Neural interpretation of European summer climate ensemble predictions

Abstract

Current state-of-the-art dynamical seasonal prediction systems still show limited skill, particularly over Europe in summer. To circumvent this, we propose a neural network-based classification of individual ensemble members at the initialisation of summer climate predictions, prior to performing a skill analysis. Different from European winter climate, largely dominated by the North Atlantic Oscillation, predictability of European summer climate has been associated with several physical mechanisms, including teleconnections with the tropics. Recent studies have shown that predictive skill improves when the dominant physical processes in a given season are identified at the initialisation of a prediction. Each of these dominant physical processes is associated with large-scale circulation patterns, often depicted by modes of Empirical Orthogonal Functions (EOF). We argue that Self-Organising Maps (SOM), a type of neural network classifier, can provide further insight on interpreting the predictive skill of mixed resolution hindcast ensemble simulations generated by MPI-ESM. This is achieved by identifying which circulation patterns over the North Atlantic-European sector (NAE) at the initialisation of hindcasts lead to more predictable states than others, their preferable transition states, and whether the spatial structure of each SOM mode play a role in shaping climate over Europe. We train SOM networks on sea level pressure fields of ERA-20C reanalysis at the initialisation of the seasonal prediction system (every May) for the period of 1900-2010, covering NAE. We compare the SOM-derived modes with circulation patterns derived from EOF analysis, and characterise each class of circulation regime. This analysis is used to distinguish classes of predictions with two different sets of MPI-ESM initialised simulations with 10 and 30 members, covering the period of 1902-2008 and 1982-2016, respectively. We then discuss the differences and advantages of performing a neural interpretation of the skill of an ensemble prediction, over traditional skill analysis.