Pith. sign in

REVIEW 1 major objections 4 minor 4 cited by

End-to-End Multi-Track Reconstruction using Graph Neural Networks at Belle II

T0 review · 1 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A graph neural network that reads Belle II drift-chamber hits directly reconstructs macroscopically displaced tracks at 85.4% combined efficiency with a 2.5% fake rate, far above the baseline algorithm's 52.2% and 4.1%.

desk verdict A credible, well-evaluated GNN end-to-end tracker for the Belle II drift chamber; the displaced-track gain holds up, but the prompt-number caveats and an unscanned fitter covariance deserve referee attention. read the letter →

arxiv 2411.13596 v2 pith:OW4T6ETF submitted 2024-11-18 physics.ins-det hep-exphysics.data-an

Lea Reuter , Giacomo De Pietro , Slavomira Stefkova , Torben Ferber , Valerio Bertacchi , Giulia Casarosa , Luigi Corona , Patrick Ecker
show 10 more authors
Alexander Glazov Yubo Han Martina Laurenza Thomas Lueck Ludovico Massaccesi Suryanarayan Mondal Bianca Scavino Stefano Spataro Christian Wessel Laura Zani
This is my paper · ORCID
classification physics.ins-dethep-exphysics.data-an PACS 29.40.Gx07.05.Mh
keywords trackfindinggraphneuralnetworksobjectcondensationdriftchamberBelleIIdisplacedverticesend-to-endreconstructionparticletracking
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a single graph neural network, the CAT Finder, can replace the conventional track-finding stage of Belle II's drift chamber reconstruction: it consumes raw wire hits with no pre-filtering, predicts the unknown number of charged tracks together with their momenta, starting points, and charges, and then hands each predicted track's hits to a standard Kalman filter. If correct, this would give a generic way to do tracking where detectors measure drift time rather than direct three-dimensional positions, which is exactly the regime where the existing Legendre-transform baseline loses efficiency for tracks from decays displaced by centimeters to a meter. The paper reports that for dark-Higgs-like displaced decays the CAT Finder achieves a combined finding-and-fitting charge efficiency of 85.4% per track with a 2.5% fake rate, compared to 52.2% and 4.1% for the baseline, and that for prompt tracks it performs comparably in the barrel while substantially better in both endcaps. This matters for experimental particle physics because displaced-vertex signatures are a primary search route for long-lived new particles.

What carries the argument

The mechanism is a graph neural network whose nodes are individual wire hits, processed by four GravNet blocks, a distance-weighted graph layer that builds edges between nodes in a learned representation space and aggregates features along those edges. Object condensation is the central identity: a loss that makes each node predict a condensation coordinate in a learned low-dimensional cluster space and a confidence value $\beta$, pulling hits of the same particle toward one high-$\beta$ condensation point while repelling different particles; after inference, thresholds on $\beta$ and on cluster-space distances ($t_\beta$, $t_d$, $t_h$) select the condensation points that define the tracks and their parameter predictions, and hits within a radius $t_h$ of a condensation point form the track's hit set for the subsequent fitter. This learned cluster space is what carries the argument, because it converts the unknown-number-of-tracks problem into a clustering problem whose solution is read off geometrically.

What would settle it

Run the trained CAT Finder on real Belle II collision data for $K^0_S \to \pi^+\pi^-$ decays, where the decay vertex is independently measured, and compare per-track finding-and-charge efficiency and fake rate in the barrel with the simulation's ~93% efficiency and ~5% fake rate; a large shortfall would show the simulated drift-time or wire-inefficiency model is not faithful for displaced tracks.

