Pith. sign in

REVIEW 3 major objections 4 minor 3 cited by

A graph neural network has been deployed in the Belle II calorimeter trigger readout, running at 8 MHz; it improves cluster positions and rejects background, though its latency is still three times too high for the actual trigger decision.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 22:56 UTC pith:KPLCEF74

load-bearing objection Genuine engineering milestone: a GNN running synchronously at 8 MHz inside a collider trigger readout chain, with bitwise-validated firmware and collision-data results; the honest caveat is that it is not yet in the L1 decision path, and that framing needs careful wording. the 3 major comments →

arxiv 2602.15118 v2 pith:KPLCEF74 submitted 2026-02-16 physics.ins-det hep-ex

Real-time graph neural networks on FPGAs for the Belle II electromagnetic calorimeter

I. Haide , M. Neu , Y. Unno , T. Justinger , V. Dajaku , F. Baptist , T. Lobmaier , J. Becker
show 14 more authors
T. Ferber H. Bae A. Beaubien J. Eppelt R. Giordano G. Heine T. Koga Y.-T. Lai K. Miyabayashi H. Nakazawa M. Remnev L. Reuter K. Unger R. van Tonder
This is my paper
classification physics.ins-det hep-ex PACS 29.40.Vj07.05.Mh
keywords graph neural networksFPGA triggercalorimeter clusteringBelle IIreal-time reconstructionobject condensationL1 triggerelectromagnetic calorimeter
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper reports the first live deployment of a graph neural network inside the trigger readout path of a collider experiment. The module, GNN-ETM, treats every calorimeter trigger cell as a graph node and, in a single pass, clusters the deposits, predicts cluster positions and energies, and scores each cluster as signal or background. On an FPGA it sustains the full Belle II rate of 8 million trigger windows per second with a deterministic 3.168 μs end-to-end latency and 100% uptime. Compared with the existing trigger logic, it improves high-energy cluster position resolution in the barrel by up to 18%, low-energy cluster purity by up to 20%, and overlapping two-photon cluster efficiency by up to 20%, while keeping energy resolution comparable and rejecting up to 70% of background clusters at 97.5% signal retention. The authors are explicit that this latency exceeds the L1 decision budget by about a factor of three, so the module currently runs in parallel with the trigger decision rather than inside it; they present the result as a first step toward GNN-based real-time triggering.

Core claim

Central claim: GNN-ETM is the first GNN-based reconstruction system to run inside the real-time trigger readout path of a collider experiment. Its network, CaloClusterNet, reads up to 32 trigger cells per event and in one pass clusters them, predicts energies and positions, and scores signal vs background. Reported gains: up to 18% better barrel position resolution for high-energy clusters, up to 20% better low-energy purity, up to 20% better efficiency for overlapping two-photon clusters, and up to 70% background rejection at 97.5% signal retention. The FPGA sustains 8 MHz at a deterministic 3.168 μs latency; the authors state this exceeds the L1 decision budget by about 3x, so the module r

What carries the argument

The central object is CaloClusterNet, a compact graph neural network that works on up to 32 trigger cells per event. Each cell is a node with position, energy, and timing features; two distance-weighted message-passing blocks build edges dynamically among the k nearest neighbours in a learned space, and a one-stage 'object condensation' head predicts, for each node, a latent cluster center, a beta value, an energy scale, a position, and a signal score. The condensation-point selection algorithm picks isolated high-beta centers in latent space and reads their predicted properties directly, so the number of clusters does not need to be known in advance. On the FPGA, the computation is mapped t

Load-bearing premise

The claim that GNN-ETM counts as a real-time GNN trigger rests on the assumption that a parallel, non-decision readout path is a valid stand-in for an active trigger, and that the planned latency reductions will bring the 3.168 μs latency inside the L1 decision budget without losing the demonstrated physics performance; neither is shown in this paper.

What would settle it

Route GNN-ETM's output into the actual Global Decision Logic and measure the end-to-end trigger decision latency at 8 MHz; if the module still exceeds the Belle II L1 decision budget (roughly one microsecond for this pipeline stage) or the physics gains vanish, the claim of a GNN operating in the real-time trigger path is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the latency budget can be recovered, Belle II's calorimeter trigger would gain up to 18% better barrel position resolution for high-energy clusters and up to 20% better cluster purity at low energies, directly improving photon-based trigger lines.
  • The signal classifier offers a way to reduce beam-background-dominated trigger rates without cutting low-energy signal: at 97.5% signal retention it rejects up to 70% of background clusters, which matters for dark-sector and axion-like-particle searches.
  • For overlapping two-photon signatures (opening angles of roughly 3–11 degrees), cluster-finding efficiency improves by up to 20 percentage points, so triggers that count close photon pairs become more efficient.
  • The demonstrated 8 MHz throughput with deterministic latency and 100% uptime shows that a GNN can run synchronously inside a collider readout chain, not merely on a bench test.
  • If the planned latency reductions—doubling the clock, removing one message-passing block, and bypassing the existing trigger module's routing—are realized, the module could move from parallel monitoring to active trigger participation.

