Dissertations, Theses, and Capstone Projects

Date of Degree

9-2025

Document Type

Doctoral Dissertation

Degree Name

Doctor of Philosophy

Program

Physics

Advisor

Matthew O'Dowd

Committee Members

Viviana Acquaviva

Luis Anchordoqui

Quinn Minor

Laurence Perreault-Levasseaur

Subject Categories

Astrophysics and Astronomy

Keywords

Quasars, AGN, gravitational lensing, dark matter, machine learning, neural networks

Abstract

Machine learning methods are well suited to handle the unprecedented data volume expected from upcoming wide-field surveys. The European Space Agency's Euclid telescope was launched in July 2023 and is expected to observe billions of galaxies at high resolution, including tens of thousands of strongly lensed galaxies. The Rubin Observatory Legacy Survey of Space and Time (LSST) will monitor tens of millions of quasars throughout its ten-year lifetime, thousands of which will be strongly lensed. This flood of data will enable measuring lens galaxy mass distributions and substructure, as well as probing quasar accretion disk structure and black hole properties through variability studies of both strongly lensed and unlensed quasars. In addition, through the analysis of time delays in gravitationally lensed quasars and by combining data from Euclid and LSST, it will be possible to precisely measure the cosmic expansion rate independently of the cosmic distance ladder. Such independent measurements are key to resolving the Hubble tension, the discrepancy between the cosmological expansion rates derived from the cosmic distance ladder and the cosmic microwave background. Realizing these science opportunities demands a deep physical understanding of gravitational lensing, quasar variability, and galaxy structure, as well as the development of scalable, autonomous techniques. Machine learning is inherently scalable and capable of modeling complex nonlinear relationships in high-dimensional feature spaces where traditional optimization methods struggle, offering a promising framework for extracting scientific insights from gravitational lensing and quasar variability studies. In this thesis, machine learning techniques are used to study quasar variability, galaxy structure, and gravitational lensing. In Chapter 2, we use convolutional neural networks (CNNs) to measure the substructure power spectrum of 23 Hubble Space Telescope images of strongly lensed galaxies using a novel method of producing uncertainties on the predictions. Our predictions can be used to constrain dark matter theories, and our method will be applicable to the tens of thousands of strong lenses expected to be found by Euclid. In Chapter 3, we apply machine learning methods to microlensed light curves of strongly lensed quasar sources. We use a CNN to predict black hole properties from simulated high-cadence light curves that could potentially be triggered by LSST. This includes a proof of concept application to archival caustic crossings of QSO 2237+0305 to estimate the black hole mass and inclination angle. In addition, we train a recurrent neural network to classify when there are high-magnification event peaks for simulated LSST microlensed quasar light curves so that such rapid follow-up can be triggered. In Chapter 4, we introduce latent stochastic differential equations (SDEs) in the context of modeling quasar variability. Latent SDEs are a physically motivated generative deep learning technique we developed to model stochastic, irregularly sampled multivariate time series and perform parameter inference. We show how latent SDEs outperform a Gaussian process regression baseline on simulated ten-year LSST light curves and can predict variability and accretion disk parameters such as the black hole mass, inclination angle, and temperature slope. We then expand upon the method by developing the first auto-differentiable accretion disk simulation, which we incorporate into the neural network architecture to tie the light curve reconstruction to the accretion disk parameter inference. This improved physically-constrained model also reconstructs the unobserved driving signal and transfer function kernels, and we show how it is robust to out-of-distribution variability. In Chapter 5, we give our conclusions and discuss future work and the application of our methods to upcoming large datasets.

Download

Share

COinS