pith. machine review for the scientific record. sign in

arxiv: 2605.04434 · v1 · submitted 2026-05-06 · 🌌 astro-ph.GA · astro-ph.IM

Recognition: 3 theorem links

· Lean Theorem

A CNN--Transformer Denoiser for low-S/N Galaxy Spectra: Stellar Population Recovery in Synthetic Tests

Suk Kim , Joon Hyeop Lee , Soo-Chang Rey
Authors on Pith no claims yet

Pith reviewed 2026-05-08 18:04 UTC · model grok-4.3

classification 🌌 astro-ph.GA astro-ph.IM
keywords galaxy spectradenoisingstellar populationsCNN-Transformerlow signal-to-noisesynthetic spectrapPXF fitting
0
0 comments X

The pith

A CNN-Transformer denoiser cuts RMS residuals in low-S/N galaxy spectra by 96.5 percent at S/N=5 in synthetic tests.

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

The paper tests whether a hybrid deep-learning model can clean noisy galaxy spectra sufficiently to recover stellar population properties like age and metallicity without spatial binning. Low signal-to-noise ratios restrict measurements in faint regions of galaxies observed by integral field unit surveys. The authors train the Enhanced U-Net Transformer on 90,000 synthetic spectra built from MILES models with added wavelength-dependent noise that mimics SAMI data. In tests on 10,000 held-out spectra the model sharply reduces residuals across the full wavelength range and in key absorption lines while preserving line shapes. Downstream fitting then yields tighter constraints on mass-weighted age and metallicity.

Core claim

The Enhanced U-Net Transformer reduces the full-spectrum RMS residual by about 96.5 percent at S/N = 5 and about 94 percent at S/N = 20 on an independent test set, with recovery rates of at least 99.8 percent; in pPXF fits the RMS scatter in mass-weighted age drops from 0.41 to 0.25 dex at S/N = 5 and from 0.32 to 0.22 dex at S/N = 10, while [M/H] scatter improves from 0.45 to 0.36 dex and from 0.32 to 0.28 dex respectively.

What carries the argument

The Enhanced U-Net Transformer (EUT), a one-dimensional hybrid CNN-Transformer architecture that learns a direct mapping from noisy input spectra to denoised outputs by combining local convolutional feature extraction with global transformer attention.

If this is right

  • Residuals in fixed windows around Ca II H, Hdelta, Hbeta, Fe I 4383, Mg b and Na D drop by more than 88 percent while line profiles remain intact.
  • At S/N = 20 the denoised and noisy inputs produce statistically consistent stellar-population fits within the synthetic-test uncertainties.
  • The method reduces the need for aggressive spatial binning to reach usable S/N, potentially preserving spatial resolution in galaxy maps.
  • Recovery rates stay above 99.8 percent across the tested S/N range, indicating stable behavior on the synthetic distribution.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the synthetic-test gains hold on real data, the approach could enable stellar-population maps at the native spatial sampling of IFU instruments rather than binned scales.
  • The same hybrid architecture might be retrained on spectra from other instruments or wavelength ranges once suitable synthetic training sets exist.
  • The reduction in age scatter from 0.41 to 0.25 dex at S/N = 5 corresponds to distinguishing stellar populations separated by roughly 1.4 Gyr instead of 2.6 Gyr at typical galaxy ages.

Load-bearing premise

That performance gains measured on synthetic spectra with injected SAMI-like noise will transfer to real observed galaxy spectra without introducing new systematic biases in the recovered stellar populations.

What would settle it

Apply the trained EUT to real low-S/N spaxels from an integral-field survey and compare the pPXF-derived ages and metallicities against independent measurements obtained from the same regions after spatial binning to much higher S/N.

Figures

Figures reproduced from arXiv: 2605.04434 by Joon Hyeop Lee, Soo-Chang Rey, Suk Kim.

