Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Quantum autoencoders can flag rare collider events in under a microsecond on an FPGA, matching classical anomaly detectors.

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

2026-08-01 10:12 UTC pith:NR6ZHN7Z

load-bearing objection A credible first FPGA-emulation study of QML anomaly detectors for HEP triggers, with reproducible code and data — but the headline performance claim is asserted rather than demonstrated, and the hardware case rests on synthesis estimates. the 4 major comments →

arxiv 2607.20302 v1 pith:NR6ZHN7Z submitted 2026-07-22 cs.LG hep-phphysics.ins-det

Classical Hardware Acceleration of Quantum Autoencoders for Real-Time Anomaly Detection in Collider Experiments

Ivan Ge , Sagar Addepalli , Abhilasha Dave , Julia Gonski This is my paper
classification cs.LG hep-phphysics.ins-det
keywords anomaly detectionquantum autoencoderFPGAtrigger systemshigh energy physicsvariational quantum circuitsfixed-point quantizationreal-time inference
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 tries to establish that quantum machine learning, emulated on classical FPGA hardware, is ready for real-time anomaly detection triggers at particle colliders. It builds two autoencoder models—a fully quantum one and a hybrid quantum-classical one—that learn the background of ordinary collision events and score events by how badly they deviate. On four beyond-Standard-Model signal benchmarks, both match or beat the reference classical sensitivity while using fewer than 100 trainable quantum parameters. After fixed-point quantization and high-level synthesis, the fully quantum model runs in 0.47 microseconds and the hybrid in 6.1 microseconds, both fitting inside one FPGA region, which the authors argue satisfies future trigger constraints. If true, quantum-style models can be deployed in today's classical trigger pipelines, and collider infrastructure becomes quantum-ready without waiting for quantum hardware.

Core claim

The central claim is that variational quantum autoencoders, trained only on background events, can serve as practical trigger-level anomaly detectors. The authors demonstrate this with two architectures: a hybrid VAE whose first encoder layer is a parameterized quantum circuit, and a fully quantum autoencoder that scores events from trash-wire measurements. Both models encode each reconstructed particle onto a single qubit via rotation gates, then entangle particles within small type-based blocks. Trained on four million simulated quark-gluon multijet background events and evaluated on four BSM signals (A-to-4-lepton, leptoquark to b-tau, charged Higgs to tau-nu, and neutral Higgs to tau-tau

What carries the argument

The load-bearing design is the block decomposition of the 19-qubit event register into particle-type sub-blocks—one qubit for missing transverse energy, four for electrons, four for muons, and two independent five-qubit blocks for the ten jets. The jet split is chosen using pairwise quantum mutual information, which recovers the original transverse-momentum ordering and keeps the largest emulated state vector at 2^5 amplitudes instead of 2^10. Each block uses single-qubit rotations for feature encoding followed by nearest-neighbor or ring CNOT entanglers, and the fully quantum model converts trash-wire excitation probabilities into background-normalized z-scores, summing the two largest as t

Load-bearing premise

The load-bearing premise is that the synthesized resource and timing estimates predict the real FPGA's behavior; the paper reports no on-chip test or timing closure at the 5.5 ns clock.

What would settle it

Run the full design on the physical FPGA with the 5.5 ns target clock and measure end-to-end latency; if the design fails timing closure or the measured latency exceeds the stated budgets (0.47 microseconds for the QAE, 6.1 microseconds for the hybrid), the central deployment claim collapses. A separate physics falsifier is to evaluate TPR at 10^-5 on the same four BSM benchmarks against the same classical baselines and check whether the quantum models genuinely remain comparable.

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

If this is right

  • Both models synthesize to latencies (0.47 microseconds for the QAE, 6.1 microseconds for the hybrid) and resource footprints that fit within one FPGA region, so trigger-level timing is plausible if the post-synthesis estimates hold.
  • Fewer than 100 trainable quantum parameters are enough to match reference classical anomaly detectors across four distinct BSM signatures, suggesting quantum autoencoders are parameter-efficient for this task.
  • The FPGA implementation is only the fastest at batch size one, the actual trigger operating point; at larger batch sizes CPU and GPU implementations win, so the advantage is specific to real-time single-event inference.
  • At low fixed-point precision, AUC becomes unreliable because score quantization compresses the background tail, whereas TPR at 10^-5 false-positive rate remains stable for the hybrid model, so deployment validation should rely on tail metrics.
  • The fully quantum model contains no classical trainable parameters, meaning the entire inference path can be expressed as a quantum circuit and classically emulated in one hardware kernel.

