A new generative AI approach to predicting chemical reactions (news.mit.edu)

Researchers at MIT have developed a novel generative AI model, called FlowER, that significantly improves the accuracy of predicting chemical reaction outcomes by explicitly incorporating fundamental physical principles such as the conservation of mass and electrons. Unlike previous AI approaches relying solely on input-output data, FlowER uses a bond-electron matrix representation to track electrons and atoms throughout the reaction process, preventing the generation of physically impossible products. This advancement addresses a key limitation of standard large language models, which may "invent" or omit atoms during predictions due to their token-based architecture. FlowER builds on a 1970s chemist Ivar Ugi’s method to represent electron distributions and was trained on over a million reactions from a U.S. Patent Office database, grounding predictions in experimentally validated data. Although still in early stages and currently limited in handling certain metal- or catalyst-involving reactions, FlowER already matches or outperforms existing models in predicting mechanistic pathways with higher validity and physical consistency. The open-source system is poised to aid chemists in drug discovery, materials science, and other fields by providing more reliable assessments of chemical reactivity and reaction mechanisms. This breakthrough marks a pivotal step toward AI systems that not only predict reaction outcomes but also elucidate underlying mechanisms in a scientifically rigorous way. The team aims to expand FlowER’s capabilities to include catalytic cycles and diverse chemistries, heralding a new era of AI-augmented chemical innovation with long-term potential to accelerate the discovery of novel reactions and synthesis routes.