A Machine Learning Approach for Prediction of Rate Constants

We report a machine learning approach to train and predict bimolecular thermal rate constants over a large temperature range. The approach uses Gaussian process (GP) regression to evaluate the difference between accurate quantum results and Eckart-corrected conventional transition state theory, mostly for collinear reactions. Training is done on a database of rate constants for 13 reaction/potential surface combinations, and testing is performed on a set of 39 reaction/potential surface combinations. Averaged over all test reactions, the GP method is within 80% of the accurate answer, whereas transition state theory (TST) is only within 330% and Eckartcorrected TST (ECK) is within 110%. In the tunneling region, GP is generally (with a few exceptions) more accurate and sometimes much more accurate. In the high-temperature recrossing region, GP is significantly more accurate than either TST or ECK, neither of which addresses the possibility of recrossing. The GP predictions for the 3D reactions O(P-3) + H-2, OH + H-2, O(P-3) + CH4, and H + CH4, for which accurate quantum results are available, provide further encouragement to the machine learning approach.