Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Machine Learning Power Week 2023: Clustering in Hadronic Calorimeters

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A one-week graduate-student machine-learning competition produced seven clustering algorithms for the ePIC forward hadronic calorimeter that all outperformed a K-means benchmark, with the best reaching a weighted V-score of 0.96692.

desk verdict A transparent workshop report whose headline ranking is undermined by its own metric; read it for the qualitative insights, not for the V-score table. read the letter →

arxiv 2508.09938 v1 pith:P33XNXQA submitted 2025-08-13 nucl-ex hep-ex

Muaz Al Halabi , Marcel Bajdel , Jeroen Peter Bormans , Hannah Bossi (ed) , Maria Calmon Behling , Florian Ehmann , Niklas Gotz , Jerome Jung
show 14 more authors
Rafet Kavak Mario Krüger Robin Lakos Annemarie Lauterbach Patrick Mccormack (ed) Akhil Mithran Daniel Murnane (ed) Mathis Nolte Tim Rogoschinski Jan Scharf Oddharak Tyagi Liv V{aa}ge (ed) Maxim Valialshchikov Stephan Wagner
This is my paper
classification nucl-exhep-ex
keywords machinelearningcrowdsourcingcalorimeterclusteringK-meansV-scoreePICElectron-IonColliderdetectorR&D
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

This paper aims to show that crowd-sourced machine-learning competitions can contribute useful reconstruction algorithms to future particle-physics facilities whose detector simulations can be released publicly. The specific test case was a one-week graduate-student competition to cluster simulated energy deposits in the forward hadronic calorimeter of the ePIC detector at the planned Electron-Ion Collider. The paper reports that all seven teams beat the provided K-means benchmark on a weighted V-score, an energy-weighted harmonic mean of cluster purity and completeness. The authors argue that this demonstrates both that students can learn machine learning in this format and that such events can give early feedback to detector design. The paper also notes that the contest's merged-event scoring metric produced artificially high scores and recommends changing the scoring in future challenges.

What carries the argument

The weighted V-score is the single number that ranks all submissions: $V = 2 \cdot \mathrm{hom} \cdot \mathrm{com} / (\mathrm{hom} + \mathrm{com})$, where homogeneity measures how much a cluster contains hits from only one particle, completeness measures how much a particle's hits are captured in one cluster, and both are weighted by hit energy. The algorithmic workhorse is K-means with a per-event estimate of the number of clusters $k$: most submissions predict $k$ from event-level features such as hit count and total energy, often via linear or quadratic regression, then run K-means in spherical coordinates $\phi,\theta$ rather than Cartesian. The graph-based submission replaces that estim

What would settle it

Recompute the weighted V-score separately for each test event and average it over events, using the same test set. The paper already reports that one algorithm's merged-event score was 0.965 while its average over 100-event subsets was 0.433; if under per-event scoring any of the seven submissions falls below the benchmark's per-event score, the paper's headline claim that every team beat the benchmark is false for that metric.

Watch

Extended reading notes

Core claim

This paper reports the results of a machine-learning "Power Week" held in July 2023, in which seven teams of graduate students developed clustering algorithms for the longitudinally segmented forward hadronic calorimeter (LFHCal) of the ePIC detector at the planned Electron-Ion Collider. The algorithms group simulated energy deposits into clusters corresponding to individual particles, a task made difficult by unknown particle multiplicity, dense hits near the beamline, and overlapping showers. The paper's central claim is that the week demonstrated crowd-sourced development works for future facilities: all seven teams beat the K-means benchmark on the competition's weighted V-score, and the

Load-bearing premise

The comparison of all seven submissions rests on a single weighted V-score computed by merging all test events into one clustering problem, a method the paper itself says gives artificially high scores; if that metric does not reflect per-event clustering quality, the ranking and the claim that every team beat the benchmark do not follow.

