Predicting characteristics of bursty bulk flows in Earth's plasma sheet using machine learning techniques

Xuedong Feng^1*Jian Yang¹Jacob Bortnik²CHIH-PING WANG²Jiang Liu³

• ¹Department of Earth and Space Sciences, Southern University of Science and Technology, Shenzhen, China
• ²Department of Atmospheric and Oceanic Sciences, College of Physical Sciences, University of California, Los Angeles, Los Angeles, California, United States
• ³epartment of Earth, Planetary, and Space Sciences, University of California Los Angeles, Department of Atmospheric and Oceanic Sciences, College of Physical Sciences, University of California, Los Angeles, Los Angeles, California, United States

Bursty bulk flows (BBFs) play a crucial role in transporting energy, mass, and magnetic flux from the Earth's magnetotail to the near-Earth region. However, their impulsive nature and small spatial scale pose significant difficulties for in-situ observations, given that only a handful number of spacecraft operate within the vast expanse of the magnetotail. Consequently, accurately predicting their behavior remains a challenging goal. In this study, we employ the XGBoost machine learning algotithm to predict the variation range of several essential BBF properties, including duration, magnetic field, plasma moments, and specific entropy parameters. The observed characteristics of a BBF are shaped by its formation in the downstream tail and its journey until it reaches the spacecraft. Therefore, we use both the background properties of the plasma sheet prior to the arrival of the BBF and the attributes of indirectly related variables during the BBF interval as inputs. Trained on 17 years of THEMIS data, we explore different input configurations. One approach involves incorporating optimal parameter combinations, utilizing as many input parameters as possible to predict upper and lower bounds of a target variable. Within this framework, we further apply the leave-one-feature-out method to quantitatively assess the contribution of each input, identifying the most dominant factor influencing BBFs in a statistical sense. Another approach involves crossinstrument prediction, leveraging measurements from a different payload. Our findings reveal that including observed background values enhances prediction accuracy by 10-20 percentage points. This study offers data-driven insights to improve BBF predictability, providing valuable guidance for future space weather monitoring and theoretical research.