Reinforcement Learning for Hamiltonian Engineering of Dipolar Coupled Spin Systems

In systems of electronic and nuclear spins, spin-spin interactions and onsite disorder can lead to a decay of the spin coherence. However, by applying a sequence of resonant pulses to the system, the effective Hamiltonian for the system can be engineered to suppress these effects and extend the coherence times of the spins. Methods such as Average Hamiltonian Theory and Floquet theory have provided a framework to generate effective pulse sequences, both analytically and using numerical methods. However, the performance of these sequences depends on the relative strengths of the disorder and interaction strength. For example, sequences that work well for nuclear spins where interactions typically dominate do not work as well for electronic spins where disorder often dominates. Additionally, different experimental errors influence sequence performance in different ways. Here we show that the reinforcement learning assisted sequence design can be tuned to the specific degree of disorder and interactions present in the experimental system of interest, while also allowing us to compensate for a broad range of experimental errors.