Pith. sign in

REVIEW 3 major objections 5 minor 42 references

The paper shows that isolated double-quantum-dot charge stability maps can be read automatically for electron occupancy with two compact convolutional networks, reaching 95.3% exact line counts on held-out devices and 93.8% end-to-end accur

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 ↗

A quality-screening CNN (94% accuracy) and a line-counting CNN (95.3% exact-count accuracy on 1,131 held-out images) automate charge-state readout of isolated double quantum dots.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid empirical contribution to isolated-mode quantum dot autotuning; the held-out-device evaluation is the real deal, and the annotator caveat is real but already half-acknowledged. the 3 major comments →

arxiv 2607.20871 v1 pith:KHWTO2CS submitted 2026-07-23 cond-mat.mes-hall cs.LG

Machine Learning for Charge State Characterization of Isolated Double Quantum Dots

Hyma Vallabhapurapu , Marco Candido , Krishna Choudhary , Paul Steinacker , Ensar Vahapoglu , Chris Escott , Wee Han Lim , Andre Saraiva
show 2 more authors
Nard Dumoulin Stuyck MengKe Feng
This is my paper
classification cond-mat.mes-hall cs.LG PACS 85.35.Gv07.05.Mh
keywords isolated-mode quantum dotscharge stability mapelectron occupancy readoutconvolutional neural networkline detectionsynthetic pre-trainingcross-device generalizationsilicon spin qubits
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 reading

The paper argues that the electron occupancy of double quantum dots operated in isolated mode can be read automatically from charge stability maps by two small convolutional networks: one that screens images for charge instability and sensor artefacts, and one that detects and counts the near-vertical charge-transition lines. On 16 devices never used in training, the line counter achieves 95.3% exact line-count accuracy, and the combined pipeline returns the correct electron count for 93.8% of clean held-out images. The paper also shows that pre-training the line counter on cheap synthetic images and fine-tuning on a small amount of hand-labelled data keeps accuracy above 90% even with only 5% of the labels, whereas training from scratch collapses. If correct, this replaces a major manual bottleneck in tuning silicon spin-qubit devices, since the models are compact and fast enough to run in a laboratory setting.

Core claim

The central claim is that the characteristic phenomenology of isolated-mode charge stability maps—fixed total charge, near-vertical transition lines, and parasitic-dot branching—is a learnable image-analysis problem that two task-specific CNNs can solve across devices. CSMClassifier, an encoder-only multi-label network, assigns each image clean/unstable/unclear labels, reaching 94% macro-averaged accuracy on 2,407 held-out images. ChargeLineNet, a U-Net with anisotropic branches, outputs a signed heatmap of line start and end points plus offset vectors that pair each end with its start, and reads off the electron count; it reaches 95.3% exact-count accuracy and 98.6% within one line on 1,131

What carries the argument

The load-bearing object is ChargeLineNet's signed-heatmap-plus-offset-vector output: positive Gaussian blobs mark each transition line's sharp start, negative blobs mark its diffuse end, and per-pixel offset vectors anchored at each end point point back to the paired start, so the decoder recovers electron occupancy by counting paired lines without a combinatorial matching step that fails when blobs merge. The network's anisotropic encoder branches (7x1, 5x5, 1x7, 3x3, dilated 3x3, and 1x15 kernels) make it orientation-selective for the near-vertical plunger-gate lines while suppressing branching features from parasitic dots, and a histogram-invariant stem removes contrast differences betwee

Load-bearing premise

The reported accuracies rest on the hand-labelled ground truth being correct and representative; the paper states that inter-annotator agreement has not been quantified, and the line-counter test set contains only the human-countable subset of images, so if annotators disagree or live data often falls outside that subset, deployed accuracy could be lower.

What would settle it

Take a random subset of held-out charge stability maps and have two or more independent trained annotators label the line counts; if per-image agreement between annotators is well below the reported 95.3% accuracy, the ground truth itself cannot pin down the claim. Alternatively, run the pipeline live on new devices and compare accepted electron counts against counts verified by subsequent Pauli-spin-blockade measurements.

