SuperSalt: equivariant neural network force fields for multicomponent molten salts system.

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Molten salts are vital for clean energy technologies, but their complex thermophysical properties are difficult to explore.
• Predicting these properties across diverse chemical compositions is a significant challenge.

Purpose of the Study:

• To develop a highly accurate machine learning interatomic potential (MLIP) for predicting molten salt properties.
• To enable efficient exploration of molten salt chemical space for clean energy applications.

Main Methods:

• Developed SuperSalt, a machine learning interatomic potential (MLIP) trained on 11-cation chloride melts.
• Utilized an integrated workflow for systems of one, two, and 11 components.
• Validated the model's accuracy and transferability across a broad chemical space.

Main Results:

• SuperSalt achieves near-density functional theory (DFT) accuracy in predicting key thermophysical properties like density, bulk modulus, thermal expansion, and heat capacity.
• The model demonstrates excellent transferability across diverse molten salt compositions.
• Bayesian optimization combined with SuperSalt accelerates the discovery of optimal salt compositions.

Conclusions:

• SuperSalt provides a powerful and efficient tool for understanding and predicting molten salt behavior.
• This MLIP accelerates the discovery of novel molten salts for clean energy applications.
• The framework allows for future extensions to more complex and multi-elemental systems.