Pith. sign in

REVIEW 5 major objections 4 minor 53 references

Cosmological super-resolution of the 21-cm signal

T0 review · 5 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Score-based diffusion models can super-resolve 21-cm brightness-temperature simulations from 12 cMpc to 3 cMpc voxels with about 0.57 mK pixel error at z=10, and one training simulation is enough.

desk verdict Solid proof-of-concept for 3D 21-cm super-resolution with diffusion models, but the SKA utilization claim outruns the noiseless-degradation evaluation. read the letter →

arxiv 2502.00852 v1 pith:ZDSC6SPA submitted 2025-02-02 astro-ph.CO astro-ph.IM

Simon Pochinda , Jiten Dhandha , Anastasia Fialkov , Eloy de Lera Acedo This is my paper · ORCID
classification astro-ph.COastro-ph.IM
keywords 21-cmcosmologysuper-resolutiondiffusionmodelsscore-basedgenerativecosmicreionizationSKA1-Lowpowerspectrumbrightnesstemperature
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper shows that score-based diffusion models can turn coarse, low-resolution boxes of the 21-cm hydrogen signal into high-resolution maps that preserve both the pixel statistics and the power spectrum of the original simulation. The best model reaches pixelwise RMSE around 0.57 mK and dimensionless power-spectrum residuals of $10^{-2}$ to $10^{-1}\ \mathrm{mK^2}$ for $128^3$, $256^3$, and $512^3$ volumes at redshift 10, with a single training simulation sufficient regardless of initial conditions. The point is practical: gigaparsec-scale high-resolution simulations are computationally expensive, but the trained network can deliver the spatial scales that SKA1-Low will observe in hours rather than tens of thousands of CPU hours. If this transfers to actual telescope maps, it would remove one of the bottlenecks to using SKA1-Low data to constrain the astrophysics of the early Universe.

What carries the argument

The load-bearing object is a conditional score-based diffusion model of the Variance Preserving Stochastic Differential Equation type, with a 3D-adapted DDPM++ denoising network. During training, a high-resolution 21-cm box is corrupted by Gaussian noise at levels drawn from the VP SDE schedule, and the network learns the score, the gradient of the log probability of the clean data, conditioned on a low-resolution box obtained by trilinear downsampling to 12 cMpc voxels and upsampling back, together with the simulation initial conditions: matter overdensity and baryon-dark matter relative velocity fields. At inference, Euler-Maruyama reverse diffusion turns random noise into a 21-cm map guided by those conditions. The conditioning inputs carry the astrophysical and cosmological information that cannot be inferred from the initial conditions alone, and the learned score propagates small-scale structure that the low-resolution box lacks.

What would settle it

Take a high-resolution 21-cm simulation, degrade it through a realistic SKA1-Low instrument model including primary beam, thermal noise, and foreground filtering to produce the conditional input, then measure pixelwise RMSE and power-spectrum residuals against the original $3\ \mathrm{cMpc}$ box; if the errors rise above the sub-mK range or above the expected noise level, the central transferability claim fails.

Watch

Extended reading notes

Core claim

The authors establish that a class of conditional score-based diffusion models, specifically Variance Preserving Stochastic Differential Equation diffusion with a 3D-adapted DDPM++-style denoiser, can super-resolve 21-cm differential brightness temperature simulations by a factor of 4 in voxel size, from 12 cMpc to 3 cMpc, while remaining faithful to both the one-point voxel distribution and the dimensionless power spectrum. At $z=10$, the best model achieves an RMSE of 0.57 mK on $512^3$ boxes and similar errors on $256^3$ and $128^3$ boxes, with power-spectrum residuals of $10^{-2}$ to $10^{-1}\ \mathrm{mK^2}$, below the expected SKA1-Low noise level. They also demonstrate that performance is essentially independent of training dataset size, which they attribute to the large simulated volumes being immune to cosmic variance, so the relevant statistics are fixed by astrophysics alone. The paper frames the result as a proof of concept that generative super-resolution can make SKA1-Low's gigaparsec spatial scales usable for astrophysical inference.

