Toward Multiscale Modeling of Ni/Molten FLiNaK Interfaces with Equivariant Machine Learning Potentials

Molten fluoride salts are attractive as heat transfer fluids and thermal energy storage media in concentrated solar power and as coolants and fuel carriers in Generation IV Molten Salt Reactors (MSRs); however, corrosion of structural alloys remains a main obstacle to their deployment. In this work, ab initio molecular dynamics (AIMD) simulations of three binary and four ternary LiF–NaF–KF compositions between 900 and 1300 K yield densities, radial distribution functions, and heat capacities in good agreement with experimental data. Simulations of Ni/FLiNaK interfaces showed no corrosion within the picosecond timescale accessible to AIMD. To extend this timescale, the MACE MH-1 foundation model was fine-tuned on the salt and interface datasets using a multihead replay strategy, with active learning to incorporate the Ni(110) surface. The resulting MLIP reproduces the AIMD heat capacities within 4.4% and extrapolates to the pure salts absent from the training data. MLIP-MD simulations reveal that fluoride coverage increases as Ni(111) < Ni(110) < Ni(kink), following the decreasing coordination of the surface nickel atoms, in agreement with a previous study on Ni–Cr alloys in FLiNaK.