Machine learning emulator for physics-based prediction of ionospheric response to solar wind variations

Physics-based simulations are important for elucidating the fundamental mechanisms behind the time-varying complex ionospheric conditions, such as field-aligned currents (FACs) and plasma convection patterns, against unprecedented solar wind variations incidents in the Earth’s magnetosphere. However, to perform a huge parameter survey for understanding the nonlinear solar wind density dependence of the FAC and convection patterns, for example, a large-scale cluster computer is not fast enough to run state-of-the-art global magnetohydrodynamic (MHD) simulations. Here we report the impressive performance of a machine-learning based surrogate model for the ionospheric outputs of a global MHD simulation, using the reservoir computing technique called echo state network (ESN). The trained ESN-based emulator is exceptionally fast to perform the parameter survey, suggesting a missing solar wind density dependence of the ionospheric polar cap potential. We discuss future directions including the promising application for the space weather forecast.