Automated classification of transients in optical time-domain surveys

Repeated sky surveys in the past decade have led to the proliferation in the discovery of transients. This has not come without its own challenges: the rate of discovery from current sky surveys greatly exceeds the human capacity to manually identify and classify newly discovered objects. The use of machine learning approaches to automate the discovery process with little to no human intervention is rapidly becoming a standard practice in surveys. In this thesis we present the use of machine learning to classify objects discovered by sky surveys observing at optical wavelengths. The Gravitational-wave Optical Transient Observer (GOTO) is a survey with the aim of searching for the optical counterparts to gravitational waves, while also scanning the night sky for transients and variable objects. We use both a machine learning and a deep learning approach to classify objects observed by GOTO using their light curves, and compare the effectiveness and limitations of both methods for photometric classification. We find that using a deep learning approach with recurrent neural networks works best to reliably classify objects using their light curves in real-time.
We investigate the use of Gaussian processes to create uniform representations of supernova light curves from different surveys. These are then used with a convolutional
neural network for classification into supernova sub-types. Future surveys will have a lack of labelled data to train classifiers. We use transfer learning to show how data from another survey can be used to train a classifier for a new survey. Machine learning is a widely used methodology, and has uses in other fields of research. We show how the classifiers developed for GOTO light curve classification can be adapted for other classification tasks, and how they perform on these tasks.