 591–598. https://do
TC2020-22399. i.org/10.1016/j.promfg.2020.02.251.
Guest, J. K., Prévost, J. H., & Belytschko, T. (2004). Achieving minimum Kallioras, N. A., & Lagaros, N. D. (2021a). DL-SCALE: A novel deep
length scale in topology optimization using nodal design vari- learning-based model order upscaling scheme for solving topol-
ables and projection functions. International Journal for Numerical ogy optimization problems. Neural Computing and Applications, 33,
Methods in Engineering, 61(2), 238–254. https://doi.org/10.1002/nm 7125–7144. https://doi.org/10.1007/s00521- 020- 05480- 8.
e.1064. Kallioras, N. A., & Lagaros, N. D. (2021b). MLGen: Generative design
Guo, T., Lohan, D. J., Cang, R., Ren, M. Y., & Allison, J. T. (2018). An indi- framework based on machine learning and topology optimiza-
rect design representation for topology optimization using vari- tion. Applied Sciences, 11(24), 12044. https://doi.org/10.3390/app1
ational autoencoder and style transfer. In Proceedings of the 2018 12412044.
AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Kallioras, N. A., Kazakis, G., & Lagaros, N. D. (2020). Accelerated topol-
Conference. https://doi.org/10.2514/6.2018-0804. ogy optimization by means of deep learning. Structural and Multi-
Guo, X., Zhang, W., & Zhong, W. (2014). Doing topology optimization disciplinary Optimization, 62(3), 1185–1212. https://doi.org/10.1007/
explicitly and geometrically—A new moving morphable compo- s00158- 020- 02545- z.
nents based framework. Journal of Applied Mechanics, 81(8), 081009. Karlsson, P., Pejryd, L., & Strömberg, N. (2021). Generative design op-
https://doi.org/10.1115/1.4027609. timization and characterization of triple periodic lattice struc-
Guo, H., Zhuang, X., & Rabczuk, T. (2021). A deep collocation tures in AlSi10Mg. In Industrializing Additive Manufacturing: Pro-
method for the bending analysis of Kirchhoff plate. arXiv preprint ceedings of AMPA2020(pp. 3–16). Springer International Publishing.
arXiv:2102.02617. https://doi.org/10.1007/978- 3- 030- 54334- 1_1.
1760 | Topology optimization via artificial intelligence: a review

Keshavarzzadeh, V., Alirezaei, M., Tasdizen, T., & Kirby, R. M. (2021). Li, H., Kafka, O. L., Gao, J., Yu, C., Nie, Y., Zhang, L., Tajdari, M., Tang,
Image-based multiresolution topology optimization using deep S., Guo, X., Li, G., Tang, S., Cheng, G., & Liu, W. K. (2019b). Clus-
disjunctive normal shape model. Computer-Aided Design, 130, tering discretization methods for generation of material perfor-
102947. https://doi.org/10.1016/j.cad.2020.102947. mance databases in machine learning and design optimization.
Keshavarzzadeh, V., Kirby, R. M., & Narayan, A. (2020). Stress-based Computational Mechanics, 64, 281–305. https://doi.org/10.1007/s004
topology optimization under uncertainty via simulation-based 66- 019- 01716- 0.
Gaussian process. Computer Methods in Applied Mechanics and En- Liang, Q. Q., Xie, Y. M., & Steven, G. P. (2000). Optimal topology se-
gineering, 365, 112992. https://doi.org/10.1016/j.cma.2020.112992. lection of continuum structures with displacement constraints.
Kim, J. E., Cho, N. K., & Park, K. (2022). Computational homoge- Computers & Structures, 77(6), 635–644. https://doi.org/10.1016/S0
nization of additively manufactured lightweight structures with 045- 7949(00)00018- 3.
multiscale topology optimization. Journal of Computational Design Limkilde, A., Evgrafov, A., & Gravesen, J. (2018). On reducing compu-
and Engineering, 9(5), 1602–1615. https://doi.org/10.1093/jcde/qwa tational effort in topology optimization: We can go at least this
c078. far!. Structural and Multidisciplinary Optimization, 58, 2481–2492.
Kim, C., Lee, J., & Yoo, J. (2021). Machine learning-combined topology https://doi.org/10.1007/s00158- 018- 2121- 1.
optimization for functionary graded composite structure design. Lin, Q., Hong, J., Liu, Z., Li, B., & Wang, J. (2018). Investigation into

