Disruption Prediction
A twin can watch for the precursors of a plasma disruption and forecast its onset in time for the machine to respond.
The event a tokamak most wants to avoid
A disruption is a sudden loss of plasma confinement in a tokamak, releasing stored energy quickly and imposing severe thermal and mechanical loads on the machine. Predicting a disruption early enough to act, whether to steer the plasma back to safety or to trigger a controlled shutdown, is one of the most important functions a plasma twin can provide.
Precursors
Disruptions are usually preceded by warning signs: growing magnetic fluctuations, changes in the plasma profiles, and drifts toward known stability limits. A twin that reconstructs the internal state can watch how close the plasma is to those limits, combining a physics understanding of stability boundaries with pattern recognition of precursor signatures.
Physics-based and data-driven prediction
- Physics-based: track the plasma's margin to known stability boundaries using the reconstructed state
- Data-driven: learn precursor patterns from histories of disruptive and non-disruptive behavior
- Hybrid: use physics to define the danger zone and data to sharpen the timing
The cost of errors
Prediction must balance two failure modes: missing a real disruption is dangerous, while too many false alarms needlessly interrupt operation. A twin quantifies this trade-off explicitly and reports its confidence, so the response, gentle correction or a hard trip, can be matched to the certainty and the time remaining. Carrying competing hypotheses, as a particle filter does, is natural here. See particle filters.
In the Kronos program
Disruption prediction applies to the Hyperion breeder, a spherical tokamak; the burner, being a tandem mirror rather than a tokamak, faces different loss mechanisms and its twin watches for those instead. Because the machines are not built, prediction methods are developed today on simulations and on published disruption data from other devices, with the honest caveat that plant-specific skill can only be established once the breeder operates near 2030. Predicting the event well is also central to protecting component life. See component-life tracking.