This thesis focuses on the problem of Faraday complexity classification using Bayesian deep learning, investigating the task in a level of detail not currently available in the literature, and highlighting a number of key topics for consideration as more observations are made in the near-future. A code-base is developed for producing some of the most realistic simulations of Faraday populations currently available, and results are found and analysed on this synthetic dataset. This research then provides the first application of deep learning methods to real Faraday rotation data by obtaining labels and associated uncertainties for 2,206 polarised radio sources observed using the Low Frequency Array (LOFAR) telescope. Successful application to the LOFAR data is enabled through careful consideration of the simulated-to-real (Sim2Real) transfer learning problem, with the model trained on synthetic data before application to real data. By using a group of human experts to label a subset of samples, a comparison is made possible between the AI predictions and human labels, including confirmation that the model trained only on simulated data functions at a near-human-level on unseen real data. Particular attention is devoted to quantifying the total, aleatoric and epistemic uncertainties associated with the predicted labels, and to the application of domain randomisation as a domain adaption technique for improving Sim2Real performance. The work in this thesis contributes towards research on of cosmic magnetism, the study of the largest physical-scale magnetic fields in the Universe. This area is set for a rapid growth of knowledge provided by new and near-future telescopes including the Square Kilometre Array. Novel insight into cosmic magnetism will be enabled through an extreme increase in the quality and quantity of data in observations of radio astronomy that utilise the Faraday effect, and this big data motivates a machine learning based approach.
|Date of Award
|31 Dec 2022
- The University of Manchester
|Jonathan Shapiro (Supervisor) & Anna Scaife (Supervisor)