arxiv: 2605.13973 · v1 · submitted 2026-05-13 · 🌌 astro-ph.GA

Recognition: 2 theorem links

· Lean Theorem

Determining star formation histories and age-metallicity relations with convolutional neural networks

Enrique Galceran , Patricia S\'anchez-Bl\'azquez , Artemi Camps-Fari\~na , M\'ed\'eric Boquien , Ralf S. Klessen , Francesco Belfiore , Daniel A. Dale , Francesca Pinna

show 3 more authors

Ivan S. Gerasimov Thomas G. Williams Hsi-An Pan

Authors on Pith no claims yet

Pith reviewed 2026-05-15 05:07 UTC · model grok-4.3

classification 🌌 astro-ph.GA

keywords star formation historyage-metallicity relationconvolutional neural networkgalaxy stellar populationsPHANGS surveyintegral field spectroscopyphotometrymachine learning

0 comments

The pith

A convolutional neural network recovers star formation histories and age-metallicity relations from combined spectra and photometry with negligible bias.

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

The paper trains a convolutional neural network on 165,000 synthetic spectra and photometric measurements to jointly infer star formation histories and metallicities in 16 age bins from PHANGS-MUSE spectroscopy and PHANGS-HST photometry. The network uses convolutional layers and attention mechanisms in a shared latent space to mitigate classical degeneracies while handling realistic noise and dust effects. It produces luminosity- and mass-weighted mean ages and metallicities with dispersions of about 0.12 dex in age and 0.03 dex in metallicity. The approach is thousands of times faster than traditional spectral fitting, enabling efficient analysis of large nearby-galaxy surveys.

Core claim

The CNN accurately recovers SFHs and age-metallicity relations over a wide range of evolutionary scenarios. The inferred luminosity- and mass-weighted mean ages and metallicities show negligible bias, with dispersions of ∼0.12 dex in age and ∼0.03 dex in metallicity. When applied to real PHANGS-MUSE and PHANGS-HST data for NGC 3627, the network produces smooth, spatially coherent maps of stellar age and metallicity that recover physically meaningful structures, including younger populations tracing the spiral arms and star-forming regions.

What carries the argument

A convolutional neural network with convolutional layers, attention mechanisms, and a shared latent space that jointly processes integral-field spectra and five-band photometry to predict star formation histories in 16 age bins along with metallicities.

If this is right

The network yields smooth, spatially coherent maps that trace younger stellar populations along spiral arms and star-forming regions in galaxies like NGC 3627.
Mean ages and metallicities are recovered with negligible bias across a broad range of SFH shapes and metallicity evolutions.
The method runs 5,000 to 20,000 times faster than conventional full spectral fitting, making it practical for large spectro-photometric surveys.
Joint use of spectroscopy and photometry improves constraints on spatially resolved star formation and metallicity evolution.

Where Pith is reading between the lines

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

The same architecture could be retrained on data from other integral-field surveys to map stellar populations across a wider range of galaxy types and environments.
Faster inference opens the possibility of applying detailed SFH recovery to statistically large samples rather than a handful of well-studied objects.
The shared latent space might allow the network to generalize to missing data modes, such as photometry-only or lower-resolution spectra, for incomplete observations.

Load-bearing premise

The 165,000 synthetic spectra and photometric measurements fully capture the degeneracies, dust attenuation, noise properties, and instrumental effects present in real observations.

What would settle it

Compare the CNN-derived age and metallicity maps for NGC 3627 directly against independent maps produced by traditional full spectral fitting codes on the same PHANGS-MUSE and PHANGS-HST data.

Figures

Figures reproduced from arXiv: 2605.13973 by Artemi Camps-Fari\~na, Daniel A. Dale, Enrique Galceran, Francesca Pinna, Francesco Belfiore, Hsi-An Pan, Ivan S. Gerasimov, M\'ed\'eric Boquien, Patricia S\'anchez-Bl\'azquez, Ralf S. Klessen, Thomas G. Williams.

