基于深度学习的内磁层等离子体密度动态演化模型  

Dynamic evolution of inner magnetospheric plasma density based on deep learning methods

作　　者：郭英杰 倪彬彬[1,2] 付松[1] 胡泽骏[3] 郭建广[4] 冯明航 周若贤 郭德宇 闫玲 马新 顾旭东 GUO YingJie;NI BinBin;FU Song;HU ZeJun;GUO JianGuang;FENG MingHang;ZHOU RuoXian;GUO DeYu;YAN Ling;MA Xin;GU XuDong(Department of Space Physics,School of Electronic Information,Wuhan University,Wuhan 430072,China;Center for Excellence in Comparative Planetology,Chinese Academy of Sciences,Hefei 230026,China;SOA Key Laboratory for Polar Science,Polar Research Institute of China,Shanghai 200136,China;National Center for Space Weather,Key Laboratory of Space Weather,China Meteorological Administration,Beijing 100081,China)

机构地区：[1]武汉大学电子信息学院空间物理系,武汉430072 [2]中国科学院比较行星学卓越创新中心,合肥230026 [3]中国极地研究中心国家海洋局极地科学重点实验室,上海200136 [4]中国气象局国家空间天气监测预警中心,北京100081

出　　处：《地球物理学报》2022年第6期1931-1939,共9页Chinese Journal of Geophysics

基　　金：国家自然科学基金(42025404,42188101,41974186,42174188,41674163);科技部重点研发项目(2018YFC1407303);中国科学院先导B计划(XDB41000000);民用航天预研项目(D020303,D020308,D020104)资助。

摘　　要：等离子体密度作为空间环境的重要参量,它的全球实时分布信息不仅对理解内磁层带电粒子时空演化过程具有重要意义,对于预报和防范灾害性空间天气过程也有着潜在应用价值.利用范阿伦双星的高质量等离子体密度观测数据,本文基于机器学习算法训练得到一个稳定的深度神经网络模型:包含五个隐藏层;激活函数包括Sigmoid和ReLU函数;以太阳风参数、地磁指数以及卫星对应的位置信息作为输入.在测试集上,该模型输出值和观测值之间的线性相关系数约为0.93,均方根误差(RMSE)约为0.3,表明该模型性能良好.通过使用该模型对2012年4月24日磁暴事件中等离子体密度的全球动态变化进行模拟,我们成功重构了磁暴期间内磁层等离子体密度的全球变化过程,包括等离子体层的侵蚀和恢复,以及羽流的形成和消失.该内磁层等离子体密度的深度神经网络模型将有助于推动内磁层波粒相互作用的深入研究.Plasma density is a fundamental parameter of space environment,the global and real-time information of which is of importance to both understandings of inner magnetospheric particle dynamics and forecast of hazardous space weather processes.Using high-quality plasma density measurements from Van Allen Probes and adopting the technique of machine learning,this study develops a stable deep neural network(DNN)model,which is featured by five hidden layers,the activation function that includes the Sigmoid and ReLU functions,and the inputs of solar wind parameters,geomagnetic indices and spacecraft location.For the test set,the output of the model and the observations generate a linear correlation coefficient of 0.93 and the root-mean-square error(RMSE)of 0.3,indicating the good performance of this DNN model.We then implement this model to simulate the global dynamic variation of plasma density in the inner magnetosphere during the course of the storm on 24 April 2012,which is found to reasonably reconstruct the spatio-temporal evolution of inner magnetospheric plasma density including the erosion and recovery of the plasmasphere and the formation and disappearance of the plasmaspheric plume.Therefore,such a DNN model can provide reasonable information of the global and real-time profile of magnetospheric plasma density and then be valuable to help deepen current understanding of the causes and consequences of wave-particle interactions in the inner magnetosphere.

关 键 词：范阿伦卫星 地球磁层 神经网络模型 等离子体密度 

分 类 号：P353[天文地球—空间物理学]