arxiv: 2604.26398 · v1 · submitted 2026-04-29 · 🌌 astro-ph.SR · astro-ph.IM

Recognition: unknown

The Phenomenological Classification of TESS Eclipsing Binaries

Shi-Qi Liu , Kai Li , Xiao-Dian Chen , Li-Heng Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-07 13:03 UTC · model grok-4.3

classification 🌌 astro-ph.SR astro-ph.IM

keywords eclipsing binariesTESSneural networklight curve classificationEA EB EWmachine learningvariable starsphotometric surveys

0 comments

The pith

A neural network classifies TESS eclipsing binaries into EA, EB, and EW types with 99 percent accuracy.

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

The paper develops a fully connected neural network to classify TESS eclipsing binaries into three phenomenological types according to the shapes of their light curves. Eclipsing binaries act as direct laboratories for measuring stellar masses, radii, and evolutionary paths, but the volume of data from modern surveys makes manual classification impractical. The authors assembled a training set of 9576 light curves by cross-matching the ASAS-SN catalog with TESS targets, processed them through a uniform pipeline, and trained the network to reach 99.23 percent accuracy on validation data and 99.03 percent on test data. They then applied the model to 20196 TESS eclipsing binaries and performed manual review to produce a final catalog containing 13376 EA, 2114 EB, and 4706 EW systems. The resulting standardized pipeline supplies a practical method for handling the large numbers of eclipsing binaries expected from future surveys.

Core claim

We first extracted eclipsing binaries from the ASAS-SN variable star catalog and cross-matched them with TESS targets. The corresponding TESS light curves were processed through a unified pipeline, resulting in a high-quality training set of 9576 eclipsing binary light curves (2801 EA, 1930 EB, and 4845 EW systems). We designed and trained a fully connected neural network that achieved accuracy of 99.23% and 99.03% on the validation and test set respectively. Applying the trained neural network to a total of 20196 TESS eclipsing binaries collected from multiple star catalogs and performing manual visual inspection, we finally obtained 13376 EA, 2114 EB, and 4706 EW systems. The standardized

What carries the argument

Fully connected neural network that takes preprocessed TESS light curves as input and outputs one of three eclipsing binary classes (EA, EB, or EW).

Load-bearing premise

The 9576 light curves used for training represent the full diversity of TESS eclipsing binary signals, and manual visual inspection can reliably correct any misclassifications produced by the network.

What would settle it

Applying the trained network to an independent collection of TESS light curves whose types have been established by independent photometric or spectroscopic analysis and finding accuracy well below 95 percent would falsify the performance claim.

Figures

Figures reproduced from arXiv: 2604.26398 by Kai Li, Li-Heng Wang, Shi-Qi Liu, Xiao-Dian Chen.

**Figure 1.** Figure 1: A histogram of the number of available sectors per target. 2.2. Light curve preprocessing All TESS light curves underwent unified preprocessing to ensure structural consistency of input data of the FCNN model: 1. Detrending: Locally weighted scatterplot smoothing (LOWESS; W. S. Cleveland 1979) was applied for light curve detrending. The smoothing parameter was adaptively determined based on stellar variabi… view at source ↗

**Figure 2.** Figure 2: Detrending process of the light curve for TIC 133966005. The upper panel shows the original light curve (blue dots) and the corresponding LOWESS-fitted trend (red line), while the lower panel displays the detrended light curve. as the primary minimum. Finally, the entire light curve was phase-shifted to align the primary minimum at phase 0. The phase range was adjusted to [−0.75, 0.25] to avoid splitting o… view at source ↗

**Figure 3.** Figure 3: Data denoising process of the light curve for TIC 391623240. The upper two panels display the denoising process in the phase domain, and the lower two panels show the corresponding results in the time domain. Blue dots represent valid retained data, while red dots represent the noise points removed during the denoising process view at source ↗

**Figure 4.** Figure 4: Representative light curves from left to right for EA, EB, and EW systems. capacity while avoiding overfitting given the dataset size. We also tested several alternative configurations, including varying the number of hidden layers, neurons per layer, and dropout rates; however, these modifications did not yield significant improvements over the adopted architecture and in some cases even led to worse perf… view at source ↗

**Figure 5.** Figure 5: Training dynamics of the FCNN model over 500 epochs. Left: Training and validation loss (logarithmic y-axis). Right: Validation accuracy. The x-axis denotes the epoch number view at source ↗

