For the two-impulsive lunar transfer problem, the trajectory can be optimized one by one but it is time-consuming. It is difficult to balance computation time and model precision, and this kind of balance is required for largescale analysis. Therefore, it is possible to apply the machine learning method to solve the problem in a fast and efficient way.
In the lunar transfer process, the spacecraft usually executes two impulsive manoeuvres from LEO to LLO. At Translunar Injection (TLI), the first impulse is assumed parallel to the parking velocity of LEO, which places the spacecraft into the LTO. At Lunar Orbit Injection (LOI), the second impulse inserts the spacecraft into LLO. It is noted that the transfer trajectory is not limited to the symmetric free-return orbit which is required by the manned spacecraft. The high-fidelity dynamics model is used for orbit propagation.
Basically, a LEO and an LLO are given to the system and the algorithm has to calculate the LTO between them. For training the algorithm they had 11000 LTOs, which they used 80% of to train the system and 20% to test the system. For results, each set of a LEO and LLO got a set of different LTOs, that satisfied the LEO and LLO. The most impressive part of this algorithm is the speed - it took 24 seconds to train the system and 0.1 seconds to implement it. Previous methods used, based on numerical optimisation, took about 20 seconds per trajectory, resulting in hundreds of hours of computing time.
The errors in this method were in an acceptable range, according to the authors. Although it seems, like the method for LTO calculation this paper has presented would be used in the further planning of Lunar missions, no actual code or test sets were presented, meaning that this algorithm would be impossible to reproduce and actually use.
Main article reference: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9185586