**Figure 2.** Figure 2: Schematic representation of our CNN Network architecture as described in Section [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Four representative examples of the recovered SFH and age–metallicity relation compared to the ground truth, using our [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Mean difference between the input and the predicted SFR as a function of Age, normalised by the mean SFR in each bin. The error bars represent the standard deviation against the mean. where wi represents either the mass fraction (for mass-weighted quantities, XW = MW) or the V-band light fraction (for lightweighted quantities, XW = LW) contributed by the stellar population in age bin i. Although the mass… view at source ↗

**Figure 5.** Figure 5: Comparison between the ground truth mean mass- and luminosity-weighted ages (top row) and metallicities (bottom row) [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗

**Figure 6.** Figure 6: Differences between the input and predicted mean log(age) as a function of the corresponding differences in [Z/H]. Points are colour-coded by the ground-truth luminosityweighted metallicity as indicated in the insets. The top (bottom) panels show luminosity-weighted (mass-weighted) quantities. Left panels display the results obtained with the CNN, while the right panels show those derived using pPXF. di… view at source ↗

**Figure 8.** Figure 8: Violin plots showing the distributions of the di [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗

**Figure 9.** Figure 9: Impact of removing photometric fluxes from the input data on the recovery of the light-weighted age. Each panel shows the [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗

**Figure 10.** Figure 10: Impact of different spectral regions on the LW-age predictions. Each panel shows the difference between the groundtruth and predicted ⟨log(Age/yr)⟩LW for different perturbations of the input spectrum. The first nine panels (from left to right and top to bottom) show the results obtained when individual wavelength segments are replaced by random values drawn from the empirical distribution of the same seg… view at source ↗

**Figure 11.** Figure 11: Same as Fig [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗

**Figure 12.** Figure 12: Impact of different spectral regions on the predictions of the mean LW metallicity. The different panels are equivalent to those in [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗

**Figure 13.** Figure 13: Top row: maps of the luminosity- and mass-weighted stellar ages and metallicities predicted by the CNN for NGC 3627. [PITH_FULL_IMAGE:figures/full_fig_p015_13.png] view at source ↗

**Figure 14.** Figure 14: Comparison of luminosity- and mass-weighted mean [PITH_FULL_IMAGE:figures/full_fig_p016_14.png] view at source ↗

read the original abstract

We aim to develop a state-of-the-art tool to infer detailed star formation histories (SFHs) and age-metallicity relations from realistic observational data, while mitigating classical degeneracies and substantially reducing computational cost. In particular, we seek to exploit the complementarity of spectroscopic and photometric data to improve constraints on the spatially resolved SFH and metallicity evolution of nearby galaxies in the PHANGS collaboration. We construct and train a convolutional neural network (CNN) that combines convolutional layers, attention mechanisms, and a shared latent space to jointly predict SFHs and metallicities in 16 age bins. The network simultaneously processes integral-field spectroscopic data from PHANGS-MUSE and five-band photometric fluxes from PHANGS-HST. Training is performed on a dataset of 165\,000 synthetic spectra and photometric measurements spanning a broad range of SFH shapes, metallicity evolution, dust attenuation, and signal-to-noise ratios representative of the observations. The CNN accurately recovers SFHs and age-metallicity relations over a wide range of evolutionary scenarios. The inferred luminosity- and mass-weighted mean ages and metallicities show negligible bias, with dispersions of $\sim0.12$ dex in age and $\sim0.03$ dex in metallicity. When applied to real PHANGS-MUSE and PHANGS-HST data for NGC\,3627, the network produces smooth, spatially coherent maps of stellar age and metallicity that recover physically meaningful structures, including younger populations tracing the spiral arms and star-forming regions. The CNN is approximately $5\times10^{3}$--$2\times10^{4}$ times faster than traditional full spectral fitting codes, providing a powerful and efficient alternative for the analysis of large spectro-photometric surveys.

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.

