Modeling Lensed Quasars with Neural Posterior Estimation: Complex Mass Models

The Vera C. Rubin Observatory is expected to discover ~10³ galaxy-scale strongly lensed Active Galactic Nucleus(AGN) systems. A machine learning method, neural posterior estimation, is used to automatically model these systems with the intention of using them for time delay cosmography, to constrain the Hubble constant and, ultimately, the properties of Dark Energy. The neural network is trained on simulated lens systems with simple mass distributions that consist of a single central deflector. The robustness of the network’s ability to make predictions for complex mass models is tested by adding a perturbing mass to the lens plane in the mock Legacy Survey of Space and Time(LSST) data test set of 100 lenses. The presence of the perturber introduces a bias in the Hubble constant of ~6%, while reducing the model precision by a factor of 1.25. This indicates the need to train the network with a more complex, realistic mass model consisting of dark and luminous sub structure.