Classifying Different Types of Solar-Wind Plasma with Uncertainty Estimations Using Machine Learning

Abstract

Decades of in-situ solar-wind measurements have clearly established the variation of solar-wind physical parameters. These variable parameters have been used to classify the solar-wind magnetized plasma into different types, leading to several classification schemes being developed. These classification schemes, while useful for understanding the solar wind’s originating processes at the Sun and early detection of space weather events, have left open questions regarding which physical parameters are most useful for classification and how recent advances in our understanding of solar-wind transients impact classification. In this work, we use neural networks trained with different solar-wind magnetic and plasma characteristics to automatically classify the solar wind in coronal hole, streamer belt, sector reversal and solar transients such as coronal mass ejections comprised of both magnetic obstacles and sheaths. Furthermore, our work demonstrates how probabilistic neural networks can enhance the classification by including a measure of prediction uncertainty. Our work also provides a ranking of the parameters that lead to an improved classification scheme with \(\sim 96\%\) accuracy. Our new scheme paves the way for incorporating uncertainty estimates into space-weather forecasting with the potential to be implemented on real-time solar-wind data.