**Figure 6.** Figure 6: Confusion matrix of the FCNN model evaluated on the test set. effectively distinguish among the EA, EB, and EW types view at source ↗

**Figure 7.** Figure 7: Autoencoder training curves: loss and MAE versus training epochs. Left panel presents the training and validation loss curves, while the right panel shows the training and validation MAE curves. However, both normal and abnormal samples contained light curves with incorrect periods (e.g., the true period was twice or half the current period). To address this issue, the scipy.signal.find peaks function in P… view at source ↗

**Figure 8.** Figure 8: Color–magnitude diagram, period–absolute-magnitude diagram, and period histogram for the eclipsing binary catalog. In the left panel, the orange elliptical area contains stars with extremely high temperatures but moderate luminosities, possibly corresponding to sdB or sdO subdwarfs in special evolutionary stages. The gray elliptical area is occupied by the white dwarf sequence, whose members remain very fa… view at source ↗

**Figure 9.** Figure 9: Kernel density distribution curves of morphological parameters for EA, EB, and EW identified in this work, derived from cross-matching with the catalog by A. Prˇsa et al. (2022). The x-axis denotes the morphological parameter, and the y-axis represents the kernel density. unify the feature space of all light curves. Subsequently, a simple and efficient FCNN model was constructed, which achieved near-perfec… view at source ↗

read the original abstract

Eclipsing binaries are crucial astrophysical laboratories for studying stellar parameters and evolutionary processes. In this study, we constructed a machine-learning-based model for systematic phenomenological classification of eclipsing binaries. We first extracted eclipsing binaries from the ASAS-SN variable star catalog and cross-matched them with TESS targets. The corresponding TESS light curves were processed through a unified pipeline, resulting in a high-quality training set of 9576 eclipsing binary light curves (2801 EA, 1930 EB, and 4845 EW systems). We designed and trained a fully connected neural network (FCNN) that achieved accuracy of 99.23% and 99.03% on the validation and test set respectively, demonstrating excellent performance. Applying the trained neural network to a total of 20196 TESS eclipsing binaries collected from multiple star catalogs and performing manual visual inspection, we finally obtained 13376 EA, 2114 EB, and 4706 EW systems. The standardized preprocessing pipeline and high-performance classifier developed in this study provide a reliable tool for the rapid automated classification of massive numbers of eclipsing binary in future 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 paper delivers a new catalog of 20k TESS eclipsing binaries typed as EA/EB/EW via an FCNN trained on ASAS-SN cross-matches, but the 99% accuracy claim only covers the training distribution and does not demonstrate reliable performance on the broader multi-catalog sample.

read the letter

The main output is a catalog of 20,196 TESS eclipsing binaries split into 13,376 EA, 2,114 EB, and 4,706 EW systems. They built it by training a fully connected neural network on 9,576 light curves from ASAS-SN targets observed by TESS, then ran the model on a larger collection drawn from several other catalogs and finished with manual visual checks. The preprocessing pipeline they describe is straightforward and reusable, and the internal validation and test accuracies of 99.23% and 99.03% are high for this kind of task. That part is solid applied work and gives people a concrete data product they can download and use for statistical studies of TESS binaries. The citation list is appropriate and points to the source catalogs without obvious gaps. The stress-test note is correct on the key limitation. Both the validation and test sets come from the same ASAS-SN cross-matched pool, so the accuracy numbers do not test generalization to the 20k targets assembled from other catalogs. Those catalogs can differ in selection, cadence, and even how subtypes are defined, which means the model could be making systematic errors on the new data. The paper relies on the subsequent manual inspection to correct those errors, but it gives no count of how many labels were changed or any measure of consistency in the visual review. Without that information, it is difficult to judge how much the final counts depend on the human step versus the network. The training set may also not capture the full range of TESS light-curve quality or binary parameters present in the larger sample. This is a standard supervised classification setup with a clear data product, so it is worth a referee's time. The work is honest about its methods and produces something usable, even if the generalization claim needs more evidence. A reader building a classifier for their own survey would find the pipeline helpful as a baseline, while someone needing a large typed TESS sample could use the catalog with the usual caution about selection effects. I would bring it to a reading group for the data product and the practical pipeline details. I would not cite it in my own work unless I needed exactly these counts. Send it to peer review rather than desk reject; the referees should focus on the validation gap and ask for the manual-correction statistics.

