A transferable active-learning strategy for reactive molecular force fields.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Predictive molecular simulations need fast, accurate, and reactive interatomic potentials.
• Machine learning (ML) can construct these potentials by fitting quantum-mechanical data.
• Current ML methods often require significant human input and large datasets.

Purpose of the Study:

• To develop an autonomous method for creating accurate Gaussian Approximation Potential (GAP) models.
• To reduce the data volume and human intervention needed for ML potential development.
• To enable routine generation of ML force fields for reactive molecular systems.

Main Methods:

• Leveraging hierarchical and active learning strategies.
• Employing separate intra- and inter-molecular fits for GAP models.
• Utilizing a prospective error metric to assess potential accuracy.

Main Results:

• Accurate GAP models were developed autonomously for diverse chemical systems.
• The method required only hundreds to a few thousand energy and gradient evaluations.
• Demonstrated applications in bulk solvents, solvated ions, metallocages, Diels-Alder reactions, and S_N2 reactions.

Conclusions:

• The developed approach enables autonomous generation of accurate ML force fields.
• This method significantly reduces the data requirements for developing reactive potentials.
• Provides a viable route for routine creation of ML potentials for complex chemical systems.