Google researchers have applied reinforcement learning to continuously recalibrate a quantum processor, enabling real‑time error correction. Using the Sycamore superconducting processor, the team trained a reinforcement‑learning system to adjust thousands of microwave control parameters while error‑correction routines ran on two logical qubits. The approach increased the ability to detect and correct errors by about 20 percent compared with fixed calibration settings.

Calibration drift, caused by heating and other factors, has traditionally required interrupting calculations to re‑measure and re‑apply optimal microwave pulse settings. Performing full recalibration mid‑algorithm is impractical, so Google integrated the learning algorithm with the error‑detection data from surface and color code error correction. Small, simultaneous perturbations to the control parameters allowed the system to explore the control space and infer adjustments that minimized errors.

Demonstrations with two logical qubits showed a 20 percent improvement in error detection and correction when reinforcement learning was active. The method works only if drift remains within the range experienced during training; frequent re‑evaluation is required, but randomizing all possible configurations during a calculation would degrade performance. Researchers note that the trade‑off between exploration and exploitation can sustain performance as long as drift is slow enough for short, simple algorithms.

While the current implementation is limited to short tasks, the work shows that a solvable calibration problem can be addressed in real time, offering a step toward larger, more stable quantum computers.