Model spread in multidecadal North Atlantic Oscillation variability connected to stratosphere–troposphere coupling

The underestimation in multidecadal variability in the wintertime North Atlantic Oscillation (NAO) by global climate models remains poorly understood. Understanding the origins of this weak NAO variability is important for making model projections more reliable. Past studies have linked the weak multidecadal NAO variability in models to an underestimated atmospheric response to the Atlantic Multidecadal Variability (AMV). We investigate historical simulations from Coupled Model Intercomparison Project Phase 6 (CMIP6) large-ensemble models and find that most of the models do not reproduce observed multidecadal NAO variability, as found in previous generations of climate models. We explore statistical relationships with physical drivers that may contribute to inter-model spread in NAO variability. There is a significant anticorrelation across models between the AMV–NAO coupling parameter and multidecadal NAO variability over the full historical period (, p<0.05). However, this relationship is relatively weak and becomes obscured when using a common period (1900–2010) and de-trending the data in a consistent way, with observations to enable a model–data comparison. This suggests that the representation of NAO–AMV coupling contributes to a modest proportion of inter-model spread in multidecadal NAO variability, although the importance of this process for model spread could be underestimated, given evidence of a systematically poor representation of the coupling in the models. We find a significant inter-model correlation between multidecadal NAO variability and multidecadal stratospheric polar vortex variability and a stratosphere–troposphere coupling parameter, which quantifies the relationship between stratospheric winds and the NAO. The models with the lowest NAO variance are associated with weaker polar vortex variability and a weaker stratosphere–troposphere coupling parameter. The two stratospheric indices are uncorrelated across models and together give a pooled R2 with an NAO variability of 0.7, which is larger than the fraction of inter-model spread related to the AMV (R2=0.3). The identification of this relationship suggests that modelled spread in multidecadal NAO variability has the potential to be reduced by improved knowledge of observed multidecadal stratospheric variability; however, observational records are currently too short to provide a robust constraint on these indices.