Pith. sign in

REVIEW 3 major objections 6 minor 80 references

Predicting the Kinematics of the Cold Circumgalactic Medium from its Morphology using Convolutional Neural Networks

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A convolutional network can infer the plane-of-sky velocities of cold circumgalactic gas from H-alpha images alone.

desk verdict Credible proof-of-concept for CNN-based CGM kinematics from morphology, but the headline accuracy is for bright pixels only and real-data transfer is unproven. read the letter →

arxiv 2608.04087 v1 pith:KRNCW6EM submitted 2026-08-04 astro-ph.GA

Connor Jennings , Earl P. Bellinger , Imad Pasha , Pieter van Dokkum , Pratik J. Gandhi This is my paper · ORCID
classification astro-ph.GA
keywords circumgalacticmediumcoldgaskinematicsplane-of-skyvelocityH-alphaemissionconvolutionalneuralnetworksUNetTNG50simulationMOTHRAtelescope
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 argues that a single two-dimensional H-alpha image of the cold gas around a Milky Way-mass galaxy contains enough hidden structure to infer the plane-of-sky motions of that gas, a velocity component that is normally impossible to measure outside our own galaxy. It trains a convolutional neural network on mock emission and velocity maps from the TNG50 cosmological simulation, and shows that the trained network recovers both radial and tangential flow directions with typical errors of 0.3–0.5 v_vir at noise levels expected for the upcoming MOTHRA telescope. If true, this would add two phase-space dimensions to CGM observations, allowing astronomers to see whether cold circumgalactic gas is falling in, orbiting, or being ejected, and to plan targeted follow-up of particular streams. The claim is conditional on simulated CGMs resembling real ones.

What carries the argument

The machinery is a UNet, an encoder-decoder convolutional network with skip connections that maps 2D images to 2D images, trained in PyTorch with an ImageNet-pretrained encoder. Inputs are 256x256 mock H-alpha maps generated from TNG50 gas cells with Cloudy emissivities; targets are the two components of the plane-of-sky cold-gas velocity, averaged along the line of sight and H-alpha-emission-weighted. Loss is mean squared error evaluated only on H-alpha-bright pixels outside a 16 kpc central mask, and Gaussian noise is added during training to simulate observational depth. The network's job is to invert the simulated radiative transfer: given where the gas glows, it must recover which way it moves.

What would settle it

Take deep H-alpha images of several local galaxies with MOTHRA, measure the plane-of-sky motions of individual CGM clouds from multi-epoch imaging or IFU velocity fields, and check whether the network's velocity predictions agree; if real CGMs are systematically clumpier or smoother than TNG50, the errors should exceed the claimed 0.3–0.5 v_vir.

Watch

Extended reading notes

Core claim

The central discovery is that there exists a non-injective but learnable mapping from H-$\alpha$ surface-brightness morphology to the line-of-sight-averaged plane-of-sky velocity field of cold (T < $10^{4}$.5 K) circumgalactic gas. A UNet, trained only on emission maps as input and on masked velocity maps as targets, learns to reproduce global flows: it correctly identifies disk rotation, radial and tangential infall, and velocity gradients in streams, while failing on polar outflows and satellite-associated gas. At zero noise the typical RMS error is 0.3–0.5 v_vir; errors stay at this level until noise reaches about $10^{-20}$ erg/s/$cm^{2}$/$arcsec^{2}$, above which the network falls back to a prior of global radial infall. Since MOTHRA is predicted to reach deeper than that threshold, the authors conclude that forthcoming H-$\alpha$ imaging can estimate two additional phase-space dimensions of the cold CGM.

Load-bearing premise

The load-bearing premise is that the cold CGM gas in real galaxies emits H-alpha with the same statistical morphology as in TNG50, so a network trained on simulated projections will transfer to real images.

Editorial extensions

If this is right

  • Deep H-alpha surveys of local galaxies could produce maps of cold-gas flow direction without spectroscopy or proper motions.
  • The recovered velocity maps add two phase-space dimensions to CGM observations, enabling targeted follow-up of inflowing or orbiting streams.
  • There is a noise floor around 10^-20 erg/s/cm^2/arcsec^2 below which the mapping degrades, so telescope depth directly controls kinematic information content.
  • The network's high-noise output can be read as a learned prior, letting observers isolate the information genuinely contributed by CGM morphology.
  • Monte Carlo Dropout uncertainty maps identify which morphological features are underrepresented, guiding where larger training sets would help.

