IonCast: A Deep Learning Framework for Forecasting Ionospheric Dynamics

TL;DR

IonCast tackles the challenge of global ionospheric TEC forecasting by integrating diverse observations and drivers with deep learning. It introduces a LSTM-based and a GraphCast-inspired GNN model to learn spatiotemporal TEC dynamics on a global 2D map, using forcings like Sun/Moon positions and orbital mechanics. The results show that the GraphCast-inspired GNN provides superior long-horizon TEC forecasts, with ablations clarifying the pivotal role of orbital mechanics and quasi-dipole coordinates, and it outperforms the empirical IRI model up to six hours. These findings demonstrate the feasibility and potential operational impact of data-driven ionospheric forecasting for space weather resilience.

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.

Figures (7)

• Figure 1: Model predictions plotted and evaluated over various forecast lead times for a moderate geomagnetic storm (G2 event). Left: Subplots show ground truth data from JPL (upper) plotted against predictions by the IonCast LSTM (middle) and IonCast GNN (lower) models at a 60-minute lead time. The colorbar has units TECU (TEC Units). Center: The joint distribution between JPL (x-axis) and predictions (y-axis) over a 12-hour long-horizon forecast. The diagonal indicates perfect prediction; points below (above) the diagonal correspond to underprediction (overprediction). The colorbar encodes the mode of the forecast time (minutes) at which the prediction occurs, highlighting the error growth across the sequence. Marginal histograms show the distributions of ground truth and predicted values throughout the forecast. Right: The plot compares the average global RMSE of the IonCast LSTM (blue), IonCast GNN (orange), and persistence model predictions across growing forecast horizons.
• Figure 2: Model predictions plotted and evaluated over various forecast lead times for a moderate geomagnetic storm (G2 event). The IonCast GNN model shown is the same as in the main body. The IonCast LSTM model is trained on a longer context window (16 15-minute cadence time steps) and a larger amount of training data (trains on every 32nd sequence of 4 hours, thus sequences skip 8 hours between start and end dates) than the 8-context and 256 date dilation model evaluated in the main text. Left: Subplots show ground truth data from JPL (upper) plotted against predictions by the IonCast LSTM (middle) and IonCast GNN (lower) models at a 60-minute lead time. The colorbar has units TECU (TEC Units). Center: The joint distribution between JPL (x-axis) and predictions (y-axis) over a 12-hour long-horizon forecast. The diagonal indicates perfect prediction; points below (above) the diagonal correspond to underprediction (overprediction). The colorbar encodes the mode of the forecast time (minutes) at which the prediction occurs, highlighting the error growth across the sequence. Marginal histograms show the distributions of ground truth and predicted values throughout the forecast. Right: The plot compares the average global RMSE of the IonCast LSTM (blue), IonCast GNN (orange), and persistence model predictions across growing forecast horizons.
• Figure 3: Model predictions plotted and evaluated over various forecast lead times for a quiet ionospheric condition in the top figure (G0 event), a moderate (G2) and a severe geomagnetic event (G4 event) in the mid and bottom figure. Left and Center panels have the same models and meaning of the one shown in Figure 1. Right: The plot on the right compares the RMSE of the IonCast LSTM (blue), IonCast GNN (orange), and persistence model predictions across growing forecast horizons, split across low ($30^{\circ}$ South to $30^{\circ}$ North), mid ($60^{\circ}$ to $30^{\circ}$ South and $30^{\circ}$ to $60^{\circ}$ North), and high latitudes ($60^{\circ}$ to $90^{\circ}$ South and $60^{\circ}$ to $90^{\circ}$ North).
• Figure 4: Solar cycle model predictions at a 360-minute lead time for a G4 level event. (Left) JPLD Ground Truth, (Middle) Model trained on solar maximum date range (2013-2015), (Right) Model trained on solar minimum date range (2018-2020).
• Figure 5: Distribution of G-level events over the data range (2010-2024). The x and y axes represent the time (years) and the intensity of the event (G-level), respectively. Each class bin in the y-axis is then divided into four segments, which correspond to the event duration, as shown in the lower part of the plot.