Climate simulation without model calibration

Many scientific disciplines use numerical models that require calibration or adjustment before they can be used. From the simplest to the most sophisticated, what these models have in common is that they are constrained by a set of empirical relationships with modifiable parameters and coefficients. 

In the case of climate models, there's no need to calibrate them using observations. 

In general, time series of observed data are divided into two periods: the first is used for calibration, while the second is used to evaluate performance. Many models are based on this approach.

Examples include ecological models (ecological niches, population dynamics), hydrological models (HSAMI, MOHYSE, HYDROTEL, CEQUEAU), bioclimatic models for crop growth and yield (CATIMO, STICS), and models for forest ecosystems (e.g. ForHyM and ForSTeM for boreal and temperate forests).
 

Weather forecasting models and physical climate models such as ESMs and RCMs are not calibrated. They take a different approach, based on fluid mechanics equations. They simulate a series of meteorological events for each of the points on their computation grid.

For a climate model, these series of events extend over several decades, or even centuries. To do this, they need to know the chemical composition of the atmosphere and the parameters for solar radiation, as well as detailed information on the planetary surface provided by geophysical data.

What distinguishes them from the calibrated empirical models mentioned above is that in no case are the time series and statistics of temperature, wind, humidity, pressure and precipitation observations or those of any other climate variable used to formulate or adjust the model’s equations. Essentially, the laws of physics, particularly those of fluid mechanics, are the basis of a physical climate model.

However, many of the physical parameterizations included in climate models are based on empirical relationships. Do they need to be calibrated? 

Physical parameterizations seek to represent, in the model, the effect of certain phenomena that are observed in nature but cannot be directly taken into account by fluid mechanics equations. Their empirical relationships have sometimes been developed through measurement campaigns and, once integrated into the model, their parameters can be adjusted to a certain extent by testing a few different values to try and improve the statistics (mean, variance, etc.) of certain simulated variables compared to observed statistics. This form of adjustment is called “tuning” and is generally much less elaborate than the calibration process described above. 

Although it’s sometimes possible to improve a very limited number of model variables by testing a few coefficient values of one or other of the empirical relationships included in the physical parameterizations, the complexity and non-linear nature of the physical equations of the climate system make it impossible to force the model to reproduce either the statistics or the chronology of observations for all variables and at every point on the grid.