Reading between the lines

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

  • Going beyond the paper: if the mapping is as general as claimed, training on a second cosmological simulation with different feedback physics would provide a direct transfer test, since the authors note TNG50 may not match real cloud morphology.
  • Editorial inference: adding line-of-sight velocity information, NII/OIII line ratios, or stellar surface-brightness maps as auxiliary inputs should reduce the aleatoric errors the authors find dominate, because these observables break degeneracies in the non-injective mapping.
  • Editorial inference: the same architecture could be applied to other emission lines or to higher-redshift CGM images, where morphology is all that is available.
  • Editorial inference: the predicted velocity maps, even imperfect, could be used to design slit or IFU placements for spectroscopic follow-up, converting a probabilistic velocity prior into an observing strategy.
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

3 major / 6 minor

Summary. The paper trains a UNet on mock H-alpha emission maps of 182 MW/M31-like galaxies from the TNG50 simulation, with the goal of predicting plane-of-sky velocities of cold CGM gas from single-band emission morphology alone. The authors generate 80 rotated/reflected views per galaxy along ten sightlines, forward-model H-alpha emissivities with Cloudy, add Gaussian noise at levels comparable to the planned MOTHRA telescope, and train on a galaxy-level train/test/validation split. They report qualitative success for inflows and rotation, RMS velocity errors of 0.3-0.5 v_vir in the zero-noise limit and at MOTHRA-relevant noise, a noise-dependence study identifying a critical noise floor, and Monte Carlo dropout estimates of epistemic uncertainty. The paper concludes that 2D emission morphology encodes a learnable, non-injective mapping to plane-of-sky kinematics, enabling new phase-space information from upcoming observations.

Significance. If the result holds, it would add two previously unobservable phase-space dimensions of cold CGM gas to what can be extracted from single emission images, with direct implications for follow-up observations with MOTHRA and similar facilities. The paper has clear strengths: a clean galaxy-level train/test/validation split, explicit noise-dependence tests, Monte Carlo dropout uncertainty analysis, hyperparameter optimization, and an honest discussion of the simulation-dependent nature of the method. The internal claim that a CNN can learn this mapping within TNG50 is well supported. The practical claim about real observations rests on two assumptions that are acknowledged but not quantitatively tested: that TNG50 CGMs resemble real CGMs, and that the reported error statistics, which are computed only on masked, H-alpha-bright pixels, transfer to the faint, detection-threshold-limited regime that MOTHRA will actually observe. The stress-test concern about the masked metric lands and is the main reason the paper requires revision before the headline claim can be accepted.

major comments (3)
  1. [Abstract; Sec. 3.2 and Fig. 5] The headline RMS errors of 0.3-0.5 v_vir are computed only on pixels that satisfy the loss mask defined in Sec. 2.4: pixels with bolometric H-alpha surface brightness below 10^-21 erg/s/cm2/arcsec2 are excluded, as is a central 16 kpc circle. The abstract and the summary bullet in Sec. 5 state the accuracy without this qualifier, implying it applies to all gas that MOTHRA would detect. Because the faintest CGM emission is precisely what a photon-starved survey will measure, the reported accuracy may substantially overstate real-world performance. Please recompute the error statistics on a detection-thresholded pixel selection (e.g., pixels above a per-pixel S/N cut in the noisy image) and explicitly state the mask whenever the 0.3-0.5 v_vir value is quoted.
  2. [Sec. 2.4 and Appendix B] The model is not trained to distinguish regions that contain cold gas from those that do not, and Fig. 8 demonstrates that on pure noise the model confidently predicts coherent flows. The paper acknowledges this in Sec. 4.1, but the full application pipeline is never tested: real MOTHRA images will require some external detection/thresholding step before the model is applied, and false positives from noise will contaminate the velocity predictions. An experiment that applies a detection threshold to noisy simulated images and then measures velocity errors on the detected pixels would close this gap. Without it, the claim that the method 'enables targeted followup' is not fully supported.
  3. [Sec. 2.1 and Sec. 4.1] The practical implication in the abstract assumes that TNG50 cold CGM morphology is statistically representative of real galaxies. This is acknowledged as the largest caveat, but it is not tested. Since the paper explicitly notes that small-scale CGM morphology varies significantly between simulations (citing Wright et al. 2024 and Medlock et al. 2025), a cross-simulation transfer test (e.g., train on TNG50 and test on FIRE-2 or FOGGIE mock maps, even in a simplified form) or a sensitivity analysis to clumping factor and cloud-size assumptions would materially support the claimed applicability. Absent such a test, the conclusions should be restricted to the simulation-internal statement.
minor comments (6)
  1. [Sec. 2.2.1 and Eqs. (2)-(3)] The notation in Eq. (2) uses \(\tilde{F}_{H\alpha}\) before it is defined in Eq. (3), and the normalization denominator in Eq. (3) appears to omit the \(T<10^{4.5}\) restriction carried by Eq. (2). Please align the notation and make the integration limits explicit.
  2. [Sec. 2.2.1] The text states that dividing the emission density by 4\pi steradian gives a brightness in [erg/s/cm3/arcsec2]; the final map after line-of-sight integration should be [erg/s/cm2/arcsec2]. Please correct the units in the sentence.
  3. [Sec. 3.2] The noise level is quoted as '10^-20 erg/s/cm2/arcsec2/\AA' in one place, but elsewhere as a surface brightness in erg/s/cm2/arcsec2. If the intended quantity is per \AA, the definition should be stated; otherwise the /\AA should be removed.
  4. [Fig. 6 caption and text] There are several typographical errors: 'Dependance' in the Fig. 6 caption, 'assosiated' in Sec. 3.1 and Sec. 3.3, 'boundry' in Sec. 1, and 'corelation' in Sec. 3.3. A careful proofread is needed.
  5. [References] Weinberger et al. 2017a and 2017b appear with identical bibliographic details (MNRAS, 465, 3291). Please verify whether these are distinct papers or a duplicate entry.
  6. [Appendix A] The best-fit model uses the largest encoder and decoder width in the search range, and the text notes this may indicate that a larger model would perform better. Since this is a potential unresolved optimization issue, briefly stating the computational budget that prevented exploring larger models would help the reader calibrate the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the velocity prediction is validated on galaxy-level held-out TNG50 data, and the target maps are not constructed from the input emission maps by any identity.

