Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Supervised Extraction of the Thermal Sunyaev$-$Zel'dovich Effect with a Three-Dimensional Convolutional Neural Network

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A supervised 3D convolutional network trained on simulated signals injected into Planck maps can extract the thermal Sunyaev–Zel'dovich effect as reliably as the standard NILC method, in both simulated tests and real cluster measurements.

desk verdict Competent, honest ML alternative to NILC for tSZ extraction, but the shared-sky train/test setup weakens the headline parity claim more than the authors concede. read the letter →

arxiv 2507.13400 v1 pith:JSYBPRG2 submitted 2025-07-16 astro-ph.IM astro-ph.CO

Cameron T. Pratt , Zhijie Qu , Joel N. Bregman This is my paper · ORCID
classification astro-ph.IMastro-ph.CO
keywords Sunyaev-ZeldovicheffectcosmicmicrowavebackgroundcomponentseparationconvolutionalneuralnetworkcurriculumlearningPlanckNILCmissingbaryons
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

The paper aims to establish that a fully supervised neural network can stand in for semi-blind component-separation methods when mapping the thermal Sunyaev–Zel'dovich (SZ) effect, the faint distortion CMB photons acquire by scattering off hot gas. The authors train a three-dimensional Attention Nested U-Net, called SZU, to turn nine Planck frequency maps into a Compton-$y$ map, using synthetic SZ signals from simulations injected into real Planck data as labeled examples. They report that SZU's reconstructed maps are comparable to those of the widely used needlet internal linear combination (NILC) method on simulated sky and on real clusters from the PSZ2 catalog. If this holds, SZ mapping gains a data-driven pipeline that does not rest on analytic assumptions about foreground correlations, advancing the search for the diffuse gas that may account for the missing baryons.

What carries the argument

The load-bearing object is SZU, a three-dimensional Attention Nested U-Net: an encoder-decoder convolutional architecture with nested skip connections and attention gates, extended so the convolution blocks operate on 9$\times$128$\times$128 multispectral cubes while only the spatial dimensions are downsampled. It is trained end-to-end with a masked mean-absolute-error loss that keeps only pixels above a signal threshold $\tilde{y}_{\rm sig}(e) = \tilde{y}_0 - 2e$, so the network first learns the strongest, sparsest SZ signals and then progressively weaker ones. The training labels come from synthetic SZ maps injected into real Planck data, which lets the model learn foreground behavior from the true sky rather than from simulated emission laws. The same nine frequency channels feed NILC, making the comparison a direct test of supervised versus analytic component separation on identical inputs.

What would settle it

Take the public SZU model and run it on an independent validation set whose simulated SZ signals include a CIB component correlated with cluster dust, then compare reconstructed y-maps with the injected signals; if bias grows with CIB intensity, the missing correlation is the deciding flaw, and if it does not, the limitation is benign in practice.

Watch

Extended reading notes

Core claim

The central claim is that end-to-end supervised learning can extract the thermal SZ signal from multifrequency Planck observations at a quality comparable to NILC. Synthetic SZ maps from cosmological simulations are converted to frequency-dependent signals, convolved with Planck beams, and superimposed on the real Planck frequency maps; the network learns the inverse mapping from those nine channels to the smoothed $y$ map. Curriculum learning, which starts the loss on strong signals and gradually admits weaker ones, markedly reduces bias for intermediate and strong signals. On a held-out simulated comparison set, SZU and NILC produce nearly identical binned pixel statistics, with differences of only a few percent in mean squared error, and their integrated $Y_{\rm SZ}$ measurements from PSZ2 clusters follow a close 1:1 relation. The paper's stated interpretation is that SZU is a viable supervised alternative to NILC for extracting thermal SZ y-maps from Planck-like data.

Load-bearing premise

The load-bearing premise is that the simulated SZ maps used as training labels are faithful enough to the real sky; in particular, they omit the known spatial and spectral correlation between the SZ signal and the cosmic infrared background from cluster dust, so the learned mapping could be biased on real data even though the paper reports NILC-comparable performance.

