Towards optimal photometric calibration of digital astronomical plates with deep learning

Bowen Huang; Haibo Yuan; Jianhai Zhao; Kai Xiao; Lin Yang; Meiting Yang; Mingyang Ma; Shiyin Shen; Yong Yu; Zhenghong Tang

REVIEW 4 major objections 4 minor 28 references

Towards optimal photometric calibration of digital astronomical plates with deep learning

T0 review · 4 major / 4 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 Deep learning doubles precision of historical plate photometry

desk verdict A credible mechanism and a new joint-calibration net, but the reported factor-of-two gain is measured on a test set that steered training; needs a clean split before I'd quote the numbers. read the letter →

arxiv 2608.01391 v1 pith:VFKAYFLU submitted 2026-08-02 astro-ph.IM

Mingyang Ma , Haibo Yuan , Lin Yang , Kai Xiao , Bowen Huang , Shiyin Shen , Zhengjun Shang , Yong Yu

show 3 more authors

Meiting Yang Zhenghong Tang Jianhai Zhao

This is my paper

classification astro-ph.IM

keywords photometriccalibrationphotographicplatesdeeplearningneuralnetworkmagnitude-positioncouplingimagequalitystellarphotometryMFF-Net

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that standard plate calibration, which treats magnitude, color, and position corrections as independent terms, leaves systematic errors when image quality varies with brightness across a plate. To fix this, the authors introduce MFF-Net, a neural network that takes magnitude, color, and pixel coordinates and outputs a single correction capturing their joint dependence. Applied to 1,200 digitized Chinese plates, it improves calibration precision from the 0.11–0.26 mag range to 0.08–0.18 mag, with the largest gains for bright stars. The result matters because historical plate archives hold over a century of sky observations, and better calibration makes them more scientifically usable.

What carries the argument

MFF-Net is a multi-layer fully connected network that takes four features—instrumental magnitude, color (BP−RP), and pixel coordinates (x, y)—and predicts the magnitude correction Δm = m_std − m_inst. It combines a fourth-order polynomial expansion in magnitude and second-order in color (matching the MYX25 formulation) with Fourier feature encoding of position, which allows the network to represent multi-scale spatial systematics. The explicit encodings let the network model the coupled, nonlinear dependencies among the input variables, directly addressing the failure of separable models.

What would settle it

Train MFF-Net on one random half of a plate's standard stars and validate on the other half, with no test-set feedback during training; if the precision gain over MYX25 shrinks or vanishes, the reported improvement is an artifact of the test set being used to drive training.

Watch

Extended reading notes

Core claim

The central discovery is that the magnitude–position coupling seen in post-calibration residuals of digitized photographic plates is driven by magnitude-dependent spatial variations in image quality (FWHM). Because the plate response depends on surface brightness rather than total flux, stars of equal total brightness but different image quality require different magnitude corrections. The paper shows through simulations that a separable model like MYX25 cannot capture this coupling, and that a neural network—MFF-Net—which jointly learns a nonlinear correction as a function of magnitude, color, and position, removes the residual structure and improves photometric precision by nearly a factor

Load-bearing premise

The reported accuracy assumes the 20% test split is a clean held-out set, but the training loop uses that same test set to decide learning-rate decay and when to stop, so generalization gains may be overstated.

Editorial extensions

If this is right

Historical plate archives such as those digitized in China can be recalibrated with MFF-Net to achieve sub-0.1 mag precision for bright stars, enabling more reliable time-domain and variability studies.
The method can be adapted to other plate collections that have sufficient standard-star coverage, potentially improving the photometric homogeneity of legacy sky surveys worldwide.
The finding that image-quality variations cause magnitude–position coupling suggests that future plate digitization efforts should record or model FWHM information to support more accurate calibration.
The ability to model cross-term corrections with a single network may inspire similar deep-learning approaches for other nonlinear detectors or non-standard photometric systems.

Reading between the lines

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

The same architecture could be applied to CCD data with spatially varying point-spread functions, where conventional flat-fielding assumes separability but PSF-induced magnitude biases may remain.
The paper's attribution of the coupling to FWHM suggests a testable physical model: if plate response depends on surface brightness, a FWHM-based correction term could be added to traditional separable models, potentially achieving MFF-Net's precision with a simpler, interpretable function.
Because the test set is used to drive learning-rate decay and early stopping, independent validation on truly held-out plates or a separate subset is needed to confirm that the reported gains generalize, a concern the paper does not explicitly address.

