pith. sign in

arxiv: 2507.19205 · v2 · submitted 2025-07-25 · 💻 cs.LG

Physics-Informed Graph Neural Networks for Transverse Momentum Estimation in CMS Trigger Systems

Md Abrar Jahin , Shahriar Soudeep , M. F. Mridha , Muhammad Mostafa Monowar , Md. Abdul Hamid This is my paper

Pith reviewed 2026-05-19 02:43 UTC · model grok-4.3

classification 💻 cs.LG
keywords Graph Neural NetworksPhysics Informed MLTransverse MomentumCMS DetectorTrigger SystemsEdge ConvolutionMessage Passing
0
0 comments X

The pith

Physics-informed graph neural networks estimate transverse momentum more accurately with over half the parameters of standard models.

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

The paper shows how to build graph neural networks that incorporate the geometry of particle detectors and physical properties of momentum distributions to improve real-time estimation of particle transverse momentum. This matters because high-energy physics experiments like CMS need fast algorithms that work under tight hardware limits and high data rates from pileup. By testing four different ways to turn detector data into graphs and adding a custom message passing layer with attention, the approach achieves better accuracy than generic deep learning while using fewer resources. If correct, it suggests that embedding domain knowledge directly into the model structure leads to more efficient and robust performance in constrained environments.

Core claim

The central claim is that a physics-informed GNN framework with four graph construction strategies encoding detector geometry and observables, combined with a novel Message Passing Layer featuring intra-message attention and gated updates, and domain-specific loss functions with pT-distribution priors, yields superior accuracy-efficiency trade-offs. Specifically, a station-informed EdgeConv model reaches a state-of-the-art MAE of 0.8525 with at least 55 percent fewer parameters than baselines such as TabNet.

What carries the argument

The station-informed EdgeConv model that uses station-as-node graph representations to capture detector geometry within the physics-informed GNN framework.

If this is right

  • Real-time pT estimation in CMS trigger systems becomes more accurate and resource-efficient.
  • The approach validates the use of tailored graph structures for physics observables in high-pileup scenarios.
  • Domain-specific loss functions incorporating pT priors improve regression performance under hardware constraints.

Where Pith is reading between the lines

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

  • Similar physics-informed graph constructions could apply to other detector-based regression tasks in particle physics.
  • Integrating these methods with hardware accelerators might further reduce latency in trigger systems.
  • This co-design of graph structure and loss could inspire analogous techniques for other constrained ML applications.

Load-bearing premise

The four graph construction strategies and domain-specific loss functions are assumed to capture detector geometry and physical observables without introducing biases that degrade performance on unseen high-pileup data.

What would settle it

Running the models on a new dataset with significantly higher pileup levels and observing that the MAE exceeds that of simpler baselines would falsify the performance claims.

Figures

Figures reproduced from arXiv: 2507.19205 by Md. Abdul Hamid, Md Abrar Jahin, M. F. Mridha, Muhammad Mostafa Monowar, Shahriar Soudeep.

Figure 1
Figure 1. Figure 1: (a) Gaussian distribution and (b) Frequency heatmap of TabNet (left) and best-performing 4 EdgeConv-based (embedding dimension: 16) custom loss function optimized model (right). The EdgeConv model (right) demonstrates significantly more concentrated residuals (lower standard deviation, mean closer to zero) and a tighter alignment between predicted and true pT values in the heatmap, indicating superior pred… view at source ↗
read the original abstract

