Elaboration of a neural-network interatomic potential for silica glass and melt

Because of its importance in various aspects of everyday life, silica is a material that has been the subject of extensive research. Studies on its amorphous phase have particularly benefited from the contribution of atomistic simulations to understand the close relationships between its structure and properties. In this context, the main difficulty lies in the compromise that had to be made between the precision of the interactions that need to be computed at an ab initio level and the important statistics required to describe disorder. With the advent of machine learning approaches, it is now possible to couple accuracy and statistics by using interatomic potentials trained on ab initio databases. This opens up unprecedented prospects for studies where calculation accuracy, system size and trajectory length are critical. In this work, we propose a machine learning potential for silica obtained from a neural network trained on a database consisting of a few hundred configurations extracted from an ab initio molecular dynamics trajectory at the Density Functional Theory (DFT) level of a silica liquid at high temperature and under pressure. We show that this potential is sufficiently accurate to describe the liquid and amorphous phases of silica, and that it is also transferable to glasses under moderate pressure and, more surprisingly, to certain crystalline phases.