Using deep learning to characterize single-exposure double-line spectroscopic binaries

Avraham Binnenfeld; Raja Giryes; Rami Nasser; Samuel Lilek; Shay Zucker

REVIEW 3 major objections 5 minor 46 references

Using deep learning to characterize single-exposure double-line spectroscopic binaries

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A deep neural network trained on simulated Gaia spectra can recover the temperature, surface gravity, metallicity, and rotation of both stars in a double-line spectroscopic binary from a single exposure — no spectral disentangling or…

desk verdict A clean proof-of-concept for single-exposure SB2 parameter extraction, but the real-Gaia claim rests on a domain shift the paper itself admits it never tests. read the letter →

arxiv 2507.12363 v1 pith:5N7UT5JA submitted 2025-07-16 astro-ph.IM astro-ph.SR

Avraham Binnenfeld , Samuel Lilek , Rami Nasser , Raja Giryes , Shay Zucker This is my paper

classification astro-ph.IMastro-ph.SR

keywords spectroscopicbinariesSB2deeplearningGaiaRVSstellarparametersspectraldisentanglingsyntheticspectrasingle-exposurespectroscopy

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 claims that a deep convolutional network can read off the stellar parameters of both components of a double-line spectroscopic binary (SB2) from a single Gaia RVS exposure, skipping the classical steps of spectral disentangling and radial-velocity extraction. This matters because SB2 characterization today requires spectra from multiple orbital phases, which many survey targets do not have. The authors train a two-stage DenseNet on 200,000 simulated RVS spectra built from PHOENIX models and report test-set errors of about 394 K in $T_{\rm eff}$, 0.48 dex in $\log g$, 0.79 dex in $[\mathrm{Fe/H}]$, and 22 km s$^{-1}$ in $v \sin i$. If the synthetic-to-real transfer holds, single-exposure surveys gain a fast, $O(1)$ way to flag and characterize SB2 candidates.

What carries the argument

The load-bearing mechanism is a two-stage DenseNet pipeline with a deliberate ordering trick: the first network predicts the two component temperatures and is trained to output them lowest-first, and the second network takes the spectrum together with those ordered temperatures and predicts $\log g$, $[\mathrm{Fe/H}]$, and $v \sin i$ in the same order, so temperature ordering acts as a surrogate for component identity and replaces the disentangling step with an attribution-by-ordering scheme. On the data side, the synthetic generator that blends PHOENIX spectra under main-sequence mass-radius relations and circular orbits defines what the network can learn, including the paper's own stated degeneracy between the component flux ratio and relative line depths.

What would settle it

Apply the trained network to real Gaia RVS single exposures of known SB2s whose component parameters have been determined independently from multi-epoch observations (for example APOGEE or LAMOST data), and compare the network's per-component $T_{\rm eff}$ and $\log g$ to those values; the central claim fails if real-data errors exceed the reported test-set errors by roughly a factor of two or more, especially in the single-digit S/N regime that dominates individual RVS exposures.

Watch

Extended reading notes

Core claim

The central discovery is that the blended spectrum of an SB2 contains enough information for a network to predict each component's intrinsic parameters directly, without ever separating the two spectra or solving for their velocities. Trained on simulated Gaia RVS spectra (PHOENIX templates, circular orbits, main-sequence flux ratios, signal-to-noise 30–350), the network predicts the effective temperature, $\log g$, $[\mathrm{Fe/H}]$, and $v \sin i$ of both stars on a held-out test set, with each star's parameters attributed to the hotter or cooler component through a temperature-ordering scheme. The paper finds the method is hardly affected by the velocity separation between the components, degrades with lower metallicity and lower S/N, and performs poorly on single-star spectra — indicating the network has learned the specifically binary structure of the spectra rather than a generic stellar parameterization.

Load-bearing premise

The network must transfer from synthetic PHOENIX spectra at S/N 30–350 to real single-exposure Gaia RVS spectra, which are often far noisier (frequently single digits) and deviate from the template and main-sequence assumptions — a gap the authors themselves flag in their S/N analysis; if that transfer fails, the reported accuracies will not hold on real data.

