Chemical reactions reach a ‘transition state’ when molecules gain enough energy for the reaction to proceed. This state is brief and hard to observe experimentally. The arrangement of these transition states can be calculated through quantum chemistry, but it is highly time-consuming. Scientists at MIT have developed a faster method using machine learning which computes these structures within a few seconds.
This model could support chemists in designing new reactions and catalysts to generate beneficial products, such as fuels or drugs. It could allow for the modelling of naturally occurring chemical reactions like those potentially driving the evolution of life on Earth. Chenru Duan, a PhD student and lead author behind this development, stated that this new model views the transition state as vitally important to understanding and designing catalysts and to comprehend how natural systems enable certain transformations.
The transition state in a chemical reaction is important as it determines the likelihood of a chemical transformation. Transition states can be calculated through density functional theory in quantum chemistry. However, this method is energy-intensive, taking potentially several days to compute. Some researchers have previously tried to utilise machine learning to determine transition state structures, but models developed until now have needed the two reactants to remain in the same orientation to each other. Any differing orientations would need separate reactions to be modelled, leading to increased computation time.
The team at MIT have developed a new computational methodology where two reactants can be placed in any orientation to each other in a diffusion model. As testing data for their model, the research team used structures of reactants, products, and transition states for 9,000 different chemical reactions. The model matched the transition state structure that pairs with the reactants and products. The new MIT approach was tested on around 1,000 reactions which it had not seen before and was able to generate 40 possible transition state solutions, computed within a few seconds for each reaction.
The model was primarily tested on reactions involving compounds with up to 23 atoms, but it was found that it could produce accurate predictions for reactions involving larger molecules. The team plans to develop their model to include other components such as catalysts. This could be useful for creating new manufacturing methods for medicines, fuels, or other beneficial compounds. The model could also have potential use in exploring potential reactions between gases on other planets or modelling early reactions that may have impacted early life on Earth.
