Empirical learning of dynamical decoupling on quantum processors

Abstract

Dynamical decoupling (DD) is a low-overhead method for quantum error suppression. Despite extensive work in DD design, finding pulse sequences that optimally decouple computational qubits on noisy quantum hardware is not well understood. In this work, we describe how learning algorithms can empirically tailor DD strategies for any quantum circuit and device. We use a genetic algorithm-inspired search to optimize DD (GADD) strategies for IBM's superconducting-qubit based quantum processors. In all observed experimental settings, we find that empirically learned DD strategies significantly improve error suppression relative to canonical sequences, with relative improvement increasing with problem size and circuit sophistication. We leverage this to study mirror randomized benchmarking on 100 qubits, GHZ state preparation on 50 qubits, and the Bernstein-Vazirani algorithm on 27 qubits. We further demonstrate that our empirical learning method finds strategies, in time constant with increasing circuit width and depth, that provide stable performance over long periods of time without retraining and generalize to larger circuits when trained on small sub-circuit structures.