Pith. sign in

REVIEW 4 major objections 4 minor 40 references

A 1,649-parameter convolutional network, compiled into FPGA qubit-control firmware, detects charge jumps in superconducting qubits with 6.19 microsecond latency, matching the efficiency of offline analysis without per-qubit tuning.

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

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

T0 review · deepseek-v4-flash

2026-08-02 02:31 UTC pith:7II6QMEG

load-bearing objection A genuinely new FPGA/CNN feasibility study for online charge-jump detection, but the central 'matched FPR' claim is not supported on the numbers, and all results are simulation-only. the 4 major comments →

arxiv 2607.14293 v1 pith:7II6QMEG submitted 2026-07-15 quant-ph cs.LG

Real-Time Detection of Charge Jumps in Superconducting Qubits with a Convolutional Neural Network

Daniel Gaytan-Villarreal , Peter Meiring , Daniel Baxter , Daniel Bowring , Grace Bratrud , Matteo Cremonesi , Giuseppe Di Guglielmo , Grace Wagner
show 1 more author
Bowen Xiao
This is my paper
classification quant-ph cs.LG
keywords charge jumpssuperconducting qubitsdilated causal CNNreal-time detectionFPGA firmwareRamsey tomographyquantum sensingquantization-aware training
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.

The paper tries to establish that charge-jump detection in superconducting qubits can move from offline, per-qubit-tuned analysis to a real-time control primitive. It claims a tiny dilated causal convolutional neural network, trained entirely on synthetic Ramsey scans and quantized to fixed-point arithmetic, runs on an FPGA in the qubit control loop at 6.19 microseconds per inference. On the benchmark task of finding jumps of 0.1–0.5 electron charges, this network matches the detection efficiency of the established offline chi-squared algorithm at the same false-positive rate, while requiring no per-qubit hyperparameter tuning. If true, this makes adaptive quantum error mitigation and in-situ response to radiation events practical.

Core claim

The central claim is that a 1,649-parameter Dilated Causal Convolutional Neural Network (DCCNN), trained only on simulated Ramsey charge-tomography scans and then quantized to 16-bit fixed-point precision, can be compiled to an FPGA block inside qubit control firmware and detect charge jumps with a mean efficiency of 0.843 ± 0.022 on magnitudes 0.1–0.5 e, statistically matching the offline chi-squared method's 0.866 ± 0.020 at a matched false-positive rate. The network processes an 80-sample scan in 6.19 microseconds and needs no post-training per-qubit calibration. This shifts charge-jump detection from a post-hoc diagnostic to an in-the-loop control primitive.

What carries the argument

The key object is the DCCNN: four one-dimensional convolution layers with kernel size 3 and dilation rates 1 and 2, giving a causal receptive field that spans the scan without extra parameters, followed by global pooling, a dense layer, and a sigmoid. The input is a two-channel 80-sample scan: the raw qubit transmission amplitude plus its first-difference trace, which makes a discontinuous phase shift stand out as a local spike. The model is trained with quantization-aware training to ap_fixed<16,6> precision and translated to FPGA firmware via a high-level synthesis workflow, with a streaming interface feeding a sliding window of scans and a control processor polling the output flag.

Load-bearing premise

The load-bearing premise is that the synthetic Ramsey scans are faithful stand-ins for real qubit data: the simulator assumes instantaneous phase jumps, Gaussian noise with qubit-specific envelopes, and Poisson jump timing, so if real charge jumps have transients, non-Gaussian noise, or correlations the simulator omits, the deployed model's efficiency will differ from the reported 0.843.

What would settle it

Run the compiled FPGA model on real Ramsey scans whose jump timestamps are known independently (e.g., from simultaneous radiation monitors or engineered bias steps) and compare its per-qubit detection efficiency against the simulated 0.843–0.866 range; a substantial shortfall would show the synthetic training distribution is not faithful.

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

If this is right

  • Charge-jump detection becomes a real-time control primitive: the control system can pause experiments, recalibrate, or branch its sequence the moment a jump is flagged.
  • A single trained network serves multiple qubit templates without per-qubit tuning, removing a source of analysis bias and manual overhead.
  • The resource footprint (with block RAM as the tightest constraint) fits comfortably in the ZCU216 FPGA alongside existing qubit control logic, suggesting room for additional on-chip processing.
  • The same fixed-latency, simulation-trained pipeline can be reused as a trigger for qubit-based particle detectors looking for radiation events.
  • The ablation result—that residual connections do not improve accuracy—supports lightweight architectures for this task, lowering the latency and cost of deployment.

