Disruption Prediction
Estimating in real time whether a plasma is approaching a sudden loss of confinement, early enough for the stack to act.
What a disruption is
A disruption is a fast, largely irreversible loss of confinement in which the plasma's stored thermal and magnetic energy dumps into the wall and structures over milliseconds. It can drive large mechanical forces, intense localized heating, and runaway electrons. Predicting disruptions with enough lead time to avoid or mitigate them is a central control problem.
The prediction task
A predictor takes live diagnostic signals and outputs a probability, or time-to-disruption, updated every control cycle. Useful prediction means calling the event with enough warning - typically tens of milliseconds - to trigger avoidance or mitigation, while keeping false alarms low enough that the machine is not needlessly terminated.
Signals that carry warning
- Growing magnetic fluctuations (locked and rotating modes)
- Rising radiated-power fraction and impurity accumulation
- Approach to density, current, or beta limits
- Loss of vertical control or shape error growth
- Profile changes such as a peaking or hollowing of the current
Physics and data together
Predictors combine physics-based proximity indicators (distance to known stability limits) with data-driven classifiers trained on labeled discharges. Physics indicators generalize but miss subtle precursors; learned models catch subtle patterns but can fail outside their training regime. Robust systems use both and weight them by confidence.
The honest limits
Disruption data are imbalanced (disruptions are rare), device-specific, and sparse for future regimes that no machine has reached. A predictor trained on one device rarely transfers cleanly to another, and a reactor's regime lies outside all existing data. Good practice treats predictions as decision support with quantified uncertainty and leans on avoidance rather than heroics near the limit.