Where Pith is reading between the lines

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

  • If the post-synthesis numbers survive place-and-route, the sub-microsecond QAE footprint is small enough that it could be embedded alongside existing classical trigger algorithms on the same FPGA, and the same 1-4-4-5-5 block-splitting strategy could extend to higher jet multiplicities by splitting further, at the cost of losing cross-block quantum correlations.
  • The optimization loop used here—minimize latency under a single-region resource constraint—is not specific to quantum kernels and could be reapplied to classical ML trigger models, potentially improving the broader trigger ML stack.
  • Because the models were trained on simulated background only, a natural test is robustness to changing detector and beam conditions; the background-normalized z-score normalization would need online re-fitting in a real trigger system.

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

4 major / 5 minor

Summary. The manuscript presents two anomaly-detection models for real-time triggering at collider experiments: a hybrid quantum-classical variational autoencoder whose first encoder layer is replaced by block-structured parameterized quantum circuits, and a fully quantum autoencoder that scores events from trash-wire measurements. Both models are trained on public LHC background data (QCD multijet events) and evaluated on four BSM signal benchmarks. The inference kernels are quantized to fixed-point formats and synthesized with HLS for an Alveo U200 FPGA; the authors report resource usage and latencies of 6.1 µs (hybrid VAE) and 0.47 µs (QAE), and claim that both fit within one FPGA SLR. The abstract and conclusions further claim that the models achieve performance comparable to state-of-the-art classical approaches and are consistent with future-collider trigger requirements.

Significance. If the performance and hardware claims held, this would be a useful proof-of-concept: it would demonstrate that small variational quantum circuits can be compiled to classical FPGA logic for low-latency, trigger-level inference, with public data and code supporting reproducibility. The architectural description is reasonably detailed, the training procedure is reproducible from the text, and the use of public datasets and a linked code repository are strengths. However, the central performance claim is currently not backed by printed quantitative comparisons, and the hardware conclusion rests on HLS estimates rather than an implemented, placed-and-routed design. Both issues are load-bearing for the paper's stated motivation.

