Data-Driven Machine Learning Models for Atomic and Molecular Collisions

Many areas of applied physics rely on atomic and molecular collision cross sections as fundamental inputs for modeling complex problems. One factor influencing the success of these models is the availability and accuracy of the cross sections, which are often required for multiple collision processes and target species over a wide range of energies. While several databases and portals are available, the supply of data remains limited due to computational and experimental constraints, as well as a high barrier of entry for non-expert users. Thus, there is a need for alternative techniques to provide estimates of collision cross sections.

We present the development of several feed-forward neural networks [Eur. Phys. J. D. 67, 130 (2013); J. Phys. B 57, 025201 (2024)] that can be used to predict collision cross sections for target species in which limited or no data exists. Our models were trained on published experimental data and used to predict electron- and proton-impact collision cross sections for atomic and molecular targets beyond those in their training sets. Our results show that these models provide reasonable estimates of the cross sections even with limited training data. In some cases, the networks were trained on as a few as 15 different targets and cross sections were predicted to within 10% for target species previously unseen by the network. Our models were then used to predict cross sections for molecular targets in which no published data was available, as well as benchmark experimental data for the biomolecule 3,4-dihydro-2H-pyran (DHP) [J. Chem. Phys. 161, 064304 (2024)]. The successful application of our machine learning algorithms indicates that these techniques are a viable alternative to provide much needed data for applied physics modeling and can successfully complement experimental measurements and traditional theoretical models.