An Accurate Machine Learning Potential for Many-Component Molten Salts

Molten salts are crucial for clean energy applications, yet exploring their thermophysical properties across diverse chemical space remains challenging. We present the development of a machine learning interatomic potential (MLIP) called SuperSalt, which targets 11-cation chloride melts (LiCl-NaCl-KCl-RbCl-CsCl-MgCl₂-CaCl₂-SrCl₂-BaCl₂-ZnCl₂-ZrCl₄) and captures the essential physics of molten salts with near-DFT accuracy. SuperSalt was fit using an efficient workflow that integrates systems of one, two, and 11 components, and can accurately predict thermophysical properties such as density, bulk modulus, thermal expansion, and heat capacity. SuperSalt was validated across a broad chemical space, demonstrating excellent transferability. Our approach includes many elements but treats a consistent type of chemistry and only one phase. In that way it provides a middle ground between typical few-element MLIPs, which must be fit for each system of interest, and the emerging Universal MLIPs, which treat most of the periodic table but often have limited accuracy for a specific study. We further illustrate how Bayesian optimization combined with SuperSalt can accelerate the discovery of optimal salt compositions with desired properties, e.g., finding a target density. SuperSalt can be easily extended to new elements and phases and represents a shift towards a more universal, efficient, and accurate modeling of molten salts for advanced energy applications.