Machine Learning Assisted Superconducting Qubit Readout – Part I: From Pulse Shaping to State Discrimination*

Qubit-state readout is a significant error source in contemporary superconducting quantum processors. The fidelity of dispersive qubit-state readout depends on the readout pulse shape and resulting phase-shifted readout signal discriminator. For a single qubit, fast and high-fidelity readout is achieved with minor changes to the rising and falling edge of a rectangular pulse, and a linear matched filter discriminator. However, in resource-efficient, frequency-multiplexed readout of multiple qubits, optimizing the readout pulse shape and discriminator becomes a computationally intensive task. Here, we experimentally demonstrate deep machine learning techniques to improve superconducting qubit readout pulse shapes and discrimination compared to conventional methods.