Conceptual

ORACLE: Hierarchical Deep-Learning Real-Time Classifier for Astronomical Transients

ORACLE is a GRU-based recurrent neural network that classifies LSST/Rubin transient and variable sources in real time along an observationally driven taxonomy, trained with a custom hierarchical cross-entropy loss so predictions stay consistent from coarse to fine classes. It concatenates the light-curve embedding with host-galaxy context (photometric redshift, offset, ellipticity, brightness) and produces high-confidence coarse classifications from as little as one photometric epoch, matching state-of-the-art fine-grained accuracy while classifying much earlier.