We develop a fidelity-informed neural pulse-compilation framework for a continuous family of single-qubit gates on a three-qubit liquid-state nuclear magnetic resonance (NMR) processor. Instead of decomposing each target unitary into a sequence of calibrated basis gates, the method learns a direct map from the axis-angle parameters of an arbitrary U_2 in SU(2) operation to a piecewise-constant radio-frequency control sequence that implements the desired transformation. Training is performed end-to-end through the time-ordered propagator of the driven Hamiltonian using global-phase-insensitive unitary fidelity as the learning signal. We show numerically that a single model generalizes across a continuous range of gate parameters and experimentally validate representative compiled pulses on a benchtop three-qubit NMR device. In addition, we analyze sensitivity to structured perturbations in Hamiltonian and control parameters by introducing a prescribed uncertainty set and performing a comparative risk-aware redesign based on right-tail Conditional Value-at-Risk (RU-CVaR). This stage produces pulse solutions with broader tolerance margins within the chosen uncertainty model. The results demonstrate continuous pulse-level gate synthesis in an experimentally accessible setting and illustrate a hardware-aware compilation strategy that can be extended to other quantum platforms. While the uncertainty model considered here is tailored to NMR, the neural compilation and risk-aware optimization framework are general and may be useful in architectures where calibration overhead, parameter drift, or control constraints make repeated per-gate optimization costly.
We develop a fidelity-informed neural pulse-compilation framework for a continuous family of single-qubit gates on a three-qubit liquid-state nuclear magnetic resonance (NMR) processor. Instead of decomposing each target unitary into a sequence of calibrated basis gates, the method learns a direct map from the axis-angle parameters of an arbitrary U_2 in SU(2) operation to a piecewise-constant radio-frequency control sequence that implements the desired transformation. Training is performed end-to-end through the time-ordered propagator of the driven Hamiltonian using global-phase-insensitive unitary fidelity as the learning signal. We show numerically that a single model generalizes across a continuous range of gate parameters and experimentally validate representative compiled pulses on a benchtop three-qubit NMR device. In addition, we analyze sensitivity to structured perturbations in Hamiltonian and control parameters by introducing a prescribed uncertainty set and performing a comparative risk-aware redesign based on right-tail Conditional Value-at-Risk (RU-CVaR). This stage produces pulse solutions with broader tolerance margins within the chosen uncertainty model. The results demonstrate continuous pulse-level gate synthesis in an experimentally accessible setting and illustrate a hardware-aware compilation strategy that can be extended to other quantum platforms. While the uncertainty model considered here is tailored to NMR, the neural compilation and risk-aware optimization framework are general and may be useful in architectures where calibration overhead, parameter drift, or control constraints make repeated per-gate optimization costly.
Real-time adaptive control of quantum systems requires rapid generation of robust, high-fidelity pulses across a continuous range of operating conditions. Standard optimization algorithms such as gradient-ascent pulse engineering (GRAPE) solve each instance independently, discarding information between runs and requiring costly reinitialization when parameters change. We present an approach to robust optimal quantum control based on model-based reinforcement learning, in which a single neural network -- embedding the Hamiltonian directly into the training pipeline -- generates robust gates across an entire family of gate configurations, without pre-computed training data. Demonstrated on a single-spin (two-level) system, the trained networks produce pulses for arbitrary rotation angles over a range of pulse durations, detunings, and field inhomogeneities in milliseconds, at fidelities comparable to multi-seed GRAPE. The framework is inherently adaptable: any parameter entering the Hamiltonian can serve as a network input, extending the approach to different systems and control settings. Beyond speed, the network reveals structure in the control landscape: it discovers the same structured phase profiles that appear in GRAPE solutions -- made identifiable through fidelity-invariant symmetry transformations -- but more consistently than independent optimization. This consistency enables smooth interpolation across the entire trained parameter space.
Control synthesis for continuously-parameterized families of quantum gates can enable critical advantages for mid-sized quantum computing applications in advance of fault-tolerance. We combine quantum optimal control with physics-informed machine learning to efficiently synthesize control surfaces that interpolate among continuously-parameterized gate families. Using optimal control as an active learning strategy to guide pretraining, we bootstrap a physics-informed neural network to achieve rapid convergence to nonlinear control surfaces sufficient for our desired gates. We find our approach is critical for enabling an expressiveness beyond linear interpolation, which is important in cases of hard quantum control. We show in simulation that by adapting our pretraining to use a few reference pulse calibrations, we can apply transfer learning to quickly calibrate our learned control surfaces when devices fluctuate over time. We demonstrate synthesis for one and two qubit gates with one or two parameters, focusing on gate families for variational quantum algorithm (VQA) ansatz. By avoiding the inefficient decomposition of VQA ansatz into basis gate sets, continuous gate families are a potential method to improve the noise robustness of VQAs in the near term. Our framework shows how accessible optimal control tools can be combined with simple machine learning to enable practitioners to achieve 3x speedups for their algorithms by going beyond the standard gate sets.
