Artificial neural networks for multi-target low-thrust missions

Viavattene, Giulia (2022) Artificial neural networks for multi-target low-thrust missions. PhD thesis, University of Glasgow.

Full text available as:
[thumbnail of 2022ViavanettePhD.pdf] PDF
Download (20MB)


Multi-target missions are an attractive solution to visit multiple bodies in a single mission, increasing the scientific return and reducing the cost, compared to multiple missions to individual targets. Designing multi-target missions represents a challenging task since it requires multiple options to be estimated, given the large number of objects which can be considered as potential targets. Low-thrust propulsion systems are preferred to rendezvous multiple targets in a mission as they allow to utilise less propellant mass than high-thrust systems to perform the same trajectory. However, low-thrust trajectories are computationally expensive to compute.

This PhD thesis proposes to use artificial neural networks (ANN), as a fast and accurate estimation method for optimal low-thrust transfers. An artificial neural network and a sequence search (SS) algorithm can be designed to find solutions to three kinds of multi-target global optimisation problems: (i) multiple active debris removal missions (MADR), (ii) multiple near-Earth asteroid rendezvous (MNR) missions, with the option of returning a sample to Earth, and (iii) multi-objective optimisation of low-thrust propulsion systems for multi-target missions. MADR missions allows for the disposal of inactive satellites and larger objects, preventing the build-up of space junk and allowing to replace ageing agents in a constellation. Similarly, MNR missions allow to reduce the cost of each NEA observation and increase the possibility of visiting multiple asteroids of interest in a single mission.

The trained ANN is employed within a SS algorithm, based on a tree-search method and breadth-first criterion, to identify multiple rendezvous sequences and select those with lowest time of flight and/or required propellant mass. To compute the full trajectory and control history, the sequences are subsequently recalculated by using an optimal control solver based on a pseudospectral method. Also, to optimise the propulsion system for a given mission, a multi-objective optimisation using a genetic algorithm is performed, where ANNs are employed to quickly estimate the cost and duration of multi-target transfers.

The results show that neural networks can estimate the duration and cost of low-thrust transfers with high accuracy, for all the three applications. Employing machine learning within a sequence search algorithm to preliminary design multitarget missions allows to significantly reduce the computational time required with respect to other most commonly used methods in the literature, while maintaining a high accuracy. Given the combinatorial nature of the problem, the benefits in terms of computational time introduced by the ANN increase exponentially with a linear increase of the number of bodies in the database.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Colleges/Schools: College of Science and Engineering > School of Engineering
Supervisor's Name: Ceriotti, Dr. Matteo and McInnes, Professor Colin
Date of Award: 2022
Depositing User: Theses Team
Unique ID: glathesis:2022-83098
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 31 Aug 2022 10:59
Last Modified: 31 Aug 2022 11:02
Thesis DOI: 10.5525/gla.thesis.83098
Related URLs:

Actions (login required)

View Item View Item


Downloads per month over past year