Andrea D’Ambrosio, University of South Florida; Andrea Scorsoglio, University of Arizona; Roberto Furfaro, University of Arizona
https://doi.org/10.64861/CKYZ7793
Keywords: Orbit determination, physics-informed neural networks, uncertainty quantification, space domain awareness
Abstract:
Accurate orbit determination (OD) is a critical capability for Space Domain Awareness, particularly in regimes where classical OD techniques exhibit degraded performance due to poor observability, strong nonlinear dynamics, and limited tracking geometry. These limitations motivate the development of novel methodologies based on scientific machine learning, which combine the flexibility of data-driven models with the physical consistency of physics-informed neural networks.
This work presents a probabilistic Physics-Informed Orbit Determination framework (BPIOD) based on a single-layer physics-informed neural network that explicitly embeds the orbital motion equations into the learning process. The OD problem is formulated as a dynamical inference task in which the target state is estimated from angle-only observations provided by ground-based telescopes. The proposed framework does not require any initial state guess or prior knowledge of the trajectory.
Contrary to existing deterministic PIOD approaches, BPIOD incorporates uncertainty quantification by modeling the network output weights as random variables sampled from learned probability distributions. During inference, BPIOD produces multiple trajectory estimates, each compatible with the available measurements. Propagating the estimated states backward and forward in time provides a family of possible past and future ephemerides, offering a basis for quantifying uncertainty in the object’s trajectory.
The proposed framework is applied to simulated and real data to establish its accuracy and reliability. By facilitating confidence bounds on reconstructed trajectories, BPIOD supports robust decision-making in space surveillance and tracking applications.
Date of Conference: September 15-18, 2026
Track: Astrodynamics