Figure 1
Figure 1. Figure 1: Schematic of the Enhanced U-Net Transformer (EUT) architecture. The network takes a synthetic galaxy spectrum and a noise-added realization as input (left) and predicts a denoised spectrum (right); the loss is evaluated against the corresponding noise-free spectrum. The encoder (blue, left) has three 1D convolutional stages with channel dimensions 1 → 96 → 192 → 384. Each stage applies a convolution (kerne… view at source ↗
Figure 2
Figure 2. Figure 2: Training (black) and validation (red) loss as a function of epoch. The losses decrease rapidly at early epochs and then approach a plateau; the minimum validation loss occurs at epoch 974. We adopt the corresponding checkpoint for subsequent analysis. still substantial, improvement of 88.4±6.9%. This lower value likely reflects that Na D is weak in a subset of the synthetic spectra, making the residuals mo… view at source ↗
Figure 3
Figure 3. Figure 3: Example of full-spectrum denoising results for a representative synthetic spectrum at four input signal-to-noise ratios (S/N = 5, 10, 15, and 20). In each panel, the top plot displays the noise-free reference (blue), the noisy input (red), and the EUT-denoised output (cyan). The bottom plot shows the residuals relative to the noise-free spectrum for both the noisy input (red) and the denoised output (cyan)… view at source ↗
Figure 4
Figure 4. Figure 4: Full-spectrum denoising statistics for 10,000 synthetic test spectra at input S/N = 5, 10, 15, and 20. (a) Box-and-whisker plots of per-spectrum RMS residuals relative to the corresponding noise-free spectra, comparing noisy in￾puts (red) and EUT-denoised outputs (cyan). (b) Distributions of RMS residuals for the denoised outputs; vertical ticks mark medians, and the legend lists the mean (µ) and standard … view at source ↗
Figure 5
Figure 5. Figure 5: Example of absorption-line recovery at input S/N = 5 and 10. Columns correspond to six diagnostic features: Ca II H, Hδ, Fe i 4383, Hβ, Mg b, and Na D. In each column, the upper panel overlays the noise-free spectrum (blue), the noisy inputs (red), and the denoised outputs (cyan). Solid lines represent S/N = 5, while dashed lines represent S/N = 10. The shaded regions mark the fixed wavelength windows used… view at source ↗
Figure 6
Figure 6. Figure 6: Absorption-line denoising statistics at input S/N = 5 for six diagnostic features (Ca II H, Hδ, Fe I 4383, Hβ, Mg b, and Na D). Measurements use the fixed wavelength windows highlighted in view at source ↗
Figure 7
Figure 7. Figure 7: Recovery of mass-weighted age from noisy and EUT-denoised spectra. Each panel compares the input age [log(age/yr); x-axis] with the value recovered by pPXF (y-axis) for 104 synthetic spectra. Colors show the logarithmic number density, and the black line indicates the one-to-one relation. The top row shows results for the noisy spectra, and the bottom row shows results for the corresponding denoised spectr… view at source ↗
Figure 8
Figure 8. Figure 8: Same as view at source ↗
read the original abstract

Stellar population measurements in integral field unit surveys are often limited by low signal-to-noise ratios (S/N) in low-surface-brightness spaxels. Using controlled synthetic experiments, we test whether deep-learning-based denoising can recover stellar population information without spatial binning. We introduce the Enhanced U-Net Transformer (EUT), a one-dimensional CNN-Transformer model trained on 90,000 synthetic spectra constructed from MILES simple stellar population models following Lee et al. (2023). Wavelength-dependent noise is injected on the fly to emulate SAMI-like data with S/N = 5-20, measured in a 4484.77-4573.12 Angstrom continuum window. On an independent test set of 10,000 spectra, EUT reduces the full-spectrum RMS residual by about 96.5 percent at S/N = 5 and about 94 percent at S/N = 20, with recovery rates of at least 99.8 percent. In fixed windows around Ca II H, Hdelta, Hbeta, Fe I 4383, Mg b, and Na D, residuals decrease by more than about 88 percent while preserving line-profile structure. In downstream pPXF fitting, the RMS scatter in recovered mass-weighted age decreases from about 0.41 to 0.25 dex at S/N = 5 and from about 0.32 to 0.22 dex at S/N = 10. For mass-weighted metallicity, [M/H], the scatter decreases from about 0.45 to 0.36 dex and from about 0.32 to 0.28 dex, respectively. At S/N = 20, denoised and noisy inputs give consistent results within the synthetic-test uncertainties. These experiments suggest that hybrid CNN-Transformer denoisers can improve low-S/N spectra for stellar population studies, although validation with observed spectra is still required.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 3 minor

