Surrogate Model
Definition
A data-driven model that approximates the input-output behavior of an expensive simulator so that large numbers of evaluations become cheap. The emphasis is on "replacing" rather than "reducing": it usually does not exploit the governing-equation structure explicitly.
What it is
A surrogate treats the simulator as a black box, y ~ f(x). You sample a training set with a limited number of real simulations, then fit a regressor — Gaussian process, neural network, random forest. Once trained, each evaluation costs milliseconds, which is what makes design optimization, sensitivity analysis and uncertainty quantification with tens of thousands of evaluations feasible.
Key points
- Boundary with ROM: ROM explicitly projects the equations (white/grey box); a surrogate is usually pure data fitting (black box). The line blurs for physics-informed surrogates.
- Sampling strategy decides success: Latin hypercube, Sobol sequences and adaptive infill (Bayesian-optimization style) beat random sampling by a wide margin.
- Curse of dimensionality: beyond roughly 20 input dimensions the required sample count explodes — run a sensitivity analysis and reduce dimension first.
- Never trust extrapolation: the model is only valid inside the training domain and gives no guarantees outside it.
Common pitfalls / boundaries
- Do not use a surrogate outside its training domain, and do not treat it as an explanation of physical law.
- Separate "test-set error" from "deployment-condition error" in your accuracy report; the latter is usually much larger.