Strong Lensing, Dark Perturbers, and Machine Learning

Abstract

Observations of the Universe generally support $\Lambda$CDM, but since distant structures $\lesssim 10^{9}\,M_\odot$ are usually too dim to observe, their properties are still poorly constrained. Competing particle models of Dark Matter (DM) make different predictions, so these scales can teach us about the nature of DM, even in the absence of a direct detection. A promising probe is to use the subtle gravitational lensing effect of these dim structures, which is directly sensitive to mass rather than luminosity. This effect is too weak to measure on its own, but a strong gravitational lens, where one background object is lensed and appears (distorted) in multiple places, provides enough additional redundancy that the different copies of the background object can be compared and analyzed for weak gravitational perturbations.

This thesis consists of three parts. (1) We study the combined statistical effect of many small gravitational perturbers. These come in two varieties, subhalos and interlopers. Subhalos physically lie within the main lens (a galactic halo), whereas interlopers are aligned by chance along the line of sight. The statistical effect due to subhalos has been studied before, but we show that interlopers are most likely the dominant component and certainly cannot be ignored in future analyses. (2) We analyze a real observation of a strong lens known to harbor a relatively large, individually detectable perturber. We reanalyze the system and find the perturber is better fit as a line-of-sight interloper rather than a subhalo as initially assumed. (3) We demonstrate how machine learning can accelerate the processing required to find individual perturbers. This is pertinent because we expect of order $10^5$ strong lenses discovered by 2030, and traditional sampling-based methods are too computationally expensive to scale to this influx of data. We demonstrate that a UNet can find individual subhalos in realistic simulated lenses, but the substructure must have a high concentration parameter.