Summary. The paper introduces the Enhanced U-Net Transformer (EUT), a hybrid 1D CNN-Transformer denoiser trained on 90,000 synthetic galaxy spectra constructed from MILES SSP models. Wavelength-dependent noise is injected on-the-fly to emulate SAMI-like observations at S/N = 5–20 (measured in the 4484.77–4573.12 Å window). On a held-out test set of 10,000 spectra, EUT reduces full-spectrum RMS residuals by ~96.5% at S/N=5 and ~94% at S/N=20 (recovery rate ≥99.8%), with >88% residual reduction in key line windows while preserving profiles. Downstream pPXF fits show reduced scatter in mass-weighted age (0.41→0.25 dex at S/N=5) and [M/H] (0.45→0.36 dex at S/N=5). The work is framed as synthetic tests and explicitly notes that real-data validation remains required.

Significance. If the synthetic-test gains generalize, the method could enable stellar-population analysis of individual low-S/N spaxels in IFU surveys without spatial binning, preserving spatial resolution in low-surface-brightness regions. Strengths include the use of an independent test set, quantitative RMS and parameter-recovery metrics, integration with the established pPXF code, and the hybrid architecture’s ability to handle both local features and global context. The explicit caveat on real-data validation is appropriate. The primary limitation is that all results rest on MILES-based synthetics with a modeled noise distribution; broader impact therefore hinges on future real-data tests.

major comments (1)
  1. [§3 (Results, pPXF subsection)] §3 (Results, pPXF subsection): the reported RMS scatter reductions for age and metallicity are quantified only at S/N=5 and S/N=10; the statement that results are “consistent within uncertainties” at S/N=20 is qualitative. Providing the corresponding numerical scatter values at S/N=20 would make the cross-S/N comparison load-bearing for the claim that denoising benefits diminish at higher S/N.
minor comments (3)
  1. [Abstract] Abstract: all improvement percentages are qualified by “about”; reporting the exact computed values (or 1-σ ranges) from the test set would improve precision and reproducibility.
  2. [§2 (Methods)] §2 (Methods): the precise functional form and parameters of the wavelength-dependent SAMI-like noise model are not fully specified; a short equation or pseudocode block would allow exact reproduction of the training distribution.
  3. [Figure captions] Figure captions (e.g., those showing example spectra): the distinction between noisy input, denoised output, and ground-truth should be stated explicitly in every relevant caption to avoid reader ambiguity.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive review and recommendation for minor revision. We address the major comment below.

read point-by-point responses

  1. Referee: [§3 (Results, pPXF subsection)] §3 (Results, pPXF subsection): the reported RMS scatter reductions for age and metallicity are quantified only at S/N=5 and S/N=10; the statement that results are “consistent within uncertainties” at S/N=20 is qualitative. Providing the corresponding numerical scatter values at S/N=20 would make the cross-S/N comparison load-bearing for the claim that denoising benefits diminish at higher S/N.