full rationale

The central claim is an empirical supervised-learning result: a UNet maps H-alpha emission morphology to plane-of-sky cold-gas velocities with 0.3-0.5 v_vir RMS error. The targets are computed from TNG50 cold-gas velocities weighted by H-alpha emissivity (Eqs. 2-3), while the inputs are the emissivity maps themselves; velocity components are independent physical fields, so the input-output relation is non-injective and learned, not imposed by definition. The train/test/validation split is performed at the galaxy level (Sec. 2.4), so test galaxies are unseen during training, and no parameters are fit to the test set. The reported errors are therefore genuine generalization errors within TNG50. The loss mask (Sec. 2.4) restricts evaluation to H-alpha-bright, non-central pixels; this is an evaluation choice that may overstate applicability to faint MOTHRA data, as the authors themselves concede in Sec. 4.1 and Appendix B, but it is not a circular reduction of the prediction to the input. The paper's own caveats, namely that simulation-based inference assumes TNG50 resembles real CGMs (Sec. 4.1) and that the model is not trained to distinguish signal from no-signal regions (Appendix B), are limitations rather than circularity. The only self-citation, 'I. Pasha in prep.' supporting the MOTHRA depth used as motivation, is not load-bearing for the learned mapping and does not make the prediction equivalent to its inputs. No circular step is present.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The paper's central error claims depend on a set of hand-chosen thresholds and ML hyperparameters, all reported in Section 2.4 and Appendix A. The physical fidelity of the forward model rests on three unverified domain assumptions: TNG50 realism, optical thinness with unity clumping factor, and a Gaussian noise model for future instruments. No new entities are introduced.

free parameters (9)
  • Decoder base width = 128
    Tuned with Optuna on the validation set; the model size affects reported accuracy.
  • Dropout fraction = 0.128
    Tuned with Optuna; part of the model's regularization and Monte Carlo dropout uncertainty.
  • Learning rate = 3e-4
    Tuned with Optuna; affects convergence and final performance.
  • Weight decay = 1.5e-6
    Tuned with Optuna; regularizer.
  • Encoder architecture = ResNet152
    Tuned with Optuna; the best fit is at the upper bound of the search range, suggesting further scaling might be needed.
  • H-alpha surface brightness threshold = 10^-21 erg/s/cm2/arcsec2
    Chosen by hand to define valid pixels for loss and error reporting; the headline RMS error applies only to pixels above this threshold.
  • Central mask radius = 8 pixels (about 16 kpc)
    Chosen by hand to exclude the unresolved host galaxy; affects which pixels are evaluated.
  • Input clipping noise floor = 10^-25 erg/s/cm2/arcsec2
    Chosen by hand; affects the dynamic range of inputs.
  • yt smoothing factor = 2
    Chosen by hand; the paper states it does not affect results because gas particles are smoothed to resolutions finer than the map resolution.
assumptions (5)
  • domain assumption TNG50 accurately represents real CGM cold gas morphology and kinematics.
    Invoked in Section 2.1 and Section 4.1 as the basis for transferring the learned mapping to observations; the paper explicitly calls this the largest caveat.
  • domain assumption The CGM is optically thin to H-alpha and the clumping factor C equals 1 within each cell.
    Explicitly assumed in Section 2.2.1 when integrating emission along the line of sight; if real clumping is larger, simulated emission would be dimmer and more uniform than real observations.
  • domain assumption Gaussian noise added to simulated images is representative of upcoming observational facilities.
    The methods in Section 2.2.1 only add per-pixel Gaussian noise, yet the abstract claims forward modeled noise properties consistent with upcoming facilities; no PSF, sky background, or instrument response is modeled.
  • domain assumption H-alpha-emission-weighted line-of-sight averaging is a meaningful definition of the target velocity field.
    Defined in Section 2.2.2; the target velocities are weighted by H-alpha emission, so the network learns to predict this specific weighted quantity, not the raw mass-weighted velocity.
  • domain assumption ImageNet-pretrained encoder weights transfer to astrophysical emission maps.
    Used in Section 2.4 as initial weights; a practical transfer learning assumption that is standard but not physically motivated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Predicting the Kinematics of the Cold Circumgalactic Medium from its Morphology using Convolutional Neural Networks." pith.science (2026). https://pith.science/paper/KRNCW6EM

