| Resumo: | <p>Modern radio telescopes are data-intensive machines, producing many TB of data every night. Amongst this deluge of data are transient and variable phenomena, whose study can shed new light on processes as varied as stellar dynamos and the accretion discs in supermassive black holes. In this work I demonstrate the applicability of different methods to the discovery of these astrophysical transients and variables coming from telescopes such as MeerKAT.</p>
<p>I first consider a standard approach to discovering transients by characterising their variability. By making use of even modest sampling with the high sensitivity and wide field of view of MeerKAT, I demonstrate how we are now able to uncover new transients almost by accident - if we exclude the vast amount of time spent planning, building and operating excellent telescopes, efficient pipelines and well- crafted observing proposals. In this work I found a stellar flare from a nearby M dwarf, which was then followed up and complemented by optical and X-ray photometry and spectroscopy, providing new insights on the system.</p>
<p>Next I built a citizen science platform in order to perform such transient searches at scale, making use of a wide range of data available in the MeerKAT archive. I detail the process of review and beta-testing that resulted in the final design of the Bursts from Space: MeerKAT project. Over 1000 volunteers took part, demonstrating a healthy appetite for further Zooniverse data releases. Volunteers discovered or recovered a wide range of phenomena, from flare stars and pulsars to scintillating AGN and transient OH maser emission. I was also able to use the known transients in our fields to understand some reasons why interesting sources may be missed and will fold this learning through to future iterations of the project. This is the first demonstration of volunteers finding radio transients in images.</p>
<p>Finally, I show how anomaly detection, an unsupervised machine learning approach, is a suitable tool for finding these variable phenomena at scale, as is required for modern astronomical surveys. I use three feature sets as applied to two anomaly detection techniques in the Astronomaly package and analyse anomaly detection performance by comparison with citizen science labels. By using transients found by citizen scientists as a ground truth I demonstrate that anomaly detection techniques can recall over half of the radio transients within 10% of the sample dataset. I find that the choice of feature set is crucial, especially when considering available resources for human inspection and follow-up. I find that active learning on ∼2% of the data improves recall by up to 10%, depending on the feature-model pair. The best performing feature-model pairs result in a factor of 5 times fewer sources requiring vetting by humans. This is the first effort to apply anomaly detection techniques to finding radio transients and shows great promise for application to other datasets, a real-time transient detection system and upcoming large surveys.</p>
|
|---|