Multiscale Modeling of Granular Materials Using Mesoscale DEM and Machine Learning Approaches

Abstract We establish the necessary framework for inputting any kind of mesostructure into multi-scale models for granular materials. Keeping intact the general statistical homogenization scheme, we propose a strategy to compute the mechanical response of the mesostructures directly with discrete element simulations of a few grains or thanks to surrogate models relying on artificial neuron networks (ANN). By applying machine learning techniques at the mesoscale (instead of the Representative Elementary Volume scale), it is indeed possible to generate the necessary learning database from discrete element simulations at a relatively cheap computational cost. We apply the meso-DEM and meso-ANN strategies to the H-model (one particular micromechanical model), and we show that they can replicate the original analytical expression of the model on biaxial tests. This work paves the way for using more complex mesostructures to account for instance for gap-graded materials.