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

Short version: this paper identifies a real failure mode in separable plate-calibration models—magnitude-dependent FWHM structure that leaves coupled residuals—and shows a joint neural net fixes it on 1,200 Chinese plates. That is a new and useful result for the plate-archive community. But the headline numbers rest on a test set that is doing double duty as the validation set, so I wouldn't trust the factor-of-two until a clean split or cross-validation confirms it.

What's good: the residual analysis is careful, the FWHM mechanism is physically plausible (surface brightness vs. total flux), and the two simulation experiments are the right way to test it. The architecture is simple and sensible: polynomial features for magnitude/color, Fourier features for position, Huber loss, magnitude-bin resampling. The claim that MFF-Net edges out MYX25 on essentially all 1,200 plates is believable in direction, and the bright-source improvement is an important check since faint stars dominate the overall scatter. They also report training cost, which is helpful.

The soft spots: first and most load-bearing, the 'test set' is used for early stopping and the learning-rate schedule (every 20 epochs, test loss and mean residual trigger LR decay; training stops when LR bottoms out). That makes the test-set precision an optimistically biased estimate of generalization. The bias could be modest or large; there's no way to tell from the paper. Second, no uncertainties are given for the 5th–95th percentile ranges, so '0.11–0.26 to 0.08–0.18' has no error bars. Third, all hyperparameters were chosen on this dataset, which is fine for a proof of concept, but it further blurs the line between fit and evaluation. Fourth, the FWHM attribution rests on simulations using a plate response function that is never specified; if that response curve came from the same plates, the experiment has a whiff of circularity. They acknowledge this only indirectly—the claim is local to one plate.

None of these sink the paper. The central argument—that separable models leave magnitude–position coupling and a joint nonlinear correction helps—holds up. But the quantitative gain needs a re-evaluation with a true held-out set and error bars.

Who this is for: anyone working with digitized plate archives, and ML-for-astro workflows generally. I'd send it to review; a good referee can demand the clean split and response-function details. I'd also bring it to a reading group focused on calibration methods.

Referee Report

4 major / 4 minor

Summary. The paper introduces MFF-Net, a neural-network-based photometric calibration model for digitized astronomical plates. The network takes instrumental magnitude, color, and pixel coordinates, applies a polynomial expansion for magnitude/color and a Fourier feature encoding for position, and predicts a magnitude correction trained with a Huber loss using standard-star residuals. Using 1,200 Chinese plates, the authors report that MFF-Net outperforms the separable MYX25 method, improving the 5th–95th percentile precision from 0.11–0.26 mag to 0.08–0.18 mag and delivering about a factor-of-two gain for bright sources. They also use simulations on one plate to argue that magnitude-dependent spatial variations in FWHM drive the coupled residuals that MYX25 misses.

Significance. If the reported gains survive an unbiased evaluation, this is a practically valuable contribution: historical plate archives contain millions of plates, and a flexible nonlinear calibrator that removes brightness–position coupling could improve their photometric utility. The paper also makes a concrete, testable physical claim (FWHM-driven coupling) and provides a broad empirical comparison across 1,200 plates, with a simple architecture and modest computational cost (tens to ~100 s per plate on one GPU). The residual maps on the example plate show a visible reduction of annular structure compared with MYX25. However, the central quantitative claims currently rest on an evaluation protocol that uses the test set for training decisions, and they lack uncertainty estimates. The accepted strength of the empirical contribution is therefore not yet established.

major comments (4)

