Machine learning techniques to probe the properties of molten salt phase change materials for thermal energy storage

Machine learning methods could advance the application of molten salt phase change materials. In this study, we employ two machine learning methods with three machine learning potential functions to investigate the local structure and thermal properties of a binary chloride salt, and the accuracy and applicability of the three machine learning potentials are assessed. The results reveal that the precision of datasets and the root mean square error of fit may not entirely capture the advantages of machine learning potentials. Deep potential methods necessitate more extensive datasets to ensure the stability of the potential function, whereas neuroevolution potential methods can yield good outcomes with fewer but finer datasets, demonstrating superiority in usability and computational speed. Both methods can accurately predict the properties of a binary salt. This work will contribute fresh perspectives to the advancement of machine learning for molten salt thermal energy storage materials.