LSTM-based multi-GNSS radio occultation fusion models global electron density

Scientists have unveiled a new deep learning framework that transforms two decades of satellite radio occultation measurements into a global, three-dimensional map of the ionosphere’s electron density, achieving accuracy that surpasses the widely used International Reference Ionosphere (IRI) empirical model when forecasting conditions beyond its training window. The study, published in Astrophysics and Space Science, harnesses long short-term memory (LSTM) networks to fuse multi-constellation GNSS radio occultation data from 2001 through 2019 with indices of solar and geomagnetic activity, and then tests how well the model can extrapolate to the year 2020 — a period it had never seen. The results suggest that recurrent deep learning architectures can capture the memory-laden, slowly evolving physics of the ionosphere better than conventional statistical or empirical approaches, with implications for satellite navigation, radar operations, and space weather forecasting.

The ionosphere, the shell of charged particles extending from roughly 60 to 1,000 kilometers above Earth, is far more than a scientific curiosity. It bends, delays, and scatters the radio signals that GPS, Galileo, BeiDou, and other navigation constellations beam to the ground, and its variability can degrade satellite communication, introduce positioning errors, and even pose risks to low-Earth-orbit spacecraft through increased atmospheric drag. Modeling the three-dimensional distribution of electron density — the fundamental quantity that governs how radio waves propagate through the charged upper atmosphere — has therefore been a decades-long challenge. Traditional empirical models such as IRI synthesize decades of ground- and space-based observations into climatological averages, but they struggle to respond dynamically to the day-to-day and hour-to-hour swings driven by solar radiation and geomagnetic storms.

The new research, led by Yutian Chen of Huaiyin Normal University together with Changzhi Zhai of China University of Geosciences (Wuhan), Dongjie Yue of Hohai University, and Xiangyu Tian of Jiangsu Ocean University, takes a fundamentally different approach. Rather than encoding climatological averages, their LSTM model learns the temporal dynamics of electron density directly from observations. The training corpus draws on radio occultation profiles from three satellite missions: COSMIC-1 (Constellation Observing System for Meteorology, Ionosphere, and Climate), CHAMP (Challenging Minisatellite Payload), and GRACE (Gravity Recovery and Climate Experiment). These data, archived by the COSMIC Data Analysis and Archive Center (CDAAC), provide vertical electron density profiles retrieved whenever a GNSS signal from GPS, GLONASS, or other constellations slices through the ionosphere as seen from a low-Earth-orbiting satellite — a technique that yields globally distributed, all-weather measurements of the ionosphere’s vertical structure.

Radio occultation works on a simple but powerful principle. As a GNSS satellite sets or rises behind Earth’s limb from the perspective of a low-Earth orbiter, the transmitted radio signals traverse progressively deeper layers of the atmosphere before reaching the receiver. The charged particles of the ionosphere introduce a frequency-dependent delay in the signals, and by precisely measuring these delays, scientists can invert them into a vertical profile of electron density from the satellite’s orbital altitude down through the ionospheric layers. Because occultations occur thousands of times per day across the entire globe, they offer a uniquely comprehensive view of the ionosphere’s three-dimensional structure — one that ground-based instruments, anchored to fixed locations, cannot match. Merging nearly two decades of such occultations from COSMIC-1, CHAMP, and GRACE produced the long, continuous record needed to train a recurrent neural network to recognize how the ionosphere evolves through the solar cycle.

The team augmented the occultation profiles with solar and geomagnetic activity indices, the external drivers that modulate ionization on timescales ranging from minutes to years. Solar indices capture the varying extreme-ultraviolet flux that creates ionospheric plasma, while geomagnetic indices quantify the disturbances that ripple through the upper atmosphere when solar wind and interplanetary magnetic field structures strike Earth’s magnetosphere. Feeding these indices alongside the occultation-derived electron densities allowed the LSTM network to learn the causal linkages between space environment and ionospheric response. The choice of architecture was deliberate: LSTM networks contain gating mechanisms — input, forget, and output gates — that regulate what information is retained or discarded across time steps. This solves the vanishing gradient problem that plagues standard recurrent neural networks (RNNs) when modeling long sequences, making LSTMs especially suited to phenomena like the ionosphere, whose state today depends on its state hours and days earlier.

To test whether the learned dynamics generalize, the researchers evaluated the models’ extrapolation capability for 2020, a year entirely absent from training. Verification relied on vertical electron density profiles measured by eight ionosondes distributed across the globe — ground-based radars that pulse high-frequency signals upward and infer electron density from the reflected echoes. The LSTM model did not merely perform adequately; it outperformed the IRI model overall, registering the lowest errors of all approaches tested. The comparison also included a gated recurrent unit (GRU) network, a streamlined relative of the LSTM with fewer internal gates, and a standard RNN. On the test set, the LSTM achieved correlation coefficients exceeding 0.9 against the reference data, beating both the GRU and the RNN and confirming that the full gating architecture confers a genuine advantage when modeling ionospheric memory.