    Authors: We agree that reporting the explicit RMS scatter values at S/N=20 will strengthen the quantitative comparison across S/N regimes. In the revised manuscript we will add the corresponding numerical values for mass-weighted age and [M/H] (both noisy and denoised cases) at S/N=20. These values are directly available from the same synthetic test suite and confirm that the differences lie within the reported uncertainties, but their inclusion will make the statement that benefits diminish at higher S/N fully load-bearing. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper reports empirical results from training a new EUT model on 90,000 synthetic spectra generated from MILES SSP templates (with noise injection) and evaluating performance metrics on a fully independent held-out test set of 10,000 spectra. Downstream stellar-population recovery is quantified by feeding both noisy and denoised spectra into the established external pPXF code and measuring scatter reductions. The single citation to Lee et al. (2023) is used only to describe the data-generation procedure and does not supply any load-bearing justification, uniqueness theorem, or fitted parameter for the reported RMS reductions or age/metallicity improvements. No equations, predictions, or central claims reduce by construction to the inputs or to self-citations; the test-set metrics are computed directly from the held-out data.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The work depends on the quality of the synthetic data generation and the transferability of results to real observations, which the abstract acknowledges requires further testing.

free parameters (2)
  • Neural network weights
    The parameters of the EUT model are learned from the training data.
  • S/N measurement window
    Specific continuum window (4484.77-4573.12 Angstrom) used to define S/N for noise injection.
axioms (2)
  • domain assumption MILES simple stellar population models can be used to generate synthetic spectra that are sufficiently realistic for training and testing a denoiser.
    Basis for the 90,000 training spectra as per Lee et al. (2023).
  • domain assumption pPXF is a reliable tool for recovering mass-weighted age and metallicity from the spectra.
    Used to measure the downstream impact of denoising.

pith-pipeline@v0.9.0 · 5670 in / 1714 out tokens · 150196 ms · 2026-05-08T18:04:03.336036+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages · 3 internal anchors

  1. [1]

    2017, MNRAS, 469, 151, doi: 10.1093/mnras/stx789

    Belfiore, F., Maiolino, R., Tremonti, C., et al. 2017, MNRAS, 469, 151, doi: 10.1093/mnras/stx789

    work page doi:10.1093/mnras/stx789 2017

  2. [2]

    Overview of the SDSS-IV MaNGA Survey: Mapping Nearby Galaxies at Apache Point Observatory

    Bundy, K., Bershady, M. A., Law, D. R., et al. 2015, ApJ, 798, 7, doi: 10.1088/0004-637X/798/1/7

    work page Pith review doi:10.1088/0004-637x/798/1/7 2015

  3. [3]

    Improving the Full Spectrum Fitting Method: Accurate Convolution with

    Cappellari, M. 2017, MNRAS, 466, 798, doi: 10.1093/mnras/stw3020

    work page internal anchor Pith review doi:10.1093/mnras/stw3020 2017

  4. [4]

    2003, MNRAS, 340, 1214, doi: 10.1046/j.1365-8711.2003.06380.x

    Cappellari, M., & Copin, Y. 2003, MNRAS, 342, 345, doi: 10.1046/j.1365-8711.2003.06541.x

    work page doi:10.1046/j.1365-8711.2003.06541.x 2003

  5. [5]

    Parametric

    Cappellari, M., & Emsellem, E. 2004, PASP, 116, 138, doi: 10.1086/381875 Cid Fernandes, R., P´ erez, E., Garc´ ıa Benito, R., et al. 2013, A&A, 557, A86, doi: 10.1051/0004-6361/201220616

    work page doi:10.1086/381875 2004

  6. [6]

    Conroy, C., & van Dokkum, P. G. 2012, ApJ, 760, 71, doi: 10.1088/0004-637X/760/1/71

    work page doi:10.1088/0004-637x/760/1/71 2012

  7. [7]

    P., D’Souza, R., Kauffmann, G., et al

    Cooper, A. P., D’Souza, R., Kauffmann, G., et al. 2013, MNRAS, 434, 3348, doi: 10.1093/mnras/stt1245

    work page doi:10.1093/mnras/stt1245 2013

  8. [8]