Real-time particle transverse momentum ($p_T$) estimation in high-energy physics demands algorithms that are both efficient and accurate under strict hardware constraints. Static machine learning models degrade under high pileup and lack physics-aware optimization, while generic graph neural networks (GNNs) often neglect domain structure critical for robust $p_T$ regression. We propose a physics-informed GNN framework that systematically encodes detector geometry and physical observables through four distinct graph construction strategies that systematically encode detector geometry and physical observables: station-as-node, feature-as-node, bending angle-centric, and pseudorapidity ($\eta$)-centric representations. This framework integrates these tailored graph structures with a novel Message Passing Layer (MPL), featuring intra-message attention and gated updates, and domain-specific loss functions incorporating $p_{T}$-distribution priors. Our co-design methodology yields superior accuracy-efficiency trade-offs compared to existing baselines. Extensive experiments on the CMS Trigger Dataset validate the approach: a station-informed EdgeConv model achieves a state-of-the-art MAE of 0.8525 with $\ge55\%$ fewer parameters than deep learning baselines, especially TabNet, while an $\eta$-centric MPL configuration also demonstrates improved accuracy with comparable efficiency. These results establish the promise of physics-guided GNNs for deployment in resource-constrained trigger systems.

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 physics-informed GNN framework for real-time p_T regression in CMS trigger systems. It defines four graph-construction strategies (station-as-node, feature-as-node, bending-angle-centric, η-centric) that encode detector geometry, pairs them with a custom Message Passing Layer containing intra-message attention and gated updates, and augments training with domain-specific losses that incorporate p_T-distribution priors. On the CMS Trigger Dataset a station-informed EdgeConv variant is reported to reach MAE 0.8525 while using ≥55 % fewer parameters than TabNet and other deep-learning baselines; an η-centric MPL configuration is also shown to improve accuracy at comparable efficiency.

Significance. If the reported accuracy and parameter counts are reproducible under controlled splits and pileup variation, the work would supply a concrete, hardware-friendly alternative to existing trigger algorithms and demonstrate that modest domain-specific graph inductive biases can yield measurable efficiency gains in a production HEP setting.

major comments (2)
  1. [Experiments] Experiments section: the central performance claim (MAE = 0.8525, ≥55 % parameter reduction) is presented without error bars, without any description of the train/validation/test split ratios or random seeds, and without an ablation that isolates the contribution of each of the four graph constructions or of the p_T-prior losses versus a standard MSE baseline. These omissions make the numerical result impossible to verify or attribute.
  2. [Results] Results and discussion: no pileup-stratified metrics or out-of-distribution splits are reported. Because the motivating claim is robustness under high instantaneous luminosity, the absence of these controls leaves open whether the physics-informed components are capturing detector geometry or merely fitting training-specific correlations.
minor comments (2)
  1. [Abstract] Abstract: the phrase “that systematically encode detector geometry and physical observables” is repeated verbatim in consecutive sentences.
  2. [Methods] Notation: the manuscript introduces “MPL” and “EdgeConv” without an explicit first-use definition or pointer to the original EdgeConv reference.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below, indicating where revisions will be made to improve clarity, reproducibility, and robustness analysis.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: the central performance claim (MAE = 0.8525, ≥55 % parameter reduction) is presented without error bars, without any description of the train/validation/test split ratios or random seeds, and without an ablation that isolates the contribution of each of the four graph constructions or of the p_T-prior losses versus a standard MSE baseline. These omissions make the numerical result impossible to verify or attribute.

    Authors: We agree that additional details are needed for reproducibility and to properly attribute performance gains. In the revised manuscript we will report error bars computed across multiple runs with distinct random seeds, explicitly document the train/validation/test split ratios and seeds, and include ablation studies that separately evaluate each of the four graph-construction strategies as well as the p_T-prior loss terms against a plain MSE baseline. These changes will make the central claims verifiable and allow clearer attribution of the reported improvements. revision: yes

  2. Referee: [Results] Results and discussion: no pileup-stratified metrics or out-of-distribution splits are reported. Because the motivating claim is robustness under high instantaneous luminosity, the absence of these controls leaves open whether the physics-informed components are capturing detector geometry or merely fitting training-specific correlations.

