Explainable deep learning for the analysis of MHD activity in locked-mode disruptive pulses

The locked mode amplitude is one of the most commonly used signals for disruption prediction in tokamaks. On a dataset from the JET baseline scenario, our results suggest that the simple application of a threshold on that signal yields a disruption predictor with more than 95% accuracy. It is well-known that mode locking is one of the main disruption causes at JET; however, it is often too late to avoid a disruption by the time it is detected. In this work, we investigate the possibility of predicting the locked mode itself and, in particular, whether it is possible to apply machine learning to identify the MHD behavior that typically precedes mode locking. For this purpose, we trained a deep learning model over MHD spectrograms. The model is a Convolutional Neural Network (CNN) that receives a time window from the spectrogram and predicts whether the locked mode amplitude will exceed a given threshold. In addition, we use Class Activation Mapping (CAM) to explain why the model arrives at a certain prediction. The results suggest that the interruption of MHD activity followed by the resurgence of a mode at the q=2 surface are strong indicators that mode locking is about to occur, which is consistent with the literature. This work also suggests that neural networks can be useful, as interpretable machine learning models, to support the analysis of MHD activity.