REINFORCEMENT LEARNING AND ORBIT-DISCOVERY ENHANCED BY SMALL-BODY PHYSICS-INFORMED NEURAL NETWORK GRAVITY MODELS

Abstract

The novel Physics-Informed Neural Network (PINN) gravity model enables accurate and computationally efficient representations of complex gravity fields. Prior work has studied the use of PINNs for gravity field modeling of large celestial bodies and asteroids. This paper explores how PINN gravity model’s accuracy and speed can be leveraged to address two problems of interest pertaining to small-body exploration: 1) augmenting the behavior of traditional spacecraft safe mode to account for the complex dynamics about small-bodies, and 2) rapid discovery of near-periodic orbits under the influence of inhomogeneous gravity fields. These problems are challenging to address due to the computationally intense algorithms needed to solve them which include reinforcement learning and boundary value problem methods. By taking advantage of the PINN gravity model’s efficiency, these once cumbersome algorithms can be evaluated orders-of-magnitude more quickly than before — enabling simulations and solvers to be run without unnecessary simplification of the gravitational dynamics. This research demonstrates that the PINN gravity model uniquely enables the training of more robust and dynamically-informed reinforcement learning agents, as well as assists traditional boundary value problem solvers to identify near-periodic orbits from arbitrary initial conditions.