Unraveling the structure and thermophysical property of heterogeneous eutectic salt by machine learning potential for solar thermal energy storage

Molten salts are an excellent high temperature phase change thermal storage material with high heat storage density. Machine learning (ML) methods with deep potential have been recognized to have tremendous potential application advantages in predicting the thermal properties of molten salt. Herein, we employ the Deep Potential GENerator (DP-GEN) active learning approach to construct and evaluate the potential function of Na₂CO₃-NaCl heterogeneous eutectic salt, and the thermophysical properties and structures of eutectic salt are comprehensively predicted and analyzed. Density and radial distribution function (RDF) are used to validate the accuracy of the simulated structure and properties, with the density simulation results showing an error of merely 2.54 % compared to experimental data. DPMD achieves a level of accuracy comparable to AIMD in simulating melt structure, with an error of just 0.92 %. Na ions and O ions primarily form two types of tetrahedral structures within the molten salt system. CO₃²⁻ exhibits a regular triangular structure, with CO bonds oscillating in a plane centered on C. As the temperature increases from 973 K to 1173 K, the thermal conductivity decreases from 0.564 W/(m·K) to 0.559 W/(m·K), the viscosity decreases from 3.454 mPa·s to 1.978 mPa·s, a trend opposite to that of the self-diffusion coefficient (D). Change in viscosity is attributed to alterations in interparticle interactions, distances and coordination relationships. This work provide a new perspective to use DP-GEN active ML strategy to precisely predict structure and thermal property of heterogeneous molten salt systems.