Spacecraft trajectory optimization is a critical aspect of space mission analysis. In recent years, there has been an increased interest within NASA in applying machine-learning algorithms to improve the performance of trajectory optimization solvers. Optimization of trajectories for spacecraft employing solar-electric propulsion is a challenging problem because it requires the solution of a nonlinear, non-convex mathematical programming problem. This problem is even more complicated when the spacecraft is located close to a planetary body. First, the low-thrust propulsion system provides a small acceleration relative to the local gravitational acceleration, making the transfer long and complex. Second, the presence of the planets shadow prohibits thrust generation by electric thrusters, thereby making the transfer multi-phase. Third, gravitationally trapped radiation degrades the spacecraft solar array that powers the electric thrusters.

This project targets development of a new, machine-learning assisted optimization tool for on-ground mission design. The automated, fast and robust nature of the proposed methodologies makes the tool suitable for onboard implementation as well. The architecture of the proposed software allows for sequential progression of fidelity by incorporating increasingly rigorous force models at different levels of trajectory optimization; this facilitates the improvement of lower-fidelity solutions, while simultaneously managing the computational complexity of the underlying problem in an automated manner. Proposed modular architecture allows for application of the proposed software in two different settings, such as preliminary mission analysis by ground personnel and onboard mission planning. The overall trajectory design is modelled as a two-level process, with the low-level trajectory optimization phase, and a high-level planning that allows for the application of machine learning techniques to trajectory optimization. The project plans to incorporate the following innovations: using dynamical coordinates in trajectory optimization, a modified state observer to estimate unmodeled acceleration, and the use of an artificial neural network for adaptive tuning of planning variables. Additionally, in the context of onboard implementation, the project will consider data-driven updates of the neural networks based on information obtained for sensors. The project will also consider the addition of atmospheric drag models for analysis of aero-capture and atmospheric entry.