pith. sign in

arxiv: 2605.29016 · v1 · pith:MVOQP32Znew · submitted 2026-05-27 · 🌌 astro-ph.IM · astro-ph.CO· cs.LG

Three-dimensional Conditional Diffusion Models for Cosmological 21 cm Lightcone Emulation

Bin Xia , John H. Wise This is my paper

Pith reviewed 2026-06-29 09:24 UTC · model grok-4.3

classification 🌌 astro-ph.IM astro-ph.COcs.LG
keywords 21 cm cosmologydiffusion modelslightcone emulationthree-dimensional emulationcosmological simulationspreprocessing methodsglobal 21 cm signal
0
0 comments X

The pith

Yeo-Johnson preprocessing plus moderate amplitude compression yields the most stable training and lowest normalized errors for three-dimensional 21 cm lightcone diffusion models.

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

The paper tests conditional diffusion models on 64 by 64 by up to 1024 voxel 21 cm lightcones generated from 21cmFAST. Memory constraints force tiny batches and the voxel values are strongly skewed, so the authors run controlled sweeps over preprocessing methods, dynamic-range compression, network depth, and training length. They train on 25,600 lightcones and compare each model against 800 independent realizations at fixed cosmology points using brightness-temperature slices, the global signal, the power spectrum, and reduced scattering coefficients. Across all tests, the choice of preprocessing dominates stability and fidelity; the combination of Yeo-Johnson transformation with moderate amplitude compression consistently ranks highest on the standard-deviation-normalized mean absolute error of the global signal and shows compatible behavior on the other diagnostics. Even the best samples remain visually plausible yet carry measurable biases in two-point and higher-order statistics, positioning the work as an initial simulation-level baseline.

Core claim

Preprocessing is the dominant factor governing stable training and physical fidelity in three-dimensional conditional diffusion models for 21 cm lightcones; among the configurations tested, Yeo-Johnson preprocessing combined with moderate amplitude compression produces the most consistently favorable trade-off according to standard-deviation-normalized mean absolute error of the global signal and qualitatively compatible results on slices, power spectra, and scattering coefficients.

What carries the argument

Conditional diffusion model applied to three-dimensional 21 cm lightcones, with preprocessing (Yeo-Johnson transformation) and amplitude compression as the primary controls on training stability and output fidelity.

If this is right

  • Future 3D emulation studies can adopt Yeo-Johnson preprocessing with moderate compression as a starting point rather than re-exploring the full preprocessing space.
  • Any claim of physical fidelity for these models must still be checked against two-point and higher-order statistics even when slices look realistic.
  • The same architecture can serve as a simulation-level baseline once more realistic observational effects are added.
  • Memory-efficient 3D training remains the main engineering barrier that future work must solve to reach larger volumes.

Where Pith is reading between the lines

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

  • If the diagnostics turn out to be incomplete, the ranking of preprocessing choices could shift once full lightcone power spectra or cross-correlations with other tracers are examined.
  • The approach could be extended to joint emulation of multiple reionization fields by conditioning on additional astrophysical parameters without changing the preprocessing recipe.
  • The residual statistical biases suggest that adding a physics-informed loss term on the power spectrum during training might close the remaining gap.

Load-bearing premise

The chosen set of image and summary-statistic diagnostics is enough to establish that generated lightcones are physically faithful for cosmological use.

What would settle it

A new configuration or loss term that produces lower MAE_std on the global signal while simultaneously reducing the reported biases in the power spectrum and scattering coefficients at fixed cosmology points.

Figures

Figures reproduced from arXiv: 2605.29016 by Bin Xia, John H. Wise.

