Physics-Informed Graph Neural Network for Flight Dynamics Modeling
Aerospace · 16.05.2026 · 13 (5) · 471
DOI: 10.3390/aerospace13050471 ↗Kurz gefasst
Fachartikel in der Zeitschrift Aerospace, online erschienen am 16.05.2026. Verfasst von Liang Ma und 4 weiteren Autor:innen. Die Arbeit nennt 18 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Flight dynamics modeling is a fundamental cornerstone of aircraft design, simulation, and control. Traditional approaches rely on aerodynamic look-up tables for numerical integration, which suffer from high data-acquisition costs, poor extrapolation capability, and difficulty in assimilating flight test data. This paper proposes an architectural integration of physics-informed neural networks (PINNs), graph neural networks (GNNs), and known flight mechanics equations for flight dynamics modeling. Without requiring aerodynamic coefficient labels, the method predicts flight state derivatives using state-transition data. The approach encodes the structural knowledge of flight mechanics equations into graph topology and a physics computation layer (PhysicsLayer), so that the neural network only needs to learn the unknown aerodynamic coefficients while all remaining physical relationships are computed by the governing equations. Using an F-16 fighter six-degree-of-freedom model as the verification platform, an ablation study involving Direct-MLP, PINN, PIGNN, and GNN is conducted. Results show that the PIGNN architecture improves single-step derivative prediction accuracy by 86.6% over Direct-MLP, 60.9% over pure PINN, and 90.8% over GNN. In 499-step (approximately 5 s) rollout state prediction, the PIGNN Core RMSE is 1.1554, with approximately linear error growth within the first 100 steps indicating well-controlled short-range error accumulation. The graph-structural prior enables the network to learn aerodynamic coefficients that closely match the F-16 reference aerodynamic database without aerodynamic coefficient supervision. The results demonstrate that combining graph-based dependency modeling with hard physical constraints is effective for interpretable flight dynamics surrogate modeling.
Zitieren
APA
Ma, L., Li, Z., Zhang, J., Li, Y., & Deng, S. (2026). Physics-Informed Graph Neural Network for Flight Dynamics Modeling. Aerospace, 13(5), 471. https://doi.org/10.3390/aerospace13050471
Vancouver
Ma L, Li Z, Zhang J, Li Y, Deng S. Physics-Informed Graph Neural Network for Flight Dynamics Modeling. Aerospace. 2026;13(5):471. doi:10.3390/aerospace13050471
BibTeX
@article{Ma2026,
author = {Ma, Liang and Li, Zhanwu and Zhang, Juntao and Li, You and Deng, Shijie},
title = {Physics-Informed Graph Neural Network for Flight Dynamics Modeling},
journal = {Aerospace},
year = {2026},
volume = {13},
number = {5},
pages = {471},
doi = {10.3390/aerospace13050471}
}
EUVIMED