Abstract The polarity inversion line (PIL) in active regions (ARs) is considered to be closely associated with solar flare eruptions. In this study, we rigorously constructed standardized data sets based on time series of different lengths using Space‐weather HMI Active Region Patches (SHARP) parameters calculated along the PIL. We compared the performance of traditional non‐sequential models and a time‐series model in solar flare prediction tasks, as well as the predictive performance of time‐series models with different input lengths within the CNN–BiLSTM–AT framework. The main findings of this study are summarized as follows: (a) SHARP parameters computed along the PIL consistently yield better prediction performance than those calculated over entire active regions. (b) In realistic and highly imbalanced prediction scenarios, the time‐series model outperforms non‐sequential models, achieving an F1 score of 0.59 for strong‐flare prediction. (c) Robustness tests and sliding‐window probability forecasts further demonstrate the practical feasibility of the proposed approach. These results provide useful guidance for data representation and model selection in solar flare forecasting.

The Global Positioning System (GPS) satellite constellation has provided long-term observations of energetic electrons in Earth’s radiation belts, but inconsistencies among different satellites limit the direct use of their combined measurements. A comparative analysis of electron flux data from year 2000 to 2020 reveals significant deviations for several satellites, particularly NS41 and NS48 in low-flux regions, as well as scattered observations from satellites such as NS74 and NS69. These discrepancies highlight the necessity of cross-calibration to ensure data consistency. To address this, we conducted the first systematic cross-calibration of energetic electron fluxes from 25 GPS satellites. Taking the 2.0 MeV average unidirectional differential electron fluxes and the ≥ 2.0 MeV omnidirectional integral electron fluxes as examples, we adopted the conjunction method in magnetic coordinates (Lm, B/B_0 ) (Lm: McIlwain L-shell parameter, B/B_0 : magnetic field ratio) and applied cubic polynomial fitting to achieve unified calibration using NS59 as the reference. For the 2.0 MeV differential electron fluxes, the Root-Mean-Square Deviation (RMSD) before calibration is on average 3.08 times larger than that after calibration, while the Correlation Coefficient (CC) increases by a factor of 1.14 on average. For the ≥ 2.0 MeV integral electron fluxes, the corresponding values are 1.68 and 1.01, respectively. This method can be extended to other energy channels and satellites. The calibrated dataset facilitates quantitative analysis and modeling of variations in the high-energy electron distribution in medium Earth orbits.

Abstract We investigate the response of thermospheric neutral density and associated Joule heating (JH) with different interplanetary structures during 73 intense geomagnetic storms. Neutral density is derived from accelerometers onboard CHAMP, GRACE‐A, SWARM‐C, and GRACE‐FO satellites during 2001–2024. We estimate the JH variation using the data from the DMSP spacecraft and the Weimer electric potential model. Four typical interplanetary structures identified as the primary drivers of intense storms (Dst ≤ −100 nT) are analyzed in terms of their influences on neutral density and JH: isolated interplanetary coronal mass ejections (ICMEs) (29% storms), successive ICMEs (44% storms), interacting ICME‐HSS structures (26% storms), and high‐speed streams (HSSs) alone (1% storm). The middle two categories represent the complex interplanetary structures. The results show that neutral density and associated JH were significantly enhanced during the May 2024 superstorm driven by successive ICMEs. The complex interplanetary structures, including successive ICMEs and interacting ICME‐HSS structures, lead to significant and prolonged neutral density enhancements during intense storms, in contrast to isolated ICMEs. During successive ICME‐driven storms, enhanced interplanetary magnetic field and solar wind (SW) are associated with increased JH, which has a strong impact on storm‐time neutral density. The interaction between ICMEs and HSSs causes the same distribution in density intensity as that during isolated ICME‐induced intense storms but substantially extends the duration of density enhancement, plausibly resulting from enhanced SW due to interacting ICME‐HSS structures.

