End-to-End Analysis of Charge Stability Diagrams with Transformers

Barnaby van Straaten; Cornelius Carlsson; Federico Fedele; Jonas Schuff; Joshua Ziegler; Lucas Schorling; Natalia Ares; Parth Girdhar; Pranav Vaidhyanathan; Rahul Marchand

REVIEW 3 major objections 5 minor 1 cited by

End-to-End Analysis of Charge Stability Diagrams with Transformers

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a transformer trained only on simulated images can detect triple points and the transition lines connecting them in real quantum dot devices across three architectures, without retraining, beating a CNN baseline.

desk verdict A useful sim-trained transformer for reading charge stability diagrams, with a real but weakly quantified performance claim. read the letter →

arxiv 2508.15710 v1 pith:YTD75DYB submitted 2025-08-21 cond-mat.mes-hall cond-mat.mtrl-scics.LGquant-ph

Rahul Marchand , Lucas Schorling , Cornelius Carlsson , Jonas Schuff , Barnaby van Straaten , Taylor L. Patti , Federico Fedele , Joshua Ziegler

show 3 more authors

Parth Girdhar Pranav Vaidhyanathan Natalia Ares

This is my paper · ORCID

classification cond-mat.mes-hallcond-mat.mtrl-scics.LGquant-ph PACS 85.35.Gv07.05.Mh

keywords chargestabilitydiagramstriplepointstransformerobjectdetectionquantumdotarraysspinqubitssimulation-to-experimenttransfervirtualgatecalibrationconnectivitygraph

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that a single transformer model, trained only on simulated images, can read experimental charge stability diagrams of semiconductor quantum dot devices as graphs: it finds triple points, where charge-transition lines cross, and the lines connecting them. That graph is the input needed for automated tuning tasks like gate virtualization, charge-state initialization, drift correction, and pulse sequencing. The model, named TRACS, is tested on three different spin-qubit device architectures and outperforms a conventional CNN object detector on triple-point detection without any retraining. If true, it makes the case that end-to-end transformer analysis could replace multi-stage, device-specific tuning pipelines and scale with quantum dot arrays.

What carries the argument

TRACS' load-bearing object is the attention-based end-to-end detector: a convolutional backbone (modified ResNet-18) turns the greyscale diagram into tokens; positional encoding preserves layout; a transformer encoder-decoder with 25 learnable object queries reasons across tokens; prediction heads emit pair confidence, coordinates, and an adjacency matrix for line connectivity between triple points. The connectivity head is itself an attention block, and the model is trained end-to-end with a Hungarian-matched loss combining binary cross-entropy, wing loss, and focal loss, on labels generated alongside 500,000 QArray simulations.

What would settle it

Run the released, simulation-only trained model on charge stability diagrams from a quantum dot platform outside the silicon/germanium family, for example GaAs or Si-MOS, scoring triple-point detection under the same 15-pixel tolerance used in the paper; if recall falls well below the reported 0.84–0.90 experimental range, the architecture-agnostic claim is bounded by the simulator's noise and capacitance model.

Watch

Extended reading notes

Core claim

TRACS is an end-to-end transformer detector that turns a charge stability diagram—a two-gate voltage sweep read by a charge sensor—into a graph of triple points (where charge-transition lines cross) and the lines connecting them. Trained only on 500,000 simulated diagrams from QArray's constant-capacitance model with added thermal broadening, white noise, telegraph noise, and latching, it is tested without retraining on experimental diagrams from three spin-qubit architectures. The paper reports triple-point recall 0.84–0.90, precision 0.92–0.94; line-connectivity recall 0.91–0.95, precision 0.85–0.94; location error within 3% of the scan range; and an average improvement over five percentag

Load-bearing premise

The transfer claim rests on QArray's simulated diagrams, with added noise and latching, being close enough to real measurements that a model trained only on them recognizes triple points and lines in devices; the three test devices are all silicon or germanium heterostructures, so how far this extends is untested.

