Abstract
Implementing low-thrust propulsion on spacecraft can be quite advantageous, resulting in lower fuel consumptions, greater payload fractions, lower launch costs, and more. Due to these advantages, low-thrust propulsion has been used as the primary and secondary means of propulsion in almost every space application, though a majority of spacecraft launched with low-thrust propulsion have been geosynchronous equatorial orbit (GEO) satellites. As the use of spacecraft launched with low-thrust propulsion can be expected to continue increasing significantly, so can the need for quick, accurate, and robust low-thrust trajectory optimization as well as autonomous low-thrust control. However, due to lower thrust magnitudes and longer transfer times, low-thrust trajectory optimization via traditional optimization techniques can be complex as well as computationally-expensive and time-consuming. Artificial neural networks, on the other hand, can provide an alternative to such techniques. However, though the use of artificial neural networks trained by supervised learning models for applications related to low-thrust trajectory design and optimization is extensive, previous research has largely focused on shallow or deep feedforward architectures with little emphasis placed on recurrent ones, even though recurrent architectures are inherently more suited to time histories. Thus, the objective of this research was to investigate the potential of recurrent artificial neural networks, specifically long short-term memory artificial neural networks, for low-thrust trajectory design and optimization with respect to their feedforward equivalents. As a majority of spacecraft launched with low-thrust propulsion have primarily been GEO satellites, the focus of this research was, primarily, on the low-thrust, orbit-raising problem for both time-optimal and fuel-optimal transfers. Overall, this dissertation will present the results and conclusions from this research as well as the scientific contributions.