Multi-domain: A dynamic way of training across domains and resolutions

The domain of weather forecasting is currently undergoing a significant transformation driven by advances in machine learning. State of the art data-driven weather models have demonstrated performance that surpasses the state of the art traditional numerical weather prediction (NWP) models, while operating at a fraction of the computational cost (Bouallegue et al., 2024). 

While global models like AIFS (Lang et al.,2024), GraphCast, and Pangu have gained a lot of attention, different flavours of high-resolution regional modeling have emerged  developed by different meteorological institutes. Stretched-grid is a global model with an increased spatial resolution and dynamics over a region of interest, a novel approach to regional modelling. This capability is demonstrated to be highly competitive and even surpass state of the art regional NWP for certain variables (Nipen et al., 2025; Nordhagen et al., 2025).

Building on the concept of stretched-grid and generalizing the idea to incorporate more high-resolution data, while avoiding intermediate fine-tuning and transfer learning steps, one can achieve high resolution predictions for any domain in Europe.  In this work, we propose a probabilistic multi-domain model, introducing a new way of training across domains and resolutions, by utilizing the concept of dynamical graphs. This is done by alternating between different global and regional data and its corresponding graph across different spatial regions, grid-types and resolutions. This capability is enabled through a flexible encoder, processor and decoder architecture. The idea is to make the model less biased towards certain grid types, terrain and dynamics induced by the training data, enabling the model to generalize across resolutions and climate zones. 

The kilometre scale model has been trained on analyses from four high resolution NWP models covering various parts of Europe, including AROME-Artic (2.5km), the MetCoOp Ensemble Prediction System (MEPS, 2.5km), UWC-West (2km) and the Austrian Reanalysis (ARA, 2.5km). We evaluate the performance of this model on another domain that was left out of the training, and show improved generalizability when compared to a model trained on fewer domains.  With one year of verification, we show improvements across parameters such as 2m temperature, mean sea level pressure, 10m wind speed and total precipitation.