Computing for Disruption Prediction

Forecasting sudden losses of plasma confinement early enough to mitigate them before they stress the machine.

What a disruption is

In a tokamak, a disruption is a rapid, uncontrolled loss of plasma confinement in which the stored thermal and magnetic energy is dumped quickly. This can impose large forces and heat loads on the machine. Predicting a disruption seconds or even tens of milliseconds early allows mitigation, such as a controlled shutdown, to soften its effects.

The prediction problem

• Inputs: many diagnostic signals evolving in time
• Output: probability the plasma will disrupt within a horizon
• Constraint: must run in real time within the control loop
• Cost asymmetry: missing a disruption is far worse than a false alarm

Physics-based versus data-driven

Some disruption precursors are understood physically, for instance approaching known operational limits in density or pressure. Others are captured statistically by machine-learning models trained on past discharges. Practical predictors combine both: physics limits as hard boundaries, learned models for the subtler patterns, with the physics providing sanity checks on the data-driven part.

def disruption_alarm(signals, model, physics_limits, threshold):
    if physics_limits.violated(signals):
        return True                        # hard limit: act regardless
    p = model.predict_proba(signals)
    return p > threshold                    # learned precursor

The transfer challenge

A model trained on one machine or one operating regime may not transfer to another, because disruptions depend on the specific configuration. This is a live research problem: predictors must be validated in the regime where they will be used, and their confidence outside that regime treated with suspicion.

Kronos framing

For a spherical tokamak like the Hyperion breeder, disruption handling is part of the control-system design developed against simulation before operation. This is design-stage work; the machine is not built, and construction begins in the second quarter of 2027. Negative triangularity (-0.30) is among the configuration choices whose stability behavior such models must capture.