Editorial extensions

If this is right

  • If crowd-sourced competitions prove effective for ePIC, future facilities can release detector simulations publicly and use external groups to help develop reconstruction algorithms before the detector design is finalised.
  • All algorithms struggle most near the beamline; 97% of cases where every solution was wrong involved hits within ±80 cm in the transverse plane, so improving that region is the clearest path to better clustering.
  • K-means with a good per-event $k$ estimate is a strong baseline for calorimeter clustering: it outperformed autoencoders, neural-regression variants, and an anti-$k_T$ approach within the week-long competition.
  • The graph-based solution's edge classifier reached an AUC of 0.998, and the authors expect that using more training data, a GNN or transformer edge model, and a better seed-selection scheme would improve it further.
  • The merged-event scoring method used in the competition should be replaced in future challenges; the paper reports that one algorithm scored 0.965 on 5000 merged events but only 0.433 when averaged over 100-event subsets.
  • A single accuracy number is not enough to evaluate a clustering algorithm for downstream physics; the authors recommend benchmarking against task-specific metrics such as reconstructed-energy resolution in addition to the V-score.

Reading between the lines

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

  • One consequence not stated directly: because the Table 1 ranking uses merged-event V-scores, the ordering of the seven submissions—and even the claim that every team beat the benchmark—should be treated as provisional until per-event average V-scores and a downstream metric are reported for each submission.
  • The paper's failure map suggests a natural testable extension that the authors only hint at: a two-stage clustering algorithm that first handles the dense beamline region separately and then clusters the rest of the detector should recover most of the roughly 10% accuracy lost there.
  • The dominance of K-means with a per-event $k$ estimate suggests that, for this dataset, the dominant source of clustering error is uncertainty in the number of clusters rather than in the geometric assignment; separating $k$-estimation error from assignment error would provide a sharper benchmark for future challenges.
  • The graph-based method's strong edge-classifier performance in just one week, with only 20,000 training events and an MLP, implies that supervised graph approaches could overtake K-means once trained on the full 100,000-event sample with a more expressive edge model.
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

4 major / 5 minor

Summary. The paper reports on an 'ML Power Week' workshop held in July 2023, in which seven graduate-student teams competed to build a clustering algorithm for the ePIC forward hadronic calorimeter at the future Electron-Ion Collider. Using a public Kaggle-style competition, participants developed solutions ranging from anti-kT jet clustering and K-means variants with learned cluster-number regressions to a graph-based seed-to-nonseed classifier. The paper presents the competition setup, describes each approach, and ranks submissions by a weighted V-score. Its central quantitative claim is that all seven teams beat the K-means benchmark, with the best score 0.96692 versus the benchmark 0.95207. The paper also draws qualitative conclusions about where algorithms fail (near the beamline) and offers feedback for future crowd-sourcing in HEP.

Significance. If the performance ranking were robust, this would be a useful demonstration that a one-week, student-led crowd-sourcing effort can generate competitive clustering solutions for a future detector, and the paper makes a fair contribution in documenting the format, dataset, and lessons learned. Strengths include the public release of the dataset and example code, the explicit discussion of metric limitations in Section 5.2, and the inclusion of several independent solution families. However, the central quantitative claim is currently not established: the Table 1 scores are computed on merged events, a protocol the authors themselves state gives artificially high values, and no uncertainties are provided on the very small score differences. As a workshop-report, the paper is readable and honest, but as a quantitative benchmark comparison it needs revision.

