Pith. sign in

REVIEW 3 major objections 5 minor 59 references

A Vision Transformer-UNet hybrid emulates 3D EoR 21-cm cubes conditioned on reionization parameters, matching simulated large-scale power spectra with mild small-scale suppression.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 12:56 UTC pith:JGHTX4KD

load-bearing objection Solid in-domain emulator with an honest limitations section, but the out-of-domain evidence for unseen initial conditions is too thin to support the field-level inference claim. the 3 major comments →

arxiv 2510.01121 v2 pith:JGHTX4KD submitted 2025-10-01 astro-ph.CO

CosmoUiT: A Vision Transformer-UNet Hybrid for Fast and Accurate Emulation of 21-cm Maps from the Epoch of Reionization

Prasad Rajesh Posture , Yashrajsinh Mahida , Suman Majumdar , Leon Noble This is my paper
classification astro-ph.CO
keywords signalreionizationcosmouitfieldemulatorinputparametersvision
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

During the Epoch of Reionization, ultraviolet light from the first galaxies ionized hydrogen in intergalactic space. The redshifted 21-cm line of neutral hydrogen is the main observational window into this process, but realistic simulations are expensive, and parameter inference needs many simulation runs. CosmoUiT is a machine-learning model trained on about 7,000 seminumerical simulations that vary three reionization parameters: minimum halo mass, ionizing efficiency, and mean free path. Given a dark-matter density field and a halo catalog at redshift 7, plus values of the three parameters, it predicts the full 3D neutral-hydrogen fraction cube, which is then converted to a 21-cm brightness temperature cube.

The model combines two ideas. A vision transformer splits each cube into small patches and uses multi-head self-attention so every patch can exchange information with every other patch, capturing global morphology. A UNet then processes the reconstructed fields with convolution and skip connections, preserving small-scale detail. The reionization parameters are injected both as extra tokens in the transformer and at the UNet bottleneck, so the output changes with the parameters instead of producing a generic field. The authors test the model on parameter sets from the same simulation grid and report voxel-wise R² around 0.91–0.94 for neutral fraction fields and higher for brightness temperature. They also test on a dark-matter realization from a different random seed; performance drops markedly, with R² around 0.37–0.54, which they acknowledge.

The main error is at the boundaries of ionized regions: the network produces gradual transitions where the simulation has sharp fronts, suppressing small-scale power and shifting bubble-size distributions. No code or data

Core claim

The abstract and Section 6 claim that CosmoUiT can "emulate the entire 3D 21-cm signal cube with high accuracy at both large and small scales" and that it produces parameter-specific predictions conditioned on the input astrophysical parameters. If true, the model maps fixed dark-matter and halo fields plus (M_h,min, N_ion, R_mfp) to ReionYuga-equivalent 3D x_HI and δT_b cubes in a single forward pass, with in-domain R² ≈ 0.91–0.94 for x_HI and better for δT_b, and power spectra matching simulations within sample variance on large scales (k ≲ 0.3 Mpc⁻¹).

Load-bearing premise

The emulator is trained on a single realization of the dark-matter and halo fields—variability across the 7,204 training samples comes only from the three reionization parameters (Section 2). The paper's field-level-inference goal assumes the learned parameter-to-ionization mapping transfers to different cosmic initial conditions; but the out-of-domain test uses only two unseen seeds and shows R² degrading to 0.37–0.54 (Section 5.3). A second load-bearing premise is that ReionYuga's excursion-set outputs are adequate stand-ins for the true 21-cm signal, so the emulator's fidelity is only as good as that seminumerical model.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces CosmoUiT, a hybrid vision-transformer/UNet emulator designed to map a fixed dark-matter density field, a fixed halo density field, and three ReionYuga reionization parameters (M_h,min, N_ion, R_mfp) to 3D neutral-fraction and 21-cm brightness-temperature cubes at redshift z=7. Training uses 7,204 ReionYuga parameter combinations, each augmented with 48 cube orientations and downsampled to 96^3, with all samples sharing a single dark-matter/halo realization. The authors report strong in-domain voxel-wise agreement (x_HI R² ≈ 0.91–0.94; δT_b R² ≈ 0.91–0.94), and good large-scale power-spectrum agreement, while identifying a 'fuzzy boundary' problem that suppresses small-scale x_HI power and biases bubble-size distributions. On two out-of-domain dark-matter/halo realizations, accuracy drops substantially (x_HI R² ≈ 0.54 and 0.49; δT_b R² ≈ 0.37 and 0.39). The stated motivation is to enable fast field-level Bayesian inference for upcoming SKA-era 21-cm tomographic observations.

