A combined neural network- and physics-based approach for modeling the plasmasphere dynamics

Plasmasphere is a torus of cold plasma surrounding the Earth and is a very dynamic region. Its dynamics is driven by space weather. Having an accurate model of the plasmasphere is very important for wave-particle interactions and radiation belt modeling. In recent years, feedforward neural networks (NNs) have been successfully applied to reconstruct the global plasmasphere dynamics in the equatorial plane [Bortnik et al., 2016, Zhelavskaya et al., 2017, Chu et al., 2017]. These neural network-based models have been able to capture the large-scale dynamics of the plasmasphere, such as plume formation and the erosion of the plasmasphere on the night side. However, NNs have one limitation. When data is abundant, NNs perform really well. In contrast, when the coverage is limited or non-existent, as during geomagnetic storms, NNs do not perform well. The reason is that since these data are underrepresented in the training set, NNs cannot learn from the limited number of examples. This limitation can be overcome by employing physics-based modeling during such intervals. Physics-based models perform stably during high geomagnetic activity time periods if initialized and configured correctly. In this work, we show the combined approach to model the global plasmasphere dynamics that utilizes advantages of both neural network- and physics-based modeling and produces accurate global plasma density reconstruction during extreme events. We present examples of the global plasma density reconstruction for a number of extreme geomagnetic storms that occured in the past including the Halloween storm in 2003. We validate the global density reconstructions by comparing them to the IMAGE EUV images of the He+ particles distribution in the Earth’s plasmasphere for the same time periods.