A Neural Network Model of Three-dimensional Magnetospheric Chorus Waves

The evolution of chorus waves is important in the inner magnetosphere since it is closely related to the loss and acceleration of radiation belt electrons. In this study, we develop neural-network-based models for upper-band chorus (UBC; 0.5 f_ce < f <  f_ce ) waves and lower-band chorus (LBC; 0.05 f_ce < f < 0.5 f_ce) waves, where f_ce is the equatorial electron gyrofrequency. We establish a root-mean-square amplitude database for both UBC and LBC using Van Allen Probe levels 2 and 3 data products from the EMFISIS payload between October 1, 2012 and January 14, 2018. Based on the database, we construct an artificial neural network with corresponding L, magnetic local time, magnetic latitude, solar wind parameters and geomagnetic indices on different time windows as model inputs. Additionally, we adopt several different feature selection techniques to determine the most important features of magnetospheric chorus waves, reduce training or running time and improve the model accuracy. Our study suggests that the model results using the machine learning technique have the great potential to highly improve current understanding of the radiation belt dynamics.