pith. sign in

arxiv: 2607.02225 · v1 · pith:XJODKBSUnew · submitted 2026-07-02 · ✦ hep-ex

Heavy-Flavor Electron Classification Using Hadronic Environment as Point Cloud

Jingyu Zhang , Wanbing He , Long Ma This is my paper

Pith reviewed 2026-07-03 02:47 UTC · model grok-4.3

classification ✦ hep-ex
keywords heavy-flavor electronscharm bottom classificationpoint cloudtransformerhadronic environmentmachine learningsemi-leptonic decaysBDT baseline
0
0 comments X

The pith

Representing the hadronic environment as a point cloud lets set-based networks classify charm-origin versus bottom-origin electrons at roughly 80 percent purity for 40 percent efficiency.

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

The paper seeks to separate electrons from semi-leptonic decays of charm hadrons from those of bottom hadrons, which is difficult because their decay topologies are similar. It does this by treating the surrounding hadrons as an unordered point cloud and feeding that representation into several set-based machine learning architectures, including Transformers. These models reach about 80 percent purity at 40 percent efficiency on simulated test data and outperform a boosted decision tree that uses hand-crafted observables. A sympathetic reader would care because better separation would sharpen measurements of heavy-flavor production and decay rates in high-energy collisions. The work finds that performance is similar across architectures, that the networks respond mainly to geometric and topological features, and that they capture discriminating information beyond a small set of manually defined variables.

Core claim

We represent the hadronic environment as a point cloud and apply set-based machine learning architectures including Transformer models to distinguish electrons originating from charm versus bottom hadron decays. Comparable performance across architectures indicates that the dominant limitation is the intrinsic similarity between charm- and bottom-related hadronic structures rather than model expressivity. At an experimentally relevant working point of approximately 40 percent efficiency the classifier achieves a purity close to 80 percent on the test dataset and significantly improves upon a hand-crafted observable BDT baseline. Relation of model response to physics observables together with

What carries the argument

Hadronic environment represented as a point cloud and processed by set-based neural network architectures such as Transformers.

If this is right

  • The classifier improves separation of charm-origin and bottom-origin electrons relative to boosted decision trees that rely on hand-crafted observables.
  • The learned representation responds primarily to geometric and topological properties of the hadronic environment.
  • Comparable results across different set-based architectures imply that physical similarity of the structures sets the performance limit.
  • The representation extracts discriminating information that is not captured by a small set of manually constructed high-level observables.

Where Pith is reading between the lines

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

  • If the simulation matches data, the method could be deployed in LHC analyses to refine heavy-flavor cross-section and decay measurements.
  • The point-cloud approach might be tested on other particle classification tasks where surrounding environment topology supplies useful discrimination.
  • Results from feature perturbation tests could suggest new hand-crafted variables that combine geometric information with existing observables.
  • Adding tracking or vertex information to the point cloud input might raise purity without requiring larger models.

Load-bearing premise

The simulated hadronic environments used for training and testing accurately capture the geometric and topological differences between charm- and bottom-origin electrons that would be present in real detector data.

What would settle it

Direct application of the trained classifier to real collision data and measurement of whether purity at 40 percent efficiency remains near 80 percent or falls substantially below that value.

Figures

Figures reproduced from arXiv: 2607.02225 by Jingyu Zhang, Long Ma, Wanbing He.

Figure 1
Figure 1. Figure 1: Schematic illustration of the dataset construction. For each selected [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic illustration of the model architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Classification performance as a function of electron transverse mo [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Receiver operating characteristic (ROC) curves for di [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: shows the distribution of the classifier score s for charm- and bottom-origin electrons. Although the two distri￾butions exhibit overlap, a visible separation is observed, in￾dicating that the model captures nontrivial information from the hadronic environment. Rather than acting as a purely bi￾nary discriminator, the score reflects continuous variations in the event topology associated with heavy-flavor d… view at source ↗
Figure 7
Figure 7. Figure 7: Dependence of the classifier score s on several observables characterizing the hadronic environment. The corresponding observable distributions for charm- and bottom-origin electrons are shown together with the score profiles. 0.5 0.6 0.7 0.8 0.9 Classification AUC BDT: handcrafted BDT: model s + handcrafted Transformer 0.6854 0.7815 0.7810 [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of classification performance obtained using di [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Performance degradation induced by feature shu [PITH_FULL_IMAGE:figures/full_fig_p006_10.png] view at source ↗
read the original abstract

