Pith. sign in

REVIEW 4 major objections 3 minor 2 cited by

ProMage: fast galaxy magnitudes emulation combining SED forward-modelling and machine learning

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A feed-forward network emulates ProSpect galaxy magnitudes to <0.02 mag for 99% of test sources, running about 10,000 times faster.

desk verdict ProMage is a solid, useful emulator for ProSpect magnitudes, but the per-mille accuracy claim is overstated and the Stage IV suitability rests on in-prior validation only. read the letter →

arxiv 2509.00150 v1 pith:F6JU3ZJC submitted 2025-08-29 astro-ph.GA astro-ph.IM

Luca Tortorelli , Silvan Fischbacher , Aaron S. G. Robotham , Céline Nussbaumer , Alexandre Refregier This is my paper
classification astro-ph.GAastro-ph.IM
keywords galaxies:stellarcontentfundamentalparametersmethods:numericalstatisticsneuralnetworkemulatorspectralenergydistributionHyperSuprime-Camforwardmodelling
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 introduces ProMage, a feed-forward neural network that reproduces the observer- and rest-frame galaxy magnitudes computed by the ProSpect spectral energy distribution code. Trained on 10^7 ProSpect runs, it predicts magnitudes from physical galaxy parameters (redshift, star formation history, dust, gas) and reaches absolute errors below 0.02 magnitudes for 99% of a 10^6-galaxy test set across the HSC g, r, i, z, y bands, roughly 10^4 times faster than ProSpect. The authors argue this speed makes feasible large-scale forward-modelling and Bayesian inference of galaxy properties for Stage IV surveys such as Euclid and Rubin-LSST.

What carries the argument

A feed-forward neural network with five hidden layers (512-256-128-64-32 neurons) mapping 11 physical ProSpect inputs (redshift, star formation history parameters, dust and gas parameters, metallicity) to a single magnitude per band, trained per band with mean squared error loss on raw magnitudes and the Alsing et al. (2020) activation function. The speed-up comes from replacing a full SPS evaluation (tens of milliseconds in ProSpect) with one matrix multiplication chain (microseconds).

What would settle it

Run ProMage on 10^6 ProSpect realisations sampled from a prior that includes AGN components or extends to z > 5, and check whether more than 1% of emulated magnitudes deviate from ProSpect by more than 0.02 mag; also compare emulated magnitudes against real HSC photometry for a sample of galaxies with independently fitted physical parameters.

Watch

Extended reading notes

Core claim

The central claim is that a modest feed-forward network can stand in for a full stellar population synthesis calculation for the purpose of computing photometric magnitudes, achieving per-mille accuracy and a four-orders-of-magnitude speed-up. ProMage predicts each band independently from the same 11 physical inputs, trained per band with a custom activation function. On a held-out test set of 10^6 sources, 99% of emulated magnitudes lie within 0.02 mag of ProSpect's values in both observer and rest frames in all five HSC bands, below typical photometric zero-point uncertainties of current surveys. The paper positions this as a practical enabler of amortised simulation-based inference and MC

Load-bearing premise

Accuracy is only demonstrated on test galaxies drawn from the same Latin hypercube prior used for training, so the emulator's performance on galaxies outside that prior—higher redshifts, AGN-inclusive SEDs, or different filter sets—is untested.

Editorial extensions

If this is right

  • ProMage can compute photometric magnitudes for billions of galaxies in minutes on a single CPU, enabling large synthetic survey realisations for weak-lensing redshift calibration.
  • It makes amortised simulation-based inference practical: networks can be trained over the emulator to invert galaxy properties from observed magnitudes.
  • Observer- and rest-frame magnitudes in HSC g, r, i, z, y are cheap enough that MCMC chains over SPS parameters become feasible at survey scale.
  • The per-band, independent design means adding new filters or expanding parameter ranges only requires retraining that band, not the whole model.

Reading between the lines

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

  • If the <0.02 mag accuracy holds at the edges of the prior, extending ProMage to galaxies with AGN or different dust geometries would likely require retraining or an added correction term; the paper notes AGN are excluded.
  • Because the emulator is fast and differentiable, it could be coupled directly to gradient-based inference or used in simulation-based calibration of survey systematics.
  • The accuracy numbers are reported on a Latin-hypercube test set; real galaxy photometry includes noise and selection effects, so end-to-end validation on observed HSC data would strengthen the generalization claim.
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 / 3 minor

Summary. The paper presents ProMage, a feed-forward neural network that emulates observer- and rest-frame magnitudes computed by the SED package ProSpect in the HSC g,r,i,z,y bands. The network is trained on 10^7 galaxies drawn from a Latin hypercube prior over 11 physical parameters (redshift, SFH, gas, dust), with per-band and per-frame networks. On a held-out 10^6 galaxy test set, the authors report absolute errors <0.02 mag for 99% of sources and <0.01 mag for 95%, a ~10^4 speed-up over ProSpect, and state that the emulator is well suited for Stage IV surveys and forward-modelling frameworks such as GalSBI-SPS.

