Surrogate Models: Concepts
Fast approximations that stand in for expensive simulations, and the rules for using them without fooling yourself.
What a surrogate is
A surrogate model approximates the input-output behavior of an expensive simulation or experiment with a cheap function fit to a set of evaluations. Once built, it answers new queries in microseconds instead of hours, enabling optimization, uncertainty propagation, and control that would otherwise be infeasible.
Common surrogate forms
- Polynomial and radial-basis interpolants
- Gaussian processes, which give calibrated uncertainty
- Neural networks for high-dimensional inputs
- Gradient-boosted trees for tabular data
The interpolation rule
A surrogate is trustworthy inside the region covered by its training points and unreliable outside it. The central discipline is knowing where you are relative to the data. Gaussian processes report growing variance away from data; other models need explicit out-of-distribution checks.
Building a good surrogate
Sampling matters as much as the model. Space-filling designs (Latin hypercube, low-discrepancy sequences) cover the input space efficiently. Active learning adds points where the surrogate is most uncertain or where they most improve a downstream objective, spending expensive evaluations wisely.
Validation
Hold out simulation runs the surrogate never saw and measure error there. Report error distributions, not just means, and test at the corners of the input domain where extrapolation risk is highest. A surrogate that is accurate on average but wrong at extremes can mislead an optimizer that pushes toward those extremes.
In fusion, surrogates stand in for transport, turbulence, MHD stability, and neutronics calculations, each covered on its own page.