pith. sign in

arxiv: 2502.19872 · v2 · pith:HBEPBAQYnew · submitted 2025-02-27 · 🪐 quant-ph

Designing a Machine Learning-Driven, Cross-Hardware Emulator for Noisy Quantum Computers with Gate-Based Protocols

Matthew Ho , Jun Yong Khoo , Adrian M. Mak , Stefano Carrazza This is my paper

Pith reviewed 2026-05-23 02:52 UTC · model grok-4.3

classification 🪐 quant-ph
keywords quantum emulationnoise modelinggate set tomographymachine learningquantum chemistryH2 moleculeunitary coupled cluster
0
0 comments X

The pith

A machine learning model trained on simulated gate set tomography data predicts real quantum device noise with 0.128 percent relative error.

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

The paper establishes a protocol that trains a supervised machine learning model exclusively on simulated gate set tomography data to infer the noise characteristics of a target quantum processor. From those inferred parameters it assembles a device-specific emulator that reproduces hardware behavior using only gate-based measurements. The central demonstration applies the resulting emulator to the unitary coupled cluster calculation of the H2 ground-state energy and obtains an expectation-value error of 0.128 percent relative to the same circuit executed on the physical device. A reader would care because such emulators let programmers test and optimize circuits offline before committing expensive hardware time, and because the method requires no pulse-level access.

Core claim

The protocol uses supervised learning on a library of simulated gate set tomography experiments to map observed gate-set data from a target device onto a compact noise model; that model is then inserted into a classical emulator. When the emulator is used to compute the unitary coupled cluster energy of H2, the expectation value differs from the value measured on the actual hardware by a relative error of 0.128 percent. The same workflow works across hardware platforms and requires only standard gate-based tomography, without pulse-level control.

What carries the argument

The supervised machine learning model that ingests gate set tomography measurement outcomes and outputs the parameters of a device-specific noise model for insertion into the emulator.

If this is right

  • Accurate device-specific emulators can be built from gate-based tomography alone, without pulse-level access.
  • The same trained model can be applied to different quantum processors to produce cross-hardware emulators.
  • Expectation values computed in the emulator for the H2 unitary coupled cluster ansatz match hardware results to 0.128 percent relative error.
  • Noise characterization and independent validation become possible using only standard gate operations.

Where Pith is reading between the lines

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

  • Programmers could iterate circuit designs entirely in simulation before any hardware run, lowering the cost of algorithm development.
  • The approach might be tested on larger molecules or different ansatzes to check whether the same error level holds.
  • Transfer of the learned mapping to devices with modestly different gate sets could be checked by retraining only the final layer.

Load-bearing premise

A supervised model trained only on simulated gate set tomography data will produce noise parameters that accurately reproduce the behavior of real quantum hardware when only gate-based measurements are supplied.

What would settle it

Execute an unseen circuit on the target hardware, run the identical circuit in the ML-constructed emulator, and observe whether the relative error in the measured expectation value exceeds a few tenths of a percent.

Figures

Figures reproduced from arXiv: 2502.19872 by Adrian M. Mak, Jun Yong Khoo, Matthew Ho, Stefano Carrazza.

Figure 1
Figure 1. Figure 1: FIG. 1. Schematic of the proposed AI-powered, generalized gate-based protocol for processing simulated and hardware GST [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Evaluation of NN-2Q using (a) unseen data with pre [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Evaluation of NN-1Q using (a) unseen data with pre [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. (a) Transpiled UCC circuits to IQM Garnet native gates for measurement in the Z, Y, and X bases. Adjusting the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. To illustrate the effects of different noise models on different single-qubit gates, a heatmap containing the 4 by 4 [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
read the original abstract

Quantum computer emulators model the behavior and error rates of specific quantum processors. Without accurate noise models in these emulators, it is challenging for users to optimize and debug executable quantum programs prior to running them on the quantum computer, as device-specific noise is not properly accounted for. To overcome this challenge, we design a machine learning(ML)-driven approach to construct approximate device-specific emulators that applies to different hardware platforms. We apply supervised ML on a pre-generated library containing simulated gate set tomography training data. The ML model then analyses gate set tomography data from a target quantum computer to predict its noise model, which is in turn used to construct the device-specific emulator. We demonstrate the effectiveness of our protocol's emulator in estimating the unitary coupled cluster energy of the H$_2$ molecule and compare the results with those from actual quantum hardware. Remarkably, our noise model captures device noise with high accuracy, achieving a percentage relative error of just 0.128\% in expectation value relative to the actual quantum hardware. Importantly, we show that even without access to pulse-level control, noise from the quantum computer can nonetheless be characterized and independently validated by our protocol.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes an ML-driven protocol to build approximate device-specific emulators for noisy quantum computers. Supervised learning is applied to a library of simulated gate-set tomography (GST) data to train a model that, given GST data from a target device, predicts a noise model; this model is then used to construct a gate-based emulator. The approach is demonstrated by estimating the unitary coupled-cluster energy of H2 on the emulator and comparing to hardware results, with a reported relative error of 0.128% in the expectation value. The authors emphasize that the method operates without pulse-level control.