major comments (4)
  1. [Table 1 and Section 5.2] The headline claim that all teams beat the benchmark rests on weighted V-scores computed after merging test events. Section 5.2 states that this gives an artificially high score: one algorithm scores 0.965 across 5000 merged events but only 0.433 when averaged over 100-event subsets. Because the inflation is likely algorithm-dependent (cluster counts and assignments grow with event multiplicity), the ordering in Table 1 and even the benchmark comparison are not established. The gaps between submissions are tiny (A vs B differ by 0.00004) and no uncertainties are reported. Please re-score on a per-event basis with bootstrapped uncertainties and reframe the central claim accordingly.
  2. [Section 4.2 and Table 1] The benchmark K-means uses a fixed k=40, chosen explicitly 'so the students could easily beat the benchmark'. This makes the 'all teams beat benchmark' result weak evidence of genuine improvement. The comparison should include a stronger baseline, for example K-means with an adaptive cluster-number estimator matching the event multiplicity, or at least a clear caveat in the text.
  3. [Section 5.1 and Figure 12] The NMI comparison shows that the top solutions differ mostly in minor ways and that 'the choice of scoring metric leaves ambiguity'. This effectively concedes that the Table 1 ranking is not unique. To support the paper's conclusions, the authors should quantify how the ranking changes under per-event scoring and under different reasonable metrics (e.g., homogeneity or completeness alone, or an energy-weighted variant).
  4. [Section 4.4 and Table 2] The graph-based solution is described in Section 4.4 as an MLP edge classifier with KNN seed-to-nonseed connections, but Table 2 labels it 'Graph neural net'. This inconsistency obscures what was actually implemented. Please harmonize the description and either implement a GNN or call the model a graph-based MLP.
minor comments (5)
  1. [Section 2.2 and 2.3] The text says 100,000 simulated events are split into training and test samples, but Section 2.3 states only 50,000 training and 10,000 test events were used. Clarify whether the remaining 40,000 events were unused or held back.
  2. [Section 2.4, Equations (1)-(3)] The marginal probabilities p_i and p_j in Eq. (3) are not explicitly defined before use. Define them as cluster and class size fractions for clarity.
  3. [Section 5.1] The sentence 'The lowest score is solution E, but the lowest scoring solution in the Kaggle competition was G' is confusing because the discussion switches from NMI to V-score. Please specify which score each clause refers to.
  4. [Table 2] The team name 'Genuine Stupididy' appears to be a typo for 'Genuine Stupidity'.
  5. [Section 4.1.1] 'pseudo jets' should be 'pseudo-jets' for consistency with standard terminology.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the paper reports a competition, and its self-citations and metric caveats do not make a load-bearing result circular.

full rationale

This paper does not claim a derivation from first principles; it reports the setup and outcomes of a one-week Kaggle-style clustering challenge. The main quantitative claim, that all seven teams beat the K-means benchmark (Table 1), rests on the weighted V-score. Section 5.2 explicitly warns that the merged-event scoring protocol gives artificially high scores (0.965 over 5000 merged events vs. 0.433 averaged over 100-event subsets), but this is a metric-validity and ranking-reliability concern, not a circular reduction: the benchmark and submissions were evaluated under the same protocol, and the authors do not present the scores as an independently verified prediction. The benchmark's k=40 was chosen 'so the students could easily beat the benchmark' (Section 4.2), which limits the strength of the claim but does not make it circular. Hyperparameters such as anti-kT R=0.4, cluster-number regressions, the GNN seed count, and the edge threshold were tuned on the evaluation metric; this is standard competition practice and is not disguised as an independent derivation. Reference [26] is a code repository authored by the organizers, but it is cited as a resource, not as a load-bearing proof or uniqueness theorem. No equation is shown to equal its own input, and no fitted parameter is renamed as a prediction. The paper is self-contained as a report with acknowledged limitations, so no significant circularity is present.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central performance comparison rests on the V-score metric, the particle-gun simulation, and ground-truth particle IDs. The paper states these assumptions and notes the metric inflates scores when merging events.

free parameters (6)
  • anti-kT jet radius R = 0.4
    Chosen by Group G to maximize V-score on the training data (Figure 5).
  • Benchmark number of clusters N = 40
    Fixed k-means cluster count for the benchmark, chosen so students could easily beat it.
  • Graph model seed count Ns = 40
    Number of highest-energy hits used as seeds in Group D's graph method.
  • Edge classifier threshold = 0.6
    Threshold on MLP edge scores in Group D, chosen to discard low-scoring edges.
  • Cluster-number regression coefficients = various
    Teams A, B, C, E and F fit regressions (linear or quadratic) to predict the number of clusters from event features; these coefficients are fitted on the training set.
  • Weighted V-score energy weighting = linear in energy
    Design choice by the organizers to weight hits by deposited energy; it changes the metric but is not fitted.
assumptions (4)
  • domain assumption The particle-gun simulation is representative of ePIC forward calorimeter data
    The dataset is generated without the electromagnetic calorimeter in front and with a simplified particle gun; real data will differ.
  • domain assumption Ground-truth particle IDs from the simulation are correct
    The V-score evaluation uses the true cluster IDs provided by the simulation; any mislabeling in the generator would propagate into the scores.
  • domain assumption The V-score (homogeneity and completeness) is an appropriate clustering quality measure
    The choice of metric is not derived from a physics requirement; the paper acknowledges it is physics-agnostic and may not map to downstream performance.
  • domain assumption K-means with Euclidean distance on spherical coordinates is a sensible baseline
    The paper uses K-means as the benchmark and most final solutions; it does not rigorously justify why this geometry is optimal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine Learning Power Week 2023: Clustering in Hadronic Calorimeters." pith.science (2026). https://pith.science/paper/P33XNXQA

