Temperature Extensible Deep Potential Model for Molten NaF–BeF2–ZrF4: Predicting Transport Properties and Local Structure

Abstract

Molten NaF–BeF₂–ZrF₄ (FNaBeZr) has garnered significant interest as one of the potential nuclear fuel carrier salts for molten salt reactors. Here, deep potential molecular dynamic (DPMD) simulations by integrating first-principles calculation, machine learning, and molecular dynamic methods are employed to systematically investigate the transport properties and local structures of molten FNaBeZr across a broad temperature range. Such computational approaches will bypass resource-intensive trial-and-error experimentation and massive high-fidelity density functional theory (DFT) calculations, which represents that the DPMD framework achieves dual optimization by substantially reducing computational costs while enabling enhanced system scalability─a critical advancement for modeling complex coordination chemistry environments and deciphering coupled ionic transport phenomena at extended spatiotemporal scales. The trained deep potential model reproduces the densities, ionic self-diffusion coefficients, and pair/cluster structures from 773 to 973 K and successfully predicts these properties as well as the ionic conductivity and shear viscosity at extended temperatures (beyond the training range). Furthermore, the quantifiable correlations between structural features (e.g., first-peak height/position and coordination number) and temperature are established, and their latent connections to transport properties are also analyzed. The robust DPMD results demonstrate the temperature extensibility of the deep potential for molten FNaBeZr and provide critical technical parameters and engineering data for future simulations of molten fluorides and for the design optimization of pump systems and reprocessing infrastructure in high-temperature environments.