AlphaGen: a new neural network for fast multi-plane gravitational lensing

Charles Wilson ( Université de Montréal )

Strong gravitational lensing has proven to be an invaluable probe of the abundance of low-mass dark matter halos, which is at the core of the current “Standard Model” of cosmology’s missing satellites problem. With upcoming experiments such as the Rubin Observatory and the Euclid space telescope projected to discover more than 100 000 new strong gravitational lens systems, much effort has been dedicated to developing novel statistical frameworks for detecting the extremely faint signature of light halo populations in this wealth of lensing data. Many such approaches require efficient and accurate simulators of the effects of line-of-sight halos, the modeling of which has historically been limited by the high computational complexity of the traditional multi-plane lensing framework. We present AlphaGen, a neural-network based modeling pipeline for the simulation of multi-plane deflection angles, conditioned on a population of line-of-sight dark matter halos and a user-defined main lens profile. The novel network architecture developed in this work is GPU-accelerated, and simulates multi-plane lensing at an HST-like resolution nearly two orders of magnitude faster than the traditional framework. I will present the result of our method being used to accurately model the line-of-sight effects in the lensing of real galaxy images from the COSMOS survey, and discuss promising future use cases of the pipeline.