Watch

Extended reading notes

Core claim

The central claim is that end-to-end multi-track reconstruction in a drift chamber is feasible and outperforms the conventional finder for the most challenging topologies: the CAT Finder simultaneously predicts the number of tracks, each track's three-momentum, starting point, and charge, and assigns the detector hits that belong to it, using object condensation in a learned cluster space. The evaluation is deliberately realistic: full GEANT4 detector simulation with digitized detector response, beam-background hits overlaid from actual collision data, and wire-efficiency maps from data-taking conditions. Against this backdrop the GNN finds and correctly charges 85.4% of tracks from displaced dark-Higgs-like decays with a 2.5% fake rate, versus 52.2% and 4.1% for the baseline; it is also significantly better in the endcaps for prompt tracks and comparable in the barrel. The paper presents this as the first end-to-end machine-learning multi-track reconstruction for a drift chamber operated in a realistic particle physics environment.

Load-bearing premise

The results stand or fall on whether the GEANT4 simulation with digitized detector response, data-overlay beam backgrounds, and the corresponding wire-inefficiency maps faithfully reproduces the real Belle II CDC response for tracks that originate far from the interaction point.

Editorial extensions

If this is right

  • Long-lived-particle searches at Belle II (e.g., dark Higgs and inelastic dark matter) would see a large efficiency jump: for a benchmark dark Higgs with $m_h=1.5\,\text{GeV}$ and $c\tau=21.5\,\text{cm}$, both tracks are reconstructed in 87.2% of events versus 44.9% with the baseline, at a lower fake rate.
  • Endcap tracking, historically the weakest region for the baseline finder, improves to near-barrel levels for both prompt and displaced tracks, increasing the usable acceptance for forward and backward physics.
  • Because the GNN outputs track parameters directly from the condensation point, the hit clustering and Kalman-filter steps can be dropped for real-time triggering, opening a route to FPGA-based end-to-end tracking.
  • The robustness test (training on low simulated background, evaluating on high background from data) shows only a few-percent efficiency loss without retraining, suggesting the model can ride out changing accelerator conditions.

Reading between the lines

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

  • The same architecture should transfer to other drift chambers and gaseous tracking detectors that, like the Belle II CDC, lack direct spatial readout and must infer positions from drift time; the network learns the wire geometry, so only the input feature encoding would need to change.
  • The paper's observation that the network infers a vertex position even without nearby hits, using the companion track's direction, hints that a GNN of this kind could provide trigger-level vertexing for displaced signatures without a separate vertex-finding step.
  • The large gap between the CAT Finder's hit efficiency for curling tracks and the conventional fitter's success on them suggests that the reported 85.4% is limited by the fitter, not the finder; a fitter that accepts multi-curler hit sets could push efficiency higher.
  • The fidelity of the simulated wire inefficiency and cross-talk is untested on real displaced tracks, so the magnitudes of the reported gains should be re-measured on data before relying on them.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

1 major / 4 minor

Summary. The manuscript presents CAT Finder, a graph neural network (GNN) based end-to-end track finder for the Belle II central drift chamber. Using wire hits as node inputs, the network predicts an unknown number of tracks via object condensation, estimates each track's momentum, charge, and starting position, and clusters hits for later fitting by GENFIT2 with a deterministic annealing filter. The study is performed on GEANT4-based full detector simulation with beam backgrounds and detector conditions overlaid from recorded high-luminosity data. The central result is a combined track finding and fitting charge efficiency of 85.4% per track with a 2.5% fake rate for dark-Higgs-like displaced decays, compared with 52.2% and 4.1% for the Belle II baseline CDC finder. Additional results cover prompt tracks, radiative muon pairs, and K0S decays, along with robustness checks against background conditions and a discussion of lessons learned during training.

Significance. If the reported performance holds, this is a valuable demonstration of an end-to-end machine-learned track finder in a realistic drift chamber environment, with substantial gains over the existing algorithm for displaced decays and for the forward/backward endcap regions. The paper's strengths include statistically independent evaluation samples (about one million events total), realistic simulation including data-overlay beam backgrounds, a documented check that removing physical training samples avoids a specific overfitting mode, an explicit robustness study across background conditions, and public code for reproduction. The central claim is an empirical benchmark rather than a derivation, so there is no circularity concern; the main risk is whether the headline fake rate is robust to an arbitrary track-fitter initialization, as detailed in the major comments.

