DNN estimation of low-thrust transfer time: Focusing on fast transfers in multi-asteroid rendezvous missions

Abstract

In the preliminary design phases of low-thrust multi-asteroid missions, there are so many alternative transfer trajectories that it is impossible to simultaneously optimize them with high fidelity. Therefore, a fast and accurate method to evaluate the cost of low-thrust transfer is significant. In this work, a low-thrust time-optimal multi-asteroid rendezvous mission is studied, where deep neural networks (DNNs) are utilized to estimate the minimum time of flight (TOF) of a single-leg transfer between every two asteroids. By restricting the initial states during the dataset generation and constructing new loss functions, the DNN prediction accuracy of fast transfers, i.e., the transfers with short TOFs, is greatly improved while the overall accuracy is not discounted. Furthermore, beam search is modified to optimize the sequences of multi-asteroid missions. The solution sequences retained after the extension of each level in beam search are optimized by an indirect method to eliminate the impact of estimation error on subsequent searches. Finally, the DNNs are well trained and employed in the modified beam search to find the optimal sequence. A time-optimal multi-rendezvous example of 10 main-belt asteroids is investigated. The results demonstrate that the DNNs can accurately estimate the minimum TOF and that the proposed methods can achieve better sequences than those previously known.