Coupled numerical and explainable machine-learning assessment of cracked diaphragm walls in anisotropic earth dams

Abstract

Abstract This study evaluates how diaphragm-wall cracking and hydraulic anisotropy jointly affect seepage discharge and downstream slope stability in earth dams. A coupled two-dimensional numerical framework was developed by transferring pore-water pressures obtained from steady-state SEEP/W analysis into SLOPE/W limit-equilibrium stability analysis. The numerical database contains 105 full-factorial scenarios combining the anisotropy ratio K y / K x = 0.0001–1.00, the crack-width ratio t / H = 0.01–0.50, and the crack-location ratio y / H = 0.25–0.75. The results show that increasing K y / K x and t / H intensifies seepage through the cracked diaphragm wall and reduces the factor of safety ( FS ), whereas increasing y / H reduces the hydraulic severity of the defect. Across the studied domain, q/(K y H) varies from 0.006 to 0.180 and FS varies from 1.979 to 0.533. Comparison with the benchmark field-scale diaphragm-wall model gave differences of 3.33% in seepage discharge and 0.27% in FS . Four tree-based models, namely random forest, extra trees, gradient boosting, and XGBoost, were developed as numerical surrogates for q /( K y H ) and FS . Gradient boosting achieved five-fold out-of-fold R² values of 0.9950 for q /( K y H ) and 0.9959 for FS . SHAP analysis identified K y / K x and t / H as the dominant controls on seepage discharge, while K y / K x was the dominant control on FS . The study quantifies the coupled effects of crack geometry and hydraulic anisotropy and provides an explainable numerical screening approach for cracked diaphragm walls within the investigated parameter ranges.