Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2607.20795

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, Haoyu Yang, Brucek Khailany, Sara Sussman, Enectali Figueroa-Feliciano, Eli M Levenson-Falk, Taylor L. Patti

quant-phSubmitted Jul 22, 2026• Announced Jul 24, 2026• 0 citations
v1 · Jul 22, 2026v2 · Jul 30, 2026
arXivHTMLPDF

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 approximately 56 ms on CPU, versus approximately 2 min for a conventional EM capacitance extraction on the same hardware, a speedup of more than 2,100 times. Batching minimizes the AI model inference overhead, reducing the runtime to 0.24 ms per sample on CPU and 2.5 microseconds per sample on GPU, resulting in speedups of 5.0 x 10^5 and 4.8 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.

Related papers

Neural-Network Inverse Design of SRF Cavities and Transmons for Bosonic Quantum Computation

Joseph Yaker, Jovan Markovic, Alessandro Reineri +2 more

Three-dimensional superconducting radio-frequency (SRF) cavities provide exceptionally long-lived electromagnetic modes and, when coupled to nonlinear elements such as transmon qubits, become promising architectures for bosonic quantum information processing. The inverse design of such systems, i.e., recovering device geometries that produce specified electromagnetic and coupling targets, is generally a one-to-many problem. The qubit-cavity coupling strength depends sensitively on both the transmon geometry and its position within the cavity's electromagnetic field. As these systems scale up and their design parameter spaces grow, the cost of conventional iterative simulation becomes prohibitive. We present two deep neural network (DNN) approaches that address this inverse-design problem at complementary levels of the design stack. The first proposes SRF cavity geometries that produce target cavity observables. The second proposes transmon qubit designs that produce target qubit-cavity parameters - the coupling rate, qubit frequency, and anharmonicity $(g, ν_q, α)$. The recovered candidate designs match the targets to within ~5% (cavity) and ~2% (transmon), confirmed by end-to-end re-simulation. Both approaches map desired device behavior directly to candidate designs, a fast alternative to the iterative simulation studies usually required.

quant-pharXiv:2607.02289•Jul 3, 2026•0 citations

Waveform-driven prediction of transmon qubit energy spectra using a physics-informed path-transformed transformer trained on Josephson junction dynamics.

Md. Ifthakhar Khan Sagor, Md. Sanwar Hossain, Md. Zillur Rahman +2 more

Extracting the energy spectrum of a superconducting transmon qubit requires time-consuming microwave spectroscopy that must be repeated following every thermal cycle or parameter drift event. We introduce the Physics-Informed Path-Transformed (PaTH) Transformer, a surrogate model trained exclusively on resistively-shunted Josephson junction (RCSJ) numerical simulations, which predicts the five lowest energy eigenvalues and the anharmonicity from a raw DC voltage waveform. The model is trained on 10,000 simulated circuits spanning the realistic transmon parameter space (EJ /EC ∈ [10.7, 300], f01/h ∈ [2.6, 13.3] GHz) and evaluated on 1,500 independent test circuits, achieving mean absolute energy error 0.886 μeV (R2 ≥ 0.998 per level) and transition-frequency spacing error 2.14%, with GPU inference completing in 8.70 ms per circuit. A curriculum-ramped physics loss incorporating Josephson monotonicity, quantization spacing, anharmonicity sign, and Lindblad-Heisenberg constraints produces physically self-consistent predictions throughout training. Treating anharmonicity as a direct sixth output resolves a gradient-scale conflict (∼25× imbalance) that prevents reliable convergence when it is inferred post-hoc from level differences. The model delivers, from a single forward pass: DRAG gate fidelity ⟨F⟩ = 0.99982 at 3.9 ns; dispersive readout shift −14.3 MHz; macroscopic quantum tunneling crossover at 56.6 mK; SQUID tunable range 5.95 GHz; and inverse-design frequency errors below 0.05 kHz. Zero-noise extrapolation across five dilution-refrigerator temperatures confirms that model accuracy already reaches the Johnson-Nyquist noise floor, with the macroscopic quantum tunneling crossover temperature providing a physics-derived accuracy benchmark independent of model architecture. Validated against eight published experimental spectra from four independent laboratories, the surrogate achieves mean f01 error 1929 MHz versus 1935 MHz for exact Hamiltonian diagonalization, demonstrating that RCSJ waveform encoding conveys device-level physics that the static ideal Hamiltonian cannot represent.

