Unveiling the dark Universe with HI and EMBER-2

TL;DR

This paper introduces EMBER-2, a novel deep learning model that predicts dark matter distributions from HI data across cosmic time, significantly improving accuracy over traditional methods and aiding future radio survey analyses.

Contribution

The paper presents EMBER-2, a new model trained on simulations to directly infer dark matter from HI observations, capturing complex relations more effectively than empirical approaches.

Findings

Accurately recovers dark matter mass fractions and surface density profiles.

Reconstructs cross-correlations at 20% accuracy down to k=100 h/cMpc.

Outperforms traditional empirical methods in key statistical measures.

Abstract

Next-generation radio telescopes will provide unprecedented data volumes of the neutral hydrogen (HI) distribution across cosmic time. The spatial and kinematic distribution of HI is a biased tracer of the underlying matter field, and as such contains information on the distribution of dark matter over a wide range of scales. Extracting dark matter properties from HI, however, is non-trivial because baryonic processes linked to galaxy formation significantly modify the HI distribution. Additionally, methods that use empirical relations, often calibrated via numerical simulations, do not use the full field-level information to model the complex relation between HI and dark matter. We use the recently introduced EMBER-2 model to directly predict dark matter distributions from HI tracers over a wide redshift range, z=0-6. After training on cosmological galaxy formation simulations run with FIRE-2, our method accurately recovers key statistics, including dark matter mass fractions, surface density profiles and cross-correlations, where the latter are reconstructed at an accuracy of 20% down to scales of k = 100 h/cMpc constituting a significant improvement over traditional approaches. The presented method may become a key ingredient in future inference pipelines as it can be readily integrated into downstream analysis tasks of radio surveys.