A semisupervised machine learning search for never-seen gravitational-wave sources
ABSTRACT By now, tens of gravitational-wave (GW) events have been detected by the LIGO and Virgo detectors. These GWs have all been emitted by compact binary coalescence, for which we have excellent predictive models. However, there might be other sources for which we do not have reliable models. Some are expected to exist but to be very rare (e.g. supernovae), while others may be totally unanticipated. So far, no unmodelled sources have been discovered, but the lack of models makes the search for such sources much more difficult and less sensitive. We present here a search for unmodelled GW signals using semisupervised machine learning. We apply deep learning and outlier detection algorithms to labelled spectrograms of GW strain data, and then search for spectrograms with anomalous patterns in public LIGO data. We searched ${\sim}13{{\ \rm per\ cent}}$ of the coincident data from the first two observing runs. No candidates of GW signals were detected in the data analyzed. We evaluate the sensitivity of the search using simulated signals, we show that this search can detect spectrograms containing unusual or unexpected GW patterns, and we report the waveforms and amplitudes for which a $50{{\ \rm per\ cent}}$ detection rate is achieved.