major comments (4)
  1. [§3.1, Fig. 5; Abstract] The central claim that the models "achieve performance comparable to state-of-the-art classical approaches" is unsupported by any quantitative evidence in the manuscript. The text reports no AUC or TPR@10^-5 values for the proposed models, and Figure 5 contains ROC curves for the hybrid VAE and QAE only, with no classical baseline curves overlaid and no numerical baseline values from Ref. [33] or from the classical VAE of Ref. [25]. Without these numbers, the sentence "Both the AUC and TPR metrics are competitive with or exceeding performance of state-of-the-art classical ML methods [33]" is an assertion rather than a result. Because this comparison is the stated motivation for the hardware study, the authors must add a direct, quantitative baseline comparison, ideally with statistical uncertainties; the public dataset and code make this readily checkable.
  2. [§3.2, Table 2; §3.3] The FPGA-feasibility claim is based on HLS post-synthesis resource and latency estimates, not on an actual placed-and-routed design or on-board measurement. The paper provides no timing-closure report at the 5.5 ns target clock, no bitstream generation, and no measured inference latency on the Alveo U200. HLS estimates can change substantially after place-and-route, especially for resource utilization and achievable clock frequency. Therefore the statement that the designs "after FPGA synthesis ... are consistent with real-time constraints of future collider trigger systems" overstates the evidence. The authors should either run the vendor backend implementation flow and report the resulting timing/resources, or explicitly rephrase the conclusion as "HLS estimates suggest" and state the assumed trigger latency budget against which 6.1 µs and 0.47 µs are judged.
  3. [§2.2.2, Eq. (7)] The QAE is described as containing "no classical trainable parameters," but the anomaly score in Eq. (7) depends on per-trash-wire normalization parameters (μ_i, σ_i), nine fitted pairs calibrated on background data. These are data-derived and directly affect the anomaly score; they should be disclosed as calibration parameters and either included in the parameter count or clearly separated from the quantum trainable parameters. This does not invalidate the physics result, but it is necessary for an accurate description of the model.
  4. [§3.2, Fig. 6] The quantization discussion reports that AUC becomes "artificially inflated" for W≥18 while TPR@10^-5 remains within ±5%, and that the deployed formats in Table 1 use different widths per datatype. However, the fixed-point sweep in Fig. 6 appears to vary a single global width W, and the relation between that sweep and the individually chosen formats in Table 1 is not stated. Please clarify which widths were swept jointly, which were held fixed, and how the final per-datatype formats were selected. The paper should also report the actual AUC and TPR@10^-5 values for the deployed formats so the quantization loss is explicitly quantified.
minor comments (5)
  1. [§3.1] Report exact AUC and TPR@10^-5 values for all four signals in the text or in the figure caption, in addition to the ROC curves. As written, Figure 5 is qualitative and the reader cannot verify the claimed competitiveness.
  2. [§2.1] The sentence "The jet bisection is based on pairwise quantum mutual information [31, 35], which recovers the original pT ordering" is confusing: if the result is the pT ordering, the role of the mutual-information calculation is unclear. Please clarify the relationship or provide the supporting calculation.
  3. [§2.3] The hybrid VAE's reconstruction loss and decoder architecture are not specified. Since the decoder is used only during training, a short description would help reproducibility, including how the quantum feature vector is decoded back to input space.
  4. [§3.2] The description of the agentic AI optimization is too vague to be reproducible. Please specify the optimization objective, search space, and stopping criteria, or state that the workflow is a tool-specific detail.
  5. [§3.3, Fig. 7] The CPU/GPU/FPGA latency comparison is qualitative in the text. Please state the actual per-event latencies at the trigger operating point (batch size 1) for at least the FPGA and the fastest classical implementation.

Circularity Check

0 steps flagged

No significant circularity: the anomaly-score evaluations are held-out tests after background-only training, and the unsupported classical-baseline comparison is an evidence gap, not a circular reduction.

full rationale

The central derivation chain is self-contained with respect to circularity. The hybrid VAE anomaly score (Eq. 6) is the CKL score on the latent mean, and the QAE score (Eq. 7) is the sum of the two largest background-normalized trash-wire z-scores; both are model outputs defined after training, not inputs used to define the signal labels. Both models are trained on background-only events, with held-out background for testing and the four BSM signal files used only at evaluation time, so the reported ROC/AUC/TPR values are not forced by construction. The z-score normalization uses fitted (mu, sigma) values, but this is a disclosed calibration of the output distribution, not a fitted parameter that is then renamed as a prediction of signal presence. The quantization sweep compares fixed-point implementations against the 32-bit baseline and reports TPR@10^-5 changes; this is a sensitivity analysis, not a circular restatement. The only self-citation is Ref. [31] for the jet mutual-information ordering that reduces to pT ordering; this is a peripheral architecture choice and does not carry the paper's central feasibility or performance claims. No uniqueness theorem or ansatz is imported from the authors' prior work in a load-bearing way. The main weakness is that the claim of being 'competitive with or exceeding state-of-the-art classical ML methods [33]' is asserted without showing classical ROC curves or AUC/TPR numbers; however, that is an unsupported-evidence or falsifiability problem, not a case where a result reduces to its own inputs by definition or by self-citation.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claims rest on trained model parameters, hand-chosen architecture and hyperparameters, synthesis-tool fidelity, and the validity of the external LHC dataset. No new particles, forces, or physical entities are introduced; the quantum objects are standard rotation gates and entanglers.

