Prediction of Solar Eruptions with Machine-Learning Algorithms Combining Physical Models and Observations

空间天气预测已经存在了几十年，但由于更多数据的出现，复杂的数据处理技术和计算能力正在达到新的高度。 用于研究空间天气的现代卫星和地面望远镜获取的数据比以往任何时候都多。 例如，太阳动力学天文台（SDO）每天获取约1.5 TB的数据。 因此，现在是时候对这个问题给出肯定的答案了：我们能否利用数据革命来有效地预测一次重大太阳耀斑的爆发？ 这个为期三年的项目的主要目的是通过汇集计算机科学家和太阳物理学家的跨学科团队来回答这个问题，以分析SDO，全球振荡网络组（GONG）和地球静止运行环境卫星（GOES）观测站所获取的数据，以及来自数值模型的数据产品。使用机器学习算法来描述和理解哪些特征表明太阳耀斑即将爆发。 以前的研究试图使用这些成分的一个子集来预测太阳耀斑，但不是全部。 尽管有大量的数据可用，但很少有空间天气研究利用数据革命来使用机器学习算法预测空间天气。这个为期三年的项目解决了太阳耀斑物理学中的未决问题。 首先，该项目旨在确定哪些特征有助于本地和全球的耀斑前签名。 虽然活动区域之间的相互作用可以触发耀斑行为，但超局部现象也可以触发耀斑行为。 第二，项目小组将探索任何特定特征的变化率如何影响太阳上的喷发活动。 最后，该项目旨在确定每个特征最能预测未来燃烧活动的时间尺度。 为了实现这些目标，项目团队将使用这三项任务中确定的相关功能，沿着可解释的机器学习算法，如状态空间模型，来预测太阳活动。空间气象预测是一项国家优先事项，被列为最近十年期调查的一个关键目标，也是《国家空间气象行动计划》的一个主要重点。 这个为期三年的项目旨在构建一个开源的、文档齐全的、统一的、可复制的和可操作的特征数据集和机器学习模型。 该项目团队将使用该数据集和机器学习模型，通过暑期实习和他们的书《统计学、数据挖掘和太阳物理学中的机器学习》来教授学生。 该项目的研究和EPO议程支持AGS部门在发现、学习、多样性和跨学科研究方面的战略目标。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Space-weather prediction has been around for many decades but is reaching new heights due to the advent of more data, sophisticated data processing techniques, and computing power. Modern satellites and ground-based telescopes designed to study space weather take more data than ever before. The Solar Dynamics Observatory (SDO), for example, acquires ~1.5 terabytes of data a day. As such, the time may be right to give an affirmative answer to the question: Can we harness the data revolution to effectively predict the onset of a major solar flare? The main purpose of this three-year project is to answer this question by bringing together an interdisciplinary team of computer scientists and solar physicists to analyze data taken by the SDO, Global Oscillation Network Group (GONG), and Geostationary Operational Environmental Satellite (GOES) observatories, as well as data products derived from numerical models, using machine-learning algorithms to characterize and understand which signatures indicate the imminent eruption of a solar flare. Previous research studies have attempted to predict solar flares using a subset of these components, but not all of them. Few space-weather studies have harnessed the data revolution to predict space-weather using machine learning algorithms despite the vast amount of data available. This three-year project addresses open questions in solar flare physics. First, the project aims to determine which features contribute to local and global pre-flare signatures. While interaction between active regions can trigger flaring behavior, hyper-local phenomena can also trigger flaring behavior. Second, the project team will explore how the rate of change of any given feature influences eruptive activity on the Sun. Finally, the project aims to identify the timescales on which each individual feature best predicts future flaring activity. To accomplish these goals, the project team will use the relevant features identified in these three tasks, along with interpretable machine learning algorithms, such as state-space models, to predict solar activity. Space weather prediction is a national priority, listed as a key goal of the most recent decadal survey and a prime focus of the National Space-Weather Action Plan. This three-year project aims to build an open-source, well-documented, unified, reproducible, and operational feature dataset and machine-learning model. The project team will use this dataset and machine-learning model to teach students both through summer internships and via their book entitled Statistics, Data Mining, and Machine Learning in Heliophysics. The research and EPO agenda of this project supports the Strategic Goals of the AGS Division in discovery, learning, diversity, and interdisciplinary research.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.