Prediction of the effective properties of matrix composites via micromechanics-based machine learning

Abstract

This study aims to integrate micromechanics-based analytical models with machine learning (ML) models to predict the effective properties of two-phase composites. A novel approach grounded in Maxwell’s effective field method (EFM) is proposed to address the accuracy limitations inherent in micromechanics-based models while minimizing the amount of data needed to fit ML models. Notably, this new approach requires only two macroscale data points to predict the effective properties. The approach is introduced for inhomogeneities of arbitrary shape, orientation, and properties and is applicable to effective thermal, electrical, elastic, and other properties. Two case studies focusing on the elasticity problem are presented to illustrate the applicability and accuracy of the new approach; one involving a particulate composite of copper reinforced with diamond particles, and the other a unidirectional composite of 3D-printed nylon reinforced with Kevlar fibers. The results of these case studies are compared with finite element models and demonstrate an excellent agreement.