Development of Deep-Network Learned Interatomic Potential for Shock Simulations of B4C

Due to recent unprecedented innovations in computing power, data-driven methods like Machine Learning (ML) have risen in popularity in the field of solid mechanics. Specifically, ML models have been fit to material potential energy surfaces (PES) due to their ability to reach ab initio accuracy with significantly reduced computational cost. Neural Networks (NNs) are deep learning methods that can leverage the flexibility of biological neural pathways to learn the PES of complex, highly covalent materials in extreme environments. The development of such interatomic potentials (IP) is non-trivial and requires further investigations. This work details the development of an NN-based IP for boron carbide (B¬4C), specifically the training data generation, model selection, and model validation. The breadth of computational and experimental literature available on B4C allows for development and thorough validation of the MLP. Preliminary results indicate a run-time speed increase of nearly two orders of magnitude in NN-based IP shock simulations as compared to traditional ReaxFF-based simulations. This increase in efficiency can radically improve the predictability and accuracy of computational investigations previously unattainable with conventional approach.