Operational Improvement of Probabilistic Nowcasting in the Nordics

This follow-up project tries to improve the probabilistic precipitation nowcasting system developed during the previous EuroHPC allocation and prepares it for operational use over the Nordic radar domain. The current Anemoi graph neural-network precipitation nowcaster already runs at operational speed and produces five ensemble members in less than two minutes on one GPU from station and radar history.

The next step is not to replace this working system, but to improve its physical realism and spatial rain representation while preserving the pointwise and onset skill needed by public forecasting services. The project will first build a dense postprocessed precipitation-analysis target with radar fields and quality-controlled official and citizen observations, following MET Norway's operational use of citizen observations.This target will be used to fine-tune the current model and test whether observation-constrained targets improve spatial coherence without degrading station-level verification.

In parallel, the project will develop lightning probability nowcasting as an auxiliary product. Initial experiments have started by predicting future lightning counts from past lightning fields and from precipitation-only inputs through a decoder. These results will be extended into calibrated lightning probability fields and evaluated as a correlated convective variable. If the radar composite alone cannot support the required realism, additional vertical radar information will be tested as a targeted extension. The project requests 42,500 node-hours, about 170,000 GPU-hours on 4-GPU nodes, and approximately 30TB of storage.