Does affine calibration change simulator outputs?

Yes, indirectly: changing the inputs changes the outputs produced by the simulator. The distinction is where the correction is applied.

Can a single offset be sufficient?

Sometimes. Setting the scale to one gives an offset-only special case, but its suitability must be evaluated from the problem and data.

Does exact agreement at calibration points guarantee accurate forecasts?

No. Interpolation, extrapolation, parameter uncertainty and model inadequacy still require assessment.

AI Glossary / Statistics and probability / Intermediate

Affine calibration

Definition

Affine calibration, in the source glossary's computer-experiment sense, improves agreement between simulated and observed responses by applying an affine transformation to one or more simulator inputs. A simple transformation replaces an input x with a + bx, combining an offset and a scale adjustment before the simulator produces its response.

What it means

The selected meaning concerns adjustment on the input side of a simulator. The simulator itself may be nonlinear, so an affine input transformation need not produce a simple affine transformation of its output. Calibration aligns predictions with observations over a relevant region.

Terminology varies across fields: sensor and vision applications also use affine calibration in other settings. NIST calls a transformation of the raw simulated response global calibration, whereas the affine calibration described here modifies input factors.

How it works

Collect observations of a physical process together with the relevant input conditions. Define which simulator inputs may be transformed and choose an affine form, such as x-star = a + bx for one factor. For multiple factors, transformations may involve several offsets and scale parameters, subject to the application's interpretation and constraints.

Estimate the calibration parameters by comparing observed responses with simulator outputs obtained from transformed inputs. An appropriate fitting criterion might weight discrepancies according to measurement uncertainty. The precise criterion is part of the calibration model, not implied by the word affine.

Check whether the parameters can be identified from the available observations and whether transformed inputs remain physically and computationally valid. Evaluate performance on observations not used for fitting. Residual disagreement can reflect measurement noise, uncertain inputs or limitations of the simulator itself. A transformation that compensates for model inadequacy over a narrow range may fail elsewhere, so agreement during calibration is not sufficient validation.

Three practical examples

1. An engineering team adjusts the offset and scale of a simulator's temperature input to investigate a systematic mismatch between predicted and measured process responses.

2. A manufacturing analyst calibrates a simulator's input factor over an experimentally supported operating range, then checks predictions at separate validation conditions.

3. A simulation team compares an input-side affine correction with an output-side correction, documenting which mechanism is being changed and why it is plausible.

Think of it this way

Think of a machine that responds to the position of a control dial. If the dial's zero point and spacing are misaligned with the machine's internal scale, an offset and scale conversion can translate the dial reading before it reaches the machine. Changing that translation is different from altering the displayed result afterward.

Worked example

Consider an original toy simulator f(z) = z squared. Suppose physical input x = 1 produces an observed response of 9, and x = 3 produces 49. Restrict transformed inputs to positive values and choose z = a + bx.

Matching the two observations requires a + b = 3 and a + 3b = 7. Subtraction gives 2b = 4, so b = 2 and a = 1. The calibrated simulator is f(1 + 2x) = (1 + 2x) squared. At x = 2, it predicts 25, whereas the unadjusted simulator f(2) predicts 4.

For comparison, an output-side affine correction c + d f(x) matching the same two observations satisfies c + d = 9 and c + 9d = 49, giving d = 5 and c = 4. At x = 2 it predicts 24, not 25. Thus two approaches can agree at calibration points but differ between them. A separate observation at x = 2 would help evaluate them; two-point agreement alone does not establish either as the correct model.

Illustrative case study

Fictional case study: Simulane compares a production simulator with a small set of physical trials. A team proposes rescaling one input to improve agreement. Before accepting the correction, analysts check the input's units, plausible range and whether an actual measurement offset could explain the mismatch.

They fit the transformation on selected trials and evaluate it on reserved conditions. Some residual patterns remain, so they document a restricted range of use and investigate missing process behavior rather than adding arbitrary corrections until every trial fits.

This fictional case illustrates calibration with validation and physical interpretation. It does not claim that an affine transformation can repair every simulator or identify the true cause of a discrepancy.

Common misconceptions and mistakes

Affine does not mean an arbitrary nonlinear correction: the input transformation has an offset-plus-linear form. Conversely, the resulting response can be nonlinear when the simulator is nonlinear. In strict mathematical terminology, a nonzero offset makes a transformation affine rather than purely linear.

A good calibration fit does not prove that the simulator's physical assumptions are correct. Different transformations or model errors may explain the same observations. Avoid confusing input-side calibration with probability calibration in classification, or assuming that a fitted correction remains reliable far beyond the tested input range.

Business use

Affine calibration offers a relatively simple way to investigate systematic simulator mismatch when a scale or offset adjustment is scientifically meaningful. It can support virtual experiments, process planning and comparisons between measured and simulated behavior.

Record the transformed factors, parameter estimates, units, fitting data and validated operating region. Preserve the original simulator configuration so the impact of the calibration is traceable. Account for uncertainty in observations and fitted parameters when decisions depend on predicted performance. If a correction requires implausible input values or changes markedly across datasets, investigate the underlying model and measurement process instead of treating the fitted parameters as unquestionable physical facts.

Check your understanding

An input transformation is z = 2 + 0.5x. What value enters the simulator when the recorded input is 8?

Answer: z = 2 + 0.5 times 8 = 6. The simulator then evaluates its response at 6; the response itself is not necessarily 6.

Related terms

Computer experiment
Affine transformation
Global calibration
simulator calibration
model discrepancy
identifiability
validation and uncertainty quantification

Learn next

compare input-side and output-side corrections for a nonlinear toy simulator. Fit both on the same observations, test them at new inputs and explain why matching training points does not make the two calibrated models equivalent.

Sources

Primary sources: NIST/SEMATECH, Glossary, affine calibration and global calibration: https://www.itl.nist.gov/div898/handbook/glossary.htm

Toman, Calibration and Uncertainty Analysis of Predictions from Computational Models, NIST: https://www.nist.gov/publications/calibration-and-uncertainty-analysis-predictions-computational-models

Previous: Affine bilateral grid (HDRnet) Next: Affine transformation