Assessing the Geoeffectiveness of Stream Interaction Regions through Physically Interpretable Machine Learning

Stream interaction regions (SIRs) are common solar wind structures that mainly drive moderate geomagnetic storms, especially during the declining and minimum phases of the solar cycle. Despite substantial progress in identifying their solar wind drivers, predicting SIR geoeffectiveness remains challenging due to the nonlinear and multivariate nature of solar wind–magnetosphere coupling. In this study, we apply an interpretable machine learning framework based on a Support Vector Machine (SVM) classifier to evaluate the geoeffectiveness of 879 SIR events from 1995 to 2019. A comprehensive feature set is constructed using OMNI solar wind data, incorporating interplanetary magnetic field (IMF) components, plasma parameters, and event-level statistical descriptors, including pre- and post-event conditions. A two-step feature reduction strategy, combining Fisher score ranking with grid search optimization, is used to identify an optimal subset of input variables. The SVM model achieves a True Skill Statistic of 0.80 and an F1-score of 0.81 on the geoeffective class (Dst < –50 nT), demonstrating strong performance in identifying moderate and stronger geomagnetic storms. SHapley Additive exPlanations values and Uniform Manifold Approximation and Projection visualization are employed to interpret and validate the model’s behavior. The analysis reveals that features related to the duration of southward IMF, the solar wind electric field, and IMF Bz components play dominant roles in the model’s decision-making, consistent with known reconnection-driven energy transfer processes. These results highlight the capability of classical machine learning models, coupled with physically meaningful feature engineering and interpretability methods, to support space weather research and improve understanding of SIR-driven geomagnetic activity.