Desk Editor's Note private letter to a colleague

The CNN recovers SFHs well on synthetics with low bias but the real-data test on NGC 3627 is only visual, so transferability remains the open question.

read the letter

The main thing here is a CNN that takes MUSE spectra plus HST photometry and spits out 16-bin SFHs plus metallicities in one forward pass. It is trained on 165k synthetics that cover a range of SFH shapes, dust, and noise, and it reports ~0.12 dex scatter in age and ~0.03 dex in metallicity with negligible bias on held-out synthetics. That speed-up (thousands of times faster than full spectral fitting) is the practical payoff if the accuracy holds up. The architecture itself—convolutional layers plus attention feeding a shared latent space—is a reasonable way to fuse the two data types and avoid separate pipelines. On the synthetic side the numbers look clean and the training distribution is stated explicitly, so the recovery metrics are at least internally consistent. The NGC 3627 maps are presented as smooth and physically plausible, with younger stars along the arms, which is encouraging but still only qualitative. The soft spot is exactly what the stress-test note flags: no quantitative comparison of the CNN output against traditional fitting on the same real spectra, and no clear test that the synthetic noise and dust models capture the residual sky, instrumental response, or complex dust geometry in the actual PHANGS observations. If those systematics shift the effective priors, the quoted dispersions could widen on real data. The paper is therefore most useful right now as a proof-of-concept for fast inference on large IFU+photometry samples, but it needs a side-by-side real-data benchmark before anyone should treat the precision numbers as ready for science. I would send it to review with a request for that comparison and for explicit checks on generalization across different galaxies or S/N regimes. It is worth a referee’s time because the speed advantage is real and the synthetic results are solid; the question is only how far they travel to observations.

Referee Report

2 major / 2 minor

Summary. The manuscript presents a convolutional neural network (CNN) trained on 165,000 synthetic spectra and five-band photometry to jointly recover star formation histories (SFHs) in 16 age bins and age-metallicity relations from combined PHANGS-MUSE spectroscopy and PHANGS-HST photometry. The network uses convolutional layers and attention mechanisms with a shared latent space. On held-out synthetic data it reports negligible bias and dispersions of ~0.12 dex in luminosity- and mass-weighted mean ages and ~0.03 dex in metallicities. When applied to NGC 3627 the CNN produces spatially coherent maps that recover expected structures such as younger populations along spiral arms.

Significance. If the synthetic recovery metrics generalize, the method would offer a 5,000- to 20,000-fold speed-up over traditional full spectral fitting while exploiting spectro-photometric complementarity to reduce classical degeneracies, enabling efficient analysis of large IFU surveys.

major comments (2)

[Application to NGC 3627 (results section)] The central claim that the CNN accurately recovers SFHs and age-metallicity relations on real data rests on visual inspection of NGC 3627 maps only. No quantitative comparison (e.g., pixel-by-pixel residuals or recovered mean ages/metallicities) against traditional fitting codes applied to the identical real spectra is reported, leaving open whether the quoted 0.12/0.03 dex dispersions survive unmodeled systematics such as residual sky lines, spatially varying instrumental response, or complex dust geometries.
[Methods (training dataset and network architecture)] The training description states that the 165,000 synthetics span a broad range of SFH shapes, metallicities, dust attenuation, and SNR, but provides insufficient detail on the train/validation/test split ratios, how noise properties were matched to real MUSE/HST observations, or any post-training generalization tests on out-of-distribution synthetics. These elements are load-bearing for the claim of negligible bias across evolutionary scenarios.

minor comments (2)

[Abstract and discussion] The speed-up range (5e3 to 2e4) is stated without specifying the hardware, number of age bins, or exact traditional code used for the comparison; a single benchmark table would clarify the claim.
[Methods] Notation for the 16 age bins and the precise definition of luminosity- versus mass-weighted means should be given explicitly in the methods to allow direct reproduction of the reported dispersions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments. We have revised the manuscript to provide additional details on the training procedure and to strengthen the discussion of real-data validation. Below we respond to each major comment.

read point-by-point responses

Referee: [Application to NGC 3627 (results section)] The central claim that the CNN accurately recovers SFHs and age-metallicity relations on real data rests on visual inspection of NGC 3627 maps only. No quantitative comparison (e.g., pixel-by-pixel residuals or recovered mean ages/metallicities) against traditional fitting codes applied to the identical real spectra is reported, leaving open whether the quoted 0.12/0.03 dex dispersions survive unmodeled systematics such as residual sky lines, spatially varying instrumental response, or complex dust geometries.

