A TPU-ENHANCED DEEP REINFORCEMENT LEARNING APPROACH FOR AUTOMATED GENERATIONS OF INTERPRETABLE MODELS FOR ENERGETIC MATERIALS ACROSS LENGTH SCALES

Machine learning has increasingly become a disruptive force that changes how computational models are generated, verified and validated. With the new experimental data afforded by technologies such as micro-CT tomographic imaging, both the amount of data and the diversity of the data types are rapidly growing. In particular, the dynamic responses of energetic materials at the macroscopic scale is attributed to a combination of mechanisms across multiple length scales, such as reactive collapse of pores, localized shear, damage, and fragmentation of crystals, frictional heating amount contacts and debonding of the crystal-binder interfaces. This complexity makes it difficult to manually derive a theoretical model that guarantees all the important multi-physical coupling mechanisms are considered properly, while simultaneously ensuring that simplifications for the sake of convenience and interpretability remains appropriate for the intended applications. The objective of this DURIP project is to advance the fundamental understanding of the dynamic responses of energetic materials at the macroscopic scale via information obtained from sub-scale MD simulations and experiments. The support of this DURIP will be used to purchase a NVIDIA high-performance computing server designed specifically for accelerating machine learning tasks.By enabling acceleration through a new Tensor Core architecture, the time used for reinforcement and supervised machine learning used for both the PIs on-going YIP project, Modeling the highrate responses of wetted granular materials across scales and the third-party replicable validation exercises utilizing 3D printers, and the new FY19 MURI project, Integrating multiscale Modeling and Experiments to Develop a Meso-Informed Predictive Capability for Explosives Safety and Performance can be significantly shortened.