Enhanced prediction of transverse mechanical behavior of unidirectional fiber-reinforced composites using new spatial descriptors based on deep neural networks

Abstract

In this study, we developed a data-driven approach using a deep neural network (DNN) to predict the nonlinear stress-strain behavior of fiber-reinforced composites under transverse tensile loading. We explored the impact of different input features characterizing the composite microstructure on the prediction accuracy of the DNN model. Specifically, we introduced a spatial descriptor representing fiber arrangement and proposed new spatial descriptors based on fiber area to improve input feature quality. Incorporating widely-used spatial descriptors such as the second-order intensity function and radial distribution function improved the prediction performance compared to previous methods. Combining spatial descriptors for both fiber distribution and alignment further enhanced prediction accuracy. To address the limitations of conventional spatial descriptors, we developed new spatial descriptors using a continuous function. Our findings demonstrate the importance of selecting appropriate input features for improved DNN model performance, even with the same dataset. Moreover, the proposed fiber area-based spatial descriptors offer insights into the micromechanical behavior of composite materials.