Master's Thesis Defense

Furkan Gonul

(Faculty advisor: Professor Dimitri Mavris)

"SURROGATE MODELING FOR LOW THRUST EARTH TO MOON SPIRAL TRAJECTORY DESIGN AND PRELIMINARY MISSION FEASIBILITY ASSESSMENT"

Tuesday, November 17

8:00 - 10:00 a.m.

Weber, CoVE

Abstract: 

Low thrust electric propulsion offers propellant savings. However, designing such
transfers is computationally expensive. Formal optimization techniques for multi revolution
spiral trajectories often take hours per trajectory and require manual initial guess
calculations which makes it difficult to explore the design space in the preliminary mission
design stage where engineers need to evaluate hundreds to thousands of candidate designs.
Machine learning surrogate models offer an approximation that can deliver mission feasibility
predictions in milliseconds rather than hours after training.

Despite its potential, the use of such surrogates for low thrust spiral Earth to Moon
trajectories has not been investigated thoroughly. This thesis asks which representation
of the transfer makes the surrogate trainable, whether a trained model with a feasibility
classifier can serve as a sizing and screening tool in preliminary mission design. Also
under which conditions a physics informed neural network behaves better than a purely data
driven one. Thesis uses a dataset of 30,000 planar transfers flown by the Q-law guidance
law in modified equinoctial elements with the Moon as a third body.

Sampling the transfer once per revolution, at perigee, is the lead contribution. It removes
the oscillation within each revolution and makes the surrogate more accurate. The
surrogate’s error in the propellant of an unseen design is 3.2 kg on average, 0.7 percent of
the spacecraft’s initial mass and one prediction takes about 3 milliseconds. Therefore, the
surrogate’s final mass is a conservative estimate of the mission and its prediction served
as a starting guess from which the optimizer converged. The physics informed loss, tested
on a second model, helps when training transfers are few or noisy. The results show that a
neural network surrogate model can be used to investigate the design space in milliseconds
for preliminary mission design.

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