Google Research has developed a reinforcement learning (RL) framework that allows quantum computers to continuously recalibrate themselves, addressing a major bottleneck in their operation. Unlike traditional methods that require halting computations for tuning, this RL agent learns from quantum error detections to dynamically adjust thousands of control parameters, stabilizing the quantum system against drift during computation. This breakthrough, published in Nature, enables a quantum computer that learns from its errors and keeps computing.

Quantum computers are sensitive analog machines prone to drift, necessitating frequent recalibration of control parameters like signal frequencies and amplitudes. Current methods require a full stop of computation, limiting the duration of useful quantum algorithms. The new RL framework leverages data from Quantum Error Correction (QEC) parity checks, which signal that an error has occurred but not its precise location. The RL agent uses these detection events as a learning signal to steer control parameters and prevent new errors.

This RL quantum control was validated on Google's Willow superconducting processor. By injecting artificial drift, the RL steering improved logical stability 3.5-fold. Even after expert calibration, RL fine-tuning further reduced the logical error rate by an additional 20%. Simulations indicate the approach scales to larger quantum computers, with training iterations independent of system size. This work, involving the Google Quantum AI team and Google Deep Mind, ushers in a new paradigm for quantum computing.