Significance. A reliable cross-hardware emulator that infers noise parameters from gate-based measurements alone would be useful for pre-execution debugging and circuit optimization. The reported 0.128% error on the H2 UCC observable is numerically small, but the single-observable, single-molecule validation does not yet establish that the inferred noise model is faithful rather than merely compensatory on that particular circuit.

major comments (3)
  1. [Abstract and demonstration section] The central empirical claim (0.128% relative error on the H2 UCC energy) is presented without any description of the ML model architecture, training procedure, data volume, cross-validation strategy, or uncertainty quantification. This information is required to evaluate whether the low error reflects genuine noise-model fidelity or an under-constrained fit.
  2. [Results / demonstration] Validation is limited to a single low-depth circuit and observable (UCC energy of H2). Agreement on this scalar does not rule out under-parameterized noise models, error cancellation, or limited test coverage; parameter-level validation against known GST parameters or tests on additional circuits/observables are absent.
  3. [Methods / protocol description] The manuscript states that the ML model is trained exclusively on simulated GST data yet is applied to real-device GST data. No analysis is provided of the domain gap between simulated and experimental GST or of how the model generalizes when only gate-based measurements are available.
minor comments (2)
  1. [Methods] Notation for the predicted noise parameters and the emulator construction step should be defined explicitly with equations rather than prose descriptions.
  2. [Abstract] The abstract claims 'high accuracy' on the basis of one scalar; a brief statement of the test-circuit depth and the number of independent hardware runs would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their detailed and constructive report. We address each major comment below and indicate where revisions have been made to the manuscript.

read point-by-point responses

  1. Referee: [Abstract and demonstration section] The central empirical claim (0.128% relative error on the H2 UCC energy) is presented without any description of the ML model architecture, training procedure, data volume, cross-validation strategy, or uncertainty quantification. This information is required to evaluate whether the low error reflects genuine noise-model fidelity or an under-constrained fit.

    Authors: We agree that these details are necessary for proper evaluation. The submitted manuscript focused on the protocol and demonstration but omitted full methodological specifications. In the revised manuscript we have added a dedicated subsection in Methods that specifies the ML model architecture, training procedure on the simulated GST library, data volume, cross-validation approach, and uncertainty quantification method. These additions allow readers to assess whether the reported error arises from faithful noise modeling. revision: yes

  2. Referee: [Results / demonstration] Validation is limited to a single low-depth circuit and observable (UCC energy of H2). Agreement on this scalar does not rule out under-parameterized noise models, error cancellation, or limited test coverage; parameter-level validation against known GST parameters or tests on additional circuits/observables are absent.

    Authors: We acknowledge the limitation of validating on a single observable and molecule. The H2 UCC energy serves as a standard benchmark for quantum chemistry emulation, and the 0.128% relative error demonstrates practical utility for cross-hardware transfer. However, we agree that scalar agreement alone does not fully exclude compensatory effects. In revision we have added an explicit limitations paragraph discussing possible error cancellation and the absence of parameter-level GST recovery, while noting that the protocol's design prioritizes observable-level fidelity for emulation purposes. Additional circuit tests were not available in the current dataset. revision: partial

  3. Referee: [Methods / protocol description] The manuscript states that the ML model is trained exclusively on simulated GST data yet is applied to real-device GST data. No analysis is provided of the domain gap between simulated and experimental GST or of how the model generalizes when only gate-based measurements are available.

