Surrogate Model

Surrogate model is an approximation of a complex, expensive, or otherwise inaccessible function, constructed to make optimization or analysis tractable. In Bayesian optimization, the surrogate is typically a Gaussian Process that replaces the expensive black-box objective with a cheap probabilistic approximation. But the concept generalizes: surrogate models appear in engineering design, climate simulation, and any domain where the true function is too costly to evaluate directly and where a cheaper stand-in can guide decision-making.

The epistemic status of a surrogate is delicate. It is not merely a computational convenience but a theory of the function it replaces — a theory whose predictions are used as if they were the function itself. When the surrogate and the true function disagree, the optimizer has learned the wrong landscape. This makes surrogate modeling a form of meta-modeling: the model is not of reality but of another model's behavior, and its accuracy depends on where it has been trained and what structure it assumes.