| Abstract Scope |
The deployment of molten salt reactors and salt-based energy systems requires a comprehensive understanding of the intrinsically linked chemistry, thermophysical properties, and thermodynamic behavior of salt mixtures over a wide range of compositions and thermodynamic conditions. This complexity demands extensive experimentation and predictive simulation, which with current analysis methods, presents significant scientific and logistical challenges. The high-temperature requirements, environmental sensitivity of salts, and their complex structure-property relationships make precise, fundamentally informed analysis difficult, costly and time-consuming. Meanwhile, predictive simulations offer atomistic-level chemical insight but are computationally intensive and rely on system-specific approximations that require careful validation.
To overcome these limitations, we propose a new framework that enables accurate, high-throughput, and chemically insightful property prediction. This approach integrates state-of-the-art machine learning-based atomistic simulations, supervised ab initio-informed property prediction, and unsupervised generative modeling to guide the prediction of molten salt properties across a wide compositional and thermodynamic landscape. Our recent results demonstrate that this method achieves experimental-level accuracy, scales across the periodic table, and reveals underlying chemistry-structure-property relationships that are interpretable and broadly applicable. The talk will present these results, discuss current limitations, and outline future opportunities for expanding the scope and impact of machine learning in molten salt research. |