Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Multi-fidelity emulator for large-scale 21 cm lightcone images: a few-shot transfer learning approach with generative adversarial network

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read A GAN trained on 30,000 small reionization simulations plus 80 large ones generates 512 Mpc 21 cm lightcones whose small-scale statistics match the simulator at the percent level.

desk verdict A promising multi-fidelity GAN emulator whose headline accuracy is likely inflated by test-set checkpoint selection; the method deserves review, but the precision claims need re-validation. read the letter →

arxiv 2502.04246 v2 pith:3ROJYW3N submitted 2025-02-06 astro-ph.IM astro-ph.CO

classification astro-ph.IMastro-ph.CO
keywords epochofreionization21cmcosmologygenerativeadversarialnetworksfew-shottransferlearningmulti-fidelityemulationscatteringtransformcosmologicallightconespowerspectrum
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 is trying to establish that a GAN emulator can produce large-volume 512 Mpc 21 cm lightcone images of the epoch of reionization without paying the full simulation cost. The proposed recipe is multi-fidelity: first train a GAN on 30,000 cheap small-scale simulations, then adapt that pretrained model with only 80 large-scale simulations using few-shot transfer learning. The authors report that the resulting images match 21cmFAST at percentage-level precision on small scales, with errors rising to a few tens of percent on very large scales, while cutting the training-data cost by one to two orders of magnitude. If the claim holds, survey-sized mock catalogs for upcoming 21 cm experiments become computationally feasible.

What carries the argument

The carrier of the argument is the few-shot transfer-learning pipeline built on StyleGAN2. The source model is trained on abundant small-scale images; then the target model is created by duplicating the generator's Constant Input layer four times along the height axis, turning an output shape of $2\times64\times512$ into $2\times256\times512$. During fine-tuning, the first two layers of the discriminator are frozen on the assumption that small-scale statistics are identical across box sizes, an extra adversarial loss is added from the small-scale discriminator acting on cut patches of generated large-scale images, and a cross-domain correspondence loss compares cosine-similarity distributions of paired small- and large-scale GAN outputs to preserve intra-batch diversity. These components together let 80 large-scale simulations calibrate the large-scale modes while the small-scale modes remain anchored by the pretrained model.

What would settle it

Cut the 80 large-scale 512 Mpc lightcones into 128 Mpc patches and compare their 21 cm power spectra and scattering-transform coefficients against dedicated 128 Mpc simulations run at the same astrophysical parameters and redshifts; a systematic offset in the small-scale statistics beyond the sample variance of the two sets would show that the frozen small-scale layers are reproducing the wrong small-scale physics.

Watch

Extended reading notes

Core claim

The paper's central claim is that a two-stage multi-fidelity GAN can emulate large-scale 21 cm lightcones from the epoch of reionization with limited high-fidelity data. In the first stage, a StyleGAN2 generator and discriminator are trained on 120,000 small-scale images from 30,000 simulations of comoving boxes roughly 128 Mpc on a side. In the second stage, the generator's initial layer is expanded to emit 512 Mpc lightcone images, and the model is fine-tuned on only 320 images from 80 large-scale simulations, using a frozen early discriminator, patch-level discrimination, and a cross-domain correspondence loss to prevent mode collapse. The paper demonstrates that the transfer-learned large-scale GAN reproduces the global brightness temperature history, the 2D power spectrum, and scattering-transform coefficients of 21cmFAST at percent-level accuracy on small scales, with large-scale errors at the level of tens of percent, and that conventional training with large-scale simulations alone would require thousands of such simulations and one to two orders of magnitude more computing time.

Load-bearing premise

The transfer pipeline assumes that the small-scale statistics of 21 cm reionization inside the 512 Mpc boxes are identical to those in the 128 Mpc boxes, so the small-scale discriminator layers can be frozen and reused; if box-size sample variance or the different number of concatenated simulation realizations changes those small-scale statistics, the frozen layers will inject a bias that only 80 large-scale images are too few to correct.

Editorial extensions

If this is right

  • Epoch-of-reionization mock catalogs spanning hundreds of megaparsecs become affordable: the full training set costs about 11,400 CPU core hours, versus an estimated 150,000 to 900,000 CPU core hours for training only on large-scale simulations.
  • The transfer-learned large-scale GAN inherits small-scale fidelity from the pretrained model, keeping power-spectrum and scattering-transform errors at the percent level on small scales and only tens of percent on the largest scales ($k \lesssim 0.02\,\mathrm{Mpc}^{-1}$).
  • Retraining on the large-scale data removes the concatenation-boundary artifacts visible in the small-scale GAN output, an improvement that a GAN trained directly on only 80 large-scale simulations fails to achieve due to mode collapse.
  • The same two-step scheme is proposed for other fidelity splits beyond box size, such as low versus high resolution, or semi-numerical versus full radiative-transfer simulations.
  • The Fréchet Scattering Distance plateau analysis suggests that at least thousands of large-scale simulations would be needed to match this emulator by conventional training, making the transfer approach one to two orders of magnitude cheaper.

