[Paper Review] Machine Learning in Orbit Estimation: a Survey

Francisco José de Cáldas, Cláudia Soares|arXiv (Cornell University)|Jul 19, 2022

Space Satellite Systems and Control4 citations

TL;DR

This survey reviews the application of machine learning (ML) techniques—particularly deep learning and probabilistic models—to improve orbit determination, orbit prediction, and atmospheric density modeling in low Earth orbit. It highlights how ML enhances accuracy by modeling non-conservative forces, reducing uncertainty, and enabling data-driven alternatives to traditional physics-based models.

ABSTRACT

Since the late 1950s, when the first artificial satellite was launched, the number of Resident Space Objects has steadily increased. It is estimated that around one million objects larger than one cm are currently orbiting the Earth, with only thirty thousand larger than ten cm being tracked. To avert a chain reaction of collisions, known as Kessler Syndrome, it is essential to accurately track and predict debris and satellites' orbits. Current approximate physics-based methods have errors in the order of kilometers for seven-day predictions, which is insufficient when considering space debris, typically with less than one meter. This failure is usually due to uncertainty around the state of the space object at the beginning of the trajectory, forecasting errors in environmental conditions such as atmospheric drag, and unknown characteristics such as the mass or geometry of the space object. Operators can enhance Orbit Prediction accuracy by deriving unmeasured objects' characteristics and improving non-conservative forces' effects by leveraging data-driven techniques, such as Machine Learning. In this survey, we provide an overview of the work in applying Machine Learning for Orbit Determination, Orbit Prediction, and atmospheric density modeling.

Motivation & Objective

• Address the limitations of traditional physics-based orbit estimation methods, which suffer from kilometer-level errors in seven-day predictions due to uncertainties in initial states and environmental forces.
• Investigate how data-driven ML techniques can improve accuracy in orbit determination and prediction by learning from observational data and reducing reliance on simplified dynamical models.
• Examine the role of machine learning in modeling thermospheric mass density, a key source of uncertainty in low Earth orbit, with a focus on replacing empirical models with data-driven alternatives.
• Evaluate the integration of uncertainty quantification techniques—such as Bayesian neural networks and Monte Carlo dropout—into ML-based orbit estimation to support collision avoidance and mission planning.
• Identify challenges in adopting ML for space operations, including interpretability, data scarcity, and the need for collaboration between aerospace and machine learning experts.

Proposed method

• Systematically survey ML applications in three core tasks: orbit determination (OD), orbit prediction (OP), and atmospheric density modeling, using a taxonomy of techniques and models.
• Analyze the use of artificial neural networks (ANNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs) for modeling orbital dynamics and estimating RSO characteristics from observational data.
• Evaluate hybrid models such as EKF-ANN and Gaussian Process (GP)-based uncertainty estimation to combine physical models with data-driven learning for improved state estimation.
• Review the application of Bayesian neural networks and deep ensembles to estimate predictive uncertainty in orbit prediction, enabling probabilistic collision assessment.
• Assess grey-box models that embed physical constraints (e.g., conservation laws) into neural networks to improve interpretability and generalization in space environment modeling.
• Examine the use of time-series and image-based models (e.g., Siamese CNNs, 3D generators) to infer RSO shape, attitude, and pose from light curves and optical data.

Experimental results

Research questions

• RQ1 How can machine learning improve the accuracy of orbit determination and prediction beyond traditional physics-based models, especially in the presence of uncertain initial states and environmental forces?
• RQ2 To what extent can data-driven models replace or enhance empirical atmospheric density models like JB2008 and DTM, and what are the performance gains in terms of accuracy and uncertainty quantification?
• RQ3 What role do uncertainty-aware ML models—such as Bayesian neural networks and deep ensembles—play in enabling reliable probability of collision (PoC) assessments for space debris mitigation?
• RQ4 How can ML techniques be used to infer unmeasured physical properties of resident space objects (e.g., mass, area-to-mass ratio, attitude) from observational data?
• RQ5 What are the key challenges in adopting ML for operational space situational awareness, and how can grey-box modeling and interdisciplinary collaboration bridge the gap between aerospace and machine learning communities?

Key findings

• ML-based orbit prediction models, particularly those using deep learning, can outperform traditional SGP4 propagators by learning complex, non-linear perturbations from data, reducing long-term prediction errors.
• Bayesian neural networks and Monte Carlo dropout have been successfully applied to estimate uncertainty in atmospheric density models, improving reliability for collision avoidance systems.
• Hybrid models such as EKF-ANN and GP-enhanced filters demonstrate improved state estimation accuracy by combining physical constraints with data-driven learning, especially under noisy or sparse observations.
• Deep learning models trained on historical HASDM data can replicate and even improve upon the 10–15% 1-sigma accuracy of legacy empirical models like JB2008 and DTM.
• Techniques such as Siamese CNNs and 3D image generators enable the inference of RSO shape and pose from light curves and optical data, reducing reliance on prior knowledge of object characteristics.
• Grey-box models that embed physical laws into neural networks show promise in balancing interpretability and performance, offering a path toward trustworthy, explainable ML in space operations.