pith. sign in

arxiv: 2605.26280 · v1 · pith:2Y4E6GRDnew · submitted 2026-05-25 · ⚛️ nucl-th

CNN-Based Online Trigger for QGP Event Selection

Olga Soloveva , Artemiy Belousov , Ivan Kisel , Elena Bratkovskaya This is my paper

Pith reviewed 2026-06-29 19:02 UTC · model grok-4.3

classification ⚛️ nucl-th
keywords CNNQGPevent selectiononline triggerPHSDUrQMDparticle histogramsheavy-ion collisions
0
0 comments X

The pith

A convolutional neural network classifies quark-gluon plasma events from reconstructed particle histograms with 83.7 percent accuracy.

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

The paper develops a convolutional-neural-network trigger for selecting events associated with quark-gluon plasma formation in real time from continuous streams of reconstructed events. Events are encoded as compact multidimensional histograms that capture particle species, momentum magnitude, and angular information. The network is first trained and evaluated within the PHSD framework using microscopic QGP-related labels, then validated on UrQMD simulations to check stability across generators. For Au+Au collisions at 30 AGeV the accuracy falls from 95.1 percent on generator-level events to 83.7 percent after full reconstruction, yet retains usable separation power for online selection, with a lightweight C++ package enabling deployment after standard tracking.

Core claim

The authors demonstrate that a CNN operating on multidimensional histograms of reconstructed particles can distinguish events associated with QGP formation, achieving 83.7 percent classification accuracy after full reconstruction in Au+Au collisions at 30 AGeV while remaining stable when the same architecture and representation are applied to UrQMD simulations, and that SHAP analysis can identify the dominant particle-species contributions to the decisions.

What carries the argument

Multidimensional histograms of particle species, momentum magnitude, and angular information, processed by a convolutional neural network for binary classification of QGP-associated events.

If this is right

  • The approach enables real-time filtering of rare QGP signatures under high data-throughput constraints in modern experiments.
  • Performance remains practically useful after full reconstruction and topology analysis.
  • The learned response is stable against generator-dependent effects when cross-checked between PHSD and UrQMD.
  • A lightweight C++ inference package supports deployment directly at the physics-analysis stage.
  • SHAP-based analysis reveals which particle species most strongly influence the network decisions.

Where Pith is reading between the lines

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

  • The method could reduce the volume of data that must be stored by discarding non-QGP events at the trigger level.
  • Similar histogram encodings and CNN architectures might be adapted for selecting other rare processes in nuclear collisions.
  • Direct comparison against actual detector output would be required to quantify any remaining simulation-to-reality gap.
  • The same representation could be tested at different beam energies or in asymmetric collision systems to map the range of applicability.

Load-bearing premise

That microscopic QGP-related labels available inside the PHSD framework constitute reliable ground truth that transfers meaningfully to UrQMD dynamics and to events after detector reconstruction.

What would settle it

Classification accuracy dropping substantially below 80 percent when the trained network is applied to real experimental data from a heavy-ion detector or to events generated by a third independent model.

Figures

Figures reproduced from arXiv: 2605.26280 by Artemiy Belousov, Elena Bratkovskaya, Ivan Kisel, Olga Soloveva.

Figure 1
Figure 1. Figure 1: FIG. 1. The QGP energy fraction from the PHSD as a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Architecture of the 3D CNN model used in this work. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Normalized mean absolute SHAP contribution for [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Distribution of the impact parameter [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Event-by-event comparison of the impact parameter [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Restricted transfer comparison of the CNN-predicted integrated QGP-strength response for PHSD and UrQMD [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Classification accuracy at different stages of data [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
read the original abstract

