A nuclear-quantum-corrected machine-learning potential reveals quantum-enhanced hydrogen segregation at general grain boundaries in alpha-iron

Abstract

Atomistic descriptions of hydrogen diffusion and trapping at defects are essential for understanding hydrogen embrittlement. As the lightest solute in metals, hydrogen exhibits nuclear quantum effects that alter these processes even at room temperature. Explicit treatment of such effects is computationally demanding, limiting large-scale simulations of complex environments. Here, we use an Fe-H machine-learning interatomic potential (MLIP) based on the performant implementation of the atomic cluster expansion (PACE), covering diverse Fe-H environments, and relabel the training configurations underpinning its transferability with quantum mean forces from centroid-constrained path-integral molecular dynamics at 300 K. This yields a nuclear-quantum-corrected PACE (NQC-PACE) without additional density functional theory calculations. At parent PACE, NQC-PACE describes nuclear quantum effects on hydrogen trapping at vacancies, dislocations, surfaces and general grain boundaries, H-H interactions, and diffusion in alpha-Fe. Grand-canonical Monte Carlo/molecular dynamics simulations show nuclear quantum effects markedly enhance hydrogen segregation at general grain boundaries and trapping behaviour in closer agreement with experimental trends. This enhancement arises from selective quantum stabilisation of open, anisotropically soft local environments. Our framework uses finite-temperature quantum mean forces to relabel the configurational space covered by an MLIP, enabling large-scale analysis of complex materials where light-element quantum effects matter.