Designing Quantum Control Sequences Using Reinforcement Learning

Quantum optimal control theory aims to manipulate quantum state evolution such that some desired quantity is optimized, such as quantum Fisher information or fidelity. This has applications in quantum metrology and quantum computing, allowing experiments to realize desired state evolution with more accuracy. One technique, borrowed from the field of nuclear magnetic resonance, seeks to manipulate state evolution using a sequence of effectively instantaneous electromagnetic pulses. However, realistic noise signals make it difficult and sometimes impossible to find analytic solutions for pulse timings. We propose a method for designing pulse sequences on qubits using reinforcement learning (RL). Specifically, we analyze two scenarios. In the first, we consider a qubit in a time-dependent, sinusoidal magnetic field of unknown amplitude and frequency, and we ask the RL agent to find the pulse sequence which, over a band of frequencies, maximizes the quantum Fisher information with respect to the amplitude. This has potential applications for axion detection schemes relying on spin precession. In the second scenario, we consider a qubit whose coupling with the environment causes decoherence, and we ask for the pulse sequence which minimizes the decoherence.