Ottaway, DavidLiu, LingqiaoLasky, Paul (Monash University)Jenner, Kendall Louise2025-05-232025-05-232024https://hdl.handle.net/2440/144699Current analysis methods for estimating the source parameters of gravitational wave signals can take from days to weeks. As gravitational wave detectors improve, current analysis methods will be insufficient to keep up with the increasing number of events detected. We investigate the use of convolutional neural networks (CNNs) and normalising flows to estimate the gravitational wave source parameters from the spectrogram of the signal. We find that a CNN can estimate the component masses of a simulated binary black hole (BBH) signal with a maximum 40% error; and that the accuracy decreases when also trying to estimate the luminosity distance and the angle between the line-of-sight and the total angular momentum of the binary system. We introduce a new architecture that uses a normalising flow conditioned on the CNNembedded spectrogram. This proves to be a much more robust approach, with the learned posteriors of the component masses capturing the injected values a majority of the time. When learning the posteriors of the component masses, luminosity distance, and θJN the model recovers the injected primary mass and θJN consistently, but under constrains the luminosity distance and over constrains the secondary mass. Overall, the posteriors produced using the convolutional neural network embedding model with a normalising flow (CNN+n.flow) were less accurate and much broader than those produced using matched filtering techniques with the Bayesian Inference Library (BILBY), but the inference time was significantly faster. This could open up opportunities for using the deep learning method to provide a proposal distribution for likelihood reweighting, offering both higher accuracy and quicker speeds for inference of gravitational wave signals.enGravitational wave astronomyconvolutional neural networksmachine learningThesis