Forecasting Ionospheric Irregularities on GNSS Lines of Sight Using Dynamic Graphs with Ephemeris Conditioning

This paper introduces IonoDGNN, a dynamic graph neural network framework that models the ionosphere as a time-evolving graph of satellite pierce points and utilizes ephemeris conditioning to predict future irregularities on unseen lines of sight, achieving significant performance improvements over persistence baselines in multi-GNSS forecasting tasks.

The Gist

The Big Picture: Predicting "Space Weather" Without a Map

Imagine the Earth is surrounded by a giant, invisible, and slightly wobbly ocean of charged particles called the ionosphere. This isn't water, but it acts like a fog that radio signals (like those from GPS satellites) have to swim through. Sometimes, this "fog" gets choppy and forms irregularities (like underwater bubbles or storms). When these happen, your GPS might glitch, your phone might lose signal, or a plane's navigation could get confused.

Scientists want to predict when these "storms" will happen so we can prepare.

The Old Way (The Grid Problem):
Most current forecasting models try to predict this by drawing a giant checkerboard (a grid) over the Earth. They guess the weather for every square on the board.

• The Problem: GPS satellites don't actually look at a grid. They look at specific points in the sky as they fly by. The "grid" method forces the data into squares, smoothing out the sharp, fast-moving bubbles that cause the real trouble. It's like trying to describe a fast-moving car by only looking at a photo of a grid overlaying the road; you miss the speed and the specific path.

The New Way (The Dynamic Graph):
This paper proposes a smarter way. Instead of a static grid, imagine the ionosphere as a living, breathing social network.

• The Nodes (People): Every time a GPS satellite sends a signal to a receiver on the ground, it creates a "connection point" (called an Ionospheric Pierce Point or IPP). Think of these as people at a party.
• The Edges (Handshakes): As satellites move, these people move, new people arrive, and others leave. The "connections" (edges) between them change constantly based on who is close to whom in the sky.
• The Model: The authors built an AI (called IonoDGNN) that watches this moving party. It doesn't force the data into a grid; it follows the actual satellites as they dance around the Earth.

The Secret Sauce: "Ephemeris Conditioning"

Here is the paper's most clever trick, which they call Ephemeris Conditioning.

The Analogy: The Predictable Dance
Imagine you are at a party, and you want to predict what will happen in the next hour. Usually, you only know what people are doing right now.

• The Twist: In space, we know exactly where every satellite will be for the next 24 hours. Their orbits are like a choreographed dance routine that never changes. We know exactly who will walk into the room and where they will stand before they even arrive.

How the AI Uses This:
The AI doesn't just look at the current party; it looks at the future dance routine.

1. It knows a new satellite (a new guest) is going to enter the room in 30 minutes.
2. Because it knows the "future graph" (who will be standing next to whom), it can prepare to make a prediction for that new guest before the guest even arrives.
3. Without this trick, the AI would be confused when a new satellite appeared, saying, "I've never seen this person before, I have no idea what they're doing!" With the trick, it says, "Ah, I know this person is walking in from the North, and they will be standing next to three people who are currently having a storm. Therefore, they will likely have a storm too."

How They Tested It

They used data from two GPS stations in Singapore that are right next to each other (like twins).

• The Labeling Game: Since they can't easily see the "bubbles" in the sky directly, they used the two stations to cross-check. If Station A sees a storm and Station B (right next to it) sees the same storm, they label it "Confirmed." If only one sees it, they assume it's a glitch and ignore it.
• The Result: The AI was incredibly good at predicting these storms up to 2 hours in advance.
  • It beat the "old way" (just guessing the weather will stay the same as it is now) by a huge margin.
  • It was especially good at handling new satellites that appeared during the forecast, thanks to the "Ephemeris Conditioning."

Why This Matters (The Takeaway)

1. No More Blurry Maps: By stopping the use of grids and using the actual satellite paths, the AI sees the "bubbles" clearly instead of blurring them out.
2. Future-Proofing: Because the AI knows the future positions of the satellites, it can predict problems for signals that haven't even been received yet.
3. Resilience: If a receiver loses data (like a phone losing signal), the AI can still guess what's happening by looking at the "neighbors" in the graph, just like you can guess the weather in your town by looking at the weather in the next town over.

In short: The authors built a GPS weather forecaster that doesn't use a static map. Instead, it watches the moving satellites, knows exactly where they are going to be, and uses that knowledge to predict space weather storms with high accuracy. It's like having a crystal ball that knows the dance moves of the entire solar system.

Technical Summary

1. Problem Statement

Current data-driven models for forecasting ionospheric irregularities (specifically those affecting Global Navigation Satellite System, or GNSS, signals) typically rely on gridded ionospheric products. These products suffer from several limitations:

• Loss of Native Structure: They impose a fixed grid on satellite–receiver lines of sight (LoS), which naturally evolve as satellites rise and set. This process involves smoothing and interpolation that can distort sharp, localized features like Equatorial Plasma Bubbles (EPBs).
• Bias Introduction: Gridded estimates depend on model assumptions and sparse observation mapping, potentially introducing method-dependent biases, particularly in low-latitude regions.
• Inability to Handle Transient Sampling: Standard models struggle to predict conditions for satellites that appear only within the forecast horizon (rising satellites) because they lack historical context for those specific LoS paths.

