Work

Explainable Deep-Learning Based Potentially Hazardous Asteroids Classification Using Graph Neural Networks

PresentedHarvard AstroAI Workshop 2025.

  • 1 Carnegie Mellon University Africa
Explainable Deep-Learning Based Potentially Hazardous Asteroids Classification Using Graph Neural Networks — featured figure

Abstract

Classifying Potentially Hazardous Asteroids (PHAs) is crucial for planetary defense. While the official designation relies on simple metrics (Earth MOID and Absolute Magnitude H), we explore if complex dynamical relationships offer a robust, interpretable hazard assessment framework. We introduce a Graph Neural Network (GNN) approach that models asteroids as nodes defined by 11 orbital and physical features. Edges are defined using the Tisserand parameter relative to Jupiter (TJ) as a proxy for dynamical similarity, allowing the GNN to capture latent relationships among asteroid groups. On a dataset with a 0.22% hazardous class, the paper reports 99% accuracy and 0.99 AUROC, alongside 78% recall and 37% F1 for hazardous asteroids after SMOTE.

Supervision Clarence Worrel

  • Space
  • AI
  • Astrophysics