Space weather dependence of decameter-scale ionospheric irregularities spatial distribution

1 Introduction

• Prior irregularity models include autoregressive, regression, and physical approaches; equatorial examples include SCINDA and GIM scintillation models.WB-MOD is another equatorial model that was modified for high latitudes.Source: §1 · p. 2
• Sparse HF radar coverage forces the use of low-parameter models, while prior work includes spherical-harmonic and IMF-conditioned approaches.The introduction also notes regression-based multiscale prediction at individual measurement locations.Source: §1 · pp. 2–3
• The paper organizes model design around selecting inputs, representing their temporal dynamics, and approximating spatial distributions where observations are absent.These are presented as three central problems in constructing spatiotemporal models of HF irregularities.Source: §1 · pp. 3–4
• Input selection is framed as either choosing parameters from physical assumptions or using data-driven analysis to rank measured parameters.The introduction also describes regression, delay-based, and multiscale approaches to representing temporal dynamics.Source: §1 · p. 3
• For spatial approximation, prior approaches include interpolation or spherical harmonics, coordinate-input models, and predictions limited to observed locations.The introduction presents these as alternatives for regions without observations.Source: §1 · p. 4
• The proposed model predicts high-latitude irregularity occurrence from SuperDARN and SECIRA data, using multiscale pattern recognition and spatial coordinates as inputs.Its input parameters are selected as a minimal set optimized for model quality.Source: §1 · p. 4

2.1 Data used

• The study uses radar data from 2010–2023 for training and 2024–2025 for validation, selecting at least two years per radar to span relatively high and low solar activity.The authors note that computing limits constrained the dataset and suggest applying the approach to a wider dataset in future.Source: §2.1 · p. 4

2.2 Problem formulation

• The prediction target is the probability of ionospheric scattering under space-weather conditions and at a location, framed to reduce effects of nonuniform radar coverage, operating intervals, and mode changes.The formulation assumes an existing irregularity will be detected under the operating conditions of the SuperDARN and SECIRA coherent-scatter HF radars.Source: §2.2 · pp. 6–7

2.3 Detection of ionospheric scattering

• The three detection algorithms yield distinct echo-probability patterns: 36C gives higher probabilities, CA lower probabilities, and FA intermediate results.CA shows the equatorial boundary of the convection zone reaching its highest latitude around 12 MLT; the authors therefore consider FA useful for analyzing common features and differences.Source: §2.3 · p. 9

2.4 Model input parameters θs

• Hourly-averaged space-weather parameters and spatial coordinates, including MLAT, MLT, and GLAT, form the model inputs.Input quantile transformation interpolates missing values using the previous good value and maps inputs to a bounded range; random augmentation is also applied.Source: §2.4 · pp. 9, 12, §2.5 · pp. 12–13

2.5 Optimal coordinate system

• Geomagnetic latitude (MLAT) is preferable to geographic latitude (GLAT) as a model input, almost regardless of the ionospheric-scattering detection method.This agrees with established approaches to analyzing high-latitude scattering regions.Source: §2.5 · p. 14

3.1 Model architecture

• The RN-NN represents sparse, variably ordered radar observations as dynamic graphs and simplifies the graph structure to a space-weather root linked to radar-data nodes.Traditional ordered-data networks and generative diffusion networks are described as unsuitable for this dataset.Source: §3.1 · p. 15
• The model combines geomagnetic coordinates at radar nodes with a recurrent history of space-weather parameters, then uses attention to predict scattering probability at each node.The space-weather node’s history passes through an LSTM-based recurrent network; node predictions are computed simultaneously and are independent of processing order.Source: §3.1 · pp. 16, 18
• Separate multi-head recurrent filters learn temporal patterns for each input parameter from a 31-day history, supporting later analysis of parameter importance.Each parameter has six experimentally selected heads; the RN and NN are trained jointly.Source: §3.1 · pp. 15, 18, §3.2 · p. 19

