Integrating GNSS-Derived Atmospheric Delays into Large Weather Foundation Models

Large weather foundation models have recently emerged as a powerful paradigm for global weather forecasting, leveraging transformer-based architectures pretrained on vast and heterogeneous Earth system datasets. Despite their success, accurately predicting moisture-related processes - particularly those associated with atmospheric water vapor and precipitation - remains a key challenge. Global Navigation Satellite System (GNSS) observations provide an independent and physically meaningful source of information on atmospheric water vapor through signal delays induced along the signal path, offering an opportunity to enhance data-driven weather models.

In this work, we investigate the integration of GNSS-derived Zenith Wet Delays (ZWDs) into Aurora, a state-of-the-art large weather foundation model based on a hierarchical vision transformer architecture. Building on Aurora’s pretrained representations, we perform full fine-tuning using ten years of ERA5 reanalysis data augmented with surface-level ZWD fields generated by the ZWDX global forecasting model. To rigorously assess the contribution of GNSS information, we conduct controlled experiments in which identical model configurations are fine-tuned both with and without the inclusion of ZWDs. Experiments are performed on two model scales, comprising approximately 250 million and 1.3 billion parameters.

To enable stable learning when introducing the additional GNSS-derived variable, we propose an adaptive loss-weight scheduling strategy that gradually increases the contribution of the ZWD loss during training. This approach allows the model to successfully learn the new variable while maintaining performance on the original atmospheric fields. The learned ZWD representations reach an accuracy comparable to that of the other variables included during pretraining.

Beyond the direct prediction of ZWDs, we analyze the influence of GNSS information on moisture-related atmospheric variables, including specific humidity from the original pretraining set and precipitation, which is added during fine-tuning alongside ZWDs. The inclusion of ZWDs leads to measurable changes in the prediction skill of these variables at the surface and, for specific humidity, throughout the atmospheric column. While the magnitude and physical interpretation of these effects are still under investigation, the results indicate that GNSS-derived information is effectively utilized by the model and influences its internal representation of atmospheric moisture.

A central objective of this research is to assess whether GNSS-informed foundation models can improve the prediction of precipitation and nowcasting of extreme weather events, where accurate moisture representation is critical. Ongoing work extends the evaluation to shorter lead times and event-based analyses. Future developments include incorporating direct GNSS station measurements instead of interpolated products and developing regional high-resolution forecasting setups to better exploit the spatial density of GNSS networks, with the ultimate goal of enhancing forecasts of localized, high-impact extreme events.