Disruption Prediction and Avoidance
A disruption is a sudden loss of plasma confinement; predicting the precursors early enough to steer away from them is a core control task.
What a disruption is
In a tokamak a disruption is a fast, uncontrolled collapse of the plasma current and thermal energy. It dumps heat onto the wall and induces large forces in the structure. Avoiding disruptions, or terminating a discharge softly when one is unavoidable, protects the machine and keeps operation continuous.
Prediction as classification
Disruptions are preceded by measurable precursors: growing magnetic modes, radiation spikes, density limits, and current-profile changes. A predictor watches these signals and estimates, moment to moment, the probability that a disruption is developing. This is a time-series classification problem where the cost of a late alarm is high and the cost of a false alarm is real but smaller.
- Physics-based indicators flag known instability boundaries.
- Machine-learning models catch patterns across many signals at once.
- The two are combined so the alarm has both interpretability and coverage.
From prediction to avoidance
A prediction is only useful if there is time to act. An early, calibrated warning lets the control system back the plasma away from the boundary by adjusting current, density, or heating. If avoidance is not possible, the same warning triggers a controlled shutdown that spreads the energy safely.
Honest framing
No predictor is perfect, so it is characterized by its warning time, its true-positive rate, and its false-alarm rate, all reported openly. The tandem-mirror burner does not disrupt the way a tokamak does, so the Hyperion breeder is where disruption work concentrates.
Why it matters
Reliable avoidance is the difference between a research device that trips often and a plant that runs, and it feeds directly into safety and licensing evidence.