Significance. If the reported performance is robust, ProMage would be a practically useful component for fast SED-based inference and synthetic catalogue generation: the architecture is simple, per-band training enables parallelization and extension, and the 10^6 source test set is substantial. The speed-up claim is plausible and well aligned with the needs of simulation-based inference. However, the headline accuracy metric is ambiguous and internally inconsistent, and the generalization claims go beyond the evidence presented. The core emulation result is credible as an in-prior interpolation statement, but the survey-scale suitability claim requires additional validation or explicit scoping.

major comments (4)
  1. [Abstract and §4] The abstract claims 'per-mille relative accuracy for 99% of sources', but §4 reports absolute errors <0.02 mag for 99% and <0.01 mag for 95%. In flux units, 0.01 mag is ~0.9% and 0.02 mag is ~1.8%, not 0.1% (per-mille). The relative accuracy metric is never defined. This is the headline quantitative claim and must be corrected or redefined (e.g., as median absolute percentage flux error or as a stated magnitude tolerance).
  2. [§3–§4] The validation is entirely in-prior: the test set is a random 10% split of the same Latin hypercube sample used for training, and the LHS is space-filling, so every test point lies in a densely sampled region of the training prior. This demonstrates interpolation, not generalization to the actual galaxy distribution. The statement that ProMage is 'well suited' for Stage IV surveys and for GalSBI-SPS (§4) is therefore not yet supported. Provide an out-of-distribution test, e.g., comparing against ProSpect on a realistic mock catalogue or on the posterior draws produced by GalSBI-SPS, or at minimum a leave-out-region test, and discuss the implied accuracy outside the prior.
  3. [§2] AGN are explicitly excluded 'for now' from the training set, yet ProSpect itself includes AGN emission and a large fraction of Stage IV galaxies host AGN, particularly at high redshift. The suitability claim in §4 is unqualified. Either extend the training to AGN-inclusive SEDs, or explicitly scope the claim to non-AGN galaxies and assess the expected impact on Stage IV applications.
  4. [Fig. 1 / §4] Quantitative results are shown only for the g and i bands; the statement that 'similar performance occurs for the other HSC optical bands' is asserted without data. Since the abstract claims per-mille accuracy 'across the g,r,i,z,y bands', please provide a table or additional panels reporting the 95th and 99th percentile absolute errors for r, z, and y in both observer and rest frames.
minor comments (3)
  1. [Table 1] The prior range for mpeak is given as '[−(2 + tlb), 13.4 − tlb]' but tlb is never defined in the text. Please define this quantity or replace with an explicit numerical range. Also, there is a typo in 'αSF,,screen'.
  2. [§2–§3] The definition of rest-frame magnitudes is not stated. Please specify how the rest-frame band is computed (e.g., filter transmission shifted to rest wavelength, K-correction conventions, and whether the same filter set is used). This is needed for reproducibility.
  3. [§4] The timing statements '10^5 sources in less than half a second' (Introduction) and '10^6 sources in under three seconds' (Results) are roughly consistent but should be accompanied by the exact hardware and timing methodology (e.g., batch size, number of repeated runs) for a fair speed comparison with ProSpect.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the accuracy claim is an out-of-sample emulator test against the same generator used to produce the training set, which is appropriate for an emulator; the survey-suitability claim is an untested extrapolation rather than a circular step.

full rationale

The paper's derivation chain is an empirical surrogate construction: (1) a 10^7-galaxy training sample is drawn from a Latin hypercube prior (Section 2), (2) ProSpect computes the target magnitudes, (3) a feed-forward network is trained on an 80% split and tested on a separate 10% split (Section 3). The test set is held out from training, so the reported absolute errors <0.02 mag / <0.01 mag are genuine out-of-sample predictions of the emulator within the prior, not the same points used for fitting. Because the emulator's definition is 'ProSpect magnitude emulator', agreement with ProSpect magnitudes is the correct target; it is self-consistency by construction but not a reasoning loop. The paper does not fit a parameter and then rename it as a prediction, invoke a uniqueness theorem from prior work, or smuggle an ansatz through citations. Self-citations to ProSpect (Robotham et al. 2020) and GALSBI-SPS (Tortorelli et al. 2025) are used as tools/frameworks, not as unverified premises for the emulator's accuracy. The genuine weaknesses are not circularity: the validation is in-prior and AGN-free, so the 'well suited for Stage IV surveys' statement extrapolates beyond the tested distribution; and the abstract's 'per-mille relative accuracy' is inconsistent with the reported 0.01–0.02 mag absolute errors (0.01 mag corresponds to ~0.9% flux, i.e. sub-percent but not per-mille). Those are external-validity and metric-precision concerns, not circular reductions.