    Authors: We acknowledge the importance of these controls for validating the motivating claim of robustness under high pileup. The CMS Trigger Dataset is derived from real collision data that already contains varying pileup levels. In the revision we will add pileup-stratified performance tables. For out-of-distribution evaluation we will report results on held-out high-pileup subsets; if the available data volume limits full OOD splits we will clearly state the constraints and the extent of the analysis performed. These additions will help demonstrate that gains arise from the physics-informed inductive biases rather than training-set correlations. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical model with external dataset validation

full rationale

The manuscript presents an empirical physics-informed GNN framework trained and evaluated on the external CMS Trigger Dataset, reporting concrete performance metrics such as MAE 0.8525. Graph construction strategies and domain-specific losses are presented as design choices that incorporate detector geometry and pT priors, not as a closed mathematical derivation that reduces to its own inputs by construction. No equations, self-citations, or fitted-parameter renamings appear in the text that would make the central claims equivalent to the inputs. The approach remains self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central performance claim rests on standard supervised learning assumptions plus domain knowledge that detector geometry and pT distributions can be usefully encoded as graph structure and loss priors; no new particles or forces are postulated.

free parameters (1)
  • pT-distribution priors
    Incorporated into domain-specific loss functions; their exact functional form and fitting procedure are not detailed in the abstract.
axioms (1)
  • domain assumption Detector geometry and physical observables can be systematically encoded through the four listed graph construction strategies without loss of critical information.
    Invoked when defining station-as-node, feature-as-node, bending-angle-centric, and η-centric representations.

pith-pipeline@v0.9.0 · 5782 in / 1452 out tokens · 48658 ms · 2026-05-19T02:43:36.876047+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages · 4 internal anchors

  1. [1]

    C. M. S. Collaboration, The CMS trigger system, Journal of Instrumentation 12 (01) (2017) P01020–P01020, arXiv:1609.02366 [physics]. doi:10.1088/1748-0221/12/01/P01020. URLhttp://arxiv.org/abs/1609.02366

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1088/1748-0221/12/01/p01020 2017

  2. [2]

    Hayrapetyan, A

    A. Hayrapetyan, A. Tumasyan, W. Adam, J. W. Andrejkovic, L. Benato, T. Bergauer, S. Chatterjee, K. Damanakis, M. Dragicevic, P. S. Hussain, Performance of the CMS high-level trigger during LHC Run 2, Journal of Instrumentation 19 (11) (2024) p11021. URLhttps://iopscience.iop.org/article/10.1088/1748-0221/19/11/P11021/meta

    work page doi:10.1088/1748-0221/19/11/p11021/meta 2024

  3. [3]

    R. F. d. Silva, M. Rynge, G. Juve, I. Sfiligoi, E. Deelman, J. Letts, F. W Ãijrthwein, M. Livny, Characterizing a High Throughput Computing Workload: The Compact Muon Solenoid (CMS) Experiment at LHC, in: S. Koziel, L. Leifsson, M. Lees, V . V . Krzhizhanovskaya, J. J. Dongarra, P. M. A. Sloot (Eds.), Proceedings of the International Conference on Computa...

    work page doi:10.1016/j.procs.2015.05.190 2015

  4. [4]

    Andreeva, S

    J. Andreeva, S. Campana, F. Fanzago, J. Herrala, High-Energy Physics on the Grid: the ATLAS and CMS Experience, J. Grid Comput. 6 (1) (2008) 3–13.doi:10.1007/S10723-007-9087-3. URLhttps://doi.org/10.1007/s10723-007-9087-3

    work page doi:10.1007/s10723-007-9087-3 2008

  5. [5]

    H. B. Newman, High energy physics - Search for higgs boson diphoton decay with CMS at LHC, in: Proceedings of the ACM/IEEE SC2006 Conference on High Performance Networking and Computing, November 11-17, 2006, Tampa, FL, USA, ACM Press, 2006, p. 59. doi:10.1145/1188455.1188517

    work page doi:10.1145/1188455.1188517 2006

  6. [6]

    Acosta, A

    D. Acosta, A. Brinkerhoff, E. Busch, A. Carnes, I. Furic, S. Gleyzer, K. Kotov, J. F. Low, A. Madorsky, J. Rorie, B. Scurlock, W. Shi, o. b. o. t. C. Collaboration, Boosted Decision Trees in the Level-1 Muon Endcap Trigger at CMS, Journal of Physics: Conference Series 1085 (4) (2018) 042042, publisher: IOP Publishing.doi:10.1088/1742-6596/1085/4/042042. U...

