Deep learning the astrometric signature of dark matter substructure

TL;DR

This paper demonstrates that deep learning, specifically ResNet-18, can effectively identify the astrometric signatures of dark matter subhalos in the Milky Way using simulated survey data, aiding in understanding dark matter distribution.

Contribution

It introduces a novel application of convolutional neural networks to detect dark matter substructure signatures in astrometric data, with methods for localization of subhalos.

Findings

ResNet-18 can classify lensed vs. unlensed quasars in simulated data.

SKA-like surveys can probe dark matter substructure in the Milky Way.

Axiomatic attribution helps localize substructures in lensing maps.

Abstract

We study the application of machine learning techniques for the detection of the astrometric signature of dark matter substructure. In this proof of principle a population of dark matter subhalos in the Milky Way will act as lenses for sources of extragalactic origin such as quasars. We train {\it ResNet-18}, a state-of-the-art convolutional neural network to classify angular velocity maps of a population of quasars into lensed and no lensed classes. We show that an SKA -like survey with extended operational baseline can be used to probe the substructure content of the Milky Way, and demonstrate how axiomatic attribution can be used to localize substructures in lensing maps.