Machine Learning-Driven Molecular Dynamics Study of LiF-SmF₃ Molten Salts

LiF-SmF₃ systems are the basic medium for the electrolytic preparation of Sm-Fe alloys by molten salt electrolysis. An in-depth analysis of its properties and local structural features is of great theoretical and practical significance for optimizing the electrolytic preparation process of Sm-Fe alloys. Owing to the long lead time and high cost of systematic measurements of the properties of high-temperature fluoride molten salts, machine learning-based calculations have become an efficient way to obtain the properties and structural features of molten salt systems. In this study, a machine learning potential model for the analysis of LiF-SmF₃ systems is developed. Its computational accuracy can reach 96.92%, with root-mean-square deviations of 5.48 × 10^-3 eV/atom and 6.23 × 10^-2 eV/Å for energy and force, respectively, indicating an accuracy comparable to that of density functional calculations. The machine learning potential model was used to perform molecular dynamics simulations of the LiF-SmF₃ system in the temperature range of 1100-1350 K. The density and viscosity of the system were calculated, and their deviations from the experimental measurements ranged from about 1.03%-2.77% and 3.36%-4.58%, respectively. The ion self-diffusion coefficients of the system were calculated, and the structural features of the system were analyzed using the radial distribution function of the ion pairs.