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

Computer Methods in Applied Mechanics and Engineering, 387, 114158. the topology optimization for conductive heat transfer based on
https://doi.org/10.1016/j.cma.2021.114158. deep learning approach. International Communications in Heat and
Kingma, D. P., & Welling, M. (2013). Auto-encoding variational bayes. Mass Transfer, 97, 103–109. https://doi.org/10.1016/j.icheatmasstr
arXiv preprint arXiv:1312.6114. ansfer.2018.07.001.
Kingma, D. P., & Welling, M. (2014). Stochastic gradient VB and the Liu, K., & Tovar, A. (2014). An efficient 3D topology optimization code
variational auto-encoder. In Proceedings of the Second International written in Matlab. Structural and Multidisciplinary Optimization, 50,
Conference on Learning Representations, ICLR(Vol. 19, p. 121). 1175–1196. https://doi.org/10.1007/s00158- 014- 1107- x.
Kollmann, H. T., Abueidda, D. W., Koric, S., Guleryuz, E., & Sobh, N. A. Liu, K., Tovar, A., Nutwell, E., & Detwiler, D. (2015). Towards non-
(2020). Deep learning for topology optimization of 2D metamate- linear multimaterial topology optimization using unsupervised
rials. Materials & Design, 196, 109098. https://doi.org/10.1016/j.ma machine learning and metamodel-based optimization. In Proceed-
tdes.2020.109098. ings of the International Design Engineering Technical Conferences and
Kudyshev, Z. A., Kildishev, A. V., Shalaev, V. M., & Boltasseva, A. (2021). Computers and Information in Engineering Conference(Vol. 57083, p.
Machine learning–assisted global optimization of photonic de- V02BT03A004). American Society of Mechanical Engineers. https:
vices. Nanophotonics, 10(1), 371–383. https://doi.org/10.1515/nano //doi.org/10.1115/DETC2015-46534.
ph- 2020- 0376. Liu, W., Wang, Z., Liu, X., Zeng, N., Liu, Y., & Alsaadi, F. E. (2017). A sur-
Kumar, T., & Suresh, K. (2020). A density-and-strain-based K- vey of deep neural network architectures and their applications.
clustering approach to microstructural topology optimization. Neurocomputing, 234, 11–26. https://doi.org/10.1016/j.neucom.201
Structural and Multidisciplinary Optimization, 61, 1399–1415. https: 6.12.038.
//doi.org/10.1007/s00158- 019- 02422- 4. Liu, G., Xiong, Y., & Rosen, D. W. (2022). Multidisciplinary design op-
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, timization in design for additive manufacturing. Journal of Com-
521(7553), 436–444. https://doi.org/10.1038/nature14539. putational Design and Engineering, 9(1), 128–143. https://doi.org/10
Lee, S., Kim, H., Lieu, Q. X., & Lee, J. (2020). CNN-based image recog- .1093/jcde/qwab073.
nition for topology optimization. Knowledge-Based Systems, 198, Liu, J., Yan, J., & Yu, H. (2021). Stress-constrained topology optimiza-
105887. https://doi.org/10.1016/j.knosys.2020.105887. tion for material extrusion polymer additive manufacturing. Jour-
Lei, X., Liu, C., Du, Z., Zhang, W., & Guo, X. (2019). Machine learning- nal of Computational Design and Engineering, 8(3), 979–993. https:
driven real-time topology optimization under moving morphable //doi.org/10.1093/jcde/qwab028.
component-based framework. Journal of Applied Mechanics, 86(1), Lynch, M. E., Sarkar, S., & Maute, K. (2019). Machine learning to aid
011004. https://doi.org/10.1115/1.4041319. tuning of numerical parameters in topology optimization. Journal
Lew, A. J., & Buehler, M. J. (2021). Encoding and exploring latent de- of Mechanical Design, 141(11), 114502. https://doi.org/10.1115/1.40
sign space of optimal material structures via a VAE-LSTM model. 44228.
Forces in Mechanics, 5, 100054. https://doi.org/10.1016/j.finmec.20 McDermott, M., Yan, T., Naumann, T., Hunt, N., Suresh, H., Szolovits,
21.100054. P., & Ghassemi, M. (2018). Semi-supervised biomedical translation
Li, M., Cheng, Z., Jia, G., & Shi, Z. (2019c). Dimension reduction and with cycle Wasserstein regression GANs. In Proceedings of the AAAI
surrogate based topology optimization of periodic structures. Conference on Artificial Intelligence(Vol. 32). https://doi.org/10.1609/
Composite Structures, 229, 111385 https://doi.org/10.1016/j.compst aaai.v32i1.11890.
ruct.2019.111385. MacQueen, J. (1967). Classification and analysis of multivariate obser-
Li, J., Gao, L., Ye, M., Li, H., & Li, L. (2022). Topology optimization of ir- vations. In Proceedings of the 5th Berkeley Symposium on Mathematical
regular flow domain by parametric level set method in unstruc- Statistics and Probability(pp. 281–297). University of California.
tured mesh. Journal of Computational Design and Engineering, 9(1), Malviya, M. (2020). A systematic study of deep generative models for
100–113. https://doi.org/10.1093/jcde/qwab071. rapid topology optimization. https://doi.org/10.31224/osf.io/9gvq
Li, B., Huang, C., Li, X., Zheng, S., & Hong, J. (2019a). Non-iterative s.
structural topology optimization using deep learning. Computer- Martínez-Frutos, J., Martínez-Castejón, P. J., & Herrero-Pérez,
Aided Design, 115, 172–180. https://doi.org/10.1016/j.cad.2019.05. D. (2017). Efficient topology optimization using GPU com-
038. puting with multilevel granularity. Advances in Engineering
Li, M., Jia, G., Cheng, Z., & Shi, Z. (2021). Generative adversarial net- Software, 106, 47–62. https://doi.org/10.1016/j.advengsoft.2017.
work guided topology optimization of periodic structures via sub- 01.009.
set simulation. Composite Structures, 260, 113254. https://doi.org/ Mirza, M., & Osindero, S. (2014). Conditional generative adversarial
10.1016/j.compstruct.2020.113254. nets. arXiv preprint arXiv:1411.1784.
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1761

Mlejnek, H. P. (1992). Some aspects of the genesis of structures. Struc- ential equations. Journal of Computational Physics, 378, 686–707.
tural Optimization, 5, 64–69. https://doi.org/10.1007/BF01744697. https://doi.org/10.1016/j.jcp.2018.10.045.
Mnih, V., Badia, A. P., Mirza, M., Graves, A., Lillicrap, T., Harley, T., Silver, Rawat, S., & Shen, M. H. (2018). A novel topology design approach
D., & Kavukcuoglu, K. (2016). Asynchronous methods for deep re- using an integrated deep learning network architecture. arXiv
inforcement learning. In Proceedings of the International Conference preprint arXiv:1808.02334.
on Machine Learning(pp. 1928–1937). PMLR. Rawat, S., & Shen, M. H. (2019a). Application of adversarial networks
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wier- for 3D structural topology optimization. SAE Technical Paper (No.
stra, D., & Riedmiller, M. (2013). Playing Atari with deep reinforce- 2019-01-0829). SAE International. https://doi.org/10.4271/2019-0
ment learning. arXiv preprint arXiv:1312.5602. 1-0829.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, Rawat, S., & Shen, M. H. H. (2019b). A novel topology optimiza-
M. G., Graves, A., Riedmiller, M., Fidjeland, A. K., Ostrovski, G., Pe- tion approach using conditional deep learning. arXiv preprint
tersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumaran, arXiv:1901.04859.
D., Wierstra, D., Legg, S., & Hassabis, D. (2015). Human-level con- Ringnér, M. (2008). What is principal component analysis?. Nature
trol through deep reinforcement learning. Nature, 518(7540), 529– Biotechnology, 26(3), 303–304. https://doi.org/10.1038/nbt0308-30
533. https://doi.org/10.1038/nature14236. 3.

