Mars, Matthijs; (2024) Learned image reconstruction for interferometric telescopes. Doctoral thesis (Ph.D), UCL (University College London).

Text Mars_10199599_thesis.pdf Access restricted to UCL open access staff until 1 June 2025. Download (16MB)

Abstract

Radio interferometry is an essential tool in astronomy; from the study of the epoch of reionisation to studying cosmic magnetism and distant galaxies. The next generation of interferometric radio telescopes, such as the Square Kilometre Array (SKA), will generate vast amounts of data, necessitating efficient and scalable image reconstruction techniques. Traditional methods like CLEAN and variational regularisation approaches rely on iterative optimisation and require numerous evaluations of the measurement operator, whose computation cost scales linearly with the number of visibilities, making these methods computationally expensive for the large data volumes anticipated from the SKA. To address these challenges, this thesis introduces two novel machine learning-based reconstruction methods: (1) a fully data-driven approach, and (2) a hybrid approach that combines data-driven learning with model-based optimisation. These methods are initially developed for the SPIDER instrument, an optical interferometer, for which the imaging problem is analogous to that of radio interferometers. The learned image reconstruction approaches use neural networks to reduce computational costs and improve reconstruction quality compared to traditional methods. The proposed approaches are then extended to radio interferometry, where varying visibility coverages present a challenge for learned reconstruction methods. To produce robust reconstruction models, specific training strategies are proposed to make the models agnostic to these variations with minimal fine-tuning. Finally, the methods are integrated within a generative framework, which efficiently generates samples from the posterior distribution to quantify uncertainties in the reconstructions, crucial for scientific interpretation. The proposed methods achieve significant improvements in computational efficiency and reconstruction quality compared to traditional methods. For the SPIDER instrument, this enables real-time imaging capability, while for radio interferometric telescopes, the methods offer efficient and high-quality reconstructions with built-in uncertainty quantification. This work sets the stage for the widespread adoption of machine learning techniques in the reconstruction pipelines of large, next-generation radio telescopes.

Download activity - last month

Export as CSVJSONXML

Download activity - last 12 months

Export as CSVJSONXML

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item