Artificial Neural Network Fine-Tuning to Increase Accuracy of Low-Thrust Space Trajectory Cost Approximators

The use of artificial neural networks (ANNs) to approximate the cost (propellant mass or time of flight) of low-thrust trajectories has become increasingly prevalent in recent years, outperforming other methods such as analytical approximations or shape-based approaches both in CPU time and accuracy. An ANN requires a set of data on which it is to be trained—in the case of low-thrust trajectory approximators this is produced by pre-optimising a large number of trajectories between randomly generated orbits. Whilst this number is substantially smaller than the total number of trajectories that the ANN will approximate in its usage period, the computational cost associated with generating this training database is still significant (typically many days). This paper proposes that instead of generating a large dataset of optimal trajectories as in a conventional method, the ANN can first be pretrained on a large dataset of sub-optimal trajectories (‘low-resolution’ model) that can be produced very quickly. Once pretrained on this sup-optimal database, a small number of optimal trajectories (the ‘high-resolution’ model that ultimately needs to be approximated) can be generated and the pretrained ANN is fine-tuned on this small dataset. It is shown that such an approach enables reaching a high accuracy of fit to the high-accuracy trajectories in a significantly shorter time than if the ANN were trained directly on them with no pretraining.