Surrogate Models and Emulators
A fast approximation trained on a slow model lets you explore vast design spaces, as long as you respect where it was trained.
Standing In for the Expensive Model
A surrogate model, also called an emulator or response surface, is a fast approximation fit to the outputs of an expensive simulation. Where one high-fidelity run is costly, a surrogate answers in an instant, enabling optimization, uncertainty quantification, and design sweeps that would be infeasible with the full model. It trades a small, bounded loss of accuracy for an enormous gain in speed.
How They Are Built
- Run the expensive model at a set of carefully chosen input points, a design of experiments.
- Fit a flexible function, such as a Gaussian process or neural network, to those points.
- Validate the fit on held-out runs the surrogate did not see.
- Report the surrogate's error and its valid input range.
The Domain of Validity
A surrogate is trustworthy only inside the region of input space where it was trained and tested. Outside that region it extrapolates without any mechanism to guide it, and its output can be confidently wrong. The single most important discipline in using surrogates is knowing and respecting the boundary of their valid domain.
Uncertainty-Aware Surrogates
Some surrogates, notably Gaussian processes, report not just a prediction but an estimate of their own uncertainty, which grows away from the training points. This lets a workflow know when it is extrapolating and fall back to the full simulation, or add a new training run, exactly where confidence is low. A surrogate that knows what it does not know is far safer than one that does not.
At Kronos
Where full physics runs are too slow to sweep, surrogates enable design exploration, but they are validated on held-out cases and used only within their tested range, with the underlying simulation as the authority whenever a new regime is reached. The fast approximation accelerates the search; it does not replace the physics.