Post-alarm time-to-disruption classification in disruptive JET discharges: comparison of SVM, LSTM, and transformer models

This work presents a systematic comparison of classical machine-learning and deep-learning models for post-alarm time-to-disruption classification in disruptive JET discharges. Using alarms generated by the validated linear disruption predictor of Rattá et al (JET Contributors) (2019 Fusion Eng. Des.146 2393–6) at a fixed working point, we train three models: support vector machines, long short-term memory networks, and a Transformer encoder. They are used to assign each alarmed disruptive pulse to one of three action-oriented remaining-time classes: mitigation (), avoidance (–), and prevention (). Within the present dataset and evaluation setup, the Transformer slightly outperforms the alternatives, and the time derivative of the locked-mode amplitude emerges as an informative feature, particularly for the prevention class. The estimator is modular in formulation, but the present validation is restricted to disruptive discharges, one upstream predictor, and one working point, without an end-to-end assessment including false alarms and missed alarms.