The most common current method to explore downscaled predictability with climate models is to use the so-called "Regional Climate Models". These are similar to GCMs, but are tuned to resolve the smaller scale climate processes. They are implemented for a given geographical region - as was illustrated for West Africa in Fig. 3.4. A forecast from a GCM is selected, of the type described in the earlier lectures. The winds and other predicted features from the GCM are used to drive the RCM to yield higher resolution atmospheric solutions. The RCM feels the large scale forcing of the GCM each day during the forecast period - and the RCM's physical and dynamical equations then find the solution of the smaller scale circulations that are forced by the GCM's large-scale forecast. The RCMs are a good practical choice to investigate dynamical downscaling relative to global models. Large sets of forecasts can be run at resolutions down to 20km or less on Personal Computer platforms, whereas global models at such resolutions require the world's most powerful computer resources. An important consideration is that the GCM, over a large set of past years, demonstrates clear skill in the large-scale circulations - for example, as demonstrated for East Africa in lecture 2. If the GCM has on average no skill in predicting the large-scale circulations, then the RCM cannot be expected to create any skill at smaller scales - garbage in garbage out.

Implementing RCMs for a region of interest requires considerable expertise and experience. Experience has shown that having the RCM boundary in a region of climatological strong convection or complex orography leads to inferior model performance. Results are generally found to be more robust for regions away from the RCM's geographical boundaries, so this is a consideration when choosing a domain to study the climate of a particular region. There are a number of specific designs for nesting the RCM inside the GCM. Two distinctions are between a so-called 'perturbation method' and a boundary forcing method. In the boundary forcing method, the RCM only receives information from the GCM at its geographical boundaries. In the perturbation method, the RCM receives the large-scale information from the GCM across its whole domain, the RCM 'perturbs' the large-scale circulations, generating the samller scale climate features. Fig. 4.5 gives an example of a perturbation method RCM running for two days. You can see how the model received the large scale information from the GCM every 6 hours, and then is allowed to forecast the development of the small scale climate features across the model domain stepping forward in short time steps, until after 6 hours, it receives another shot of the large-scale information from the GCM. The RCM then, once again, forecasts the next six hours until the next input from the GCM. Once the model is implemented, the researcher now has the challenge of tuning the model's schemes to represent the key processes in the region as best as possible. One example is different schemes for creating convection and clouds. Another challenge for the researcher is the resolution at which to run the model. Higher resolution does not necessarily mean better performance, and this is another feature for the researcher to use their experience in testing different configurations before arriving at one that seems robust and justifies making a large set of experiements to explore and evaluate predictability.

One aspect of climate modelling that was not emphasized in lecture 2 was the representation of the land surface. However, for downscaling, the land surface representation in models becomes particularly critical. More advance land models are now being coupled to RCMs, even allowing some degree of vegetation development and feedback between the evolving climate and the biosphere. The exchange of heat, moisture and momentum between the land surface and the atmosphere can be critical for local scale circulations. As with the possibility of orography inducing sub-regional scale rainfall anomlies, so the potential for land surface gradients, or land-lake contrasts also exists (e.g. Fig. 4.6).

A set of RCM simulations made for East Africa for 1970-95 clearly shows the improvement in ability to at least capture the climatological pattern of rainfall in East Africa (Fig. 4.7). This improvement could lead to better simulation of the large-scale processes themselves, especially when they involve tight gradients, such as the north-south transition from wet to dry in the Sahel region of West Africa, and this itself could enhance prediction skill. For the new type of downscaled forecast information, a more relevant question is whether this improved ability to capture the climatological mean rainfall translate into an ability to see processes like those described in Fig. 4.2? This is still a research question, but indications are that at least some benefit is gained, and examples of processes like those described in Fig. 4.2 are emerging. For example, Fig. 4.8 shows how an RCM has simulated a pattern that seems consistent with the type of process suggested in Fig.4.2. Here, anomalous southwesterly winds predicted across Taiwan for the El Niño year (relative to the other case study year), lead to generally wetter conditions across Taiwan for the El Niño year, but with a pocket of relatively drier conditions simulated in the northeastern parts of the island, to the lee of the mountains. In addition, there are examples of improved ability to simulate the year-to-year variations in daily rainfall distributions using RCMs. For the West Africa example, the forecast for the dry year 1983 contains more realistic dry spell structure through the season (see Fig. 4.4).

To make a comprehensive evaluation of the skill of a RCM as a downscaling forecast tool requires a large suite of experiments. Reliable skill evaluations are difficult if less than about 30 years are available for analysis. Furthermore, there is a need to incorporate the probabilistic nature of seasonal prediction - one way to do this is to drive the RCM with a number of the ensemble members from the GCM. For Northeastern Brazil, such a set of experiments were undertaken as a collaboration between FUNCEME (Ceara State, Brazil) and IRI. Ten ensemble members from the ECHAM4.5 GCM were used to drive a regional climate model (the NCEP Regional Spectral Model). The correlation skill between the ensemble mean rainfall and observed rainfall is shown in Fig. 4.9. A substantial region has a correlation skill score greater than 0.7. It is important to appreciate that the skill achieved by the GCM at its scale (250km resolution) is here of similar magnitude (i.e. correlation skill greater than 0.7). The potential of the RCM to enhance skill above a simple statistical transformation of the GCM prediction is a separate question and one that requires careful statistical treatment. It is often easier to demonstrate improvement in the raw output RCM forecast information in terms of ability to predict the statistics of weather in the season. For example, Fig 4.10 evaluates the ability of the RCM to predict aspects of the daily rainfall statistics for the season, such as the number of dry days and the number of dry spells in each rainy season. The results are a demonstration that in some situations, the direct output of the model may be adequate for information about such features. However, further improvements will often be possible through some form of statistical correction.