Where Pith is reading between the lines

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

  • The authors leave implicit that the same single-pass design could replace the baseline's fixed isolated-cluster logic entirely, since object condensation avoids the duplicate-hit ambiguity and the region-gap problem without a separate clustering step.
  • A natural testable extension is to make the signal classifier thresholds adapt continuously to beam-background level (e.g., out-of-time crystal counts); the paper's rate-vs-background curves already show the rate slope flattening when the classifier is applied.
  • The 32-trigger-cell input cap was validated for December 2024 background conditions; at higher SuperKEKB luminosities the fraction of truncated events could grow, so future runs should re-check the occupancy distribution before relying on the cap.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper presents GNN-ETM, an FPGA implementation of a graph neural network (GravNet + object condensation) that performs calorimeter trigger-cell clustering, cluster parameter regression, and signal/background classification for the Belle II ECL. The system is deployed on a UT4 board with a custom preprocessing stage, a GNN dataflow accelerator, postprocessing, and a Belle2Link readout subsystem. It sustains 8 MHz throughput with an end-to-end latency of 3.168 μs, and the authors present a three-level validation chain (QKERAS, C-simulation, RTL simulation) with bitwise agreement between C-simulation and hardware on cosmic-ray data. Physics performance is evaluated on simulation and December 2024 collision data, reporting improved angular resolution, purity at low energy, overlap efficiency, and signal/background separation. The authors explicitly state in Sec. 6.3 and Sec. 8 that the current latency exceeds the L1 trigger decision budget and that the GDL interface is unused; the system is operated synchronously as a parallel readout path.

Significance. If the results hold, the paper makes an important contribution: it is, to my knowledge, the first GNN-based reconstruction system deployed on FPGAs and operated synchronously in a collider's real-time readout path. The engineering methodology—mixed-precision quantization-aware training, HLS/Chisel co-design, three-level bitwise simulation, and deployment on real hardware—is rigorous and reproducible, with code links provided. The physics improvements are plausible and supported by a large set of simulation and collision-data studies. However, the significance is currently overstated: because the system does not meet the L1 latency budget and its GDL connection is unused, it is a synchronous readout-path monitor rather than an active trigger participant. The paper's central 'trigger module' claim therefore needs reframing or additional feasibility evidence.

major comments (3)
  1. [Sec. 6.3, Fig. 11, Sec. 8] The end-to-end latency of 3.168 μs is reported to exceed the maximum allowed for an active L1 trigger decision by approximately a factor of three, and the interface to the Global Decision Logic 'exists but is currently unused' (Fig. 11 caption). Section 8 reiterates that the system 'does not satisfy the latency constraints of the Belle II L1 trigger system required to participate in the overall trigger decision.' Nevertheless, the title, abstract, and introduction repeatedly call GNN-ETM a 'trigger module' and claim 'first operation ... within the real-time trigger readout path.' This conflates a synchronous readout-path monitor with an active trigger module. The planned latency reductions (doubling the clock, removing one GravNet block, eliminating ICN-ETM routing overhead) are only listed, not implemented or evaluated. This is load-bearing for the central claim; please either reframe t
  2. [Abstract vs. Sec. 8] There are two versions of the central claim: the abstract block states 'a first step towards GNN-based real-time reconstruction on FPGAs in a collider trigger,' while the full-text abstract and Sec. 8 state 'the first operation of a Graph Neural Network-based reconstruction system implemented on FPGAs within the real-time trigger readout path' and 'the first GNN-based reconstruction algorithm operating in a real-time particle physics environment.' These are materially different claims. The stronger version is not supported by the current system, which does not participate in the L1 decision. The manuscript should be harmonized to use one, precisely qualified claim, and the 'first operation ... within the trigger readout path' phrasing should be used only if the GDL interface is active or if 'readout path' is explicitly defined to exclude decision logic.
  3. [Sec. 7.1.1–7.1.3, Sec. 7.2.3, Abstract] The headline improvements—18% position resolution improvement, 20% cluster purity gain, 20% overlap-efficiency improvement, and up to 70% background rejection at 97.5% signal retention—are all obtained on specifically constructed simulation samples (Single Photon Sample, 2S1B isolated events, Overlap Diphoton Sample). The abstract reports these numbers without the 'simulation-only' qualification, and the collision-data sections show smaller or differently quantified gains (e.g., up to 20% rate reduction for a hypothetical two-cluster trigger line in Sec. 7.2.3). To avoid overstatement, the abstract and summary should clearly attribute these specific percentages to simulation, with data results described at the level demonstrated.
minor comments (4)
  1. [Table 3] The last column header reads 'ICN-ETM(MC)' but should be 'ICN-ETM(Data)' based on the context and values.
  2. [Sec. 7.1.1] Typo: 'ooffline reconstructed cluster' should be 'offline reconstructed cluster'.
  3. [Throughout] Several references are formatted inconsistently, e.g., 'Sec. Section 5.1' and 'Eq.,5.1'. Please standardize.
  4. [Sec. 6.3] The phrase 'reconfigure the module chain to eliminate the additional latency introduced by routing data through the ICN-ETM' is vague; a sentence explaining the alternative routing would help.

