Pith. sign in

REVIEW 2 major objections 4 minor 38 references

Neural inverse design maps qubit targets to transmon layouts with sub-1% error

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

2026-08-03 00:43 UTC pith:WWYAUPYK

load-bearing objection Solid in-simulation demonstration of tandem inverse design for a transmon layout; the claims are honestly scoped, but the abstract's 'comparable to fabrication uncertainty' wording outruns the evidence. the 2 major comments →

arxiv 2607.20795 v2 pith:WWYAUPYK submitted 2026-07-22 quant-ph

Component-Level Inverse Design of Transmon Qubits Using Neural Networks

Olivia Seidel , Firas Abouzahr , Abhishek Chakraborty , Sadman Ahmed Shanto , Saikat Das , Daniel Baxter , Jonathan Asaadi , Nicola Pancotti
show 6 more authors
Haoyu Yang Brucek Khailany Sara Sussman Enectali Figueroa-Feliciano Eli M Levenson-Falk Taylor L. Patti
This is my paper
classification quant-ph
keywords transmon qubitsinverse designneural networksforward surrogatequbit frequencyanharmonicityelectromagnetic simulationsmall-data learning
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.

Designing a superconducting qubit usually means a slow, hit-and-miss loop: draw a geometry, simulate it, extract capacitances, compute the qubit's frequency and anharmonicity, and adjust. This paper tries to invert that loop: it trains a small neural network to take the two target Hamiltonian parameters — the qubit transition frequency and the anharmonicity — and output three geometry dimensions of a cross-claw transmon layout directly. To avoid the one-to-many ambiguity of the inverse map, the inverse network is trained through a frozen forward surrogate, so the loss is computed on the reconstructed Hamiltonian rather than on geometry labels. Against a conventional electromagnetic solver, 97% of the generated designs are usable, with mean errors of 0.73% in frequency and 1.58% in anharmonicity, and a single query takes about 60 ms compared with roughly two minutes for a conventional extraction. The claim is that component-level inverse design is practical even with only about 1,900 simulated training samples.

Core claim

The central result is that a tandem neural-network pipeline can map target Hamiltonian parameters to component-level layout parameters with accuracy comparable to simulation uncertainty. The inverse network receives the desired qubit frequency fq and anharmonicity α and outputs three scaled layout parameters: coupling-claw length, ground-plane spacing, and transmon cross length. A frozen forward network, trained on the same simulated data to map geometry back to (fq, α), reconstructs the Hamiltonian from the proposed geometry, and the inverse network is trained by the mean absolute difference between requested and reconstructed parameters in normalized coordinates, plus a soft penalty that k

What carries the argument

The load-bearing mechanism is the tandem inverse-plus-surrogate architecture. A compact inverse multilayer perceptron (one hidden layer, 64 units, 387 trainable parameters) maps the two Hamiltonian targets to three scaled geometry parameters; a frozen forward multilayer perceptron (736 hidden units, 4,418 parameters) maps geometry back to the Hamiltonian. Training the inverse network through the frozen surrogate with a Hamiltonian-space mean-absolute-error loss avoids a known failure mode of inverse problems: when many geometries give nearly the same response, a geometry-matching objective averages incompatible solutions into invalid designs. A range penalty keeps outputs within the min-max-

Load-bearing premise

The accuracy claim is validated only against the same electromagnetic-solver workflow that generated the training data — not against fabricated devices — and because the quoted simulation-to-measurement and fabrication uncertainties are also at the percent level, a solver that systematically misses real device behavior would erase the reported advantage; additionally, the inverse map is only demonstrated along the one-parameter curve set by a fixed Josephson energy, so arbitr

