Performance of two machine-learning interatomic potentials for studying hydrogen isotope diffusion in vanadium

Vanadium (V) is a candidate structural material for liquid lithium tritium breeder blankets. There is a likelihood for the tritium generated in the lithium to diffuse across interfaces into the V structural material. It is crucial to model the transport in the surrounding material to enable accurate accounting of the generated tritium within the breeder blanket assembly. In this work, we develop two machine-learning interatomic potentials (MLIPs) – an atomic cluster expansion (ACE) potential and a deep-learning neural network (DeepMD) potential – to study the behavior of hydrogen (H) isotopes in irradiated V. Specifically, we compare the formation and migration energies of interstitial H and the incremental binding energy of H clusters in V monovacancies predicted by the two potentials to density functional theory (DFT) calculations. We perform molecular dynamics (MD) simulations using the potentials to study the temperature-dependent de-trapping rate of each H atom in a monovacancy, and the effective diffusivity of H in V. Monovacancies act as H traps, reducing the effective diffusivity compared to defect-free V. We also perform path integral molecular dynamics (PIMD) using the ACE potential to analyze the effect of nuclear quantum effects on hydrogen diffusion, specifically at low temperatures. To study the difference between the H isotopes, we include the zero-point energy contributions to the DFT energetics, and perform classical MD and PIMD explicitly for the different isotopes. Diffusivity is found to drop with increasing H-isotope mass, with the differences increasing at low temperatures. We also compare the computational cost of using the ACE and DeepMD potentials, and discuss the performance of a set of graphical atomic cluster expansion (GRACE) universal machine-learning interatomic potential foundation models for studying H in V, finding that the accuracy of the latter is strongly influenced by the source training database.