Lightning and precipitation forecast by means of hybrid deterministic and AI-based tools

Mediterranean coastal regions are regularly affected by sudden heavy precipitation events leading to very dangerous flash floods. Severe convection prediction, being the result of many mutually interacting multiscale processes, not yet completely understood and modeled, is still a major challenge for numerical weather prediction (NWP) systems. In recent times, artificial intelligence (AI) emerged as a powerful tool for handling vast amounts of data and extracting patterns and relationships that might be challenging to identify through traditional fully-deterministic algorithms. In the framework of the AIxtreme (Physics-based AI for predicting extreme weather and space weather events) project, a suite of AI-based techniques is being developed to calibrate numerical models based on the physics of the atmosphere, with the aim of anticipating the occurrence of extreme weather events and supporting decisions of civil protection agencies.A first significant result of the project is the development of a deep learning framework, named FlashNet, able to forecast lightning flashes up to 48 h ahead in terms of probability of occurrence. FlashNet is capable to find an optimal mapping of meteorological features predicted two days ahead by the state-of-the-art numerical weather prediction model by the European Centre for Medium-range Weather Forecasts (ECMWF) into lightning flash occurrence. The prediction skill of the resulting AI-enhanced algorithm turns out to be significantly higher than that of the fully deterministic algorithm employed in the ECMWF model. A remarkable Recall peak of about 95% within the 0-24 h forecast interval is obtained. This performance surpasses the 85% achieved by the ECMWF model at the same precision of the AI algorithm.A second tool, in an advanced stage of development, is designed for forecasting extreme precipitation events. A neural network is trained with features from a ECMWF large-scale model and observations from the rainfall network, to predict the occurrence of an extreme precipitation event and its cumulative within 3 hours up to 48 h ahead. The network thus designed is able to improve the prediction of the raw model at any point of the ECMWF model grid, surpassing, at the level of classification and regression indices, the local-scale non-hydrostatic model MOLOCH.