Editorial extensions

If this is right

Tuning routines that consume connectivity graphs can be reused across device architectures, because TRACS abstracts each charge stability diagram into the same graph structure.
Inference in tens of milliseconds on a CPU enables closed-loop drift-correction loops and much faster automated tuning than multi-stage pipelines.
The capacitively coupled slopes of detected lines, reported with median errors of 0.015–0.039, feed directly into gate virtualization.
Point and line detection happen in a single forward pass, removing pre- and post-processing stages from the tuning pipeline.
The architecture supports a two-stage paradigm of simulation pre-training followed by experimental fine-tuning, leaving a clear path for further transfer.

Reading between the lines

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

The three test devices are all silicon or germanium heterostructures, so the architecture-agnostic claim rests on one materials family; a direct test on, say, GaAs or Si-MOS dots is the natural next check.
Because the model outputs an explicit adjacency matrix, it could be reused for graph-level tuning tasks the paper does not run, such as detecting multi-dot charge configurations rather than only isolated triple points.
The paper deliberately avoids fine-tuning; fine-tuning on a small experimental set would reveal how much of the generalization comes from the transformer itself versus the simulator's realism.
Ablating each simulator augmentation—thermal broadening, white noise, telegraph noise, latching—would show which ingredient is responsible for the sim-to-real transfer.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Read the paper. Here's the short version: it's a genuine step forward for automated quantum dot tuning. TRACS is a DETR-style transformer that takes a charge stability diagram and outputs triple point locations plus an adjacency matrix for line connectivity. The connectivity head is new as far as I know, and it's a natural fit because you need the graph structure of the CSD for virtualization and pulse sequencing. They train on 500k simulated diagrams from QArray (their own simulator) and test on experimental data from three different devices — Si/SiGe accumulation, Ge/SiGe depletion, Ge/SiGe accumulation — with no retraining. That's a real sim-to-real transfer result.

The numbers: recall 0.84–0.90 vs YOLO's 0.76–0.84, precision mostly better except Device B where YOLO edges it. They report a >5 percentage point average improvement. That's plausible, but the evidence is thinner than the prose. Test sets are 60–72 diagrams each. No error bars, no significance test, and the word "significantly" is doing ordinary-language work. On these sample sizes one missed triple point can move recall by a couple points. The stress-test note is right to flag that. That said, the direction is consistent across devices and the simulated test set shows a 6-point recall gap (0.98 vs 0.92), so I don't think the headline claim is a mirage — it's just not rigorously sized.

Other soft spots: only YOLOv1 as baseline, which is old and not designed for this task; no comparison to MAVIS despite it being the obvious modular alternative. Code and data are promised only upon publication, which is frustrating for reproducibility. The simulator and model are from the same group, but since test data are independent experimental measurements, that's not a circularity problem in the central claim.

Who's this for? Groups working on spin-qubit tuning automation, and ML folks interested in sim-to-real with transformers. It deserves a serious referee — I'd send it to review and ask for error bars, a significance test, and a release of the code/data in the revision. I'd cite it if the code lands.

Referee Report

3 major / 5 minor

Summary. The paper introduces TRACS, a transformer-based object-detection model for charge stability diagrams (CSDs). TRACS uses a ResNet-18-style CNN front-end followed by a DETR-style transformer and prediction heads that output triple-point locations, confidence scores, and pairwise connectivity. The model is trained exclusively on 500,000 simulated CSDs generated with the QArray simulator and is evaluated, without retraining, on three experimental data sets (Devices A, B, C with 67, 72, and 60 CSDs, respectively) and a simulated test set of 128 CSDs. The central claim is that TRACS outperforms a YOLOv1 CNN baseline on triple-point detection and also predicts line connectivity, with high recall/precision and fast inference, supporting device- and architecture-agnostic automated tuning.

Significance. If the performance and cross-architecture generalization claims hold, TRACS would be a practically valuable step toward automated tuning of semiconductor spin-qubit arrays. The simulated-only training strategy, the direct graph-structured output (triple points plus connectivity), and the test on three distinct experimental devices are notable strengths. The paper also promises public release of data and code. However, the central quantitative comparison against YOLOv1 currently rests on small test sets with no uncertainty quantification, and several generalizability claims go beyond what the three similar spin-qubit test devices can support. With added statistical rigor and more careful wording, the contribution would be solid.

major comments (3)

