Pith. sign in

REVIEW 4 major objections 4 minor 41 references

Search for $t\bar tt\bar tW$ Production at $\sqrt{s} = 13$ TeV Using a Modified Graph Neural Network at the LHC

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a physics-informed, quantum-enhanced graph neural network markedly outperforms boosted decision trees and XGBoost in separating ultra-rare ttttW events from huge backgrounds, reporting a significance of 0.174 and…

desk verdict The reported 0.174 significance is arithmetically incompatible with the paper's own yields, so the central claim fails; only the architecture sketch is salvageable. read the letter →

arxiv 2507.23723 v1 pith:2G6ZAQPB submitted 2025-07-31 hep-ex hep-ph

Syed Haider Ali , Ashfaq Ahmad , Muhammad Saiel , Nadeem Shaukat This is my paper · ORCID
classification hep-exhep-ph
keywords four-topproductionttttWgraphneuralnetworkquantummachinelearningphysics-informedlosseventclassificationLHCfullyhadronicchannel
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that a modified hybrid graph neural network—combining graph convolution layers, a six-qubit quantum encoder, and a physics-informed jet-multiplicity loss—can separate the ultra-rare Standard Model process ttttW from its much larger backgrounds in the fully hadronic decay channel. Using Monte Carlo events reweighted to 350/fb, the authors report a signal significance of 0.174 and ROC-AUC of 0.974, compared with 0.148 and 0.913 for a boosted decision tree and 0.149 and 0.920 for XGBoost. If the comparison holds, this would make the network a more sensitive event-selection tool for a channel that is uniquely sensitive to four-fermion SMEFT operators. The paper also argues, through an ablation study, that the physics-informed loss and the quantum branch each add to the performance.

What carries the argument

The load-bearing object is the modified GNN itself: a dual-stream architecture in which node-level jet kinematics (pT, eta, phi) are processed by two GINEConv layers, while event-wide observables are processed in parallel by an MLP and by a six-qubit quantum circuit with RY angle encoding, trainable RZ phases, and nearest-neighbour CNOT entanglement. A cross-attention block fuses the local graph embedding with the two global embeddings, and the output is a two-class log-softmax. The Jet Multiplicity Loss is the physics-informed regularizer: it penalises predicted jet and b-jet counts that deviate from the expected hadronic decay pattern for signal versus background, using 1-$\sigma$ tolerances measured from the simulated samples.

What would settle it

Recompute Z = S/sqrt(S+B) from the paper's own stated expectations: S = 2.32 signal events and B ≈ 160 million background events (Section II.C) gives Z ≈ 0.00018, not 0.174; one can settle the claim by checking whether the quoted significance comes from the reweighted yields or from the balanced test-set counts in the confusion matrix.

Watch

Extended reading notes

Core claim

The paper's central claim is that a physics-informed, quantum-enhanced graph neural network achieves better signal-background discrimination for fully hadronic ttttW production at 13 TeV than conventional multivariate classifiers. The authors report that the GNN reaches a significance of 0.174, a recall of 0.957, and a ROC-AUC of 0.974, exceeding the BDT (0.148, 0.914, 0.913) and XGBoost (0.149, 0.924, 0.920). They attribute the gain to three components working together: GINEConv graph layers that capture local jet correlations, a six-qubit variational circuit that encodes global event observables in an entangled quantum feature space, and a cross-attention fusion that weights the quantum stream more heavily, with about 60% of the attention. The physics-informed Jet Multiplicity Loss regularises the model toward the expected jet counts (roughly 4 b-jets and 10 light jets for signal, fewer for backgrounds).

Load-bearing premise

The reported numbers all rest on the assumption that the LO Monte Carlo samples, processed through Pythia8 and Delphes and reweighted by cross-section, faithfully reproduce the real jet multiplicities and kinematics of the fully hadronic ttttW final state and its backgrounds, and that the significance was computed from the luminosity-normalised yields.

Editorial extensions

If this is right

  • If the reported classification performance holds, the GNN would become a viable event-selection step for ttttW searches in the fully hadronic channel, where the final state has roughly 4 b-jets and 10 light jets.
  • A 17.6% relative improvement in significance over BDT (0.174 versus 0.148) would directly translate into increased sensitivity to dimension-six four-fermion operators, since ttttW production is especially sensitive to qqtt contact interactions.
  • The ablation study's stepwise gains (base GNN 0.926, with physics loss 0.933, with global MLP 0.951, with quantum circuit 0.957 ROC-AUC) imply that each architectural component contributes positively and could be transferred to other rare-event searches.
  • The 60/40 attention split toward the quantum token, if robust, suggests that quantum feature maps can provide complementary information to classical global features in collider event classification.

