ML for Density-Limit Prediction
Learned models estimate proximity to the operational density limit, where excessive fueling drives radiative collapse and disruption.
The density limit
Tokamaks cannot be fueled arbitrarily dense. Beyond an operational density limit, edge cooling, detachment, and radiation drive the plasma toward collapse and disruption. The classic empirical bound scales with plasma current and inverse size, but the true limit depends on many factors and varies by scenario.
Why prediction is hard
The density limit is not a sharp line but a regime where several processes interact: edge radiation, detachment dynamics, and magnetohydrodynamic activity. A single empirical formula captures the trend but not the scenario-specific margin, so a data-driven model that uses more of the plasma state can give a sharper proximity estimate.
- Inputs: density relative to the empirical limit, radiated fraction, edge conditions
- Outputs: probability of imminent density-limit disruption or margin to the limit
- Use: warn a controller to reduce fueling before collapse
Coupling to control
A proximity estimate is actionable: as the plasma approaches the limit, the controller can reduce fueling, add heating, or adjust the edge. This makes density-limit prediction a component of broader disruption avoidance rather than a standalone metric.
Cautions
Because the limit is scenario- and device-dependent, models trained on one machine transfer imperfectly, and rare limit-disruptions create class imbalance. Predictions should carry uncertainty and be backed by the conservative empirical bound as a floor.
Density control is part of the operating envelope for any breeder concept, including Hyperion at its design current of 9.86 MA. In Kronos design work these predictors are developed against simulated scenarios; they act on modeled plasmas because the hardware is not yet built, and their estimates are computational.