Significance. If the in-domain accuracy and parameter conditioning hold, CosmoUiT is a useful step toward field-level emulation of 21-cm EoR signals, and the comparison with CosmoViT and CosmoUNet usefully demonstrates why transformer conditioning at an early stage helps. The paper reports standard metrics, physically sensible summary statistics (BSD, power spectrum), and openly discusses the fuzzy-boundary limitation. The main strength is the clean architecture ablation and the consistently reported in-domain metrics. However, the field-level inference goal rests on generalization to unseen initial conditions, and the evidence for that is currently thin and partly contradicted by the reported out-of-domain numbers. The central claims therefore need substantial qualification or additional experiments before they can support the stated application.

major comments (3)
  1. [Section 5.3, Abstract, Section 6] The field-level inference claim is not supported by the out-of-domain results. With one training realization, the only transfer evidence is two unseen-seed tests: x_HI R² = 0.54 and 0.49, and δT_b R² = 0.37 and 0.39 (Figs. 9–10), with errors no longer confined to ionization boundaries. Section 5.3 nonetheless concludes that the model 'has learned a generalized mapping,' and the Section 6 bullet claims it 'performs reliably well' on unseen seeds. These statements are inconsistent with the reported metrics. Please provide a quantitative multi-seed evaluation (N > 10 unseen realizations with mean and scatter), or train on multiple initial-condition realizations; otherwise the abstract's unqualified 'high accuracy' claim should be restricted to the training realization.
  2. [Section 4.2] The validation split is underspecified in a way that affects the headline R² = 0.94. The text says the 7,204 × 48 orientation-augmented dataset 'was split into 80% training and 20% validation subsets, ensuring coverage across the full range of reionization parameter values.' If the split is random over augmented samples, the same parameter combination appears in both training and validation in different orientations, so the validation score measures orientation invariance, not parameter generalization. Since all samples also share the same DM/halo field, the validation cannot test field-level generalization at all. Please state explicitly whether the split is disjoint in parameter combination and initial condition, and if so, report validation metrics for held-out parameter combinations only.
  3. [Section 5.1.3 / Fig. 6 / Abstract] The abstract's 'high accuracy at both large and small scales' is stronger than the quantitative evidence. For x_HI, Fig. 6 shows a systematic suppression for k > 0.3 Mpc⁻¹, described in the text as a mild underprediction of less than an order of magnitude. This may be an acceptable limitation, but it should be stated in the abstract and conclusions rather than as an unqualified success. In addition, the text claims the predicted x_HI power spectrum 'falls within the error bars of the simulated field,' but the figure does not show any error bars or sample-variance estimate; please display the estimated sample variance or revise the wording.
minor comments (5)
  1. [Table 1] The table heading says 'Corrected summary of the CosmoUiT architecture'; the word 'Corrected' appears to be an artifact and should be removed.
  2. [Section 2] The parameter sampling is described as 'uniformly gridded parameter space' with 7,204 sets, but the grid spacing or sampling strategy is not given. Please include this for reproducibility.
  3. [Eq. (5.1)] The Fourier transform definition omits the normalization factor commonly used in power-spectrum conventions. Since the subsequent power spectra are dimensionless and ratio-plotted, this is not fatal, but a consistent convention should be stated.
  4. [Section 5.3] The statement that MSE increased 'by nearly a factor of six' for the unseen-seed examples is correct for the two displayed cases, but it would be helpful to report the distribution over a larger set of unseen seeds rather than only two examples.
  5. [General] No code or data availability statement is provided. Given the importance of reproducibility for emulator papers, please state whether the training code and trained weights will be released.

