An Integrated Framework for Arch Dam Shape Optimization Using Stratified Conditional Sampling and Gaussian Process Surrogates

Shape optimization of arch dams is essential for balancing structural safety and economic efficiency, yet remains computationally intensive due to costly finite element analyses and strict geometric constraints. Conventional sampling techniques often yield infeasible designs that undermine surrogate model fidelity. This study proposes an integrated framework combining Stratified Conditional Latin Hypercube Sampling (SC-LHS), automated modeling, and Gaussian Process (GP) surrogate models. SC-LHS incorporates hierarchical constraints to eliminate infeasible samples during generation, while a Python-driven workflow automates the process from parameterization to simulation. Coupling the GP surrogate with NSGA-II enables efficient Pareto front exploration. The results indicate that SC-LHS is superior to standard LHS, Constrained LHS, and Sobol sequences with rejection in terms of feasibility rate and space-filling metrics. The optimal compromise solution reduces dam volume by 10.4% and tensile zone volume by 15.2% compared to the initial design. This framework effectively reconciles economic and safety objectives, offering a robust methodology for complex hydraulic structure design.