Authors: We agree that quantitative validation on real data would strengthen the presentation. The primary performance claims (negligible bias and quoted dispersions) are based on the held-out synthetic test set, while the NGC 3627 maps serve as an illustrative application demonstrating spatially coherent, physically plausible results. In the revised manuscript we have added a limited quantitative comparison: for a random subset of 500 spaxels we ran both the CNN and STARLIGHT, finding that the recovered luminosity-weighted mean ages and metallicities agree to within 0.15 dex and 0.05 dex rms, respectively. We also added a dedicated paragraph discussing the impact of unmodeled systematics (sky residuals, instrumental response, dust geometry) and why a perfect pixel-by-pixel match is not expected given differing modeling assumptions between the two approaches. revision: partial
Referee: [Methods (training dataset and network architecture)] The training description states that the 165,000 synthetics span a broad range of SFH shapes, metallicities, dust attenuation, and SNR, but provides insufficient detail on the train/validation/test split ratios, how noise properties were matched to real MUSE/HST observations, or any post-training generalization tests on out-of-distribution synthetics. These elements are load-bearing for the claim of negligible bias across evolutionary scenarios.

Authors: We thank the referee for highlighting these omissions. The revised Methods section now explicitly states the split ratios (70 % training, 15 % validation, 15 % test) and describes how realistic noise was injected: we used the per-spaxel variance spectra from the PHANGS-MUSE cubes and the photometric uncertainties from PHANGS-HST to generate noise realizations that match the observed SNR distributions. We have also added a new subsection reporting generalization tests on out-of-distribution synthetic spectra (including extreme bursty SFHs and metallicities outside the main training range), which show only modest degradation (age dispersion increases to 0.18 dex, metallicity to 0.05 dex) while bias remains negligible. revision: yes

Circularity Check

0 steps flagged

No significant circularity; recovery metrics computed on independent held-out synthetics

full rationale

The paper generates 165000 synthetic spectra and photometry from parametrized SFH shapes, metallicity evolution, dust curves, and noise models that are specified separately from the CNN architecture and loss. The reported negligible bias and dispersions (~0.12 dex age, ~0.03 dex metallicity) are measured by comparing CNN outputs against the known ground-truth labels of a held-out test subset of those synthetics; these statistics are therefore empirical performance measures rather than quantities defined by construction from fitted parameters or network weights. Application to real NGC 3627 data is presented only as qualitative maps of coherent structures with no quantitative self-referential benchmark that would close a loop. No load-bearing self-citations, uniqueness theorems, or ansatzes imported from prior author work are invoked to justify the central claims. The derivation from training distribution to reported accuracies therefore remains independent of the outputs themselves.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim depends on the assumption that synthetic training data adequately represent real observations and that the network generalizes beyond the training distribution; no new physical entities are introduced.

free parameters (1)

Number of age bins
SFH is discretized into 16 fixed age bins whose boundaries are chosen by the authors rather than derived from data.

axioms (1)

domain assumption Synthetic spectra and photometry generated from assumed SFH shapes, metallicity evolution, and dust laws accurately reproduce the statistical properties of real PHANGS observations.
All training and reported performance metrics rest on this coverage assumption.

pith-pipeline@v0.9.0 · 5673 in / 1420 out tokens · 67736 ms · 2026-05-15T05:07:37.264939+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We construct and train a convolutional neural network (CNN) that combines convolutional layers, attention mechanisms, and a shared latent space to jointly predict SFHs and metallicities in 16 age bins... Training is performed on a dataset of 165,000 synthetic spectra...
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The CNN accurately recovers SFHs and age-metallicity relations... dispersions of ∼0.12 dex in age and ∼0.03 dex in metallicity.

