Toward a Generalizable Prediction Model of Molten Salt Mixture Density with Chemistry-Informed Transfer Learning

Area of Science:

• Materials Science
• Chemical Engineering
• Computational Chemistry

Background:

• Accurate thermophysical properties of molten salts are crucial for designing applications.
• Existing databases have data gaps and experimental measurements are challenging due to high temperatures and corrosivity.
• Current Redlich-Kister (RK) models lack generality as they require subcomponent data for each new system.

Purpose of the Study:

• To address data sparsity and improve generalizability in predicting molten salt density.
• To develop a novel approach using deep neural networks (DNNs) for molten salt property prediction.
• To outperform existing methods in molten salt density prediction.

Main Methods:

• A transfer learning procedure was employed to train deep neural networks (DNNs).
• The DNNs were trained using a combination of semi-empirical relationships (RK models), molten salt thermal properties database, and JARVIS database descriptors.
• The models predicted the density of molten salts.

Main Results:

• Deep neural networks (DNNs) achieved an R-squared value over 0.99.
• The mean absolute percentage error for DNN predictions was under 1%.
• The DNN approach demonstrated superior performance compared to alternative methods.

Conclusions:

• The proposed DNN transfer learning method effectively predicts molten salt density.
• This approach overcomes limitations of existing models by enhancing generalizability and addressing data sparsity.
• The findings pave the way for more accurate and efficient design of molten salt applications.