The paper proposes a shift from grid-based representations to a dynamic graph formulation that operates directly on the evolving set of Ionospheric Pierce Points (IPPs).

2. Methodology: IonoDGNN

The authors introduce IonoDGNN, a Dynamic Graph Neural Network designed to model the ionosphere as a time-varying graph where nodes represent active satellite–receiver LoS links.

A. Dynamic Observation Graph

• Nodes: Each node corresponds to a visible IPP at a specific 5-minute timestep. The node set $V_t$ and edge set $E_t$ change dynamically as satellites rise and set.
• Edges: Nodes are connected via $K$-nearest neighbors (K-NN) based on Haversine distance on a 450 km thin-shell model. Edge features include geodetic distance, coordinate differences, and solar time separation.
• Node Features:
  • Observed ($x_{obs}$): Historical ROTI (Rate of TEC Index), TEC rate, Signal-to-Noise Ratio (SNR), and space weather indices (Dst, F10.7).
  • Ephemeris-derived ($x_{eph}$): IPP coordinates (geographic and magnetic), elevation angle, and cyclical time encodings. These are computable from satellite ephemerides for future timesteps.

B. Ephemeris Conditioning (Key Innovation)

Since satellite orbits are near-deterministic (predictable up to 24 hours ahead with high accuracy), the authors construct the entire forecast-horizon graph in advance before any future observations are made.

• Mechanism: The model conditions its predictions on the known future graph topology ($G_{t+k}$).
• Benefit: This allows the model to reason about the spatial structure of the ionosphere at future times, even for new nodes (IPPs from satellites that rise during the forecast window) which have no historical observation data.
• Implementation: Historical embeddings are concatenated with future node encodings (derived solely from ephemeris). For new nodes, the history embedding is a zero-vector, and context is recovered via spatial message passing from neighboring observed nodes.

C. Architecture

The model consists of two modules:

1. History Module: Processes observed graphs ($t-23$ to $t$) using Graph Attention Networks (GAT) and temporal self-attention to generate node embeddings.
2. Prediction Module: Processes future graphs ($t+1$ to $t+24$) using ephemeris-derived features. It fuses historical embeddings with future node encodings and applies graph attention over the forecast topology.

3. Data and Experimental Setup

• Dataset: Multi-GNSS data (GPS, Galileo, BeiDou, QZSS, GLONASS) from a co-located receiver pair in Singapore (NTUS and SIN1) spanning January 2023 to April 2025 (Solar Cycle 25 high activity).
• Labeling Strategy: A zero-baseline QC approach was used. Events are labeled "Confirmed" if both receivers detect overlapping irregularities, "Invalid" if only one does (likely artifact), and "Unverified" if data is missing. This ensures high-quality ground truth for training.
• Task: Binary probabilistic classification of ROTI-defined irregularities (event vs. quiet) at a 5-minute cadence up to 2 hours ahead.
• Metrics: Brier Skill Score (BSS), ROC-AUC, PR-AUC, Probability of Detection (POD), and False Alarm Ratio (FAR).

4. Key Results

The model was evaluated against a persistence baseline and several ablation variants (without graph, without conditioning, without history).

• Overall Performance: IonoDGNN achieved a BSS of 0.49 and PR-AUC of 0.75, representing a 35% improvement in BSS and 52% in PR-AUC over the persistence baseline.
• Handling New Nodes (Ephemeris Conditioning):
  • For nodes appearing only in the forecast horizon (no history), the full model achieved a ROC-AUC of 0.95.
  • Without ephemeris conditioning, performance on these new nodes collapsed to near-random (ROC-AUC 0.52).
• Spatial Context: Removing graph message passing reduced BSS by ~22%, confirming that spatial relationships between simultaneous IPPs are critical for event discrimination.
• Lead Time: The model outperforms persistence significantly as lead time increases. At 2 hours, IonoDGNN maintains a BSS of ~0.35, while persistence drops to ~0.05.
• Robustness to Data Loss: Under simulated coverage dropout (up to 80%), the model retained predictive skill on affected nodes by leveraging spatial message passing from observed neighbors.

5. Significance and Contributions

1. Novel Representation: This is the first work to model ionospheric forecasting as a dynamic graph on native LoS observations, eliminating the need for intermediate gridding and the associated biases.
2. Ephemeris Conditioning: The paper introduces a new paradigm of conditioning forecasts on known future sampling structures. This solves the "cold start" problem for rising satellites and enables predictions on LoS paths that have never been observed in the training history.
3. Operational Resilience: The framework naturally handles variable observation density and data gaps (common in real-world GNSS networks) through spatial aggregation, offering a robust alternative to grid-based gap-filling methods.
4. Generalizability: The approach is not tied to specific observables (ROTI) and can be extended to other GNSS-derived metrics (scintillation indices, TEC) or heterogeneous sensor networks.

Conclusion

The paper demonstrates that treating GNSS observations as a dynamic graph with known future topology is a superior approach for forecasting ionospheric irregularities. By leveraging ephemeris conditioning, the model can accurately predict events on new satellite paths and maintain skill under data scarcity, offering a viable and more physically consistent alternative to traditional grid-based forecasting models.