The world of artificial intelligence has made astronomical leaps in various fields, but there seems to be an Achilles heel when it comes to predicting chemical reaction outcomes. These underwhelming results can often be traced back to the lack of an association with basic physics principles, specifically the conservation of mass and electrons. But this might soon change, thanks to the efforts of the Massachusetts Institute of Technology (MIT).
A recent study spearheaded by MIT researchers, was capable of designing an AI model named FlowER (Flow matching for Electron Redistribution) that ingeniously incorporates physical constraints into its predictions. “The prediction of reaction outcomes is a very important task. If you want to make a new drug, you need to know how to make it”, says former MIT postdoc, Joonyoung Joung, who’s now an assistant professor at Kookmin University in South Korea.
Though capable, existing chemical Large Language Models (LLMs) are found wanting in one respect – without proper constraints placed upon them, they tend to “invent” atoms in a manner that overtly disregards physics laws. The MIT team sought to remedy this by ensuring their AI system, FlowER, can meticulously track every atom and electron from the beginning to the end of the reaction.
Their solution was found in a four-decade-old concept: a matrix-based representation developed by chemist Ivar Ugi in the 1970s. With this utility, the model can efficiently monitor both atoms and electrons throughout a reaction.
In its infancy, FlowER has already started showing signs that it’s a cut above the rest. According to Connor Coley, the senior author and an MIT professor, the AI model rivals or even outmatches existing systems in predicting standard reaction mechanisms, all while maintaining physical validity.
Yet, the researchers didn’t stop at theoretical successes. Ensuring their AI model aligned closer to reality, they validated their findings with experimental data sourced from patent literature. “We’re imputing mechanisms from experimental data, and that’s not something that has been done and shared at this kind of scale before” Coley points out.
FlowER is currently available as open-source software on GitHub for those who wish to utilize it. This includes a dataset created by Joung that meticulously details the mechanistic steps of known reactions, a resource believed to be first of its kind.
The applications of this AI method are nothing short of far-reaching. While FlowER still has some way to go in perfecting its predictions, especially involving metal-based or catalytic reactions, the ongoing research is anticipated to bear fruit that will be beneficial across various sectors: medicinal chemistry, materials science, combustion, atmospheric chemistry, and electrochemistry.
As Coley puts it, “We’ve just scratched the surface. A lot of the excitement is in using this kind of system to help discover new complex reactions and help elucidate new mechanisms.”
Read more about the research on the MIT News website.
So while the gap between AI and fundamental physics might have been a longstanding issue, projects like FlowER from MIT represent a hopeful future where AI is deeply integrated with science, helping unlock its full potential in diverse fields like chemistry.
This website uses cookies.