Circularity Check

0 steps flagged

No significant circularity: the emulator is trained and evaluated on held-out ReionYuga outputs, not on a quantity derived from its own fit.

full rationale

The paper's central claim is that CosmoUiT can emulate 3D neutral-fraction and 21-cm brightness-temperature fields from fixed dark-matter/halo fields plus three reionization parameters. This is a supervised emulation task: the target fields are produced by the ReionYuga seminumerical code (Section 2), and the network is trained to map inputs to those outputs. The validation (Sections 4.2 and 5) compares predictions to simulation outputs for held-out parameter combinations and, in Section 5.3, for two unseen initial-condition realizations. There is no step in which a fitted parameter is renamed as a prediction: the reionization parameters are inputs, not fitted outputs, and the reported R^2/MSE/SSIM and power-spectrum metrics are computed against independently generated simulation fields. The self-citations to ReionYuga [43-45] identify the code used to create training labels; this is a tool citation rather than an appeal to a uniqueness theorem or an unverified prior result, and it is not load-bearing for the emulator's in-domain accuracy, which is measured against held-out simulation outputs. The out-of-domain degradation (x_HI R^2 dropping to 0.54 and 0.49, and delta_Tb R^2 to 0.37 and 0.39) is a genuine generalization weakness, and the 48-orientation augmentation split could potentially mix augmented copies of the same parameter set across train and validation, but neither issue makes the derivation equivalent to its inputs by construction. The paper also candidly states limitations (single redshift, fixed initial-condition training seed, no foregrounds/systematics) in Section 6. Overall, this is an empirical performance claim about a trained surrogate, not a first-principles derivation that reduces to its own inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard ML architecture choices plus the domain assumption that ReionYuga's seminumerical output is the appropriate target. There are no newly invented physical entities. The most fragile element is the single-realization training assumption, which the paper itself flags and only weakly tests.

free parameters (6)
  • Input cube downsampling factor = 96^3 from native 384^3
    Training data are downsampled to 96^3 to fit memory; small-scale accuracy claims are relative to this reduced resolution, not the native simulation resolution.
  • Vision transformer patch size = 8×8×8
    Hand-chosen patch size that balances local detail versus global context in the attention mechanism.
  • Embedding/projection dimension = 512
    All patch tokens and parameter tokens are projected to 512 dimensions; a hand-chosen architectural capacity.
  • Transformer encoder depth and heads = 8 layers, 8 attention heads
    Hand-chosen model depth/capacity; no sensitivity analysis is reported.
  • UNet feature map schedule = 32→64→128→256→512, then decoding
    Hand-chosen encoder/decoder widths; affects the balance of large- and small-scale reconstruction.
  • Training schedule = 60 epochs, Adam lr=1e-4, batch size 16
    Hand-chosen hyperparameters; validation loss is shown but no ablation or sensitivity study is provided.
axioms (4)
  • domain assumption ReionYuga excursion-set ionization model produces physically adequate training labels for the EoR 21-cm signal
    All training and validation labels come from ReionYuga; the emulator's fidelity is judged against this seminumerical model, not against radiative-transfer simulations or observations.
  • domain assumption The DM-only PM simulation and FoF halo catalog at z=7 provide sufficient input information
    The model maps DM/halo fields plus parameters to the ionization field; this assumes the DM and halo fields contain the relevant environmental information for reionization morphology.
  • ad hoc to paper A single fixed realization of the DM/halo fields is sufficient to learn a parameter-to-field mapping that generalizes to other initial conditions
    Training uses one seed and the paper's own out-of-domain test shows substantial degradation (R² 0.37–0.54) on two unseen seeds, so this assumption is load-bearing and only weakly tested.
  • domain assumption Downsampling from 384^3 to 96^3 preserves the scales relevant for the claimed accuracy
    The model outputs 96^3 cubes; all power-spectrum and morphology comparisons are made at this reduced resolution, so claims about 'small scales' are limited to what remains after downsampling.

pith-pipeline@v1.3.0-alltime-deepseek · 24022 in / 9841 out tokens · 75781 ms · 2026-08-04T12:56:50.986713+00:00 · methodology

