EUVIMEDEUVIMED entdecken
Crossref

Residual-compensated physics-reinforced neural networks for reduced-order modeling with incomplete governing equations

Honglin Wan, Yuanhong Chen, Xiang Sun, Zhen Gao

Physics of Fluids · 01.11.2026 · 38 (11)

DOI: 10.1063/5.0353589 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Physics of Fluids, online erschienen am 03.11.2026. Verfasst von Honglin Wan und 3 weiteren Autor:innen. Die Arbeit nennt 58 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Shandong Provincial Key Laboratory of Stochastic System Control and Scientific Computing.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

For parametrized partial differential equation (PDE) dynamical systems with partially known governing equations, we propose a residual-compensated physics-reinforced neural network (RC-PRNN) for reduced-order modeling. The method constructs a low-dimensional representation using proper orthogonal decomposition (POD) and learns the mapping from time and parameters to the corresponding reduced-order coefficients. To address modeling errors caused by incomplete physics, an auxiliary residual-compensation network is introduced to learn an effective residual correction for the incomplete reduced-order dynamics as a function of time, parameters, and reduced-order coefficients. The learned compensation is then embedded into the reduced-order equations to form a corrected physics constraint. RC-PRNN is trained by combining the data mismatch of labeled POD coefficients with a compensated reduced-order residual loss term. Additional residual points in the target time-parameter domain are further used to enrich the physical constraints without requiring extra high-fidelity full-order simulations. Numerical experiments demonstrate that, for parametrized dynamical systems with complex incomplete physics, RC-PRNN achieves higher prediction accuracy than conventional PRNN and purely data-driven methods. Moreover, for high-dimensional problems, the proposed method achieves significantly lower computational cost than PDE-preserved neural networks, providing an accurate and efficient framework for physics-informed reduced-order modeling with incomplete governing equations.

Förderung

  • National Natural Science Foundation of China
  • Shandong Provincial Key Laboratory of Stochastic System Control and Scientific Computing

Zitieren

APA

Wan, H., Chen, Y., Sun, X., & Gao, Z. (2026). Residual-compensated physics-reinforced neural networks for reduced-order modeling with incomplete governing equations. Physics of Fluids, 38(11). https://doi.org/10.1063/5.0353589

Vancouver

Wan H, Chen Y, Sun X, Gao Z. Residual-compensated physics-reinforced neural networks for reduced-order modeling with incomplete governing equations. Physics of Fluids. 2026;38(11). doi:10.1063/5.0353589

BibTeX
@article{Wan2026,
  author = {Wan, Honglin and Chen, Yuanhong and Sun, Xiang and Gao, Zhen},
  title = {Residual-compensated physics-reinforced neural networks for reduced-order modeling with incomplete governing equations},
  journal = {Physics of Fluids},
  year = {2026},
  volume = {38},
  number = {11},
  doi = {10.1063/5.0353589}
}
Kostenlos starten →