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

Murthy, S. K. (1998). Automatic construction of decision trees from Rozvany, G. I., Zhou, M., & Birker, T. (1992). Generalized shape opti-
data: A multi-disciplinary survey. Data Mining and Knowledge Dis- mization without homogenization. Structural optimization, 4(3-4),
covery, 2, 345–389. https://doi.org/10.1023/A:1009744630224. 250–252. https://doi.org/10.1007/BF01742754.
Napier, N., Sriraman, S. A., Tran, H. T., & James, K. A. (2020). An ar- Samaniego, E., Anitescu, C., Goswami, S., Nguyen-Thanh, V. M., Guo,
tificial neural network approach for generating high-resolution H., Hamdia, K., Zhuang, X., & Rabczuk, T. (2020). An energy ap-
designs from low-resolution input in topology optimization. Jour- proach to the solution of partial differential equations in compu-
nal of Mechanical Design, 142(1), 011402. https://doi.org/10.1115/1. tational mechanics via machine learning: Concepts, implemen-
4044332. tation and applications. Computer Methods in Applied Mechanics
Nie, Z., Lin, T., Jiang, H., & Kara, L. B. (2021). TopologyGAN: Topol- and Engineering, 362, 112790. https://doi.org/10.1016/j.cma.2019.1
ogy optimization using generative adversarial networks based on 12790.
physical fields over the initial domain. Journal of Mechanical Design, Sasaki, H., & Igarashi, H. (2019a). Topology optimization accelerated
143(3), 031715. https://doi.org/10.1115/1.4049533. by deep learning. IEEE Transactions on Magnetics, 55(6), 1–5. https:
Oh, S., Jung, Y., Kim, S., Lee, I., & Kang, N. (2019). Deep generative de- //doi.org/10.1109/TMAG.2019.2901906.
sign: Integration of topology optimization and generative models. Sasaki, H., & Igarashi, H. (2019b). Topology optimization of IPM mo-
Journal of Mechanical Design, 141(11), 111405. https://doi.org/10.111 tor with aid of deep learning. International Journal of Applied Electro-
5/1.4044229. magnetics and Mechanics, 59(1), 87–96. https://doi.org/10.3233/JAE-
Oh, S., Jung, Y., Lee, I., & Kang, N. (2018). Design automation by inte- 171164.
grating generative adversarial networks and topology optimiza- Schulman, J., Levine, S., Abbeel, P., Jordan, M., & Moritz, P. (2015). Trust
tion. In Proceedings of the International Design Engineering Techni- region policy optimization. In Proceedings of the International Con-
cal Conferences and Computers and Information in Engineering Confer- ference on Machine Learning(pp. 1889–1897). PMLR.
ence(Vol. 51753, p. V02AT03A008). American Society of Mechani- Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O.
cal Engineers. (2017). Proximal policy optimization algorithms. arXiv preprint
Park, S. M., Park, S., Park, J., Choi, M., Kim, L., & Noh, G. (2021). Design arXiv:1707.06347.
process of patient-specific osteosynthesis plates using topology Sethian, J. A., & Wiegmann, A. (2000). Structural boundary design
optimization. Journal of Computational Design and Engineering, 8(5), via level set and immersed interface methods. Journal of Compu-
1257–1266. https://doi.org/10.1093/jcde/qwab047. tational Physics, 163(2), 489–528. https://doi.org/10.1006/jcph.200
Patel, J., & Choi, S. K. (2012). Classification approach for reliability- 0.6581.
based topology optimization using probabilistic neural networks. Settles, B. (2012). Active learning. Springer. https://doi.org/10.1007/97
Structural and Multidisciplinary Optimization, 45(4), 529–543. https: 8- 3- 031- 01560- 1.
//doi.org/10.1007/s00158- 011- 0711- 2. Sharpe, C., & Seepersad, C. C. (2019). Topology design with
Qian, C., & Ye, W. (2021). Accelerating gradient-based topology opti- conditional generative adversarial networks. In Proceedings
mization design with dual-model artificial neural networks. Struc- of the International Design Engineering Technical Conferences
tural and Multidisciplinary Optimization, 63, 1687–1707. https://doi. and Computers and Information in Engineering Conference(Vol.
org/10.1007/s00158- 020- 02770- 6. 59186, p. V02AT03A062). American Society of Mechanical
Qiu, C., Du, S., & Yang, J. (2021a). A deep learning approach for effi- Engineers.
cient topology optimization based on the element removal strat- Shen, M. H. H., & Chen, L. (2019). A new CGAN technique
egy. Materials & Design, 212, 110179. https://doi.org/10.1016/j.ma for constrained topology design optimization. arXiv preprint
tdes.2021.110179. arXiv:1901.07675.
Qiu, Z., Li, Q., Liu, S., & Xu, R. (2021b). Clustering-based concurrent Shin, S., Shin, D., Kim, M., Ryu, H., & Kang, N. (2021). Machine learning-
topology optimization with macrostructure, components, and based topology optimization: A review. In Proceedings of the 2021
materials. Structural and Multidisciplinary Optimization, 63, 1243– World Congress on Advances in Structural Engineering and Mechanics
1263. https://doi.org/10.1007/s00158- 020- 02755- 5. (ASEM21). IASEM, KAIST, KTA, SNU DAAE.
Rade, J., Balu, A., Herron, E., Pathak, J., Ranade, R., Sarkar, S., & Krish- Sigmund, O. (1997). On the design of compliant mechanisms using
namurthy, A. (2020). Physics-consistent deep learning for struc- topology optimization. Journal of Structural Mechanics, 25(4), 493–
tural topology optimization. arXiv preprint arXiv:2012.05359. 524. https://doi.org/10.1080/08905459708945415.
Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics-informed Sigmund, O. (2001a). A 99 line topology optimization code written in
neural networks: A deep learning framework for solving for- Matlab. Structural and Multidisciplinary Optimization, 21(2), 120–127.
ward and inverse problems involving nonlinear partial differ- https://doi.org/10.1007/s001580050176.
1762 | Topology optimization via artificial intelligence: a review

