Deep Reinforcement Learning for Robust Dynamical Decoupling

Preservation of quantum coherence is of great interest for quantum information and sensing. Standard examples of dynamical decoupling protocols for qubit systems compose a train of pulsed control operations and periods of free evolution. We pose this problem in a reinforcement learning

paradigm, by allowing a neural network agent to determine the choice of pulses for a spin system with dipolar interactions, magnetic disorder and control errors. Perhaps unsurprisingly, we find that machine-learning based searches fail to consider varying noise environments unless such inho-

mogeneities are explicitly included during training. To address this issue, we employ Monte-Carlo sampling to generate an array of spin systems, each with parameters sampled from user-defined noise distributions. We study the statistics of these measurement observables during unitary evolution, using them to inform our search for robust decoupling sequences.