[Table I and Appendix C] The central claim that TRACS outperforms YOLOv1 by >5 percentage points on average for recall and precision is not statistically supported. The experimental test sets contain only 67, 72, and 60 CSDs; Table I reports point estimates at a single 0.5 confidence threshold with no confidence intervals, no per-CSD variance, and no significance test. On Device B, YOLOv1 precision (0.95) actually exceeds TRACS (0.94), so the reported average advantage is not uniform. With 60-72 images, one or two detection errors can shift recall/precision by several percentage points, so sampling noise could explain the observed gap. The authors should provide bootstrap confidence intervals, per-CSD metric distributions, and/or a paired significance test (e.g., McNemar or Wilcoxon) before using the word 'significantly'.
[Abstract, §III, and Conclusion] The terms 'device- and architecture-agnostic,' 'wide range of measured CSDs,' and the general claim of robust generalization are stronger than the evidence supports. All three experimental devices are semiconductor spin-qubit devices based on Si/SiGe or Ge/SiGe heterostructures with charge-sensing readout; they do not sample a wide range of materials or qubit platforms. The QArray simulation space, while flexible, is also defined by parameters chosen by the authors. Please qualify the generalizability claims to the tested architectures, or substantially expand the experimental test set to genuinely dissimilar device families.
[§II, §III, and Appendix A] The YOLOv1 comparison is under-specified. The text says 'keeping the CNN from the first stage of TRACS,' but YOLOv1 has its own internal CNN architecture, so it is unclear whether YOLO was retrained with the same simulated data, the same input resolution, the same training budget, and the same object-matching criterion. Without these details, the reader cannot assess whether the comparison is a fair head-to-head between equivalent training conditions. Please specify the YOLO training procedure, hyperparameters, and all evaluation parameters, or the outperformance claim is not well grounded.

minor comments (5)

[Abstract and §III] The word 'significantly' in the abstract is used in the ordinary-language sense. Since no statistical significance testing is reported, either add a test or replace 'significantly' with a more neutral qualifier.
[§III, paragraph after Table I] The '45% reduction in false negatives and 19% reduction in false positives' is stated without showing the underlying counts. Provide the aggregate TP/FP/FN numbers or a table footnote so readers can verify these percentages.
[Fig. 2(d)] The violin plots are only defined for true-positive detections and do not state the number of points per device. Please indicate the sample sizes and, if possible, show the per-CSD error distributions separately from aggregate detection metrics.
[Data Availability] The statement that code and data 'will be made public upon final publication' is helpful, but providing a repository link or an anonymized data/code artifact during review would strengthen reproducibility.
[Table I, rows 5-8] Line-detection results are reported only for TRACS, with no baseline. The paper does not claim YOLO is competitive for lines, but the absence of any comparator means the line-detection performance cannot be benchmarked against prior art; consider adding a simple baseline or a qualitative comparison with prior line-extraction methods.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: training on independently published QArray simulation and testing on held-out experimental CSDs is a self-contained generalization result.

full rationale

The paper's derivation chain is: (1) a standard DETR-based architecture, (2) supervised training on 500,000 charge stability diagrams generated by QArray, an externally published GPU-accelerated simulator, and (3) evaluation on held-out experimental CSDs from three device architectures, compared with YOLO trained on the same simulated data. No experimental data were used for training, for setting simulation constants, or for post-hoc fitting; the reported recall/precision numbers are direct forward-pass evaluations on held-out test sets. The only self-citation touching the training pipeline is QArray (refs. 51, 52), but this is a public, code-released simulator, and the paper's central generalization claim is independently evidenced by the experimental test results themselves. There is no equation in which a predicted quantity is defined from the fitted quantity, no fitted-parameter-renamed-as-prediction step, and no load-bearing uniqueness or ansatz imported via self-citation. The paper's use of 'significantly' is informal and the small experimental test sets are a statistical robustness concern, not a circularity concern. The derivation therefore does not reduce to its inputs by construction.

