A two-stage deep learning framework for predicting crack patterns and mechanical properties of unidirectional composites with void defects

Abstract

Void defects in unidirectional composites critically govern crack initiation and propagation, leading to substantial degradation of transverse mechanical properties. To accurately characterize the influence of void defects on composites, this study proposes a novel two-stage deep learning framework that integrates U-net for crack pattern predictions and convolutional neural network for predicting stiffness and strength of unidirectional composites with void defects. To improve accuracy of mechanical property prediction, the innovative feature-fusion mechanism utilizes both material microstructures and corresponding crack patterns generated by the crack prediction network as input features. For the training of deep learning framework, a comprehensive dataset, generated through micromechanical modeling, contains randomly distributed fibers, inter-fiber voids, matrix voids, and resin-rich areas. The proposed framework achieves high-precision predictions of crack patterns and mechanical properties while significantly reducing computational costs, demonstrating strong potential for applications in material design.