Deep learning in the SKA era : patterns in the SNR population with unsupervised ML methods

Abstract

The Square Kilometre Array precursors are releasing the first data of their large-field continuum surveys. The complexity of such datasets makes clear that deep learning is the primary solution for handling an overwhelming volume of data also in the radio astronomy field. Within this framework, our research group is taking a forefront position in various research initiatives aimed at assessing the effectiveness of ML techniques on survey data from ASKAP and MeerKAT. In this work we show how an unsupervised multi-stage pipeline is able to discover physically meaningful clusters within the heterogeneous Supernova Remnant (SNR) population: a convolutional autoencoder extracts features from multiwavelength imagery of a SNR sample; then an unsupervised clustering process operates on the latent space to identify patterns. Despite a large number of outliers, we were able to find a new classification system, in which most clusters relate to the presence of certain features regarding not only the morphology but also the relative weight of the different frequencies.