Reading between the lines

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

  • Using the paper's own expected yields (S ≈ 2.32 events, B ≈ 1.6 × 10^8 events at 350/fb), the formula Z = S/sqrt(S+B) gives Z ≈ 0.00018, about three orders of magnitude below the quoted 0.174; this suggests the reported significance was not computed from the full-luminosity yields.
  • The confusion-matrix counts (TP = 1915, FP = 193) appear to come from an equal-sized test sample rather than the reweighted physics yields, so the ROC-AUC and recall numbers are classifier-performance metrics on a balanced set and are not directly equivalent to an LHC search significance.
  • A real search would also need to account for systematic uncertainties, trigger efficiencies, and the mistag rate of light jets as b-jets; the fast Delphes simulation cannot capture these detector effects, so the true sensitivity at CMS may differ.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The manuscript reports a search for fully hadronic ttttW production at sqrt(s)=13 TeV using 350/fb of Delphes-simulated Monte Carlo events, and introduces a hybrid graph neural network that combines GINEConv layers, a six-qubit quantum encoder, cross-attention fusion, and a 'physics-informed' jet-multiplicity loss. The authors claim the GNN achieves a signal significance S/sqrt(S+B) of 0.174 and ROC-AUC of 0.974, outperforming BDT (0.148, 0.913) and XGBoost (0.149, 0.920). The paper also presents an ablation study of the architecture components and an attention analysis intended to show that the quantum branch receives about 60% of the attention weight.

Significance. The topic is timely: four-top associated production is a rare Standard Model process with potential sensitivity to new physics, and ML classifiers are standard tools for such searches. The paper is clearly written in places, and the stepwise ablation of model components (Table IV) is a reasonable study design. However, the central numerical claim is contradicted by the paper's own equations: with the stated yields of 2.32 signal and ~1.6e8 background events, Eq. (1) gives Z=1.8e-4, not 0.174, and the confusion-matrix efficiencies do not rescue the number. The cross-attention equation is also structurally degenerate. Because the headline significance, the claimed 17.6% improvement over BDT, and the attention-interpretation all rest on these unsupported numbers, the manuscript does not currently establish its main conclusions. There is no code or data release, and no systematic uncertainties are incorporated, so the reported performance cannot be independently checked or interpreted as an LHC search sensitivity.

major comments (4)
  1. [Section II.C and Section IV.B, Eq. (1) and Fig. 7] The reported significance of 0.174 cannot be reproduced from the paper's own yields and efficiencies. With S=2.32 and B=1.6e8 in Eq. (1), Z=2.32/sqrt(2.32+1.6e8)=1.8e-4. Using the confusion matrix in Fig. 7, e_s=1915/2000=0.957 and e_b=193/1000=0.193, so the selected yields are S=2.22 and B=3.1e7, giving Z~4e-4. Reaching Z=0.174 with S=2.22 would require B~161 events, i.e., a background efficiency of order 1e-6 at 0.957 signal efficiency, which is inconsistent with FP=193 on a 1000-event background validation set. The retained background yield after the optimal cut is never stated. The abstract and Table V therefore present a significance that is incompatible with the paper's own formulas and figures.
  2. [Section III.B.2, Eq. (12)] The cross-attention mechanism is degenerate as written. G and Q are single vectors in R^hidden_dim, so the softmax is taken over a one-element set and is identically 1; hence A = W_V L, independent of the query and key. This contradicts the text's claim that the mechanism 'weighs' the local particle interactions against global event properties, and it is inconsistent with the stated use of 8 attention heads, which requires multiple query-key pairs. The attention-score interpretation in Fig. 12 is therefore not supported by the model equation. The formula needs to be rewritten with genuine multi-token or multi-head attention before the architecture claims can be evaluated.
  3. [Section IV.A and IV.B, Fig. 6] The reported significance is obtained by scanning a cut on the GNN output and taking the maximum on what appears to be the same sample used for evaluation. No train/validation/test split is described, and the paper does not state the number of events per process used for the confusion matrices in Figs. 7 and 9. Selecting the operating point on the evaluation sample makes the quoted 0.174 and the claimed 17.6% improvement over BDT optimistically biased. A valid comparison requires a pre-specified threshold or an independent test set, with the operating point chosen on a validation sample only.
  4. [Section III.C.1 and Section IV.A] The Jet Multiplicity Loss is presented as a physics-informed constraint, but its parameters are fitted to the evaluation data. The tolerances toll and tolb are set to one standard deviation of the jet-count distributions in the same simulated samples used to evaluate the classifier, and the loss weights are dynamically tuned to target precision/recall values of 0.85/0.95 chosen by the authors. The ablation gains in Table IV therefore conflate physics regularization with hyperparameter tuning on the test distribution. The tolerances should be derived from independent principles or treated as hyperparameters and validated on a held-out set before any claim of a physics-informed improvement can be made.
minor comments (4)
  1. [Section V vs Table V] The conclusion lists XGBoost's ROC-AUC as 0.913, while Table V and the abstract give 0.920; Section IV.C also mentions 0.96 vs 0.91 for the ROC comparison. Please reconcile all reported performance numbers.
  2. [Section II.C and Table II] The text says the CSV file has 21 columns in total, but Table II lists 23 variables including the target. Please clarify the exact feature count and the mapping to the table.
  3. [Equations (15)-(16)] These equations contain stray commas and unmatched parentheses (e.g., 'if ytrue = 1, ,'), and the surrounding text uses inconsistent notation for the signal process (tbar t tbar t W vs ttttW-). Please proofread the formulas and notation.
  4. [Section III.B.2 and Fig. 12] Once Eq. (12) is corrected, Fig. 12 and the related text on attention scores need to be revised, because with the current equation there is no meaningful distribution of attention scores between the classical and quantum tokens.