Editorial extensions

If this is right

SB2 candidates can be characterized immediately from a single exposure, without solving for an orbit and without multi-epoch follow-up for a first physical readout.
Because prediction error is nearly independent of the velocity separation between the components, the method works for strongly blended, low-separation binaries where classical disentangling is hardest.
The reported $v \sin i$ and $[\mathrm{Fe/H}]$ accuracies are on the same scale as Gaia's own formal errors for single stars, making the outputs usable as survey-level estimates, though the authors advise against model-sensitive uses of the $v \sin i$ values.
The tool's poor performance on single-star spectra makes an SB2 detector a necessary front end, and the authors recommend pairing the two stages into a single pipeline.
The approach can be transplanted to other large spectroscopic surveys such as APOGEE, LAMOST, GALAH, and 4MOST, where targets with one or a few exposures pose the same limitation.

Reading between the lines

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

Given the paper's own flux-ratio/line-depth degeneracy caveat, the network is effectively fitting relative, flux-scaled spectra, so on real stars the temperature ordering and relative parameters may transfer far better than absolute calibration, which likely inherits systematic offsets from the PHOENIX main-sequence mass-radius assumption.
The separation-independence result suggests a cheap test: retrain the identical architecture on APOGEE or LAMOST-MRS single exposures, and if it generalizes, those surveys gain the same one-epoch SB2 readout without any Gaia-specific mask.
Nothing in the method requires deciding first that a target is an SB2; the $O(1)$ runtime makes it natural to run the network inside a survey pipeline as a triage layer that feeds candidates to multi-epoch RV follow-up, reversing today's 'classify first, characterize later' order.

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.

Desk Editor's Note

read the letter

Quick take: this paper does one thing and does it honestly — it shows a DenseNet can extract component Teff, log g, [Fe/H], vsini from a single synthetic blended RVS spectrum without disentangling. That's new relative to Zhang et al. 2022 and Jing et al. 2024, which only detect SB2s, and to El-Badry/Kovalev tools that fit multi-epoch or other instruments. The internal evaluation is clean: proper train/test split, 200k spectra, metrics reported both overall and broken out by S/N, metallicity, and RV separation. They also flag the main caveat themselves: Sect. 4.2 says real single RVS exposures often have S/N in the single digits, far below their 30–350 training range. So the reader's stress-test is on the mark — the abstract's expectation of usefulness for Gaia analysis is not demonstrated.

Soft spots, in order of size. First, the domain shift is load-bearing. Everything is in-distribution on PHOENIX + main-sequence mass-radius relationships. No real spectra, no low-S/N test, no instrument effects like continuum normalization, sky residuals, or CCD gaps. Figure 5 shows degradation toward S/N 30 but never probes the realistic single-digit regime. The claim that this will be useful for Gaia is therefore an extrapolation, not a result. Second, no code or data were released despite the abstract promising simulated data will be made available; that makes the reproducibility claim uncheckable. Third, no baseline comparison — I'd like to see how a simple simultaneous fit of two PHOENIX templates to the same single exposure does, to justify the DNN. On the plus side, the paper doesn't oversell: they explicitly say not to use v sin i for precise work, they discuss the flux-ratio degeneracy, and they warn about SB2s with very different metallicities. That's honest.

The temperature-ordering convention is a bit ad hoc but not a flaw — it's a labeling choice and the differences are reported separately.

My verdict: proof-of-concept, not more. The central argument — that a CNN can learn to parameterize SB2 components from a single idealized spectrum — holds up on the internal evidence. The real-Gaia usefulness claim does not. That's a referee-fixable gap, not a fatal one: add low-S/N experiments, show performance extrapolating or retrain to S/N ~5–20, test on a handful of real RVS spectra if any exist, and release code. The paper deserves peer review, but the review should push on the transfer question hard.

I'd bring this to reading group as a useful example of honest ML-in-astro pipeline evaluation, but I wouldn't cite it yet for real-data applications.

