Machine Learning Driven Forward Prediction and Inverse Design for 4D Printed
Hierarchical Architecture with Arbitrary Shapes
Liuchao Jina,b,c , Shouyi Yub,c , Jianxiang Chengb,c , Haitao Yec,d , Xiaoya Zhaie , Jingchao Jiangf , Kang Zhanga ,
Bingcong Jiang,b,c , Mahdi Bodaghih , Qi Geb,c,∗, Wei-Hsin Liaoa,i,∗
a Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Hong Kong, China
b Shenzhen Key Laboratory of Soft Mechanics & Smart Manufacturing, Southern University of Science and Technology, Shenzhen, 518055, China
c Department of Mechanical and Energy Engineering, Southern University of Science and Technology, Shenzhen, 518055, China
d Department of Mechanical Engineering, City University of Hong Kong, Kowloon, Hong Kong, China
e School of Mathematical Sciences, University of Science and Technology of China, Hefei, 230026, China
f Department of Engineering, University of Exeter, Exeter, United Kingdom
g School of Mechanical Engineering, Tongji University, Shanghai, 200092, China
h Department of Engineering, School of Science and Technology, Nottingham Trent University, Nottingham, NG11 8NS, UK
i Institute of Intelligent Design and Manufacturing, The Chinese University of Hong Kong, Hong Kong, China
Abstract
The forward prediction and inverse design of 4D printing have primarily focused on 2D rectangular surfaces or plates,
leaving the challenge of 4D printing parts with arbitrary shapes underexplored. This gap arises from the difficulty
of handling varying input sizes in machine learning paradigms. To address this, we propose a novel machine
learning-driven approach for forward prediction and inverse design tailored to 4D printed hierarchical architectures
with arbitrary shapes. Our method encodes non-rectangular shapes with special identifiers, transforming the design
domain into a format suitable for machine learning analysis. Using Residual Networks (ResNet) for forward prediction
and evolutionary algorithms (EA) for inverse design, our approach achieves accurate and efficient predictions and
designs. The results validate the effectiveness of our proposed method, with the forward prediction model achieving
a loss below 10−2 mm, and the inverse optimization model maintaining an error near 1 mm, which is low relative to
the entire shape of the optimized model. These outcomes demonstrate the capability of our approach to accurately
predict and design complex hierarchical structures in 4D printing applications.
Keywords: 4D printing, machine learning, inverse design, hierarchical architecture, design optimization, residual
network, evolutionary algorithm
1. Introduction
4D printing, an evolution of traditional 3D printing [1–5], has gained significant attention in the past decades due
to its revolutionary potential in construction [6], textile [7, 8], automotive [9], aerospace industry [10], and biomedical
applications [11–15]. Unlike conventional additive manufacturing processes that yield static objects, 4D printing
empowers the creation of dynamic structures capable of self-transformation over time in response to external stimuli
[16, 17], such as heat [18–24], light [25, 26], humidity [27, 28], pH [29, 30], and electric or magnetic fields [31–34].
Advances in 4D printing depend on the development of robust forward prediction and inverse design methods [35–
37]. Forward prediction involves predicting how a 4D-printed structure will behave when subjected to specific stimuli,
such as temperature fluctuations, moisture absorption, or mechanical deformation [38]. In contrast, inverse design
focuses on optimizing the material composition and structural configuration of a component to obtain predefined
performance criteria or a desired response [39]. These two aspects of 4D printing design (forward prediction and
∗ Correspondingauthors
Email addresses: geq@[Link] (Qi Ge), whliao@[Link] (Wei-Hsin Liao)
Preprint submitted to Applied Materials Today August 27, 2024
inverse optimization) form the cornerstone of efficient structure development, providing insights into material behavior
and guiding the design process to achieve desired results.
In traditional design workflows, finite element analysis (FEA) has served as the primary tool for predicting the
behavior of 4D printed structures. However, the reliance on FEA poses several limitations, including time-consuming
simulations and computational burdens, particularly in inverse design tasks requiring millions of iterations. Hamel
et al. [40] proposed a machine-learning approach that combines the finite element method with an evolutionary
algorithm to design active composite structures capable of achieving target shape-shifting responses. Their method
optimizes the distribution of passive and active materials within equally sized voxel units to attain the desired shape
changes. However, the reliance on finite element analysis for forward prediction poses computational challenges
due to its time-consuming nature. Similarly, Athinarayanarao et al. [41] addressed the inverse design problem of
4D printing by presenting a computational framework based on finite element analysis and evolutionary algorithms.
Their approach optimizes both the materials distribution and material layout within a design space through topology
optimization, offering a highly capable tool for designing 4D-printed active composites. Nonetheless, like Hamel et
al. [40]’s approach, the utilization of FEA for forward prediction presents computational bottlenecks.
arning Enabled Forward
Le Pre
hine dic
tio
ac n
M Hierarchical
Architectures
Design Behavior
Mechanical 4D
Performance Printing
n
Ev
io
lu
iz
at
tio
o
m
na
ry O p ti
A lg se
o rith er
m s E na ble d In v
Figure 1: Schematic diagram of the proposed loop pipeline including machine learning-enabled forward prediction and evolutionary algorithms
enabled inverse optimization.
To overcome these challenges and expedite the design process, machine learning emerges as a compelling alterna-
tive [5]. By leveraging machine learning algorithms, it becomes possible to establish relationships between material
allocation and desired responses, facilitating fast and accurate predictions. The integration of machine learning into
4D printing design workflows aims to streamline the forward prediction and inverse optimization processes, offering
a data-driven approach to structure development. Sun et al. [42] introduced a novel machine learning and evolu-
tionary algorithm-based approach for designing 4D-printed active composite structures. They employed a recurrent
neural network (RNN) based machine learning model for forward shape-change prediction and then use evolutionary
algorithms to solve the inverse problem of finding the optimal design. Despite the efficiency demonstrated by their
2
approach for multiple target shapes, it focuses primarily on forward prediction and optimization of 2D rectangular sur-
faces, limiting its applicability to arbitrary shapes. Similarly, Sun et al. [43] proposed an integrated machine learning
and sequential subdomain optimization approach for ultrafast inverse design of 4D-printed active composite struc-
tures. However, like previous studies, its focus is limited to forward prediction and optimization of 2D rectangular
surfaces, restricting its applicability to arbitrary shapes. Sun et al. [44] also focused on inverse design for active com-
posite plates which were still rectangle. Their approach combines machine learning (ML) with gradient-descent (GD)
and evolutionary algorithms (EA) to efficiently determine the material distribution necessary for achieving desired 3D
shape changes. Besides, Jin et al. [45] designed a residual neural network-enabled forward prediction and genetic
algorithm-based inverse optimization for the design and optimization of 4D-printed rectangular hierarchical architec-
ture. In addition, Zhao et al. [46] proposed an optimization method for grayscale digital light processing (DLP) 3D
printing based on machine learning and evolutionary algorithms. Their approach integrates automated finite element
model-based evaluation and a machine learning model based on recurrent neural networks to efficiently predict and
optimize grayscale distributions for desired deformations. However, similar to previous approaches, it is tailored for
2D rectangular surfaces and lacks versatility in handling arbitrary shapes.
