Improving and Extending the Rate Coefficient Database for Inelastic CS-H2 and SiO-H2 Collisions Using Gaussian Process Regression

Machine learning has become a powerful yet computationally efficient tool in solving physics problems. While “deep learning” algorithms have provided accurate models for complex interpolation problems, they have fallen short in their ability to extrapolate. However, supervised machine learning algorithms, such as Gaussian Process Regression (GPR), have succeeded where deep learning has failed in this respect. In this work, we use GPR to improve rate coefficients for CO-H2 and SiO-H2 collision systems by training our model using both limited, but accurate 6-dimensional closed-coupling (6DCC) results with more expansive, but approximate 5-dimensional coupled-states (5DCS) rate coefficient data. Our approach is two-fold. First, our model improves the accuracy of the 5DCS approximate data to be more comparable with accurate 6DCC data. Second, we use the model to extrapolate rate coefficient data for states in which 6DCC calculations have not been performed due to computational cost. In addition to improving and extending the 6DCC database our studies have shown that the GPR machine learning algorithm actually starts to predict the exponential energy gap law with no empirical information.