Real-Time GPS ionospheric TEC Mapping using Multi-Modal Deep Learning: Bridging Solar Imagery and Ionospheric Physics

The ionosphere's Total Electron Content (TEC) is a critical parameter for Global Navigation Satellite System (GNSS) positioning, particularly Precise Point Positioning (PPP), satellite communications, and space weather monitoring. While the International GNSS Service (IGS) provides a baseline for global ionospheric maps (GIMs) errors, these products are generated post-facto and lack the temporal resolution needed for real-time forecasting during rapidly evolving solar events. Accurate short-term TEC prediction remains challenging due to the complex, nonlinear coupling between solar extreme ultraviolet (EUV) radiation and ionospheric dynamics.

Here, we present a multi-modal deep learning framework that integrates full-disk solar imagery from NASA's Solar Dynamics Observatory (SDO)/Atmospheric Imaging Assembly (AIA) with historical TEC observations to generate global TEC forecasts 15 minutes ahead. The model employs a U-Net architecture with Feature-wise Linear Modulation (FiLM), enabling solar EUV intensities across three wavelength channels (94 Å, 131 Å, 171 Å) to dynamically condition ionospheric feature extraction. By predicting TEC residuals with the standard IGS grid (2.5° × 5° geographic resolution), the framework directly learns solar-driven perturbations while preserving spatial coherence through geometry-aware inputs, encoding solar zenith angle and local solar time. The training dataset spans between 2010-2018 and comprises approximately 770 solar flare events (C, M, and X-class). Stratified sampling across flare classes ensures robust model performance under diverse space weather conditions. We implement comprehensive data preprocessing, including exposure normalization, disk masking, and logarithmic intensity scaling.

The operational concept leverages SDO's continuous, near-real-time AIA data availability. The model ingests the current TEC state, combined with the latest solar EUV imagery, to predict TEC at t+15 minutes. These predictions can then serve as input for the subsequent forecast step, creating an autoregressive chain where each iteration combines new AIA observations (available with ~ 15-minute latency) with the previously predicted TEC state. This sliding window approach enables continuous TEC nowcasting without reliance on ground-based GNSS processing delays.