pith. sign in

arxiv: 1907.00132 · v1 · pith:LGBKU7RQnew · submitted 2019-06-29 · 🌌 astro-ph.SR

Searching for Hot Subdwarf Stars in LAMOST DR1-II. Pure spectroscopic identification method for hot subdwarfs

Zhenxin Lei , Yude Bu , Jingkun Zhao , P\'eter N\'emeth , Gang Zhao This is my paper

Pith reviewed 2026-05-25 13:08 UTC · model grok-4.3

classification 🌌 astro-ph.SR
keywords hot subdwarf starsLAMOST surveymachine learningHELM algorithmspectroscopic identificationhelium sequencessdB starsNLTE model atmospheres
0
0 comments X

The pith

Hierarchical extreme learning machine identifies hot subdwarf stars from LAMOST spectra alone.

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

The paper shows that the hierarchical extreme learning machine algorithm can classify hot subdwarf stars directly from their observed spectra in the LAMOST DR1 survey. After training on spectral examples, the method filters out characteristic features without any photometric input. The authors report 56 identified stars, with atmospheric parameters derived from fits to hydrogen and helium lines using NLTE models. Five of these are helium-rich while the rest are helium-poor sdB, sdO, and sdOB types. This approach also reproduces the two distinct helium sequences previously noted in the temperature-abundance diagram.

Core claim

The HELM algorithm, trained suitably on spectral data, reliably identifies hot subdwarf stars in LAMOST DR1 from spectroscopy alone, producing a sample of 56 stars whose derived parameters confirm the two helium sequences in the Teff-log(nHe/nH) plane.

What carries the argument

The hierarchical extreme learning machine (HELM) algorithm that classifies objects by operating directly on observed spectroscopy to isolate spectral properties.

If this is right

  • HELM works without supplementary photometric data for classification.
  • The same trained method applies to searching for other objects with clear spectral features.
  • The sample contains five He-rich stars with log(nHe/nH) > -1 and 51 He-poor stars.
  • The two helium sequences reported by Edelmann et al. appear in the new data.

Where Pith is reading between the lines

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

  • Later LAMOST data releases could be processed with the same HELM setup to expand the known hot subdwarf population.
  • Pure spectral selection may avoid biases that photometric pre-selection introduces in other surveys.
  • The algorithm could be retrained on different wavelength ranges or resolution to target rarer subtypes.

Load-bearing premise

The training set must contain spectral examples representative of hot subdwarfs versus other stars present in the LAMOST survey.

What would settle it

Follow-up high-resolution spectroscopy or independent classification showing that a large fraction of the 56 candidates lack the atmospheric parameters of hot subdwarfs.

Figures

Figures reproduced from arXiv: 1907.00132 by Gang Zhao, Jingkun Zhao, P\'eter N\'emeth, Yude Bu, Zhenxin Lei.

