Ionospheric Forecasts Across Geomagnetic Conditions: Comparing Vision, Graph, and Recurrent Models

Ionospheric variability affects Global Navigation Satellite Systems (GNSS) accuracy and radio frequency communications, with consequences for near-Earth and aviation operations. As society's reliance on space-based infrastructure grows, forecasting ionospheric variability -- especially during geomagnetic storms -- becomes increasingly important. We present a spatiotemporal deep learning framework for short-term forecasting of global Total Electron Content (TEC), trained on 14 years of historical global ionospheric maps (GIMs). We compare a GraphCast-inspired autoregressive Graph Neural Network (GNN), an autoregressive Vision Transformer (ViT), and a Convolutional Long Short-Term Memory (ConvLSTM) model. We evaluate the models on 650 held-out event windows, including 42 geomagnetic storm events, across storm levels, geographic regions, local-time bands, and forecast horizons up to 3 hours. The ViT and GNN outperform the ConvLSTM and persistence baselines, with aggregate RMSE below 3 TECU over 3-hour rollouts. These results establish an event-stratified benchmark for global TEC forecasting.