    2011, MNRAS, 418, 467, doi: 10.1111/j.1365-2966.2011.19497.x

    Croom, S. M., Lawrence, J. S., Bland-Hawthorn, J., et al. 2012, MNRAS, 421, 872, doi: 10.1111/j.1365-2966.2011.20365.x 14

    work page doi:10.1111/j.1365-2966.2011.20365.x 2012

  9. [9]

    2017, A&A, 603, A60, doi: 10.1051/0004-6361/201630240 Garc´ ıa-Lorenzo, B., M´ arquez, I., Barrera-Ballesteros, J

    Frontera-Pons, J., Sureau, F., Bobin, J., & Le Floc’h, E. 2017, A&A, 603, A60, doi: 10.1051/0004-6361/201630240 Garc´ ıa-Lorenzo, B., M´ arquez, I., Barrera-Ballesteros, J. K., et al. 2015, A&A, 573, A59, doi: 10.1051/0004-6361/201423485

    work page doi:10.1051/0004-6361/201630240 2017

  10. [10]

    2025, Open Astronomy, 34, 20240010, doi: 10.1515/astro-2024-0010

    Gebran, M., Bentley, I., Brienza, R., & Paletou, F. 2025, Open Astronomy, 34, 20240010, doi: 10.1515/astro-2024-0010

    work page doi:10.1515/astro-2024-0010 2025

  11. [11]

    2000 , month = feb, pages =

    Girardi, L., Bressan, A., Bertelli, G., & Chiosi, C. 2000, A&AS, 141, 371, doi: 10.1051/aas:2000126

    work page doi:10.1051/aas:2000126 2000

  12. [12]

    2017, MNRAS, 465, 688, doi: 10.1093/mnras/stw2719 Gonz´ alez Delgado, R

    Goddard, D., Thomas, D., Maraston, C., et al. 2017, MNRAS, 465, 688, doi: 10.1093/mnras/stw2719 Gonz´ alez Delgado, R. M., Cid Fernandes, R., Garc´ ıa-Benito, R., et al. 2014, ApJL, 791, L16, doi: 10.1088/2041-8205/791/1/L16 Gonz´ alez Delgado, R. M., Garc´ ıa-Benito, R., P´ erez, E., et al. 2015, A&A, 581, A103, doi: 10.1051/0004-6361/201525938

    work page doi:10.1093/mnras/stw2719 2017

  13. [13]

    F., et al

    Husemann, B., Jahnke, K., S´ anchez, S. F., et al. 2013, A&A, 549, A87, doi: 10.1051/0004-6361/201220582

    work page doi:10.1051/0004-6361/201220582 2013

  14. [14]

    2019, ApJL, 881, L30, doi: 10.3847/2041-8213/ab365f

    Jia, P., Huang, Y., Cai, B., & Cai, D. 2019, ApJL, 881, L30, doi: 10.3847/2041-8213/ab365f

    work page doi:10.3847/2041-8213/ab365f 2019

  15. [15]

    2009 , pages =

    Kriek, M., van Dokkum, P. G., Labb´ e, I., et al. 2009, ApJ, 700, 221, doi: 10.1088/0004-637X/700/1/221

    work page doi:10.1088/0004-637x/700/1/221 2009

  16. [16]

    R., Cherinka, B., Yan, R., et al

    Law, D. R., Cherinka, B., Yan, R., et al. 2016, AJ, 152, 83, doi: 10.3847/0004-6256/152/4/83

    work page doi:10.3847/0004-6256/152/4/83 2016

  17. [17]

    H., Pak, M., Jeong, H., & Oh, S

    Lee, J. H., Pak, M., Jeong, H., & Oh, S. 2023, MNRAS, 521, 4207, doi: 10.1093/mnras/stad814

    work page doi:10.1093/mnras/stad814 2023

  18. [18]

