Abstract Plasma convection at high latitudes is a key manifestation of magnetosphere–ionosphere coupling and serves as a critical indicator of space weather conditions. The Super Dual Auroral Radar Network (SuperDARN) provides essential data on ionospheric convection. Based on the dependence of ionospheric convection on six near‐Earth parameters (Bx, By, Bz, Vx, Pd, AE), we constructed three deep learning models using 2024 SuperDARN data: the multi‐layer perceptron (MLP), the MLP enhanced with multi‐head attention (MHA‐MLP), and the residual MLP (Res‐MLP). These models are designed to predict high‐latitude ionospheric plasma convection. The models are evaluated on an independent data set based on root mean square error (RMSE), structural similarity index (SSIM), and linear correlation coefficient (R). We also compared and analyzed the predicted distributions of cross‐polar cap potential and vortex spacing with SuperDARN observations. Furthermore, we conducted a qualitative analysis of the Res‐MLP's predictions under different typical space weather conditions. The Res‐MLP effectively mitigates the vanishing gradient problem in deep networks through residual connections and achieved the best prediction performance. On the independent testing set, Res‐MLP achieved an average SSIM of 0.54, R of 0.88, and RMSE of 5.64 kV. The experimental results indicate that Res‐MLP is an efficient method that can improve the convection prediction performance. This method efficiently predicts high‐latitude ionospheric convection patterns using only six near‐Earth space parameters plus the Universal Time, significantly reducing reliance on historical observations and improving real‐time performance. The model can be used for rapid, data‐driven assessment of high‐latitude space weather conditions and short‐term convection monitoring.
Abstract To address the complex spatiotemporal characteristics of ionospheric total electron content (TEC), a bidirectional convolutional long short‐term memory network with a dynamic cross‐attention mechanism for spatiotemporal feature weighting (CASF‐BiConvLSTM) is proposed and further combined with the sliding interquartile range (SIQR) method, referred to as CASF‐SIQR, for pre‐earthquake TEC anomaly detection. TEC data from the Center for Orbit Determination in Europe (CODE) are divided into training (1999–2014), validation (2015), and testing (2016–2018) sets. Compared with the BiConvLSTM baseline, the proposed CASF‐BiConvLSTM achieves consistent improvements across annual, seasonal, and latitudinal variations, with an average root mean square error reduction of about 21% during 2016–2018. For TEC anomaly detection, the CASF‐SIQR framework achieves significantly lower false rates during solar and geomagnetic quiet periods, underscoring the importance of accurate TEC prediction in effectively suppressing false anomalies. During the Jiuzhaigou and Luding earthquakes, more TEC anomalies were identified than with the traditional SIQR method, with detection rates improved by 22.23% and 13.29%, respectively. During the pre‐earthquake periods, the detection rates reach 40.83% and 32.78%, representing improvements of 23.61% and 13.73% over SIQR. During the strong geomagnetic storm on 26 August 2018, the validation reveals that the proposed framework can utilize the spatiotemporal characteristics of detected TEC anomalies to distinguish geomagnetic storm disturbances from pre‐earthquake anomalies. These results demonstrate the framework's greater sensitivity and effectiveness across geophysical conditions, offering valuable insights for TEC modeling, prediction, and detection of potential earthquake‐related anomalies.
Abstract This paper presents a major update to the PyIRI model, a fully vectorized Python implementation of the International Reference Ionosphere (IRI). The update reformulates the IRI core coefficients using spherical harmonics in Quasi‐Dipole (QD) latitude and Magnetic Local Time coordinates, replacing the legacy geographic‐coordinate‐based basis functions of the CCIR (Consultative Committee on International Radio) and URSI (International Union of Radio Science) models of the foF2 critical plasma frequency parameter. The resulting coefficients remain valid over decadal timescales, with re‐derivation recommended every 5 years to account for magnetic pole drift. This new formulation removes the dependence on the modified dip angle, enabling a more transparent and physically interpretable representation of ionospheric climatology. Diurnal variability is expressed through real Fourier series, with the optimal number of spatial and temporal coefficients determined via reconstruction error analysis. The coefficients are stored in standardized NetCDF format and validated against IRI‐2020. Artificial small‐scale oscillations present in the original IRI climatology are substantially reduced in the new formulation, leading to an ionosphere more consistent with the SAMI physics‐based model. The updated PyIRI model also includes previously unavailable IRI parameters, such as the SHU‐2015 and AMTB2013 options for hmF2, and the B0 and B1 F2‐layer thickness and shape parameters. Spherical harmonic coefficients are extracted identically across the foF2 CCIR, foF2 URSI, hmF2 SHU‐2015, hmF2 AMTB2013, M(3000)F2, and foEs parameters.
