Autoencoder latent space sensitivity to material structure in convergent-beam low energy electron diffraction

Abstract

The convergent-beam low energy electron diffraction technique has been proposed as a novel method to gather local structural and electronic information from crystalline surfaces during low-energy electron microscopy. However, the approach suffers from high complexity of the resulting diffraction patterns. We show that Convolutional Autoencoders trained on CBLEED patterns achieve a highly structured latent space. The latent space is then used to estimate structural parameters with sub-angstrom accuracy. The low complexity of the neural networks enables real time application of the approach during experiments with low latency.