Machine Learning Based Guidance for Optimal Spacecraft De-Orbiting

Recently, Artificial Neural Networks (ANNs) have shown excellent performances in approximating solutions to optimal control problems, minimizing or maximizing a cost function. Some successful outcomes of this approach in the Space Engineering field regard interplanetary transfers, orbital maneuvering and landing, while no work focuses on the de-orbiting. This scenario inspired the authors to propose a novel application of this approach to a spacecraft de-orbiting, where high precision is required to ensure the satellite landing in a desired location. The current work presents a novel guidance system for Low Earth Orbit (LEO) based on ANNs. The training set relies on optimal de-orbiting solutions generated with an optimization algorithm developed by the authors. An innovative procedure for the training-set generation managed the computational cost associated with the resolution of a large number of optimal control problems. The possibility of ensuring high precision in reaching the target while guaranteeing minimum time and limited computational cost makes this algorithm extremely powerful and innovative, with many potential applications in space missions. A conclusive Monte Carlo analysis of over 500 cases further demonstrated the algorithm's robustness and generalization in a wide range of conditions.