A Machine Learning Approach for Computationally Efficient Prediction of State-to-State Transition Cross Sections for Molecular Collision Processes

Abstract A neural network model is adopted to develop a machine learning approach for the prediction of rotational state-to-state transition cross sections for inelastic molecule + molecule collisions. This method is applied to H2O + H2 system that includes 200 rotational states of the water molecule as a target and all states of H2 projectile up to j = 8. It is shown that with a relatively small subset of transitions used for training and validation of the neural network, one can have a fast and reliable prediction of cross sections and rate coefficients with a mean relative error of ∼7%, which is comparable to the accuracy of the method used to generate the original dataset. This demonstrates that the calculation of rate coefficients based on cross sections predicted by the machine learning approach may dramatically reduce the CPU effort associated with numerical solution of the molecular scattering problem, which opens ample opportunities for accelerating the generation of collisional databases for astrochemical and other applications.