Abstract/Summary

To describe the dynamics of the atmospheric circulation variability often the circulation is divided into a limited number of so-called atmospheric circulation regimes, which characterise the low-frequency variability in the dynamics. These recurrent and persistent circulation patterns can be identified in different regional domains and time periods, where the focus in this thesis is on the wintertime Euro-Atlantic sector. A central challenge here is to accurately capture the regime variability signal. To this end several novel methods are introduced and analysed in this study. Existing methods mostly identify four circulation regimes over the Euro-Atlantic sector. The common approach to regime identification is to apply a k-means clustering algorithm to principal component (PC) data, often after applying a low-pass time filter. We find that using gridpoint data instead of PC data gives an optimal number of regimes of six instead of four. Furthermore, a time-regularised clustering algorithm is proposed to identify the persistent regime dynamics. This regularised approach increases the persistence of the regimes compared to a standard k-means clustering algorithm, with the dynamics being less affected by noise. The use of a low-pass filter leads to a bias in the regime frequencies, while the regularised method does not. To study non-stationary regime dynamics signals on (sub-)seasonal and interannual timescales an ensemble-regularised k-means clustering algorithm is proposed. This approach couples the information within an ensemble of model hindcast data, allowing to identify a more pronounced non-stationary regime signal. On interannual timescales this signal is dominated by the El Ni˜no Southern Oscillation (ENSO) and it is found to be predictable on seasonal timescales for two of the six regimes considered. The two regularised clustering methods discussed require the selection of a constraint parameter, which can be made using for example information criteria. In the final part a Bayesian approach to regime assignment, which does not require such a parameter selection, is proposed. This method uses Bayes theorem to incorporate prior information to obtain a better informed probabilistic regime assignment. It leads to more persistent regime dynamics, even compared to the time-regularised clustering method, and helps to identify a pronounced interannual variability signal comparable to that obtained using the ensemble-regularised clustering method. The three introduced and investigated approaches help in identifying robust regime variability signals, significantly improving on existing methods. These techniques can be adapted to suit different regime variability signal questions and are not limited to circulation regimes in terms of applicability. The identified regime signals raise questions for future research directions.