Electrons from semi-leptonic decays of charm (D) and bottom (B) hadrons are important probes in high-energy collisions, while their separation remains challenging due to the similarity of the underlying decay topologies. In this work, we represent the hadronic environment as a point cloud and investigate a hadron-based approach for distinguishing charm- and bottom-origin electrons using several set-based machine learning architectures, including Transformer models. Comparable performance is observed across different architectures, indicating that the dominant limitation originates from the intrinsic similarity between charm- and bottom-related hadronic structures rather than model expressivity. At an experimentally relevant working point corresponding to approximately 40% efficiency, the classifier achieves a purity close to 80% on the test dataset and significantly improves the classification performance relative to a hand-crafted observable BDT baseline. By studying the relation between the model response and physics-motivated observables, together with feature perturbation tests, we find that the learned representation is primarily sensitive to geometric and topological properties of the hadronic environment. Comparisons with high-level observables further suggest that the learned representation captures nontrivial discriminating information beyond a small set of manually constructed variables.

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 proposes representing the hadronic environment around electrons as a point cloud and applies set-based ML architectures (including Transformers) to classify electrons from charm versus bottom hadron decays. It reports that multiple architectures yield comparable performance, with an operating point at ~40% efficiency achieving ~80% purity on a simulated test set, outperforming a BDT baseline using hand-crafted observables; the model is shown to be sensitive primarily to geometric and topological features via response studies and perturbation tests.

Significance. If the simulation-to-data fidelity holds, the approach could offer a data-driven alternative for heavy-flavor electron tagging that captures information beyond a small set of high-level variables. The observation that performance saturates across architectures and is limited by intrinsic charm-bottom similarity rather than model capacity is a useful physics insight. The work demonstrates the viability of point-cloud methods in this context but currently lacks the validation steps needed for direct experimental application.

major comments (2)
  1. [Abstract] Abstract and results: the headline performance numbers (~40% efficiency, ~80% purity, improvement over BDT) are obtained exclusively on a simulated test dataset with no reported real-data validation, closure tests, or unfolding; given that the learned features are geometric/topological and the claim is framed as 'experimentally relevant,' this is load-bearing for the central claim.
  2. [Abstract] Abstract: no information is provided on dataset size, training/validation/test splits, or systematic uncertainties arising from simulation modeling (detector response, material budget, underlying event); these details are required to assess whether the reported purity is robust.
minor comments (1)
  1. The manuscript would benefit from explicit statements on how the point-cloud representation handles variable multiplicity and detector acceptance effects.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the scope and reporting of our simulation-based study. We address each point below, clarifying the methodological focus while proposing targeted revisions to improve transparency.

read point-by-point responses

  1. Referee: [Abstract] Abstract and results: the headline performance numbers (~40% efficiency, ~80% purity, improvement over BDT) are obtained exclusively on a simulated test dataset with no reported real-data validation, closure tests, or unfolding; given that the learned features are geometric/topological and the claim is framed as 'experimentally relevant,' this is load-bearing for the central claim.

    Authors: We agree that all reported metrics are obtained on a simulated test set, as ground-truth labels from the generator are required for supervised training and evaluation. The study is framed as a proof-of-concept demonstration of point-cloud methods rather than a ready-to-deploy experimental tool. We will revise the abstract to explicitly state that results are from Monte Carlo simulation and add a new paragraph in the discussion section outlining the additional validation steps (closure tests, data-driven calibration) needed before experimental application. revision: yes

  2. Referee: [Abstract] Abstract: no information is provided on dataset size, training/validation/test splits, or systematic uncertainties arising from simulation modeling (detector response, material budget, underlying event); these details are required to assess whether the reported purity is robust.