Circularity Check

3 steps flagged · score 6.0 of 10

Reported GNN significance of 0.174 is a test-set-optimized maximum, not the output of Eq. (1) with the paper's own yields; physics-informed loss tolerances and cross-section inputs are also self-referential.

  1. fitted input called prediction [Section IV.B and Section II.C, Eq. (1)]
    "The significance, defined as S/√(S+B), where S and B are the signal and background yields, respectively, was optimized by determining an optimal cut on the GNN output."

    The paper's Eq. (1) defines significance as Z = S/√(S+B), with S and B the reweighted yields. Section II.C states S≈2.32 and B≈1.6e8, but Section IV.B reports Z=0.174 after optimizing a cut on the GNN output on the same test set used for the confusion matrix. The cut is a fitted parameter, so the reported significance is the maximum of a scan rather than a pre-specified prediction. Using the paper's own yields and confusion-matrix efficiencies (ε_s=0.957, ε_b=0.193) in Eq. (1) gives Z≈4e-4, not 0.174, so the headline number is not the output of the paper's stated derivation chain.

  2. self definitional [Section III.C.1, Jet Multiplicity Loss]
    "These tolerance values (toll, tolb) are set to one standard deviation of the jet-count distributions observed in our simulated samples."

    The Jet Multiplicity Loss is introduced as a physics-informed constraint 'derived from event decay dynamics,' but its tolerances are calibrated on the same simulated samples used for training and for the reported evaluation. The constraint is therefore defined by the data it constrains: the loss penalizes deviations from the MC sample's own statistical spread, making the 'physics' term a self-referential regularizer rather than an independently derived physical law. Any performance gain attributed to this loss reflects consistency with the evaluation sample's statistics, not an external physics input.

1 more flagged steps
  1. self citation load bearing [Section II.A, Table I, and reference [1]]
    "The t¯tt¯tW− signal is generated at LO using MadGraph5 aMC@NLO v2.9.15 [28]. The NLO calculation of 6.6+2.4−2.6 ab was used to normalize the simulation [1]."

