Deep potential-driven molecular dynamics unveiling composition–structure–property correlations in LiF–BeF2–UF4 molten salt system

Molten fluoride salts are critical heat carriers and fuel solvents for advanced molten salt reactors. Yet, their complex atomic interactions and composition-dependent structural evolution remain challenging to characterize over a wide range of temperatures and compositions. Here, we develop a highly transferable and accurate machine learning interatomic potential for the ternary LiF–BeF2–UF4 system, enabling large-scale molecular dynamics simulations with near-density functional theory accuracy across UF4 concentrations from 0 to 25 mol. %. Trained from ab initio molecular dynamics data via an active learning loop, this potential achieves low root-mean-square errors for energy and forces and exhibits excellent cross-concentration generalization. Using this model, we reveal that U4+ acts as a network modifier that competes for F− ligands and promotes the formation of an extended Be–F–U ionic network, which governs the macroscopic physicochemical properties. With increasing UF4 content, the formation of interconnected networks is found to be correlated with reduced ion mobility, lower thermal conductivity, and higher shear viscosity, suggesting a mechanistic link that awaits further quantitative validation. We further quantify the coordination environments, cage correlation lifetimes, cluster network evolution, ion diffusion activation energies, and thermophysical properties, establishing a clear composition–structure–property relationship at the atomic scale. This work provides a robust computational framework for fluoride fuel salts and offers mechanistic insights critical for the design and optimization of thorium-based molten salt reactor fuels. The developed deep potential is transferable, scalable, and expected to serve as a reliable tool for future multiscale studies of complex molten salt systems.