Referee Report

3 major / 5 minor

Summary. The paper proposes and evaluates a deep-learning approach to estimate Teff, log g, [Fe/H], and vsini of both components of a double-line spectroscopic binary (SB2) from a single mock Gaia RVS spectrum. The authors generate 200,000 synthetic SB2 spectra from PHOENIX models, train separate DenseNet models (one for temperatures and one for the remaining parameters), and report RMSE and MAE on a 20,000-spectrum held-out test set. They also analyze performance as a function of metallicity, S/N, and RV separation, and they test the model on single-star spectra as a control. The central claim is that a single-epoch spectrum suffices for approximate component characterization without spectral disentangling or RV extraction, and that the tool will be useful for analyzing upcoming Gaia RVS data.

Significance. If the in-distribution results transferred to real RVS observations, the method would be a fast and useful tool for preliminary SB2 characterization in surveys with few epochs per target. The paper has real strengths: the test set is properly held out; the reported metrics are internally consistent; the paper includes an explicit single-star control; and the authors are transparent about the synthetic nature of the data and about several modeling assumptions. However, all quantitative results are in-distribution, and the paper itself notes that real single RVS exposures often have S/N in the single digits, far below the S/N=30\u2013350 range used in training. Until either the model is evaluated in the realistic regime or the claims are scaled back, the contribution is a proof-of-concept rather than a demonstrated tool for Gaia analysis.

major comments (3)

[\u00a74.2, Fig. 5] The paper's own Sect. 4.2 states that real single RVS exposures frequently have S/N in the single digits, but the training, validation, and test sets use S/N in the range 30\u2013350, and Fig. 5 does not probe below 30. Because the error curves are still rising as S/N decreases toward 30, the reported RMSEs in Tables 4 and 5 cannot be extrapolated to the regime in which the tool is actually proposed to be used. Please add an explicit evaluation at S/N values representative of real RVS single exposures (e.g., 3\u201320), or, if such an evaluation is not feasible with the current simulator, revise the abstract and Discussion to present the result as a proof-of-concept on idealized spectra and to avoid stating that the tool is expected to be useful for real Gaia analysis without qualification.
[\u00a72] The simulation description does not state that the PHOENIX spectra were convolved with an instrumental line-spread function corresponding to R\u224811,000. PHOENIX library spectra are typically provided at substantially higher resolving power, so the synthetic spectra may contain line profiles sharper than real RVS spectra. If the convolution was omitted, the network could exploit information that is not present in the real data, making even the idealized test metrics optimistic. Please state explicitly whether and how the RVS instrumental profile was applied, and if it was not, repeat the evaluation with an appropriate convolution.
[\u00a72, \u00a75] The simulation sets the component flux ratio using PHOENIX main-sequence mass\u2013radius\u2013flux relations, and the Discussion correctly notes the degeneracy between flux ratio and line depths (Sect. 5). For real SB2s, which can contain evolved or otherwise non-main-sequence components, the assumed flux ratio may be systematically wrong. The paper does not test the sensitivity of the network to violations of the main-sequence assumption. Please add a robustness experiment with off-main-sequence flux ratios, or explicitly restrict the applicability claim to main-sequence components.

minor comments (5)

[\u00a73] The 'lower temperature first' ordering convention resolves the component-label ambiguity only when the two temperatures differ by more than the typical Teff error; for near-equal temperature pairs the ordering is arbitrary and the assignment of the remaining parameters to a specific component is not meaningful. A sentence acknowledging this and quantifying the fraction of test pairs with |\u0394Teff| below the reported RMSE would help.
[\u00a74] The comparison of v sin i and [Fe/H] MAEs to Gaia single-star formal precisions (Katz et al. 2023; Recio-Blanco et al. 2023) is made without giving the compared numerical values; please provide the quantitative figures so the reader can assess the comparison.
[\u00a74.2] The statement that the highest S/N bucket shows 'marginally or slightly worse results' should be checked for statistical significance; with 20,000 test spectra this could reflect a real effect or a small number of outliers.
[\u00a74.1] The sentence explaining why [Fe/H] prediction improves at low metallicity is confusing: an 'almost complete absence of metal lines' would seem to make the metallicity harder, not easier, to constrain once the spectrum becomes nearly line-free; please clarify the reasoning or add a diagnostic.
[\u00a75] The Discussion says the method is 'effective ... including in a low S/N regime', but the lowest simulated S/N is 30 while Sect. 4.2 notes real single exposures often have S/N in the single digits; please harmonize the wording with the actual tested range.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the method is a standard supervised-learning benchmark on held-out simulated spectra; the only self-citations are contextual and non-load-bearing.

