Human‐Interpretable SEP Event Prediction Using White‐Box Explainable AI

Abstract Solar energetic particle (SEP) events pose radiation risks to astronauts and space systems, making their prediction essential. Although various prediction models have been developed, most machine learning (ML)‐based models lack interpretability; that is, it is difficult to explain the reasons behind individual predictions. In this study, we employ a novel explainable artificial intelligence (XAI), Wide Learning (WL), for SEP prediction. WL, developed by Fujitsu Research, is based on emerging pattern mining algorithms. A unique capability of WL is extracting effective combinations of feature conditions for classification. We classified solar flares (SFs) based on whether they led to SEP events with 10 MeV and 10 pfu observed by the Geostationary Operational Environmental Satellite (GOES). To better detect and characterize higher‐risk events (100 pfu), we also developed a three‐class classification model. We created 57 pre‐ and post‐flare features. On the class‐imbalanced data set, our three‐class model achieved a true skill statistic (TSS) of 0.54 0.12, while the binary model reached 0.79 0.11, comparable to state‐of‐the‐art performance. WL revealed that X‐ray and radio integrated intensities, and magnetic free energies, physics‐based eruptive SF predictors from three‐dimensional extrapolated magnetic fields, were key contributors to classification. The effectiveness of the extracted conditions was verified in a human‐interpretable manner; this is the essence of white‐box XAI, demonstrating their value for future operational use. Furthermore, analyses of false‐negative high‐risk SEP events suggested that identifying reasons for correct/incorrect predictions based on prediction rationales extracted by WL has the potential to further improve model predictive performance.