WeatherGenerator Applications

Weather affects nearly every aspect of society, influencing weather-dependent industries, emergency management, and the general public. Accurate weather information is essential for informed decision-making, and has traditionally been provided by physics-based numerical weather prediction (NWP) models. The WeatherGenerator (WG), an EU-funded project under HORIZON-INFRA-2024-TECH-01-03 (Project ID 101187947), aims to develop a generative foundation model of the Earth system. Pre-trained on a vast and diverse collection of datasets—including reanalyses, satellite and radar observations, and in-situ measurements—the WG will learn a rich latent representation of the earth system. Downstream applications that rely on weather information can directly access this representation, enabling them to predict application-specific parameters using lightweight neural networks instead of building complex models from scratch. The main objective is to leverage the WG to develop machine-learning (ML) solutions for 23  high-impact applications, spanning weather and climate prediction, renewable energy, water, human health, food security, and the biosphere. Each application will fine-tune a pre-trained WG model using application-specific data. Compute resources are needed for repeated cycles of hyper-parameter tuning and validation of the models, playing a key role in guiding the core development of the WG in this phase of the project.