Gravitational-Wave Searches for Cosmic String Cusps in Einstein Telescope Data using Deep Learning

Gravitational-wave searches for cosmic strings are currently hindered by the presence of detector glitches, some classes of which strongly resemble cosmic string signals. This confusion greatly reduces the efficiency of searches. A deep-learning model is proposed for the task of distinguishing between gravitational wave signals from cosmic string cusps and simulated blip glitches in design sensitivity data from the future Einstein Telescope. The model is an ensemble consisting of three convolutional neural networks, achieving an accuracy of 79%, a true positive rate of 76%, and a false positive rate of 18%. This marks the first time convolutional neural networks have been trained on a realistic population of Einstein Telescope glitches. On a dataset consisting of signals and glitches, the model is shown to outperform matched filtering, specifically being better at rejecting glitches. The behaviour of the model is interpreted through the application of several methods, including a novel technique called waveform surgery, used to quantify the importance of waveform sections to a classification model. In addition, a method to visualise convolutional neural network activations for one-dimensional time series is proposed and used. These analyses help further the understanding of the morphological differences between cosmic string cusp signals and blip glitches. Because of its classification speed in the order of magnitude of milliseconds, the deep-learning model is suitable for future use as part of a real-time detection pipeline. The deep-learning model is transverse and can therefore potentially be applied to other transient searches.