Figure 1
Figure 1. Figure 1: Summary of scalar MAE metrics across the hyperparameter sweep. Markers show MAErel, MAEstd, and MAEσ for each indexed run listed in [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of selected 3D runs against 21cmFAST (dotted line) using the global signal and its resid￾ual, ∆⟨Tb⟩. The upper panel shows the median ⟨Tb⟩ evolu￾tion for representative runs spanning different preprocessing choices, amplitude scalings, and residual-block depths, and the lower panel shows the corresponding deviation from the 21cmFAST median. the 2D transform tests contained in the blue shaded re￾… view at source ↗
Figure 3
Figure 3. Figure 3: Brightness-temperature slice comparison for the representative 3D run (Yeo–Johnson preprocessing, A = 0.1, one residual block per level in the encoder, 240 epochs). For each conditioning point, the upper two strips show independent 21cmFAST realizations and the lower two strips show diffusion-generated realizations [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Global-signal (left) and power-spectrum (right) diagnostics for the representative 3D run shown in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Reduced scattering-coefficient comparison for the representative 3D run shown in Figures 3 and 4. The top panel compares the second-order coefficients S2 across multiscale index pairs (j1, j2) and relative orientation classes, while the lower panels show the corresponding residual channels. the image-space residuals: once sharp boundaries are smoothed and low Tb interiors are broadened, the largest impact … view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of preprocessing choices using lightcone voxel value distributions. The two panels show voxel value probability density functions in transformed space (left) and raw brightness-temperature space (right) for 21cmFAST targets and diffusion-generated samples. narrows the transformed dynamic range that the diffu￾sion model must learn. The same figure also clarifies why transformed-space improvement … view at source ↗
read the original abstract

We investigate conditional diffusion modeling for three-dimensional 21 cm lightcone emulation, focusing on cubes with a sky-plane size of $64\times64$ and a line-of-sight depth up to 1024 cells. Relative to earlier 2D studies, the 3D setting is substantially harder because memory limits enforce very small micro-batches while the underlying voxel distribution is highly skewed and long tailed. We perform controlled comparisons across preprocessing choices, dynamic-range compression settings, architecture depth, and training duration using $25{,}600$ training lightcones and validation ensembles at fixed parameter points. For validation, each reference parameter point contains 800 21cmFAST realizations with independent initial conditions, and we use 800 samples per model and per reference set for the reported ensemble comparisons. We evaluate generated lightcones with complementary diagnostics in both image and summary-statistic spaces: brightness-temperature slices, the global signal, the power spectrum, and reduced scattering coefficients. Across the tested configurations, preprocessing is the dominant factor governing stable training and the resulting physical fidelity. Among the configurations explored here, Yeo-Johnson preprocessing combined with moderate amplitude compression gives the most consistently favorable trade-off, with the strongest quantitative support coming from rankings based on the standard-deviation-normalized mean absolute error ($\mathrm{MAE}_{\rm std}$) of the global signal and qualitatively compatible behavior in the complementary diagnostics. At the same time, visually plausible 3D samples still retain measurable biases in two-point and higher-order statistics. We therefore view the present work as a simulation-level baseline for three-dimensional 21 cm emulation and for future studies that incorporate more realistic observational effects.

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

1 major / 1 minor

Summary. The manuscript investigates conditional diffusion models for emulating three-dimensional 21 cm lightcones (64x64 sky plane, up to 1024 cells along the line of sight) from 21cmFAST simulations. It conducts controlled comparisons of preprocessing choices (including Yeo-Johnson), amplitude compression settings, architecture depth, and training duration using 25,600 training lightcones and validation ensembles of 800 independent realizations per reference parameter point. Generated samples are evaluated via brightness-temperature slices, global signal, power spectrum, and reduced scattering coefficients. The central claim is that Yeo-Johnson preprocessing combined with moderate amplitude compression provides the most favorable trade-off, with strongest quantitative support from rankings on standard-deviation-normalized MAE of the global signal and qualitatively compatible behavior in the other diagnostics; the work positions itself as a simulation-level baseline, noting that even the best samples retain measurable biases in two-point and higher-order statistics.

Significance. If the chosen diagnostics are jointly sufficient to establish physical fidelity for cosmological applications, the work provides a useful baseline for 3D 21 cm emulation under memory constraints and skewed voxel distributions. The controlled experimental design with fixed validation ensembles and multiple complementary diagnostics is a strength, as is the explicit acknowledgment of remaining biases in higher-order statistics.

major comments (1)
  1. [Abstract] Abstract: The central ranking of configurations (Yeo-Johnson + moderate compression as best) rests on the assumption that the four diagnostics—brightness-temperature slices, global signal (via MAE_std), power spectrum, and reduced scattering coefficients—are jointly sufficient to detect biases that would affect downstream cosmological use cases such as parameter inference. The abstract itself states that 'visually plausible 3D samples still retain measurable biases in two-point and higher-order statistics,' yet provides no quantitative test (e.g., bispectrum comparison or propagation of residuals to parameter constraints) demonstrating that the selected diagnostics are sensitive to the biases that matter for 21 cm cosmology.
minor comments (1)
  1. [Abstract] The abstract reports 25,600 training lightcones and 800 validation realizations per reference set but does not specify how many distinct reference parameter points were used or how the 800 samples per model were generated (e.g., same random seeds or independent).