Reading between the lines

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

  • A natural extension the authors do not test is transferring from large-scale semi-numerical simulations to small-scale full radiative-transfer simulations, where the small-scale statistics differ substantially, so the number of high-fidelity calibration images would likely need to grow.
  • The frozen-layer assumption could be stress-tested by varying the large-scale box size, for example comparing transfer from 256 Mpc to 512 Mpc targets, to see how many high-fidelity images are required as the small-scale sample variance changes.
  • The reported tens-of-percent errors on very large scales set a practical floor for analyses that depend on those modes, so survey-focused studies using this emulator should down-weight or mask $k \lesssim 0.02\,\mathrm{Mpc}^{-1}$ scales.
  • Because the method only requires two statistically related image sets, it should transfer to other generative backbones such as diffusion models or flow matching, a direction the paper explicitly mentions as future work.
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

4 major / 6 minor

Summary. The paper presents a multi-fidelity emulation approach for large-scale 21 cm lightcone images during the epoch of reionization. A StyleGAN2 emulator is first trained on 120,000 image samples from 30,000 small-scale (128 Mpc) 21cmFAST simulations, and is then adapted to generate large-scale (512 Mpc) lightcone images via few-shot transfer learning using only 320 image samples from 80 large-scale simulations. The transfer uses a patchy-level discriminator, cross-domain correspondence, frozen low-level discriminator layers, and an extra small-scale adversarial loss. The authors validate the emulator on five held-out parameter sets with statistics including the global 21 cm brightness temperature, the 2D power spectrum, and scattering transform coefficients, and they report percentage-level precision on small scales with errors rising to a few tens of percent on very large scales. They further estimate that the multi-fidelity approach reduces the training-data computational cost by one to two orders of magnitude relative to training a large-scale GAN exclusively on large-scale simulations.

Significance. If the reported precision and cost savings hold, this is a valuable contribution to 21 cm cosmology: it would make large-volume EoR mock catalogs far more affordable, and it provides a concrete demonstration of multi-fidelity GAN transfer learning for astrophysical image emulation. The paper ships public code (two GitHub repositories) and includes several commendable validation steps: a large small-scale test set with 500 realizations, explicit mode-collapse diagnostics at the pixel and feature levels, and a baseline comparison showing that conventional GAN training with only 80 large-scale simulations fails badly. The central claim is plausible and the methodological framework is worth publishing, provided the quantitative precision and cost claims survive a cleaner validation protocol.

major comments (4)
  1. [Appendix A.4 and Section 5.5] The large-scale GAN checkpoint is selected using the test set, and the same test set is then used to report the accuracy that the abstract and Table 1 advertise. Appendix A.4 states that the FSD with j=2,4,6 is monitored every 200 iterations and that 1,400 iterations is 'optimal', while Section 5.5 explicitly says that the FSD is computed from scattering-transform coefficients evaluated at test-set reionization parameters. No separate validation split is described anywhere in the paper. Consequently, the checkpoint that produced Figures 9-12 and Table 1 was chosen as the one whose scattering statistics most closely match these specific test realizations, so the reported relative errors are selected rather than predicted, and the percentage-level precision claim in the abstract is not yet supported. Because FSD is a smooth metric over checkpoints, selecting its minimum on the evaluation set inflates apparent agreement. The authors should hold out a validation set (e.g., one parameter set or an independent set of realizations), select the checkpoint on that validation set, and only then quote test-set errors.
  2. [Section 5.5] The computational-savings estimate is based on a post hoc and not fully matched comparison. First, the claim that 80 large-scale simulations is the smallest number that keeps a comparable small-scale FSD is the result of testing {10,20,40,80} and then choosing 80 on the same test-set FSD used for the final evaluation; this is post hoc selection on the test set and should be replaced by a validation-based selection or reported as a search result with the associated uncertainty. Second, the grey dotted line in Figure 16 represents the large-scale FSD computed with ST coefficients j=2,4,6, whereas the blue baseline curve for the small-scale GAN uses j=0,2,4; these are different coefficient sets, so the horizontal comparison in Figure 16 and the factor-of-four scaling argument leading to the '4000-20000 large-scale simulations' estimate in Section 5.5 are not like-for-like. The one-to-two order-of-magnitude cost claim in the abstract and Table 2 therefore rests on an unsupported quantitative comparison.
  3. [Section 3.3 and Figures 9-12] The large-scale test set consists of only four realizations per parameter set (Section 3.3), yet the relative-error curves in Figures 9-12 and the entries in Table 1 are reported without error bars or any sampling-variance estimate. With four realizations, the test-set mean itself has substantial cosmic variance, especially on the largest scales where the paper quotes errors of 20-30%; therefore the direction and magnitude of the reported discrepancies are not established at the claimed precision. The authors should provide bootstrap or realization-based error bars on epsilon_rel, or increase the number of test realizations, and propagate this uncertainty into Table 1 and the abstract's precision statement.
  4. [Section 2.2] The transfer-learning pipeline assumes that 'the small-scale information in both training sets is identical', which justifies freezing the first two discriminator layers and adding the L_adv,s term in Eq. (9). However, the small-scale set (128 Mpc boxes, eight concatenated realizations per lightcone) and the large-scale set (512 Mpc boxes, two concatenated realizations per lightcone) differ in box size and lightcone construction; sample variance on 128 Mpc scales and stitching artifacts could make the small-scale statistical content of the two training sets differ. This assumption is never tested. The authors should directly compare small-scale statistics (e.g., the 2D power spectrum at k > 0.05 Mpc^-1 or ST coefficients with j=0,2) between the small-scale and large-scale training sets; if a non-negligible discrepancy exists, the frozen layers would inject a systematic bias that the 80 large-scale images are too few to correct.
minor comments (6)
  1. [Table 1] Table 1 reports only the first three parameter sets, although Section 3.3 states that five test sets are used and that the conclusions are based on all five; please include all five sets or explicitly justify the omission.
  2. [Section 4.1, Eq. (11)] The statement that 'a cutoff is performed when Stats_test is close to zero' should specify the exact threshold used, since the reported relative errors depend sensitively on this cutoff.
  3. [Figure 16] The caption and text should state explicitly that lower FSD is better, describe the normalization of the coefficients, and explain that the grey dashed and dotted lines are horizontal values from the trained large-scale GAN rather than curves.
  4. [Section 5.5] There is a typo in 'we tested {10,20,40,80} largen-scale simulations'; it should read 'large-scale'.
  5. [References] Several bibliographic entries are duplicated (Mesinger et al. 2011 appears twice, Murray et al. 2020 appears twice, and Kannan et al. 2021/2022 appears twice); these should be merged.
  6. [Introduction and Appendix A.4] There are encoding artifacts in the text, e.g., '21cmF AST' and 'Frech´ et Inception score', which should be cleaned up in the final version.