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

The central claim does not introduce new physical entities or conserved quantities. The free parameters are ordinary ML hyperparameters and under-specified simulation bounds. The principal unproven input is the realism of simulated training data, which is common to sim-to-real ML and should be tested more broadly.

free parameters (4)

Loss weights: alpha2, alpha3, beta1, w1, w2, w3, gamma = alpha2=0.05, alpha3=0.5, beta1=0.5, w1=0.3, w2=w3=1, gamma=2
Chosen by hand to balance confidence, localization, and adjacency losses; affects model behavior but not fitted to experimental outcomes.
QArray simulation parameter bounds = not specified
Uniform sampling bounds are described only as 'selected to span a wide range of experimental conditions'; the exact values are not given, so the training distribution is under-specified.
Number of object queries n_q = 25
Fixed hyperparameter determining the maximum number of triple-point pairs the model can output.
Optimizer settings = AdamW lr=2e-5, batch 256, 225 epochs
Standard training choices, not experimental fits.

assumptions (3)

domain assumption QArray's constant-capacitance model produces charge stability diagrams whose visual features (line patterns, triple points, noise) are representative of experimental devices.
Invoked in Methods when all 500,000 training diagrams are generated; sim-to-real transfer is the load-bearing premise for all experimental results.
domain assumption The ground-truth triple-point labels and connectivity labels from QArray correspond to the physically correct interpretation of the experimental diagrams.
Supervision and evaluation both rely on these labels; if QArray labels differ from expert annotation, reported precision/recall would be mis-measured.
standard math Hungarian matching for bipartite assignment between predictions and ground truth is a valid surrogate for detection correctness.
Used in the loss function (Appendix A) and is standard in DETR-style object detection.

how reviews work

0 comments

Cite this review

Pith. "Pith review of End-to-End Analysis of Charge Stability Diagrams with Transformers." pith.science (2026). https://pith.science/paper/YTD75DYB

@misc{pith2026250815710,
  author       = {Pith},
  title        = {Pith review of: End-to-End Analysis of Charge Stability Diagrams with Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YTD75DYB}},
  note         = {Machine review of arXiv:2508.15710}
}

read the original abstract

Transformer models and end-to-end learning frameworks are rapidly revolutionizing the field of artificial intelligence. In this work, we apply object detection transformers to analyze charge stability diagrams in semiconductor quantum dot arrays, a key task for achieving scalability with spin-based quantum computing. Specifically, our model identifies triple points and their connectivity, which is crucial for virtual gate calibration, charge state initialization, drift correction, and pulse sequencing. We show that it surpasses convolutional neural networks in performance on three different spin qubit architectures, all without the need for retraining. In contrast to existing approaches, our method significantly reduces complexity and runtime, while enhancing generalizability. The results highlight the potential of transformer-based end-to-end learning frameworks as a foundation for a scalable, device- and architecture-agnostic tool for control and tuning of quantum dot devices.

Figures

Figures reproduced from arXiv: 2508.15710 by the authors.

**Figure 2.** FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Action-Factored Multi-Agent Reinforcement Learning for Scalable Quantum Device Tuning
cs.LG 2026-07 conditional novelty 6.0 of 10

Online action-space factorization via Kalman-refined cross-capacitance lets shared multi-agent policies zero-shot tune larger quantum-dot arrays with near-constant steps.

Reference graph

Works this paper leans on

6 extracted references · 2 canonical work pages · cited by 1 Pith paper

[2]

Attention is all you need,

σ(i) represents the per- mutation of labels and object tokens based on the min- imal cost matching via the Hungarian algorithm. The double subscripts ci = 1indicate that the sum only in- cludes terms fulfilling this condition. Lwing has width ω = 0.05 and curvature ϵ = 0.01. We use the AdamW optimizer with a learning rate of 2e-5, a learning rate schedule...

work page 2017
[2016]

Distri- bution transformers: Fast approximate bayesian inference with on-the-fly prior adaptation,

book in preparation for MIT Press. 7U. Kamath, K. Graham, and W. Emara,Transformers for ma- chine learning: a deep dive (Chapman and Hall/CRC, 2022). 8G. Whittle, J. Ziomek, J. Rawling, and M. A. Osborne, “Distri- bution transformers: Fast approximate bayesian inference with on-the-fly prior adaptation,” arXiv preprint arXiv:2502.02463 (2025). 9R. Roy and...

arXiv 2022
[2017]

12-spin-qubitarraysfabricatedona300mmsemi- conductor manufacturing line,

pp. 17–32. 42H. C. George, M. T. Mądzik, E. M. Henry, A. J. Wagner, M. M. Islam, F. Borjans, E. J. Connors, J. Corrigan, M. Curry, M. K. Harper, et al.,“12-spin-qubitarraysfabricatedona300mmsemi- conductor manufacturing line,” Nano Letters25, 793–799 (2024). 43J. Saez-Mollejo, D. Jirovec, Y. Schell, J. Kukucka, S. Calcaterra, D. Chrastina, G. Isella, M. R...

