The 25th July of 2023 Maria Teresa Mateu Lucena presented her PhD thesis: Understanding gravitational wave data with Bayesian inference and machine learning.
Title: Understanding gravitational wave data with Bayesian inference and machine learning.
Author: Maria Teresa Mateu Lucena
Directors: Dr. Sascha Husa
Abstract: One of the essential fields in gravitational wave (GW) data analysis is the development of accurate and computationally efficient waveform models, which are used in Bayesian inference methods to estimate the parameters of the astrophysical source of an event. Phenomenological waveform models are one of the main waveform modelling approaches used for such analyses. These waveform models have reached high accuracy and efficiency in analysing the data of the third and fourth observational runs of the LIGO-Virgo-Kagra detector network. However, new and more robust detectors are under development, and current detectors’ sensitivity is constantly increasing. As a result, there will be a significant increase in the number of detections, which will present new difficulties in processing the huge number of gravitational wave signals detected and extracting the most valuable scientific information from them. This requires the development of more accurate and faster waveform models, and here is where machine learning algorithms come into play. The main goal of this thesis is to understand the parameter estimation samplers’ operation and how different waveform models of the current generation, impact the convergence of the results. Furthermore, after understanding how important the accuracy of such waveform models is to achieving better convergence of the results and avoiding systematic errors, in the second part of this thesis, we use machine learning techniques to continue the improvement of the current generation of phenomenological waveform models, in particular the ringdown part of the waveform.
In the first part of this thesis, I study in detail the impact on parameter estimation of the improvements in the precessing treatment and the addition of subdominant modes in the fourth generation of phenomenological models, compared to the previous generation. This study is performed as a re-analysis of the parameter estimation of the highly unequal mass event, GW190412 since the effect of sub-dominant spherical harmonic modes is stronger in unequal-mass systems and all the binary black hole coalescences of the first GW catalog. In the second part of the thesis, I also perform a study to improve the ringdown part of the last generation of phenomenological waveform models by improving the current fits of the remnant properties of binary black hole systems. In order to do that, I extend the previous work performed by Haegel et al. (2020) in order to find a general methodology to build a deep neural network (DNN) with the desired accuracy and efficiency, preventing underfitting and overfitting. The work has been first applied to non-precessing binary black hole systems due to the simplicity of the parameter space and then extended, as a first attempt, towards precessing systems. Deeper studies have to be applied in the precessing parameter space, but preliminary results significantly improve the current approximations used in the phenomenological waveform models. Such methodology will also be expanded toward the whole construction of phenomenological waveform models and will be used to fit the phenomenological coefficients.