    The normalization of every yield and significance depends on the signal and background cross-sections, which the paper cites only to [1], an unpublished Master's thesis by co-author M. Saiel. The expected signal yield of 2.32 events is σ×L with σ=6.6 ab from [1], and the background yield of about 1.6e8 events sums the Table I cross-sections also from [1]. The central numerical claims therefore reduce to a self-citation chain: the load-bearing physics inputs are not externally verified, machine-checked, or otherwise independent values, but come from the authors' own unpublished work. This does not invalidate the ML comparison itself, but makes the quoted cross-sections and expected yields unverified, author-supplied inputs.

full rationale

The most significant circular element is the headline significance. Equation (1) defines Z = S/√(S+B) in terms of reweighted yields, and Section II.C states S≈2.32 and B≈1.6e8. Section IV.B then reports Z=0.174 after 'determining an optimal cut on the GNN output' on the same test set that yields the confusion matrix. Because the cut is chosen to maximize the significance on that same test set, the reported value is a fitted maximum, not a prediction from Eq. (1). Plugging the paper's own yields and confusion-matrix efficiencies into Eq. (1) gives Z≈4e-4, so the stated 0.174 is not reproducible from the paper's stated inputs. This is a clear case of a fitted input being presented as a performance prediction. Additional circularity arises from the Jet Multiplicity Loss: its tolerances are set to one standard deviation of the same simulated samples used for training and evaluation, so the 'physics-informed' regularization is defined by the data it constrains. Finally, the normalization of all yields rests on cross-sections cited only to [1], an unpublished Master's thesis by co-author M. Saiel; the numerical claims therefore depend on a self-citation chain for their load-bearing inputs. The architectural components and ROC curves are direct empirical outputs with genuine ML content, so the paper is not entirely circular, but the central significance claim partially reduces to test-set optimization and self-cited inputs, warranting a score of 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central performance claim depends on Monte-Carlo-derived jet-count tolerances, a test-set-fitted decision threshold, and cross-sections from an author's thesis rather than a peer-reviewed source. These inputs are not independently benchmarked against real data, and no new physical entities are introduced.

free parameters (3)
  • Decision threshold on GNN output = 0.51
    The optimal cut is found by maximizing S/sqrt(S+B) on the test set (Section IV.B). The reported significance is the maximum over this cut, so it is a fitted quantity rather than a pre-specified selection.
  • Jet multiplicity loss tolerances (toll, tolb) = signal: 1.668, 1.147; background: 1.8, 1.2
    Set to one standard deviation of the jet-count distributions in the same simulated samples used for training and evaluation (Section III.C.1). These Monte-Carlo-derived values control the physics-informed loss.
  • Physics-informed loss dynamic weighting targets = precision 0.85, recall 0.95, weight range 0.05 to 0.3, EMA decay 0.9 and 0.99
    Hand-chosen training hyperparameters (Section IV.A). They affect the final classifier and the reported metrics but are not the main claim.
assumptions (4)
  • domain assumption The NLO cross-section for ttttW and the LO cross-sections for backgrounds, quoted from [1], are correct.
    All expected yields and the reported significance are normalized using these cross-sections (Section II.C). [1] is an unpublished Master's thesis by one of the authors.
  • domain assumption MadGraph5, Pythia8, and Delphes provide a faithful model of signal and background kinematics and jet multiplicities for the fully hadronic channel.
    All training, validation, and performance claims rest on these simulated samples (Section II.A).
  • domain assumption Fully hadronic ttttW decays produce approximately 4 b-jets and 10 light jets, while backgrounds produce fewer jets, as given in Table III.
    This is the physical basis of the Jet Multiplicity Loss and of the expected jet counts used in training (Section III.C.1).
  • standard math Standard automatic differentiation and gradient-based optimization of the hybrid quantum-classical model is valid.
    Used throughout Section IV.A to train the GNN; no formal verification is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Search for $t\bar tt\bar tW$ Production at $\sqrt{s} = 13$ TeV Using a Modified Graph Neural Network at the LHC." pith.science (2026). https://pith.science/paper/2G6ZAQPB

@misc{pith2026250723723,
  author       = {Pith},
  title        = {Pith review of: Search for $t\bar tt\bar tW$ Production at $\sqrts = 13$ TeV Using a Modified Graph Neural Network at the LHC},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2G6ZAQPB}},
  note         = {Machine review of arXiv:2507.23723}
}
abstract