Therefore, in this paper, we present a novel approach to forward prediction and inverse design for 4D printed
hierarchical architecture with arbitrary shapes, driven by machine learning techniques. Our methodology aims to
address the limitations of traditional design methods by harnessing the power of machine learning to predict structure
behavior and optimize design parameters. By encoding non-rectangular shapes with unique identifiers and leveraging
Residual Networks (ResNet) for forward prediction and evolutionary algorithms (EA) for inverse design, our pro-
posed framework demonstrates promising results in predicting and designing complex hierarchical structures in 4D
printing applications. The schematic diagram of the proposed loop pipeline is shown in Figure 1. On the one hand,
machine learning-enabled forward prediction is used to efficiently figure out the relationship between the design and
the behavior of the hierarchical architectures, which answers the question of how to predict the behavior of this design
when subjected to stimuli. On the other hand, the evolutionary algorithms enabled inverse optimization is applied to
inverse design the hierarchical architecture, which answers the question that given desired behavior, how to design the
part so that the designed part can reach the desired behavior when subjected to stimuli. Through empirical validation
and case studies, we showcase the efficacy of our approach in achieving precise and efficient design outcomes, paving
the way for the advancement of 4D printing technology.
The remainder of this paper is organized as follows. In Section 2, we will discuss the modelling and digital
representation of 4D printed hierarchical architecture with arbitrary shapes to facilitate training and optimization
by machine learning algorithms. Section 3 will present the methodology of the forward prediction method based
on machine learning. In Section 4, we will illustrate the inverse optimization method for 4D printed hierarchical
architecture with arbitrary shapes using evolutionary algorithms. Section 5 will present the results obtained from our
approach. Finally, in Section 6, we will draw conclusions based on our findings and discuss potential future works in
this field.
2. Modelling and Digital Representation of Hierarchical Architecture
To facilitate machine learning training and understanding, we employ digital representation to encode the material
allocation within the hierarchical architecture. For instance, in Figure 2, depicting a hierarchical architecture with
two distinct materials-blue and red-we assign numerical labels for representation. Specifically, we designate ”1” to
represent the blue material and ”2” to denote the red material. This digital representation enables the entire structure
to be inputted into machine learning algorithms for analysis and learning purposes.
It’s important to note that within this representation, we introduce the definitions of active and passive materials to
enhance clarity and differentiation between them. Active materials, in the context of 4D printing, refer to those with a
significant thermal expansion ratio, meaning they undergo considerable dimensional changes in response to changes in
temperature. These materials exhibit dynamic behavior, expanding or contracting when subjected to external stimuli
such as heat. In contrast, passive materials either have no thermal expansion ratio or exhibit minimal dimensional
changes in response to temperature variations.
The concept of hierarchical architecture with active and passive materials harnesses the unique properties of these
materials to achieve the 4D printing effect. By strategically combining active and passive materials within a hierar-
chical structure, dynamic systems capable of self-transformation over time can be created. When exposed to external
3
Triangle Circle Pentagon
Material
Allocation
Active Material
Passive Material
Digital
Representation
Active Material
Passive Material
No Material
Figure 2: The demonstration of digital representation of active and passive materials with arbitrary shapes.
stimuli such as heat or moisture, the active materials undergo significant dimensional changes, while the passive ma-
terials remain relatively unchanged. This contrast in behavior between the active and passive components drives the
shape-shifting capabilities of the hierarchical architecture, resulting in the observed 4D printing effect. The hierar-
chical arrangement ensures that specific regions of the structure respond differently to external stimuli, allowing for
precise control over the overall shape transformation process.
The existing research commonly employs matrices of 0s and 1s to digitally represent the passive and active materi-
als for data inputs of entire rectangular parts [40, 42, 43], facilitating machine learning analyses. However, challenges
emerge when dealing with non-rectangular shapes, leading to either non-uniform input data for recurrent neural net-
works or non-rectangular input data for convolutional neural networks. These discrepancies in input data format pose
obstacles to the application of machine learning techniques, highlighting the need for solutions to ensure uniformity
and compatibility with neural network architectures.
To address the challenge posed by non-rectangular shapes in machine learning models, we propose a solution
wherein we fill the gaps between the non-rectangular shape and a rectangular shape with zeros. Subsequently, we
assign non-zero numbers to represent the material-containing parts of the non-rectangular shape. This approach
ensures that the input data fed into the machine learning model becomes uniform and rectangular, thereby facilitating
the training process and enabling efficient analysis and prediction.
As depicted in Figure 2, the nonrectangular shapes such as triangles, circles, and pentagons represent the regions
requiring material allocation optimization. Conventional approaches may use binary values (0 for passive materials,
1 for active materials) to represent such shapes, leading to irregular input data shapes, posing challenges for machine
learning models. Our proposed solution involves enclosing these irregular shapes within a rectangular frame, filling
the empty spaces with zeros, and assigning numeric values (e.g., 1 for passive materials, 2 for active materials) to
the corresponding regions. Additional numeric values can be utilized to represent further material variations. This
method ensures that the input data shape fed into the machine learning model becomes uniform or rectangular, thereby
enhancing model compatibility and efficiency.
4
3. Forward Prediction Methods
In the context of forward prediction methods for hierarchical architecture, the primary objective revolves around
harnessing machine learning algorithms to establish correlations between the allocation of different materials within
components and the resultant behavior of the printed parts. Traditionally, finite element analysis has served as the pre-
dominant method for forward prediction, involving intricate simulations to replicate real-world scenarios. However,
the inherent time-consuming nature of FEA, particularly in the context of inverse design tasks, presents a signifi-
cant bottleneck. With inverse design necessitating millions of cases for thorough forward prediction evaluation, the
limitations of FEA become apparent, especially when dealing with large or complex parts. Consequently, the ad-
vancement of inverse design methodologies is hindered by the computational overhead associated with traditional
FEA approaches. To address this challenge, our proposal advocates for the integration of machine learning as a viable
alternative to FEA. By leveraging machine learning, which can assimilate data from FEA outcomes or experiment
data in reality, predictive models can be developed to establish robust relationships between material allocation and
behavior. These models offer a substantially faster alternative to traditional FEA simulations, thereby expediting the
forward prediction process and facilitating the exploration of diverse material configurations.
In this paper, we utilize ResNet, short for Residual Networks, as our forward prediction model due to its re-
markable capability to capture intricate patterns and features within complex datasets. ResNet is a deep learning
architecture that was introduced to address the vanishing gradient problem in deep convolutional neural networks
[47]. The vanishing gradient problem occurs when training deep networks with many layers, as the gradients can
become extremely small, leading to slow convergence or even a complete halt in learning. ResNet introduces skip
connections or shortcut connections that enable the flow of gradients directly from earlier layers to later layers, by-
passing a few layers in between. These skip connections allow for the training of deeper networks by alleviating the
vanishing gradient problem. By propagating the gradients more effectively, ResNet enables the training of models
with hundreds or even thousands of layers. The skip connections in ResNet are implemented through the use of
residual blocks. A residual block consists of a convolutional layer followed by batch normalization and a non-linear
activation function. The output from the block is then added to the input of the block, forming the skip connection.