What would settle it

Fabricate a batch of, say, twenty inverse-designed cross-claw transmons with the predicted geometry parameters, measure fq and α at millikelvin temperatures, and compare to the requested targets: if the measured deviations systematically exceed the percent-level simulation-to-measurement uncertainty cited in the paper, the reported accuracy is a property of the simulator, not of the design method. Also, request a target pair that lies off the fixed-Josephson-energy curve; if the model returns a geometry anyway, that output is unsupported and can be checked for whether it hits the target.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Candidate transmon geometries can be generated and screened in milliseconds rather than minutes, so large design sweeps become feasible; the paper's nearest-neighbor stress test screens 50,000 candidates in seconds.
  • The 97% solver-valid geometry rate implies that automated inverse design can feed an existing design loop with mostly usable starting points, with only a small fraction needing rejection or repair.
  • Because the reported errors sit at or below the expected simulation-to-measurement uncertainty, the pipeline's accuracy is not the limiting factor in device design for this layout class.
  • The small-data result (accuracy plateaus after about half of roughly 1,900 samples) suggests the same tandem approach can be transferred to other parameterized qubit layouts without requiring huge simulation datasets.
  • The speedup compounds in iterative loops: what would take over 1,500 hours of solver time for 50,000 candidates can be pre-screened by the surrogate in seconds before committing a few designs to full simulation.

Where Pith is reading between the lines

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

  • With the Josephson energy held fixed at 10 nH, frequency and anharmonicity are locked to a one-dimensional curve in target space, so the demonstrated inverse map is effectively one-dimensional; a testable extension is to vary the junction inductance as well, which would require targets spread over a two-dimensional region.
  • The tandem Hamiltonian-space loss should transfer to other target quantities — coupling strengths, readout frequencies, or multi-qubit parameters — whenever a forward simulation can be learned; the same one-to-many structure will appear there.
  • The three failed geometries all violated layout rules (metal overlap or missing ground spacing), which suggests a hard geometric constraint layer or a physics-informed penalty could raise the usable rate toward 100% without retraining the physics loss.
  • The reported 2,000x speedup compares a single neural query with a single conventional extraction; the true end-to-end gain depends on how many iterations the designer would otherwise run, and the paper's own figures suggest the gain grows with batch size and with the number of candidates screened.

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

2 major / 4 minor

Summary. The paper presents a neural-network inverse-design workflow for a cross-claw planar transmon. A forward surrogate MLP is trained on SQuADDS simulation records to map three geometry parameters (claw length, ground spacing, cross length) to qubit frequency fq and anharmonicity α. A compact inverse MLP is then trained with the surrogate frozen, with the loss computed in Hamiltonian space. The inverse map is validated by rendering predicted geometries and re-extracting fq and α with Ansys Q3D and scqubits under SQuADDS settings. On a held-out split of the 1,934-sample SQuADDS cross-claw dataset, the pipeline achieves mean percent errors of 0.73% (fq) and 1.58% (α) over 97 valid designs, with a 97% solver-valid geometry rate; a single query takes ~60 ms versus ~2 min for EM extraction. The authors also report small-data scaling and a nearest-neighbor stress test. A stated limitation is that EJ is fixed, so targets lie on a one-dimensional curve in (fq, α) space.

Significance. If the results are interpreted as in-simulation validation, the paper makes a credible and useful contribution: it demonstrates a small-data tandem inverse-design pipeline for a realistic transmon layout, with careful data splitting, training-set-only normalization, EM-solver-in-the-loop validation, and a nearest-neighbor baseline. The code/data availability statements support reproducibility. The main caveat is that the validation ground truth is generated by the same EM-solver workflow that produced the training data, so the reported accuracy characterizes consistency with that solver rather than device-level accuracy; the paper's claim of parity with fabrication uncertainty goes beyond the evidence. The fixed-EJ limitation also means the demonstrated inverse map has one effective degree of freedom. These caveats do not undermine the in-simulation demonstration but should be reflected in the abstract.