Figure 1
Figure 1. Figure 1: Normalized spectra near the Hδ line in three different types of stars. The blue dashed curve is the fitting profile of Hδ line. The values of D0.2 and fm for each star are showed. 3.1 Excluding BHB stars and WDs from our sample BHB stars are horizontal branch stars bluer than the RR Lyrae instability strip in the color-magnitude diagram (CMD). These stars present effective temperatures in the range of abou… view at source ↗
Figure 2
Figure 2. Figure 2: Panel (a): the distribution of BHB stars, hot subdwarfs and WDs in the D0.2 − fm diagram. Panel (b): Our hot subdwarf candidates selected by the HELM algorithm in the D0.2 − fm diagram. The red dashed line is a clear boundary between BHB stars and hot subdwarfs at D0.2 =17.0 Å. y = n − a exp" − [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Four normalized spectra of hot subdwarf stars with different spectral types identified in this study. Best-fitting synthetic spectra are over plotted by a red dashed line on each spectra. From top to bottom, a He-sdOB, sdOB, sdB and sdO star is presented respectively. Some H Balmer lines and important He I and He II lines marked at the bottom of the figure. with weak or no obvious Hδ lines (e.g., He-sdO, H… view at source ↗
Figure 4
Figure 4. Figure 4: The distribution of 74 selected subdwarf candidates in the HRD of Gaia DR2. 57 stars (marked with blue triangles) locate in the subdwarf region, and 12 stars (denoted by yellow squares) are distributed along the MS region, while the position of 6 stars (represented by red circles) correspond to the WD sequence. 2018b), while blue triangles, yellow squares and red circles are the common stars in our sample.… view at source ↗
Figure 5
Figure 5. Figure 5: Panel (a): Comparisons between the atmospheric parameters obtained in this study and the ones from Geier et al. (2017). Panel (b): Comparisons between the atmospheric parameters obtained in this study and the ones from Nemeth et al. (2012). ´ models not only with H and He composition but also include C, N and O composition. Furthermore, the observed spectra in our sample (obtained in LAMOST survey) are dif… view at source ↗
Figure 6
Figure 6. Figure 6: Teff -log g diagram for for the 56 hot subdwarf stars identified in this study. Stars with log(nHe/nH) ≤ −2.2 are marked with filled circles, stars with −2.2 < log(nHe/nH) < −1.0 are represented by open triangles, while stars with log(nHe/nH) ≥ −1.0 are showed by open squares. The thick solid line denotes the He-MS from Paczynski (1971), and the two dashed lines represent ZAHB and TAHB from Dorman et al. (… view at source ↗
Figure 7
Figure 7. Figure 7: Teff -log(nHe/nH) diagram for the 56 hot subdwarf stars identified in this study. The red dashed line denotes the solar He abundance. The dotted line and solid line are the linear regression line fitted by Edelmann et al. (2003), while the dot-dashed line is the best-fitting line for the He-poor sequence in Nemeth et al. (2012). ´ Diamonds denote the stars for which we just obtained the upper limit of log(… view at source ↗
Figure 8
Figure 8. Figure 8: The logg − log (nHe/nH) plane for the 56 hot subdwarf stars identified in this study, the red dashed line marks the solar He abundance for reference. Diamonds denote the stars for which we just obtained the upper limit of log(nHe/nH) (see [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
read the original abstract

Employing a new machine learning method, named hierarchical extreme learning machine (HELM) algorithm, we identified 56 hot subdwarf stars in the first data release (DR1) of the Large Sky Area Multi-Object Fibre Spectroscopic Telescope (LAMOST) survey. The atmospheric parameters of the stars are obtained by fitting the profiles of hydrogen (H) Balmer lines and helium (He) lines with synthetic spectra calculated from non-Local Thermodynamic Equilibrium (NLTE) model atmospheres. Five He-rich hot subdwarf stars were found in our sample with their log(nHe/nH) > -1 , while 51 stars are He-poor sdB, sdO and sdOB stars. We also confirmed the two He sequences of hot subdwarf stars found by Edelmann et al. (2003) in Teff - log(nHe/nH) diagram. The HELM algorithm works directly on the observed spectroscopy and is able to filter out spectral properties without supplementary photometric data. The results presented in this study demonstrate that the HELM algorithm is a reliable method to search for hot subdwarf stars after a suitable training is performed, and it is also suitable to search for other objects which have obvious features in their spectra or images.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The manuscript applies the hierarchical extreme learning machine (HELM) algorithm to LAMOST DR1 spectra to identify 56 hot subdwarf stars without photometric data. Atmospheric parameters are derived by fitting Balmer and helium line profiles with NLTE synthetic spectra, yielding 5 He-rich stars (log(nHe/nH) > -1) and 51 He-poor sdB/sdO/sdOB stars. The work confirms the two helium sequences previously reported by Edelmann et al. (2003) in the Teff-log(nHe/nH) plane and asserts that HELM is reliable for hot subdwarfs (and other objects with clear spectral features) after suitable training.

Significance. If the HELM classifications are shown to be robust, the method supplies a photometry-independent route to enlarge samples of hot subdwarfs in large spectroscopic surveys, enabling statistical studies of their formation channels and atmospheric evolution. The confirmation of the two He sequences is consistent with earlier work but does not constitute a new result.

