Physics-Informed Learning of Effective Error Processes from Limited Noisy Transmon Measurements for Robust QAOA Reliability
Ebrahim Khaleghian 1 , Özgür E. Müstecaplıoğlu · 2026
Read Physics-Informed Learning of Effective Error Processes from Limited Noisy Transmon Measurements for Robust QAOA Reliability by Ebrahim Khaleghian 1 , Özgür E. Müstecaplıoğlu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-00353.txt.
Introduction
Khaleghian and Müstecaplıoğlu learn small error maps for quantum chips from scant, noisy reads. Their test rig mimics a transmon chip in code. Bits are three-grade units with flaws in steer, loss, leak, and read. The learner sees just picked local probes of that rig.
Learned maps take a plain shift form for each bit's state ball. Edge scraps from pair probes add the next layer of truth. Tests ask if the maps can clone and fix warps in QAOA cost views. Plain ruled fits tie nets in the three-bit scale-up. Pair probes give the clearest edge gains at full spend. A fix based on the maps cuts land error tenfold in the best test.
Catalog record
| CREATOR | Ebrahim Khaleghian 1 , Özgür E. Müstecaplıoğlu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.00353
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Physics-Informed Learning of Effective Error Processes from Limited Noisy Transmon Measurements for Robust QAOA Reliability Ebrahim Khaleghian 1 , Özgür E. Müstecaplıoğlu Abstract We present a physics-informed pipeline for learning effective quantum error processes from finite-shot measurements generated by hidden transmon-like simulators. Each physical transmon is modeled as an imperfect qutrit, while the learner only receives limited tomography data rather than microscopic Hamiltonian parameters. The learned representations are compact effective models: local affine Bloch channels for each qubit and, in the three-qubit extension, pairwise residuals that capture correlated errors. The learned error models are evaluated operationally by their ability to mitigate the cost landscape of the Quantum Approximate Optimization Algorithm (QAOA) for MaxCut. A two-qubit proof of concept shows that a neural-network approach can infer a full 24-parameter effective channel from only 12 local tomography values and improve QAOA landscape reliability by about 20.4 × 20.4\times . A scaled three-qubit study shows that local structured learning still strongly improves QAOA reliability: at K = 18 K=18 local measurements, Ridge regression and the neural-network approach reduce QAOA mean absolute error from about 0.1775 0.1775 to 0.0269 0.0269 and 0.0306 0.0306 , respectively. Pair probes substantially improve correlated-error identifiability, reducing pair-residual L2 error from about 1.731 1.731 to 1.122 1.122 . These results support effective error-process learning as a hardware-aware route toward more reliable variational quantum algorithms. I Introduction Noisy intermediate-scale quantum devices implement algorithms through imperfect physical dynamics rather than ideal gates. In superconducting circuits, the observed error process is shaped by weak anharmonicity, coherent detuning and coupling, residual interactions, dissipation, dephasing, leakage to noncomputational levels, readout assignment errors, and finite sampling noise Koch2007Transmon ; Krantz2019Guide ; Blais2021CircuitQED ; Motzoi2009DRAG . Quantum error mitigation (QEM) attempts to reduce the bias of noisy expectation values without full fault tolerance Temme2017QEM ; Endo2018PracticalQEM ; Cai2023QEMReview . However, many mitigation approaches require either additional circuit executions, noise-scaling assumptions, or training data that are not directly tied to a physically structured device model. This work asks a more operational question. Can limited, finite-shot measurements from a hidden transmon-like device be used to learn effective error processes that improve the reliability of a target algorithm? The target algorithm is the Quantum Approximate Optimization Algorithm (QAOA) for MaxCut Farhi2014QAOA ; Blekos2024QAOAReview ; Zhou2020QAOA . The physical simulator is intentionally hidden from the learner: it knows the Hamiltonian, open-system dynamics, readout error, and finite-shot sampling process, but the learning model receives only local and pairwise measurement features. The learning target is not microscopic hardware-parameter recovery. Instead, the target is a compact operational representation of the effective error process. The paper is organized as a progression from a clean two-qubit proof of concept to a three-qubit hardware-aware extension. In two qubits, partial local tomography is deliberately underdetermined: the learner receives only 12 finite-shot tomography values but must infer a full 24-parameter local channel. This validates the idea that a neural model can act as a learned physical prior over a structured family of hardware-generated error processes. In three qubits, the setting becomes more realistic: local channels have 36 parameters, local tomography has 54 possible settings, and correlated errors from two edges must be considered. The three-qubit results therefore shift the emphasis from