Deep Equivariant Drag Coefficient Models for Satellite Re-Entry and Collision Analysis

Abstract: Drag modeling remains one of the most prominent sources of uncertainty in low- earth orbit (LEO) satellite operations. Machine learning has made numerical drag models more efficient by approximating atmospheric particle interactions, helping to reduce the vast computational cost. However, typical machine learning methods fail to leverage the symmetries associated with these particle interactions. We introduce the Deep Equivariant Drag Model (DEDM), a neural network that uses symmetry constraints to predict drag coefficients more accurately, especially with limited training data. We demonstrate our model’s ability to help predict orbital lifetimes and very close encounters between satellites in LEO.

(Camera-ready coming soon)