Improving Thermospheric Density Predictions in Low-Earth Orbit With Machine Learning

Thermospheric density is one of the main sources of uncertainty in the estimation of satellites' position and velocity in low-Earth orbit. This has negative consequences in several space domains, including space traffic management, collision avoidance, re-entry predictions, orbital lifetime analysis, and space object cataloging. In this paper, we investigate the prediction accuracy of empirical density models (e.g., NRLMSISE-00 and JB-08) against black-box machine learning (ML) models trained on precise orbit determination-derived thermospheric density data (from CHAMP, GOCE, GRACE, SWARM-A/B satellites). We show that by using the same inputs, the ML models we designed are capable of consistently improving the predictions with respect to state-of-the-art empirical models by reducing the mean absolute percentage error (MAPE) in the thermospheric density estimation from the range of 40%-60% to approximately 20%. As a result of this work, we introduce Karman: an open-source Python software package developed during this study. Karman provides functionalities to ingest and preprocess thermospheric density, solar irradiance, and geomagnetic input data for ML readiness. Additionally, it facilitates developing and training ML models on the aforementioned data and benchmarking their performance at different altitudes, geographic locations, times, and solar activity conditions. Through this contribution, we offer the scientific community a comprehensive tool for comparing and enhancing thermospheric density models using ML techniques. Accurately modeling the density of the thermosphere is pivotal for spacecraft operations such as collision avoidance, re-entry prediction, and orbital lifetime analysis. In this study, our aim is twofold. First, we want to study and compare the performance of data-driven machine learning (ML) models in predicting thermospheric density data against standard empirical models used in the field, which are used as baseline. By training ML models using precise orbit determination-derived satellite data, we show that they can achieve significant performance improvement compared to empirical models, with a reduction of 61% in the mean absolute percentage error. Second, we also provide the community with a shared software framework that supports the ingestion of solar irradiance, geomagnetic, and thermospheric density data, as well as a training and benchmarking framework to develop ML models. This framework allows researchers and operators to both train their ML models and to compare them at different periods of the solar cycle, geomagnetic storm conditions, geographical locations, and times. Machine learning (ML) models can significantly outperform existing physics-based thermospheric neutral density models on exactly the same inputs ML models can improve over NRLMSISE-00 and JB-08 empirical density models by 61% and 39% respectively in the mean absolute percentage error The software allows the creation of ML-ready data for training and benchmarking new models, supporting solar irradiance and geomagnetic data