The process of adjusting a model's parameters so its outputs match observed data, such as measured river flow or groundwater levels.
Model calibration is the process of adjusting a model's parameters until its output matches observations as closely as possible. In hydrology this usually means tuning soil storage, infiltration, roughness or recession parameters so that simulated streamflow fits gauged flow, or so that simulated heads in a groundwater model fit measured well levels. Calibration needs three choices. First, the data: a record long enough to include wet and dry years, with a warm-up period at the start so model stores settle before scoring begins. Second, an objective function that scores the fit, such as root mean square error (RMSE), Nash-Sutcliffe efficiency, Kling-Gupta efficiency or percent bias. Third, a search method, from manual trial and error to automatic algorithms such as SCE-UA or DDS, which test thousands of parameter sets. For example, observed flows on three days are 10, 20 and 30 m³/s. Parameter set A gives 12, 18 and 33, with squared errors 4, 4 and 9, so RMSE = √(17 / 3) = 2.38 m³/s. Set B gives 10, 25 and 28, with squared errors 0, 25 and 4, so RMSE = √(29 / 3) = 3.11 m³/s. Set A fits better by this score. A calibrated model must then be validated on data not used for fitting. The usual way is a split-sample test: calibrate on one period, test on another. A model that scores well in calibration but poorly in validation has been overfitted. The choice of score shapes the result. Squared-error scores favour high flows, so a model tuned for floods may be poor at low flows; scoring log-transformed flows, or several objectives at once, can help. Many parameter sets often fit equally well, a problem called equifinality, so a single best set can hide large uncertainty. Calibration also pushes errors in rainfall and flow data into the parameters.