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

If this is right

  • The reported 93.8% end-to-end occupancy accuracy on clean held-out images implies that isolated-mode DQD charge-state readout can be automated in the low-occupancy regime relevant to Pauli-spin-blockade search.
  • Because both models are validated on 16 entirely held-out devices spanning deliberate design differences, the learned features transfer across devices rather than memorizing device-specific noise.
  • Synthetic pre-training plus fine-tuning keeps accuracy above 90% even with 5% of the labels, implying that adapting the pipeline to a new device generation or sensor design needs far less hand-labelling than training from scratch.
  • With a combined footprint of 6.5 MB and inference under 60 ms on a consumer CPU, the models are light enough to sit inside a measure-infer-adjust tuneup loop on standard laboratory hardware.
  • The classical baseline comparison (best 61% exact-count accuracy) implies that the task genuinely requires learned orientation selectivity; a hand-engineered line counter cannot distinguish plunger-gate lines from parasitic-dot branching.

Where Pith is reading between the lines

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

  • The authors evaluate offline, not in a live loop; an implication is that in deployment, flagged images would trigger re-measurement rather than being scored as errors, so the practical yield could differ from the reported 93.8%—likely higher if re-measurement succeeds, but that is not measured here.
  • Since multidot tuning typically sweeps two plunger gates at a time, the same DQD line counter could plausibly apply to pairwise charge stability maps from larger arrays; the paper notes the DQD data basis but does not test this directly.
  • The paper suggests a vision-language model could automate labelling; a testable extension is to replace human annotations with VLM-proposed start/end coordinates and verify that downstream accuracy holds, which would also quantify inter-annotator agreement.
  • Because fine-tuned accuracy plateaus between 20% and 100% of labels, an untested implication is that broadening the labelled set across device geometries and artefact types will improve generalization more than adding more images from the same devices.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper presents two compact CNN models for automated analysis of charge stability maps (CSMs) from isolated-mode double quantum dots: CSMClassifier, a multi-label quality gate (clean/unstable/unclear), and ChargeLineNet, a signed-heatmap plus offset-vector line detector that counts charge-transition lines to read electron occupancy. Both models are trained on 16 SiMOS DQD devices and evaluated on 16 fully held-out devices against hand-labelled ground truth. The reported results are 94% macro-averaged quality classification accuracy on 2,407 held-out images, 95.3% exact line-count accuracy on the 1,131-image human-countable subset, and 93.8% end-to-end occupancy accuracy on the 693 clean held-out images that carry a reliable count. The paper also demonstrates that synthetic pre-training provides strong label efficiency during fine-tuning, reports a tuned classical baseline (61% exact count), and shows that the deployed models are compact and fast enough for lab hardware.

Significance. If the quantitative claims hold, this is a useful and timely contribution. Automated tuneup of quantum dot devices is an active bottleneck, and the isolated-mode regime is genuinely less studied than the reservoir-coupled case. The evaluation design has real strengths: held-out devices are used for all headline metrics, bootstrap confidence intervals are reported for the classifier, the line-count accuracy is shown to be insensitive to decoding knobs, and the classical baseline is tuned on a training-only subsample on the same cross-device footing. The label-efficiency experiment with synthetic pre-training is also a clear and practically relevant result. The main weakness is that every headline accuracy is scored against hand labels whose reliability is not demonstrated, and the line-count evaluation population is itself defined by the annotators' ability to count the images. This is a correctness risk rather than a presentation issue.

