Justyn R Maund

Time-domain astronomy is entering a new era as wide-field surveys with higher cadences allow for more discoveries than ever before. The field has seen an increased use of machine learning and deep learning for automated classification of transients into established taxonomies. Training such classifiers requires a large enough and representative training set, which is not guaranteed for new future surveys such as the Vera Rubin Observatory, especially at the beginning of operations. We present the use of Gaussian processes to create a uniform representation of supernova light curves from multiple surveys, obtained through the Open Supernova Catalog for supervised classification with convolutional neural networks. We also investigate the use of transfer learning to classify light curves from the Photometric LSST Astronomical Time Series Classification Challenge (PLAsTiCC) data set. Using convolutional neural networks to classify the Gaussian process generated representation of supernova light curves from multiple surveys, we achieve an Area Under the Receiver Operating Characteristic curve (AUC) score of 0.859 for classification into Types Ia, Ibc, and II. We find that transfer learning improves the classification accuracy for the most under-represented classes by up to 18 per cent when classifying PLAsTiCC light curves, and is able to achieve an AUC score of 0.946 ± 0.001 when including photometric redshifts for classification into six classes (Ia, Iax, Ia-91bg, Ibc, II, and SLSN-I). We also investigate the usefulness of transfer learning when there is a limited labelled training set to see how this approach can be used for training classifiers in future surveys at the beginning of operations.