Referee Report

2 major / 1 minor

Summary. The manuscript presents a machine-learning pipeline for phenomenological classification of TESS eclipsing binaries into EA, EB, and EW types. A training set of 9,576 light curves is assembled by cross-matching the ASAS-SN variable-star catalog with TESS targets (2,801 EA, 1,930 EB, 4,845 EW). A fully connected neural network is trained and reported to reach 99.23% accuracy on a validation set and 99.03% on a test set. The model is then applied to 20,196 TESS eclipsing binaries drawn from multiple catalogs; manual visual inspection of the predictions yields a final catalog containing 13,376 EA, 2,114 EB, and 4,706 EW systems. A standardized light-curve preprocessing pipeline is also described.

Significance. If the reported classification accuracy generalizes, the resulting catalog of more than 20,000 TESS eclipsing binaries would constitute a useful resource for studies of stellar parameters, mass-radius relations, and evolutionary pathways. The standardized preprocessing pipeline and the demonstration of an FCNN classifier could be adopted or extended by future all-sky photometric surveys. The work therefore has clear archival and methodological value provided the generalization claim is substantiated.

major comments (2)

[Abstract] Abstract: the stated validation and test accuracies (99.23% and 99.03%) are obtained exclusively on random or unspecified splits of the 9,576 ASAS-SN cross-matched light curves. No description is supplied of the train/validation/test partitioning strategy, whether the split was stratified by morphological class to address the EW majority, the precise light-curve features fed to the FCNN, or any quantitative error analysis (confusion matrix, per-class precision/recall). Because the subsequent application is to a 20,196-object sample assembled from heterogeneous catalogs whose selection functions and possible label conventions differ from ASAS-SN, the internal accuracy figures do not by themselves establish reliable performance on the target population.
[Abstract] Abstract: after the FCNN is applied to the 20,196 multi-catalog TESS targets, the final published counts are obtained only after an additional manual visual-inspection step. The manuscript does not report how many objects had their automatic label changed by this inspection, nor does it provide any measure of inter-inspector agreement or a subsample that was inspected independently. Without these statistics it is impossible to quantify the residual error rate or to judge whether the training-set distribution was sufficiently representative.

minor comments (1)

[Abstract] Abstract: the sentence 'a total of 20196 TESS eclipsing binaries collected from multiple star catalogs' is followed by final counts that sum exactly to 20,196. It would be clearer to state explicitly whether any targets were discarded during the manual-inspection stage and, if so, on what criteria.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their careful and constructive review. We address the major comments point by point below and indicate the changes we will make to the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the stated validation and test accuracies (99.23% and 99.03%) are obtained exclusively on random or unspecified splits of the 9,576 ASAS-SN cross-matched light curves. No description is supplied of the train/validation/test partitioning strategy, whether the split was stratified by morphological class to address the EW majority, the precise light-curve features fed to the FCNN, or any quantitative error analysis (confusion matrix, per-class precision/recall). Because the subsequent application is to a 20,196-object sample assembled from heterogeneous catalogs whose selection functions and possible label conventions differ from ASAS-SN, the internal accuracy figures do not by themselves establish reliable performance on the target population.

Authors: We agree that the abstract omits these methodological details. We will revise the manuscript to explicitly describe the train/validation/test partitioning strategy, confirm whether the split was stratified by class, specify the precise light-curve features input to the FCNN, and include a confusion matrix together with per-class precision and recall. We also agree that the reported accuracies on the ASAS-SN cross-matched sample do not by themselves demonstrate reliable performance on the heterogeneous 20,196-object target sample; the subsequent manual visual inspection step was performed precisely to mitigate this limitation. revision: yes
Referee: [Abstract] Abstract: after the FCNN is applied to the 20,196 multi-catalog TESS targets, the final published counts are obtained only after an additional manual visual-inspection step. The manuscript does not report how many objects had their automatic label changed by this inspection, nor does it provide any measure of inter-inspector agreement or a subsample that was inspected independently. Without these statistics it is impossible to quantify the residual error rate or to judge whether the training-set distribution was sufficiently representative.

