Active Learning for Calibrating Entangling Gates via Surrogate-Based Optimization

Technical Summary: Active Learning for Calibrating Entangling Gates via Surrogate-Based Optimization

Problem Statement
High-fidelity quantum gate operations are essential for near-term quantum advantage and fault-tolerant error correction. However, achieving high fidelity requires precise calibration of classical control parameters defining the system's Hamiltonian. While theoretical models exist, the actual Hamiltonian implemented on hardware often deviates from the ideal due to unknown, potentially nonlinear transfer functions in the classical control infrastructure. Traditional calibration methods, such as sequential parameter sweeps or optimization relying on precise forward models (e.g., Bayesian calibration with accurate theoretical models), face significant limitations. These include poor scalability with system size, high computational costs for classical simulations, and the inability to handle hardware drift without periodic recalibration. Furthermore, quantum systems operate in a noisy, low-data regime, rendering data-hungry approaches like neural networks impractical. The core challenge is to identify optimal control parameters efficiently, tolerating measurement noise and operating without an exact physical model of the experimental system.

Methodology
The authors propose an active learning framework based on Bayesian Optimization (BO) utilizing a Gaussian Process (GP) as a surrogate model. This approach is designed to navigate the calibration landscape using a limited number of noisy measurements.

• Surrogate Modeling: The framework employs a GP to model the unknown calibration landscape $f(x)$, where $x$ represents the vector of control parameters (Rabi frequency $\Omega$, sideband detuning $\delta$, and center-line detuning $\omega_{cl}$). Unlike standard GPs, this implementation uses a heteroskedastic noise model. The measurement noise $\epsilon_i$ is modeled as a zero-mean Gaussian with variance dependent on the control settings and the number of measurement shots $N$, reflecting the physics of quantum projection noise.
• Bayesian Optimization Loop: The optimization process iterates through three steps:
  1. Surrogate Update: The GP posterior is updated with new experimental data (control parameters and measured scores), optimizing hyperparameters (length scales, signal variance, noise scaling) via type-II maximum likelihood.
  2. Active Learning Acquisition: An acquisition function, specifically the Upper Confidence Bound (UCB), selects the next experimental configuration. This function balances exploitation (sampling regions with high predicted fidelity) and exploration (sampling regions with high uncertainty) to efficiently locate the global optimum.
  3. Physical Evaluation: The selected parameters are applied to the quantum system (simulated in this work), and the resulting gate fidelity score is measured.
• Simulation Environment: The method is validated using a simulated trapped-ion Mølmer-Sørensen (MS) gate. The simulation uses IonSim.jl to model two $^{40}\text{Ca}^+$ ions, incorporating off-resonant carrier excitation and square pulse effects while neglecting decoherence to isolate the impact of quantum projection noise. The calibration sequence involves initializing qubits in $|gg\rangle$, applying two MS gates, and measuring the probability of the $|ee\rangle$ state ($P_{ee}$) as the score.

Key Contributions

• Model-Free Calibration: The paper demonstrates a calibration protocol that does not require a validated forward model of the quantum system, relying instead on a data-driven GP surrogate.
• Heteroskedastic Noise Handling: The authors integrate a noise model where observation variance depends on the control parameters and measurement shots, acknowledging that quantum projection noise varies across the landscape.
• Active Learning Efficiency: The framework utilizes active learning to accelerate the discovery of optimal parameters, showing that the GP can effectively "denoise" the system by enforcing spatial smoothness, even when trained on noisy data.
• Quantification of Limits: The study explicitly characterizes the fidelity saturation limits imposed by finite-shot quantum projection noise, distinguishing between algorithmic convergence and physical measurement limits.

Results
The authors validated the approach through numerical simulations of the trapped-ion MS gate:

• Surrogate Accuracy: The GP surrogate successfully modeled the control landscape. Increasing the number of training points ($m$) and measurement shots ($N$) improved accuracy. Notably, for large $m$, the GP mean could filter out uncorrelated projection noise, achieving Root Mean Square Errors (RMSE) below the mean noise level of the data.
• Optimization Convergence:
  • Parameter Range: Reducing the initial uncertainty range of parameters (e.g., from $\pm 10$ kHz to $\pm 1$ kHz) significantly decreased the time to convergence and reduced variance in the optimization path. However, the optimizer remained robust enough to find the global optimum even in full-scale search domains ($r=1.0$).
  • Shot Budget: The optimization fidelity saturated based on the number of shots per evaluation ($N$). The saturation followed the expected scaling of $1 - Q \approx Q_\infty + \gamma N^{-1/2}$. Even with a low budget of $N=100$ shots, the optimizer reliably reached a fidelity floor of $\sim 10^{-3}$, though further iterations could not improve fidelity below the finite-shot limit.
  • Parameter Precision: The frequency parameters ($\omega_{cl}$ and $\delta$) were converged to within tens of Hz of the ideal values, while the Rabi frequency ($\Omega$) converged to the ideal ratio with minimal variance.
• Visualization: The active learning process was shown to intelligently sample the landscape, concentrating measurements around the optimal region and narrowing uncertainty around the frequency peaks, while the amplitude landscape remained broader due to lower sensitivity in the chosen calibration sequence.

Significance and Claims
The paper establishes the utility of active learning and surrogate models for quantum calibration and control. The authors claim that their model-free approach provides a "lightweight route to automated gate maintenance," capable of handling calibration drift and operating without detailed physics models. The results suggest that Gaussian Process Bayesian Optimization is a practical tool for noisy, expensive quantum-control tasks. The work highlights that while resource-efficient calibration is possible, the ultimate fidelity is bounded by the quantum projection noise inherent in the measurement process. The authors position this method as a foundational step toward characterizing, calibrating, and controlling scaling quantum systems, noting that future work should address real-world nonidealities and multi-fidelity approaches combining classical simulations with quantum measurements.