    Authors: The full manuscript provides dataset sizes, split ratios, and training details in the Methods section, but we acknowledge the abstract should be self-contained. We will update the abstract with concise statements on dataset scale and splits. For simulation systematics, we will add a short subsection discussing sensitivity to variations in detector response and underlying event modeling, while noting that a complete systematic evaluation requires experimental data and is beyond the current scope. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical ML performance on held-out simulation with no derivation chain

full rationale

The paper reports classification metrics (efficiency, purity, improvement over BDT) obtained by training set-based models on simulated point-cloud data and evaluating on a held-out test set. No equations, functional forms, or uniqueness theorems are claimed; the headline numbers are direct empirical outputs of standard supervised training rather than quantities derived from or fitted to the reported results themselves. No self-citation load-bearing steps, ansatz smuggling, or renaming of known results appear in the provided text. The result is therefore self-contained as a simulation-based benchmark.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central performance claim rests on the assumption that simulation faithfully reproduces real hadronic environments and that the chosen point-cloud features are sufficient; no explicit free parameters, axioms, or invented entities are stated in the abstract.

pith-pipeline@v0.9.1-grok · 5725 in / 1178 out tokens · 21194 ms · 2026-07-03T02:47:15.300986+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

38 extracted references · 34 canonical work pages · 11 internal anchors

  1. [1]

    Extraction of Heavy-Flavor Transport Coefficients in QCD Matter

    A. Beraudo, et al., Extraction of Heavy-Flavor Transport Coefficients in QCD Matter, Nucl. Phys. A 979 (2018) 21–86. arXiv:1803.03824, doi:10.1016/j.nuclphysa.2018.09.002

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/j.nuclphysa.2018.09.002 2018

  2. [2]

    X. Dong, V . Greco, Heavy quark production and proper- ties of Quark–Gluon Plasma, Prog. Part. Nucl. Phys. 104 (2019) 97–141. doi:10.1016/j.ppnp.2018.08.001

    work page doi:10.1016/j.ppnp.2018.08.001 2019

  3. [3]

    M. He, H. van Hees, R. Rapp, Heavy-quark dif- fusion in the quark–gluon plasma, Prog. Part. Nucl. Phys. 130 (2023) 104020. arXiv:2204.09299, doi:10.1016/j.ppnp.2023.104020

    work page doi:10.1016/j.ppnp.2023.104020 2023

  4. [4]

    J. Y . Zhang, Y . J. Song, L. Ma, W. H. Ma, H. Z. Huang, Effect of initial geometry fluctuations on the heavy fla- vor flow using a multiphase transport model, Phys. Rev. C 112 (4) (2025) 044901. doi:10.1103/2xt8-bh6x

    work page doi:10.1103/2xt8-bh6x 2025

  5. [5]

    J. Zhao, K. Zhou, S. Chen, P. Zhuang, Heavy fla- vors under extreme conditions in high energy nuclear collisions, Prog. Part. Nucl. Phys. 114 (2020) 103801. arXiv:2005.08277, doi:10.1016/j.ppnp.2020.103801

    work page doi:10.1016/j.ppnp.2020.103801 2020

  6. [6]

    Azimuthal anisotropy of heavy-flavour decay electrons in p-Pb collisions at $ \sqrt{s_{\rm NN}}$ = 5.02 TeV

    S. Acharya, et al., Azimuthal Anisotropy of Heavy-Flavor Decay Electrons inp-Pb Colli- sions at √sNN =5.02 TeV, Phys. Rev. Lett. 122 (7) (2019) 072301. arXiv:1805.04367, doi:10.1103/PhysRevLett.122.072301

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1103/physrevlett.122.072301 2019

  7. [7]

