Deep Learning methods to study the Auxiliary Channels in the Advanced Virgo Detector

Gravitational Waves (GW) are a prediction of Einsteins General theory of relativity and were first detected by the advanced-LIGO interferometers in 2015, almost 100 years after they were theorized. The detection of Gravitational Waves is affected by the background noise, which can be non-stationary and non-Gaussian. Short duration noise bursts, also called ”glitches” can be quite detrimental to the search, since they can affect data quality and mimic the gravitational waves signal itself. To assert the quality of the measurements, thousands of sensors monitor the state of the detector continuosly, but the large amount of Data and the complex nature of the couplings between channels render any kind of analysis quite challenging. Deep Learning offers an opportunity to efficiently analyze and handle the large amounts of data related to the Auxiliary Channels. In particular, Variational Autoencoders (VAE) can help in the task of reducing the dimensionality of the data without losing any relevant information, by projecting the samples on a lower dimensional manyfold, called the latent space, on which the analysis can be run instead. The aim of this thesis was to apply Deep Learning tools such as VAEs to the study of the Auxiliary Channels to better characterize and gain further insight in to the noise of the Advanced Virgo detector. In this thesis we exploit the latent space representation of a multi-channel Variational Autoencoder trained on Auxiliary Channels spectrograms, coupled with a clustering algorithm to find patterns in this complex data landscape. The main focus was to study the response of the interferometer’s mirrors suspensions, called Super Attenuators, to microseismic events, and to explore the correlation with data quality and low frequency glitches in the main channel. The results are promising for the characterization of the performances of Super Attenuators during microseisms and in identifying the causes of some glitches, which is a step towards improving the detector sensitivity in the low frequency domain.