Page 23 - IJAMD-2-3

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

2D periodic structures, offering new avenues for material eliminating the noise to obtain the original image, as
design with tailored mechanical properties. shown in Figure 10A. This process is based on a Markov
Furthermore, a GAN model, referred to as IRT-GAN, chain, where each step refers only to the previous one,
was developed for automated defect detection in composite enabling stable training through the repeated learning
materials using IRT. By utilizing simulated large- of small changes. As a relatively recent and advanced
126
scale numerical datasets with various defects, IRT-GAN DL approach, diffusion models demonstrate high
was trained to generate segmentation images of defects performance in generating realistic 2D and 3D composite
automatically. This method demonstrated improved microstructures by learning from large datasets of actual
128,129
detection performance for both glass fiber- and CFRPs, material structures. As shown in Figure 10B, Lyu and
128
highlighting the ability of GANs to enhance quality control Ren accurately reconstructed various complex 2D and
in composite manufacturing processes. 3D microstructures of composite materials, such as random
textures and chessboard structures, using a diffusion
Conventional GANs generate data solely based on model. The generated images showed distributions highly
random noise without control over the output. On the other consistent with the original structures when evaluated
hand, conditional GANs (cGANs) introduce additional using indicators, such as the two-point correlation
conditions to enable the generation of goal-oriented function, lineal-path function, and Fourier descriptor.
data tailored to specific conditions. Yang et al. utilized Furthermore, conditional generation of 3D structures
122
cGANs to enhance stress and strain field predictions, corresponding to target permeability ranges was also
as conventional GANs have been shown to generate conducted, demonstrating the feasibility of performance-
arbitrary outputs following simple data distributions. As driven inverse material design. As illustrated in Figure 10C,
shown in Figure 9B, they accurately predicted both global Bastek and Kochmann proposed a diffusion model-based
130
mechanical properties (e.g., stiffness and resilience) and generative framework to design metamaterial structures
localized stress concentration phenomena corresponding that satisfy target stress-strain responses with nonlinear
to specific composite microstructures and loading mechanical behavior. They employed a pixel-based 2D
conditions, achieving computation times of less than one microstructural representation as the design parameter,
second. Similarly, as shown in Figure 9C, Gupta et al. enabling the expression of nonlinear physical phenomena,
110
developed a CNN-based GAN model to predict principal such as buckling, internal contact, and plastic deformation.
stress distributions by incorporating factors such as the After training, the model was evaluated on previously
microstructure of chopped carbon fiber epoxy composites, unseen target responses, achieving a normalized root mean
material properties, and time-dependent loading square error as low as 1.5%, thereby demonstrating its high
conditions. These studies present cGAN models as strong accuracy and effectiveness for inverse design applications.
candidates for generating and predicting the physical
131
behavior of composites based on their microstructure. A diffusion-based model is also used by Huang et al.
for data amplification for non-destructive structural
Overall, the integration of GANs into composite
material design, optimization, and defect detection health monitoring of CFRP composite plates using Lamb
wave damage signals. It is demonstrated that the diffusion
presents numerous advantages, including enhanced model outperformed VAE and GAN methods in terms of
accuracy in microstructure modeling, multi-objective the diversity and quality of the generated samples. With
material optimization, and quality inspection. As these
technologies continue to evolve, they hold the potential ongoing advancements in ML/DL, diffusion models are
to revolutionize the development and manufacture of poised to play a crucial role in the future of composite
composite materials, leading to higher performance, material design and optimization.
reduced costs, and improved efficiency in industries such Normalizing flows (NFs) are a type of generative
as automotive, aerospace, and civil engineering. model that consists of a chain of parameterized invertible
mappings to obtain the likelihood of a new sample by
3.3. Other generative models transforming the input probability distribution into a
In addition to VAEs and GANs, diffusion-based deep well-defined probability distribution, such as the normal
generative models are being increasingly explored as distribution. NFs perform sampling, reconstruction,
generative models in the field of composite material and probability density estimation stably through their
127
research. Diffusion models consist of a forward phase invertible structure, which leads to fewer unstable
in which noise is gradually applied to the original image training issues compared to GANs and reduces blurry
until it resembles Gaussian noise, and the reverse phase samples often seen in VAEs. Stochastic inverse design of
in which the model learns the process of progressively microstructures for woven CMCs with tailored anisotropic

Volume 2 Issue 3 (2025) 17 doi: 10.36922/IJAMD025210016