MIT’s FlowER AI predicts chemical reactions while conserving physics

MIT researchers have developed FlowER (Flow matching for Electron Redistribution), a new generative AI system that predicts chemical reactions while maintaining fundamental physical constraints like conservation of mass and electrons. The breakthrough addresses a critical limitation in existing AI models that often violate basic scientific principles by creating or destroying atoms during reaction predictions, potentially revolutionizing drug discovery and materials science.

The big picture: Previous AI attempts at reaction prediction have struggled because they ignore fundamental physics, leading to scientifically impossible outcomes where atoms appear or disappear during reactions.

• Large language models like ChatGPT use computational “tokens” to represent atoms, but without proper constraints, “the LLM model starts to make new atoms, or deletes atoms in the reaction,” according to lead researcher Joonyoung Joung, a recent MIT postdoc now at Kookmin University in South Korea.
• “This is kind of like alchemy,” Joung explains, rather than being grounded in real scientific understanding.

How it works: FlowER uses a 1970s method developed by chemist Ivar Ugi that employs a bond-electron matrix to track every electron throughout a chemical reaction.

• The system represents electrons in a matrix format, using nonzero values for bonds or lone electron pairs and zeros for their absence.
• “That helps us to conserve both atoms and electrons at the same time,” says Mun Hong Fong, former MIT software engineer now at Duke University.
• Unlike traditional models that only examine inputs and outputs, FlowER tracks “all the chemicals, and how the chemicals are transformed” throughout the entire reaction process.

Key performance metrics: The model was trained on over one million chemical reactions from U.S. Patent Office databases and demonstrates significant improvements over existing approaches.

• FlowER matches or outperforms current systems in finding standard mechanistic pathways while generalizing to previously unseen reaction types.
• “Using the architecture choices that we’ve made, we get this massive increase in validity and conservation, and we get a matching or a little bit better accuracy in terms of performance,” says senior author Connor Coley, the Class of 1957 Career Development Professor at MIT.

Current limitations: The system represents a proof of concept with specific constraints that researchers acknowledge.

• The training data lacks certain metals and catalytic reactions, limiting the breadth of chemistries the model can handle.
• “We certainly acknowledge that there’s a lot more expansion and robustness to work on in the coming years,” Coley notes.

What’s next: The team plans to expand the model’s capabilities to include more complex chemical systems.

• Researchers are “quite interested in expanding the model’s understanding of metals and catalytic cycles” since most current reactions don’t include these elements.
• Long-term goals include using the system “to help discover new complex reactions and help elucidate new mechanisms.”

Open access impact: The entire FlowER system is available as open-source software through GitHub, including models, data, and a comprehensive dataset of known reaction mechanisms.

• “I think we are one of the pioneering groups making this dataset, and making it available open-source, and making this usable for everyone,” Fong explains.
• The model could potentially benefit medicinal chemistry, materials discovery, combustion research, atmospheric chemistry, and electrochemical systems.

What they’re saying: The research team emphasizes the importance of grounding AI predictions in experimental validation.

• “What’s unique about our approach is that while we are using these textbook understandings of mechanisms to generate this dataset, we’re anchoring the reactants and products of the overall reaction in experimentally validated data from the patent literature,” Coley explains.
• “We’re imputing them from experimental data, and that’s not something that has been done and shared at this kind of scale before.”