Calibration of atmospheric density model based on Gaussian Processes

Abstract

Neutral mass density is presently the predominant uncertain term among all the factors affecting the atmospheric drag, which is the dominant perturbation force for space objects at low altitude. The current best density estimation performance is often achieved by empirical models that can be limited by their assumed, parametric formulations of regression. This paper presents a density estimation framework that integrates information from empirical models, environment conditions, and satellite measurement data. Different from existing frameworks, the new integration mechanism is based on Gaussian Processes (GPs) which are nonlinear, non-parametric regression methods. The method can estimate both current and future densities. Furthermore, it will provide uncertainty quantifications in its estimates through GPs' underlying Bayesian inference. Simulations are designed to test the hypothesis that the new framework is valuable to improve the nowcast performance of the empirical models, and to predict future density. Empirical models including NRLMSISE-00 and JB2008 and accelerometer-inferred densities from satellite CHAMP are used for the study. The new method is shown to achieve better Pearson correlation coefficient (R), root mean squared error (RMSE), and mean ratio from the empirical models when the density estimation is tested on both the missing data and future densities, which are, respectively, within and following the GP's training period. Together with providing quality uncertainty estimations, the proposed framework has the great potential to reduce the estimation errors from the empirical model and provide an effective means to estimate density for a satellite.