pith. sign in

arxiv: 2603.01880 · v2 · submitted 2026-03-02 · 🌌 astro-ph.HE

Estimating the peak energy of Swift gamma-ray bursts using supervised machine learning

Wan-Peng Sun , Si-Yuan Zhu , Da-Ling Ma , Fu-Wen Zhang This is my paper

Pith reviewed 2026-05-15 17:21 UTC · model grok-4.3

classification 🌌 astro-ph.HE
keywords gamma-ray burstspeak energymachine learningSuperLearnerSwift BATFermi GBMKonus-WindEp estimation
0
0 comments X

The pith

A machine learning ensemble estimates peak energy Ep for Swift gamma-ray bursts from BAT data alone, achieving 0.72 correlation with direct measurements.

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

The paper develops a SuperLearner ensemble of supervised models to predict the peak energy Ep of gamma-ray bursts seen only by Swift/BAT. It trains on Swift observational features and uses Ep values from 516 bursts also detected by Fermi/GBM or Konus-Wind as labels. Cross-validation over 100 runs of five-fold splits yields an average Pearson r of 0.72 between estimated and observed Ep, and the estimates appear closer to true values than earlier Bayesian approaches. The model is then applied to produce Ep values for 650 additional Swift bursts. This expands the sample available for studying prompt emission physics and energy origins in GRBs.

Core claim

The central claim is that an ensemble of supervised machine learning algorithms, trained on Swift/BAT light-curve and spectral features with Ep labels from joint Fermi or Konus detections, produces reliable Ep estimates for Swift-only bursts; cross-validation shows a mean correlation r = 0.72, and the resulting values for 650 bursts are presented as statistically useful for constraining emission mechanisms.

What carries the argument

The SuperLearner ensemble that stacks multiple supervised algorithms to map Swift/BAT observational features onto Ep values.

If this is right

  • The catalog of GRBs with estimated Ep grows by 650 events.
  • These estimates supply additional data points for statistical tests of prompt emission models.
  • The ensemble method is reported to outperform previous Bayesian Ep predictions.
  • Larger samples become available for studying correlations between Ep and other burst properties.
  • New constraints on GRB energy origins become testable with the expanded set.

Where Pith is reading between the lines

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

  • The same feature set and training strategy could be tested on other incomplete high-energy transient catalogs.
  • If the 0.72 correlation persists in truly out-of-distribution bursts, the model might support rapid Ep estimates during ongoing Swift observations.
  • Incorporating additional Swift data products not used here might raise the correlation above 0.72.
  • The approach offers a template for imputing missing physical parameters in other astrophysical surveys.
  • Further validation on simulated bursts with known Ep would test whether the reported accuracy holds when ground truth is fully controlled.

Load-bearing premise

The Ep values measured by Fermi/GBM and Konus-Wind are accurate, unbiased ground-truth labels and the chosen Swift features contain enough information to predict Ep outside the training set.

What would settle it

A fresh sample of GRBs jointly detected by Swift and Fermi/GBM whose measured Ep values show a Pearson correlation with the model's predictions that is substantially below 0.72.

Figures

Figures reproduced from arXiv: 2603.01880 by Da-Ling Ma, Fu-Wen Zhang, Si-Yuan Zhu, Wan-Peng Sun.

Figure 1
Figure 1. Figure 1: Distributions of the four input quantities, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Correlation heatmap of various parameters in the training [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Feature importance scores trained by the random forest [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Optimization plot of hyperparameters in the random forest algorithm. Panels (a) and (b) show the variations in RMSE and [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Optimization of XGBoost hyperparameters. Panels (a) and (b) show the dependence of the RMSE and the Pearson correlation [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Relation between the estimated peak energy ( [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Relation between the estimated peak energy ( [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distributions of root mean squared error (left) and Pearson correlation coefficient (right) over 100 runs of five-fold cross [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of the peak energy distribution of the 516 [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Relationship between 𝐸p,z and 𝐸iso of BAT GRBs. Deep blue (light blue) dots represent SGRBs in the training (gener￾alization) set, and orange (green) dots represent LGRBs in the training (generalization) set. The solid red line represents the best-fit line for all LGRBs, and the dashed line represents the 2𝜎 confidence interval. 46 47 48 49 50 51 52 53 54 log Liso (erg/s) 1 2 3 4 lo g E p, z (k e V) 2 zon… view at source ↗
Figure 13
Figure 13. Figure 13: Relationship between 𝐸p,z and 𝐿iso of BAT GRBs. The caption is the same as that of [PITH_FULL_IMAGE:figures/full_fig_p009_13.png] view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of the observed peak energy with the esti [PITH_FULL_IMAGE:figures/full_fig_p011_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Comparison of the peak energies estimated from the [PITH_FULL_IMAGE:figures/full_fig_p012_16.png] view at source ↗
read the original abstract