Circularity Check

0 steps flagged

No circular derivation; the FPGA deployment and physics benchmarks rest on independent training, hardware validation, and measured comparisons.

full rationale

The paper's derivation chain is self-contained. GNN-ETM is trained with standard supervised targets from offline basf2 reconstruction (Sec. 5.1), and the reported efficiency, purity, and resolutions are measured against those same offline clusters as the reference. This is the conventional ground truth for trigger reconstruction, not a circular prediction: the ICN-ETM baseline is evaluated on the identical events, and the claimed improvements (e.g., cross-gap clustering and overlapping-cluster splitting) are capabilities learned from offline labels and then demonstrated on independent simulated samples and collision data (Secs. 7.1-7.2). The energy bias correction (Sec. 5.2) fits a shift to the mean of the uncorrected resolution distribution and then reports the width; removing a bias does not fabricate resolution and is disclosed. The signal classifier threshold is tuned to a fixed 97.5% signal retention (GNN-ETM97.5, Sec. 7.1.2), so retention is an operating-point constraint, while the quoted background rejection up to 70% is a measured consequence after fixing that threshold, not a fitted target. Hardware validation is anchored by bitwise agreement between C-simulation and RTL simulation and by cosmic-ray data on hardware (Sec. 6.4), independent of the physics training targets. Self-citations such as [39] for prior FPGA graph building are contextual and not load-bearing; no uniqueness theorem, renamed ansatz, or construction-by-definition step appears. The paper's own limitations—the GDL interface 'exists but is currently unused' (Fig. 11) and latency 'exceeds the maximum allowed for an active L1 trigger decision by approximately a factor of three' (Sec. 6.3), repeated in Sec. 8—undercut the 'trigger module' framing, but that is a scope-of-claim issue, not circularity.

Axiom & Free-Parameter Ledger

9 free parameters · 7 axioms · 0 invented entities

The central claim rests primarily on domain assumptions about the fidelity of GEANT4 simulation and offline reconstruction as ground truth, plus a set of disclosed tuning parameters (thresholds, correction functions, hyperparameters). The paper introduces no new physical entities. The free parameters are honest tuning choices, but several—particularly the signal classifier thresholds and energy corrections—directly set the values of the headline performance metrics.

free parameters (9)
  • Energy bias correction f_corr(E_trg) = energy-dependent, per detector region
    Fitted via double-sided Crystal Ball means on simulated single photons (Sec. 5.2, Fig. 6); applied to both ICN-ETM and GNN-ETM energies before resolution reporting. Different corrections affect the headline energy resolution comparison.
  • Signal classifier thresholds tsig = 0.730 (forward), 0.615 (barrel), 0.465 (backward)
    Tuned on the Uniform Photon Sample to achieve 97.5% signal efficiency in 0.15-0.25 GeV (Sec. 7.1.2, Fig. 26). The GNN-ETM97.5 background-rejection claims depend on this choice.
  • Beta condensation cut t_beta = 0.04
    Chosen post-training to balance efficiency and purity (Sec. 5.5); directly affects cluster selection.
  • Latent distance cut t_d = 0.3
    Fixed with negligible impact on performance (Sec. 5.5); used in condensation point isolation.
  • Exponential weighting f_exp = 10
    Hyperparameter optimized over 1-10 range (Table 2); controls message-passing separation power.
  • Number of nearest neighbours k = 8
    Hyperparameter optimized over 2-8 range (Table 2); controls graph connectivity in GravNet.
  • Training sample balancing parameters a, b, lambda = a=5.0, b=32.6, lambda=3
    Determined from simulation studies to balance signal/background cluster energy and multiplicity in the PoissonUniformPhotonSample (Sec. 5.1). These shape the classifier's training distribution and could bias energy-based separation.
  • Pruning rate = 40%
    Low-magnitude pruning applied to fit FPGA resources; causes up to 10 percentage points efficiency loss in the backward endcap (Sec. 5.4).
  • Layer-wise quantization bit widths = 8-16 bit, Q3.5 to Q10.10
    Chosen by hand to satisfy FPGA resource and latency constraints (Sec. 5.4, Fig. 10); affects numerical fidelity and physics performance.
axioms (7)
  • domain assumption GEANT4 simulation accurately models the Belle II ECL response and beam backgrounds.
    All training and most evaluation uses GEANT4 + basf2 simulated events with overlaid beam backgrounds (Sec. 5.1). If the simulation is wrong, the training labels and performance estimates are wrong.
  • domain assumption Offline basf2 reconstructed clusters are a valid ground truth for trigger-level clustering.
    Training labels and evaluation reference are derived from offline reconstruction (Sec. 5.1). The paper does not validate offline reconstruction against a higher-truth standard.
  • domain assumption Simulated beam backgrounds from 2021 conditions are representative of 2024 data-taking conditions.
    Training uses beam background approximating 2021 luminosity 1.06e34 cm^-2 s^-1 (Sec. 5.1); the paper validates against Runs A-F with varying OOTC but does not retrain on 2024 backgrounds.
  • ad hoc to paper Events with more than 32 active trigger cells are negligible and can be truncated without bias.
    The model is optimized for up to 32 TCs, and inputs exceeding 32 are 'truncated without ordering, resulting in an arbitrary cut-off' (Sec. 5.3). The paper shows <0.002% of events exceed 32 in the evaluated runs, but the truncation behavior is otherwise unmodeled.
  • domain assumption The signal definition based on simulated energy fraction (>20% from a single particle) is appropriate.
    Signal labels for classifier training use simulation truth (Sec. 5.1). This is simulation-dependent and cannot be directly checked in collision data.
  • standard math Object condensation loss functions produce meaningful clusters with the chosen architecture.
    The model relies on the object condensation framework [28] to map nodes to variable-number clusters; the paper does not re-derive or justify the loss beyond citation.
  • domain assumption The 100 MeV trigger-cell energy threshold and 250 ns trigger window are fixed inputs.
    The GNN operates on trigger cells already thresholded at 100 MeV by the existing readout (Sec. 3). This discards low-energy information that offline reconstruction sees.