free parameters (5)
  • Hybrid VAE trainable parameters (87 quantum rotation angles, 630 classical weights) = trained values not reported; counts only
    Performance parity claim depends on the fit to 4M QCD background; no weights or seed information are provided.
  • QAE trainable parameters (20 single-qubit rotations, 3 delta_block offsets) = trained values not reported
    The QAE has 23 trainable quantum parameters; the reported anomaly-detection performance depends on this fit.
  • Per-trash-wire z-score normalization (mu_i, sigma_i), 9 pairs = not reported; fitted on background
    Equation (7) defines z_i using fitted mu_i/sigma_i, and the final QAE score sums the top two positive z-scores, so this calibration is load-bearing.
  • FPGA fixed-point formats (q_ae, mu, m, z, s) = Table 1: hybrid q_ae <10,3>, mu <18,6>, m <20,8>, z <18,4>, s <20,5>; QAE q_ae <14,3>, mu <18,6>, s <32,15>
    Chosen via quantization sweep; resource/latency and post-quantization performance claims depend on these widths. The text notes AUC artifacts at some widths.
  • Hand-chosen hyperparameters (latent dim 3, beta cycle 0.1-0.8, lr 1e-4/1e-3, batch 512/1024, early stopping, epoch caps) = as stated in Sec. 2.3
    These were chosen by hand and not swept systematically; they affect the reported results.
axioms (6)
  • standard math PennyLane state-vector simulation correctly computes the expectation values used for training and scoring.
    The paper relies on the simulator's correctness for all quantum observables; no independent verification is given.
  • domain assumption The LHC dataset from Ref. [33] (background plus four BSM signals) is a valid surrogate for real detector data and the classical SOTA benchmarks in that reference are valid.
    The central performance comparison and all signal injections come from this external dataset.
  • domain assumption One-particle-one-qubit encoding (Ref. [34]) and the clipped KL anomaly score (Ref. [25]) are suitable building blocks for anomaly detection.
    The models are constructed from these prior components without re-deriving their validity for this trigger setting.
  • ad hoc to paper The 10-jet register can be split into two independent 5-qubit blocks, with two-wire correlators recovering enough cross-jet information for competitive anomaly detection.
    This hardware-driven split is introduced in Sec. 2.1; all performance claims are for this split architecture, not a fully entangled 10-qubit jet register.
  • ad hoc to paper HLS post-synthesis resource and latency estimates on the Alveo U200 are a faithful proxy for real hardware timing.
    The FPGA feasibility claim rests on synthesis numbers in Table 2; no placed-and-routed or on-board validation is described.
  • domain assumption Background-only training and evaluation on full signal files yields unbiased anomaly-detection metrics.
    This is standard unsupervised AD methodology, but no signal contamination check on the background sample is discussed.

pith-pipeline@v1.3.0-alltime-deepseek · 9744 in / 17785 out tokens · 158393 ms · 2026-08-01T10:12:35.183910+00:00 · methodology

0 comments
read the original abstract

Quantum machine learning (QML) algorithms in high energy physics (HEP) can efficiently represent and leverage long-range, high-order correlations in high-dimensional collider data, potentially with fewer parameters and favorable scaling relative to classical models. Deployment of QML in real-time collider applications such as trigger systems requires the ability to emulate and compile quantum circuits classically, then synthesize the resulting quantum gates onto low-latency hardware accelerators, namely field-programmable gate arrays (FPGAs). We present a study of variational quantum autoencoder models for real-time anomaly detection triggers in modern collider experiments. The models achieve performance comparable to state-of-the-art classical approaches and, after FPGA synthesis, satisfy resource usage and timing constraints consistent with trigger applications in future colliders. This work provides one of the first FPGA implementations of QML models for HEP triggers, enabling higher-capability models in today's classical data acquisition pipelines while advancing quantum readiness of collider experiment infrastructure.

discussion (0)

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

Reference graph

Works this paper leans on

42 extracted references · 11 canonical work pages

  1. [1]

    Journal of Instrumentation19(05), 05063 (2024) https://doi.org/10.1088/1748-0221/19/05/ p05063

    ATLAS Collaboration: The ATLAS experiment at the CERN Large Hadron Collider: a description of the detector configuration for Run 3. Journal of Instrumentation19(05), 05063 (2024) https://doi.org/10.1088/1748-0221/19/05/ p05063

    work page doi:10.1088/1748-0221/19/05/ 2024

  2. [2]

