Machine Learning Based Modeling of Thermospheric Mass Density

Abstract In this study, we propose a machine learning based approach to construct an empirical model of thermospheric mass densities, based on the MultiLayer Perceptron and bi‐directional Long Short‐Term Memory for ensemble learning model (MBiLE). The MBiLE model was trained by using only the thermospheric mass density from Swarm C satellite at ∼450 km altitude. To assess the performance of the MBiLE model, the model predictions were compared with observations from several satellites, namely, the Swarm C, the Challenging Minisatellite Payload (CHAMP) and the Gravity Field and Steady‐State Ocean Circulation Explorer (GOCE) satellites. The determination coefficients (R 2 ) for the three satellites are 0.98, 0.99, and 0.98, respectively. The MBiLE model predicts the thermospheric mass density well not only at Swarm C altitude but also at lower altitudes. Earlier empirical models based on multivariate least‐square‐fitting approach failed to achieve this good altitude generalization (e.g., Liu et al., 2013, https://doi.org/10.1002/jgra.50144 ; Xiong et al., 2018a, https://doi.org/10.5194/angeo‐2018‐25 ). Further tests have been made by checking the MBiLE model prediction deviations in relation to magnetic local time, day of year, solar flux level, and magnetic activities. No obvious dependences are found for these parameters. Comparing with the NRLMSIS‐2.0 model, the MBiLE model improves prediction accuracy by 91%, 66%, and 56% at the three satellites altitudes. The results indicate that the MBiLE model has the ability to predict well the thermospheric mass density over a wide altitude range, for example, from 224 to 528 km, offering potential for atmospheric research applications.