Efficient Discrimination of Multiplexed Superconducting-Qubit States using Next-Generation Reservoir Computing

Quantum processors need rapid, high-fidelity qubit-state discrimination to achieve a computational advantage over classical systems. Traditional methods can perform quick state discrimination but often face challenges with crosstalk during multi-qubit readout, especially for frequency-multiplexed superconducting qubits. While recent neural network approaches enhance fidelity by detecting crosstalk, they are typically slow and challenging to scale for larger systems. We introduce a hybrid approach that blends the efficiency of traditional techniques with the accuracy of machine learning using a next-generation reservoir computing model. This approach avoids the computational expense of typical neural network activation functions, is highly parallelizable, and can learn in real time to adapt to changing conditions. We demonstrate its effectiveness in discriminating qubit states in a five-qubit experimental dataset and compare its computational efficiency and fidelity with existing methods.