Circularity Check

1 steps flagged · score 5.0 of 10

Large-scale GAN checkpoint is selected by FSD on the test parameter sets, so the headline ST/FSD precision figures are partly selected rather than independently predicted.

  1. fitted input called prediction [Appendix A.4 and Section 5.5 (Eq. 13), used in Figures 11-12 and Table 1]
    "For the large-scale GAN, we adopt the FSD with j = 2, 4, 6 to capture the performance on large scales and monitor the FSD after every 200 iterations. We found that at 1,400 iterations it gives the optimal performance with 80 large-scale training simulations. ... For this baseline analysis (small-scale GAN), we use ST coefficients with scales j = 0, 2, 4, computed for three distinct patches along the redshift axis using test set reionization parameters, consistent with the coefficients presented in Figures 6 and 7."

    The early-stopping checkpoint of the large-scale GAN is selected as the iteration minimizing FSD computed from scattering coefficients at the five test parameter sets, and the same test sets and ST/FSD statistics are then reported as validation in Figures 11-12, Table 1, and Section 5.5. No separate validation split is described. Choosing the minimum of a smooth metric on the evaluation set forces the reported FSD and ST relative errors below what a typical held-out checkpoint would give, so the small-scale percentage-level ST agreement and the large-scale ~20% ST error are partly selected rather than predicted. Global-signal, power-spectrum, and mode-collapse checks remain independent, so the circularity is partial.

full rationale

The emulator is fitted to the 21cmFAST simulation distribution by construction, which is the normal meaning of an emulator and not itself circular; the defensible non-circular claim is generalization to held-out parameter sets, and the paper does test this with five test parameter sets not used for gradient training. The main circularity is that the large-scale GAN's final iteration was chosen by monitoring FSD during training (Appendix A.4), and the FSD is computed from scattering coefficients at the same test-set reionization parameters used later for the reported ST errors and FSD values (Section 5.5). This makes the ST/FSD precision figures selected rather than independently predicted, and the post hoc choice of 80 large-scale simulations from {10,20,40,80} further tailors the cost-savings claim. However, the transfer-learning construction, the small-scale GAN validation on 500 independent realizations, the power-spectrum and global-signal comparisons, and the mode-collapse diagnostics retain independent content, and the paper's self-citations (Diao & Mao 2023; FSD from Zhao et al. 2023) are not load-bearing uniqueness arguments. Score 5 reflects partial circularity in the headline large-scale precision claim rather than a fully self-referential derivation.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no physical entities. Its free parameters are ML design choices, the most consequential being the 80-simulation training set size and the 1,400-iteration checkpoint, both selected post hoc. The central scientific assumption is that small-scale statistics transfer from 128 Mpc to 512 Mpc boxes, which is plausible for a homogeneous reionization model but not independently verified.

