Data-Driven Modeling of Polar IonosphericElectrodynamics Using Convolutional Neural Networks

Polar ionospheric electrodynamics are an integral part of understanding the Sun to Earth connection. Fusing data from multiple measurement sources and creating more accurate, consistent, and complete descriptions of electrodynamics for research in this area is incredibly important, yet a classically difficult task. One of the main issues is the constantly changing state of the ionospheric electrodynamics. The recently developed AMGeO (Assimilative Mapping of Geospace Observations) open-source software utilizes a Bayesian inferential framework to solve the spatial prediction problem posed by the need to fill the gaps left in sparsely and irregularly sampled data. AMGeO offers a robust approximation to this inferential problem and produces snapshot maps of polar ionospheric electrodynamics states.

This paper aims to expand the scope of the Bayesian inferential problem solved in AMGeO to a spatiotemporal prediction problem. Solving the spatiotemporal problem will require a state transition model that takes a current state and predicts a state at a future time step. The state transition model is being learned from large volumes of AMGeO maps using Convolutional Neural Networks. The improvements gained by this expansion will be evaluated on various performance metrics.