Deep learning with physics-guided event screening for sawtooth recognition on the EXL-50U spherical torus

Abstract This paper presents a supervised learning framework for automatic sawtooth recognition based on multi- channel AXUV signals from the EXL-50U spherical torus. A neural network first classifies signal segments into normal sawtooth waves, inverted sawtooth waves, and non-sawtooth waves. The segment predictions are subsequently merged into continuous sawtooth-active intervals and screened using a physics-guided strategy that combines model confidence, temporal periodicity, and multi-channel consistency. This design reduces frag- mented false positives and improves the continuity of interval-level outputs in noisy experimental environments. Experimental results demonstrate that the proposed method achieves a segment-level classification accuracy of 98.49% and a macro-F1 score of 98.48% on a 794-segment test set separated by discharge. For 10 consecutive historical discharges containing 21450 manually reviewed segments, the model achieves 97.93% accuracy and an 88.72% macro-F1 score. At the event-score threshold S ≥ 80, the full method matches all seven expert- annotated sawtooth-active intervals, with no false-positive or missed intervals. In addition, the framework can generate standardized shot-wise event summaries and channel–time overview plots, thereby supporting efficient post-discharge analysis and providing a technical foundation for future statistical studies of sawtooth dynamics and their potential interactions with other magnetohydrodynamic instabilities.