Orbit Determination Using Physics-Informed Neural Networks

作者：Gus E. Loshelder, Rohan Sood, Weihua Su · 年份：2025 · DOI：10.2514/6.2025-1931 · 被引用次数：5 · 研究领域：Model Reduction and Neural Networks、Fault Detection and Control Systems、Control Systems and Identification

This paper explores the use of a physics-informed neural network (PINN) to determine the orbits of negligible-mass space objects, such as satellites, vehicles, and asteroids. PINNs are a powerful class of machine learning algorithms that train a neural network on domain-specific governing equations, typically differential equations. These equations are incorporated as constraints during training, serving as a substitute or supplement to training data. In recent years, PINNs have seen increasingly widespread use in science and engineering since their inception. They represent a versatile merger of computational numerical modeling and data-driven machine learning. In this study, the PINN framework is adapted to characterize the motion of a body in orbit around the Earth by using the orbital differential equation of motion as a training constraint, allowing the body's motion to be modeled given its initial conditions. Moreover, this work aims to establish the feasibility of applying this technique to two-body and n-body systems involving objects of negligible mass. The results demonstrate a strong potential for PINNs in orbital mechanics, further supporting the notion that PINNs are a robust and widely applicable computational tool.