Probabilistic Solar Proxy Forecasting with Neural Network Ensembles

Accurate prediction of satellite orbital state is essential for space traffic management and space domain awareness. In this work, we investigate the impact of atmospheric drag uncertainty on predicted orbital state, focusing on two sources of space weather uncertainty: thermosphere driver uncertainty and thermosphere model uncertainty. We begin by presenting a new probabilistic model for the most used thermosphere solar driver, F10.7, and evaluate its performance against operational forecast models. We then evaluate methods for propagating driver uncertainty through the stochastic HASDM-ML model for thermosphere mass density. Finally, we perform orbital case studies to demonstrate the impact of driver and model uncertainty on predicted orbital state covariance. The importance of quantifying and propagating atmospheric drag uncertainty for accurate satellite orbit prediction is highlighted. By explicitly modelling and incorporating uncertainties in the both the thermosphere drivers and model, we aim to provide more reliable estimates of thermosphere characteristics and provide predicted orbital state and associated uncertainty.