Transfer learning, alternative approaches, and visualization of a convolutional neural network for retrieval of the internuclear distance in a molecule from photoelectron momentum distributions

TL;DR

This paper explores deep learning methods, especially convolutional neural networks, for accurately determining internuclear distances in molecules from photoelectron momentum data, emphasizing transfer learning and interpretability.

Contribution

It introduces transfer learning for CNNs in molecular distance retrieval, compares with other methods, and uses occlusion sensitivity for feature analysis.

Findings

CNN with transfer learning outperforms other methods in transferability

Support-vector machines and decision trees show limited transferability

Occlusion sensitivity reveals key features used by the neural network

Abstract

We investigate the application of deep learning to the retrieval of the internuclear distance in the two-dimensional H$_2^{+}$ molecule from the momentum distribution of photoelectrons produced by strong-field ionization. We study the effect of the carrier-envelope phase on the prediction of the internuclear distance with a convolutional neural network. We apply the transfer learning technique to make our convolutional neural network applicable to distributions obtained for parameters outside the ranges of the training data. The convolutional neural network is compared with alternative approaches to this problem, including the direct comparison of momentum distributions, support-vector machines, and decision trees. These alternative methods are found to possess very limited transferability. Finally, we use the occlusion-sensitivity technique to extract the features that allow a neural network to take its decisions.