Artificial intelligence for operational nowcasting of extreme meteorological phenomena and of parameters for energy production from renewable sources

This PhD thesis collects studies and analyses of data assimilation in meteorological and climate systems. It examines both data-driven methods using artificial intelligence and hybrid approaches combining neural networks with physical models. In particular, artificial intelligence techniques are explored to improve the resolution of atmospheric model outputs, with applications in urban-scale climate impact analysis. Then, the use of neural networks for Data Assimilation is analyzed. By exploiting hybrid models, in fact, it is possible to improve predictions of chaotic systems, inference on unmeasured states and estimation of complete states of the system starting from partial measurements. In the second part, instead, we connect data assimilation with the theory of synchronization of dynamical systems. We show how synchronization can help to estimate key system parameters using optimization algorithms, valuable for improving the performance and efficiency of renewable energy systems. We also focus on space-time chaos and examine the simulation of chaotic systems on tree networks. We highlight how computational limitations, such as machine precision, can affect synchronization and lead to unexpected behaviors in simulations. Overall, the thesis demonstrates how the combination of AI and physical models improves predictions and understanding of complex climate and energy systems.