Hailstorms represent one of the most damaging convective phenomena, with severe consequences for agriculture, infrastructure, and society. Their prediction remains challenging due to the multiscale nature of the processes involved, from mesoscale dynamics to microphysical growth mechanisms. In this study, Artificial Intelligence (AI) is employed to identify atmospheric conditions favorable to hail over the Continental United States (CONUS). Vertical atmospheric profiles from radiosonde soundings serve as predictors, while hail occurrences are derived from NOAA ground-based observations, allowing the problem to be framed as a binary classification (hail vs. no-hail).
An Extreme Gradient Boosting (XGBoost) model was trained using a balanced strategy to mitigate class imbalance and tested on the original unbalanced dataset. The model exhibited robust performance, successfully discriminating hail-producing environments. To assess the robustness and transferability of the approach, a comparison was performed between results obtained using real radiosonde data and synthetic soundings extracted from the GFS model. The analysis highlights a trade-off between model skill and data availability, suggesting opportunities for application in regions with limited sounding coverage, such as Italy.
Finally, SHAP-based interpretability analysis confirmed the physical consistency of key predictors and offered new insights into their role in hail formation.

Accurate representation of atmospheric dynamics at convection scale remains a major challenge for numerical models and a key factor in operational weather predictions. Reliable initial conditions, generated through data assimilation using observations from multiple platforms, are essential to improve forecasts in deep convection environments. In this study, the ICOsahedral Non-hydrostatic (ICON) model is applied at convection-permitting scale over Italy, following the operational configuration of Arpae Emilia-Romagna and adopted by the ItaliaMeteo Agency. ICON is coupled with the Local Ensemble Transform Kalman Filter (LETKF) within the Kilometre-scale Ensemble Data Assimilation (KENDA) system. Focusing on a poorly predicted extreme convective storm in the Marche region, we show the positive impact of convection-scale data assimilation based on conventional and radar observations. Nevertheless, precipitation remains substantially underestimated when relying solely on these datasets. We highlight the crucial role of low-level moisture convergence in convection initiation and the significant undersampling of humidity in conventional data. To address this, we investigate the added value of humidity-sensitive microwave radiances from polar satellites, still rarely employed in limited-area models worldwide. Assimilation of clear-sky observations from the Microwave Humidity Sounder (MHS) leads to notable improvements in precipitation forecasts. To extend further the rich multi-platform dataset, infrared all-sky radiances in water vapor absorption channels from the geostationary Meteosat Second Generation SEVIRI instrument are integrated, providing higher spatial and temporal resolution. The relative contributions of these observation types are analyzed, including their positive effects on surface and upper-level variables and convective indices. This study also supports the future operational assimilation of satellite radiances in the Arpae and ItaliaMeteo system.

Extreme precipitation events represent a growing global challenge, particularly affecting the Mediterranean basin, where their frequency and intensity continue to rise. Accurate numerical weather prediction models are crucial to effectively forecast these events; however, operational models often struggle with capturing the precise magnitude and location of intense convective storms. This limitation mainly arises from coarse model resolution and reliance on convection parameterizations. This study explores the potential benefits of hectometric-scale numerical modelling (500-meter resolution) by employing the ICON model to investigate the devastating floods that impacted Italy’s Marche region in September 2022 and another hail-driven case study affecting the Black forest and the Swabian-Jura region in Germany in August 2023. Through sensitivity tests and ensemble simulations, we demonstrate substantial improvements in the representation of extreme precipitation events. Our results show that hectometric-scale simulations, combined with explicit convection and the Smagorinsky turbulence scheme, significantly enhance the accuracy and realism of precipitation forecasts compared to traditional coarser-resolution operational setups. Further, a focus on perturbations of microphysical cloud parameters is analysed. The study highlights the crucial influence of initial and boundary conditions and choice of microphysics schemes on forecast quality, with specific ensemble members clearly outperforming others due to their accurate representation of critical atmospheric features, such as wind convergence and moisture transport. These findings underline the necessity and added value of very high-resolution ensemble modelling approaches, coupled with advanced turbulence parameterizations, for improving the predictability of extreme convective rainfall and hail.

Negli ultimi decenni, il cambiamento climatico ha portato a un aumento significativo della frequenza e dell’intensità di eventi meteorologici estremi, come forti precipitazioni e piene improvvise, con conseguente incremento del rischio idrogeologico e della vulnerabilità di ecosistemi e infrastrutture urbane. Tali fenomeni, caratterizzati da forte variabilità spaziale e temporale, risultano particolarmente impattanti nelle aree urbane, dove la copertura impermeabile del suolo, l’alta densità abitativa e la presenza di infrastrutture critiche amplificano le conseguenze di allagamenti ed esondazioni.

Il sistema idraulico di Milano rappresenta un caso emblematico: corsi d’acqua naturali e canali artificiali si intrecciano strettamente con il tessuto urbano. In particolare, le piene del Fiume Seveso causano allagamenti ricorrenti nel quartiere Niguarda, a nord della città, provocando danni diffusi a persone, infrastrutture e mobilità.

In questo scenario, la capacità di prevedere con precisione variabili meteorologiche e idrologiche a brevissimo termine risulta fondamentale per gestire il rischio e sviluppare sistemi di allerta tempestivi. Questo studio propone l’impiego di modelli di machine learning, come LDCast e GPTCast, sviluppati dalla Fondazione Bruno Kessler di Trento, per la previsione radar in chiave nowcasting. Le stime prodotte da questi modelli vengono poi utilizzate sia come input per modelli idrologici fisicamente basati sia in algoritmi di intelligenza artificiale sviluppati dal Politecnico di Milano.

L’obiettivo dello studio è valutare le prestazioni complessive del sistema previsionale e dimostrare come esso possa rappresentare un importante passo avanti nell’implementazione di sistemi di allerta a brevissimo termine.