Physics-informed orbit determination for X-GEO space situational awareness

The increasing number of artificial space objects launched in the X-GEO region poses a new challenge for the space situational awareness. The need of having accurate observations and performing precise orbit determination of those objects is becoming critical to build and maintain a X-GEO space catalog. For this reason, this paper adapts the physics-informed orbit determination (PIOD) technique to X-GEO objects with real angle-only observations. The methodology relies on the powerful capabilities of physics-informed neural networks, a machine learning framework that combines the available observed data with the knowledge of the physics, to perform a physically-consistent data regression. The Cartesian state of the object is approximated through single layer feed-forward neural networks trained via Extreme Learning Machine. To incorporate the physics in the training loss, a high-fidelity orbital dynamics model, comprising non-spherical gravitational of the Earth and the third body perturbations, is exploited. The PIOD technique is applied to real observations of three objects in the X-GEO regions: 2020 SO, which is a Centaur upper stage; the rocket body 59228, and the Falcon 9 rocket body that carried the lunar lander NOVA-C. PIOD shows very good accuracy, with observation residuals in the order of arcseconds, and comparable or better results with respect to the batch least squares, with the advantage of not requiring any initial guess and a-priori information of the objects’ orbit.