    Masciocchi, Investigation of charm and beauty pro- duction via semileptonic decays of heavy-flavour hadrons in pp at 7 TeV and Pb–Pb at 2.76 TeV with AL- ICE, J

    S. Masciocchi, Investigation of charm and beauty pro- duction via semileptonic decays of heavy-flavour hadrons in pp at 7 TeV and Pb–Pb at 2.76 TeV with AL- ICE, J. Phys. G 38 (2011) 124069. arXiv:1109.6436, doi:10.1088/0954-3899/38/12/124069

    work page doi:10.1088/0954-3899/38/12/124069 2011

  8. [8]

    B. B. Abelev, et al., Measurement of electrons from semileptonic heavy-flavor hadron decays inppcollisions at √s=2.76 TeV, Phys. Rev. D 91 (1) (2015) 012001. arXiv:1405.4117, doi:10.1103/PhysRevD.91.012001

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1103/physrevd.91.012001 2015

  9. [9]

    Measurement of electrons from beauty hadron decays in pp collisions at $\sqrt{s} = 7$ TeV

    B. Abelev, et al., Measurement of electrons from beauty hadron decays inppcollisions at √s= 7 TeV, Phys. Lett. B 721 (2013) 13–23, [Erratum: Phys.Lett.B 763, 507–509 (2016)]. arXiv:1208.1902, doi:10.1016/j.physletb.2013.01.069

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/j.physletb.2013.01.069 2013

  10. [10]

    Averbeck, Open heavy flavor measurements with the PHENIX experiment at RHIC, J

    R. Averbeck, Open heavy flavor measurements with the PHENIX experiment at RHIC, J. Phys. G 31 (2005) S259–S266. arXiv:nucl-ex/0410007, doi:10.1088/0954- 3899/31/4/032

    work page doi:10.1088/0954- 2005

  11. [11]

    Heavy Quark Production in p+p and Energy Loss and Flow of Heavy Quarks in Au+Au Collisions at sqrt(s_NN)=200 GeV

    A. Adare, et al., Heavy Quark Production inp+p and Energy Loss and Flow of Heavy Quarks in Au+Au Collisions at √sNN =200 GeV, Phys. Rev. C 84 (2011) 044905. arXiv:1005.1627, doi:10.1103/PhysRevC.84.044905

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1103/physrevc.84.044905 2011

  12. [12]

    Adare, et al., Single electron yields from semileptonic charm and bottom hadron decays in Au+Au collisions at√sNN =200 GeV, Phys

    A. Adare, et al., Single electron yields from semileptonic charm and bottom hadron decays in Au+Au collisions at√sNN =200 GeV, Phys. Rev. C 93 (3) (2016) 034904. arXiv:1509.04662, doi:10.1103/PhysRevC.93.034904

    work page doi:10.1103/physrevc.93.034904 2016

  13. [13]

    Acharya, et al., Measurement of electrons from semileptonic heavy-flavour hadron decays at midrapid- ity in pp and Pb-Pb collisions at √sNN =5.02 TeV, Phys

    S. Acharya, et al., Measurement of electrons from semileptonic heavy-flavour hadron decays at midrapid- ity in pp and Pb-Pb collisions at √sNN =5.02 TeV, Phys. Lett. B 804 (2020) 135377. arXiv:1910.09110, doi:10.1016/j.physletb.2020.135377

    work page doi:10.1016/j.physletb.2020.135377 2020

  14. [14]

    Acharya, et al., Measurement of beauty and charm production in pp collisions at √s=5.02 TeV via non- prompt and prompt D mesons, JHEP 05 (2021) 220

    S. Acharya, et al., Measurement of beauty and charm production in pp collisions at √s=5.02 TeV via non- prompt and prompt D mesons, JHEP 05 (2021) 220. arXiv:2102.13601, doi:10.1007/JHEP05(2021)220

    work page doi:10.1007/jhep05(2021)220 2021

  15. [15]

