We address the problem of visualization of the electronic band dispersion from angle-resolved photoemission spectroscopy spectra as a semantic segmentation task. The goal is to label each spectrum’s pixel to one of two possible classes. The first class represents the electronic dispersion pixel, and the other one is the “background” (broadening, distortion of the dispersion by intrinsic and extrinsic processes, noises, and the influence of the apparatus function). The models U-Net and DeepLabV3 were applied to this problem. Both models were tested on the experimentally obtained spectra and showed efficiency in electronic dispersion visualization without manual inspection and pre-denoising. The DeepLab network effectively identifies weak zones, but it’s prone to adding artifacts and struggles with highly noisy spectra (regions). In contrast, U-Net is less effective at identifying weak zones, but it doesn’t introduce artifacts and performs better with higher noise levels.
Pustovit et al. (Sun,) studied this question.