The process of identifying the fleeting chemical transition states that occur during reactions could be significantly sped up thanks to a machine learning system developed by researchers from MIT. At present, these states can be calculated using quantum chemistry, but this process is time and computing power intensive, often taking days to calculate a single transition state.
Now, researchers led by MIT associate professor of chemistry and chemical engineering, Heather Kulik, have created a machine learning model that can calculate the structure of transition states within seconds. The model, based on a diffusion model, allows the team to represent two reactants in any arbitrary orientation to each other. This allows the team to consider different orientations of the same reactants but within the same reaction, saving a significant amount of time.
Transition state data from 9,000 different chemical reactions was provided for their model in terms of reactant, product and transition state structures. The developed model was accurate to within 0.08 angstroms when tested on 1,000 reactions it hadn’t previously encountered, capable of generating 40 potential solutions for each transition state in just a few seconds.
This model could be used to help design new catalysts, develop new synthesis methods for useful compounds, or model natural reactions. In the future, its creators plan to expand the model’s capabilities to include catalyst incorporation in order to better understand how much a particular catalyst could accelerate a reaction. Other potential applications include modeling the simple reactions that may have occurred during the early evolution of life on Earth and the gaseous interactions occurring on other planets.
This research is announced as a “significant step forward” by Jan Halborg Jensen, a University of Copenhagen chemistry professor not involved in the research.