ML for Plasma State Classification
Classifiers label the plasma regime in real time, distinguishing confinement modes and operational phases to inform control decisions.
Knowing the regime
A discharge passes through distinct phases and confinement regimes, from low-confinement to high-confinement operation, with transitions that change how the plasma responds to actuators. Automatically classifying the current regime lets a controller apply the right strategy for the state it is in.
Signals of a regime
- Edge behavior, such as the pedestal signature of high-confinement mode
- Fluctuation levels in magnetic and radiation signals
- Stored energy and confinement trends
- Divertor and edge-density conditions
A supervised classifier trained on labeled discharges maps these signals to a regime label. Because regimes are defined by combinations of features rather than a single threshold, learned classifiers often outperform hand-set rules, particularly near transitions.
Transition detection
The most valuable and hardest task is detecting the transition itself, such as the sudden step into high-confinement mode, with low latency. Sequence models that track the evolving state can flag a transition as it happens, which matters because control priorities shift at the boundary.
Interpretability
Operators need to trust a regime label, so interpretable features and confidence scores are valued over opaque outputs. A classifier that also reports why it chose a label, and how confident it is, integrates better with human oversight and safety logic.
Regime awareness supports scenario control for concepts like the Hyperion breeder. These classifiers are trained on simulated and archival data ahead of construction; they label modeled plasma states because the device is not yet built, and their labels are computational inferences to be validated on hardware.