Generation and correction of machine learning interatomic potential for simulation of liquid metal corrosion with near experimental accuracy: A study for iron corrosion in liquid lead

Abstract

Molecular dynamics using a machine-learning (ML) potential trained by density functional theory (DFT) calculations is an emerging computational tool that enables accurate atomistic simulations of complex phenomena. Using iron corrosion in liquid lead as a test case, we show that although an as-trained ML potential still has significant error in simulating iron solubility due to the propagation of DFT errors, a simple correction can realize near experimental accuracy. This study provides a basic framework for the construction, correction, and use of ML potentials to facilitate their advanced and widespread applications for accurate atomistic simulations on liquid metal corrosion.