Modern high-rate experiments require rare physics signatures to be identified in real time from continuous streams of reconstructed events under stringent data-throughput and storage constraints. We present a convolutional-neural-network-based trigger concept for selecting events associated with quark-gluon plasma (QGP) formation. Events are encoded as compact multidimensional histograms of reconstructed particle content, including particle species, momentum magnitude, and angular information. The method is first evaluated within the Parton-Hadron-String Dynamics (PHSD) framework, where microscopic QGP-related labels are available. As an independent validation, the same event representation and network architecture are applied to Ultra-relativistic Quantum Molecular Dynamics (UrQMD) simulations, providing a distinct description of the collision dynamics. Cross-checks between PHSD and UrQMD are used to assess the stability of the learned response against generator-dependent effects and to quantify model-transfer robustness. For realistic deployment, a lightweight C++ inference package, ANN4FLES, is employed at the physics-analysis stage after tracking and topology reconstruction. For Au+Au collisions at 30 AGeV, the classification accuracy decreases from 95.1% on generator-level PHSD events to 83.7% after full reconstruction, while retaining practical separation power for online event selection. SHAP-based interpretability analysis is used to identify the dominant particle-species contributions to the network decision.

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 / 2 minor

Summary. The manuscript proposes a CNN-based online trigger for selecting QGP-associated events in high-rate heavy-ion collisions. Events are encoded as multidimensional histograms of reconstructed particle species, momenta, and angles. The network is trained on PHSD simulations using internal microscopic QGP labels, achieving 95.1% accuracy at generator level for Au+Au collisions at 30 AGeV; the same architecture yields 83.7% accuracy after full detector reconstruction. Cross-validation on UrQMD events and SHAP interpretability analysis are used to assess model-transfer robustness, with deployment via the lightweight ANN4FLES C++ package.

Significance. If the central performance claims hold under an observable definition of QGP, the work would address a practical need for real-time rare-event selection under data-throughput constraints at facilities such as FAIR or the LHC. The compact histogram representation, explicit post-reconstruction testing, cross-generator stability check, and provision of an inference package constitute concrete strengths that could facilitate experimental adoption.

major comments (2)
  1. [Abstract] Abstract: the reported accuracies (95.1% generator-level PHSD, 83.7% post-reconstruction) are defined exclusively with respect to PHSD-internal microscopic QGP labels. Because these labels are not directly observable and encode model-specific parton-hadron dynamics, the UrQMD cross-check only demonstrates inter-generator stability rather than providing an independent, falsifiable ground truth; this directly limits the strength of the 'practical separation power' claim for real data.
  2. [Abstract] Abstract: no statistical uncertainties, error bars, or dataset-size information accompany the accuracy figures, and no baseline comparison to simpler classifiers or conventional trigger algorithms is presented. These omissions make it impossible to judge whether the quoted performance exceeds what could be achieved with conventional observables.
minor comments (2)
  1. The manuscript would benefit from an explicit diagram of the CNN architecture and input histogram binning scheme to allow reproduction.
  2. Training hyperparameters, optimizer choice, and regularization details are not stated, hindering assessment of potential overfitting to generator-specific features.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below and indicate the corresponding revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the reported accuracies (95.1% generator-level PHSD, 83.7% post-reconstruction) are defined exclusively with respect to PHSD-internal microscopic QGP labels. Because these labels are not directly observable and encode model-specific parton-hadron dynamics, the UrQMD cross-check only demonstrates inter-generator stability rather than providing an independent, falsifiable ground truth; this directly limits the strength of the 'practical separation power' claim for real data.

    Authors: We agree that the QGP labels are internal to the PHSD model and not directly observable. The UrQMD cross-validation tests robustness to differences in collision modeling between generators but does not supply an independent experimental ground truth. This does limit the strength of claims about separation power in real data. In the revised manuscript we will update the abstract to state explicitly that accuracies refer to model-specific labels and to moderate the phrasing of the 'practical separation power' claim. revision: yes

  2. Referee: [Abstract] Abstract: no statistical uncertainties, error bars, or dataset-size information accompany the accuracy figures, and no baseline comparison to simpler classifiers or conventional trigger algorithms is presented. These omissions make it impossible to judge whether the quoted performance exceeds what could be achieved with conventional observables.

