Gravitational-wave Signal Detection with Machine Learning

Abstract

The field of Gravitational Wave (GW) astronomy has only just started and with the construction of third-generation detectors there is a lot of work being done in developing faster and more robust data analysis methods. Machine learning has emerged as a powerful addition or even alternative for the data analysis involved in GW detection. In this thesis we explore two separate machine learning methods for two different type of GW signals. The first being a Convolution Neural Network (CNN) for the detection of GW from core-collapse supernovae (CCSNe). Our results here corroborate those of previous work done, where we were able to identify signals from a phenomenological template bank and numerical 3D simulations of CCSNe with similar efficiencies and false alarm rates. For the second method we construct two fully connected neural networks that are equivalent to the traditional matched filtering approach based on previous work for the detection of GW signals generated by Compact Binary Coalescences (CBCs). These networks can theoretically approach the performance of an optimal classifier with knowledge of a prior of the parameter distribution. By training these networks on training data consisting of real LIGO data and simulated injections of GW signals we can outperform the basic matched filter and even approach the theoretically predicted optimal classifier.