🔭 astrophysics

Unveiling the dark Universe with HI and EMBER-2

This paper demonstrates that the EMBER-2 model, trained on FIRE-2 cosmological simulations, can accurately reconstruct dark matter distributions from neutral hydrogen (HI) tracers across redshifts 0–6 by overcoming baryonic complexities and outperforming traditional empirical methods in recovering key statistical properties.

Original authors: Mauro Bernardini, Robert Feldmann, Daniel Anglés-Alcázar, Philipp Denzel, Jindra Gensior

Published 2026-03-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mauro Bernardini, Robert Feldmann, Daniel Anglés-Alcázar, Philipp Denzel, Jindra Gensior

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Seeing the Invisible

Imagine you are in a dark room filled with invisible furniture (the Dark Matter). You can't see the furniture itself, but you can see the dust motes dancing in the air around it (the Neutral Hydrogen, or HI).

For a long time, astronomers have tried to figure out where the invisible furniture is just by looking at the dust. The problem? The dust doesn't just sit there; it gets blown around by fans, swept up by brooms, and rearranged by the people in the room (galaxy formation processes like stellar winds and black holes). This makes it very hard to guess the shape of the furniture just by looking at the dust.

This paper introduces a new "AI detective" called EMBER-2. It's a machine learning model that looks at the dust (HI) and instantly figures out exactly where the invisible furniture (Dark Matter) is, even when the dust has been messed around with.

The Problem: The "Bad Map"

Traditionally, scientists tried to map the invisible furniture using simple rules.

The Old Way: They would say, "If there is a big pile of dust, there must be a big chair underneath." They used simple math formulas based on simulations.
The Flaw: This is like trying to guess the layout of a messy bedroom just by looking at a single sock. It misses the complex details. The relationship between the dust and the furniture is messy and non-linear. Simple rules fail to capture the full picture, especially on small scales.

The Solution: The "Super-Translator" (EMBER-2)

The authors built a neural network (a type of AI) called EMBER-2. Think of it as a super-advanced translator that has studied millions of "rooms" in a video game simulation.

The Training: They fed the AI a massive amount of data from a supercomputer simulation called FIRE-2. In this simulation, they knew exactly where the invisible furniture was and exactly where the dust was.
The Learning: The AI learned the complex, messy relationship between the two. It learned that "Oh, when the dust swirls this way, it's because a heavy table is there, even if the dust looks light."
The Magic: Now, when you show the AI a real map of the dust (HI) from the universe, it can "translate" it back into a map of the invisible furniture (Dark Matter) with incredible accuracy.

How Good Is It? (The Results)

The paper tests this AI by comparing its "guesses" against the actual simulation data (the ground truth).

The "X-Ray" Vision: Figure 1 in the paper shows a side-by-side comparison. The AI's reconstruction looks almost identical to the real simulation. It can see the big structures (like galaxy clusters) and the tiny details (like individual halos) with surprising clarity.
The "Mass" Check: The AI correctly predicts how much "invisible furniture" exists in different areas. It gets the weight of the dark matter right, whether it's a tiny speck or a massive cluster.
The "Speed" Test: The AI can do this mapping across a huge range of time (from 6 billion years ago to today) and space.
The Accuracy: The AI is accurate to within 20% even on very small scales. This is a huge improvement over the old "simple rule" methods, which often get lost in the details.

Why Does This Matter?

We are about to get a flood of new data from giant radio telescopes (like the SKA). These telescopes will map the "dust" (HI) across the entire universe in 3D.

Before: We had the dust map, but we couldn't reliably turn it into a dark matter map because the physics was too complicated.
Now: With EMBER-2, we can take that new, massive dust map and instantly generate a high-quality map of the Dark Matter.

The Future: A Multi-Tool Approach

The authors suggest that this isn't just a one-trick pony.

Adaptability: Just like you can train a dog to do different tricks, this AI can be retrained to handle different types of "noise" (like the static on a radio) or even added new senses (like adding starlight data from optical telescopes).
Speed: It works fast enough (seconds per map) to be used in real-time analysis pipelines.

The Bottom Line

This paper is like handing astronomers a pair of smart glasses. Before, they could only see the dust in the dark room. Now, with EMBER-2, they can look at the dust and instantly see the invisible Dark Matter structure underneath, revealing the true skeleton of our Universe.

1. Problem Statement

The physical nature of dark matter (DM) remains one of the most significant open questions in astrophysics. While the $\Lambda$ CDM model successfully describes large-scale structure, directly inferring DM properties from observational data is challenging.

Limitations of Current Tracers: Weak Lensing (WL) requires massive statistical samples and precise modeling of systematic effects (e.g., baryonic feedback).
The HI Connection: Neutral Hydrogen (HI) serves as a complementary tracer of the underlying DM distribution. However, the relationship between HI and DM is complex and non-linear, heavily modified by baryonic processes (stellar winds, AGN feedback, ram pressure stripping) that redistribute matter on Mpc scales.
Methodological Gaps: Traditional approaches, such as Halo Occupation Distribution (HOD) models and Bias Expansion (BE), rely on empirical relations or a limited set of fitting parameters. These methods often discard the full field-level information stored in the data, failing to capture the complex interplay between HI and DM, particularly on small scales ( $<1$ Mpc).

