✨ TL;DR
This paper introduces IonoDGNN, a dynamic graph neural network that forecasts ionospheric irregularities by modeling satellite pierce points as graph nodes with time-varying connectivity, conditioned on predictable future satellite positions. The approach outperforms persistence baselines by 35-52% and enables forecasting on lines of sight that don't yet exist in the observation period.
Traditional data-driven ionospheric forecasting models rely on gridded representations that fail to preserve the actual time-varying sampling structure of satellite-based measurements. As satellites move, the locations where they pierce the ionosphere change continuously, creating a dynamic observation geometry. Existing grid-based approaches must interpolate these measurements onto fixed spatial grids, losing information about the actual measurement structure and making it difficult to forecast conditions along lines of sight that only appear in the future. This is particularly problematic for predicting ionospheric irregularities that affect Global Navigation Satellite System (GNSS) signals, where accurate forecasts along specific satellite-receiver paths are needed. The challenge is compounded by the need to forecast irregularities for satellites that rise during the forecast horizon—satellites not yet visible at prediction time but whose trajectories are known from ephemeris data. Standard approaches struggle with this scenario because they cannot naturally incorporate information about future observation geometry into their predictions.
The authors model the ionosphere as a dynamic graph where nodes represent ionospheric pierce points (IPPs) and edges connect spatially or temporally related measurements. The graph topology evolves as satellites move, naturally capturing the time-varying observation geometry. The key innovation is ephemeris conditioning: because satellite trajectories are predictable from orbital mechanics, the future graph structure over the forecast horizon can be constructed in advance and explicitly provided to the model. This allows the neural network to condition its predictions on where satellites will be, enabling forecasts for lines of sight that don't yet exist in the observation window. The model, called IonoDGNN, uses graph neural networks to perform spatial message passing between connected IPPs, allowing information to propagate across the observation network. The architecture processes multi-GNSS data from co-located receivers in Singapore, forecasting Rate of TEC Index (ROTI) irregularities as binary probabilistic classifications at 5-minute intervals up to 2 hours ahead. The model is trained and evaluated on data spanning January 2023 through April 2025, with the task formulated as predicting whether irregularities will occur at each node in the future graph.
What the paper shows.
IonoDGNN achieves a Brier Skill Score of 0.49 and precision-recall area under the curve of 0.75 when forecasting ionospheric irregularities up to 2 hours ahead at 5-minute cadence. These metrics represent improvements of 35% in BSS and 52% in PR-AUC compared to persistence baselines. The performance gains are larger at longer lead times, indicating the model's ability to capture temporal evolution beyond simple extrapolation. For the critical case of satellites rising during the forecast horizon, ephemeris conditioning proves essential, achieving a receiver operating characteristic AUC of 0.95 compared to only 0.52 without this conditioning. Under simulated coverage dropout scenarios, the model maintains predictive skill on affected nodes by leveraging spatial message passing from neighboring observations in the graph structure.
The evaluation is conducted on data from a single geographic location (Singapore) with co-located receivers, which may limit generalizability to other regions with different ionospheric characteristics or receiver configurations. The paper does not extensively discuss computational costs or scalability to larger networks with many receivers, nor does it address how the model might perform in regions with sparser GNSS coverage. The 2-hour forecast horizon, while operationally useful, is relatively short compared to some space weather forecasting applications that may require longer lead times. The binary classification formulation may not capture the full range of irregularity severity, and the paper does not discuss how the model handles extreme or rare ionospheric events that may be underrepresented in the training data.
✨ Generated by Claude · Apr 21, 2026 · Read the PDF for authoritative content.