The simultaneous production of four top quarks in association with a ($W$) boson at $(\sqrt{s} = 13)$ TeV is an rare SM process with a next-to-leading-order (NLO) cross-section of $(6.6^{+2.4}_{-2.6} {ab})$\cite{saiel}. Identifying this process in the fully hadronic decay channel is particularly challenging due to overwhelming backgrounds from $t\bar{t}, t\bar{t}W, t\bar{t}Z$, and triple-top production processes. This study introduces a modified physics informed Neural Network, a hybrid graph neural network (GNN) enhancing event classification. The proposed model integrates Graph layers for particle-level features, a custom Multi Layer Perceptron(MLP) based global stream with a quantum circuit and cross-attention fusion to combine local and global representations. Physics-informed Loss function enforce jet multiplicity constraints, derived from event decay dynamics. Benchmarked against conventional methods, the GNN achieves a signal significance $(S/\sqrt{S+B})$ of $0.174$ and ROC-AUC of 0.974, surpassing BDT's significance of $0.148$ and ROC of $0.913$, while Xgboost achieves a significance of $0.149$ and ROC of $0.920$. The classification models are trained on Monte Carlo (MC) simulations, with events normalized using cross-section-based reweighting to reflect their expected contributions in a dataset corresponding to $350\;$fb$^{-1}$ of integrated luminosity. This enhanced approach offers a framework for precision event selection at the LHC, leveraging high dimensional statistical learning and physics informed inference to tackle fundamental HEP challenges, aligning with ML developments.

Figures

Figures reproduced from arXiv: 2507.23723 by the authors.

Figure 1
Figure 1. FIG. 1. Jet Multiplicities of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Schematic Diagram for data conversion from [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. FIG. 4. Quantum encoder circuit: each classical feature [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: FIG. 5. The ROC AUC curves comparison of the GNN [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Cut efficiency and optimal cut for GNN, [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 9
Figure 9. Figure 9: FIG. 9. Confusion matrix for BDT, showing true labels [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 8
Figure 8. Figure 8: FIG. 8. GNN output distributions for signal and back [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 10
Figure 10. Figure 10: FIG. 10. BDT output distributions for signal and back [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: FIG. 11. Background rejection versus signal efficiency for [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: FIG. 12. Histogram showing the distribution of atten [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 26 canonical work pages

  1. [1]

    Figure 2 illustrates this graph representation, where nodes are categorized into jet nodes (blue) and b-jet nodes (orange), with edges 6 FIG

    Representing Tabular Data as Graphs Converting tabular particle physics data into a graph format is a crucial step before we begin the training for the GNN. Figure 2 illustrates this graph representation, where nodes are categorized into jet nodes (blue) and b-jet nodes (orange), with edges 6 FIG. 2. Schematic Diagram for data conversion from tabular to g...

    work page

  2. [2]

    Architecture Design for the modified-GNN The architecture of the modified Graph Neural Network (GNN), depicted in Figure 3, is designed to exploit the graph representation of t¯tt¯tW − events by integrating classical graph neural network tech- niques with quantum computing principles, captur- ing both local particle interactions and global event propertie...

    work page

  3. [3]

    Jet Multiplicity Loss The Jet Multiplicity Loss enforces consistency be- tween the predicted jet multiplicity from the Monte Carlo (MC) simulations and expected jet multiplic- ities from the fully hadronic decay of t¯tt¯tW − events compared to background processes. In the fully hadronic channel, the signal is characterized by a high jet multiplicity due t...

    work page 1915

  4. [4]

    CMS and Atlas Collaborations, Combination of measurements of the top quark mass from data col- lected by the Atlas and CMS experiments at√s = 7 and 8 TeV, arXiv preprint arXiv:2402.08713 (2024). 13

    arXiv 2024

  5. [5]

    Saiel, Feasibility study of four top quarks in asso- ciation with Gauge Boson at CMS , Master’s thesis, Quaid-I-Azam university (2024)

    M. Saiel, Feasibility study of four top quarks in asso- ciation with Gauge Boson at CMS , Master’s thesis, Quaid-I-Azam university (2024)

    work page 2024

  6. [6]

    Abachi et al

    S. Abachi et al. (D0), Observation of the top quark, Phys. Rev. Lett. 74, 2632 (1995), arXiv:hep- ex/9503003

    work page arXiv 1995

  7. [7]

    Abe et al

    F. Abe et al. (CDF), Observation of top quark pro- duction in ¯pp collisions, Phys. Rev. Lett. 74, 2626 (1995), arXiv:hep-ex/9503002

    arXiv 1995

  8. [8]

    Tumasyan, et al., CMS Collaboration, Observa- tion of four top quark production in proton-proton collisions at √s = 13 TeV, Phys

    A. Tumasyan, et al., CMS Collaboration, Observa- tion of four top quark production in proton-proton collisions at √s = 13 TeV, Phys. Lett. B847 (2023)

    work page 2023

Show all 41 references
  1. [9]

