Uncloaking hidden repeating fast radio bursts with unsupervised machine learning

The origins of fast radio bursts (FRBs), astronomical transients with millisecond timescales, remain unknown. One of the difficulties stems from the possibility that observed FRBs could be heterogeneous in origin; as some of them have been observed to repeat, and others have not. Due to limited observing periods and telescope sensitivities, some bursts may be misclassified as non-repeaters. Therefore, it is important to clearly distinguish FRBs into repeaters and non-repeaters, to better understand their origins. In this work, we classify repeaters and non-repeaters using unsupervised machine learning, without relying on expensive monitoring observations. We present a repeating FRB recognition method based on the Uniform Manifold Approximation and Projection (UMAP). The main goals of this work are to: (i) show that the unsupervised UMAP can classify repeating FRB population without any prior knowledge about their repetition, (ii) evaluate the assumption that non-repeating FRBs are contaminated by repeating FRBs, and (iii) recognise the FRB repeater candidates without monitoring observations and release a corresponding catalogue. We apply our method to the Canadian Hydrogen Intensity Mapping Experiment Fast Radio Burst (CHIME/FRB) database. We found that the unsupervised UMAP classification provides a repeating FRB completeness of 95 per cent and identifies 188 FRB repeater source candidates from 474 non-repeater sources. This work paves the way to a new classification of repeaters and non-repeaters based on a single epoch observation of FRBs.