Page 9 - IJAMD-2-3

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

breakthroughs in DNN research and applications. Because
of their deep architecture, DNNs excel at capturing intricate
patterns and relationships, making them particularly
effective in handling nonlinear data and solving complex
problems. When trained with sufficient, high-quality data,
highly accurate predictions can be made for the given
inputs, making it a powerful tool for data-driven analysis
in various domains.
With the advancement of DNN technology, its
applications in the field of composite materials have been
steadily expanding. Traditional analytical methods often
rely on various assumptions and simplifications that require
extensive computational resources and time. However, by
learning from large-scale datasets, DNNs can efficiently
and accurately predict the behavior and properties of
composite materials, overcoming the limitations of
traditional methods in terms of speed and precision.

DNNs have been effectively utilized to predict
the nonlinear responses and complex characteristics
Figure 2. Machine learning applications in composite materials research: of composite structures, such as composite springs,
prediction, generation, and automation
Abbreviation: AI: Artificial intelligence. composite tubes, and bistable composite structures,
which are difficult to analyze analytically. 29,30,33 Zhang
33
2. Predictive models for composite et al. applied a DNN-based surrogate model to analyze
materials the behavior of deployable bistable composite structures
with C-cross sections, using finite element analysis (FEA)
Predictive ML models are designed to estimate output data for training and subsequently performing structural
properties or behaviors based on the given input data by optimization. Similarly, as shown in Figure 3B, Hong et al.
29
identifying complex patterns and relationships. In the utilized DNNs to predict ground reaction forces using self-
field of composite materials, predictive models play a sensed capacitance data from composite ankle springs,
crucial role as traditional experimental and numerical enabling the extraction of running parameters for exo-
methods for property prediction can be time-consuming robot applications. Wang et al. developed a DNN-based
30
and computationally expensive, especially when dealing surrogate model to predict the mechanical properties of
with the anisotropic and nonlinear nature of composites. braided-textile reinforced tubular structures. Figure 3C
By leveraging ML models trained on extensive datasets, shows the performance and error of the predicted
predictive models enable rapid and reliable predictions, results (peak force, mean crushing force, displacement
significantly enhancing efficiency in material design, corresponding to peak force, and effective compression
performance evaluation, and failure analysis. This section stroke). In addition, as shown in Figure 3D, a DNN was
will explore different ML models and techniques used for utilized to predict the transverse mechanical response of
prediction in composite materials research. unidirectional fiber-reinforced composites, incorporating
the effects of defects, such as matrix voids and fiber-matrix
2.1. Deep neural networks (DNNs)
debonding. Their approach used an experimentally
31
An artificial neural network (ANN) is a type of ML validated discrete element method model to generate
technology inspired by the human nervous system. It training data of crack initiation and propagation at the
serves as a mathematical model for learning information particle-level in representative volume elements (RVEs)
from the given data and recognizing patterns. 24,25 A DNN, of the composite. The trained DNN predicted the initial
as shown in Figure 3A, is a specialized form of ANN crack location and the corresponding stress at the onset
that incorporates multiple hidden layers, allowing for of cracking with accuracies of 82% and 94%, respectively.
the extraction of complex features from data. 26-28 While DNNs were integrated with the Abaqus FEA code by
the concept of ANNs has been around for decades, Tao et al. to learn composite constitutive laws, ensuring
34
the advancement of GPU technology has enabled the physical consistency and enabling accurate, data-driven
practical implementation of DNNs, leading to significant structural predictions. Furthermore, DNNs were used for

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