Reinforcement Learning Unlocks Continuous Quantum Computing by Self-Correction

Quantum error correction (QEC) stands as the cornerstone strategy for safeguarding quantum computers from the inherent noise and instability of their surrounding environments. However, the efficacy of QEC hinges on errors remaining sufficiently rare, a condition that necessitates constant adaptation of the computer's control parameters to fluctuating environmental conditions. The prevailing approach to address this challenge involves halting the entire quantum computation for recalibration, a method that is fundamentally incompatible with the extended runtimes anticipated for future, more complex quantum algorithms. This bottleneck has long posed a significant hurdle to the practical realization of robust quantum computing. In a groundbreaking advancement, researchers have unveiled a novel framework that directly tackles this critical limitation by seamlessly integrating calibration with computation. This innovative approach redefines the role of the quantum error correction process, granting it a dual function. Beyond its traditional task of detecting and correcting errors in the logical quantum state, its error-detection events are now ingeniously repurposed as a dynamic learning signal. This learning signal is fed into a reinforcement learning (RL) agent, which is then trained to continuously adjust and steer the quantum computer's control parameters. This real-time, adaptive control mechanism ensures the stabilization of the quantum system throughout the computation, eliminating the need for disruptive pauses. Essentially, the quantum computer learns from its own errors as they occur, allowing it to maintain optimal performance without interruption. The efficacy of this pioneering framework was experimentally demonstrated on a Willow superconducting processor. The results were remarkably promising, showcasing a significant 3.5-fold improvement in the logical stability of the surface code when subjected to injected drift. Furthermore, by synthesizing a full suite of technological advancements, the team achieved record-breaking performance for both surface and color codes, reporting average logical error rates per cycle of 7.72(9) × 10−4 and 8.19(14) × 10−3, respectively. These figures represent a substantial leap forward in the reliability of quantum operations. Beyond the experimental success, numerical simulations conducted on large codes, involving tens of thousands of control parameters, confirmed the robust scalability of this reinforcement learning framework. Crucially, these simulations revealed an optimization speed that remains independent of the system size, a vital characteristic for future large-scale quantum processors. This work thus heralds a new paradigm in quantum computing: a self-correcting quantum computer that learns autonomously from its errors and operates without interruption, paving the way for truly continuous and reliable quantum computation.