Gravitational wave search with Machine Learning

Abstract

We train two Deep Learning (DL) classifiers, based on VGG19, using Gravitational Waves (GW) simulated data. The first of them (from now on, N-S) is capable of distinguishing between plain noise and a simulated signal with injected noise. The second one (from now on, L-U) is trained with noisy GW signals and is able to tell if said signals are lensed GW or not. For our simulations, we consider Binary Black Holes (BBH) systems with spinless members that have masses between 10 and 80 solar masses. The luminosity distances detector-source lie between 500 and 3500 Mpc and we discard events with Signal-to-Noise Ratio (SNR) smaller than 5 or bigger than 50. We feed our models with images of the Q transform of these strains of data, finding that VGG19 performs well in both classifications: when classifying the test sets, N-S achieved an accuracy of 100%, while L-U achieved an accuracy of 98%. The conclusion of this work is that it ratifies the potential GW has, not only in the detection of GW signals, but also in the study of other predicted effects such as lensing.