    W., & Bovy, J

    Leung, H. W., & Bovy, J. 2024, MNRAS, 527, 1494, doi: 10.1093/mnras/stad3015

    work page doi:10.1093/mnras/stad3015 2024

  19. [19]

    2010, ApJ, 725, 2312, doi: 10.1088/0004-637X/725/2/2312

    Burkert, A. 2010, ApJ, 725, 2312, doi: 10.1088/0004-637X/725/2/2312 P´ erez, E., Cid Fernandes, R., Gonz´ alez Delgado, R. M., et al. 2013, ApJL, 764, L1, doi: 10.1088/2041-8205/764/1/L1

    work page doi:10.1088/0004-637x/725/2/2312 2010

  20. [20]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    Ronneberger, O., Fischer, P., & Brox, T. 2015, arXiv e-prints, arXiv:1505.04597, doi: 10.48550/arXiv.1505.04597 S´ anchez, S. F., Kennicutt, R. C., Gil de Paz, A., et al. 2012, A&A, 538, A8, doi: 10.1051/0004-6361/201117353 S´ anchez, S. F., Barrera-Ballesteros, J. K., Colombo, D., et al. 2021, MNRAS, 503, 1615, doi: 10.1093/mnras/stab442 S´ anchez-Bl´ az...

    work page internal anchor Pith review doi:10.48550/arxiv.1505.04597 2015

  21. [21]

    Santhanam, G. K. 2017, MNRAS, 467, L110, doi: 10.1093/mnrasl/slx008

    work page doi:10.1093/mnrasl/slx008 2017

  22. [22]

    2023, MNRAS, 526, 3037, doi: 10.1093/mnras/stad2709

    Scourfield, M., Saintonge, A., de Mijolla, D., & Viti, S. 2023, MNRAS, 526, 3037, doi: 10.1093/mnras/stad2709

    work page doi:10.1093/mnras/stad2709 2023

  23. [23]

    Attention Is All You Need

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

    work page internal anchor Pith review doi:10.48550/arxiv.1706.03762 2017

  24. [24]

    J., Lucas, P

    Vazdekis, A., S´ anchez-Bl´ azquez, P., Falc´ on-Barroso, J., et al. 2010, MNRAS, 404, 1639, doi: 10.1111/j.1365-2966.2010.16407.x

    work page doi:10.1111/j.1365-2966.2010.16407.x 2010

  25. [25]

    , keywords =

    Vojtekova, A., Lieu, M., Valtchanov, I., et al. 2021, MNRAS, 503, 3204, doi: 10.1093/mnras/staa3567

    work page doi:10.1093/mnras/staa3567 2021

  26. [26]

    and Cappellari, Michele and Bershady, Matthew A

    Westfall, K. B., Cappellari, M., Bershady, M. A., et al. 2019, AJ, 158, 231, doi: 10.3847/1538-3881/ab44a2

    work page doi:10.3847/1538-3881/ab44a2 2019

  27. [27]

    M., Maraston, C., Thomas, D., et al

    Wilkinson, D. M., Maraston, C., Thomas, D., et al. 2015, MNRAS, 449, 328, doi: 10.1093/mnras/stv301

    work page doi:10.1093/mnras/stv301 2015

  28. [28]

    Worthey, G., & Ottaviani, D. L. 1997, ApJS, 111, 377, doi: 10.1086/313021

    work page doi:10.1086/313021 1997

  29. [29]

    2024, A&A, 683, A163, doi: 10.1051/0004-6361/202347994

    Zhang, M., Wu, F., Bu, Y., et al. 2024, A&A, 683, A163, doi: 10.1051/0004-6361/202347994

    work page doi:10.1051/0004-6361/202347994 2024

  30. [30]

    R., Heneka, C., et al

    Zhong, F., Napolitano, N. R., Heneka, C., et al. 2025, MNRAS, 543, 691, doi: 10.1093/mnras/staf1482

    work page doi:10.1093/mnras/staf1482 2025