Isolating the effects of model uncertainty on ensemble reanalysis data and their relation to North Atlantic flow regimes

    The level of uncertainty of reanalysis datasets varies greatly based on the quality and amount of available observations and the uncertainty of physical parameterizations used in the background forecast model. Ensemble data assimilation (EDA) schemes are used to quantifiy this combined uncertainty. However, isolating the effects of observational and model uncertainties based on a given ensemble reanalysis is not straightforward. Here, we use the 9 member EDA ensemble produced for the ECMWF 5th Generation Reanalysis product (ERA5) to investigate synoptic scale model uncertainty and its connection to the occurrence of specific weather regimes.

    To control for ensemble spread caused by observation uncertainty - especially on long time scales - we devise grid-point-wise statistical models for the logarithmic ensemble variance with temporal predictors. We use a binary segmentation algorithm to objectively identify change points in ensemble spread time-series caused by abrupt changes in the observation system.

    The set of statistical models allows for statements about the relative impact of changes in the observation system on the total background forecast uncertainty between different grid-points. After filtering out the impact of changes in the observation uncertainty, we obtain a long time series of model uncertainty estimates, which we analyze climatologically with respect to flow characteristics, regime structure and impact of physical parameterizations. This provides regions of high model uncertainty for the respective regimes as well as differences in the role of model uncertainty among the regimes.