pith. sign in

arxiv: 2605.21012 · v1 · pith:VFKG2ELJnew · submitted 2026-05-20 · 🌌 astro-ph.CO

Reconstruction of Reionization Histories from 21 cm Power-Spectrum Evolution with Artificial Neural Networks

Yu-Le Wang , Hayato Shimabukuro This is my paper

Pith reviewed 2026-05-21 02:20 UTC · model grok-4.3

classification 🌌 astro-ph.CO
keywords reionization history21 cm power spectrumneural networkinverse mappingneutral fractionredshift evolutioncosmological observations
0
0 comments X

The pith

A neural network learns to reconstruct the neutral hydrogen history during cosmic reionization from the redshift evolution of the 21 cm power spectrum at fixed scales.

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

The authors test whether measurements of how the 21 cm power spectrum changes with redshift at one chosen spatial scale hold enough detail to recover the full timeline of when the universe's hydrogen went from neutral to ionized. They generate many example histories inside a simple three-parameter simulation setup and train a basic neural network to perform the inverse step from those power-spectrum curves back to the neutral-fraction curve. A sympathetic reader would care because this offers a potential shortcut for extracting reionization details from radio observations that future telescopes will deliver, without requiring complete three-dimensional imaging. The network recovers the central time of reionization more reliably than the length of the transition period. The results also hold when a basic model of telescope noise is added.

Core claim

The redshift evolution of the fixed-k dimensionless 21 cm power spectrum contains sufficient information to reconstruct reionization histories with artificial neural networks. Using a restricted three-parameter model family, a compact feed-forward network learns the inverse mapping from power-spectrum trajectories to the neutral-fraction history. Representative tests recover the midpoint redshift with MAE of 0.0046 and RMSE of 0.0100, while the duration yields MAE of 0.0302 and RMSE of 0.0378, indicating stronger information about timing than about transition width. The reconstruction remains stable under an idealized foreground-free noise model.

What carries the argument

A compact feed-forward neural network trained to perform the inverse mapping from the redshift evolution of the fixed-k 21 cm power spectrum to the neutral-fraction history x_HI(z).

If this is right

  • The midpoint redshift of reionization is recovered with high accuracy.
  • The duration of the reionization transition is recovered with moderate accuracy.
  • Fixed-scale power spectrum evolution provides stronger constraints on the timing than on the width of reionization.
  • The reconstruction is stable when an idealized thermal-plus-sample-variance noise model is included.

Where Pith is reading between the lines

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

  • If real observations match the patterns in the training set, this network approach could serve as a rapid estimator for reionization parameters from power-spectrum data alone.
  • Training on a wider range of reionization models might allow the network to handle more realistic histories beyond the three-parameter family.
  • Combining this method with data at multiple spatial scales could reduce errors in estimating the duration of reionization.

Load-bearing premise

Reionization histories are generated inside a narrow three-parameter family of a semi-numerical simulation model, and the redshift evolution of the power spectrum at one fixed scale contains enough information to recover the neutral-fraction history inside that family.

What would settle it

Generate power-spectrum evolution tracks from reionization models that use more than three parameters or different physical assumptions, apply the trained network, and verify whether the errors on the recovered midpoint and duration stay below 0.01 and 0.04 respectively.

Figures

Figures reproduced from arXiv: 2605.21012 by Hayato Shimabukuro, Yu-Le Wang.

Figure 1
Figure 1. Figure 1: Comparison between a more traditional reionization history and a late-reionization history motivated [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Redshift evolution of the dimensionless power spectrum [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative MSE learning curves during model optimization. Early stopping is applied using the [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Reconstructed reionization histories using fixed- [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distribution of correlation coefficients for reconstructions from fixed- [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representative reconstruction of the neutral-fraction history from noisy fixed- [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: MAE distribution for the reconstructed reionization history in the transition interval using noisy [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: RMSE distribution for the reconstructed reionization history in the transition interval using noisy [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distribution of the signed midpoint error, [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of the signed duration error, [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Scatter plot of midpoint recovery for the reconstructed reionization history. The lower panel shows [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Scatter plot of duration recovery for the reconstructed reionization history. The lower panel shows [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
read the original abstract

