Technical Note: Sub-millimetre real-time GNSS precipitable water vapor conversion using AI weather forecast foundation models

Abstract. Global Navigation Satellite Systems (GNSSs) networks provide high-quality, high-temporal-resolution (5-min) tropospheric delay products at over 20,000 stations worldwide. However, the vertical meteorological information remains inaccessible at most sites due to the lack of in-situ meteorological sensors, which makes it challenging to convert high-accuracy tropospheric products to accurate precipitable water vapor (PWV). The continued development of empirical models has provided feasible solutions, but their inherent modelling errors continue to restrict the accuracy of real-time PWV conversion, and the current optimal method is still limited to the 2–3 mm level. We propose a real-time high-accuracy GNSS PWV conversion method based on artificial intelligence weather forecast foundation models. This innovative approach efficiently calculates path-integrated zenith hydrostatic delay (ZHD) and weighted mean temperature (Tm) at any global location locally within seconds, bypassing inaccuracies inherent in empirical ZHD and Tm models, significantly reducing PWV errors. The method is evaluated on three representative foundation models, Pangu-Weather, GraphCast, and FengWu, showing root-mean-square errors of real-time GNSS PWV less than 1 mm (refer to 5-day forecast time, for the 5-day delay in ERA5 reported by ECMWF). It offers a robust solution for real-time GNSS PWV conversion and has the potential to facilitate the assimilation of the increasing GNSS observations into global numerical weather prediction. In turn, employing more accurate atmospheric states that assimilate more GNSS observations as input can further enhance the method itself.