The science of climate modelling

Dr Michael Davey divides his time between UCL Mathematics and the Met Office’s Hadley Centre, the UK’s foremost climate change research centre. His research aims to link the theoretical and practical aspects of long-range climate prediction.

Here Dr Davey explains how meteorologists use a range of strategies and models as they attempt to understand what will happen to our climate in the future.

Satellite picture of a tropical storm

“The climate predictions described in the most recent series of Intergovernmental Panel on Climate Change reports (IPCC) draw on information from a wide variety of climate models. The most comprehensive ones are known as Coupled General Circulation Models (CGCMs), and 23 such models were used for the IPCC Fourth Assessment Report (AR4).

Each model is a particular configuration of equations that represent the physical processes and fluid motions of the climate system – necessarily an approximation to the real world – that are then converted into numerical models for use on computers. There is no perfect or ideal configuration, and the various centres that develop such models make differing choices for their models. The models are assessed by testing their ability to represent present-day and past climate, and then used to produce projections for future climate under various standard scenarios. (The scenarios themselves require differing assumptions about greenhouse gas emissions depending on forecasts of economic growth and energy sources, but that is another story.)

For a given scenario, each separate model produces a somewhat different outlook. The results provide samples, albeit limited in number, of possible future outcomes. This is an example of the ‘multi-model’ approach to making predictions, which is being increasingly used in meteorology. The individual contributions are members of an ‘ensemble’ of values. From these values typically a mean value (‘ensemble mean’) is calculated, along with some statistical measure of the spread to indicate the scatter. For some purposes, such as estimation of the likelihood of relatively high or low values, further statistical analysis is carried out.

Effectively the members of the ensemble are used to estimate properties of a probability distribution. Such probability distributions apply to many variables: for example, mean winter temperature in London presently has a probability distribution known from observations accumulated over recent decades. That distribution will shift as climate alters, and for practical applications (e.g. energy consumption, health service management) it is important to estimate such changes.

Within an individual CGCM the representation of physical processes also requires the selection of a number of parameters whose value has some degree of uncertainty. This could be due to the nature of the process or limitations in its measurement. Thus it is useful to extend the ensemble sample by including members for which several combinations of values of key parameters are employed. Such an approach was included in the Met Office contributions to AR4, and the methodology is now widely used. An added benefit is that the sensitivity of climate to such parameters can be assessed. This parameter perturbation approach has been (and is being) used in a project called ‘climateprediction.net’ by which modified GCMs can be run on personal computers to produce data from thousands of ensemble members for analysis.

The ensemble concept also applies to single models. The evolution of any forecast depends on its starting point, and the chaotic nature of the weather ensures that in GCMs small differences (determined, for example, by the accuracy of observational instruments and their coverage) grow and influence the variability on timescales from hours to decades. For climate prediction – that is, the prediction of climate statistics averaged over a few decades – the precise starting point is not so important, as models with quite substantial initial differences typically settle into states with similar climate statistics. However the effect is not negligible and it is wise to make climate predictions with the same model and several different initial conditions.

For predictions on shorter range timescales, days to seasons ahead, for which the initial state is vital, the ensemble strategy is likewise used to sample future conditions, typically using several 10s of members. In this case the effectiveness of the strategy can be assessed directly by making retrospective forecasts extending over recent decades and checking the results against observed behaviour. Over time it has been conclusively demonstrated that single and multi-model ensembles add skill and value to forecasts.

A large international project, appropriately called ENSEMBLES, funded by the European Commission, is coming to an end. In this project state-of-the-art CGCMs from several European partners have been used to make and assess predictions on timescales from seasons to decades ahead, using the multi-model ensemble strategy. The results, which include use of the data in applications such as agriculture and health, are now being analysed.  

The ensembles strategy does pose a problem to forecasters with finite computational resources. Ever more sophisticated GCMs that resolve increasing geographic detail require correspondingly increased computer power to run each sample of the future. Thus there is a competition between model complexity, model varieties and ensemble sizes, but what is the ideal balance of these factors? Don't put all your eggs in one basket, but just how many eggs and how many baskets are needed? Undoubtedly the answer depends on the range of the forecast (next month? next decade?), and probably also on the quantity that is to be predicted!”

Image: Satellite picture of a tropical storm