free parameters (4)
  • Number of large-scale training simulations = 80
    Selected post hoc as the smallest number (among 10, 20, 40, 80) that keeps the small-scale FSD comparable to the small-scale GAN (Section 5.5).
  • Large-scale GAN training iterations = 1,400
    Chosen as the iteration with optimal FSD during training (Appendix A.4).
  • Relative error cutoff near zero = not specified
    A cutoff on Stats_test is applied to avoid division by near-zero denominators, but the threshold is not given (Section 4.1).
  • Anchor region radius = not specified
    The anchor region in parameter space for the patchy discriminator is described as 'a small radius' but its value is not reported (Section 2.2).
assumptions (4)
  • domain assumption 21cmFAST semi-numerical simulation is an adequate model of the EoR 21 cm signal
    The emulator aims to reproduce 21cmFAST output, so 21cmFAST's accuracy as a physical model is assumed (Section 3.1).
  • domain assumption Small-scale statistics are identical between small-scale and large-scale simulation boxes
    This justifies freezing the first two discriminator layers and applying transfer learning (Section 2.2: 'the small-scale information in both training sets is identical').
  • ad hoc to paper The five test parameter sets are representative of the full parameter space
    Test coverage is limited to five LHS-style points; the paper claims generic conclusions from these (Section 3.3).
  • standard math FSD is a valid metric for GAN convergence and model selection
    Used for monitoring training and selecting checkpoints (Appendix A.4); this is a standard but heuristic ML practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-fidelity emulator for large-scale 21 cm lightcone images: a few-shot transfer learning approach with generative adversarial network." pith.science (2026). https://pith.science/paper/3ROJYW3N

@misc{pith2026250204246,
  author       = {Pith},
  title        = {Pith review of: Multi-fidelity emulator for large-scale 21 cm lightcone images: a few-shot transfer learning approach with generative adversarial network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3ROJYW3N}},
  note         = {Machine review of arXiv:2502.04246}
}
read the original abstract

Emulators using machine learning techniques have emerged to efficiently generate mock data matching the large survey volume for upcoming experiments, as an alternative approach to large-scale numerical simulations. However, high-fidelity emulators have become computationally expensive as the simulation volume grows to hundreds of megaparsecs. Here, we present a {\it multi-fidelity} emulation of large-scale 21~cm lightcone images from the epoch of reionization, which is realized by applying the {\it few-shot transfer learning} to training generative adversarial networks (GAN) from small-scale to large-scale simulations. Specifically, a GAN emulator is first trained with a huge number of small-scale simulations, and then transfer-learned with only a limited number of large-scale simulations, to emulate large-scale 21~cm lightcone images. We test the precision of our transfer-learned GAN emulator in terms of representative statistics including global 21~cm brightness temperature history, 2D power spectrum, and scattering transform coefficients. We demonstrate that the lightcone images generated by the transfer-learned GAN emulator can reach the percentage level precision in most cases on small scales, and the error on large scales only increases mildly to the level of a few tens of per cent. Nevertheless, our multi-fidelity emulation technique saves a significant portion of computational resources that are mostly consumed for generating training samples for GAN. On estimate, the computational resource by training GAN completely with large-scale simulations would be one to two orders of magnitude larger than using our multi-fidelity technique. This implies that our technique allows for emulating high-fidelity, traditionally computationally prohibitive, images in an economic manner.

Figures

Figures reproduced from arXiv: 2502.04246 by the authors.

Figure 1
Figure 1. An illustration of the StyleGAN2 generator architecture. Our generator consists of a mapping network f, which modifies the convolution kernel according to astrophysical parameters and random vectors, and a synthesis network g, which generate images progressively, with noise injection for multiple times. commonly used ResNet (He et al. 2015) architecture. Our generator consists of two parts. First, a mapping network … view at source ↗
Figure 2
Figure 2. An illustration of CDC. We generate a set of samples with both small-scale GAN and large-scale GAN, calculate the similarity between each image pair generated by the same GAN, normalize the similarity vector of each GAN with softmax, and compute the KL-divergence as the CDC. Given a well-behaved small-scale StyleGAN2 emula￾tor, the network structure should be modified to enable the generation of large-scale images, … view at source ↗
Figure 3
Figure 3. Examples of the emulated images using the small-scale GAN (left), in comparison with the simulated images using 21cmFAST in the test set (right). In each panel, we show the 21 cm brightness temperature (Tb) field (the upper half) and the matter overdensity (δm) field (the lower half). The LoS is along the x-axis. evolution is clearly visible in the GAN samples, and the density field shows a clearer web structure as … view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: (Top) the global 21 cm signal emulated with the small-scale GAN, with the mean (dashed line) and 2σ scatter (shallow shaded region). We show the results with different values of reionization parameters (in different colors). Each set of reionization parameter is calcul…
Figure 5
Figure 5. Figure 5: The 2D power spectrum of the 21 cm signal emulated with the small-scale GAN with the mean (dashed line) and 2σ scatter (shallow shaded region). We show the results with different values of reionization parameters (in different colors). Each set of reionization paramete…
Figure 6
Figure 6. Figure 6: The ST coefficient S1(j, l) of the 21 cm signal emulated with the small-scale GAN (dashed line) and that of the test set images (solid line) at three representative red￾shifts of the center slice (from top to bottom) z = 7.933, 9.384, and 11.221, respectively. We show …
Figure 8
Figure 8. Figure 8: Same as [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Same as [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Same as [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Same as [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 13
Figure 13. Figure 13: Visualization of the emulated images using the large-scale GAN (top), in comparison with the simulated images using 21cmFAST (bottom). Here we show four different realizations with the same parameters (log10 ζ, log10 Tvir) = (1.7, 5.0). Each realization was computed u…
Figure 14
Figure 14. Figure 14: The standard deviation of the 21 cm brightness temperature map for each pixel over 1,024 image samples of the large-scale GAN (top), in comparison with the simulated images using 21cmFAST (bottom) [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: The 2σ scatter of the ST coefficient S2(j1, l1, j2, l2) of the 21 cm signal emulated with the large-scale GAN (dashed line for the mean and error bar for 2σ scatter) and that of the simulation test set images (solid line for the mean and shaded region for 2σ scatter) …
Figure 16
Figure 16. Figure 16: FSD for the small-scale GAN as a function of the number of training simulations (blue solid line). For comparison, the grey dashed line shows the FSD calculated using small-scale ST coefficients (j = 0, 2, 4) for our trained large-scale GAN, while the grey dotted line…
Figure 17
Figure 17. Figure 17: An illustration of a ResNet block. A. DETAILS OF GAN ARCHITECTURE AND TRAINING CONFIGURATIONS In this appendix, we present the network structure in detail. A.1. Generator For our small-scale GAN, as described in Section 2, the generator consists of a mapping network f…
Figure 18
Figure 18. Figure 18: FSD with j = 0, 2, 4 for the small-scale GAN with 10,000 simulations as a function of the training itera￾tions. at 40,000 iterations it reaches best performance. For the large-scale GAN, we adopt the FSD with j = 2, 4, 6 to capture the performance on large scales and …
Figure 19
Figure 19. Figure 19: Same as [PITH_FULL_IMAGE:figures/full_fig_p021_19.png]
Figure 20
Figure 20. Figure 20: Same as [PITH_FULL_IMAGE:figures/full_fig_p021_20.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Small-scale Lyman alpha forest cosmology with PRIYA: Constraints from XQ100 and KODIAQ-SQUAD one-dimensional flux power spectra

