Machine Learned Neural Networks for the Prediction of Final Ro-vibrational State Distributions of an Atom-Diatom Collision and its Isotopic Variants

This presentation will discuss the use of neural networks to predict the final ro-vibrational state distributions of the atom-diatom collision of calcium and the hydrogen molecule. Furthermore, the prediction of these distributions will be extended to the reactions of calcium with diatomic deuterium and calcium with diatomic tritium. Neural networks were trained using data generated by quasi-classical trajectory (QCT) simulations for collision energies in the hyperthermal regime (5,000 - 50,000 K) and the resulting reactive rotational and vibrational quantum number probability distributions were the targets. The target distributions have the advantage of characterizing the population of the rotational and vibrational states of the product diatom as well as the probability of a reactive process occurring for a given initial state over a range of impact parameters. The featurization of the models include the initial state (the collision energy, initial rotational quantum number, and the initial vibrational quantum number) and information about the reactant diatomic molecule, such as its internal energies. We show that with this featurization and QCT collision data the reaction of calcium with hydrogen and its isotopic variants can be described in terms of its final ro-vibrational state distributions using machine learned neural networks.