Fei Yang, Kaiwen Xu, Ruixian Hao, Ruohan Liu, Lei Wang, Zi Wang, Dashan Zhou
ABSTRACT Accurate estimation and forecasting of zenith tropospheric delay (ZTD) are vital for satellite Earth observation, atmospheric monitoring, and meteorological applications. Ensemble forecasts could provide relevant atmospheric parameters that can be used for ZTD forecasting, but achieving high-accuracy, real-time ZTD prediction remains challenging due to the uncertainty of the initial conditions and the model resolution. With the development of large-scale models, artificial intelligence-based (AI-based) weather models also provide these parameters for ZTD forecasting with enhanced computational efficiency. A real-time calculation and comparative evaluation of high-accuracy ZTD forecasts was conducted, in which the prediction methods based on the ensemble and AI-based forecasts, i.e., ECMWF-IFS and ECMWF-AIFS, were compared in detail. Using global IGS-ZTD as a reference, ECMWF-AIFS demonstrated superior error distribution characteristics compared to the widespread systematic overestimation of ECMWF-IFS, achieving a regression slope of 0.995 versus 0.992. Spatially, ECMWF-AIFS significantly optimized global consistency by concentrating RMSE within the 0-15 mm range, effectively mitigating error fluctuations in low-latitude and equatorial regions. Temporally, both types of data maintained continuous and stable forecast capabilities throughout the 15-day forecast cycle; ECMWF-IFS maintained an RMSE below 30 mm during the first 6 days, while ECMWF-AIFS improved initial forecast accuracy by 35.2% and controlled the RMSE at 38.73 mm even on Day 15. In complex tropical coastal environments characterized by high water vapor variability, ECMWF-AIFS exhibited exceptional robustness, reducing the RMSE to 17.58 mm. These results indicate that AI-based weather models can effectively suppress error accumulation and correct systematic biases, demonstrating significant potential for real-time, high-precision tropospheric delay forecasting.