The error analysis revealed instructive spatial structure. Root mean square error (RMSE) followed a clear altitude-dependent pattern, peaking between 250 and 350 kilometers — the region surrounding the F2 layer, where electron density reaches its maximum and where the physics of plasma production, loss, and transport are most vigorously contested. Errors also increased at low latitudes, a reflection of the extraordinary complexity of equatorial ionospheric dynamics, where the eastward electric field drives the equatorial plasma fountain, lifting plasma upward and outward to create the equatorial ionization anomaly — twin crests of enhanced density flanking the magnetic equator. These regions have long challenged both empirical and physics-based models, and the finding that deep learning errors concentrate there identifies precisely where future modeling efforts, whether through richer input features or hybrid physics-informed architectures, should focus.

Independent validation came from an entirely different observing system: the European Space Agency’s Swarm-A satellite, which carries instruments that measure in-situ electron density along its polar orbit at approximately 450 kilometers altitude. The team compared model predictions against Swarm-A measurements under both geomagnetically quiet and disturbed conditions. During quiet periods, correlations between the models and the satellite observations were strong across the board, with R values exceeding 0.8 for all three deep learning architectures, and RMSEs remaining below 0.04 × 10⁶ electrons per cubic centimeter. When geomagnetic storms roiled the ionosphere, correlations stayed comparably high — a noteworthy result, since storm-time prediction is where empirical models most often fail — but the RMSEs rose above 0.06 × 10⁶ electrons per cubic centimeter, confirming that storm-time ionospheric variability remains the discipline’s toughest test. That the deep learning models preserved their correlation structure even during disturbances hints that the networks had internalized, at least partially, how geomagnetic forcing reshapes the global plasma distribution.

The significance of the work extends beyond a single model benchmark. The ionosphere is coupled to space weather in ways that turn solar eruptions into terrestrial technological hazards: geomagnetic storms can disrupt high-frequency communications, degrade satellite navigation accuracy, endanger crewed and uncrewed spaceflight, and, through thermospheric heating and expansion, increase drag on satellites in low Earth orbit. Accurate, global, three-dimensional electron density models are a prerequisite for mitigating these risks, and the demonstrated ability of an LSTM trained on historical occultations to extrapolate into unseen years suggests a path toward operational, data-driven ionospheric forecasting. The authors note that all data underlying the study are publicly available — occultation records from CDAAC, ionosonde measurements from NOAA’s National Centers for Environmental Information, Swarm-A observations from ESA, and geomagnetic indices from NASA’s OMNIWeb — lowering the barrier for other groups to build upon the framework.

The study also situates itself within a rapidly growing body of machine learning applications to ionospheric science. Prior efforts have used artificial neural networks to model the F2 layer’s critical frequency, forecast total electron content, and construct three-dimensional climatological models such as ANNIM-3D, which was trained on long-term ionospheric observations. Related work by overlapping author groups has applied ConvLSTM networks to reconstruct ionospheric disturbances from hurricanes and to characterize the global ionospheric consequences of the January 2022 Tonga volcanic eruption. What distinguishes the new study is its combination of multi-mission GNSS occultation data spanning nearly two decades, a genuinely three-dimensional electron density target, rigorous out-of-sample extrapolation testing against independent ground truth, and a systematic head-to-head comparison of recurrent architectures under both quiet and storm-time conditions.

Challenges remain before such models can fully replace or complement physics-based ionospheric modeling in operations. The increased errors at low latitudes and at F-region heights reflect the intrinsic difficulty of capturing equatorial electrodynamics from data alone, and storm-time error growth indicates that extreme events — precisely those with the greatest societal impact — still elude the model’s grasp. The training record, ending in 2019, also spans a specific phase of the solar cycle, and how well the network generalizes to the approaching maximum of Solar Cycle 25 remains to be tested. Nevertheless, the study demonstrates convincingly that recurrent deep learning can distill the memory-driven behavior of the ionosphere from observational data into a predictive tool that outperforms the field’s standard empirical benchmark, marking a meaningful step toward data-driven, global, three-dimensional space weather forecasting.

As constellations of GNSS and low-Earth-orbit satellites continue to multiply, the volume of radio occultation data available for training such models will only grow, promising ever richer representations of the ionosphere’s structure and variability. If deep learning architectures like the one demonstrated here can be coupled with real-time data streams and solar wind observations, the prospect of minute-scale, global ionospheric nowcasting — and with it, more resilient navigation, communication, and satellite operations — moves from aspiration toward engineering reality.