major comments (2)
  1. [Abstract; §6] The statement that the pipeline errors (0.73% fq, 1.58% α) are 'comparable to or below the fabrication and simulation-to-measurement uncertainty expected for academic-process transmon devices' is not supported by the evidence in the manuscript. All validation is performed against Ansys Q3D using the SQuADDS settings (Section 3.2), which is the same solver workflow that generated the training labels. The reported 0.005%/0.01% reproduction of SQuADDS confirms consistency with that database, not accuracy against fabricated hardware. Section 6 cites literature-level percent discrepancies, but no device measurement is presented. The abstract and Section 6 should be revised to state explicitly that validation is simulation-level, and the parity-with-fabrication claim should be removed or clearly qualified as conditional on the solver being representative of hardware.
  2. [§2.1; Abstract] Because EJ is fixed at 10 nH, fq and α are both determined by the single free quantity EC; the targets used in this work all lie on the one-dimensional curve of Eq. (2), and the inverse map is effectively from one Hamiltonian degree of freedom to three geometry parameters. The paper is explicit about this in Section 2.1, but the abstract presents the model as mapping target fq and α to geometry without the fixed-EJ caveat. An arbitrary (fq, α) pair is outside the demonstrated regime. Please add this scope limitation to the abstract and to the list of contributions so that readers do not over-generalize the headline numbers.
minor comments (4)
  1. [§5.1; Appendix A] The mean percent-error statistics are computed over the 97 usable EM-validated designs after removing 3 geometry-invalid candidates. This conditional reporting should be stated more prominently; the abstract should use phrasing such as 'conditional on the 97% of designs that pass the geometry-rule check' so that the 3% failure rate is not hidden.
  2. [§5.2] The batched speedup factors (3.9×10^7 and 4.6×10^7) compare a per-sample batched inference cost with a single EM extraction. This is not a workload-equivalent comparison; the single-query 60 ms versus ~2 min comparison is the fair one. Please add a sentence clarifying that the batched figure is an arithmetic per-sample speedup under the assumption that 2048 independent designs are needed.
  3. [Fig. 3(a)] The axis label 'fq ®' appears to be a rendering artifact and should be corrected to 'f_q' (and similarly check α labels). The box-plot axis labels should also be checked for readability in the final PDF.
  4. [§3.1; §5.1] The paper notes that ground spacing is sampled on a coarse discrete grid. This is relevant to the inverse-model predictions, which can request values between grid points; a brief reminder in Section 5.1 would help readers interpret the error distributions, especially for designs that lie off-grid in sground.

Circularity Check

0 steps flagged

No circularity: the central accuracy claims rest on held-out EM-solver validation and a Hamiltonian-space loss, not on a fitted input renamed as a prediction; same-solver training/validation is a scope limitation, not a tautology.

full rationale

The paper's derivation chain is not circular. The inverse model is trained with a loss (Eq. 5) that compares requested and reconstructed Hamiltonian parameters through a frozen forward surrogate, so the objective is not defined in terms of the predicted geometry. The headline errors (0.73% fq, 1.58% alpha) are obtained by generating 100 candidates from the 291-record held-out test set and validating them through the conventional Ansys Q3D/scqubits workflow, as stated in Section 5.1. Because the test records are not used during training or model selection, the reported errors are an honest generalization measurement within the simulator, not a statistically forced echo of the training fit. The paper explicitly discloses the two main scope limits: with EJ fixed, (fq, alpha) targets lie on the one-dimensional curve of Eq. (2) (Section 2.1), which narrows the demonstrated regime but does not make the inverse map tautological; and validation uses the same Q3D solver settings that generated the SQuADDS training data (Section 3.2), which limits external validity to simulator consistency rather than fabricated-device accuracy. The abstract's 'comparable to fabrication and simulation-to-measurement uncertainty' claim is a literature-calibrated contextual statement, supported in Section 6 by cited percent-level discrepancies (including Ref. [7] by co-authors); it is not the derivation of the pipeline accuracy, and removing it would not change the central results. SQuADDS overlap with the present authors is a data-source provenance issue, not a load-bearing self-citation chain. No equation or fitted parameter is shown to reduce to its own input by construction, so no circular step is present.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The central claim rests on a small number of physics domain assumptions (transmon Hamiltonian, EM solver fidelity) and on fitted model design choices (architecture, penalty weight, fixed EJ). No new physical entities are postulated. The most load-bearing assumption is that the EM-solver-defined Hamiltonian is the correct target mapping, since no fabricated-device validation is performed.

