Page 27 - IJAMD-2-3

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

region, as shown in Figure 11B. However, challenges remain on the use of a VAE combined with Bayesian neural
in detecting more subtle defects, such as loose tows and networks (BNN) to quantify damage in composite plates.
twists, where boundary detection becomes more difficult This semi-supervised learning model helps with structural
due to minimal height variations or uneven distribution health monitoring by effectively handling the uncertainty
of defects. One of the reasons for this limitation is that and measurement errors prevalent in damage detection.
CNNs focus on learning local regions, making it difficult The VAE-BNN model achieved high accuracy for several
to capture global relationships. To address this limitation, damage types, such as cracks and delamination, offering
many studies have explored the use of other DL models. improved robustness against noise compared to traditional
methods.
For instance, Liu et al. integrated a transformer-based
144
neural network with a CNN for automatic delamination These studies exemplify the growing impact of
detection in composite curved structures. This automation models in improving quality control in
transformer-based model can learn correlations among composite material manufacturing. The presence of
elements within the entire image through a self-attention defects, such as voids or wrinkles, can significantly
mechanism, enabling the detection of defects with high compromise the mechanical properties of composite parts.
accuracy by considering patterns and morphological By automating defect detection and enhancing defect
information. Moreover, by utilizing enhanced terahertz characterization, ML/DL technologies are transforming
time-domain spectroscopy signals for non-destructive the production of composite materials, ensuring better
testing and imaging of hidden delamination defects, quality, reducing production times, and achieving cost
the transformer-based approach demonstrated superior savings in the long run. Moreover, the integration of
accuracy and generalization performance compared to ML/DL allows for continuous monitoring and real-time
traditional models, highlighting its potential for real- adjustments to further enhance the precision and efficiency
time defect detection in complex composite structures. of the production process.
This method could be particularly beneficial in industries 4.2. Machine learning/deep learning-driven process
such as aerospace, where structural integrity is critical. In automation
addition, the potential for ML/DL-driven automation in
the quality control of the RTM process was demonstrated The production of composite materials has traditionally
by Wang et al. The PixelRNN model, an image-based relied on the expertise of skilled professionals. While the
142
neural network architecture, was used to predict mold work of these skilled experts remains an effective method
filling patterns with high accuracy given the resin for producing composite materials, the manufacturing
injection location, as shown in Figure 11C. This model was process is highly dependent on external environments
particularly effective in handling nonlinear filling patterns and conditions and susceptible to human error, requiring
and significantly reduced computational costs compared advanced technical skills and labor-intensive efforts. In
to traditional physics-based simulations. This highlights addition, some processes may generate harmful chemicals
the increasing role of ML/DL in process monitoring and that might have negative impacts on the health of workers.
automatic detection of dry spots in resin impregnation Due to these limitations, there has been an
processes. increasing interest in automation in composite material
In addition, as described in Sections 3.2 and 3.3, manufacturing. However, simple mechanical automation
generative models, such as GANs, VAEs, and diffusion is insufficient considering the complex shapes and variable
118
126
models, were also used to detect and classify defects manufacturing conditions associated with composite
131
in composite materials. In contrast to the methods of materials. In other words, advanced AI-based automation
defect detection, which require data from both normal technology with the flexibility to adapt to diverse shapes
and defective samples, Szarski and Chauhan developed and complex conditions is essential for composite material
146
a model based on NFs to obtain the image likelihood of manufacturing.
a new sample using the knowledge learned from input These innovations are particularly important in
images of only normal samples. Considering the rarity manufacturing processes, such as AFP, RTM, and composite
and variety of defects, collecting training data is expensive curing, where high precision and efficiency are critical. 127-129
and time-consuming, and supervised learning proves to be A hybrid framework combining spatial-temporal analysis,
difficult. Instead of detecting the presence of defects, the thermography, and ML algorithms was recently developed
model determines whether the sample is different from a for process monitoring and defect detection in AFP. In
147
set of normal samples, without requiring any data from this framework, an extensive thermal image database was
defective samples. Furthermore, Zhang et al. focused used to train the model, achieving an impressive accuracy
118

Volume 2 Issue 3 (2025) 21 doi: 10.36922/IJAMD025210016