Machine Learning Approach to Multi-scale Modeling of Granular Materials

Machine Learning Approach to Multi-scale Modeling of Granular Materials aims to address the computational cost of multi-scale models by using machine learning to develop surrogate models for the behavior of granular materials. The discrete nature of these materials leads to complex bulk behavior dependent on the local physics of particles in crucial areas. The current literature has taken phenomenological or multi-scale approaches to capture these complexities. In phenomenological approaches, we assume the existence of explicit stress-strain equations, allowing for efficient simulations at the cost of uncertainty from the ’simple’ equations. For multi-scale approaches, simulations of individual grains are coupled with bulk-scale simulations, leading to more certainty at the cost of computation time. We will train a neural network on computational stress-strain data of granular materials to construct an accurate replacement of discrete-scale simulations.