Surrogate Models in Code Workflows
Surrogate models approximate expensive physics codes with fast statistical or machine-learned functions, enabling scans that would otherwise be infeasible.
Replacing the expensive part
Some physics modules, gyrokinetic turbulence, full-wave heating, edge simulation, are too slow to call thousands of times inside optimization or uncertainty loops. A surrogate model is a fast approximation trained on a set of full-code runs that reproduces the input-output mapping at a tiny fraction of the cost, standing in for the expensive code where speed is essential.
Common surrogate forms include Gaussian-process regression, which also returns a predictive uncertainty, polynomial or spline fits, and neural networks for high-dimensional mappings.
Training and validation
A surrogate is only trustworthy within the region spanned by its training data. Building one requires a designed set of full runs, held-out validation to measure accuracy, and clear bounds on where it may be used. Extrapolation beyond the training envelope is unreliable and must be guarded against.
Uses
Surrogates enable real-time-capable transport models, fast optimization over many parameters, and affordable uncertainty sampling. The surrogate's own approximation error must be folded into any uncertainty budget so it does not hide behind the speed gain.
Design relevance
For the Hyperion breeder, surrogates of turbulent transport allow broad scans of the operating space that full gyrokinetics could not cover, accelerating design iteration. Their predictions are cross-checked against the full codes at selected points, keeping the simulation evidence honest ahead of construction.
- Fast approximation of an expensive code
- Gaussian processes, polynomials, neural nets
- Valid only within the training envelope
- Enables optimization and uncertainty at scale