Residual learning method for Mars aerobraking autonomous predictive guidance

Due to the extended duration and highly nonlinear Mars aerobraking dynamics, state prediction via numerical integration is computationally prohibitive for predictive guidance. In this paper, a residual learning method for autonomous predictive guidance is proposed. The orbital state of the spacecraft governed by high-fidelity Mars aerobraking dynamics is demonstrated to consist of a simplified two-body component and a residual component. The solution for the simplified two-body component is first analytically derived, while the residual component arising from the nonspherical gravitational perturbations and aerodynamic accelerations is compensated by a neural network. With its specially designed structure, the neural network achieves better physical consistency with the integration process. Compared to directly learning the orbital states, it is proved that learning the residual mitigates the vanishing gradient problem, thereby improving prediction accuracy. Leveraging the residual learning method, a predictive guidance method is developed. By predicting the spacecraft’s orbital states over future revolutions, the velocity increments are optimized to maximize aerodynamic deceleration while ensuring compliance with the heat flux constraint. The guidance method is also proved to be asymptotically stable. Finally, the effectiveness of the proposed method is validated through Mars aerobraking numerical simulations.