Abstract Coronal mass ejections (CMEs) are among the key solar eruptive activities, triggering space weather disturbances. Thus, forecasting their geoeffectiveness has become a research focus. This study constructs a model to recommend similar events for forecasting the geoeffectiveness of CMEs. The input parameters are optimized via feature dimensionality reduction, while the cosine similarity algorithm, combined with logistic regression, is employed to perform four tasks: binary classification of whether a CME will reach Earth, prediction of travel time, assessment of the intensity of geomagnetic disturbances, and matching of historical similar events. Experimental results show that the model achieves an F1‐score of 0.43 for binary classification, with a mean absolute error (MAE) of 13.75 hr for travel time prediction and a MAE of 1.68 for the maximum Kp value (Kp max) on the test set. The findings indicate that the model not only achieves staged multi‐task forecasting for the geoeffectiveness of CMEs but also enhances the interpretability of forecasting results through similar event recommendation, providing pragmatic decision support for space weather warnings.

The geodynamics and plate tectonics of the South China Sea (SCS)-Taiwan region since Miocene times are uncertain because the former extent and tectonic configuration of the subducted easternmost SCS along the Manila trench is uncertain. Here we unravel the regional kinematic context from main offshore constraints including published unfolding of the Manila slab from seismic tomography, which provides insight on restoring the subducted part of the SCS. We reconstruct a whole northern SCS continent-ocean boundary (COB) that consists of a northeastern SCS COB segment (called ‘S3’), trending N070° that roughly parallels the present SCS shelf; a 350-km long ~N-S trending segment S2 that steps north to Hualien; and, a third segment S1 that extends from east of Hualien beneath the Ryukyu subduction zone trending N085° that ends near Miyako Island in the Ryukyus.We demonstrated that the two-plate kinematic model is the best framework to explain the existing data. The boundary between Eurasia and Philippine Sea plate is a ~NS oriented left-lateral lithospheric shear fault called the Manila transcurrent fault (MTF). The MTF initiated ~18 Ma at the onset of the tear and progressively moved eastward, creating the intra-oceanic Luzon arc. The MTF ancestor was a N337° oriented left-lateral shear fault, while the HB-PSP/EU motion was changing to N307°, allowing the HB-PSP plate to subduct between the two lips of a westward propagating tear fault until ~7 Ma. Since ~7 Ma, the MTF-PFZ constantly moved westward and 23° clockwise rotated from N337° to ~NS, which began collision ~7 Ma ago along the EU margin. Plate kinematic reconstructions from ~18 Ma to Present are synthesized in terms of continental or oceanic nature of the main PSP-HB and EU entities before their subduction that provide new understanding on Taiwan, PSP-SCS kinematics, and regional histories. This work is supported by Key projects of the Chinese National Natural Science Foundation (contracts 91958212, 42106078).

With the increasing number of large constellations, it is crucial to accurately predict satellite positions and movements using upper atmosphere models driven by geomagnetic indices. Machine learning (ML) can quickly provide geomagnetic index predictions. However, previous research using ML to forecast Stream Interaction Region-driven (SIR-driven) geomagnetic storms primarily follows two approaches: using real-time observations at the L1 point with historical geomagnetic indices, which lacks solar source information, or using solar images as input. To forecast SIR-driven storms over longer horizons (3 days), a neural network model is constructed by incorporating coronal magnetic field physical parameters and the coronal hole features extracted from solar images (Pch ${P}_{\text{ch}}$). We first select the appropriate combination of coronal magnetic parameters for forecasting geomagnetic indices. Using the Explainable AI technique, we find that the model can learn the propagation characteristics of interactions between high- and low-speed solar wind from historical coronal parameters, which relate to the physical process of SIR formation, providing a physical basis for the approach. We then introduce Pch ${P}_{\text{ch}}$ and the historical Kp index to develop an optimal model. Based on accuracy evaluations across different activity levels and solar cycle phases, the model has proven to provide reasonable forecast results for SIR-driven events over the next 2-3 days. The model outperforms an empirical model and two ML models in terms of event detection ability for horizons longer than one day. This study offers a new approach for 3-day geomagnetic disturbance forecasting using a deep neural network.

