Analyzing Data Quality and Decay in Mega-Constellations: A Physics-Informed Machine Learning Approach

TL;DR

This work assesses the reliability of publicly available Starlink ephemerides for Space Situational Awareness by comparing them to high-fidelity Orekit propagations and by training physics-informed Neural Ordinary Differential Equations to recover latent non-conservative accelerations. It demonstrates that a non-trivial portion of trajectory discrepancies arises from data limitations and internal propagation choices, while showing that NODEs with physical priors can achieve accurate predictions and reveal regime-dependent residual forces, especially during deorbiting. The study provides a data-driven decay model and quantifies the magnitude and directionality of non-conservative accelerations, enabling better interpretation of publicly released orbital data. Overall, the framework advances SSA by combining rigorous physics with data-driven inference to address data quality and decay dynamics in mega-constellations.

Abstract

In the era of mega-constellations, the need for accurate and publicly available information has become fundamental for satellite operators to guarantee the safety of spacecrafts and the Low Earth Orbit (LEO) space environment. This study critically evaluates the accuracy and reliability of publicly available ephemeris data for a LEO mega-constellation - Starlink. The goal of this work is twofold: (i) compare and analyze the quality of the data against high-precision numerical propagation. (ii) Leverage Physics-Informed Machine Learning to extract relevant satellite quantities, such as non-conservative forces, during the decay process. By analyzing two months of real orbital data for approximately 1500 Starlink satellites, we identify discrepancies between high precision numerical algorithms and the published ephemerides, recognizing the use of simplified dynamics at fixed thresholds, planned maneuvers, and limitations in uncertainty propagations. Furthermore, we compare data obtained from multiple sources to track and analyze deorbiting satellites over the same period. Empirically, we extract the acceleration profile of satellites during deorbiting and provide insights relating to the effects of non-conservative forces during reentry. For non-deorbiting satellites, the position Root Mean Square Error (RMSE) was approximately 300 m, while for deorbiting satellites it increased to about 600 m. Through this in-depth analysis, we highlight potential limitations in publicly available data for accurate and robust Space Situational Awareness (SSA), and importantly, we propose a data-driven model of satellite decay in mega-constellations.

Figures (12)

• Figure 1: Graph showing the evolution of satellites' semi major axis throughout the data collection period. The semi major axis values were recorded for each orbit every $2$ days and represent the mean value during that time. The blue toned lines correspond to satellites that remained relatively stable, with minor variations due to periodic manoeuvres. The orange-toned lines represent satellites that began deorbiting and ultimately reentered the atmosphere after the collection period.
• Figure 2: Histogram of the RMSE values between the Orekit Propagation and the Starlink ephemerides of $300$ satellites. These values were separated into three equal sized groups based on their RMSE.
• Figure 3: Evolution of position RMSE for Satellite $48104$. The peak in the plot marks a $\sim$1.9 h manoeuvre, during which both RMSE and covariance determinants spike sharply. A second marked change occurs near the $48$ h mark, consistent with a switch in Starlink’s internal propagation model.
• Figure 4: Evolution of position covariance determinant for Satellite $48104$. The peak in the plot marks a $\sim$1.9 h manoeuvre, during which both RMSE and covariance determinants spike sharply. A second marked change occurs near the $48$ h mark, consistent with a switch in Starlink’s internal propagation model.
• Figure 5: Schematic of the Neural ODE architecture used for satellite trajectory prediction. The input variables consist of the satellite's initial state $\mathbf h$ and the prediction time, along with the initial angular momentum, space weather features and short history of past states. The Neural Network outputs the acceleration, $a_p$, which is then added to the two body acceleration with $J_2$ and $J_3$ correction ($a_{total}$). This output is integrated over time using an ODE solver to produce the predicted state at a future time step. The ODE solver queries the Neural Network at each integration step ($\lambda$) to obtain the time derivatives, which are then used to update the state vector. The process is repeated until the prediction time is reached. The output of the NODE is the predicted position and velocity of the satellite at the specified time step.