Page 11 - IJAMD-2-3

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

permeability prediction, in which statistical RVEs with and material behavior, making CNNs essential when
randomized fiber distributions were generated specifically considering image-based composite material problems.
for training DNNs to address the stochastic nature of CNNs have been successfully applied to various aspects
fibrous microstructures. 35 of composite material analysis. 45,46,50-52 Key characteristics,
These examples highlight the significant impact of such as the fiber arrangement, fiber diameter, and resin
DNNs on the research of composite materials. DNNs enable content, can be extracted from microstructure images using
accurate predictions and efficient calculations, capturing CNNs, which can then be used to predict the mechanical
complex structural responses that are challenging to model behavior and failure mechanisms of composite materials.
with traditional methods. By learning from large datasets, As shown in Figure 4B, CNN-based models have been
DNNs provide a reliable alternative or complement employed to predict the stacking angles of fiber-reinforced
to numerical methods, such as FEA, offering data- composites from cross-sectional images. Extracting
45
driven insights and improved speed and precision when composite stacking information is particularly valuable for
optimizing composite materials and structures. However, the reverse-engineering of composite material structures,
DNNs face significant limitations when there is insufficient as the performance of fiber-reinforced composites is highly
data or when the model encounters conditions different sensitive to stacking angles.
from those in the training dataset. In such cases, transfer U-Net architectures, which are built on CNNs, have
learning techniques serve as a powerful tool to overcome been used to predict stress fields within composite
these challenges. 36-38 By applying the knowledge (such as microstructures by using stress maps generated from FEA
network weights and feature representations) obtained as training data. U-Net models provide a computationally
50
from a pre-trained model to a new related problem, transfer efficient alternative to traditional simulations, reducing
learning accelerates training, improves performance, and the computational burden of multi-scale modeling while
enables effective learning even with limited data. For maintaining high levels of predictive accuracy.
example, transfer learning has been applied to predict
the behavior of composite pressure vessels by combining In addition to stress field and stacking angle predictions,
analytical and numerical data, significantly enhancing CNNs have been applied to estimate the mechanical
both accuracy and computational efficiency. As presented properties of composite materials, such as the transverse
32
in Figure 3E, the approach involves pre-training on a large- modulus, tensile strength, and fracture toughness, based
scale analytical dataset with relatively low fidelity but low on microstructural images using finite element simulation
51
cost, and then fine-tuning with a smaller-scale numerical results as training data. By training on a large dataset of
dataset that offers higher fidelity but at a higher cost. This RVEs, these models can capture complex material behaviors
strategy enables the rapid assessment of structural integrity and improve the accuracy of property estimation.
under various loading conditions, reducing the need for Moreover, as shown in Figure 4C, CNN-based
time-consuming finite element simulations. approaches have been successfully implemented for
efficient prediction of the 3D permeability of fibrous
2.2. Convolutional neural networks (CNNs)
microstructures by integrating 2D image-based learning
A CNN is a type of DL model specifically designed for with circuit analogy models. The permeability is a crucial
46
image processing and analysis. 39-41 Unlike general DNNs, factor in composite material manufacturing since it affects
CNNs are highly effective at handling image-based data, the ability of the fibrous reinforcement to achieve complete
making them particularly advantageous for visual data- impregnation during the resin infusion process. The use
related problems. As illustrated in Figure 4A, CNNs use of CNNs not only accelerates the permeability prediction
convolutional and pooling operations to efficiently extract process but also enhances the overall understanding of
important features from input data, reducing the need for fluid flow in porous composite media.
manual feature engineering. CNNs have demonstrated These studies highlight the transformative potential
exceptional performance in various vision tasks, of CNNs in composite materials research, offering new
including image classification, object detection, and facial possibilities for advanced composite structures through
recognition. 42-44 enhanced image-based analysis and prediction. However,
In the field of composite materials, microstructure despite the effectiveness of CNNs, one limitation is the
analysis and fiber orientation prediction are often difficult challenge of acquiring sufficient training data. Specifically,
to conduct using conventional analytical methods or collecting high-quality images of composite materials is
general DNNs. 48,49 Taking advantage of the superior pattern difficult, which limits the model’s learning process. In addition,
recognition capabilities of CNNs, image data of composite CNN-based approaches require extensive pre-processing of
materials can be utilized to predict microstructures image data, increasing the complexity of implementation.

Volume 2 Issue 3 (2025) 5 doi: 10.36922/IJAMD025210016