The Prediction of Short‐Term Atmospheric Density in Lower Thermosphere

Abstract Accurate short‐term prediction of atmospheric density in the lower thermosphere (85–125 km) is essential for precise orbit determination and satellite drag mitigation, particularly during geomagnetic storms. Although physics‐based models such as NRLMSIS 2.1 provide empirical estimates, their effectiveness is limited under extreme space weather. This study introduces a deep learning framework based on a hybrid ConvLSTM‐Encoder‐Decoder architecture to forecast global atmospheric density with high spatiotemporal resolution. The model uses multi‐frame input sequences to capture spatial and temporal dynamics, leveraging logarithmically transformed density data from NRLMSIS 2.1. Two variants are developed: a “Normal model” trained on solar minimum (2019) and maximum (2024) conditions, and a “Storm model” trained exclusively on 102 geomagnetic storm events from 2020 to 2023. Evaluation on four independent 2025 storm events show the Storm model significantly outperforms the Normal model under disturbed conditions—especially at higher altitudes—achieving up to a 55.86% reduction in global average error at 125 km. It also exhibits enhanced stability in long‐horizon predictions, maintaining positive R 2 where the Normal model fails. Results indicate data set diversity, not just volume, is critical for robust generalization. Geomagnetic storm impacts show distinct altitude‐dependent patterns: antisymmetric polar enhancements at 85 km evolve into equatorially concentrated increases at 105 km, with perturbations generally below a few percent. These findings highlight the potential of specialized deep learning approaches to improve short‐term forecasting and support safer low‐Earth orbit operations.