Skip to content Skip to footer

The computational model accurately represents the hard-to-detect transitional phases of chemical reactions.

An MIT research team has developed an approach that quickly calculates the structure of transition states fundamental in chemical reactions – the fleeting and typically unobservable point that determines whether a reaction proceeds. This new machine learning-based model could assist in developing new reactions and catalysts for creating materials like fuels or drugs, and might help model natural chemical reactions such as those that led to the evolution of life on Earth. The old method for calculating these states required substantial time and computing power, while the new model uses learning algorithms trained on thousands of past chemical reactions to make future predictions. It was accurate within 0.08 angstroms (one hundred-millionth of a centimeter) when compared to structures generated using the more cumbersome quantum techniques.

Typically, it is important for chemists to know the structure of the transition state in order to design catalysts because it helps to understand how natural systems bring about specific transformations. Currently, transition states are calculated using a quantum chemistry method known as density functional theory. But, this method needs massive computing power, taking many hours or even days to calculate just one transition state.

The current crop of machine learning models require the two reagents of the reaction to be thought of as one entity that maintains its relative orientation during the reaction. This, however, increases computation time as they cannot account for different orientations and therefore consider those as separate reactions. This makes machine learning training more difficult as well as less accurate.

The new method developed by the MIT team overcomes these limitations by considering two reactants in any arbitrary orientation to each other, using a diffusion model. As a result, it can learn which types of processes are most likely to generate a particular outcome. It was trained using structures of reactants, products and transition states that had been calculated using quantum computation methods for 9,000 different chemical reactions. The entire computational process using this model takes just a few seconds for each reaction.

Another benefit of this new approach is that it could be expanded to include other components such as catalysts, which could help investigate how much a catalyst would speed up a reaction. This could be useful when developing new processes for creating pharmaceuticals, fuels, or other useful compounds. Other potential applications of these models could be understanding chemical interactions on other planets, or modeling simple reactions that might have taken place during the early evolution of life on Earth.

Leave a comment

0.0/5