@misc{pith2026250809938,
  author       = {Pith},
  title        = {Pith review of: Machine Learning Power Week 2023: Clustering in Hadronic Calorimeters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P33XNXQA}},
  note         = {Machine review of arXiv:2508.09938}
}
read the original abstract

In both high-energy physics and industry applications, a crowd-sourced approach to difficult problems is becoming increasingly common. These innovative approaches are ideal for the development of future facilities where the simulations can be publicly distributed, such as the Electron-Ion Collider (EIC). In this paper, we discuss a so-called ``Power Week" where graduate students were able to learn about machine learning while also contributing to an unsolved problem at a future facility. Here, the problem of interest was the clustering of the forward hadronic calorimeter in the foreseen electron-proton/ion collider experiment (ePIC) detector at the EIC. The different possible approaches, developed over the course of a single week, and their performance are detailed and summarised. Feedback on the format of the week and recommendations for future similar programs are provided in the hopes to inspire future learning opportunities for students that also serve as a crowd-sourced approaches to unsolved problems.

Figures

Figures reproduced from arXiv: 2508.09938 by the authors.

Figure 1
Figure 1. 3D rendering of the design of the LFHCal (left) and a zoomed-in rendering of a single 8M module [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Distribution of the logarithm of the energy distribution for all hits in the training sample (left) as [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Distribution of hits for various clusters in the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Example of clustering dijets in the ATLAS detector. Highly energetic cells act as seeds for clusters, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Performance metrics (left) and relative energy resolution (right) as a function of ’jet’ radius [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Illustration of the cluster assignment for a specific event with a radius of [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Distribution of the cluster number in a subset of the training data. The number of hits and energy [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Number of clusters versus the total energy in an event including fits determined by linear regression [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: (Left) Hits and their corresponding cluster identifications for an example event passed as input to [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Example graphs constructed on showering events. (Left) The outcome of the first stage, where [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: (Left) The ratio of hits for which n solutions got the correct solution aggregated over 1000 events [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Heatmap of the normalised mutual information score. [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 23 canonical work pages

  1. [1]

    The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics

    G. Kasieczka et al., “The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics”, Rept. Prog. Phys. 84 (2021), no. 12, 124201, doi:10.1088/1361-6633/ac36b9, arXiv:2101.08320

    arXiv 2020

  2. [2]

    Icecube - neutrinos in deep ice

    A. C. L. H. P. E. R. Ørsøe Sohier Dane, “Icecube - neutrinos in deep ice”, 2023. https://kaggle.com/competitions/icecube-neutrinos-in-deep-ice

    work page 2023

  3. [3]

    CaloChallenge 2022: A Community Challenge for Fast Calorimeter Simulation

    O. Amram et al., “CaloChallenge 2022: A Community Challenge for Fast Calorimeter Simulation”, arXiv:2410.21611

    arXiv 2022

  4. [4]

    F AIR Universe HiggsML Uncertainty Challenge Competition

    W. Bhimji et al., “F AIR Universe HiggsML Uncertainty Challenge Competition”, arXiv:2410.02867

    arXiv

  5. [5]

    The Dark Machines Anomaly Score Challenge: Benchmark Data and Model Independent Event Classification for the Large Hadron Collider

    T. Aarrestad et al., “The Dark Machines Anomaly Score Challenge: Benchmark Data and Model Independent Event Classification for the Large Hadron Collider”, SciPost Phys. 12 (2022), no. 1, 043, doi:10.21468/SciPostPhys.12.1.043, arXiv:2105.14027

    arXiv 2022

  6. [6]