pith-pipeline@v1.3.0-alltime-deepseek · 39771 in / 12186 out tokens · 125890 ms · 2026-08-02T22:56:23.657687+00:00 · methodology

0 comments
read the original abstract

We present the development and evaluation of a real-time Graph Neural Network-based trigger module for the electromagnetic calorimeter of the Belle~II experiment at the SuperKEKB collider. The algorithm processes calorimeter trigger cells as graph nodes to perform clustering, feature extraction, and per-cluster signal classification with deterministic latency. The model predicts cluster positions and energies and provides a signal classification score, enabling a more flexible clustering strategy than the baseline trigger algorithm. Implemented on an FPGA and integrated into the Belle~II trigger readout infrastructure for synchronous operation, the system sustains the MHz trigger throughput with an end-to-end latency of $3.168\,\mu$s. The performance is evaluated on simulated events and collision data. The energy resolution is comparable to the baseline trigger, while the position resolution for high-energy clusters improves by up to 18% in the central detector region. Cluster purity increases by up to 20% at low energies for isolated clusters, and cluster efficiency improves by up to 20% for overlapping clusters. The signal classifier enables additional background suppression at fixed signal retention. These results demonstrate a first step towards GNN-based real-time reconstruction on FPGAs in a collider trigger. While the end-to-end latency exceeds the trigger decision budget, the system already sustains full operational conditions with 100% uptime.

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. RTL Fault Injection of a Deployed Graph Neural Network Trigger for Belle II

    hep-ex 2026-07 conditional novelty 6.5

    Inter-stage liveness monitoring of GNN-ETM reveals MTTF estimates up to 78.7% lower than output-only observation and ranks Preprocessing and GNN stages as top hardening targets.

  2. RTL Fault Injection of a Deployed Graph Neural Network Trigger for Belle II

    hep-ex 2026-07 conditional novelty 6.0

    Inter-stage liveness monitoring of a deployed L1 GNN trigger estimates MTTF up to 78.7% lower than output-only monitoring, revealing a detection-latency bias in current FPGA verification.

  3. Reconfigurable Computing Challenge: Real-Time Graph Neural Networks for Online Event Selection in Big Science

    cs.AR 2026-05 unverdicted novelty 5.0

    Hybrid FPGA-AI Engine deployment of a dynamic GNN for Belle II trigger achieves 2.94M events/s throughput at 7.15us latency with 53% better throughput and DSP usage reduced from 99% to 19%.

Reference graph

Works this paper leans on

71 extracted references · 30 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Abe et al.,Belle II Technical Design Report, Tech

    T. Abe et al.,Belle II Technical Design Report, Tech. Rep. KEK-REPORT-2010-1 (2010). [2]SuperKEKBcollaboration,SuperKEKB Collider,Nucl. Instrum. Meth. A907(2018) 188 [1809.01958]

    Pith/arXiv arXiv 2010

  2. [3]

    Natochii et al.,Measured and projected beam backgrounds in the Belle II experiment at the SuperKEKB collider,Nucl

    A. Natochii et al.,Measured and projected beam backgrounds in the Belle II experiment at the SuperKEKB collider,Nucl. Instrum. Meth. A1055(2023) 168550 [2302.01566]

    Pith/arXiv arXiv 2023

  3. [4]

    Fabbrichesi, E

    M. Fabbrichesi, E. Gabrielli and G. Lanfranchi,The Dark Photon,2005.01515

    Pith/arXiv arXiv 2005

  4. [5]

