Active learning meets metadynamics: automated workflow for reactive machine learning interatomic potentials.

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Machine-learned interatomic potentials (MLIPs) offer a cost-effective alternative to ab initio molecular dynamics (AIMD) for atomistic simulations.
• Current MLIP applications in reaction modeling are limited by the need for extensive training datasets to cover high-energy transition states.

Purpose of the Study:

• To present a data-efficient, automated workflow for training MLIPs for reaction modeling.
• To reduce the data requirement for MLIP training, needing only a few initial configurations without prior transition state knowledge.

Main Methods:

• Combined automated active learning with well-tempered metadynamics for targeted exploration of configuration space.
• Utilized data-efficient architectures like the linear Atomic Cluster Expansion.
• Applied the workflow to diverse organic reactions: S_N2, methyl shift, and glycosylation.

Main Results:

• Successfully trained accurate and stable MLIPs for all tested organic reactions.
• Demonstrated the workflow's effectiveness across different chemical environments and reaction complexities.
• Validated the strategy's versatility in modeling reactive processes.

Conclusions:

• The proposed automated workflow significantly enhances the efficiency of MLIP training for reaction modeling.
• This approach lowers the barrier to entry for using MLIPs in complex chemical simulations.
• The method shows broad applicability for accurately simulating reactive processes.