AI reconstruction of European temperature and precipitation anomalies from Euro-Atlantic weather regimes

Euro-Atlantic weather regimes (WRs) act as critical drivers of European weather variability and extreme events. While previous research extensively details the correlational impacts of these quasi-stationary large-scale circulation patterns, explicitly predicting ground-level meteorological variables, such as surface temperature and precipitation, directly from WR indices remains an underexplored challenge.

In this contribution, we introduce an AI model that maps Euro-Atlantic WR indices to monthly anomalies in European 2-meter temperature and precipitation, explicitly bridging the circulation–surface link at seasonal time scales. Drawing on ERA5 reanalysis data (1940–2024), we derive seven year-round and four seasonal (DJF/JJA) WRs from 500 hPa geopotential height (Z500) fields using EOF analysis and k-means clustering. A residual neural network then processes these monthly WR indices alongside calendar data to reconstruct the corresponding surface anomaly fields across Europe.

Our model demonstrates high anomaly correlation and low mean absolute error, particularly during the winter months, and significantly outperforms traditional linear WR-composite reconstructions. Furthermore, when driven by WR indices predicted by the bias-corrected ECMWF SEAS5, the AI framework matches or exceeds SEAS5's ensemble mean across most deterministc metrics.

A critical question for operational implementation is the minimum accuracy required in the predicted WR indices for our AI framework to outperform baseline dynamical forecasts. To address this, we perform a sensitivity analysis by incrementally introducing artificial error into the input WR indices, simulating varying levels of forecast skill. By tracking how the AI model's output quality declines as the inputs worsen, we pinpoint the minimum WR forecast accuracy required for our framework to outperform direct ECMWF SEAS5 predictions of European winter and summer anomalies.

Ultimately, our findings reveal that a substantial portion of the spatial structure within European monthly anomalies can be directly inferred from the low-frequency Euro-Atlantic regime state. This research motivates the broader application of AI and hybrid methodologies to enhance regime predictability, ultimately advancing sub-seasonal to seasonal (S2S) forecasts of European weather and associated climate risks.