major comments (3)
  1. [Section II (data), Table III, and Section III.C.d (Limitations)] The ground-truth labels are the load-bearing reference for every headline accuracy, yet inter-annotator agreement is not quantified. The line-detection validation set is defined as the 1,131 images 'that a human could reliably count,' and Section III.C.d concedes that hand-labelling is 'prone to mistakes and to inconsistency' and that inter-annotator agreement has not been formally quantified. Because the same labelling process defines both the training targets and the evaluation set, a model can score highly by matching one annotator's conventions without implying robust cross-annotator or deployment performance. Please add a second-annotation audit on a random subset of the held-out images (e.g., 200-300 images), reporting per-class agreement for CSMClassifier and exact-count agreement for ChargeLineNet, and quantify how often the 'human-countable' decision itself varies between annota
  2. [Section III.C.b and Section II (pipeline evaluation)] The end-to-end '93.9% of all held-out images' claim is ambiguous and not directly measured as stated. Only 693 of the 2,407 held-out images are clean with a ground-truth count; non-clean images have no line-count ground truth, so their 'correct rejection' is scored by construction rather than by the line detector. The 93.8% clean-image figure is also based on 693 images, but no confidence interval is given for it and the exact scoring rule (what happens to images the gate wrongly passes, and how non-clean images enter the denominator) is not specified. Please state the scoring rule explicitly, report the 95% bootstrap CI for the 93.8% figure, and give a pipeline-level breakdown (clean images correctly counted, clean images rejected by the gate, non-clean images incorrectly passed).
  3. [Section III.B.a, Fig. 7a] The per-occupancy accuracy bars for the 6- and 7-line bins are based on n=53 and n=21 images, respectively, and the text appropriately cautions that these are indicative. Nevertheless, the figure plots them as solid bars on the same scale as well-sampled bins. Please add error bars or a marked 'low-n' annotation to these bins so that the reader is not visually misled by the apparent flatness of the per-count accuracy curve.
minor comments (5)
  1. [Appendix D2, Eq. (D4)] The loss expression uses 'SmoothL1' without defining it. Please add one line defining smooth-ℓ1 and note the default β value used.
  2. [Section II.A.a / Appendix D1] The per-class positive weights w_k in Eq. (D1) are described but their actual values are not reported. Since the class distribution is highly imbalanced (Table I), giving the weights used in training would improve reproducibility.
  3. [Section III.B.a] The phrase 'roughly 34 percentage-point improvement over the synthetic baseline' is correct (95.3 - 61.4 = 33.9), but earlier in the Introduction the same value is attached to the classical Hough baseline (61%). This coincidence is intentional, but the wording in the Introduction should be checked so that readers do not confuse the two baselines.
  4. [Appendix H] The freeze-mode section states that 258,659 of 935,283 parameters are trainable (about 28%), which matches the text 'only around a quarter.' However, the statement 'the task-specific decoder and output head are retrained' is slightly incomplete because the encoder is frozen, so the trained portion includes the decoder and head only. A one-sentence clarification would remove ambiguity.
  5. [Section III.C.d] The discussion of deployment latency correctly separates forward-pass time from decoding time, but the CPU latency for ChargeLineNet is quoted as 'under 50ms' with no figure or hardware details for the decoding step beyond '~2ms'. Please state whether this is a wall-clock measurement or an estimate.

Circularity Check

0 steps flagged

No significant circularity: empirical held-out benchmark with independent hand labels; self-citations are infrastructural, not load-bearing.

full rationale

This paper is an empirical benchmark rather than a derivation. CSMClassifier and ChargeLineNet are trained on images from 16 devices and evaluated on images from 16 fully held-out devices against hand-labelled ground truth; no reported metric is the result of fitting a parameter to the evaluation set. The pipeline figure (93.8%) is explicitly the product of the two stages' independent accuracies (0.96 x 0.979), a decomposition rather than a circular reuse of the output as input. The only self-citations are to the authors' earlier automated cryogenic probing platform and prior U-Net segmentation work [27], used to describe data collection and motivation; the cited work does not contain either model, its training labels, or the accuracy numbers, and the U-Net architecture itself is credited to Ronneberger et al. The synthetic pre-training comparison is controlled on the same held-out set, and the label-efficiency experiment varies the training fraction while keeping the evaluation set fixed and external to training. The paper's own stated limitation that inter-annotator agreement has not been formally quantified, and that ChargeLineNet is evaluated on the 'human-countable' subset of held-out images, is a legitimate concern about ground-truth reliability and test-set representativeness, but it is not circularity: the hand labels are external to the model and are not defined by the model's outputs. No equation, fitted parameter, or self-citation is used as evidence for the central predictive claims in a way that reduces those claims to their inputs. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The central claims rest on the quality of human annotations and on the transferability of the synthetic generator, not on new physics or invented entities. Free parameters are standard ML hyperparameters and decode thresholds, which the paper shows do not drive the headline numbers (decode-configuration spread <1%).

