Abstract
Weak interlayers influence seepage and damage evolution in roller-compacted concrete (RCC) gravity dams. This study develops a two-dimensional hydro–mechanical–damage (HMD) coupled finite-element model incorporating damage-dependent hydraulic conductivity, assessed through benchmark comparisons and local sensitivity analyses. A dataset of 1500 simulations generated using Latin hypercube sampling is used to compare four surrogate models: multilayer-perceptron–artificial-neural-network (MLP-ANN), XGBoost, random forest, and support vector models. Numerical comparisons indicate that damage-dependent conductivity promotes interlayer damage extension. Increasing upstream water level amplifies plastic deformation and crest displacement, whereas reduced interlayer tensile strength increases cracking susceptibility. For samples classified as Normal, MLP-ANN achieves the highest average test R2 of 0.965 across the equivalent interlayer damage-length ratio, log-transformed equivalent plastic strain, and log-transformed crest displacement. For numerically defined High-risk classification, MLP-ANN achieves recall of 0.939 and an F1-score of 0.805 at a probability threshold of 0.50, with a ROC-AUC of 0.979 and a PR-AUC of 0.828. KernelSHAP identifies upstream water level as the dominant attribution variable, with interlayer tensile strength, hydraulic conductivity, downstream slope, and cohesion providing secondary contributions. The framework supports interpretable numerical response prediction and adverse-parameter screening within the investigated parameter space.