Key facts
- Google researchers have developed a method to calibrate quantum processors in real-time.
- The system uses reinforcement learning to adjust control algorithms based on error correction data.
- This approach aims to address the issue of calibration drift during complex quantum computations.
- Tests showed a 20% improvement in error detection and correction capabilities.
- The method was successfully demonstrated in real-time on a large error-corrected qubit.
Google has developed a novel approach to calibrate quantum processors in real-time, addressing a significant challenge in achieving reliable quantum computation. Traditional calibration methods, which involve testing microwave pulse frequencies and amplitudes to minimize error rates, cannot be performed concurrently with complex calculations. This limitation leads to 'drift,' where hardware settings deviate from optimal parameters over time, compromising accuracy for long and intricate algorithms.
The new system leverages reinforcement learning, utilizing the same data employed for quantum error correction to continuously adjust control algorithms. The researchers found that errors detected during the error correction process can also indicate calibration failures. By deliberately introducing small perturbations to control parameters, the reinforcement learning system explores the control space and learns how to adjust these parameters to minimize specific errors. This process can occur in parallel with the existing error detection and correction mechanisms managing logical qubits.
Initial tests on a system with two logical qubits demonstrated a 20 percent increase in the ability to detect and correct errors when the reinforcement learning system was active. A key challenge remains ensuring the system's effectiveness when drift causes significant deviations from its trained state. The researchers showed that by constantly re-evaluating parameter effectiveness, the trade-off between exploration and exploitation can favor frequent, albeit sub-optimal, error correction, which ultimately prevents larger problems caused by drift. This method was successfully demonstrated in real-time on a large error-corrected qubit with approximately 40,000 parameters.