Friday

Modern synoptic sky surveys have proven transformative for time-domain astrophysics, only possible by high-performance machine learning classification to efficiently sift astrophysical sources from artifacts in the vast quantities of data these projects generate. With upcoming step changes in survey capability (e.g. Rubin Observatory's LSST) further increasing this data rate, continued improvements are required to continue to efficiently leverage the unprecedented discovery stream these projects provide.

We present recent progress on source classification in difference images (Killestein et al., 2021), applying Bayesian deep learning to the incoming candidate stream of the Gravitational-wave Optical Transient Observer (GOTO) survey. Our classifier provides uncertainty-aware predictions to assist human interpretation, and can be trained to state-of-the-art (1% FPR, 1.5% FNR) performance with minimal human labelling via a fully-automated data generation procedure. A key improvement is the implementation of a novel data-driven augmentation scheme to generate realistic synthetic supernovae to train on, yielding a CNN classfier uniquely optimised for locating extragalactic transient sources. This provides significant improvements in the recovery of faint and nuclear transients of key interest, and the kilonovae that form the principal science goal of the GOTO project.

We will also showcase early results from the next-generation GOTO source classifier, leveraging contextual information to provide rich, multi-label, hierarchical classification -- reducing human vetting effort in the short-term, with the overarching aim of fully autonomous, real-time triggering of follow-up resources.