    Electron Ion Collider: The Next QCD Frontier: Understanding the glue that binds us all

    A. Accardi et al., “Electron Ion Collider: The Next QCD Frontier: Understanding the glue that binds us all”, Eur. Phys. J. A 52 (2016), no. 9, 268, doi:10.1140/epja/i2016-16268-9, arXiv:1212.1701

    arXiv 2016

  7. [7]

    Snowmass 2021 White Paper: Electron Ion Collider for High Energy Physics

    R. Abdul Khalek et al., “Snowmass 2021 White Paper: Electron Ion Collider for High Energy Physics”, arXiv:2203.13199

    arXiv 2021

  8. [8]

    Artificial Intelligence for the Electron Ion Collider (AI4EIC)

    C. Allaire et al., “Artificial Intelligence for the Electron Ion Collider (AI4EIC)”, Comput. Softw. Big Sci. 8 (2024) 5, doi:10.1007/s41781-024-00113-4 , arXiv:2307.08593

    arXiv 2024

Show all 45 references
  1. [9]

    Hgs-hire power week - an epic competition

    D. M. H. B. L. V. P. McCormack, “Hgs-hire power week - an epic competition”, 2023. https://kaggle.com/competitions/hgs-hire-power-week-an-epic-competition

    2023

  2. [10]

    Scikit-learn: Machine Learning in Python

    F. Pedregosa et al., “Scikit-learn: Machine Learning in Python ”, Journal of Machine Learning Research 12 (2011) 2825–2830

    2011

  3. [11]

    Particle-flow reconstruction and global event description with the CMS detector

    CMS Collaboration, “Particle-flow reconstruction and global event description with the CMS detector”, JINST 12 (2017), no. 10, P10003, doi:10.1088/1748-0221/12/10/P10003, arXiv:1706.04965

    2017 arXiv

  4. [12]

    Fuzzy V-measure - an evaluation method for cluster analyses of ambiguous data

    J. Utt, S. Springorum, M. K¨ oper, and S. Schulte im Walde, “Fuzzy V-measure - an evaluation method for cluster analyses of ambiguous data”, in Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC‘14) , N. Calzolari et al., eds., pp. 581...

    2014

  5. [13]

    V-measure: A conditional entropy-based external cluster evaluation measure

    A. Rosenberg and J. Hirschberg, “V-measure: A conditional entropy-based external cluster evaluation measure”, in Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL) , J. Eisner, ...

    2007

  6. [14]

    The CMS Experiment at the CERN LHC

    CMS Collaboration, “The CMS Experiment at the CERN LHC”, JINST 3 (2008) S08004, doi:10.1088/1748-0221/3/08/S08004

    2008 doi

  7. [15]

    The ATLAS experiment at the CERN Large Hadron Collider

    ATLAS Collaboration, “The ATLAS experiment at the CERN Large Hadron Collider”, Journal of Instrumentation 3 (aug, 2008) S08003, doi:10.1088/1748-0221/3/08/S08003

    2008 doi

  8. [16]

    Particle-flow reconstruction and global event description with the CMS detector

    CMS Collaboration, “Particle-flow reconstruction and global event description with the CMS detector”, Journal of Instrumentation 12 (oct, 2017) P10003–P10003, doi:10.1088/1748-0221/12/10/p10003

    2017 doi

  9. [17]

    Jet reconstruction and performance using particle flow with the ATLAS detector

    ATLAS Collaboration, “Jet reconstruction and performance using particle flow with the ATLAS detector”, The European Physical Journal C 77 (jul, 2017) doi:10.1140/epjc/s10052-017-5031-2

    2017 doi

  10. [18]

    Topological cell clustering in the ATLAS calorimeters and its performance in LHC Run 1

    ATLAS Collaboration, “Topological cell clustering in the ATLAS calorimeters and its performance in LHC Run 1”, Eur. Phys. J. C 77 (2017) 490, doi:10.1140/epjc/s10052-017-5004-5 , arXiv:1603.02934. 17

    2017 arXiv

  11. [19]

    Calorimeter Clustering Algorithms: Description and Performance

    W. Lamplet al., “Calorimeter Clustering Algorithms: Description and Performance”, technical report, CERN, Geneva, 2008. All figures including auxiliary figures are available at https://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PUBNOTES/ATL-LARG-PUB-2008-002