3.2 Model training and simplification

• Training addresses incomplete radar coverage by adding zero-occurrence samples at low magnetic latitudes; training data from 2010–2023 precede validation data from 2024–2025.The model typically converges in fewer than 50 epochs.Source: §3.2 · pp. 19–20
• The 10–72-hour lead-time range has the highest errors and should not be used; nowcasting has the highest quality, with comparable performance through two hours.Figure 4 distinguishes nowcast/short-term, mid-term, and long-term regions; forecast error increases with lead time.Source: Figure 4 · p. 21, §3.2 · p. 21
• The eight selected space-weather parameters perform no worse than the original model on validation data over 0–9 hours, while further input reduction lowers quality.The selected set is ap, Kp, Psw, Esw, f Bz, Tsw, AE, and F10.7.Source: §3.2 · p. 23

• The observation-probability model predicts signal presence on an MLAT–MLT grid, using an RN-NN trained on binary maps of whether any radar signal was received.The grid spans MLAT 40–90 and MLT 0–24; zero marks no received signal and one marks at least one signal.Source: §4 · pp. 23, 25
• Predicted observation probability has two peaks near 55 and 80 MLAT, a minimum between them, and is zero below 50 MLAT.The plotted mean and standard deviation use roughly equal numbers of randomly selected profiles at each Kp level to address dataset imbalance.Source: §4 · p. 25
• Because predicted observation probability varies little with space-weather state and MLT, the subsequent analysis uses a stationary latitude profile averaged across available conditions.Typical P(OBS|MLAT) ranges from 0.2 to 0.45 over 55–90 MLAT.Source: §4 · pp. 25–26

5 Resulting model P(IS|θs) and discussion

• The final probability model multiplies the tabulated radar observation probability by the probability that an observed scattered signal is ionospheric scattering.The authors state that the model’s significant parameters do not contradict known physical observations or parameters controlling substorm development and high-latitude convection.Source: §5 · pp. 26, 29
• During the August 11–13, 2024 storm, nowcast predictions agreed well with observations and tracked a daytime probability decrease near local noon.The validation case also showed equatorward boundary movement alongside increasing geomagnetic disturbance.Source: §5 · p. 26, Figure 8 · p. 27
• As Kp, AE, and Esw increase, the maximum latitude of the equatorial boundary moves equatorward.The relationship is among three dependency patterns reported for the 10–12 MLT sector.Source: §5 · p. 26, Figure 9 · p. 28
• As F10.7, Tsw, and f Bz increase, the maximum latitude of the equatorial boundary moves poleward.The relationship is reported for the nowcast model in the 10–12 MLT sector.Source: §5 · p. 26, Figure 9 · p. 28
• Psw and ap show no unambiguous direct correlation with boundary latitude, although their strong dependence suggests temporal dynamics affect the result.The paper also describes Esw as having a strong trigger effect.Source: §5 · pp. 26, 29

6 Conclusion

• The proposed data-driven, multiscale model uses a neural network inspired by graph neural networks and recognizes patterns in space-weather parameter dynamics.It predicts the probability of high-latitude decameter ionospheric irregularities from SuperDARN and SECIRA radar data and space-weather parameters.Source: §6 · p. 29
• The model’s optimal performance uses MLT and MLAT coordinates and combines geomagnetic and solar indices with solar-wind and interplanetary magnetic-field parameters.The data cover 11 radars from 2010–2025, and all input parameters use hourly means.Source: §6 · p. 29
• Acceptable forecast quality extends to 7–9 hours, with the highest quality at zero lead time.f Bz and F10.7 are significant only for nowcasting; AE and Tsw are significant above 2 hours and can be neglected for short-term predictions.Source: §6 · pp. 29–30
• The reported statistical relationships between boundary maximum latitude and model parameters show good correlation with previously known effects.The passage does not quantify the relationships.Source: §6 · p. 30