Item

Abstract

Recurrent large-scale atmospheric circulation patterns, or teleconnections, exert a prominent effect on the Euro-Atlantic surface climate. In summer, teleconnections are amongst the main drivers for high-impact climatic processes such as heatwaves, and hence several relevant socio-economic sectors could benefit from their credible seasonal prediction. However, dynamical climate models show limited capability to reproduce summer teleconnections. This problem is further compounded by the complex physical mechanisms influencing their predictability, which are still not well understood. While conventional statistical tools offer only a limited assessment of these physical mechanisms, artificial intelligence (AI) outperforms these tools, learning complex relationships from data and thereby advancing physical understanding. Here, I promote the combination of observations and dynamical climate modelling with AI to overcome some of these limitations and to achieve improved predictions of European summer climate a season ahead. I implement this novel AI-dynamical approach in two complementary steps: I first refine the assessment of summer teleconnections in observations and a model, and then I apply this knowledge to improve Euro-Atlantic summer seasonal climate predictions. I use the AI classifier Self-Organising Maps (SOM) to characterise the observed and modelled variability of the two dominant Euro-Atlantic summer teleconnections in the 20th century: the summer North Atlantic Oscillation (NAO) and summer East Atlantic Pattern (EA). I find that while the ensemble dynamical prediction system can reproduce summer NAO and EA spatial features, it shows limited model performance in reproducing their frequency of occurrence. I use SOM to illustrate that the seasonal predictability of summer teleconnections is associated with North Atlantic sea surface temperatures (SST), however this influence varies in intensity with time and is more relevant for summer EA than for summer NAO. I go beyond standard forecast practices by applying these SST predictors to constrain the credibility of summer climate predictions in a dynamical ensemble prediction system. I show that, particularly for years during which EA dominates, summer climate predictions up to 4 months ahead can be significantly improved in parts of Europe using these SST predictors. With the use of an AI causal inference tool I find that although extratropical NA SST in spring show a causal link with EA in the second half of the 20th century in observations, the evaluated dynamical ensemble prediction shows limited performance to reproduce this causal link. However, I do find that those ensemble simulations that reproduce this causal link show improved surface climate prediction credibility over those that do not. Overall, my findings promote the use of combined AI-dynamical approach to improve seasonal predictions and could even be applied operationally, benefiting actual seasonal predictions of European summer climate.