Qubit-state discrimination techniques for accurate quantum error correction

Quantum error correction codes rely on the rapid and high-fidelity qubit-state readout. Qubit-state readout is often among the most error-prone operations for superconducting quantum processors. For instance, qubit-state transitions during readout and noise added to the measurement signal can make readout signals cumbersome to classify. To address these challenges, we recently demonstrated a readout scheme composed of two techniques: a shelving technique to mitigate the error from state transitions and a two-tone readout signal to increase the readout-signal distinguishability. Within readout times of 140 ns, we achieve fidelities in excess of 99.5%. Building on this result, we first evaluate to what extent the readout technique is quantum-non-demolition (QND), an essential requirement for quantum error correction protocols. We find that the method is QND even if we excite the qubits to higher states during readout. Next, we investigate how to exploit the uncertainty in qubit-state discrimination. A feedforward neural network (FNN) classifier used to post-process the measurement result readily offers confidence information in the qubit-state assignment. We explore how to improve minimum weight matching decoders by incorporating this information in the weights of graph edges that correspond to stabilizer measurement errors.