PLoS ONE•Sep 4, 2026•0 citations

Multimode Duffing–Jaynes–Cummings Hamiltonian modeling of a coupled transmon chip layout

Nongmeikapam Brajabidhu Singh, Joseph L. Pachuau, Anish Kumar Saha

This paper presents an electromagnetic-to-Hamiltonian workflow for modeling a two-qubit coupled transmon superconducting device circuit. The device layout, comprising two coupled transmon qubits with individual readout resonators, is designed and parameterized in Qiskit Metal and analyzed using 3D eigenmode simulations in Ansys HFSS. Adaptive-pass refinement converges the eigenfrequencies and electromagnetic fields, yielding reliable modal data for quantization. Energy-participation ratio values from HFSS are processed through the pyEPR–Qiskit Metal flow to extract effective capacitances, anharmonicities, and qubit–resonator couplings. These parameters define a multimode Duffing–Jaynes–Cummings Hamiltonian that captures the transmon nonlinearities and their interactions with the resonator modes. Frequency-convergence checks and coupling diagnostics verify a consistent, physically accurate mapping from the layout geometry to the device circuit-level model. The Hamiltonian workflow offers a reproducible and scalable framework for modeling and analyzing superconducting transmon architectures in quantum chip design.

IEEE International Conference on AI Engineering and Innovations (AIEI)•Mar 26, 2026•0 citations

Transmon Qubit Modeling and Characterization for Dark Matter Search

Roberto Moretti, Danilo Labranca, Pietro Campana +21 more

This study presents the design, simulation, and experimental characterization of a superconducting transmon qubit circuit prototype for potential applications in dark matter detection experiments. We describe a planar circuit design featuring two non-interacting transmon qubits, one with fixed frequency and the other flux tunable. Finite-element simulations were employed to extract key Hamiltonian parameters and optimize component geometries. The qubit was fabricated and then characterized at 20 mK, allowing for a comparison between simulated and measured qubit parameters. Good agreement was found for transition frequencies and anharmonicities (within 1% and 10% respectively) while coupling strengths exhibited larger discrepancies (30%). We discuss potential causes for measured coherence times falling below expectations (T1 ∼1-2 μs) and propose strategies for future design improvements. Notably, we demonstrate the application of a hybrid 3D-2D simulation approach for energy participation ratio evaluation, yielding a more accurate estimation of dielectric losses. This work represents an important first step in developing planar Quantum Non-Demolition (QND) single-photon counters for dark matter searches, particularly for axion and dark photon detection schemes.

arXiv:2409.05988•IEEE Transactions on Quantum Engineering•Nov 14, 2025•4 citations

SQChip-EM: A Qiskit-Metal-Based Layout-to-EM Simulation Dataset for Superconducting Quantum Chips

Yu Peng, Junyu Chen, Yuhan Chen +5 more

Translating superconducting quantum chip layouts into physically meaningful simulation records is a critical step in quantum hardware development, yet current datasets rarely capture the electromagnetic (EM) effects involved in this process. We propose SQChip-EM, an open layout to EM simulation dataset based on Qiskit Metal that links parameterized layout designs to Ansys HFSS and Q3D simulation results. Unlike abstract circuit datasets, SQChip-EM preserves layout dependent EM evidence through GDSII files, raw simulation outputs, and JSON annotations covering frequencies, anharmonicity, dispersive shifts, and coupling parameters. These annotations form reusable simulation level ''virtual chip'' records for downstream workflows such as initialization, calibration tuning, and control simulation. SQChip-EM also includes three illustrative benchmark tasks, including design to metric regression, specification feasibility classification, and inverse design retrieval, together with one downstream workflow use case. By releasing data, code, and evaluation protocols, SQChip-EM provides a standardized data resource for ML assisted quantum electronic design automation.

ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)•Aug 6, 2026•0 citations

Machine Learning-Based Predictive Modeling for Designing Transmon Superconducting Qubits

Ferris Prima Nugraha, Qiming Shao

Scaling of superconducting-based quantum computers has been possible since the invention of transmon qubits due to their low charge noise sensitivity. However, pursuing fault-tolerant and computationally powerful quantum processors may require incorporating more qubits, which may amplify the complexity of the large design space exploration and simulation. Inspired by the recent CMOS Machine Learning assisted electronic design automation to solve similar issues, we claim this work is the first to demonstrate applying machine learning in superconducting circuit design. To start, we attempt to predict the characteristics of individual transmon qubits with a machine learning-based approach based on the simulation data collected with relevant software frameworks like Qiskit Metal and ANSYS Electronics. Our approach aims to perform forward modeling to predict the qubit engineering parameters such as frequency. On the other hand, the machine learning-based model can also infer, through inverse modeling, the candidate sets of design space parameters like the device dimensions satisfying a particular specification while obeying the geometrical constraint of the transmon types. This work paves an alternative way for future applications in designing and automating superconducting quantum computing circuits.

IEEE International Conference on Quantum Computing and Engineering (QCE)•Sep 17, 2023•3 citations
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 approximately 56 ms on CPU, versus approximately 2 min for a conventional EM capacitance extraction on the same hardware, a speedup of more than 2,100 times. Batching minimizes the AI model inference overhead, reducing the runtime to 0.24 ms per sample on CPU and 2.5 microseconds per sample on GPU, resulting in speedups of 5.0 x 10^5 and 4.8 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.
Three-dimensional superconducting radio-frequency (SRF) cavities provide exceptionally long-lived electromagnetic modes and, when coupled to nonlinear elements such as transmon qubits, become promising architectures for bosonic quantum information processing. The inverse design of such systems, i.e., recovering device geometries that produce specified electromagnetic and coupling targets, is generally a one-to-many problem. The qubit-cavity coupling strength depends sensitively on both the transmon geometry and its position within the cavity's electromagnetic field. As these systems scale up and their design parameter spaces grow, the cost of conventional iterative simulation becomes prohibitive. We present two deep neural network (DNN) approaches that address this inverse-design problem at complementary levels of the design stack. The first proposes SRF cavity geometries that produce target cavity observables. The second proposes transmon qubit designs that produce target qubit-cavity parameters - the coupling rate, qubit frequency, and anharmonicity (g,νq,α)(g, ν_q, α)(g,νq​,α). The recovered candidate designs match the targets to within ~5% (cavity) and ~2% (transmon), confirmed by end-to-end re-simulation. Both approaches map desired device behavior directly to candidate designs, a fast alternative to the iterative simulation studies usually required.
Extracting the energy spectrum of a superconducting transmon qubit requires time-consuming microwave spectroscopy that must be repeated following every thermal cycle or parameter drift event. We introduce the Physics-Informed Path-Transformed (PaTH) Transformer, a surrogate model trained exclusively on resistively-shunted Josephson junction (RCSJ) numerical simulations, which predicts the five lowest energy eigenvalues and the anharmonicity from a raw DC voltage waveform. The model is trained on 10,000 simulated circuits spanning the realistic transmon parameter space (EJ /EC ∈ [10.7, 300], f01/h ∈ [2.6, 13.3] GHz) and evaluated on 1,500 independent test circuits, achieving mean absolute energy error 0.886 μeV (R2 ≥ 0.998 per level) and transition-frequency spacing error 2.14%, with GPU inference completing in 8.70 ms per circuit. A curriculum-ramped physics loss incorporating Josephson monotonicity, quantization spacing, anharmonicity sign, and Lindblad-Heisenberg constraints produces physically self-consistent predictions throughout training. Treating anharmonicity as a direct sixth output resolves a gradient-scale conflict (∼25× imbalance) that prevents reliable convergence when it is inferred post-hoc from level differences. The model delivers, from a single forward pass: DRAG gate fidelity ⟨F⟩ = 0.99982 at 3.9 ns; dispersive readout shift −14.3 MHz; macroscopic quantum tunneling crossover at 56.6 mK; SQUID tunable range 5.95 GHz; and inverse-design frequency errors below 0.05 kHz. Zero-noise extrapolation across five dilution-refrigerator temperatures confirms that model accuracy already reaches the Johnson-Nyquist noise floor, with the macroscopic quantum tunneling crossover temperature providing a physics-derived accuracy benchmark independent of model architecture. Validated against eight published experimental spectra from four independent laboratories, the surrogate achieves mean f01 error 1929 MHz versus 1935 MHz for exact Hamiltonian diagonalization, demonstrating that RCSJ waveform encoding conveys device-level physics that the static ideal Hamiltonian cannot represent.