Authors: We agree that the number of label changes during visual inspection should be reported. We will add this information to the revised manuscript. The visual inspection was performed collaboratively by the author team; no formal inter-inspector agreement study or independently inspected subsample was conducted, so we cannot supply those quantitative statistics. We consider the ASAS-SN training distribution representative of the morphological classes, but acknowledge that catalog-specific selection effects may remain and were addressed through the inspection step. revision: partial

standing simulated objections not resolved

Quantitative measure of inter-inspector agreement and results from an independently inspected subsample, as these were not performed.

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper extracts a training set of 9576 ASAS-SN cross-matched TESS light curves, trains an FCNN, reports measured accuracy on internal validation/test splits drawn from the same pool, applies the model to an independent collection of 20196 TESS targets assembled from multiple catalogs, and performs manual visual inspection to produce final counts. No equations, fitted parameters, or self-citations reduce the reported accuracies or catalog sizes to quantities defined by construction from the same inputs. The manual inspection step is presented as an external correction, and the pipeline contains no self-definitional loops, fitted-input-as-prediction reductions, or load-bearing self-citations that would force the central claims.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the representativeness of the ASAS-SN cross-matched training set, the sufficiency of the chosen light-curve features for type discrimination, and the reliability of post-hoc manual inspection; no new physical entities or ad-hoc constants are introduced.

free parameters (1)

Neural network hyperparameters
Number of layers, hidden units, learning rate, and regularization strength chosen or tuned to reach the reported accuracy on the validation set.

axioms (1)

domain assumption Extracted light-curve features from the unified preprocessing pipeline are sufficient to distinguish EA, EB, and EW morphological types
Invoked when the FCNN is trained and applied without additional physical modeling.

pith-pipeline@v0.9.0 · 5508 in / 1499 out tokens · 86499 ms · 2026-05-07T13:03:13.694910+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

36 extracted references · 32 canonical work pages · 2 internal anchors

[1]

1991, A&A Rv, 3, 91, doi: 10.1007/BF00873538

Andersen, J. 1991, A&A Rv, 3, 91, doi: 10.1007/BF00873538

work page doi:10.1007/bf00873538 1991
[2]

J., Kirk, J., Lam, K

Armstrong, D. J., Kirk, J., Lam, K. W. F., et al. 2016, MNRAS, 456, 2260, doi: 10.1093/mnras/stv2836 12

work page doi:10.1093/mnras/stv2836 2016
[3]

C., Kulkarni, S

Bellm, E. C., Kulkarni, S. R., Graham, M. J., et al. 2019, PASP, 131, 018002, doi: 10.1088/1538-3873/aaecbe

work page doi:10.1088/1538-3873/aaecbe 2019
[4]

Kepler Planet-Detection Mission: Introduction and First Results.Science2010,327, 977

Borucki, W. J., Koch, D., Basri, G., et al. 2010, Science, 327, 977, doi: 10.1126/science.1185402

work page doi:10.1126/science.1185402 2010
[5]

S., & Lowe, D

Broomhead, D. S., & Lowe, D. 1988, Complex Systems, 2, 321

1988
[6]

, keywords =

Chen, X., Wang, S., Deng, L., et al. 2020, ApJS, 249, 18, doi: 10.3847/1538-4365/ab9cae

work page doi:10.3847/1538-4365/ab9cae 2020
[7]

2002, A&A, 386, 763, doi: 10.1051/0004-6361:20020258

Clarke, D. 2002, A&A, 386, 763, doi: 10.1051/0004-6361:20020258

work page doi:10.1051/0004-6361:20020258 2002
[8]

Cleveland, W. S. 1979, Journal of the American Statistical Association, 74, 829, doi: 10.1080/01621459.1979.10481038

work page doi:10.1080/01621459.1979.10481038 1979
[9]

V., Gramajo, L

Daza-Perilla, I. V., Gramajo, L. V., Lares, M., et al. 2023, MNRAS, 520, 828, doi: 10.1093/mnras/stad141

work page doi:10.1093/mnras/stad141 2023
[10]

2025, AJ, 169, 202, doi: 10.3847/1538-3881/adb846

Ding, X., Ji, K., Cheng, Q., et al. 2025, AJ, 169, 202, doi: 10.3847/1538-3881/adb846

work page doi:10.3847/1538-3881/adb846 2025
[11]

