An Integrated Neural Network and Finite Element Method for Defect Characterization and Strength Prediction of Unidirectional Composites

Abstract This study proposes an integrated approach that combines neural networks and finite element analysis for robust defect recognition and accurate prediction of mechanical strength in unidirectional composite materials. ABAQUS is employed for simulating random distributions of fibers and pores, along with variations in mechanical properties. Numerical simulations, based on the Monte Carlo method, provide authentic mechanical performance data as output labels for the neural network. Two methods for recognizing geometric features are employed: 1) Image recognition using a two-point cross-correlation algorithm, refined through techniques like feature selection and principal component analysis for efficient extraction of essential geometric features. 2) Convolutional neural networks with attention mechanisms for improved capturing and recognition of features. Additional descriptors such as kurtosis, skewness, and eccentricity are incorporated for a comprehensive analysis of the influence of pore morphology on mechanical performance. Deep learning techniques reveal concealed patterns, leading to highly precise mechanical property models. This innovative approach for defect characterization in unidirectional composite materials yields rapid and accurate results, with predicted tensile strength errors consistently below 10%.