free parameters (3)
  • Fixed Josephson energy EJ (10 nH inductance) = 10 nH
    EJ is held fixed across all samples, which reduces the (fq, α) target space to a one-dimensional curve. This is a design choice that makes the inverse problem under-determined; the paper does not explore the impact of varying EJ.
  • Range penalty weight λ_range = 1
    Weight on Lrange in Eq. (5)/Appendix D; selected via Keras Tuner. It affects whether predicted geometries stay inside the [0,1] scaled design box and therefore the reported 97% usable-geometry rate.
  • Inverse MLP architecture (1 hidden layer, 64 units) and surrogate width (736 units) = 64 / 736
    Selected via hyperparameter sweep. These model choices directly affect the reported error values; they are fitted to the data rather than derived from physics.
axioms (3)
  • domain assumption Transmon Hamiltonian can be described by the Duffing oscillator approximation with fq and α as the two parameters (Eq. 1, Eq. 2)
    Standard physics for transmons, invoked in Section 2.1. The inverse design targets are defined within this model, so the workflow inherits its approximation accuracy.
  • domain assumption Ansys Q3D EM solver with SQuADDS settings accurately predicts the capacitance matrix and hence the Hamiltonian parameters of the cross-claw transmon layout
    Both the training data (SQuADDS) and the final validation use this solver. The paper confirms its own pipeline reproduces SQuADDS values to 0.005%, but this is internal consistency with the same simulation framework, not independent experimental verification.
  • domain assumption The three varied geometry parameters (ℓclaw, sground, ℓcross) are sufficient to parametrize the layout variations in the dataset; all other layout settings are held fixed
    The paper restricts the demonstration to a single cross-claw layout with three free parameters. Generalization to more of the Quantum Metal parameter space is not tested, and the fixed-EJ condition further limits the explored target manifold.

pith-pipeline@v1.3.0-alltime-deepseek · 20101 in / 10806 out tokens · 100054 ms · 2026-08-03T00:43:45.670666+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Component-Level Inverse Design of Transmon Qubits Using Neural Networks." pith.science (2026). https://pith.science/paper/WWYAUPYK

@misc{pith2026260720795,
  author       = {Pith},
  title        = {Pith review of: Component-Level Inverse Design of Transmon Qubits Using Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WWYAUPYK}},
  note         = {Machine review of arXiv:2607.20795}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Designing a superconducting qubit to realize specific Hamiltonian parameters typically requires iterating through a time and compute-intensive forward loop in which the designer chooses a layout geometry, simulates it, extracts circuit parameters such as capacitances, and refines the geometry. We study the inverse version of this task using a neural-network workflow that maps target Hamiltonian parameters directly to component-level layout parameters, which we subsequently demonstrate on a planar transmon layout. During training, we pair the inverse model with a frozen forward surrogate model and evaluate the loss in Hamiltonian space rather than in layout-parameter space. In validation against a conventional EM solver, 97% of generated designs produce usable geometries, and the inverse-plus-surrogate pipeline reaches mean percent errors of 0.73% for qubit frequency and 1.58% for anharmonicity, comparable to or below the fabrication and simulation-to-measurement uncertainty expected for academic-process transmon devices of this type. A single pipeline query takes ~60 ms on CPU, versus ~2 min for a conventional EM capacitance extraction on the same hardware, a speedup of approximately 2,000x. Batching minimizes the AI model inference overhead, reducing the runtime to 3.1 microseconds per sample on CPU and 2.6 microseconds per sample on GPU at a batch size of 2048, resulting in speedups of 3.9 x 10^7 and 4.6 x 10^7, respectively, relative to a single conventional CPU EM extraction. Our results indicate that component-level inverse design usefully extends and complements conventional EM simulation, including for small datasets on the order of 1,000 samples.

