MIT researchers have developed an approach based on machine learning that can calculate transition states of chemical reactions within seconds. The structures of these transition states, a temporary condition in the middle of a chemical reaction, can typically only be calculated using techniques based on quantum chemistry – a process that can be extremely time-consuming. The newly developed model could help chemists design new reactions and catalysts for the generation of beneficial products and for modelling natural chemical reactions.
The team’s model learns to represent two or more reactants in arbitrary orientations based on training data from around 9,000 different chemical reactions. Once trained, the model can propose a transition state structure for new reactant and product pairs. It was accurate to within 0.08 angstroms when tested against transition states from unseen reactions generated using quantum techniques.
The researchers primarily trained their model on reactions involving compounds with a relatively small number of atoms, with a total of 23 atoms for the entire system. However, they discovered that the model could also make accurate predictions for reactions involving larger molecules.
The team intends to develop the model further to include other components, such as catalysts. This could be beneficial for developing new ways to generate pharmaceuticals, fuels, and other useful compounds, particularly when the synthesis includes multiple chemical steps.
This model could also be applied to explore potential gas interactions on other planets. Alternatively, it could be used in modeling simple reactions that may have occurred during the early evolution of life on Earth.
Jan Halborg Jensen, a professor of chemistry at the University of Copenhagen who was uninvolved in the research, hailed the development as a significant step forward in predicting chemical reactivity, potentially automating one of the hardest tasks in the field.