[§3, Training strategy] The text states: 'Every 20 epochs, we evaluate on the test set; if the mean residual over all sources falls below 0.025 and the test loss has decreased by less than 2x10^-4 since the previous evaluation, we reduce the learning rate by a factor of 0.2. Training terminates once the learning rate falls below 2x10^-6.' This explicitly uses the test set for learning-rate scheduling and early stopping, so the test set functions as a validation set. The reported precision on that set—the basis of the Abstract, Section 1, and Fig. 9—is therefore an optimistically biased estimate of generalization. The magnitude of the bias is unknown but potentially non-negligible given the network's flexibility and the repeated every-20-epoch decisions. The authors must re-run the evaluation with a genuinely held-out split (or a three-way split) and report the untouched test-set metrics.
[§4 / Fig. 9] No uncertainties are reported on any of the precision values. The headline numbers (0.11–0.26 mag vs 0.08–0.18 mag) are presented as point estimates over 1,200 plates, but there are no error bars, confidence intervals, or plate-to-plate scatter statistics. It is also unclear what 'precision' means exactly (per-plate RMS? median absolute residual?) and how the 5th–95th percentile range is computed. Without uncertainty estimates, the claimed improvements—especially the 'factor of two' for bright sources—cannot be distinguished from noise, particularly given the known contamination described above. The authors should provide bootstrapped confidence intervals and specify the exact statistic plotted in Fig. 9.
[§2, Simulation experiments] The attribution of coupled residuals to magnitude-dependent FWHM rests on simulations that use 'the plate response curve' to convert standard magnitudes to simulated instrumental magnitudes. The response curve is never specified: is it taken from the MYX25 calibration, fitted independently from standard stars, or assumed from first principles? If the response curve is derived from the very residuals being explained, the experiment is partly circular. Please state the functional form and provenance of the response curve, and ideally demonstrate that the simulation result is robust to reasonable variations in that curve.
[§3, Hyperparameters] The authors write that 'All hyperparameters were selected empirically on our dataset' and that MFF-Net is 'stable over a reasonable range of settings,' but no details or supporting experiments are given. Because the test set is already used for training decisions, the hyperparameter selection may also be influenced by the same set, further weakening the independence of the evaluation. The authors should describe the hyperparameter search, including which set was used for model selection, and provide ablation or stability evidence for the claimed robustness.

minor comments (4)

[Abstract / §3] The phrase 'Tests on 1,200 digitized Chinese plates' is slightly misleading: the evaluation is performed on a 20% star-level split per plate, not on 1,200 independent plate-level test sets. Please clarify this in the abstract or methods.
[Fig. 9] The definition of bright sources as m_inst < (m_min + m_max)/2 is arbitrary and plate-dependent. A sensitivity check (e.g., using the 25th percentile, or a fixed absolute magnitude cut when available) would strengthen the bright-source claim.
[Eq. (4)] The summation index appears as 'NM k=0' in the text; it should be 'N' or 'N-1' with a clear convention. The text also says '8 + 4×(N+1)' for the encoded feature dimension; please verify that this matches Eq. (4) as typeset.
[§2] The statement that 'when the image quality is spatially uniform, the residual maps remain correspondingly uniform across magnitude' is presented without quantitative support. If this is based on visual inspection, say so; if it is a statistical test, report it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the calibration target is external, though the 'test set' doubles as a validation set for early stopping, which is a separate evaluation-bias concern.

full rationale

The paper's central claim is a supervised regression: MFF-Net maps instrumental magnitude, color, and pixel coordinates to external standard magnitudes from the BEST catalog. The reported residuals are computed against those external labels, so the calibration result is not a tautology or a fit renamed as a prediction. The baseline MYX25 is a prior paper by the same group, but it functions as a comparison method, not as a load-bearing justification; no uniqueness theorem or untested ansatz is imported from it. The FWHM attribution is tested via forward simulations that inject FWHM-dependent surface-brightness effects and show that MYX25 reproduces the observed residual pattern; this is a controlled experiment, not a circular fit. The one notable methodological weakness is that the 'test set' is used for the learning-rate schedule and early stopping (Section 3: 'Every 20 epochs, we evaluate on the test set... Training terminates once the learning rate falls below 2e-6'), which makes the reported 5th-95th percentile precision an optimistically biased estimate of generalization. However, this is a validation leak, not a definitional circularity, and does not make the derivation equivalent to its inputs by construction. The FWHM simulation's plate response curve is not fully specified, but this is a reproducibility limitation rather than a circular step. Overall, the derivation chain is self-contained and the central claim has independent empirical content.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method is an empirical fit to standard-star data; the main free parameters are hyperparameters chosen by hand on the evaluation dataset. No new physical entities are introduced.

free parameters (6)

