Deep Neural Networks for High-fidelity Measurement of Multiqubit Circuits.

Superconducting qubits are a leading candidate for fault-tolerant quantum computing; however, it is challenging to maintain high measurement fidelity as systems are scaled to large size. In this work, we perform numerical simulations to benchmark various Deep Neural Network (DNN) architectures on the task of multiplexed dispersive measurement of superconducting qubits. We compare the robustness of the different state assignment approaches against three sources of measurement infidelity: added measurement noise, qubit relaxation during measurement, and state initialization errors. We find that transformer and convolutional neural network architectures increase readout fidelity relative to conventional thresholding and that these approaches are robust against labeling error in the training datasets. In addition, we calculate the theoretical limit for readout fidelity and demonstrate that the transformer approach provides assignment fidelity approaching the theoretical limit.