This paper presents an electromagnetic-to-Hamiltonian workflow for modeling a two-qubit coupled transmon superconducting device circuit. The device layout, comprising two coupled transmon qubits with individual readout resonators, is designed and parameterized in Qiskit Metal and analyzed using 3D eigenmode simulations in Ansys HFSS. Adaptive-pass refinement converges the eigenfrequencies and electromagnetic fields, yielding reliable modal data for quantization. Energy-participation ratio values from HFSS are processed through the pyEPR–Qiskit Metal flow to extract effective capacitances, anharmonicities, and qubit–resonator couplings. These parameters define a multimode Duffing–Jaynes–Cummings Hamiltonian that captures the transmon nonlinearities and their interactions with the resonator modes. Frequency-convergence checks and coupling diagnostics verify a consistent, physically accurate mapping from the layout geometry to the device circuit-level model. The Hamiltonian workflow offers a reproducible and scalable framework for modeling and analyzing superconducting transmon architectures in quantum chip design.
This study presents the design, simulation, and experimental characterization of a superconducting transmon qubit circuit prototype for potential applications in dark matter detection experiments. We describe a planar circuit design featuring two non-interacting transmon qubits, one with fixed frequency and the other flux tunable. Finite-element simulations were employed to extract key Hamiltonian parameters and optimize component geometries. The qubit was fabricated and then characterized at 20 mK, allowing for a comparison between simulated and measured qubit parameters. Good agreement was found for transition frequencies and anharmonicities (within 1% and 10% respectively) while coupling strengths exhibited larger discrepancies (30%). We discuss potential causes for measured coherence times falling below expectations (T1 ∼1-2 μs) and propose strategies for future design improvements. Notably, we demonstrate the application of a hybrid 3D-2D simulation approach for energy participation ratio evaluation, yielding a more accurate estimation of dielectric losses. This work represents an important first step in developing planar Quantum Non-Demolition (QND) single-photon counters for dark matter searches, particularly for axion and dark photon detection schemes.
Translating superconducting quantum chip layouts into physically meaningful simulation records is a critical step in quantum hardware development, yet current datasets rarely capture the electromagnetic (EM) effects involved in this process. We propose SQChip-EM, an open layout to EM simulation dataset based on Qiskit Metal that links parameterized layout designs to Ansys HFSS and Q3D simulation results. Unlike abstract circuit datasets, SQChip-EM preserves layout dependent EM evidence through GDSII files, raw simulation outputs, and JSON annotations covering frequencies, anharmonicity, dispersive shifts, and coupling parameters. These annotations form reusable simulation level ''virtual chip'' records for downstream workflows such as initialization, calibration tuning, and control simulation. SQChip-EM also includes three illustrative benchmark tasks, including design to metric regression, specification feasibility classification, and inverse design retrieval, together with one downstream workflow use case. By releasing data, code, and evaluation protocols, SQChip-EM provides a standardized data resource for ML assisted quantum electronic design automation.
Scaling of superconducting-based quantum computers has been possible since the invention of transmon qubits due to their low charge noise sensitivity. However, pursuing fault-tolerant and computationally powerful quantum processors may require incorporating more qubits, which may amplify the complexity of the large design space exploration and simulation. Inspired by the recent CMOS Machine Learning assisted electronic design automation to solve similar issues, we claim this work is the first to demonstrate applying machine learning in superconducting circuit design. To start, we attempt to predict the characteristics of individual transmon qubits with a machine learning-based approach based on the simulation data collected with relevant software frameworks like Qiskit Metal and ANSYS Electronics. Our approach aims to perform forward modeling to predict the qubit engineering parameters such as frequency. On the other hand, the machine learning-based model can also infer, through inverse modeling, the candidate sets of design space parameters like the device dimensions satisfying a particular specification while obeying the geometrical constraint of the transmon types. This work paves an alternative way for future applications in designing and automating superconducting quantum computing circuits.