AI for Weather & Climate
Foundational AI models for every timescale, from the next hour to the next decades
Climate change is increasing how often and how hard extreme weather events hit Italy and Europe. Intense thunderstorms, hailstorms, flash floods and heatwaves put people, infrastructure and economic activity at risk, especially in complex mountain areas like the Alps. To respond, we need forecasts that are more accurate and more detailed, and that can be produced fast enough to support real-time decisions.
FBK researches and develops artificial intelligence models for the next generation of weather and climate prediction. We train them on decades of re-analysis data, numerical simulations and observations. The resulting models match or improve on traditional forecasting skill across timescales, at a fraction of the computational cost of numerical simulations. We work closely with civil-protection agencies, meteorological services and companies, and we take our models all the way from research to operational services.
What we do
AI nowcasting and short-term forecasting
We build generative AI models for precipitation nowcasting (0–6 hours) that predict how thunderstorms and hailstorms will evolve. The models combine weather radar with other sources, including satellite, lightning detection and numerical weather prediction models. They produce probabilistic forecasts as ensembles of possible scenarios, at very high resolution: down to 0.5–1 km and every 5 minutes. We are now extending them to the short-term range, up to 24 hours. Our research also covers hail probability estimation, storm-cell tracking, and blending AI nowcasts with numerical weather prediction.
AI downscaling
We develop generative AI models that turn coarse meteorological fields into kilometre-scale ones. They are trained on decades of paired low- and high-resolution simulations, re-analyses and observations. This lets us reconstruct local variables such as temperature, precipitation and wind at a much lower computational cost than dynamical downscaling. We tailor each training setup to its application, whether that is downscaling seasonal forecasts, climate projections or other numerical simulations.
Seasonal forecasting
We are exploring AI approaches to seasonal forecasting based on weather regimes. Our models learn the complex relationships between Euro-Atlantic atmospheric circulation and surface conditions in order to reconstruct European temperature and precipitation anomalies. Early results point to computationally efficient alternatives to traditional seasonal forecasting.
Weather early-warning systems
We turn our AI models into operational, real-time early-warning systems. Our platform takes in data from many sources and sends alerts through several channels. It is built to fit into the existing workflows of regional and national civil-protection agencies.
Major projects and initiatives
IT4LIA AI Factory – Earth vertical
FBK leads the Earth vertical of IT4LIA, the Italian AI Factory supported by EuroHPC and hosted by Cineca. The vertical develops foundational and generative AI models for weather, climate and the environment, from local nowcasting to climate projections, along with AI-ready datasets and services. Researchers, public institutions and companies can use them for civil protection, agriculture, insurance, utilities and renewable energy.
ICSC – National Research Centre for High Performance Computing, Big Data and Quantum Computing (Spoke 4 “Earth & Climate”)
In Spoke 4 of ICSC, FBK develops generative AI models for downscaling climate and weather simulations. These models produce high-resolution information to support impact studies and adaptation strategies, using Italy’s national HPC infrastructure.
AI nowcasting for the Alpine region – Trento and Bolzano
In the Autonomous Province of Trento, FBK’s nowcasting models feed a fully automated, multi-channel early-warning system that Meteotrentino and the provincial civil protection use. It combines weather stations, radar data and AI nowcasts.
In the Autonomous Province of Bolzano, FBK and the Civil Protection Agency’s Provincial Functional Centre are jointly developing NowcastingBZ, an AI nowcasting system built for the Alpine terrain of South Tyrol. Its main features:
- It brings together radar (including the new dual-polarisation radar), satellite, lightning and the ICON numerical model in one multi-source dataset.
- It uses state-of-the-art generative models such as LDCast and GPTCast to give probabilistic forecasts of precipitation and lightning at 0.5 km and 5-minute resolution, up to about 90 minutes ahead.
- To make up for the new radar’s short history, it uses transfer learning: models are first trained on the long MeteoSwiss radar archive, then fine-tuned on local data.
- It adds hail probability, storm-cell tracking and blending with numerical weather prediction, and it runs in real time on FBK infrastructure.
Scientific collaborations
Our research is carried out in active scientific collaboration with leading European and Italian meteorological services and research institutions:
- ItaliaMeteo – Italian National Agency for Meteorology and Climatology
- Arpae Emilia-Romagna
- Meteotrentino and the Civil Protection Agency of the Autonomous Province of Bolzano
- MeteoSwiss – Federal Office of Meteorology and Climatology
- ECMWF – European Centre for Medium-Range Weather Forecasts
- DMI – Danish Meteorological Institute
- RMI – Royal Meteorological Institute of Belgium
- CIMA Research Foundation
- CMCC – Euro-Mediterranean Center on Climate Change
- CNR – National Research Council of Italy
- ENEA – Italian National Agency for New Technologies, Energy and Sustainable Economic Development
If you are already a subscriber of this newsletter, you can request to update your preferences or to cancel your subscription.
We've sent you an email message, which contains all details to confirm your address to complete the signup to our newsletter.
Something went wrong during this request, please try again and make sure all information is entered correctly.