Review history (2 revisions) →

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

38 extracted references · 6 linked inside Pith

  1. [1]

    https:// qiskit-community.github.io/ qiskit-metal/

    Qiskit Metal. https:// qiskit-community.github.io/ qiskit-metal/. Accessed 2026-07-21 (2026)

    2026

  2. [2]

    Zenodo (2021)

    Minev, Z.K., McConkey, T.G., Drys- dale, J., Shah, P., Wang, D., Facchini, M., Harper, G., Blair, J., Zhang, H., Lanzillo, N., Mukesh, S., Shanks, W., Warren, C., Gambetta, J.M.: Qiskit Metal: An Open-Source Frame- work for Quantum Device Design and Analysis. Zenodo (2021). https: //doi.org/10.5281/zenodo.4618154

    doi:10.5281/zenodo.4618154 2021

  3. [3]

    Physical Review A69(6), 062320 (2004) https://doi.org/10

    Blais, A., Huang, R.-S., Wallraff, A., Girvin, S.M., Schoelkopf, R.J.: Cavity quantum electrodynamics for superconducting electrical circuits: An architecture for quantum com- putation. Physical Review A69(6), 062320 (2004) https://doi.org/10. 1103/PhysRevA.69.062320

    2004

  4. [4]

    Physical Review Letters108(24), 240502 (2012) https://doi.org/10

    Nigg, S.E., Paik, H., Vlastakis, B., Kirchmair, G., Shankar, S., Frun- zio, L., Devoret, M.H., Schoelkopf, R.J., Girvin, S.M.: Black-box super- conducting circuit quantization. Physical Review Letters108(24), 240502 (2012) https://doi.org/10. 1103/PhysRevLett.108.240502

    2012

  5. [5]

    npj Quantum Information7(1), 131 (2021) https://doi.org/10.1038/ s41534-021-00461-8

    Minev, Z.K., Leghtas, Z., Mund- hada, S.O., Christakis, L., Pop, I.M., Devoret, M.H.: Energy-participation quantization of josephson circuits. npj Quantum Information7(1), 131 (2021) https://doi.org/10.1038/ s41534-021-00461-8

    2021

  6. [6]

    Quantum5, 583 (2021) https://doi.org/10.22331/ q-2021-11-17-583

    Groszkowski, P., Koch, J.: scqubits: a python package for supercon- ducting qubits. Quantum5, 583 (2021) https://doi.org/10.22331/ q-2021-11-17-583

    2021

  7. [7]

    Quantum8, 1465 (2024) https://doi

    Shanto, S., Kuo, A., Miyamoto, C., Zhang, H., Maurya, V., Vlachos, E., Hecht, M., Shum, C.W., Levenson- Falk, E.M.: SQuADDS: a validated design database and simulation work- flow for superconducting qubit design. Quantum8, 1465 (2024) https://doi. org/10.22331/q-2024-09-09-1465

    doi:10.22331/q-2024-09-09-1465 2024

  8. [8]

    Materials for Quantum Technol- ogy5(2), 022003 (2025) https:// doi.org/10.1088/2633-4356/ade10d arXiv:2411.16967 [quant-ph]

    Levenson-Falk, E.M., Shanto, S.A.: A review of design concerns in superconducting quantum circuits. Materials for Quantum Technol- ogy5(2), 022003 (2025) https:// doi.org/10.1088/2633-4356/ade10d arXiv:2411.16967 [quant-ph]

    Pith/arXiv arXiv 2025

  9. [9]