Where Pith is reading between the lines

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

  • If the synthetic-to-real gap is small, the same compile-and-deploy pipeline could be extended to detect other time-series anomalies in qubit control, such as two-level-system shifts or readout drifts, with the first-difference channel providing a generic discontinuity signature.
  • The 6.19-microsecond figure is per 80-sample scan; for truly low-latency feedback, a future variant could shorten the window or classify streamingly sample-by-sample, at some cost in efficiency.
  • Because the model learns jump signatures from synthetic data, it could be retrained on real labeled charge-jump traces once enough are collected, potentially pushing efficiency above the current chi-squared parity.
  • The network's reliance on the first-difference channel suggests that a feature-engineering baseline might achieve similar results, but the CNN's advantage is avoiding per-qubit thresholds and providing a uniform decision rule.

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 / 4 minor

Summary. The paper proposes a dilated causal convolutional neural network (DCCNN) for detecting radiation-induced charge jumps in Ramsey charge-tomography scans of superconducting qubits. The network is trained on synthetic scans generated from NEXUS qubit templates, quantized to ap_fixed<16,6> via QKeras, and compiled to FPGA firmware with hls4ml for integration into QICK on the ZCU216. The central quantitative claim is that, at a matched false-positive rate, the quantized DCCNN reaches a mean detection efficiency of 0.843 ± 0.022 versus 0.866 ± 0.020 for the established χ2 algorithm on |Δq| ∈ [0.1, 0.5] e, with a reported per-inference latency of 6.19 μs and 1,649 parameters, and with no per-qubit hyperparameter tuning. The paper also ablates residual connections by comparing the DCCNN to a TCN and reports FPGA resource utilization.

Significance. If the headline result holds, this is a useful engineering contribution: it demonstrates an end-to-end path from simulated training data to FPGA firmware for an online charge-jump detector, with modest resource overhead and latency in the few-microsecond range, making the method a plausible control-loop primitive for QICK-based experiments. The paper is commendably transparent about training configuration, quantization procedure, resource utilization, and per-qubit results, and the TCN-vs-DCCNN ablation is a clean way to isolate the effect of residual connections. The significance is qualified, however, by an internal inconsistency in the central matched-FPR comparison and by the fact that all quantitative detection claims are evaluated only on the authors' synthetic simulator; the deployed-detector efficiency on real charge-jump data remains unmeasured.

major comments (4)
  1. [III C / Table VII / Abstract] The claim of a 'matched false-positive rate' is not supported by the manuscript. Section III C states that the χ2 algorithm was tuned under a 5% FPR constraint, while the DCCNN's FPR after quantization is 4% (6% before quantization). Table VII uses the quantized DCCNN, so the comparison is at 4% vs 5% FPR. Since detection efficiency generally increases with allowable FPR, the χ2 algorithm may have a permissive advantage, and the small 0.023 mean-efficiency gap could close or reverse at a common FPR. Please re-evaluate at a strictly matched FPR (e.g., threshold sweep for the DCCNN and/or a 4% FPR constraint for χ2), or provide an ROC-style comparison that justifies the 1-point mismatch as negligible.
  2. [IV C / Table IX] The per-layer latency values in Table IX are inconsistent with the reported total. The sum of the listed minimum per-layer cycles is approximately 6,951 cycles, whereas the total latency is given as 1,860 cycles. If the layers are scheduled with overlap or the per-layer numbers are standalone HLS estimates, this should be stated explicitly; as written, the table appears to contradict the 6.19 μs headline latency and prevents the reader from reproducing the end-to-end figure.
  3. [Table VII, Q4 (SO) row] The per-qubit comparison in Table VII shows a substantial efficiency gap for Q4 in shield-open mode: DCCNN 0.810 ± 0.019 vs χ2 0.883 ± 0.018, a difference of roughly 2.7 combined standard deviations. This undercuts the claim of 'comparable efficiency across all five qubit templates' and suggests that the no-tuning advantage has a per-qubit cost. Please discuss this row explicitly and qualify the abstract/conclusion accordingly, or provide evidence that the discrepancy is within expected statistical scatter.
  4. [II B / III C / Conclusions] All quantitative performance claims are measured on synthetic scans generated from the same templates and jump-injection methodology used to train the network. The generator assumes instantaneous phase discontinuities, uniform jump magnitudes, Gaussian noise, and Poisson jump times; if real charge jumps have settling transients, non-Gaussian noise, or correlated multi-jump behavior, the deployed efficiency will differ. The abstract and conclusions should state clearly that the reported efficiencies are simulation-based, and the paper would be strengthened by a small validation set of measured scans or an explicit, prominent limitation statement.
minor comments (4)
  1. [II B] The paragraph beginning 'Each scan is converted to a two-channel input...' is repeated verbatim in the manuscript. Please remove the duplicate.
  2. [Figure 2 and surrounding text] The body text says the bottom scan contains a jump of ΔV = +0.227 V, while the caption says Δq = +0.417 e at gate voltage −0.184 V. The units and values should be reconciled.
  3. [III A vs III B] Tables III and IV report per-class metrics without specifying the jump-magnitude range. Tables V–VII use |Δq| ∈ [0.1, 0.5] e, but the earlier tables appear to include smaller jumps from [0.01, 0.5] e. Please state the magnitude range used for each table so the metrics are comparable.
  4. [Conclusions] The phrase 'supervision-free at inference time' is potentially misleading; the network is trained with labels. Consider rewording to 'does not require labels during inference' or similar.

