Surrogate Model

A surrogate model, also called a metamodel or emulator, is a fast stand-in for a slow model. It is built by running the detailed model many times over a range of inputs and then fitting a simpler model that maps those inputs to the outputs of interest. Once trained, it can give an answer in a fraction of a second where the full model might take minutes or hours. Common surrogate types include polynomial response surfaces, Gaussian process models, random forests and neural networks. Another family simplifies the physics itself, for example a coarse-grid or reduced-order version of a hydraulic model. Water examples are many. A 2D flood model of a city may take hours per storm; a surrogate trained on hundreds of simulated storms can estimate flood depths at key locations fast enough for real-time warnings, or for testing thousands of scenarios. In groundwater management, a surrogate of a groundwater flow model can sit inside an optimisation that chooses pumping rates. In model predictive control of sewers or supply networks, surrogates make it possible to solve the control problem every few minutes. What it needs: a trusted detailed model, and a well-designed set of training runs that covers the conditions the surrogate will face. Sampling designs such as Latin hypercube sampling help spread runs across the input space. Limits: a surrogate is only as good as the model it copies, so it inherits that model's errors. It is reliable only inside the range of its training runs. Outside it, for a storm larger than any in the training set or a changed network layout, it can give poor answers without any warning. When the real system changes, the surrogate must be retrained. Always test it on runs it has not seen before trusting it.