Abstract Forecasting thermospheric density remains a critical challenge for satellite drag modeling and space situational awareness. Building on our previously introduced Reduced Order Probabilistic Emulator framework which compresses high‐dimensional Thermosphere‐Ionosphere‐ Electrodynamics General Circulation Model outputs into a low‐dimensional latent space using a convolutional orthogonal autoencoder—this work focuses on modeling the temporal dynamics of the latent thermospheric state. We employ both interpretable dynamical system identification techniques (Sparse Identification of Nonlinear Dynamics and Dynamic Mode Decomposition with control) and deep learning architectures (Long Short‐Term Memory, Gated Recurrent Units, and Transformer) to learn the evolution of these latent coefficients. To enhance model resilience during geomagnetic storms, we introduce a storm‐tuning strategy that fine‐tunes pre‐trained models on high‐ K p intervals, leading to improved accuracy during storm onset and recovery phases. We also implement a simple equal‐weighted ensemble of top‐performing models, which achieves Mean Absolute Percentage Error as low as 3.15%, significantly improving robustness across varying space weather conditions. Model performance is assessed using Superposed Epoch Analysis, revealing consistent error reduction across altitude and latitude during disturbed periods. These advances represent a significant step toward operational, data‐driven thermospheric forecasting under both nominal and storm‐time conditions.
Tapedia et al. (2026) studied this question.