Using new observations from the Chinese Meridian Project (CMP), this study examines the characteristics of neutral winds in the East Asian sector during the Mother's Day super‐intense storm in May 2024, primarily focusing on its effects on the disturbances over China and adjacent areas. It is the first time that new measurements from three Dual‐Channel Optical Interferometers (DCOIs) are utilized to analyze storm‐time neutral winds at an altitude of 250 km in northern China. By developing all northern‐hemisphere Super Dual Auroral Radar Network radars including six Chinese Dual Auroral Radar Network (CN‐DARN) radars, the newly derived ionospheric convection pattern and its impacts on neutral winds can be well analyzed. The results show that a strong equatorward wind with a maximum amplitude of ∼400 m/s in the meridional component was observed for the first time during the storm main phase. In the East Asian sector, a negative ionospheric storm over China and adjacent areas was accompanied by this enhancement in equatorward wind in the night of May 10. Additionally, ionospheric convection expanded to 43° magnetic latitude (MLAT) with eastward ion velocities exceeding 800 m/s around 50° MLAT. This can strengthen zonal wind in northern China, producing a notable eastward surge of ∼230 m/s measured by new DCOIs in the dawnside sub‐auroral region. Wave‐like oscillations in neutral winds were observed by multi‐DCOI stations, which were associated with the storm‐time Traveling Atmospheric Disturbances (TADs). During the storm recovery phase, a high‐level total electron content cluster shifted from eastern China to the central regions, which may be attributed to enhanced ∑[O]/[N 2 ].

The rapid development of low‐Earth orbit (LEO) satellites brings increased attention to spacecraft collisions, space debris, orbital decay, and satellite reentry. Neutral density and associated drag force on the satellite orbits elevate space risks, significantly determined by space weather disturbances, particularly geomagnetic storms. On 3 September 2024, the Australian Binar‐2, 3, and 4 satellites were deployed, but their actual lifetimes were only around 2 months, much shorter than designed, only 20% that of the nearly identical Binar 1 satellite launched in 2022. For the first time, we analyze the premature reentries of the Binar‐2, 3, and 4 satellites to unveil the severe space weather impact on their lifetimes, especially the influence of accuracy of medium‐ and long‐term space weather prediction on satellite lifetime designs. Our findings reveal that the premature reentries of Binar‐2, 3, and 4 satellites were caused by enhanced neutral density due to much higher solar and geomagnetic activities than predicted. The actual satellite lifetimes align well with estimations based on observed space parameters, whereas large deviations occur when using predicted parameters. The widely used predictions for medium‐ and long‐term space weather underestimated the F10.7 index and sunspot numbers during Solar Cycle 25 (SC25), especially during the peak time, leading to discrepancies in the designed satellite lifetimes. Our results illustrate the importance of medium‐to long‐term space weather forecasting for the lifespan of LEO satellites.