0 comments
read the original abstract

The observation of the redshifted 21-cm signal from the intergalactic medium will probe the epoch of reionization (EoR) with unprecedented detail. Various simulations are being developed and used to predict and understand the nature and morphology of this signal. However, these simulations are computationally very expensive and time-consuming to produce in large numbers. To overcome this problem, an efficient field-level emulator of this signal is required. However, the EoR 21-cm signal is highly non-Gaussian; therefore, capturing the correlations between different scales of this signal, which is directly related to the evolution of the reionization, with the neural network is quite difficult. Here, we introduce CosmoUiT, a UNet integrated vision transformer-based architecture, to overcome these difficulties. CosmoUiT emulates the 3D cubes of 21-cm signal from the EoR, for a given input dark matter density field, halo density field, and reionization parameters. CosmoUiT uses the multi-head self-attention mechanism of the transformer to capture the long-range dependencies and convolutional layers in the UNet to capture the small-scale variations in the target 21-cm field. Furthermore, the training of the emulator is conditioned on the input reionization parameters such that it gives a fast and accurate prediction of the 21-cm field for different sets of input reionization parameters. We evaluate the predictions of our emulator by comparing various statistics (e.g., bubble size distribution, power spectrum) and morphological features of the emulated and simulated maps. We further demonstrate that this vision transformer-based architecture can emulate the entire 3D 21-cm signal cube with high accuracy at both large and small scales.

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

59 extracted references · 53 linked inside Pith

  1. [1]

    Fan, M.A

    X. Fan, M.A. Strauss, R.H. Becker, R.L. White, J.E. Gunn, G.R. Knapp et al.,Constraining the Evolution of the Ionizing Background and the Epoch of Reionization with z~6 Quasars. II. A Sample of 19 Quasars,The Astronomical Journal132(2006) 117 [astro-ph/0512082]

    Pith/arXiv arXiv 2006

  2. [2]

    McGreer, A

    I.D. McGreer, A. Mesinger and X. Fan,The first (nearly) model-independent constraint on the neutral hydrogen fraction at z∼6,Monthly Notices of the Royal Astronomical Society415 (2011) 3237 [1101.3314]

    Pith/arXiv arXiv 2011

  3. [3]

    McGreer, A

    I.D. McGreer, A. Mesinger and V. D’Odorico,Model-independent evidence in favour of an end to reionization by z≈6,Monthly Notices of the Royal Astronomical Society447(2015) 499 [1411.5375]

    Pith/arXiv arXiv 2015

  4. [4]

    Ouchi, K

    M. Ouchi, K. Shimasaku, H. Furusawa, T. Saito, M. Yoshida, M. Akiyama et al.,Statistics of 207 LyαEmitters at a Redshift Near 7: Constraints on Reionization and Galaxy Formation Models,The Astrophysical Journal723(2010) 869 [1007.2961]

    Pith/arXiv arXiv 2010

  5. [5]

    Robertson, R.S

    B.E. Robertson, R.S. Ellis, S.R. Furlanetto and J.S. Dunlop,Cosmic Reionization and Early Star-forming Galaxies: A Joint Analysis of New Constraints from Planck and the Hubble Space Telescope,The Astrophysical Journal Letters802(2015) L19 [1502.02024]

    Pith/arXiv arXiv 2015

  6. [6]

    Zheng, J

    Z.-Y. Zheng, J. Wang, J. Rhoads, L. Infante, S. Malhotra, W. Hu et al.,First Results from the Lyman Alpha Galaxies in the Epoch of Reionization (LAGER) Survey: Cosmological Reionization at z∼7,The Astrophysical Journal Letters842(2017) L22 [1703.02985]

    Pith/arXiv arXiv 2017

  7. [7]

    Komatsu, K.M

    E. Komatsu, K.M. Smith, J. Dunkley, C.L. Bennett, B. Gold, G. Hinshaw et al.,Seven-year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Cosmological Interpretation, The Astrophysical Journal Supplement192(2011) 18 [1001.4538]

    Pith/arXiv arXiv 2011

  8. [8]

    Planck Collaboration, R. Adam, N. Aghanim, M. Ashdown, J. Aumont, C. Baccigalupi et al., Planck intermediate results. XLVII. Planck constraints on reionization history,Astronomy& Astrophysics596(2016) A108 [1605.03507]

    Pith/arXiv arXiv 2016

  9. [9]

    Aghanim, Y

    Planck Collaboration, N. Aghanim, Y. Akrami, M. Ashdown, J. Aumont, C. Baccigalupi et al., Planck 2018 results. VI. Cosmological parameters,Astronomy&Astrophysics641(2020) A6 [1807.06209]

