Google uses error correction data to auto-calibrate quantum processors mid-calculation

Quantum computers face a stubborn practical problem: even after qubits are carefully calibrated, the control hardware—classical computers and microwave sources kept outside the refrigeration—drifts from its original settings due to heat and other random factors. Superconducting qubits, which Google and other companies rely on, are especially sensitive to these shifts in microwave pulse wavelengths and amplitudes. Today's systems are short-lived enough that drift is not yet a critical issue, but the algorithms researchers ultimately want to run will need hours or longer of continuous, stable operation.

Google's insight is elegant: error correction already continuously monitors the quantum hardware by measuring a subset of qubits to detect and characterize errors. Those same measurements reveal the fingerprints of calibration failures—the researchers note that "errors from imperfect calibrations produce detectable syndromes just like all other errors." Rather than trying to manually separate the two types of error, the team used reinforcement learning to probe the control space during the calculation itself. The system applies small, intentional perturbations to roughly 1,000 control parameters (or up to 40,000 in the larger test) and observes how each adjustment changes the pattern of detected errors. If calibration drift begins to cause problems, the system can then apply the corrections it has learned are effective—all without halting the computation.

The trade-off is subtle: because the system must keep exploring different parameter settings to stay adaptive, it deliberately operates slightly away from optimal for brief moments. The key finding is that this short-term sub-optimality is outweighed by the prevention of larger errors caused by unaddressed drift. Simulations confirmed the trade-off worked, provided drift remained slow. This is not a fix for current quantum computers, which are too error-prone and short-lived for the technique to matter. But it demonstrates that a problem Google knows will become critical can be managed—an important step toward the reliable, long-running quantum systems the field needs.