This addition operation allows the network to learn the residual or the difference between the input and the desired
output. The ResNet structure we have constructed is visually depicted in Figure 3, providing a clear representation of
its architectural layout and components. In the subsequent subsections, we embark on a detailed exploration of the
ResNet model, elucidating its operational process and underlying methodologies.
3.1. Data Acquisition
The data acquisition process for the machine learning model relies on the utilization of Abaqus simulation soft-
ware to generate the requisite datasets. This process unfolds in several key stages. Firstly, we employ Python scripting
to automate the creation of Abaqus input files tailored to our specific requirements. These input files encapsulate the
parameters and conditions necessary for simulating the behavior of our hierarchical architectural structures. Through
this automation, we generate a substantial volume of input files, ensuring an ample dataset for subsequent analysis and
model training. Subsequently, we initiate Abaqus simulations in parallel using Python, a crucial step that accelerates
the data generation process. By harnessing parallel computing capabilities, we can execute multiple simulations con-
currently, thereby maximizing computational efficiency and expediting the acquisition of output databases. During the
simulation phase, Abaqus calculates and records various output metrics, including deformation data, which are inte-
gral to our machine learning task. Upon completion of the simulations, we access the output databases containing the
pertinent information generated during the computational experiments. To extract the requisite data for our machine
learning models, we employ Python scripts to read and parse the output databases generated by the Abaqus simula-
tions. This step allows us to systematically retrieve the material allocation data and corresponding deformation data
for each simulation case, facilitating the assembly of comprehensive datasets for analysis and model training. Upon
completion of the data acquisition process, we possess the necessary material allocation data and deformation data
required to train and validate our machine learning models. These datasets serve as the foundation for the machine
learning model, enabling it to explore the complex relationships between material allocation and structural behavior
within hierarchical architectural designs.
Here, we present the details for the finite element analysis. For meshing. we employed a detailed meshing strategy
to enhance the accuracy of our results. Specifically, each voxel in our model was subdivided into a 10 × 10 mesh. This
5
Forward Prediction Model
Active Material
Passive Material
Input: 2: Active Material
Output: Displacement
Material Allocation 1: Passive Material
Deformed Shape Curve / Field
0: No Material
Input Data
x
3 × 3 Conv
3 × 3 Conv
Batch Norm
Batch Norm
ResNet Model
2 × 2 MaxPool
ReLU 1 × 1 Conv
ResNet Block 3 × 3 Conv
5× Batch Norm
f (x)
ResNet Block
+
Global AvePool
ReLU
ResNet Block
Output Data
Figure 3: Forward prediction model constructed by ResNet to predict the deformation based on the material allocation.
6
means that every voxel, which represents a discrete volume element in the simulation space, is further divided into a
grid of 100 smaller finite elements.
The boundary conditions are essential for ensuring realistic simulations in finite element method. In our simulation
setup using Abaqus, we implemented a fixed boundary condition on the left edge of the model during the temperature
increasing step. This boundary condition restricts displacement in the x-direction to zero, simulating a scenario where
the left edge of the structure is fixed or clamped. The entire part is subjected to a temperature change of 30◦ C.
In this paper, we used TPU as the passive material and SMP (with a glass transition temperature of 55 degrees
Celsius) from SMP Technology, Japan, as the active material. The materials were then printed using a dual-nozzle
Polarbear 3D printer. The material properties for the active material are Young’s modulus of 3.26 MPa, Poisson’s ratio
of 0.35, and coefficient of thermal expansion of −0.01 (1/◦ C). Conversely, the passive material has Young’s modulus
of 15.4 MPa, Poisson’s ratio of 0.35, and coefficient of thermal expansion of 0.0 (1/◦ C). These properties delineate
the mechanical and thermal characteristics of the materials used in the beam. For simplification and efficiency, both
materials were assigned elastic properties. We used two-state properties to simulate the performance of the printed
part, including the properties at low and high temperatures. The thermal expansion coefficient is calculated by the
difference in length at low and high temperatures. The Young’s modulus and the Poisson’s ratio are taken from the
properties at the high temperature because most of the deformation happens above the glass transition temperature.
For TPU, although the real thermal expansion coefficient is positive, the shrinkage of the part due to residual stress and
strain after 3D printing effectively compensates for the thermal expansion. This compensation results in a near-zero
effective thermal expansion coefficient. Therefore, for simplification, using a zero thermal expansion coefficient for
TPU in the simulation is a reasonable approximation to reflect the actual behavior observed in the printed samples.
For SMP, the real thermal expansion coefficient is also positive. However, the large amount of residual stress and
strain that exists post-3D printing leads to significant shrinkage of the part when the temperature increases and these
stresses and strains are released. To accurately represent this effect in our simulations, we regard the thermal expansion
coefficient as negative. This adjustment allows us to model the thermal behavior of SMP more accurately, reflecting
the observed shrinkage due to the release of residual stresses.
3.2. Data Preprocessing
The first step in preparing the data for the ResNet model is data preprocessing. This involves normalizing both the
input and output data to ensure they are in a suitable range for training. For the input data, which represents active (2),
passive (1), and no (0) allocation matrices, we apply a normalization equation: xn = x−1, where x is the input data and
xn is normalized input data. This converts the original values of 0, 1, and 2 to −1, 0, and 1 respectively, allowing for
better convergence during training. On the other hand, the output data is normalized based on the mean and standard
deviation. By using the equation yn = (y − ymean )/y std , where y is the output data, yn is the normalized output data, and
ymean and y std represents the mean and the standard deviation of all output data, respectively, we transform the output
data to have a mean of 0 and a standard deviation of 1, ensuring consistency and stability in the training process.
After normalizing the data, the next step is to split it into training and validation sets. The purpose of this split is
to evaluate the performance of the model on unseen data and prevent overfitting. We allocate 80% of the normalized
data to the training set and the remaining 20% to the validation set. This division ensures that the model is trained on
a significant amount of data while still having a separate set for evaluation.
3.3. Model Architecture
The ResNet model architecture is designed to capture complex patterns and features in the data. As demonstrated
in Figure 3, it consists of convolution layers, max-pooling, and residual blocks. The input layer starts with a convolu-
tion layer that has 4 filters and a (3, 3) kernel size, followed by max-pooling to downsample the data. The core of the
model consists of five residual blocks, each containing convolution layers with 8 filters. These blocks contribute to
the depth and complexity of the model, enabling it to extract intricate features from the data. Finally, the output layer
is a fully connected layer with a linear activation function. This layer shapes the model’s output to the desired form,
ready for prediction or classification.
7
3.4. Hyperparameters
To train the ResNet model effectively, several hyperparameters need to be specified. The learning rate is set to
0.0005, determining the step size for adjusting the model weights during optimization. A suitable learning rate ensures
that the model converges efficiently without overshooting or getting stuck in local minima. The batch size is set to 50,
which determines the number of samples processed in each mini-batch during training. A larger batch size can lead
to faster convergence, but it requires more memory. The max epochs parameter is set to 500, defining the maximum
number of training iterations. This parameter prevents the model from training indefinitely and helps control training
time. Additionally, a loop decay factor of 0.6 per 25 epochs is applied to decrease the learning rate gradually over
time, allowing the model to fine-tune its performance.
