Physics-Informed Neural Networks for Gravitational Potential Estimation

Machine learning is continuously evolving through new research and development of network architectures. One notable recent advancement is physics-informed neural networks (PINNs), which are capable of solving complex differential equations and integrating physical laws directly into the learning process. These networks have proven to be very successful for modeling various physics-related problems, such as fluid dynamics, thermodynamics, and structural mechanics. In this study, we will be using a PINN to address the gravity field modeling problem by learning gravitational potentials from position and acceleration data. Accurate gravity field modeling is essential when traveling in space. Mission planning, navigation and control, all rely on such data, especially when dealing with irregular density celestial bodies. PINNs may prove to be a good alternative to traditional methods, such as spherical harmonics or polyhedral models, as they struggle with scalability or generalization in varying orbital regimes. Specifically, we will be focusing on learning the gravity field generated from a heterogeneous-density asteroid, the 433-Eros, using the PINN-GM-III. This study also analyzes various modifications aimed at improving the original model. All trained networks are assessed using five distinct evaluation metrics to check network performance at different orbital regimes and altitudes.