    Pith/arXiv arXiv 2018

  10. [10]

    Furlanetto, S.P

    S.R. Furlanetto, S.P. Oh and F.H. Briggs,Cosmology at low frequencies: The 21 cm transition and the high-redshift Universe,Physics Reports433(2006) 181 [astro-ph/0608032]

    Pith/arXiv arXiv 2006

  11. [11]

    Pritchard and A

    J.R. Pritchard and A. Loeb,21 cm cosmology in the 21st century,Reports on Progress in Physics75(2012) 086901 [1109.6012]

    Pith/arXiv arXiv 2012

  12. [12]

    Paciga, J.G

    G. Paciga, J.G. Albert, K. Bandura, T.-C. Chang, Y. Gupta, C. Hirata et al.,A simulation-calibrated limit on the H I power spectrum from the GMRT Epoch of Reionization experiment,Monthly Notices of the Royal Astronomical Society433(2013) 639 [1301.5906]

    Pith/arXiv arXiv 2013

  13. [13]

    Abdurashidova, J.E

    HERA Collaboration, Z. Abdurashidova, J.E. Aguirre, P. Alexander, Z.S. Ali, Y. Balfour et al., First Results from HERA Phase I: Upper Limits on the Epoch of Reionization 21 cm Power Spectrum,The Astrophysical Journal925(2022) 221 [2108.02263]

    Pith/arXiv arXiv 2022

  14. [14]

    Mertens, M

    F.G. Mertens, M. Mevius, L.V.E. Koopmans, A.R. Offringa, S. Zaroubi, A. Acharya et al., Deeper multi-redshift upper limits on the epoch of reionisation 21 cm signal power spectrum – 31 – from LOFAR between z = 8.3 and z = 10.1,Astronomy&Astrophysics698(2025) A186 [2503.05576]

    Pith/arXiv arXiv 2025

  15. [15]

    Barry, M

    N. Barry, M. Wilensky, C.M. Trott, B. Pindor, A.P. Beardsley, B.J. Hazelton et al.,Improving the Epoch of Reionization Power Spectrum Results from Murchison Widefield Array Season 1 Observations,The Astrophysical Journal884(2019) 1 [1909.00561]

    Pith/arXiv arXiv 2019

  16. [16]

    Koopmans, J

    L. Koopmans, J. Pritchard, G. Mellema, J. Aguirre, K. Ahn, R. Barkana et al.,The Cosmic Dawn and Epoch of Reionisation with SKA, inAdvancing Astrophysics with the Square Kilometre Array (AASKA14), p. 1, Apr., 2015, DOI [1505.07568]

    Pith/arXiv arXiv 2015

  17. [17]

    Gehlot, F.G

    B.K. Gehlot, F.G. Mertens, L.V.E. Koopmans, M.A. Brentjens, S. Zaroubi, B. Ciardi et al., The first power spectrum limit on the 21-cm signal of neutral hydrogen during the Cosmic Dawn at z = 20-25 from LOFAR,Monthly Notices of the Royal Astronomical Society488 (2019) 4271 [1809.06661]

    Pith/arXiv arXiv 2019

  18. [18]

    Mertens, M

    F.G. Mertens, M. Mevius, L.V.E. Koopmans, A.R. Offringa, G. Mellema, S. Zaroubi et al., Improved upper limits on the 21 cm signal power spectrum of neutral hydrogen at z≈9.1 from LOFAR,Monthly Notices of the Royal Astronomical Society493(2020) 1662 [2002.07196]

    Pith/arXiv arXiv 2020

  19. [19]

    Abdurashidova, J.E

    Z. Abdurashidova, J.E. Aguirre, P. Alexander, Z.S. Ali, Y. Balfour, R. Barkana et al.,HERA Phase I Limits on the Cosmic 21 cm Signal: Constraints on Astrophysics and Cosmology during the Epoch of Reionization,The Astrophysical Journal924(2022) 51 [2108.07282]

    Pith/arXiv arXiv 2022

  20. [20]

    Ceccotti, A.R

    E. Ceccotti, A.R. Offringa, F.G. Mertens, L.V.E. Koopmans, S. Munshi, J.K. Chege et al., First upper limits on the 21-cm signal power spectrum of neutral hydrogen atz= 9.16from the LOFAR 3C196 field,arXiv e-prints(2025) arXiv:2504.18534 [2504.18534]

    arXiv 2025

  21. [21]

