Seamless Weather Forecasting through Machine Learning

Delivering seamless weather forecasts has long been a central objective, reflecting both user expectations and sustained development efforts. At present, multiple specialized forecasting systems are operated across different lead times and subsequently combined to produce integrated products and services. Because this integration is often tailored to specific applications, inconsistencies may arise between forecast products, complicating their interpretation and potentially reducing their overall utility. In response, MeteoSwiss has initiated a project to develop a seamless forecasting system designed to support a wide range of applications, including public weather forecasts, impact-based warnings, and climate services. Part of the project explores the possibility to use data-driven forecasts as recent advances in Machine Learning-based forecasting methods offer new opportunities to achieve genuine seamlessness directly within a unified forecasting framework. The envisioned system is designed to deliver ensemble forecasts with lead times of up to ten days while enabling high-frequency updates on the order of ten minutes to address nowcasting requirements. Beyond system development, the project also aims to transition this new approach into operations and progressively replace components of the current baseline systems.

This contribution summarizes progress achieved during the first two years of the project. Development of the ML-based forecast system has taken place within the Anemoi framework, in close collaboration with broader European initiatives in data-driven prediction. As an initial demonstrator, a deterministic forecast with a five-day horizon, 1 km spatial resolution, and hourly temporal resolution has been implemented, running every six hours. The system employs a stretched-grid configuration and has been trained using ERA5 data, a 20-year regional ICON reanalysis dataset, and several months of operational ICON analyses. Evaluation shows that, in terms of deterministic skill scores, the demonstrator equals or outperforms the operational ICON limited-area model. Current developments focus on enabling more frequent forecast updates and incorporating observational data during inference to support nowcasting, representing the next key milestone toward operational deployment.

We conclude by summarizing the main insights gained to date—spanning scientific and methodological advances as well as technical, operational, and organizational experience—and by outlining the principal challenges that remain on the path toward full operational implementation.