Quantum-tailored machine-learning characterization of quantum processors

Standard methods for characterizing quantum devices require resources that grow exponentially with system size. As an alternative to these unscalable techniques, one can use readily available knowledge about the device in order to acquire information about a quantum state more efficiently. Here, we introduce a machine-learning architecture for inferring and interpreting the dynamics of a quantum device from time-series measurement data. Our recurrent architecture leverages quantum-mechanical structure in its design to interpret limited measurement data from quantum devices more accurately. We treat both numerically generated and experimental data for the case of a driven superconducting transmon qubit probed by a weak microwave tone implementing heterodyne measurement. This work shows how a physics-inspired machine-learning model can improve the learning of device-specific quantum dynamics and parameters. These can then be used to characterize and calibrate superconducting quantum processors.