Abstract
The lack of tangible evidence for non-gravitational interactions between dark and visible sectors drives the need for exploring new avenues of inferring dark matter properties through purely gravitational probes. In particular, addressing small-scale distribution of dark matter could lead to valuable new insights into its particle nature, either confirming predictions of the collisionless cold dark matter hypothesis or favouring models with suppressed small-scale matter power spectrum. In this work we propose a novel machine learning approach for constraining the abundance of galactic dark matter subhaloes through the analysis of Milky Way's stellar field that has been only recently mapped with sufficient coverage thanks to the Gaia mission. Our method is based on convolutional neural networks which represent a powerful tool for identifying characteristic perturbations in spatial maps of stellar number density and velocity distribution moments. For generating the training data we develop a robust and computationally efficient algorithm, capable of generating mock stellar fields from an arbitrary underlying phase-space distribution of stars. By preforming a preliminary study of the outlined approach on synthetic datasets we demonstrate that sensitivities down to (or even below) $10^8 M_\odot$ could be reached. Furthermore, our results show that the accuracy of the advocated technique crucially depends on the kinematic properties of mapped stars and could be further improved by applying it to abundant stellar populations with particularly low velocity dispersion, such as the galactic thin disc stars.