Circularity Check

0 steps flagged

No construction-level circularity; the network is evaluated on held-out synthetic scans and the χ2 baseline is a comparator, not a fitted input, though real-data validation is deferred and the FPR matching is imperfect.

full rationale

The paper's derivation chain is not circular at the level of equations or definitions. The DCCNN is trained on synthetic Raman scans generated from templates and noise envelopes taken from the prior NEXUS work (Ref. [17]) and then evaluated on a separately held-out set of scans produced by the same generator. Held-out evaluation on the training distribution is a genuine classification test: the network could have failed to detect injected jumps, and its measured efficiency (0.843) is not forced by the training objective or by the generator. The χ2 algorithm is imported as a baseline and is itself tuned on the training set to a 5% FPR constraint, but the DCCNN efficiency is not a fit to that baseline; it is an independent forward-pass measurement. The reliance on Ref. [17] for templates, jump statistics, and evaluation methodology is a self-citation, but Ref. [17] is a published experimental study providing measured templates and observed jump parameters, so it is independent evidence rather than an unverified assumption imported as a uniqueness result. The paper itself states the main outstanding limitation: 'Future work will focus on deploying the DCCNN on real-time data acquisition with operating qubit hardware, validating the simulation-trained model on measured scans,' which means all quantitative claims are simulation-only. That is a validation gap, not a circular reduction. Separately, the abstract's 'matched false-positive rate' is not strictly supported by the text: the χ2 is tuned to 5% FPR while the quantized DCCNN operates at 4% FPR (6% before quantization). This affects the interpretation of the efficiency comparison (0.866 vs 0.843) but is an operating-point/correctness concern, not a circularity. No step reduces by definition, no fitted parameter is renamed as a prediction, and no self-citation forbids alternatives. Score 1 reflects the minor self-citation dependence without any construction-level circularity.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The experimental loop is closed inside the paper: templates and statistics come from the group's prior Ref [17], the synthetic generator, the trained network, and the tuned baseline are all the authors' own. There is no external real-data benchmark. Hardware numbers are from compilation/synthesis, not from a running board.