free parameters (3)
  • Sigmoid decision threshold τ for CSMClassifier = 0.5
    Used to decide active quality labels; reported results at τ=0.5; per-class accuracy varies with threshold (Fig. 6a) but the model is not fitted to the validation set.
  • Heatmap detection thresholds (noise-anchored median + 5σ; 0.25 signal ceiling; 0.5 hysteresis confirmation) = 0.25 / 0.5
    Hand-tuned in Appendix F; paper reports <1% spread in line-count accuracy across five decode configurations, so these are not load-bearing for the headline numbers.
  • Minimum line length and angle filter thresholds = configurable, not numerically specified
    Used in decoding (Appendix F); part of the post-processing pipeline.
axioms (3)
  • domain assumption Hand-labelled ground truth is accurate and consistent across annotators.
    The held-out validation metrics assume the human labels are correct; the paper notes inter-annotator agreement is unquantified (Sec. IIIC.d).
  • domain assumption The synthetic image generator captures the features of experimental CSMs relevant to line counting.
    The label-efficiency result depends on transfer from synthetic to real data; the simulator is parameterised to reproduce salient features (Appendix I).
  • standard math Standard CNN training and stochastic optimisation behave as expected (convergence, generalisation).
    The paper relies on standard deep learning machinery.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine Learning for Charge State Characterization of Isolated Double Quantum Dots." pith.science (2026). https://pith.science/paper/KHWTO2CS

@misc{pith2026260720871,
  author       = {Pith},
  title        = {Pith review of: Machine Learning for Charge State Characterization of Isolated Double Quantum Dots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KHWTO2CS}},
  note         = {Machine review of arXiv:2607.20871}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Scaling semiconductor quantum dot arrays toward fault-tolerant quantum computing requires efficient tuneup of spin qubits, a process that depends on the analysis of charge stability maps (CSMs) and remains largely manual. While machine learning has been widely applied to CSM analysis in reservoir-coupled devices, automated tuning in the increasingly important isolated-mode regime has received limited attention. In isolated-mode CSMs, charge transitions appear as near-vertical lines, making them well suited to compact, task-specific models. We present two convolutional neural networks with fewer than one million parameters, trained on CSMs collected from 32 silicon metal-oxide-semiconductor (SiMOS) double-quantum-dot devices measured at approximately 1 K using an automated cryogenic probing system. Sixteen devices were used for training and sixteen were held out to evaluate cross-device generalization against hand-labeled ground truth. CSMClassifier identifies charge instability and sensor artifacts, achieving 94% macro-averaged accuracy across three quality classes on 2,407 held-out images. ChargeLineNet localizes charge-transition lines and determines electron occupancy, achieving 95.3% exact line-count accuracy on 1,131 held-out images. Combined into a single pipeline, the models correctly determine electron occupancy for 93.8% of clean held-out images. Pre-training on synthetic images substantially improves label efficiency. Fine-tuning the pre-trained model on limited experimental data maintains over 90% accuracy, whereas training from scratch degrades significantly under the same conditions. Together, the two models occupy only 6.5 MB and process images in less than 60 ms on standard laboratory hardware, demonstrating a practical path toward scalable, automated characterization and tuneup of quantum-dot devices.

Figures

Figures reproduced from arXiv: 2607.20871 by Andre Saraiva, Chris Escott, Ensar Vahapoglu, Hyma Vallabhapurapu, Krishna Choudhary, Marco Candido, MengKe Feng, Nard Dumoulin Stuyck, Paul Steinacker, Wee Han Lim.