    work page doi:10.1088/1742-6596/1085/4/042042 2018

  7. [7]

    J. C. Sekhar, R. Priyanka, A. K. Nanda, P. J. Josephson, M. J. D. Ebinezer, T. K. Devi, Stochastic gradient boosted distributed decision trees security approach for detecting cyber anomalies and classifying multiclass cyber-attacks, Comput. Secur. 151 (2025) 104320. doi: 10.1016/J.COSE.2025.104320. URLhttps://doi.org/10.1016/j.cose.2025.104320

    work page doi:10.1016/j.cose.2025.104320 2025

  8. [8]

    Khataei, K

    A. Khataei, K. Bazargan, TreeLUT: An Efficient Alternative to Deep Neural Networks for Inference Acceleration Using Gradient Boosted Decision Trees, in: A. Putnam, J. Li (Eds.), Proceedings of the 2025 ACM/SIGDA International Symposium on Field Programmable Gate Arrays, FPGA 2025, Monterey, CA, USA, 27 February 2025 - 1 March 2025, ACM, 2025, pp. 14–24.do...

    work page doi:10.1145/3706628.3708877 2025

  9. [9]

    Mahmood, A

    T. Mahmood, A. Rehman, T. Saba, T. J. Alahmadi, M. Tufail, S. A. O. Bahaj, Z. Ahmad, Enhancing Coronary Artery Disease Prognosis: A Novel Dual-Class Boosted Decision Trees Strategy for Robust Optimization, IEEE Access 12 (2024) 107119–107143. doi:10.1109/ ACCESS.2024.3435948

    work page arXiv 2024

  10. [10]

    Antoniuk, S

    P. Antoniuk, S. K. Zielinski, H. Lee, Ensemble width estimation in HRTF-convolved binaural music recordings using an auditory model and a gradient-boosted decision trees regressor, EURASIP J. Audio Speech Music. Process. 2024 (1) (2024) 53. doi:10.1186/ S13636-024-00374-2. URLhttps://doi.org/10.1186/s13636-024-00374-2

    work page doi:10.1186/s13636-024-00374-2 2024

  11. [11]

    C. S. Griffiths, J. M. Lebert, J. Sollini, J. K. Bizley, Gradient boosted decision trees reveal nuances of auditory discrimination behavior, PLoS Comput. Biol. 20 (4) (2024) 1011985.doi:10.1371/JOURNAL.PCBI.1011985. URLhttps://doi.org/10.1371/journal.pcbi.1011985

    work page doi:10.1371/journal.pcbi.1011985 2024

  12. [12]

    C. E. Brown, Fast Machine Learning in the CMS Level-1 Trigger for the High-Luminosity LHC, Ph.D. thesis, Imperial College London, presented 25 Jul 2023 (2023). URLhttps://cds.cern.ch/record/2875830

    work page arXiv 2023

  13. [13]

    Hariri, Graph Neural Network Architectures for Fast Simulation and Muon Momentum Inference at the CMS Detector, Ph.D

    A. Hariri, Graph Neural Network Architectures for Fast Simulation and Muon Momentum Inference at the CMS Detector, Ph.D. thesis, American University of Beirut, presented 10 Dec 2020 (2021). URLhttps://cds.cern.ch/record/2758631

    work page arXiv 2020

  14. [14]

    Duarte, S

    J. Duarte, S. Han, P. Harris, S. Jindariani, E. Kreinar, B. Kreis, J. Ngadiuba, M. Pierini, R. Rivera, N. Tran, Fast inference of deep neural networks in FPGAs for particle physics, Journal of instrumentation 13 (07) (2018) P07027, publisher: IOP Publishing. URLhttps://iopscience.iop.org/article/10.1088/1748-0221/13/07/P07027/meta

    work page doi:10.1088/1748-0221/13/07/p07027/meta 2018

  15. [15]

