Machine learning reveals properties of molten salts

A surge in demand for renewable energy has led to a need for improved thermal energy storage. Molten salts, with high stability and energy storage potential, have emerged as a leading option.

To better understand these salts’ properties and microstructures, Tian et al. adopted a machine learning-enhanced method that integrates the advantages of traditional empirical potential functions and first-principles molecular dynamics. The researchers applied the method, called the Deep Potential GENerator (DPGEN) enhanced sampling method, to a binary carbonate system for the first proof of concept.

“This work provides a solid theoretical foundation for optimizing the performance and facilitating the application development of molten salt materials,” said author Heqing Tian.

The authors started by using ab initio molecular dynamics to form an initial dataset that was used to train a preliminary model. Using the DPGEN algorithm, they iterated to a final potential function and used guided molecular dynamics simulations to calculate various properties of the molten salts. An analysis of the results showed the relationships between the salt’s properties and its microstructure.

“Our findings not only illuminate the mechanisms underlying the temperature-dependent effects on molten salt properties, but also delve into the interactions between ions and microstructural changes, offering scientific evidence for the design and development of novel molten salts,” Tian said. “Furthermore, this study showcases the immense potential of the DPGEN method in simulating complex systems, thereby offering theoretical support for research in related fields.”

The results, the authors said, will also deepen the understanding of molten salt systems and can provide guidance for developing future thermal energy storage media and electrolytes needed for rare earth material electrolysis.

Source: “A theoretical study of thermal properties and structural evolution in binary carbonates phase change material: Machine learning-enhanced sampling strategy,” by Heqing Tian, Wenguang Zhang, and Chaxiu Guo, Journal of Chemical Physics (2024). The article can be accessed at https://doi.org/10.1063/5.0219401 .