Recently, the research group of Shanghai Institute of Applied Physics (SINAP),Chinese Academy of Sciences (CAS) has made progress in the prediction of properties and structures of molten salts. The microscope structure, thermophysical properties and structure-property relationship of binary magnesium chloride based molten salts are described in detail using machine learning and molecular dynamics simulation techniques. The article entitled "Development of Deep Potential of Molten MgCl2-NaCl and MgCl2-KCl Sales Driven by Machine Learning" has been published in the journal of “ACS Applied Materials & Interfaces”. Xu Tingrui, doctoral student of SINAP, is the first author, and Associate Prof. Xuejiao Li & Prof. Zhongfeng Tang are the corresponding authors.
Molten MgCl2-NaCl (MN) and MgCl2-KCl (MK) salts, as high-performance and low-cost heat transfer fluids and storage media, are expected to play important roles in the fourth generation molten salt reactor system and the third generation concentrated solar power system. The heat capacity, density, viscosity, and thermal conductivity of molten salt are directly related to the design of thermal-hydraulic systems, while experimental measurements cannot avoid significant errors caused by the reactivity and corrosiveness of molten salts at high temperatures. First principles molecular dynamics (FPMD) simulations can predict the properties of molten salts without force field models, but the extendable spatiotemporal scales are limited, resulting in larger prediction errors for time-correlated properties. Classical molecular dynamics simulations can to some extent eliminate scale effects, but require inputting empirical potential parameters. The author drew their strengths and complements weaknesses, first used deep learning method to obtain the interaction potentials of two molten salts based on the FPMD simulation results, and then conducted larger spatiotemporal scale deep potential molecular simulations (DPMD). And the calculation protocol for the viscosity and thermal conductivity of molten salt was optimized using time decomposition method and eliminating the thermoelectric coupling effect, respectively. Finally, the correlation between properties and local structures of two molten salts was explained. Research has shown that MK has higher specific heat capacity and better thermal storage performance, and MN has higher thermal conductivity and lower viscosity, which performs better in heat transfer. Through structural analysis such as radial distribution function, potential of mean forces, and coordination number, the important influence of Mg-Cl bonds in the properties of molten salts was revealed. Compared to MN, Mg-Cl bonds bind more tightly in MK, and the energy barrier required for anions to jump out of the first coordination shell is higher, that’s why MK has higher specific heat capacity and lower transport performance. In addition, the deep potential model obtained in this study demonstrates a high degree of generalizability, which not only accurately simulates the properties of molten salts at dataset temperatures, but also enables successful prediction of properties at non-dataset temperatures. This technological advancement exemplifies the prospective utility of machine learning within the realm of molten salt chemistry, particularly pertaining to the thermodynamics characteristics of intricate molten salt systems.
This work was financially supported by the National Key R&D Program, the National Natural Science Foundation of China and Young Potential Program of SINAP, CAS.