We investigate whether the redshift evolution of the fixed-$k$ dimensionless 21 cm power spectrum, $\Delta^2_{21}(k, z)$, contains sufficient information to reconstruct reionization histories $x_{\mathrm{HI}}(z)$ with artificial neural networks. Using semi-numerical realizations generated within a restricted three-parameter 21cmFAST model family, we train a compact feed-forward network to learn the inverse mapping from power-spectrum trajectories to the neutral-fraction history over $6 \le z \le 15$. For $k = 0.1$, $0.5$, and $1.0\ h\ \mathrm{Mpc}^{-1}$, representative tests on an independent test set show that the midpoint redshift $z_{50}$ is recovered more accurately than the duration $\Delta z = z_{75} - z_{25}$: $z_{50}$ is reconstructed with MAE = 0.0046 and RMSE = 0.0100, whereas $\Delta z$ yields MAE = 0.0302 and RMSE = 0.0378. This result indicates that fixed-$k$ power-spectrum evolution carries stronger information about the timing of reionization than about the detailed width of the transition within the adopted prior. We further test an idealized foreground-free SKA1-Low-like thermal-plus-sample-variance noise model and find that the reconstruction remains stable in the favorable signal-to-noise regime considered here. These results demonstrate that neural networks can serve as prior-dependent inverse mapping for reconstructing reionization histories from 21 cm power-spectrum evolution.

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

0 major / 3 minor

Summary. The paper claims that the redshift evolution of the fixed-k dimensionless 21 cm power spectrum Δ²₂₁(k, z) contains sufficient information to reconstruct reionization histories x_HI(z) with artificial neural networks. Using semi-numerical realizations from a restricted three-parameter 21cmFAST model family, a compact feed-forward network is trained to map power-spectrum trajectories to the neutral-fraction history over 6 ≤ z ≤ 15. On an independent test set, the midpoint redshift z₅₀ is recovered with MAE = 0.0046 and RMSE = 0.0100, while the duration Δz = z₇₅ - z₂₅ yields MAE = 0.0302 and RMSE = 0.0378. The reconstruction remains stable under an idealized foreground-free SKA1-Low-like noise model. The work demonstrates that neural networks can serve as prior-dependent inverse mappings for this task.

Significance. If the reported mapping holds, the paper provides a clear quantitative demonstration that fixed-k 21 cm power-spectrum evolution carries usable information about reionization timing (z₅₀) within the adopted three-parameter prior, with weaker but still measurable constraints on transition duration (Δz). The explicit scoping to a prior-dependent inverse, the independent test-set metrics, and the idealized noise stability check are strengths that allow direct assessment of information content. This contributes a concrete example of machine-learning-assisted inversion in 21 cm cosmology, useful for guiding analysis of future SKA data under similar model assumptions.

minor comments (3)
  1. [Abstract and §3] The abstract and methods should explicitly state the total number of simulations, the parameter ranges sampled for training versus test sets, and the precise network architecture (number of layers, neurons, activation functions) to allow reproducibility and assessment of overfitting risk.
  2. [Results] Clarify whether the reported MAE/RMSE values are computed on the full x_HI(z) trajectory or only on the summary parameters z₅₀ and Δz; if the former, include an example reconstruction plot with residuals.
  3. [Noise test subsection] The idealized noise model is described as 'foreground-free'; add a brief statement on how residual foregrounds or other systematics might affect the claimed stability, even if only qualitatively.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive summary and significance assessment, which accurately reflect the scope and results of our manuscript. We appreciate the recommendation for minor revision. No specific major comments were raised in the provided report, so we have no point-by-point responses or revisions to propose at this stage. We remain available to address any additional questions or suggestions from the editor.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper scopes its central claim explicitly to a prior-dependent inverse mapping learned by a neural network inside a restricted three-parameter 21cmFAST model family. Training data and independent test data are both drawn from the same forward model, and the reported MAE/RMSE values simply quantify how much information the fixed-k power-spectrum trajectories carry for x_HI(z) within that prior; no load-bearing step reduces by construction to a self-definition, a fitted parameter renamed as a prediction, or a self-citation chain. The derivation is therefore self-contained as a standard supervised inversion exercise on simulated data.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the 21cmFAST three-parameter model family being an adequate representation of reionization physics for training purposes and on the assumption that fixed-k power spectrum trajectories encode the necessary information about x_HI(z) within that family.

