During a chemical reaction, molecules gain energy until they reach a position termed the transition state. Existing at a level of energy where the reaction has no choice but to proceed, the transition state’s transient nature makes it extremely difficult to observe experimentally. Its structures can be calculated using quantum chemistry-based techniques, but these are time-consuming. Now, MIT researchers have developed a machine learning-based alternative that can provide these calculations more swiftly, within just a few seconds.
The new model has numerous potential applications. It could help chemists design new reactions and catalysts for the creation of fuels, drugs, and other useful products. It could also model natural chemical reactions, such as those potentially involved in the evolution of life on Earth.
The concept of the transition state is integral to understanding the likelihood of a given chemical reaction’s occurrence. Determined by the formation likelihood of a transition state, this probability can be calculated using density functional theory, a quantum chemistry method. However, due to the significant computing power such calculations require, determining a single transition state can take hours or even days.
Some researchers are attempting to leverage machine learning models to identify transition state structures, but existing models require viewing two reactants as a unit where the reactants maintain the same orientation concerning each other. This restriction means any other potential orientations have to be modeled as separate reactions, which increases the computation time.
The MIT team’s new computational approach overcomes this, enabling the representation of two reactants in any arbitrary orientation concerning each other. Their diffusion model, trained on structures of reactants, products, and transition states based on quantum computation methods for 9,000 different chemical reactions, can predict the most likely processes to yield a particular outcome. In testing, it generated solutions accurate to within 0.08 angstroms compared to those produced by quantum techniques and completed the computational process in just seconds per reaction.
Although primarily trained on reactions involving relatively small compounds comprising up to 23 atoms, the model showed it can make accurate predictions for larger molecular reactions. The researchers plan to expand their model to encompass other components, such as catalysts and investigate their impact on reaction speed.
An exciting application of this model could be exploring interactions between other planet’s gases, or modeling the simple reactions potentially involved in life’s early evolution on Earth. Jan Halborg Jensen, a professor of chemistry at the University of Copenhagen not involved in the research, describes the team’s work as “a significant step forward in predicting chemical reactivity”, signaling a potential end to various fields’ current bottleneck. The research was funded by the U.S. Office of Naval Research and the National Science Foundation.