3.5. Training
Once the data and model architecture are prepared, the training process begins. The model is compiled using the
Adam optimizer, a popular optimization algorithm that adapts the learning rate for each parameter. Mean squared
error (MSE) is chosen as the loss function, which measures the dissimilarity between the predicted output and the
actual output. Additionally, early stopping is employed to monitor the validation loss. If the validation loss fails to
improve for a set number of consecutive epochs, the training process is halted to prevent overfitting. Throughout the
training process, the model learns patterns and relationships within the input data and its corresponding output, aiming
to make accurate predictions or classifications on new, unseen data.
4. Inverse Design Methods
Inverse optimization is employed to determine the optimal material allocation within the hierarchical architecture
to ensure that the resulting printed part exhibits predefined performance characteristics. In essence, it involves working
backward from the desired performance requirements to identify the material distribution that will best achieve those
objectives.
In this paper, we employ an evolutionary algorithm to perform inverse optimization for 4D printed hierarchical
architectures with arbitrary shapes. The evolutionary algorithm is a computational optimization technique inspired by
the process of natural selection and evolution in biology. It mimics the principles of natural evolution to search for
optimal solutions to complex problems iteratively. Unlike topology optimization, which often relies on gradient-based
methods and can get trapped in local optima [48–50], evolutionary algorithms perform a global search. This ability
to explore the entire solution space increases the likelihood of finding the true global optimum, especially in complex
and high-dimensional problems. At the core of the evolutionary algorithm is a population of candidate solutions,
represented as individuals or chromosomes. Each individual in the population encodes a potential solution to the
optimization problem. In our case, the individuals represent different configurations of the 4D printed hierarchical
architecture with arbitrary shapes.
The steps for the evolutionary algorithms are depicted in Figure 4. The algorithm begins by generating an initial
population comprising 5,000 individuals. Each individual represents a potential solution for the material allocation and
stimuli distribution in 4D printing. These solutions are randomly generated to provide a diverse range of configurations
for the 4D-printed structures. Random initialization helps explore a wide range of design possibilities and prevents
the algorithm from getting stuck in local optima.
Once the initial population is generated, each individual undergoes fitness evaluation using a fitness function. The
fitness function quantifies the performance of the 4D-printed structure based on the Root Mean Square Error (RMSE)
between the predicted and target coordinates. The RMSE serves as a measure of how well the material allocation
and stimuli distribution align with the desired deformation. Lower RMSE values indicate better alignment and higher
fitness. The fitness evaluation provides a numerical measure of each individual’s effectiveness in achieving the desired
deformation.
After fitness evaluation, the population is sorted based on the fitness scores. The top 10% of individuals, which
exhibit the highest fitness scores, are selected as elite individuals. These elite individuals possess favorable character-
istics in terms of material allocation and stimuli distribution that contribute to achieving the desired deformation. By
preserving the best-performing individuals, the algorithm ensures that promising solutions are carried over to the next
generation, increasing the chances of finding optimal or near-optimal solutions.
8
Initial Population Inverse Optimization Model
Forward Prediction Model (ResNet)
Fitness Function f (x)
Selection Crossover Mutation
Parents
Parents
Elite Randomly one column
Randomly select several to mutate
columns to exchange Children (change 0 to 1 and 1 to 0)
Children
New
Population
2: Active Material
Input: Displacement Output:
1: Passive Material
Target Deformation Curve / Field Optimized Material Allocation
0: No Material
Active Material
Passive Material
Figure 4: Inverse optimization model constructed by EA to optimize material allocation based on the predefined deformation.
9
The remaining 70% of the population is used for crossover operations. Crossover is a genetic operation that
combines genetic material from two parent individuals to create two offspring. In this implementation, a two-point
crossover is employed. Two random crossover columns, referred to as a and b, are selected. The genetic material
between these columns, which is circled in green in Figure 4, is exchanged between the parents, creating two children.
This process introduces diversity and potential improvements into the population by exploring different combinations
of genetic material. By combining genetic information from different individuals, crossover helps in the exploration
of the solution space and can lead to the discovery of novel and better solutions.
The remaining 20% of the population is allocated for mutation operations. Mutation introduces random changes
to the genetic material of individual members within the population. It helps introduce diversity and prevents the
algorithm from prematurely converging to suboptimal solutions. In this implementation, a random column in the
digital chromosome is selected, and the corresponding bit is flipped (from 2 to 1 or vice versa), as illustrated in Figure
4. This operation adds randomness to the population and aids in exploring the solution space by introducing small,
random variations to the genetic material.
The combination of elite individuals, offspring generated through crossover, and individuals subjected to muta-
tion forms the new population for the next generation. This new population represents the updated generation and
undergoes the same process of fitness evaluation, selection of elite individuals, crossover, and mutation. By iteratively
refining the population, the algorithm explores the design space and aims to converge towards optimal or near-optimal
solutions for the 4D printing design optimization problem.
The iterative process continues until a satisfactory level of convergence or desired performance is achieved. Con-
vergence refers to the point where the algorithm has reached a stable state, and further iterations do not significantly
improve the solutions. The stopping criteria for convergence can be predefined based on the desired level of per-
formance or a specific number of generations. Once the algorithm converges or achieves the desired performance,
the best solution(s) found in the final population can be selected as the optimal or near-optimal solutions for the 4D
printing design optimization problem.
5. Results and Discussions
In this section, we present the results of both the forward prediction and inverse optimization models developed
for the 4D-printed hierarchical architecture with arbitrary shape, as outlined in the previous sections.
Fixed Design Domain Random Design Domain
Active Material
Passive Material
Design Domain
Design Boundary
Design active and passive materials distribution Design active and passive materials distribution
within the fixed design domain with the changing design domain
within the design boundary
Figure 5: Two scenarios for inverse design of 4D printed hierarchical architecture with arbitrary shapes.
There are two scenarios for the inverse design of 4D printed hierarchical architecture with arbitrary shapes as
illustrated in Figure 5. The first scenario is the fixed design domain with material allocation. In this scenario, the
design domain remains fixed, which is defined within the green line in Figure 5, and the objective is to determine
the optimal allocation of active and passive materials within this design domain. The goal is to find the material
distribution that achieves the desired deformation or shape change when subjected to the appropriate stimuli. The
10
evolutionary algorithm is utilized to search for the optimal allocation of materials within the given design domain. By
iteratively evaluating and refining the population of material allocations, the algorithm aims to converge towards the
optimal or near-optimal solution that maximizes the desired deformation while satisfying any imposed constraints.
