Flow matching for reaction pathway generation

Area of Science:

• Computational Chemistry
• Chemical Reaction Engineering
• Machine Learning in Chemistry

Background:

• Elucidating reaction mechanisms is crucial for chemical research and development.
• Traditional methods for generating transition states (TSs) and products are often inefficient and require manual intervention.
• Existing diffusion and sequence-based models offer improvements but have limitations in control and efficiency.

Purpose of the Study:

• To introduce MolGEN, a novel conditional flow-matching framework for efficient generation of transition states and reaction products.
• To enable template-free generative exploration of chemical reaction networks.
• To reduce the reliance on extensive quantum-chemistry calculations.

Main Methods:

• Utilized a conditional flow-matching framework with deterministic optimal transport.
• Mapped Gaussian priors to chemical distributions for generative tasks.
• Employed a unified backbone for both transition state and product sampling.

Main Results:

• MolGEN demonstrated improved transition state geometry and barrier-height prediction compared to diffusion models, with sub-second sampling.
• Achieved competitive top-k accuracy for reaction product generation while ensuring mass and electron balance.
• Significantly reduced quantum-chemistry evaluations for the γ-ketohydroperoxide decomposition network (12 vs. 1156).
• Identified a lower-barrier pathway in the studied reaction network.

Conclusions:

• MolGEN offers a highly efficient and accurate method for generating transition states and products, accelerating reaction mechanism elucidation.
• The framework enables template-free exploration of reaction networks, reducing computational burden.
• MolGEN represents a significant advancement in computational chemistry for reaction discovery and mechanism studies.