Load-bearing premise

The low-resolution conditional input is always a clean trilinear downsampling of the same high-resolution simulation, never a map degraded by SKA1-Low's beam, noise, and foregrounds.

Editorial extensions

If this is right

  • A single 3 cMpc-resolution simulation is enough to train the super-resolver for fixed astrophysics, cutting the training-data requirement by orders of magnitude relative to emulator-based approaches.
  • The same model generalizes across box sizes from $128^3$ to $512^3$, so a network trained on a small cube can be applied to a gigaparsec-scale $512^3$ volume.
  • Sub-mK pixel error and power-spectrum residuals below SKA1-Low noise mean the super-resolved outputs can be used as surrogates for full high-resolution simulations in summary-statistic analyses.
  • Sampling a $512^3$ volume takes hours on GPUs rather than the weeks of CPU time needed for a direct simulation, making large-volume exploration feasible.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension the paper leaves implicit is parameter-conditioned generation: conditioning on astrophysical parameters rather than fixed values would let SKA1-Low data directly constrain reionization astrophysics.
  • Because the model was only tested at $z=10$ and on emission-only (nonnegative) brightness temperatures, the same architecture would need explicit handling of negative signal to generalize to other redshifts.
  • The single-simulation sufficiency result should be stress-tested across different astrophysical parameter sets; the paper's cosmic-variance explanation predicts it holds whenever the training box is large enough.
  • If instrument realism breaks the noiseless-downsampling assumption, a cheap fix would be to train on degraded low-resolution inputs via forward modeling, preserving the super-resolution gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper trains conditional score-based diffusion models (VP SDE, DDPM++ architecture) to super-resolve 3D 21-cm brightness temperature cubes from 21CMSPACE. Low-resolution inputs are produced by trilinear downsampling of the same HR simulation to 12 cMpc voxels and upsampling back; models are conditioned on these LR maps and on initial-condition fields. Six models are trained at two channel-base multipliers and three training-set fractions (1, 28, 56 simulations). The best model (DM8.1) is evaluated at 128^3, 256^3, and 512^3 volumes at z = 10 and achieves pixelwise RMSE about 0.56-0.57 mK and dimensionless power-spectrum residuals of 10^-2 to 10^-1 mK^2. The authors conclude that a single training simulation suffices regardless of initial conditions and that the network allows utilization of all spatial scales covered by SKA1-Low.

Significance. As a simulation-level proof of concept, this is a useful demonstration: the results are measured on held-out simulations from the same code, across three box sizes, with 24 rotations, and the paper reports an explicit compute comparison (4.7 GPU-hours for sampling vs 45,000 CPU-hours for the 512^3 simulation). Simultaneously matching the pixel PDF, power spectrum, and obtaining sub-mK RMSE is a worthwhile benchmark for 21-cm super-resolution. The main weakness is that all evaluations use a noiseless, perfectly aligned trilinear degradation of the same simulation code and require initial-condition fields as conditional inputs; the SKA1-Low applicability claim therefore goes beyond the current evidence. The paper would be publishable after reframing the claims to simulation-based emulation and/or adding instrument-realistic tests.

