Risk-aware framework development for disruption prediction: Alcator C-Mod survival analysis

Disruptions pose a significant threat to the operation of future high-performance tokamaks. Over the course of a discharge, the plasma control system must be able to predict disruption onset with enough warning such that mitigation systems can be triggered. Risk-aware frameworks allow better prioritization of actuator response, enabling not only prediction of disruptions but also performance optimization. Determining the least hazardous actions to avoid disruptions is comparable with the time-to-event predictions often used in healthcare for selecting treatments given some mortality risk. This framework is called survival analysis, and there have been many tools developed which we aim to apply to disruption prediction. Using the open-source Auton-Survival package [1] and data from Alcator C-Mod, we have benchmarked performance of the binary classifier model Disruption Prediction with Random Forests [2] against several survival analysis models including Cox Proportional Hazards, Deep Survival Machines, and Kaplan-Meier [3]. We find that while the survival analysis models have similar accuracy with respect to binary classifiers, they allow robust instability predictions at greater time horizons.

[1] C. Nagpal et al., Proceedings of Machine Learning Research 182 (2022)

[2] C. Rea et al., Nuclear Fusion 59, 096016 (2019)

[3] R. A. Tinguely et al., Plasma Physics and Controlled Fusion 61, 095009 (2019)