An autoencoder neural network integrated into gravitational-wave burst searches to improve the rejection of noise transients

The gravitational-wave (GW) detector data are affected by short-lived instrumental or terrestrial transients, called 'glitches', which can simulate GW signals. Mitigation of glitches is particularly difficult for algorithms which target generic sources of short-duration GW transients (GWT), and do not rely on GW waveform models to distinguish astrophysical signals from noise, such as coherent WaveBurst (cWB). This work is part of the long-term effort to mitigate transient noises in cWB, which led to the introduction of specific estimators, and a machine-learning based signal-noise classification algorithm. Here, we propose an autoencoder neural network, integrated into cWB, that learns transient noises morphologies from GW time-series. We test its performance on the glitch family known as 'blip'. The resulting sensitivity to generic GWT and binary black hole mergers significantly improves when tested on LIGO detectors data from the last observation period (O3b). At false alarm rate of one event per 50 years the sensitivity volume increases up to 30% for signal morphologies similar to blip glitches. In perspective, this tool can adapt to classify different transient noise classes that may affect future observing runs, enhancing GWT searches.