Editorial extensions

If this is right

  • SZU offers a supervised pipeline that does not require NILC's choices of wavelet basis and spatial window function, the two hyperparameters the analytic method depends on.
  • Curriculum learning should be preferred when the scientific goal is unbiased intermediate and strong SZ signals, while models trained without it give slightly lower total error on weak signals.
  • For stacking many weak SZ signals, the bias-variance trade-off matters: once stacking suppresses statistical noise, a low-bias extraction such as the curriculum-trained SZU becomes important for accurate results.
  • SZU and NILC handle foregrounds in a similar way on the PSZ2 sample, since their residuals show no trend with Galactic latitude.
  • Because SZU uses only the nine Planck channels, its performance should improve with future surveys offering higher spectral sampling, which the paper identifies as the key to better component separation.

Reading between the lines

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

  • A natural extension the paper leaves implicit is to train SZU on simulations that include the SZ-CIB correlation; the authors' own limitation note predicts this would change the learned mapping, and measuring that change would quantify how much real dusty clusters bias the current model.
  • SZU's strategy of injecting signals into a real-sky background effectively teaches it to separate signal from the specific foregrounds present in Planck; the same architecture could be retrained on next-generation frequency maps without algorithmic change, a cheaper route to improved y-maps than refining analytic component separation.
  • The masked cluster regions, where predictions would be most scientifically useful, are exactly where the model was never trained; testing SZU on independent synthetic skies with clusters present would reveal whether the ambient-region assumption is safe.
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 / 5 minor

Summary. The paper presents SZU, a three-dimensional Attention Nested U-Net for extracting thermal Sunyaev-Zel'dovich (SZ) maps from Planck PR4 frequency maps. The network is trained end-to-end with supervised learning on simulated full-sky y-maps from Han et al. (2021) injected into real Planck maps, using curriculum learning, pixel masking above a signal threshold, and masking of known galaxy clusters. The authors compare SZU with NILC on 2100 simulated validation patches and on PSZ2 cluster fluxes, reporting comparable performance, and discuss bias-variance tradeoffs and future improvements.

Significance. If the reported parity with NILC holds on truly independent sky, SZU would be a viable supervised alternative to NILC for tSZ extraction, with the practical advantage of end-to-end training and curriculum learning for bias control. The paper gives a detailed description of the architecture and training procedure, and the code is publicly available, which are concrete strengths. The simulated comparison set and PSZ2 flux test are useful falsifiable anchors. However, the current evaluation has several load-bearing limitations, including shared real-sky backgrounds between training and test, model selection on the same validation sets used for the headline comparison, noisy labels due to the real SZ background, and omitted CIB-SZ correlations. These issues must be resolved before the central parity claim is fully supported.

