Joel Bowman

A Machine Learning Approach for Prediction of Rate Constants

The calculation of thermal rate constants (coefficients) is of longstanding and current interest. Much of the current interest is in using ring polymer dynamics, including instanton, as well as semi-classical VPT2 methods. These powerful methods require considerable, if not complete, knowledge of high-dimensional potential energy surfaces. Thus, the computational effort scales steeply with the number of degrees of freedom. I will present a novel and very efficient machine learning approach to train and predict bimolecular thermal rate constants over a large temperature range. The approach is based on correcting conventional transition state theory with standard (and trivial to calculate) Eckart tunneling. An extensive database of exact quantum calculations of rate constants for collinear reactions is used for training using Gaussian Process Regression (GPR). Testing is done on a separate database of rate constants with very encouraging results.