Machine learning models for atom-diatom reactions across isotopologues

TL;DR

This paper demonstrates that neural networks can accurately predict the outcomes of atom-diatom reactions across different isotopologues, including unseen reactants, by learning from related chemical reactions.

Contribution

It introduces a neural network model capable of generalizing reaction outcomes across isotopologues, including unseen reactants, in the context of atom-diatom reactions.

Findings

Neural networks accurately predict ro-vibrational state distributions.

Models can generalize to reactions with unseen isotopologues.

The approach is relevant for buffer gas chemistry applications.

Abstract

This work shows that feed-forward neural networks can predict the final ro-vibrational state distributions of inelastic and reactive processes of the reaction of Ca $+$ H2 $\rightarrow$ CaH $+$ H in the hyperthermal regime, relevant for buffer gas chemistry. Furthermore, these models can be extended to the isotopologues of the reaction involving deuterium and tritium. In addition, we develop a neural network model that can learn across the chemical space based on the isotopologues of hydrogen. The model can predict the outcome of a reaction whose reactants have never been seen. This is done by training on the Ca $+$ H2 and Ca $+$ T2 reactions and subsequently predicting the Ca $+$ D2 reaction.