Magnetopause location modeling using machine learning: inaccuracy due to solar wind parameter propagation

Abstract

An intrinsic limitation of empirical models of the magnetopause location is a predefined magnetopause shape and assumed functional dependences on relevant parameters. We overcome this limitation using a machine learning approach (artificial neural networks), allowing us to incorporate general, purely data-driven dependences. For the training and testing of the developed neural network model, a data set of about 15,000 magnetopause crossings identified in the THEMIS A-E, Magion 4, Geotail, and Interball-1 satellite data in the subsolar region is used. A cylindrical symmetry around the direction of the impinging solar wind is assumed, and solar wind dynamic pressure, interplanetary magnetic field magnitude, cone angle, clock angle, tilt angle, and corrected Dst index are considered as parameters. The effect of these parameters on the magnetopause location is revealed. The performance of the developed model is compared with other empirical magnetopause models. Finally, we demonstrate and discuss the inaccuracy of magnetopause models due to the inaccurate information about the impinging solar wind parameters based on measurements near the L1 point. This inaccuracy imposes a theoretical limit on the precision of magnetopause predictions, a limit that our model closely approaches.