Thermospheric Density Prediction Using a Deep Evidential Model-Based Framework

At low altitudes, atmospheric drag is the primary perturbation force for satellites. Among the factors influencing this drag, neutral mass density remains the most uncertain. Current leading density prediction models, such as NRLMSISE and JB, are empirical and derive from satellite, rocket, and radar observation data. Yet, constrained by their parametric formulas, these models can produce errors exceeding 100% during high solar and geomagnetic activities. While physics-based models like the Thermospheric General Circulation Models and Global Ionosphere Thermosphere Model offer insights into atmospheric physics, they are currently hampered by uncertain input and boundary conditions and haven't surpassed empirical models in performance.

Our recent work introduces a deep evidential model-based framework that synergizes empirical models, accelerometer-inferred density, and geomagnetic and solar indices. This framework, tested for on-orbit density prediction and global density construction, outperforms the GPs model from our prior research. It not only predicts thermospheric density with high accuracy during both quiet and storm periods but also offers superior uncertainty predictions. Notably, our model distinguishes between aleatoric and epistemic uncertainties, enhancing its reliability and confidence levels.