    Ferber, C

    T. Ferber, C. Garcia-Cely and K. Schmidt-Hoberg,Belle II sensitivity to long–lived dark photons, Phys. Lett. B833(2022) 137373 [2202.03452]

    Pith/arXiv arXiv 2022

  5. [6]

    Jaeckel and A.V

    J. Jaeckel and A.V. Phan,Searching dark photons using displaced vertices at Belle II – with backgrounds,JHEP08(2024) 062 [2312.12522]

    arXiv 2024

  6. [7]

    Duerr, T

    M. Duerr, T. Ferber, C. Hearty, F. Kahlhoefer, K. Schmidt-Hoberg and P. Tunney,Invisible and displaced dark matter signatures at Belle II,JHEP02(2020) 039 [1911.03176]

    Pith/arXiv arXiv 2020

  7. [8]

    I.Adachietal.,SearchforaDarkHiggsBosonProducedinAssociationwithInelasticDarkMatterat the Belle II Experiment,Phys. Rev. Lett.135(2025) 131801 [2505.09705]

    arXiv 2025

  8. [9]

    Dolan et al.,Revised constraints and Belle II sensitivity for visible and invisible axion-like particles,JHEP12(2017) 094 [1709.00009]

    M.J. Dolan et al.,Revised constraints and Belle II sensitivity for visible and invisible axion-like particles,JHEP12(2017) 094 [1709.00009]

    Pith/arXiv arXiv 2017

  9. [10]

    Abudinén et al.,Search for Axion-Like Particles produced in𝑒+𝑒− collisions at Belle II,Phys

    F. Abudinén et al.,Search for Axion-Like Particles produced in𝑒+𝑒− collisions at Belle II,Phys. Rev. Lett.125(2020) 161806 [2007.13071]

    Pith/arXiv arXiv 2020

  10. [11]

    Lai et al.,Design of the Global Reconstruction Logic in the Belle II Level-1 Trigger system,Nucl

    Y.T. Lai et al.,Design of the Global Reconstruction Logic in the Belle II Level-1 Trigger system,Nucl. Instrum. Meth. A1078(2025) 170577 [2503.02192]

    Pith/arXiv arXiv 2025

  11. [12]

    S. Kim, I. Lee, Y. Unno and B. Cheon,Status of the electromagnetic calorimeter trigger system at Belle II.,J. Phys. Conf. Ser.928(2017) 012022

    2017

  12. [13]

    Canudas et al.,Graph Clustering: A Graph-Based Clustering Algorithm for the Electromagnetic Calorimeter in LHCb,The European Physical Journal C83(2023)

    N.V. Canudas et al.,Graph Clustering: A Graph-Based Clustering Algorithm for the Electromagnetic Calorimeter in LHCb,The European Physical Journal C83(2023)

    2023

  13. [14]

    Valsecchi,Deep Learning Techniques for Energy Clustering in The CMS ECAL,Journal of Physics: Conference Series2438(2023) 012077

    D. Valsecchi,Deep Learning Techniques for Energy Clustering in The CMS ECAL,Journal of Physics: Conference Series2438(2023) 012077

    2023

  14. [15]

    Wemmer et al.,Photon Reconstruction in the Belle II Calorimeter Using Graph Neural Networks, Comput

    F. Wemmer et al.,Photon Reconstruction in the Belle II Calorimeter Using Graph Neural Networks, Comput. Softw. Big Sci.7(2023) 13 [2306.04179]

    Pith/arXiv arXiv 2023

  15. [16]

    Simkina,Machine Learning Techniques for Calorimetry,Instruments6(2022) 47

    P. Simkina,Machine Learning Techniques for Calorimetry,Instruments6(2022) 47

    2022

  16. [17]

    Belayneh et al.,Calorimetry With Deep Learning: Particle Simulation and Reconstruction for Collider Physics,The European Physical Journal C80(2019)

    D.T. Belayneh et al.,Calorimetry With Deep Learning: Particle Simulation and Reconstruction for Collider Physics,The European Physical Journal C80(2019)

    2019

  17. [18]

    Boldyrev, V

    A. Boldyrev, V. Chekalina and F. Ratnikov,Machine Learning Approach to Boosting Neutral Particles Identification in the LHCb Calorimeter,J. Phys. Conf. Ser.1525(2020) 012096

    2020

  18. [19]

    Novosel et al.,Identification of light leptons and pions in the electromagnetic calorimeter of Belle II,Nucl

    A. Novosel et al.,Identification of light leptons and pions in the electromagnetic calorimeter of Belle II,Nucl. Instrum. Meth. A1056(2023) 168630 [2301.05074]

    Pith/arXiv arXiv 2023

  19. [20]

    Shlomi, P

    J. Shlomi, P. Battaglia and J. Vlimant,Graph Neural Networks in Particle Physics,Machine Learning: Science and Technology2(2021) 021001

    2021

  20. [21]

