Machine-Learned Pedestal Prediction
Neural models predict the height and width of the edge transport barrier, a key driver of overall confinement in high-confinement plasmas.
The pedestal and why it matters
In high-confinement operation a steep pressure gradient forms at the plasma edge, called the pedestal. Its height sets a boundary condition for the core profiles, so overall performance is strongly sensitive to it. Predicting pedestal structure from engineering parameters is therefore central to scenario planning.
Physics baseline
The established physics picture links the pedestal to two constraints: peeling-ballooning magnetohydrodynamic stability limiting the pressure gradient, and a model for the pedestal width. Evaluating these constraints requires stability calculations that are too slow for rapid scans, motivating a learned surrogate.
Learned surrogate
- Inputs: plasma current, magnetic field, shape parameters, density, and heating power
- Outputs: pedestal height, width, and top pressure
- Training: databases of stability-code results and, where available, experimental measurements
A network trained on these mappings returns pedestal predictions instantly, letting designers sweep shaping and current to find configurations with favorable edge stability. Because shape strongly affects edge stability, this is especially relevant for concepts using strong shaping.
Caveats
Pedestal physics involves edge-localized instabilities and their control, which add complexity beyond a single steady-state height. A learned model captures trends within its training envelope but cannot substitute for dedicated stability analysis of a specific configuration. Extrapolation to unbuilt regimes is uncertain and must be flagged.
For the negative-triangularity Hyperion breeder concept, edge behavior differs from conventional positively-shaped plasmas, so pedestal surrogates trained mostly on conventional data are used cautiously and cross-checked against configuration-specific stability studies. The prediction is a design-screening aid, not a performance guarantee for hardware that is not yet built.