Analysing and combining atmospheric general circulation model simulations forced by prescribed SST: tropical response

The ECHAM 3.2 (T21), ECHAM 4 (T30) and LMD (version 6, grid-point resolution with 96 longitudes × 72 latitudes) atmospheric general circulation models were integrated through the period 1961 to 1993 forced with the same observed Sea Surface Temperatures (SSTs) as compiled at the Hadley Centre. Three runs were made for each model starting from different initial conditions. The large-scale tropical inter-annual variability is analysed to give a picture of the skill of each model and of some sort of combination of the three models. To analyse the similarity of model response averaged over the same key regions, several widely-used indices are calculated: Southern Oscillation Index (SOI), large-scale wind shear indices of the boreal summer monsoon in Asia and West Africa and rainfall indices for NE Brazil, Sahel and India. Even for the indices where internal noise is large, some years are consistent amongst all the runs, suggesting inter-annual variability of the strength of SST forcing. Averaging the ensemble mean of the three models (the super-ensemble mean) yields improved skill. When each run is weighted according to its skill, taking three runs from different models instead of three runs of the same model improves the mean skill. There is also some indication that one run of a given model could be better than another, suggesting that persistent anomalies could change its sensitivity to SST. The index approach lacks flexibility to assess whether a model’s response to SST has been geographically displaced. We focus on the first mode in the global tropics, found through singular value decomposition analysis, which is clearly related to El Niño/Southern Oscillation (ENSO) in all seasons. The Observed-Model and Model-Model analyses lead to almost the same patterns, suggesting that the dominant pattern of model response is also the most skilful mode. Seasonal modulation of both skill and spatial patterns (both model and observed) clearly exists with highest skill (between tropical Pacific SST and tropical rainfall) and reproducibility amongst the runs in December-February, and least skill/reproducibility in March-May and June-August. The differences between each model suggest that a simple linear regression combination of each GCM’s prediction indices will be improved upon by combination methods that take account of the errors in the spatial teleconnection structures generated by the GCM.