Prediction of Toughness in Heterogeneous Materials using Machine Learning

Composites or heterogeneous materials have a wide range of applications in the fields of aerospace, structural materials, biomedical, energy conversion devices, high-temperature materials, etc. But the effective properties of composites, in particular, toughness and strength are hard to predict as a function of their constituent properties. By definition, heterogeneous materials contain different shapes, sizes, compositions, and material or phase distributions at different length scales, making iterative experiments inefficient, time-consuming, and expensive to use for this purpose. In this work, the toughness of heterogeneous brittle material is investigated using a combination of continuum scale simulations and machine learning techniques. Our goal is to predict the configurations (or different structural and material parameters) of the composite that optimize its effective toughness. The continuum simulations are based on the variational phase-field modeling of fracture, and we applied the novel surfing boundary condition, which allows adequate time for the crack to propagate, enabling evaluation of the critical energy release rate or toughness. In particular, we focus on understanding the criteria for deflection vs. penetration for a number of composite configurations and identify the structural or material descriptors that define the criteria. Using the descriptors, we develop supervised machine learning models for predicting the toughness of the composite. The models are trained using data sets, collected from continuum simulations and are used to ascertain the toughness-structure correlation for a number of unexplored variants of the configurations. This talk will discuss the findings and highlight the need for applying machine learning tools to predict the toughness of a composite, for an arbitrary choice of structural and material parameters.