ML for Edge-Localized-Mode Prediction
Machine learning forecasts edge-localized modes and supports control strategies that mitigate the transient heat loads they impose on plasma-facing components.
What an ELM is
An edge-localized mode (ELM) is a repetitive burst that expels energy and particles from the pedestal region of a high-confinement plasma. Large ELMs deposit intense transient heat loads on divertor and wall components, so predicting and controlling them protects hardware and sustains steady operation.
Prediction from signals
ELM onset leaves precursors in magnetic, radiation, and edge-density diagnostics. Time-series models, including recurrent and convolutional networks, learn to flag an imminent ELM from these signals with tens of milliseconds of lead time, enough to trigger a mitigation actuator.
- Magnetic pickup coils sensing edge fluctuations
- Divertor and edge radiation signals
- Edge density and temperature from fast diagnostics
- Derived features such as fluctuation amplitude and frequency drift
Control coupling
Prediction is only useful if paired with an actuator: resonant magnetic perturbation coils, pellet pacing that triggers small frequent ELMs, or adjustments to fueling and shaping. A learned predictor can gate these actuators or provide a control signal, and reinforcement learning has been explored to tune pacing schedules.
Practical limits
ELM behavior is device- and scenario-specific, so predictors trained on one machine transfer imperfectly to another. Class imbalance is a challenge because large ELMs are relatively rare, requiring careful resampling and cost weighting so the model does not simply predict the common case.
For design of any burning-plasma device, controlling edge transients is a component-protection requirement. In Kronos concepts these questions belong to the deuterium-tritium breeder line rather than the tandem-mirror burner, and any ELM-control scheme is a simulation and design study until validated on operating hardware, which does not yet exist.