Figure 1
Figure 1. Figure 1: Overview of the double-quantum-dot (DQD) device and its charge stability map. (a) Representative Scanning [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: CSMClassifier architecture visualised on real CSM images, one row per quality class (clean, unstable, unclear). Each row passes a hand-labelled experimental CSM of that class through the trained network and shows the actual intermediate feature maps (only the single highest-variance, i.e. most active, channel is shown at each stage; the remaining channels are omitted to avoid clutter): input → histogram-in… view at source ↗
Figure 3
Figure 3. Figure 3: ChargeLineNet architecture visualised on a real CSM image. A hand-labelled experimental CSM is passed through the trained network and the actual intermediate feature maps are shown left to right (only the single highest-variance, i.e. most active, channel is shown at each stage; the remaining channels are omitted to avoid clutter): input → histogram-invariant stem (LCN ∥ InstanceNorm) → three anisotropic e… view at source ↗
Figure 4
Figure 4. Figure 4: Representative CSM images from the three quality classes, one class per column (three examples each). [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: ChargeLineNet outputs for three example CSM images (one per row). Column 1 (a, d, g, j): preprocessed input CSM. Column 2 (b, e, h, k): predicted signed heatmap (channel 0), with positive peaks marking line start points and negative peaks marking line end points (colour bar). Column 3 (c, f, i, l): decoded charge-transition lines overlaid on the input, comparing the ground-truth annotation (GT) with the mo… view at source ↗
Figure 6
Figure 6. Figure 6: CSMClassifier performance on the 2,407-image held-out validation set. (a) Per-class accuracy as a function of the sigmoid decision threshold τ ; the dashed vertical line marks τ = 0.5, and the legend reports the per-class accuracy at that threshold (clean 0.95, unstable 0.96, unclear 0.92). (b–d) Binary confusion matrices at τ = 0.5 for the clean, unstable, and unclear classes respectively, each annotated … view at source ↗
Figure 7
Figure 7. Figure 7: Performance of the three ChargeLineNet training variants on the 1,131-image held-out validation set. The variants are: synthetic, trained on synthetic data only; fine-tuned, i.e. pre-trained on synthetic data and then adapted on experimental data with the encoder frozen; and scratch, trained on experimental data from scratch. (a) Exact line-count accuracy broken down by ground-truth electron occupancy (0–7… view at source ↗
Figure 8
Figure 8. Figure 8: Data efficiency of synthetic pre-training for [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: CSMClassifier encoder-only architecture (see section II A). Spatial dimensions after each stage are shown below each block. The histogram-invariant stem (LCN ∥ InstanceNorm) is identical to ChargeLineNet. The stem projects to 16 channels. Four encoder stages use isotropic 3×3 residual blocks (two convolutions with batch normalisation, ReLU, and a 1×1 projection shortcut) followed by 2×2 max-pooling, progre… view at source ↗
Figure 10
Figure 10. Figure 10: ChargeLineNet encoder–decoder architecture (see section II B). Spatial dimensions after each stage are shown below each block. The histogram-invariant stem (LCN ∥ InstanceNorm) removes global contrast variation. Three encoder levels use six-branch anisotropic blocks (7×1, 5×5, 1×7, 3×3, dilated 3×3, 1×15) with DropPath (p=0.1) and 2×2 max-pooling. The bottleneck adds Dropout2d (p=0.2). Three decoder level… view at source ↗
Figure 11
Figure 11. Figure 11: Response of the six anisotropic branches in the first [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Representative synthetic CSM images produced by the parameterised image simulator and used to pre-train [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative outcomes of the classical line counter (probabilistic Hough transform with the bilateral-denoise [PITH_FULL_IMAGE:figures/full_fig_p023_13.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

42 extracted references · 2 linked inside Pith

  1. [1]

    Twoparallelbranches process the single input channel:

    Shared Histogram-Invariant Stem Both models begin with the same histogram-invariant stem, which renders the network insensitive to global brightness and contrast differences between training (simu- lated)andtest(experimental)data. Twoparallelbranches process the single input channel:

  2. [2]

    This removes any DC offset and gain variation while preserving edge structure

    Local contrast normalisation (LCN).A fixed 9 × 9Gaussian kernel (σ = 1.5px) blurs the input; theblurredimageissubtractedfromtheoriginaland the result is divided by the local standard deviation. This removes any DC offset and gain variation while preserving edge structure

  3. [3]

    InstanceNorm2d nor- malises each image independently to zero mean and unit variance, providing a complementary global contrast normalisation

    Instance normalisation. InstanceNorm2d nor- malises each image independently to zero mean and unit variance, providing a complementary global contrast normalisation. The two branch outputs are concatenated and projected via a3×3convolution followed bybatch normalisation and ReLU. The only difference between the two models is the projection width:c 1 = 16c...

  4. [4]

    Exact” and “±1

    is averaged over the blob region, weighted by blob intensity, to obtain a robust estimate of the end-to-start displacement. The predicted start position is computed as the blob centroid plus the averaged offset. This offset- predicted start is then used as the seed for a snapping step: a corridor is searched along the offset direction and the seed is snap...