free parameters (8)
  • Synthetic jump rate λ = 1.1 mHz
    Chosen to match Ref [17]'s simulated rate; controls jump/no-jump class balance in training and evaluation.
  • Jump magnitude distribution = U[0.01, 0.5] e
    Injection range used in Sec II.B; headline efficiency is only on the [0.1,0.5] e subset, so low-magnitude jumps affect training but not the reported metric.
  • Network hyperparameters = 8/16 filters, kernel 3, dilations 1 and 2, depth 4
    Hand-chosen architecture (Sec II.A); no systematic hyperparameter search beyond residual ablation.
  • Fixed-point precision = ap_fixed<16,6>
    Selected from a three-point bit-width sweep as highest accuracy meeting latency/resource constraints (Sec II.A/IV).
  • Decision threshold = not reported (implicit)
    No explicit classification threshold is given; DCCNN FPR is measured as 6% (float) and 4% (quantized), which the chi2 baseline is tuned to match.
  • Chi2 FPR constraint = 5%
    Baseline is tuned on the training set to 5% FPR to approximately match the DCCNN's observed FPR; exact matching is not achieved.
  • Training hyperparameters = lr 1e-3/1e-4, batch 256, patience 40
    Standard QAT choices; not physics parameters but required to reproduce training.
  • Scan length N = 80 samples
    Determines receptive field and input shape; inherited from the NEXUS measurement protocol (Sec II.B).
axioms (6)
  • domain assumption Gaussian noise with per-template envelope σ̂(θ) describes the measurement noise in Ramsey scans.
    Assumed in Sec II.B for synthetic data; if real noise is non-Gaussian or correlated, detection efficiencies will shift.
  • domain assumption A charge jump is an instantaneous discontinuity in template phase with uniformly sampled magnitude, random sign, and uniform scan position.
    The jump injection model in Sec II.B; real jumps may have settling dynamics or correlated structure.
  • domain assumption Templates and noise envelopes measured for Q1–Q3 and Q4 (SC/SO) at NEXUS span the variation seen across other qubit devices.
    The claim of no per-qubit tuning assumes new qubits will be drawn from a similar template family; used implicitly in Sec III/Conclusions.
  • domain assumption Vitis HLS C-synthesis latency and post-place-and-route resource utilization predict on-hardware performance.
    Section IV.C reports latency 1,860 cycles (6.19 μs) from synthesis, not from a running FPGA; the 'real-time' claim depends on this.
  • domain assumption The QICK baseline resource counts of Ref [20] are accurate and representative for the ZCU216 firmware.
    Resource increments in Table VIII are measured relative to that baseline; a different QICK build would give different headroom.
  • domain assumption The χ2 algorithm implementation reproduces the behavior of Ref [17] after tuning on the same synthetic training set.
    Section III.C; fair comparison depends on faithful reimplementation.

pith-pipeline@v1.3.0-alltime-deepseek · 11675 in / 19060 out tokens · 189846 ms · 2026-08-02T02:31:16.441642+00:00 · methodology

0 comments
read the original abstract

Ionizing radiation from cosmic rays and gammas can induce discontinuous jumps in the environmental charge of superconducting qubits (charge jumps), causing correlated errors that challenge fault-tolerant quantum computing while simultaneously providing a detection signature for quantum sensing applications. Current detection methods operate offline, introducing latency incompatible with in-the-loop qubit control. In this paper, an online detector of charge jumps for superconducting qubits, based on a dilated causal convolutional neural network (DCCNN) designed for in-the-loop deployment on the Quantum Instrumentation Control Kit (QICK) platform, is presented. The network is trained on synthetic Ramsey tomography scans generated from qubit templates measured at the Northwestern Experimental Underground Site (NEXUS) at Fermilab, and translated to FPGA firmware via hls4ml with ap_fixed$\langle 16,6 \rangle$ quantization, reaching a per-inference latency of $6.19 \mu$s on the Zynq UltraScale+ RFSoC ZCU216. At this operating point the DCCNN matches the detection efficiency of the established offline $\chi^2$ algorithm ($0.843 \pm 0.022$ vs. $0.866 \pm 0.020$ on $|\Delta q| \in [0.1, 0.5] e$ at matched false-positive rate), while requiring no per-qubit hyperparameter tuning. This shifts charge-jump detection from a post-hoc diagnostic to a control-loop primitive, enabling adaptive protocols that respond to radiation-induced events in situ, with applications to quantum-computing error mitigation and to the use of superconducting qubits as particle detectors.

Figures

Figures reproduced from arXiv: 2607.14293 by Bowen Xiao, Daniel Baxter, Daniel Bowring, Daniel Gaytan-Villarreal, Giuseppe Di Guglielmo, Grace Bratrud, Grace Wagner, Matteo Cremonesi, Peter Meiring.