    Aaltonen et al

    T. Aaltonen et al. (CDF), First Observation of Elec- troweak Single Top Quark Production, Phys. Rev. Lett. 103, 092002 (2009), arXiv:0903.0885 [hep-ex]

    2009 arXiv

  2. [10]

    V. M. Abazov et al. (D0), Observation of Single Top Quark Production, Phys. Rev. Lett. 103, 092001 (2009), arXiv:0903.0850 [hep-ex]

    2009 arXiv

  3. [11]

    van Beekveld, A

    M. van Beekveld, A. Kulesza, and L. M. Valero, Threshold resummation for the production of four top quarks at the lhc (2025), arXiv:2212.03259 [hep- ph]

    2025 arXiv

  4. [12]

    Brivio and M

    I. Brivio and M. Trott, The standard model as an effective field theory, Physics Reports 793, 1–98 (2019)

    2019

  5. [13]

    Tumasyan, et al., CMS Collaboration, Evidence for four-top quark production in proton-proton col- lisions at √s = 13 TeV, Phys

    A. Tumasyan, et al., CMS Collaboration, Evidence for four-top quark production in proton-proton col- lisions at √s = 13 TeV, Phys. Lett. B 844, 138076 (2023)

    2023

  6. [14]

    Leonardo, Search and measurement of the b → µµ rare processes with lhc run i data, Nuclear and Particle Physics Proceedings , 1308 (2016)

    N. Leonardo, Search and measurement of the b → µµ rare processes with lhc run i data, Nuclear and Particle Physics Proceedings , 1308 (2016)

    2016

  7. [15]

    Zhang, Constraining qqtt operators from four- top production: a case for enhanced EFT sensitivity, Chinese Physics C 42, 023104 (2018)

    C. Zhang, Constraining qqtt operators from four- top production: a case for enhanced EFT sensitivity, Chinese Physics C 42, 023104 (2018)

    2018

  8. [16]

    Lalchand, Extracting more from boosted deci- sion trees: A high energy physics case study (2020), arXiv:2001.06033 [stat.ML]

    V. Lalchand, Extracting more from boosted deci- sion trees: A high energy physics case study (2020), arXiv:2001.06033 [stat.ML]

    2020 arXiv

  9. [17]

    Grzadkowski, M

    B. Grzadkowski, M. Iskrzy´ nski, M. Misiak, and J. Rosiek, Dimension-six terms in the standard model lagrangian, Journal of High Energy Physics 2010, 10.1007/jhep10(2010)085 (2010)

    2010 doi

  10. [18]

    R. E. Schapire, A brief introduction to boosting, in International Joint Conference on Artificial Intelli- gence (1999)

    1999

  11. [19]

    Chen and C

    T. Chen and C. Guestrin, Xgboost: A scalable tree boosting system, CoRR abs/1603.02754 (2016), 1603.02754

    2016 arXiv

  12. [20]

    Builtjes, S

    L. Builtjes, S. Caron, P. Moskvitina, C. Nellist, R. D. Austri, R. Verheyen, and Z. Zhang, Atten- tion to the strengths of physical interactions: Trans- former and graph-based event classification for par- ticle physics experiments, Semantic Scholar (2022)

    2022

  13. [21]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, Attention is all you need (2023), arXiv:1706.03762 [cs.CL]

    2023 arXiv

  14. [22]

    Fukushima, Neocognitron: A self-organizing neu- ral network model for a mechanism of pattern recog- nition unaffected by shift in position, Biological Cy- bernetics 36, 193 (1980)

    K. Fukushima, Neocognitron: A self-organizing neu- ral network model for a mechanism of pattern recog- nition unaffected by shift in position, Biological Cy- bernetics 36, 193 (1980)

    1980

  15. [23]

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Learning representations by back-propagating er- rors, nature 323, 533 (1986)

    1986

  16. [24]

    Kaltsas, Constrained hamiltonian systems and physics informed neural networks: Hamilton-dirac neural nets, Semantic Scholar (2024)

    D. Kaltsas, Constrained hamiltonian systems and physics informed neural networks: Hamilton-dirac neural nets, Semantic Scholar (2024)

    2024

  17. [25]

    Zhang, S

    H. Zhang, S. Lin, X. Zhang, Y. Wang, Y. Z. C. of Artificial Intelligence, Automation, H. University, F. of Electronic, I. Engineering, X. J.-L. University, S. of Electrical Engineering, S. University, S. of Con- trol, C. Engineering, N. University, S. of Automa- tion, E. Engin...

