Initializing a climate simulation
Physical climate models, like weather forecasting models, are based on fluid mechanics equations translated into computer code.
In order to produce a simulation that will show changes in the various climate variables over time, these models must know the characteristics of the land surface and the composition of the atmosphere, as well as the initial value (at the model’s time zero) of the various variables such as the soil water content, the snow cover on the ground, the air temperature and the wind. This is called model initialization.
Unlike weather forecasting models, climate models are not typically initialized using observations.
To explain this important characteristic, let’s do a more detailed comparison with their cousins, the weather forecasting models.
In order to predict weather events for the next few days as accurately as possible, weather forecasting models need to be initialized with a very accurate picture of the observed atmosphere (and even the ocean), which is called the analysis. The analysis provides initial values that will allow the fluid mechanics equations to predict the actual sequence of weather events for a few days. Beyond this time, the “predictive” aspect of the model is no longer possible due to the combined effect of the chaotic nature of the equations and the inaccuracies arising from both the observations and the model.
For this reason, a climate model, which far exceeds this time frame of a few days when used to simulate the recent past over several decades (e.g. the period 1960-2014 chosen for CORDEX-CMIP6), cannot reproduce the exact sequence of observed meteorological events. However, the various statistical moments of the climate variables simulated by the model are expected to be similar to those derived from the observed climate. This is no longer a forecast, but rather a climate simulation. To produce such a simulation, the climate model does not need to know exactly what the states of the atmosphere and ocean were at the initial time. This means that there is no need to initialize it with observations.
Since the equations of the climate model still require initial values, what do we use to initialize a climate simulation?
Although unnecessary, observations can still be used, but very different values, such as null values or average values from another climate simulation, will do just as well. All of these alternatives produce essentially the same climate.
Once information about the Earth’s surface and the composition of its atmosphere is provided to the climate model and initial values are assigned to the various variables, the model begins calculations to produce a sequence of virtual weather conditions spanning several decades. When we look at these time series, we see that the climate model takes a certain amount of time (but we’re talking about model time, not real time here) to gradually produce values similar to those that could be found in reality.
This transition period, the time taken to reach a state of statistical equilibrium, is called the spin-up. It is typically a few days for the atmosphere, while it can reach several years for the soil and even several hundred or thousands of years for the ocean and parts of the cryosphere. When the spin-up period is over, the various statistics in the climate simulation stabilize. The model then reaches its own equilibrium and its own climate.
If we were to repeat the exercise with other initial values, we would notice that once the spin-up was completed, the simulation would be different but would more or less converge toward the same climate. For long simulations, the choice of initial values therefore has almost no impact on the results of the climate model.
On the other hand, the use of realistic initial values is advantageous because it can considerably reduce the spin-up time by initializing the model as close as possible to its equilibrium state. This can be done with both observed and simulated climate values. This is a way to substantially increase the useful portion of the climate simulation and to considerably reduce the computation time, because the simulated data that is part of the spin-up is always subtracted when analyzing the results of a simulation.
Advanced concept
The predictability of the ocean can extend to a few years, compared to a few days for the atmosphere. This is what a new line of research called “decadal prediction” seeks to exploit. In order to predict certain climate variables that are strongly influenced by the ocean over several years, a climate model is initialized with very precise observations of the atmosphere, the ocean surface and even the deep ocean. Decadal prediction can be seen as a hybrid between weather forecasting and climate simulation. Decadal prediction is still in the experimental stage.