The second scenario is the random design domain with topology optimization. In the second scenario, the design
domain is not fixed, and each location within the structure can be allocated as no material, active material, or passive
material. This scenario resembles topology optimization, where the goal is to determine the optimal distribution of
material throughout the entire structure. The evolutionary algorithm is employed to explore the design space and find
the optimal material distribution that achieves the desired deformation or shape change. This scenario offers a higher
degree of design freedom, as each location can be assigned a specific material allocation based on certain constraints
and objectives. The algorithm iteratively evaluates and evolves the population of material distributions, aiming to
converge towards the optimal or near-optimal solution that maximizes the desired deformation while satisfying the
given constraints.
(a) Optimized Material Allocation for Parabola Curve
U, U2 (mm)
+5.030e+00
−4.289e+00
−1.361e+01
−2.293e+01
−3.225e+01 Active Material Passive Material No Material
−4.157e+01
−5.089e+01
−6.020e+01
−6.952e+01
−7.884e+01
−8.816e+01
−9.748e+01 Simulation of Optimized Results
−1.068e+02
(b) (c) (d)
RMSE (mm)
Generation
Figure 6: The results for the inverse design of 4D-printed hierarchical architecture with arbitrary shape (parabola case). (a). Optimized results
obtained from the evolutionary algorithms and corresponding simulation results of the parts with the optimized material allocation. (b). The
training loss and the validation loss for the forward prediction model. (c). The fitness for the best individual in each generation. (d). The
comparison between the target deformation and simulated deformation.
5.1. Fixed Design Domain
To demonstrate the effectiveness of the forward prediction and inverse optimization approach for 4D-printed hier-
archical architectures with arbitrary shapes in fixed design domain, we will present two examples.
In the first example, we consider a non-rectangular beam composed of two opposite triangles. The objective is
to design the material allocation within the beam structure so that it deforms to a predefined parabolic curve when
subjected to the appropriate stimuli. To achieve this, we utilize a ResNet (Residual Neural Network) for forward
prediction, which can accurately predict the deformation behavior of the beam for different material allocations.
11
We generate a dataset of 20,000 instances, each consisting of a different material allocation and the corresponding
resulting deformation. This dataset is used to train the ResNet model.
Once the ResNet model is trained, we employ an evolutionary algorithm for inverse optimization. The EA explores
the design space of material allocations within the non-rectangular beam to find the optimal or near-optimal solution
that deforms the beam to the predefined parabolic curve. The EA iteratively evaluates and evolves the population
of material allocations, utilizing the fitness function based on the RMSE between the predicted and target parabolic
curves. Through the iterative process of selection, crossover, and mutation, the EA converges towards an optimized
material allocation that achieves the desired deformation behavior.
The desired curve for this case is that the irregular-shaped beam will deform into a parabola with an end displace-
ment of −100 mm. The results for the parabola curve are shown in Figure 6. Figure 6(a) displays the optimized results
obtained from the evolutionary algorithms and the corresponding simulation results of the parts with the optimized
material allocation, which met our expectations and requirements. In Figure 6(b), the training loss and the validation
loss for the forward prediction model are depicted, indicating that the forward prediction model can predict the defor-
mation based on the digital material allocation very accurately. Figure 6(c) shows the fitness for the best individual
in each generation, demonstrating that the evolutionary algorithms can optimize material allocation to achieve the
desired configuration. Lastly, Figure 6(d) presents the comparison between the target deformation and the simulated
deformation, indicating that the inverse design we developed for 4D printed hierarchical architecture has performed
well.
In the second example, we aim to achieve the closing and opening motion of a complex non-rectangular hand
shape using 4D printing. Similar to the previous example, we utilize a ResNet for forward prediction to predict the
resulting deformation of the hand shape for different material allocations. A dataset of 20,000 instances is generated,
consisting of different material allocations and the corresponding hand-shape deformations.
The results for the hand shape are shown in Figure 7. In Figure 7(a), the training loss and the validation loss for the
forward prediction model are depicted, indicating that the forward prediction model can establish a robust relationship
between the material allocation and behavior, with loss reaching below 10−2 . Figure 7(b) shows the fitness for the best
individual in each generation, demonstrating the good performance of our forward prediction and inverse optimization
model. Figure 7(c) displays the optimized results obtained from the evolutionary algorithms and Figure 7(d) illustrates
corresponding simulation results of the parts with the optimized material allocation. Figure 7(e) and (g) give a real
hand demonstration of close and open hand movements. Finally, the results are validated by experiment as shown
in Figure 7(f) and (h). The simulation results and experimental results align with the target shape, showing that the
optimized result can replicate the movements demonstrated by the real hand.
From the above two examples, it’s evident that our design method exhibits good universality for 4D-printed hier-
archical architecture with irregular shapes, where the design domain is fixed. By employing the method of utilizing 0,
1, and 2 to represent different material allocations, we have effectively overcome the previous limitation where only
rectangular structures could be optimized using 1 and 2. This expansion of the application of machine learning in
forward prediction and inverse optimization has been successful in achieving satisfactory results.
5.2. Random Design Domain
In this subsection, we delve into the inverse design of 4D-printed hierarchical architecture with random design
domains. Despite the randomness of these design domains, certain restrictions must be adhered to to ensure the
structural integrity and functionality of the printed parts, which is explored in Figure 8. Firstly, we enforce the
constraint that there cannot be any ”isolated islands” within the structure, where an area is surrounded entirely by
empty space, which is circled with blue boxes in Figure 8. Such isolated regions serve no purpose and can compromise
the overall structural integrity. Secondly, we prohibit the existence of ”diagonal lands,” characterized by a 2 × 2 matrix
where one diagonal contains material while the other diagonal does not. This constraint is circled with yellow boxes
in the picture. This configuration introduces unpredictability and instability into the system, rendering it unsuitable
for practical applications. Finally, along the direction of the optimized curve, it is essential to have a continuous
”continent” connecting one end of the curve to the other, which is illustrated by the green line in Figure 8. The
green line represents the break between these two ”continents”. This constraint ensures the coherence of the part and
prevents it from being fragmented into disjointed segments, maintaining its functionality and usability. By adhering to
these constraints, we aim to optimize the design of 4D-printed hierarchical architectures with random design domains,
ensuring their reliability and effectiveness in real-world applications.
12
(a) (b)
(c) (d) U, U2 (mm)
Active Material +7.915e+00
+6.530e+00
Passive Material +5.146e+00
+3.761e+00
No Material +2.376e+00
Simulation +9.908e−01
−3.942e−01
−1.779e+00
Optimized −3.164e+00
Material Allocation for −4.549e+00
Hand Shape −5.934e+00
−7.319e+00
−8.704e+00
(e) Hand Closed (f) Printed Part (g) Hand Opened (h) Heated Part
1 cm 1 cm
Figure 7: The results for the inverse design of 4D-printed hierarchical architecture with arbitrary shape (hand shape). (a). The training loss and the
validation loss for the forward prediction model. (b). The fitness for the best individual in each generation. (c). Optimized results obtained from
the evolutionary algorithms. (d). Corresponding simulation results of the part with the optimized material allocation. (e). Demonstration of closed
hand. (f). Printed part using two-nozzle FDM machine. (g). Demonstration of open hand. (h). Shape morphing after heating.
Isolated Island
Diagonal Land
Discontinuous
Continent
Figure 8: The constraints of material allocation for the random design domain.
