In a breakthrough study at MIT, researchers have used machine learning (ML) to calculate the ephemeral transition state in chemical reactions, representing a significant step forward for computational chemistry. The transition state occurs when molecules in a reaction gain energy to the point where the reaction becomes irreversible. Researchers have struggled to observe this pivotal process as it is so transient.
Traditionally, the structures of these transition states could be derived from complex quantum chemistry methods, such as density functional theory. However, these techniques require significant amounts of computing power and can take vast amounts of time. Some scientists have employed ML models to unearth these transition state structures, but these models viewed two reactants as a single entity, maintaining their orientation to each other. This approach failed to account for the flexibility of molecules and impedes the accuracy of calculations.
The team from MIT, led by Heather Kulik, Associate Professor of Chemistry and Chemical Engineering, and Chenru Duan, a PhD student at the institute, have addressed these limitations through a novel computational approach. Their ML model employed an innovative diffusion technique to represent two reactants in any arbitrary orientation with respect to each other. Enlisting quantum computation methods, the researchers used structures of over 9,000 different chemical reactions as training data for their model. The tool then learned the distribution pattern of reactant, product, and transition state structures, enabling it to generate a transition state structure from new reactant-product pairs.
Tests of the model yielded results accurate to within 0.08 angstroms compared to transition state structures obtained using quantum techniques, demonstrating an exceptional level of precision. Unlike the conventional method, which can take days, this groundbreaking technique delivers results within a matter of seconds. While primarily training their ML on reactions involving compounds of up to 23 atoms, the researchers found the model could accurately predict reactions of larger compounds.
Looking ahead, the MIT team plans to augment their model to include catalysts, which might shed light on ways to expedite reactions. This could provide a meaningful contribution to the development of new processes for generating pharmaceuticals, fuels and other beneficial compounds. Furthermore, the model could facilitate the exploration of chemical interactions that might occur between gases discovered on other planets, or simulate simple reactions that could have occurred during the inception of life on Earth.
This pioneering investigation represents a significant forward leap in predicting chemical reactivity, according to Jan Halborg Jensen, a chemistry professor at the University of Copenhagen. He noted the study could eliminate key challenges that inhibit important fields such as computational catalyst and reaction discovery. The research was funded by the U.S. Office of Naval Research and the National Science Foundation.