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on the sufficiency of our diagnostics. We address the major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central ranking of configurations (Yeo-Johnson + moderate compression as best) rests on the assumption that the four diagnostics—brightness-temperature slices, global signal (via MAE_std), power spectrum, and reduced scattering coefficients—are jointly sufficient to detect biases that would affect downstream cosmological use cases such as parameter inference. The abstract itself states that 'visually plausible 3D samples still retain measurable biases in two-point and higher-order statistics,' yet provides no quantitative test (e.g., bispectrum comparison or propagation of residuals to parameter constraints) demonstrating that the selected diagnostics are sensitive to the biases that matter for 21 cm cosmology.

    Authors: We agree that the manuscript does not include explicit quantitative tests such as bispectrum comparisons or propagation of emulation residuals through parameter inference pipelines to demonstrate that the chosen diagnostics fully capture all biases relevant to downstream 21 cm cosmological applications. The power spectrum directly addresses two-point statistics while reduced scattering coefficients target higher-order information, but these are not exhaustive. The work is explicitly framed as a simulation-level baseline under memory and distribution constraints, with the abstract already noting remaining measurable biases. We have revised the abstract to clarify that the configuration ranking is based on the reported diagnostics and that further targeted validation (e.g., for specific inference tasks) would be required for full cosmological use. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical ranking on independent held-out validation ensembles

full rationale

The paper conducts controlled empirical comparisons of preprocessing, compression, and architecture choices for 3D diffusion models. Training uses 25,600 lightcones; validation employs separate 800-realization ensembles with independent initial conditions. Reported metrics (MAE_std on global signal, plus qualitative checks on slices/power spectrum/scattering coefficients) are computed post-training on these held-out sets and are not inputs to model fitting or parameter selection. No self-definitional equations, no fitted quantities renamed as predictions, no load-bearing self-citations, and no uniqueness theorems invoked. The derivation chain consists of standard training + evaluation steps that remain independent of the final ranking claim.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the reliability of 21cmFAST as ground truth and on the sufficiency of the chosen summary statistics to measure fidelity. No new physical entities are postulated. Preprocessing choices are explored rather than derived.

free parameters (2)
  • Yeo-Johnson transformation parameter
    Chosen among preprocessing options to achieve best MAE_std ranking; not derived from first principles.
  • amplitude compression setting
    Moderate compression level selected after exploring dynamic-range options; affects training stability on skewed data.
axioms (1)
  • domain assumption 21cmFAST realizations with independent initial conditions accurately represent the target distribution of 21 cm lightcones
    The paper uses these realizations both for training (25,600 lightcones) and for validation ensembles (800 per reference point).

pith-pipeline@v0.9.1-grok · 5832 in / 1621 out tokens · 36936 ms · 2026-06-29T09:24:31.897372+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

34 extracted references · 26 canonical work pages · 1 internal anchor

  1. [1]

    E., Alexander, P., et al

    Abdurashidova, Z., Aguirre, J. E., Alexander, P., et al. 2022a, ApJ, 925, 221, doi: 10.3847/1538-4357/ac1c78

    work page doi:10.3847/1538-4357/ac1c78

  2. [2]

    E., Alexander, P., et al

    Abdurashidova, Z., Aguirre, J. E., Alexander, P., et al. 2022b, ApJ, 924, 51, doi: 10.3847/1538-4357/ac2ffc

    work page doi:10.3847/1538-4357/ac2ffc

  3. [3]

    A., Aghanim, N., Arnaud, M., et al

    Ade, P. A., Aghanim, N., Arnaud, M., et al. 2016, Astronomy & Astrophysics, 594, A13

    2016

  4. [4]

    2020, Journal of Machine Learning Research, 21, 1

    Andreux, M., Angles, T., Exarchakis, G., et al. 2020, Journal of Machine Learning Research, 21, 1

    2020

  5. [5]

