Student Name: Furkan Gonul

Advisor: Dr. Dimitri Mavris

Milestone: MS Thesis Proposal

Degree Program: Aerospace Engineering

Title: Surrogate Modeling for Low Thrust Earth to Moon Spiral Trajectory Design and Preliminary Mission Feasibility Assessment

Abstract: Transfers from Earth to the Moon are becoming increasingly attractive due to the need for a permanent base on the Moon for interplanetary missions and cargo deliveries. Low thrust electric propulsion offers substantial propellant savings for such a trajectory. However, designing such transfers is computationally expensive, and there is insufficient real data from real missions for study. Formal optimization techniques for multi revolution spiral trajectories often take hours per trajectory and require comprehensive manual initial guess calculations. This makes it very difficult to explore the design space iteratively in the preliminary mission design stage, where engineers need to evaluate hundreds to thousands of combinations of candidate parameters. Machine learning surrogate models, however, offer a promising approximation that can deliver mission feasibility predictions in milliseconds rather than hours after training. Despite its potential, the use of machine learning surrogates for low thrust spiral Earth to Moon trajectories has not been investigated due to the spectral bias nature of these trajectories. Therefore, several key questions remain unanswered: Which state representation makes the surrogate model trainable? When will a physics informed neural network behave better than a purely data driven one, and can a single trained model actually be used in preliminary mission design? This thesis addresses these gaps by using a heuristic trajectory dataset in modified equinoctial elements and systematically comparing purely data driven and physics informed neural networks. It also investigates methods to improve speed and accuracy, such as perigee sampling and architecture that applies Fourier features to oscillatory channels. Preliminary findings indicate that the MEE formulation decreases the effect of spectral bias and provides trainable data for the neural network. Perigee based sampling significantly improves accuracy and training time, as it also reduces the effect of spectral bias. On clean data, the physics informed neural network reduces accuracy, but under noise, it matches the accuracy of the purely data driven network while yielding far more physically consistent predictions. The results show that a neural network surrogate model can be used to investigate the design space in milliseconds for preliminary mission design. 

Date and time: 2026-05-08, 09:00 - 11:00 am

Location: Collaborative Visualization Environment (CoVE) Weber SST II

Committee:
Dr. Dimitri Mavris (advisor), School of Aerospace Engineering
Prof. Graeme Kennedy , School of Aerospace Engineering
Dr. Tristan Sarton du Jonchay , School of Aerospace Engineering