A Neural Network Model for Predicting the Solar Energetic Particle Events

太阳高能粒子（SEP）在日冕和太阳风中的加速和输运过程中种子粒子的来源和注入是太阳物理学中一个知之甚少的问题，因此准确预测SEP通量是一项艰巨的任务。 这个为期三年的项目旨在研究日冕中种子种群的分布及其对SEPs产生的影响。 该项目团队将开发一个用于预测SEP事件的机器学习（ML）模型，该模型将基于新罕布什尔州大学开发的高能粒子辐射环境模块（EPREM）代码。 该模型将利用GOES、STEREO和帕克太阳探测器卫星的数据，并采用开源ML库。 该项目预计将促进我们对太阳日冕中种子粒子的起源和分布的理解，这些粒子影响整个日光层中SEP的加速和运输。 该项目将为替代模型开发必要的技术和算法，这些替代模型将在空间气象预报方面有各种应用。 由一位职业生涯中期的女性PI领导的研究调查将涉及新罕布什尔州大学的研究生。 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。开发用于准确SEP预测的ML模型的挑战是缺乏足够的观察到的SEP事件数据库来训练和验证ML模型，这被称为“类不平衡”问题。 规避这一困难的一种方法是采用代理模型：即，在SEP事件的合成/模拟数据上训练ML模型，然后使用观察到的SEP数据优化和验证模型。 在这个为期3年的项目中，该团队将使用EPREM代码模拟主要的SEP事件，以探索种子种群的参数空间;他们将考虑为研究选择的事件的事件前，事件期间和事件后的光谱。 项目团队还将通过纳入这些种子群体参数来模拟SEP通量的时间序列，以便训练ML模型，然后使用可用的SEP事件数据库进行测试。 该项目将利用LSTM等深度学习技术以及PCA、tSNE和autoencoder等分类和降维技术。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

The origin and injection of seed particles in the acceleration and transport of solar energetic particles (SEPs) in the Sun’s corona and the solar wind is still a poorly understood problem in heliophysics, thus making it a formidable task to predict with accuracy the SEP flux for the sake of reliable space weather forecast. This 3-year project aims to investigate the distribution of seed population in the solar corona and its influence in the production of SEPs. The project team will develop a Machine Learning (ML) model for the prediction of SEP events, which will be based on the Energetic Particle Radiation Environment Module (EPREM) code developed at the University of New Hampshire. The model will utilize data from the GOES, STEREO, and the Parker Solar Probe satellites, and it will adopt open-source ML libraries. This project is expected to advance our understanding of the origin and distribution of seed particles in the Sun’s corona, which influence the acceleration and transport of SEPs throughout the heliosphere. The project will develop the necessary technique and an algorithm for surrogate models that will have diverse applications in space weather forecasting. The research investigations, led by a mid-career female PI, will involve graduate students at the University of New Hampshire. The research and EPO agenda of this project supports the Strategic Goals of the AGS Division in discovery, learning, diversity, and interdisciplinary research.The challenge in developing a ML model for an accurate SEP prediction is the lack of sufficient database of observed SEP events to train and validate the ML model, which is known as the “class imbalance” problem. One way to circumvent this difficulty is to employ surrogate models: that is, train the ML model on synthetic/simulated data of SEP events, and then optimize and validate the model using the observed SEP data. During this 3-year project, the team will simulate major SEP events using the EPREM code in order to explore the parameter space for the seed population; they will consider the pre-, during, and post-event spectra of the events selected for the study. The project team will also simulate time-series of SEP fluxes by incorporating these seed population parameters in order to train the ML model and then use the available SEP events database for testing. The project will make use of deep learning techniques such as LSTM as well as classification and dimensionality reduction techniques such as PCA, tSNE and autoencoder.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.