Gamma-ray bursts (GRBs) are among the most energetic explosive phenomena in the Universe, and their peak energy ($E_{\rm p}$) is a key physical quantity for understanding the prompt emission mechanism. However, due to the limited energy coverage of the Swift satellite, a large fraction of Swift GRBs lack reliable peak energy measurements. Therefore, developing an accurate and efficient method for estimating $E_{\rm p}$ is of great importance. In this work, we propose a method based on the SuperLearner framework that integrates multiple supervised machine learning algorithms to estimate the $E_{\rm p}$ of Swift/BAT GRBs. We used the Swift/BAT observational data from December 2004 to September 2022 as training features, and adopted the peak energies of 516 GRBs jointly detected by Swift and either Fermi/GBM or Konus-Wind as training labels. After training and testing multiple supervised models, the final SuperLearner ensemble yields a more robust and reliable predictive model. In 100 iterations of five-fold cross-validation, the estimated $E'_{\rm p}$ values show a tight correlation with the observed $E_{\rm p}$, with an average Pearson correlation coefficient of $r = 0.72$. Compared with previous Bayesian estimates, our model provides estimations that are likely closer to the true values. Based on the trained model, we further estimated the peak energies of 650 Swift GRBs, significantly increasing the number of GRBs with estimated peak energies and providing new statistical support for constraining GRB emission mechanisms and energy origins.

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

3 major / 2 minor

Summary. The manuscript proposes a SuperLearner ensemble of supervised machine learning models to estimate the peak energy Ep of Swift/BAT gamma-ray bursts. Using observational features from Swift/BAT data for 516 GRBs jointly detected with Fermi/GBM or Konus-Wind as training labels, the model achieves an average Pearson correlation r = 0.72 in 100 iterations of five-fold cross-validation. The trained model is then applied to estimate Ep for 650 additional Swift GRBs.

Significance. If the reported correlation generalizes beyond the joint-detection training distribution, this work would substantially expand the sample of GRBs with Ep estimates, providing new statistical power for constraining prompt emission mechanisms. The ensemble approach and repeated cross-validation procedure are positive aspects that support robustness within the joint-detection sample.

major comments (3)
  1. [Abstract] Abstract: the claim that the model provides estimations 'likely closer to the true values' than previous Bayesian estimates lacks any quantitative head-to-head comparison (e.g., RMSE or correlation on the same 516-object sample) or independent hold-out validation.
  2. [Methods] Methods/Results: the training set is conditioned on joint detection by Fermi/GBM or Konus-Wind; no Kolmogorov-Smirnov test or covariate-shift diagnostic is reported to verify that the Swift/BAT feature distributions (fluence, T90, hardness ratios) of the 650 target bursts lie within the training support.
  3. [Results] Results: the average r = 0.72 is presented without per-iteration standard deviation, without comparison to a simple linear or random-forest baseline on identical features, and without assessment of whether the correlation remains stable when the training set is restricted to the Swift-only fluence range.
minor comments (2)
  1. [Abstract] Abstract and §2: the specific Swift/BAT features fed to the SuperLearner and the identities/weights of the base learners are not enumerated.
  2. [Figures] Figure captions: axis labels and units for the Ep vs. E'p scatter plots should be stated explicitly (e.g., keV in observer frame).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which have helped us identify areas for improvement in clarity and robustness. We address each major comment below and have revised the manuscript accordingly where changes are warranted.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the model provides estimations 'likely closer to the true values' than previous Bayesian estimates lacks any quantitative head-to-head comparison (e.g., RMSE or correlation on the same 516-object sample) or independent hold-out validation.

    Authors: We agree that the statement lacks supporting quantitative evidence from a direct comparison on the same sample. We will revise the abstract to remove this claim entirely and instead highlight only the cross-validation Pearson correlation and the expanded sample size. No head-to-head comparison was conducted in the original analysis. revision: yes

  2. Referee: [Methods] Methods/Results: the training set is conditioned on joint detection by Fermi/GBM or Konus-Wind; no Kolmogorov-Smirnov test or covariate-shift diagnostic is reported to verify that the Swift/BAT feature distributions (fluence, T90, hardness ratios) of the 650 target bursts lie within the training support.

