Fidelity-informed neural pulse compilation of a continuous family of quantum gates with uncertainty-margin analysis

1. Problem Statement

Quantum algorithms are typically realized on physical hardware by decomposing target unitaries into sequences of discrete, calibrated basis gates. This approach faces significant challenges:

• Inefficiency for Continuous Gates: Variational quantum algorithms and analog-inspired ansätze often require continuously parameterized unitaries. Decomposing these into discrete gates increases circuit depth and compilation overhead.
• Calibration Burden: Optimizing a unique control pulse for every specific gate parameter is computationally expensive and prone to calibration drift.
• Robustness: Standard optimal control pulses are often "fragile," meaning small perturbations in hardware parameters (e.g., magnetic field drift, amplitude errors) can cause significant fidelity loss.

The paper addresses the need for a continuous pulse-level compiler that maps arbitrary single-qubit gate parameters directly to control waveforms, while also ensuring these pulses are robust against structured hardware uncertainties.

2. Methodology

A. Physical System and Control Task

• Platform: A three-qubit liquid-state Nuclear Magnetic Resonance (NMR) processor (SpinQ Triangulum Mini).
• Hamiltonian: The system is governed by a drift Hamiltonian ($H_0$) involving Zeeman splittings (chemical shifts) and scalar (Ising-type) spin-spin couplings, controlled by a single global radio-frequency (RF) field ($H_c(t)$).
• Target: Implement an arbitrary single-qubit unitary $U_2 \in SU(2)$ on the first qubit while leaving the other two qubits unchanged. The target is parameterized by axis-angle coordinates $(\gamma, \theta, \phi)$.
• Metric: Global-phase-insensitive unitary fidelity is used as the performance metric.

B. Fidelity-Informed Neural Compiler

Instead of solving an optimal control problem for each gate, the authors train a neural network to act as a continuous compiler:

• Input: The gate parameters $(\gamma, \theta, \phi)$, encoded via trigonometric features ($\cos, \sin$).
• Output: A piecewise-constant RF control sequence (phase modulation) applied over $T$ time slices.
• Training Mechanism:
  • End-to-End Differentiation: The network is trained through the physical time-ordered propagator of the driven Hamiltonian.
  • Loss Function: $L = 1 - F(U_{target}, U_{simulated})$.
  • Gradient Propagation: Gradients are computed using the Fréchet derivative of the matrix exponential and backpropagated through the time-ordered product (similar to GRAPE algorithms), allowing the network to learn the physics directly without labeled pulse data.

C. Uncertainty-Aware Redesign (Risk-Aware Optimization)

To address hardware imperfections, the authors introduce a Risk-Aware training stage:

• Prescribed Uncertainty Set: They define a set of structured perturbations including chemical shift noise, scalar coupling variations, RF amplitude/phase bias, and timing jitter.
• Risk Metric (RU-CVaR): Instead of minimizing average error, they optimize for Right-tail Conditional Value-at-Risk (RU-CVaR). This metric penalizes the "worst-case" scenarios within the uncertainty set (the tail of the loss distribution).
• Objective: The goal is to find control solutions that lie in "flatter" regions of the control landscape, ensuring high fidelity even when parameters deviate from the nominal model.

3. Key Contributions

1. Continuous Neural Compiler: Development of a single neural model capable of mapping a continuous range of $SU(2)$ gate parameters directly to executable RF pulses, eliminating the need for discrete decomposition or per-gate optimization.
2. End-to-End Physics Learning: Demonstration of training a neural network directly through the quantum mechanical propagator, ensuring the learned pulses are consistent with the specific device Hamiltonian.
3. Risk-Aware Robustness Framework: Introduction of a RU-CVaR-based optimization strategy that explicitly trades off peak nominal performance for broader tolerance margins against structured uncertainties.
4. Experimental Validation: Successful implementation of the compiled pulses on a real benchtop 3-qubit NMR device, validating the physical realizability of the neural synthesis.

4. Results

Numerical Results

• Generalization: The trained model achieved mean and median fidelities >99% across a dense grid of unseen gate parameters in the nominal setting, demonstrating smooth generalization rather than memorization.
• Sensitivity Analysis: Nominal pulses were found to be highly sensitive to detuning (frequency shifts), global amplitude scaling, and timing distortions.
• Robustness Improvement: Under RU-CVaR optimization (specifically with risk level $\alpha=0.2$):
  • The controller sacrificed a negligible amount of peak nominal fidelity.
  • It significantly broadened the tolerance margins for the most detrimental error channels (detuning, amplitude, timing).
  • The fidelity curves became "flatter," indicating reduced fragility.

Experimental Results

• Hardware Implementation: The authors executed compiled pulses on the SpinQ Triangulum Mini (based on a $C_2F_3I$ molecule).
• Validation:
  • State Tomography: Reconstructed density matrices showed a fidelity of 92% for a representative gate, confirming the pulses are physically viable despite experimental noise (state preparation, readout errors, coherence limits).
  • Spectral Analysis: By varying the azimuthal angle parameter, the team verified that the phase of the output NMR spectrum varied linearly as predicted, confirming the correct implementation of the continuous gate family.

5. Significance and Outlook

• Hardware-Aware Compilation: The work bridges the gap between abstract quantum algorithms and physical control, offering a strategy to reduce calibration overhead in variational algorithms.
• Scalability: While demonstrated on NMR, the framework is platform-agnostic. It is particularly relevant for architectures like superconducting qubits or Rydberg arrays, where parameter drift and control constraints make repeated per-gate optimization costly.
• Robustness Paradigm: The use of RU-CVaR provides a practical, computationally efficient method to design "robust-by-design" pulses without requiring full knowledge of the experimental noise distribution, only a structured uncertainty model.
• Future Work: The authors suggest extending this approach to multi-qubit targets, larger gate families, and other quantum platforms to further validate the transferability of the risk-aware compilation strategy.

In summary, this paper presents a proof-of-principle for a neural network that acts as a universal, continuous pulse compiler for quantum gates, enhanced with a risk-aware layer to ensure robustness against realistic hardware imperfections, all validated on a physical quantum processor.