full rationale

The paper's derivation chain is a supervised-learning pipeline: synthetic SB2 spectra are generated from PHOENIX models with known labels, a DenseNet is trained on 150,000 examples, and metrics are reported on a held-out 20,000-spectrum test set drawn from the same generator. This is in-distribution evaluation, not circularity: the test labels are not used in training, and the network must learn the mapping from the blended spectrum to the sorted component parameters. The temperature-ordering convention in Sect. 3 ('the temperature values in the target output array were reordered from lowest to highest') is a labeling device to resolve the inherent component ambiguity; it does not supply the target values, because the network still has to predict them from the input spectrum. The flux-ratio prescription using PHOENIX main-sequence masses and radii in Sect. 2 is a physical modeling choice, not a fit to the test outputs. The only self-citations (e.g., 'Similarly to the simulation presented by Binnenfeld et al. (2020)' and contextual mentions of Zucker & Giryes 2018 and Zucker & Mazeh 1994) are non-load-bearing; no uniqueness theorem or prior result is invoked to force the architecture or forbid alternatives. Section 4.2 honestly notes that real single RVS exposures often have S/N in the single digits while the simulated range is 30–350; this is an external-validity limitation, explicitly acknowledged, not a circular reduction. No step reduces the reported predictions to their inputs by construction.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

The central claim depends on a chain of simulation assumptions. The most important is that PHOENIX synthetic spectra, broadened and blended under main-sequence assumptions, represent real Gaia RVS observations well enough for a network trained on them to transfer. The chosen S/N range (30-350) is optimistic relative to single-exposure RVS conditions.

free parameters (3)

Stellar parameter ranges in simulation = Teff 3500-6500 K; log g 3-6 dex; [Fe/H] -4 to +1 dex; vsini 0-100 km/s
Hand-selected bounds defining the training domain; no external justification given for these limits, and the grid-aligned interpolation restricts coverage.
Signal-to-noise ratio range = 30-350
Training and test noise range; this is more optimistic than real single-exposure RVS data, which the paper states often reach only single digits.
Maximum RV separation K1+K2 = 10-90 km/s
Orbital semi-amplitude range used to simulate blended spectra; chosen without justification.

assumptions (7)

domain assumption PHOENIX synthetic spectra reliably represent real Gaia RVS stellar spectra
The network is trained exclusively on PHOENIX-based synthetic spectra; if this library is inaccurate for real stars, the model cannot transfer. Invoked throughout Sects. 2 and 5.
ad hoc to paper Interpolating one PHOENIX grid axis at a time preserves physical consistency
The paper avoids multi-axis interpolation to prevent artifacts, but this restricts the simulated parameter space to grid-aligned combinations and may not reflect real continuous parameter coupling. Section 2.
domain assumption Main-sequence mass-radius relation assumed for component masses, radii, and flux ratios
K1, K2 and the intensity ratio are derived from PHOENIX main-sequence properties; real SB2 components may be evolved or off the main sequence. Section 2 and discussion of flux-ratio degeneracy in Sect. 5.
domain assumption Circular orbits assumed for all simulated systems
Eccentric orbits are not simulated; real SB2s can have eccentric orbits and the spectral line separation pattern differs. Section 2.
ad hoc to paper The temperature-ordering convention (lower first) resolves the component identity ambiguity
The network is forced to output temperatures in ascending order; this imposes an arbitrary labeling that may not correspond to the physical primary/secondary in all cases. Section 3.
domain assumption Synthetic training distribution is representative of future real Gaia RVS data
The model is never tested on real spectra; the expectation of transferability is stated but not validated. Sects. 4.2 and 5.
domain assumption The added noise model approximates real Gaia RVS noise statistics
Paper only specifies a S/N range, not the noise distribution; real detector noise has wavelength-dependent and Poissonian characteristics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Using deep learning to characterize single-exposure double-line spectroscopic binaries." pith.science (2026). https://pith.science/paper/5N7UT5JA

