Modeling Asteroid Gravitational Potential with Physics-Informed Neural Networks

Gravity models are essential tools for representing the gravitational field exerted by celestial bodies: this thesis proposes a Physics-Informed Neural Network (PINN), trained on the asteroid 433 Eros, that significantly decreases the computational expense and memory footprint compared to traditional analytic methods. The model is built on the PINN-GM-III architecture, and modifies its structure and physics-informed loss function to improve performance for irregular celestial bodies. Each architectural change undergoes extensive testing, with the best configuration achieving an average error reduction of 76.68% across all metrics compared to the polyhedral analytic model.