  5. [5]

    Loss and D

    D. Loss and D. P. DiVincenzo, Quantum computation with quantum dots, Phys. Rev. A57, 120 (1998)

    1998

  6. [6]

    Elzerman, R

    J. Elzerman, R. Hanson, L. Willems van Beveren, B. Witkamp, L. Vandersypen, and L. P. Kouwenhoven, Single-shot read-out of an individual electron spin in a quantum dot, nature430, 431 (2004)

    2004

  7. [7]

    J. R. Petta, A. C. Johnson, J. M. Taylor, E. A. Laird, A. Yacoby, M. D. Lukin, C. M. Marcus, M. P. Hanson, and A. C. Gossard, Coherent manipulation of coupled electron spins in semiconductor quantum dots, Science 309, 2180 (2005)

    2005

  8. [8]

    S. J. Angus, A. J. Ferguson, A. S. Dzurak, and R. G. Clark, Gate-defined quantum dots in intrinsic silicon, Nano letters7, 2051 (2007)

    2051

  9. [9]

    C. H. Yang, R. Leon, J. Hwang, A. Saraiva, T. Tanttu, W. Huang, J. Camirand Lemyre, K. W. Chan, K. Tan, F. E. Hudson,et al., Operation of a silicon quantum processor unit cell above one kelvin, Nature580, 350 (2020)

    2020

  10. [10]

    Petit, H

    L. Petit, H. Eenink, M. Russ, W. Lawrie, N. Hendrickx, S. Philips, J. Clarke, L. Vandersypen, and M. Veldhorst, Universal quantum logic in hot silicon qubits, Nature580, 355 (2020)

    2020

  11. [11]

    Dumoulin Stuyck, A

    N. Dumoulin Stuyck, A. Saraiva, W. Gilbert, J. Ci- fuentes Pardo, R. Li, C. C. Escott, K. De Greve, S. Voinigescu, D. J. Reilly, and A. S. Dzurak, Cmos com- patibility of semiconductor spin qubits, Nature Reviews Electrical Engineering , 300 (2026)

    2026

  12. [12]

    S. K. Bartee, W. Gilbert, K. Zuo, K. Das, T. Tanttu, C. H. Yang, N. Dumoulin Stuyck, S. J. Pauka, R. Y. Su, W. H. Lim,et al., Spin-qubit control with a milli-kelvin cmos chip, Nature643, 382 (2025)

    2025

  13. [13]

    Acuna, T

    Members of the HRL Quantum Team, Collaborators: Michael Abraham, E. Acuna, T. S. Adams, M. Akmal, M. R. Alfaro, I. Alvarado, J. Amontree, C. Andrews, R. W. Andrews, M. Antcliffe, A. R. Aséncio, R. M. Avila Ba- tres, C. D. Baringer, D. W. Barnes, K. M. Beech, R. G. Blakey, Z. T. Bloom, A. J. Bluestone, J. Z. Blumoff, M. G. Borselli, K. A. Bose, B. Boyd, J....

    Pith/arXiv arXiv 2026

  14. [14]

    Steinacker, N

    P. Steinacker, N. Dumoulin Stuyck, W. H. Lim, T. Tanttu, M.Feng, S.Serrano, A.Nickl, M.Candido, J.D.Cifuentes, E. Vahapoglu,et al., Industry-compatible silicon spin- qubit unit cells exceeding 99% fidelity, Nature646, 81 (2025)

    2025

  15. [15]

    Nickl, N

    A. Nickl, N. Dumoulin Stuyck, P. Steinacker, J. D. Ci- fuentes, S. Serrano, M. Feng, E. Vahapoglu, F. E. Hud- son, K. W. Chan, S. Kubicek, J. Jussot, Y. Canvel, S. Beyne, Y. Shimura, R. Loo, C. Godfrin, B. Raes, S. Baudot, D. Wan, A. Laucht, C.-H. Yang, W. H. Lim, A. Saraiva, C. C. Escott, K. De Greve, A. S. Dzurak, and T. Tanttu, Eight-qubit operation of...