@misc{pith2026250712363,
  author       = {Pith},
  title        = {Pith review of: Using deep learning to characterize single-exposure double-line spectroscopic binaries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5N7UT5JA}},
  note         = {Machine review of arXiv:2507.12363}
}

read the original abstract

Distinguishing the component spectra of double-line spectroscopic binaries (SB2s) and extracting their stellar parameters is a complex and computationally intensive task that usually requires observations spanning several epochs that represent various orbital phases. This poses an especially significant challenge for large surveys such as Gaia or LAMOST, where the number of available spectra per target is often not enough for a proper spectral disentangling. We present a new approach for characterizing SB2 components from single-exposure spectroscopic observations. The proposed tool uses deep neural networks to extract the stellar parameters of the individual component spectra that comprise the single exposure, without explicitly disentangling them or extracting their radial velocities. The neural networks were trained, tested, and validated using simulated data resembling Gaia RVS spectra, which will be made available to the community in the coming Gaia data releases. We expect our tool to be useful in their analysis.

Figures

Figures reproduced from arXiv: 2507.12363 by the authors.

**Figure 1.** Example of a characteristic 20-epoch RV time series of an SB2 system generated in the process described in Sect. 2. The red circles mark the primary RV, and the blue triangles show the RV of the secondary. Top panel: RVs plotted against observation times. Bottom panel: Same RV time-series, phase-folded by the known 81-day period. Teff, log g, or [Fe/H]) while keeping the other parameters fixed at their original gri… view at source ↗

**Figure 2.** Underlying spectra of the two components from the simulated system presented in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Simulated against predicted values in the test set for all predicted parameters. The models were generally successful in assessing the approximate value of the given parameters. Some parameters were more easily learned than others, however. While our model is clearly able to make strong predictions for effective temperature and log g, the model outputs for v sin i and metallicity exhibited a higher variance. It is … view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: √ MSE and MAE loss vs. metallicity for all stellar parameters. to reach S/N values of about 20 or slightly higher (Gaia Collaboration et al. 2023b), single RVS exposures usually have a much lower S/N, often in the single digits, depending on the magnitude and spectral…

**Figure 6.** Figure 6: √ MSE and MAE loss vs. component RV separation for all stellar parameters. we generated a dedicated test set consisting of single-star spectra and processed it using our models. The results indicate a poor performance: For all parameters, the square root of the MSE an…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 32 canonical work pages

[1]

M., Lim, P

Astropy Collaboration, Price-Whelan, A. M., Lim, P. L., et al. 2022, ApJ, 935, 167 Astropy Collaboration, Price-Whelan, A. M., Sip˝ocz, B. M., et al. 2018, AJ, 156, 123 Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33

2022
[2]

& Gies, D

Bagnuolo, William G., J. & Gies, D. R. 1991, ApJ, 376, 266

work page 1991
[3]

Bailer-Jones, C. A. L., Irwin, M., Gilmore, G., & von Hippel, T. 1997, MNRAS, 292, 157

work page 1997
[4]

2019, arXiv e-prints [arXiv:1904.07248]

Baron, D. 2019, arXiv e-prints [arXiv:1904.07248]

arXiv 2019
[5]

2020, A&A, 642, A146

Binnenfeld, A., Shahaf, S., & Zucker, S. 2020, A&A, 642, A146

work page 2020
[6]

Brooks, R. A. & Chiro, G. D. 1976, Phys. Med. Biol., 21, 689

work page 1976
[7]

2018, A&A, 616, A5

Cropper, M., Katz, D., Sartoretti, P., et al. 2018, A&A, 616, A5

work page 2018
[8]

2012, Res

Cui, X.-Q., Zhao, Y .-H., Chu, Y .-Q., et al. 2012, Res. Astron. Astrophys., 12, 1197

2012

Show all 46 references

[9]

P., et al

Czesla, S., Schröter, S., Schneider, C. P., et al. 2019, PyA: Python astronomy- related packages

2019
[10]

