A team of Google researchers has demonstrated that a quantum processor can use reinforcement learning to continuously adjust its own control parameters during a calculation, replacing the traditional stop-and-recalibrate workflow with a real-time adaptive system. The findings were published in Nature and represent a significant step toward making large-scale quantum computers practical for real-world problems like drug discovery and cryptography.
The challenge stems from a mundane source: drift. Superconducting qubits are controlled by microwave pulses generated by external hardware that can heat up or fluctuate over time. Currently, when calibration drifts, engineers stop the computation and recalibrate from scratch — an approach that is impossible during a long, complex algorithm. Google's team realized that the error-detection data already being collected during error-corrected computation contains information about calibration drift. By applying reinforcement learning — an AI technique where the system explores small perturbations to its roughly 1,000 control parameters and scores each change by its effect on error rates — the quantum processor can infer which adjustments minimize errors and apply them in real time without stopping.
In the experiment, the system managed two logical qubits using different error-correction schemes (a surface code and a color code) and achieved a 20 percent improvement in error detection and correction when the reinforcement learning loop was active. Importantly, the approach handles a critical trade-off: the system must explore sub-optimal configurations to learn, but the aggregate performance across all tested settings still outperformed running without any adaptive calibration. The researchers also demonstrated that the system can continuously re-evaluate its strategy, meaning it can adapt to drift over extended computations without human intervention.
Knowledge takeaway: Google's quantum processor uses reinforcement learning to self-calibrate during computation by treating error-correction data as feedback, achieving 20% better error detection; this solves the "drift" problem that previously forced computation stoppages for recalibration; the approach works across different error-correction schemes and can adapt continuously, removing a major bottleneck on the path to practical quantum computing.