Disruption Prediction

Prediction gives the control system advance warning of an approaching disruption so it can avoid or mitigate in time.

The prediction task

Disruption prediction answers a real-time question: is this discharge heading toward a disruption, and how soon? A useful predictor must give enough warning for the control system to act, with few false alarms, since a false alarm that triggers mitigation needlessly ends a good discharge.

Physics-based indicators

Some predictors track physical proximity to known limits: a growing locked-mode amplitude, a density approaching the limit, a shrinking safety-factor margin, or rising radiated power fraction. These are interpretable and tie directly to a cause, which makes the appropriate response clear.

Data-driven predictors

Machine-learning predictors are trained on many past discharges, both disruptive and healthy, to recognize the signatures that precede a disruption. They can catch combinations of signals that no single physics indicator flags. Their weakness is that they may not generalize to conditions unlike their training data, a serious concern for a new machine with little operating history.

Warning time versus accuracy

There is a trade-off between how early a warning comes and how accurate it is. Earlier warnings are less certain; later ones are surer but leave less time to act. Predictors are tuned to give the response system the lead time it needs, biased toward catching real disruptions even at the cost of some false alarms.

In the Kronos program

For the Hyperion breeder, prediction feeds the avoidance-first strategy: an early warning prompts the off-normal layer to restore margin or terminate safely, reserving mitigation for late, high-confidence warnings. Because a new machine lacks operating history, initial predictors lean on interpretable physics indicators validated in simulation, with data-driven methods added as operating experience accumulates.