Physics-Informed Metalearning for Rapid Quantum Device Characterization

Scaling modular quantum processors requires efficient calibration methods that can rapidly characterize new hardware modules. Current black-box approaches require extensive experimental data for each new device, creating prohibitive overhead for systems with many modes. We present progress on a physics-informed metalearning framework using Neural Ordinary Differential Equations (Neural ODEs) that shifts this data burden from experiment to simulation. By meta-training on simulated trajectories spanning realistic device parameter variations, the model learns shared dynamical structure across device contexts. The neural network is architecturally constrained to output Hamiltonian coefficients for physical operators, ensuring quantum-mechanically valid predictions while maintaining interpretability. We demonstrate Hamiltonian recovery on simulated multi-qubit systems with tunable couplers, showing that the framework captures the relationship between control inputs and quantum dynamics. This approach establishes groundwork for rapid few-shot adaptation to new devices using only sparse experimental measurements.