Fourier frequency count N (Eq. 5) = 2, 3, or 4 depending on standard-star count S
Chosen by hand to set spatial resolution of the Fourier feature encoding; affects the network's ability to model small-scale flat-field structure.
Polynomial expansion degree for magnitude = 4
Fourth order in m_inst, matching MYX25; selected to match the earlier formulation.
Polynomial expansion degree for color = 2
Second order in BP-RP; matching MYX25.
Huber loss threshold delta = 0.2
Set to the typical precision of MYX25; chosen empirically.
Hidden layer width and count = 128 neurons per layer; total count unspecified
Chosen empirically; exact number of layers is not stated, only that the first three use Leaky ReLU.
Learning rate schedule thresholds = initial 2e-4, decay factor 0.2, mean residual threshold 0.025, test-loss decrease 2e-4, terminate at 2e-6
Chosen empirically; the test set is used to trigger decay, a methodological concern.

assumptions (4)

domain assumption Photographic plates respond to surface brightness rather than total flux
Invoked in Section 2 to explain how FWHM variations produce magnitude-dependent spatial residuals; used to justify the network's inputs and the simulation experiments.
domain assumption The BEST standard-star catalog provides accurate reference magnitudes for calibration
The network and MYX25 are both trained and evaluated against these magnitudes; errors in the catalog would transfer to all results.
domain assumption A neural network with inputs (m_inst, BP-RP, x, y) can represent the true correction function
The method assumes the coupled systematics can be expressed as a function of these four features; if the true dependence involves other unobserved variables, the network cannot fully correct it.
domain assumption The plate response curve used in the simulations is representative of the actual plate
In Experiment 1 and 2, simulated instrumental magnitudes are derived from an unspecified plate response curve; the attribution of residuals to FWHM depends on this curve.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards optimal photometric calibration of digital astronomical plates with deep learning." pith.science (2026). https://pith.science/paper/VFKAYFLU

@misc{pith2026260801391,
  author       = {Pith},
  title        = {Pith review of: Towards optimal photometric calibration of digital astronomical plates with deep learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VFKAYFLU}},
  note         = {Machine review of arXiv:2608.01391}
}

read the original abstract

Photometric calibration of digitized photographic plates is commonly modeled with separable magnitude-, color-, and position-dependent terms, but this separability can break down when image quality varies across the field in a magnitude-dependent way, leaving coupled spatial systematics in the residuals. We introduce a deep-learning calibration framework, the Multi-Feature Fused Network (MFF-Net), which takes instrumental magnitude, color, and pixel coordinates as input and learns a single nonlinear correction that jointly captures their coupled dependencies. Tests on 1{,}200 digitized Chinese plates show that MFF-Net consistently outperforms the MYX25 method (Ma et al. 2025), improving the 5th--95th percentile precision from 0.11--0.26~mag to 0.08--0.18~mag and delivering an approximately factor-of-two gain for bright sources. The learned correction largely removes the magnitude--position coupling seen in post-calibration residual maps, enabling higher-precision plate photometry and more reliable use of large historical plate archives.

Figures

Figures reproduced from arXiv: 2608.01391 by the authors.

**Figure 1.** Calibration residuals for plate ZT5810N973001 as functions of instrumental magnitude, color, and pixel coordinates. The upper row shows results from MYX25, with a larger overall error of 0.24 mag; the lower row shows results from MFF-Net, reducing the error to 0.12 mag [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** MYX25 calibration residuals (left) and measured FWHM (right) for plate ZT5810N973001, shown in nine magnitude bins. The magnitude range and number of stars are labeled in the left panels, which share a common color scale. The right panels use individual color scales to highlight spatial structure (red: higher values; blue: lower values). to FWHM; because FWHM is often position-dependent, this produces the annular pa… view at source ↗

**Figure 3.** MYX25 calibration residuals (left) and FWHM of the simulated data from Experiment 1, plotted as in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 5.** Figure 5: MFF-Net architecture. unchanged. We adopt the Huber loss (Eq. 6) with transition threshold δ = 0.2, which matches the typical precision of MYX25. The Huber loss combines the advantages of mean squared error (MSE) and mean absolute error (MAE): it remains sensitive t…

**Figure 6.** Figure 6: MFF-Net correction as a function of magnitude (left) and color (right), evaluated at the field center of plate ZT5810N973001. In the left panel, curves correspond to different colors; in the right panel, curves correspond to different magnitudes [PITH_FULL_IMAGE:figur…

