Long-term Dynamical Evolution and Ejection of Near-Earth Asteroids

Abstract Long-term integrations of asteroid orbits with high-accuracy numerical integrators are essential for understanding dynamical evolution and ejection from the solar system, but are computationally expensive. Here, we investigate the dynamical behavior of asteroids and explore machine learning (ML) and deep learning approaches as efficient, scalable alternatives for classifying long-term dynamical outcomes. While the ML classifiers are trained on initial orbital elements, the convolutional neural network is trained on recurrence plots derived from short-period numerical integrations generated with the MERCURY integrator. Ensemble tree models perform strongly on the ephemeris input, and the neural network captures temporal signatures of chaotic motion with comparable or slightly improved accuracy. Backward integrations reveal a partial overlap between forward- and reverse-ejected sets, illustrating time-asymmetric behavior in chaotic regions; these backward results are interpreted only as diagnostic probes rather than reconstructions of past histories. Nonejected asteroids largely correspond to known dynamical groups, underscoring the constraining role of the initial orbital configuration. These methods provide scalable frameworks to complement numerical integrations and inform prioritization for detailed long-term dynamical studies, with implications for planetary-defence analyses.