During a chemical reaction, molecules gain energy until they reach what is known as the transition state — a point at which the reaction must proceed. This state is extremely short-lived and nearly impossible to observe experimentally. Its structures can be calculated using quantum chemistry techniques, but these methods are very time-consuming.
Recently, a team of MIT researchers has developed a new approach, utilizing machine learning, to calculate these structures much more rapidly. This machine learning model can potentially calculate transition state structures within a few seconds. This could be beneficial to chemists in designing new reactions and catalysts, useful in the creation of products like fuels or drugs. It could also be used to model chemical reactions that occur naturally and may have contributed to Earth’s progression of life.
The probability of any chemical reaction occurring is considerably influenced by the likelihood of the formation of the transition state. “The transition state helps to determine the likelihood of a chemical transformation happening… the transition state and how favorable that is determines how likely we are to get from the reactant to the product,” says Heather Kulik, an associate professor of chemistry and chemical engineering at MIT.
Currently, transition states are calculated using quantum chemistry methods, which require significant computing power and time. Some researchers have experimented with machine-learning models to discern transition state structures. Despite progress, these models currently see two reactants as one single entity in which the reactants keep the same orientation relative to each other.
To overcome this limitation, the MIT team developed a computational approach that allows two reactants to be represented in any arbitrary orientation relative to each other, using a diffusion model. This allowed them to calculate transition states for 9,000 different chemical reactions.
When tested on about 1,000 chemical reactions, the model was able to generate transition state solutions with high accuracy. Moreover, these solutions were estimated in just a few seconds of computation time for each reaction, which paled in comparison to the time consumed using traditional quantum methods.
Even though the researchers mainly trained their model on smaller reactions, they discovered it could also make accurate predictions for reactions involving larger molecules.
The researchers now aim to integrate other components such as catalysts into their model. This could potentially assist in the creation of new methods for manufacturing pharmaceuticals, fuels, and other valuable compounds. Another potential use for this model would be to explore interactions that could occur between gases on other planets or to model early-life Earth reactions.
Overall, this study represents an important advancement in predicting chemical reactivity. According to Jan Halborg Jensen, a professor of chemistry at the University of Copenhagen, it addresses one of the hardest parts of chemical reactivity prediction: finding the transition state and the associated barrier, a significant challenge in the automation of computational catalyst and reaction discovery. This work was funded by the U.S. Office of Naval Research and the National Science Foundation.
