Deep learning identification of SST teleconnections driving early-winter North Atlantic climate

Víctor Galván Fraile , Irene Polo, Marta Martín-Rey, Belén Rodríguez-Fonseca, Magdalena A. Balmaseda

Seasonal predictability over the North Atlantic-European (NAE) sector is strongly modulated by the background climate state, particularly in early winter. In this season, different ENSO teleconnections have been reported before and after the 1990s. However, these studies rely on linear analysis, and the reasons for this lack of stationarity and its implications for seasonal forecasting have not yet been fully assessed. Here, we evaluate the non-stationarity of early winter ENSO teleconnections over the NAE using a non-linear framework, the Neural Network forecast application (NN4CAST). The NN4CAST model predicts November-December sea level pressure anomalies from October sea surface temperature anomalies. The neural network achieves skill comparable to the ECMWF SEAS5 dynamical system, and even exceeds it in certain regions like the NAE, despite relying only on October SST predictors. The model detects a period with significant improvement in skill over the North Atlantic Oscillation centres of action during 1970-1990s, specially for the Iceland region, whereas for the East Atlanticn pattern, the improvement comes in recent decades, namely 1990-2020s. These changes in predictability might be driven by an intensification of the Rossby wave source over Southeast Asia as well as a southward displacement of the North Atlantic jet stream during the latter period. Our results highlight the value of explainable artificial intelligence for identifying state-dependent sources of seasonal predictability and for bridging data-driven approaches with physical understanding of climate teleconnections.