The coastal zone, a critical interface between land and sea, is characterized by intense human activity and development, making it a key area for multidisciplinary research. However, achieving accurate shallow subsurface detection in these areas, including nearshore waters, remains challenging due to limitations in current exploration technologies and the distinct challenges of marine and land-based survey methods. This study tackles this issue by investigating the western coastline of Dong'ao Island, Zhuhai. A network of 22 node seismograph stations were deployed across varied coastal environments—hillside, beach, and seafloor—to systematically analyze seismic ambient noise characteristics in each setting. The research aims to enhance understanding of seismic noise in different coastal contexts, contributing to improved techniques for shallow subsurface detection in coastal zones. Key findings from the analysis can be summarized as follows:Submarine stations exhibit higher seismic ambient noise energy levels compared to their hillside and beach counterparts. Dominant frequency bands concentrate within the ranges of 2~10 s and 0.5~0.01 s. Sources of this noise are primarily attributed to mechanical disturbances originating from coastal and offshore maritime activities, alongside anthropogenic influences such as nearby human activities and road traffic. High-frequency seismic background noise greater than 1 Hz in the three environments is rich in information and balanced in signal, which meets the needs of shallow strata imaging in coastal zones. Employing F-K inversion techniques, a two-dimensional shear wave velocity profile was successfully generated for the study area, delineating the depth of the basement interface beneath the sedimentary layer. The viability of utilizing ambient noise for probing coastal zones was substantiated via comparison with results obtained from the horizontal-to-vertical spectral ratio (HVSR) method and corroborated by adjacent drilling data. This outcome underscores the potential of passive seismic methodologies for investigating complex coastal geophysical structures. To derive high-order mode dispersion curves characterized by energy concentration, comparative experiments were conducted with varying operational parameters for data acquisition. Subsequently, a joint inversion of both high-order and low-order modes was performed, yielding higher-resolution and more accurate velocity structure imaging results beneath the coastal zone. The successful acquisition of two-dimensional shallow shear wave velocity profiles in the western coastal zone of Dong'ao Island unequivocally validates the feasibility of employing passive source node seismograph exploration technology to eliminate blind spots in shallow strata exploration within coastal zones. This approach transcends the limitations imposed by traditional exploration techniques, achieving seamless integration of land-to-sea seismic exploration. It is anticipated that this research will furnish robust technical support for engineering projects and resource development initiatives in coastal regions. This research was granted by the National Natural Science Foundation of China (No. 42106078) and the Guangzhou Science and Technology Plan Project (No. 2023A04J0243).

Forecasting the dynamic variations of high-energy electrons in geostationary orbit remains a frontier challenge in space physics and is critical for the protection of geostationary satellites. In this study, we develop Transformer-based models to predict >= ${\ge} $2 MeV electron daily fluences and explore three distinct data-centric enhancement strategies to improve forecast accuracy: data augmentation (using Generative Adversarial Network and VAE networks), pre-training on higher-resolution data, and Few-shot Learning (FSL). Using training data from 2012 to 2015 and testing data from 2016 to 2018 from the GOES-15 satellite, the baseline Transformer model without enhancement achieved prediction efficiency (PE) values of 0.877 and 0.885 for the optimal two- and three-parameter inputs. Data augmentation provided only marginal gains, likely due to the inability of synthetic data to fully capture the complex physical variability of real observations. In contrast, pre-training yielded improved performance, with PE values of 0.912 and 0.906 for the best two- and three-parameter combinations. FSL proved to be the most effective enhancement strategy, achieving PE of 0.925 or 0.918 with (F, Vsw, Kp) or (F, AE) as inputs (F denotes the log10 of the >= ${\ge} $2 MeV electron daily fluence) and significantly boosting performance during relativistic electron enhancement events. It successfully predicts 96.9% of days with these events and issues first-day alerts for 41 out of 66 multi-day events. Comparative analyses with previous models confirm that the proposed enhancement strategies, particularly FSL and pre-training, offer substantial gains in forecasting accuracy for >= ${\ge} $2 MeV electron fluences, providing solutions for data-limited scenarios in space weather prediction.

In this study, we develop models to predict the log10 of ≥2 MeV electron fluxes with 5-minute resolution at the geostationary orbit using the Long Short-Term Memory (LSTM) and transformer neural networks for the next 1-hour, 3-hour, 6-hour, 12-hour, and 1-day predictions. The data of the GOES-10 satellite from 2002 to 2003 are the training set, the data in 2004 are the validation set, and the data in 2005 are the test set. For different prediction time scales, different input combinations with 4 days as best offset time are tested and it is found that the transformer models perform better than the LSTM models, especially for higher flux values. The best combinations for the transformer models for next 1-hour, 3-hour, 6-hour, 12-hour, 1-day predictions are (log10 Flux, MLT), (log10 Flux, Bt, AE, SYM-H), (log10 Flux, N), (log10 Flux, N, Dst, Lm), and (log10 Flux, Pd, AE) with PE values of 0.940, 0.886, 0.828, 0.747, and 0.660 in 2005, respectively. When the low flux outliers of the ≥2 MeV electron fluxes are excluded, the prediction efficiency (PE) values for the 1-hour and 3-hour predictions increase to 0.958 and 0.900. By evaluating the prediction of ≥2 MeV electron daily and hourly fluences, the PE values of our transformer models are 0.857 and 0.961, respectively, higher than those of previous models. In addition, our models can be used to fill the data gaps of ≥2 MeV electron fluxes.

