Inferring Astrophysics and Dark Matter Properties from 21cm Tomography using Deep Learning

TL;DR

This paper introduces a 3D convolutional neural network to infer astrophysical and dark matter properties from 21cm tomography data, achieving high accuracy even with realistic noise and foreground effects.

Contribution

It presents a novel deep learning approach for joint inference of astrophysics and dark matter parameters from 3D 21cm light-cones, outperforming traditional summary statistics.

Findings

High-fidelity parameter recovery ($R^2>0.78$) for astrophysics and dark matter parameters.

Robust inference under realistic noise and foreground conditions, especially for $ ext{WDM mass}$.

Transfer learning shows robustness for some parameters but increased bias for dark matter properties.

Abstract

21cm tomography opens a window to directly study astrophysics and fundamental physics of early epochs in our Universe's history, the Epoch of Reionisation (EoR) and Cosmic Dawn (CD). Summary statistics such as the power spectrum omit information encoded in this signal due to its highly non-Gaussian nature. Here we adopt a network-based approach for direct inference of CD and EoR astrophysics jointly with fundamental physics from 21cm tomography. We showcase a warm dark matter (WDM) universe, where dark matter density parameter $\Omega_\mathrm{m}$ and WDM mass $m_\mathrm{WDM}$ strongly influence both CD and EoR. Reflecting the three-dimensional nature of 21cm light-cones, we present a new, albeit simple, 3D convolutional neural network for efficient parameter recovery at moderate training cost. On simulations we observe high-fidelity parameter recovery for CD and EoR astrophysics ($R^2>0.78-0.99$), together with DM density $\Omega_\mathrm{m}$ ($R^2>0.97$) and WDM mass ($R^2>0.61$, significantly better for $m_\mathrm{WDM}<3-4\,$keV). For realistic mock observed light-cones that include noise and foreground levels expected for the Square Kilometre Array, we note that in an optimistic foreground scenario parameter recovery is unaffected, while for moderate, less optimistic foreground levels (occupying the so-called wedge) the recovery of the WDM mass deteriorates, while other parameters remain robust against increased foreground levels at $R^2>0.9$. We further test the robustness of our network-based inference against modelling uncertainties and systematics by transfer learning between bare simulations and mock observations; we find robust recovery of specific X-ray luminosity and ionising efficiency, while DM density and WDM mass come with increased bias and scatter.