Detection of interconnections between unstable resonant orbits via machine learning

In this paper, deep neural networks and extreme learning machine are applied as efficient substitutes for the unsupervised approach to the identification of interconnections between unstable resonant orbits, which was presented earlier. This phenomenon induces multiple transitions between resonances at specific energies, giving rise to a multi-energy multi-label classification problem. The learning strategy is addressed by subdividing it into a series of independent binary classification problems, in which Bayesian optimisation and genetic algorithms are used to fine-tune the architectures of the trained models, prioritising precision over recall to enhance the reliability of the predictions. The constructed hybrid composite multi-classifier shows overall a promising ability to predict new interconnections, given the small dataset used, and paves the way for examining alternative data-driven techniques. Two case study scenarios, illustrated for the CR3BP Jupiter-Europa system, demonstrate the usefulness of developing predictive tools to assist in preliminary resonant flyby designs that require rapid consideration of multiple solutions.