@misc{pith2026260804087,
  author       = {Pith},
  title        = {Pith review of: Predicting the Kinematics of the Cold Circumgalactic Medium from its Morphology using Convolutional Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KRNCW6EM}},
  note         = {Machine review of arXiv:2608.04087}
}
abstract

We present a novel approach to predicting plane-of-sky velocities of cold gas clouds in the circumgalactic medium (CGM) of galaxies. The method uses a convolutional neural network (CNN) trained on simulated emission maps derived from the TNG50 cosmological simulation, with forward modeled noise properties consistent with upcoming observational facilities. Using 182 Milky Way/Andromeda analog galaxies, we generate emission maps in H$\alpha$ using $\texttt{Cloudy}$ models, as well as line-of-sight averaged 2D velocity maps. Using a UNet architecture, we train the CNN to take emission maps as input and return plane-of-sky velocity maps as output, which cannot be observationally constrained using traditional methods. Qualitatively, the model is generally able to infer the true overall flow direction. We quantify the effects of Gaussian noise on the network's training and predictive power. At depths expected to be probed by forthcoming telescopes such as MOTHRA, the network has a typical RMS error for the plane-of-sky velocity direction of $0.3-0.5 v_{vir}$. This implies that 2D emission maps of sufficient depths will be able to estimate two additional phase space dimensions of cold CGM gas, enabling targeted followup and a better understanding of overall CGM flows.

Figures

Figures reproduced from arXiv: 2608.04087 by the authors.

Figure 1
Figure 1. Stellar mass surface density, Hα emission, and cold gas velocity of TNG50 subid 392277. Σ⋆ is plotted on top of Hα emission. A quiver plot is overlayed on the Hα emission to show the velocities associated with cold gas structures in the CGM. match two output arrays to the true plane-of-sky ve￾locity of cold gas in the simulated CGM. We use pro￾jections of cold CGM gas from the IllustrisTNG50 (D. Nelson et al. 2019; … view at source ↗
Figure 2
Figure 2. Top: Quiver plot of predicted (blue) and true (white) values for the plane-of-sky velocity of cold gas around a galaxy in our test set (TNG50 subid 552581), overlayed on its Hα emission. Bottom Left: Signed errors in the predicted vθ and vR over the regions of the CGM that would have been used to determine loss if the galaxy were in our training set. Bottom Right: Standard deviations of Monte Carlo Dropout draws, hi… view at source ↗
Figure 3
Figure 3. Quiver plots of predicted (blue) and true (white) values for the plane-of-sky velocity of cold gas around six galaxies in our test set, chosen to showcase a range of CGM morphologies. TNG50 subids are shown for each. Subids 487742 and 427211 show the response of the model to “messy” CGMs with multiple satellites. For these, the model has moderate predictive power for the broader, more cohesive flows but smooths over… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Dependence of model predictions on increasing Gaussian noise. Each column shows the results for a specific noise level. Top: input Hα map, true velocities (white) and predicted velocities (blue). Middle and Bottom: signed errors in vθ and vR. For noise ϵ ∼ N(0, σ = 10−…
Figure 5
Figure 5. Figure 5: Dependence of the signed error in vθ and vR on Gaussian noise. Error distributions were calculated using the set of all Hα-bright, non-central pixels that satisfy the loss mask (Sec. 2.4) in the set of test galaxies. Lines trace the 5th, 16th, 50th, 84th and 95th perce…
Figure 6
Figure 6. Figure 6: Dependance of root mean squared (RMS) errors and Monte Carlo dropout (MCD) deviation on pixel bright￾ness for the zero noise test set of galaxy views. Error widths are the 16th and 84th percentiles of the RMS and MCD deviation over the test sample. The RMS errors are s…
Figure 7
Figure 7. Figure 7: Results of the model applied to a TNG50 galaxy outside of the MW/M31 sample. This galaxy is a massive quiescent, and was selected as an example due to having a comparable CGM morphology to the galaxy shown in [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Out-of-distribution test assessing the behavior of the model when inputs are not CGM data. Top Left: output from input of pure random noise, with the same standard deviation and mean as the training data set. Other plots: output from input of structured, non-astronomic…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

80 extracted references · 8 canonical work pages

  1. [1]

    2019, in The 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2623–2631

    Akiba, T., Sano, S., Yanase, T., Ohta, T., & Koyama, M. 2019, in The 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2623–2631

    2019

  2. [2]

    2021, MNRAS, 506, 5129, doi: 10.1093/mnras/stab1958

    Byrohl, C., Nelson, D., Behrens, C., et al. 2021, MNRAS, 506, 5129, doi: 10.1093/mnras/stab1958

    doi:10.1093/mnras/stab1958 2021

  3. [3]

