A machine learning based approach to reaction rate estimation

Chemical kinetic models are vital to accurately predicting phenomena in a wide variety of fields from combustion to atmospheric chemistry to electrochemistry. However, building an accurate chemical kinetic model requires the efficient and accurate estimation of many reaction rate coefficients for many reaction classes with highly variable amounts of available training data. Current techniques for fast automatic rate estimation tend to be poorly optimized and tedious to maintain and extend. We have developed a machine learning algorithm for automatically training subgraph isomorphic decision trees (SIDT) to predict rate coefficients for arbitrary reaction types. This method is fully automatic, scalable to virtually any dataset size, human readable, can incorporate qualitative chemical knowledge from experts and provides detailed uncertainty information for estimates. The accuracy of the algorithm is tested against the state of the art rate rules scheme in the RMG-database for five selected reaction families. The SIDT method is shown to significantly improve estimation accuracy across all reaction families and considered statistics. The estimator uncertainty estimates are validated against actual errors. Development of a machine learning algorithm for automatically training subgraph isomorphic decision trees (SIDT) to predict reaction rate coefficients.