    Authors: The training uses simulated GST to learn a mapping from gate-based measurements to noise parameters, which is then applied to experimental GST. GST is constructed to extract device-independent parameters from gate sequences, providing a natural bridge. We agree an explicit domain-gap analysis is valuable. The revised manuscript includes a new paragraph in Methods that discusses the assumptions of the simulation-to-experiment transfer, the role of GST standardization in reducing domain shift, and empirical checks on generalization to real gate-based data. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical validation on held-out hardware data is independent of training inputs

full rationale

The paper trains a supervised ML model exclusively on simulated GST data to map GST measurements to noise parameters, then applies the trained model to real-device GST data to infer a noise model for the emulator. The central claim (0.128% relative error on H2 UCC energy) is an empirical comparison between the emulator output and actual hardware measurements on a held-out circuit and observable. This constitutes an external test rather than a reduction by construction: the ML prediction step operates on real GST inputs that were never seen during training, and the reported error is not a fitted quantity but a post-hoc match on device data. No self-definitional equations, fitted-input predictions, or load-bearing self-citations appear in the abstract or described protocol. The derivation chain remains self-contained against external hardware benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities are identifiable from the provided text.

pith-pipeline@v0.9.0 · 5745 in / 1088 out tokens · 27369 ms · 2026-05-23T02:52:53.022751+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. From Characterization To Construction: Generative Quantum Circuit Synthesis from Gate Set Tomography Data

    quant-ph 2026-05 unverdicted novelty 6.0

    A generative QMLC framework tokenizes GST data, embeds it via curriculum-trained set-vision transformers into a context-aware latent space, and uses diffusion models to synthesize circuits conditioned on desired measu...

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    a sufficiently comprehensive noise model with a set of input noise parameters capable of emulating quantum hardware noise,

    work page

  2. [2]

    a set of gate-based circuits to execute on quan- tum hardware that serve to collect characterization data,

    work page

  3. [3]

    In this work, we concretize this generalized protocol by making a specific choice for each key component to construct the GST-trained heuristic (GTH) noise model

    a machine learning framework to predict the noise parameters corresponding to the device based on its characterization data. In this work, we concretize this generalized protocol by making a specific choice for each key component to construct the GST-trained heuristic (GTH) noise model. A schematic of the protocol is summarized in Fig. 1, and detailed in ...

    work page

  4. [4]

    Define ⃗λ := ( λd, λa, λf , λr), representing single- qubit depolarizing noise, amplitude damping noise, dephasing noise, and readout noise parameters, re- spectively

    work page

  5. [5]

    Define a discrete range for ⃗λ

    work page

  6. [6]

    1) to generate single-qubit simulated GST data

    For each value in the range for ⃗λ, use single-qubit noise model M⃗λ (Eq. 1) to generate single-qubit simulated GST data. The full dataset is then used to train NN-1Q

    work page

  7. [7]

    Perform single-qubit GST for qubit qi on hardware, input the results into the trained NN-1Q to obtain predicted noise ⃗λqi

    work page

  8. [8]

    Repeat Step 4 on qubit qj (which is coupled to qi) to get ⃗λqj

    work page

  9. [9]

    2) where ⃗ζ represents the two-qubit depolarizing noise

    Construct two-qubit composite noise model N ⃗λqi , ⃗λqj ,⃗ζ (Eq. 2) where ⃗ζ represents the two-qubit depolarizing noise

    work page

  10. [10]

    Define a discrete range for ⃗ζ

    work page

  11. [11]

    The full dataset is used to train NN-2Q

    For each value in the range for ⃗ζ, use N⃗λqi ,⃗λqj ,⃗ζ to generate two-qubit simulated GST data. The full dataset is used to train NN-2Q

    work page

  12. [12]

    Perform two qubit GST for qubit qi and qj on hard- ware, input the results into the trained NN-2Q to obtain the predicted two-qubit depolarizing noise ⃗ζqi,qj

    work page

  13. [13]

    predicted vs true

    Finally, form the GTH noise model Nqi,qj := N⃗λqi ,⃗λqj ,⃗ζqi ,qj . 5 We provide a more detailed explanation of the training procedure, including the choice of step sizes for ⃗λ and ⃗ζ, in Appendix C. E. T raining and evaluating neural networks for prediction The third and final component of our protocol is a machine learning framework that can be used to...

    work page

  14. [14]

    [29], we provide a detailed description of the UCC method elsewhere in Appendix A, and shall only provide the important experimental de- tails here

    Unitary Coupled Cluster energy of the H 2 molecule We present here a test of Nqi,qj on an example quan- tum chemistry calculation, namely the unitary coupled cluster (UCC) energy of the H2 molecule with atomic sep- aration of 0.70 ˚Ausing the STO-6G basis set, as demon- strated by O’Malley et al. [29], we provide a detailed description of the UCC method e...

    work page 2097

  15. [15]

    Quantum Computing in the NISQ era and beyond,

    J. Preskill, “Quantum Computing in the NISQ era and beyond,” Quantum, vol. 2, p. 79, Aug. 2018

    work page 2018

  16. [16]