major comments (1)
  1. [Section 6.3, Table 4] The combined track finding and fitting fake rate of 2.5% (barrel 2.12% after fitting) is a headline claim, but it depends on the GENFIT2 deterministic annealing filter, whose hit-acceptance thresholds scale with the initial covariance matrix. All initial covariance entries for CAT Finder are set to 0.1, and the manuscript states only that the impact on track finding efficiency is negligible and that performance optimization is left to future work. The post-fit fake rate in the barrel drops from 5.12% (track finding) to 2.12% (after fitting), so the fitter is performing substantial fake rejection. A different diagonal covariance scale could plausibly change the DAF's hit rejection and thus the post-fit fake rate and charge efficiency. The paper should either scan the initial covariance scale (e.g., over factors from 0.1 to 10) and report efficiency, fake rate, and charge efficiency, or provide a principled initialization (for example, a covariance derived from the GNN parameter predictions) and show that the headline 2.5% fake rate is stable under this choice.
minor comments (4)
  1. [Section 7.1.1, Fig. 7, Table 3] The main prompt-track comparison removes curling tracks from the evaluation, so the claim of significantly better forward/backward performance is specific to the non-curling subset. This restriction should be stated more prominently in Section 7.1.1 and in the captions of Fig. 7 and Table 3, since the integrated prompt numbers are not directly comparable to a full inclusive sample.
  2. [Appendix D, Table D1] The caption of Table D1 states that the muon-pair results are for low beam background, while Section 7.1.2 and Fig. D5 describe these samples as using high data beam backgrounds; this inconsistency should be corrected.
  3. [Section 6.3, covariance text] Please clarify the units and structure of the initial covariance matrix (diagonal versus full, units of momentum and position entries), and briefly state what the Baseline Finder uses for its initial covariance so the reader can judge the fairness of the comparison.
  4. [Section 5, Eqs. (5) and (9)] The relationship between the track charge efficiency and the wrong-charge rate could be stated more explicitly: a matched track with the wrong charge is counted in the wrong-charge rate but not in the charge efficiency, so the two metrics do not sum to unity; making this explicit would avoid confusion when reading the tables.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor test-set tuning of the track-finding working point; the headline dark-Higgs result is independent.

  1. fitted input called prediction [Section 6.4 (Eqs. 12-13) and Section 7.1.3 / Table 5]
    "The three tracking hyperparameters tβ, td and th are optimized using samples from category 2 and K0S → π+π− events with the K0S momentum pointing in the CDC barrel region (see Section 4 for details). ... The working point w = (tβ, td, th) is chosen so that first εtrk(wi)+ptrk(wi) ≥ εtrk(baseline)+ptrk(baseline), and then max_i [εtrk(wi)category 2 + εtrk(wi)K0S→π+π−]."

    The K0S → π+π− sample is used both to select the working point (tβ = 0.3, td = 0.3, th = 0.15) via Eq. (13) and then to report the K0S performance numbers in Table 5. The optimization explicitly maximizes the K0S track-finding-and-fitting efficiency, so the K0S efficiency and fake rate presented in Section 7.1.3 are not an independent test of the algorithm: they are partially tuned on the same sample. This is a mild form of fitting a choice to the data and then presenting the resulting performance as a measured result. The headline claim for dark-Higgs-like displaced decays (Table 4) is not affected, because the working point was not optimized on the h → μ+μ− sample, so the central comparison to the baseline remains independent.

full rationale

The paper is an empirical benchmark study, not a derivation, and the central claim is an efficiency/fake-rate comparison on simulation samples. The headline dark-Higgs result (85.4% vs 52.2% efficiency, 2.5% vs 4.1% fake rate) is measured on a statistically independent sample that was not used for GNN training or hyperparameter optimization. The only identifiable circular element is the working-point selection in Section 6.4: the hyperparameters tβ, td, and th are chosen partly by maximizing efficiency on the K0S → π+π− barrel sample, and the same sample is later used for the K0S performance evaluation in Table 5. This biases the K0S numbers upward but does not affect the dark-Higgs evaluation or the overall conclusion that CAT Finder improves displaced-track finding. There is no self-definitional equation structure, no renamed known result, and no load-bearing self-citation chain; the cited object condensation and GravNet methods are external prior work. The un-scanned GENFIT2 initial covariance matrix is a robustness gap, not a circularity, so it is not counted in the score.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physics entities are introduced. The central claim rests on the fidelity of the detector simulation, the label matching procedure, the metric definitions, and the fitted neural network weights and tracking thresholds.

