Accelerating Data Set Population for Training Machine Learning Potentials with Automated System Generation and...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine Learning Interatomic Potentials (MLIPs) significantly enhance molecular dynamics simulations but require extensive, high-quality training data.
• Current data generation methods for MLIPs are often computationally expensive and require substantial user intervention.

Purpose of the Study:

• To introduce Strategic Configuration Sampling (SCS), an active learning framework designed to automate and optimize the generation of training data sets for MLIPs.
• To develop a method for constructing compact and comprehensive data sets efficiently, reducing the burden of ab initio calculations.

Main Methods:

• SCS utilizes automated workflows for generating and exploring systems through collections of molecular dynamics (MD) simulations with automatically configured parameters.
• Employs 'collaging' to dynamically assemble initial geometries from preceding simulation runs, enabling exploration of complex atomic environments.
• Incorporates parallel execution of exploration workflows, resource allocation based on computational complexity, and leverages pretrained MLIP models to guide MD simulations.

Main Results:

• Demonstrates the ability to generate compact and comprehensive data sets for MLIP training through automated, active learning.
• Case studies confirm the framework's versatility and effectiveness in accelerating MLIP deployment across various materials science applications.
• SCS provides a fully open-source, high-throughput solution for data generation, reducing the need for initial data sets.

Conclusions:

• Strategic Configuration Sampling (SCS) offers a robust and automated solution for generating high-quality training data for MLIPs.
• The framework significantly accelerates the application of MLIPs in materials science by streamlining the data acquisition process.
• SCS represents a key advancement in making MLIPs more accessible and efficient for scientific research.