    ATLAS Collaboration: Technical Design Report for the Phase-II Upgrade of the ATLAS TDAQ System https://doi.org/10.17181/CERN.2LBB.4IAL

    doi:10.17181/cern.2lbb.4ial

  3. [3]

    Technical report, CERN, Geneva (2021)

    Collaboration, C.: The Phase-2 Upgrade of the CMS Data Acquisition and High Level Trigger. Technical report, CERN, Geneva (2021). This is the final version of the document, approved by the LHCC. https://cds.cern.ch/record/2759072

    arXiv 2021

  4. [4]

    https://iml-wg.github.io/HEPML-LivingReview/

    HEP ML Community: A Living Review of Machine Learning for Particle Physics. https://iml-wg.github.io/HEPML-LivingReview/

  5. [5]

    Nature549(7671), 195–202 (2017) https://doi.org/ 10.1038/nature23474

    Biamonte, J., Wittek, P., Pancotti, N., Rebentrost, P., Wiebe, N., Lloyd, S.: Quantum machine learning. Nature549(7671), 195–202 (2017) https://doi.org/ 10.1038/nature23474

    doi:10.1038/nature23474 2017

  6. [6]

    Science376(6598), 1182–1186 (2022) https://doi.org/ 10.1126/science.abn7293 11

    Huang, H.-Y., Broughton, M., Cotler, J., Chen, S., Li, J., Mohseni, M., Neven, H., Babbush, R., Kueng, R., Preskill, J., McClean, J.R.: Quantum advantage in learning from experiments. Science376(6598), 1182–1186 (2022) https://doi.org/ 10.1126/science.abn7293 11

    doi:10.1126/science.abn7293 2022

  7. [7]

    Nachman, B., Provasoli, D., Jong, W.A., Bauer, C.W.: Quantum algorithm for high energy physics simulations. Phys. Rev. Lett.126, 062001 (2021) https:// doi.org/10.1103/PhysRevLett.126.062001

    work page doi:10.1103/physrevlett.126.062001 2021

  8. [8]

    npj Quantum Inf.7, 111 (2021) https://doi.org/10.1038/s41534-021-00443-w arXiv:2004.13747 [stat.ML]

    Felser, T., Trenti, M., Sestini, L., Gianelle, A., Zuliani, D., Lucchesi, D., Mon- tangero, S.: Quantum-inspired machine learning on high-energy physics data. npj Quantum Inf.7, 111 (2021) https://doi.org/10.1038/s41534-021-00443-w arXiv:2004.13747 [stat.ML]

    Pith/arXiv arXiv 2021

  9. [9]

    https://arxiv.org/abs/2509.24383

    Linne, K.C., Uemura, S., Ji, Y., Zang, A., Federico, M.D., Quaranta, O., Cancelo, G., Roy, D.: SQuaD: Smart Quantum Detection for Photon Recognition and Dark Count Elimination (2026). https://arxiv.org/abs/2509.24383

    arXiv 2026

  10. [10]

    In: Proceedings of the IEEE International Conference on Computer Design

    Khalid, A.U., Zilic, Z., Radecka, K.: Fpga emulation of quantum circuits. In: Proceedings of the IEEE International Conference on Computer Design. ICCD ’04, pp. 310–315. IEEE Computer Society, USA (2004)

    2004

  11. [11]

    In: 2024 31st IEEE International Conference on Electronics, Circuits and Systems (ICECS), pp

    Conti, C., Volpe, D., Graziano, M., Zamboni, M., Turvani, G.: AMARETTO: Enabling Efficient Quantum Algorithm Emulation on Low-Tier FPGAs. In: 2024 31st IEEE International Conference on Electronics, Circuits and Systems (ICECS), pp. 1–4 (2024). https://doi.org/10.1109/ICECS61496.2024.10848965

    arXiv 2024

  12. [12]

    In: 2017 Eight Argentine Symposium and Conference on Embedded Systems (CASE), pp

    Silva, A., Zabaleta, O.G.: FPGA quantum computing emulator using high level design tools. In: 2017 Eight Argentine Symposium and Conference on Embedded Systems (CASE), pp. 1–6 (2017). https://doi.org/10.23919/SASE-CASE.2017. 8115369

    work page doi:10.23919/sase-case.2017 2017

  13. [13]

