GNSS Zenith Wet Delay Integration Lifts Aurora Model Severe Precipitation Skill 8.8% at 99th Percentile

The arXiv preprint demonstrates the first assimilation of Zenith Wet Delay, a direct column moisture metric from GNSS signal delays, into a machine-learning weather foundation model. Researchers fine-tuned Aurora on ZWD alongside its native variables then evaluated six-hour accumulated precipitation. The approach exploits ZWD's all-weather availability to supply information that conventional satellite and in-situ observations only infer indirectly. Performance scaled with event severity, indicating the data corrects a systematic bias in MLWMs that under-represent heavy rain.

Precipitation power spectra improved at synoptic and planetary scales, showing the model now better resolves moisture convergence patterns that drive organized convective systems. This gain complements existing numerical weather prediction assimilation of ZWD, which has operated for decades, yet had remained absent from leading foundation models despite documented deficiencies in extreme rainfall. The result suggests ML architectures can exploit raw observational constraints more effectively than parameterized physics when moisture is explicitly supplied.

Operational rollout within months could sharpen local forecasts used for aviation, agriculture, and flood alerts. Remaining questions include generalization across GNSS networks with varying density and whether similar gains appear for other moisture-sensitive variables such as convective available potential energy. A multi-model replication using GraphCast or FourCastNet would strengthen causal claims about ZWD's unique value.

Next steps involve real-time cycling experiments and verification against independent rain-gauge networks to confirm the reported ETS improvements hold outside the training distribution.