Must the transformation be linear?

No. A linear function is a simple example, not the definition.

Is this the same as changing simulator inputs?

No. This glossary sense adjusts the simulated response.

Does better calibration prove causality?

No. Agreement with observations does not establish the correctness of causal mechanisms.

AI Glossary / Statistics and probability / Intermediate

Global calibration

Definition

Global calibration, in the computer-experiment sense used by the NIST engineering-statistics glossary, means developing a function that transforms a simulator's raw response to improve its agreement with empirical observations. The adjustment operates on the output of the simulator. This entry uses that specific meaning, because the same phrase can describe different calibration tasks in other fields, including sensor networks and machine-learning probability assessment.

What it means

A simulation can capture useful relationships while consistently predicting values on the wrong scale. Instead of immediately rebuilding the simulator, an analyst may learn how its outputs correspond to observed measurements. The resulting transformation becomes an additional component of the prediction process.

For example, a simulator's energy-use estimate might be adjusted upward to match observed consumption over a studied operating range. The calibration does not make the simulator's internal mechanisms correct, nor does it guarantee performance outside that range.

The word global should not be read as a promise of worldwide applicability, perfect agreement, or a mathematically global optimum. The important distinction here is the transformation of the simulated response.

How it works

Collect matched cases: simulator output s_i and empirical response y_i for the same relevant conditions. Choose a transformation g and estimate its parameters from these pairs. A simple illustrative model is y_i = a + b*s_i + error_i, producing adjusted prediction g(s) = a + b*s.

Examine discrepancies and residual patterns, and use a transformation whose complexity is justified by the data and application. Nonlinear adjustments may be appropriate, but increasing flexibility can fit noise rather than persistent disagreement.

Assess performance on observations reserved from fitting, preferably covering the intended operating conditions. Track measurement uncertainty, calibration-parameter uncertainty, simulator variability where present, and remaining discrepancy. Updating the simulator may invalidate a previously fitted adjustment.

NIST distinguishes this output transformation from its glossary's affine calibration, which modifies simulator inputs through linear transformations. An output adjustment can itself be affine mathematically without belonging to that input-adjustment definition.

Three practical examples

Energy planning: adjust simulated building consumption using matched observed readings, then test the adjustment on later operating periods before using it in forecasts.
Manufacturing: compare a simulation's predicted response with measured trial outcomes and investigate whether an output transformation improves agreement over the tested range.
AI-assisted engineering: calibrate outputs from a simulation feeding a decision-support system, retaining the raw output and adjustment separately so their contributions remain auditable.

Think of it this way

Think of a printer that produces a useful image but consistently renders its brightness incorrectly. A correction applied to the finished output can improve the result without changing the underlying drawing. That correction must be checked on other images; it cannot restore detail that the original image never contained.

Worked example

Take an original teaching example with simulator outputs s = 10, 20, 30, 40 and matched empirical responses y = 14, 23, 35, 44. Fit the output transformation y = a + b*s by ordinary least squares.

The means are 25 and 29. The sum of squared deviations of s is 500, and the sum of cross-products of centered s and y is 510. Therefore b = 510/500 = 1.02 and a = 29 - 1.02*25 = 3.5. The calibrated predictions are 13.7, 23.9, 34.1, and 44.3.

Their residuals are 0.3, -0.9, 0.9, and -0.3, giving squared-error sum 1.8. Unadjusted simulator outputs have squared-error sum 4² + 3² + 5² + 4² = 66 on these same cases. Direct least-squares calculation independently confirms both coefficients and sums.

For a new simulator output of 25, the adjusted prediction is 3.5 + 1.02*25 = 29. If a genuinely held-out empirical response is 30, the calibrated absolute error is 1, compared with 5 for the raw output. That single check is encouraging but insufficient to establish general reliability. The training-error reduction alone is not validation, and a simulator output far beyond 10-40 would require additional scrutiny.

Illustrative case study

In a fictional factory, a simulation predicts cooling duration consistently below observed duration. Engineers collect paired simulation and measurement records across a planned set of conditions. They fit a simple output correction and preserve separate records of raw and adjusted predictions.

On a later trial, the correction improves agreement for ordinary production settings but fails for an unusual material. The team limits its use to the supported conditions and investigates the missing material effect. This fictional case illustrates why a useful calibration can coexist with important model limitations; it does not claim an observed commercial improvement.

Common misconceptions and mistakes

Common mistakes include confusing global calibration with probability calibration, assuming the correction validates the simulator's physics, and fitting and assessing the transformation on exactly the same records.

Another mistake is treating empirical observations as error-free truth. Their measurement process may contribute uncertainty or systematic disagreement. A fitted correction cannot establish which component caused the mismatch without additional evidence.

An output-only map also cannot distinguish two situations that produce the same raw simulator output but require different corrections. That limitation may motivate a richer discrepancy model or revision of the simulator itself. Such extensions should be described explicitly rather than hidden inside an overly broad label.

Business use

Use this approach when paired empirical evidence supports an output adjustment and the resulting predictions help a defined decision. Document the simulator version, calibration dataset, transformation, supported conditions, and validation results. Business owners should understand when the adjustment needs review, especially after equipment changes, changed operating practices, or substantial shifts in input conditions.

Check your understanding

Answer: A simulator gives the same output for two materials, but observed responses differ systematically by material.

Can one deterministic function of that output alone produce both required corrections?

Answer: No. It receives the same input in both cases. Additional information or a revised simulation is needed to represent the distinction.

Related terms

Calibration
Computer experiment
Affine calibration
Validation
simulator
model discrepancy
extrapolation
measurement uncertainty

Learn next

Learn the difference between fitting and validation, then study discrepancy modeling and uncertainty propagation. Compare output correction with parameter calibration and input adjustment using examples that preserve their different meanings.

Sources

NIST/SEMATECH Engineering Statistics Handbook, Glossary, global calibration and affine calibration: https://www.itl.nist.gov/div898/handbook/glossary.htm — authoritative terminology used here. NIST, Data analysis and model validation: https://www.itl.nist.gov/div898/handbook/mpc/section3/mpc365.htm — calibration-curve fitting and residual checks. NIST, How do we Use the Model for Extrapolation?: https://www.itl.nist.gov/div898/handbook/pri/section5/pri599c.htm — need for confirmatory observations. The worked example is an original output-calibration illustration, not a reproduction of a NIST experiment.

Previous: Glitch token Next: Global constraint