    Shlomi, P

    J. Shlomi, P. Battaglia, J.-R. Vlimant, Graph neural networks in particle physics, Machine Learning: Science and Technology 2 (2) (2020) 021001, publisher: IOP Publishing. URLhttps://iopscience.iop.org/article/10.1088/2632-2153/abbf9a/meta

    work page doi:10.1088/2632-2153/abbf9a/meta 2020

  16. [16]

    C. Sun, J. Ngadiuba, M. Pierini, M. Spiropulu, Fast Jet Tagging with MLP-Mixers on FPGAs, CoRR abs/2503.03103, arXiv: 2503.03103 (2025).doi:10.48550/ARXIV.2503.03103. URLhttps://doi.org/10.48550/arXiv.2503.03103

    work page doi:10.48550/arxiv.2503.03103 2025

  17. [17]

    A. Stein, Novel jet flavour tagging algorithms exploiting adversarial deep learning techniques with efficient computing methods and preparation of open data for robustness studies, PhD Thesis, RWTH Aachen University, Germany (2024). URLhttps://publications.rwth-aachen.de/record/991721

    work page 2024

  18. [18]

    Lange, S

    T. Lange, S. Nandan, J. Pata, L. Tani, C. Veelken, Tau lepton identification and reconstruction: A new frontier for jet-tagging ML algorithms, Comput. Phys. Commun. 298 (2024) 109095.doi:10.1016/J.CPC.2024.109095. URLhttps://doi.org/10.1016/j.cpc.2024.109095

    work page doi:10.1016/j.cpc.2024.109095 2024

  19. [19]

    Iiyama, G

    Y . Iiyama, G. Cerminara, A. Gupta, J. Kieseler, V . Loncar, M. Pierini, S. R. Qasim, M. Rieger, S. Summers, G. Van Onsem, Distance-weighted graph neural networks on FPGAs for real-time particle reconstruction in high energy physics, Frontiers in big Data 3 (2021) 598927, publisher: Frontiers Media SA. URLhttps://www.frontiersin.org/articles/10.3389/fdata...

    work page doi:10.3389/fdata.2020.598927/full 2021

  20. [20]

    Brignone, Making acceleration more amenable with novel high-level synthesis techniques for FPGAs, PhD Thesis, Polytechnic University of Turin, Italy (2025)

    G. Brignone, Making acceleration more amenable with novel high-level synthesis techniques for FPGAs, PhD Thesis, Polytechnic University of Turin, Italy (2025). URLhttps://hdl.handle.net/11583/2997456

    work page 2025

  21. [21]

    Bosio, G

    R. Bosio, G. Brignone, T. Urso, M. T. Lazarescu, L. Lavagno, P. Pasini, Low-Power Subgraph Isomorphism at the Edge Using FPGAs, IEEE Access 13 (2025) 67127–67135.doi:10.1109/ACCESS.2025.3560405

    work page doi:10.1109/access.2025.3560405 2025

  22. [22]

    S. H. Hozhabr, R. Giorgi, A Survey on Real-Time Object Detection on FPGAs, IEEE Access 13 (2025) 38195–38238.doi:10.1109/ACCESS. 2025.3544515

    work page doi:10.1109/access 2025

  23. [23]

    Lahti, T

    S. Lahti, T. D. HÃd’mÃd’lÃd’inen, High-Level Synthesis for FPGAs - A Hardware Engineer’s Perspective, IEEE Access 13 (2025) 28574– 21 28593.doi:10.1109/ACCESS.2025.3540320

    work page doi:10.1109/access.2025.3540320 2025

  24. [24]

    J. A. Jones, Development of Trigger and Control Systems for CMS, Ph.D. thesis, Imperial Coll., London (2007). URLhttps://cds.cern.ch/record/1037625

    work page arXiv 2007

  25. [25]

