(38a) Deep Learning of Activation Energies and Automated Reaction Dataset Generation

Automated mechanism generation is an efficient tool for rapidly constructing large-scale chemical reaction mechanisms, which are necessary for realistic modeling of complex processes, such as combustion and oxidation, soot formation, or pyrolysis. Nonetheless, hard-coded reaction templates employed by mechanism generation software do not cover the set of all possible reactions that might be important. Moreover, only a limited quantity of high-quality data is available for the estimation of kinetic rates during mechanism generation, which may result in poor estimates and the erroneous exclusion of important reactions. In response, we have developed methods to automatically generate a large database of chemical reactions and designed a template-free deep learning model capable of providing a quick assessment of potentially important reactions that are missing from a reaction mechanism.

We use activation energy as a temperature-independent measure to gauge the potential importance of new reactions and train a machine learning model to estimate the energies by relying on a reaction encoding which predominantly focuses on the atoms involved in the reactive center. In order to predict kinetically relevant activation energies as well as being able to tell which reactions have barriers that are so large that they cannot possibly be relevant, the training data for the machine learning model must span a more diverse set of reactions than is currently available in the literature. We use an automated transition state finding algorithm to generate a diverse training set of tens of thousands of reactions, and thus create an activation energy prediction model that can quantify the importance of proposed reactions.