Machine Learning for MHD Stability Prediction
Surrogates estimate magnetohydrodynamic stability boundaries fast enough for scenario design and real-time avoidance of unstable operating points.
Stability sets the limits
Magnetohydrodynamic (MHD) stability determines how much pressure and current a plasma can hold before large-scale modes grow and degrade or terminate confinement. Evaluating stability rigorously means running eigenvalue solvers over candidate equilibria, which is too slow to sweep large scenario spaces or to run in real time.
Learned stability boundaries
- Predict growth rates or stable-unstable labels from equilibrium features
- Estimate margins to key limits, such as the pressure and current bounds
- Identify the dominant unstable mode number
- Flag proximity to a resistive or ideal instability
A surrogate trained on stability-code results maps equilibrium and profile descriptors to these outputs instantly. This lets designers avoid unstable regions during scenario search and lets controllers steer away from limits during operation.
Real-time avoidance
Coupled with fast equilibrium reconstruction, a stability surrogate supports active avoidance: as the plasma approaches an unstable boundary, the controller adjusts current, pressure, or shaping to retreat. This is disruption avoidance at the level of the underlying instability rather than its precursors.
Cautions
Stability can depend sensitively on details of the current and pressure profiles that a coarse descriptor omits, so surrogates must capture the relevant features and carry uncertainty. Near a boundary, small input errors can flip a stable-unstable prediction, so conservative margins are used.
Stability limits shape the operating envelope for concepts like the Hyperion breeder at its design current and field. In Kronos design work these surrogates are developed against stability codes ahead of construction; they assess modeled equilibria because the machine is not yet built, and results are computational estimates cross-checked with full solvers.