Effects of training machine-learning potentials for radiation damage simulations using different pseudopotentials

Machine learning (ML) has become a commonplace approach in the development of interatomic potentials for molecular dynamics simulations, and its use also for radiation effect modeling is increasing. In this work, we investigate the effects of training ML potentials on density functional theory data calculated with different pseudopotentials in nickel. We look in detail at the differences that appear in radiation damage simulations. The use of a ”harder” pseudopotential with semicore electrons has a direct impact on the short-range interactions, which in turn has implications on the radiation damage simulations. We find that despite these differences, the average threshold displacement energy is quite similar (40–50 eV for Ni). However, we find significant differences in the cumulative damage predicted by massively overlapping cascade simulations and compare them with Rutherford Backscattering Spectroscopy/channeling experiments. We also investigate approaches to modify the repulsive pair interactions after training the potentials and discuss the feasibility of such approaches. • The average threshold displacement energy quite similar (40–50 eV). • Equilibrium properties very similar. • Significant differences with regards to radiation damage simulations. • Semicore pseudopotential results closest to experiments in RBS/c simulations. • Modifications of the repulsive pair interactions after training were explored.