Short-duration Backward Integrations as Phase-space Diagnostics: A Sequence-based Machine Learning Approach to Near-Earth Asteroid Triage

Abstract Understanding the dynamical pathways by which small bodies are transported into near-Earth space remains central to planetary dynamics, impact risk assessment, and source-region inference. Here we combine high-precision N -body integrations with sequence-based machine learning to test whether short orbital segments contain useful dynamical diagnostics for predicting long-term outcomes of near-Earth asteroids (NEAs). We use 1 Myr backward integrations with the M ercury N -body integrator to define ensemble labels (ejection versus long-lived Earth- or Mars-crossing behaviour) and train recurrent networks—gated recurrent unit (GRU) and long short-term memory (LSTM)—on the initial 0.2 Myr of each backward trajectory. Importantly, we treat backward segments as phase-space diagnostics, not unique reconstructions of individual pasts, and interpret results at the ensemble level. GRU and LSTM models using semimajor axis and eccentricity reach test accuracies of ≈86% (ROC AUC ≈0.94); eccentricity-only inputs approach the two-parameter performance. Post hoc attributions (LIME, SHAP) identify early semimajor axis variations and late eccentricity growth as the dominant predictors, consistent with resonant diffusion and secular forcing in the underlying dynamics. We demonstrate robustness through clone ensembles and parameter sensitivity tests. Short backward segments therefore provide a computationally efficient and physically grounded diagnostic to triage NEAs for targeted, higher-fidelity dynamical follow-up.