2. Methodology

The authors propose a data-driven Machine Learning (ML) approach to perform an "inverse mapping": predicting the underlying 2D dark matter density and velocity fields directly from 2D projected HI tracers.

Model Architecture (EMBER-2): The study utilizes EMBER-2, a Conditional Generative Adversarial Network (cGAN) introduced in a related work (Bernardini et al. 2025).
- Input ( $x$ ): Projected HI surface density ( $\Sigma_{HI}$ ) and line-of-sight radial velocity ( $v_{HI}$ ).
- Output ( $y$ ): Reconstructed Dark Matter surface density ( $\Sigma_{dm}$ ) and radial velocity ( $v_{dm}$ ).
- Mechanism: The generator ( $G$ ) learns the conditional probability distribution $p(y|x, z)$ to produce DM samples that match the true distribution, while a discriminator ( $D$ ) attempts to distinguish between real and generated samples. The model is trained across a redshift range of $z = 0 - 6$ .
Training Data:
- Simulations: Cosmological hydrodynamical simulations from the FIREbox suite (specifically FB30 and FB15), run with the FIRE-2 galaxy formation physics model using the GIZMO code.
- Resolution: The simulations use a box length of 30 cMpc/ $h$ (FB30) with $1024^3$ particles. Data is projected onto $1024^2$ grids with a spatial resolution of $\sim 43$ ckpc.
- Dataset Construction: Particle data is aggregated into slabs of 1.5 Mpc/ $h$ . The model is trained on projections from two axes of the FB30 simulation (using 3-fold cross-validation) and tested on the third axis of FB30 and all axes of the independent FB15 simulation (generated with a different random seed).
Preprocessing: The model uses scaling functions to normalize density and velocity fields. Data augmentation includes random cropping ( $128^2$ patches), rotation, and flipping.

3. Key Contributions

Inverse Mapping Framework: Unlike previous works that predict HI from DM, this paper demonstrates the feasibility of reconstructing the Dark Matter field from HI tracers across a wide redshift range ( $z=0-6$ ).
Field-Level Information Utilization: The method leverages the full spatial and kinematic information of the HI field, avoiding the information loss inherent in HOD or simple bias expansion models.
Adaptability: The fully convolutional architecture allows the model to capture correlations across a broad range of scales, from large-scale cosmic web structures down to individual halo profiles.
Speed: The inference time is on the order of seconds per map, making it suitable for integration into downstream analysis pipelines for future radio surveys (e.g., SKA).

4. Results

The performance of EMBER-2 was evaluated using multiple statistical metrics comparing the reconstructed maps against the ground-truth simulations:

Visual Reconstruction: Qualitative comparisons (Figure 1) show high structural similarity between simulated and reconstructed DM fields, accurately capturing density and velocity patterns across different dynamic ranges.
Mass Function: The model accurately reproduces the structure mass function ( $\phi$ ) of dark matter halos across all redshifts, particularly for masses $\log M_{dm} \leq 12$ .
HI to DM Ratios: The reconstructed ratio of Dark Matter mass to HI mass ( $M_{dm}/M_{HI}$ ) as a function of peak HI column density ( $N_{HI}$ ) matches simulations with excellent agreement. The model correctly captures the scatter for various systems: Intergalactic Medium (IGM), Lyman Limit Systems (LLSs), sub-DLAs, and DLAs.
Radial Profiles: The surface density profiles ( $\Sigma_{dm}/\Sigma_{HI}$ ) for LLSs, sDLAs, and DLAs at $z=0$ are reconstructed with correct median amplitude and slope, both in the inner and outer regions of halos.
Cross-Correlations:
- The model reconstructs cross-correlation coefficients between HI and DM fields with errors less than 20% for both surface density ( $\Sigma$ ) and kinetic energy density ( $\pi$ ).
- This accuracy holds for scales up to $k = 100 \, h/\text{cMpc}$ .
- This represents a significant improvement over traditional HOD/Bias methods, which typically struggle to achieve $<10\%$ accuracy at $k \lesssim 1 \, h/\text{cMpc}$ and fail at smaller scales.
- Note: A slight deviation ( $\sim 25\%$ ) was observed at $z=0$ for the FB30 test set, attributed to the small number of massive halos in the simulation volume, rather than a model failure (confirmed by the FB15 test set).

5. Significance and Outlook

Future Surveys: With the advent of next-generation radio telescopes like the Square Kilometre Array (SKA) and pathfinders like MeerKAT, vast amounts of HI data will become available. EMBER-2 provides a critical tool to convert this HI data into high-fidelity Dark Matter maps.
Baryonic Marginalization: By learning the complex mapping from simulations that include realistic baryonic feedback, the method can effectively marginalize over baryonic effects, which are a major source of systematic error in DM inference.
Pipeline Integration: The speed and accuracy of the model allow it to be integrated into analysis frameworks (e.g., KARABO) to simulate instrumental noise and systematics, facilitating joint modeling of simulated and observational data.
Multi-wavelength Potential: The framework can be extended to include additional input channels, such as stellar light from optical surveys (LSST, ELT), potentially further enhancing reconstruction accuracy.

In conclusion, this work establishes a proof-of-concept that deep learning models trained on cosmological simulations can successfully invert the HI-to-DM relationship, offering a powerful new avenue for probing the dark sector of the Universe.