major comments (4)
  1. [Secs. 2.2 and 4.3] The train and test patches are not independent in sky background. As described in Sec. 2.2, the 700 patch center positions are identical across all 100 y-map realizations, and the 67:33 split is applied to the resulting 70,000 samples; a random split therefore places patches at the same sky positions in both training and test, always with the same fixed Planck background maps. The validation metrics in Table 2 and Figure 5 thus measure performance on sky that the network has already seen during training, up to different injected SZ realizations. The defense in Sec. 4.3 that different injected SZ maps force generalization is not sufficient: a network can partially memorize the fixed background and subtract it, inflating its apparent performance. The statement that strict angular separation would degrade performance because the model would not have seen the full spatial variability of the contamination is effectively an admission that the model depends on the training-sky background. To support the parity claim against the untrained NILC method, the authors should evaluate SZU on angularly disjoint test regions, or on an independent foreground simulation, and show that performance is preserved.
  2. [Secs. 2.4.3 and 3.2] The comparison set used for the headline SZU-versus-NILC result is the same validation set used for model selection. Section 2.4.3 states that the best model is chosen by rank-ordering errors and biases on validation sets, and Section 3.2 states that the comparison uses 'three validation datasets taken from one of the cross-validations.' This means the SZU model was selected to minimize error and bias on the very patches used in Figure 5, while NILC is an analytical method that receives no such selection. The reported parity may therefore partly reflect selection bias in favor of SZU. The authors should report performance on a held-out test set that was never used for model selection, or at least quantify how much the rank-based selection affects the comparison.
  3. [Sec. 2.3 and Eq. (3)] The training labels are not clean maps of the total SZ signal. The loss in Eq. (3) uses the injected simulated SZ map as ground truth, but the input frequency maps also contain the real sky's SZ signal, including unresolved clusters and the cosmic SZ background discussed in Sec. 2.2. Known clusters are masked and weak pixels are excluded, but residual real SZ remains in the input and is effectively treated as noise to be ignored. NILC, by contrast, attempts to recover the total SZ signal. Consequently, the simulated comparison in Figure 5 may penalize NILC for correctly recovering real SZ and reward SZU for learning to suppress it, so the parity claim is not as clean as stated. The authors should quantify the level of residual real SZ in their training patches, or run a controlled test where a known additional SZ component is present in the input but not in the target.
  4. [Sec. 4.4] The training simulations omit the well-established spectral and spatial correlations between the SZ signal and the cosmic infrared background (CIB). As the authors acknowledge, CIB from cluster member galaxies produces SZ-like signals and is a significant contaminant for unresolved Planck sources. Because the network is trained end-to-end on labels that lack this correlation, the learned mapping may be biased on real clusters with significant dust emission. This is load-bearing for the real-world leg of the parity claim: the PSZ2 YSZ agreement in Figure 6 could reflect shared systematics between SZU and NILC rather than accurate extraction. A quantitative test, such as injecting a correlated CIB component into simulated test data or checking residuals against dust tracers for the PSZ2 sample, is needed to establish that the real-world agreement is not an artifact of the missing correlation.
minor comments (5)
  1. [Sec. 2.4.2] The curriculum schedule in Eq. (4) is described as decreasing ysig by alpha = 2 per epoch, with a floor at ysig = 10; the text should state explicitly that for y0 = 60 the floor is reached at epoch 25 and remains fixed thereafter, since Figure 3 shows training out to 60 epochs.
  2. [Sec. 4.1, Eq. (7)] The signal-to-noise ratio formula uses 'Bp' in the numerator and 'B' in the denominator; these symbols are not defined consistently. Please clarify whether Bp is the bias of positive-selection pixels and how it relates to B.
  3. [Sec. 2.2] The sentence 'the center positions of the 700 images were identical across all realizations' is important for reproducibility; please also state explicitly whether the 67:33 split is performed per realization, across all samples, or by sky position, since this determines whether the test set is angularly disjoint from the training set.
  4. [Sec. 2.5.1] The NILC implementation omits the analysis/synthesis split and smooths in pixel space rather than harmonic space; a short justification is given, but the text says 'differences were negligible' without showing a comparison. A reference to a validation of this simplified NILC would be helpful.
  5. [Figure 6] In the left panel of Figure 6, the full sample includes negative YSZ values that are unphysical; the authors explain that these may arise from foreground overcorrection, but the text would benefit from stating how many of the 1600+ PSZ2 sources remain after the theta500 and latitude cuts and how many have negative YSZ.

Circularity Check

0 steps flagged · score 2.0 of 10

No definitional circularity; shared-sky train/test overlap is a generalization concern, not an equation-level equivalence.

