Dataset for Machine Learning and Deep Learning Classification of Near-Earth Asteroid Ejection

This dataset contains two complementary data formats for classifying the ejection outcome of near-Earth asteroids (NEAs) using machine learning (ML) and deep learning (DL) methods. The data are derived from numerically integrated NEA trajectories over 1 Myr. The ML_Data folder contains tabulated orbital features extracted at the initial epoch, including semi-major axis, eccentricity, inclination, argument of perihelion, longitude of ascending node, mean anomaly, perihelion distance, aphelion distance, focal distance, and asteroid type. The class_1 file contains asteroids that were ejected during the 1 Myr integration, while class_2 contains asteroids that remained bound. For ML training, the classes should be balanced before training, with approximately 15,000 objects per class. The DL dataset contains recurrence-plot images generated from 0.2 Myr orbital time-series data and is organized into Train/class_1, Train/class_2, Test/class_1, and Test/class_2 directories. Here, class_1 represents the positive class (ejected asteroids) and class_2 represents the negative class (non-ejected asteroids). For CNN training, the training set should be split into train, validation, and test subsets in an approximate 80:10:10 ratio to support early stopping and reliable model selection. The trained CNN model weights (best_model.h5) are provided in this Zenodo record to enable direct reproducibility of the deep-learning results. The implementation of the machine-learning models, along with their corresponding trained weights and training pipeline, is available in the associated GitHub repository at https://github.com/chetan-hub-02/asteroid-dynamics-ml-dl.