We report an M9.3 flare and filaments activities from NOAA Active Region 11261 that are strongly modulated by the 3D magnetic skeleton. Magnetic field extrapolation from the vector magnetic field suggests complex magnetic connectivity and the existence of a high coronal null point southeast of the active region. A small filament over the inversed V-shaped polarity inversion line erupted and resulted in the M9.3 flare associated with a weak ejection in the EUV hot channel and the formation of a relatively large filament. Both the weak ejection and the eruption of the large filament were toward the southeast. Comparative analyses have disclosed the following new facts. First, the trajectory of looptop hard X-ray emission provides solid evidence that the magnetic reconnection site propagated up toward the coronal null point as the flare and filaments erupted. Second, the EVU observations show coronal mass ejection-like eruption features in the ejection region of the magnetic skeleton. Third, the closed fan confined the west end of the large filament and the corresponding flare ribbons. We demonstrate a spatiotemporal relationship between the magnetic skeleton and the flare filament activity. We conclude that the magnetic skeleton can modulate and determine almost all the characteristics of the studied activity in the corresponding scale.

Energetic particle precipitation is the major source of electron production that controls the ionospheric Pedersen and Hall conductances at high latitudes. Typically, the ionospheric conductances are estimated using either theoretical or empirical equations. The former method requires several ionospheric and thermospheric parameters as inputs. By contrast, empirical equations are simple, such as Robinson formulas and Galand formulas that have been widely used. In this study, we evaluate the empirical formulas of ionospheric conductances during four different types of auroral precipitation conditions based on 63 conjugate events observed by DMSP and EISCAT. The conductances calculated from the DMSP data with the empirical formulas are compared with those based on EISCAT measurements with the standard equations. The best correlation between these two is found when the empirical Robinson formulas are used in the presence of diffuse electron precipitation without ions. In the presence of ion precipitation, the correlation coefficients are smaller, but the correlation improves when the Galand formulas are used to estimate the contribution of ion precipitation to the conductances. For the condition of pure ion precipitation, the ionospheric conductances are increased up to 2-7 S for Pedersen and 2.5-10 S for Hall conductances. The increase is larger for a higher geomagnetic AE index. Overall, the empirical formulas applied to the DMSP particle spectra underestimate the ionospheric conductances.

We present an updated evaluation of SE Asian geodynamics that includes the interactions of the South China Sea (SCS) marginal basin with surrounding plates since the end of SCS spreading 20.5-18 Ma. Newly available Ar39/ Ar40 ages of SCS oceanic crust drilled at IODP U1431 near the SCS East basin extinct spreading center are older than 18 Ma. Conversely, the oldest ages of the Luzon arc and forearc at Taiwan's Lanyu island, Coastal range and Lichi melange are 17-18 Ma, suggesting that onset of the Manila subduction zone may have begun a few m.y. earlier. Before -20.5 Ma, the northern part of the Manila transcurrent fault (MTF), considered as the western boundary of the Ryukyu subduction zone, was a left-lateral lithospheric-scale shear zone. From -20.5-18 Ma to -7 Ma, this portion of MTF was connected to the Manila trench. Since -7 Ma, the MTF extended into the Taiwan Longitudinal valley and continued southwards to north Luzon island as near-vertical, left-lateral shear zone. Today, south of -24 degrees N, the MTF protrudes down to 30 km depths and terminates above the deeper Manila slab. Since -7 Ma, the whole MTF shifted 400 km westward with respect to Eurasia and rotated -23 degrees clockwise to become oriented -NS north of 16 degrees N latitude. We identify a tear fault in the Eurasian (EU) plate north of the Ryukyu trench that is located south of the Myako and Yonaguni islands. Since -10 Ma, the tear continuously progressed westward within EU crust, with the Philippine Sea plate progressively subducting northwestward between the two lips of the tear fault. A RFF (ridge-fault-fault) triple junction was active in the EU crust before 20.5 Ma, from 10 to 7 Ma, and since 2 Ma. This triple junction was always located on the MTF with one branch of the MTF on each side of the triple junction, and the third branch being the spreading center.

