Deep Learning for Modelling and Separating Gravitational Wave Signals, Glitches, and Noise

Since the first direct detection of gravitational waves (GWs) in 2015, the LIGOVirgo–KAGRA collaboration has observed hundreds of events, marking a new era in astrophysics. As detector sensitivity improves, the rate of GW detections continues to rise, enhancing the scientific output of the collaboration. However, a key challenge is the presence of non-Gaussian noise transients in GW detectors, called glitches, which can mimic or obscure genuine signals, degrading the parameter estimation of GW sources and increasing false alarms. If glitch rates remain unchanged, the growing number of detections will naturally lead to more instances of glitchsignal overlap. Traditional glitch mitigation methods are often computationally intensive and struggle to scale with the growing volume of GW data. This challenge becomes more pressing with the advent of next-generation detectors–such as the Einstein Telescope and Cosmic Explorer–which are expected to detect hundreds of events per day. As a result, the need for fast, accurate, and scalable analysis methods is becoming increasingly critical. This thesis leverages deep learning as a fast, data-driven approach to glitch mitigation and signal reconstruction, aiming to improve the accuracy and efficiency of GW data analysis. This thesis presents two main contributions: (1) deep learning models for denoising and reconstructing GW signals and glitches, and (2) generative models for simulating realistic glitches and waveforms. We first introduce DeepExtractor, a framework that learns to subtract background noise and recover transient signals and glitches without templates. Extending the DeepExtractor approach across a multiple-detector network and incorporating signal models during training improves the estimation of GW source parameters in glitch-contaminated data by allowing for glitch removal even when they overlap with astrophysical events. To support data augmentation and simulationbased testing, we develop Derivative GAN (DVGAN) and its class-conditional variant cDVGAN, which incorporate a derivative-based discriminator to improve training stability and sample fidelity. We further advance cDVGAN to generate diverse, highquality glitch samples spanning a wide range of realistic detector noise morphologies, all within a single user-controlled model. Together, these methods provide scalable and accurate alternatives to traditional pipelines, enhancing the robustness of GW data analysis as detectors become increasingly sensitive and their data more complex.