Machine-Learning Surrogates for Plasma Transport
Plasma transport is expensive to compute yet needed everywhere; learned surrogates make it fast enough for design and control.
The transport bottleneck
How heat and particles move across a magnetized plasma, its transport, sets confinement and therefore performance. Computing transport from first principles is expensive, often the slowest part of a whole-device model. Yet transport predictions are needed throughout design, scenario planning, and control, where the full calculation is too slow.
The surrogate approach
- Run high-fidelity transport calculations across many plasma conditions.
- Train a model to predict transport fluxes from local plasma parameters.
- Embed the fast surrogate wherever transport is needed.
- Re-check against the full calculation where the surrogate is uncertain.
Why it works
Transport fluxes depend mainly on local gradients and parameters, a relationship a well-trained model can capture. A surrogate trained on high-fidelity data returns a flux in microseconds instead of hours, with characterized error, making whole-device modeling and scenario optimization practical.
Staying honest
A transport surrogate is valid only across the conditions it was trained on. Outside that range it is flagged and the full calculation is used, so the speedup never silently produces wrong physics. Its validated range travels with it, as with any reduced model.
For Kronos
Transport surrogates support closing the Hyperion design point, where confinement is central, and they feed the fast models used in control. The burner has its own transport physics and its own surrogates.
Verification
Surrogate transport predictions are checked against the full model continuously, part of automated verification, so confidence in them is earned and maintained.