A neural-ODE based quantum control method minimizes leakage outside the computational basis while preserving high fidelity, and is validated as a practical, hardware-compatible tool for high-performance control in superconducting circuits.

The core idea is to model the control field as a continuous-time function produced by trainable neural networks, eliminating predefined bases or discrete parameterizations and enabling smooth, hardware-friendly waveforms.

Training relies on a differentiable simulation of quantum dynamics where the loss function measures the gap between implemented and target operations, guiding gradients to optimize pulse amplitudes and phases for both qubits and qutrits.

The neural-ODE framework achieves parity measurements with fidelity above 99.9% across a detuning range of roughly ±10 MHz, increasing robustness to noise, detuning, and drive-amplitude fluctuations compared with rectangular or DRAG pulses.

Compared with traditional methods like GRAPE or CRAB, this approach avoids basis expansions and discretization, producing smooth pulses that directly interface with hardware and potentially scale to multiqubit gates.

A team led by Marko Kuzmanić experimentally demonstrated a neural-ODE quantum control method achieving 99.95% accuracy for a π/2 gate in superconducting transmon circuits, outperforming standard pulse designs.

Experimentally, the method shows robustness in parity-measurement sequences, Wigner tomography, and bosonic QED contexts, highlighting its potential for error mitigation, state reconstruction, and advanced sensing applications.