Surrogate Model

A surrogate model is an inexpensive, often statistical or machine‑learning based approximation that stands in for a costly-to-evaluate objective function. Instead of repeatedly running a time‑consuming simulation, physical experiment, or high‑dimensional computation, the surrogate learns a mapping from input variables to output performance using a limited set of true evaluations. Once trained, it can be queried cheaply, providing estimated objective values that guide further exploration.

The value of surrogate models lies in their ability to dramatically accelerate iterative processes such as design optimisation, hyperparameter search, or policy learning. By offering quick approximations, they enable algorithms to evaluate many candidate solutions, estimate gradients, or assess uncertainty without incurring the full expense of the original system. The resulting efficiency gains make previously infeasible optimisation problems tractable and allow practitioners to allocate resources toward promising regions of the design space.

Surrogate modelling appears in fields ranging from aerospace engineering—where CFD simulations are replaced by response‑surface models—to automated machine learning, where Bayesian optimisation employs Gaussian process surrogates. It is also common in materials discovery, robotics control tuning, and any domain where each true evaluation carries a significant computational or experimental cost.