A Multilayer Perceptron to Extend and Improve Diatomic Molecule Collision Rate Coefficients

Abstract A data-driven approach to improve the accuracy of state-resolved molecular collision data for the relaxation of vibrationally and rotationally excited molecules is developed. We apply the approach to the SiO–H 2 collision system, which is crucial for reliable astrophysical environment modeling of SiO emission. The most accurate quantum scattering calculation method employs the coupled-channel formalism, which is computationally expensive. Approximate approaches, generally referred to as decoupling approximations, are significantly more computationally efficient, but less accurate. To bridge this accuracy gap, we use a multilayer perceptron to predict the difference between the accurate and approximate results. Our model is trained on a carefully selected and preprocessed data set, which includes both approximate and accurate calculation results. We show that our approach significantly improves the accuracy of the approximate calculation results, as the relative rms errors from the predicted results are orders of magnitude smaller than those from the approximate calculation results, benchmarked against the accurate calculations. The present approach can be applied to similar problems in other fields to improve the accuracy of computationally efficient approximate methods, reducing computation time and enhancing automation.