    Acharya, F

    A. Acharya, F. Mertens, B. Ciardi, R. Ghara, L.V.E. Koopmans and S. Zaroubi,Revised LOFAR upper limits on the 21-cm signal power spectrum at z≈9.1 using machine learning and gaussian process regression,Monthly Notices of the Royal Astronomical Society534(2024) L30 [2408.10051]

    Pith/arXiv arXiv 2024

  22. [22]

    Mellema, L

    G. Mellema, L. Koopmans, H. Shukla, K.K. Datta, A. Mesinger and S. Majumdar,HI tomographic imaging of the Cosmic Dawn and Epoch of Reionization with SKA, inAdvancing Astrophysics with the Square Kilometre Array (AASKA14), p. 10, Apr., 2015, DOI [1501.04203]

    Pith/arXiv arXiv 2015

  23. [23]

    Schmit and J.R

    C.J. Schmit and J.R. Pritchard,Emulation of reionization simulations for Bayesian inference of astrophysics parameters using neural networks,Monthly Notices of the Royal Astronomical Society475(2018) 1213 [1708.00011]

    Pith/arXiv arXiv 2018

  24. [24]

    Jennings, C.A

    W.D. Jennings, C.A. Watkinson, F.B. Abdalla and J.D. McEwen,Evaluating machine learning techniques for predicting power spectra from reionization simulations,Monthly Notices of the Royal Astronomical Society483(2019) 2907 [1811.09141]

    Pith/arXiv arXiv 2019

  25. [25]

    Tiwari, A.K

    H. Tiwari, A.K. Shaw, S. Majumdar, M. Kamran and M. Choudhury,Improving constraints on the reionization parameters using 21-cm bispectrum,Journal of Cosmology and Astroparticle Physics2022(2022) 045 [2108.07279]

    Pith/arXiv arXiv 2022

  26. [26]

    Sikder, R

    S. Sikder, R. Barkana, I. Reis and A. Fialkov,Emulation of the cosmic dawn 21-cm power spectrum and classification of excess radio models using an artificial neural network,Monthly Notices of the Royal Astronomical Society527(2024) 9977 [2201.08205]

    Pith/arXiv arXiv 2024

  27. [27]

    Breitman, A

    D. Breitman, A. Mesinger, S.G. Murray, D. Prelogović, Y. Qin and R. Trotta,21CMEMU: an emulator of 21CMFAST summary observables,Monthly Notices of the Royal Astronomical Society527(2024) 9833 [2309.05697]

    Pith/arXiv arXiv 2024

  28. [28]

    Mahida, S.K

    Y. Mahida, S.K. Yadav, S. Majumdar, L. Noble, C. Shekhar Murmu, S. Dasgupta et al.,From ANN to BNN: Inferring Reionization Parameters using Uncertainty-aware Emulators of 21-cm Summaries,arXiv e-prints(2025) arXiv:2508.13261 [2508.13261]. – 32 –

    arXiv 2025

  29. [29]

    Gillet, A

    N. Gillet, A. Mesinger, B. Greig, A. Liu and G. Ucci,Deep learning from 21-cm tomography of the cosmic dawn and reionization,Monthly Notices of the Royal Astronomical Society484 (2019) 282 [1805.02699]

    Pith/arXiv arXiv 2019

  30. [30]

    Hortúa, L

    H.J. Hortúa, L. Malagò and R. Volpi,Constraining the reionization history using bayesian normalizing flows,Machine Learning: Science and Technology1(2020) 035014

    2020

  31. [31]

    Hassan, S

    S. Hassan, S. Andrianomena and C. Doughty,Constraining the astrophysics and cosmology from 21 cm tomography using deep learning with the SKA,Monthly Notices of the Royal Astronomical Society494(2020) 5761 [1907.07787]

    Pith/arXiv arXiv 2020

  32. [32]

    X. Zhao, Y. Mao, C. Cheng and B.D. Wandelt,Simulation-based Inference of Reionization Parameters from 3D Tomographic 21 cm Light-cone Images,The Astrophysical Journal926 (2022) 151 [2105.03344]