**Figure 7.** Figure 7: MFF-Net correction as a function of pixel coordinates for plate ZT5810N973001, shown for stars of different magnitudes at fixed BP − RP = 1. This work is supported by the National Natural Science Foundation of China (NSFC 12222301, 12173007) and the National Key Basic…

**Figure 8.** Figure 8: MFF-Net calibration residual of ZT5810N973001, plotted as in the left panel of [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

**Figure 9.** Figure 9: Photometric precision comparison (in magnitudes) for 1,200 plates (each with > 1, 000 reference stars). σMYX: MYX25 on all sources; σMFF: MFF-Net on all sources; σ Bright MFF : MFF-Net on bright sources only; σ Bright MYX : MYX25 on bright sources only. Bright sources …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 17 canonical work pages

[1]

T., Mildenhall, B., Tancik, M., et al.\ 2021, 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 580

Barron, J. T., Mildenhall, B., Tancik, M., et al.\ 2021, 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 580. doi:10.1109/ICCV48922.2021.00580

arXiv 2021
[2]

& Murphy, S.\ 2012, , 124, 912, 140

Bessell, M. & Murphy, S.\ 2012, , 124, 912, 140. doi:10.1086/664083

doi:10.1086/664083 2012
[3]

J., Shanks, T., & Croom, S

Boyle, B. J., Shanks, T., & Croom, S. M.\ 1995, , 276, 1, 33. doi:10.1093/mnras/276.1.33

work page doi:10.1093/mnras/276.1.33 1995
[4]

M., Ratcliffe, A., Parker, Q

Croom, S. M., Ratcliffe, A., Parker, Q. A., et al.\ 1999, , 306, 3, 592. doi:10.1046/j.1365-8711.1999.02560.x

arXiv 1999
[5]

doi:10.1051/0004-6361/202348793

Enke, H., Tuvikene, T., Groote, D., et al.\ 2024, , 687, A165. doi:10.1051/0004-6361/202348793

work page doi:10.1051/0004-6361/202348793 2024
[7]

C., MacGillivray, H

Hambly, N. C., MacGillivray, H. T., Read, M. A., et al.\ 2001, , 326, 4, 1279. doi:10.1111/j.1365-2966.2001.04660.x

arXiv 2001
[8]

A., Templeton, M., Terrell, D., et al.\ 2016, VizieR Online Data Catalog, 2336

Henden, A. A., Templeton, M., Terrell, D., et al.\ 2016, VizieR Online Data Catalog, 2336. II/336

work page 2016
[9]

Hoeg, E., B \"a ssgen, G., Bastian, U., et al.\ 1997, , 323, L57

work page 1997

Show all 28 references

[10]

doi:10.3847/1538-4365/ad18b1

Huang, B., Yuan, H., Xiang, M., et al.\ 2024, , 271, 1, 13. doi:10.3847/1538-4365/ad18b1

2024 doi
[11]

doi:10.1002/asna.201813515

Hudec, R.\ 2018, Astronomische Nachrichten, 339, 5, 408. doi:10.1002/asna.201813515

2018 doi
[12]

doi:10.1002/asna.201913676

Hudec, R.\ 2019, Astronomische Nachrichten, 340, 7, 690. doi:10.1002/asna.201913676

2019 doi
[13]

C., & Driffield, V

Hurter, F. C., & Driffield, V. C. \ 1890, J. Soc. Chem. Ind., 9, 5, 455. doi:10.1002/jctb.5000090508

doi
[14]

Gaia Collaboration, Brown, A. G. A., Vallenari, A., et al.\ 2021, , 650, C3. doi:10.1051/0004-6361/202039657e

2021 doi
[15]

Gaia Collaboration, Vallenari, A., Brown, A. G. A., et al.\ 2023, , 674, A1. doi:10.1051/0004-6361/202243940

2023 doi
[16]

Grindlay, J., Tang, S., Simcoe, R., et al.\ 2009, Preserving Astronomy's Photographic Legacy: Current State and the Future of North American Astronomical Plates, 410, 101

2009
[17]

Kovalevsky, J., Lindegren, L., Perryman, M. A. C., et al.\ 1997, , 323, 620

1997
[18]

European Space Agency (ESA). 1997. The Hipparcos and Tycho Catalogues. Report No. ESA SP-1200, Noordwijk, The Netherlands: ESA Publications Division