    Authors: We acknowledge that statistical uncertainties and baseline comparisons are missing. In the revision we will add binomial or bootstrap error estimates to the accuracy numbers together with the sizes of the training and test samples. We will also include a comparison against a simple multiplicity-based trigger to provide context for the CNN performance. revision: yes

Circularity Check

0 steps flagged

No circularity: standard ML train/test on independent generators and reconstruction

full rationale

The paper reports CNN classification accuracies measured on held-out events from PHSD (training generator) and UrQMD (independent validation generator), plus after detector reconstruction. These are empirical performance metrics on separate data splits and models; no derivation chain, equations, or self-citations reduce the reported numbers to fitted parameters or prior results by construction. The use of internal PHSD labels for supervision is a modeling assumption, not a self-referential loop in any derivation.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

Abstract supplies no explicit numerical free parameters beyond standard neural-network weights; the central claim rests on domain assumptions about simulation labels and histogram sufficiency rather than new invented entities.

free parameters (1)
  • CNN weights and biases
    Fitted during supervised training on PHSD-labeled events to produce the reported classification accuracies.
axioms (2)
  • domain assumption Microscopic QGP-related labels inside PHSD constitute reliable supervised ground truth
    Training and label generation rely on this premise as stated in the abstract.
  • domain assumption Multidimensional histograms of particle species, momentum, and angle preserve the information needed to distinguish QGP events
    This encoding choice is the input representation used for both training and inference.

pith-pipeline@v0.9.1-grok · 5779 in / 1558 out tokens · 60135 ms · 2026-06-29T19:02:13.137661+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

46 extracted references · 20 canonical work pages · 6 internal anchors

  1. [1]

    PHSD The Parton–Hadron–String Dynamics (PHSD) ap- proach provides a microscopic off-shell transport de- scription of relativistic heavy-ion collisions based on the Kadanoff–Baym equations in first-order gradient expan- sion [18–22]. It describes the full nonequilibrium evo- lution of the reaction, from the initial hard scatterings and string formation thr...

  2. [2]

    core” is most naturally identified with the deconfined QGP domain, while the “corona

    URQMD To explore the robustness of the CNN approach beyond the PHSD description, we employed the Ultra-relativistic Quantum Molecular Dynamics (UrQMD) model [23–26] in two different variations. UrQMD provides a micro- scopic description of heavy-ion collisions in terms of hadronic and string degrees of freedom and is therefore well suited as an independen...

  3. [3]

    Friman, C

    B. Friman, C. Hohne, J. Knoll, S. Leupold, J. Randrup, R. Rapp, and P. Senger, eds.,The CBM physics book: Compressed baryonic matter in laboratory experiments, Lect. Notes Phys., Vol. 814 (2011) pp. 1–980

    2011

  4. [4]

    Sergeev, E

    F. Sergeev, E. Bratkovskaya, I. Kisel, and I. Vassiliev, Int. J. Mod. Phys. A35, 2043002 (2020)

    2020

  5. [5]

    Belousov, I

    A. Belousov, I. Kisel, and R. Lakos, Algorithms16, 383 (2023)

    2023

  6. [6]

    Kisel, R

    I. Kisel, R. Lakos, and G. Zischka, Algorithms18, 229 (2025)

    2025

  7. [7]

    Accardiet al., Eur

    A. Accardiet al., Eur. Phys. J. A52, 268 (2016)

    2016

  8. [8]

    Science Requirements and Detector Concepts for the Electron-Ion Collider: EIC Yellow Report

    R. Abdul Khaleket al., Nucl. Phys. A1026, 122447 (2022), arXiv:2103.05419 [physics.ins-det]

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  9. [9]

    Pang, Nucl

    L.-G. Pang, Nucl. Phys. A1005, 121972 (2021)

    2021

  10. [10]

    Zheng and J

    S. Zheng and J. Liu, Symmetry16, 1426 (2024)

    2024

  11. [11]

    L.-G. Pang, K. Zhou, N. Su, H. Petersen, H. St¨ ocker, and X.-N. Wang, Nature Commun.9, 210 (2018)

    2018

  12. [12]

    Y.-L. Du, K. Zhou, J. Steinheimer, L.-G. Pang, A. Mo- tornenko, H.-S. Zong, X.-N. Wang, and H. St¨ ocker, Eur. Phys. J. C80, 516 (2020)

