Flow Matching for Reaction Pathway Generation

Elucidating reaction mechanisms requires efficient generation of transition states (TSs), products, and complete reaction networks. While diffusion models and sequence-based generators accelerate TS and product sampling, respectively, current pipelines still depend on manual enumeration for at least one of these steps. Moreover, the stochastic dynamics underlying diffusion models can be inefficient and difficult to control. We show that flow matching, a deterministic ordinary differential equation formulation, can replace stochastic differential equation-based diffusion for reaction generation, and introduce MolGEN, a conditional flow-matching framework that learns an optimal-transport path mapping Gaussian priors to target chemical distributions. On benchmarks used by TSDiff, OA-ReactDiff, and ReactOT, MolGEN improves TS geometry accuracy and barrier-height prediction while enabling sub-second sampling. For product generation, it achieves competitive top-$k$ accuracy and avoids mass/electron-balance violations common to sequence models. Crucially, MolGEN unifies TS and product sampling within a single backbone, enabling reaction network exploration driven by generative inference rather than templates or repeated quantum-chemistry TS searches. In a realistic test on the $\gamma$-ketohydroperoxide decomposition network, MolGEN yields more valid/intended TSs than string-based methods while drastically reducing the number of quantum-chemistry evaluations from 1,156 to 12. Furthermore, MolGEN uncovers a novel pathway with a barrier lower than the reported minimum, demonstrating an accurate and efficient framework for generative reaction-network exploration.