Work Structure OverviewThe overall project is organised in four Themes. The first Theme puts the datasets that are needed together and makes them usable for machine learning in an efficient and scalable manner. The second theme is building the core WeatherGenerator model.The third Theme is working on twenty-two applications that are developed by the project partners. The fourth Theme coordinates the project, organises hackatons and dissemination workshops, and provides services to externals to get started using the WeatherGenerator for their own work.Cover Photo of - Christian Leesig - Work structure overview visual Open Modal Close Modal
Collecting the DataTheme 1 focuses on gathering and preparing the data used by the WeatherGenerator. Because this model is fully data-driven, having broad and reliable datasets is essential to make sure forecasts can be generated accurately.Different weather tasks require different kinds of data and time scales. The work in this theme combines many sources into a usable foundation for the model.Cover Photo of - Ilaria Iuise - Collecting the data visual Open Modal Close Modal
Building the ModelTheme 2 develops the core WeatherGenerator model itself. This work connects the prepared datasets with the machine learning architecture and the latent-space timestepping needed to generate forecasts.The focus is on making the system flexible enough to work across different inputs while staying robust for real applications.Cover Photo of - Martin Schultz - Building the model visual Open Modal Close Modal
Applications in PracticeTheme 3 brings the model into practice through a broad range of applications developed by project partners. These use cases help validate the approach and show where the WeatherGenerator can support real decision-making contexts.The result is a set of practical demonstrations that connect research development with user needs.Cover Photo of - Thomas Nipen - Applications in practice visual Open Modal Close Modal
Coordination & OutreachTheme 4 coordinates the overall project and supports the wider community around it. This includes workshops, hackathons, dissemination activities, and services for external users who want to understand and adopt the WeatherGenerator.It ensures the project remains collaborative, visible, and usable beyond the core consortium.Cover Photo of - Florentine Weber - Coordination and outreach visual Open Modal Close Modal
Representation LearningWhat is representation learning and how does it support smarter weather forecasting?Sophie Xhonneux from explores exactly that in the second WeatherGenerator Science Explainer.The atmosphere is measured continuously by satellites, weather stations, and balloons, producing vast, overlapping datasets. Rather than training separate models on all of this data, WeatherGenerator learns a compressed, unified representation that captures the underlying physics of the atmosphere, enabling more efficient and powerful forecasting across applications.Watch the video to hear Sophie explain the principles behind representation learning and its fundamentals.Cover Photo of - Sophie Xhonneux - Collecting the data visual Open Modal Close Modal
Energy-Aware ForecastingIn the first instalment of the WeatherGenerator Science Explainer series, Even Nordhagen from explains how the WeatherGenerator goes beyond traditional weather models, directly predicting energy - relevant variables like wind power output and reservoir inflow, without the need for separate downstream models. With a built-in long-term memory, the system can track slow-building processes like snowpack and soil moisture, and ultimately learn the full chain from atmospheric conditions to electricity markets.Cover Photo of - Even Nordhagen - Work structure overview visual Open Modal Close Modal