    2020, arXiv e-prints, arXiv:2004.10710, doi: 10.48550/arXiv.2004.10710

    Caldeira, J., & Nord, B. 2020, arXiv e-prints, arXiv:2004.10710, doi: 10.48550/arXiv.2004.10710

    arXiv doi:10.48550/arxiv.2004.10710 2020

  4. [4]

    J., & Miniati, F

    Cantalupo, S., Porciani, C., Lilly, S. J., & Miniati, F. 2005, ApJ, 628, 61, doi: 10.1086/430758

    doi:10.1086/430758 2005

  5. [5]

    X., et al

    Chang, Y.-L., Lan, T.-W., Prochaska, J. X., et al. 2025, arXiv e-prints, arXiv:2512.03845, doi: 10.48550/arXiv.2512.03845

    doi:10.48550/arxiv.2512.03845 2025

  6. [6]

    2014, MNRAS, 438, 1435, doi: 10.1093/mnras/stt2288

    Chen, H.-W., Gauthier, J.-R., Sharon, K., et al. 2014, MNRAS, 438, 1435, doi: 10.1093/mnras/stt2288

    doi:10.1093/mnras/stt2288 2014

  7. [7]

    Chen, H.-W., & Zahedy, F. S. 2026, in Encyclopedia of

    2026

  8. [8]

    4, 370–400, doi: 10.1016/B978-0-443-21439-4.00059-6

    Astrophysics, Volume 4, Vol. 4, 370–400, doi: 10.1016/B978-0-443-21439-4.00059-6

    doi:10.1016/b978-0-443-21439-4.00059-6

Show all 80 references
  1. [9]

    S., Boettcher, E., et al

    Chen, H.-W., Zahedy, F. S., Boettcher, E., et al. 2020, MNRAS, 497, 498, doi: 10.1093/mnras/staa1773

    2020 doi

  2. [10]

    2023, ApJL, 955, L25, doi: 10.3847/2041-8213/acf85b

    Chen, H.-W., Qu, Z., Rauch, M., et al. 2023, ApJL, 955, L25, doi: 10.3847/2041-8213/acf85b

    2023 doi

  3. [11]

    C., Erb, D

    Chen, Y., Steidel, C. C., Erb, D. K., et al. 2021, MNRAS, 508, 19, doi: 10.1093/mnras/stab2383

    2021 doi

  4. [12]

    G., & Kravtsov, A

    Conroy, C., van Dokkum, P. G., & Kravtsov, A. 2015, ApJ, 803, 77, doi: 10.1088/0004-637X/803/2/77 de los Rios, M., Mart ´ ınez, H. J., Coenda, V., et al. 2021, MNRAS, 500, 1784, doi: 10.1093/mnras/staa3339

    2015 doi

  5. [13]

    F., Genel, S., et al

    DeFelippis, D., Bouch´ e, N. F., Genel, S., et al. 2021, ApJ, 923, 56, doi: 10.3847/1538-4357/ac2cbf

    2021 doi

  6. [14]

    2006, MNRAS, 368, 2, doi: 10.1111/j.1365-2966.2006.10145.x

    Dekel, A., & Birnboim, Y. 2006, MNRAS, 368, 2, doi: 10.1111/j.1365-2966.2006.10145.x

    2006

  7. [15]

    2009, in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 248–255, doi: 10.1109/CVPR.2009.5206848

    Deng, J., Dong, W., Socher, R., et al. 2009, in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 248–255, doi: 10.1109/CVPR.2009.5206848

    2009

  8. [16]

    2017, arXiv e-prints, arXiv:1704.03416, doi: 10.48550/arXiv.1704.03416

    Dijkstra, M. 2017, arXiv e-prints, arXiv:1704.03416, doi: 10.48550/arXiv.1704.03416

    2017 doi arXiv

  9. [17]

    2021, in International Conference on Learning Representations

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., et al. 2021, in International Conference on Learning Representations. https://openreview.net/forum?id=YicbFdNTTy

    2021

  10. [18]

    2025, ApJ, 985, 44, doi: 10.3847/1538-4357/adc922

    Dutta, S., Muzahid, S., Schaye, J., et al. 2025, ApJ, 985, 44, doi: 10.3847/1538-4357/adc922

    2025 doi

  11. [19]

    Fabian, A. C. 2012, ARA&A, 50, 455, doi: 10.1146/annurev-astro-081811-125521 Faucher-Gigu` ere, C.-A. 2020, MNRAS, 493, 1614, doi: 10.1093/mnras/staa302 Faucher-Gigu` ere, C.-A., Lidz, A., Zaldarriaga, M., &

    2012 doi

  12. [20]

    2009, ApJ, 703, 1416, doi: 10.1088/0004-637X/703/2/1416 Faucher-Gigu` ere, C.-A., & Oh, S

    Hernquist, L. 2009, ApJ, 703, 1416, doi: 10.1088/0004-637X/703/2/1416 Faucher-Gigu` ere, C.-A., & Oh, S. P. 2023, ARA&A, 61, 131, doi: 10.1146/annurev-astro-052920-125203

