Machine Learning of Optimal Low-Thrust Transfers Between Near-Earth Objects

Abstract

During the initial phase of space trajectory planning and optimization, it is common to have to solve large dimensional global optimization problems. In particular continuous low-thrust propulsion is computationally very intensive to obtain optimal solutions. In this work, we investigate the application of machine learning regressors to estimate the final spacecraft mass $m_f$ after an optimal low-thrust transfer between two Near Earth Objects instead of solving the corresponding optimal control problem (OCP). Such low thrust transfers are of interest for several space missions currently being developed such as NASA’s NEA Scout.

Previous work has shown machine learning to greatly improve the estimation accuracy in the case of short transfers within the main asteroid belt. We extend this work to cover also the more complicated case of multiple-revolution transfers in the near Earth regime. In the process, we reduce the general OCP of solving for $m_f$ to a much simpler OCP of determining the maximum initial spacecraft mass $m^*$ for which the transfer is feasible. This information, along with readily available information on the orbit geometries, is sufficient to learn the final mass $m_f$ for the same transfer starting with any initial mass $m_i$. This results in a significant reduction of the computational cost compared to solving the full OCP.