free parameters (6)
  • Beta threshold tBeta = 0.3
    Chosen in Section 6.4 by maximizing track finding and fitting charge efficiency on category 2 and Ks samples; controls which nodes seed condensation points.
  • Condensation distance threshold td = 0.3
    Chosen in the same optimization; minimum separation between condensation points in the learned cluster space.
  • Hit assignment radius th = 0.15
    Chosen in the same optimization; radius in cluster space within which nodes are assigned to a track.
  • GNN hyperparameters = N=4, k=54, momentum=0.77, S=4, LL width=126, CS=3, LL2 width=16
    Optimized with Weights and Biases on a 6% low-background subset using validation loss (Section 6.4, Tab. 2).
  • GNN trainable weights = 797,812 parameters
    Learned from the training sample via object condensation and parameter prediction losses; the performance claim depends on these fitted weights.
  • Track matching criteria = epsilon_hit > 0.05, purity > 0.66, nhits >= 7
    Defined in Section 5 to relate found tracks to simulated particles; these thresholds affect both efficiency and fake rate for both algorithms.
assumptions (5)
  • domain assumption GEANT4 simulation plus digitized detector response and data-overlay beam background faithfully models the Belle II CDC for displaced tracks.
    Invoked in Section 4; all efficiencies are measured on simulation, so the realism of the simulation is the bridge to real detector performance.
  • domain assumption The hit-to-particle matching rule (first hit in time, at most one particle per hit) provides correct training labels.
    Section 4 states 'Each CDC hit is matched to up to one simulated particle... first hit in time'; incorrect labels would degrade measured performance.
  • domain assumption The metric definitions in Section 5 are a fair basis for comparing CAT Finder and Baseline Finder.
    Thresholds such as purity > 0.66 and nhits >= 7 are common to both algorithms but are arbitrary and affect the reported numbers.
  • domain assumption Object condensation trained on up to 15 particles per event generalizes to an unknown number of tracks at test time.
    The model learns track multiplicity from the training distribution; the evaluation samples must lie within the learned regime.
  • domain assumption GENFIT2 with a single pion mass hypothesis and a fixed initial covariance of 0.1 is a fair track fitter for both algorithms.
    Section 6.1 and 6.3; the CAT Finder does not provide an optimized initial covariance, which may understate its combined finding and fitting efficiency.

how reviews work

0 comments
Cite this review

Pith. "Pith review of End-to-End Multi-Track Reconstruction using Graph Neural Networks at Belle II." pith.science (2026). https://pith.science/paper/OW4T6ETF

@misc{pith2026241113596,
  author       = {Pith},
  title        = {Pith review of: End-to-End Multi-Track Reconstruction using Graph Neural Networks at Belle II},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OW4T6ETF}},
  note         = {Machine review of arXiv:2411.13596}
}
read the original abstract

We present the study of an end-to-end multi-track reconstruction algorithm for the central drift chamber of the Belle II experiment at the SuperKEKB collider using Graph Neural Networks for an unknown number of particles. The algorithm uses detector hits as inputs without pre-filtering to simultaneously predict the number of track candidates in an event and their kinematic properties. In a second step, we cluster detector hits for each track candidate to pass to a track fitting algorithm. Using a realistic full detector simulation including beam-induced backgrounds and detector noise taken from actual collision data, we find significant improvements in track finding efficiencies for tracks in a variety of different event topologies compared to the existing baseline algorithm used in Belle II. For events with a hypothetical long-lived massive particle with a mass in the GeV-range decaying uniformly along its flight direction into two charged particles, the GNN achieves a combined track finding and fitting efficiency of 85.4% with a fake rate of 2.5%, compared to 52.2% and 4.1% for the baseline algorithm. This is the first end-to-end multi-track machine learning algorithm for a drift chamber detector that has been utilized in a realistic particle physics environment.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Transformer-Based Approach to Enhance Positron Tracking Performance in MEG II

    hep-ex 2025-12 conditional novelty 6.0 of 10

    A Transformer-based hit classifier improves MEG II positron tracking efficiency and resolution, yielding an expected ~10% gain in μ→eγ sensitivity.