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

93 extracted references · 93 canonical work pages · 7 internal anchors

[1]

G., Groves, B

Allen, M. G., Groves, B. A., Dopita, M. A., Sutherland, R. S., & Kewley, L. J. 2008, ApJS, 178, 20–55

work page 2008
[2]

R., La Barbera, F., et al

Angthopo, J., Granett, B. R., La Barbera, F., et al. 2024, A&A, 690, A198

work page 2024
[3]

Anwar, O., Groves, B., Cortese, L., & Watts, A. B. 2025, PASA, 42, e102 Asa’d, R. S., Vazdekis, A., Cerviño, M., et al. 2017, MNRAS, 471, 3599

work page 2025
[4]

V ., Cid Fernandes, R., Stasi´nska, G., et al

Asari, N. V ., Cid Fernandes, R., Stasi´nska, G., et al. 2007, MNRAS, 381, 263

work page 2007
[5]

1994, A&AS, 106, 275

Bertelli, G., Bressan, A., Chiosi, C., Fagotto, F., & Nasi, E. 1994, A&AS, 106, 275

work page 1994
[6]

2019, A&A, 622, A103

Boquien, M., Burgarella, D., Roehlly, Y ., et al. 2019, A&A, 622, A103

work page 2019
[7]

2001, Machine Learning, 45, 5

Breiman, L. 2001, Machine Learning, 45, 5

work page 2001
[8]

2019, MNRAS, 483, 529

Cabayol, L., Sevilla-Noarbe, I., Fernández, E., et al. 2019, MNRAS, 483, 529

work page 2019
[9]

C., et al

Calzetti, D., Armus, L., Bohlin, R. C., et al. 2000, ApJ, 533, 682

work page 2000
[10]

2012, pPXF: Penalized Pixel-Fitting stellar kinematics extraction, Astrophysics Source Code Library, record ascl:1210.002

Cappellari, M. 2012, pPXF: Penalized Pixel-Fitting stellar kinematics extraction, Astrophysics Source Code Library, record ascl:1210.002

work page 2012
[11]

2023, MNRAS, 526, 3273

Cappellari, M. 2023, MNRAS, 526, 3273

work page 2023
[12]

A., Clayton, G

Cardelli, J. A., Clayton, G. C., & Mathis, J. S. 1989, ApJ, 345, 245

work page 1989
[13]

C., McLure, R

Carnall, A. C., McLure, R. J., Dunlop, J. S., & Davé, R. 2018, MNRAS, 480, 4379

work page 2018
[14]

J., Cardiel, N., Gorgas, J., et al

Cenarro, A. J., Cardiel, N., Gorgas, J., et al. 2001, MNRAS, 326, 959

work page 2001
[15]

2001, ApJ, 554, 1274

Chabrier, G. 2001, ApJ, 554, 1274

work page 2001
[16]

T., Thilker, D

Chandar, R., Barnes, A. T., Thilker, D. A., et al. 2025, AJ, 169, 150

work page 2025
[17]

& Fall, S

Charlot, S. & Fall, S. M. 2000, ApJ, 539, 718

work page 2000
[18]

J., Aragón-Salamanca, A., et al

Cheng, T.-Y ., Conselice, C. J., Aragón-Salamanca, A., et al. 2020, MNRAS, 493, 4209 Cid Fernandes, R., Gu, Q., Melnick, J., et al. 2004, MNRAS, 355, 273 Cid Fernandes, R., Mateus, A., Sodré, L., Stasi ´nska, G., & Gomes, J. M. 2005, MNRAS, 358, 363 Cid Fernandes, R., Pérez, E., García Benito, R., et al. 2013, A&A, 557, A86 da Cunha, E., Charlot, S., & El...

work page 2020
[19]

A., Helou, G., Magdis, G

Dale, D. A., Helou, G., Magdis, G. E., et al. 2014, ApJ, 784, 83

work page 2014
[20]

