Title: Towards Comprehensive Modeling of the Earth-Sun Magnetic Interaction using Machine Learning

Date: Thursday, May 4th

Time: 2:30 - 4:00 PM (EST)

Location: https://gatech.zoom.us/j/8105749379?pwd=YUliTXlFOEx6TlNWc2Rma3Jjc2x6UT09

Charles Topliff

Machine Learning PhD Student

Electrical & Computer Engineering
Georgia Institute of Technology

Committee

1 Dr. Morris Cohen (Advisor), Electrical and Computer Engineering, Georgia Tech

2 Dr. Mark Davenport (Co-advisor), Electrical and Computer Engineering, Georgia Tech

3 Dr. Matthieu Bloch, Electrical and Computer Engineering, Georgia Tech

4 Dr. David Anderson, Electrical and Computer Engineering, Georgia Tech

5 Dr. Jacob Bortnik, Department of Atmospheric and Oceanic Sciences, UCLA

Abstract

This thesis aims to comprehensively model the earth-sun magnetic interaction using machine learning techniques. The space science community has developed models of different regions of the space environment which are well understood as standalone components. In recent years, machine learning techniques have emerged as a way to address the shortcomings of these models by leveraging the wealth of data describing the space environment that has accumulated since the onset of the space race. Earth's magnetic field (i.e. the geomagnetic field) is heavily influenced by solar activity through a coupling mechanism known as the solar wind. The solar wind is a stream of charged particles that carries the interplanetary magnetic field through interplanetary space and ultimately influences geomagnetic activity, which can result in disruptions to societal infrastructure such as the power grid and satellite communications. If we can forecast these disruptions, then this damage can be mitigated through precautionary measures. Our research seeks to increase the lead time of geomagnetic activity forecasts to mitigate damage to these systems. In Aim 1, we improve direct geomagnetic index prediction with an LSTM trained to predict geomagnetic indices using measurements of the solar wind near Earth. In Aim 2, we look further back in the chain of events to predict the solar wind directly using solar image data and convolution autoencoders. In particular, we focus on solar wind streams coming from coronal holes, which are responsible for a large portion of the variance of the solar wind. In Aim 3, we aim to improve prediction of solar wind during times when the solar wind is under the influence of anomalies such as interplanetary coronal mass ejections by augmenting our coronal hole model with a model taking sequences of images as an input.