Figure 1
Figure 1. Figure 1: FIG. 1. Proposed architectures: TCN (left) and DCCNN [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Example synthetic Ramsey scans generated for Qubit 1 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

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

40 extracted references · 3 canonical work pages

  1. [1]

    Danilin, N

    S. Danilin, N. Nugent, and M. Weides, Quantum sensing with tunable superconducting qubits: optimization and speed-up (2024), arXiv:2211.08344 [quant-ph]

    Pith/arXiv arXiv 2024

  2. [2]

    Bonomoet al., First operation of superconducting quantum bits as particle detectors, PoS14YRM2025, 001 (2026)

    C. Bonomoet al., First operation of superconducting quantum bits as particle detectors, PoS14YRM2025, 001 (2026). 8

    2026

  3. [3]

    Ramanathan, B

    K. Ramanathan, B. J. Sandoval, J. E. Parker, L. M. Joshi, A. D. Beyer, P. M. Echternach, S. Rosenblum, and S. R. Golwala, Quantum parity detectors: A qubit-based particle-detection scheme with mev thresholds for rare- event searches, APS Open Science1, 10.1103/kqd2-spb1 (2026)

    work page doi:10.1103/kqd2-spb1 2026

  4. [4]

    A. V. Dixit, S. Chakram, K. He, A. Agrawal, R. K. Naik, D. I. Schuster, and A. Chou, Searching for dark matter with a superconducting qubit, Physical Review Letters 126, 10.1103/physrevlett.126.141302 (2021)

    doi:10.1103/physrevlett.126.141302 2021

  5. [5]

    C. Fink, C. Salemi, B. Young, D. Schuster, and N. Kurin- sky, Superconducting quasiparticle-amplifying transmon: A qubit-based sensor for mev-scale phonons and sin- gle terahertz photons, Physical Review Applied22, 10.1103/physrevapplied.22.054009 (2024)

    doi:10.1103/physrevapplied.22.054009 2024

  6. [6]

    Linehan, I

    R. Linehan, I. Hernandez, D. J. Temples, S. Q. Dang, D. Baxter, L. Hsu, E. Figueroa-Feliciano, R. Khatiwada, K. Anyang, D. Bowring, G. Bratrud, G. Cancelo, A. Chou, R. Gualtieri, K. Stifter, and S. Sussman, Estimating the energy threshold of phonon-mediated superconduct- ing qubit detectors operated in an energy-relaxation sensing scheme, Physical Review ...

    doi:10.1103/phys- 2025

  7. [7]

    Linehan, T

    R. Linehan, T. Trickle, C. R. Conner, S. Ghosh, T. Lin, M. Sholapurkar, and A. N. Cleland, Listening for new physics with quantum acoustics (2024), arXiv:2410.17308 [hep-ph]

    Pith/arXiv arXiv 2024

  8. [8]

    E. Celi, R. Linehan, P. M. Harrington, M. Li, H. D. Pinckney, K. Serniak, W. D. Oliver, J. A. Formaggio, E. Figueroa-Feliciano, and D. Baxter, Measuring quasi- particle dynamics for particle impact reconstruction in a superconducting qubit chip (2026), arXiv:2604.13176 [quant-ph]

    Pith/arXiv arXiv 2026

  9. [9]

    Cardaniet al., Disentangling the sources of ionizing radiation in superconducting qubits, Eur

    L. Cardaniet al., Disentangling the sources of ionizing radiation in superconducting qubits, Eur. Phys. J. C83, 94 (2023), arXiv:2211.13597 [quant-ph]

    Pith/arXiv arXiv 2023

  10. [10]

    C. D. Wilenet al., Correlated charge noise and relaxation errors in superconducting qubits, Nature594, 369 (2021), arXiv:2012.06029 [quant-ph]

    Pith/arXiv arXiv 2021

  11. [11]

    Thorbeck, A

    T. Thorbeck, A. Eddins, I. Lauer, D. T. McClure, and M. Carroll, Two-level-system dynamics in a supercon- ducting qubit due to background ionizing radiation, PRX Quantum4, 10.1103/prxquantum.4.020356 (2023)

    work page doi:10.1103/prxquantum.4.020356 2023

  12. [12]

    S. L. Glashow, Partial-symmetries of weak interactions, Nuclear Physics22, 579 (1961)

    1961

  13. [13]

    Weinberg, A Model of Leptons, Phys

    S. Weinberg, A Model of Leptons, Phys. Rev. Lett.19, 1264 (1967)

    1967

  14. [14]