    Duarte and J.-R

    J. Duarte and J.-R. Vlimant,Graph Neural Networks for Particle Tracking and Reconstruction, 2012.01249. – 46 –

    Pith/arXiv arXiv 2012

  21. [22]

    DeZoort et al.,Graph neural networks at the Large Hadron Collider,Nature Rev

    G. DeZoort et al.,Graph neural networks at the Large Hadron Collider,Nature Rev. Phys.5(2023) 281

    2023

  22. [23]

    Wang et al.,Dynamic Graph CNN for Learning on Point Clouds,1801.07829

    Y. Wang et al.,Dynamic Graph CNN for Learning on Point Clouds,1801.07829

    Pith/arXiv arXiv

  23. [24]

    Qasim et al.,Learning Representations of Irregular Particle-Detector Geometry With Distance-Weighted Graph Networks,Eur

    S.R. Qasim et al.,Learning Representations of Irregular Particle-Detector Geometry With Distance-Weighted Graph Networks,Eur. Phys. J. C79(2019) 608 [1902.07987]. [25]CMS HGCAL, CALICE AHCALcollaboration,Using graph neural networks to reconstruct charged pion showers in the CMS High Granularity Calorimeter,JINST19(2024) P11025 [2406.11937]. [26]Exa.TrkXco...

    Pith/arXiv arXiv 2019

  24. [27]

    Redmon et al.,You Only Look Once: Unified, Real-Time Object Detection, in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    J. Redmon et al.,You Only Look Once: Unified, Real-Time Object Detection, in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 779–788, 2016, DOI

    2016

  25. [28]

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

    J. Kieseler,Object condensation: one-stage grid-free multi-object reconstruction in physics detectors, graph and image data,Eur. Phys. J. C80(2020) 886 [2002.03605]

    Pith/arXiv arXiv 2020

  26. [29]

    Qasim et al.,End-to-end multi-particle reconstruction in high occupancy imaging calorimeters with graph neural networks,Eur

    S.R. Qasim et al.,End-to-end multi-particle reconstruction in high occupancy imaging calorimeters with graph neural networks,Eur. Phys. J. C82(2022) 753 [2204.01681]

    Pith/arXiv arXiv 2022

  27. [30]

    Duarte et al.,Fast inference of deep neural networks in FPGAs for particle physics,JINST13 (2018) P07027 [1804.06913]

    J. Duarte et al.,Fast inference of deep neural networks in FPGAs for particle physics,JINST13 (2018) P07027 [1804.06913]

    Pith/arXiv arXiv 2018

  28. [31]

    10.5281/zenodo.1201549

    FastML Team,fastmachinelearning/hls4ml, 2023. 10.5281/zenodo.1201549

    doi:10.5281/zenodo.1201549 2023

  29. [32]

    Blott et al.,FINN-R: An end-to-end deep-learning framework for fast exploration of quantized neural networks,TRETS11(2018)

    M. Blott et al.,FINN-R: An end-to-end deep-learning framework for fast exploration of quantized neural networks,TRETS11(2018)

    2018

  30. [33]

    Kohne et al.,Realization of a second level neural network trigger for the H1 experiment at HERA,Nucl

    J.K. Kohne et al.,Realization of a second level neural network trigger for the H1 experiment at HERA,Nucl. Instrum. Meth. A389(1997) 128

    1997

  31. [34]

    Afanasyev et al.,The Multilevel trigger system of the DIRAC experiment,Nucl

    L. Afanasyev et al.,The Multilevel trigger system of the DIRAC experiment,Nucl. Instrum. Meth. A 491(2002) 376 [hep-ex/0202045]

    Pith/arXiv arXiv 2002

  32. [35]

    Bähr et al.,The neural network first-level hardware track trigger of the Belle II experiment,Nucl

    S. Bähr et al.,The neural network first-level hardware track trigger of the Belle II experiment,Nucl. Instrum. Meth. A1073(2025) 170279 [2402.14962]. [36]CMScollaboration,Testing a Neural Network for Anomaly Detection in the CMS Global Trigger Test Crate during Run 3,JINST19(2024) C03029 [2312.10009]

    Pith/arXiv arXiv 2025

  33. [37]

    Iiyama et al.,Distance-Weighted Graph Neural Networks on FPGAs for Real-Time Particle Reconstruction in High Energy Physics,Front

    Y. Iiyama et al.,Distance-Weighted Graph Neural Networks on FPGAs for Real-Time Particle Reconstruction in High Energy Physics,Front. Big Data3(2020) 598927 [2008.03601]

    Pith/arXiv arXiv 2020

  34. [38]

    Que et al.,LL-GNN: Low Latency Graph Neural Networks on FPGAs for High Energy Physics, ACM Trans

    Z. Que et al.,LL-GNN: Low Latency Graph Neural Networks on FPGAs for High Energy Physics, ACM Trans. Embed. Comput. Syst.23(2024) 1 [2209.14065]

    Pith/arXiv arXiv 2024

  35. [39]

    Neu et al.,Real-Time Graph Building on FPGAs for Machine Learning Trigger Applications in Particle Physics,Comput

