Hybrid optimization and AI-driven surrogate model for seepage parameters inversion in complex dam foundations

•A hybrid algorithm and AI surrogate model approach is proposed for seepage inversion.•Improved SLE-CLHS technique ensures uniform coverage in constrained parameter space.•Improved BKA employs dynamic switching and random learning for robust optimization.•The proposed approach replaces conventional...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Journal of hydrology (Amsterdam) Ročník 664; s. 134484
Hlavní autoři: Zheng, Changhai, Cen, Weijun, Liu, Bo, Qian, Jiahao, Ding, Yuxuan, Mo, Chongxun
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier B.V 01.01.2026
Témata:
Coefficient of permeability
Gated recurrent unit
Improved black-winged kite algorithm
Parameter inversion
Seepage calculation
Surrogate model
Improved black-winged kite algorithm
Gated recurrent unit
Parameter inversion
Seepage calculation
Surrogate model
Coefficient of permeability
ISSN:0022-1694
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:•A hybrid algorithm and AI surrogate model approach is proposed for seepage inversion.•Improved SLE-CLHS technique ensures uniform coverage in constrained parameter space.•Improved BKA employs dynamic switching and random learning for robust optimization.•The proposed approach replaces conventional FEM, boosting efficiency by 129.5 times. Accurate estimation of seepage parameters is critical for analyzing seepage behavior in concrete face rockfill dams (CFRDs), particularly under complex geological conditions where foundation permeability remains a challenge. Conventional approaches combining optimization algorithms with finite element method (FEM) simulations are often limited by computational efficiency and accuracy. To address these limitations, a novel approach is proposed that integrates improved sampling techniques, optimized high-performance algorithms, and an enhanced machine learning-based surrogate model for efficient and precise seepage parameter inversion. In this approach, parameter space exploration is optimized by successive local enumeration constrained Latin hypercube sampling (SLE-CLHS) to ensure comprehensive and balanced coverage. Global and local optimization efficiency is enhanced by an improved black-winged kite algorithm (IBKA) incorporating dynamic switching and stochastic learning. FEM is replaced by a gated recurrent unit (GRU) neural network, capturing the seepage field’s spatiotemporal dynamics. Engineering case studies demonstrate a maximum absolute error of 1.12 m and a maximum relative error of 0.97 % in seepage pressure, which are within engineering tolerances. The proposed approach enhances computational efficiency by 129.5 times compared to FEM-based inversion and achieves over 53.67 % higher accuracy under equal time constraints, confirming its reliability and practical applicability. Superior performance is also observed against popular surrogate models (e.g., backpropagation neural network, extreme learning machine) in terms of accuracy and generalization, with inverted coefficients of permeability yielding seepage fields closely aligned with measured data, confirming its reliability for complex dam foundations. This integrated framework provides a robust solution for seepage parameter identification in complex geological environments, offering significant value for dam safety assessment.
ISSN:0022-1694
DOI:10.1016/j.jhydrol.2025.134484