    astro-ph.CO 2025-09 conditional novelty 6.0 of 10

    The XQ100 Lyman-alpha forest power spectrum, modeled with the PRIYA simulations, gives cosmological constraints consistent with eBOSS and Planck and constrains the IGM thermal history without external temperature data...

Reference graph

Works this paper leans on

87 extracted references · 14 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter doi edition editor eprint howpublished institution journal key month number organization pages publisher school series title misctitle type volume year version url label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts ...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION format.url url empty "" new.block "" url * "" * if FUNCTION format.eprint eprint empty "" archivePrefix empty "" archivePrefix "arXiv" = new.block " " eprint * " " * new.block " " eprint * " " * if if if FUNCTION format.doi doi empty "" " " doi * " " * if FUNCTION format.pid doi empty eprint empty ur...

  3. [3]

    Jk`0]r,Y r Yx> ie! /ʇ BI g s ZsS 1O;uyu S6p e Lc⢇GX ]CӲ sЪ DXU t5&VGɻ7&VM

    thebibliography [1] 20pt to REFERENCES 6pt =0pt -12pt 10pt plus 3pt =0pt =0pt =1pt plus 1pt =0pt =0pt -12pt =13pt plus 1pt =20pt =13pt plus 1pt \@M =10000 =-1.0em =0pt =0pt 0pt =0pt =1.0em @enumiv\@empty 10000 10000 `\.\@m \@noitemerr \@latex@warning Empty `thebibliography' environment \@ifnextchar \@reference \@latexerr Missing key on reference command E...

  4. [4]

    2020, , 641, A1, 10.1051/0004-6361/201833880

    Aghanim, N., Akrami, Y., Arroja, F., et al. 2020, , 641, A1, 10.1051/0004-6361/201833880

  5. [5]

    2019, , 629, A115, 10.1051/0004-6361/201834975

    Allys , E., Levrier , F., Zhang , S., et al. 2019, , 629, A115, 10.1051/0004-6361/201834975

  6. [6]

    2019, , 10.1093/mnras/stz1960

    Alsing, J., Charnock, T., Feeney, S., & Wandelt, B. 2019, , 10.1093/mnras/stz1960

  7. [7]

    2020, Journal of Machine Learning Research, 21, 1

    Andreux, M., Angles, T., Exarchakis, G., et al. 2020, Journal of Machine Learning Research, 21, 1. http://jmlr.org/papers/v21/19-047.html

  8. [8]

    2022, arXiv e-prints, arXiv:2211.05000

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

Show all 87 references
  1. [9]

    2024, in 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS '24) (ACM), 10.1145/3620665.3640366

    Ansel, J., Yang, E., He, H., et al. 2024, in 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS '24) (ACM), 10.1145/3620665.3640366

  2. [10]

    Bevins , H. T. J., Fialkov , A., de Lera Acedo , E., et al. 2022, Nature Astronomy, 6, 1473, 10.1038/s41550-022-01825-6

  3. [11]

    D., Rogers, A

    Bowman, J. D., Rogers, A. E. E., Monsalve, R. A., Mozdzen, T. J., & Mahesh, N. 2018, Nature, 555, 67, 10.1038/nature25792

  4. [12]

    G., et al

    Breitman , D., Mesinger , A., Murray , S. G., et al. 2024, , 527, 9833, 10.1093/mnras/stad3849

  5. [13]

    2019, , 885, 23, 10.3847/1538-4357/ab43e6

    Chen, Z., Xu, Y., Wang, Y., & Chen, X. 2019, , 885, 23, 10.3847/1538-4357/ab43e6

  6. [14]

