Speaker
Description
As human space activities, such as the Artemis program, become increasingly ambitious, ensuring safety in the space environment has become more critical than ever. In response, Fujitsu Ltd. and Nagoya University are conducting joint research on space weather for future lunar and deep-space exploration. A key focus of this research is solar energetic particle (SEP) events, which are mostly triggered by solar flares (SFs) and coronal mass ejections (CMEs), and can pose significant risks to human health and space systems.
In our previous work, we built a binary classification model using Fujitsu’s explainable AI, Wide Learning (WL), to predict whether SFs are associated with SEP events (ESWW2024, CD5.2 Kato et al.). In this study, we extended our model to a three-class classification model with WL, which classifies SFs as follows: Class 0, not associated with SEP events (flux < 10 pfu); Class 1, associated with SEP events below the S2 threshold (10 pfu ≤ flux < 100 pfu); and Class 2, associated with SEP events at or above the S2 threshold (flux ≥ 100 pfu), based on the NOAA S-scale (≥ 10 MeV particles). The purpose of this extension is twofold: to enhance the detection of higher-risk events and to identify the characteristics of these events.
We created a data catalogue from X-, M-, and C-class SFs observed during Solar Cycle 24 and 25 (up to the end of June 2024). The catalogue consists of 57 features derived from GOES/XRS X-ray data, SDO/HMI magnetic field data, the κ-scheme, a physics-based flare prediction scheme developed by Kusano et al. (2020) that is based on three-dimensional extrapolated magnetic fields of solar active regions, and other data. Due to the class imbalance between positive (Class 1 and 2) and negative (Class 0) samples, we created two types of datasets from the catalogue: (a) an imbalanced dataset (positive:negative = 1:30), and (b) a balanced dataset (positive:negative = 1:2).
Our three-class model yielded true skill statistic (TSS) of (a) 0.54 ± 0.12 and (b) 0.57 ± 0.11, while the binary model achieved (a) 0.79 ± 0.11 and (b) 0.76 ± 0.09, which are comparable to those reported for a similar binary model. We also performed a leave-one-out analysis for Class 2 SEP events and identified possible reasons for the correct/incorrect predictions. In addition, analysis of the decision rationales extracted from WL revealed that the X-ray integrated flux was a key feature in the classification, particularly for Class 2. The free energy within high free-energy regions (HiFER) was also identified as an important contributor to the classification. This suggests that incorporating physics-based features, such as the free energy in HiFER, could significantly improve the predictability, especially for pre-flare SEP prediction, compared to utilizing solely the directly observed data. We will discuss the potential for developing a step-by-step SEP forecasting system from pre-flare to SEP onset, as well as the prospects for creating new physics-based predictors based on insights from our WL-based model.
| Do you plan to attend in-person or online? | In-person |
|---|