  2. Estimating the track-reconstruction efficiency in phenomenological proposals of long-lived-particle searches

    hep-ex 2025-01 conditional novelty 6.0 of 10

    TrackEff estimates Belle II track-reconstruction efficiency by geometrically counting drift-chamber hits, with default thresholds tuned to a Belle II tau-tau tracking efficiency measurement.

  3. Physics Instrument Design with Reinforcement Learning

    physics.ins-det 2024-12 conditional novelty 6.0 of 10

    A reinforcement learning agent that sequentially places detector components produced calorimeter and spectrometer designs that outperform hand-made baselines in simulation.

  4. Real-Time Analysis of Unstructured Data with Machine Learning on Heterogeneous Architectures

    physics.data-an 2025-08 conditional novelty 5.0 of 10

    A graph neural network (ETX4VELO) reconstructs LHCb VELO tracks with performance comparable to the production 'search by triplet' algorithm while running end to end in the GPU-based first-level trigger, with additiona...

Reference graph

Works this paper leans on

52 extracted references · 42 canonical work pages · cited by 4 Pith papers

  1. [1]

    Ferber, C

    T. Ferber, C. Garcia-Cely, and K. Schmidt- Hoberg. Belle II sensitivity to long–lived dark photons. Phys. Lett. B , 833:137373, 2022

    work page 2022

  2. [2]

    Duerr et al

    M. Duerr et al. Invisible and displaced dark matter signatures at Belle II. J. High Energ. Phys., 02:039, 2020

    work page 2020

  3. [3]

    Duerr et al

    M. Duerr et al. Long-lived Dark Higgs and Inelastic Dark Matter at Belle II. J. High Energ. Phys., 04:146, 2021

    work page 2021

  4. [4]

    Natochii et al

    A. Natochii et al. Beam Background Expec- tations for Belle II at SuperKEKB. 2022. https://arxiv.org/abs/2203.05731

    arXiv 2022

  5. [5]

    Shlomi, P

    J. Shlomi, P. Battaglia, and J.-R. Vlimant. Graph neural networks in particle physics. Mach. Learn.: Sci. Technol. , 2(2):021001, 2020

    work page 2020

  6. [6]

    Wang et al

    Y. Wang et al. Dynamic Graph CNN for Learning on Point Clouds. 2018. https: //arxiv.org/abs/1801.07829

    arXiv 2018

  7. [7]

    S. R. Qasim et al. Learning representa- tions of irregular particle-detector geometry with distance-weighted graph networks. Eur. Phys. J. C , 79(7):608, 2019

    work page 2019

  8. [8]

    Wemmer et al

    F. Wemmer et al. Photon Reconstruction in the Belle II Calorimeter Using Graph Neural Networks. Comput Softw Big Sci , 7:13, 2023

    work page 2023

Show all 52 references
  1. [9]

    Kieseler

    J. Kieseler. Object condensation: one- stage grid-free multi-object reconstruction in physics detectors, graph and image data.Eur. Phys. J. C , 80(9):886, 2020

    2020

  2. [10]

    S. R. Qasim et al. End-to-end multi-particle reconstruction in high occupancy imaging calorimeters with graph neural networks. Eur. Phys. J. C , 82(8):753, 2022

    2022

  3. [11]

    Amrouche et al

    S. Amrouche et al. The Tracking Machine Learning challenge : Accuracy phase. In The NeurIPS ’18 Competition , pages 231–264, 2019

    2019

  4. [12]