    2021, arXiv e-prints, arXiv:2112.01288

    Cheng , S., & M \'e nard , B. 2021, arXiv e-prints, arXiv:2112.01288. 2112.01288

  7. [15]

    2020, , 499, 5902, 10.1093/mnras/staa3165

    Cheng, S., Ting, Y.-S., M \' e nard, B., & Bruna, J. 2020, , 499, 5902, 10.1093/mnras/staa3165

  8. [16]

    2024, arXiv e-prints, arXiv:2407.03523, 10.48550/arXiv.2407.03523

    Choudhury , M., Ghara , R., Zaroubi , S., et al. 2024, arXiv e-prints, arXiv:2407.03523, 10.48550/arXiv.2407.03523

  9. [17]

    R., Parsons, A

    DeBoer, D. R., Parsons, A. R., Aguirre, J. E., et al. 2017, Publications of the Astronomical Society of the Pacific, 129, 045001, 10.1088/1538-3873/129/974/045001

  10. [18]

    2023, in Fortieth ICML Machine Learning for Astrophysics workshop, 12

    Diao , K., & Mao , Y. 2023, in Fortieth ICML Machine Learning for Astrophysics workshop, 12

  11. [19]

    D., et al

    D'Odorico , V., Ba \ n ados , E., Becker , G. D., et al. 2023, , 523, 1399, 10.1093/mnras/stad1468

  12. [20]

    M., van der Horst , A

    Fausey , H. M., van der Horst , A. J., Tanvir , N. R., et al. 2024, arXiv e-prints, arXiv:2412.09732, 10.48550/arXiv.2412.09732

  13. [21]

    M., Mellema , G., Iliev , I

    Friedrich , M. M., Mellema , G., Iliev , I. T., & Shapiro , P. R. 2012, , 421, 2232, 10.1111/j.1365-2966.2012.20449.x

  14. [22]

    R., & Oh, S

    Furlanetto, S. R., & Oh, S. P. 2016, , 457, 1813, 10.1093/mnras/stw104

  15. [23]

    R., Oh, S

    Furlanetto, S. R., Oh, S. P., & Briggs, F. H. 2006, Physics Reports, 433, 181, 10.1016/j.physrep.2006.08.002

  16. [24]

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

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

  17. [25]

    2019, , 484, 282, 10.1093/mnras/stz010

    Gillet, N., Mesinger, A., Greig, B., Liu, A., & Ucci, G. 2019, , 484, 282, 10.1093/mnras/stz010

  18. [26]

    Gnedin , N. Y. 2014, , 793, 29, 10.1088/0004-637X/793/1/29

  19. [27]

    2021, in APS Meeting Abstracts, Vol

    Gonzalez Morales , A., & Dark Energy Spectroscopic Instrument Collaboration . 2021, in APS Meeting Abstracts, Vol. 2021, APS April Meeting Abstracts, Z08.001

  20. [28]

    J., Pouget-Abadie , J., Mirza , M., et al

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

  21. [29]

    2017, Proceedings of the International Astronomical Union, 12, 18, 10.1017/s1743921317011103

    Greig, B., & Mesinger, A. 2017, Proceedings of the International Astronomical Union, 12, 18, 10.1017/s1743921317011103

  22. [30]

    Greig , B., Ting , Y.-S., & Kaurov , A. A. 2022, arXiv e-prints, arXiv:2207.09082. 2207.09082

  23. [31]

    R., Millman , K

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

  24. [32]

    2020, , 494, 5761, 10.1093/mnras/staa1151

    Hassan, S., Andrianomena, S., & Doughty, C. 2020, , 494, 5761, 10.1093/mnras/staa1151

  25. [33]

    2017, Proceedings of the International Astronomical Union, 12, 47–51, 10.1017/S1743921317010833

    Hassan, S., Liu, A., Kohn, S., et al. 2017, Proceedings of the International Astronomical Union, 12, 47–51, 10.1017/S1743921317010833

  26. [34]

    Hassan, S., Liu, A., Kohn, S., & Plante, P. L. 2018, , 10.1093/mnras/sty3282

  27. [35]

    2022, , 937, 83, 10.3847/1538-4357/ac8b09

    Hassan, S., Villaescusa-Navarro, F., Wandelt, B., et al. 2022, , 937, 83, 10.3847/1538-4357/ac8b09

  28. [36]

    2015, Deep Residual Learning for Image Recognition, arXiv, 10.48550/ARXIV.1512.03385

    He, K., Zhang, X., Ren, S., & Sun, J. 2015, Deep Residual Learning for Image Recognition, arXiv, 10.48550/ARXIV.1512.03385

  29. [37]

    K., et al

    Hirling , P., Bianco , M., Giri , S. K., et al. 2023, arXiv e-prints, arXiv:2311.01492, 10.48550/arXiv.2311.01492

  30. [38]

    2020, arXiv e-prints, arXiv:2006.11239, 10.48550/arXiv.2006.11239

    Ho , J., Jain , A., & Abbeel , P. 2020, arXiv e-prints, arXiv:2006.11239, 10.48550/arXiv.2006.11239

