Uncertainty-driven dynamics for active learning of interatomic potentials.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning (ML) models trained on quantum simulations yield accurate interatomic potentials.
• Active learning (AL) iteratively builds diverse datasets by selecting configurations based on ML model uncertainty.

Purpose of the Study:

• To develop a strategy for rapidly discovering configurations that significantly enhance ML training datasets.
• To improve the efficiency of active learning in exploring complex chemical configuration spaces.

Main Methods:

• Introduced uncertainty-driven dynamics for active learning (UDD-AL).
• Modified the potential energy surface in molecular dynamics to prioritize regions of high model uncertainty.
• Applied UDD-AL to sample glycine conformations and acetylacetone proton transfer.

Main Results:

• UDD-AL efficiently identifies and samples configurations that meaningfully augment training data.
• Demonstrated effective exploration of chemically relevant configuration spaces.
• Showcased improved sampling compared to traditional dynamical methods.

Conclusions:

• UDD-AL accelerates the generation of high-quality datasets for ML potentials.
• The method enables efficient exploration of complex chemical landscapes, including those inaccessible to standard simulations.
• UDD-AL holds promise for advancing materials discovery and chemical process simulation.