Denvir Higgins, Lawrence Livermore National Laboratory; Kerianne Pruett, Lawrence Livermore National Laboratory; Travis Yeager, Lawrence Livermore National Laboratory; Michael Schneider, Lawrence Livermore National Laboratory
Keywords: SOM, cislunar, unsupervised machine learning, maps
Abstract:
Cislunar orbits occupy an immense volume of phase-space, spanning from GEO out to where the Earth-Moon system is gravitationally dominant. Most existing cislunar models focus on small regions in this phase-space and target individual missions and assume simplified gravitational models (e.g., CR3BP). However, in this regime chaos can dominate dynamics, requiring n-body simulations for understanding and characterizing cislunar orbits (which becomes increasingly more important with increasing orbital timescales). Additionally, TLE’s, which are commonly used as a method for summarizing orbital states (often used to understand similarity in close orbits), fail to properly capture orbital properties when 3- or n-body dynamics are important. Machine learning methods can be applied to complicated datasets to aid humans in understanding and characterizing these orbits. Due to the complexity of cislunar orbits, there are not obvious classes or labels that one could apply to similar orbits, making supervised learning (i.e., labeled) methods a non-viable solution. Unsupervised learning (i.e., unlabeled) will be the superior approach for understanding orbits in this space, potentially providing labels for things that were previously unknown, and that we may not know we are looking for. To begin understanding orbits in this complex space, we simulate one million orbits (considering n-body dynamics) over varying timescales utilizing high-performance computing resources at Lawrence Livermore National Laboratory (LLNL) and leveraging an LLNL-developed space situational awareness python package (SSAPy). Using this library of high-fidelity cislunar orbits we generate self-organizing maps (SOMs; an unsupervised machine learning method that creates a feature space where like features are grouped together) to characterize these orbits. Through these mappings we infer properties for different cislunar orbits (e.g., predicted lifetimes, end-of-life outcomes, and potential mission relevant orbits). Additionally, these methods can aid in identifying new families or classes of orbits and find new stable regions in cislunar phase-space, opening up new avenues for future cislunar mission orbits.
Date of Conference: September 17-20, 2024
Track: Cislunar SDA