A team of researchers from the Massachusetts Institute of Technology (MIT) has developed a machine learning model that can quickly calculate the structures of transition states in chemical reactions. These fleeting moments occur when molecules have gained enough energy to proceed with a reaction, but are notoriously difficult to study due to their ephemeral nature. Quantum chemistry can calculate these states, but this is a time-consuming process. The MIT team’s model, however, can reportedly do so within a few seconds, potentially helping chemists in the design of new reactions and catalysts for everything from fuel to pharmaceutical development.
Understanding the structures of transition states is beneficial for understanding and developing catalysts or understanding natural system processes. Chenru Duan PhD ’22 is the lead author of the published paper, with contributions from MIT graduate student Haojun Jia and Cornell University graduate student Yuanqi Du.
Traditional quantum chemistry methods for calculating transition states require huge computing power, taking hours or even days to calculate a single transition state. Machine learning models developed previously had to consider two reactants as a single entity “If the reactant molecules are rotated, then the reaction is considered a different one. The MIT team’s computational approach, based on a diffusion model, allows for reactants in any orientation.
The MIT team’s model was trained on structures of reactants, products and transition states for 9,000 different chemical reactions. It learned the distribution patterns of the three structures coexist, and could rapidly generate a transition state structure for reactants and products. It was tested on 1,000 unfamiliar reactions. Predicted states were comparable in accuracy to quantum techniques, with only a tiny discrepancy in predicted results and the entire process taking just seconds.
The model was primarily trained on reactions involving compounds with a relatively small number of atoms, but also provided accurate predictions for reactions involving larger molecules. Kulik commented that even in bigger systems or enzyme-catalysed systems, the model predicts the likely atomic rearrangements accurately.
The team plans to enhance their model by incorporating catalysts, potentially aiding the development of new procedures for pharmaceutical, fuel and other useful compounds production. The model could replace quantum chemistry in these calculations. Investigating gas interactions on other planets and modelling simple early reactions that may have occurred during the early evolution of life on Earth could also benefit from this model. Jan Halborg Jensen, a University of Copenhagen professor of chemistry not involved in the study, said the solution could remove a significant bottleneck inhibiting many vital fields such as computational catalyst and reaction discovery.