free parameters (1)
  • three parameters of the 21cmFAST model family
    Reionization histories are drawn from a restricted three-parameter space whose specific parameters and ranges are not enumerated in the abstract but define the prior over which the network is trained.
axioms (1)
  • domain assumption Semi-numerical realizations generated by 21cmFAST accurately capture the redshift evolution of the 21 cm power spectrum during reionization.
    All training and test data are produced by this code; any mismatch between the code and real physics would propagate directly into the learned inverse mapping.

pith-pipeline@v0.9.0 · 5819 in / 1664 out tokens · 69512 ms · 2026-05-21T02:20:38.334520+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages

  1. [1]

    2001, Phys

    Barkana, R., & Loeb, A. 2001, Phys. Rep., 349, 125

    work page 2001

  2. [2]

    2019, arXiv e-prints, arXiv:1912.12699

    Braun, R., Bonaldi, A., Bourke, T., Keane, E., & Wagg, J. 2019, arXiv e-prints, arXiv:1912.12699

    work page arXiv 2019

  3. [3]

    2004, Astrophysical Journal, 602, 1

    Chen, X., & Miralda-Escud´e, J. 2004, Astrophysical Journal, 602, 1

    work page 2004

  4. [4]

    2008, Astrophysical Journal, 684, 18

    Chen, X., & Miralda-Escud´e, J. 2008, Astrophysical Journal, 684, 18

    work page 2008

  5. [5]

    2018, Phys

    Dayal, P., & Ferrara, A. 2018, Phys. Rep., 780, 1

    work page 2018

  6. [6]

    R., Parsons, A

    DeBoer, D. R., Parsons, A. R., Aguirre, J. E., et al. 2017, PASP, 129, 045001

    work page 2017

  7. [7]

    E., Hall, P

    Dewdney, P. E., Hall, P. J., Schilizzi, R. T., & Lazio, T. J. L. W. 2009, Proceedings of the IEEE, 97, 1482

    work page 2009

  8. [8]

    L., & Keating, B

    Fan, X., Carilli, C. L., & Keating, B. 2006, ARA&A, 44, 415

    work page 2006

  9. [9]

    Field, G. B. 1958, Proceedings of the IRE, 46, 240

    work page 1958

  10. [10]

    Field, G. B. 1959, Astrophysical Journal, 129, 536

    work page 1959

  11. [11]

    R., & Briggs, F

    Furlanetto, S. R., & Briggs, F. H. 2004, New Astronomy Reviews, 48, 1039

    work page 2004

  12. [12]

    R., Oh, S

    Furlanetto, S. R., Oh, S. P., & Briggs, F. H. 2006, Phys. Rep., 433, 181

    work page 2006

  13. [13]

    R., Zaldarriaga, M., & Hernquist, L

    Furlanetto, S. R., Zaldarriaga, M., & Hernquist, L. 2004, ApJ, 613, 1

    work page 2004

  14. [14]

    2015, Monthly Notices of the Royal Astronomical Society, 449, 4246

    Greig, B., & Mesinger, A. 2015, Monthly Notices of the Royal Astronomical Society, 449, 4246

    work page 2015

  15. [15]

    Haslam, C. G. T., Salter, C. J., Stoffel, H., & Wilson, W. E. 1982, A&AS, 47, 1

    work page 1982

  16. [16]

    Hirata, C. M. 2006, Monthly Notices of the Royal Astronomical Society, 367, 259

    work page 2006

  17. [17]

    1989, Neural Networks, 2, 359

    Hornik, K., Stinchcombe, M., & White, H. 1989, Neural Networks, 2, 359

    work page 1989

  18. [18]

    2025, ApJS, 278, 33

    Kageura, Y ., Ouchi, M., Nakane, M., et al. 2025, ApJS, 278, 33

    work page 2025

  19. [19]

    2015, in Advancing Astrophysics with the Square Kilometre Array (AASKA14), 1 La Plante, P., & Ntampaka, M

    Koopmans, L., Pritchard, J., Mellema, G., et al. 2015, in Advancing Astrophysics with the Square Kilometre Array (AASKA14), 1 La Plante, P., & Ntampaka, M. 2019, The Astrophysical Journal, 880, 110

    work page 2015

  20. [20]

    R., Peng Oh, S., et al

    Lidz, A., Furlanetto, S. R., Peng Oh, S., et al. 2011, The Astrophysical Journal, 741, 70

    work page 2011

  21. [21]

    Madau, P., Haardt, F., & Rees, M. J. 1999, ApJ, 514, 648

    work page 1999

  22. [22]

