SHINE: Understanding the Relationships of Photospheric Vector Magnetic Field Parameters in Solar Flare Occurrences using Graph-based Machine Learning Models

点击翻译按钮获取中文摘要

This project advances interdisciplinary research connecting heliophysics and computer science. Solar observations from NASA and NOAA observatories will be used to train machine learning classifiers for creating a new solar flare prediction model. Solar flares are intense localized eruptions of electromagnetic radiation in the Sun’s atmosphere, and the prediction of solar flares is essential because of their potential hazardous impacts on today’s technology-driven society. The project supports an early-career faculty member and will encourage underrepresented minority students to explore data science and space weather research through the Native American Summer Mentorship Program at Utah State University (USU). Two graduate students and one undergraduate student will be supported and the PI will offer a distance learning course for rural students within the USU system.This research leverages the rich connectivity information of graph data for flare prediction from multivariate time series (MVTS)-represented solar active region data. The research will be centered around two science questions. (1) How to leverage the time series similarities of the magnetic field parameters for the prediction of flares? (2) What are the most important magnetic field parameters (and their time series similarities) that maximally distinguish multiple flare classes? In Task 1, the team will transform MVTS instances to parameter graphs, where the nodes denote the magnetic field parameters, and edges denote the univariate time series similarities of the node pairs. Novel graph neural network (GNN) models will be designed that can achieve better flare prediction performance on a benchmark MVTS dataset. In Task 2, features will be extracted from the parameter graphs at node-level, edge-level, and subgraph-level, and ranked by the most important graph features. In Task 3, the team will model the MVTS dataset as a single graph, where nodes will denote AR temporal segments and edges will denote MVTS similarities of the node pairs. Unsupervised clustering will be applied with community detection algorithms to find the magnetically homogeneous active region temporal segments, and clustering-based features will be used to support flare prediction.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.