full rationale

The paper's central claim is that SZU, a 3D Attention Nested U-Net trained end-to-end on simulated SZ injections into Planck maps, performs comparably to NILC. The derivation chain is: (i) external Han et al. (2021) simulations provide y-maps; (ii) these are converted to frequency signals with the standard tSZ spectral law and added to PR4 maps; (iii) the network is trained to map the nine frequency channels to the injected y with a masked loss and curriculum learning; (iv) performance is measured against the injected labels and against NILC and PSZ2. No step defines the output in terms of the model's own fitted parameters: the labels are independent of the network, NILC is an external analytical method, and the PSZ2 comparison uses catalog positions and masses. The only self-citations (Pratt et al. 2024 for the NILC Gamma=10 setting and for y-map preprocessing) are methodological choices, not load-bearing derivations. Section 4.3 acknowledges that train and test patches share the same fixed Planck background; this is a genuine independence concern, and the authors' rebuttal that different injected SZ realizations force generalization is debatable because a fixed position-dependent background can in principle be memorized across realizations. However, this is a data-splitting/leakage issue rather than a definitional circularity: the model's output is not, by the paper's equations, identical to a fitted input, and the real-world comparison is anchored to an external method and catalog. Similarly, using validation data for the comparison set is a statistical double-dipping concern, not an equation-level reduction. No 'prediction' is a renamed fitted parameter, and no load-bearing step reduces to a self-citation chain.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the realism of simulated labels and the equivalence of the NILC comparison, neither of which is independently benchmarked beyond the paper's own controls.

free parameters (5)
  • Training signal threshold y_sig_min = 10 (scaled y, ~2.9e-6 in y)
    Pixels with scaled SZ below 10 are masked from the loss; chosen from the superposition bias test (Fig. 1) where fractional bias reaches ~1%.
  • Curriculum starting threshold y0 = 60 or 10 (scaled y)
    y0=60 implements curriculum learning, y0=10 starts at the final threshold; chosen arbitrarily to test the effect on bias and variance.
  • Curriculum decay rate alpha = 2 per epoch
    The loss threshold y_sig decreases by alpha each epoch (Eq. 4); fixed for all models.
  • Cluster mask radius = 2 R500
    Known clusters from MCXC and PSZ2 are masked out to 2R500 to avoid superposing simulated and real SZ signals; chosen based on cluster extent.
  • NILC spatial parameter Gamma = 10
    Adopted from Pratt et al. (2024) as the value that yielded unbiased NILC solutions; affects the NILC comparison baseline.
assumptions (5)
  • domain assumption The Han et al. (2021) synthetic SZ y-maps are statistically representative of the real SZ sky.
    These maps are the only source of labels; the paper notes they lack CIB correlations and local systems, so this representativeness is imperfect.
  • standard math The SZ spectral dependence g(nu) is known exactly and is used to convert y-maps to frequency maps (Eq. 2).
    Standard physical law; underpins both the injected signals and the NILC extraction.
  • domain assumption The real Planck maps have negligible SZ signal in the unmasked regions where synthetic signals are injected, relative to the injected signal above the threshold.
    The superposition approach relies on this to keep label noise manageable; the true SZ background is largely unknown, so this is an assumption.
  • domain assumption Contamination properties in masked cluster regions resemble those in unmasked training regions.
    Explicitly stated in Section 4.3: 'We assume the contamination properties present in the masked regions are similar to their ambient regions that were available during training.'
  • domain assumption The custom NILC implementation is equivalent to the standard NILC.
    Stated in Section 2.5.1: 'the differences between the original process and the one used here were negligible.'

how reviews work

0 comments
Cite this review

Pith. "Pith review of Supervised Extraction of the Thermal Sunyaev$-$Zel'dovich Effect with a Three-Dimensional Convolutional Neural Network." pith.science (2026). https://pith.science/paper/JSYBPRG2

@misc{pith2026250713400,
  author       = {Pith},
  title        = {Pith review of: Supervised Extraction of the Thermal Sunyaev$-$Zel'dovich Effect with a Three-Dimensional Convolutional Neural Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JSYBPRG2}},
  note         = {Machine review of arXiv:2507.13400}
}
abstract

The thermal Sunyaev$-$Zel'dovich (SZ) effect offers a unique probe of the hot and diffuse universe that could help close the missing baryon problem. Traditional extractions of the SZ effect, however, exhibit systematic noise that may lead to unreliable results. In this work, we provide an alternative solution using a three-dimensional Attention Nested U-Net trained end-to-end with supervised learning. Our labeled data consists of simulated SZ signals injected into $\textit{Planck}$ frequency maps, allowing our model to learn how to extract SZ signals in the presence of realistic noise. We implement a curriculum learning scheme that gradually exposed the model to weaker SZ signals. The absence/presence of curriculum learning significantly impacted the amount of bias and variance present in the reconstructed SZ signal. The results from our method were comparable to those from the popular $\textit{needlet internal linear combination}$ (NILC) method when evaluated on simulated data as well as real-world SZ signals. We conclude by discussing future avenues for advancing machine learning extractions of SZ signals.