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

The model contains trained weights and hand-chosen priors/hyperparameters that set the domain of the accuracy claim. It introduces no new physical entities; it is a computational surrogate for an existing SED code.

free parameters (4)
  • Neural network weights and biases of five dense layers = Fitted on 10^7 ProSpect samples
    The entire predictive mapping is learned; accuracy is a property of these fitted values.
  • Activation function gamma and beta = Initialized 1.0 and 0.1, optimized during training
    These activation parameters are listed in Section 3 as optimized and affect fit quality.
  • Training hyperparameters = lr 10^-3, batch size 64, 200 epochs max, early stopping 20
    Chosen by hand; no sensitivity analysis is shown, so they are part of the unvalidated recipe.
  • Input prior ranges (Table 1) = e.g., redshift in [0,5], mSFR in [-3,4]
    Choosing these bounds defines the training distribution and therefore the domain where the accuracy claim holds.
assumptions (4)
  • domain assumption ProSpect output is the correct ground-truth magnitude for training and testing
    All labels come from ProSpect; any ProSpect systematics are inherited (Section 2).
  • domain assumption The Table 1 Latin hypercube prior covers the parameter space relevant to Stage IV galaxy populations
    Accuracy is only measured in-distribution; no external or real galaxy validation is provided.
  • domain assumption The chosen feed-forward architecture can represent the ProSpect mapping to the claimed accuracy
    Empirically supported on the test set but no theoretical guarantee or architecture sensitivity analysis.
  • domain assumption Underlying SPS ingredients (MIST isochrones, C3K spectra, Chabrier IMF, MAPPINGS-III, Charlot and Fall dust) are valid descriptions of galaxies
    ProMage inherits these via ProSpect; cited in Section 2 as the physical basis for inference applications.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ProMage: fast galaxy magnitudes emulation combining SED forward-modelling and machine learning." pith.science (2026). https://pith.science/paper/F6JU3ZJC

@misc{pith2026250900150,
  author       = {Pith},
  title        = {Pith review of: ProMage: fast galaxy magnitudes emulation combining SED forward-modelling and machine learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F6JU3ZJC}},
  note         = {Machine review of arXiv:2509.00150}
}
abstract

We present ProMage, a feed-forward neural network that emulates the computation of observer- and rest-frame magnitudes from the generative galaxy SED package ProSpect. The network predicts magnitudes conditioned on input galaxy physical properties, including redshift, star formation history, gas and dust parameters. ProMage accelerates magnitude computation by a factor of $10^4$ compared to ProSpect, while achieving per-mille relative accuracy for $99\%$ of sources in the test set across the $g,r,i,z,y$ Hyper Suprime-Cam bands. This acceleration is key to enabling fast inference of galaxy physical properties in next-generation Stage IV surveys and to generating large catalogue realisations in forward-modelling frameworks such as GalSBI-SPS.

Figures

Figures reproduced from arXiv: 2509.00150 by the authors.

Figure 1
Figure 1. Figure showing the performance of PROMAGE in emulating galaxy magnitudes computed with PROSPECT. The left panels refer to the g-band, while the right panels to the i-band, with upper and lower panels referring to observer (‘obs’) and rest-frame (‘rest’) magnitudes, respectively. We report both the histograms of the prediction accuracy and their distribution as function of the true input magnitude. Red and blue bands… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. GalSBI: Forward Modelling Galaxy Clustering and Population

    astro-ph.CO 2026-06 unverdicted novelty 6.0 of 10

    GalSBI extended with optimal transport subhalo matching and SBI to forward-model galaxy population plus clustering, validated against DES Y3 and HSC data with public code release.

  2. Machine Learning Techniques for Astrophysics and Cosmology: Photometric Redshifts

    astro-ph.IM 2026-05 unverdicted novelty 3.0 of 10

    AI techniques for photometric redshift estimation have converged and are now limited by the size, systematics, and selection effects in spectroscopic training samples rather than by methodology.