    M. Neu et al.,Real-Time Graph Building on FPGAs for Machine Learning Trigger Applications in Particle Physics,Comput. Softw. Big Sci.8(2024) 8 [2307.07289]

    Pith/arXiv arXiv 2024

  36. [40]

    J. Kvapil et al.,Intelligent experiments through real-time AI: Fast Data Processing and Autonomous Detector Control for sPHENIX and future EIC detectors,PoSICHEP2024(2025) 1033 [2501.04845]

    Pith/arXiv arXiv 2025

  37. [41]

    Kou et al.,The Belle II Physics Book,PTEP2019(2019) 123 C01

    E. Kou et al.,The Belle II Physics Book,PTEP2019(2019) 123 C01

    2019

  38. [42]

    Ikeda,Development of the CsI(Tl) Calorimeter for the Measurement of CP Violation at KEK B-Factory, Ph.D

    H. Ikeda,Development of the CsI(Tl) Calorimeter for the Measurement of CP Violation at KEK B-Factory, Ph.D. thesis, Nara Women’s University, 1999. – 47 –

    1999

  39. [43]

    Aulchenko et al.,Time and energy reconstruction at the electromagnetic calorimeter of the Belle-II detector,JINST12(2017) C08001

    V. Aulchenko et al.,Time and energy reconstruction at the electromagnetic calorimeter of the Belle-II detector,JINST12(2017) C08001

    2017

  40. [44]

    Beylin et al.,Study of the radiation hardness of CsI(Tl) scintillation crystals,Nucl

    D.M. Beylin et al.,Study of the radiation hardness of CsI(Tl) scintillation crystals,Nucl. Instrum. Meth. A541(2005) 501 [physics/0403136]

    Pith/arXiv arXiv 2005

  41. [45]

    Kovalenko et al.,Luminosity online monitor for the Belle II detector,Nucl

    E. Kovalenko et al.,Luminosity online monitor for the Belle II detector,Nucl. Instrum. Meth. A1079 (2025) 170614

    2025

  42. [46]

    Cheon et al.,Electromagnetic calorimeter trigger at Belle,Nucl

    B.G. Cheon et al.,Electromagnetic calorimeter trigger at Belle,Nucl. Instrum. Meth. A494(2002) 548. [47]GEANT4 collaborationcollaboration,GEANT4: A Simulation Toolkit,Nucl.Instrum.Meth.A506 (2003) 250

    2002

  43. [48]

    Belle II Analysis Software Framework (basf2)

    T. Kuhr et al.,The Belle II Core Software,Computing and Software for Big Science.3(2019) . [49]Belle IIcollaboration, Belle II Collaboration, “Belle II Analysis Software Framework (basf2).” https://doi.org/10.5281/zenodo.5574115

    doi:10.5281/zenodo.5574115 2019

  44. [50]

    Liptak et al.,Measurements of beam backgrounds in SuperKEKB Phase 2,Nucl

    Z.J. Liptak et al.,Measurements of beam backgrounds in SuperKEKB Phase 2,Nucl. Instrum. Methods Phys. Res. A1040(2022) 167168 [2112.14537]

    Pith/arXiv arXiv 2022

  45. [51]

    Natochii et al.,Beam Background Expectations for Belle II at SuperKEKB,2203.05731

    A. Natochii et al.,Beam Background Expectations for Belle II at SuperKEKB,2203.05731

    Pith/arXiv arXiv

  46. [52]

    Jadach, B.F.L

    S. Jadach, B.F.L. Ward and Z. Was,The Precision Monte Carlo event generator KK for two fermion final states in electron positron collisions,Comput. Phys. Commun.130(2000) 260 [hep-ph/9912214]

    Pith/arXiv arXiv 2000

  47. [53]

    Balossini, C.M

    G. Balossini, C.M. Carloni Calame, G. Montagna, O. Nicrosini and F. Piccinini,Matching perturbative and parton shower corrections to Bhabha process at flavour factories,Nucl. Phys. B758 (2006) 227 [hep-ph/0607181]

    Pith/arXiv arXiv 2006

  48. [54]

    Gaiser,Charmonium Spectroscopy From Radiative Decays of the𝐽/𝜓and𝜓′, Ph.D

    J.E. Gaiser,Charmonium Spectroscopy From Radiative Decays of the𝐽/𝜓and𝜓′, Ph.D. thesis, SLAC, 1982

    1982

  49. [55]

    Skwarnicki,A study of the radiative CASCADE transitions between the Upsilon-Prime and Upsilon resonances, Ph.D

    T. Skwarnicki,A study of the radiative CASCADE transitions between the Upsilon-Prime and Upsilon resonances, Ph.D. thesis, Cracow, INP, 1986

    1986

  50. [56]

    Chollet et al., “Keras.”https://keras.io, 2015

    F. Chollet et al., “Keras.”https://keras.io, 2015

    2015

  51. [57]