    In: 2022 13th International Conference on Information and Communica- tion Technology Convergence (ICTC), pp

    Hong, Y., Jeon, S., Park, S., Kim, B.-S.: Quantum Circuit Simulator based on FPGA. In: 2022 13th International Conference on Information and Communica- tion Technology Convergence (ICTC), pp. 1909–1911 (2022). https://doi.org/10. 1109/ICTC55196.2022.9952408

    arXiv 2022

  14. [14]

    In: Proceedings of the 2008 IEEE Computer Society Annual Symposium on VLSI

    Aminian, M., Saeedi, M., Zamani, M.S., Sedighi, M.: Fpga-based circuit model emulation of quantum algorithms. In: Proceedings of the 2008 IEEE Computer Society Annual Symposium on VLSI. ISVLSI ’08, pp. 399–404. IEEE Computer Society, USA (2008). https://doi.org/10.1109/ISVLSI.2008.43 . https://doi.org/10.1109/ISVLSI.2008.43

    work page doi:10.1109/isvlsi.2008.43 2008

  15. [15]

    https://arxiv.org/abs/2510.07110

    Duy, T.V., Vu, T.H., Le, V.T.D., Pham, H.L., Nakashima, Y.: HPQEA: A Scal- able and High-Performance Quantum Emulator with High-Bandwidth Memory for Diverse Algorithms Support (2025). https://arxiv.org/abs/2510.07110

    arXiv 2025

  16. [16]

    IEEE Access (2025) arXiv:2506.01029 [quant-ph]

    Lagostina, L., Volpe, D., Zamboni, M., Turvani, G.: AEQUAM: Accelerating quantum algorithm validation through FPGA-based emulation. IEEE Access (2025) arXiv:2506.01029 [quant-ph]. arXiv:2506.01029

    Pith/arXiv arXiv 2025

  17. [17]

    Quantum Engineering2025(1), 9017796 (2025) https://doi.org/10.1155/que2/9017796

    Bennakhi, A., Byrd, G.T., Franzon, P.: Q2SV: A high-level synthesis approach for 12 state vector quantum simulation. Quantum Engineering2025(1), 9017796 (2025) https://doi.org/10.1155/que2/9017796

    work page doi:10.1155/que2/9017796 2025

  18. [18]

    Reviews in Physics12, 100091 (2024) https://doi.org/10.1016/ j.revip.2024.100091

    Belis, V., Odagiu, P., Aarrestad, T.K.: Machine learning for anomaly detection in particle physics. Reviews in Physics12, 100091 (2024) https://doi.org/10.1016/ j.revip.2024.100091

    arXiv 2024

  19. [19]

    ATLAS Collaboration: Dijet Resonance Search with Weak Supervision Using√s= 13 TeVppCollisions in the ATLAS Detector. Phys. Rev. Lett.125, 131801 (2020) https://doi.org/10.1103/PhysRevLett.125.131801 arXiv:2005.02983 [hep- ex]

    Pith/arXiv arXiv 2020

  20. [20]

    ATLAS Collaboration: Anomaly detection search for new resonances decaying into a Higgs boson and a generic new particleXin hadronic final states using√s= 13 TeVppcollisions with the ATLAS detector. Phys. Rev. D108, 052009 (2023) https://doi.org/10.1103/PhysRevD.108.052009 arXiv:2306.03637 [hep-ex]

    Pith/arXiv arXiv 2023

  21. [21]

    ATLAS Collaboration: Search for new physics in final states with semivisible jets or anomalous signatures using the ATLAS detector. Phys. Rev. D112, 012021 (2025) https://doi.org/10.1103/44zp-mh1q arXiv:2505.01634 [hep-ex]

    Pith/arXiv arXiv 2025

  22. [22]

    ATLAS Collaboration: Search for New Phenomena in Two-Body Invariant Mass Distributions Using Unsupervised Machine Learning for Anomaly Detection at√s= 13 TeV with the ATLAS Detector. Phys. Rev. Lett.132, 081801 (2024) https://doi.org/10.1103/PhysRevLett.132.081801 arXiv:2307.01612 [hep-ex]

    Pith/arXiv arXiv 2024

  23. [23]

