The performance of 61 global climate models participating in CMIP5 and 6 is evaluated for the Southern Hemisphere extratropics in terms of typical regional-scale atmospheric circulation patterns. These patterns are known to be linked with a number of key variables in atmospheric physics and chemistry and provide an overarching concept for model evaluation. First, hemispheric-wide error and ranking maps are provided for each model and regional details are described. Then, the results are compared with those obtained in a companion study for the Northern Hemisphere. For most models, the average error magnitude and ranking position is similar on both hemispheres, ruling out systematic tuning towards either of the two. CMIP6 models perform better on average than CMIP5 models and the interactive simulation of more climate system components does not deteriorate the results for most model families. Better performance is associated with higher resolution in the atmosphere, following a non-linear relationship.

The ability of global climate models to reproduce recurrent regional atmospheric circulation types is introduced as an overarching concept to explore potential dependencies between these models. If this approach is applied on a sufficiently large spatial domain, the similarity of the resulting error pattern can be compared from one model to another. By computing a pattern correlation matrix for a large multi-model ensemble from the Coupled Model Intercomparison Project Phases 5 and 6, groups of comparatively strong correlation coefficients are obtained for those models working with similar atmospheric components. Thereby, frequent shared error patterns are found within in the ensemble, which also occur for nominally different atmospheric component models. The error pattern correlation coefficients describing these similarities are nearly unrelated to model performance and can be used as statistical dependency weights.

The ability of global climate models to reproduce recurrent regional atmospheric circulation types is introduced as an overarching concept to explore potential dependencies between these models. If this approach is applied on a sufficiently large spatial domain, the similarity of the resulting error pattern can be compared from one model to another. By computing a pattern correlation matrix for a large multi-model ensemble from the Coupled Model Intercomparison Project, groups of comparatively strong correlation coefficients are obtained for those models working with similar atmospheric components. Thereby, frequent shared error patterns are found within in the ensemble, which also occur for nominally different atmospheric component models.