MIT researchers have developed a new generative AI approach called FlowER (Flow matching for Electron Redistribution) to predict chemical reactions with improved accuracy and reliability. This system incorporates fundamental physical principles, such as the laws of conservation of mass, to ensure that predictions are physically realistic.
FlowER uses a matrix to represent electrons in a reaction, tracking bonds and lone electron pairs to conserve atoms and electrons simultaneously. This approach enables the model to predict reaction outcomes with high accuracy, matching or outperforming existing approaches in finding standard mechanistic pathways and generalizing to previously unseen reaction types.
The FlowER model has potential applications in medicinal chemistry, materials discovery, combustion, and atmospheric chemistry. In medicinal chemistry, for instance, predicting reactions can help with drug discovery and development. In materials discovery, forecasting reactions can aid in materials synthesis and design.
While FlowER shows promise, it has limitations. The model was trained on over a million chemical reactions from a U.S. Patent Office database, but this data does not include certain metals and catalytic reactions. Researchers plan to expand the model's understanding of metals and catalytic cycles and further improve its accuracy and robustness.