2024, AJ, 167, 192, doi: 10.3847/1538-3881/ad3048 D’Isanto, A., & Polsterer, K

Ding, X., Song, Z., Wang, C., & Ji, K. 2024, AJ, 167, 192, doi: 10.3847/1538-3881/ad3048 Gaia Collaboration, Prusti, T., de Bruijne, J. H. J., et al. 2016, A&A, 595, A1, doi: 10.1051/0004-6361/201629272

work page doi:10.3847/1538-3881/ad3048 2024
[12]

Classification of Periodic Variable Stars from TESS

Gao, X., Chen, X., Wang, S., & Liu, J. 2025, ApJS, 276, 57, doi: 10.3847/1538-4365/ad9dd6 G lowacki, M., Soszy´ nski, I., Udalski, A., et al. 2024, AcA, 74, 241, doi: 10.32023/0001-5237/74.4.1

work page doi:10.3847/1538-4365/ad9dd6 2025
[13]

2016, Deep

Goodfellow, I., Bengio, Y., & Courville, A. 2016, Deep

2016
[14]

Reducing the dimensionality of data with neural networks

Hinton, G. E., & Salakhutdinov, R. R. 2006, Science, 313, 504, doi: 10.1126/science.1127647

work page doi:10.1126/science.1127647 2006
[15]

B., Sobeck, C., Haas, M., et al

Howell, S. B., Sobeck, C., Haas, M., et al. 2014, PASP, 126, 398, doi: 10.1086/676406 Ivezi´ c,ˇZ., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111, doi: 10.3847/1538-4357/ab042c

work page doi:10.1086/676406 2014
[16]

S., Stanek, K

Jayasinghe, T., Kochanek, C. S., Stanek, K. Z., et al. 2018, MNRAS, 477, 3145, doi: 10.1093/mnras/sty838

work page doi:10.1093/mnras/sty838 2018
[17]

2012, MNRAS, 420, 1825, doi: 10.1111/j.1365-2966.2011.19805.x

Jiang, D., Han, Z., Ge, H., Yang, L., & Li, L. 2012, MNRAS, 421, 2769, doi: 10.1111/j.1365-2966.2011.20323.x

work page doi:10.1111/j.1365-2966.2011.20323.x 2012
[18]

Kallrath, J., & Milone, E. F. 2009, Eclipsing Binary Stars: Modeling and Analysis, 2nd edn., Astronomy and Astrophysics Library (New York: Springer), doi: 10.1007/978-1-4419-0699-1

work page doi:10.1007/978-1-4419-0699-1 2009
[19]

Adam: A Method for Stochastic Optimization

Kingma, D. P., & Ba, J. 2014, arXiv e-prints, arXiv:1412.6980, doi: 10.48550/arXiv.1412.6980

work page internal anchor Pith review doi:10.48550/arxiv.1412.6980 2014
[20]

, keywords =

Kostov, V. B., Powell, B. P., Fornear, A. U., et al. 2025, ApJS, 279, 50, doi: 10.3847/1538-4365/ade2d8 Kov´ acs, G., Zucker, S., & Mazeh, T. 2002, A&A, 391, 369, doi: 10.1051/0004-6361:20020802

work page doi:10.3847/1538-4365/ade2d8 2025
[21]

Lafler, J., & Kinman, T. D. 1965, ApJS, 11, 216, doi: 10.1086/190116

work page doi:10.1086/190116 1965
[22]

2020, AJ, 159, 189, doi: 10.3847/1538-3881/ab7cda

Li, K., Kim, C.-H., Xia, Q.-Q., et al. 2020, AJ, 159, 189, doi: 10.3847/1538-3881/ab7cda

work page doi:10.3847/1538-3881/ab7cda 2020
[23]

2025, ApJS, 277, 51, doi: 10.3847/1538-4365/adba63

Li, K., & Wang, L.-H. 2025, ApJS, 277, 51, doi: 10.3847/1538-4365/adba63

work page doi:10.3847/1538-4365/adba63 2025
[24]

2019, MNRAS, 485, 4588, doi: 10.1093/mnras/stz715 Lightkurve Collaboration, Cardoso, J

Li, K., Xia, Q.-Q., Michel, R., et al. 2019, MNRAS, 485, 4588, doi: 10.1093/mnras/stz715 Lightkurve Collaboration, Cardoso, J. V. d. M., Hedges, C., et al. 2018, Lightkurve: Kepler and TESS time series analysis in Python,, Astrophysics Source Code Library, record ascl:1812.013 http://ascl.net/1812.013

work page doi:10.1093/mnras/stz715 2019
[25]