Figures

Figures reproduced from arXiv: 2507.13400 by the authors.

Figure 1
Figure 1. The effect of superimposing two SZ signals. (left) Averages of superimposed pixel values binned by the strength of the injected SZ signal. (right) Fractional biases and errors of the superimposed signal as a function of injected signal strength. In both panels, the upper x-axis shows the scaled SZ signal strength. The vertical orange line denotes were the scaled SZ signal equals 10. convolutional neural networks. Ea… view at source ↗
Figure 2
Figure 2. All cutouts are centered around the Galactic coor￾dinates (l, b) = (286.97◦ , -73.04◦ ) where the x-axis is Galactic latitude and y-axis Galactic longitude. (top) Panel of the SZ signals injected into the Planck frequency maps. Notice that y˜ has negative values since the median pixel value was sub￾tracted for a local background correction. (middle) A mask of known galaxy clusters provided by the PSZ2 and MCXC catal… view at source ↗
Figure 3
Figure 3. Errors, biases, and noise curves as a function of training epoch in the left, middle, and right panels respectively. The curves are the medians of the three cross-validation sets for the small model. Solid and dashed lines represent the results for y˜0 values of 60 and 10 respectively. Red curves monitor the performance of the strong pixels (˜y > 60) and green for the weak signals (10 < y <˜ 30). below this limit we… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Example of extractions for a patch of sky centered around the same coordinates in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: (top) Recovered pixel values for SZU (black) and NILC (red) for the comparison sample. Data are binned by the true values (˜y) and red data are slightly offset to the right for visual purposes. (bottom) Statistics of the biases (dot￾ted), standard deviations (solid), a…
Figure 6
Figure 6. Figure 6: Comparisons of extracted YSZ values for SZU and NILC using the PSZ2 catalog. The left panel shows the relation between the two methods for the full sample while the middle panel provides a zoomed-in version for the positive signals. The dashed black line in these plots…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 9 canonical work pages

  1. [1]

    W., Piffaretti, R., et al

    Arnaud, M., Pratt, G. W., Piffaretti, R., et al. 2010, A&A, 517, A92, doi: 10.1051/0004-6361/200913416

    doi:10.1051/0004-6361/200913416 2010

  2. [2]

    1999, Physics Reports, 310, 97, doi: https://doi.org/10.1016/S0370-1573(98)00080-5

    Birkinshaw, M. 1999, Physics Reports, 310, 97, doi: https://doi.org/10.1016/S0370-1573(98)00080-5

    doi:10.1016/s0370-1573(98)00080-5 1999

  3. [3]

    2008, Statistical Methodology, 5, 307, doi: 10.1016/j.stamet.2007.10.003

    Aghanim, N. 2008, Statistical Methodology, 5, 307, doi: 10.1016/j.stamet.2007.10.003

    work page doi:10.1016/j.stamet.2007.10.003 2008

  4. [4]

    Bregman, J. N. 2007, ARA&A, 45, 221, doi: 10.1146/annurev.astro.45.051806.110619

    arXiv 2007

  5. [5]

    2008, IEEE Journal of Selected Topics in Signal Processing, 2, 735, doi: 10.1109/JSTSP.2008.2005346

    Cardoso, J.-F., Le Jeune, M., Delabrouille, J., Betoule, M., & Patanchon, G. 2008, IEEE Journal of Selected Topics in Signal Processing, 2, 735, doi: 10.1109/JSTSP.2008.2005346

    work page arXiv 2008

  6. [6]

    1991, A&A, 245, L21

    Cavaliere, A., Menci, N., & Setti, G. 1991, A&A, 245, L21

    work page 1991

  7. [7]

