Machine learning regression for estimating characteristics of low-thrust transfers

Abstract

In this thesis, a methodology for training machine learning algorithms to predict the fuel and time costs of low-thrust trajectories between two objects is developed. In order to demonstrate the methodology, experiments and hypotheses were devised. The first experiment identified that a direct method was more efficient than an indirect method for solving low-thrust trajectories. The second experiment, an offshoot of the first, found that the Sims-Flanagan method as implemented in the Python library PyKEP would be the most efficient manner of creating the training data. The training data consisted of the orbital elements of both the departure and arrival bodies, as well as the fuel and time-of-flight associated with a transfer between those bodies. A total of 7,218 transfers made up the training data. After creating the training data, the third and final experiment could be conducted, to see if machine learning methods could accurately predict fuel and time costs of low-thrust trajectory for a larger design space that had been investigated in previous literature. As such, the training data consisted of transfers, generated using a space-filling Latin Hypercube design of experiments, between bodies of highly varying orbital elements. The departure and arrival bodies’ semimajor axis and inclination differ much more than in previous literature. It was found that all the machine learning regression methods analyzed greatly outperformed the Lambert predictor, a predictor based on the impulsive thrust assumption. The accuracy of the time-of-flight prediction was close to that of the mass prediction when considering the mean absolute error of the expended propellant mass.