    Authors: This is a valid point regarding potential covariate shift. We will add Kolmogorov-Smirnov tests (and report p-values) comparing the distributions of fluence, T90, and hardness ratios between the 516 training GRBs and the 650 target GRBs. We will also include a brief discussion of the results and any limitations if shifts are detected. revision: yes

  3. Referee: [Results] Results: the average r = 0.72 is presented without per-iteration standard deviation, without comparison to a simple linear or random-forest baseline on identical features, and without assessment of whether the correlation remains stable when the training set is restricted to the Swift-only fluence range.

    Authors: We will report the standard deviation of the Pearson r across the 100 cross-validation iterations. We will add direct comparisons to linear regression and random forest baselines trained on the identical feature set. We will also include an additional stability check by restricting the training set to the fluence range observed in Swift-only bursts and re-evaluating the correlation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; ML pipeline uses independent labels and held-out CV

full rationale

The paper trains a SuperLearner ensemble on Swift/BAT features to predict Ep labels taken from independent Fermi/GBM and Konus-Wind measurements on 516 joint detections. Performance is reported via 100 iterations of five-fold cross-validation on held-out folds, yielding r = 0.72; this metric is computed on data never seen during training and therefore cannot reduce to the fitted parameters by construction. The subsequent estimates for the 650 additional Swift bursts are out-of-sample predictions from the trained model. No self-citation chain, self-definitional loop, or renaming of a fitted quantity as a prediction appears in the derivation. The central claim therefore rests on external labels and standard cross-validation rather than on tautological reuse of its own inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on a trained ensemble whose internal parameters are learned from data; the only explicit external inputs are the Fermi/Konus labels and the assumption that BAT features are predictive.

free parameters (1)
  • SuperLearner hyperparameters and base-learner weights
    All algorithm-specific tuning parameters and stacking weights are fitted during training and not reported.
axioms (2)
  • domain assumption Fermi/GBM and Konus-Wind Ep measurements are accurate and unbiased ground truth for the Swift population.
    Invoked when treating joint detections as training labels.
  • domain assumption The selected BAT observational features are sufficient to predict Ep.
    Required for the model to generalize beyond the training set.

pith-pipeline@v0.9.0 · 5592 in / 1340 out tokens · 46301 ms · 2026-05-15T17:21:23.815822+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

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

  • IndisputableMonolith/Cost/FunctionalEquation washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    we propose a method based on the SuperLearner framework that integrates multiple supervised machine learning algorithms to estimate the Ep of Swift/BAT GRBs... average Pearson correlation coefficient of r = 0.72

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

52 extracted references · 52 canonical work pages

  1. [1]

    & Razzaque, S

    Aldowma, T. & Razzaque, S. 2024, MNRAS, 529, 2676

    work page 2024

  2. [2]

    2006, Monthly Notices of the Royal Astronomical Society, 372, 233

    Amati, L. 2006, Monthly Notices of the Royal Astronomical Society, 372, 233

    work page 2006

  3. [3]

    2002, A&A, 390, 81

    Amati, L., Frontera, F., Tavani, M., et al. 2002, A&A, 390, 81

    work page 2002

  4. [4]

    1993, ApJ, 413, 281 Bošnjak, Ž., Götz, D., Bouchet, L., Schanne, S., & Cordier, B

    Band, D., Matteson, J., Ford, L., et al. 1993, ApJ, 413, 281 Bošnjak, Ž., Götz, D., Bouchet, L., Schanne, S., & Cordier, B. 2014, A&A, 561, A25

    work page 1993

  5. [5]

    2001, Mach

    Breiman, L. 2001, Mach. Learn., 45, 5–32

    work page 2001

  6. [6]

    R., Kocevski, D., Bloom, J

    Butler, N. R., Kocevski, D., Bloom, J. S., & Curtis, J. L. 2007, ApJ, 671, 656

    work page 2007

  7. [7]

    I., Firmani, C., Avila-Reese, V., et al

    Cabrera, J. I., Firmani, C., Avila-Reese, V., et al. 2007, MNRAS, 382, 342

    work page 2007

  8. [8]

    2024, MNRAS, 527, 4272

    Chen, J.-M., Zhu, K.-R., Peng, Z.-Y., & Zhang, L. 2024, MNRAS, 527, 4272

    work page 2024

  9. [9]