The potential of quantum computing is constrained by the susceptibility of qubits to noise and crosstalk during multi-qubit gate operations. Existing strategies, such as hardware isolation and dynamical decoupling, face limitations in scalability, feasibility, and robustness against complex noise sources. Here we show a physics-guided neural control framework that generates robust control pulses for superconducting transmon qubit systems, specifically targeting crosstalk mitigation. By combining a hardware-aware parameterization with a Hamiltonian-informed objective that accounts for condition-dependent crosstalk distortions, the framework steers the search toward smooth and physically realizable pulses while exploring high-dimensional control landscapes. Numerical simulations for the controlled-Z gate demonstrate superior fidelity and pulse smoothness compared to the Krotov and GRAPE baselines under matched constraints. The results show consistent and practically meaningful improvements in both nominal and perturbed conditions, with pronounced gains in worst-case fidelity, supporting physics-guided neural control as a route to robust control on near-term transmon devices. Quantum computers are highly sensitive to unwanted crosstalk between qubits, especially during multi-qubit gate operations. Here, the authors show that physics-guided neural control can generate smooth, hardware-compatible pulses that make two-qubit gates more robust against crosstalk.
Control synthesis for continuously-parameterized families of quantum gates can enable critical advantages for mid-sized quantum computing applications in advance of fault-tolerance. We combine quantum optimal control with physics-informed machine learning to efficiently synthesize control surfaces that interpolate among continuously-parameterized gate families. Using optimal control as an active learning strategy to guide pretraining, we bootstrap a physics-informed neural network to achieve rapid convergence to nonlinear control surfaces sufficient for our desired gates. We find our approach is critical for enabling an expressiveness beyond linear interpolation, which is important in cases of hard quantum control. We show in simulation that by adapting our pretraining to use a few reference pulse calibrations, we can apply transfer learning to quickly calibrate our learned control surfaces when devices fluctuate over time. We demonstrate synthesis for one and two qubit gates with one or two parameters, focusing on gate families for variational quantum algorithm (VQA) ansatz. By avoiding the inefficient decomposition of VQA ansatz into basis gate sets, continuous gate families are a potential method to improve the noise robustness of VQAs in the near term. Our framework shows how accessible optimal control tools can be combined with simple machine learning to enable practitioners to achieve 3x speedups for their algorithms by going beyond the standard gate sets.
Achieving scalable quantum computing demands high-fidelity operations capable of mitigating population leakage into non-computational states. Physics-Informed Neural Networks (PINNs) have recently emerged as a powerful paradigm to unify quantum hardware characterization (inverse problems) and pulse engineering (direct problems), laying the foundational architecture for autonomous quantum processors. However, standard PINN frameworks face severe numerical bottlenecks, such as spectral bias, when attempting to simultaneously solve highly oscillatory multi-level dynamics and optimize continuous control fields under strict global phase constraints. In this work, we propose an enhanced PINN scheme for quantum optimal control (PINNQOC) that circumvents these limitations by incorporating Fourier feature embeddings, dynamic epoch normalization, and an informed pre-training routine. To rigorously evaluate its performance, we systematically benchmark our framework against two premier continuous control solvers: the first-order Krotov method and the second-order Projection Operator Newton Method for Trajectory Optimization (PRONTO). These techniques are applied to implement multiple quantum gates on a truncated three-level fluxonium qubit and a four-level Nitrogen-Vacancy center coupled to a Carbon-13 nuclear spin. Our advanced PINNQOC approach successfully suppresses population leakage while achieving gate fidelities exceeding 99.9%, matching the efficacy of traditional solvers. Finally, we provide a comprehensive analysis of computational times, iteration efficiency, and mean leakage, highlighting the distinct trade-offs and avenues for embedding physics-guided machine learning into automated quantum hardware pipelines.
Quantum gate design is often represented as pulse optimization, although the physical object that implements a gate is the full controlled evolution generated by the pulse. Here we use physics-informed neural networks to represent single-qubit gate design at this evolution level: the control fields, the Bloch-state trajectories, and the total duration are learned together under the Bloch equation. This changes the optimized object from pulse amplitudes to a differentiable physical process whose structure can be inspected and refined. For rotation gates, the optimized evolutions recover the physical organization expected for bounded single-qubit control, with no prescribed pulse ansatz or duration scan. For a geometric gate, the representation identifies localized bottlenecks in maintaining the geometric condition and turns this diagnosis into feedback, reducing the residual path error while preserving high fidelity. Thus physics-informed learning is used not only to synthesize gates, but also to make optimized quantum controls physically readable, diagnosable, and locally refinable. This process-level view may be especially useful for adapting gates to hardware-specific, task-specific, and locally varying experimental constraints.