major comments (5)
  1. [§2.1, §3.2, Abstract] The low-resolution conditional input is generated from the HR simulation by trilinear downsampling to 12 cMpc voxels and trilinear upsampling, with no thermal noise, primary beam, point-spread function, foreground residuals, or line-of-sight/transverse anisotropy. All reported RMSE and power-spectrum residuals therefore measure performance on this specific synthetic degradation. The abstract and Section 4 claim that the network 'allows us to utilize all spatial scales covered by SKA1-Low' and Section 3.2 states residuals are 'below the expected noise level of SKA1-Low'; neither claim is supported without evaluating on instrument-realistic LR inputs. Please either add such an evaluation or restrict the conclusions to simulation-based super-resolution.
  2. [§2.1, §2.2, §4] The model is conditioned on the simulation initial conditions (matter overdensity and baryon-dark matter relative velocity fields). These fields are not observable by SKA1-Low, and the paper does not explain how they would be obtained for an observed volume. The statement in Section 4 that future parameter-conditioned generation would allow SKA1-Low scales to constrain astrophysics does not resolve this, because the described model at inference still requires initial-condition inputs. The SKA applicability claim needs either a protocol for constructing these inputs from observations or a demonstration that the LR 21-cm map alone suffices.
  3. [§3.1] The comparison of per-voxel RMSE of simulated 21-cm maps with REACH (25 mK) and SARAS 3 (213 mK) RMSE conflates different observables: those experiments constrain sky-averaged global 21-cm spectra, not tomographic brightness-temperature maps or spatial power spectra. This comparison should be removed or replaced with a quantity relevant to 21-cm power-spectrum measurements (e.g., noise on Δ^2_21 or sample variance).
  4. [§3.1] The conclusion that performance is independent of training-set size is inferred from overlapping 95% confidence intervals in Figure 1. Overlap of CIs is not a formal test of equivalence; the data are consistent with a modest dependence that the experiment is underpowered to detect. Please report the estimated difference between training fractions with a confidence interval or an equivalence bound. This matters because the 'single simulation suffices' claim is a headline result.
  5. [§3.2] The 512^3 evaluation uses a single simulation (with 24 rotations), and the 256^3 evaluation uses 8 subcubes of that same simulation. The quoted confidence intervals therefore quantify rotation and subcube variability, not field-to-field cosmic variance. Please state the effective number of independent realizations explicitly and avoid presenting these as independent field realizations.
minor comments (4)
  1. [Introduction, §3.2] There are typos: 'gravitaitonal' in the Introduction and 'quantitiy' in Section 3.2.
  2. [§2.2, §3.1] Please clarify whether reported RMSE values are computed after rescaling outputs back to physical mK units; the text says the HR target is normalized to the LR mean and standard deviation, but the un-normalization step and the exact RMSE definition are not specified.
  3. [Figure 1] '95 percentile confidence interval' should read '95% confidence interval'; also specify how the CIs were computed (e.g., bootstrap over which samples).
  4. [General] No code or data availability statement is included; given the reproducibility value of the trained models and preprocessing pipeline, please indicate where code, trained weights, and data will be released.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the super-resolution outputs are learned functions evaluated on held-out simulation boxes, not identities derived from the conditional inputs.

full rationale

The paper's central claim is that a score-based diffusion model can super-resolve 21-cm boxes. Training pairs are (LR conditional, HR target), where the LR conditional is generated from the HR target by trilinear downsampling and upsampling (Section 2.1). This is a standard supervised super-resolution setup; the model output is not defined as the LR input, and test RMSE values (Sections 3.1 and 3.2) are computed against HR targets that are not supplied as conditioning. The result therefore does not reduce by construction to its inputs. The claim that a single training simulation suffices is supported by an argument about cosmic variance and identical astrophysical parameters; whether that argument is correct is a scientific judgement, not circularity. Citations to 21CMSPACE (Refs. [8-11, 3, 7, 32-34, 26, 13, 14, 39]) and to REACH (Ref. [4]) include overlapping authors, but they are normal data-source and instrument references; they do not carry the argument that the trained model transfers to SKA1-Low. The SKA-utilization claim assumes the noiseless trilinear degradation used in training is representative of the instrument response; that is an external-validity limitation (noted in Section 2.1: 'generated from the HR simulations by a trilinear downsampling...'), not a circular derivation. No step in the paper's own equations or citation chain makes the reported accuracy equivalent to a fitted input.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical entities. Its central claim rests on the fidelity of 21CMSPACE as ground truth, on a cosmic-variance argument, and on the assumption that a synthetic trilinear downsampling represents SKA1-Low observations. The only hand-chosen numbers are standard ML hyperparameters and the LR resolution.

