Surrogate Modeling of Thermospheric Mass Density Using Dimensionality Reduction and Machine Learning Regression

We present a unified surrogate modeling framework for reconstructing and forecasting thermospheric mass density using dimensionality reduction and data-driven regressors. A full-resolution daily 3D mass density dataset from the Whole Atmosphere Model with Ionosphere Plasmasphere Electrodynamics (WAM-IPE) model. Input drivers include real sequences of geomagnetic index Kp and solar flux F10.7. Dimensionality reduction was applied using Incremental PCA, Autoencoders (AEs), Sparse AEs, β-VAEs, and Variational Autoencoders, reducing the latent dimensionality by over four orders of magnitude. We compared two typical surrogate models—Polynomial Chaos Expansion (PCE) and Long Short-Term Memory networks (LSTM)—on the latent space using both reconstruction RMSE and time-dependent forecasting performance. Our pipeline enables full decoding back to physical 3D space and robust forecasting given unseen driver sequences. This work offers a scalable solution for surrogate modeling of thermospheric state under operational and data assimilation constraints.