Data-driven characterisation of variable source light curves with multimodal representation learning

Abstract

Surveys such as the Zwicky Transient Facility (ZTF) and the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) produce light curves for hundreds of millions of variables. Their characterisation traditionally relies on supervised classifiers with hand-crafted features and curated labels, limiting them to predefined taxonomies and individual surveys. We present a general, survey-agnostic framework that characterises variable sources without labelled training data, producing low-dimen…