  31. [39]

    Ho, M.-F., Bird, S., & Shelton, C. R. 2021, , 509, 2551, 10.1093/mnras/stab3114

  32. [40]

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

  33. [41]

    2023, , 958, 21, 10.3847/1538-4357/acfcb5

    Jacobus , C., Harrington , P., & Luki \'c , Z. 2023, , 958, 21, 10.3847/1538-4357/acfcb5

  34. [42]

    2021, Experimental Astronomy, 51, 193, 10.1007/s10686-020-09697-2

    Jishnu Nambissan , T., Subrahmanyan , R., Somashekar , R., et al. 2021, Experimental Astronomy, 51, 193, 10.1007/s10686-020-09697-2

  35. [44]

    2022, , 511, 4005, 10.1093/mnras/stab3710

    Kannan , R., Garaldi , E., Smith , A., et al. 2022, , 511, 4005, 10.1093/mnras/stab3710

  36. [45]

    2020, arXiv e-prints, arXiv:2006.06676, 10.48550/arXiv.2006.06676

    Karras , T., Aittala , M., Hellsten , J., et al. 2020, arXiv e-prints, arXiv:2006.06676, 10.48550/arXiv.2006.06676

  37. [46]

    2020, in Proc

    Karras, T., Laine, S., Aittala, M., et al. 2020, in Proc. CVPR

  38. [47]

    2000, Biometrika, 87, 1, 10.1093/biomet/87.1.1

    Kennedy, M., & O'Hagan, A. 2000, Biometrika, 87, 1, 10.1093/biomet/87.1.1

  39. [48]

    2015, in Proceedings of Advancing Astrophysics with the Square Kilometre Array PoS ( AASKA 14) (Sissa Medialab), 10.22323/1.215.0001

    Koopmans, L., Pritchard, J., Mellema, G., et al. 2015, in Proceedings of Advancing Astrophysics with the Square Kilometre Array PoS ( AASKA 14) (Sissa Medialab), 10.22323/1.215.0001

  40. [49]

    2022, A very early onset of massive galaxy formation, arXiv, 10.48550/ARXIV.2207.12446

    Labbe, I., van Dokkum, P., Nelson, E., et al. 2022, A very early onset of massive galaxy formation, arXiv, 10.48550/ARXIV.2207.12446

  41. [50]

    Li , Y., Ni , Y., Croft , R. A. C., et al. 2021, Proceedings of the National Academy of Science, 118, e2022038118, 10.1073/pnas.2022038118

  42. [51]

    Lipman , Y., Chen , R. T. Q., Ben-Hamu , H., Nickel , M., & Le , M. 2022, arXiv e-prints, arXiv:2210.02747, 10.48550/arXiv.2210.02747

  43. [52]

    List , F., Bhat , I., & Lewis , G. F. 2019, , 490, 3134, 10.1093/mnras/stz2759

  44. [53]

    List , F., & Lewis , G. F. 2020, , 493, 5913, 10.1093/mnras/staa523

  45. [54]

    2012, Communications on Pure and Applied Mathematics, 65, 1331, https://doi.org/10.1002/cpa.21413

    Mallat, S. 2012, Communications on Pure and Applied Mathematics, 65, 1331, https://doi.org/10.1002/cpa.21413

  46. [55]

    D., Beckman, R

    McKay, M. D., Beckman, R. J., & Conover, W. J. 2000, Technometrics, 42, 55

  47. [56]

    Mellema, G., Koopmans, L. V. E., Abdalla, F. A., et al. 2013, Experimental Astronomy, 36, 235, 10.1007/s10686-013-9334-5

  48. [57]

    2024, , 683, A24, 10.1051/0004-6361/202347591

    Meriot , R., & Semelin , B. 2024, , 683, A24, 10.1051/0004-6361/202347591

  49. [58]

    2018, Which Training Methods for GANs do actually Converge?, arXiv, 10.48550/ARXIV.1801.04406

    Mescheder, L., Geiger, A., & Nowozin, S. 2018, Which Training Methods for GANs do actually Converge?, arXiv, 10.48550/ARXIV.1801.04406

  50. [60]

    2011, , 411, 955, 10.1111/j.1365-2966.2010.17731.x

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

  51. [61]

    2020, Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs, arXiv, 10.48550/ARXIV.2002.10964

    Mo, S., Cho, M., & Shin, J. 2020, Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs, arXiv, 10.48550/ARXIV.2002.10964

  52. [62]

    2024, arXiv e-prints, arXiv:2405.13680, 10.48550/arXiv.2405.13680

    Montero-Camacho , P., Li , Y., & Cranmer , M. 2024, arXiv e-prints, arXiv:2405.13680, 10.48550/arXiv.2405.13680

  53. [63]

    F., & Wyithe, J

    Morales, M. F., & Wyithe, J. S. B. 2010, Annual Review of , 48, 127, 10.1146/annurev-astro-081309-130936

  54. [65]

    G., Greig, B., Mesinger, A., et al

    Murray, S. G., Greig, B., Mesinger, A., et al. 2020, Journal of Open Source Software, 5, 2582, 10.21105/joss.02582