major comments (2)
  1. [Abstract / Methods] Abstract and §3 (or equivalent methods section): The central claim that HELM is a reliable classifier after suitable training is unsupported by any reported details on training-set construction (size, selection of known sdB/sdO versus LAMOST contaminants, class balance) or quantitative performance (accuracy, precision, recall, cross-validation scores, or false-positive rate on held-out LAMOST-like spectra). This information is load-bearing for the reliability of the 56 identifications.
  2. [Abstract / Results] Abstract and results section: No cross-check of the HELM-selected candidates against independent hot-subdwarf catalogs or any estimate of contamination rate is presented. The subsequent NLTE parameter fitting occurs after classification and therefore cannot validate the upstream HELM decisions; if the training distribution differs from the survey in S/N, wavelength coverage, or contaminant mix, the identifications rest on an untested assumption.
minor comments (1)
  1. [Abstract] Abstract: the phrase 'suitable training' is repeated without elaboration; a one-sentence summary of training-set provenance would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight areas where the manuscript can be strengthened. We respond to each major comment below and will revise the manuscript to address the concerns.

read point-by-point responses

  1. Referee: [Abstract / Methods] Abstract and §3 (or equivalent methods section): The central claim that HELM is a reliable classifier after suitable training is unsupported by any reported details on training-set construction (size, selection of known sdB/sdO versus LAMOST contaminants, class balance) or quantitative performance (accuracy, precision, recall, cross-validation scores, or false-positive rate on held-out LAMOST-like spectra). This information is load-bearing for the reliability of the 56 identifications.

    Authors: We agree that the current manuscript does not provide sufficient detail on the HELM training procedure to fully support the reliability claim. In the revised version, we will add a dedicated subsection to the methods describing the training-set construction (including size, selection of known hot subdwarfs and LAMOST contaminants, and class balance) along with quantitative performance metrics such as accuracy, precision, recall, and cross-validation scores. This will directly address the load-bearing nature of this information for the 56 identifications. revision: yes

  2. Referee: [Abstract / Results] Abstract and results section: No cross-check of the HELM-selected candidates against independent hot-subdwarf catalogs or any estimate of contamination rate is presented. The subsequent NLTE parameter fitting occurs after classification and therefore cannot validate the upstream HELM decisions; if the training distribution differs from the survey in S/N, wavelength coverage, or contaminant mix, the identifications rest on an untested assumption.

    Authors: The referee correctly identifies that no independent cross-check or contamination-rate estimate is included, and that the downstream NLTE fitting cannot validate the HELM classification step. In revision we will perform and report a cross-match against existing hot-subdwarf catalogs to quantify overlap and provide an estimate of contamination. We will also add an explicit discussion of the assumptions regarding training versus survey distributions (S/N, wavelength coverage, and contaminant mix) and any associated limitations. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper applies the HELM algorithm (described as a new machine learning method) to LAMOST spectra after suitable training to select 56 candidates, then fits atmospheric parameters via NLTE synthetic spectra on those candidates and confirms known He sequences from external literature (Edelmann et al. 2003). No equations, self-definitional relations, fitted inputs renamed as predictions, or load-bearing self-citations appear in the text. The classification step is independent of the downstream parameter fits, and the method is presented as externally applicable rather than reducing to its own outputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities are described beyond standard NLTE model fitting.

pith-pipeline@v0.9.0 · 5772 in / 957 out tokens · 28479 ms · 2026-05-25T13:08:07.079566+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages · 2 internal anchors

  1. [1]

    1992, AJ, 103, 267

    Beers, Timothy C., Preston, George W., Shectman, Stephen A., et al. 1992, AJ, 103, 267

    work page 1992

  2. [2]

    M., Cassisi, S., D’Antona, F., et al

    Brown, T. M., Cassisi, S., D’Antona, F., et al. 2016, ApJ, 822, 44

    work page 2016

  3. [3]

    2017, ApJS, 233, 2 (Paper I)

    Bu, Yude., Lei, Zhenxin., Zhao, Gang., et al. 2017, ApJS, 233, 2 (Paper I)

    work page 2017

  4. [4]

    2009, Ap&SS, 320, 261

    Catelan, M. 2009, Ap&SS, 320, 261

    work page 2009

  5. [5]

    2013, MNRAS, 434, 186

    Chen, Xuefei., Han, Zhanwen., Deca, Jan., et al. 2013, MNRAS, 434, 186

    work page 2013

  6. [6]

    J., Hewett, P

    Clewley, L., Warren, S. J., Hewett, P. C., et al. 2002, MNRAS, 337, 87

    work page 2002

  7. [7]

    M., Morales-Rueda, L., Marsh, T

    Copperwheat, C. M., Morales-Rueda, L., Marsh, T. R., et al. 2011, MNRAS, 415, 1381

    work page 2011

  8. [8]

