Meso-scale simulation of energetic materials. I. A method for generating synthetic microstructures using deep feature representations

Abstract

The response of a wide class of heterogeneous energetic materials (HEs) to loads is determined by dynamics at the meso-scale, i.e., by physicochemical processes in their underlying microstructure. Structure-property-performance (S-P-P) linkages for such materials can be developed in a multi-scale framework, connecting the physics and thermophysical properties at the meso-scale to response at the macro-scale. Due to the inherent stochasticity of the microstructure, ensembles of microstructures are required to conduct meso-scale simulations to establish S-P-P linkages. Here, a deep neural network-based method called deep feature representation is applied to generate a range of material microstructures from heterogeneous energetic materials to metal foams and metallic mixtures. The method allows for the generation of stochastic microstructures using a single real microstructure as the input and is not limited to low packing density or topological complexity of solids. In its application to pressed energetic materials, we show that qualitative and quantitative features of real (i.e., imaged) microstructures are captured in the synthetic microstructures. Therefore, a stochastic ensemble of synthetic microstructures can be created for use in reactive meso-scale simulations to relate the microstructures of HEs to their performance. While the focus is on pressed HE microstructures, we also show that the method is general and useful for generating microstructures for in silico experiments for a wide range of composite/multiphase materials, which can be used to establish S-P-P linkages.