    2026

  16. [16]

    J. P. Zwolak, J. M. Taylor, R. W. Andrews, J. Benson, G. W. Bryant, D. Buterakos, A. Chatterjee, S. Das Sarma, M. A. Eriksson, E. Greplová, M. J. Gullans, F. Hader, T. J. Kovach, P. S. Mundada, M. Ramsey, T. Rasmussen, B. Severin, A. Sigillito, B. Undseth, and B. Weber, Data needs and challenges for quantum dot devices automation, NPJ Quantum Inf.10(2024)

    2024

  17. [17]

    J. P. Zwolak and J. M. Taylor, Colloquium: Advances in automation of quantum dot devices control, Reviews of modern physics95, 011006 (2023)

    2023

  18. [18]

    Samaha, A

    P. Samaha, A. Torki, Y. Renaud, S. Fiette, E. Chan- 25 rion, P.-A. Mortemousque, and Y. Beilliard, Automatic charge state tuning of 300 mm silicon quantum dots using neural network segmentation of charge stability diagram, Machine Learning: Science and Technology (2026)

    2026

  19. [19]

    Y. Muto, M. R. Zielewski, M. Shinozaki, K. Noro, and T. Otsuka, Automatic detection of single-electron regime and virtual gate definition in quantum dots using u-net and clustering, Scientific Reports16, 8161 (2026)

    2026

  20. [20]

    S. S. Kalantre, J. P. Zwolak, S. Ragole, X. Wu, N. M. Zimmerman, M. Stewart Jr, and J. M. Taylor, Machine learning techniques for state recognition and auto-tuning in quantum dots, npj Quantum Information5, 6 (2019)

    2019

  21. [21]

    Durrer, B

    R. Durrer, B. Kratochwil, J. V. Koski, A. J. Landig, C. Reichl, W. Wegscheider, T. Ihn, and E. Greplova, Automated tuning of double quantum dots into specific charge states using neural networks, Physical Review Applied13, 054019 (2020)

    2020

  22. [22]

    J. P. Zwolak, T. McJunkin, S. S. Kalantre, J. Dodson, E. MacQuarrie, D. Savage, M. Lagally, S. Coppersmith, M. A. Eriksson, and J. M. Taylor, Autotuning of double- dot devices in situ with machine learning, Physical review applied13, 034075 (2020)

    2020

  23. [23]

    Schuff, M

    J. Schuff, M. J. Carballido, M. Kotzagiannidis, J. C. Calvo, M. Caselli, J. Rawling, D. L. Craig, B. van Straaten, B. Severin, F. Fedele,et al., Fully autonomous tuning of a spin qubit, Nature Electronics , 1 (2026)

    2026

  24. [24]

    Hader, F

    F. Hader, F. Fuchs, S. Fleitmann, K. Havemann, B. Scherer, J. Vogelbruch, L. Geck, and S. van Waasen, Automated charge transition detection in quantum dot charge stability diagrams, IEEE Trans. Quantum Eng.6, 1 (2025)

    2025

  25. [25]

    Darulová, S

    J. Darulová, S. J. Pauka, N. Wiebe, K. W. Chan, G. C. Gardener, M. J. Manfra, M. C. Cassidy, and M. Troyer, Autonomous tuning and charge-state detection of gate- defined quantum dots, Phys. Rev. Appl.13(2020)

    2020

  26. [26]

    V. Yon, B. Galaup, C. Rohrbacher, J. Rivard, A. Morel, D. Leclerc, C. Godfrin, R. Li, S. Kubicek, K. D. Greve, et al., Experimental online quantum dots charge autotun- ing using neural networks, Nano Letters25, 3717 (2025)

    2025

  27. [27]

    M. P. R. Losert, D. Denora, B. van Straaten, M. Chan, S. D. Oosterhout, L. Stehouwer, G. Scappucci, M. Veld- horst, and J. P. Zwolak, Automated electrostatic charac- terization of quantum dot devices in single- and bilayer heterostructures (2026), arXiv:2601.00067 [cond-mat.mes- hall]

    arXiv 2026

  28. [28]

    Díaz Moreno, R

    V. Díaz Moreno, R. P. Khalili, D. Schug, P. J. Walsh, and J. P. Zwolak, Benchmarking machine learning models for multi-class state recognition in double quantum dot data (2025), arXiv:2511.22451 [cs.CV]

    arXiv 2025

  29. [29]

    A. C. Johnson, C. M. Marcus, M. P. Hanson, and A. C. Gossard, Charge sensing of excited states in an isolated double quantum dot, Phys. Rev. B Condens. Matter Mater. Phys.71(2005)

