Tokamak disruption alarms improve by predicting events in fixed time horizons
Horizon-Aware Early Event Prediction for Tokamak Disruption Alarms
Machine Learning
Summary
Safely running tokamaks, devices for controlled nuclear fusion, requires predicting dangerous disruptions ahead of time. The authors studied different ways to predict disruptions within specific future time windows, rather than predicting a full timeline to disruption. They tested three methods on data from three different tokamaks and found that directly predicting the chance of disruption within a fixed future period performed best on two devices. The results suggest that simpler, horizon-focused predictions may be more effective for alarm systems than more complex full-timeline models in this context.
What this means in practice
- •For tokamak operation teams: Improve early warning alarm settings by using horizon-focused disruption prediction models tailored to specific tokamak devices.
- •For fusion control system engineers: Integrate horizon-aware event prediction methods into control systems to better handle disruption risks during tokamak operation.
Authors
Takeshi Koshizuka, Takaharu Yaguchi
Abstract
Reliable disruption prediction is essential for the safe operation of future tokamaks. Existing full-distribution survival methods model the complete residual time-to-disruption distribution, whereas operational decisions primarily depend on disruption risk within a finite prediction horizon. This mismatch motivates introducing Early Event Prediction (EEP) objectives into survival-based disruption prediction. We take Deep Survival Machines (DSM) as the full-distribution baseline and propose applying two established EEP methods to tokamak disruption prediction: Temporal Label Smoothing (TLS), which directly predicts disruption probability within a finite horizon, and survTLS, which additionally models the event-time distribution within that horizon. Using a common causal encoder, we compare these methods on DIII-D, Alcator C-Mod, and EAST. We distinguish threshold-free deadline ranking from validation-selected fixed-policy alarm performance and evaluate prediction horizons and encoder architectures. TLS achieves the best mean alarm performance on DIII-D and EAST, whereas all methods perform poorly on Alcator C-Mod. survTLS does not consistently outperform DSM, suggesting that directly learning horizon-level event probability is more effective than modeling detailed within-horizon event-time distributions in the present setting. Finally, the selected prediction horizons and encoder-ablation results vary across devices, reflecting differences in disruption characteristics.