Scientists at MIT have devised a machine learning-based method that can rapidly calculate the transitional states of molecules during a chemical reaction. The transient nature of these states has made observation particularly challenging. Understanding these states is key to developing catalysts or deciphering how natural systems induce specific changes.
The MIT team constructed their computational approach utilizing a diffusion model, which is capable of discerning which processes are most likely to result in a particular outcome. This model was trained on structures of transition states, reactants, and products associated with 9,000 distinct chemical reactions. The researchers found the model’s predictions were accurate to within one hundred millionth of a centimeter compared to those acquired through quantum techniques, with the computation process taking just a few seconds per reaction.
The model was primarily trained on reactions with a limited number of atoms – no more than 23 atoms for the entire system. However, accurate predictions were also produced for reactions involving larger molecules, offering impressive coverage for different atom rearrangement scenarios. The team now intends to enhance their model to include other components such as catalysts.
The innovative machine-learning approach could have implications for generating pharmaceuticals and fuels or understanding biochemical reactions on earth and possible reactions on other planets. Jan Halborg Jensen, a chemistry professor at the University of Copenhagen, praised the development as a significant advance in predicting chemical reactivity and potentially overcoming one of the main obstacles holding back computational catalyst and reaction discovery.
