Estimating Electron-Molecule Cross Sections using a Data-Driven Machine Learning Model

Electron-molecule collision cross sections play a pivotal role in many areas of applied physics, but detailed measurements and state-of-the-art theoretical calculations are often too difficult to perform for highly complex molecules or for a wide variety of molecular targets. Consequently, the needed data sets are often unavailable or incomplete. Machine learning algorithms may be able to provide reasonable estimates for molecular targets that are beyond the reach of theoretical models and experimental measurements and help fill the gap in available cross section data. We present results from a feed-forward neural network that was trained on published experimental data and show that it provides reasonable estimates of electron-molecule collision cross sections for molecular targets beyond those in its training set. We test our model by comparing its predictions with measured cross sections, and our results show that with training on as a few as 15 molecular targets, cross sections can be predicted to within 10% for many molecules. The success of our simple model with relatively few training data sets indicates that machine learning algorithms can successfully complement experiment measurements and traditional theoretical models and are a viable alternative to provide much needed data for applied physics modeling.