    Pith/arXiv arXiv 2022

  33. [33]

    Prelogović, A

    D. Prelogović, A. Mesinger, S. Murray, G. Fiameni and N. Gillet,Machine learning astrophysics from 21 cm lightcones: impact of network architectures and signal contamination, Monthly Notices of the Royal Astronomical Society509(2022) 3852 [2107.00018]

    Pith/arXiv arXiv 2022

  34. [34]

    Neutsch, C

    S. Neutsch, C. Heneka and M. Brüggen,Inferring astrophysics and dark matter properties from 21 cm tomography using deep learning,Monthly Notices of the Royal Astronomical Society511 (2022) 3446 [2201.07587]

    Pith/arXiv arXiv 2022

  35. [35]

    Schosser, C

    B. Schosser, C. Heneka and T. Plehn,Optimal, fast, and robust inference of reionization-era cosmology with the 21cmPIE-INN,SciPost Physics Core8(2025) 037 [2401.04174]

    Pith/arXiv arXiv 2025

  36. [36]

    A. Ore, C. Heneka and T. Plehn,SKATR: A self-supervised summary transformer for SKA, SciPost Physics18(2025) 155 [2410.18899]

    Pith/arXiv arXiv 2025

  37. [37]

    Chardin, G

    J. Chardin, G. Uhlrich, D. Aubert, N. Deparis, N. Gillet, P. Ocvirk et al.,A deep learning model to emulate simulations of cosmic reionization,Monthly Notices of the Royal Astronomical Society490(2019) 1055 [1905.06958]

    Pith/arXiv arXiv 2019

  38. [38]

    Korber, M

    D. Korber, M. Bianco, E. Tolley and J.-P. Kneib,PINION: physics-informed neural network for accelerating radiative transfer simulations for cosmic reionization,Monthly Notices of the Royal Astronomical Society521(2023) 902 [2208.13803]

    Pith/arXiv arXiv 2023

  39. [39]

    Zhao, Y.-S

    X. Zhao, Y.-S. Ting, K. Diao and Y. Mao,Can diffusion model conditionally generate astrophysical images?,Monthly Notices of the Royal Astronomical Society526(2023) 1699 [2307.09568]

    Pith/arXiv arXiv 2023

  40. [40]

    Bharadwaj and P.S

    S. Bharadwaj and P.S. Srikant,HI fluctuations at large redshifts: III — Simulating the signal expected at GMRT,Journal of Astrophysics and Astronomy25(2004) 67 [astro-ph/0402262]

    Pith/arXiv arXiv 2004

  41. [41]

    Mondal, S

    R. Mondal, S. Bharadwaj, S. Majumdar, A. Bera and A. Acharyya,The effect of non-Gaussianity on error predictions for the Epoch of Reionization (EoR) 21-cm power spectrum.,Monthly Notices of the Royal Astronomical Society449(2015) L41 [1409.4420]

    Pith/arXiv arXiv 2015

  42. [42]

    Davis, G

    M. Davis, G. Efstathiou, C.S. Frenk and S.D.M. White,The evolution of large-scale structure in a universe dominated by cold dark matter,The Astrophysical Journal292(1985) 371

    1985

  43. [43]

    Choudhury, M.G

    T.R. Choudhury, M.G. Haehnelt and J. Regan,Inside-out or outside-in: the topology of reionization in the photon-starved regime suggested by Lyαforest data,Monthly Notices of the Royal Astronomical Society394(2009) 960 [0806.1524]

    Pith/arXiv arXiv 2009

  44. [44]

    Majumdar, G

    S. Majumdar, G. Mellema, K.K. Datta, H. Jensen, T.R. Choudhury, S. Bharadwaj et al.,On the use of seminumerical simulations in predicting the 21-cm signal from the epoch of reionization,Monthly Notices of the Royal Astronomical Society443(2014) 2843 [1403.0941]

    Pith/arXiv arXiv 2014

  45. [45]

    Mondal, S

    R. Mondal, S. Bharadwaj and S. Majumdar,Statistics of the epoch of reionization (EoR) 21-cm signal - II. The evolution of the power-spectrum error-covariance,Monthly Notices of the Royal Astronomical Society464(2017) 2992 [1606.03874]. – 33 –