W., & Dambre, J

Dieleman, S., Willett, K. W., & Dambre, J. 2015, MNRAS, 450, 1441 Domínguez Sánchez, H., Coelho, P., Bruzual, G., et al. 2026, A&A, 705, A219 Domínguez Sánchez, H., Margalef, B., Bernardi, M., & Huertas-Company, M. 2022, MNRAS, 509, 4024 Domínguez Sánchez, H., Martin, G., Damjanov, I., et al. 2023, MNRAS, 521, 3861

work page 2015
[21]

A guide to convolution arithmetic for deep learning

Dumoulin, V . & Visin, F. 2016, A Guide to Convolution Arithmetic for Deep Learning, arXiv preprint arXiv:1603.07285

work page internal anchor Pith review Pith/arXiv arXiv 2016
[22]

2022, A&A, 659, A191

Emsellem, E., Schinnerer, E., Santoro, F., et al. 2022, A&A, 659, A191

work page 2022
[23]

A., O’Briain, T., et al

Fabbro, S., Venn, K. A., O’Briain, T., et al. 2018, MNRAS, 475, 2978

work page 2018
[24]

Farahani, A., Pourshojae, B., Rasheed, K., & Arabnia, H. R. 2021, arXiv e-prints, arXiv:2104.02144

work page arXiv 2021
[25]

2019, Journal of Machine Learning Re- search, 20, 1

Fisher, A., Rudin, C., & Dominici, F. 2019, Journal of Machine Learning Re- search, 20, 1

work page 2019
[26]

2018, MNRAS, 478, 2633

Ge, J., Yan, R., Cappellari, M., et al. 2018, MNRAS, 478, 2633

work page 2018
[27]

2014, MNRAS, 445, 175

Genel, S., V ogelsberger, M., Springel, V ., et al. 2014, MNRAS, 445, 175

work page 2014
[28]

2000, A&AS, 141, 371

Girardi, L., Bressan, A., Bertelli, G., & Chiosi, C. 2000, A&AS, 141, 371

work page 2000
[29]

& Bengio, Y

Glorot, X. & Bengio, Y . 2010, in Proceedings of Machine Learning Research, V ol. 9, Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, ed. Y . W. Teh & M. Titterington (Chia Laguna

work page 2010
[30]

2025, ApJ, submitted

Graham, G., Dale, D., & Chase, S. 2025, ApJ, submitted

work page 2025
[31]

D., Silva, D., Rayner, J., et al

Gregg, M. D., Silva, D., Rayner, J., et al. 2006, in The 2005 HST Calibration Workshop: Hubble After the Transition to Two-Gyro Mode, ed. A. M. Koeke- moer, P. Goudfrooij, & L. L. Dressel, 209

work page 2006
[32]

C., Lee, J

Hannon, S., Whitmore, B. C., Lee, J. C., et al. 2023, MNRAS, 526, 2991

work page 2023
[33]

& Robertson, B

Hausen, R. & Robertson, B. E. 2020, ApJS, 248, 20

work page 2020
[34]

Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification

He, K., Zhang, X., Ren, S., & Sun, J. 2015, arXiv e-prints, arXiv:1502.01852

work page internal anchor Pith review Pith/arXiv arXiv 2015
[35]

F., Somerville, R

Hopkins, P. F., Somerville, R. S., Hernquist, L., et al. 2006, ApJ, 652, 864

work page 2006
[36]

2015, ApJS, 221, 8

Huertas-Company, M., Gravet, R., Cabrera-Vives, G., et al. 2015, ApJS, 221, 8

work page 2015
[37]

H., & Pernet, E

Iglesias-Navarro, P., Huertas-Company, M., Martín-Navarro, I., Knapen, J. H., & Pernet, E. 2024, A&A, 689, A58

work page 2024
[38]

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Ioffe, S. & Szegedy, C. 2015, arXiv e-prints, arXiv:1502.03167

work page internal anchor Pith review Pith/arXiv arXiv 2015
[39]

G., Gawiser, E., Faber, S

Iyer, K. G., Gawiser, E., Faber, S. M., et al. 2019, ApJ, 879, 116

work page 2019
[40]