    2012, RAA, 12, 1197

    Cui, Xiang-Qun., Zhao, Yong-Heng., Chu, Yao-Quan., et al. 2012, RAA, 12, 1197

    work page 2012

  9. [9]

    1993, ApJ, 419, 596

    Dorman, Ben., Rood, Robert T., & O’Connell, Robert W. 1993, ApJ, 419, 596

    work page 1993

  10. [10]

    S., Jeffery, C

    Drilling, J. S., Jeffery, C. S., Heber, U., et al. 2013, A&A, 551, 31

    work page 2013

  11. [11]

    2003, A&A, 400, 939 20

    Edelmann, H., Heber, U., Hagen, H.-J., et al. 2003, A&A, 400, 939 20

    work page 2003

  12. [12]

    Gaia Data Release 2. Summary of the contents and survey properties

    Eisenstein, Daniel J., Liebert, James., Harris, Hugh C., et al. 2006, ApJS, 167, 40 Gaia Collaboration, Brown, A., Vallenari, A., et al. 2018a, arXiv:1804.09365 Gaia Collaboration, Babusiaux, C., van Leeuwen, F., et al. et al. 2018b, arXiv:1804.09378

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  13. [13]

    2011, A&A, 530, 28

    Geier, S., Hirsch, H., Tillich, A., et al. 2011, A&A, 530, 28

    work page 2011

  14. [14]

    2015, Science, 347, 1126

    Geier, S., Frst, F., Ziegerer, E., et al. 2015, Science, 347, 1126

    work page 2015

  15. [15]

    H., N´emeth, P., et al

    Geier, S., Østensen, R. H., N´emeth, P., et al. 2017, A&A, 600, 50

    work page 2017

  16. [16]

    1974, ApJS, 28, 157

    Greenstein, Jesse L., & Sargent, Anneila I. 1974, ApJS, 28, 157

    work page 1974

  17. [17]

    Han, Z., Podsiadlowski, Ph., Maxted, P. F. L., et al. 2002, MNRAS, 336, 449

    work page 2002

  18. [18]

    Han, Z., Podsiadlowski, Ph., Maxted, P. F. L., et al. 2003, MNRAS, 341, 669

    work page 2003

  19. [19]

    Han, Z., Podsiadlowski, Ph., & Lynas-Gray, A. E. 2007, MNRAS, 380, 1098

    work page 2007

  20. [20]

    2009, ARA&A, 47, 211

    Heber, U. 2009, ARA&A, 47, 211

    work page 2009

  21. [21]

    2016, PASP, 128, 2001

    Heber, U. 2016, PASP, 128, 2001

    work page 2016

  22. [22]

    2006, Neurocomputing, 70, 489

    Huang, G., Zhu, Q., & Siew, C. 2006, Neurocomputing, 70, 489

    work page 2006

  23. [23]

    1995, ApJ, 439, 875

    Hubeny, I., Lanz, T. 1995, ApJ, 439, 875

    work page 1995

  24. [24]

    A brief introductory guide to TLUSTY and SYNSPEC

    Hubeny, I., & Lanz, T. 2017, arXiv:1706.01859

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  25. [25]

    2015, MNRAS, 450, 3514

    Kawka, A., Vennes, S., O’Toole, S., et al. 2015, MNRAS, 450, 3514

    work page 2015

  26. [26]

    O., Pelisoli, I., Koester, D., et al

    Kepler, S. O., Pelisoli, I., Koester, D., et al. 2015, MNRAS, 446, 4078

    work page 2015

  27. [27]

    O., Pelisoli, I., Koester, D., et al

    Kepler, S. O., Pelisoli, I., Koester, D., et al. 2016, MNRAS, 455, 3413

    work page 2016

  28. [28]

    1995, ApJ, 439, 905

    Lanz, T., Hubeny, I. 1995, ApJ, 439, 905

    work page 1995

  29. [29]

    2007, ApJS, 169, 83

    Lanz, T., Hubeny, I. 2007, ApJS, 169, 83

    work page 2007

  30. [30]

    2015, MNRAS, 449, 2741

    Lei, Zhenxin., Chen, Xuemei., Zhang, Fenghui., et al. 2015, MNRAS, 449, 2741

    work page 2015

  31. [31]