1997
[19]

U.\ 1992, , 104, 340

Landolt, A. U.\ 1992, , 104, 340. doi:10.1086/116242

1992 doi
[20]

M., Sturch, C

Lasker, B. M., Sturch, C. R., Lopez, C., et al.\ 1988, , 68, 1. doi:10.1086/191282

1988 doi
[21]

doi:10.1088/0004-6256/140/4/1062

Laycock, S., Tang, S., Grindlay, J., et al.\ 2010, , 140, 4, 1062. doi:10.1088/0004-6256/140/4/1062

2010 doi
[22]

doi:10.3847/1538-4365/adec7e

Ma, M., Yuan, H., Xiao, K., et al.\ 2025, , 280, 1, 18. doi:10.3847/1538-4365/adec7e

2025 doi
[23]

P., Tancik, M., et al.\ 2020, arXiv:2003.08934

Mildenhall, B., Srinivasan, P. P., Tancik, M., et al.\ 2020, arXiv:2003.08934. doi:10.48550/arXiv.2003.08934

2020 doi arXiv
[24]

Hurter, F. I G. \ 1954, Apr., 1, 3, 131. doi:10.2307/1504975

1954 doi
[25]

E., Hinton, G

Rumelhart, D. E., Hinton, G. E., & Williams, R. J.\ 1986, , 323, 6088, 533. doi:10.1038/323533a0

1986 doi
[26]

L., Lasker, B

Russell, J. L., Lasker, B. M., McLean, B. J., et al.\ 1990, , 99, 2059. doi:10.1086/115484

1990 doi
[27]

doi:10.1088/1674-4527/ad339d

Shang, Z.-J., Yu, Y., Wang, L.-L., et al.\ 2024, Research in Astronomy and Astrophysics, 24, 5, 055010. doi:10.1088/1674-4527/ad339d

2024 doi
[28]

doi:10.1086/671760

Tang, S., Grindlay, J., Los, E., et al.\ 2013, , 125, 929, 857. doi:10.1086/671760

2013 doi
[29]

doi:10.1038/s41597-026-06590-z

Xiao, K., Huang, Y., Yuan, H., et al.\ 2026, Nature Scientific Data, 13, 1, 265. doi:10.1038/s41597-026-06590-z

2026 doi

Pith tools

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

[1] [1]

T., Mildenhall, B., Tancik, M., et al.\ 2021, 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 580

Barron, J. T., Mildenhall, B., Tancik, M., et al.\ 2021, 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 580. doi:10.1109/ICCV48922.2021.00580

arXiv 2021

[2] [2]

& Murphy, S.\ 2012, , 124, 912, 140

Bessell, M. & Murphy, S.\ 2012, , 124, 912, 140. doi:10.1086/664083

doi:10.1086/664083 2012

[3] [3]

J., Shanks, T., & Croom, S

Boyle, B. J., Shanks, T., & Croom, S. M.\ 1995, , 276, 1, 33. doi:10.1093/mnras/276.1.33

work page doi:10.1093/mnras/276.1.33 1995

[4] [4]

M., Ratcliffe, A., Parker, Q

Croom, S. M., Ratcliffe, A., Parker, Q. A., et al.\ 1999, , 306, 3, 592. doi:10.1046/j.1365-8711.1999.02560.x

arXiv 1999

[5] [5]

doi:10.1051/0004-6361/202348793

Enke, H., Tuvikene, T., Groote, D., et al.\ 2024, , 687, A165. doi:10.1051/0004-6361/202348793

work page doi:10.1051/0004-6361/202348793 2024

[6] [7]

C., MacGillivray, H

Hambly, N. C., MacGillivray, H. T., Read, M. A., et al.\ 2001, , 326, 4, 1279. doi:10.1111/j.1365-2966.2001.04660.x

arXiv 2001

[7] [8]

A., Templeton, M., Terrell, D., et al.\ 2016, VizieR Online Data Catalog, 2336

Henden, A. A., Templeton, M., Terrell, D., et al.\ 2016, VizieR Online Data Catalog, 2336. II/336

work page 2016

[8] [9]

Hoeg, E., B \"a ssgen, G., Bastian, U., et al.\ 1997, , 323, L57

work page 1997