Characterizing In-Situ Solar Wind Observations Using Clustering Methods

The Sun’s atmosphere is a hot, high pressure, partially ionized gas, that overcomes the Sun’s gravity to create a flow of plasma that expands out into the solar system: a phenomena that is called the solar wind. The solar wind comprises the continuous and dynamic streams of plasma released from the Sun. Some of the physical properties of these plasma streams (composition, charge state, proton entropy) can be determined in the solar corona and remain non-evolving as the solar wind parcels expand into the heliosphere. These properties can be used to connect in-situ measurements to their coronal origins. Determining how solar wind parcels differ based their coronal origins will reveal the nature of physical processes (such as heating and acceleration) involved in their formation. Studies up to this point have largely relied upon statistical methods, characterizing the wind into groups such as fast and slow wind. Other methods have been used to detect signatures that represent transient events, such as interplanetary coronal mass ejections (ICMEs). The boundaries representing physical distinctions between the groups usually have included an aspect of them that was subjective. In the past few years, there has been a growing push to use machine learning in the field of heliophysics, in the form of (including but not limited to) predicting conditions in the solar wind/time series regression, identifying coronal features/events in solar images, and to serve as a tool to reduce subjective bias when determining solar wind groups. Here, we examine how machine learning, applied to several case studies of the solar wind, has the potential to link the physical properties measured in-situ to the origin of the wind at the Sun. We evaluate the robustness of the results of two different methods, comparing their performance. We discuss caveats that may arise when applying such techniques. Finally, we identify their strengths and define what aspects would be beneficial in continuing to develop machine learning approaches to this field.