pith. sign in

arxiv: 2605.29876 · v1 · pith:OAEAI5QRnew · submitted 2026-05-28 · 🌌 astro-ph.CO · astro-ph.GA

Exploring the High-Redshift 21-cm Signal via Self-Consistent Simulations using Artificial Neural Network Emulation

Colton R. Feathers , Eli Visbal , Steven Murray , Ryan Hazlett , Yin-Zhe Ma This is my paper

Pith reviewed 2026-06-29 06:06 UTC · model grok-4.3

classification 🌌 astro-ph.CO astro-ph.GA
keywords 21-cm brightness temperaturecosmic dawnstar formation calibrationneural network emulationsemi-numeric simulationspopulation III starsHERA observationsdark matter halo mergers
0
0 comments X

The pith

Self-consistent neural network simulations of cosmic dawn predict detectable 21-cm signals at redshifts below 25.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper develops a semi-numeric simulation of the cosmic dawn that uses a neural network to emulate star formation rates calibrated directly from hydrodynamic simulations. The model incorporates stochastic dark matter halo merger histories and large-scale density fluctuations to generate predictions for the 21-cm brightness temperature. It shows that including these elements produces earlier and more intense star formation than previous models, with Population II stars dominating the signal at lower redshifts. The results indicate that observations with instruments like HERA could detect the absorption feature and thereby constrain the timing between different populations of early stars.

Core claim

The central discovery is that calibrating small-scale star formation to hydrodynamic simulations via neural network emulation, combined with a critical halo mass for star formation that accounts for molecular hydrogen self-shielding and stochastic merger histories, leads to earlier star formation and higher star formation rates throughout the cosmic dawn epoch. This produces a 21-cm brightness temperature that is detectable at redshifts less than or equal to 25 with 1080 hours of HERA observations, and the absence of a detection at redshifts greater than or equal to 20 would imply a delay time between Population III and Population II star formation of at least 30 million years.

What carries the argument

Neural network emulation of star formation efficiency from AEOS and Renaissance hydrodynamic simulations, embedded in a semi-numeric code tracking large-scale fluctuations and stochastic halo mergers.

If this is right

  • More efficient Population II star formation produces stronger Lyman-alpha coupling and thus a shallower, wider absorption trough in the 21-cm signal.
  • Population II star formation dominates the 21-cm power spectrum at redshifts below 20 and on smaller scales, while Population III dominates at redshifts above 34 and on larger scales at intermediate redshifts.
  • The delay period between Population III and Population II star formation has a significant impact on the 21-cm brightness temperature.
  • Including dark matter halo merger histories is required to accurately model the transition between the two populations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Future work could test whether incorporating additional small-scale physics such as varying initial mass functions alters the dominance of different populations in the power spectrum.
  • The model's emphasis on stochastic histories suggests that ensemble predictions may show larger variance in the 21-cm signal than mean-field approaches.
  • These simulations could inform the design of next-generation radio arrays by highlighting the redshift ranges where the signal is most sensitive to early star formation physics.

Load-bearing premise

The neural network trained on the AEOS and Renaissance simulations accurately reproduces the small-scale star formation efficiency and its dependence on local conditions when placed inside the semi-numeric code with stochastic mergers.

What would settle it

A non-detection of the 21-cm absorption signal at redshifts between 20 and 25 after 1080 hours of HERA integration under moderate foregrounds would indicate that the delay time between Population III and II star formation is at least 30 million years or that the star formation rates are lower than modeled.

Figures

Figures reproduced from arXiv: 2605.29876 by Colton R. Feathers, Eli Visbal, Ryan Hazlett, Steven Murray, Yin-Zhe Ma.

