Physics-Aware Convolutional Neural Networks for Modelling Energetic Material in the Strong Shock Regime

Simulating energetic materials (EM) typically requires enormous computational resources due to complex dynamics processes such as strong shocks and nonlinear reactions. Artificial Intelligence (AI) has a promise to reduce simulation time from weeks on supercomputers to seconds on desktops with minimal impact on accuracy. This work promotes a novel deep learning algorithm, Physics-aware Recurrent Convolutional Neural Network (PARCv2), capable of modeling EM thermo-mechanics in strong shock regimes without explicitly requiring the governing equations. We show that compared to existing state-of-art machine learning models, ours is capable of not only predicting shock patterns, hotspot area, hotspot temperature and hotspot growth rate to higher accuracy, but also better at generalizing into unseen void shapes as well. We will explore advantages and challenges of inductive bias in modeling extreme dynamics, and propose a number of avenues of improvements that would further expand the application of AI into modeling of fast transient physics problems.