Speaker
Description
Strong Lensing is a phenomenon predicted by Einstein’s General Relativity in which the light emitted by a distant source is deflected by a massive object in its path, causing the image of the distant object to appear magnified and distorted. This process carries information about the dark matter distribution and the underlying cosmology in galaxies and galaxy clusters, making Strong Lensing a valuable probe of a few different astrophysical phenomena. Currently, the small number of known lenses, in the range of hundreds confirmed and modeled, prevents robust statistical analyses of these objects. However, future surveys such as the Vera Rubin Observatory and the ESA Euclid are expected to greatly increase the number of known strong lenses. Current modeling methods are relatively slow and require supervision, making them unsuitable for the volume of data expected in the next few years. As a result, methods based on Deep Learning have been proposed as alternatives for parameter inference in a likelihood-free fashion. In this work, we leverage Simulation-Based Inference methods to obtain posterior distributions for a few strong lensing parameters. We make use of Deep Learning techniques and, in particular, normalizing flows as density estimators to approach this problem under a Bayesian framework. We trained our models to infer Einstein Radius, lens galaxy velocity dispersion, and redshift of both lens and source objects on realistic wide field DECam-based simulated griz-band images, therefore DES-like, including realistic effects such as PSF and noise accordingly. Our top-performing models are able to generate well-calibrated posterior distributions with up to 83% median precision in inferred values, and show median fractional deviations of 5.5% for measurements of Einstein radius on simulated images. Our current focus is on achieving comparable results in real data.
