The process of adjusting model parameters and, when appropriate, inputs or bias-correction terms so that model predictions better agree with observational data, often including validation steps and consideration of predictive performance over target conditions.

Principle

Calibration aligns model outputs with data by optimizing parameters or adding corrective terms while guarding against overfitting through cross-validation, regularization, or withholding validation data; it recognizes both parameter uncertainty and structural model error.

Demonstration

Tuning parameters of a climate model to reproduce observed seasonal temperature cycles, adjusting emission-scaling factors in an air-quality model to match concentration measurements, or calibrating sensor bias in an instrument model before assimilation into predictions.

Misapplication

Over-tuning parameters to match a particular dataset (data dredging) so the model fits noise rather than signal, or using calibration to mask structural model deficiencies rather than confronting and revising model structure.

Consequence

A well-calibrated model provides improved predictive accuracy within the calibration domain and clearer uncertainty estimates; it supports reliable forecasting and policy-relevant simulations but requires careful validation to ensure transferability beyond calibration data.

Reversal

Model validation or falsification: instead of adjusting parameters to fit the data, the model is tested and potentially rejected, or alternative model structures are considered; model averaging may be used rather than single-model calibration.

Boundary

Focuses on parameter/input adjustment and practical alignment with data for predictive purposes; excludes pure model-selection (deciding between alternative structures), and differs from parameter estimation in that calibration is often pragmatic, iterative, and oriented to predictive utility rather than purely statistical optimality.

Semantic Tension

Tension exists between calibration and overfitting on one side and rigorous statistical estimation on the other; calibration is pragmatic and performance-driven, while statistical parameter estimation emphasizes principled uncertainty quantification and identifiability diagnostics.

Synthesis

Model calibration is the practical adjustment of parameters and corrective terms to bring model outputs into agreement with observations, balancing fit and generalization through validation, regularization, and explicit treatment of uncertainty and structural error.