Convolutional neural network versus persistence for forecasting syndrome rate threshold crossings

Abstract Calibration drift makes the reliability of quantum processors change over time, motivating tools that forecast degradation from data already produced during error correction. We test whether a convolutional neural network (CNN) can predict, from eight cycles of syndrome statistics, whether the next cycle's nonzero-syndrome rate in a simulated three-qubit bit-flip code will exceed a fixed threshold. Fifteen independent drifting-noise trajectories (6,000 cycles) were split by trajectory into training, validation, and test sets, so no cycle was shared between sets. On three held-out trajectories the CNN reached 89.6% accuracy and 98.0% recall for threshold crossings, while a last-cycle persistence baseline reached 91.2% accuracy. Short-horizon crossings are therefore predictable under slow drift, but the CNN shows no accuracy advantage over persistence. The results show why trajectory-disjoint evaluation and simple temporal baselines are needed when assessing machine-learning monitors for quantum error correction.