    Amazon Braket

    “Amazon Braket.” https://aws.amazon.com/braket/. Accessed: 2025-02-19

    work page 2025

  17. [17]

    IBM Quantum

    “IBM Quantum.” https://quantum.ibm.com/. 2021

    work page 2021

  18. [18]

    Quantinuum

    “Quantinuum.” https://www.quantinuum.com/

    work page

  19. [19]

    Benchmarking a trapped-ion quantum computer with 30 qubits,

    J.-S. Chen, E. Nielsen, M. Ebert, V. Inlek, K. Wright, V. Chaplin, A. Maksymov, E. P´ aez, A. Poudel, P. Maunz, et al., “Benchmarking a trapped-ion quantum computer with 30 qubits,” Quantum, vol. 8, p. 1516, 2024

    work page 2024

  20. [20]

    Aquila: Quera’s 256-qubit neutral-atom quantum computer,

    J. Wurtz, A. Bylinskii, B. Braverman, J. Amato-Grill, S. H. Cantu, F. Huber, A. Lukin, F. Liu, P. Weinberg, J. Long, et al., “Aquila: Quera’s 256-qubit neutral-atom quantum computer,” arXiv preprint arXiv:2306.11727 , 2023

    work page arXiv 2023

  21. [21]

    Boothby, P

    K. Boothby, P. Bunyk, J. Raymond, and A. Roy, “Next- generation topology of d-wave quantum processors,” arXiv preprint arXiv:2003.00133 , 2020

    work page arXiv 2003

  22. [22]

    A variational eigenvalue solver on a photonic quantum processor,

    A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’brien, “A variational eigenvalue solver on a photonic quantum processor,” Nature communications, vol. 5, no. 1, p. 4213, 2014

    work page 2014

  23. [23]

    A Quantum Approximate Optimization Algorithm

    E. Farhi, J. Goldstone, and S. Gutmann, “A quantum approximate optimization algorithm,” arXiv preprint arXiv:1411.4028, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  24. [24]

    Variational quantum algorithms,

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, et al., “Variational quantum algorithms,” Na- ture Reviews Physics, vol. 3, no. 9, pp. 625–644, 2021

    work page 2021

  25. [25]

    Simulating noisy variational quantum eigen- solver with local noise models,

    J. Zeng, Z. Wu, C. Cao, C. Zhang, S.-Y. Hou, P. Xu, and B. Zeng, “Simulating noisy variational quantum eigen- solver with local noise models,” Quantum Engineering , vol. 3, no. 4, p. e77, 2021

    work page 2021

  26. [26]

    Nisq computing: where are we and where do we go?,

    J. W. Z. Lau, K. H. Lim, H. Shrotriya, and L. C. Kwek, “Nisq computing: where are we and where do we go?,” AAPPS bulletin , vol. 32, no. 1, p. 27, 2022. 9

    work page 2022

  27. [27]

    Nonlinear quantum dynamics in superconducting nisq processors,

    M. Umer, E. Mastorakis, S. Evangelou, and D. G. Ange- lakis, “Nonlinear quantum dynamics in superconducting nisq processors,” arXiv preprint arXiv:2403.16426, 2024

    work page arXiv 2024

  28. [28]

    Computing electronic correlation en- ergies using linear depth quantum circuits,

    C. H. Chee, A. M. Mak, D. Leykam, P. Barkoutsos, and D. G. Angelakis, “Computing electronic correlation en- ergies using linear depth quantum circuits,” Quantum Science and Technology, vol. 9, p. 025003, 2024

    work page 2024

  29. [29]

    Randomized benchmarking of quantum gates,

    E. Knill, D. Leibfried, R. Reichle, J. Britton, R. B. Blakestad, J. D. Jost, C. Langer, R. Ozeri, S. Sei- delin, and D. J. Wineland, “Randomized benchmarking of quantum gates,” Physical Review A—Atomic, Molec- ular, and Optical Physics , vol. 77, no. 1, p. 012307, 2008

    work page 2008

  30. [30]

    Char- acterizing quantum gates via randomized benchmark- ing,

    E. Magesan, J. M. Gambetta, and J. Emerson, “Char- acterizing quantum gates via randomized benchmark- ing,” Physical Review A—Atomic, Molecular, and Op- tical Physics, vol. 85, no. 4, p. 042311, 2012

    work page 2012

  31. [31]

    Introduction to Quantum Gate Set Tomography

    D. Greenbaum, “Introduction to quantum gate set to- mography,” arXiv preprint arXiv:1509.02921 , 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  32. [32]