Figure 1
Figure 1. Figure 1: Results of our fiducial simulation framework. Top panel: The average 21-cm brightness temperature, 𝛿𝑇𝑏(𝑧), with 1𝜎 standard deviation. Second panel: Redshift evolution of the 21-cm power, Δ 2 (𝑘), for various 𝑘-modes. Third panel: A light cone of the cumulative total (PopII + PopIII) stellar mass formed. Fourth panel: A light cone of the Lyman-𝛼 background intensity, 𝐽𝛼. Fifth panel: The observed 21-cm lig… view at source ↗
Figure 2
Figure 2. Figure 2: Central slices of our fiducial simulation volume at various redshifts. From top to bottom, the rows depict 1. the cumulative PopIII + PopII stellar mass, 𝑀★, 2. the Ly-𝛼 background intensity, 𝐽𝛼, 3. the 21-cm brightness temperature, 𝛿𝑇b, and 4. its differential with respect to the average. From left to right, each column depicts a different redshift step from 𝑧 = 40 − 15. The remaining four panels of [PIT… view at source ↗
Figure 3
Figure 3. Figure 3: A comparison of the average results of each simulation framework. From top to bottom, panels compare the resulting PopIII SFRD, PopII SFRD, the ratio of the two SFRDs, the LW background intensity (𝐽LW), and the Ly-𝛼 background intensity (𝐽𝛼) averaged across all cells. Within the SFRD ratio plot, vertical dotted lines correspond to the redshift at which each simulation experiences the transition from PopIII… view at source ↗
Figure 4
Figure 4. Figure 4: The average error of our NNs by simulation framework as a function of redshift. Each curve represents the average error of 100 randomly sampled cells from the large-scale simulation. Left panel: The average error of our bursty (thick lines) and steady (thin lines, if applicable) PopII emulations. Right panel: The average error of our PopIII emulations. In both panels, we show the redshift-averaged error of… view at source ↗
Figure 5
Figure 5. Figure 5: Left: A comparison of the average IGM 𝑇S (solid) and 𝑇K (dashed) resulting from each of our simulation frameworks, with the CMB temperature represented by the dotted black line. Right: The corresponding 𝛿𝑇b(𝑧) compared with previously published 21-cm predictions. For comparison, we include the claimed EDGES detection (gray dashed, Bowman et al. 2018) as well as various theoretical predictions from Liu et a… view at source ↗
Figure 6
Figure 6. Figure 6: A comparison of the dimensionless 21-cm power spectra resulting from each simulation framework. From top left to bottom right, we compare the resulting Δ 2 (𝑘) at various redshifts from 𝑧 = 40 − 15. At each redshift, the top panel shows the power spectra of each semi-numeric simulation, while the bottom panels depict the corresponding ratio of each model with respect to the fiducial. We also show the resul… view at source ↗
Figure 7
Figure 7. Figure 7: Redshift evolution of the 21-cm 𝛿𝑇b power on four various 𝑘-modes, as denoted by the title. The top subplot of each panel displays the power evolution of each simulation, and the bottom subplot shows the ratio of power with respect to the fiducial model. We also show the power evolutions of M22 AllGalaxies model resulting from 21cmFAST (gray dot-dashed), as well as HERA sensitivity curves obtained from 21c… view at source ↗
read the original abstract

We present a novel, self-consistent, semi-numeric Cosmic Dawn (CD) simulation in which small-scale star formation (SF) is calibrated to the \emph{AEOS} and \emph{Renaissance} hydrodynamic simulations. SF proceeds within dark matter (DM) halos via neural network emulation while considering large-scale fluctuations in density and feedback. We translate the resulting 3D distribution of galaxies into predictions for the 21-cm brightness temperature, \Tb, and power spectrum, \PS. We simulate several unique realizations to study the impact of varying astrophysics on \Tb, finding that more efficient Population II (PopII) SF largely yields stronger Lyman-$\alpha$ coupling, resulting in a shallower and wider absorption trough. However, we find that PopII SF dominates \PS\ at $z \lesssim 20$ and on smaller scales at intermediate redshifts ($k \gtrsim 0.2\ \mathrm{Mpc^{-1}}$ at $z \simeq 34-20$) while Population III (PopIII) SF dominates \PS\ at $z\gtrsim34$ and on larger scales at intermediate redshifts. Compared with previous works, we find that the combination of hydrodynamic SF calibration, a critical halo mass for SF considering \Htwo\ self-shielding, and stochastic DM halo merger histories results in both earlier SF and higher SF rates across CD. Further, we find that the delay period separating PopIII and PopII SF (\tdelay) significantly impacts \Tb, and that one must include DM halo merger histories to properly account for this transition. Finally, we find our fiducial \Tb\ to be detectable at $z\lesssim25$ with 1080 hours of HERA observations under moderate foreground assumptions, and the lack of such a detection at $z \gtrsim 20$ would suggest \tdelay\ $\gtrsim$ 30 Myr.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces a semi-numeric Cosmic Dawn simulation that emulates small-scale star formation via a neural network trained on the AEOS and Renaissance hydrodynamic runs, incorporating large-scale density/feedback fluctuations and stochastic DM halo merger histories. It generates 3D galaxy distributions to predict the 21-cm brightness temperature Tb and power spectrum PS, finding that the setup produces earlier and higher star-formation rates than prior work, with Pop II dominating PS at z ≲ 20 (and smaller scales at intermediate z) while Pop III dominates at z ≳ 34 (and larger scales intermediately); the fiducial Tb is stated to be detectable at z ≲ 25 with 1080 hours of HERA data under moderate foregrounds, and non-detection at z ≳ 20 would imply tdelay ≳ 30 Myr.