AbstractSatellite drag coefficients are crucial for determining the neutral mass densities that affect spacecraft operations in the thermosphere. Many studies typically utilize a constant drag coefficient of 2.2 to calculate the neutral density. However, due to the variability of space environment, uncertainties in the drag coefficient can lead to significant systematic discrepancies in neutral density measurements. Satellite drag coefficient may fluctuate in the thermosphere under various geomagnetic activities and altitudes. For the first time, we calculate the spherical satellite drag coefficient using data from the “Orbital Atmospheric Density Detection Experimental Satellite,” referred to as the QX satellite. Our findings reveal that the drag coefficient can be estimated by thermospheric temperature and density, which are dependent on geomagnetic activity and altitude. At an altitude of ∼510 km, drag coefficients are adjusted to around 2.425, instead of the constant value of 2.2. Furthermore, the drag coefficient may decrease due to the significant influence of increasing geomagnetic activity, such as geomagnetic storms, on thermospheric density and temperature. These estimates of the drag coefficient can also be used to reduce discrepancies when deducing the ballistic coefficient. Consequently, using the estimated drag coefficient can accurately determine the QX‐derived neutral density, which agrees well with the density from Swarm‐B satellite.

The current operational needs of space weather forecasting strongly require accurate predic-tions of the future 3-day Kp index.Such forecasts involve a multitude of predictors,including physical parameters observed at the Earth-Sun L1 point and historical characteristics of the Kp index.Therefore,previous research primarily relied on statistical or empirical methods for prediction.However,the com-plex coupling of multiple parameters during geomagnetic storm events has made it challenging to quanti-fy the contributions of various predictors to Kp index forecasting over a 3-day timescale,hindering fur-ther improvements in forecast accuracy.This study builds a 3-day Kp index forecasting model based on neural network modeling and utilizes explainable AI(Artificial Intelligence)algorithm,specifically the in-tegrated gradient algorithm,to quantify the contributions of individual predictor.The research results indicate that the southward interplanetary magnetic field contributes significantly to Kp index predic-tion,accounting for 37.15%of all factors,making it the primary contributor.Following this,solar wind speed contributes 15.73%,underscoring the model's ability to capture parameters aligned with physical characteristics as the primary predictive factors during training.The contribution of historical character-istics of Kp index(recurrence characteristics)gradually increases with the forecasting horizon and reach-es 68.06%at a lead time of 3-day.This substantiates the strong predictive capabilities of the AI model in forecasting geomagnetic storm events induced by high-speed solar wind streams originating from coronal holes.Furthermore,this study conducts contribution analysis on two significant geomagnetic storm events that occurred in 2015 and 2017.It reveals that the predominant predictors contributing to each event differ.This underscores the model's capability to accurately capture the complex coupling of multi-ple parameters in geomagnetic storm forecasting.In conclusion,this research demonstrates that employ-ing explainable AI algorithms can help quantify the contributions of various predictive factors to Kp in-dex forecasting to some extent.This has the potential to enhance further research and improvements in 3-day Kp index AI forecasting models.