    2016, MNRAS, 463, 3449

    Lei, Zhenxin., Zhao, Gang., Zeng, Aihua., et al. 2016, MNRAS, 463, 3449

    work page 2016

  32. [32]

    2015, JARS, 9, 097296

    Li, J., Du, Q., Li, W., et al. 2015, JARS, 9, 097296

    work page 2015

  33. [33]

    2015, RAA, 15, 1095

    Luo, A.-Li., Zhao, Yong-Heng., Zhao, Gang., et al. 2015, RAA, 15, 1095

    work page 2015

  34. [34]

    2016, ApJ, 818, 202

    Luo, Yang-Ping., N´emeth, P., Liu, Chao., et al. 2016, ApJ, 818, 202

    work page 2016

  35. [35]

    2014, Mathematical Problems in Engineering, 2014, 1

    Mao, L., Zhang, L., Liu, X., et al. 2014, Mathematical Problems in Engineering, 2014, 1

    work page 2014

  36. [36]

    Maxted, P. F. L., Heber, U., Marsh, T. R., et al. 2001, MNRAS, 326, 139

    work page 2001

  37. [37]

    2010, Neurocomputing, 73, 1906

    Minhas, R., Baradarani, A., Seifzadeh,S., et al. 2010, Neurocomputing, 73, 1906

    work page 2010

  38. [38]

    S., et al

    Moehler, S., Richtler, T., de Boer, K. S., et al. 1990, A&AS, 86, 53

    work page 1990

  39. [39]

    A., Lisker, T., et al

    Napiwotzki, R., Karl, C. A., Lisker, T., et al. 2001, Ap&SS, 291, 321 N´emeth, P., Østensen, R., Tremblay, P., et al. 2014, ASPC, 481, 95 N´emeth, P., Kawka, A., & Vennes, S. 2012, MNRAS, 427, 2180 O’Connell, Robert W. 1999, ARA&A, 37, 603 21 Østensen, R. H. 2006, Baltic, 15, 85 Østensen, R. H., Silvotti, R., Charpinet, S., et al. 2010, MNRAS, 409, 1740 P...

    work page 2001

  40. [40]

    2004, AJ, 127, 899

    Sirko, Edwin., Goodman, Jeremy., Knapp, Gillian R., et al. 2004, AJ, 127, 899

    work page 2004

  41. [41]

    2015, ITNN, 27, 809

    Tang, J., Deng, C., & Huang, G. 2015, ITNN, 27, 809

    work page 2015

  42. [42]

    2015, ApJ, 696, 1755

    Tremblay, P.-E., Bergeron, P. 2015, ApJ, 696, 1755

    work page 2015

  43. [43]

    2011, MNRAS, 410, 2095

    Vennes, S., Kawka, A & Nmeth, P. 2011, MNRAS, 410, 2095

    work page 2011

  44. [44]

    2009, MNRAS, 395, 847

    Wang, B., Meng, X., Chen, X., et al. 2009, MNRAS, 395, 847

    work page 2009

  45. [45]

    2017, A&A, 599, 54

    Xiong, H., Chen, X., Podsiadlowski, Ph., et al. 2017, A&A, 599, 54

    work page 2017

  46. [46]

    X., Rix, H

    Xue, X. X., Rix, H. W., Zhao, G., et al. 2008, ApJ, 684, 1143

    work page 2008

  47. [47]

    2000, AJ, 120, 1579

    York, Donald G., Adelman, J., Anderson, John E., et al. 2000, AJ, 120, 1579

    work page 2000

  48. [48]

    Zhang, Xianfei., & Jeffery, C. S. 2012, MNRAS, 419, 452

    work page 2012

  49. [49]

    Simon., et al

    Zhang, Xianfei., Hall, Philip D.., Jeffery, C. Simon., et al. 2017, ApJ, 835, 242

    work page 2017

  50. [50]

    2006, ChJAA, 6, 265

    Zhao, Gang., Chen, Yu-Qin., Shi, Jian-Rong., et al. 2006, ChJAA, 6, 265

    work page 2006

  51. [51]

    2012, RAA, 12, 723 22

    Zhao, Gang., Zhao, Yong-Heng., Chu, Yao-Quan., et al. 2012, RAA, 12, 723 22

    work page 2012