Interpretable ML-Based Forecasting of CMEs Associated with Flares

Abstract

Coronal mass ejections (CMEs) that cause geomagnetic disturbances on the Earth can be found in conjunction with flares, filament eruptions, or independently. Though flares and CMEs are understood as triggered by the common physical process of magnetic reconnection, the degree of association is challenging to predict. From the vector magnetic field data captured by the Helioseismic and Magnetic Imager (HMI) onboard the Solar Dynamics Observatory (SDO), active regions are identified and tracked in what is known as Space Weather HMI Active Region Patches (SHARPs). Eighteen magnetic field features are derived from the SHARP data and fed as input for the machine-learning models to classify whether a flare will be accompanied by a CME (positive class) or not (negative class). Since the frequency of flare accompanied by CME occurrence is less than flare alone events, to address the class imbalance, we have explored the approaches such as undersampling the majority class, oversampling the minority class, and synthetic minority oversampling technique (SMOTE) on the training data. We compare the performance of eight machine-learning models, among which the Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA) model perform best with True Skill Score (TSS) around 0.78?±?0.09 and 0.8?±?0.05, respectively. To improve the predictions, we attempt to incorporate the temporal information as an additional input parameter, resulting in LDA achieving an improved TSS of 0.92?±?0.04. We utilize the wrapper technique and permutation-based model interpretation methods to study the significant SHARP parameters responsible for the predictions made by SVM and LDA models. This study will help develop a real-time prediction of CME events and better understand the underlying physical processes behind the occurrence.