Neural network-driven framework for efficient microstructural modeling of particle-enriched composites

Simulating materials with complex, variable microstructures like nanoparticle-enriched matrices presents significant challenges for FEM, particularly in defining constitutive equations across diverse loading conditions and internal structures. These complexities often make purely physics-based FEM simulations impractical for larger macroscopic domains. We present a novel approach that significantly reduces the dimensionality of the FEM data, while preserving the ability to interpolate and extrapolate across various nanoparticle configurations. Our method combines raw data projection, neural network surrogates, and convolutional autoencoders to geometrically normalize the data in a reduced-dimensional space. This reduction enables efficient training and prediction, while retaining flexibility for different microstructural variations. Compared to direct training on unprocessed data, our approach reduces dimensionality and computational costs without sacrificing accuracy, allowing robust virtual testing of particle-enriched composites. This framework can be applied to other multiscale composite systems for optimized material design.