    2005

  30. [30]

    Bertrand, H

    B. Bertrand, H. Flentje, S. Takada, M. Yamamoto, S. Tarucha, A. Ludwig, A. D. Wieck, C. Bäuerle, and T. Meunier, Quantum manipulation of two-electron spin states in isolated double quantum dots, Phys. Rev. Lett. 115, 096801 (2015)

    2015

  31. [31]

    Candido, E

    M. Candido, E. Vahapoglu, J. Dehollain, K. Simoes, H. Vallabhapurapu, K. Chan, T. Tanttu, P. Steinacker, W. Gilbert, S. Serrano,et al., Investigation of 300mm pro- cess simos spin qubit device uniformity with automated cryogenic probing, in2025 IEEE International Electron Devices Meeting (IEDM)(IEEE, 2025) pp. 1–4

    2025

  32. [32]

    Ziegler, T

    J. Ziegler, T. McJunkin, E. S. Joseph, S. S. Kalantre, B. Harpt, D. E. Savage, M. G. Lagally, M. A. Eriksson, J. M. Taylor, and J. P. Zwolak, Toward robust autotuning of noisy quantum dot devices, Phys. Rev. Appl.17, 024069 (2022)

    2022

  33. [33]

    Oquab, L

    M. Oquab, L. Bottou, I. Laptev, and J. Sivic, Learning and transferring mid-level image representations using convolutional neural networks, in2014 IEEE Conference on Computer Vision and Pattern Recognition(2014) pp. 1717–1724

    2014

  34. [34]

    Czischek, V

    S. Czischek, V. Yon, M.-A. Genest, M.-A. Roux, S. Ro- chette, J. Camirand Lemyre, M. Moras, M. Pioro-Ladrière, D. Drouin, Y. Beilliard, and R. G. Melko, Miniaturizing neural networks for charge state autotuning in quantum dots, Mach. Learn. Sci. Technol.3, 015001 (2022)

    2022

  35. [35]

    Darulová, M

    J. Darulová, M. Troyer, and M. C. Cassidy, Evaluation of synthetic and experimental training data in supervised machine learning applied to charge-state detection of quantum dots, Machine Learning: Science and Technology 2, 045023 (2021)

    2021

  36. [36]

    N. S. Lai, W. Lim, C. Yang, F. Zwanenburg, W. Coish, F. Qassemi, A. Morello, and A. Dzurak, Pauli spin block- ade in a highly tunable silicon double quantum dot, Sci- entific reports1, 110 (2011)

    2011

  37. [37]

    A. E. Seedhouse, T. Tanttu, R. C. Leon, R. Zhao, K. Y. Tan, B. Hensen, F. E. Hudson, K. M. Itoh, J. Yoneda, C. H. Yang,et al., Pauli blockade in silicon quantum dots with spin-orbit control, PRX quantum2, 010303 (2021)

    2021

  38. [38]

    X. Zhou, D. Wang, and P. Krähenbühl, Objects as points, arXiv preprint arXiv:1904.07850 (2019)

    Pith/arXiv arXiv 1904

  39. [39]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox, U-net: Con- volutional networks for biomedical image segmentation, inInternational Conference on Medical image computing and computer-assisted intervention(Springer, 2015) pp. 234–241

    2015

  40. [40]

    S. Cao, Z. Zhang, A. Agarwal, G. Bratrud, N. R. Beysen- gulov, D. C. Cole, A. Gómez Frieiro, E. O. Glen, H. Hsu, G. Huang, R. Jow, G. Shaji, T. Lubowe, L. Zhu, L. Man- tilla Calderón, N. Pancotti, J. Pendleton, B. Severin, C. E. Staub, S. Sussman, A. Vepsäläinen, N. R. Vora, Y. Xu, V. Bernales, D. Bowring, E. Kyoseva, I. Rungger, G. Se- meghini, S. Stanwy...

    2026

  41. [41]

    K. He, X. Zhang, S. Ren, and J. Sun, Deep residual learning for image recognition, inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)(2016) pp. 770–778

    2016

  42. [42]

    Szegedy, W

    C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabi- novich, Going deeper with convolutions, inProceedings of the IEEE conference on computer vision and pattern recognition(2015) pp. 1–9

    2015

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.