    Madau, P., Meiksin, A., & Rees, M. J. 1997, ApJ, 475, 429

    work page 1997

  23. [23]

    Mao, Y ., D’Aloisio, A., Zhang, J., & Shapiro, P. R. 2013, Phys. Rev. D, 88, 081303 22 Wang & Shimabukuro

    work page 2013

  24. [24]

    2008, Phys

    Mao, Y ., Tegmark, M., McQuinn, M., Zaldarriaga, M., & Zahn, O. 2008, Phys. Rev. D, 78, 023529

    work page 2008

  25. [25]

    McQuinn, M., Zahn, O., Zaldarriaga, M., Hernquist, L., & Furlanetto, S. R. 2006, ApJ, 653, 815

    work page 2006

  26. [26]

    2019, The Cosmic 21-cm Revolution; Charting the first billion years of our universe

    Mesinger, A. 2019, The Cosmic 21-cm Revolution; Charting the first billion years of our universe

    work page 2019

  27. [27]

    2011, MNRAS, 411, 955

    Mesinger, A., Furlanetto, S., & Cen, R. 2011, MNRAS, 411, 955

    work page 2011

  28. [28]

    F., & Wyithe, J

    Morales, M. F., & Wyithe, J. S. B. 2010, Annual Review of Astronomy and Astrophysics, 48, 127

    work page 2010

  29. [29]

    2020, The Journal of Open Source Software, 5, 2582

    Murray, S., Greig, B., Mesinger, A., et al. 2020, The Journal of Open Source Software, 5, 2582

    work page 2020

  30. [30]

    Nair, V ., & Hinton, G. E. 2010, Omnipress

    work page 2010

  31. [31]

    2019, MNRAS, 484, 933

    Park, J., Mesinger, A., Greig, B., & Gillet, N. 2019, MNRAS, 484, 933

    work page 2019

  32. [32]

    R., Pober, J

    Parsons, A. R., Pober, J. C., Aguirre, J. E., et al. 2012, ApJ, 756, 165 Planck Collaboration, Aghanim, N., Akrami, Y ., et al. 2020, A&A, 641, A6

    work page 2012

  33. [33]

    R., & Loeb, A

    Pritchard, J. R., & Loeb, A. 2012, Reports on Progress in Physics, 75, 086901

    work page 2012

  34. [34]

    Robertson, B. E. 2022, ARA&A, 60, 121

    work page 2022

  35. [35]

    E., Hinton, G

    Rumelhart, D. E., Hinton, G. E., & Williams, R. J. 1986, Nature, 323, 533

    work page 1986

  36. [36]

    Scott, D., & Rees, M. J. 1990, MNRAS, 247, 510

    work page 1990

  37. [37]

    2025, Research in Astronomy and Astrophysics, 25, 085017

    Shimabukuro, H. 2025, Research in Astronomy and Astrophysics, 25, 085017

    work page 2025

  38. [38]

    2023, PASJ, 75, S1

    Shimabukuro, H., Hasegawa, K., Kuchinomachi, A., Yajima, H., & Yoshiura, S. 2023, PASJ, 75, S1

    work page 2023

  39. [39]

    2022, Research in Astronomy and Astrophysics, 22, 035027

    Shimabukuro, H., Mao, Y ., & Tan, J. 2022, Research in Astronomy and Astrophysics, 22, 035027

    work page 2022

  40. [40]

    2017, MNRAS, 468, 3869

    Shimabukuro, H., & Semelin, B. 2017, MNRAS, 468, 3869

    work page 2017

  41. [41]

    2024, MNRAS, 527, 9977

    Sikder, S., Barkana, R., Reis, I., & Fialkov, A. 2024, MNRAS, 527, 9977

    work page 2024

  42. [42]

    Umeda, M

    Umeda, H., Ouchi, M., Kageura, Y ., et al. 2025, arXiv e-prints, arXiv:2504.04683

    work page arXiv 2025

  43. [43]

    2025, The Astrophysical Journal Supplement Series, 277, 37

    Umeda, H., Ouchi, M., Kikuta, S., et al. 2025, The Astrophysical Journal Supplement Series, 277, 37

    work page 2025

  44. [44]

    Wouthuysen, S. A. 1952, Astronomical Journal, 57, 31

    work page 1952

  45. [45]

    R., & Hernquist, L

    Zaldarriaga, M., Furlanetto, S. R., & Hernquist, L. 2004, Astrophysical Journal, 608, 622

    work page 2004