    2020, ApJ, 902, 56, doi: 10.3847/1538-4357/abb403

    Chiang, Y.-K., Makiya, R., M´ enard, B., & Komatsu, E. 2020, ApJ, 902, 56, doi: 10.3847/1538-4357/abb403

    doi:10.3847/1538-4357/abb403 2020

  8. [8]

    F., Le Jeune, M., et al

    Delabrouille, J., Cardoso, J. F., Le Jeune, M., et al. 2009, A&A, 493, 835, doi: 10.1051/0004-6361:200810514

    doi:10.1051/0004-6361:200810514 2009

Show all 33 references
  1. [9]

    J., Aurlien, R., et al

    Galloway, M., Andersen, K. J., Aurlien, R., et al. 2023, A&A, 675, A3, doi: 10.1051/0004-6361/202243137 G´ orski, K. M., Hivon, E., Banday, A. J., et al. 2005, ApJ, 622, 759, doi: 10.1086/427976

    2023 doi

  2. [10]

    2022, JCAP, 2022, 030, doi: 10.1088/1475-7516/2022/01/030

    Guzman, E., & Meyers, J. 2022, JCAP, 2022, 030, doi: 10.1088/1475-7516/2022/01/030

    2022 doi

  3. [11]

    2021, PhRvD, 104, 123521, doi: 10.1103/PhysRevD.104.123521

    Han, D., Sehgal, N., & Villaescusa-Navarro, F. 2021, PhRvD, 104, 123521, doi: 10.1103/PhysRevD.104.123521

    2021 doi

  4. [12]

    2019, JCAP, 2019, 039, doi: 10.1088/1475-7516/2019/02/039

    Khatri, R. 2019, JCAP, 2019, 039, doi: 10.1088/1475-7516/2019/02/039

    2019 doi

  5. [13]

    M., Cardoso, J

    Leach, S. M., Cardoso, J. F., Baccigalupi, C., et al. 2008, A&A, 491, 597, doi: 10.1051/0004-6361:200810116

    2008 doi

  6. [14]

    2014, arXiv e-prints, arXiv:1409.5185, doi: 10.48550/arXiv.1409.5185

    Lee, C.-Y., Xie, S., Gallagher, P., Zhang, Z., & Tu, Z. 2014, arXiv e-prints, arXiv:1409.5185, doi: 10.48550/arXiv.1409.5185

    2014 doi arXiv

  7. [15]

    2020, Computers & Graphics, 90, 11, doi: https://doi.org/10.1016/j.cag.2020.05.003

    Li, C., Tan, Y., Chen, W., et al. 2020, Computers & Graphics, 90, 11, doi: https://doi.org/10.1016/j.cag.2020.05.003

    2020 doi

  8. [16]

    2022, arXiv e-prints, arXiv:2207.02376, doi: 10.48550/arXiv.2207.02376

    Li, R., Wang, X., Huang, G., et al. 2022, arXiv e-prints, arXiv:2207.02376, doi: 10.48550/arXiv.2207.02376

    2022 doi arXiv

  9. [17]

    McCarthy, F., & Hill, J. C. 2023, arXiv e-prints, arXiv:2307.01043, doi: 10.48550/arXiv.2307.01043

    2023 doi arXiv

  10. [18]

    C., Coulton, W

    McCarthy, F., Hill, J. C., Coulton, W. R., & Hogg, D. W. 2024, arXiv e-prints, arXiv:2404.03557, doi: 10.48550/arXiv.2404.03557

    2024 doi arXiv

  11. [19]

    B., Bartlett, J

    Melin, J. B., Bartlett, J. G., & Delabrouille, J. 2006, A&A, 459, 341, doi: 10.1051/0004-6361:20065034

    2006 doi

  12. [20]

    W., Pointecouteau, E., & Melin, J

    Piffaretti, R., Arnaud, M., Pratt, G. W., Pointecouteau, E., & Melin, J. B. 2011, A&A, 534, A109, doi: 10.1051/0004-6361/201015377 Planck Collaboration, Ade, P. A. R., Aghanim, N., et al. 2013, A&A, 554, A140, doi: 10.1051/0004-6361/201220247 Planck Collaboration, Ade, P. A. R...

