The impact of vertical velocity as a variable on predictability of convective events in data-driven weather prediction

Data-driven weather prediction is rapidly emerging as a transformative approach in modern forecasting. While these models often achieve improved traditional skill scores, such as root mean square error (RMSE) for near-surface variables like 2m temperature, they frequently exhibit overly smoothed spatial structures. This smoothing of spatial fields limits physical interpretability and reduces skill in representing localized extreme events, particularly convective phenomena associated with heavy precipitation. 

In this study, we investigate whether augmenting training datasets with adding vertical velocity (w) as a diagnostic variable can improve the representation and predictability of convective extremes. Since w represents convective updrafts, using the variable as a diagnostic can increase the similarity of data-driven models to the traditional, physics-based NWP models. Using the anemoi framework in a stretched-grid, limited-area configuration, we train models on ERA5 data and the Austrian regional re-analysis (ARA) at a horizontal grid spacing of 2.5km. Vertical velocity at multiple atmospheric levels is introduced as an additional diagnostic predictor. 

Model performance is evaluated using case studies of a strong summertime convective events, a known challenge for both physics-based and data-driven forecasting systems. We compare simulations from (i) a baseline data-driven model, (ii) the new  version including vertical velocity, and (iii) our physics-based numerical weather prediction system (CLAEF-AA). We will analyse whether incorporating vertical velocity improves the representation of convective structures and provides added value for extreme event prediction. We discuss implications for model design and outline future directions for integrating physically meaningful diagnostics into data-driven weather prediction systems.