Deep Learning Based Nonlinear Dimensionality Reduction for Emulators of Numerical Thermosphere Density Models

Modeling and forecasting of atmospheric drag for space objects in low Earth orbit (LEO) is a critical challenge for space situational awareness and environment safety and sustainability. The largest source of dynamics error or uncertainty affecting drag is the thermospheric density. Current operations use empirical models that lack fidelity and are deterministic resulting in unrealistic state and covariance
$v_{rel}$
predictions. For more than two decades, numerical physics-based models of thermospheric density have been touted as the next big thing for drag modeling. However, the computational cost combined with the lack of mature algorithms for data assimilation (or
$C_{D}$
model-data fusion) has not yet allowed them to make impact in operations. The research community has seen a recent trend towards the development of reduced order models or emulators to overcome these limitations for enabling operational deployment of numerical density models. Dimensionality reduction is an important first step in this process. We build upon previous work in the community to design a nonlinear dimensionality reduction approach based in a deep convolutional autoencoder (CAE). We develop a new architecture that employs an attention module, spatial loss scaling, and weighted sampling to optimize performance and overcome data imbalance. We also employ an orthogonality constraint for enabling robust data assimilation as the next step. Results show that we achieve performance similar in terms of reconstruction error (−2%) to that of the robust but linear principal component analysis (PCA) approach while significantly improving performance during nonlinear periods of geomagnetic storms.