Sigmund, O. (2001b). Design of multiphysics actuators using topology Watkins, C. J., & Dayan, P. (1992). Q-learning. Machine Learning, 8, 279–
optimization–Part II: Two-material structures. Computer Methods 292. https://doi.org/10.1007/BF00992698.
in Applied Mechanics and Engineering, 190(49-50), 6605–6627. https: Wen, F., Jiang, J., & Fan, J. A. (2019). Progressive-growing of generative
//doi.org/10.1016/S0045- 7825(01)00252- 3. adversarial networks for metasurface optimization. arXiv preprint
Sigmund, O. (2007). Morphology-based black and white filters for arXiv:1911.13029.
topology optimization. Structural and Multidisciplinary Optimization, Wen, F., Jiang, J., & Fan, J. A. (2020). Robust freeform metasurface
33(4-5), 401–424. https://doi.org/10.1007/s00158- 006- 0087- x. design based on progressively growing generative networks. ACS
Sigmund, O., & Torquato, S. (1997). Design of materials with extreme Photonics, 7(8), 2098–2104. https://doi.org/10.1021/acsphotonics.0
thermal expansion using a three-phase topology optimization c00539.
method. Journal of the Mechanics and Physics of Solids, 45(6), 1037– White, D. A., Arrighi, W. J., Kudo, J., & Watts, S. E. (2019). Multiscale
1067. https://doi.org/10.1016/S0022- 5096(96)00114- 7. topology optimization using neural network surrogate models.
Sim, E. A., Lee, S., Oh, J., & Lee, J. (2021). GANs and DCGANs for gener- Computer Methods in Applied Mechanics and Engineering, 346, 1118–
ation of topology optimization validation curve through cluster- 1135. https://doi.org/10.1016/j.cma.2018.09.007.
ing analysis. Advances in Engineering Software, 152, 102957. https: Wu, Z., Wang, S., Xiao, R., & Yu, L. (2020). A local solution approach for
//doi.org/10.1016/j.advengsoft.2020.102957. level-set based structural topology optimization in isogeometric