    2008

  12. [20]

    The Iterative Clustering framework for the CMS HGCAL Reconstruction

    CMS Collaboration, “The Iterative Clustering framework for the CMS HGCAL Reconstruction”, J. Phys. Conf. Ser. 2438 (2023), no. 1, 012096, doi:10.1088/1742-6596/2438/1/012096

    2023 doi

  13. [21]

    Performance of the ALICE Electromagnetic Calorimeter

    ALICE Collaboration, “Performance of the ALICE Electromagnetic Calorimeter”, JINST 18 (2023), no. 08, P08007, doi:10.1088/1748-0221/18/08/P08007, arXiv:2209.04216

    2023 arXiv

  14. [22]

    Graph Clustering: a graph-based clustering algorithm for the electromagnetic calorimeter in LHCb

    N. V. Canudas, M. Calvo G´ omez, X. Vilas ´ ıs-Cardona, and E. G. Rib´ e, “Graph Clustering: a graph-based clustering algorithm for the electromagnetic calorimeter in LHCb”, Eur. Phys. J. C 83 (2023), no. 2, 179, doi:10.1140/epjc/s10052-023-11332-1 , arXiv:2212.11061

    2023 arXiv

  15. [23]

    The Phase-2 Upgrade of the CMS Endcap Calorimeter

    CMS Collaboration, “The Phase-2 Upgrade of the CMS Endcap Calorimeter”, technical report, CERN, Geneva, 2017. doi:10.17181/CERN.IV8M.1JY2

    2017 doi

  16. [24]

    Object condensation: one-stage grid-free multi-object reconstruction in physics detectors, graph, and image data

    J. Kieseler, “Object condensation: one-stage grid-free multi-object reconstruction in physics detectors, graph, and image data”, The European Physical Journal C 80 (sep, 2020) doi:10.1140/epjc/s10052-020-08461-2

    2020 doi

  17. [25]

    GNN-based end-to-end reconstruction in the CMS phase 2 high-granularity calorimeter

    S. Bhattacharya et al., “GNN-based end-to-end reconstruction in the CMS phase 2 high-granularity calorimeter”, Journal of Physics: Conference Series 2438 (feb, 2023) 012090, doi:10.1088/1742-6596/2438/1/012090

    2023 doi

  18. [26]

    ePIC Clustering 2023

    P. McCormack, D. Murnane, H. Bossi, and L. Vage, “ePIC Clustering 2023”. https://github.com/wpmccormack/ePIC_Clustering_2023/tree/master, 2023

    2023

  19. [27]

    Performance of the electromagnetic pixel calorimeter prototype epical-2

    J. Alme et al., “Performance of the electromagnetic pixel calorimeter prototype epical-2”, Journal of Instrumentation 18 (01, 2023) P01038, doi:10.1088/1748-0221/18/01/P01038

    2023 doi

  20. [28]

    The anti- kt jet clustering algorithm

    M. Cacciari, G. P. Salam, and G. Soyez, “The anti- kt jet clustering algorithm”, JHEP 04 (2008) 063, doi:10.1088/1126-6708/2008/04/063, arXiv:0802.1189

    2008 arXiv

  21. [29]

    FastJet user manual

    M. Cacciari, G. P. Salam, and G. Soyez, “FastJet user manual”, The European Physical Journal C 72 (mar, 2012) doi:10.1140/epjc/s10052-012-1896-2

    2012 doi

  22. [30]

    Applied Predictive Modeling

    M. Kuhn and K. Johnson, “Applied Predictive Modeling”. Springer, 2013. ISBN 978-1-4614-6848-6

    2013

  23. [31]

    Deep Learning

    I. Goodfellow, Y. Bengio, and A. Courville, “Deep Learning”. MIT Press, 2016. http://www.deeplearningbook.org

    2016

  24. [32]

    A density-based algorithm for discovering clusters in large spatial databases with noise

    M. Ester et al., “A density-based algorithm for discovering clusters in large spatial databases with noise”, in Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, KDD’96, p. 226–231. AAAI Press, 1996