    Pith/arXiv arXiv 2017

  46. [46]

    Furlanetto, M

    S.R. Furlanetto, M. Zaldarriaga and L. Hernquist,The Growth of H II Regions During Reionization,The Astrophysical Journal613(2004) 1 [astro-ph/0403697]

    Pith/arXiv arXiv 2004

  47. [47]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez et al.,Attention Is All You Need,arXiv e-prints(2017) arXiv:1706.03762 [1706.03762]

    Pith/arXiv arXiv 2017

  48. [48]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner et al.,An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,arXiv e-prints (2020) arXiv:2010.11929 [2010.11929]

    Pith/arXiv arXiv 2020

  49. [49]

    Gündüç,Tensor-to-Image: Image-to-Image Translation with Vision Transformers,arXiv e-prints(2021) arXiv:2110.08037 [2110.08037]

    Y. Gündüç,Tensor-to-Image: Image-to-Image Translation with Vision Transformers,arXiv e-prints(2021) arXiv:2110.08037 [2110.08037]

    Pith/arXiv arXiv 2021

  50. [50]

    Ronneberger, P

    O. Ronneberger, P. Fischer and T. Brox,U-net: Convolutional networks for biomedical image segmentation, inMedical Image Computing and Computer-Assisted Intervention – MICCAI 2015, (Cham), pp. 234–241, Springer International Publishing, 2015

    2015

  51. [51]

    H. Xu, L. Xiang, H. Ye, D. Yao, P. Chu and B. Li,Permutation Equivariance of Transformers and Its Applications,arXiv e-prints(2023) arXiv:2304.07735 [2304.07735]

    Pith/arXiv arXiv 2023

  52. [52]

    Tsimenidis,Limitations of Deep Neural Networks: a discussion of G

    S. Tsimenidis,Limitations of Deep Neural Networks: a discussion of G. Marcus’ critical appraisal of deep learning,arXiv e-prints(2020) arXiv:2012.15754 [2012.15754]

    Pith/arXiv arXiv 2020

  53. [53]

    S.K. Giri, G. Mellema, K.L. Dixon and I.T. Iliev,Bubble size statistics during reionization from 21-cm tomography,Monthly Notices of the Royal Astronomical Society473(2018) 2949 [1706.00665]

    Pith/arXiv arXiv 2018

  54. [54]

    Mesinger and S

    A. Mesinger and S. Furlanetto,Efficient Simulations of Early Structure Formation and Reionization,The Astrophysical Journal669(2007) 663 [0704.0946]

    Pith/arXiv arXiv 2007

  55. [55]

    Lin, S.P

    Y. Lin, S.P. Oh, S.R. Furlanetto and P.M. Sutter,The distribution of bubble sizes during reionization,Monthly Notices of the Royal Astronomical Society461(2016) 3361 [1511.01506]

    Pith/arXiv arXiv 2016

  56. [56]

    T.-Y. Lu, C.A. Mason, A. Hutter, A. Mesinger, Y. Qin, D.P. Stark et al.,The reionizing bubble size distribution around galaxies,Monthly Notices of the Royal Astronomical Society528 (2024) 4872 [2304.11192]

    Pith/arXiv arXiv 2024

  57. [57]

    Bolton and M.G

    J.S. Bolton and M.G. Haehnelt,The nature and evolution of the highly ionized near-zones in the absorption spectra of z~= 6 quasars,Monthly Notices of the Royal Astronomical Society 374(2007) 493 [astro-ph/0607331]

    Pith/arXiv arXiv 2007

  58. [58]

    Mao,RESEARCH PAPER: Ionization state of cosmic hydrogen by early stars and quasars,Research in Astronomy and Astrophysics9(2009) 665

    X.-C. Mao,RESEARCH PAPER: Ionization state of cosmic hydrogen by early stars and quasars,Research in Astronomy and Astrophysics9(2009) 665

    2009

  59. [59]

    Bianco, S.K

    M. Bianco, S.K. Giri, I.T. Iliev and G. Mellema,Deep learning approach for identification of H II regions during reionization in 21-cm observations,Monthly Notices of the Royal Astronomical Society505(2021) 3982 [2102.06713]. – 34 –

    Pith/arXiv arXiv 2021