D., Leja, J., Conroy, C., & Speagle, J

Johnson, B. D., Leja, J., Conroy, C., & Speagle, J. S. 2021, ApJS, 254, 22

work page 2021
[41]

E., Ksoll, V

Kang, D. E., Ksoll, V . F., Itrich, D., et al. 2023, A&A, 674, A175

work page 2023
[42]

M., Yoon, S.-J., & Yi, S

Kaviraj, S., Rey, S.-C., Rich, R. M., Yoon, S.-J., & Yi, S. K. 2007, MNRAS, 381, L74

work page 2007
[43]

Khan, A., Sohail, A., Zahoora, U., & Qureshi, A. S. 2020, Artificial Intelligence Review, 53, 5455

work page 2020
[44]

Kingma, D. P. & Ba, J. 2014, arXiv e-prints, arXiv:1412.6980

work page internal anchor Pith review Pith/arXiv arXiv 2014
[45]

Kotz, S., Balakrishnan, N., & Johnson, N. L. 2000, Continuous multivariate dis- tributions. vol. 1, models and applications, V ol. 1 (A Wiley-Interscience Pub- lication John Wiley and Sons)

work page 2000
[46]

Krizhevsky, A., Sutskever, I., & Hinton, G. E. 2017, Commun. ACM, 60, 84–90

work page 2017
[47]

F., Reissl, S., Klessen, R

Ksoll, V . F., Reissl, S., Klessen, R. S., et al. 2024, A&A, 683, A246

work page 2024
[48]

2015, Nat, 521, 436

LeCun, Y ., Bengio, Y ., & Hinton, G. 2015, Nat, 521, 436

work page 2015
[49]

1998, Proceedings of the IEEE, 86, 2278

Lecun, Y ., Bottou, L., Bengio, Y ., & Haffner, P. 1998, Proceedings of the IEEE, 86, 2278

work page 1998
[50]

C., Sandstrom, K

Lee, J. C., Sandstrom, K. M., Leroy, A. K., et al. 2023, ApJ, 944, L17

work page 2023
[51]

C., Whitmore, B

Lee, J. C., Whitmore, B. C., Thilker, D. A., et al. 2022, ApJS, 258, 10

work page 2022
[52]

C., Johnson, B

Leja, J., Carnall, A. C., Johnson, B. D., Conroy, C., & Speagle, J. S. 2019, ApJ, 876, 3

work page 2019
[53]

K., Hughes, A., Liu, D., et al

Leroy, A. K., Hughes, A., Liu, D., et al. 2021, ApJS, 255, 19

work page 2021
[54]

M., Kewley, L

Levesque, E. M., Kewley, L. J., & Larson, K. L. 2010, AJ, 139, 712

work page 2010
[55]

2022, IEEE Transactions on Neural Net- works and Learning Systems, 33, 6999

Li, Z., Yang, W., Peng, S., & Liu, F. 2022, IEEE Transactions on Neural Net- works and Learning Systems, 33, 6999

work page 2022
[56]

L., Kawata, D., Sánchez-Blázquez, P., Ferreras, I., & Symeonidis, M

Liew-Cain, C. L., Kawata, D., Sánchez-Blázquez, P., Ferreras, I., & Symeonidis, M. 2021, MNRAS, 502, 1355

work page 2021
[57]

C., Acquaviva, V ., Thomas, P

Lovell, C. C., Acquaviva, V ., Thomas, P. A., et al. 2019, MNRAS, 490, 5503

work page 2019
[58]

2020, Proceedings of the National Academy of Science, 117, 7052

Lu, L., Dao, M., Kumar, P., et al. 2020, Proceedings of the National Academy of Science, 117, 7052

work page 2020
[59]

2015, PASP, 127, 16

Magris, G., Mateu, J., Mateu, C., et al. 2015, PASP, 127, 16

work page 2015
[60]

2024, MN- RAS, 531, 2864

Maksymowicz-Maciata, M., Spiniello, C., Martín-Navarro, I., et al. 2024, MN- RAS, 531, 2864

work page 2024
[61]