    1996

  25. [33]

    Ensemble methods in machine learning

    T. G. Dietterich, “Ensemble methods in machine learning”, in Multiple Classifier Systems , pp. 1–15. Springer Berlin Heidelberg, Berlin, Heidelberg, 2000

    2000

  26. [34]

    Inertia-based indices to determine the number of clusters in k-means: An experimental evaluation

    A. Rykov, R. C. De Amorim, V. Makarenkov, and B. Mirkin, “Inertia-based indices to determine the number of clusters in k-means: An experimental evaluation”, IEEE Access 12 (2024) 11761–11773, doi:10.1109/ACCESS.2024.3350791

    2024

  27. [35]

    Graph Neural Networks for Particle Reconstruction in High Energy Physics detectors

    X. Juet al., “Graph Neural Networks for Particle Reconstruction in High Energy Physics detectors”, arXiv:2003.11603 [hep-ex, physics:physics] (June, 2020). arXiv: 2003.11603

    2003 arXiv

  28. [36]

    ATLAS ITk Track Reconstruction with a GNN-based pipeline

    ATLAS Collaboration, “ATLAS ITk Track Reconstruction with a GNN-based pipeline”, technical report, CERN, Geneva, 2022

    2022

  29. [37]

    Graph Neural Network Jet Flavour Tagging with the ATLAS Detector

    ATLAS, “Graph Neural Network Jet Flavour Tagging with the ATLAS Detector”, technical report, CERN, Geneva, 2022. All figures including auxiliary figures are available at https://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PUBNOTES/ATL-PHYS-PUB-2022-027. 18

    2022

  30. [38]

    Flavour tagging with graph neural networks with the ATLAS detector

    ATLAS Collaboration, A. Duperrin, “Flavour tagging with graph neural networks with the ATLAS detector”, in 30th International Workshop on Deep-Inelastic Scattering and Related Subjects . 6, 2023. arXiv:2306.04415

    2023 arXiv

  31. [39]

    Graph neural networks for low-energy event classification & reconstruction in IceCube

    R. A. et al., “Graph neural networks for low-energy event classification & reconstruction in IceCube”, Journal of Instrumentation 17 (nov, 2022) P11003, doi:10.1088/1748-0221/17/11/p11003

    2022 doi

  32. [40]

    GNN-based end-to-end reconstruction in the CMS Phase 2 High-Granularity Calorimeter

    CMS Collaboration, “GNN-based end-to-end reconstruction in the CMS Phase 2 High-Granularity Calorimeter”, Technical Report 1, CERN, Geneva, 2023. doi:10.1088/1742-6596/2438/1/012090, arXiv:2203.01189, 5 pages, 5 figures, proceedings for the 20th International Workshop on Advan...

    2023 arXiv

  33. [41]

    Graph Neural Networks in Particle Physics: Implementations, Innovations, and Challenges

    S. Thais et al., “Graph Neural Networks in Particle Physics: Implementations, Innovations, and Challenges”, in Snowmass 2021 . 3, 2022. arXiv:2203.12852

    2021 arXiv

  34. [42]

    Graph neural networks in particle physics

    J. Shlomi, P. Battaglia, and J.-R. Vlimant, “Graph neural networks in particle physics”, Machine Learning: Science and Technology 2 (jan, 2021) 021001, doi:10.1088/2632-2153/abbf9a

    2021 doi

  35. [43]

    An improved algorithm for finding the strongly connected components of a directed graph

    D. J. Pearce, “An improved algorithm for finding the strongly connected components of a directed graph”, in Technical report, Victoria University, NZ . 2005

    2005

  36. [44]

    SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python

    P. Virtanen et al., “SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python”, Nature Methods 17 (2020) 261–272, doi:10.1038/s41592-019-0686-2

    2020 doi

  37. [45]

    Information theoretic measures for clusterings comparison: Variants

    N. Vinh, J. Epps, and J. Bailey, “Information theoretic measures for clusterings comparison: Variants”, Properties, Normalization and Correction for Chance 18 (2009). 19 Name Algorithm, N - finding Method Inputs to find N Coordinate System Kmeans Inputs Init. Sample W eight En...

    2009

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.