Page 21 - IJAMD-2-3

International Journal of AI for
Materials and Design AI applications in composite materials

space, and the decoder reconstructs the original data from demonstrated high performance even with a small dataset.
the input data. Specifically, to prevent the model from In addition, in a β-VAE, a hyperparameter β is introduced
merely replicating the input, VAEs employ a probabilistic to adjust the weight of the KL divergence term in the
approach by encoding the input into the parameters of a standard VAE, thereby promoting statistical independence
Gaussian distribution (i.e., mean and standard deviation), among latent variables. Jiang et al. predicted the residual
120
and reconstructing the data by minimizing a loss function stiffness degradation caused by fatigue damage in CMCs
composed of a reconstruction loss and a Kullback-Leibler with the aid of β-VAE. By combining β-VAE with neural
(KL) divergence term. As a result, VAEs are highly useful ordinary differential equations, they captured the latent
for generating, transforming, and compressing data, as dynamics of fatigue damage progression and reconstructed
well as creating new samples. the full stiffness degradation curves over fatigue cycles.
VAEs can serve as a powerful tool in the field of composite Moreover, their model enabled accurate predictions even
materials to address data scarcity issues and generate new under new loading conditions through latent variable
designs. 115-120 For instance, a VAE-based approach was interpolation and partial data retraining.
used for generating and optimizing composite layups with These studies illustrate the versatility of VAEs in
specified properties. This method allows for increased enhancing the design and performance of composite
117
flexibility in designing complex layup structures and materials. By enabling the generation of complex
outperforms traditional optimization methods, providing material configurations and reducing the reliance on
faster results and expanding the design space. extensive experimental datasets, VAEs are anticipated to
Moreover, as shown in Figure 8B, VAEs were applied revolutionize the design and optimization of composite
to the design and material selection process in the structures in various industries, including aerospace,
development of thermoset shape memory polymers automotive, and civil engineering.
(SMPs) with high glass transition temperatures. Teimouri
and Li applied a VAE-based approach coupled with 3.2. Generative adversarial networks (GANs)
115
transfer learning to explore the vast chemical design space VAEs are powerful tools for data generation, but the
of SMPs. Their model optimized the glass temperature quality of the generated data is relatively low, particularly
of SMPs, which was experimentally verified to be highly for applications that require high-resolution images or
accurate, thus enhancing the material development process complex patterns. They also struggle to capture fine details
by reducing time and resource consumption. and intricate variations in real-world data, often leading to
Furthermore, as presented in Figure 8C, VAEs have reconstruction errors, especially with high-dimensional
also been utilized in the design of metamaterial arrays or complex structures. On the other hand, GANs can
for impact protection. To address the high computational address these limitations by implicitly estimating the
cost of traditional optimization methods, such as genetic probability distribution, flexibly learning highly complex
algorithms and Bayesian optimization, Wang et al. and nonlinear data distributions in high-dimensional
116
employed VAEs to generate unit cell designs and graded spaces. GANs employ two competing neural networks: a
arrays with desired performance. The generated graded generator, which creates data that mimics real samples,
metamaterial demonstrated a 97.25% similarity to and a discriminator, which distinguishes real data from
121
the target performance and significantly delayed the generated data. The structure of a GAN is depicted
transmission of impact energy through the material in Figure 9A. Through this competitive process, both
compared to conventional designs. networks progressively improve, refining the quality of
the generated data. Composite material analysis involves
However, conventional VAEs rely solely on a
continuous Gaussian latent space and focus on learning capturing details of the intricate microstructures and
variations from data consisting of high-resolution
static distributions, which makes it difficult to represent images and complex patterns. Consequently, GANs can
the complex mechanisms of composite materials and be a particularly powerful tool in the field of composite
limits their ability to predict dynamic changes. Therefore, materials.
research is ongoing to apply extended VAE models to
resolve these challenges. For instance, the vector quantized- In recent studies, GANs have been used to synthesize
VAE (VQ-VAE) constructs a discrete latent space, allowing realistic 2D microstructures of unidirectional fiber-
the generation of new data through diverse combinations reinforced composites. Guo et al. applied a deep
123
and thus enabling more reliable results. Using VQ-VAE, convolutional GAN to generate 2D transverse
Wang et al. predicted the 3D damage field of composite microstructures of fiber reinforced composites. Their
119
laminates under various low-velocity impact conditions and approach captured key microstructural features, such

Volume 2 Issue 3 (2025) 15 doi: 10.36922/IJAMD025210016