EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Fatigue life prediction of flexible spacecraft solar arrays under multi-source loading via a physics-informed dual-stream neural network

Minghang Li, Hanwen Pu, Xingzhi Yang, Geng Chen

Frontiers in Mechanical Engineering · 25.06.2026 · 12

DOI: 10.3389/fmech.2026.1869741 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Frontiers in Mechanical Engineering, online erschienen am 25.06.2026. Verfasst von Minghang Li und 3 weiteren Autor:innen. Die Arbeit nennt 68 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Introduction

Flexible spacecraft solar arrays are subjected to multi-source thermo-vibro-mechanical loading in orbit. High-fidelity finite element analysis (FEA) is computationally prohibitive in high-cycle fatigue workflows. Purely data-driven surrogates are prone to overfitting in limited-data regimes and lack constraints rooted in basic mechanical principles. To address these challenges, this paper proposes Phy-Dual-Net, a physics-informed dual-stream neural network for multi-source fatigue life prediction.

Methods

A dataset of 450 loading cases is constructed using FEA. Rainflow counting is applied to each load sequence to extract dominant stress cycles, to which Goodman mean stress correction is then applied to compute a physics-informed feature S phy , embedding a physically grounded damage reference into the network input. An asymmetric dual-stream architecture is adopted: one stream captures cyclic loading effects from a temporal perspective, while the other aggregates macroscopic statistics to constrain fatigue damage accumulation. A hybrid loss function combining data-driven regression with physics-based regularization and a monotonicity penalty guides the network to obey fundamental fatigue damage laws.

Results

On the testing set ( N = 135 ), Phy-Dual-Net, enhanced by test-time augmentation and deep ensemble inference, achieves R 2 = 0.9966 with every prediction confined within the factor-of-two scatter band. Across five independent training runs, the model yields R 2 = 0.9904 ± 0.0010 with stable convergence.

Discussion

Compared with the computational cost of high-fidelity FEA, Phy-Dual-Net reduces the expense of fatigue life evaluation while preserving physically consistent responses across the full fatigue life range.

Zitieren

APA

Li, M., Pu, H., Yang, X., & Chen, G. (2026). Fatigue life prediction of flexible spacecraft solar arrays under multi-source loading via a physics-informed dual-stream neural network. Frontiers in Mechanical Engineering, 12. https://doi.org/10.3389/fmech.2026.1869741

Vancouver

Li M, Pu H, Yang X, Chen G. Fatigue life prediction of flexible spacecraft solar arrays under multi-source loading via a physics-informed dual-stream neural network. Frontiers in Mechanical Engineering. 2026;12. doi:10.3389/fmech.2026.1869741

BibTeX
@article{Li2026,
  author = {Li, Minghang and Pu, Hanwen and Yang, Xingzhi and Chen, Geng},
  title = {Fatigue life prediction of flexible spacecraft solar arrays under multi-source loading via a physics-informed dual-stream neural network},
  journal = {Frontiers in Mechanical Engineering},
  year = {2026},
  volume = {12},
  doi = {10.3389/fmech.2026.1869741}
}
Kostenlos starten →