neural networks to a more general claim: structured learning methods, including regularized linear models and neural models, can learn useful effective error processes, while pairwise correlated errors require targeted pair probes. The remainder of the paper is organized as follows. Section II positions the work with respect to superconducting-transmon hardware, quantum process characterization, learning-based error mitigation, and adaptive measurement ideas. Section III gives an overview of the proposed architecture, from hidden physical data generation to effective error learning and QAOA-level mitigation. Section IV describes the qutrit transmon simulator and the noise taxonomy. Section V introduces the local tomography and pair-probe measurement protocols, while Section VI defines the effective error representations used for learning. Section VII explains the QAOA/MaxCut benchmark. Section VIII describes the learning models, evaluation metrics, and the model-based bias-correction rule. Sections IX–XI present the numerical results. Section XII outlines future directions. Section XIII concludes the work. The appendices provide additional implementation and diagnostic material. Appendix A summarizes the numerical simulator, hidden Hamiltonian and noise-parameter distributions, and model architectures. Appendix B collects supplementary two-qubit diagnostics. Appendix C gives additional three-qubit local-learning and QAOA diagnostics, while Appendix D reports exploratory robustness and stress-test figures for different physically motivated noise families. II Related work The physical side of the work is rooted in superconducting transmon qubits. The transmon was introduced as a charge-insensitive superconducting qubit derived from the Cooper-pair box Koch2007Transmon , and modern superconducting-qubit engineering is reviewed in Ref. Krantz2019Guide . Circuit QED provides the broader control and readout context Blais2021CircuitQED . Because transmons are weakly anharmonic oscillators, leakage beyond the logical two-level subspace is a central control issue; DRAG-style pulse shaping was introduced partly to suppress leakage in weakly nonlinear qubits Motzoi2009DRAG . The measurement and characterization side is related to quantum process tomography and gate set tomography. Gate set tomography (GST) provides detailed, predictive and self-consistent characterization of quantum logic operations Nielsen2021GST ; simultaneous GST has also been used to study crosstalk in multiqubit systems Rudinger2021SimGST . Compressed sensing tomography and compressed process tomography show that structural assumptions can reduce measurement burden Gross2010Compressed ; Rodionov2014CompressedQPT . The present work differs from full GST and compressed tomography because it does not aim at a complete characterization of an arbitrary process. It learns a low-dimensional, task-oriented effective error model from limited local and pair measurements. Machine learning has also been used for quantum state and process tomography. Tensor-network process tomography Torlai2023TensorQPT , neural-network-assisted state and process tomography with limited data Gaikwad2024NNTomography , and Kraus-operator learning Ahmed2023KrausLearning all support the broader idea that structure can be learned from incomplete measurement data. Learning-based QEM and Clifford-data regression frame mitigation as a supervised or regression problem Strikis2021LearningQEM ; Czarnik2021CDR . Recent work on practical ML-QEM compares linear models, random forests, multilayer perceptrons, and graph neural networks across larger circuit families Liao2024MLQEM . Finally, adaptive tomography and active measurement selection provide a natural future direction. Active-learning quantum tomography Lange2023ActiveQST , neural adaptive tomography Quek2021AdaptiveNN , and active process-tomography proposals Yang2025ActiveQPT motivate a closed-loop extension in which the next hardware measurement is chosen from model uncertainty. III Architecture overview Figure 1 summarizes the architecture. The hidden physical layer generates data from a qutrit transmon simulator. The learner observes finite-shot local tomography and, in the three-qubit extension, pair probes. The learning layer predicts local affine channels and pair residuals. These effective error models are then used to predict and mitigate QAOA cost landscapes. Figure 1: Overall architecture. The simulator contains hidden qutrit transmon physics, Hamiltonian-level imperfections, open-system noise, leakage, readout assignment error, and finite-shot sampling. The learner sees only limited measurement data and predicts effective error processes. The final implemented experiments use fixed local and pair measurement budgets. IV Hidden transmon physical model and noise taxonomy Figure 2 summarizes the noise architecture. IV.1 Qutrit transmon truncation Each physical transmon is truncated to three levels, |0\rangle , |1\rangle , and |2\rangle . The computational subspace is \{|0\rangle,|1\rangle\} , while |2\rangle allows leakage to be represented explicitly. This qutrit representation makes the hidden device more