IonCast: A Deep Learning Framework for Forecasting Ionospheric Dynamics

AI summary

Overview

Research area: Space weather / ionospheric physics combined with deep learning — specifically, global forecasting of ionospheric Total Electron Content (TEC) using graph neural networks and recurrent architectures.

Technical level: Advanced. The paper assumes familiarity with graph neural networks, autoregressive sequence modeling, message passing on icosahedral meshes, and ionospheric/space-weather terminology (TEC, GIM, SYM-H, Kp, solar wind drivers).

Scope: The paper introduces IonCast, a suite of deep learning models — a GraphCast-inspired graph neural network (IonCast GNN) and a convolutional LSTM baseline (IonCast LSTM) — that forecast global TEC autoregressively from heterogeneous space- and ground-based data, and benchmarks them against persistence and the empirical IRI model.

What This Paper Is About

The ionosphere (roughly 50 to 1500 km altitude) varies in ways that degrade GNSS accuracy, high-frequency communications, and aviation and power-grid operations, yet existing empirical models such as IRI and physics-based models such as GITM carry well-documented limitations. The authors ask whether modern deep learning architectures can learn global TEC dynamics directly from data, combining two-dimensional global maps of TEC with one-dimensional solar and geomagnetic drivers. Their goal is a modular, global nowcast-and-forecast framework that can complement or potentially substitute existing empirical and physics-based approaches.

Key Contributions

1. IonCast, a modular deep learning suite for global TEC forecasting, consisting of IonCast LSTM (a convolutional encoder–decoder with an LSTM bottleneck) and IonCast GNN (a GraphCast-inspired encoder–processor–decoder architecture built on the NVIDIA PhysicsNeMo framework in PyTorch).
2. A curated heterogeneous dataset combining 2D vertical TEC global maps from JPL GIM at 15-minute cadence with 1D solar-wind and geomagnetic driver time series, planetary activity indices, solar irradiance proxies, quasi-dipole magnetic coordinates, and orbital mechanics features (Sun/Moon ephemerides). The authors describe it as one of the most comprehensive datasets assembled for this type of analysis, and note that including auxiliary spatial features was not attempted in previous ML applications.
3. A like-for-like evaluation against persistence and IRI on held-out storm-time and quiet conditions spanning the NOAA geomagnetic storm scale (G0 through G5), with 10% of storms at each scale held out from training.
4. Ablation studies isolating the contribution of each input data class, and a separate study of how training on different solar cycle phases (solar maximum 2013–2015 vs. solar minimum 2018–2020 vs. the full 2010–2024 dataset) affects skill.

Main Findings

• IonCast GNN beats persistence at almost all lead times. The performance gain over persistence grows as the forecast horizon increases, and the GNN's RMSE stabilizes from forecast hours 6 to 12.
• IonCast GNN outperforms IRI up to a 6-hour forecast horizon. The authors note this is not a perfect comparison because IRI is not a forecast model. The reported IRI RMSE values in the appendix comparison are 5.32 (G0 event, 2018-04-20), 6.24 (G2 event, 2015-09-07), and 16.59 (G4 event, 2023-04-23), with IonCast LSTM and GNN values given at 1, 6, and 12 hours averaged over the full 48-time-step (12-hour) period.
• The LSTM degrades at long horizons. IonCast LSTM has a bias toward over-predicting TEC as the forecast proceeds and struggles at longer horizons compared with the GNN.
• Orbital mechanics and quasi-dipole magnetic coordinates are the single most valuable input class. In the ablation (Table 1, RMSE in TECU averaged over a 12-hour forecast on the validation set): JPLD alone 22.4 ± 3.2; JPLD + F10.7 23.9 ± 10.2; JPLD + F10.7, S10.7, M10.7, JB2008 13.3 ± 4.4; JPLD + Ap & Kp 12.7 ± 3.2; JPLD + Bx/By/Bz & vx/vy/vz (OMNIWeb) 15.5 ± 12.6; JPLD + Orbital Mechanics + Quasi-Dipole 9.2 ± 4.5 (best); JPLD + All (non-residual target) 18.8 ± 10.7; JPLD + All 10.7 ± 4.5.
• Auxiliary channels help, but not uniformly. Every data stream other than JPLD + F10.7 outperformed the model trained on TEC input alone. The F10.7-only model performed poorly, which the authors attribute to F10.7 being produced at daily cadence — too slow to inform 15-minute-resolution forecasts over a 12-hour horizon. The model trained on orbital mechanics plus quasi-dipole coordinates even outperformed the model trained on all data streams; the authors suggest this may be because forcings are excluded from the loss, reducing the number of channels over which the loss is computed and letting the model focus on the TEC channel.
• Without orbital mechanics and magnetic coordinate channels, forecasts drift. Models trained without those channels struggle to account for the apparent motion of the ionosphere over horizons longer than 4 hours, producing spatial drift and rising RMSE.
• Residual targets help substantially. Using a residual target (predicting x_T+1 = x_T + x̂_predicted) provided a notable improvement over the non-residual variant (18.8 ± 10.7).
• Solar cycle training window matters (Table 3). The solar maximum model (2013–2015) performs comparably to the full-dataset model, achieving the lowest RMSE for All Events (8.63 ± 4.34), G0 (5.79 ± 0.53), G2 (6.57 ± 1.57), and G4 (11.36 ±

Authors’ abstract

The ionosphere is a critical component of near-Earth space, shaping GNSS accuracy, high-frequency communications, and aviation operations. For these reasons, accurate forecasting and modeling of ionospheric variability has become increasingly relevant. To address this gap, we present IonCast, a suite of deep learning models that include a GraphCast-inspired model tailored for ionospheric dynamics. IonCast leverages spatiotemporal learning to forecast global Total Electron Content (TEC), integrating diverse physical drivers and observational datasets. Validating on held-out storm-time and quiet conditions highlights improved skill compared to persistence. By unifying heterogeneous data with scalable graph-based spatiotemporal learning, IonCast demonstrates how machine learning can augment physical understanding of ionospheric variability and advance operational space weather resilience.