arXiv 2024
[2020]

Probing quantum devices with radio-frequency reflectometry,

pp. 213–229. 48F. Vigneau, F. Fedele, A. Chatterjee, D. Reilly, F. Kuemmeth, M. F. Gonzalez-Zalba, E. Laird, and N. Ares, “Probing quantum devices with radio-frequency reflectometry,” Applied Physics Re- views 10 (2023). 49Z.-H. Feng, J. Kittler, M. Awais, P. Huber, and X.-J. Wu, “Wing loss for robust facial landmark localisation with convolutional neural...

arXiv 2023
[2022]

A survey of transformers,

pp. 15340–15359. 3T. Lin, Y. Wang, X. Liu, and X. Qiu, “A survey of transformers,” AI open 3, 111–132 (2022). 4A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” (2021), ...

arXiv 2022
[2025]

Transformer models and end-to-end learning frameworks are rapidly revolutionizing the field of artificial intelligence. In this work, we apply object detection transformers to analyze charge stability diagrams in semiconductor quantum dot arrays, a key task for achieving scalability with spin-based quantum comput- ing. Specifically, our model identifies t...

work page Pith review arXiv 2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [2]

Attention is all you need,

σ(i) represents the per- mutation of labels and object tokens based on the min- imal cost matching via the Hungarian algorithm. The double subscripts ci = 1indicate that the sum only in- cludes terms fulfilling this condition. Lwing has width ω = 0.05 and curvature ϵ = 0.01. We use the AdamW optimizer with a learning rate of 2e-5, a learning rate schedule...

work page 2017

[2] [2016]

Distri- bution transformers: Fast approximate bayesian inference with on-the-fly prior adaptation,

book in preparation for MIT Press. 7U. Kamath, K. Graham, and W. Emara,Transformers for ma- chine learning: a deep dive (Chapman and Hall/CRC, 2022). 8G. Whittle, J. Ziomek, J. Rawling, and M. A. Osborne, “Distri- bution transformers: Fast approximate bayesian inference with on-the-fly prior adaptation,” arXiv preprint arXiv:2502.02463 (2025). 9R. Roy and...

arXiv 2022

[3] [2017]

12-spin-qubitarraysfabricatedona300mmsemi- conductor manufacturing line,

pp. 17–32. 42H. C. George, M. T. Mądzik, E. M. Henry, A. J. Wagner, M. M. Islam, F. Borjans, E. J. Connors, J. Corrigan, M. Curry, M. K. Harper, et al.,“12-spin-qubitarraysfabricatedona300mmsemi- conductor manufacturing line,” Nano Letters25, 793–799 (2024). 43J. Saez-Mollejo, D. Jirovec, Y. Schell, J. Kukucka, S. Calcaterra, D. Chrastina, G. Isella, M. R...

arXiv 2024

[4] [2020]

Probing quantum devices with radio-frequency reflectometry,

pp. 213–229. 48F. Vigneau, F. Fedele, A. Chatterjee, D. Reilly, F. Kuemmeth, M. F. Gonzalez-Zalba, E. Laird, and N. Ares, “Probing quantum devices with radio-frequency reflectometry,” Applied Physics Re- views 10 (2023). 49Z.-H. Feng, J. Kittler, M. Awais, P. Huber, and X.-J. Wu, “Wing loss for robust facial landmark localisation with convolutional neural...

arXiv 2023

[5] [2022]

A survey of transformers,

pp. 15340–15359. 3T. Lin, Y. Wang, X. Liu, and X. Qiu, “A survey of transformers,” AI open 3, 111–132 (2022). 4A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” (2021), ...

arXiv 2022

[6] [2025]

Transformer models and end-to-end learning frameworks are rapidly revolutionizing the field of artificial intelligence. In this work, we apply object detection transformers to analyze charge stability diagrams in semiconductor quantum dot arrays, a key task for achieving scalability with spin-based quantum comput- ing. Specifically, our model identifies t...

work page Pith review arXiv 2025