free parameters (3)
  • Low-resolution voxel size = 12 cMpc (4x downsampling from 3 cMpc HR voxels)
    Hand-chosen conditioning scale representing the spatial resolution of the LR input; it defines the super-resolution task and affects achievable accuracy.
  • Number of reverse diffusion steps = 100
    Hand-chosen sampling hyperparameter; more steps improve sample quality at higher compute cost, and the paper does not justify this specific value.
  • Channel base multiplier (CBM) = 4 or 8
    Network width hyperparameter; the paper compares CBM=4 and CBM=8 and chooses the larger for the best model.
assumptions (3)
  • domain assumption 21CMSPACE simulations are faithful models of the 21-cm signal
    The paper treats the semi-numerical simulation output as ground truth HR data (Section 2.1); all metrics are relative to this code, not to observations.
  • domain assumption Gigaparsec volumes are large enough that cosmic variance is negligible, so summary statistics depend only on the fixed astrophysical and cosmological parameters
    Stated in Section 3.1 to explain why one simulation may suffice; if cosmic variance were significant, the single-training-simulation claim would fail.
  • ad hoc to paper Trilinear downsampling to 12 cMpc followed by trilinear upsampling is a valid proxy for the spatial response of SKA1-Low
    The LR conditional input is generated this way (Section 2.1); the SKA1-Low applicability claim assumes this degradation matches real observations, but no beam or noise model is included.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cosmological super-resolution of the 21-cm signal." pith.science (2026). https://pith.science/paper/ZDSC6SPA

@misc{pith2026250200852,
  author       = {Pith},
  title        = {Pith review of: Cosmological super-resolution of the 21-cm signal},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZDSC6SPA}},
  note         = {Machine review of arXiv:2502.00852}
}
abstract

In this study, we train score-based diffusion models to super-resolve gigaparsec-scale cosmological simulations of the 21-cm signal. We examine the impact of network and training dataset size on model performance, demonstrating that a single simulation is sufficient for a model to learn the super-resolution task regardless of the initial conditions. Our best-performing model achieves pixelwise $\mathrm{RMSE}\sim0.57\ \mathrm{mK}$ and dimensionless power spectrum residuals ranging from $10^{-2}-10^{-1}\ \mathrm{mK^2}$ for $128^3$, $256^3$ and $512^3$ voxel simulation volumes at redshift $10$. The super-resolution network ultimately allows us to utilize all spatial scales covered by the SKA1-Low instrument, and could in future be employed to help constrain the astrophysics of the early Universe.

Figures

Figures reproduced from arXiv: 2502.00852 by the authors.

Figure 1
Figure 1. Left: Models of varying network size (CBM ∈ {4, 8}) trained on 1.25%/35%/70% (1/28/56 models) of the dataset. Right: RMSE distribution of the model with CBM = 8 and trained on 1.25% of the dataset. We train models using the Adam optimizer [22] with an initial learning rate of 10−3 . After the first 1000 epochs, the learning rate is reduced to 10−4 . Validation checks are conducted every 50 epochs after the model pas… view at source ↗
Figure 2
Figure 2. Row 1 shows slices from the 21-cm HR 5123 , 2563 , and 1283 targets. Row 2 shows slices from sampled models at each scale using DM8.1. Each sample is independently generated from the model, rather than being zoomed-in sections of the largest cube. Row 3 presents histograms of the voxel distributions, and Row 4 compares the 21-cm dimensionless power spectrum of HR targets and SR predictions across different spatial s… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 25 canonical work pages

  1. [1]

