AI and the World's Hard Problems · Astronomy & space · arXiv · 2026-10-22
Data-driven characterisation of variable source light curves with multimodal representation learning
This framework characterizes variable-source light curves without labelled training data by combining a Transformer on raw observations with a convolutional network on periodograms. On 37,006 ZTF sources, a linear classifier using the learned representations achieved 77.1% balanced accuracy across 14 classes and 90.5% across broad categories. Combining the two branches improved balanced accuracy by up to 14.9 percentage points over the best single branch, while concatenating representations from multiple bands raised it to 86.1%.