    2020

  13. [13]

    Kvasiuk, E

    Y. Kvasiuk, E. Zabrodin, L. Bravina, I. Didur, and M. Frolov, JHEP07, 133

  14. [14]

    Mallick, S

    N. Mallick, S. Prasad, A. N. Mishra, R. Sahoo, and G. G. Barnaf¨ oldi, Phys. Rev. D105, 114022 (2022), arXiv:2203.01246 [hep-ph]

    work page arXiv 2022

  15. [15]

    Boyda, S

    D. Boyda, S. Cal` ı, S. Foreman, L. Funcke, D. C. Hack- ett, Y. Lin, G. Aarts, A. Alexandru, X.-Y. Jin, B. Lu- cini, and P. E. Shanahan, inSnowmass 2021(2022) arXiv:2202.05838 [hep-lat]

    work page arXiv 2021

  16. [16]

    K. Zhou, L. Wang, L.-G. Pang, and S. Shi, Prog. Part. Nucl. Phys.135, 104084 (2024), arXiv:2303.15136 [hep- ph]

    work page arXiv 2024

  17. [17]

    W.-B. He, L. Ma, L.-G. Pang, H.-C. Song, and K. Zhou, Nucl. Sci. Tech.34, 88 (2023), arXiv:2303.06752 [hep-ph]

    work page arXiv 2023

  18. [18]

    Omana Kuttan, J

    M. Omana Kuttan, J. Steinheimer, K. Zhou, A. Redel- bach, and H. Stoecker, Particles4, 47 (2021)

    2021

  19. [19]

    Bleicher and E

    M. Bleicher and E. Bratkovskaya, Prog. Part. Nucl. Phys. 122, 103920 (2022)

    2022

  20. [20]

    Parton transport and hadronization from the dynamical quasiparticle point of view

    W. Cassing and E. L. Bratkovskaya, Phys. Rev. C78, 034919 (2008), arXiv:0808.0022 [hep-ph]

    work page internal anchor Pith review Pith/arXiv arXiv 2008

  21. [21]

    Cassing, Eur

    W. Cassing, Eur. Phys. J. ST168, 3 (2009)

    2009

  22. [22]

    E. L. Bratkovskaya, W. Cassing, V. P. Konchakovski, and O. Linnyk, Nucl. Phys. A856, 162 (2011)

    2011

  23. [23]

    Moreau, O

    P. Moreau, O. Soloveva, L. Oliva, T. Song, W. Cassing, and E. Bratkovskaya, Phys. Rev. C100, 014911 (2019), arXiv:1903.10257 [nucl-th]

    work page arXiv 2019

  24. [24]

    A. W. R. Jorge, T. Song, Q. Zhou, and E. Bratkovskaya, Phys. Rev. C111, 064904 (2025), arXiv:2503.05253 [nucl-th]

    work page arXiv 2025

  25. [25]

    S. A. Basset al., Prog. Part. Nucl. Phys.41, 255 (1998), arXiv:nucl-th/9803035

    work page internal anchor Pith review Pith/arXiv arXiv 1998

  26. [26]

    Bleicheret al., J

    M. Bleicheret al., J. Phys. G25, 1859 (1999), arXiv:hep- 15 ph/9909407

    work page arXiv 1999

  27. [27]

    Core-corona separation in the UrQMD hybrid model

    J. Steinheimer and M. Bleicher, Phys. Rev. C84, 024905 (2011), arXiv:1104.3981 [hep-ph]

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  28. [28]

    Omana Kuttan, A

    M. Omana Kuttan, A. Motornenko, J. Steinheimer, H. Stoecker, Y. Nara, and M. Bleicher, Eur. Phys. J. C82, 427 (2022), arXiv:2201.01622 [nucl-th]

    work page arXiv 2022

  29. [29]

    Linnyk, E

    O. Linnyk, E. L. Bratkovskaya, and W. Cassing, Prog. Part. Nucl. Phys.87, 50 (2016)

    2016

  30. [30]

    Soloveva, A

    O. Soloveva, A. Palermo, and E. Bratkovskaya, Phys. Rev. C110, 034908 (2024)

    2024

  31. [31]