    R. Barkana. The rise of the first stars: Supersonic streaming, radiative feedback, and 21-cm cosmology. Phys. Rep., 645:1–59, July 2016. doi: 10.1016/j.physrep.2016.06.006

    doi:10.1016/j.physrep.2016.06.006 2016

  2. [2]

    H. T. J. Bevins, S. Heimersheim, I. Abril-Cabezas, A. Fialkov, E. de Lera Acedo, W. Handley, S. Singh, and R. Barkana. Joint analysis constraints on the physics of the first galaxies with low-frequency radio astronomy data. Mon. Not. R. Astron. Soc. , 527(1):813–827, Jan. 2024. doi: 10.1093/mnras/stad3194

    doi:10.1093/mnras/stad3194 2024

  3. [3]

    Cohen, A

    A. Cohen, A. Fialkov, and R. Barkana. The 21-cm BAO signature of enriched low-mass galaxies during cosmic reionization. Mon. Not. R. Astron. Soc. , 459(1):L90–L94, June 2016. doi: 10.1093/mnrasl/slw047

    doi:10.1093/mnrasl/slw047 2016

  4. [4]

    de Lera Acedo, D

    E. de Lera Acedo, D. I. L. de Villiers, N. Razavi-Ghods, W. Handley, A. Fialkov, A. Magro, D. Anstey, H. T. J. Bevins, R. Chiello, J. Cumner, A. T. Josaitis, I. L. V . Roque, P. H. Sims, K. H. Scheutwinkel, P. Alexander, G. Bernardi, S. Carey, J. Cavillot, W. Croukamp, J. A. Ely, T. Gessey-Jones, Q. Gueuning, R. Hills, G. Kulkarni, R. Maiolino, P. D. Meer...

    2022

  5. [5]

    P. E. Dewdney, P. J. Hall, R. T. Schilizzi, and T. J. L. W. Lazio. The square kilometre array. Proceedings of the IEEE, 97(8):1482–1496, 2009. doi: 10.1109/JPROC.2009.2021005

    arXiv 2009

  6. [6]

    Dhariwal and A

    P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, editors,Advances in Neural Infor- mation Processing Systems, volume 34 of NIPS 2021, pages 8780–8794. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/paper_files/paper/2021/file/ 49ad23d1ec9fa...

    work page 2021

  7. [7]

    Fialkov and R

    A. Fialkov and R. Barkana. Signature of excess radio background in the 21-cm global signal and power spectrum. Mon. Not. R. Astron. Soc. , 486(2):1763–1773, June 2019. doi: 10.1093/ mnras/stz873. 6

    work page 2019

  8. [8]

    Fialkov, R

    A. Fialkov, R. Barkana, D. Tseliakhovich, and C. M. Hirata. Impact of the relative motion between the dark matter and baryons on the first stars: semi-analytical modelling. Mon. Not. R. Astron. Soc., 424(2):1335–1345, Aug. 2012. doi: 10.1111/j.1365-2966.2012.21318.x

    arXiv 2012

Show all 53 references
  1. [9]

    Fialkov, R

    A. Fialkov, R. Barkana, E. Visbal, D. Tseliakhovich, and C. M. Hirata. The 21-cm signature of the first stars during the Lyman-Werner feedback era. Mon. Not. R. Astron. Soc. , 432(4): 2909–2916, July 2013. doi: 10.1093/mnras/stt650

    2013 doi

  2. [10]

    Fialkov, R

    A. Fialkov, R. Barkana, A. Pinhas, and E. Visbal. Complete history of the observable 21 cm signal from the first stars during the pre-reionization era. Mon. Not. R. Astron. Soc. , 437(1): L36–L40, Jan. 2014. doi: 10.1093/mnrasl/slt135

    2014 doi

  3. [11]

    Fialkov, R

    A. Fialkov, R. Barkana, and E. Visbal. The observable signature of late heating of the Universe during cosmic reionization. Nature, 506(7487):197–199, Feb. 2014. doi: 10.1038/nature12999