    2009 doi

  13. [21]

    J., Chatzikos, M., Guzm´ an, F., et al

    Ferland, G. J., Chatzikos, M., Guzm´ an, F., et al. 2017, RMxAA, 53, 385, doi: 10.48550/arXiv.1705.10877

    2017 doi arXiv

  14. [22]

    2017, MNRAS, 467, 4802, doi: 10.1093/mnras/stx398

    Fumagalli, M., Haardt, F., Theuns, T., et al. 2017, MNRAS, 467, 4802, doi: 10.1093/mnras/stx398

    2017 doi

  15. [23]

    2016, PMLR, arXiv:1506.02142, doi: 10.48550/arXiv.1506.02142

    Gal, Y., & Ghahramani, Z. 2016, PMLR, arXiv:1506.02142, doi: 10.48550/arXiv.1506.02142

    2016 doi arXiv

  16. [24]

    G., Cantalupo, S., Lilly, S., et al

    Gallego, S. G., Cantalupo, S., Lilly, S., et al. 2018, MNRAS, 475, 3854, doi: 10.1093/mnras/sty037

    2018 doi

  17. [25]

    L., et al

    Gaspari, M., McDonald, M., Hamer, S. L., et al. 2018, ApJ, 854, 167, doi: 10.3847/1538-4357/aaaa1b

    2018 doi

  18. [26]

    2024, MNRAS, 527, 10038, doi: 10.1093/mnras/stad3784

    Villaescusa-Navarro, F., & Angl´ es-Alc´ azar, D. 2024, MNRAS, 527, 10038, doi: 10.1093/mnras/stad3784

    2024 doi

  19. [27]

    Gronke, M., & Oh, S. P. 2020, MNRAS, 492, 1970, doi: 10.1093/mnras/stz3332

    2020 doi

  20. [28]

    F., et al

    Guo, Y., Bacon, R., Bouch´ e, N. F., et al. 2023, Nature, 624, 53, doi: 10.1038/s41586-023-06718-w

    2023 doi

  21. [29]

    P., et al

    Gupta, A., Miotello, A., Williams, J. P., et al. 2024, A&A, 683, A133, doi: 10.1051/0004-6361/202348007

    2024 doi

  22. [30]

    K., Danelljan, M., & Sch¨ on, T

    Gustafsson, F. K., Danelljan, M., & Sch¨ on, T. B. 2019, arXiv e-prints, arXiv:1906.01620, doi: 10.48550/arXiv.1906.01620

    2019 doi arXiv

  23. [31]

    R., Millman, K

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

    2020 doi

  24. [32]

    F., Wetzel, A., Kereˇ s, D., et al

    Hopkins, P. F., Wetzel, A., Kereˇ s, D., et al. 2018, MNRAS, 480, 800, doi: 10.1093/mnras/sty1690

    2018 doi

  25. [33]

    A., et al

    Huang, Y.-H., Chen, H.-W., Shectman, S. A., et al. 2021, MNRAS, 502, 4743, doi: 10.1093/mnras/stab360

    2021 doi

  26. [34]

    Hui, L., & Gnedin, N. Y. 1997, MNRAS, 292, 27, doi: 10.1093/mnras/292.1.27

    1997 doi

  27. [35]

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

    2007 doi

  28. [36]

    2019, Segmentation Models Pytorch,, https://github.com/qubvel/segmentation models.pytorch GitHub

    Iakubovskii, P. 2019, Segmentation Models Pytorch,, https://github.com/qubvel/segmentation models.pytorch GitHub

    2019

  29. [37]

    G., Muzahid, S., Churchill, C

    Kacprzak, G. G., Muzahid, S., Churchill, C. W., Nielsen, N. M., & Charlton, J. C. 2015, ApJ, 815, 22, doi: 10.1088/0004-637X/815/1/22 17

    2015 doi

  30. [38]

    H., & Hernquist, L

    Katz, N., Weinberg, D. H., & Hernquist, L. 1996, ApJS, 105, 19, doi: 10.1086/192305

    1996 doi

  31. [39]

    Selim, I. M. 2017, arXiv e-prints, arXiv:1709.02245, doi: 10.48550/arXiv.1709.02245

    2017 doi arXiv

  32. [40]

    P., & Ba, J

    Kingma, D. P., & Ba, J. 2014, arXiv preprint arXiv:1412.6980

    2014 arXiv

  33. [41]

    2017, in Advances in Neural Information Processing Systems, ed

    Lakshminarayanan, B., Pritzel, A., & Blundell, C. 2017, in Advances in Neural Information Processing Systems, ed. I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, & R. Garnett, Vol. 30 (Curran Associates, Inc.). https://proceedings.neurips.cc/paper f...