    Amrouche et al

    S. Amrouche et al. The Tracking Machine Learning Challenge: Throughput Phase. Comput. Softw. Big Sci. , 7(1):1, 2023

    2023

  5. [13]

    Choma et al

    N. Choma et al. Track Seeding and Labelling with Embedded-space Graph Neural Net- works. 2020. https://arxiv.org/abs/2007. 00149

    2020

  6. [14]

    Caillou et al

    S. Caillou et al. Novel fully-heterogeneous GNN designs for track reconstruction at the HL-LHC. EPJ Web Conf. , 295:09028, 2024

    2024

  7. [15]

    Ju et al

    X. Ju et al. Performance of a geometric deep learning pipeline for HL-LHC particle tracking. Eur. Phys. J. C , 81(10):876, 2021

    2021

  8. [16]

    Lieret et al

    K. Lieret et al. High Pileup Particle Tracking with Object Condensation. 2023. https:// arxiv.org/abs/2312.03823

    2023 arXiv

  9. [17]

    Correia et al

    A. Correia et al. Graph Neural Network- Based Track Finding in the LHCb Vertex Detector. 2024. https://arxiv.org/abs/2407. 12119

    2024

  10. [18]

    Caillou et al

    S. Caillou et al. Physics Performance of the ATLAS GNN4ITk Track Reconstruction Chain. EPJ Web Conf. , 295:03030, 2024

    2024

  11. [19]

    Akram and X

    A. Akram and X. Ju. Track Reconstruction using Geometric Deep Learning in the Straw Tube Tracker (STT) at the PANDA Exper- iment. 2022. https://arxiv.org/abs/2208. 12178

    2022

  12. [20]

    Jia et al

    X. Jia et al. BESIII track reconstruction algo- rithm based on machine learning. EPJ Web Conf., 295:09006, 2024

    2024

  13. [21]

    Kaneko et al

    F. Kaneko et al. Extracting Signal Electron Trajectories in the COMET Phase-I Cylin- drical Drift Chamber Using Deep Learning

  14. [22]

    Lieret and G

    K. Lieret and G. DeZoort. An Object Condensation Pipeline for Charged Particle Tracking at the High Luminosity LHC. EPJ Web Conf. , 295:09004, 2024

    2024

  15. [23]

    Huang et al

    A. Huang et al. A Language Model for Par- ticle Tracking. 2024. https://arxiv.org/abs/ 2402.10239. 33

    2024 arXiv

  16. [24]

    Caron et al

    S. Caron et al. TrackFormers: In Search of Transformer-Based Particle Tracking for the High-Luminosity LHC Era. 2024. https:// arxiv.org/abs/2407.07179

    2024 arXiv

  17. [25]

    T. Keck. FastBDT: A Speed-Optimized Multivariate Classification Algorithm for the Belle II Experiment. Comput. Softw. Big Sci., 1(1):2, 2017

    2017

  18. [26]

    B¨ ahr et al

    S. B¨ ahr et al. The Neural Network First- Level Hardware Track Trigger of the Belle II Experiment. 2024. https://arxiv.org/abs/ 2402.14962

    2024 arXiv

  19. [27]

    A Living Review of Machine Learning for Particle Physics

    HEP ML Community. A Living Review of Machine Learning for Particle Physics. https: //iml-wg.github.io/HEPML-LivingReview/, 2025

    2025

  20. [28]

    Abe et al

    T. Abe et al. Belle II Technical Design Report. 2010. https://arxiv.org/abs/1011. 0352

    2010

  21. [29]

    Bertacchi et al

    V. Bertacchi et al. Track finding at Belle II. Comput. Phys. Commun. , 259:107610, 2021

    2021

  22. [30]

    The Belle II Physics Book

    E Kou et al. The Belle II Physics Book. Prog. Theor. Exp. Phys., 2019:123C01, 2019. [Erra- tum: Prog. Theor. Exp. Phys. 2020, 029201 (2020)]

    2020

  23. [31]

    Agostinelli et al

    S. Agostinelli et al. GEANT4 – a sim- ulation toolkit. Nucl. Instrum. Meth. A , 506:250–303, 2003

    2003

  24. [32]