    2014 doi

  4. [12]

    S. R. Furlanetto, S. P. Oh, and F. H. Briggs. Cosmology at low frequencies: The 21 cm transition and the high-redshift Universe. Phys. Rep. , 433(4-6):181–301, Oct. 2006. doi: 10.1016/j.physrep.2006.08.002

    2006 doi

  5. [13]

    Gessey-Jones, N

    T. Gessey-Jones, N. S. Sartorio, A. Fialkov, G. M. Mirouh, M. Magg, R. G. Izzard, E. de Lera Acedo, W. J. Handley, and R. Barkana. Impact of the primordial stellar initial mass function on the 21-cm signal. Mon. Not. R. Astron. Soc. , 516(1):841–860, Oct. 2022. doi: 10.1093/mn...

    2022 doi

  6. [14]

    Gessey-Jones, A

    T. Gessey-Jones, A. Fialkov, E. de Lera Acedo, W. J. Handley, and R. Barkana. Signatures of cosmic ray heating in 21-cm observables. Mon. Not. R. Astron. Soc. , 526(3):4262–4284, Dec

  7. [15]

    Gessey-Jones, S

    T. Gessey-Jones, S. Pochinda, H. T. J. Bevins, A. Fialkov, W. J. Handley, E. de Lera Acedo, S. Singh, and R. Barkana. On the constraints on superconducting cosmic strings from 21-cm cosmology. Mon. Not. R. Astron. Soc., 529(1):519–536, Mar. 2024. doi: 10.1093/mnras/stae512

    2024 doi

  8. [16]

    Goodfellow, J

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K. Weinberger, editors, Advances in Neural Information Processing Systems , volume ...

    2014

  9. [17]

    Abdurashidova, T

    HERA Collaboration, Z. Abdurashidova, T. Adams, J. E. Aguirre, P. Alexander, Z. S. Ali, R. Baartman, Y . Balfour, R. Barkana, A. P. Beardsley, G. Bernardi, T. S. Billings, J. D. Bowman, R. F. Bradley, D. Breitman, P. Bull, J. Burba, S. Carey, C. L. Carilli, C. Cheng, S. Choudh...

  10. [18]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, editors, Advances in Neural Informa- tion Processing Systems , volume 33 of NIPS 2020 , pages 6840–6851. Curran Associates, Inc., 2020. URL...

    2020

  11. [19]

    doi: 10.3847/1538-4357/acaf50

    doi

  12. [20]

    Karras, M

    T. Karras, M. Aittala, T. Aila, and S. Laine. Elucidating the design space of diffusion-based generative models. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems , volume 35, pages 26565–26577. Cur...

    2022

  13. [21]

    Karchev, N

    K. Karchev, N. Anau Montel, A. Coogan, and C. Weniger. Strong-Lensing Source Reconstruc- tion with Denoising Diffusion Restoration Models. In 36th Conference on Neural Information Processing Systems: Workshop on Machine Learning and the Physical Sciences , 11 2022

    2022

  14. [22]

    D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. arXiv e-prints, art. arXiv:1412.6980, Dec. 2014. doi: 10.48550/arXiv.1412.6980

    2014 doi arXiv

  15. [23]

    H. D. Kaur, N. Gillet, and A. Mesinger. Minimum size of 21-cm simulations. Monthly Notices of the Royal Astronomical Society , 495(2):2354–2362, 05 2020. ISSN 0035-8711. doi: 10.1093/mnras/staa1323. URL https://doi.org/10.1093/mnras/staa1323

    2020 doi

  16. [24]

    Legin, M

    R. Legin, M. Ho, P. Lemos, L. Perreault-Levasseur, S. Ho, Y . Hezaveh, and B. Wandelt. Posterior sampling of the initial conditions of the universe from non-linear large scale structures using score-based generative models. Mon. Not. R. Astron. Soc. , 527(1):L173–L178, Jan. 20...

