Surrogate Models for Transport
Learned stand-ins for turbulent transport calculations that make integrated modeling and control feasible.
The transport bottleneck
Turbulent transport sets how fast heat and particles leak across the confining field. Physics-based transport models compute local fluxes from local gradients, but the accurate ones are too slow to run inside an integrated whole-device simulation or a controller. Surrogates replace them.
What is being approximated
A transport surrogate maps local plasma parameters (gradients, temperature ratios, collisionality, safety factor, geometry) to turbulent heat and particle fluxes. Trained on many runs of a quasilinear or gyrokinetic model, it reproduces the flux response fast enough for repeated evaluation.
- Inputs: normalized gradients, Ti/Te, q, magnetic shear, collisionality
- Outputs: ion and electron heat flux, particle flux
- Use: integrated modeling, scenario optimization, control
Stiffness and thresholds
Turbulent transport is stiff: flux rises sharply once a gradient exceeds a critical threshold. A surrogate must capture this near-threshold behavior, or integrated simulations built on it will predict wrong profiles. Training data must be dense around the threshold, and the model must not smooth it away.
Validation and limits
Compare surrogate fluxes to the underlying physics model on held-out points, focusing on the threshold region and on parameter corners. Surrogates trained on one turbulence regime should not be used in another without retraining. Outside the training envelope, flux predictions can be badly wrong even when they look smooth.
Design use
Transport surrogates let designers scan many operating points for a device such as the Kronos breeder without rerunning full transport codes each time. Results are model predictions for a design under study, cross-checked against the source physics.