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

Sosnovik, I., & Oseledets, I. (2019). Neural networks for topology op- analysis. Journal of Computational Design and Engineering, 7(4), 514–
timization. Russian Journal of Numerical Analysis and Mathemati- 526. https://doi.org/10.1093/jcde/qwaa001.
cal Modelling, 34(4), 215–223. https://doi.org/10.1515/rnam-2019- Wu, Z., & Xiao, R. (2022). A topology optimization approach to struc-
0018. ture design with self-supporting constraints in additive manu-
Strömberg, N. (2019). A generative design optimization approach for facturing. Journal of Computational Design and Engineering, 9(2), 364–
additive manufacturing. In Proceedings of the Sim-AM 2019: II In- 379. https://doi.org/10.1093/jcde/qwac004.
ternational Conference on Simulation for Additive Manufacturing(pp. Xia, M., Yang, S., & Ho, S. L. (2017). A new topology optimization
130–141). CIMNE. methodology based on constraint maximum-weight connected
Strömberg, N. (2020). Efficient detailed design optimization of topol- graph theorem. IEEE Transactions on Magnetics, 54(3), 1–4. https:
ogy optimization concepts by using support vector machines and //doi.org/10.1109/TMAG.2017.2757001.
metamodels. Engineering Optimization, 52(7), 1136–1148. https://do Xie, Y. M., & Steven, G. P. (1993). A simple evolutionary procedure for
i.org/10.1080/0305215X.2019.1646258. structural optimization. Computers & Structures, 49(5), 885–896. ht
Sun, H., & Ma, L. (2020). Generative design by using exploration ap- tps://doi.org/10.1016/0045- 7949(93)90035- C.
proaches of reinforcement learning in density-based structural Xie, Y. M., & Steven, G. P. (1997). Basic evolutionary structural
topology optimization. Designs, 4(2), 10. https://doi.org/10.3390/ optimization. In Evolutionary structural optimization(pp. 12–29).
designs4020010. Springer. https://doi.org/10.1007/978- 1- 4471- 0985- 3_2.
Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduc- Xue, L., Liu, J., Wen, G., & Wang, H. (2021). Efficient, high-resolution
tion. MIT Press. topology optimization method based on convolutional neural
Takahashi, Y., Suzuki, Y., & Todoroki, A. (2019). Convolutional neu- networks. Frontiers of Mechanical Engineering, 16(1), 80–96. https:
ral network-based topology optimization (CNN-TO) by estimat- //doi.org/10.1007/s11465- 020- 0614- 2.
ing sensitivity of compliance from material distribution. arXiv Yago, D., Cante, J., Lloberas-Valls, O., & Oliver, J. (2022). Topology op-
preprint arXiv:2001.00635. timization methods for 3D structural problems: A comparative
Ulu, E., Zhang, R., & Kara, L. B. (2016). A data-driven investigation and study. Archives of Computational Methods in Engineering, 29(3), 1525–
estimation of optimal topologies under variable loading configu- 1567. https://doi.org/10.1007/s11831- 021- 09626- 2.
rations. Computer Methods in Biomechanics and Biomedical Engineer- Yamasaki, S., Yaji, K., & Fujita, K. (2021). Data-driven topology design
ing: Imaging & Visualization, 4(2), 61–72. https://doi.org/10.1080/21 using a deep generative model. Structural and Multidisciplinary Op-
681163.2015.1030775. timization, 64(3), 1401–1420. https://doi.org/10.1007/s00158- 021- 0
Van Engelen, J. E., & Hoos, H. H. (2020). A survey on semi-supervised 2926-y.
learning. Machine Learning, 109(2), 373–440. https://doi.org/10.100 Yang, X. Y., Xie, Y. M., Steven, G. P., & Querin, O. M. (1999). Bidirec-
7/s10994- 019- 05855- 6. tional evolutionary method for stiffness optimization. AIAA Jour-
Van Hasselt, H., Guez, A., & Silver, D. (2016). Deep reinforcement nal, 37(11), 1483–1488. https://doi.org/10.2514/2.626.
learning with double q-learning. In Proceedings of the AAAI Con- Yildiz, A. R., Öztürk, N., Kaya, N., & Öztürk, F. (2003). Integrated op-
ference on Artificial Intelligence(Vol. 30). https://doi.org/10.1609/aa timal topology design and shape optimization using neural net-
ai.v30i1.10295. works. Structural and Multidisciplinary Optimization, 25(4), 251–260.
Wang, L., Chan, Y. C., Ahmed, F., Liu, Z., Zhu, P., & Chen, W. (2020). Deep https://doi.org/10.1007/s00158- 003- 0300- 0.
generative modeling for mechanistic-based learning and design Yu, Y., Hur, T., Jung, J., & Jang, I. G. (2019). Deep learning for de-
of metamaterial systems. Computer Methods in Applied Mechanics termining a near-optimal topological design without any itera-
and Engineering, 372, 113377. https://doi.org/10.1016/j.cma.2020.1 tion. Structural and Multidisciplinary Optimization, 59(3), 787–799.
13377. https://doi.org/10.1007/s00158- 018- 2101- 5.
Wang, M. Y., Wang, X., & Guo, D. (2003). A level set method for struc- Zehnder, J., Li, Y., Coros, S., & Thomaszewski, B. (2021). NTopo: Mesh-
tural topology optimization. Computer Methods in Applied Mechanics free topology optimization using implicit neural representations.
and Engineering, 192(1-2), 227–246. https://doi.org/10.1016/S0045- Advances in Neural Information Processing Systems, 34, 10368–10381.
7825(02)00559-5. Zhang, Z., Li, Y., Zhou, W., Chen, X., Yao, W., & Zhao, Y. (2021a).
Wang, C., Yao, S., Wang, Z., & Hu, J. (2021). Deep super-resolution neu- TONR: An exploration for a novel way combining neural network
ral network for structural topology optimization. Engineering Op- with topology optimization. Computer Methods in Applied Mechanics
timization, 53(12), 2108–2121. https://doi.org/10.1080/0305215X.2 and Engineering, 386, 114083. https://doi.org/10.1016/j.cma.2021.1
020.1846031. 14083.
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1763

Zhang, C., Liu, J., Yuan, Z., Xu, S., Zou, B., Li, L., & Ma, Y. (2021b). A
novel lattice structure topology optimization method with ex-
treme anisotropic lattice properties. Journal of Computational De-
sign and Engineering, 8(5), 1367–1390. https://doi.org/10.1093/jcde
/qwab051.
Zhang, Y., Peng, B., Zhou, X., Xiang, C., & Wang, D. (2019). A deep con-
volutional neural network for topology optimization with strong
generalization ability. arXiv preprint arXiv:1901.07761.
Zhao, J., Mathieu, M., & LeCun, Y. (2016). Energy-based generative ad-
versarial network. arXiv preprint arXiv:1609.03126.
Zheng, L., Kumar, S., & Kochmann, D. M. (2021). Data-driven topology
optimization of spinodoid metamaterials with seamlessly tun-
able anisotropy. Computer Methods in Applied Mechanics and Engi-
neering, 383, 113894. https://doi.org/10.1016/j.cma.2021.113894.
Zhou, Y., & Saitou, K. (2017). Topology optimization of composite

