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

该项目推进了连接太阳物理学和计算机科学的跨学科研究。来自NASA和NOAA天文台的太阳观测将用于训练机器学习分类器，以创建新的太阳耀斑预测模型。太阳耀斑是太阳大气中强烈的局部电磁辐射爆发，太阳耀斑的预测至关重要，因为它们对当今科技驱动的社会具有潜在的危险影响。该项目支持一名职业生涯早期的教师，并将通过犹他州立大学(USU)的美国原住民暑期导师计划，鼓励未被充分代表的少数族裔学生探索数据科学和空间气象研究。两名研究生和一名本科生将得到支持，PI将在USU系统内为农村学生提供远程学习课程。本研究利用多变量时间序列(MVTS)表示的太阳活动区数据中丰富的图形数据连通性信息进行耀斑预测。这项研究将围绕两个科学问题展开。(1)如何利用磁场参数的时间序列相似性来预测耀斑？(2)最重要的磁场参数(及其时间序列相似性)是什么，最大限度地区分多种耀斑类别？在任务1中，团队将MVTS实例转换为参数图，其中节点表示磁场参数，边表示节点对的单变量时间序列相似性。将设计新的图神经网络(GNN)模型，在基准MVTS数据集上实现更好的耀斑预测性能。在任务2中，将从节点级、边级和子图级的参数图中提取特征，并根据图中最重要的特征进行排序。在任务3中，团队将MVTS数据集建模为单个图，其中节点将表示AR时间段，边将表示节点对的MVTS相似性。无监督聚类将与社区检测算法一起应用，以找到磁性均匀的活动区域时间段，基于聚类的特征将用于支持耀斑预测。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

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.