    Olsen, G

    B. Olsen, G. Petterson, W. Schneider, A method for integration of particle trajectories in an experimentally determined magnetic field, Nuclear Instruments and Methods 41 (2) (1966) 325–330, publisher: Elsevier. URLhttps://www.sciencedirect.com/science/article/pii/0029554X6690019X

    work page arXiv 1966

  26. [26]

    P. Das, C. M. S. Collaboration, An overview of the trigger system at the CMS experiment, Physica Scripta 97 (5) (2022) 054008, publisher: IOP Publishing. URLhttps://iopscience.iop.org/article/10.1088/1402-4896/ac6302/meta

    work page doi:10.1088/1402-4896/ac6302/meta 2022

  27. [27]

    High Luminosity Large Hadron Collider HL-LHC

    G. Apollinari, O. Bruening, T. Nakamoto, L. Rossi, High Luminosity Large Hadron Collider HL-LHC, arXiv:1705.08830 [physics] (2015). doi:10.5170/CERN-2015-005.1. URLhttp://arxiv.org/abs/1705.08830

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.5170/cern-2015-005.1 2015

  28. [28]

    R. Liu, P. Calafiura, S. Farrell, X. Ju, D. T. Murnane, T. M. Pham, Hierarchical graph neural networks for particle track reconstruction, CoRR abs/2303.01640 (2023).arXiv:2303.01640,doi:10.48550/ARXIV.2303.01640. URLhttps://doi.org/10.48550/arXiv.2303.01640

    work page doi:10.48550/arxiv.2303.01640 2023

  29. [29]

    Odyurt, N

    U. Odyurt, N. Dobreva, Z. Wolffs, Y . Zhao, A. Ferrer-Sánchez, R. R. de Austri Bazan, J. D. Martín-Guerrero, A. L. Varbanescu, S. Caron, Novel approaches for ml-assisted particle track reconstruction and hit clustering, CoRR abs/2405.17325 (2024). arXiv:2405.17325, doi:10.48550/ARXIV.2405.17325. URLhttps://doi.org/10.48550/arXiv.2405.17325

    work page doi:10.48550/arxiv.2405.17325 2024

  30. [30]

    F. Ma, F. Liu, W. Li, Jet tagging algorithm of graph network with Haar pooling message passing, Physical Review D 108 (7) (2023) 072007. doi:10.1103/PhysRevD.108.072007. URLhttps://link.aps.org/doi/10.1103/PhysRevD.108.072007

    work page doi:10.1103/physrevd.108.072007 2023

  31. [31]

    Aarrestad, V

    T. Aarrestad, V . Loncar, N. Ghielmetti, M. Pierini, S. Summers, J. Ngadiuba, C. Petersson, H. Linander, Y . Iiyama, G. Di Guglielmo, Fast convolutional neural networks on FPGAs with hls4ml, Machine Learning: Science and Technology 2 (4) (2021) 045015, publisher: IOP Publishing. URLhttps://iopscience.iop.org/article/10.1088/2632-2153/ac0ea1/meta

    work page doi:10.1088/2632-2153/ac0ea1/meta 2021

  32. [32]

    and Di Guglielmo, Giuseppe and Duarte, Javier and Loncar, Vladimir , title =

    B. Hawks, D. Plotnikov, N. Tran, K. Tame-Narvaez, M. M. Rahimifar, H. E. Rahali, A. C. Therrien, G. D. Guglielmo, J. Duarte, V . Loncar, wa-hls4ml and lui-gnn: A Benchmark and GNN based Surrogate Model for hls4ml Resource and Latency Estimation, in: A. Putnam, J. Li (Eds.), Proceedings of the 2025 ACM/SIGDA International Symposium on Field Programmable Ga...

    work page doi:10.1145/3706628.3708827 2025

  33. [33]

    G. D. Guglielmo, B. Du, J. Campos, A. Boltasseva, A. V . Dixit, F. Fahim, Z. Kudyshev, S. Lopez, R. Ma, G. N. Perdue, N. Tran, O. Yesilyurt, D. Bowring, End-to-end workflow for machine learning-based qubit readout with QICK and hls4ml, CoRR abs/2501.14663, arXiv: 2501.14663 (2025).doi:10.48550/ARXIV.2501.14663. URLhttps://doi.org/10.48550/arXiv.2501.14663

    work page doi:10.48550/arxiv.2501.14663 2025

  34. [34]

