Computing Library › Ml For Fusion
Ml For Fusion

Disruption Early-Warning Systems

An end-to-end disruption warning system fuses diagnostics, a trained predictor, and a mitigation trigger under strict real-time latency budgets.

From predictor to system

A classifier that recognizes disruption precursors is only one part of a working early-warning system. The deployed system must acquire diagnostics, compute features, run inference, and trigger mitigation, all within a latency budget short enough to act before the disruption develops.

Real-time pipeline

Kronos motion — time to disruption

Latency and determinism dominate the engineering. The pipeline typically runs on real-time controllers or field-programmable gate arrays so that worst-case timing is bounded, not merely fast on average.

Mitigation actuators

On a positive warning the system fires a mitigation actuator, most commonly massive gas injection or shattered-pellet injection, to radiate the stored energy uniformly and avoid concentrated heat loads and runaway electrons. The warning threshold is tuned to give these actuators enough time to act.

Reliability engineering

Because a single missed disruption can damage the machine, warning systems favor conservative thresholds, redundant signals, and graceful fallback to simple physics-based limits if the model is uncertain or a sensor fails. Continuous monitoring of prediction confidence feeds this safety logic.

For a first-of-a-kind device the system must operate credibly with little device-specific training data, so it blends transferable physics indicators with learned components and is validated in simulation and on surrogate data first. In Kronos design planning this is part of the breeder safety case and remains a design study, since construction of the Hyperion concept is scheduled to begin in Q2 2027 and no plasma hardware yet exists.