Regional Data-driven Weather Modeling with a Global Stretched-grid

Data-driven models (DDMs) are rapidly emerging as competitive alternatives to conventional numerical weather prediction (NWP) models, producing accurate weather forecasts at a fraction of the computational cost. Here, we extend the Artificial Intelligence Forecasting System (AIFS), a global DDM developed by the European Centre for Medium-range Weather Forecasts (ECMWF) to support high-resolution regional forecasting needs at the Norwegian Meteorological Institute (MET Norway) for a domain covering most of the Nordic countries. This model is global with a stretched grid, dedicating high resolution over the Nordic region of interest (2.5 km) and lower resolution (0.25°) elsewhere. The model is based on graph neural networks, an architecture that naturally affords arbitrary multi-resolution grid configuration. The model is pre-trained on 40 years of global ERA5 data, and further fine-tuned on 3 years of operational analyses from the MetCoOp ensemble prediction system (MEPS). The model is developed to produce short-range deterministic forecasts. We target the data-driven model for use in MET Norway's public weather forecasting app Yr (https://www.yr.no) and the evaluation of the model is done keeping in mind our end-users and their needs. The model is evaluated against surface synoptic observations covering Norway. We will show that the data driven model at 2.5 km resolution for many evaluation metrics outperforms the current operational NWP model.