    S. L. Glashow, J. Iliopoulos, and L. Maiani, Weak Inter- actions with Lepton-Hadron Symmetry, Phys. Rev. D2, 1285 (1970)

    1970

  15. [15]

    Gell-Mann, A schematic model of baryons and mesons, Phys

    M. Gell-Mann, A schematic model of baryons and mesons, Phys. Lett.8, 214 (1964)

    1964

  16. [16]

    Zweig, An SU(3) model for strong interaction symme- try and its breaking

    G. Zweig, An SU(3) model for strong interaction symme- try and its breaking. Version 2, inDEVELOPMENTS IN THE QUARK THEORY OF HADRONS. VOL. 1. 1964 - 1978, edited by D. B. Lichtenberg and S. P. Rosen (1964) p. 22

    1964

  17. [17]

    Bratrudet al., Measurement of correlated charge noise in superconducting qubits at an underground facility, Na- ture Commun.16, 9906 (2025), arXiv:2405.04642 [quant- ph]

    G. Bratrudet al., Measurement of correlated charge noise in superconducting qubits at an underground facility, Na- ture Commun.16, 9906 (2025), arXiv:2405.04642 [quant- ph]

    Pith/arXiv arXiv 2025

  18. [18]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, Imagenet classification with deep convolutional neural networks, inAdvances in Neural Information Processing Systems, Vol. 25, edited by F. Pereira, C. Burges, L. Bottou, and K. Weinberger (Curran Associates, Inc., 2012) pp. 1097— -1105

    2012

  19. [19]

    Dinget al.(QICK), Experimental advances with the QICK (Quantum Instrumentation Control Kit) for super- conducting quantum hardware, Phys

    C. Dinget al.(QICK), Experimental advances with the QICK (Quantum Instrumentation Control Kit) for super- conducting quantum hardware, Phys. Rev. Res.6, 013305 (2024), arXiv:2311.17171 [quant-ph]

    arXiv 2024

  20. [20]

    Di Guglielmoet al., End-to-End Workflow for Machine- Learning-Based Qubit Readout With QICK and hls4ml, IEEE Trans

    G. Di Guglielmoet al., End-to-End Workflow for Machine- Learning-Based Qubit Readout With QICK and hls4ml, IEEE Trans. Quantum Eng.6, 1 (2025), arXiv:2501.14663 [quant-ph]

    Pith/arXiv arXiv 2025

  21. [21]

    Lienhard, A

    B. Lienhard, A. Veps¨ al¨ ainen, L. C. G. Govia, C. R. Hof- fer, J. Y. Qiu, D. Rist` e, M. Ware, D. Kim, R. Winik, A. Melville, B. Niedzielski, J. Yoder, G. J. Ribeill, T. A. Ohki, H. K. Krovi, T. P. Orlando, S. Gustavs- son, and W. D. Oliver, Deep neural network discrimina- tion of multiplexed superconducting qubit states (2021), arXiv:2102.12481 [quant-ph]

    Pith/arXiv arXiv 2021

  22. [22]

    Navarathna, T

    R. Navarathna, T. Jones, T. Moghaddam, A. Kulikov, R. Beriwal, M. Jerger, P. Pakkiam, and A. Fedorov, Neu- ral networks for on-the-fly single-shot state classification, Applied Physics Letters119, 10.1063/5.0065011 (2021)

    work page doi:10.1063/5.0065011 2021

  23. [23]

    jae Jwa, G

    Y. jae Jwa, G. D. Guglielmo, L. Arnold, L. Carloni, and G. Karagiorgi, Real-time inference with 2d convo- lutional neural networks on field programmable gate arrays for high-rate particle imaging detectors (2022), arXiv:2201.05638 [physics.ins-det]

    Pith/arXiv arXiv 2022

  24. [24]

    V. Yon, B. Galaup, C. Rohrbacher, J. Rivard, A. Morel, D. Leclerc, C. Godfrin, R. Li, S. Kubicek, K. D. Greve, E. Dupont Ferrier, Y. Beilliard, R. G. Melko, and D. Drouin, Experimental online quantum dots charge autotuning using neural networks, Nano Letters25, 3717– (2025)

    2025

  25. [25]

    Samaha, A

    P. Samaha, A. Torki, Y. Renaud, S. Fiette, E. Chanrion, P.-A. Mortemousque, and Y. Beilliard, Automatic charge state tuning of 300 mm fdsoi quantum dots using neural network segmentation of charge stability diagram (2026), arXiv:2604.13662 [cond-mat.mes-hall]

