Physics-Informed Neural-Network-Based Generation of Composite Representative Volume Elements with Non-Uniform Distribution and High-Volume Fractions.

To reduce the reliance on large training sets for representative volume element (RVE) generation using machine learning, this work presents a novel approach based on physics-informed neural network (PINN) to generate RVEs for unidirectional fiber-reinforced composites with non-uniform fiber distributions and high-volume fractions. The method embeds physical constraints including fiber non-overlap, volume fraction, and boundary conditions directly into the neural network's loss function. This integration eliminates the need for large training datasets, which is typically required by traditional machine learning methods. Moreover, it achieves volume fractions exceeding 0.8, surpassing the jamming limit of conventional generation techniques. Exhaustive statistical measurements taken at different scales confirm that the proposed method could accurately reproduce local fiber distribution patterns in realistic microstructures while maintaining complete randomness at larger scales. Finite element analysis was employed on the generated RVEs to predict the elastic properties and damage behavior that taking into account the interfacial debonding and nonlinear damage in matrix. The predictions of both macroscopic mechanical properties (elastic properties and strength) and microscopic damage patterns show good agreement with experimental results. The proposed PINN-based framework provides an efficient and reliable tool for computational micromechanics of polymer matrix composites.