AL4GAP: Active learning workflow for generating DFT-SCAN accurate machine-learning potentials for combinatorial...

Area of Science:

• Computational materials science
• Chemical physics

Background:

• Machine learning interatomic potentials (MLIPs) offer a way to overcome the limitations of ab initio simulations.
• Efficient parameterization of MLIPs for complex systems like molten salts remains a challenge.

Purpose of the Study:

• To present AL4GAP, an automated active learning workflow for generating multicomposition Gaussian approximation potentials (GAP).
• To enable the creation of accurate MLIPs for diverse molten salt mixtures.

Main Methods:

• AL4GAP utilizes combinatorial chemical space definition, empirical parameterization for configurational sampling, active learning with density functional theory (DFT) calculations (SCAN functional), and Bayesian optimization for hyperparameter tuning.
• The workflow supports 11 cations and 4 anions, including heavy elements.

Main Results:

• Successfully generated five independent GAP models for binary molten salt mixtures (e.g., LiCl-KCl, NaCl-CaCl2, KCl-NdCl3, CaCl2-NdCl3, KCl-ThCl4).
• Achieved DFT-SCAN accuracy in predicting molten salt structures.
• Captured characteristic intermediate-range ordering in multivalent cationic melts.

Conclusions:

• AL4GAP provides a high-throughput method for generating accurate MLIPs for multicomposition molten salts.
• The generated GAP models can reliably predict the structural properties of diverse molten salt mixtures.