An assessment of scale-dependent variability and bias in global prediction models

The paper presents a method for the scale-dependent validation of the spatio-temporal variability in global weather or climate models and for their bias quantification in relation to dynamics. The method provides a relationship between the bias and simulated spatial and temporal variance by a model in comparison with verifying reanalysis data. For the low resolution (T30L8) subset of ERA-20C data, it was found that 80–90% (depending on season) of the global interannual variance is at planetary scales (zonal wavenumbers k = 0−3), and only about 1% of the variance is at scales with k > 7 . The reanalysis is used to validate a T30L8 GCM in two configurations, one with the prescribed sea-surface temperature (SST) and another using a slab ocean model. Although the model with the prescribed SST represents the average properties of surface fields well, the interannual variability is underestimated at all scales. Similar to variability, model bias is strongly scale dependent. Biases found in the experiment with the prescribed SST are largely increased in the experiment using a slab ocean, especially in k = 0 , in scales with missing variability and in seasons with poorly simulated energy distribution. The perfect model scenario (a comparison between the GCM coupled to a slab ocean vs. the same model with prescribed SSTs) shows that the representation of the ocean is not critical for synoptic to subsynoptic variability, but essential for capturing the planetary scales.