Advancing Artificial Intelligence in Astronomy

Abstract

This thesis addresses different questions on the use of artificial intelligence (AI) for the processing and analysis of current and upcoming radio surveys, including those specifically pertaining to the analysis of large data volumes expected from future facilities, such as the SKA. This work presents spectro-polarimetric data with a polarised detection limit of ∼ 20 μJy/bm made by the MIGHTEE collaboration with the MeerKAT array, an SKA precursor. Multiple accompanying catalogues are presented alongside a description of the calibration and imaging approaches used. Analysing this data at scale is extremely challenging. To address analyses at this scale and beyond, this work also presents an improvement to how Masked Auto Encoding, a modern AI approach to pre-training, can be employed for science data. This approach, coined frequency masked auto encoding (FMAE), is inspired by the spatial frequency sampling of radio interferometry. FMAE is expected to enable large scale pre-training (i.e. foundation models) for various astronomical data, including images, spectra, and time series data. Finally, this work presents an approach to derive a semantic taxonomy of plain English terms which are scientifically useful. This approach is applied to radio galaxy morphologies in the EMU survey as part of the Radio Galaxy Zoo EMU collaboration. These terms can be used by any English-speaking party without any technical training. It is shown that the taxonomy can recover existing astrophysical populations and enable morphological outliers to be probed. This method is domain agnostic and may find use in fields where the language used is complex but core to the associated analysis. Collectively this body of work contributes to a broad array of radio astronomy science by producing leading spectro-polarimetric data products, enabling large scale AI pre-training through FMAE, and proposing and a applying a method to derive a plain English semantic taxonomy for radio galaxy morphologies.

Date of Award	1 Aug 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Paddy Leahy (Supervisor) & Anna Scaife (Supervisor)