13
(a) Optimized Material Allocation for Parabola Curve
U, U2 (mm)
+7.555e+01
+6.889e+01
+6.224e+01
+5.558e+01
+4.892e+01 Active Materials Passive Materials No Materials
+4.227e+01
+3.561e+01
+2.895e+01 Simulation of Optimized Results
+2.230e+01
+1.564e+01
+8.984e+00
+2.328e+00
−4.328e+00
(b) (c) (d)
Figure 9: The results for the inverse design of 4D-printed hierarchical architecture with arbitrary shape for random design domain. (a). Optimized
results obtained from the evolutionary algorithms and corresponding simulation results of the parts with the optimized material allocation. (b).
The training loss and the validation loss for the forward prediction model. (c). The fitness for the best individual in each generation. (d). The
comparison between the target deformation and simulated deformation.
14
In order to verify the optimization process for 4D-printed hierarchical architectures with random irregular design
domains, we conduct the following example. Firstly, for each column in the design domain, we randomly select a
number between 1 and 4 to determine the quantity of materials to be filled. Starting from the centerline of the column,
we then distribute the selected number of materials evenly in both upward and downward directions. By repeating
this process for each column, we ensure that the three aforementioned constraints are not violated. Indeed, while
the method described provides a straightforward approach to address the constraints, there are numerous alternative
strategies that can be explored.
Based on this method, we conducted optimization of the 4D-printed hierarchical architecture with random design
domains. Our task is to distribute the materials, represented by 0, 1, and 2, within the design domain of 8 by 64 using
the aforementioned method. The objective is to design a component that, upon receiving external stimulation, will
deform into a parabolic shape with an end displacement of 80 mm. The outcomes are depicted in Figure 9. Figure
9(a) showcases the optimized results obtained from the genetic algorithms alongside the corresponding simulation
outcomes of the parts with the optimized material allocation. The optimized results obtained from the genetic algo-
rithms reveal a precisely tailored material allocation within the 4D-printed hierarchical architecture. This allocation
demonstrates a seamless integration of passive and active materials, resulting in a structure that exhibits the desired
shape-shifting behavior. The corresponding simulation results validate the effectiveness of the optimized material
distribution, showcasing the accurate realization of the intended design. Figure 9(b) illustrates the training loss and
validation loss for the forward prediction model. The training loss and validation loss for the forward prediction model
depict a robust learning process, with both losses converging to low values. This convergence indicates that the for-
ward prediction model has successfully captured the complex relationship between material allocation and structural
behavior. The consistently low losses affirm the model’s ability to accurately predict deformations based on digital
material inputs. Figure 9(c) represents the fitness for the best individual in each generation. The fitness plot for the
best individual in each generation showcases the evolutionary algorithm’s iterative optimization process. With each
generation, the RMSE steadily decreases, indicating progressive refinement of the material allocation strategy. Figure
9(d) provides a comparison between the target deformation and simulated deformation. The comparison between the
target deformation and simulated deformation highlights the efficacy of the inverse design methodology. The simu-
lated deformation closely matches the target deformation, demonstrating the accuracy and reliability of the designed
hierarchical architecture. This alignment between the desired and simulated outcomes validates the effectiveness of
the proposed approach in achieving the intended structural behavior.
Compared with the fixed design domain, the random design domain not only has a higher degree of freedom,
but can also include the fixed design domain. For instance, consider the two opposing triangles illustrated in Figure
6. If optimization is required for these triangles, the initial population in the evolutionary algorithm can be set to
comprise these two opposing triangles. Subsequently, these triangles can be inputted into the random design domain’s
forward prediction model for evaluation. The evaluations obtained are then used for selection, crossover, and mutation
operations. Over successive generations, the shapes of each generation of offspring will be two opposing triangles. In
other words, for random design domain algorithms, any shape within the design boundaries can be forward predicted
and inverse optimized.
5.3. Discussion
The hierarchical structure studied in our research presents a design space smaller than 1,000 units, which is
modest compared to many topology optimization (TO) problems in the literature that involve significantly larger
design spaces, often exceeding 100 by 100 dimensions. The primary challenge in such large-scale optimization
problems is the computational intensity and the time required for each inverse optimization process. However, our
method is designed to be scalable. The hierarchical approach we use allows us to break down the larger problem into
smaller, more manageable sub-problems. This decomposition not only simplifies the complexity of each optimization
task but also enables parallel processing, which significantly enhances computational efficiency. By distributing the
computational load across multiple CPU cores, we can perform numerous simulations concurrently, thereby reducing
the overall time required for optimization.
The optimization outcomes presented in Figures 6, 7, and 9 demonstrate that the shapes activated through our
approach closely align with the intended target shapes at a macroscopic level. However, a closer inspection at the
microscopic level reveals that the surfaces of the deformed materials do not exhibit the same smoothness as the
target shapes. This discrepancy primarily arises due to the size of the pixels used in the design. Larger pixels can
15
result in less smooth curves, giving the surface a rough appearance. This issue could potentially pose a problem in
practical applications where surface smoothness is critical. To address the observed unsmoothness, we propose a
couple of strategies. First, by decreasing the size of the pixels used in the design, we can achieve finer resolution
and smoother curves. This approach involves increasing the computational resolution of the design process, which,
although computationally more intensive, would result in a more accurate representation of the target shapes. Smaller
pixels allow for more detailed and precise control over the material distribution, thereby enhancing the smoothness
of the deformed surfaces. Additionally, the surface quality can also be enhanced by optimizing the parameters of the
FDM printing process. Adjustments such as layer height, extrusion width, and print speed can significantly impact
the surface finish. For instance, reducing the layer height can lead to finer layers being deposited, which smoothens
the surface. Similarly, optimizing the extrusion width and print speed can help in achieving better layer adhesion and
reducing surface roughness. By fine-tuning these parameters, it is possible to produce smoother surfaces even with
the existing design resolutions.
6. Conclusions and Future Works
In this paper, we have presented a novel approach for forward prediction and inverse design tailored to 4D printed
hierarchical architecture with arbitrary shapes. By integrating machine learning techniques, specifically Residual
Networks for forward prediction and evolutionary algorithms for inverse design, our methodology addresses the chal-
lenges posed by non-rectangular shapes in traditional design workflows. Our approach enables the prediction of
complex hierarchical structures’ behavior and the optimization of material allocation to achieve desired shape trans-
formations efficiently. Through empirical validation and case studies, we have demonstrated the effectiveness and
versatility of our proposed framework in accurately predicting and designing 4D printed structures with arbitrary
shapes, paving the way for advancements in 4D printing technology.
This research contributes to the broader field of 4D printing design optimization by offering a data-driven approach
that leverages machine learning to streamline the design process. By encoding non-rectangular shapes with unique
identifiers and employing ResNet for forward prediction, our methodology overcomes the limitations of traditional
design methods and enhances the efficiency and accuracy of shape prediction. Furthermore, the use of evolutionary
algorithms for inverse design enables the optimization of material allocation for achieving predefined performance
criteria, facilitating the creation of adaptive and functional structures. Overall, our work underscores the potential of
machine learning-driven approaches in advancing the capabilities of 4D printing technology, opening new avenues for
designing complex and dynamic structures for various applications across industries.