    2024 doi

  17. [25]

    Koopmans, J

    L. Koopmans, J. Pritchard, G. Mellema, J. Aguirre, K. Ahn, R. Barkana, I. van Bemmel, G. Bernardi, A. Bonaldi, F. Briggs, A. G. de Bruyn, T. C. Chang, E. Chapman, X. Chen, B. Courty, P. Dayal, A. Ferrara, A. Fialkov, F. Fiore, K. Ichiki, I. T. Illiev, S. Inoue, V . Jelic, M. J...

    2015

  18. [26]

    M. Magg, I. Reis, A. Fialkov, R. Barkana, R. S. Klessen, S. C. O. Glover, L.-H. Chen, T. Hartwig, and A. T. P. Schauer. Effect of the cosmological transition to metal-enriched star formation on the hydrogen 21-cm signal. Mon. Not. R. Astron. Soc. , 514(3):4433–4449, Aug. 2022....

    2022 doi

  19. [27]

    H. Li, Y . Yang, M. Chang, S. Chen, H. Feng, Z. Xu, Q. Li, and Y . Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47–59,

  20. [28]

    O. S. D. O’Hara, F. Dulwich, E. de Lera Acedo, J. Dhandha, T. Gessey-Jones, D. Anstey, and A. Fialkov. Understanding spectral artefacts in SKA-LOW 21-cm cosmology experiments: the impact of cable reflections. arXiv e-prints, art. arXiv:2402.04008, Feb. 2024. doi: 10.48550/ arX...

    2024 doi arXiv

  21. [29]

    Pochinda, T

    S. Pochinda, T. Gessey-Jones, H. T. J. Bevins, A. Fialkov, S. Heimersheim, I. Abril-Cabezas, E. de Lera Acedo, S. Singh, S. Sikder, and R. Barkana. Constraining the properties of Population III galaxies with multiwavelength observations. Mon. Not. R. Astron. Soc. , 531(1):1113...

    2024 doi

  22. [30]

    Mesinger, editor

    A. Mesinger, editor. The Cosmic 21-cm Revolution . 2514-3433. IOP Publishing, 2019. ISBN 978-0-7503-2236-2. doi: 10.1088/2514-3433/ab4a73. URL https://dx.doi.org/10.1088/ 2514-3433/ab4a73

    2019 doi

  23. [31]

    J. R. Pritchard and A. Loeb. 21 cm cosmology in the 21st century. Reports on Progress in Physics, 75(8):086901, Aug. 2012. doi: 10.1088/0034-4885/75/8/086901. 8

    2012 doi

  24. [32]

    I. Reis, A. Fialkov, and R. Barkana. High-redshift radio galaxies: a potential new source of 21-cm fluctuations. Mon. Not. R. Astron. Soc. , 499(4):5993–6008, Dec. 2020. doi: 10.1093/ mnras/staa3091

    2020

  25. [33]

    Podell, Z

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, and R. Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations , 2024. URL https://openreview. net...

    2024

  26. [34]

    I. Reis, R. Barkana, and A. Fialkov. Shot noise and scatter in the star formation efficiency as a source of 21-cm fluctuations. Mon. Not. R. Astron. Soc. , 511(4):5265–5273, Apr. 2022. doi: 10.1093/mnras/stac411

    2022 doi

  27. [35]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, June 2022

    2022

  28. [36]

    I. Reis, A. Fialkov, and R. Barkana. The subtlety of Ly α photons: changing the expected range of the 21-cm signal. Mon. Not. R. Astron. Soc. , 506(4):5479–5493, Oct. 2021. doi: 10.1093/mnras/stab2089

    2021 doi

  29. [37]

    Saharia, J

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence , 45 (4):4713–4726, 2023. doi: 10.1109/TPAMI.2022.3204461

    2023

  30. [38]

