Surveys with next generation instruments such as the Australian Square Kilometre Array Pathfinder (ASKAP) and the Large Synoptic Survey Telescope (LSST) will produce an unprecedented amount of data. Time-domain astronomy, especially in the radio waveband, stands to benefit enormously from the wide field of view and fast survey speeds of these instruments. In order to maximise the scientific returns offered by these instruments, it will be necessary to overcome the big data challenges. This thesis explores the use of machine learning techniques to automatically classify transient and variable astronomical sources, and presents a detailed study of one radio variable source, CU Virginis.
CU Virginis is a main sequence star that displays pulsar-like periodic radio pulses. I observed CU Virginis with the Australia Telescope Compact Array and detected 100% circularly polarised pulses twice in a rotation period. The pulse arrival times show a frequency dependence that can be explained using a geometric model in which electron cyclotron maser emission is refracted through cold plasma in the magnetosphere.
The first of the two classification problems addressed in this thesis involves the ‘online’ classification of data streams from the processing pipelines of time-domain surveys such as the Variable and Slow Transients survey (VAST) on ASKAP. First, I simulated the radio light curves of eight types of transients and variables that the VAST survey will likely find. Then, using these simulated light curves, I performed classification on the light curve data stream using a committee of Random Forest classifiers.
The second classification task is more general and involves automatically classifying variable sources from the second XMM-Newton serendipitous source catalogue (2XMM). I trained a Random Forest classifier using 873 known variable sources, with features derived from time series, spectra, and other multi-wavelength contextual information. This resulted in a classifier with a 10-fold classification accuracy of ∼97%. I then applied the trained classification model to 411 unknown variable 2XMM sources to produce a probabilistically classified catalogue.
Machine-learned classification will become an increasingly important part of an astronomer’s toolkit in the big survey era.