Future subsequent ballistic coefficient estimation using deep learning for uncontrolled reentry object prediction

Abstract

Uncontrolled reentries of space objects pose increasing risks to populated regions as the Low Earth Orbit (LEO) environment becomes more congested. Accurate reentry forecasting remains challenging due to rapidly evolving atmospheric drag, uncertainties in orbital elements near end-of-life, and the lack of recent tracking data as reentry approaches. To address these limitations, we propose a hybrid physics–data-driven framework that estimates and forecasts a satellite’s time-varying ballistic coefficient (BC) using a dual-encoder sequence-to-sequence Long Short-Term Memory (Seq2Seq LSTM) architecture. Historical BC profiles are derived from Two-Line Element (TLE) records using a physics-based semi-major axis decay model, combined with space weather and orbital parameter histories spanning an entire solar cycle. The trained deep learning model is integrated into a high-order Python-based orbital propagator employing a Dormand–Prince 8(5,3) integrator with atmospheric drag and zonal harmonics up to degree 6. BC predictions are updated every three hours during propagation in alignment with geomagnetic index availability. The method is evaluated on three representative uncontrolled reentry events—Cosmos-482 Descent Craft, NEOWISE Spacecraft, and CZ-4B R/B. Results show close agreement with authoritative prediction reports and observations, yielding reentry-time uncertainties within 11.4 min for Cosmos-482, 7.8 min for NEOWISE, and 13.2 min for CZ-4B, and showing better performance against the constant-BC reentry prediction model. These findings demonstrate that dynamically coupling physics-based BC estimation with deep learning forecasts could substantially improve reentry prediction robustness and accuracy. The proposed framework represents a promising pathway toward integrating deep learning into operational Space Situational Awareness (SSA) and Space Traffic Management (STM) systems.

Keywords

Bibliographic record

BibTeX Citation

@article{Andaristiyan2026futuresubsequent,
  author = {Andaristiyan, M.K. and Machmudah, A. and Poetro, R.E. and Hasbi, W. and Dutykh, D.},
  title = {Future subsequent ballistic coefficient estimation using deep learning for uncontrolled reentry object prediction},
  journal = {Acta Astronaut.},
  year = {2026},
  volume = {248},
  pages = {736--749},
  doi = {10.1016/j.actaastro.2026.06.022},
  abstract = {Uncontrolled reentries of space objects pose increasing risks to populated regions as the Low Earth Orbit (LEO) environment becomes more congested. Accurate reentry forecasting remains challenging due to rapidly evolving atmospheric drag, uncertainties in orbital elements near end-of-life, and the lack of recent tracking data as reentry approaches. To address these limitations, we propose a hybrid physics–data-driven framework that estimates and forecasts a satellite’s time-varying ballistic coefficient (BC) using a dual-encoder sequence-to-sequence Long Short-Term Memory (Seq2Seq LSTM) architecture. Historical BC profiles are derived from Two-Line Element (TLE) records using a physics-based semi-major axis decay model, combined with space weather and orbital parameter histories spanning an entire solar cycle. The trained deep learning model is integrated into a high-order Python-based orbital propagator employing a Dormand–Prince 8(5,3) integrator with atmospheric drag and zonal harmonics up to degree 6. BC predictions are updated every three hours during propagation in alignment with geomagnetic index availability. The method is evaluated on three representative uncontrolled reentry events—Cosmos-482 Descent Craft, NEOWISE Spacecraft, and CZ-4B R/B. Results show close agreement with authoritative prediction reports and observations, yielding reentry-time uncertainties within 11.4 min for Cosmos-482, 7.8 min for NEOWISE, and 13.2 min for CZ-4B, and showing better performance against the constant-BC reentry prediction model. These findings demonstrate that dynamically coupling physics-based BC estimation with deep learning forecasts could substantially improve reentry prediction robustness and accuracy. The proposed framework represents a promising pathway toward integrating deep learning into operational Space Situational Awareness (SSA) and Space Traffic Management (STM) systems.},
  keywords = {uncontrolled reentry, reentry prediction, deep learning, ballistic coefficient, Seq2Seq LSTM},
  publisher = {Elsevier BV},
  issn = {0094-5765}
}