Generating Minimal Training Sets for Machine Learned Potentials.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Training accurate machine-learned interatomic potentials (MLPs) typically requires extensive datasets.
• Ab initio calculations, while accurate, are computationally expensive, limiting dataset size.
• Existing methods for selecting training data can be inefficient.

Purpose of the Study:

• To develop a novel and efficient workflow for training MLPs using minimal data.
• To reduce the computational cost associated with generating training datasets for MLPs.
• To enable the use of more accurate quantum mechanical calculations in materials modeling.

Main Methods:

• Introduced Random Network Distillation (RND), a nonstandard neural network workflow.
• Coupled RND with a Density Functional Theory (DFT) workflow, using classical methods for initial data generation.
• Selected a minimal subset of configurations for expensive ab initio calculations.

Main Results:

• Demonstrated efficacy by constructing MLPs for molten salts KCl and NaCl.
• Achieved accurate models with minimal datasets as small as 32 configurations.
• Reduced required dataset sizes by at least one order of magnitude compared to alternatives.

Conclusions:

• RND significantly reduces computational overhead for training data generation.
• The method provides a pathway to utilizing more accurate quantum mechanical calculations.
• RND offers a more comprehensive starting point for active learning procedures in materials science.