Skyward AI: Advancing Astronomy with Intelligent Machines




Bialek, Spencer

Journal Title

Journal ISSN

Volume Title



This dissertation represents the work I did in integrating advanced machine learning techniques into three important challenges that the field of astronomy currently faces. Firstly, I tackled the emerging concern of contamination from low earth orbit satellites in the upcoming massive spectroscopic sky surveys. With the imminent launch of several hundred thousand satellites, there’s a potential for significant contamination in wide field, multi-fiber spectrographs. I employed a multi-staged approach to gauge the feasibility and constraints of pinpointing and reducing the impact of such contamination in a WEAVE-like stellar spectral survey. By crafting a series of convolutional-network based architectures, I managed to identify and separate stellar spectra that were artificially tainted with satellite (solar-like) spectra. My findings revealed a promising capability to flag a majority of contaminated sources and reconstruct the clean spectra with minimal error. This work offers a suite of machine learning strategies that can be harnessed to enhance stellar parameters for contaminated spectra in the WEAVE stellar spectroscopic survey and similar endeavours. In my second project, I introduced a novel solution to the well-studied problem of atmospheric turbulence compromising the clarity of astronomical images. By training a U-Net on simulated observations, I demonstrated how a sequence of short-exposure observations of a stellar field can be transformed into a turbulence- and noise-free image. This approach significantly boosts angular resolution over arbitrarily wide fields while preserving flux to a lower signal-to-noise than an averaged stack, without compromising the astrometric stability in the resultant image. It is technically simple as well, keeping costs of implementing and maintaining such a system low. Lastly, I explored the potential of self-supervised learning in extracting meaningful representations of galaxies from millions of unlabelled sources. Recognizing the power of self-supervised methods, particularly SimCLR, I aimed to validate their utility for the UNIONS Survey. My efforts were geared towards automating the clustering and classification of galaxy types, refining photometric redshift estimations, and leveraging these techniques to unearth rare astronomical phenomena such as ultra-faint dwarf galaxies, gravitational lenses, and merging galaxies. The initial results show that, by using a query galaxy image, the fully trained SimCLR model can successfully find similar types of galaxies using a self-similarity search in a database of millions of galaxies. Throughout these projects, I have combined machine learning and astronomical research, presenting innovative solutions to pressing challenges in the field. Each endeavour reflects my dedication to leveraging the capabilities of machine learning to propel astronomical discoveries forward, offering fresh perspectives and tools to address longstanding and emerging issues in the discipline.



machine learning, astronomy