2016, A&A, 594, A68 de Jong, R

Dafonte, C., Fustes, D., Manteiga, M., et al. 2016, A&A, 594, A68 de Jong, R. S., Agertz, O., Berbel, A. A., et al. 2019, The Messenger, 175, 3 De Silva, G. M., Freeman, K. C., Bland-Hawthorn, J., et al. 2015, MNRAS, 449, 2604

2016
[11]

2009, in 2009 IEEE Conference on Com- puter Vision and Pattern Recognition, 248–255

Deng, J., Dong, W., Socher, R., et al. 2009, in 2009 IEEE Conference on Com- puter Vision and Pattern Recognition, 248–255

2009
[12]

2022, AJ, 163, 237

Dvash, E., Peleg, Y ., Zucker, S., & Giryes, R. 2022, AJ, 163, 237

2022
[13]

2022, MNRAS, 512, 5620

El-Badry, K., Seeburger, R., Jayasinghe, T., et al. 2022, MNRAS, 512, 5620

2022
[14]

2018b, MNRAS, 476, 528 Gaia Collaboration, Prusti, T., de Bruijne, J

El-Badry, K., Ting, Y .-S., Rix, H.-W., et al. 2018b, MNRAS, 476, 528 Gaia Collaboration, Prusti, T., de Bruijne, J. H. J., et al. 2016, A&A, 595, A1 Gaia Collaboration, Vallenari, A., Brown, A. G. A., et al. 2023a, A&A, 674, A1 Gaia Collaboration, Vallenari, A., Brown, A. G. ...

2016
[15]

2022, Open As- tron., 31, 38

Gebran, M., Connick, K., Farhat, H., Paletou, F., & Bentley, I. 2022, Open As- tron., 31, 38

2022
[16]

2012, The Messenger, 147, 25

Gilmore, G., Randich, S., Asplund, M., et al. 2012, The Messenger, 147, 25

2012
[17]

2004, PAICz, 92, 15

Hadrava, P. 2004, PAICz, 92, 15

2004
[18]

R., Millman, K

Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357

2020
[19]

1993, in Astronomical Society of the Pacific Conference Series, V ol

Hill, G. 1993, in Astronomical Society of the Pacific Conference Series, V ol. 38, New Frontiers in Binary Star Research, ed. K.-C. Leung & I.-S. Nha, 127

1993
[20]

Huang, G., Liu, Z., van der Maaten, L., & Weinberger, K. Q. 2018, in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2261

2018
[21]

O., Wende-von Berg, S., Dreizler, S., et al

Husser, T. O., Wende-von Berg, S., Dreizler, S., et al. 2013, A&A, 553, A6

2013
[22]

2024, arXiv e-prints, arXiv:2411.03994

Jing, Y ., Mao, T.-X., Wang, J., Liu, C., & Chen, X. 2024, arXiv e-prints, arXiv:2411.03994

2024 arXiv
[23]

2019, Open Astron., 28, 68

Kassounian, S., Gebran, M., Paletou, F., & Watson, V . 2019, Open Astron., 28, 68

2019
[24]

2010, in EAS Publications Series, V ol

Katz, D., Cropper, M., Meynadier, F., et al. 2010, in EAS Publications Series, V ol. 45, EAS Publications Series, ed. C. Turon, F. Meynadier, & F. Arenou, 189–194

2010
[25]

2023, A&A, 674, A5

Katz, D., Sartoretti, P., Guerrier, A., et al. 2023, A&A, 674, A5

2023
[26]

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

2014 arXiv
[27]

R., Stassun, K

Kounkel, M., Covey, K. R., Stassun, K. G., et al. 2021, AJ, 162, 184

2021
[28]

2022, MNRAS, 517, 356

Kovalev, M., Chen, X., & Han, Z. 2022, MNRAS, 517, 356

2022
[29]

& Straumit, I

Kovalev, M. & Straumit, I. 2022, MNRAS, 510, 1515

2022
[30]

2024, MNRAS, 527, 521

Kovalev, M., Zhou, Z., Chen, X., & Han, Z. 2024, MNRAS, 527, 521

2024
[31]

Krizhevsky, A., Sutskever, I., & Hinton, G. E. 2012, in Advances in Neural Infor- mation Processing Systems, ed. F. Pereira, C. Burges, L. Bottou, & K. Wein- berger, V ol. 25 (Curran Associates, Inc.)

