Bayesian galactic cartography via Gaussian processes: dark matter and interstellar gas : towards a new era of data-driven high-dimensional modelling in astrophysics

RWTH Publications (RWTH Aachen)

Abstract

Looking at the night sky, we observe the Milky Way in every direction. However, even with the influx of large quantities of high-quality data in the past decades, our understanding of the three-dimensional structure of our Galaxy remains limited. This dissertation addresses this challenge in a methodological and data-driven manner, applying Bayesian inference techniques tailored to high-dimensional problems combined with Gaussian process priors to reconstruct key Galactic components from observational data.The first part focuses on determining the dark matter density in the solar neighbourhood using stellar kinematic data from Gaia DR3 and radial velocities from various spectroscopic surveys. As stars move through the Galactic disk, the statistical properties of their motions encode information about the underlying gravitational potential, and thus the mass distribution, including dark matter. Instead of fitting restrictive parametric models to data, we build a bottom-up probabilistic forward model, employing Gaussian processes to flexibly capture smooth variations in the velocity statistics of a stellar tracer population with height above the Galactic plane. We obtain a robust constraint on the local dark matter density, ${\rho_\mathrm{DM} = 0.0117 \pm 0.0035\, \mathrm{M}_\odot\,\mathrm{pc}^{-3}}$, along with insights into the systematic effects of modelling choices, particularly concerning symmetry assumptions.The second part presents a novel framework for reconstructing the three-dimensional distribution of neutral atomic and molecular hydrogen gas in the Milky Way. The radiation emitted by the interstellar gas carries information about its spatial distribution and kinematics, but deprojecting spectroscopic line survey data into three-dimensional maps is an inherently ill-posed inverse problem. Our Bayesian approach jointly reconstructs both gas constituents within a unified probabilistic model using three-dimensional Gaussian process priors on a spherical grid, while accounting for radiative transfer effects. Crucially, we employ a common, spatially varying velocity field that couples the kinematic information of both species, eliminating artefacts associated with fixed velocity models. As a result of the reconstruction, we obtain a set of spatially coherent gas maps, revealing key features of global and local Galactic structure, such as disc warping and flaring, as well as the local bubble.Together, these projects demonstrate the power of forward-modelling and high-dimensional Bayesian inference with Gaussian process priors for extracting maximal information from astronomical surveys while rigorously quantifying uncertainties.

Authors 1

1. Laurin Söding corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen