Climate science has long relied on simulated and projected events to populate the statistics of climate variables, using models to generate large synthetic populations of events from which the moments of a climate distribution (mean, variance, and especially its tail) can be estimated, precisely because short observational records cannot sample rare extremes with any statistical robustness. This study applies that same strategy to Mediterranean hurricanes, or "Medicanes": rare, high-impact tropical-like cyclones bringing intense rainfall to densely populated coastal regions, for which too few real cases exist to build a reliable climatology. Building on a statistical-deterministic downscaling methodology used for over a decade to translate global climate model output into storm-scale hazard information , from early applications on CMIP3 simulations (Romero and Emanuel 2013) to CMIP5-based projections (Romero and Emanuel 2017) , we generate thousands of synthetic medicane tracks from reanalysis data and global climate models. Coupling these tracks with a physics-based tropical cyclone rainfall algorithm produces spatially and temporally resolved precipitation fields, from which the full statistical distribution of medicane rainfall and its associated return periods , can be estimated, even for magnitudes far beyond anything recorded to date. The approach is validated against reanalysis, satellite, and radar observations, and sensitivity analyses probe how synthetic sample size and algorithmic choices affect the resulting statistics. Under the RCP8.5 scenario, the resulting hazard distribution shifts markedly: 250-year rainfall along the Adriatic coast increases by up to 160 mm, and some 500-year events are projected to exceed 800 mm of total rainfall. By showing how synthetic event generation can be used to populate and characterize the statistics of a rare climate hazard, this work offers a methodological blueprint directly relevant to AI- and data-driven approaches now being developed to model climate extremes for which observations alone will always be insufficient.

Saraceni, M., L. Silvestri, P. Bongioannini Cerlini, R. Romero, and K. Emanuel, 2026: Estimating Medicane Precipitation Hazard. Journal of Climate, in press, https://doi.org/10.1175/JCLI-D-25-0364.1.

Artificial Intelligence can play an important role in climate-change mitigation and adaptation, provided that its decisions are reliable, resource-efficient, and safe to deploy in the real world. This talk presents selected research activities carried out at the Intelligent Systems Lab (ISLa) of the University of Verona on the development of safe and efficient AI methods for intelligent cyber-physical systems. The first part focuses on energy management in smart buildings and energy networks. I will discuss three representative lines of work: time-series forecasting for load prediction in district-heating networks; probabilistic planning under uncertainty for HVAC control; and safe reinforcement learning for hybrid pump systems, renewable-energy systems, and building energy management. These examples illustrate how predictive models and sequential decision-making techniques can support more efficient and sustainable energy use while respecting operational and safety constraints. The second part considers environmental monitoring with unmanned surface vehicles (USVs). I will briefly present research on time-series segmentation for state-model generation, anomaly detection, and safe reinforcement learning for autonomous motion planning. In this setting, AI must operate with incomplete information, noisy data, changing environmental conditions, and potentially costly failures. Across these applications, the common goal is to develop AI systems that are not only accurate, but also efficient, robust, interpretable, and safe - key requirements for deploying AI in support of climate change mitigation.