Nowcasting of Thunderstorm Hazards with Deep Learning: Performance Report of the First Convective Season in Operations

Thunderstorms pose serious risks to life and property through hazards such as lightning, heavy rainfall, hail, and strong winds. These events develop rapidly and affect localized areas, making timely and accurate short-term forecasts essential for early warnings to the public, emergency services, and infrastructure operators. On the sub-hourly scale, nowcasting - statistical forecasting based on the most recent observations - offers high spatial and temporal precision. In particular, deep learning models can effectively perform nowcasting by learning to approximate physical processes that are implicitly embedded in diverse observational datasets. These models can generate accurate, multi-hazard predictions in seconds, making them ideal for operational early warning systems.COALITION-4 is a deep learning nowcasting model utilizing an advanced encoder-forecaster model to nowcast thunderstorm-related hazards such as accumulated precipitation, lightning occurrence, and hail probability up to one hour in advance. It leverages recurrent convolutional layers and integrates a range of predictor datasets, including radar observations from the Swiss dual-polarization radar network and Météorage lightning data. Input data undergo extensive preprocessing and data augmentation to improve generalization. The model is trained using GPU-accelerated optimization with an adaptive learning rate. The inference is performed for the entire domain of Switzerland.Recently, COALITION-4 has been operationally deployed including continuous monitoring of product completeness and quality as well as fall back options to degraded modes when input data are incomplete. In the convective season 2025, the system is evaluated and compared to the current operational thunderstorm nowcasting system by forecasters at the MeteoSwiss forecasting centers who in turn inform different key users about thunderstorm hazards: cantonal authorities, civil protection agencies and fire brigades use the information to optimize response strategies during and after thunderstorms. In aviation, accurate short-term lightning forecasts can improve safety and efficiency by guiding temporary suspensions of airport ground operations. We will present both qualitative and quantitative comparisons with current operational thunderstorm nowcasting system of MeteoSwiss, focusing on warning lead times, spatial accuracy, and consistency of forecasts across multiple hazards under diverse convective conditions. This assessment and forecaster feedback informs the next steps in model development, including enhancing forecast quality, usability, interpretability and integration into operational workflows.