Global weather forecasting models based on Artificial Intelligence (AI), have demonstrated skill surpassing the best conventional weather forecasting models, requiring only a fraction of the computational cost.
In this project, the Norwegian Meteorological Institute (MET) collaborates with the European Centre for Medium-Range Weather Forecasts (ECMWF) and meteorological institutes across Europe to develop a purely data-driven, high-resolution regional weather forecasting model.
The new model is named Bris, after the Norwegian word for "light wind." The goal of the project is for Bris to become MET Norway’s best weather forecasting model, delivering more accurate forecasts to our 10 million unique weekly users on Yr.
Innovative architecture and advanced training
Bris is based on graph neural networks (GNNs) and employs an innovative stretched-grid architecture. This architecture allocates higher spatial resolution to regions of interest while maintaining lower resolution elsewhere on the globe. This design allows weather patterns to seamlessly move into the region and enables accurate forecasting over the complex topography of Scandinavia. The model architecture is flexible, allowing the integration of multiple observation sources, such as crowd-sourced data, to enhance forecast accuracy.
The current version of Bris features 246 million trainable parameters, and essential to its development has been the supercomputing facilities provided by Sigma2. The model has been trained on 128 GPUs in parallel on the LUMI supercomputer.
10-day global forecast in just 2-3 minutes
One year into the project, Bris is already producing temperature forecasts for Norway that outperform the state-of-the-art forecasts on Yr. The model is currently semi-operational, providing updated 10-day forecasts four times per day. Unlike conventional physics-based models, once Bris is trained, it can produce a 10-day global forecast in just 2-3 minutes on a single high-end GPU.
