Xiliang Lian1,2,Mathieu Salanne1,2,3
Sorbonne Université, CNRS1,Réseau sur le Stockage Electrochimique de l’Energie (RS2E)2,Institut Universitaire de France (IUF)3
Xiliang Lian1,2,Mathieu Salanne1,2,3
Sorbonne Université, CNRS1,Réseau sur le Stockage Electrochimique de l’Energie (RS2E)2,Institut Universitaire de France (IUF)3
Machine learning potential (MLP) has emerged as a promising method to approach the potential energy surface with high accuracy and efficiency. However, the versatility and efficiency of such a method compared with classical interaction potential are rarely investigated. Using BaSnF<sub>4</sub>, a prospective solid electrolyte for fluoride ion batteries, and an auxiliary simpler system NaF, which holds a rock salt structure, we show how an MLP can capture the subtle interactions of Sn lone pairs while a polarizable force field fails. The accuracy of our MLP is validated by computing vibrational properties such as phonon dispersion and equation of states and comparing them with the results obtained from density functional theory and MLP demonstrates excellent agreement with density functional theory. The MLP also exhibits significantly boosted computational efficiency compared with the reference ab initio molecular dynamics method. Furthermore, from large-scale machine learning molecular dynamics simulation with BaSnF<sub>4</sub>, we investigated the two-dimensional fluoride ion motion between Ba-Sn and Sn-Sn layers and showed how Sn atoms modulate the fluoride ion diffusivity.