Multiscale machine-learning interatomic potentials for ferromagnetic and liquid iron

Abstract

Abstract A large and increasing number of different types of interatomic potentials exist, either based on parametrised analytical functions or machine learning. The choice of potential to be used in a molecular dynamics simulation should be based on the affordable computational cost and required accuracy. We develop and compare four interatomic potentials of different complexity for iron: a simple machine-learned embedded atom method (EAM) potential, a potential with machine-learned two- and three-body-dependent terms, a potential with machine-learned EAM and three-body terms, and a Gaussian approximation potential with the smooth overlap of atomic positions descriptor. All potentials are trained to the same diverse database of body-centred cubic and liquid structures computed with density functional theory. The first three potentials are tabulated and evaluated efficiently using cubic spline interpolations, while the fourth one is implemented without additional optimisation. The four potentials span three orders of magnitude in computational cost. We compare and discuss the advantages of each potential in terms of transferability and the balance between accuracy and computational cost.