Identification of waves in plasma using convolutional neural networks
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
The solar wind is a supersonic plasma composed primarily of protons and electrons that propagates outward from the Sun and whose extent defines the limits of the heliosphere. Because the solar wind is weakly collisional, long-range electromagnetic interactions play vital roles in energy transfer that vary depending on the properties of a given region of the solar wind. Fluctuations in the electromagnetic field can lead to energy transfer between the field and the plasma particles, with a substantial net heating of the solar wind observed in spacecraft data. While it is generally agreed that the solar wind is turbulent and that dissipation of turbulent fluctuations leads to heating of particle populations, the particular dissipation mechanisms that lead to the observed heating are not well understood. Circularly-polarized electromagnetic waves near ion kinetic scales have been observed to heat the solar wind through wave-particle interactions, which makes ion-scale waves a promising candidate for dissipation and heating. However, obtaining a large dataset of waves is difficult due to the time-expensive nature of existing wave identification methods. This work presents a novel approach to wave identification in magnetic field data using machine learning, specifically 1D convolutional neural networks, that is far less computationally expensive than existing methods. A training set is created using a combination of Wind spacecraft data labeled manually through existing wave-identification methods and synthetic data generated using a superposition of randomly-phased turbulent fluctuations and circularly-polarized waves. The convolutional neural network is trained to identify waves in six-second intervals of high-pass filtered Wind magnetic field and is tested rigorously using multiple distinct manually-labeled datasets. A high-accuracy iteration of the network is created that can readily identify circularly-polarized waves in Wind spacecraft data. The network is used to classify eighteen years of Wind spacecraft data (2005-2022) based on its circularly-polarized wave content, yielding the longest-duration dataset of ion-scale waves in the solar wind thus far. Average solar wind parameters associated with wave-containing intervals are examined, and it is shown that waves come about most often in the fast solar wind and at relatively high plasma beta compared to the rest of the solar wind. Estimates for wave parameters, including the wave amplitude, the wavevector direction, and the magnetic compression associated with waves are given. Based on the angle between the average magnetic field and the estimated wavevector direction in each interval, two distinct wave populations are identified corresponding to quasi-parallel and quasi-perpendicular propagation. Waves are relatively common when the solar wind is unstable to temperature anisotropy instabilities, such as the firehose and mirror instabilities; however, the solar wind is not often unstable to these instabilities, indicating that waves cannot only be generated locally by temperature anisotropy instabilities. The number of waves identified per day has notable periodicities that match harmonics of the solar rotation frequency; the most common sources of waves are therefore likely to be dependent on solar rotation. The wave dataset is separated based on the solar cycle phase, and the rising and declining phases are found to contain 50\% more waves proportionally compared to solar maximum and solar minimum. The large difference between solar cycle phases persists across similar categories of solar wind across phases, indicating that the primary wave generation mechanisms change throughout the solar cycle and are most prevalent during the rising and declining phases.
Description
Keywords
Supersonic plasma, Ion-scale waves, Wind spacecraft data