    2011 arXiv

  13. [21]

    Pratt, C. T. 2025,, https://doi.org/10.5281/zenodo.15570140 doi: 10.5281/zenodo.15570140

    2025 doi

  14. [22]

    T., Qu, Z., Bregman, J

    Pratt, C. T., Qu, Z., Bregman, J. N., & Miller, C. J. 2024, ApJ, 964, 122, doi: 10.3847/1538-4357/ad24f9

    2024 doi

  15. [23]

    2011, MNRAS, 410, 2481, doi: 10.1111/j.1365-2966.2010.17624.x

    Remazeilles, M., Delabrouille, J., & Cardoso, J.-F. 2011, MNRAS, 410, 2481, doi: 10.1111/j.1365-2966.2010.17624.x

    2011

  16. [24]

    2015, arXiv e-prints, arXiv:1505.04597, doi: 10.48550/arXiv.1505.04597

    Ronneberger, O., Fischer, P., & Brox, T. 2015, arXiv e-prints, arXiv:1505.04597, doi: 10.48550/arXiv.1505.04597

    2015 doi arXiv

  17. [25]

    2010, ApJ, 709, 920, doi: 10.1088/0004-637X/709/2/920

    Sehgal, N., Bode, P., Das, S., et al. 2010, ApJ, 709, 920, doi: 10.1088/0004-637X/709/2/920

    2010 doi

  18. [26]

    M., Smith, B

    Shull, J. M., Smith, B. D., & Danforth, C. W. 2012, ApJ, 759, 23, doi: 10.1088/0004-637X/759/1/23

    2012 doi

  19. [27]

    Stevens, T. S. W., van Gorp, H., Meral, F. C., et al. 2023, arXiv e-prints, arXiv:2302.05290, doi: 10.48550/arXiv.2302.05290

    2023 doi arXiv

  20. [28]

    A., & Zeldovich, Y

    Sunyaev, R. A., & Zeldovich, Y. B. 1970, Ap&SS, 7, 3, doi: 10.1007/BF00653471

    1970 doi

  21. [29]

    A., & Zeldovich, Y

    Sunyaev, R. A., & Zeldovich, Y. B. 1972, Comments on Astrophysics and Space Physics, 4, 173

    1972

  22. [30]

    2017, MNRAS, 469, 2821, doi: 10.1093/mnras/stx949

    Thorne, B., Dunkley, J., Alonso, D., & Næss, S. 2017, MNRAS, 469, 2821, doi: 10.1093/mnras/stx949

    2017 doi

  23. [31]

    2017, arXiv e-prints, arXiv:1706.03762, doi: 10.48550/arXiv.1706.03762

    Vaswani, A., Shazeer, N., Parmar, N., et al. 2017, arXiv e-prints, arXiv:1706.03762, doi: 10.48550/arXiv.1706.03762

    2017 doi arXiv

  24. [32]

    2018, arXiv e-prints, arXiv:1807.10165, doi: 10.48550/arXiv.1807.10165 SZ-UNet 15

    Zhou, Z., Mahfuzur Rahman Siddiquee, M., Tajbakhsh, N., & Liang, J. 2018, arXiv e-prints, arXiv:1807.10165, doi: 10.48550/arXiv.1807.10165 SZ-UNet 15

    2018 doi arXiv

  25. [33]

    2019, Journal of Open Source Software, 4, 1298, doi: 10.21105/joss.01298 16 Pratt et al

    Zonca, A., Singer, L., Lenz, D., et al. 2019, Journal of Open Source Software, 4, 1298, doi: 10.21105/joss.01298 16 Pratt et al. APPENDIX A. DOMAIN ADAPTATION In this work, we utilized synthetic SZ signals to train our models by superimposing them with the Planck frequency dat...

    2019 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.