(134a) Automating the Construction of Predictive Kinetic Models for the Combustion of Halogenated Hydrocarbons, with Rmg

From refrigerant fluids in car air conditioners to flame suppressants in airplane cargo holds, halogenated hydrocarbons (HHCs) have many applications where it’s crucial to know how they will burn. Predicting the combustion behavior from first principles will greatly facilitate the search for effective and safe compounds, as we strive to find compounds with a lower global warming potential. But it’s not easy! Some HHCs suppress flames under certain conditions but act as fuels in slightly different conditions. Detailed kinetic models should be able to explain, and ideally predict, such bizarre behavior.

We are extending the open-source Reaction Mechanism Generator (RMG), largely developed by the Green Group at MIT, to include halogen chemistry. The principal challenges are a combinatorial explosion of the number of possible structures and pathways when multiple halogens are considered simultaneously, and a dire scarcity of training data. To tackle the former problem, we are exploring graph neural network methods to predict thermochemistry and kinetics instead of using decision trees, but these methods can exacerbate the problem of insufficient training data. Semi-supervised and transfer learning methods offer promise, learning how best to represent molecules without needing to know much about them. To tackle the challenge of data scarcity, we are using our automated processes for performing quantum chemistry calculations of molecules and transition states, as well as our tools for importing literature data. Results include difluoromethane (refrigerant R32) and 2-bromo-3,3,3-trifluoropropene (flame suppressant 2-BTP).