    2017

  34. [42]

    2017, A&A, 608, A8, doi: 10.1051/0004-6361/201731480

    Leclercq, F., Bacon, R., Wisotzki, L., et al. 2017, A&A, 608, A8, doi: 10.1051/0004-6361/201731480

    2017 doi

  35. [43]

    2019, ApJ, 877, 4, doi: 10.3847/1538-4357/ab184e

    Schaye, J. 2019, ApJ, 877, 4, doi: 10.3847/1538-4357/ab184e

    2019 doi

  36. [44]

    Marasco, A., de Blok, W. J. G., Maccagni, F. M., et al. 2025, A&A, 697, A86, doi: 10.1051/0004-6361/202453172 Maur ´ ıcio, J., Domingues, I., & Bernardino, J. 2023, Applied Sciences, 13, doi: 10.3390/app13095521

    2025 doi

  37. [45]

    2016, MNRAS, 455, 2994, doi: 10.1093/mnras/stv2508

    McPartland, C., Ebeling, H., Roediger, E., & Blumenthal, K. 2016, MNRAS, 455, 2994, doi: 10.1093/mnras/stv2508

    2016 doi

  38. [46]

    2025, ApJ, 980, 61, doi: 10.3847/1538-4357/ada442

    Medlock, I., Neufeld, C., Nagai, D., et al. 2025, ApJ, 980, 61, doi: 10.3847/1538-4357/ada442

    2025 doi

  39. [47]

    2024, ApJ, 974, 161, doi: 10.3847/1538-4357/ad6cd1

    Melso, N., Schiminovich, D., Sitaram, M., et al. 2024, ApJ, 974, 161, doi: 10.3847/1538-4357/ad6cd1

    2024 doi

  40. [48]

    Burchett, J. N. 2021, MNRAS, 507, 4445, doi: 10.1093/mnras/stab2177

    2021 doi

  41. [49]

    2018, MNRAS, 475, 624, doi: 10.1093/mnras/stx3040

    Nelson, D., Pillepich, A., Springel, V., et al. 2018, MNRAS, 475, 624, doi: 10.1093/mnras/stx3040

    2018 doi

  42. [50]

    2019, MNRAS, 490, 3234, doi: 10.1093/mnras/stz2306

    Nelson, D., Pillepich, A., Springel, V., et al. 2019, MNRAS, 490, 3234, doi: 10.1093/mnras/stz2306

    2019 doi

  43. [51]

    2020, MNRAS, 498, 2391, doi: 10.1093/mnras/staa2419

    Nelson, D., Sharma, P., Pillepich, A., et al. 2020, MNRAS, 498, 2391, doi: 10.1093/mnras/staa2419

    2020 doi

  44. [52]

    2014, ApJL, 783, L20, doi: 10.1088/2041-8205/783/1/L20

    Pakmor, R., Marinacci, F., & Springel, V. 2014, ApJL, 783, L20, doi: 10.1088/2041-8205/783/1/L20

    2014 doi

  45. [53]

    2019, arXiv e-prints, arXiv:1912.01703, doi: 10.48550/arXiv.1912.01703

    Paszke, A., Gross, S., Massa, F., et al. 2019, arXiv e-prints, arXiv:1912.01703, doi: 10.48550/arXiv.1912.01703

    2019 doi arXiv

  46. [54]

    S., Corlies, L., Tumlinson, J., et al

    Peeples, M. S., Corlies, L., Tumlinson, J., et al. 2019, ApJ, 873, 129, doi: 10.3847/1538-4357/ab0654

    2019 doi

  47. [55]

    2021, MNRAS, 508, 4667, doi: 10.1093/mnras/stab2779

    Pillepich, A., Nelson, D., Truong, N., et al. 2021, MNRAS, 508, 4667, doi: 10.1093/mnras/stab2779

    2021 doi

  48. [56]

    2018, MNRAS, 473, 4077, doi: 10.1093/mnras/stx2656

    Pillepich, A., Springel, V., Nelson, D., et al. 2018, MNRAS, 473, 4077, doi: 10.1093/mnras/stx2656

    2018 doi

  49. [57]

    2019, MNRAS, 490, 3196, doi: 10.1093/mnras/stz2338

    Pillepich, A., Nelson, D., Springel, V., et al. 2019, MNRAS, 490, 3196, doi: 10.1093/mnras/stz2338

    2019 doi

  50. [58]

    2024, MNRAS, 535, 1721, doi: 10.1093/mnras/stae2165

    Pillepich, A., Sotillo-Ramos, D., Ramesh, R., et al. 2024, MNRAS, 535, 1721, doi: 10.1093/mnras/stae2165

    2024 doi

  51. [59]

    E., Peek, J

    Putman, M. E., Peek, J. E. G., & Joung, M. R. 2012, ARA&A, 50, 491, doi: 10.1146/annurev-astro-081811-125612

    2012 doi

  52. [60]

    H., Raiˇ cevi´ c, M., & Schaye, J

    Rahmati, A., Pawlik, A. H., Raiˇ cevi´ c, M., & Schaye, J. 2013, MNRAS, 430, 2427, doi: 10.1093/mnras/stt066