    Koch, J., Yu, T.M., Gambetta, J., Houck, A.A., Schuster, D.I., Majer, J., Blais, A., Devoret, M.H., Girvin, S.M., Schoelkopf, R.J.: Charge- insensitive qubit design derived from the Cooper pair box. Phys. Rev. A 76(4), 042319 (2007) https://doi.org/ 27 10.1103/PhysRevA.76.042319

    doi:10.1103/physreva.76.042319 2007

  10. [10]

    Applied Physics Reviews6(2), 021318 (2019) https://doi.org/10.1063/1.5089550

    Krantz, P., Kjaergaard, M., Yan, F., Orlando, T.P., Gustavsson, S., Oliver, W.D.: A quantum engineer’s guide to superconducting qubits. Applied Physics Reviews6(2), 021318 (2019) https://doi.org/10.1063/1.5089550

    doi:10.1063/1.5089550 2019

  11. [11]

    Annual Review of Condensed Matter Physics11, 369– 395 (2020) https://doi.org/10.1146/ annurev-conmatphys-031119-050605

    Kjaergaard, M., Schwartz, M.E., Braum¨ uller, J., Krantz, P., Wang, J.I.-J., Gustavsson, S., Oliver, W.D.: Superconducting qubits: Current state of play. Annual Review of Condensed Matter Physics11, 369– 395 (2020) https://doi.org/10.1146/ annurev-conmatphys-031119-050605

    2020

  12. [12]

    Reviews of Modern Physics93(2), 025005 (2021) https://doi.org/10.1103/ RevModPhys.93.025005

    Blais, A., Grimsmo, A.L., Girvin, S.M., Wallraff, A.: Circuit quan- tum electrodynamics. Reviews of Modern Physics93(2), 025005 (2021) https://doi.org/10.1103/ RevModPhys.93.025005

    2021

  13. [13]

    Science 339(6124), 1169–1174 (2013) https: //doi.org/10.1126/science.1231930

    Devoret, M.H., Schoelkopf, R.J.: Superconducting circuits for quan- tum information: An outlook. Science 339(6124), 1169–1174 (2013) https: //doi.org/10.1126/science.1231930

    doi:10.1126/science.1231930 2013

  14. [14]

    Acta Numerica 19, 451–559 (2010) https://doi.org/ 10.1017/S0962492910000061

    Stuart, A.M.: Inverse problems: A bayesian perspective. Acta Numerica 19, 451–559 (2010) https://doi.org/ 10.1017/S0962492910000061

    doi:10.1017/s0962492910000061 2010

  15. [15]

    Inverse Problems25(11), 115008 (2009) https://doi.org/10.1088/0266-5611/ 25/11/115008

    Cotter, S.L., Dashti, M., Robinson, J.C., Stuart, A.M.: Bayesian inverse problems for functions and appli- cations to fluid mechanics. Inverse Problems25(11), 115008 (2009) https://doi.org/10.1088/0266-5611/ 25/11/115008

    doi:10.1088/0266-5611/ 2009

  16. [16]

    https://arxiv

    Taghizadeh, L., J¨ ungel, A.: Bayesian inversion for the identification of the doping profile in unipolar semicon- ductor devices (2024). https://arxiv. org/abs/2408.11485

    Pith/arXiv arXiv 2024

  17. [17]

    Statistical Sci- ence28(3), 424–446 (2013)

    Cotter, S.L., Roberts, G.O., Stuart, A.M., White, D.: Mcmc methods for functions: modifying old algorithms to make them faster. Statistical Sci- ence28(3), 424–446 (2013)

    2013

  18. [18]

    arXiv preprint arXiv:2010.08895 (2020)

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A.: Fourier neu- ral operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895 (2020)

    Pith/arXiv arXiv 2010

  19. [19]

    arXiv preprint arXiv:2504.20278 (2025)

