A Deep Ensemble Transformer Model for Global Ionosphere Prediction and Uncertainty Quantification

Abstract Accurate global ionospheric forecasting is important for various purposes, from geophysical research to practical applications, including real‐time precise positioning and navigation. Existing studies primarily focus solely on deterministic predictions and often overlook uncertainty quantification. In this study, we propose a deep ensemble transformer model to predict global ionospheric maps (GIMs) one day ahead while providing reliable uncertainty estimates. The model is trained on GIMs from the Center for Orbit Determination in Europe (CODE) spanning two solar cycles and evaluated across years with varying solar activity. The model achieves root mean squared errors (RMSE) ranging from 1.25 TECU to 4.96 TECU from solar minimum to solar maximum periods. The predicted uncertainties capture the spatial and temporal variability of actual prediction errors, with coverage probabilities (68.40%–71.23%) and higher‐order intervals (, ) consistent with Gaussian distributions. The application of the forecast GIMs in single‐frequency precise point positioning further demonstrates their quality, showing an improvement of up to 7.1% in positioning accuracy over using CODE predicted GIMs. Incorporating predictive uncertainty into observation weighting further enhances the positioning performance by up to 6.3%. The model generalizes well to other GIMs with different accuracy and temporal resolutions, including CODE rapid GIMs, enabling its use in real‐time applications.