Solar sail trajectory cost estimation with transfer learning

Many future space missions will require evaluating a large number of orbital transfers to determine an optimal combination thereof (for instance for asteroid exploration or debris removal), likely using low-thrust propulsion such as solar sailing. Optimising individual trajectories is already a non-trivial process so evaluating millions for a multi-target mission can become prohibitively computationally expensive. This cost is further amplified if the spacecraft model changes as the mission design progresses, meaning that a previously computed set of trajectories may need to be re-evaluated. Artificial neural networks are often used to approximate the cost of optimal trajectories with a very fast inference time, but require upfront computational, including generating pre-optimised trajectories for training. This paper considers a scenario where two solar sail models need to be evaluated, and ANNs are trained to approximate the duration of minimum-time transfers. It is shown that an ANN does not need to be trained afresh (requiring an extensive training database) for both models—rather, an ANN trained to a good accuracy on one model can be fine-tuned on a smaller training database for the other model (and arbitrarily many others), saving a significant amount of time on database generation.