Equivariant Neural Network Force Fields for 11-Cation Chloride Molten Salts System

Molten salts are a promising class of ionic liquids for clean energy applications, such as nuclear and solar energy. However, efficient and accurate evaluation of salt properties from a fundamental, microscopic perspective remains challenging. Recently, machine learning interatomic potentials (MLIPs) have emerged as a crucial tool in materials science, combining near ab initio accuracy with the computational efficiency of classical force fields. Standard MLIPs are typically fit to a few chemical species, and recently, there have emerged so-called universal potentials, which often fit 50 or more chemical species. We target a middle ground that treats 10-20 elements with similar chemistry and phases with a goal of achieving almost ab initio accuracy for a large composition space. We describe this type of MLIP for molten salts as a 'SuperSalts' potential. We have developed an efficient workflow to generate such an MLIP for all liquid phase compositions of 11-cation chloride melts, namely LiCl-NaCl-KCl-RbCl-CsCl-MgC2-CaCl2-SrCl2-BaCl2-ZnCl2-ZrCl4. Extensive validations indicate that the SuperSalts potential can predict densities, bulk moduli, radial distribution functions, coordination numbers, potentials of mean force, specific heat capacities, viscosities, and thermal conductivities of multicomponent molten salt systems with near-DFT accuracy, maintaining good transferability across different chemical systems. The fitting uses ~70,000 ab initio training configurations, which is a large training database. However, we believe this SuperSalts potential approach is dramatically more efficient than fitting all the suballoys separately (there are 561 suballoys up to quaternaries (2047 total suballoys)), both in terms of the total number of ab initio calculations required, training time, and human time. The SuperSalts potential also serves as a consistent, validated, one-stop resource. This SuperSalts potential joins other recent results in suggesting a paradigm shift from empirical, semiempirical, and ab initio approaches to a more efficient and accurate machine learning potential approach in molten salt modeling. However, our findings further suggest that a SuperSalts potential could provide a single foundational MLIP for the majority of molten salts of interest.