    ATLAS Collaboration: Weakly supervised anomaly detection for resonant new physics in the dijet final state using proton–proton collisions at √s= 13 TeV with the ATLAS detector. Phys. Rev. D112, 072009 (2025) https://doi.org/10.1103/ 2yq5-vj59 arXiv:2502.09770 [hep-ex]

    arXiv 2025

  24. [24]

    Reports on Progress in Physics88(6), 067802 (2025) https://doi.org/10.1088/1361-6633/ add762 arXiv:2412.03747 [hep-ex]

    CMS Collaboration: Model-agnostic search for dijet resonances with anoma- lous jet substructure in proton-proton collisions at √s= 13 TeV. Reports on Progress in Physics88(6), 067802 (2025) https://doi.org/10.1088/1361-6633/ add762 arXiv:2412.03747 [hep-ex]

    Pith/arXiv arXiv 2025

  25. [25]

    Nature Machine Intelligence4(2), 154–161 (2022) https://doi.org/10

    Govorkova, E., Puljak, E., Aarrestad, T., James, T., Loncar, V., Pierini, M., Pol, A.A., Ghielmetti, N., Graczyk, M., Summers, S., Ngadiuba, J., Nguyen, T.Q., Duarte, J., Wu, Z.: Autoencoders on field-programmable gate arrays for real-time, unsupervised new physics detection at 40 MHz at the Large Hadron Collider. Nature Machine Intelligence4(2), 154–161 ...

    2022

  26. [26]

    https:// arxiv.org/abs/1312.6114 13

    Kingma, D.P., Welling, M.: Auto-Encoding Variational Bayes (2022). https:// arxiv.org/abs/1312.6114 13

    Pith/arXiv arXiv 2022

  27. [27]

    https://arxiv

    Ludmir, J.Z., Rebello, S., Ruiz, J., Patel, T.: Quorum: Zero-Training Unsuper- vised Anomaly Detection using Quantum Autoencoders (2025). https://arxiv. org/abs/2504.13113

    Pith/arXiv arXiv 2025

  28. [28]

    Physical Review D105(9) (2022) https://doi.org/10.1103/physrevd.105.095004

    Ngairangbam, V.S., Spannowsky, M., Takeuchi, M.: Anomaly detection in high- energy physics using a quantum autoencoder. Physical Review D105(9) (2022) https://doi.org/10.1103/physrevd.105.095004

    doi:10.1103/physrevd.105.095004 2022

  29. [29]

    https://arxiv.org/abs/2407.07961

    Duffy, C., Hassanshah, M., Jastrzebski, M., Malik, S.: Unsupervised Beyond- Standard-Model Event Discovery at the LHC with a Novel Quantum Autoencoder (2024). https://arxiv.org/abs/2407.07961

    Pith/arXiv arXiv 2024

  30. [30]

    Machine Learning: Science and Technology6(4), 045001 (2025) https://doi.org/10.1088/2632-2153/ae0243

    Puljak, E., Pierini, M., Garcia-Saez, A.: Tensor network for anomaly detection in the latent space of proton collision events at the LHC. Machine Learning: Science and Technology6(4), 045001 (2025) https://doi.org/10.1088/2632-2153/ae0243

    doi:10.1088/2632-2153/ae0243 2025

  31. [31]

    https://arxiv.org/abs/2603.26604

    Addepalli, S., Bhattarai, P., Dave, A., Gonski, J.: Hardware-Aware Tensor Net- works for Real-Time Quantum-Inspired Anomaly Detection at Particle Colliders (2026). https://arxiv.org/abs/2603.26604

    arXiv 2026

  32. [32]

    https://arxiv.org/abs/2601.10801

    Coppi, A., Puljak, E., Borella, L., Jaschke, D., Rico, E., Pierini, M., Pazzini, J., Triossi, A., Montangero, S.: Towards Tensor Network Models for Low-Latency Jet Tagging on FPGAs (2026). https://arxiv.org/abs/2601.10801

    arXiv 2026

  33. [33]

