Electron flow matching: a generative framework for mass- and electron-conserving reaction mechanism prediction
발표자
정준영 (국민대학교)
연구책임자
정준영 (국민대학교)
공동저자
정준영 (국민대학교)
초록
내용
Accurate prediction of chemical reactivity requires adherence to mass and electron conservation, which is often neglected by data-driven reaction models. We present FlowER (Flow matching for Electron Redistribution), a generative framework that formulates reaction prediction as electron redistribution using the bond–electron matrix. FlowER explicitly conserves atoms and electrons at each elementary step and aligns with the arrow-pushing formalism. Built on the flow matching paradigm, it learns probabilistic trajectories from reactants to products and generates full mechanistic pathways while enforcing physical constraints. FlowER achieves performance comparable to sequence-based models, avoids unphysical hallucinations, and demonstrates strong data efficiency on unseen reaction classes. Its mechanistic fidelity enhances interpretability and enables integration with quantum chemical analysis.