Statistical Analysis of Plasmaspheric Erosion and Refilling - Machine Learning Approach
Abstract
The density and composition of Earth's Plasmasphere strongly influences wave growth and propagation, as well as energetic particle scattering. This affects Earth-to-space communication and killer electron behavior. Previous empirical plasma density models have been based on statistical averages and are limited in their capability to make accurate predictions of the dynamic state of Earth's plasmasphere. Recent advances in machine learning techniques allow us to more accurately quantify complex global processes and nonlinear responses to driving conditions during geomagnetic storms. We present a three-dimensional dynamic electron density model using a feedforward neural network trained on electron density data from multiple satellites. The model takes spacecraft location as well as time series of solar wind and geomagnetic indices obtained from NASA's OMNI database as inputs. We conducted a comprehensive analysis to find the best model inputs and parameters. The results from the analysis indicate that the optimal input parameters are AE, Solar Lyman Alpha, KP, and DST indices with varying time lengths. The best positional parameters are L shell, sin(MLT), cos(MLT), and cos^3(MLAT). We optimized the neural network architecture and hyperparameters to avoid overfitting and further improve density prediction performance. The model can predict out-of-sample data with a correlation coefficient of 0.96. We carried out event and statistical analysis of plasmaspheric erosion and refilling on the equatorial plane and along the same flux tube during different geomagnetic storms. We will show interesting results about plasmaspheric dynamics that cannot be obtained using spacecraft observations, e.g. time dependent and two-stage refilling. The results of the ML-based plasmaspheric density model are generally consistent with previous studies, but show new intricacies when compared to statistically averaged models. This model demonstrates the potential for machine learning techniques to be utilized in understanding physics and insight discovery, as well as advancing state-of-the-art space weather prediction.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMNG52A0166L