    A. M. M. Scaife. Big telescope, big data: towards exascale with the Square Kilometre Array. Philosophical Transactions of the Royal Society of London Series A , 378(2166):20190060, Mar

  31. [39]

    Rouhiainen, M

    A. Rouhiainen, M. Münchmeyer, G. Shiu, M. Gira, and K. Lee. Superresolution emulation of large cosmological fields with a 3D conditional diffusion model. Phys. Rev. D, 109(12):123536, June 2024. doi: 10.1103/PhysRevD.109.123536

    2024 doi

  32. [40]

    Singh, N

    S. Singh, N. T. Jishnu, R. Subrahmanyan, N. Udaya Shankar, B. S. Girish, A. Raghunathan, R. Somashekar, K. S. Srivani, and M. Sathyanarayana Rao. On the detection of a cosmic dawn signal in the radio background. Nature Astronomy, 6:607–617, Feb. 2022. doi: 10.1038/ s41550-022-01610-5

    2022

  33. [41]

    M. J. Smith, J. E. Geach, R. A. Jackson, N. Arora, C. Stone, and S. Courteau. Realistic galaxy image simulation via score-based generative models. Mon. Not. R. Astron. Soc. , 511(2): 1808–1818, Apr. 2022. doi: 10.1093/mnras/stac130

    2022 doi

  34. [42]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In F. Bach and D. Blei, editors, Proceedings of the 32nd International Conference on Machine Learning , volume 37 of Proceedings of Machine Learning...

    2015

  35. [43]

    Sikder, R

    S. Sikder, R. Barkana, A. Fialkov, and I. Reis. Strong 21-cm fluctuations and anisotropy due to the line-of-sight effect of radio galaxies at cosmic dawn. Mon. Not. R. Astron. Soc. , 527(4): 10975–10985, Feb. 2024. doi: 10.1093/mnras/stad3847

    2024 doi

  36. [44]

    Vahdat, K

    A. Vahdat, K. Kreis, and J. Kautz. Score-based generative modeling in latent space. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, editors,Advances in Neural Information Processing Systems, volume 34, pages 11287–11302. Curran Associates, Inc., 2021. ...

    2021

  37. [45]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin. Attention is all you need. In I. Guyon, U. V . Luxburg, S. Ben- gio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Process...

    2017

  38. [46]

    Visbal, R

    E. Visbal, R. Barkana, A. Fialkov, D. Tseliakhovich, and C. M. Hirata. The signature of the first stars in atomic hydrogen at redshift 20. Nature, 487(7405):70–73, July 2012. doi: 10.1038/nature11177

    2012 doi

  39. [47]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based gener- ative modeling through stochastic differential equations. In International Conference on Learn- ing Representations, 2021. URL https://openreview.net/forum?id=PxTIG12RRHS

    2021

  40. [48]

    Yang and S

    R. Yang and S. Mandt. Lossy image compression with conditional diffusion models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neural Information Processing Systems, volume 36 of NIPS 2023, pages 64971–64995. Curran Associates, Inc...

    2023

  41. [49]

    Zhao, Y .-S

    X. Zhao, Y .-S. Ting, K. Diao, and Y . Mao. Can diffusion model conditionally generate astrophysical images? Mon. Not. R. Astron. Soc. , 526(2):1699–1712, Dec. 2023. doi: 10.1093/ mnras/stad2778. 10

    2023

  42. [51]

    X. Wang, R. Girshick, A. Gupta, and K. He. Non-local neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2018

    2018

  43. [2020]

    doi: 10.1098/rsta.2019.0060

    2019

  44. [2022]

    doi: https://doi.org/10.1016/j.neucom.2022.01.029

    ISSN 0925-2312. doi: https://doi.org/10.1016/j.neucom.2022.01.029. URL https: //www.sciencedirect.com/science/article/pii/S0925231222000522

    2022 doi

  45. [2023]

    doi: 10.1093/mnras/stad3014

    doi

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.