Predicting hydrogen diffusion in nickel–manganese random alloys using machine learning interatomic potentials
Kazuma Ito, Naoki Matsumura, Yuto Iwasaki, Yasufumi Sakai, Misaho Yamamura, Tomohiko Omura, Junichiro Yamabe, Hisao Matsunaga
Abstract
Abstract To advance carbon neutrality, structural materials for high-pressure hydrogen environments must be designed based on fundamental principles. However, the atomic-scale complexity of random alloys hinders the development of interatomic potentials that can accurately reproduce hydrogen behavior influenced by alloying elements. This study develops a machine-learning interatomic potential (MLIP) for the Ni–Mn–H ternary system by efficiently sampling training data through an active learning strategy that combines atomic-force uncertainty and structural descriptors of diverse atomic environments. Molecular dynamics simulations employing the constructed MLIP quantitatively reproduce the experimentally observed non-monotonic dependence of the hydrogen diffusion coefficient on the Mn content. Two competing Mn-addition effects are found: increased and decreased activation energies from repulsive Mn–H interactions and lattice expansion, respectively, the balance of which shifts with the Mn content and governs the diffusion behavior. This approach enables accurate prediction of hydrogen diffusion in random alloys and provides atomic-level insights into alloying effects.