A., Rosales-Ortega, F

Marino, R. A., Rosales-Ortega, F. F., Sánchez, S. F., et al. 2013, A&A, 559, A114 Martín-Navarro, I., Vazdekis, A., Falcón-Barroso, J., et al. 2018, MNRAS, 475, 3700

work page 2013
[62]

C., Thilker, D

Maschmann, D., Lee, J. C., Thilker, D. A., et al. 2024, ApJS, 273, 14

work page 2024
[63]

UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

McInnes, L., Healy, J., & Melville, J. 2018, arXiv e-prints, arXiv:1802.03426

work page internal anchor Pith review Pith/arXiv arXiv 2018
[64]

2023, AJ, 166, 74

Melchior, P., Liang, Y ., Hahn, C., & Goulding, A. 2023, AJ, 166, 74

work page 2023
[65]

2009, A&A, 507, 1793

Noll, S., Burgarella, D., Giovannoli, E., et al. 2009, A&A, 507, 1793

work page 2009
[66]

F., & Jimenez, R

Panter, B., Heavens, A. F., & Jimenez, R. 2003, MNRAS, 343, 1145

work page 2003
[67]

2023, A&A, 673, A147

Pessa, I., Schinnerer, E., Sanchez-Blazquez, P., et al. 2023, A&A, 673, A147

work page 2023
[68]

2004, ApJ, 612, 168

Pietrinferni, A., Cassisi, S., Salaris, M., & Castelli, F. 2004, ApJ, 612, 168

work page 2004
[69]

R., Tout, C

Pols, O. R., Tout, C. A., Eggleton, P. P., & Han, Z. 1995, MNRAS, 274, 964

work page 1995
[70]

Robotham, A. S. G., Bellstedt, S., Lagos, C. d. P., et al. 2020, MNRAS, 495, 905

work page 2020
[71]

2015, A&A, 583, A60

Ruiz-Lara, T., Pérez, I., Gallart, C., et al. 2015, A&A, 583, A60

work page 2015
[72]

E., Hinton, G

Rumelhart, D. E., Hinton, G. E., & Williams, R. J. 1986, in Nat, V ol. 323 (Macmillan Education), 533–536

work page 1986
[73]

Saleh, R. A. & Ehsanes Saleh, A. K. M. 2022, arXiv e-prints, arXiv:2208.04564 Salvador-Rusiñol, N., Vazdekis, A., La Barbera, F., et al. 2020, Nature Astron- omy, 4, 252 Sánchez, S. F., Pérez, E., Sánchez-Blázquez, P., et al. 2016, Rev. Mexicana As- tron. Astrofis., 52, 21 Sánchez-Blázquez, P., Ocvirk, P., Gibson, B. K., Pérez, I., & Peletier, R. F. 2011,...

work page arXiv 2022
[74]

A., Bower, R

Schaye, J., Crain, R. A., Bower, R. G., et al. 2015, MNRAS, 446, 521

work page 2015
[75]

F., Jimenez, R., & Panter, B

Tojeiro, R., Heavens, A. F., Jimenez, R., & Panter, B. 2007, MNRAS, 381, 1252

work page 2007
[76]

2009, ARA&A, 47, 371

Tolstoy, E., Hill, V ., & Tosi, M. 2009, ARA&A, 47, 371

work page 2009
[77]

A., Dale, D

Turner, J. A., Dale, D. A., Lee, J. C., et al. 2021, MNRAS, 502, 1366

work page 2021
[78]

A., Singh, H

Valdes, F., Gupta, R., Rose, J. A., Singh, H. P., & Bell, D. J. 2004, ApJS, 152, 251 Article number, page 19 of 25 A&A proofs:manuscript no. aa58401-25

work page 2004
[79]

Attention Is All You Need

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

work page internal anchor Pith review Pith/arXiv arXiv 2017
[80]

2016, MNRAS, 463, 3409

Vazdekis, A., Koleva, M., Ricciardelli, E., Röck, B., & Falcón-Barroso, J. 2016, MNRAS, 463, 3409

work page 2016

Showing first 80 references.