Molecular dynamics of high pressure tin phases: Empirical and machine learned interatomic potentials

Understanding the nature of phase transitions of tin under high pressure requires atomic scale modeling, where the simulation accuracy is largely determined by the interatomic potential (IAP) used. We have compared the current state of classical IAPs for tin and have found that only low pressure phases are well represented. However, machine learned interatomic potentials (ML-IAPs) have shown improved accuracy compared to traditional potentials. We report the development of two different training sets, domain expertise (DE) and active learning (AL), for fitting a tin ML-IAP. Each methodology has particular benefits, with DE better reproducing material properties like cold curves and AL more efficiently spanning descriptor space more efficiently. In future work, we plan to further compare the effect of both the training set and type of ML-IAP used on the accuracy of the ML-IAP on tin property calculations and extrapolative modeling performance.