Scientists from the Massachusetts Institute of Technology have used machine learning to expedite calculations of transition states in chemical reactions, a process that could support the invention of new reactions and catalysts with applications in fuels, pharmaceuticals and understanding the origins of life. Using a method known as density functional theory to compute transition states can prove extremely time-consuming, although efforts to leverage machine learning to determine structures can still demand high-levels of calculation to account for the different possible orientations of two reactants. The MIT researchers, however, devised a computational model in which two reactants can be represented in any arbitrary orientation relative to each other. Training data for the model included reactant, product and transition state structures for 9,000 reactions previously calculated using quantum computation methods. The model can then use new reactants to produce a transition state structure. The MIT scientists tested the model, asking it to generate 40 solutions to each transition state for 1,000 different reactions, then using a “confidence model” to calculate the most likely states to ensue. The model’s results were shown to be consistent with structures produced using quantum approaches, and took just seconds to calculate.
Skip to content
Skip to footer