2012
[32]

2020, arXiv e-prints, arXiv:2005.07210

Liu, C., Fu, J., Shi, J., et al. 2020, arXiv e-prints, arXiv:2005.07210

2020 arXiv
[33]

N., Omukai, K., Matsumoto, T., & Inutsuka, S.-I

Machida, M. N., Omukai, K., Matsumoto, T., & Inutsuka, S.-I. 2009, MNRAS, 399, 1255

2009
[34]

R., Schiavon, R

Majewski, S. R., Schiavon, R. P., Frinchaboy, P. M., et al. 2017, AJ, 154, 94

2017
[35]

2013, The Messenger, 154, 47

Randich, S., Gilmore, G., & Gaia-ESO Consortium. 2013, The Messenger, 154, 47

2013
[36]

A., et al

Recio-Blanco, A., de Laverny, P., Palicio, P. A., et al. 2023, A&A, 674, A29

2023
[37]

P., Järvinen, S., & Weber, M

Sablowski, D. P., Järvinen, S., & Weber, M. 2019, A&A, 623, A31

2019
[38]

2023, A&A, 674, A6

Sartoretti, P., Marchal, O., Babusiaux, C., et al. 2023, A&A, 674, A6

2023
[39]

2020, Astrophysics Source Code Library, ascl:2007.022

Shahaf, S., Binnenfeld, A., Mazeh, T., & Zucker, S. 2020, Astrophysics Source Code Library, ascl:2007.022

2020
[40]

Simon, K. P. & Sturm, E. 1994, A&A, 281, 286

1994
[41]

2020, A&A, 643, A122

Skoda, P., Podsztavek, O., & Tvrdík, P. 2020, A&A, 643, A122

2020
[42]

E., et al

Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, Nat. Methods, 17, 261 von Hippel, T., Storrie-Lombardi, L. J., Storrie-Lombardi, M. C., & Irwin, M. J. 1994, MNRAS, 269, 97

2020
[43]

H., Nikolaou, N., Coronica, P., et al

Yip, K. H., Nikolaou, N., Coronica, P., et al. 2020, Lect. Notes Comput. Sci., 322–338

2020
[44]

2022, ApJS, 258, 26

Zhang, B., Jing, Y .-J., Yang, F., et al. 2022, ApJS, 258, 26

2022
[45]

& Giryes, R

Zucker, S. & Giryes, R. 2018, AJ, 155, 147

2018
[46]

& Mazeh, T

Zucker, S. & Mazeh, T. 1994, ApJ, 420, 806 Article number, page 8 of 8

1994

Pith tools

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

[1] [1]

M., Lim, P

Astropy Collaboration, Price-Whelan, A. M., Lim, P. L., et al. 2022, ApJ, 935, 167 Astropy Collaboration, Price-Whelan, A. M., Sip˝ocz, B. M., et al. 2018, AJ, 156, 123 Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33

2022

[2] [2]

& Gies, D

Bagnuolo, William G., J. & Gies, D. R. 1991, ApJ, 376, 266

work page 1991

[3] [3]

Bailer-Jones, C. A. L., Irwin, M., Gilmore, G., & von Hippel, T. 1997, MNRAS, 292, 157

work page 1997

[4] [4]

2019, arXiv e-prints [arXiv:1904.07248]

Baron, D. 2019, arXiv e-prints [arXiv:1904.07248]

arXiv 2019

[5] [5]

2020, A&A, 642, A146

Binnenfeld, A., Shahaf, S., & Zucker, S. 2020, A&A, 642, A146

work page 2020

[6] [6]

Brooks, R. A. & Chiro, G. D. 1976, Phys. Med. Biol., 21, 689

work page 1976

[7] [7]

2018, A&A, 616, A5

Cropper, M., Katz, D., Sartoretti, P., et al. 2018, A&A, 616, A5

work page 2018

[8] [8]

2012, Res

Cui, X.-Q., Zhao, Y .-H., Chu, Y .-Q., et al. 2012, Res. Astron. Astrophys., 12, 1197

2012