Data-Driven Weather Prediction for the Greater Alpine Region Using the Anemoi Framework

Machine learning has been transforming weather prediction at an unprecedented pace, driven by advances in modelling approaches, rapid progress enabled by open-source collaboration, and growing availability of high-quality atmospheric datasets across global to regional scales. One prominent example at the crossroads of these developments is the anemoi framework, which supports both model development and practical application in data-driven weather prediction. 

In this contribution, we present our experiences applying data-driven weather prediction using the anemoi framework for the Greater Alpine Region centred on Austria. Specifically, we use a multi-stage training pipeline to reduce training costs while learning atmospheric dynamics from reanalysis datasets with varying spatiotemporal coverage and resolution using a deterministic architecture. In addition, we extend this transfer learning approach to increase temporal resolution of the forecasting model and to fine-tune it using other datasets complementing reanalysis datasets.

Our data-driven models achieve consistently better scores compared to operational numerical weather prediction models running in-house, although capturing extreme values and spatial structures of predicted fields at high resolution remains a challenge. We analyse our models with increasing temporal resolution from 6- to 3- hours and show that error growth with increasing lead time is not a critical issue within the 72-hours lead time required for our models. We also discuss options for extending the set of predicted variables by fine-tuning with complementary datasets and present initial results. Finally, we provide an initial assessment of our model for real-time predictions over the convective season of 2026 to discuss the path towards operationalisation. These findings aim to support future design choices and contribute to a clearer understanding of the capabilities and limitations of data-driven weather prediction models.