    2022, arXiv e-prints, arXiv:2211.05000, doi: 10.48550/arXiv.2211.05000

    Andrianomena, S., Villaescusa-Navarro, F., & Hassan, S. 2022, arXiv e-prints, arXiv:2211.05000, doi: 10.48550/arXiv.2211.05000

    work page doi:10.48550/arxiv.2211.05000 2022

  6. [6]

    2001, PhR, 349, 125, doi: 10.1016/S0370-1573(01)00019-9

    Barkana, R., & Loeb, A. 2001, PhR, 349, 125, doi: 10.1016/S0370-1573(01)00019-9

    work page doi:10.1016/s0370-1573(01)00019-9 2001

  7. [7]

    2020, Monthly Notices of the Royal Astronomical Society, 499, 5902, doi: 10.1093/mnras/staa3165

    Cheng, S., Ting, Y.-S., Ménard, B., & Bruna, J. 2020, Monthly Notices of the Royal Astronomical Society, 499, 5902, doi: 10.1093/mnras/staa3165

    work page doi:10.1093/mnras/staa3165 2020

  8. [8]

    R., Parsons, A

    DeBoer, D. R., Parsons, A. R., Aguirre, J. E., et al. 2017, PASP, 129, 045001, doi: 10.1088/1538-3873/129/974/045001

    work page doi:10.1088/1538-3873/129/974/045001 2017

  9. [9]

    2021, Advances in neural information processing systems, 34, 8780

    Dhariwal, P., & Nichol, A. 2021, Advances in neural information processing systems, 34, 8780

    2021

  10. [10]

    2023, arXiv e-prints, arXiv:2307.04976, doi: 10.48550/arXiv.2307.04976

    Diao, K., & Mao, Y. 2023, arXiv e-prints, arXiv:2307.04976, doi: 10.48550/arXiv.2307.04976

    work page doi:10.48550/arxiv.2307.04976 2023

  11. [11]

    2025, ApJ, 992, 173, doi: 10.3847/1538-4357/ae0325

    Diao, K., & Mao, Y. 2025, ApJ, 992, 173, doi: 10.3847/1538-4357/ae0325

    work page doi:10.3847/1538-4357/ae0325 2025

  12. [12]

    R., Oh, S

    Furlanetto, S. R., Oh, S. P., & Briggs, F. H. 2006, Physics reports, 433, 181

    2006

  13. [13]

    2021, Monthly Notices of the Royal Astronomical Society, 504, 4716, doi: 10.1093/mnras/stab1158

    Gagnon-Hartman, S., Cui, Y., Liu, A., & Ravanbakhsh, S. 2021, Monthly Notices of the Royal Astronomical Society, 504, 4716, doi: 10.1093/mnras/stab1158

    work page doi:10.1093/mnras/stab1158 2021

  14. [14]

    Generative Adversarial Networks

    Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., et al. 2014, arXiv e-prints, arXiv:1406.2661, doi: 10.48550/arXiv.1406.2661

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1406.2661 2014

  15. [15]

    2017, Proceedings of the International Astronomical Union, 12, 1821, doi: 10.1017/S1743921317011103

    Greig, B., & Mesinger, A. 2017, Proceedings of the International Astronomical Union, 12, 1821, doi: 10.1017/S1743921317011103

    work page doi:10.1017/s1743921317011103 2017

  16. [16]

    2018, MNRAS, 477, 3217, doi: 10.1093/mnras/sty796

    Greig, B., & Mesinger, A. 2018, MNRAS, 477, 3217, doi: 10.1093/mnras/sty796

    work page doi:10.1093/mnras/sty796 2018

  17. [17]

    Greig, B., Ting, Y.-S., & Kaurov, A. A. 2023, MNRAS, 519, 5288, doi: 10.1093/mnras/stac3822

    work page doi:10.1093/mnras/stac3822 2023

  18. [18]

    Array programming with NumPy,

    Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357, doi: 10.1038/s41586-020-2649-2

    work page doi:10.1038/s41586-020-2649-2 2020

  19. [19]

    2020, Advances in neural information processing systems, 33, 6840

    Ho, J., Jain, A., & Abbeel, P. 2020, Advances in neural information processing systems, 33, 6840

    2020

  20. [20]

