Multiscale machine-learning interatomic potentials for ferromagnetic and liquid iron

TL;DR

This paper develops and compares four multiscale machine-learning interatomic potentials for iron, covering various complexities and computational costs, to evaluate their accuracy, transferability, and efficiency for modeling ferromagnetic and liquid iron.

Contribution

It introduces and systematically compares four different machine-learning potentials for iron, highlighting their trade-offs in accuracy and computational efficiency.

Findings

The potentials vary significantly in complexity and cost.

All potentials are trained on the same DFT database.

The paper discusses transferability and applicability of each potential.

Abstract

We develop and compare four interatomic potentials 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 SOAP descriptor. All potentials are trained to the same diverse database of body-centered cubic and liquid structures computed with density functional theory. The four presented potentials represent different levels of complexity and span three orders of magnitude in computational cost. The first three potentials are tabulated and evaluated efficiently using cubic spline interpolations, while the fourth one is implemented without additional optimization. We compare and discuss the advantages of each implementation, transferability and applicability in terms of the balance between required accuracy versus computational cost.