Publications of the Astronomical Society of the Pacific , author =

Masci, F. J., Laher, R. R., Rusholme, B., et al. 2019, PASP, 131, 018003, doi: 10.1088/1538-3873/aae8ac

work page internal anchor Pith review doi:10.1088/1538-3873/aae8ac 2019
[26]

2010, New Astronomy, 15, 433, doi: https://doi.org/10.1016/j.newast.2009.12.002 Mr´ oz, P., Udalski, A., Skowron, J., et al

Minniti, D., Lucas, P. W., Emerson, J. P., et al. 2010, NewA, 15, 433, doi: 10.1016/j.newast.2009.12.002

work page doi:10.1016/j.newast.2009.12.002 2010
[27]

Nair, V., & Hinton, G. E. 2010, in Proceedings of the 27th International Conference on International Conference on Machine Learning, ICML’10 (Madison, WI, USA: Omnipress), 807–814

2010
[28]

2016, Acta Astronomica, 66, 421, doi: 10.48550/arXiv.1612.06394

Pawlak, M., Soszy´ nski, I., Udalski, A., et al. 2016, AcA, 66, 421, doi: 10.48550/arXiv.1612.06394 Prˇ sa, A., Kochoska, A., Conroy, K. E., et al. 2022, ApJS, 258, 16, doi: 10.3847/1538-4365/ac324a

work page doi:10.48550/arxiv.1612.06394 2016
[29]

B., Zhang, B., Soonthornthum, B., et al

Qian, S. B., Zhang, B., Soonthornthum, B., et al. 2015, AJ, 150, 117, doi: 10.1088/0004-6256/150/4/117

work page doi:10.1088/0004-6256/150/4/117 2015
[30]

R., Winn, J

Ricker, G. R., Winn, J. N., Vanderspek, R., et al. 2015, Journal of Astronomical Telescopes, Instruments, and Systems, 1, 014003, doi: 10.1117/1.JATIS.1.1.014003

work page Pith review doi:10.1117/1.jatis.1.1.014003 2015
[31]

Rucinski, S. M. 1992, AJ, 103, 960, doi: 10.1086/116118

work page doi:10.1086/116118 1992
[32]

Learning representations by back-propagating errors.Nature1986,323, 533–536

Rumelhart, D. E., Hinton, G. E., & Williams, R. J. 1986, Nature, 323, 533, doi: 10.1038/323533a0 Sch¨ olkopf, B., Platt, J. C., Shawe-Taylor, J. C., Smola, A. J., & Williamson, R. C. 2001, Neural Comput., 13, 1443–1471, doi: 10.1162/089976601750264965

work page doi:10.1038/323533a0 1986
[33]

2025, PASP, 137, 044503, doi: 10.1088/1538-3873/adc5a2

Shan, Y., Chen, J., Zhang, Z., et al. 2025, PASP, 137, 044503, doi: 10.1088/1538-3873/adc5a2 Soszy´ nski, I., Pawlak, M., Pietrukowicz, P., et al. 2016, AcA, 66, 405, doi: 10.48550/arXiv.1701.03105

work page doi:10.1088/1538-3873/adc5a2 2025
[34]

Salakhutdinov, R. 2014, J. Mach. Learn. Res., 15, 1929–1958

2014
[35]

OGLE-IV: Fourth Phase of the Optical Gravitational Lensing Experiment

Udalski, A., Szyma´ nski, M. K., & Szyma´ nski, G. 2015, AcA, 65, 1, doi: 10.48550/arXiv.1504.05966 ˇCokina, M., Maslej-Kreˇ sˇ n´ akov´ a, V., Butka, P., & Parimucha, ˇS. 2021, Astronomy and Computing, 36, 100488, doi: 10.1016/j.ascom.2021.100488

work page Pith review doi:10.48550/arxiv.1504.05966 2015
[36]

2025, ApJ, 986, 19, doi: 10.3847/1538-4357/add159

Wang, L.-H., Li, K., Gao, X., Guo, Y.-N., & Sun, G.-Y. 2025, ApJ, 986, 19, doi: 10.3847/1538-4357/add159

work page doi:10.3847/1538-4357/add159 2025