Geostationary satellites are exposed to harsh space weather conditions, including ≥2 MeV electrons from the Earth’s radiation belts. To predict ≥2 MeV electron daily fluences at 75°W and 135°W at geostationary orbit for the following three days, long short-term memory (LSTM) network models have been developed using various parameter combinations. Based on the prediction efficiency (PE) values, the most suitable time step of inputs and best combinations of two or three input parameters of models for predictions are recommended. The highest PE values for the following three days with three input parameters were 0.801, 0.658 and 0.523 for 75°W from 1995 to August 2010, and 0.819, 0.643 and 0.508 for 135°W from 1999 to 2010. Based on yearly PE values, the performances of the above models show the solar cycle dependence. The yearly PE values are significantly inversely correlated with the sunspot number, and they vary from 0.606 to 0.859 in predicting the following day at 75°W from 1995 to 2010. We have proven that the poor yearly PE is related to relativistic electron enhancement events, and the first day of events is the most difficult to predict. Compared with previous models, our models are comparable to the top performances of previous models for the first day, and significantly improve the performance for second and third days.

The deflection of Solar Coronal Mass Ejections (CMEs) near the Sun may be the consequence of interaction between a CME and a coronal hole or the solar wind (Gopalswamy et al., 2009a, Gopalswamy et al., 2009b). In this study, 124 halo-CMEs that originate from 40 active regions are analyzed to deduce whether multiple CMEs from the same active region are deflected in the same direction, as well as to find the accuracy of predicting the deflection direction of a CME according to the ambient large-scale magnetic field configuration. It was found that at least 73% (29 groups) were significantly deflected. Also of the 16 groups with multiple significantly deflected CMEs from the same active regions, the deflection direction for the CMEs in 88% (14 groups) was consistent with each other. The magnetic field configuration was computed from synoptic maps of magnetic field from SOHO/MDI and SDO/HMI using a Potential Field Source Surface (PFSS) model. We have performed the error analysis for the calculation of the ambient magnetic field. After excluding the cases with high error, of the remaining 23 significantly deflected groups, the deflection of 91% (21 groups) was consistent with the ambient magnetic field configuration. Among them, 86% of the groups were deflected toward the Heliospheric Current Sheet (HCS), the boundary between the magnetic field polarities, and 14% toward Pseudo-Streamers (PS), the boundary between the same-polarity magnetic field regions. It was found to be in good agreement with previous studies.

Solar flares are a kind of violent solar eruptive activity phenomenon and an important warning device of space weather disturbance. In space weather forecasting, flare forecasting is an important forecast content. This paper proposes a flare prediction model based on long and short-term memory neural network, which uses the time sequence of magnetic field changes in the solar active region in the past 24 h to construct samples, and analyzes the time series evolution of magnetic field characteristics through the long and short-term memory neural network to predict whether ≥M-level flares will occur in the next 48 h. This paper uses a data set for all active region samples from May 2010 to May 2017, and selects 10 magnetic field characteristic parameters of SDO/HMI SHARP. In the modeling process, six feature parameters with high weight, gain rate, and coverage rate were selected as input parameters through XGBoost method. Through test comparison, the false report rate and accuracy rate of the model are similar to the traditional machine learning model, and the accuracy rate and critical success index are better than the traditional machine learning model, which are 0.7483 and 0.7402, respectively. The overall effect of the model is better than that of the traditional machine learning model.

It is well accepted that the physical properties obtained from the solar magnetic field observations of active regions (ARs) are related to solar eruptions. These properties consist of temporal features that might reflect the evolution process of ARs, and spatial features that might reflect the graphic properties of ARs. In this study, we generated video data sets with timescales of 1 day and image data sets of the SHARP radial magnetic field of the ARs from 2010 May to 2020 December. For the ARs that evolved from “quiet” to “active” and erupted the first strong flares in 4 days, we extract and investigate both the temporal and spatial features of ARs from videos, aiming to capture the evolution properties of their magnetic field structures during their transition process from “quiet” (non–strong flaring) to “active” (strong flaring). We then conduct a comparative analysis of the model performance by video input and single-image input, as well as of the effect of the model performance variation with the prediction window up to 3 days. We find that for those ARs that erupted the first strong flares in 4 days, the temporal features that reflect their evolution from “quiet” to “active” before the first strong flares can be recognized and extracted from the video data sets by our network. These features turn out to be important predictors that can effectively improve strong-flare prediction, especially by reducing the false alarms in a nearly 2 day prediction window.