Unveiling Quasars in Gaia DR3: Multimodal Classification and Redshift Estimation of BP/RP Sources with Deep Learning

Abstract We present a multimodal deep learning framework that classifies stars, galaxies, and quasars and accurately estimates quasar redshifts using Gaia DR3’s low-resolution BP/RP (XP) spectra, alongside astrometric, photometric, and CatWISE mid-infrared data. Our hierarchical classification model achieves an overall accuracy of 99.90%. Evaluating class-specific performance under a one-versus-rest framework, a precision of 99.61% and a recall of 99.83% are achieved for quasars. Furthermore, our spectral redshift regression model, SpecNet-Z, attains an rms error of 0.0682, a normalized median absolute deviation ( σ NMAD ) of 0.0048, and a catastrophic outlier rate of only 0.42%, demonstrating superior performance in the template-fitting approach in Gaia DR3. Applying this framework to the 219 million sources with public XP spectra, we produce the XP-Quasars catalog, containing 123,342 high-purity quasar candidates. This catalog includes 3763 new candidates not listed in the official GDR3 QSO catalog. Our work provides a robust and scalable method that serves as a valuable resource for current active galactic nuclei and cosmological studies and as a vital precursor for the analysis of the complete XP spectra dataset anticipated in Gaia DR4. The XP-Quasar catalog and associated source code are publicly available via Zenodo ( https://doi.org/10.5281/zenodo.17960118 ) and GitHub ( https://github.com/Shijinghang/xp-quasars ).