Superconducting Qubit Readout Using Next-Generation Reservoir Computing

Quantum processors require rapid, high-fidelity measurements of many qubits to reach the readout error thresholds demanded by quantum error correction and achieve a computational advantage over classical computers. While superconducting qubits are among the leading modalities toward a useful quantum processor, their readout chain remains a key challenge for performance and scalability. Traditional approaches to processing measurement data cannot capture crosstalk in frequency-multiplexed readout, which is widely used to reduce the resource overhead per qubit. Recent methods address this challenge using neural networks to improve state-discrimination fidelity, but they are computationally expensive to train and evaluate, leading to increased latency and poor scalability as the number of qubits grows. We present an alternative machine-learning approach based on next-generation reservoir computing that constructs polynomial features from the measurement signals and maps them to the corresponding qubit states. This method supports real-time training and fast evaluation with low computational cost, making it adaptable and scalable to larger devices. We demonstrate the performance of our approach on single- and five-qubit experimental datasets and compare its computational complexity and fidelity with existing methods. Our results highlight that lightweight, hardware-efficient machine-learning models, together with scalable control electronics, will be crucial to realize useful large-scale quantum computers.