    2013 doi

  53. [61]

    2026, arXiv e-prints, arXiv:2602.23416, doi: 10.48550/arXiv.2602.23416

    Ramesh, R., Nelson, D., Fielding, D., & Br¨ uggen, M. 2026, arXiv e-prints, arXiv:2602.23416, doi: 10.48550/arXiv.2602.23416

    2026 doi

  54. [62]

    2023a, MNRAS, 518, 5754, doi: 10.1093/mnras/stac3524

    Ramesh, R., Nelson, D., & Pillepich, A. 2023a, MNRAS, 518, 5754, doi: 10.1093/mnras/stac3524

    doi

  55. [63]

    2023b, MNRAS, 522, 1535, doi: 10.1093/mnras/stad951

    Ramesh, R., Nelson, D., & Pillepich, A. 2023b, MNRAS, 522, 1535, doi: 10.1093/mnras/stad951

    doi

  56. [64]

    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

  57. [65]

    S., Bignone, L

    Rosito, M. S., Bignone, L. A., Tissera, P. B., & Pedrosa, S. E. 2023, A&A, 671, A19, doi: 10.1051/0004-6361/202244707

    2023 doi

  58. [66]

    2016, ApJ, 833, 39, doi: 10.3847/1538-4357/833/1/39

    Schroetter, I., Bouch´ e, N., Wendt, M., et al. 2016, ApJ, 833, 39, doi: 10.3847/1538-4357/833/1/39

    2016 doi

  59. [67]

    J., Kenney, J

    Souchereau, H. J., Kenney, J. D. P., J´ achym, P., et al. 2025, ApJ, 988, 72, doi: 10.3847/1538-4357/adde47

    2025 doi

  60. [68]

    2010, MNRAS, 401, 791, doi: 10.1111/j.1365-2966.2009.15715.x

    Springel, V. 2010, MNRAS, 401, 791, doi: 10.1111/j.1365-2966.2009.15715.x

    2010

  61. [69]

    R., Kaufmann, T., Bullock, J

    Stewart, K. R., Kaufmann, T., Bullock, J. S., et al. 2011, ApJL, 735, L1, doi: 10.1088/2041-8205/735/1/L1

    2011 doi

  62. [70]

    2021, ApJ, 916, 4, doi: 10.3847/1538-4357/ac033c

    Ferrari, F. 2021, ApJ, 916, 4, doi: 10.3847/1538-4357/ac033c

    2021 doi

  63. [71]

    S., & Werk, J

    Tumlinson, J., Peeples, M. S., & Werk, J. K. 2017, ARA&A, 55, 389, doi: 10.1146/annurev-astro-091916-055240

    2017 doi

  64. [72]

    J., Smith, B

    Turk, M. J., Smith, B. D., Oishi, J. S., et al. 2011, ApJS, 192, 9, doi: 10.1088/0067-0049/192/1/9

    2011 doi

  65. [73]

    2020, arXiv e-prints, arXiv:2008.02627, doi: 10.48550/arXiv.2008.02627

    Verdoja, F., & Kyrki, V. 2020, arXiv e-prints, arXiv:2008.02627, doi: 10.48550/arXiv.2008.02627

    2020 doi arXiv

  66. [74]

    M., Donahue, M., Bryan, G

    Voit, G. M., Donahue, M., Bryan, G. L., & McDonald, M. 2015, Nature, 519, 203, doi: 10.1038/nature14167

    2015 doi

  67. [75]

    M., Meece, G., Li, Y., et al

    Voit, G. M., Meece, G., Li, Y., et al. 2017, ApJ, 845, 80, doi: 10.3847/1538-4357/aa7d04

    2017 doi

  68. [76]

    2020, ApJS, 248, 32, doi: 10.3847/1538-4365/ab908c

    Weinberger, R., Springel, V., & Pakmor, R. 2020, ApJS, 248, 32, doi: 10.3847/1538-4365/ab908c

    2020 doi

  69. [78]

    2017b, MNRAS, 465, 3291, doi: 10.1093/mnras/stw2944 18

    Weinberger, R., Springel, V., Hernquist, L., et al. 2017b, MNRAS, 465, 3291, doi: 10.1093/mnras/stw2944 18

    doi

  70. [79]

    C., Sanderson, R

    Wetzel, A., Hayward, C. C., Sanderson, R. E., et al. 2023, ApJS, 265, 44, doi: 10.3847/1538-4365/acb99a

    2023 doi

  71. [80]

    J., Somerville, R

    Wright, R. J., Somerville, R. S., Lagos, C. d. P., et al. 2024, MNRAS, 532, 3417, doi: 10.1093/mnras/stae1688

    2024 doi

  72. [81]

    F., Wisotzki, L., et al

    Zabl, J., Bouch´ e, N. F., Wisotzki, L., et al. 2021, MNRAS, 507, 4294, doi: 10.1093/mnras/stab2165

    2021 doi

Pith tools

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