    Y. Xu, P. Moreau, T. Song, M. Nahrgang, S. A. Bass, and E. Bratkovskaya, Phys. Rev. C96, 024902 (2017), arXiv:1703.09178 [nucl-th]

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  32. [32]

    Soloveva, P

    O. Soloveva, P. Moreau, L. Oliva, V. Voronyuk, V. Kireyeu, T. Song, and E. Bratkovskaya, Particles3, 178 (2020), arXiv:2001.05395 [nucl-th]

    work page arXiv 2020

  33. [33]

    Belousov,A QGP trigger based on convolutional neural network for the CBM experiment, Ph.D

    A. Belousov,A QGP trigger based on convolutional neural network for the CBM experiment, Ph.D. thesis, Goethe U., Frankfurt (main) (2025)

    2025

  34. [34]

    Noharaet al., inProceedings of the 10th ACM Inter- national Conference on Bioinformatics, Computational Biology and Health Informatics(2019)

    Y. Noharaet al., inProceedings of the 10th ACM Inter- national Conference on Bioinformatics, Computational Biology and Health Informatics(2019)

    2019

  35. [35]

    Shrikumar, P

    A. Shrikumar, P. Greenside, and A. Kundaje, Pro- ceedings of the 34th International Conference on Ma- chine Learning, PMLR 70:3145-3153, 2017 (2017), arXiv:1704.02685 [cs.LG]

    work page arXiv 2017

  36. [36]

    Experimental and Theoretical Challenges in the Search for the Quark Gluon Plasma: The STAR Collaboration's Critical Assessment of the Evidence from RHIC Collisions

    J. Adamset al.(STAR), Nucl. Phys. A757, 102 (2005), arXiv:nucl-ex/0501009

    work page internal anchor Pith review Pith/arXiv arXiv 2005

  37. [37]

    S. A. Bass, A. Bischoff, C. Hartnack, J. A. Maruhn, J. Reinhardt, H. Stoecker, and W. Greiner, J. Phys. G 20, L21 (1994)

    1994

  38. [38]

    David, M

    C. David, M. Freslier, and J. Aichelin, Phys. Rev. C51, 1453 (1995)

    1995

  39. [39]

    S. A. Bass, A. Bischoff, J. A. Maruhn, H. Stoecker, and W. Greiner, Phys. Rev. C53, 2358 (1996), arXiv:nucl- th/9601024

    work page arXiv 1996

  40. [40]

    Omana Kuttan, J

    M. Omana Kuttan, J. Steinheimer, K. Zhou, A. Redel- bach, and H. Stoecker, Phys. Lett. B811, 135872 (2020), arXiv:2009.01584 [hep-ph]

    work page arXiv 2020

  41. [41]

    F. Li, Y. Wang, H. L¨ u, P. Li, Q. Li, and F. Liu, J. Phys. G47, 115104 (2020), arXiv:2008.11540 [nucl-th]

    work page arXiv 2020

  42. [42]

    F. Li, Y. Wang, Z. Gao, P. Li, H. L¨ u, Q. Li, C. Y. Tsang, and M. B. Tsang, Phys. Rev. C104, 034608 (2021)

    2021

  43. [43]

    Al-Turany and D

    M. Al-Turany and D. Bertini, inComputing in High En- ergy and Nuclear Physics (CHEP 2006), MACMILLAN Advanced Research Series, Vol. 1 (2006) pp. 170–171

    2006

  44. [44]

    Agostinelliet al.(GEANT4), Nucl

    S. Agostinelliet al.(GEANT4), Nucl. Instrum. Meth. A 506, 250 (2003)

    2003

  45. [45]

    Allisonet al., IEEE Trans

    J. Allisonet al., IEEE Trans. Nucl. Sci.53, 270 (2006)

    2006

  46. [46]

    Senger, Particles3, 320 (2020), arXiv:2004.11214 [nucl-ex]

    P. Senger, Particles3, 320 (2020), arXiv:2004.11214 [nucl-ex]

    work page arXiv 2020