MIT researchers have developed a machine learning-based technique that can rapidly calculate the structures of fleeting transition states during chemical reactions. Identifying and understanding these quasi-instantaneous moments, when molecules have collected enough energy to proceed with reaction, is crucial to fields such as catalyst design and natural system research. With traditional quantum chemistry-based techniques, it takes a huge amount of time and computing power to calculate one transition state, making it difficult to observe them experimentally.
However, the MIT team’s alternative approach can compute these states in a few seconds. Using a model referred to as a diffusion model, the researchers have found they can represent two reactants in arbitrary orientations to each other. The representation was trained on structures of reactants, products, and transition states calculated via quantum methods from 9,000 different chemical reactions. Once trained, the model can be presented with new reactant and product structures and asked to generate plausible transition state structures.
The scientists tested their model on approximately 1,000 unseen reactions, asking it to generate 40 different transition state candidates for each reaction. They then used a confidence model to rank the states most likely to occur. The results were approximately 0.08 angstroms (one hundred-millionth of a centimeter) accurate compared to quantum-generated transition state structures. Each computation took a few seconds.
The research team believes their model is accurate even for larger molecules containing over 23 atoms. They plan to develop its capabilities to incorporate catalysts which would facilitate the investigation of speeding up chemical reactions for generating pharmaceuticals, fuels, or other products.
Additionally, the new approach has the potential to simulate the interactions between gases on other planets or to represent the simple reactions that may have transpired during the early evolution of life on Earth. Jan Halborg Jensen, a chemistry professor at the University of Copenhagen, unaffiliated with the research, has hailed it as a revolutionary step forward for predicting chemical reactivity.