Significance. If the neural-network transfer holds, the framework supplies a route to self-consistent 21-cm forecasts that include stochastic merger histories and hydro-calibrated efficiencies, enabling exploration of astrophysical variations across multiple realizations and potential constraints on the Pop III–Pop II transition delay from future observations.

major comments (3)
  1. [Methods (NN emulation section)] Methods (neural-network emulation and embedding procedure): the central claim of earlier/higher SF rates and the resulting Tb/PS evolution rests on the NN (trained on limited-volume AEOS/Renaissance runs) accurately reproducing small-scale SF efficiency, its density/feedback dependence, and the tdelay transition once placed inside the semi-numeric code that adds stochastic large-scale mergers; no explicit validation or domain-shift test for this transfer is described, which directly affects the reported Pop II/Pop III dominance and HERA thresholds.
  2. [Results (Tb and HERA section)] Results (Tb evolution and detectability statements): the assertion that the fiducial model is detectable at z ≲ 25 with 1080 h HERA under moderate foregrounds, and that non-detection at z ≳ 20 implies tdelay ≳ 30 Myr, treats tdelay as an adjustable parameter whose specific value is not derived from first principles within the simulation; this makes the quantitative threshold sensitive to the chosen foreground model and tdelay calibration rather than an internal prediction.
  3. [Discussion (comparison paragraph)] Comparison to prior work (discussion of SF rates): the statement that hydrodynamic calibration plus critical halo mass for H2 self-shielding plus stochastic mergers produces earlier SF is presented as a combined effect, yet the manuscript does not isolate the contribution of each ingredient via controlled runs within the same code, leaving the load-bearing attribution to the full combination unquantified.
minor comments (2)
  1. Notation: the power spectrum is denoted PS throughout; explicit definition (e.g., as Δ²(k) or the dimensionless form) in the text and figure captions would improve clarity.
  2. The abstract and methods refer to the simulation as 'self-consistent,' but the SF calibration is performed against external hydrodynamic volumes; a brief caveat on the remaining dependence on those training data would be useful.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed review of our manuscript. We address each major comment point by point below, indicating where revisions will be made to improve clarity and robustness.

read point-by-point responses

  1. Referee: [Methods (NN emulation section)] Methods (neural-network emulation and embedding procedure): the central claim of earlier/higher SF rates and the resulting Tb/PS evolution rests on the NN (trained on limited-volume AEOS/Renaissance runs) accurately reproducing small-scale SF efficiency, its density/feedback dependence, and the tdelay transition once placed inside the semi-numeric code that adds stochastic large-scale mergers; no explicit validation or domain-shift test for this transfer is described, which directly affects the reported Pop II/Pop III dominance and HERA thresholds.

    Authors: We agree that the manuscript would benefit from explicit validation of the NN transfer. In the revised version we will add a dedicated subsection presenting cross-checks of SF efficiencies between the embedded NN and the original hydro runs in overlapping density and redshift regimes, along with a brief discussion of domain-shift considerations. This directly supports the robustness of the Pop II/Pop III dominance results. revision: yes

  2. Referee: [Results (Tb and HERA section)] Results (Tb evolution and detectability statements): the assertion that the fiducial model is detectable at z ≲ 25 with 1080 h HERA under moderate foregrounds, and that non-detection at z ≳ 20 implies tdelay ≳ 30 Myr, treats tdelay as an adjustable parameter whose specific value is not derived from first principles within the simulation; this makes the quantitative threshold sensitive to the chosen foreground model and tdelay calibration rather than an internal prediction.

