MIT researchers develop FlowER, a generative AI model predicting chemical reactions with higher accuracy by conserving mass and electrons.

Massachusetts, United States:  MIT researchers have developed a new generative AI model, FlowER (Flow matching for Electron Redistribution), that predicts chemical reactions while adhering to fundamental physical laws such as the conservation of mass and electrons. The breakthrough, published in Nature on August 20, significantly improves the accuracy and reliability of reaction outcome predictions.

“The prediction of reaction outcomes is a very important task,” says lead author Joonyoung Joung. Unlike earlier models that often “make new atoms, or delete atoms in the reaction,” FlowER uses a bond-electron matrix representation to explicitly track all atoms and electrons, ensuring physically realistic outputs.

Developed by a team led by MIT’s Connor Coley, the system was trained on more than a million reactions from the U.S. Patent Office database. Early results show that FlowER matches or outperforms existing approaches in identifying mechanistic pathways while maintaining scientific validity.

The open-source platform is freely available on GitHub, providing access to both the model and datasets. “It’s all open source,” notes co-author Mun Hong Fong.

Though still in early stages and limited in its coverage of metal-based and catalytic reactions, FlowER has broad potential applications in drug discovery, materials science, atmospheric chemistry, and electrochemical systems. “We’re incredibly excited about the fact that we can get such reliable predictions of chemical mechanisms,” Coley says.

The project was supported by the Machine Learning for Pharmaceutical Discovery and Synthesis consortium and the National Science Foundation.

According to TechSci Research, the development of MIT’s FlowER model holds significant benefits for the chemical industry by improving the accuracy and reliability of reaction predictions. By ensuring strict adherence to physical principles such as conservation of mass and electrons, the system minimizes errors often seen in conventional AI-driven models, enabling more realistic outcomes. This advancement can accelerate drug discovery, materials development, and process optimization by allowing researchers to map reaction pathways with greater precision and reduced experimental trial-and-error. Its open-source availability also promotes wider adoption and collaborative innovation across academia and industry. Additionally, the ability to generalize predictions to previously unseen reaction types offers strong potential for uncovering new chemical processes, enhancing efficiency, and reducing costs in R&D. Over time, this approach could become a critical tool for chemical manufacturers seeking to streamline product development while maintaining scientific rigor and reliability.