Fast machine learned interatomic potential for hydrogen-induced embrittlement in α-Fe

Abstract

In this work, we present a machine-learned interatomic potential for the $α$-Fe-H system based on the tabulated Gaussian Approximation Potential (tabGAP) formalism. Trained on a Density Functional Theory (DFT) dataset of atomic configurations, energies, forces, and virials, the potential is designed for simulations on the mechanisms of hydrogen embrittlement (HE), the issue of H-induced acceleration of mechanical failure of metals. The proposed potential is shown to outperform the widely used classical and machine-learned interatomic potentials in fundamental properties of the $α$-Fe-H system. We show that the tabGAP model reproduces H-point defect properties, H-dislocation interaction, H-H interaction, and elastic constants with nearly DFT-level accuracy at a computational cost that is competitive with the efficient classical Embedded Atom Method (EAM) potentials. As an application of the tabGAP model we simulate the effect of H on the mobility of the $\tfrac{1}{2}\langle 111\rangle$ screw dislocation, which show that the model predicts an enhancement of the mobility of the screw dislocation via trapped H atoms lowering the energy barrier of kink-pair nucleation, resulting in a decreased critical shear stress of dislocation motion at $300\,\mathrm{K}$.