Giacomo Acciarini, University of Surrey; Edward Brown, University of Cambridge; Christopher Bridges, University of Surrey; At?l?m Güne? Baydin, University of Oxford; Thomas Berger, University of Colorado / Space Weather Technology, Research, and Education Center (SWx TREC); Madhulika Guhathakurta, NASA Headquarters
Keywords: Karman, Applied Machine Learning, Space Weather, Thermospheric Density Estimation, Benchmark
Abstract:
Recently SpaceX lost 40 satellites due to a geomagnetic storm that increased atmospheric drag, preventing satellites from leaving safe mode and forcing them to deorbit [1]. More accurate estimation of thermospheric density and precise modeling of the Sun’s influence on the Earth’s atmosphere and in turn, drag, is essential in planning and running missions and maintaining a safe space environment [2]. Solar and geomagnetic activity dominates the behavior of the density of the thermosphere and significantly influences the estimation and prediction of spacecraft motion in low-Earth orbit (LEO). Currently, we are heading into the maximum of Solar Cycle 25 as the amount of operational satellites in LEO is growing rapidly, with planned mega-constellations making for an increasingly complex operating environment. The defense sector, space agencies, and commercials operators heavily rely on thermospheric neutral density models to estimate satellite trajectories for a variety of safety-critical tasks in LEO, including satellite collision avoidance, space traffic management, and re-entry; however, these models are currently suboptimal.
In this work, we develop an improved machine learning (ML) thermospheric neutral density model, and a software package, called Karman, which can be used both for hosting state-of-the-art empirical and ML thermospheric density models, as well as for benchmarking their performances at different conditions (e.g. in terms of geomagnetic storm strength, altitude, solar irradiance, etc.). This provides the community with a software framework that can be used as a proving ground by the research and operational communities for any thermospheric density model.
In particular, our contribution focuses on four different aspects:
A unified analysis-ready dataset that gathers solar irradiance, geomagnetic, and thermospheric density inputs, which span across the last 20 years. This supports the main solar irradiance (both raw measurements and proxies) and geomagnetic indices used as inputs to model the thermospheric density, as well as NRLMSISE-00 and JB-08 empirical model predictions, and precise orbit determination (POD)-derived thermospheric density.
Feed-forward neural network models that process those inputs and provide an improved thermospheric density estimation.
A benchmarking framework where users can test their own density models, as well as our provided ML-based ones, against POD-derived and empirical models’ thermospheric density. This provides a unified framework to investigate the models’ performance at different altitude ranges, geomagnetic and solar irradiance conditions and using different metrics (such as root mean squared error and mean absolute percentage error).
Experiments to show how this framework can be used in an effective way to both benchmark models, and also study the main features that drive thermospheric density changes. We compare the performance of NRLMSISE-00 and JB-08 against machine learning models trained on exactly the same inputs, showing that our ML models can consistently outperform the empirical models’ accuracy, by 20-40% in the mean absolute percentage error. State-of-the-art solutions for estimating thermospheric density involve the use of solar proxies (such as F10.7cm radio emission) for estimating the Sun’s influence on the thermosphere. However, studies suggest that direct EUV irradiance data would be essential to improve thermospheric density estimations [3,4]. We show how our model can be used to easily set up experiments that support extra inputs, therefore allowing users to validate such hypotheses.
[1] Dang, T., Li, X., Luo, B., Li, R., Zhang, B., Pham, K., Ren, D., Chen, X., Lei, J. and Wang, Y., 2022. Unveiling the space weather during the Starlink satellites destruction event on 4 February 2022. Space weather, 20(8), p.e2022SW003152.
[2] Berger, T. E., Holzinger, M. J., Sutton, E. K., and Thayer, J. P. Flying through uncertainty.Space weather, 18(1), \https://doi.org/10.1029/2019SW002373.
[3] Chamberlin, P.C., Woods, T.N. and Eparvier, F.G., 2008. Flare irradiance spectral model (FISM): Flare component algorithms and results. Space Weather, 6(5). doi: https://doi.org/ 10.1029/2020SW002588
[4] Vourlidas, A. and Bruinsma, S., 2018. EUV irradiance inputs to thermospheric density models: Open issues and path forward. Space Weather, 16(1), pp.5-15.
Date of Conference: September 19-22, 2023
Track: Atmospherics/Space Weather