Characterization and Calibration of Quantum Processors using Machine Learning

The effort required for quantum system control, particularly in terms of measurements during calibration, needs to be low enough to compensate for system parameter changes or fast enough to allow for periodic re-calibration. Complete system characterization can provide accurate numerical models for gate calibration. On the other hand, model-free learning control, although costly and laborious, represents a data-driven calibration technique. As quantum systems increase in size, these methods become more measurement-intensive. Here, we use a technique to extract the quantum processor dynamics and map them to a known Hamiltonian and a neural network capturing the unknown dynamics. This combined approach, known as residual modeling, allows for a comprehensive representation of the experimentally observed dynamics with minimal effort. This characterization technique enables low-effort calibration of the closed-system model. Subsequently, a reinforcement-learning gate-calibration agent is equipped with the acquired model and initiated using a numerically derived guess. Such a low-resource approach could enable scaling up quantum processors calibrated by reinforcement learning agents and sidestep the looming bottleneck of quantum control.