Assessing the Predictive Capability of Solar Surface Magnetic Fields for Flare Associated Coronal Mass Ejections based on Machine Learning Approaches

Patel, Nabdeep (2026) Assessing the Predictive Capability of Solar Surface Magnetic Fields for Flare Associated Coronal Mass Ejections based on Machine Learning Approaches. Masters thesis, Indian Institute of Science Education and Research Kolkata.

Text (MS Dissertation of Nabdeep Patel (24MR001)) Nabdeep.pdf - Submitted Version Restricted to Repository staff only Download (8MB)

Abstract

Solar flares and coronal mass ejections (CMEs) are among the most energetic phenomena occurring on the Sun and play a vital role in influencing space weather. Understanding the physical conditions that distinguish eruptive flares (flares associated with CMEs) and confined flares (only flares) remains a major challenge in solar physics. In this thesis, we investigate the predictive capability of photospheric magnetic field properties, characterized by SHARP (Space-weather HMI Active Region Patch) parameters derived from SDO/HMI vector magnetograms, using machine learning methods. We began by revisiting previous studies on flare-CME classification using SHARP parameters that reported high predictive performance. We identified a methodological issue: the use of class-wise normalization during preprocessing, which introduces information of class labels into the training process and leads to data leakage. On reproducing these methods under the same preprocessing setup, we obtain a high True Skill Statistic (TSS) of ∼ 0.73 ± 0.25, comparable to the reported TSS of ∼ 0.8 ± 0.2 in Bobra & Ilonidis (2016). This slight variation in TSS is due to differences in random seeding, which wasn’t applied while training their dataset. However, when the preprocessing is corrected, applying normalization without class separation, within each cross-validation fold, the predictive performance decreases significantly to ∼ 0.19 ± 0.18, indicating limited predictive capability. We further investigated the dataset by plotting scatter plots and probability distribution functions, which reveal a strong overlap of the datapoints for the eruptive and the confined flares for all SHARP parameters. Also, we showed that increasing model complexity by including a three-class classification or incorporating temporal features into the dataset, or even increasing the dataset, does not lead to substantial improvement in predictive performance. Our results demonstrate that photospheric magnetic fields alone are insufficient to reliably distinguish between eruptive and confined flares. This suggests that the key physical processes governing CME initiation are not fully captured by photospheric observations, and incorporating coronal magnetic field properties is essential for improving the predictive capability. v

Actions (login required)

View Item