Efficient exploration of reaction pathways using reaction databases and active learning.

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Accurate and fast simulation of chemical reactions is crucial in computational chemistry.
• Machine learning interatomic potentials (MLIPs) offer near-quantum mechanical accuracy at reduced computational cost.
• The performance of MLIPs heavily relies on the quality and relevance of their training datasets.

Purpose of the Study:

• To develop and demonstrate novel approaches for training reactive MLIPs using reaction pathway information.
• To enable accurate prediction of reaction pathways and transition states for various chemical reactions.
• To establish efficient methods for constructing training sets that improve MLIP accuracy and transferability.

Main Methods:

• Training reactive MLIPs using datasets containing reactant, product, and transition state structures.
• Implementing an active learning procedure that utilizes only reaction endpoint structures to train MLIPs.
• Validating MLIPs against reference quantum mechanical calculations and established force fields like ReaxFF.

Main Results:

• Accurate prediction of reaction pathways and transition states for SN2 reactions using reaction datasets.
• Development of an active learning procedure that predicts reaction barriers within 20 meV of reference methods, even with limited data.
• Successful application of the active learning procedure to complex nucleophilic aromatic substitution and proton transfer reactions.

Conclusions:

• Two effective strategies for training reactive MLIPs based on reaction pathway data have been demonstrated.
• The active learning approach provides a data-efficient method for developing accurate MLIPs without requiring prior knowledge of reaction pathways.
• These methods facilitate the creation of large, transferable reactive MLIPs for diverse chemical simulations.