    2026, A&A, 706, A15, doi: 10.1051/0004-6361/202555896

    Hothi, I., Allys, E., Semelin, B., & Meriot, R. 2026, A&A, 706, A15, doi: 10.1051/0004-6361/202555896

    work page doi:10.1051/0004-6361/202555896 2026

  21. [21]

    Hunter, J. D. 2007, Computing in Science & Engineering, 9, 90, doi: 10.1109/MCSE.2007.55

    work page doi:10.1109/mcse.2007.55 2007

  22. [22]

    2015, SKA

    Koopmans, L., Pritchard, J., Mellema, G., et al. 2015, SKA

    2015

  23. [23]

    List, F., & Lewis, G. F. 2020, Monthly Notices of the Royal Astronomical Society, 493, 5913, doi: 10.1093/mnras/staa523

    work page doi:10.1093/mnras/staa523 2020

  24. [24]

    Gruppuso, F

    Mesinger, A., Furlanetto, S., & Cen, R. 2011, MNRAS, 411, 955, doi: 10.1111/j.1365-2966.2010.17731.x

    work page doi:10.1111/j.1365-2966.2010.17731.x 2011

  25. [25]

    F., & Wyithe, J

    Morales, M. F., & Wyithe, J. S. B. 2010, ARA&A, 48, 127, doi: 10.1146/annurev-astro-081309-130936

    work page doi:10.1146/annurev-astro-081309-130936 2010

  26. [26]

    2019, in Advances in Neural Information Processing Systems 32 Prelogovi´ c, D., & Mesinger, A

    Paszke, A., Gross, S., Massa, F., et al. 2019, in Advances in Neural Information Processing Systems 32 Prelogovi´ c, D., & Mesinger, A. 2024, A&A, 688, A199, doi: 10.1051/0004-6361/202449309 Prelogovi´ c, D., Mesinger, A., Murray, S., Fiameni, G., &

    work page doi:10.1051/0004-6361/202449309 2019

  27. [27]

    2021, Monthly Notices of the Royal Astronomical Society, 509, 3852, doi: 10.1093/mnras/stab3215

    Gillet, N. 2021, Monthly Notices of the Royal Astronomical Society, 509, 3852, doi: 10.1093/mnras/stab3215

    work page doi:10.1093/mnras/stab3215 2021

  28. [28]

    R., & Loeb, A

    Pritchard, J. R., & Loeb, A. 2012, Reports on Progress in Physics, 75, 086901 15

    2012

  29. [29]

    2024, PhRvD, 109, 123536, doi: 10.1103/PhysRevD.109.123536

    Lee, K. 2024, PhRvD, 109, 123536, doi: 10.1103/PhysRevD.109.123536

    work page doi:10.1103/physrevd.109.123536 2024

  30. [30]

    J., Goeke, R., Bowman, J

    Tingay, S. J., Goeke, R., Bowman, J. D., et al. 2013, PASA, 30, e007, doi: 10.1017/pasa.2012.007 van Haarlem, M. P., Wise, M. W., Gunst, A. W., et al. 2013, A&A, 556, A2, doi: 10.1051/0004-6361/201220873

    work page doi:10.1017/pasa.2012.007 2013

  31. [31]

    2021, ApJ, 907, 44, doi: 10.3847/1538-4357/abd245

    Villanueva-Domingo, P., & Villaescusa-Navarro, F. 2021, ApJ, 907, 44, doi: 10.3847/1538-4357/abd245

    work page doi:10.3847/1538-4357/abd245 2021

  32. [32]

    Zhao, X., Mao, Y., Cheng, C., & Wandelt, B. D. 2022, The Astrophysical Journal, 926, 151, doi: 10.3847/1538-4357/ac457d

    work page doi:10.3847/1538-4357/ac457d 2022

  33. [33]

    Zhao, X., Mao, Y., Zuo, S., & Wandelt, B. D. 2024, ApJ, 973, 41, doi: 10.3847/1538-4357/ad5ff0

    work page doi:10.3847/1538-4357/ad5ff0 2024

  34. [34]

    2023, Monthly Notices of the Royal Astronomical Society, 526, 1699

    Zhao, X., Ting, Y.-S., Diao, K., & Mao, Y. 2023, Monthly Notices of the Royal Astronomical Society, 526, 1699

    2023