Thermal Property and Structure of Toxic Fluoride Molten Salt for Efficient Nuclear Energy Application Using Machine Learning Molecular Dynamics

Molten salt reactors, as a leading candidate for Generation IV nuclear energy systems, exhibit thermal performance critically dependent on the thermal properties of fluoride molten salts. Addressing the challenges of severe corrosivity and beryllium toxicity encountered in high-temperature experiments on the NaF-BeF₂ system, this study pioneers the application of active learning-enhanced sampling strategies to the fluoride molten salt system. A high-precision deep potential model spanning a broad temperature range (773-1173 K) was constructed. Through seven rounds of iterative optimization using the deep potential generator (DP-GEN) framework, the potential function was optimized, enabling the correlation of an atomic-scale microstructure with macroscopic thermal properties. The results reveal that the molten salt exhibits an amorphous structure characterized by short-range order and long-range disorder. Be²⁺ ions dominate the formation of [BeF₄]^2- tetrahedra and polymeric clusters such as [Be₂F₇]^3- and [Be₃F₁₀]^4-. The Be-F bond demonstrates significantly higher stability than the Na-F bond. However, reduced energy barriers at elevated temperatures intensify cluster dissociation. The DP model also successfully predicted the key thermal properties, including density (error within 1.62%), specific heat capacity (error within 6.79%), diffusion coefficients, viscosity, and thermal conductivity. Its accuracy significantly surpasses that of ab initio molecular dynamics simulations. Furthermore, the model elucidates the microscopic mechanisms underlying property variations, providing an atomic-scale theoretical foundation for molten salt reactor design.