    Authors: The detectability statements are presented for the fiducial tdelay choice, which is calibrated to hydro results rather than derived from first principles. We will revise the text to explicitly state the conditional nature of these thresholds, emphasize dependence on the foreground model, and add a short sensitivity discussion showing how the HERA horizon changes with tdelay. This clarifies the model-dependent character without altering the core claims. revision: partial

  3. Referee: [Discussion (comparison paragraph)] Comparison to prior work (discussion of SF rates): the statement that hydrodynamic calibration plus critical halo mass for H2 self-shielding plus stochastic mergers produces earlier SF is presented as a combined effect, yet the manuscript does not isolate the contribution of each ingredient via controlled runs within the same code, leaving the load-bearing attribution to the full combination unquantified.

    Authors: We acknowledge that the manuscript does not isolate the individual contributions of the three ingredients. Performing a full set of controlled runs would require substantial additional resources. In revision we will expand the discussion to articulate the physical motivation for treating them as a combined, self-consistent package and will note the limitation on attribution. This provides a more balanced presentation while preserving the central narrative. revision: partial

Circularity Check

0 steps flagged

No significant circularity; external hydro calibration and input parameters keep derivation independent

full rationale

The paper calibrates small-scale SF efficiency via NN emulation to external AEOS and Renaissance hydrodynamic simulations, then embeds this in a semi-numeric code with stochastic merger histories and standard cosmological initial conditions to predict Tb and PS. tdelay is explicitly an input parameter varied across realizations rather than a fitted output. No self-definitional equations, fitted inputs renamed as predictions, or load-bearing self-citations appear in the abstract or described methods. The central claims on earlier/higher SF rates and HERA detectability thresholds follow from the model structure without reducing to the inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claims rest on the accuracy of the NN emulator trained on two specific hydro runs, the validity of the critical halo mass threshold that incorporates H2 self-shielding, and the assumption that stochastic merger histories dominate the PopIII-to-PopII transition timing. No new particles or forces are introduced.

free parameters (2)
  • tdelay
    Delay period separating PopIII and PopII star formation; its value directly controls the width of the absorption trough and the quoted 30 Myr detection threshold.
  • critical halo mass for SF
    Mass threshold below which star formation is suppressed, incorporating H2 self-shielding; calibrated but treated as a fixed input for the semi-numeric runs.
axioms (2)
  • standard math Standard Lambda-CDM cosmology and linear density fluctuations on large scales provide the correct initial conditions for halo assembly.
    Invoked when generating the 3D distribution of galaxies from dark matter halos.
  • domain assumption The neural network trained on AEOS and Renaissance runs generalizes to the density and feedback environments realized in the semi-numeric code.
    Central modeling choice stated in the abstract.

pith-pipeline@v0.9.1-grok · 5906 in / 1745 out tokens · 22795 ms · 2026-06-29T06:06:31.764009+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

2 extracted references · 1 canonical work pages

  1. [1]

    Ahn, K., & Shapiro, P. R. 2021, ApJ, 914, 44, doi:10.3847/1538-4357/ abf3bf Ahn, K., Xu, H., Norman, M. L., Alvarez, M. A., & Wise, J. H. 2015, ApJ, 802, 8, doi:10.1088/0004-637X/802/1/8 Allison, A. C., & Dalgarno, A. 1969, ApJ, 158, 423, doi:10.1086/150204 Armengaud, E., Palanque-Delabrouille, N., Yèche, C., Marsh, D. J. E., & Baur, J. 2017, MNRAS, 471, ...

    work page doi:10.3847/1538-4357/ 2021

  2. [2]

    by summing over the main contributors to collisional excitation at high redshifts: collisions between a hydrogen atom and an electron (eH), a proton (pH), and a second hydrogen atom (HH). The coefficient is thus 𝑥c =𝑥 HH c +𝑥 eH c +𝑥 pH c = 𝑇★ 𝐴10𝑇𝛾 [𝑛HI𝜅HH 10 (𝑇K)+𝑛 e𝜅eH 10 (𝑇K)+𝑛 p𝜅pH 10 (𝑇K)], (A1) where𝑛 𝑖 is the number density of species𝑖, and𝜅𝑖 10 i...

    2006