Multi-Stage Mamba-Based Architecture for Fast and Scalable Superconducting Qubit Readout

Reliable qubit readout is a critical bottleneck in superconducting quantum computing. Current readout processes are high-latency and error-prone, particularly in frequency-multiplexed architectures where signal crosstalk between neighboring qubits degrades fidelity. Existing machine learning approaches, primarily based on Feedforward Neural Networks (FNNs), struggle with scalability due to their large parameter counts and inability to efficiently handle the temporal dynamics of qubit relaxation errors. This paper addresses these challenges by proposing a lightweight, Mamba-based architecture designed for fast and accurate single-shot state discrimination.

The researchers utilize the Mamba model, a Selective State-Space Model (SSM) known for its linear computational complexity relative to sequence length. The proposed system employs a two-stage pipeline: a primary classifier that performs initial state discrimination, followed by a specialized error detector that identifies and corrects relaxation-induced errors. By integrating these components into an end-to-end neural network, the model avoids the overhead of traditional matched filters while preserving full trace information. The architecture uses a linear expansion layer to process I/Q traces, followed by Mamba blocks that capture temporal dependencies, and a final classification head to output the qubit state.

The proposed model demonstrates significant improvements over state-of-the-art methods. The lightweight version reduces parameter size by 49.6% compared to previous benchmarks, while the optimal configuration achieves a geometric mean readout fidelity of 0.911. Crucially, the model remains robust at short readout durations (500 ns), maintaining a fidelity of 0.893. This efficiency is particularly valuable for Mid-Circuit Measurement (MCM) and Quantum Error Correction (QEC), where the authors report a 26% reduction in logical error rates compared to prior work, facilitating more reliable fault-tolerant quantum computing.