    & Guestrin, C

    Chen, T. & Guestrin, C. 2016, in Proceedings of the 22nd ACM SIGKDD In- ternational Conference on Knowledge Discovery and Data Mining, KDD ’16 (New York, NY, USA: Association for Computing Machinery), 785–794 Cui,X.-H.,Liang,E.-W.,&Lu,R.-J.2005,ChineseJ.Astron.Astrophys.,5,151

    work page 2016

  10. [10]

    G., Bhardwaj, S., Cook, C., et al

    Dainotti, M. G., Bhardwaj, S., Cook, C., et al. 2025, ApJS, 277, 31

    work page 2025

  11. [11]

    G., Bogdan, M., Narendra, A., et al

    Dainotti, M. G., Bogdan, M., Narendra, A., et al. 2021, ApJ, 920, 118

    work page 2021

  12. [12]

    G., Taira, E., Wang, E., et al

    Dainotti, M. G., Taira, E., Wang, E., et al. 2024, ApJS, 271, 22

    work page 2024

  13. [13]

    2025 [arXiv:2510.26355]

    Du, G.-H., Li, T.-N., Ling, J.-L., et al. 2025 [arXiv:2510.26355]

    work page arXiv 2025

  14. [14]

    2024, Astronomy and Computing, 48, 100851

    Fotopoulou, S. 2024, Astronomy and Computing, 48, 100851

    work page 2024

  15. [15]

    2009, A&A, 496, 585

    Ghirlanda, G., Nava, L., Ghisellini, G., Celotti, A., & Firmani, C. 2009, A&A, 496, 585

    work page 2009

  16. [16]

    Ghirlanda, G., Nava, L., Ghisellini, G., Firmani, C., & Cabrera, J. I. 2008, Mon. Not. Roy. Astron. Soc., 387, 319

    work page 2008

  17. [17]

    2014, ApJS, 211, 12

    Gruber, D., Goldstein, A., Weller von Ahlefeld, V., et al. 2014, ApJS, 211, 12

    work page 2014

  18. [18]

    2013, ApJ, 770, 32

    Guiriec, S., Daigne, F., Hascoët, R., et al. 2013, ApJ, 770, 32

    work page 2013

  19. [19]

    2024, Eur

    Han, T., Jin, S.-J., Zhang, J.-F., & Zhang, X. 2024, Eur. Phys. J. C, 84, 663

    work page 2024

  20. [20]

    2026, European Physical Journal C, 86, 8

    Han, T., Zhang, J.-F., & Zhang, X. 2026, European Physical Journal C, 86, 8

    work page 2026

  21. [21]

    2019, Nature Communications, 10, 1504

    Ito, H., Matsumoto, J., Nagataki, S., et al. 2019, Nature Communications, 10, 1504

    work page 2019

  22. [22]

    & Singal, J

    Jones, E. & Singal, J. 2020, PASP, 132, 024501

    work page 2020

  23. [23]

    & Wang, Y

    Li, L. & Wang, Y. 2024, arXiv e-prints, arXiv:2412.08226

    work page arXiv 2024

  24. [24]

    2015, ApJ, 813, 116

    Liang, E.-W., Lin, T.-T., Lü, J., et al. 2015, ApJ, 813, 116

    work page 2015

  25. [25]

    & Sari, R

    Lithwick, Y. & Sari, R. 2001, ApJ, 555, 540

    work page 2001

  26. [26]

    2023, MNRAS, 518, 1629

    Luo, J.-W., Zhu-Ge, J.-M., & Zhang, B. 2023, MNRAS, 518, 1629

    work page 2023

  27. [27]

    Minaev, P. Y. & Pozanenko, A. S. 2020, MNRAS, 492, 1919

    work page 2020

  28. [28]

    G., Sarkar, M., et al

    Narendra, A., Dainotti, M. G., Sarkar, M., et al. 2025, A&A, 698, A92

    work page 2025

  29. [29]

    J., Dainotti, M

    Narendra, A., Gibson, S. J., Dainotti, M. G., et al. 2022, ApJS, 259, 55 Pe’er, A. & Ryde, F. 2011, ApJ, 732, 49

    work page 2022

  30. [30]

    2004, Reviews of Modern Physics, 76, 1143

    Piran, T. 2004, Reviews of Modern Physics, 76, 1143

    work page 2004

  31. [31]

    2021, ApJ, 913, 60

    Poolakkil, S., Preece, R., Fletcher, C., et al. 2021, ApJ, 913, 60

    work page 2021

  32. [32]

