A Meta-Learning Framework for Multitask Reverberation Mapping in Active Galactic Nuclei

Abstract

The Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) is expected to observe active galactic nuclei (AGN) at sky densities of approximately 1000-4000 per sq. deg, enabling photometric reverberation mapping on an unprecedented scale. We present a meta-learning framework for AGN photometric reverberation mapping based on Attentive Latent Neural Processes (ALNP), developed by the SER-SAG-S1 directable software in-kind team for LSST. The framework clusters AGN light curves with similar topologies using Self-Organizing Maps and combines ALNPs with Mixture Density Models to learn light-curve structure, supermassive black hole (SMBH) properties, and accretion-disk transfer functions in an unsupervised manner. We evaluate the framework on simulated AGN light curves spanning a range of cadences and transfer functions, as well as on real data from the Zwicky Transient Facility. The learned latent representations encode information on both transfer functions and SMBH parameters. Relative to ensemble-trained baseline regressors, including Gaussian-process models, the framework improves light-curve reconstruction by 60-70%. The transfer function recovery improves by approximately 35% relative to the training prior in a low-variability cluster, while recovery of intrinsic SMBH and red-noise parameters improves by approximately 34%. We further demonstrate that models trained on simulated data can be applied to real AGN light curves. These results indicate that ALNP-based representations provide a flexible and scalable approach to photometric reverberation mapping and are well suited to the diverse AGN population expected from LSST and future time-domain surveys.