The future work would be to extend the design and optimization methods from 2D surfaces to 3D plates or objects.
While the current focus has been on optimizing material allocation for 2D surfaces, moving towards 3D structures
would introduce additional challenges and complexities. Designing 3D plates or objects with specific deformation
behaviors requires considering not only the material allocation in the surface plane but also the material distribution
in the depth dimension. This would involve optimizing the material properties and arrangement in three dimensions
to achieve desired deformation characteristics.
Acknowledgements
W. L. acknowledges Research Grants Council (C4074-22G), Hong Kong Special Administrative Region, China,
and The Chinese University of Hong Kong (Project ID: 3110174). Q. G. acknowledges the National Natural Sci-
ence Foundation of China (No. 12072142), the Key Talent Recruitment Program of Guangdong Province (No.
2019QN01Z438), and the support by the Science, Technology and Innovation Commission of Shenzhen Munici-
pality under grant no. ZDSYS20210623092005017. M. B. acknowledges the support by the UK Engineering and
Physical Sciences Research Council (EPSRC) (grant number EP/Y011457/1). For the purpose of open access, the
author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version
arising from this submission.
Data Availability Statement
The data supporting the findings of this study are available within the article.
16
References
[1] X. Zhai, L. Jin, J. Jiang, A survey of additive manufacturing reviews, Materials Science in Additive Manufacturing 1 (4) (2022) 21.
[2] A. Mahmood, T. Akram, C. Shenggui, H. Chen, Revolutionizing manufacturing: a review of 4D printing materials, stimuli, and cutting-edge
applications, Composites Part B: Engineering (2023) 110952.
[3] B. M. B. dos Santos, G. Littlefair, S. Singamneni, From 3D to 4D printing: A review, Materials Today: Proceedings (2023).
[4] G. Adam, A. Benouhiba, K. Rabenorosoa, C. Clévy, D. J. Cappelleri, 4D printing: Enabling technology for microrobotics applications,
Advanced Intelligent Systems 3 (5) (2021) 2000216.
[5] L. Jin, X. Zhai, K. Wang, K. Zhang, D. Wu, A. Nazir, J. Jiang, W.-H. Liao, Big data, machine learning, and digital twin assisted additive
manufacturing: a review, Materials & Design (2024) 113086.
[6] G. A. Pacillo, G. Ranocchiai, F. Loccarini, M. Fagone, Additive manufacturing in construction: A review on technologies, processes, materi-
als, and their applications of 3D and 4D printing, Material Design & Processing Communications 3 (5) (2021) e253.
[7] M. C. Biswas, S. Chakraborty, A. Bhattacharjee, Z. Mohammed, 4D printing of shape memory materials for textiles: Mechanism, mathemat-
ical modeling, and challenges, Advanced Functional Materials 31 (19) (2021) 2100257.
[8] J. P. Manaia, F. Cerejo, J. Duarte, Revolutionising textile manufacturing: a comprehensive review on 3D and 4D printing technologies,
Fashion and Textiles 10 (1) (2023) 20.
[9] A. Raina, M. I. U. Haq, M. Javaid, S. Rab, A. Haleem, 4D printing for automotive industry applications, Journal of The Institution of
Engineers (India): Series D (2021) 1–9.
[10] B. Zhang, H. Li, J. Cheng, H. Ye, A. H. Sakhaei, C. Yuan, P. Rao, Y.-F. Zhang, Z. Chen, R. Wang, et al., Mechanically robust and UV-curable
shape-memory polymers for digital light processing based 4D printing, Advanced Materials 33 (27) (2021) 2101298.
[11] Y. Wang, H. Cui, T. Esworthy, D. Mei, Y. Wang, L. G. Zhang, Emerging 4D printing strategies for next-generation tissue regeneration and
medical devices, Advanced Materials 34 (20) (2022) 2109198.
[12] Q. Ge, Z. Chen, J. Cheng, B. Zhang, Y.-F. Zhang, H. Li, X. He, C. Yuan, J. Liu, S. Magdassi, et al., 3D printing of highly stretchable hydrogel
with diverse UV curable polymers, Science Advances 7 (2) (2021) eaba4261.
[13] L. Zeenat, A. Zolfagharian, Y. Sriya, S. Sasikumar, M. Bodaghi, F. Pati, 4D printing for vascular tissue engineering: Progress and challenges,
Advanced Materials Technologies 8 (23) (2023) 2300200.
[14] A. Aufa, Z. Ismail, M. Z. Hassan, Emerging trends in 4D printing of hydrogels in the biomedical field: A review, Materials Today: Proceed-
ings (2023).
[15] T. Gu, T. Ji, H. Bi, K. Ding, H. Sun, W. Zhai, Z. Ren, Y. Wei, M. Xu, 4D printed and multi-stimulus responsive shape memory polymer
nanocomposites developed on hydrogen bonding–metal-phenolic sacrificial network: Application for hazardous chemical operations soft
robots, Applied Materials Today 35 (2023) 102009.
[16] L. Luo, F. Zhang, L. Wang, Y. Liu, J. Leng, Recent advances in shape memory polymers: Multifunctional materials, multiscale structures,
and applications, Advanced Functional Materials 34 (14) (2024) 2312036.
[17] S. Yan, F. Zhang, L. Luo, L. Wang, Y. Liu, J. Leng, Shape memory polymer composites: 4D printing, smart structures, and applications,
Research 6 (2023) 0234.
[18] W. Qiu, X. He, Z. Fang, Y. Wang, K. Dong, G. Zhang, X. Xu, Q. Ge, Y. Xiong, Shape-tunable 4D printing of LCEs via cooling rate
modulation: Stimulus-free locking of actuated state at room temperature, ACS Applied Materials & Interfaces 15 (40) (2023) 47509–47519.
[19] Y. Wang, H. Ye, J. He, Q. Ge, Y. Xiong, Electrothermally controlled origami fabricated by 4D printing of continuous fiber-reinforced
composites, Nature Communications 15 (1) (2024) 2322.
[20] F. Wang, F. Luo, Y. Huang, X. Cao, C. Yuan, 4D printing via multispeed fused deposition modeling, Advanced Materials Technologies 8 (2)
(2023) 2201383.
[21] I. Akbar, M. El Hadrouz, M. El Mansori, M. Tarfaoui, Thermomechanical shape memory testing of 4D printed novel material rhombus-shape
structure, Applied Materials Today 33 (2023) 101876.
[22] B. C. Kholkhoev, K. N. Bardakova, A. N. Nikishina, Z. A. Matveev, Y. M. Efremov, A. A. Frolova, A. A. Akovantseva, E. N. Gorenskaia,
N. A. Verlov, P. S. Timashev, et al., 4D-printing of mechanically durable high-temperature shape memory polymer with good irradiation
resistance, Applied Materials Today 36 (2024) 102022.
