Towards a data-driven weather model for forecasting on-demand extremes at hectometric scale

The domain of weather forecasting is currently undergoing a significant transformation driven by advances in machine learning, where Data-Driven Models (DDMs) have demonstrated equal or superior performance compared to traditional Numerical Weather Prediction models in predicting various variables, while operating at a fraction of the computational cost (Bouallegue et al., 2024). Probablistic DDMs have the potential to provide a computationally cheap solution for ensemble modelling at hectometric scale (with a spatial resolution of 500 to 750 m), motivating Task 330141 of the Destination Earth Weather-Induced Extremes Digital Twin (Extremes DT, DE330) project (ECWMF, 2024). An important step towards hectometric ensemble modelling is high-resolution (km-scale) modelling over a regional domain, for which Nordhagen et al. (2025) have already shown promising results for the Nordic area with their Bris CRPS-FFT model, a model using a stretched-grid approach, transfer learning and a Continuous Ranked Probability Score (CRPS) loss function with Fast Fourier Transform (FFT) to combine good model skill with spatial coherence, also at the smaller scales. On-demand extreme forecasting involves potentially high-impact events, which requires the flexibility to respond fast to triggered events and provide a tailored forecast on the domain of interest, including crucial uncertainty information. In this work, we combine the kilometer-scale multi-domain training approach -based on dynamical graph training- with the Bris CRPS-FFT approach, with the aim of providing on-demand ensemble forecasts of extremes at a hectometric scale. Probablistic multi-domain DDMs have the potential to provide the forecasting speed, the domain flexibility and the uncertainty quantification necessary to handle this complex task. Following Nipen et al. (2024), a global model at 0.25-degree resolution is pre-trained (stage A and B), after which transfer learning is applied to integrate the higher-resolution data in stage C. In this stage, dynamical graph training is employed, where multiple regional datasets are alternated as batch input to increase the generalizability of the model. In the final stage, we finetune the model on 200+ hectometric datasets of the Extremes DT triggered by forecast extreme events, resulting in the first on-demand autoregressive ensemble model of its type at hectometric scale. The results reflect the potential of multi-domain training combined with transfer learning for on-demand highly flexible domains with sparse data availability.