MIT researchers unveiled an open-source generative AI system that predicts chemical reaction mechanisms while enforcing core physical laws such as conservation of mass and electrons. The model, dubbed FlowER and detailed in Nature, represents reactions using a bond–electron matrix to track electron redistribution, reducing the spurious atom gains and losses that can plague language-model approaches. Trained on more than a million patent-derived reactions, the tool matches or tops current methods on standard benchmarks and can generalize to unfamiliar reaction types, the team said. Still, the system has gaps—especially for metal-containing and catalytic chemistries—and will need broader data to become an industry workhorse. Funded by the Machine Learning for Pharmaceutical Discovery and Synthesis consortium and the National Science Foundation, FlowER is positioned as a practical aid for mapping pathways in drug discovery, materials science, combustion, and electrochemistry, with the authors framing this release as a proof-of-concept step toward more ambitious reaction invention.
Related articles:
ASKCOS: Open-source AI for chemical synthesis planning
Flow Matching for Generative Modeling
Highly accurate protein structure prediction with AlphaFold
ChemCrow: Augmenting large-language models with chemistry tools
Scalable discovery of materials with GNoME
