Evaluating Dynamic-Thermodynamic Coupling in Seasonal Forecasts: Linking Atmosphere Blocking to Mediterranean Heatwaves via Deep Learning

The Mediterranean basin is a highly vulnerable climate change hotspot where severe summer heatwaves are predominantly driven by persistent atmospheric blocking (e.g., Omega blocks) over the Euro-Atlantic sector. While Dynamical Seasonal Forecast Systems (SFS) are crucial for early warning, they frequently exhibit biases in maintaining these low-frequency blocking ridges, casting doubt on whether their temperature forecasts are dynamically consistent or merely the result of thermodynamic tuning.

In this study, we introduce a novel, physics-informed 3D Convolutional Neural Network (CNN) to evaluate the dynamic-thermodynamic coupling in SFS models. Unlike standard AI architectures, our model utilizes a Sequential "Macro-to-Micro" spatial funnel (scaling from 19x19 to 7x7 spatial kernels) combined with a Convolutional Block Attention Module (CBAM). This architecture forces the network to first isolate the planetary-scale stationary wave before analyzing embedded synoptic transient eddies, mimicking the causal fluid dynamics of blocking maintenance.

Trained using a self-adapting focal loss on normalized anomalies of ERA5 reanalysis data, the deep ensemble creates a highly robust, bias-free "AI Blocking Index." We apply this ERA5-trained ensemble directly to the seasonal hindcast anomalies of selected C3S models [ECMWF SEAS5 and CMCC]. By cross-referencing the AI-detected blocks within the SFS troposphere against the SFS lower-tropospheric thermodynamic forecasts (specifically the 850hPa-layer Temperature fields over the Mediterranean basin), we bypass surface-level boundary noise to quantify the pure internal consistency of the dynamical models. Ultimately, this framework is designed to highlight potential divergences between SFS air-mass temperatures and physical circulation, serving as an independent diagnostic tool to identify model drift and bias-correct seasonal extremes.

Acknowledgements: This research was supported by the PREVENT project that has received funding from the EU Horizon Europe framework programme (grant no. 101081276)