    Y . Lu, X. Qi, Y . Li, M. Lai, Y . Zhao, Z. Li, H. Li, Q. Wang, Automatic Implementation of Large-Scale CNNs on FPGA Cluster Based on HLS4ML, in: IEEE International Symposium on Parallel and Distributed Processing with Applications, ISPA 2024, Kaifeng, China, October 30 - Nov. 2, 2024, IEEE, 2024, pp. 1080–1087.doi:10.1109/ISPA63168.2024.00143

    work page doi:10.1109/ispa63168.2024.00143 2024

  35. [35]

    Z. Que, H. Fan, M. Loo, H. Li, M. Blott, M. Pierini, A. D. Tapper, W. Luk, LL-GNN: Low Latency Graph Neural Networks on FPGAs for High Energy Physics, ACM Trans. Embed. Comput. Syst. 23 (2) (2024) 17:1–17:28.doi:10.1145/3640464

    work page doi:10.1145/3640464 2024

  36. [36]

    G.-W. Kim, D. Kim, J. Moon, H. Liu, T. Khan, A. Iyer, D. Kim, A. Akella, OMEGA: A Low-Latency GNN Serving System for Large Graphs, CoRR abs/2501.08547, arXiv: 2501.08547 (2025).doi:10.48550/ARXIV.2501.08547. URLhttps://doi.org/10.48550/arXiv.2501.08547

    work page doi:10.48550/arxiv.2501.08547 2025

  37. [37]

    Zhang, H

    B. Zhang, H. Zeng, V . K. Prasanna, GraphAGILE: An FPGA-Based Overlay Accelerator for Low-Latency GNN Inference, IEEE Trans. 22 Parallel Distributed Syst. 34 (9) (2023) 2580–2597.doi:10.1109/TPDS.2023.3287883

    work page doi:10.1109/tpds.2023.3287883 2023

  38. [38]

    X. Liu, J. Chen, Q. Wen, A Survey on Graph Classification and Link Prediction based on GNN, arXiv:2307.00865 [cs] (Jul. 2023). doi:10.48550/arXiv.2307.00865. URLhttp://arxiv.org/abs/2307.00865

    work page doi:10.48550/arxiv.2307.00865 2023

  39. [39]

    T. Chen, S. Bian, Y . Sun, Are Powerful Graph Neural Nets Necessary? A Dissection on Graph Classification, arXiv:1905.04579 [cs] (Jun. 2020).doi:10.48550/arXiv.1905.04579. URLhttp://arxiv.org/abs/1905.04579

    work page doi:10.48550/arxiv.1905.04579 1905

  40. [40]

    Y . Zhu, L. Tong, G. Li, X. Luo, K. Zhou, Focusedcleaner: Sanitizing poisoned graphs for robust gnn-based node classification, IEEE Transactions on Knowledge and Data Engineering 36 (6) (2023) 2476–2489, publisher: IEEE. URLhttps://ieeexplore.ieee.org/abstract/document/10285406/

    work page arXiv 2023

  41. [41]

    Y . Xie, Y . Liang, M. Gong, A. K. Qin, Y .-S. Ong, T. He, Semisupervised graph neural networks for graph classification, IEEE Transactions on Cybernetics 53 (10) (2022) 6222–6235, publisher: IEEE. URLhttps://ieeexplore.ieee.org/abstract/document/9764654/

    work page arXiv 2022

  42. [42]

    A comprehensive guide to the physics and usage of PYTHIA 8.3

    C. Bierlich, S. Chakraborty, N. Desai, L. Gellersen, I. Helenius, P. Ilten, L. LÃ˝ unnblad, S. Mrenna, S. Prestel, C. T. Preuss, T. SjÃ˝ ustrand, P. Skands, M. Utheim, R. Verheyen, A comprehensive guide to the physics and usage of pythia 8.3 (2022).arXiv:2203.11601. URLhttps://arxiv.org/abs/2203.11601