    M. S. Abdallah, et al., Evidence of Mass Ordering of Charm and Bottom Quark Energy Loss in Au+Au Col- lisions at RHIC, Eur. Phys. J. C 82 (12) (2022) 1150, [Er- ratum: Eur.Phys.J.C 83, 455 (2023)]. arXiv:2111.14615, doi:10.1140/epjc/s10052-022-11003-7

    work page doi:10.1140/epjc/s10052-022-11003-7 2022

  16. [16]

    Acharya, et al., Measurement of electrons from heavy- flavour hadron decays as a function of multiplicity in p- Pb collisions at √sNN =5.02 TeV, JHEP 02 (2020) 077

    S. Acharya, et al., Measurement of electrons from heavy- flavour hadron decays as a function of multiplicity in p- Pb collisions at √sNN =5.02 TeV, JHEP 02 (2020) 077. arXiv:1910.14399, doi:10.1007/JHEP02(2020)077

    work page doi:10.1007/jhep02(2020)077 2020

  17. [17]

    Entanglement and Chaos near critical point in strongly coupled Gauge theory

    S. Acharya, et al., Azimuthal correlations of heavy-flavor hadron decay electrons with charged particles in pp and p–Pb collisions at √sNN √sNN √sNN =5.02 TeV, Eur. Phys. J. C 83 (8) (2023) 741. arXiv:2303.00591, doi:10.1140/epjc/s10052- 023-11835-x

    work page doi:10.1140/epjc/s10052- 2023

  18. [18]

    Thomas, F

    D. Thomas, F. Colamaria, Recent Findings from Heavy-Flavor Angular Correlation Measurements in Hadronic Collisions, Universe 10 (3) (2024) 109. arXiv:2403.01035, doi:10.3390/universe10030109

    work page doi:10.3390/universe10030109 2024

  19. [19]

    Heavy-flavour and quarkonium production in the LHC era: from proton-proton to heavy-ion collisions

    A. Andronic, et al., Heavy-flavour and quarkonium pro- duction in the LHC era: from proton–proton to heavy- ion collisions, Eur. Phys. J. C 76 (3) (2016) 107. arXiv:1506.03981, doi:10.1140/epjc/s10052-015-3819-5

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1140/epjc/s10052-015-3819-5 2016

  20. [20]

    Measurement of Bottom versus Charm as a Function of Transverse Momentum with Electron-Hadron Correlations in p+p Collisions at sqrt(s)=200 GeV

    A. Adare, et al., Measurement of Bottom versus Charm as a Function of Transverse Momentum with Electron- Hadron Correlations inp +pCollisions at √s=200 GeV, Phys. Rev. Lett. 103 (2009) 082002. arXiv:0903.4851, doi:10.1103/PhysRevLett.103.082002

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1103/physrevlett.103.082002 2009

  21. [21]

    P. T. Komiske, E. M. Metodiev, J. Thaler, Energy Flow Networks: Deep Sets for Particle Jets, JHEP 01 (2019)

    2019

  22. [22]

    arXiv:1810.05165, doi:10.1007/JHEP01(2019)121. 7

    work page doi:10.1007/jhep01(2019)121 2019

  23. [23]

    L. Liu, X. Liu, J. Gao, W. Chen, J. Han, Understanding the difficulty of training transformers, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP), 2020, pp. 5747–5763

    2020

  24. [24]

    Machine Learning in High Energy Physics Community White Paper

    K. Albertsson, et al., Machine Learning in High En- ergy Physics Community White Paper, J. Phys. Conf. Ser. 1085 (2) (2018) 022008. arXiv:1807.02876, doi:10.1088/1742-6596/1085/2/022008

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1088/1742-6596/1085/2/022008 2018

  25. [25]

    Karagiorgi, G

    G. Karagiorgi, G. Kasieczka, S. Kravitz, B. Nachman, D. Shih, Machine Learning in the Search for New Fun- damental Physics (12 2021). arXiv:2112.03769

    work page arXiv 2021

  26. [26]

