Amplifying the Chirp: Using Deep Learning (U-Nets) to filter signal from noise in LIGO data

Abstract

The direct detection of gravitational waves by LIGO has heralded a new era for astronomy and physics. Typically the gravitational waves observed by LIGO are dominated by noise. In this work we use Deep Convolutional Neural Networks (specifically U-Nets) to filter a clean signal from noisy data. We present two realizations of U-Net filters, the Noise2Clean U-Net filter which is trained using noisy and clean realizations of the same signal, as well as Noise2Noise U-Net which is trained on two separate noisy realization of the same signal. We find that the U-Nets successfully filter signal from noise. We also benchmark the performance of U-Nets by using them to detect the binary presence or absence of gravitational wave signals in data.