    Yang, H., Azizzadenesheli, K., Ren, H.: Deep generative prior for first order inverse optimization. arXiv preprint arXiv:2504.20278 (2025)

    arXiv 2025

  20. [20]

    IEEE Reviews in Biomedical Engineering (2024)

    Huang, J., Wu, Y., Wang, F., Fang, Y., Nan, Y., Alkan, C., Abraham, D., Liao, C., Xu, L., Gao, Z., et al.: Data-and physics-driven deep learning based reconstruction for fast mri: Fundamentals and methodolo- gies. IEEE Reviews in Biomedical Engineering (2024)

    2024

  21. [21]

    arXiv preprint arXiv:2402.11722 (2024)

    Long, D., Xu, Z., Yuan, Q., Yang, Y., Zhe, S.: Invertible fourier neural operators for tackling both forward and inverse problems. arXiv preprint arXiv:2402.11722 (2024)

    Pith/arXiv arXiv 2024

  22. [22]

    Peurifoy, J., Shen, Y., Jing, L., Yang, Y., Cano-Renteria, F., DeLacy, B.G., Joannopoulos, J.D., Tegmark, M., Soljacic, M.: Nanophotonic particle simulation and inverse design using artificial neural networks. Sci. Adv. 4(6), 4206 (2018) https://doi.org/10. 1126/sciadv.aar4206

    2018

  23. [23]

    ACS Photonics5(4), 1365– 1369 (2018) https://doi.org/10.1021/ acsphotonics.7b01377

    Liu, D., Tan, Y., Khoram, E., Yu, 28 Z.: Training deep neural networks for the inverse design of nanophotonic structures. ACS Photonics5(4), 1365– 1369 (2018) https://doi.org/10.1021/ acsphotonics.7b01377

    2018

  24. [24]

    Molesky, S., Lin, Z., Piggott, A.Y., Jin, W., Vuckovic, J., Rodriguez, A.W.: Inverse design in nanopho- tonics. Nat. Photonics12(11), 659– 670 (2018) https://doi.org/10.1038/ s41566-018-0246-9

    2018

  25. [25]

    npj Quantum Information7(1), 49 (2021) https://doi.org/10.1038/ s41534-021-00382-6

    Menke, T., H¨ ase, F., Gustavsson, S., Kerman, A.J., Oliver, W.D., Aspuru-Guzik, A.: Automated design of superconducting circuits and its application to 4-local couplers. npj Quantum Information7(1), 49 (2021) https://doi.org/10.1038/ s41534-021-00382-6

    2021

  26. [26]

    In: 2023 IEEE International Conference on Quan- tum Computing and Engineering (QCE), pp

    Nugraha, F.P., Shao, Q.: Machine learning-based predictive modeling for designing transmon supercon- ducting qubits. In: 2023 IEEE International Conference on Quan- tum Computing and Engineering (QCE), pp. 1360–1368. IEEE, Belle- vue, W A, USA (2023). https://doi. org/10.1109/QCE57702.2023.00154

    arXiv 2023

  27. [27]

    Solid-State Electronics 201, 108582 (2023) https://doi.org/ 10.1016/j.sse.2022.108582

    Wong, H.Y., Dhillon, P., Beck, K.M., Rosen, Y.J.: A simulation methodol- ogy for superconducting qubit read- out fidelity. Solid-State Electronics 201, 108582 (2023) https://doi.org/ 10.1016/j.sse.2022.108582

    arXiv 2023

  28. [28]

    In: 2024 International Conference on Simu- lation of Semiconductor Processes and Devices (SISPAD), pp

    Lu, A., Wong, H.Y.: Rapid simu- lation framework for superconduct- ing qubit readout system inverse design and optimization. In: 2024 International Conference on Simu- lation of Semiconductor Processes and Devices (SISPAD), pp. 1– 4 (2024). https://doi.org/10.1109/ SISPAD62626.2024.10733335

    arXiv 2024

  29. [29]

    https: //arxiv.org/abs/2607.02289

    Yaker, J., Markovic, J., Reineri, A., Kurkcuoglu, D.M., Zorzetti, S.: Neural-network inverse design of SRF cavities and transmons for bosonic quantum computation (2026). https: //arxiv.org/abs/2607.02289