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  43. [43]

    Michon, C

    A. Michon, C. Poulliat, A. M. Cipriano, Graph neural networks versus gated recurrent units only for approximate bayesian MU-MIMO detectors, in: 25th IEEE International Workshop on Signal Processing Advances in Wireless Communications, SPAWC 2024, Lucca, Italy, September 10-13, 2024, IEEE, 2024, pp. 561–565.doi:10.1109/SPAWC60668.2024.10694358. URLhttps://...

    work page doi:10.1109/spawc60668.2024.10694358 2024

  44. [44]

    W. L. Hamilton, R. Ying, J. Leskovec, Inductive representation learning on large graphs, in: Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, Curran Associates Inc., Red Hook, NY , USA, 2017, pp. 1025–1035

    work page 2017

  45. [45]

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, J. M. Solomon, Dynamic Graph CNN for Learning on Point Clouds, ACM Transactions on Graphics 38 (5) (Oct. 2019).doi:10.1145/3326362. URLhttps://doi.org/10.1145/3326362

    work page doi:10.1145/3326362 2019

  46. [46]

    S. Ã. Arik, T. Pfister, TabNet: Attentive Interpretable Tabular Learning, Proceedings of the AAAI Conference on Artificial Intelligence 35 (8) (2021) 6679–6687.doi:10.1609/aaai.v35i8.16826. URLhttps://ojs.aaai.org/index.php/AAAI/article/view/16826

    work page doi:10.1609/aaai.v35i8.16826 2021

  47. [47]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, PyTorch: an imperative style, high-performance deep learning library, Curran Associates Inc., Red Hook, NY , USA, 2019

    work page 2019

  48. [48]

    M. Fey, J. E. Lenssen, Fast graph representation learning with pytorch geometric (2019).arXiv:1903.02428. URLhttps://arxiv.org/abs/1903.02428

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  49. [49]

    Zheng, M

    D. Zheng, M. Wang, Q. Gan, Z. Zhang, G. Karypis, Learning graph neural networks with deep graph library, in: Companion Proceedings of the Web Conference 2020, WWW ’20, Association for Computing Machinery, New York, NY , USA, 2020, pp. 305–306.doi:10.1145/ 3366424.3383111. URLhttps://doi.org/10.1145/3366424.3383111

    work page doi:10.1145/3366424.3383111 2020

  50. [50]

    Rozemberczki, P

    B. Rozemberczki, P. Scherer, Y . He, G. Panagopoulos, A. Riedel, M. Astefanoaei, O. Kiss, F. Beres, G. López, N. Collignon, R. Sarkar, Pytorch geometric temporal: Spatiotemporal signal processing with neural machine learning models, in: Proceedings of the 30th ACM International Conference on Information & Knowledge Management, CIKM ’21, Association for Co...

    work page doi:10.1145/3459637.3482014 2021

  51. [51]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg, J. Vanderplas, 23 A. Passos, D. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, Scikit-learn: Machine learning in python, J. Mach. Learn. Res. 12 (null) (2011) 2825–2830

    work page 2011

  52. [52]

    J. Xu, Z. Li, B. Du, M. Zhang, J. Liu, Reluplex made more practical: Leaky relu, in: 2020 IEEE Symposium on Computers and Communications (ISCC), 2020, pp. 1–7.doi:10.1109/ISCC50000.2020.9219587

    work page doi:10.1109/iscc50000.2020.9219587 2020

  53. [53]

    Feng, A new deep-neural-network–based missing transverse momentum estimator, and its application to w recoil, Ph.D

    Y . Feng, A new deep-neural-network–based missing transverse momentum estimator, and its application to w recoil, Ph.D. thesis, University of Maryland, College Park (May 2020). URLhttps://cds.cern.ch/record/2744871?ln=en 24

    work page arXiv 2020