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

structures with data-driven resin filling time manufacturing con-
straint. Structural and Multidisciplinary Optimization, 55(6), 2073–
2086 https://doi.org/10.1007/s00158- 016- 1628- 6.
Zhuang, X., Guo, H., Alajlan, N., Zhu, H., & Rabczuk, T. (2021a). Deep
autoencoder based energy method for the bending, vibration, and
buckling analysis of Kirchhoff plates with transfer learning. Eu-
ropean Journal of Mechanics-A/Solids, 87, 104225. https://doi.org/10
.1016/j.euromechsol.2021.104225.
Zhuang, C., Xiong, Z., & Ding, H. (2021b). Temperature-constrained
topology optimization of nonlinear heat conduction problems. Figure A1: TO of MBB half beam using four different method.
Journal of Computational Design and Engineering, 8(4), 1059–1081.
https://doi.org/10.1093/jcde/qwab032.
(0 ≤ xmin ≤ xe ≤ 1). xmin is greater than 0, but it is the minimum al-
lowable relative density of the empty element, which is to prevent
Appendix 1: Background of TO the singularity of the FE matrices. It can also solve the problem in
which an element with zero density cannot be re-densed.
Appendix 1 introduces the representative TO methodologies that
According to Bendsoe and Sigmund (2003) and Sigmund
are widely used in ML for TO studies as a foundation.
(2001b), the basic foliation of SIMP for minimizing compliance can
A typical TO problem is defined as follows. Find the material
be expressed as follows:
distribution that minimizes the objective function F, subject to a
volume constraint G0 ≤ 0 and possibly M other constraints Gi ≤ 
Ne
0, i = 1 ...M. The material distribution is described by the density min c (x ) = U T KU = uTe (Ee (xe ) k0 ) ue
e=1 (A.2)
variable ρ(x) that can take either 0 (void) or 1 (solid material) at any s.t.
V (x )
=f KU = F 0 ≤ xe ≤ 1, e = 1, . . . , Ne
V0
point in the design domain  (Bendsøe & Sigmund, 1999; Sigmund,
2001a; Sigmund & Torquato, 1997). where U is a displacement vector; K is a global stiffness matrix; c(x)
TO can be expressed as follows: is the compliance; ue is an element displacement vector; k0 is an
element stiffness matrix; f is the volume fraction; Ne is the number
of elements; xe is the design variable (i.e., density) of element e;
min F = F (u (ρ ) , ρ) = ∫ f (u (ρ ) , ρ) dV
and V(x) and V0 are the material volume and the volume of design
s.t. G0 (ρ ) = ∫ ρ (x ) dV − V0 ≤ 0
 (A.1) domain, respectively.
Gi (u (ρ ) , ρ) ≤ 0, j = 1, . . . , M As the relative density of the material is continuous, and ac-
ρ (x ) = 0or1, ∀x ∈  cordingly, the Young’s modulus of the material may change con-
where the state field u satisfies a linear or non-linear state equa- tinuously.
tion. The modulus of elasticity E(xe ) for each element is calculated
This section introduces the density-based method, ESO by power law, reflecting the concept of penalty factor P for Young’s
method, LSM, and MMC method among various TO methods. Fig- modulus of solid material E0 and the relative density of the ma-
ure A1 shows the results of optimizing MBB half beam, a repre- terial xe . Emin is the stiffness of soft (void) material (non-zero to
sentative benchmark problem of TO, by using the four methods. avoid singularity of the stiffness matrix). E0 is Young’s modulus
of solid material.
A1.1. Density-based method The relationship between the density of elements in Modi-
fied SIMP and Young’s modulus is expressed as follows (Sigmund,
The SIMP method is the most representative density-based
2007):
method, which obtains the optimal material distribution for spe-
cific load and boundary conditions in a given design space (Bend-
E (xe ) = Emin + xPe (E0 − Emin ) (A.3)
søe & Kikuchi, 1988; Rozvany et al., 1992). Bendsøe (1989) men-
tioned that “shape optimization in its most general setting should Penalty factor p reduces the effect of gray elements (elements
consist of a determination for every point in space whether there with medium density) on total stiffness (i.e., numerical experi-
is material in that point or not.” In the SIMP method, the relative ments indicate that a penalty factor value of p = 3 is suitable.).
density of each element has a density value that varies between In order for the optimizer to find the correct search direction
the minimum value xmin and 1 rather than a value of 0,1 binary in gradient-based optimization, sensitivity analysis of the objec-
1764 | Topology optimization via artificial intelligence: a review

tive function and constraints for the design variable is required in from the structure to find the optimal structure. BESO, an ad-
each iteration. This optimization algorithm for maximizing stiff- vanced method of the ESO, was introduced in Yang et al. (1999),
ness performs sensitivity analysis to determine the effect of ele- which removes unnecessary materials while adding necessary
materials. The optimization problem that minimizes compliance
ment density changes on the objective function and constraints.
with the most basic ESO is defined as follows:
Therefore, sensitivity analysis can be expressed as a differential of
min c (x ) = U T KU
the objective function and the constraint for the element density V (x ) (A.11)
s.t. V0
=f KU = F xe = [ 0, 1] , e = 1, . . . , Ne
under the setting that all elements have unit volume.
At this time, the criteria for removing elements are based on the
∂C
= − p(xe ) p−1 [ue ]T [Ke ] [ue ] (A.4) sensitivity number, and the sensitivity number can be obtained
d∂
N  through the following equation (Liang et al., 2000).
∂V ∂  e

= xe ve = 1 (A.5) 1 T
∂ xe ∂ xe αie =
u Ki ui (A.12)
e=1 2 i
The OC method is a method commonly used to update design Through this, the sensitivity number is calculated to remove
variables in TO problems, and is most commonly applied to the elements with elements lower than the criteria.
density-based method. The method of updating design variables

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

BESO is a method of removing elements as the process of ESO
through OC is given by while adding elements to the required area. Huang and Xie (2007)
⎧ η
⎨max (0, xe − m ) i f xe Be ≤ max (0, xe − m ) introduced a complementary BESO method that solves the mesh
xe = min (1, xe + m) i f xe Bηe ≥ min (1, xe + m)
new
(A.6) dependency problem without converging. The nodal sensitivity
⎩
xe Bηe otherwise number is used to solve the checkerboard pattern which is the
where m denotes a positive move-limit, η denotes a numerical average of the element sensitivity connected to the jth node and
damping coefficient, Be follows the formula can be obtained as follows:
− (∂ c/∂ xe ) 
M
Be = (A.7) α nj = ωi αie (A.13)
λ (∂ V/∂ xe )
i=1
In addressing the checkerboard patterns and grayscale issues, where ωi denotes the weight value of the ith element, and is de-
filtering methods, such as density filter and sensitivity filter, are fined as follows:
presented, which allow the optimal design to be derived by con-  
1 ri j
sidering the connection between elements instead of performing ωi = 1 − M (A.14)
M−1 i=1 ri j
sensitivity analysis independently for each element.
The basic filters applied to TO are the sensitivity filter and the where ri j is the distance from the center of the ith element to the
density filter, which consider the connection between elements jth node.
through the method of weighted average with neighboring ele- Then, the nodes affecting the sensitivity of the ith element may
ments of sensitivity or element density, respectively. The neighbor- be found through rmin . The nodes in this domain contribute to
hood element is determined based on the distance from the center the calculation of the ith element’s improved sensitivity number,
of the reference element, and the maximum distance considered which is calculated as follows:
K
j=1 ω ri j α j
as a neighbor is a parameter set by the user as mesh-independent n

