Development of physically informed neural network (PINN) interatomic potentials

The present work focuses on the development of a novel class of interatomic potentials known as a "physically-informed neural network" (PINN) potentials. This potential format combines the high level of flexibility inherent to artificial neural networks (ANNs) with the transferability associated with physically inspired analytic potential models. Currently we focus on single component, covalently bonded systems including silicon (Si) and germanium (Ge), however, the PINN model can generally be applied to any multicomponent metallic or covalent system. Finally we also discuss the application of these potentials to the study of various thermal properties in both bulk phase and two dimensional Si and Ge.