Is an emulator the same as the simulator?

No. It is a statistical representation trained on simulator outputs.

Can a deterministic simulator have a probabilistic emulator?

Yes. The probability can express uncertainty about unevaluated function values.

Does its interval include all real-world uncertainty?

No. Input uncertainty, discrepancy, measurement error and other sources require appropriate additional modeling.

AI Glossary / Statistics and probability / Advanced

Gaussian process emulator

Definition

A Gaussian process emulator is a statistical surrogate that learns the input-output behavior of a simulator using a Gaussian process model. It predicts outputs at inputs where the original simulator has not been run and provides a model-based distribution describing uncertainty about those outputs. Its purpose is often to make repeated exploration of an expensive simulation practical. The uncertainty concerns the simulator's unknown response between evaluated settings; it does not automatically describe how accurately that simulator represents the real world.

What it means

Suppose one engineering simulation takes an hour, but a design investigation needs thousands of evaluations. An emulator can learn from a carefully selected collection of runs and supply faster predictions elsewhere. A Gaussian process describes relationships among function values through a mean function and covariance kernel. This lets observations at one setting inform predictions at related settings. The simulator may be deterministic even though the emulator uses probability: the probability represents uncertainty about a function that has not been fully evaluated, rather than randomness in each simulator run.

How it works

Define the relevant input region and output, select training settings, and run the simulator. Choose a mean, kernel and treatment of numerical error or stochastic variation, then condition the GP on the results. Estimate or integrate uncertain parameters as appropriate. Reserve additional simulator runs for checking predictive errors and interval calibration. Refine the design where uncertainty matters to the intended decision. Under a noiseless model with fixed parameters and nonsingular covariance, predictions interpolate the training outputs. Adding a noise or smoothing term changes this property, so its meaning should be documented.

Three practical examples

Engineering design: approximate simulated energy consumption across candidate equipment settings before selecting a smaller set for full evaluation.
Sensitivity analysis: investigate which uncertain inputs contribute most to variation in a simulator's output, while retaining emulator uncertainty.
Scenario exploration: compare possible demand and capacity configurations in an expensive operational simulator. Each application needs validation over the input region actually used.

Think of it this way

Think of an emulator as an informed map drawn from a limited set of surveyed locations. It estimates what lies between them and marks where the map is uncertain. A precise map of a simulation's landscape can still differ from the physical landscape. Surveying more simulator locations improves knowledge of the simulation, but does not by itself correct flaws in the simulation's description of reality.

Worked example

Consider a deliberately small, noiseless example with zero prior means. The simulator value at training input a and at new input b has prior covariance matrix [[4, 2], [2, 4]]. This is valid: its eigenvalues are 6 and 2. Observe f(a) = 6. Gaussian conditioning gives E[f(b) | f(a)] = (2/4) × 6 = 3 and Var[f(b) | f(a)] = 4 − 2²/4 = 3. Thus the conditional standard deviation is √3, about 1.732. A central 95% conditional Gaussian interval is approximately −0.395 to 6.395. At a itself, the noiseless model returns mean 6 and variance zero. These calculations assume the supplied covariance and mean are known. Zero uncertainty at a means the simulator value there is known; it makes no assertion about the corresponding real-world outcome.

Illustrative case study

Fictional case study: Alder Systems builds an emulator for a costly thermal simulator. Its first model predicts held-out runs well near ordinary operating settings but misses sharp changes near a control threshold. The team adds simulator runs around that threshold and revisits the kernel assumptions before using the emulator in design comparisons. It separately compares selected simulated temperatures with physical measurements. The exercise distinguishes inaccurate emulation from inaccurate simulation and prevents a narrow emulator interval from being presented as a complete statement of physical uncertainty.

Common misconceptions and mistakes

Common mistakes include treating a surrogate's speed as evidence of accuracy, validating only at training inputs, and assuming a smooth kernel can represent abrupt transitions without modification. Another error is combining uncertainty about simulator outputs with model discrepancy without stating how each is modeled. Extrapolation beyond the training region can be especially dependent on prior assumptions. Fixed fitted parameters can understate uncertainty relative to an analysis that integrates parameter uncertainty. Dense GP calculations can also become expensive as the training set grows; emulation is not computationally free.

Business use

Businesses use GP emulators to investigate design tradeoffs, prioritize costly simulations and explore sensitivity before committing resources. Evaluate them against the decision they support: small average errors may still conceal mistakes near a feasibility boundary. Keep a record of simulator version, training domain, transformations, validation runs and uncertainty assumptions. Require direct simulation checks for consequential shortlisted configurations. Savings depend on the cost of building and maintaining the emulator compared with the simulator evaluations it replaces; there is no universal acceleration factor.

Check your understanding

Why does the worked example have zero variance at a but positive variance at b?

Answer: The noiseless observation reveals the simulator value at a exactly. The specified correlation provides only partial information about b. Neither statement establishes that the simulator matches reality.

Related terms

Learn next

compare simulator holdout errors with predicted standard deviations, then examine how alternative kernels and training designs affect a decision boundary. Study the distinction between uncertainty about an unevaluated simulator and discrepancy between simulation and observation before attempting calibration.

Decision boundary

Sources

Sources: A. O'Hagan, Bayesian Analysis of Computer Code Outputs: A Tutorial, https://tonyohagan.co.uk/academic/pdf/BACCO-tutorial.pdf (emulation and uncertainty distinctions). C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for Machine Learning, chapter 2, https://gaussianprocess.org/gpml/chapters/RW2.pdf (Gaussian conditioning and regression). The numerical example and fictional case are original illustrations.

Previous: Gaussian Process Classification Next: Gaussian Process Regression