radius. On this basis, the convolutional operator is defined as fol- αi = K (A.15)

j=1 ω ri j
lows:
where K is the total number of nodes in the filter domain of rmin ,
H f = rmin − dist (e, f ) (A.8) and ω(ri j ) is defined as a linear weight factor as follows:
where rmin means mesh-independent radius dist(e, f ) is the dis- ω ri j = rmin − ri j ( j = 1, 2, . . . , K) (A.16)
tance between elements and f is the element whose distance
from the center of e is less than rmin . The sensitivity filter defined In addition to the solid element, the sensitivity number of the
using this convolution operator is expressed as follows (Sigmund void element is calculated, and elements with a sensitivity num-
& Torquato, 1997): ber less than αdel
th
among the solid elements are removed, and ele-
ments with a sensitivity number greater than αadd
th
among the void
∂c 1  N
∂c elements are added.
= N Hfxf (A.9)
∂ xe xe f =1 H f f =1 ∂x f
A1.3. Level-set method
In the case of a density filter, the weighted average concept is
The LSM is used to capture the free boundaries of a structure in
the same as that of a sensitivity filter, but there is a difference that
a linear elastic structure, resulting in a crisp design. Therefore,
physical density is used instead of sensitivity, and it is defined as
the results require less postprocessing. This method is fundamen-
follows (Bourdin, 2001; Bruns & Tortorelli, 2001):
N tally different from shape optimization because it not only shifts
f =1 H f x f structural boundaries but also includes formation, extinction, and
x˜ e = N (A.10)
f =1 H f
merging of void regions that define phase designs (Sethian &
Wiegmann, 2000).
Accordingly, the original density corresponds to the design vari-
In the LSM, the boundary is expressed as a zero-level curve
able, the filtered density corresponds to the physical density, and
of the scalar function . The movements and merging of bound-
the sensitivity analysis to the design variable is modified for the
aries are operated by this scalar function and regulated by physi-
physical density (Guest et al., 2004).
cal problems and optimization conditions. The LSM implicitly ex-
presses the boundaries and allows a smooth boundary to be ex-
A1.2. ESO method
pressed in a 3D space from a zero function value.
ESO methods were first introduced in Xie and Steven (1993, 1997).
The methodology proceeds by removing unnecessary materials S = x : φ (x, 0 ) = k (A.17)
 Journal of Computational Design and Engineering, 2023, 10(4), 1736–1766 | 1765

where k is an arbitrary value containing a zero function value and variables with di = (x0i, y0i , Li , ti , θi )T , i = 1, . . . nc. Ud is the admissi-
x is an arbitrary coordinate on the curved surface φ. As the struc- ble sets that d belongs to. Moreover, gi , i = 1, ..., l are the consid-
tural optimization changes over time, Equation (A.17) can be ex- ered constraint functionals.
pressed as level set function dynamically changing as follows:

S (t ) = x (t ) : (x (t ) , t ) = k (A.18) Appendix 2: Background of ML

Differentiating this expression with time and applying the chain DL models can be categorized into four types: supervised learn-
rule defines the “Hamilton–Jacobi-type” equation that defines the ing, unsupervised learning, semi-supervised learning, and RL. Su-
initial value for the time-dependent function φ. pervised learning is a method of training by using labeled data.
∂ (x, t ) dx In the case of unsupervised learning, unlabeled data are used for
+∇ (x, t ) = 0, (x, t = 0 ) = 0 (x ) (A.19)
∂t dt training and training is done by finding undefined features or pat-
In solving this equation, let ddxt be the movement of a point on terns from input data. Semi-supervised learning is a method of
a surface driven by the objective of the optimization. Then, as a using a mixture of supervised and unsupervised learning meth-
solution of a PDE on φ, the optimal structural boundary can be ods, including training a larger unlabeled dataset by using small

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

expressed as follows (Wang et al., 2003): amounts of labeled data or by utilizing a proper combination of
supervised and unsupervised learning models. RL is a different
∂ (x, t ) concept from supervised/unsupervised learning, which does not
≡ −∇ (x ) (x, ) (A.20)
∂t require data or labels separately, receives a reward for its action on
where (x, ) denotes the “speed vector” of the level set and it de- the agent while the training progresses, and the model is trained
pends on the objective of the optimization. Proper vector obtains to build a strategy for action in the direction of increasing this
an appropriate value from the descent direction of the objective reward.
through sensitivity analysis.

