A group of MIT researchers has developed a new machine learning model which rapidly calculates the structure of transition states during chemical reactions. This fleeting moment is a crucial “point of no return” in reactions. Although this transition state is vital to understanding the pathway of the reaction, it has been notoriously difficult to observe due to its brief duration and the significant energy required to calculate it using quantum chemistry. This new technique streamlines the process, speeding up calculations from several hours or even days to a matter of seconds.
Associate Professor of Chemistry and Chemical Engineering at MIT, Heather Kulik, who is the senior author of the study, stated that understanding the transition state structure is a critical starting point for creating catalysts or understanding how natural systems enact transformations. This method could, therefore, help researchers to devise new reactions and catalysts to produce useful substances such as fuels or drugs, or to model natural chemical reactions.
Forging ahead with traditional quantum chemistry methods requires a vast amount of computing power and a significant investment of time. The team at MIT has addressed this with a novel computational approach using a diffusion model. This model can comprehend the different potential orientations of two reactants independently, making the machine-learning training aspect quicker and more accurate. “Once the model learns the underlying distribution of how these three structures coexist, we can give it new reactants and products, and it will try to generate a transition state structure that pairs with those reactants and products,” explains Chenru Duan, PhD ‘22 and lead author of the study.
The model was tested on 1,000 previously unencountered reactions, generating 40 potential solutions for each. These solutions were then ranked in terms of likelihood by a “confidence model,” and the results yielded an impressive degree of accuracy.
Currently, the model is designed to work with compounds containing up to 23 atoms, but it can also make accurate predictions with larger molecules. The MIT team intends to extend the abilities of the model to include other factors such as catalysts, which could be beneficial in the creation of new procedures for generating pharmaceuticals, fuels, or other useful compounds.
“This problem is holding back many important fields such as computational catalyst and reaction discovery, and this is the first paper I have seen that could remove this bottleneck” says Jan Halborg Jensen, a professor of chemistry at the University of Copenhagen, who was not involved in the research. The findings could fundamentally change and speed up the process of chemical reactivity prediction, opening up a whole new world of possibilities for chemical discovery. This model could also be used to explore the interactions between gases on other planets or to simulate the early evolution of life on earth.
The study was funded by the U.S. Office of Naval Research and the National Science Foundation and appears in Nature Computational Science.
