Semi-supervised morphological classification of fast radio bursts from the second CHIME/FRB catalogue
Understanding the morphology of fast radio bursts (FRB), and whether all sources repeat, are key challenges that are becoming more tractable given the increase in data from surveys such as the Canadian Hydrogen Intensity Mapping Experiment FRB project (CHIME/FRB). We present a Convolutional Autoencoder unsupervised classifier for separating the CHIME/FRB data into morphological classes. This data-driven approach is more reproducible than visual inspection, since groupings are learned from the data itself and not subject to differences between expert annotations. While most bursts occupy a similar area of morphological parameter space, we identify three classes of bursts separate from the general FRB population. While one class contains bursts with short bandwidth, and downward-drifting sub-burst structure, the characteristic bursts of other classes have very short temporal width, and occupy the entire CHIME observing band. We identify two distinct subgroups of temporally short, simple broadband bursts; one with minimal scattering and the other with higher scattering. As an additional output of our classifier, we provide a binary FRB repeatability classification, and train the classifications on simulations that mimic the first FRB catalogue from CHIME/FRB. We are able to correctly identify 86$\%$ of repeater bursts. We find that our approach is able to independently recover the downward linear drifting burst morphologies previously defined through visually inspection. Overall, we find that although there exists FRB subgroups with higher or lower proportion of repeaters, there is substantial overlap between the morphological properties of repeaters and one-off bursts consistent with previous studies.