Quantum error correction can constantly recalibrate a processor
Google researchers have developed a method using reinforcement learning to adjust quantum processor control parameters in real time, improving error correction by 20%.
A new approach to quantum computer calibration
One obstacle to practical quantum computing is hardware calibration, especially for superconducting qubits like the transmons used in Google's Sycamore processor. These qubits have subtle individual variations, and the microwave control hardware can drift from initial settings due to thermal fluctuations. Previously, computations were halted for recalibration when drift was detected, but this is impractical for long algorithms.
Reinforcement learning solution
Google's team discovered that error correction data can also identify calibration issues. Using reinforcement learning, the system explores different configurations of roughly 1,000 control parameters, scoring their effectiveness at minimizing errors. This allows simultaneous calibration and error correction.
Results and future outlook
Tests on two logical qubits using different error correction codes showed a 20% improvement in error detection and correction when the reinforcement learning system was active. Although current hardware runs only short algorithms where drift is not yet a problem, the work proves the concept for future long-duration computations. The research was published in Nature in 2026.