    Demonstration of qubit operations below a rigorous fault tolerance threshold with gate set tomography,

    R. Blume-Kohout, J. K. Gamble, E. Nielsen, K. Rudinger, J. Mizrahi, K. Fortier, and P. Maunz, “Demonstration of qubit operations below a rigorous fault tolerance threshold with gate set tomography,” Nature communications, vol. 8, no. 1, p. 14485, 2017

    work page 2017

  33. [33]

    Gate set tomogra- phy,

    E. Nielsen, J. K. Gamble, K. Rudinger, T. Scholten, K. Young, and R. Blume-Kohout, “Gate set tomogra- phy,” Quantum, vol. 5, p. 557, Oct. 2021

    work page 2021

  34. [34]

    Modeling and simulating the noisy behavior of near-term quantum computers,

    K. Georgopoulos, C. Emary, and P. Zuliani, “Modeling and simulating the noisy behavior of near-term quantum computers,” Physical Review A, vol. 104, no. 6, p. 062432, 2021

    work page 2021

  35. [35]

    A methodology to select and adjust quantum noise models through emulators: bench- marking against real backends,

    J. Bravo-Montes, M. Bastante, G. Botella, A. del Barrio, and F. Garc´ ıa-Herrero, “A methodology to select and adjust quantum noise models through emulators: bench- marking against real backends,” EPJ Quantum Technol- ogy, vol. 11, no. 1, p. 71, 2024

    work page 2024

  36. [36]

    Digital- analog quantum algorithm for the quantum fourier trans- form,

    A. Martin, L. Lamata, E. Solano, and M. Sanz, “Digital- analog quantum algorithm for the quantum fourier trans- form,” Physical Review Research, vol. 2, no. 1, p. 013012, 2020

    work page 2020

  37. [37]

    “IQM.” https://www.meetiqm.com/

    work page

  38. [38]

    Technology and per- formance benchmarks of iqm’s 20-qubit quantum com- puter,

    L. Abdurakhimov, J. Adam, H. Ahmad, O. Aho- nen, M. Algaba, G. Alonso, V. Bergholm, R. Beriwal, M. Beuerle, C. Bockstiegel, et al., “Technology and per- formance benchmarks of iqm’s 20-qubit quantum com- puter,” arXiv preprint arXiv:2408.12433 , 2024

    work page arXiv 2024

  39. [39]

    M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information . 2010

    work page 2010

  40. [40]

    Bosonic quantum codes for amplitude damping,

    I. L. Chuang, D. W. Leung, and Y. Yamamoto, “Bosonic quantum codes for amplitude damping,” Physical Review A, vol. 56, no. 2, pp. 1114–1125, 1997

    work page 1997

  41. [41]

    Qibo: a framework for quan- tum simulation with hardware acceleration,

    S. Efthymiou, S. Ramos-Calderer, C. Bravo-Prieto, A. P´ erez-Salinas, D. Garc´ ıa-Mart´ ın, A. Garcia-Saez, J. I. Latorre, and S. Carrazza, “Qibo: a framework for quan- tum simulation with hardware acceleration,” Quantum Science and Technology, vol. 7, no. 1, p. 015018, 2021

    work page 2021

  42. [42]

    GTN training data and weights

    “GTN training data and weights.” https://github.com/ qiboteam/gstdata4ml

    work page

  43. [43]

    Scalable quantum simulation of molecular energies,

    P. J. J. O’Malley, R. Babbush, I. D. Kivlichan, J. Romero, J. R. McClean, R. Barends, J. Kelly, P. Roushan, A. Tranter, N. Ding, B. Campbell, Y. Chen, Z. Chen, B. Chiaro, A. Dunsworth, A. G. Fowler, E. Jef- frey, E. Lucero, A. Megrant, J. Y. Mutus, M. Neeley, C. Neill, C. Quintana, D. Sank, A. Vainsencher, J. Wen- ner, T. C. White, P. V. Coveney, P. J. Lo...

    work page 2016

  44. [44]

    Fermionic quantum computa- tion,

    S. Bravyi and A. Kitaev, “Fermionic quantum computa- tion,” Annals of Physics , vol. 298, p. 210, 2002. 10 Appendix A: Unitary Coupled Cluster For a molecular system of clamped nuclei (Born-Oppenheimer approximation), the electronic Hamiltonian in second quantization using creation and annihilation operators ˆa† and ˆa is written as ˆHelec = X pq hpqˆa† p...

    work page 2002