Members of the Coley Group at MIT have developed a novel generative AI system called FlowER (Flow matching for Electron Redistribution) that significantly improves the prediction of chemical reactions by embedding fundamental physical laws—like conservation of mass and electrons—into the model. Traditional AI models, including large language models, often fail to respect these constraints, sometimes "creating" or "deleting" atoms in their predictions. FlowER addresses this by explicitly tracking electrons throughout a reaction using a bond-electron matrix, ensuring that predictions remain chemically and physically plausible. The work was reported in the journal Nature.

The system, trained on over a million reactions from U.S. patent data, is still in its early stages but already matches or outperforms existing models in predicting reaction mechanisms. While it currently lacks coverage of metal-catalyzed reactions, the team sees this as a promising foundation for future expansion. Importantly, the model and its datasets are open source, enabling broader use in fields like medicinal chemistry, materials science, and atmospheric chemistry. This work, supported by the NSF and the Machine Learning for Pharmaceutical Discovery and Synthesis consortium, represents a significant step toward AI-assisted discovery of new chemical reactions and mechanisms.

The MIT SuperCloud and Lincoln Laboratory Supercomputing Center provided HPC resources for the researchers’ results.