Chemists often struggle to predict the outcome of a chemical reaction as it depends on the so-called “transition state,” a fleeting moment into which molecules enter and from which they can never return unchanged. The challenge lies in the fact that the transition state is extremely ephemeral and difficult to capture in real-world experiments.
This transition state can, in theory, be calculated using methods drawn from quantum chemistry. However, these are time-consuming and computing power-hungry, taking several days at a time for a single result. In recent years, scientists have started to use machine-learning models to predict transition states more efficiently. However, the existing models are somewhat limited as they treat the two reactants of a chemical reaction as a single entity, meaning they need to stick to one orientation during the reaction – an assumption which is unrealistic in many real-world scenarios.
Researchers from the Massachusetts Institute of Technology (MIT) and others have developed a fresh approach based on machine learning, which manages to calculate the structure of transition states extremely quickly – in seconds, in fact. This could significantly streamline and speed up both the design of new, useful chemical reactions and catalysts and the modeling of naturally-occurring chemical reactions, which play a key role in the evolution of life on earth.
The team’s new computational strategy enabled them to treat the two reactants as separate entities that can assume any orientation relative to each other. They used a diffusion model capable of identifying which processes are most likely to bring about a given result in order to capture this more fluid situation. The new model proved accurate, with errors as small as 8×10-10 cm, when tested on reactions involving never-seen-before compounds. Remarkably, the computational process took only a few seconds per reaction, making this an extremely fast and efficient method.
Although the research was primarily conducted on relatively simple compounds made of no more than 23 atoms, the model was equally successful when applied to larger, more complex molecules. The researchers now plan to extend their rapid prediction model to make it include other variables, such as the presence of catalysts. This will enable them to determine how much a catalyst would accelerate a reaction, guiding the design new processes for producing pharmaceuticals, fuels, or other useful compounds, and helping to model the type of simple reactions that might have occurred on the early Earth, or on other planets.
The groundbreaking nature of the research was confirmed by Professor Jan Halborg Jensen of the University of Copenhagen. Identifying the transition state of molecules, he said, is the most important, but also the most challenging task in predicting the outcome of chemical reactions. The MIT team’s work removes a significant bottleneck in this area, he argued, and might therefore significantly advance the field. The research was sponsored by the U.S. Office of Naval Research and the National Science Foundation.