Development of machine learning potential for hexagonal boron nitride with strictly local equivariant deep neural network

Development of machine learning interatomic potentials is critical for performing accurate large-length and long-time scale simulations of materials. In this work, we utilized the Allegro model, a strictly local equivariant deep neural network interatomic potential architecture, for the development of machine learning potential for hexagonal boron nitride (h-BN) with defects and grain boundaries. We used about 30,000 images that are generated with ab initio molecular dynamic simulations at 500, 1000, 1500 K of h-BN with and without point defects as our training dataset. The developed potential is able to predict potential energies and forces of h-BN systems with and without defects with a mean absolute error (MAE) of 4 meV/atom and 60 meV/Å, respectively. It also reproduces phonon dispersion curves and density of states that are comparable with that obtained from density functional theory calculations. Results will be presented for molecular dynamics simulations of h-BN with parallel and anti-parallel 4|8 grain boundaries with tens of thousands of atoms, employing the machine learned potential that validate its accuracy and computational efficiency.