Reference graph

Works this paper leans on

21 extracted references · 17 canonical work pages · cited by 2 Pith papers

  1. [1]

    2006, astro-ph/0609591

    Albrecht, A., Bernstein, G., Cahn, R., et al. 2006, astro-ph/0609591. Euclid Collaboration, Mellier, Y ., Abdurrouf, et al. 2025, Astronomy & Astrophysics, 697, A1. Ivezi´c, ˇZ., Kahn, S. M., Tyson, J. A., et al. 2019, The Astrophysical Journal, 873, 2,

    arXiv 2006

  2. [3]

    Newman, J. A. & Gruen, D. 2022, Annual Review of Astronomy and Astrophysics, 60,

    work page 2022

  3. [5]

    2020, Proceedings of the National Academy of Science, 117, 48, 30055

    Cranmer, K., Brehmer, J., & Louppe, G. 2020, Proceedings of the National Academy of Science, 117, 48, 30055. Thorp, S., Peiris, H. V ., Jagwani, G., et al. 2025, arXiv:2506.12122. Robotham, A. S. G. & Bellstedt, S. 2025, RAS Techniques and Instruments, 4, rzaf019. Bellstedt, S. & Robotham, A. S. G. 2025, Monthly Notices of the Royal Astronomical Society, 540, 3,

    arXiv 2020

  4. [7]

    2025, arXiv:2505.21610

    Tortorelli, L., Fischbacher, S., Gr¨un, D., et al. 2025, arXiv:2505.21610. Conroy, C. 2013, Annual Review of Astronomy and Astrophysics, 51, 1,

    arXiv 2025

  5. [8]

    G., et al

    Conroy, C., Villaume, A., van Dokkum, P. G., et al. 2018, The Astrophysical Journal, 854, 2,

    work page 2018

  6. [80]

    2024, Astronomy & Astrophysics, 689, A144

    Tortorelli, L., McCullough, J., & Gruen, D. 2024, Astronomy & Astrophysics, 689, A144

    work page 2024

  7. [83]

    2019, , arXiv:1912.01703

    Paszke, A., Gross, S., Massa, F., et al. 2019, , arXiv:1912.01703. Wright, A. H., Kuijken, K., Hildebrandt, H., et al. 2024, Astronomy & Astrophysics, 686, A170. Crenshaw, J. F., Kalmbach, J. B., Gagliano, A., et al. 2024, The Astronomical Journal, 168, 2,

    arXiv 2019

  8. [111]

    2022, The Astronomical Journal, 164, 5,

    DESI Collaboration, Abareshi, B., Aguilar, J., et al. 2022, The Astronomical Journal, 164, 5,

    work page 2022

Show all 21 references
  1. [139]

    2003, The Publications of the Astronomical Society of the Pacific, 115, 809,

    Chabrier, G. 2003, The Publications of the Astronomical Society of the Pacific, 115, 809,

    2003

  2. [207]

    S., Agertz, O., Berbel, A

    de Jong, R. S., Agertz, O., Berbel, A. A., et al. 2019, The Messenger, 175,

    2019

  3. [363]

    2025, Journal of Cosmology and Astroparticle Physics, 2025, 6,

    Fischbacher, S., Kacprzak, T., Tortorelli, L., et al. 2025, Journal of Cosmology and Astroparticle Physics, 2025, 6,

    2025

  4. [393]

    E., & White, M

    Conroy, C., Gunn, J. E., & White, M. 2009, The Astrophysical Journal, 699, 1,

    2009

  5. [486]

    Robotham, A. S. G., Bellstedt, S., Lagos, C. del P., et al. 2020, Monthly Notices of the Royal Astronomical Society, 495, 1,

    2020

  6. [712]

    & Fall, S

    Charlot, S. & Fall, S. M. 2000, The Astrophysical Journal, 539, 2,

    2000

  7. [718]

    A., Helou, G., Magdis, G

    Dale, D. A., Helou, G., Magdis, G. E., et al. 2014, The Astrophysical Journal, 784, 1,

    2014

  8. [763]

    P., Robotham, A

    Driver, S. P., Robotham, A. S. G., Bland-Hawthorn, J., et al. 2013, Monthly Notices of the Royal Astronomical Society, 430, 4,

    2013

  9. [905]

    P., Chaves-Montero, J., Alarcon, A., et al

    Hearin, A. P., Chaves-Montero, J., Alarcon, A., et al. 2023, Monthly Notices of the Royal Astronomical Society, 521, 2,

    2023

  10. [1741]

    2020, The Astrophysical Journal Supplement Series, 249, 1,

    Alsing, J., Peiris, H., Leja, J., et al. 2020, The Astrophysical Journal Supplement Series, 249, 1,

    2020

  11. [2622]

    Bellstedt, S., Robotham, A. S. G., Driver, S. P., et al. 2021, Monthly Notices of the Royal Astronomical Society,, 503, 3,

    2021

  12. [2703]

    2016, The Astrophysical Journal Supplement Series, 222, 1,

    Dotter, A. 2016, The Astrophysical Journal Supplement Series, 222, 1,

    2016

  13. [3309]

    M., Kewley, L

    ProMage: fast galaxy magnitudes emulation 5 Levesque, E. M., Kewley, L. J., & Larson, K. L. 2010, The Astronomical Journal, 139, 2,

    2010

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.