Deep Learning Assisted Readout of In Situ Frequency Tunable Superconducting Qubits

Dispersive readout of superconducting qubits is one of the most critical components of superconducting quantum computing systems. Traditional dispersive readout schemes typically employ digital signal processing technique to demodulate multiple measurements into clouds in the IQ plane, determining the qubit state through projection and threshold analysis. However, when addressing tunable qubits, this method is limited to single-frequency measurements; attempting to measure at different frequencies without adjusting the readout frequency accordingly can compromise readout fidelity. To address this, this paper proposes a deep learning-based readout scheme that can directly train and predict the readout signals corresponding to the ground and excited states of the same superconducting qubit at different frequencies under a fixed readout frequency. A supervised learning approach is utilized to construct a deep learning model based on the multilayer perceptron. The trained model can directly predict the qubit state at different frequencies within a large frequency range. Moreover, the deep learning model can not only distinguish between the ground and excited states of the qubit but also leverage its capability to handle regression problems to test any superposition state of the qubit, which has been verified through Rabi oscillation tests. Experimental results indicate that this approach can effectively discern the state within a 300 MHz tuning range of the qubit frequency, achieving a readout fidelity exceeding 92% under various magnetic flux bias and a Rabi measurement fidelity above 96%. The findings of this study underscore the efficacy of the deep learning scheme in augmenting the convenience of qubit readout processes, concurrently preserving high fidelity, which is imperative for the advancement of quantum computing technologies.