Comparing the performance of machine learning methods in a surrogate approach to optimize low-thrust interplanetary trajectories