    Agarap,Deep Learning using Rectified Linear Units (ReLU),1803.08375

    A.F. Agarap,Deep Learning using Rectified Linear Units (ReLU),1803.08375

    Pith/arXiv arXiv

  52. [58]

    Coelho et al.,Automatic heterogeneous quantization of deep neural networks for low-latency inference on the edge for particle detectors,Nature Machine Intelligence3(2021) 675

    C.N. Coelho et al.,Automatic heterogeneous quantization of deep neural networks for low-latency inference on the edge for particle detectors,Nature Machine Intelligence3(2021) 675

    2021

  53. [59]

    Zhu and S

    M. Zhu and S. Gupta,To prune, or not to prune: exploring the efficacy of pruning for model compression,1710.01878

    Pith/arXiv arXiv

  54. [60]

    Experiment Tracking with Weights and Biases

    L. Biewald, “Experiment Tracking with Weights and Biases.” https://www.wandb.com, 2020

    2020

  55. [61]

    AMBA AXI-Stream Protocol Specification

    Arm Limited, “AMBA AXI-Stream Protocol Specification.” https://developer.arm.com/documentation/ihi0051/latest, 2021

    2021

  56. [62]

    Institute of Electrical and Electronics Engineers, New York, NY, USA,IEEE Standard for a Versatile Backplane Bus: VMEbus, 1987

    1987

  57. [63]

    Sun et al.,Belle2Link: A Global Data Readout and Transmission for Belle II Experiment at KEK, Physics Procedia37(2012) 1933

    D. Sun et al.,Belle2Link: A Global Data Readout and Transmission for Belle II Experiment at KEK, Physics Procedia37(2012) 1933

    2012

  58. [64]

    Yamada et al.,Data Acquisition System for the Belle II Experiment,IEEE Transactions on Nuclear Science62(2015) 1175

    S. Yamada et al.,Data Acquisition System for the Belle II Experiment,IEEE Transactions on Nuclear Science62(2015) 1175. – 48 –

    2015

  59. [65]

    Bachrach et al.,Chisel: Constructing Hardware in a Scala Embedded Language, inProceedings of the 49th Annual Design Automation Conference, (San Francisco California), pp

    J. Bachrach et al.,Chisel: Constructing Hardware in a Scala Embedded Language, inProceedings of the 49th Annual Design Automation Conference, (San Francisco California), pp. 1216–1225, ACM, June, 2012, DOI

    2012

  60. [66]

    FuseSoC

    O. Kindgren, “FuseSoC.”https://github.com/olofk/fusesoc, 2015

    2015

  61. [67]

    Invited Paper: A Scalable Approach to IPManagement with FuseSoC

    O. Kindgren, “Invited Paper: A Scalable Approach to IPManagement with FuseSoC.” Presented at the Workshop on Open Source Design Automation (OSDA), 2019, 2019

    2019

  62. [68]

    Vitis Unified Software Platform

    AMD, “Vitis Unified Software Platform.” https://www.amd.com/en/products/software/adaptive-socs-and-fpgas/vitis.html, 2025

    2025

  63. [69]

    Vivado Design Suite

    AMD, “Vivado Design Suite.” https://www.amd.com/en/products/software/adaptive-socs-and-fpgas/vivado.html, 2025

    2025

  64. [70]

    ModelSim HDL simulator

    AMD, “ModelSim HDL simulator.”https://eda.sw.siemens.com/en-US/ic/modelsim/, 2025

    2025

  65. [71]

    cocotb: Python-based chip (RTL) verification

    S. Hodgson et al., “cocotb: Python-based chip (RTL) verification.” https://github.com/cocotb/cocotb, 2025

    2025

  66. [72]

    cocotb-axi: Python-based chip (RTL) verification

    A. Forencich, “cocotb-axi: Python-based chip (RTL) verification.” https://github.com/alexforencich/cocotbext-axi, 2023

    2023

  67. [73]

    I.Adachietal.,Measurementoftheintegratedluminosityofdatasamplescollectedduring2019-2022 by the Belle II experiment,Chin. Phys. C49(2025) 013001 [2407.00965]

    Pith/arXiv arXiv 2022

  68. [74]

    Code for the GNN-ETM Training and Evaluation

    I. Haide et al., “Code for the GNN-ETM Training and Evaluation.” https://github.com/ihaide/gnnetm-software, 2026

    2026

  69. [75]

    Code for the Quantized GravNet Implementation

    M. Neu et al., “Code for the Quantized GravNet Implementation.” https://github.com/ihaide/qgravnet, 2026

    2026

  70. [76]

    Custom QKeras Fork

    M. Neu and I. Haide, “Custom QKeras Fork.” https://github.com/ihaide/qkeras, 2026

    2026

  71. [77]

    Code for the GNN-ETM Hardware Implementation

    M. Neu et al., “Code for the GNN-ETM Hardware Implementation.” https://github.com/marcneu/pcnhlslib/tree/gnnetm, 2026. – 49 –

    2026