Bhargav Joshi, Digantara; Saikat Majumder, Digantara; Thamim Ansari, Digantara; Tanveer Ahmed, Digantara
Keywords: SSA, Space Weather, Generative AI, Diffusion Model
Abstract:
Thermospheric mass density variations driven by space weather significantly impact satellite drag in Low-Earth Orbit (LEO). This was starkly demonstrated in February 2022 when enhanced atmospheric drag during minor geomagnetic storms led to the loss of 38 Starlink satellites. With the increasing number of LEO satellites and the approaching peak of Solar Cycle 25, satellite operations face heightened risks from extreme space weather events, such as the G5-level May 2024 Gannon storm. Accurate atmospheric density modeling under such conditions is thus crucial for precise orbit prediction and the safety of spaceflight.
Over the past few decades, various empirical and physics-based thermospheric density models have been developed and evaluated for drag calculations. However, due to the dynamic nature of space weather and atmospheric variability, no single model has emerged as the definitive operational standard. Empirical models typically exhibit uncertainties of approximately 10–15% under nominal conditions, while physics-based models struggle with inherent assumptions, resulting in uncertainty quantification, a key challenge. Advances in AI offer a promising solution: diffusion models, a class of Generative AI, iteratively refine noisy data to generate realistic representations. Their probabilistic nature allows for inherent uncertainty quantification through multiple samplings of the stochastic denoising process.
This paper presents a diffusion-based model for accurate thermospheric density prediction, designed to enhance satellite operators’ decision-making, particularly during geomagnetic storms. The model incorporates space weather data from NASA’s OMNI 2 dataset, including solar wind, magnetic field and plasma measurements, as well as X-ray flux data from the GOES satellite network. Training data consists of thermospheric densities derived from ESA’s open-source datasets, which include satellite-based measurements from CHAMP, GRACE-1, GRACE-2, GRACE-FO, SWARM-A, SWARM-B, and SWARM-C.
The model leverages space weather data as covariates for conditional denoising. Attention-based Transformer networks, such as PatchTST, have outperformed traditional methods like Dynamic Mode Decomposition in capturing nonlinear temporal relationships. Therefore, a multi-head Transformer architecture is employed to learn complex historical space weather patterns and guide the diffusion process. The model is trained using a Continuous Ranked Probability Score (CRPS)-based quantile loss, enabling probabilistic forecasting. Instead of producing single-point predictions, it generates a probability distribution, improving robustness and reliability in operational settings.
The model performance is evaluated using metrics such as Mean Squared Error (MSE) and Mean Absolute Error (MAE) for the predicted values and CRPS for the predicted uncertainties. The evaluations are made across various classes of space weather events, covering quiet-time periods to all severity levels of NOAA G-Scale geomagnetic storms.
The paper also describes the model’s performance against the current operational empirical and physics-based density models, including U.S. Naval Research Laboratory’s NRLMSISE-00 and NRLMSIS2.0, SET’s JB2008, and NOAA SWPC’s WAM-IPE, focusing on density estimation and trajectory prediction.
The model presented in this paper seeks to provide better space weather prediction capabilities in the field of Space Situational Awareness (SSA) by leveraging the cutting-edge developments in the field of Generative AI. As the space-based assets continue to grow, there is a growing need for more advanced, accurate and robust space weather models which can aid in mission planning including launches, on-orbit operations and re-entries.
Date of Conference: September 16-19, 2025
Track: Atmospherics/Space Weather