Spatiotemporal inhomogeneity of accuracy degradation in AI weather forecast foundation models: A GNSS perspective
Abstract
Artificial intelligence weather forecast foundation models can infer precise global atmospheric state forecasts on user devices at speeds over 10,000 times faster than the operational Integrated Forecasting System, making increasingly significant contributions to geodetic applications represented by global navigation satellite systems (GNSS). Existing evaluations often concentrate on particular events or comparisons with physical models and therefore are not comprehensive. This study obtains global forecast results of foundation models for 2022, derives GNSS tropospheric delay through numerical integration, and calculates mean deviation, mean absolute error, and root mean square error. These metrics are used to analyze spatiotemporal inhomogeneity in accuracy degradation of Huawei Cloud Pangu-Weather, Google DeepMind GraphCast, and Shanghai AI Lab FengWu, evaluating how it changes with forecast time and identifying best-performing models across regions and durations. Results indicate that considering topography during training enhances accuracy at high altitudes and reveal facilitating influence between closely related atmospheric variables such as precipitation and water vapor.
Tong Liu
Postdoctoral Fellow
Research interests include GNSS ionospheric monitoring, ionospheric impacts on precise positioning, and mitigation strategies.