Fast and accurate Fe-H machine-learning interatomic potential for elucidating hydrogen embrittlement mechanisms

Abstract

Understanding the mechanisms of hydrogen embrittlement (HE) is essential for advancing next-generation high-strength steels, thereby motivating the development of highly accurate machine-learning interatomic potentials (MLIPs) for the Fe-H binary system. However, the substantial computational expense associated with existing MLIPs has limited their applicability in practical, large-scale simulations. In this study, we construct a new MLIP within the Performant Implementation of the Atomic Cluster Expansion (PACE) framework, trained on a comprehensive HE-related dataset generated through a concurrent-learning strategy. The resulting potential achieves density functional theory-level accuracy in reproducing a wide range of lattice defects in alpha-Fe and their interactions with hydrogen, including both screw and edge dislocations. More importantly, it accurately captures the deformation and fracture behavior of nanopolycrystals containing hydrogen-segregated general grain boundaries-phenomena not explicitly represented in the training data. Despite its high fidelity, the developed potential requires computational resources only several tens of times greater than empirical potentials and is more than an order of magnitude faster than previously reported MLIPs. By delivering both a high-precision and computationally efficient potential, as well as a generalizable methodology for constructing such models, this study significantly advances the atomic-scale understanding of HE across a broad range of metallic materials.