Machine learning for cosmic explosions

Astrophysical transients are some of the most dynamic and fascinating events in the universe. These are sources that change in brightness over time - sometimes rapidly, sometimes gradually -and they include everything from exploding stars to flaring black holes. Supernovae, in particular, are not only spectacular explosions marking the end of a star’s life, but also crucial tools for understanding the cosmos, from the life cycle of galaxies to the expansion of the universe. The Vera C. Rubin observatory will detect millions of transients, producing light curves similar to the figures below. However, these light curves alone can't definitively identify the type of source that produced them. For that, machine learning is required.

My early work on transient classification established a benchmark for machine learning applied to photometric light curves. Machine learning classification is essential for the LSST alert stream, and while this field has advanced rapidly, my pioneering work established its foundation.  In Lochner et al., ApJS, 2016, which has over 250 citations, we systematically compared machine learning algorithms and feature extraction techniques, introduced standard metrics, and emphasised the importance of representative training sets—practices that are now standard in astronomy.

To manage the anticipated 20 million events per night alert stream from the Vera C. Rubin Observatory, the LSST community has developed brokers: pieces of software that run multiple algorithms to classify and tag incoming alerts. ANTARES is one of the first brokers to be developed and currently runs on the Zwicky Transient Facility alert stream. In Narayan et al. (ApJS, 2018), cited more than 60 times, we incorporated the methodology and approach to classification developed in Lochner et al. (2016) into the ANTARES broker. The algorithm incorporates novel wavelet-based techniques to extract features, enabling accurate classification even with sparse data.

Since my early publications, the field of transient classification has grown rapidly, driven by the increasing demand for accurate algorithms as new telescopes come online. I contributed to the development of the Photometric LSST Astronomical Time-series Classification Challenge (PLAsTiCC): a innovative public challenge to develop classifiers for LSST time-series data (Allam et al., ArXiv, 2018, Malz et al., AJ, 2019, Hložek et al., ApJS, 2023). My early work influenced the winning algorithms and helped advance transient classification methodologies

The rapid increase in radio transient detections from MeerKAT has highlighted the growing need for advanced machine learning approaches tailored to radio data. In Sooknunan et al. (MNRAS, 2021), we adapted the techniques originally developed for optical transient classification to analyse radio light curves. This study not only demonstrated the effectiveness of these methods in a new domain but also showcased how incorporating complementary optical data can significantly enhance classification accuracy.