  55. [66]

    P., Oesch, P

    Naidu, R. P., Oesch, P. A., Setton, D. J., et al. 2022, Schrodinger's Galaxy Candidate: Puzzlingly Luminous at z 17 , or Dusty/Quenched at z 5 ?, arXiv, 10.48550/ARXIV.2208.02794

  56. [67]

    2021, , 507, 1021, 10.1093/mnras/stab2113

    Ni , Y., Li , Y., Lachance , P., et al. 2021, , 507, 1021, 10.1093/mnras/stab2113

  57. [68]

    2021, arXiv e-prints, arXiv:2104.06820

    Ojha , U., Li , Y., Lu , J., et al. 2021, arXiv e-prints, arXiv:2104.06820. 2104.06820

  58. [69]

    R., Liu, A., Aguirre, J

    Parsons, A. R., Liu, A., Aguirre, J. E., et al. 2014, , 788, 106, 10.1088/0004-637x/788/2/106

  59. [70]

    R., & Loeb, A

    Pritchard, J. R., & Loeb, A. 2012, Reports on Progress in Physics, 75, 086901, 10.1088/0034-4885/75/8/086901

  60. [71]

    1998, , 299, 1097, 10.1046/j.1365-8711.1998.01845.x

    Scoccimarro, R. 1998, , 299, 1097, 10.1046/j.1365-8711.1998.01845.x

  61. [72]

    2017, , 468, 3869, 10.1093/mnras/stx734

    Shimabukuro, H., & Semelin, B. 2017, , 468, 3869, 10.1093/mnras/stx734

  62. [73]

    2024, , 527, 9977, 10.1093/mnras/stad3699

    Sikder , S., Barkana , R., Reis , I., & Fialkov , A. 2024, , 527, 9977, 10.1093/mnras/stad3699

  63. [74]

    P., et al

    Song , Y., Sohl-Dickstein , J., Kingma , D. P., et al. 2020, arXiv e-prints, arXiv:2011.13456, 10.48550/arXiv.2011.13456

  64. [75]

    J., Pandey , S., et al

    Sui , C., Bartlett , D. J., Pandey , S., et al. 2024, arXiv e-prints, arXiv:2410.14623, 10.48550/arXiv.2410.14623

  65. [76]

    J., Goeke, R., Bowman, J

    Tingay, S. J., Goeke, R., Bowman, J. D., et al. 2013, , 30, e007, 10.1017/pasa.2012.007

  66. [77]

    Tr \"o ster , T., Ferguson , C., Harnois-D \'e raps , J., & McCarthy , I. G. 2019, , 487, L24, 10.1093/mnrasl/slz075

  67. [78]

    P., Wise, M

    van Haarlem, M. P., Wise, M. W., Gunst, A. W., et al. 2013, , 556, A2, 10.1051/0004-6361/201220873

  68. [79]

    2022, arXiv e-prints, arXiv:2201.01300

    Villaescusa-Navarro , F., Genel , S., Angl \'e s-Alc \'a zar , D., et al. 2022, arXiv e-prints, arXiv:2201.01300. 2201.01300

  69. [80]

    H., & Cen, R

    Wise, J. H., & Cen, R. 2009, , 693, 984

  70. [81]

    2025, arXiv e-prints, arXiv:2501.06296, 10.48550/arXiv.2501.06296

    Yang , Y., Bird , S., & Ho , M.-F. 2025, arXiv e-prints, arXiv:2501.06296, 10.48550/arXiv.2501.06296

  71. [82]

    Yiu , T. W. H., Fluri , J., & Kacprzak , T. 2022, , 2022, 013, 10.1088/1475-7516/2022/12/013

  72. [83]

    2021, , 506, 357, 10.1093/mnras/stab1718

    Yoshiura , S., Shimabukuro , H., Hasegawa , K., & Takahashi , K. 2021, , 506, 357, 10.1093/mnras/stab1718

  73. [84]

    2025, The Open Journal of Astrophysics, 8, E13, 10.33232/001c.129471

    Zhang , X., Lachance , P., Dasgupta , A., et al. 2025, The Open Journal of Astrophysics, 8, E13, 10.33232/001c.129471

  74. [85]

    2024, , 528, 281, 10.1093/mnras/stad3940

    Zhang , X., Lachance , P., Ni , Y., et al. 2024, , 528, 281, 10.1093/mnras/stad3940

  75. [86]

    Zhao, X., Mao, Y., Cheng, C., & Wandelt, B. D. 2022 a , , 926, 151, 10.3847/1538-4357/ac457d

  76. [87]

    Zhao, X., Mao, Y., & Wandelt, B. D. 2022 b , , 933, 236, 10.3847/1538-4357/ac778e

  77. [88]

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

  78. [89]

    2023, , 526, 1699, 10.1093/mnras/stad2778

    Zhao , X., Ting , Y.-S., Diao , K., & Mao , Y. 2023, , 526, 1699, 10.1093/mnras/stad2778

  79. [90]

    D., Bosman , S

    Zhu , Y., Becker , G. D., Bosman , S. E. I., et al. 2021, , 923, 223, 10.3847/1538-4357/ac26c2

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.