Superconducting qubit readout pulse optimization using deep reinforcement learning

Abstract

Quantum computers promise to solve specific problems significantly faster than classical computers. Superconducting quantum processors comprising more than 50 qubits can achieve quantum supremacy, the ability to outperform existing classical computers for particular problems. However, to build a useful quantum computer, the quantum processor's constituent components such as their control and readout must be very well-calibrated. Qubit-state readout of contemporary superconducting quantum processors is a significant error source. In an efficient, frequency-multiplexed readout of multiple qubits, effects such as drive cross-talk increase the complexity of optimal readout pulse shapes, requiring computationally intensive methods to discover high-fidelity pulse shapes. In this thesis, we extend existing readout optimization methods to work in multi-qubit environments and present a new pulse shaping optimization module using deep reinforcement learning. Compared to conventional readout methods in a simulated environment, we are able to reduce required readout pulse lengths by over 63% in single-qubit environments and by over 57% in multi-qubit environments. In addition to discussing how the deep reinforcement learning pulse shaping module will be used in experimental contexts, we also evaluate the future generalized use of deep reinforcement learning methods in quantum computing.