CMIP6 simulations overestimate historical decadal temperature variability over most land areas

A robust understanding of the potential range of Earth system dynamics is essential for effectively simulating future climate change. Previous studies have reported increasing discrepancies in modelled temperature variability from global to local scale, and beyond decadal timescales, based on paleoclimate reconstructions. The instrumental record is most complete for the last 145 years. This limits a spatio-temporal assessment of historical temperature variability to multidecadal timescales at the upper end.  To this day, model-observation comparisons of regional climate variability have mostly focused on sea surface temperature. 

Here, we compare historical near-surface air temperatures from an ensemble of 50 CMIP6 models with similar initial conditions and two single-model initial-condition large ensembles (SMILE) with reanalysis and observation datasets. Following a robust like-for-like approach, all datasets are interpolated to a common grid of about 2.8 degrees and compared over the period of 1880 to 2015. Spectral analysis and filters reveal the structure of temperature variability over different spatial and temporal scales. Specifically, we focus on temperature variability on timescales of 10 to 30 years from global to local scale.  

On the global scale, models consistently display higher temperature variance in bands from 10 to 30 years than reanalysis data. Masking the analysis to regions with a consistent observational record confirms this trend. On the local scale, observed temperature variability can deviate substantially from the mean of stacked model standard deviation fields. For example, observed temperature variability in Europe lies in the lower tail of the model distribution. Vice versa, observed temperature in the southern Atlantic is representative of the model distributions' upper tail. Consistently over the multi-model ensemble and two SMILEs, decadal temperature variability is overestimated on land, but underestimated over the ocean. Nevertheless, there are exceptions to this pattern. For example, in the northern Atlantic, modelled variability overestimates observations consistent with the literature. Overall, these regional inconsistencies suggest that multiple, regionally heterogeneous processes are involved.