Machine-learning potential reveals the origin of hydrogen embrittlement at general grain boundaries in α-Fe

Development of high-strength structural alloys is crucial for realizing a carbon-neutral society. A common issue in many alloys is hydrogen embrittlement accompanied by cracking at general grain boundaries (GBs), which is characterized by lack of crystallographic symmetry. Because experimentally analyzing the effect of hydrogen on GBs is challenging, accurate atomistic simulations are essential for understanding and suppressing hydrogen embrittlement. In this study, we developed a highly accurate and transferable machine learning interatomic potential (MLIP) for Fe-H by acquiring comprehensive and efficient learning data via simultaneous learning. Our MLIP accurately describes the density functional theory results for various lattice defects in α-Fe and their interactions with hydrogen, general GBs with hydrogen segregation that are not explicitly included in the learning data, and their deformation and fracture behavior. Large-scale molecular dynamics simulations using the developed MLIP reveals that hydrogen suppresses dislocation emissions from general GBs and promotes their fracture, supporting experimental suggestions. In contrast, for general grain boundaries, where deformation twins are responsible for plasticity, the influence of hydrogen is minimal. This study contributes to the development of high-strength alloys by providing a robust MLIP construction methodology and new insights into hydrogen embrittlement mechanisms.