5:15 PM - 6:45 PM
[PEM11-P11] Predicting Solar Energetic Particle Event Occurrences Using Explainable AI
Keywords:Solar Energetic Particles, Solar Flares, Explainable AI
Fujitsu Ltd. and Nagoya University are jointly researching space weather to ensure the safety of human activities which are expanding to the Moon and Mars. Solar Energetic Particle (SEP) Events, which occur in conjunction with solar flares (SFs) and coronal mass ejections (CMEs), have effects on both humans and space systems.
We are conducting a classification task using Wide Learning, an explainable Artificial Intelligence (AI) developed by Fujitsu Research, to explore the conditions under which SFs are accompanied by SEP Events. Wide Learning, originally developed in the field of Discovery Science and based on emerging pattern mining algorithms, enables us to conduct classification tasks and to search for exhaustive conditions.
We created 57 features from Soft X-ray measurements observed by GOES, remote sensing vector magnetic fields observed by SDO, and the physics-based flare predictive scheme based on the three-dimensional extrapolated magnetic fields of solar active regions developed by Kusano et al. (2020). We classified SFs that meet the condition of SEPs > 10 MeV, > 10 pfu (cm2 s sr) in the NOAA SWPC database during Solar Cycle 24 as positive samples, and all other SFs as negative samples.
Due to the class imbalance of positive/negative samples, we fixed the positive examples and undersampled the negative samples to achieve a 1:3 ratio of positive to negative samples, conducting 10 trials with replacements.
Additionally, due to the flare class imbalance of positive/negative samples, we conducted learning and prediction for cases (a) where negative samples were randomly sampled, and (b) where a constant 1:3 ratio of negative to positive samples was maintained for each C, M, and X flare class.
Our model demonstrates a True Skill Statistic (TSS) for the discrimination capability of SEP-positive SFs of approximately 0.7 for case (a), and about 0.4 for case (b). We also identified multiple useful conditions for predicting positive/negative samples, based on numerical ranges and combinations of each feature. In case (a), the X-ray peak intensity emerged as a significant weighted hypothesis for positive examples, while in case (b), the flare duration and flare history emerged as significant weighted hypotheses. For negative examples, combinations of magnetic field features emerged as significant weighted hypotheses.
These results suggest the potential for predicting SEP Event occurrences with step-by-step alerts, and the ability to reference past cases that align with identified conditions. We will discuss the potential for new space weather forecasts using these numerical ranges and combination conditions, as well as the applicability of past similar cases to these hypotheses, and future prospects.
We are conducting a classification task using Wide Learning, an explainable Artificial Intelligence (AI) developed by Fujitsu Research, to explore the conditions under which SFs are accompanied by SEP Events. Wide Learning, originally developed in the field of Discovery Science and based on emerging pattern mining algorithms, enables us to conduct classification tasks and to search for exhaustive conditions.
We created 57 features from Soft X-ray measurements observed by GOES, remote sensing vector magnetic fields observed by SDO, and the physics-based flare predictive scheme based on the three-dimensional extrapolated magnetic fields of solar active regions developed by Kusano et al. (2020). We classified SFs that meet the condition of SEPs > 10 MeV, > 10 pfu (cm2 s sr) in the NOAA SWPC database during Solar Cycle 24 as positive samples, and all other SFs as negative samples.
Due to the class imbalance of positive/negative samples, we fixed the positive examples and undersampled the negative samples to achieve a 1:3 ratio of positive to negative samples, conducting 10 trials with replacements.
Additionally, due to the flare class imbalance of positive/negative samples, we conducted learning and prediction for cases (a) where negative samples were randomly sampled, and (b) where a constant 1:3 ratio of negative to positive samples was maintained for each C, M, and X flare class.
Our model demonstrates a True Skill Statistic (TSS) for the discrimination capability of SEP-positive SFs of approximately 0.7 for case (a), and about 0.4 for case (b). We also identified multiple useful conditions for predicting positive/negative samples, based on numerical ranges and combinations of each feature. In case (a), the X-ray peak intensity emerged as a significant weighted hypothesis for positive examples, while in case (b), the flare duration and flare history emerged as significant weighted hypotheses. For negative examples, combinations of magnetic field features emerged as significant weighted hypotheses.
These results suggest the potential for predicting SEP Event occurrences with step-by-step alerts, and the ability to reference past cases that align with identified conditions. We will discuss the potential for new space weather forecasts using these numerical ranges and combination conditions, as well as the applicability of past similar cases to these hypotheses, and future prospects.