    Mondal, L

    S. Mondal, L. Mastrolorenzo, Machine learning in high energy physics: a review of heavy-flavor jet tagging at the LHC, Eur. Phys. J. ST 233 (15-16) (2024) 2657–

    2024

  27. [27]

    arXiv:2404.01071, doi:10.1140/epjs/s11734-024- 01234-y

    work page doi:10.1140/epjs/s11734-024-

  28. [28]

    Deep Sets

    M. Zaheer, S. Kottur, S. Ravanbakhsh, B. Poczos, R. Salakhutdinov, A. Smola, Deep Sets (3 2017). arXiv:1703.06114

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  29. [29]

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, J. M. Solomon, Dynamic Graph CNN for Learning on Point Clouds (1 2018). arXiv:1801.07829

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  30. [30]

    J. Lee, Y . Lee, J. Kim, A. Kosiorek, S. Choi, Y . W. Teh, Set transformer: A framework for attention-based permutation-invariant neural networks, in: International conference on machine learning, PMLR, 2019, pp. 3744– 3753

    2019

  31. [31]

    Carleo, I

    G. Carleo, I. Cirac, K. Cranmer, L. Daudet, M. Schuld, N. Tishby, L. V ogt-Maranto, L. Zdeborová, Ma- chine learning and the physical sciences, Rev. Mod. Phys. 91 (4) (2019) 045002. arXiv:1903.10563, doi:10.1103/RevModPhys.91.045002

    work page doi:10.1103/revmodphys.91.045002 2019

  32. [32]

    Kasieczka, et al., The LHC Olympics 2020 a com- munity challenge for anomaly detection in high en- ergy physics, Rept

    G. Kasieczka, et al., The LHC Olympics 2020 a com- munity challenge for anomaly detection in high en- ergy physics, Rept. Prog. Phys. 84 (12) (2021) 124201. arXiv:2101.08320, doi:10.1088/1361-6633/ac36b9

    work page doi:10.1088/1361-6633/ac36b9 2020

  33. [33]

    Deep Learning and its Application to LHC Physics

    D. Guest, K. Cranmer, D. Whiteson, Deep Learn- ing and its Application to LHC Physics, Ann. Rev. Nucl. Part. Sci. 68 (2018) 161–181. arXiv:1806.11484, doi:10.1146/annurev-nucl-101917-021019

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1146/annurev-nucl-101917-021019 2018

  34. [34]

    Radovic, M

    A. Radovic, M. Williams, D. Rousseau, M. Kagan, D. Bonacorsi, A. Himmel, A. Aurisano, K. Terao, T. Wongjirad, Machine learning at the energy and inten- sity frontiers of particle physics, Nature 560 (7716) (2018) 41–48. doi:10.1038/s41586-018-0361-2

    work page doi:10.1038/s41586-018-0361-2 2018

  35. [35]

    Louppe, K

    G. Louppe, K. Cho, C. Becot, K. Cranmer, QCD-Aware Recursive Neural Networks for Jet Physics, JHEP 01 (2019) 057. arXiv:1702.00748, doi:10.1007/JHEP01(2019)057

    work page doi:10.1007/jhep01(2019)057 2019

  36. [36]

    H. Qu, L. Gouskos, ParticleNet: Jet Tagging via Par- ticle Clouds, Phys. Rev. D 101 (5) (2020) 056019. arXiv:1902.08570, doi:10.1103/PhysRevD.101.056019

    work page doi:10.1103/physrevd.101.056019 2020

  37. [37]

    H. Qu, C. Li, S. Qian, Particle Transformer for Jet Tagging (2 2022). arXiv:2202.03772

    work page arXiv 2022

  38. [38]

    Y . Zhu, H. Liang, Y . Wang, H. Qu, C. Zhou, M. Ruan, Par- ticleNet and its application on CEPC jet flavor tagging, Eur. Phys. J. C 84 (2) (2024) 152. arXiv:2309.13231, doi:10.1140/epjc/s10052-024-12475-5. 8

    work page doi:10.1140/epjc/s10052-024-12475-5 2024