    Pith/arXiv arXiv 2026

  26. [26]

    S. Bai, J. Z. Kolter, and V. Koltun, An empirical eval- uation of generic convolutional and recurrent networks for sequence modeling, CoRRabs/1803.01271(2018), 1803.01271

    Pith/arXiv arXiv 2018

  27. [27]

    Cholletet al., Keras,https://keras.io(2015)

    F. Cholletet al., Keras,https://keras.io(2015)

    2015

  28. [28]

    C. N. Coelho, A. Kuusela, S. Li, H. Zhuang, J. Ngadiuba, T. K. Aarrestad, V. Loncar, M. Pierini, A. A. Pol, and S. Summers, Automatic heterogeneous quantization of deep neural networks for low-latency inference on the edge for particle detectors, Nature Machine Intelligence3, 675 (2021)

    2021

  29. [29]

    C. N. Coelhoet al., QKeras: a quantization deep learn- ing library for Tensorflow Keras, https://github.com/ google/qkeras(2019)

    2019

  30. [30]

    Jacob, S

    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, Quan- tization and training of neural networks for efficient integer-arithmetic-only inference (2017), arXiv:1712.05877 [cs.LG]

    Pith/arXiv arXiv 2017

  31. [31]

    J. Chen, Y. Gai, Z. Yao, M. W. Mahoney, and J. E. Gon- zalez, A statistical framework for low-bitwidth training of deep neural networks, inAdvances in Neural Infor- mation Processing Systems (NeurIPS), Vol. 33 (Curran Associates, Inc., 2020) pp. 1823–1834

    2020

  32. [32]

    J. Chen, L. Zheng, Z. Yao, D. Wang, I. Stoica, M. W. Mahoney, and J. E. Gonzalez, Actnn: Reducing training 9 memory footprint via 2-bit activation compressed train- ing, inProceedings of the 38th International Conference on Machine Learning (ICML), Proceedings of Machine Learning Research, Vol. 139 (PMLR, 2021) pp. 1803–1813

    2021

  33. [33]

    is an open-source qubit control platform built on Xil- inx RFSoC field-programmable gate arrays (FPGAs) [ 34]. Originally introduced in 2022 [ 35] and subsequently up- graded to support higher-bandwidth RFSoC chips [ 19], QICK provides a unified hardware and software framework for superconducting qubit control and readout, requiring only external amplific...

    2022

  34. [34]

    Fermi National Accelerator Laboratory, QICK: Quan- tum Instrumentation Control Kit, https://github.com/ openquantumhardware/qick (2024), open-source soft- ware, accessed 2026-07-15

    2024

  35. [35]

    QICK: Quantum Instrumentation Control Kit, https: //docs.qick.dev, accessed: 2026-01-06

    2026

  36. [36]

    Stefanazziet al., The QICK (Quantum Instrumen- tation Control Kit): Readout and control for qubits and detectors, Rev

    L. Stefanazziet al., The QICK (Quantum Instrumen- tation Control Kit): Readout and control for qubits and detectors, Rev. Sci. Instrum.93, 044709 (2022), arXiv:2110.00557 [quant-ph]

    Pith/arXiv arXiv 2022

  37. [37]

    Duarteet al., Fast inference of deep neural networks in FPGAs for particle physics, JINST13(07), P07027, arXiv:1804.06913 [physics.ins-det]

    J. Duarteet al., Fast inference of deep neural networks in FPGAs for particle physics, JINST13(07), P07027, arXiv:1804.06913 [physics.ins-det]

    Pith/arXiv arXiv

  38. [38]

    Krantz, M

    P. Krantz, M. Kjaergaard, F. Yan, T. P. Orlando, S. Gus- tavsson, and W. D. Oliver, A quantum engineer’s guide to superconducting qubits, Applied Physics Reviews6, 021318 (2019)

    2019

  39. [39]

    (2017), ver- sion 4.0

    AXI Reference Guide (UG1037), Xilinx Inc. (2017), ver- sion 4.0

    2017

  40. [40]

    (2011), aRM IHI 0022E

    AMBA AXI and ACE Protocol Specification, ARM Ltd. (2011), aRM IHI 0022E

    2011