Abstract Geomagnetic disturbances (GMDs) are rapid changes in the magnetic field of the Earth that may drive geomagnetically induced currents (GICs), known to cause damage to infrastructure such as power grids and pipelines. Globally, GIC measurements are often hard to obtain; therefore, GMDs are used as a proxy for GICs. GMDs have been previously shown to occur on different timescales which are related to different formation mechanisms such as magnetic storms, substorm activities and sudden commencements. 1 min cadence ground‐magnetometer data has been available for decades and are the basis of many studies showing the frequency and intensity of GMDs. In recent years one‐second cadence data have become available from a network of magnetometers through the SuperMAG portal. We take the 1‐s data, averaged 10 s data, and compare the number of GMDs identified with the number found from the 1‐minute data to test the GMDs related to the different timescales. We find that 1‐s data consistently record higher amplitude GMDs as well as recording overall more GMDs, thus showing the importance of 1‐s cadence data. This is most important for GMDs between 00 and 03 magnetic local time and at low geomagnetic latitudes (>50°) as these are the locations where difference between the number of GMD detections between the 1‐s data and the 1‐min data is the largest. However, the 1‐s data have many artefacts due to unreliable data points in them leading a 10‐s median being more reliable at correctly detecting GMDs.
This study develops a semi-empirical framework for Dst index forecasting by integrating the PPMLR-MHD model of Earth's magnetosphere with four established empirical models (Burton, UCB, AK2, and Wang). The framework partitions the physics: global magnetopause and tail currents are simulated by MHD, while the ring current contribution is provided empirically. Performance was evaluated using Prediction Efficiency (PE) and Correlation Coefficient (CC) across 18 storms categorized as moderate (-50 > Dst >= -100 nT), intense (-100 > Dst >= -200 nT), and super (Dst < -200 nT). Key results show that all hybrid models improve correlation compared to their standalone empirical counterparts, but the optimal model for amplitude accuracy is intensity-dependent. The "MHD + Burton" model excels for moderate storms (avg. PE = 0.71, CC = 0.93). Conversely, for superstorms, only the "MHD + Wang" model is reliable (avg. PE = 0.86, CC = 0.95), while "MHD + Burton" fails. The superior performance of "MHD + Wang" under extreme conditions is attributed to its physics-based, solar wind dynamic pressure-dependent decay time. This work validates a practical, intensity-adaptive hybrid approach that combines physical fidelity with computational efficiency for improved operational Dst forecasting. Plain Language Summary Predicting the Dst index, a key measure of geomagnetic storm strength, is vital for protecting technology and infrastructure from space weather. We introduce a new forecasting method that combines a computer model simulating Earth's magnetic environment (an MHD model) with different statistical models of the ring current-a major storm driver. Testing on recent storms shows that no single combination works best for all events. For common, moderate storms, pairing the MHD model with the simpler Burton model yields the most accurate forecasts. However, for rare but extreme superstorms, the MHD model must be paired with the more complex Wang model to avoid severe overpredictions and produce reliable forecasts. This is because the Wang model better represents how the ring current behaves under extreme solar wind conditions. Our findings provide forecasters with a clear, adaptive strategy: use the MHD + Burton model for typical conditions and switch to MHD + Wang when extreme solar activity is expected. This practical approach balances detailed physics with speed, offering a promising tool for real-time warnings of severe space weather.
