(ii) Atmospheric GCMs forced with Forecast SST

This class of method is used by many numerical modeling centers making operational seasonal climate forecasts. The atmospheric GCMs have been demonstrated to skillfully simulate interannual climate variations (most widely verified are rainfall and temperature) in many tropical and some mid-latitude regions when they are forced with the observed time-varying SST (as we explored in lecture 2). This finding leads to a so-called "two-tiered" forecast method whereby forecasts are made of the expected SST fields, and these SST fields are then used to force the GCM through the coming season to generate the seasonal climate forecast (e.g. Fig. 3.3 shows a schematic of the forecast system at IRI as of 2002). Ensembles of forecasts attempt to capture the range of possible outcomes. The ensembles are generated by slightly altering the initial atmospheric starting point of each forecast.

The SST forecasts can be generated in a variety of ways. For the tropical Pacific, coupled models can be used with moderate to good skill as described in the previous section. For the Atlantic and Indian Ocean, statistical methods are generally considered to be the best that can be done at this stage, and for lead times of a few months, the best statistical method is often some form of persistence of the prevailing SST anomalies. The IRI forecast system (as of 2002) runs forecasts with atmospheric GCMs that are driven by two SST anomaly scenarios - (i) based on persistence everywhere, (ii) a forecast of tropical Pacific SST from a coupled model, and statistical forecasts for tropical Atlantic and Indian Oceans. The ensemble of predictions from the atmospheric GCMs attempt to provide an indication of the range of seasonal climate patterns that are possible given the prescribed SST - that is, to provide a basis for making probability estimates of particular climate outcomes, analogous to making a set of ball drops and determining the probability that a ball will drop into the left or right hand divide (Fig. 2.4). One limitation of the current system is that the possible range of SST to expect is not explicitly considered, and this may lead to underestimates of the uncertainty (the analogy with the ball drop would be concern over whether the wind that operated during the first 10 drops, would continue unchanged so that we can use the sample of 10 to estimate the likelihood of the 11th ball drop being pushed into the right hand divide by the wind).

While atmospheric GCMs have been quite widely verified in their ability to respond to the observed SST, their performance in forecast mode when they are using forecast SST has been less quantified. Thus, so far, in using the output from these methods, it has usually been necessary to use judgment in assessing the level of skill reduction expected from the lack of certainty regarding the SST fields during the forecast period. On the other hand, the demonstration that a GCM accurately responds to observed time varying SST is a very large step in demonstrating that the GCM has a strong scientific basis as a seasonal forecast tool and can be expected to contain skilful forecast information at least at short lead times of a month or season. Information on the skill of models when forced with forecast SST is now being more widely generated.

This two-tiered approach to seasonal prediction may in some circumstances actually underestimate the potential accuracy of seasonal forecasts based on the coupling of the ocean and the atmosphere. This underestimation may be quite substantial in some regions and seasons. The reason is the lack of opportunity for the atmosphere to influence the ocean as the model steps forward in time to make the forecast. This is a problem because in some situations in the real climate system, it is the atmosphere that is primarily driving the ocean. However, applying this two-tiered modeling approach to seasonal prediction, there is only opportunity for the ocean to drive the atmosphere. Consider an ocean region where during a given season, the SST anomalies were created by atmospheric forcing (for example, the trade winds were particularly strong due to remote atmospheric forcing). An atmospheric GCM feels this SST anomaly and may respond quite strongly, but most likely in error, since the SST was actually generated by the observed atmospheric anomaly. The response in the model could damage the skill of the model both locally and, through teleconnection processes, in other regions as well. Global CGCMs can theoretically resolve this problem, but are hindered by lack of observations of the sub-surface ocean and their inherent extra complexity which still leaves many research challenges in model development. A promising intermediate solution is to run atmospheric GCMs coupled to a very simple representation of the upper layers of the ocean, to allow some feedback of the predicted atmospheric circulations onto the sea-surface temperature, and avoid unrealistic situations that damage model skill. This approach may yield noticeable advances in seasonal prediction skill for some regions and seasons.