Empirical Modeling of the Plasmasphere Dynamics Using Neural Networks

Abstract

We present the PINE (Plasma density in the Inner magnetosphere Neural network‐based Empirical) model ‐ a new empirical model for reconstructing the global dynamics of the cold plasma density distribution based only on solar wind data and geomagnetic indices. Utilizing the density database obtained using the NURD (Neural‐network‐based Upper hybrid Resonance Determination) algorithm for the period of 1 October 2012 to 1 July 2016, in conjunction with solar wind data and geomagnetic indices, we develop a neural network model that is capable of globally reconstructing the dynamics of the cold plasma density distribution for 2≤L≤6 and all local times. We validate and test the model by measuring its performance on independent data sets withheld from the training set and by comparing the model‐predicted global evolution with global images of He+ distribution in the Earth's plasmasphere from the IMAGE Extreme UltraViolet (EUV) instrument. We identify the parameters that best quantify the plasmasphere dynamics by training and comparing multiple neural networks with different combinations of input parameters (geomagnetic indices, solar wind data, and different durations of their time history). The optimal model is based on the 96 h time history of Kp, AE, SYM‐H, and F10.7 indices. The model successfully reproduces erosion of the plasmasphere on the nightside and plume formation and evolution. We demonstrate results of both local and global plasma density reconstruction. This study illustrates how global dynamics can be reconstructed from local in situ observations by using machine learning techniques.

The Challenge of Machine Learning in Space Weather: Nowcasting and Forecasting

“…There are several other areas where machine learning has been applied in a less systematic way but that are nonetheless promising for a data‐driven approach. Plasmaspheric electron density estimation has been proposed in Zhelavskaya et al, (, ). Concerning the ionosphere‐thermosphere region, ionospheric scintillation has been modeled in Jiao et al (), Lima et al (), Linty et al (), McGranaghan et al (), and Rezende et al ().…”

Systematic Analysis of Machine Learning and Feature Selection Techniques for Prediction of the Kp Index

“…Training is an optimization procedure, in which the weights (the internal parameters of the network) are tuned using the training set of data so that the difference between the network output and the actual target variable is minimal. A description of FNNs applied to space physics problems can be found in Chu, Bortnik, Li, Ma, Angelopoulos, et al (), Chu, Bortnik, Li, Ma, Denton, et al (), Zhelavskaya et al (), and Zhelavskaya et al (). In this work, we use the MATLAB Deep Learning Toolbox to train neural networks (https://mathworks.com/products/deep-learning.html).…”

“…The Kp index is one of the most widely used global measures of geomagnetic activity. It is used as an input to many scientific applications, including the parameterization of ionospheric ion outflow (Yau et al, ) and aurora particle precipitation (Emery et al, ) in the ionosphere, thermosphere (Bruinsma et al, ), hot plasma particle density (Denton et al, ; Korth et al, ), cold plasma density in the plasmasphere (Goldstein et al, ; Maynard & Chen, ; Pierrard et al, ; Zhelavskaya et al, ), plasmapause location (Carpenter & Anderson, ), and radiation belt models and wave parameterizations (Agapitov et al, ; Brautigam & Albert, ; Orlova et al, ; Ozeke et al, ; Shprits et al, ) in magnetospheric physics, among others. It is therefore important to predict the Kp index accurately in order to produce most reliable forecasts in the aforementioned areas.…”

Quantitative Assessment of Radiation Belt Modeling

“…Research efforts also focused on developing better specifications of the environment that are required for common modeling approaches. These included machine learning techniques to construct the global 10.1029/2018JA026414 Journal of Geophysical Research: Space Physics distributions of plasma density (Chu et al, 2017;Zhelavskaya et al, 2017), plasmapause location on a global scale using particle tracing simulations (Goldstein et al, 2014), seed electron populations from the IMPTAM model (Ganushkina et al, 2015), and an empirical model of particle fluxes at geosynchronous orbit based on LANL/GEO (Los Alamos National Laboratory geosynchronous) data (Denton et al, 2016). The specification of both the configuration of the magnetic field and its outer boundary also saw a newfound level of attention within the RB community.…”