Machine learning approaches to extend and improve state-resolved molecular collision data

Given a set of rate coefficients for state-to-state transitions in molecular collisions, it is often desirable to (i) extend the set to a wider range of initial and final quantum numbers at the same level of accuracy and (ii) improve the accuracy of the rate coefficients when a subset of more reliable data is available. This talk will address both of these goals for molecular collision systems which require a computationally expensive quantum scattering formulation to obtain accurate data. We introduce machine learning approaches that are applied to inelastic diatom-diatom scattering for a system where varying amounts of rigorous six-dimensional close-coupling (6DCC) and approximate five-dimensional coupled states (5DCS) results are available. Goal (i) is achieved using a Gaussian process model to extend the 5DCS data set without compromising accuracy beyond the level of approximation. Goal (ii) is achieved using a neural network to improve the large database obtained in goal (i) augmented by a relatively small number of 6DCC calculations to correct the errors introduced by the 5DCS approximation. The machine learning approaches used in this work are general and may be applicable to similar problems in other fields in order to improve the accuracy of computationally efficient approximate methods, reducing computation time and enhancing automation.