    Scientific Data9(1), 118 (2022) https://doi.org/10.1038/s41597-022-01187-8

    Govorkova, E., Puljak, E., Aarrestad, T., Pierini, M., Wo´ zniak, K.A., Ngadiuba, J.: LHC physics dataset for unsupervised New Physics detection at 40 MHz. Scientific Data9(1), 118 (2022) https://doi.org/10.1038/s41597-022-01187-8

    work page doi:10.1038/s41597-022-01187-8 2022

  34. [34]

    Physical Review D112(7) (2025) https://doi.org/10.1103/l8y2-87vq

    Bal, A., Klute, M., Maier, B., Oughton, M., Pezone, E., Spannowsky, M.: One particle - one qubit: Particle physics data encoding for quantum machine learning. Physical Review D112(7) (2025) https://doi.org/10.1103/l8y2-87vq

    doi:10.1103/l8y2-87vq 2025

  35. [35]

    In: Proceedings of the Machine Learning and the Physical Sciences Workshop at NeurIPS 2022 (2022)

    Acharya, A., Rudolph, M., Chen, J., Miller, J., Perdomo-Ortiz, A.: Qubit seri- ation: Improving data-model alignment using spectral ordering. In: Proceedings of the Machine Learning and the Physical Sciences Workshop at NeurIPS 2022 (2022). Workshop at the 36th Conference on Neural Information Processing Systems (NeurIPS 2022)

    2022

  36. [36]

    https://arxiv.org/abs/1811.04968

    Bergholm, V., Izaac, J., Schuld, M., Gogolin, C., Ahmed, S., Ajith, V., Alam, M.S., Alonso-Linaje, G., AkashNarayanan, B., Asadi, A., Arrazola, J.M., Azad, U., Banning, S., Blank, C., Bromley, T.R., Cordier, B.A., Ceroni, J., Delgado, A., Matteo, O.D., Dusko, A., Garg, T., Guala, D., Hayes, A., Hill, R., Ijaz, A., Isacsson, T., Ittah, D., Jahangiri, S., J...

    Pith/arXiv arXiv 2022

  37. [37]

    https: //arxiv.org/abs/1412.6980

    Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization (2017). https: //arxiv.org/abs/1412.6980

    Pith/arXiv arXiv 2017

  38. [38]

    Zenodo (2021)

    Aarrestad, T., Govorkova, E., Ngadiuba, J., Puljak, E., Pierini, M., Wozniak, K.A.: Unsupervised New Physics detection at 40 MHz: Training Dataset. Zenodo (2021). https://doi.org/10.5281/zenodo.5046389

    work page doi:10.5281/zenodo.5046389 2021

  39. [39]

    Zenodo (2021)

    Aarrestad, T., Govorkova, E., Ngadiuba, J., Puljak, E., Pierini, M., Wozniak, K.A.: Unsupervised New Physics detection at 40 MHz:A→4ℓSignal Benchmark Dataset. Zenodo (2021). https://doi.org/10.5281/zenodo.5046446

    work page doi:10.5281/zenodo.5046446 2021

  40. [40]

    Zenodo (2021)

    Aarrestad, T., Govorkova, E., Ngadiuba, J., Puljak, E., Pierini, M., Woz- niak, K.A.: Unsupervised New Physics detection at 40 MHz:h 0 →τ τSignal Benchmark Dataset. Zenodo (2021). https://doi.org/10.5281/zenodo.5061633

    work page doi:10.5281/zenodo.5061633 2021

  41. [41]

    Zenodo (2021)

    Aarrestad, T., Govorkova, E., Ngadiuba, J., Puljak, E., Pierini, M., Woz- niak, K.A.: Unsupervised New Physics detection at 40 MHz:h + →τ νSignal Benchmark Dataset. Zenodo (2021). https://doi.org/10.5281/zenodo.7152617

    work page doi:10.5281/zenodo.7152617 2021

  42. [42]

    Zenodo (2021)

    Aarrestad, T., Govorkova, E., Ngadiuba, J., Puljak, E., Pierini, M., Woz- niak, K.A.: Unsupervised New Physics detection at 40 MHz:LQ→bτSignal Benchmark Dataset. Zenodo (2021). https://doi.org/10.5281/zenodo.5055454 15

    work page doi:10.5281/zenodo.5055454 2021