Climate and weather variability in the North Atlantic region is determined largely by the North Atlantic Oscillation (NAO). The potential for skillful seasonal forecasts of the winter NAO using an ensemble‐based dynamical prediction system has only recently been demonstrated. Here we show that the winter predictability can be significantly improved by refining a dynamical ensemble through subsampling. We enhance prediction skill of surface temperature, precipitation, and sea level pressure over essential parts of the Northern Hemisphere by retaining only the ensemble members whose NAO state is close to a “first guess” NAO prediction based on a statistical analysis of the initial autumn state of the ocean, sea ice, land, and stratosphere. The correlation coefficient between the reforecasted and observation‐based winter NAO is significantly increased from 0.49 to 0.83 over a reforecast period from 1982 to 2016, and from 0.42 to 0.86 for a forecast period from 2001 to 2017. Our novel approach represents a successful and robust alternative to further increasing the ensemble size, and potentially can be used in operational seasonal prediction systems.

The north Atlantic subpolar gyre (SPG) has been widely implicated as the source of large-scale changes in the subpolar marine environment. However, inconsistencies between indices of SPG-strength have raised questions about the active role SPG-strength and size play in determining water properties in the eastern subpolar North Atlantic (ENA). Here, by analyzing various SPG indices derived from observations and a global coupled model, we show that the choice of the SPG index dictates the interpretation of SPG strength-salinity relationship in the ENA. Variability in geostrophic currents derived from observed hydrography and model based Lagrangian trajectories reveal zonal shifts of advective pathways in the enA and meridional shifts in the western intergyre region. Such shifts in advective pathways are manifestations of variability in the size and strength of the SPG, and they impact salinity by modulating the proportion of subpolar and subtropical waters reaching the enA. SpG indices based on subsurface density and principal component analysis of sea surface height variability capture these shifts in advective pathways, and are therefore best suited to describe SPG-salinity relationship in the enA. our results establish the dynamical constraints on the choice of the SpG index and emphasize that SPG indices should be cautiously interpreted.

On interannual to decadal time scales, memory in the Earth's climate system resides to a large extent in the slowly varying heat content of the ocean, which responds to fast atmospheric variability and in turn sets the frame for large‐scale atmospheric circulation patterns. This large‐scale coupled atmosphere–ocean feedback is generally well represented in today's Earth system models. This may fundamentally change when data assimilation is used to bring such models close to an observed state to initialize interannual to decadal climate predictions. Here, we review how the large‐scale coupled atmosphere–ocean feedback is preserved in common approaches to construct such initial conditions, with the focus on the initialized ocean state. In a set of decadal prediction experiments, ranging from an initialization of atmospheric variability only to full‐field nudging of both atmosphere and ocean, we evaluate the variability and predictability of the Atlantic meridional overturning circulation, of the Atlantic multidecadal variability and North Atlantic subpolar gyre sea surface temperatures. We argue that the quality of initial conditions for decadal predictions should not purely be assessed by their closeness to observations, but also by the closeness of their respective predictions to observations. This prediction quality may depend on the representation of the simulated large‐scale atmosphere–ocean feedback. This article is categorized under: Climate Models and Modeling > Knowledge Generation with Models

Five initialization and ensemble generation methods are investigated with respect to their impact on the prediction skill of the German decadal prediction system "Mittelfristige Klimaprognose" (MiKlip). Among the tested methods, three tackle aspects of model-consistent initialization using the ensemble Kalman filter, the filtered anomaly initialization, and the initialization method by partially coupled spin-up (MODINI). The remaining two methods alter the ensemble generation: the ensemble dispersion filter corrects each ensemble member with the ensemble mean during model integration. And the bred vectors perturb the climate state using the fastest growing modes. The new methods are compared against the latest MiKlip system in the low-resolution configuration (Preop-LR), which uses lagging the climate state by a few days for ensemble generation and nudging toward ocean and atmosphere reanalyses for initialization. Results show that the tested methods provide an added value for the prediction skill as compared to Preop-LR in that they improve prediction skill over the eastern and central Pacific and different regions in the North Atlantic Ocean. In this respect, the ensemble Kalman filter and filtered anomaly initialization show the most distinct improvements over Preop-LR for surface temperatures and upper ocean heat content, followed by the bred vectors, the ensemble dispersion filter, and MODINI. However, no single method exists that is superior to the others with respect to all metrics considered. In particular, all methods affect the Atlantic Meridional Overturning Circulation in different ways, both with respect to the basin-wide long-term mean and variability and with respect to the temporal evolution at the 26 • N latitude.

We analyze the time dependency of decadal hindcast skill in the North Atlantic subpolar gyre within the time period 1961–2013. We compare anomaly correlation coefficients and temporal interquartile ranges of total upper ocean heat content and sea surface temperature for three differently initialized sets of hindcast simulations with the global coupled model MPI-ESM. All initializations use weakly coupled assimilation with the same full value nudging in the atmospheric component and different assimilation techniques for oceanic temperature and salinity: (1) ensemble Kalman filter assimilating EN4 observations and HadISST data, (2) nudging of anomalies to ORAS4 reanalysis, (3) nudging of full values to ORAS4 reanalysis. We find that hindcast skill depends strongly on the evaluation time period, with higher hindcast skill during strong multiyear trends, especially during the warming in the 1990s and lower hindcast skill in the absence of such trends. Differences between the prediction systems are more pronounced when investigating any 20-year subperiod within the entire hindcast period. In the ensemble Kalman filter initialized hindcasts, we find significant correlation skill for up to 5–8 lead years, albeit along with an overestimation of the temporal interquartile range. In the hindcasts initialized by anomaly nudging, significant correlation skill for lead years greater than two is only found in the 1980s and 1990s. In the hindcasts initialized by full value nudging, correlation skill is consistently lower than in the hindcasts initialized by anomaly nudging in the first lead years with re-emerging skill thereafter. The Atlantic meridional overturning circulation reacts on the density changes introduced by oceanic nudging, this limits the predictability in the subpolar gyre in the first lead years. Overall, we find that a model-consistent assimilation technique can improve hindcast skill. Further, the evaluation of 20 year subperiods within the full hindcast period provides essential insights to judge the success of both the assimilation and the subsequent hindcast quality