Development of a Neural Network Potential for Modeling Molten LiCl/KCl Salts: Bridging Efficiency and Accuracy

Abstract

Optimizing molten salts for molten salt reactors and concentrated solar power can be challenging due to limited experimental data. To tackle this, we utilize neural network potentials (NNPs) for the atomistic modeling of molten salts and use the widely popular LiCl/KCl salts as prototype systems. Based on the results reported herein, the NNP exhibits remarkable accuracy and is similar to density functional theory calculations. The reliability of the NNP was due to a rigorous approach to acquiring training data, which covered atomic configurations at different temperatures and pressures for pure LiCl, pure KCl, and LiCl−KCl (58.8% mol LiCl) systems. It was observed that the NNP reasonably reproduced experimental physical properties of molten LiCl/KCl salts across various compositions and temperatures and microstructures that are similar to highly accurate first-principles molecular dynamics. Furthermore, the NNP was employed to calculate diffusion coefficients of molten LiCl−KCl salts, for which no current experimental data are available. From this, we verify the NNP by reporting the well-known Chemla effect in molten LiCl−KCl systems. We further employ the use of the NNP to predict the phase diagram of the LiCl−KCl system by using solid−liquid coexistence simulations. The robustness and versatility of the NNP reported in this study demonstrate the promising potential of the developed NNP in overcoming the longstanding trade-offs between computational efficiency and accuracy in the MD simulations of molten salts.