High-fidelity Gravitational Field Reconstruction of Asteroids via Multisensor Fusion and Physics-informed Neural Networks

Accurate gravitational field reconstruction of small bodies is essential for ensuring the safety and reliability of spacecraft during close-proximity operations. However, the inherent challenges posed by asteroids, such as their irregular topography and heterogeneous internal density distributions, introduce a certain degree of challenge to the application of conventional modeling methods. In this work, we propose a high-fidelity framework for gravitational field reconstruction that integrates multisensor data fusion, physics-informed neural network constraints, and a dynamic altitude-adaptive sampling strategy. Our method leverages 3D asteroid shape information derived from optical imagery to distinguish grid points located inside or outside the asteroid via ray-intersection detection, thereby improving computational efficiency and the accuracy of internal density estimation. Additionally, by embedding gravitational field equations into the neural network training process, our model enforces physical consistency and improves interpretability. Furthermore, a dynamic sampling mechanism adaptively allocates computational resources based on altitude-specific reconstruction errors, focusing the learning process on regions requiring higher accuracy. Experimental results on both synthetic and real-world data sets, demonstrate the proposed method’s superior reconstruction accuracy across various altitudes, its robustness to noise, and its capacity for effective transfer learning via pretraining on low-fidelity models of asteroid surface shells. These advancements provide a promising avenue for precise gravitational modeling in future deep-space exploration missions.