Towards a machine learning classification model of asteroids' co-orbital regimes

The aim of this work is to classify co-orbital motion at different timescales using a machine learning approach. Asteroids that move, on average, in a 1:1 mean motion resonance with a given planet, under the assumptions of the restricted three-body problem, are of special interest, because they follow very stable orbital configurations. The capability of classifying this kind of dynamics in an automatic way is of paramount importance to understand key mechanisms in the solar system but also for space exploration and exploitation missions. Starting from the analysis of Tadpole, Quasi-Satellite and Horseshoe regimes on a medium timescale, we apply features extraction and machine learning algorithms to the timeseries of the relative angle between the asteroid and the planet. The dataset is composed of pseudo-real timeseries computed by means of the JPL Horizons system (thus considering a full dynamical model) and simulated data computed by means of the REBOUND software. In a second step, the process is enriched to also be able to classify relative transitions between different co-orbital motions and between resonant and non-resonant dynamics, thus also considering much longer timescales. To this end, different algorithms are evaluated, starting from classical statistical approaches to more advanced deep learning ones.