During a chemical reaction, molecules gain energy until they reach a point known as the transition state, a pivotal moment where the reaction must proceed. The structures of these states can be determined using quantum chemistry methods, but these calculations are time-intensive. To tackle this issue, a team of MIT researchers developed a machine learning-based model that can calculate these structures rapidly, within seconds. This breakthrough could assist chemists in creating new catalysts and reactions, and model essential natural chemical reactions.
The transition state during a chemical reaction signifies the energy peak needed for the reaction to continue. Understanding the structure of this state is crucial both for designing catalysts and for understanding natural transformations. However, calculating transition states using density functional theory, a quantum chemistry method, requires significant computing power and a lot of time, sometimes even days. While some researchers have attempted to use machine learning to discover transition state structures, these models still lack accuracy and efficiency.
To address this, the MIT researchers developed a computational approach using a diffusion model that is capable of representing two reactants in any orientation. The model was trained with the structures of reactants, products, and transition states determined through quantum computations for 9,000 distinct chemical reactions. The model could then generate a relevant transition state structure for given reactants and products. When tested on 1,000 unseen reactions, the model generated 40 possible solutions for each transition state and predicted which states were most likely to occur. These solutions matched the quantum generated structures with remarkable accuracy and in a significantly shorter time-frame.
Primarily trained on reactions involving smaller compounds, the model also demonstrated accurate predictions for larger molecules. The researchers now aim to augment their model to include catalysts and thus help ascertain how much a particular catalyst could expedite a reaction. This could be particularly beneficial for developing new procedures for producing pharmaceuticals, fuels, and other valuable compounds. Another potential application is modelling the interactions between gases on other planets or the early evolution of life on Earth.
The research was funded by the U.S. Office of Naval Research and the National Science Foundation. This novel method indicates considerable progress in predicting chemical reactivity, automating the hardest task in the field: finding the transition state of a reaction and the related barrier. As a result, this could eliminate a significant bottleneck in computational catalyst and reaction discovery.