    Pith/arXiv arXiv 2026

  30. [30]

    https://scikit-learn.org/ stable/modules/generated/sklearn

    scikit-learn developers: MinMaxS- caler. https://scikit-learn.org/ stable/modules/generated/sklearn. preprocessing.MinMaxScaler.html. Accessed 2026-05-10 (2026)

    2026

  31. [31]

    Neural Networks2(5), 359– 366 (1989) https://doi.org/10.1016/ 0893-6080(89)90020-8

    Hornik, K., Stinchcombe, M., White, H.: Multilayer feedforward net- works are universal approxima- tors. Neural Networks2(5), 359– 366 (1989) https://doi.org/10.1016/ 0893-6080(89)90020-8

    1989

  32. [32]

    https://github.com/ keras-team/keras-tuner (2019)

    O’Malley, T., Bursztein, E., Long, J., Chollet, F., Jin, H., Invernizzi, L., et al.: KerasTuner. https://github.com/ keras-team/keras-tuner (2019)

    2019

  33. [33]

    arXiv:2411.10406v3 (2026)

    Mohseni, M., Scherer, A., John- son, K.G., Wertheim, O., Otten, M., Anand, N., Aadit, N.A., Alexeev, Y., Ben-Shach, G., Bresniker, K.M., Camsari, K.Y., Chapman, B., Chat- terjee, S., Chowdhury, S., Dagnew, G.A., Dvir, T., Esposito, A., Fahim, F., Ferguson, M., Fiorentino, M., Gaj- jar, A., Gratsea, K., Gyawali, G., Heiter, C., Kavaki, A.H.Z., Khalid, A., ...

    Pith/arXiv arXiv 2026

  34. [34]

    In: Verification and Val- idation, vol

    Fong, J.T., Marcal, P.V., Rains- berger, R., Ma, L., Heckert, N.A., Filliben, J.J.: Finite element method solution uncertainty, asymptotic solu- tion, and a new approach to accuracy assessment. In: Verification and Val- idation, vol. 40795, pp. 001–12001 (2018). American Society of Mechan- ical Engineers

    2018

  35. [35]

    https://arxiv

    Sommers, D., Degnan, Z., Gau- tam, D., Chen, Y.-H., Chiu, C.-C., Fedorov, A., Pakkiam, P.: Open- Source Highly Parallel Electromag- netic Simulations for Superconduct- ing Circuits (2025). https://arxiv. org/abs/2511.01220

    arXiv 2025

  36. [36]

    Quantum Science and Technol- ogy9(2), 025006 (2024) https://doi

    Muthusubramanian, N., Finkel, M., Duivestein, P., Zachariadis, C., Meer, S.L.M., Veen, H.M., Beek- man, M.W., Stavenga, T., Bruno, A., DiCarlo, L.: Wafer-scale unifor- mity of Dolan-bridge and bridgeless Manhattan-style Josephson junctions for superconducting quantum proces- sors. Quantum Science and Technol- ogy9(2), 025006 (2024) https://doi. org/10.108...

    doi:10.1088/2058-9565/ad199c 2024

  37. [37]

    Springer, New York (2009)

    Hastie, T., Tibshirani, R., Friedman, J.: The Elements of Statistical Learn- ing: Data Mining, Inference, and Prediction, 2nd edn. Springer, New York (2009)

    2009

  38. [38]

    https://scikit-learn.org/ stable/modules/model evaluation

    scikit-learn developers: Scikit- learn User Guide: Regression metrics. https://scikit-learn.org/ stable/modules/model evaluation. html#regression-metrics. Accessed 2026-05-10 (2026) 30

    2026