Machine learning applications for transient discovery facing Rubin Science

Abstract

The Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST) will revolutionise time-domain astronomy by producing millions of photometric alerts per night, requiring scalable and robust machine-learning (ML) methods for both variability characterisation and rapid transient identification. This thesis develops ML approaches to address two key challenges in this data-intensive regime: modelling sparsely sampled nuclear variability and identifying rare extragalactic transients in real time. First, we investigate active galactic nucleus (AGN) variability using a Stochastic Recurrent Neural Network (SRNN) applied to decade-long light curves simulated with Continuous Auto-Regressive Moving Average (CARMA) models under realistic LSST cadences. This work presents the first application of SRNNs in astronomy for reconstructing LSST-like AGN light curves and recovering underlying variability parameters. We show that long-term variability amplitudes can be recovered with reasonable accuracy, while decorrelation timescales remain difficult to constrain for both SRNN and CARMA models, particularly in the presence of long seasonal gaps. An evaluation of proposed Wide-Fast-Deep cadence strategies demonstrates that the structure of seasonal gaps is the dominant factor affecting reconstruction performance. Second, we present NEEDLE, a hybrid ML classifier designed for the early identification of rare transients, specifically hydrogen-poor superluminous supernovae (SLSNe-I) and tidal disruption events (TDEs). NEEDLE combines convolutional and dense neural networks with multi-modal inputs, including image cutouts, early-time photometry, and host-galaxy contextual information. Despite extreme class imbalance, the classifier achieves high completeness and is optimised for rapid spectroscopic follow-up. NEEDLE is deployed on the LSST:UK alert broker Lasair and has successfully enabled real-time identification of multiple SLSNe and infant TDEs. Finally, this thesis demonstrates that tailored data restoration and augmentation strategies are essential for robust ML performance in alert-stream environments, highlighting the importance of data quality control in maximising the scientific return of LSST.