    D., Briggs, M

    Preece, R. D., Briggs, M. S., Mallozzi, R. S., et al. 2000, ApJS, 126, 19 Rodriguez,J.-V.,Rodriguez-Rodriguez,I.,&Woo,W.L.2022,WILEYINTER- DISCIPLINARY REVIEWS-DATA MINING AND KNOWLEDGE DIS- COVERY, 12

    work page 2000

  33. [33]

    D., Barbier, L., et al

    Sakamoto, T., Barthelmy, S. D., Barbier, L., et al. 2008, ApJS, 175, 179

    work page 2008

  34. [34]

    D., Baumgartner, W

    Sakamoto, T., Barthelmy, S. D., Baumgartner, W. H., et al. 2011, ApJS, 195, 2 Sakamoto,T.,Sato,G.,Barbier,L.,etal.2006,inAAS/HighEnergyAstrophysics

    work page 2011

  35. [35]

    & Nemiroff, R

    Shahmoradi, A. & Nemiroff, R. J. 2010, MNRAS, 407, 2075

    work page 2010

  36. [36]

    2025, ApJ, 980, 185

    Sun, W.-P., Zhang, J.-G., Li, Y., et al. 2025, ApJ, 980, 185

    work page 2025

  37. [37]

    2026, ApJ, 998, 339

    Sun, W.-P., Zhang, Y.-K., Zhang, J.-G., et al. 2026, ApJ, 998, 339

    work page 2026

  38. [38]

    S., Frederiks, D

    Svinkin, D. S., Frederiks, D. D., Aptekar, R. L., et al. 2016, ApJS, 224, 10

    work page 2016

  39. [39]

    2017, ApJ, 850, 161

    Tsvetkova, A., Frederiks, D., Golenetskii, S., et al. 2017, ApJ, 850, 161

    work page 2017

  40. [40]

    2021, ApJ, 908, 83 van der Laan Mark J., C, P

    Tsvetkova, A., Frederiks, D., Svinkin, D., et al. 2021, ApJ, 908, 83 van der Laan Mark J., C, P. E., & E., H. A. 2007, Statistical Applications in Genetics and Molecular Biology, 6, 1

    work page 2021

  41. [41]

    J., Qin, Y., Zhang, B., & Liang, E

    Virgili, F. J., Qin, Y., Zhang, B., & Liang, E. 2012, MNRAS, 424, 2821 von Kienlin, A., Meegan, C. A., Paciesas, W. S., et al. 2020, ApJ, 893, 46

    work page 2012

  42. [42]

    2025, Universe, 11, 355

    Wang, Z.-N., Qiang, D.-C., & Yang, S. 2025, Universe, 11, 355

    work page 2025

  43. [43]

    Wei, D. M. & Gao, W. H. 2003, MNRAS, 345, 743

    work page 2003

  44. [44]

    2023, A&A, 673, A20

    Xu, F., Huang, Y.-F., Geng, J.-J., et al. 2023, A&A, 673, A20

    work page 2023

  45. [45]

    2004, ApJ, 609, 935

    Yonetoku, D., Murakami, T., Nakamura, T., et al. 2004, ApJ, 609, 935

    work page 2004

  46. [46]

    2007, ApJ, 655, L25

    Zhang, B., Zhang, B.-B., Liang, E.-W., et al. 2007, ApJ, 655, L25

    work page 2007

  47. [47]

    J., et al

    Zhang, B., Zhang, B.-B., Virgili, F. J., et al. 2009, ApJ, 703, 1696

    work page 2009

  48. [48]

    Zhang, B. et al. 2007, Astrophys. J., 655, 989

    work page 2007

  49. [49]

    2012, ApJ, 750, 88

    Zhang, F.-W., Shao, L., Yan, J.-Z., & Wei, D.-M. 2012, ApJ, 750, 88

    work page 2012

  50. [50]

    B., Zhang, C

    Zhang, Z. B., Zhang, C. T., Zhao, Y. X., et al. 2018, PASP, 130, 054202 Zhu,S.-Y.,Deng,H.-Y.,Zhang,F.-W.,Mo,Q.-Z.,&Tam,P.-H.T.2025a,MNRAS, 541, 3236

    work page 2018

  51. [51]

    2023, ApJ, 950, 30

    Zhu, S.-Y., Liu, Z.-Y., Shi, Y.-R., et al. 2023, ApJ, 950, 30

    work page 2023

  52. [52]

    2024, MNRAS, 532, 1434 Article number, page 12

    Zhu, S.-Y., Sun, W.-P., Ma, D.-L., & Zhang, F.-W. 2024, MNRAS, 532, 1434 Article number, page 12

    work page 2024