Deep learning for core-collapse supernova gravitational wave signals and noise transients

In this thesis we present proof of concept for a search and classification pipeline which implements three separate deep learning architectures on top of an event trigger generator, the Wavelet Detection Filter. We first demonstrate the ability of the full pipeline on a binary classification task between neutrino-driven core-collapse supernovae signals and transient noises, known as glitches, using 1-D and 2-D convolutional neural networks, that process whitened time series and whitened spectrograms, attaining ' 90% accuracies. We then prove the ability of a merged model in a multilabel classification task involving different core-collapse supernova models and two transient noise classes, with accuracies that fall just short of 90%. Finally, we extend the multilabel classification scheme to real O2 data from three detectors separately and source distances of 1 kpc, introducing a recurrent long short-term memory network and three additional neutrino-driven core-collapse supernova models. Using a merged model, we achieve ∼ 99% total accuracy for LIGO Livingston and LIGO Hanford and ' 90% sensitivities for the most energetic models in the Virgo dataset. No false alarm affected classification of the merged model, despite 7913 noise triggers present in the LIGO datasets, while the performance on Virgo was limited by the small training set available.