min C (u, ) = ∫ Eε (u ) ε (u ) H ( ) d A2.1. Supervised learning


s.t. ∫ H () d ≤ V f The biggest characteristic of supervised learning is that labels ex-
(A.21)
 ist together with the input. Based on the labeled dataset, pre-
∫ Eε (u ) ε (v ) H ( ) d = ∫ bvH ( ) d + ∫ fv s u| = 0∀v ∈ U
  s
D
dictions are initiated for new data. Supervised learning is largely
divided into classification and regression. Classification includes
A1.4. MMC method binary classification, which classifies data into one of two cate-
In contrast to the conventional pixel-based and node point-based gories, and multiclass classification, which classifies data into one
methods, MMC method can organize an optimal structure by de- of several classes. Regression is used to predict continuous val-
termining shape characteristic parameters, such as shape, length, ues based on the features of the dataset. The representative algo-
thickness, and direction, and the layout of these components (Guo rithms for supervised learning include KNN, SVM, decision tree,
et al., 2014). random forest, and NN.
The MMC method can incorporate numerous geometric infor- The KNN is a method used to perform classification, which is
mation for defined problems, and express quite complex and di- based on the principle that instances within a dataset typically
verse structure topologies with a small number of components. exist close to other instances with similar properties. (Cover &
In the MMC method, components can overlap each other, and the Hart, 1967). SVM is a method used for classification and regres-
overlapping regions on the other components do not have a struc- sion to maximize the margin of the data class and thereby create
tural effect. the largest distance between the separated hyperplane and the
The area in which the component exists may be represented by instances on both sides (Burges, 1998). A decision tree is a tree
a level set function. that sorts and classifies instances based on the feature, and each
node in the decision tree represents the feature of the instance to
{∅ (x ) > 0, i f x ∈  ∅ (x ) = 0, i f x ∈ ∂ ∅ (x ) < 0, i f x ∈ D\ , be classified, and each branch represents the value that the node
(A.22) can predict. Instances are classified and sorted by corresponding
features, starting at the root node (Murthy, 1998). Random forest is
where an ensemble learning method used for classification and regres-
 
cosθ (x − x0 ) + sinθ (y − y0 ) m sion, formed by a combination of several decision trees (Breiman,
∅ (x, y ) =
L/2 2001). ANN is a concept that mathematically models the struc-
  ture of human brain neurons and has since developed into a DNN
−sinθ (x − x0 ) + cosθ (y − y0 ) m
+ −1 (A.23) that deepens the hidden layer of ANN (Liu et al., 2017). Among
t/2
DNNs, the most representative CNN consists of a convolutional
where m is a relatively large even integer number (we take m = 6 layer, non-linearity layer, pooling layer, and fully connected layer,
in this study). The structural component can move, dilate/shrink and shows good performance in image data application problems
and rotate in the design domain by changing the values of x0, y0, (Albawi et al., 2017).
L, t and θ .
On the basis of the discussion above, the MMC-based TO prob-
A2.2. Unsupervised learning
lem is defined as follows:
Unsupervised learning refers to a training method in which a
Find d = (d1 , . . . , dnc )T
model receives input but has no label for it. Therefore, the model
min I = I (d ) (A.24)
clusters according to the similarity based on the input data and
s.t. gi (d ) ≤ 0, i = 1, . . . , l , d ⊂ Ud
learns the characteristics and patterns of the data by itself.
where nc represents the total number of components involved in Unsupervised learning is commonly used for clustering, di-
the optimization problem, d = (d1 , . . . dnc )T is the vector of design mensionality reduction, and generative models. Clustering is a
1766 | Topology optimization via artificial intelligence: a review

method of grouping data with similar features into the same clus- training the decoder to be used as a generative model, training is
ter. As an example, K-means is the simplest and most popular al- performed by attaching an encoder.
gorithm for iterative and hill climbing clustering algorithms. The
K-means is an algorithm that groups given data into k clusters, A2.3. Reinforcement learning
which are clustered based on the center of each cluster and the RL is a type of ML that makes decisions through the interaction
average distance of the data. K-means were originally designed to of agents and environments (Sutton & Barto, 2018). RL is usually
find the optimal partition for dividing data, but k-means generally based on the MDP, but it does not require explicit models.
do not converge to the optimal partition (MacQueen, 1967). The agent takes an action at in the current time t, which returns
Dimensionality reduction is the process of extracting high- state st and reward Rt from the environment ε. The algorithm is
dimensional data into low-dimensions, which still compresses trained through trial-and-error so that the agent determines the
meaningful features of the data. For example, PCA is a mathemat- action in the direction of maximizing the total reward and returns
ical algorithm that reduces the dimension of data while main- 
Rt = γ k · Rt+k received by the agent in the state of the given en-
taining the variations of most of the data, and the dimension is k
vironment (where γ ∈ (0, 1 ), γ is a discount factor that reflects
reduced by finding a basis that maximizes the variance (Jolliffe,
future rewards in the current state). The two main approaches

Downloaded from https://academic.oup.com/jcde/article/10/4/1736/7223974 by guest on 02 August 2023

2002). Dimensionality reduction allows samples to be represented
to the RL framework are the value functions based approach and
in relatively low dimension and these reduced samples with sim-
the policy search based approach. The value function method is
ilar features are most likely to be gathered together in the latent
based on the method of estimating the expected return value in a
space (Ringnér, 2008).
given state, and deep Q network (DQN) and double-DQN the repre-
GAN, the most common generative model, consists of two NNs,
sentative method (Mnih et al., 2013, 2015; Van Hasselt et al., 2016).
a generator and a discriminator. The generator generates new
The policy method finds the gradient of the policy and updates
data based on existing data as the reference, and the discrimi-
the parameters in the direction of increasing the reward. Advan-
nator aims to distinguish between the real data and the data gen-
tage actor–critic and asynchronous advantage actor–critic are the
erated by the generator, allowing the generator to generate data
most representative methods by using these policy updates (Mnih
similar to the actual data (Goodfellow et al., 2020). By expanding
et al., 2016). Furthermore, as a hybrid actor critic method that ap-
this, cGAN uses additional information as a condition for the gen-
plies the concept of both methods, Schulman et al. (2015) proposed
erator and the discriminator (Mirza & Osindero, 2014). VAE is also
trust region policy optimization, and Schulman et al. (2017) pro-
a generative model that generates new data similar to the exist-
posed PPO.
ing data from a latent space, which contains features of the exist-
ing data (Kingma & Welling, 2014). Therefore, for the purpose of

Received: October 16, 2022. Revised: June 30, 2023. Accepted: July 1, 2023
© The Author(s) 2023. Published by Oxford University Press on behalf of the Society for Computational Design and Engineering. This is an Open Access article distributed
under the terms of the Creative Commons Attribution-NonCommercial License (https://creativecommons.org/licenses/by-nc/4.0/), which permits non-commercial re-use,
distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact [email protected]