[23] E. Tekay, B. Aybakan, V. U. Aslan, T. Orhun, 4D Printing of novel poly (ethylene-vinyl acetate)/poly (butyl methacrylate-co-isobutyl
methacrylate) shape memory polymer blends, Applied Materials Today 38 (2024) 102199.
[24] K. Zhang, Q. Gao, J. Jiang, M. Chan, X. Zhai, L. Jin, J. Zhang, J. Li, W.-H. Liao, High energy dissipation and self-healing auxetic foam by
integrating shear thickening gel, Composites Science and Technology (2024) 110475.
[25] L. Zhang, X. Huang, T. Cole, H. Lu, J. Hang, W. Li, S.-Y. Tang, C. Boyer, T. P. Davis, R. Qiao, 3D-printed liquid metal polymer composites
as NIR-responsive 4D printing soft robot, Nature Communications 14 (1) (2023) 7815.
[26] C. Deng, Y. Liu, X. Fan, B. Jiao, Z. Zhang, M. Zhang, F. Chen, H. Gao, L. Deng, W. Xiong, Femtosecond laser 4D printing of light-driven
intelligent micromachines, Advanced Functional Materials 33 (11) (2023) 2211473.
[27] K. Kim, Y. Guo, J. Bae, S. Choi, H. Y. Song, S. Park, K. Hyun, S.-K. Ahn, 4D printing of hygroscopic liquid crystal elastomer actuators,
Small 17 (23) (2021) 2100910.
[28] J. A. Sol, L. G. Smits, A. P. Schenning, M. G. Debije, Direct ink writing of 4D structural colors, Advanced Functional Materials 32 (30)
(2022) 2201766.
[29] M. Nadgorny, Z. Xiao, C. Chen, L. A. Connal, Three-dimensional printing of pH-responsive and functional polymers on an affordable desktop
printer, ACS Applied Materials & Interfaces 8 (42) (2016) 28946–28954.
[30] B. Narupai, P. T. Smith, A. Nelson, 4D printing of multi-stimuli responsive protein-based hydrogels for autonomous shape transformations,
Advanced Functional Materials 31 (23) (2021) 2011012.
[31] J. Simińska-Stanny, M. Nizioł, P. Szymczyk-Ziółkowska, M. Brożyna, A. Junka, A. Shavandi, D. Podstawczyk, 4D printing of patterned
multimaterial magnetic hydrogel actuators, Additive Manufacturing 49 (2022) 102506.
17
[32] E. R. Espı́ndola-Pérez, J. Campo, C. Sánchez-Somolinos, Multimodal and multistimuli 4D-printed magnetic composite liquid crystal elas-
tomer actuators, ACS Applied Materials & Interfaces 16 (2) (2023) 2704–2715.
[33] A. Y. Chen, E. Pegg, A. Chen, Z. Jin, G. X. Gu, 4D printing of electroactive materials, Advanced Intelligent Systems 3 (12) (2021) 2100019.
[34] H. Zhu, Y. He, Y. Wang, Y. Zhao, C. Jiang, Mechanically-guided 4D printing of magnetoresponsive soft materials across different length
scale, Advanced Intelligent Systems 4 (3) (2022) 2100137.
[35] Z. Ding, C. Yuan, X. Peng, T. Wang, H. J. Qi, M. L. Dunn, Direct 4D printing via active composite materials, Science Advances 3 (4) (2017)
e1602890.
[36] D. Jin, Q. Chen, T.-Y. Huang, J. Huang, L. Zhang, H. Duan, Four-dimensional direct laser writing of reconfigurable compound microma-
chines, Materials Today 32 (2020) 19–25.
[37] B. B. Samal, A. Jena, S. K. Varshney, C. S. Kumar, FDM 4D printing: A low-cost approach of shape programming and assessing the shape
memory properties using angle measurement methods in hot water actuation testing apparatus, Materials Today: Proceedings (2023).
[38] S. M. D. Tezerjani, M. S. Yazdi, M. H. Hosseinzadeh, The effect of 3D printing parameters on the shape memory properties of 4D printed
polylactic acid circular disks: An experimental investigation and parameters optimization, Materials Today Communications 33 (2022)
104262.
[39] X. Peng, G. Liu, J. Wang, J. Li, H. Wu, S. Jiang, B. Yi, Controllable deformation design for 4D-printed active composite structure: optimiza-
tion, simulation, and experimental verification, Composites Science and Technology 243 (2023) 110265.
[40] C. M. Hamel, D. J. Roach, K. N. Long, F. Demoly, M. L. Dunn, H. J. Qi, Machine-learning based design of active composite structures for
4D printing, Smart Materials and Structures 28 (6) (2019) 065005.
[41] D. Athinarayanarao, R. Prodhon, D. Chamoret, H. J. Qi, M. Bodaghi, J.-C. André, F. Demoly, Computational design for 4D printing of
topology optimized multi-material active composites, npj Computational Materials 9 (1) (2023) 1.
[42] X. Sun, L. Yue, L. Yu, H. Shao, X. Peng, K. Zhou, F. Demoly, R. Zhao, H. J. Qi, Machine learning-evolutionary algorithm enabled design for
4D-printed active composite structures, Advanced Functional Materials 32 (10) (2022) 2109805.
[43] X. Sun, L. Yu, L. Yue, K. Zhou, F. Demoly, R. R. Zhao, H. J. Qi, Machine learning and sequential subdomain optimization for ultrafast
inverse design of 4D-printed active composite structures, Journal of the Mechanics and Physics of Solids (2024) 105561.
[44] X. Sun, L. Yue, L. Yu, C. T. Forte, C. D. Armstrong, K. Zhou, F. Demoly, R. R. Zhao, H. J. Qi, Machine learning-enabled forward prediction
and inverse design of 4D-printed active plates, Nature Communications 15 (1) (2024) 5509.
[45] L. Jin, X. Zhai, J. Jiang, K. Zhang, W.-H. Liao, Optimizing stimuli-based 4D printed structures: a paradigm shift in programmable material
response, in: Sensors and Smart Structures Technologies for Civil, Mechanical, and Aerospace Systems 2024, Vol. 12949, SPIE, 2024, pp.
321–332.
[46] B. Zhao, M. Zhang, L. Dong, D. Wang, Design of grayscale digital light processing 3D printing block by machine learning and evolutionary
algorithm, Composites Communications 36 (2022) 101395.
[47] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, 2016, pp. 770–778.
[48] Y. Wei, P. Huang, Z. Li, P. Wang, X. Feng, Design of active materials distributions for four-dimensional printing based on multi-material
topology optimization, Smart Materials and Structures 30 (9) (2021) 095002.
[49] X. Zhai, Y. Gai, L. Jin, W.-H. Liao, F. Chen, P. Hu, Isogeometric topology optimization of auxetic materials based on moving morphable
components method, Advanced Topics in Mechanics of Materials, Structures and Construction: AToMech1-2023 31 (2023) 172.
[50] A. Zolfagharian, M. Denk, M. Bodaghi, A. Z. Kouzani, A. Kaynak, Topology-optimized 4D printing of a soft actuator, Acta Mechanica
Solida Sinica 33 (2020) 418–430.
18