    2024

  18. [26]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, Physics-informed neural networks: A deep learn- ing framework for solving forward and inverse prob- lems involving nonlinear partial differential equa- tions, Journal of Computational physics 378, 686 (2019)

    2019

  19. [27]

    Farea, O

    A. Farea, O. Yli-Harja, and F. Emmert-Streib, Understanding physics-informed neural networks: Techniques, applications, trends, and challenges, AI 10.3390/ai5030074 (2024)

    2024 doi

  20. [28]

    Alwall, R

    J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer, H.-S. Shao, T. Stelzer, P. Torrielli, and M. Zaro, The automated compu- tation of tree-level and next-to-leading order differ- ential cross sections, and their matching to parton shower simulations, Jou...

    2014

  21. [29]

    W. Hu, B. Liu, J. Gomes, M. Zitnik, P. Liang, V. Pande, and J. Leskovec, Strategies for pre-training graph neural networks (2020), arXiv:1905.12265 [cs.LG]

    2020 arXiv

  22. [30]

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka, How powerful are graph neural networks? (2019), arXiv:1810.00826 [cs.LG]

    2019 arXiv

  23. [31]

    Schuld and N

    M. Schuld and N. Killoran, Quantum machine learn- ing in feature hilbert spaces, Physical review letters 122, 040504 (2019)

    2019

  24. [32]

    S. Ovyn, X. Rouby, and V. Lemaitre, Delphes, a framework for fast simulation of a generic collider experiment (2010), arXiv:0903.2225 [hep-ph]

    2010 arXiv

  25. [33]

    A. M. Sirunyan, A. Tumasyan, W. Adam, F. Am- brogi, T. Bergauer, J. Brandstetter, M. Dragicevic, J. Er¨ o, A. Escalante Del Valle, M. Flechl, et al. , Search for the production of four top quarks in the single-lepton and opposite-sign dilepton final states in proton-proton col...

    2019

  26. [34]

    A. M. Sirunyan, A. Tumasyan, W. Adam, E. Asi- lar, T. Bergauer, J. Brandstetter, E. Brondolin, M. Dragicevic, J. Er¨ o, M. Flechl, et al. , Search for standard model production of four top quarks in proton–proton collisions at √s = 13 TeV, Physics Letters B 772, 336 (2017)

    2017

  27. [35]

    Bierlich, S

    C. Bierlich, S. Chakraborty, N. Desai, L. Gellersen, I. Helenius, P. Ilten, L. L¨ onnblad, S. Mrenna, S. Prestel, C. T. Preuss, T. Sj¨ ostrand, P. Skands, M. Utheim, and R. Verheyen, A comprehensive guide to the physics and usage of pythia 8.3 (2022), arXiv:2203.11601 [hep-ph]

    2022 arXiv

  28. [36]

    V. B. et al, Pennylane: Automatic differentiation of hybrid quantum-classical computations (2022), arXiv:1811.04968 [quant-ph]

    2022 arXiv

  29. [37]

    Brun and F

    R. Brun and F. Rademakers, Root—an object ori- ented data analysis framework, Nuclear instruments and methods in physics research section A: accelera- tors, spectrometers, detectors and associated equip- ment 389, 81 (1997). 14

    1997

  30. [38]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Van- derplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and ´E. Duchesnay, Scikit-learn: Machine learning in python, Journal of machine learn...

    2011

  31. [39]

    Clevert, T

    D.-A. Clevert, T. Unterthiner, and S. Hochreiter, Fast and accurate deep network learning by expo- nential linear units (elus) (2016), arXiv:1511.07289 [cs.LG]

    2016 arXiv

  32. [41]

    D. P. Kingma and J. Ba, Adam: A method for stochastic optimization (2017), arXiv:1412.6980 [cs.LG]

    2017 arXiv

  33. [807]

    This yields a precision of T P/(T P+ F P) = 1915/(1915 + 193) ≈ 0.908 and an accuracy of (T P+ T N)/(T P+ T N+ F P+ F N) = (1915 + 807)/(1915 + 807 + 193 + 85)≈ 0.906. The high re- call of 0.957 ( T P/(T P+ F N) = 1915/(1915 + 85)) underscores the model’s ability to capture mo...

    1915

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.