    Kuhr et al

    T. Kuhr et al. The Belle II Core Software. Computing and Software for Big Science. , 3(1), 2019

    2019

  25. [34]

    Jadach, B

    S. Jadach, B. F. L. Ward, and Z. W¸ as. The precision Monte Carlo event generator KK for two-fermion final states ine+e− collisions. Comput. Phys. Commun. , 130:260, 2000

    2000

  26. [35]

    Alwall et al

    J. Alwall et al. The automated computation of tree-level and next-to-leading order differ- ential cross sections, and their matching to parton shower simulations. J. High Energ. Phys., 07:079, 2014

    2014

  27. [36]

    R. L. Workman et al. Review of Par- ticle Physics. Prog. Theor. Exp. Phys. , 2022:083C01, 2022

    2022

  28. [37]

    Z. J. Liptak et al. Measurements of beam backgrounds in SuperKEKB Phase 2. Nucl. Instrum. Methods Phys. Res. A, 1040:167168, 2022

    2022

  29. [38]

    Hoppner et al

    C. Hoppner et al. A Novel Generic Frame- work for Track Fitting in Complex Detector Systems. Nucl. Instrum. Methods Phys. Res. A, 620:518–525, 2010

    2010

  30. [39]

    Rauch and T

    J. Rauch and T. Schl¨ uter. GENFIT - a Generic Track-Fitting Toolkit.J. Phys. Conf. Ser., 608(1):012042, 2015

    2015

  31. [40]

    Bilka et al

    T. Bilka et al. Implementation of GENFIT2 as an experiment independent track-fitting framework. 2019. https://arxiv.org/abs/ 1902.04405

    2019 arXiv

  32. [41]

    GenFit/GenFit

    Jojosito et al. GenFit/GenFit. https://doi. org/10.5281/zenodo.10301439

    doi

  33. [42]

    Alexopoulos et al

    T. Alexopoulos et al. Implementation of the Legendre Transform for track segment reconstruction in drift tube chambers. Nucl. Instrum. Methods Phys. Res. A, 592:456–462, 2008

    2008

  34. [43]

    Glazov et al

    A. Glazov et al. Filtering tracks in discrete detectors using a cellular automaton. Nucl. Instrum. Methods Phys. Res. A, 329:262–268, 1993

    1993

  35. [44]

    Fey and J

    M. Fey and J. E. Lenssen. Fast Graph Repre- sentation Learning with PyTorch Geometric

  36. [45]

    M. Lin, Q. Chen, and S. Yan. Network In Network. 2014. https://arxiv.org/abs/1312.4400. 34

    2014 arXiv

  37. [46]

    Ioffe and C

    S. Ioffe and C. Szegedy. Batch Normal- ization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. 2015. https://arxiv.org/abs/1502.03167

    2015 arXiv

  38. [47]

    Clevert, T

    D.-A. Clevert, T. Unterthiner, and S. Hochre- iter. Fast and Accurate Deep Network Learn- ing by Exponential Linear Units (ELUs)

  39. [48]

    L. Biewald. Experiment Tracking with Weights and Biases, 2020. Software available from https://wandb.ai/site/

    2020

  40. [49]

    Neu et al

    M. Neu et al. Real-Time Graph Building on FPGAs for Machine Learning Trigger Appli- cations in Particle Physics. Comput. Softw. Big Sci. , 8(1):8, 2024

    2024

  41. [50]

    and others

    Reuter, L. and others. Code for the paper ”End-to-End Multi-Track Reconstruc- tion using Graph Neural Networks at Belle II”. https://doi.org/10.5281/zenodo. 14181605, 2025. 35 Appendix A Data set event displays Typical event displays showing examples of the different training ...

    2025 doi

  42. [2016]

    https://arxiv.org/abs/1511.07289

    arXiv

  43. [2019]

    https://arxiv.org/abs/1903.02428

    1903 arXiv

  44. [2024]

    https://arxiv.org/abs/2408.04795

    arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.