Surrogate Model
A fast approximation trained to mimic an expensive simulation or experiment.
Definition
A surrogate model (or emulator) is a cheap-to-evaluate function fit to the inputs and outputs of an expensive process, such as a high-fidelity simulation. Once trained, it predicts the process's output in a fraction of the time.
Active learning makes surrogates efficient: rather than sampling the expensive process at random, the surrogate's own uncertainty guides where to gather the next data point, concentrating effort where the model is least sure. This closes the loop between the surrogate and the simulation it emulates.
A surrogate is trustworthy only within the region its training data covers, and its greatest danger is confident extrapolation beyond that region. Uncertainty-aware surrogates, such as Gaussian processes and model ensembles, address this by reporting where they are unsure, which both guards against overreach and directs new data collection to where it most improves the model. This uncertainty is as valuable as the prediction itself.
Common forms
- Gaussian processes: smooth fits with built-in uncertainty.
- Neural networks: flexible, scale to many inputs.
- Polynomial and radial-basis models for low dimensions.
Why it matters
Surrogates make design optimization, sensitivity analysis, and uncertainty quantification tractable when each true evaluation is slow. Their validity depends on staying within the region where they were trained; extrapolation is unreliable.
Fusion connection
Surrogates are central to Kronos design work: they let engineers explore the Hyperion parameter space quickly, with every promising candidate confirmed against full first-principles simulation before it informs the design.