Automated Development of Molten Salt Machine Learning Potentials: Application to LiCl

Area of Science:

• Computational materials science
• Physical chemistry
• Energy applications

Background:

• Accurate in silico modeling of molten salts is essential for developing carbon-free energy technologies.
• High computational cost of quantum mechanical methods limits the study of highly polarizable molten salts.

Purpose of the Study:

• To develop an accelerated and automated methodology for generating accurate interatomic potentials for molten salts.
• To enable large-scale simulations of molten salts for improved understanding and design.

Main Methods:

• Integration of configurational sampling with classical force fields and active learning.
• Automated generation of Gaussian approximation potentials (GAP) with reduced ab initio calculations (O(100)).
• Parametrization and application of a molten lithium chloride (LiCl) GAP model.

Main Results:

• Achieved a 19,000-fold speedup compared to Ab Initio Molecular Dynamics (AIMD) for molten LiCl.
• Generated a highly accurate GAP model for molten LiCl.
• Provided new physical insights into molten LiCl's coordination structure and validated predictions for properties like density and ionic conductivity.

Conclusions:

• The developed active learning approach significantly reduces the computational cost for creating accurate molten salt models.
• This methodology facilitates in silico understanding and design of molten salts for diverse applications.
• Lowers the barrier for simulating molten salts across the periodic table.