Three-dimensional probabilistic stability analysis of an earth dam using an active learning metamodeling approach

The probabilistic stability analysis of earth dams is usually performed within two-dimensional (2D) computational models; thus, the 3D effect is ignored. Such a simplification could lead to biased estimates for the failure probability of dams especially for those located in narrow valleys. This article attempts to provide insights into the dam 3D probabilistic analysis by investigating the reliability of a real earth dam with field measurements and comparing the 3D results with the 2D ones. It is found that using a 3D computational model in a probabilistic analysis can give smaller estimates for the dam failure probability compared to the analyses based on a 2D section model. For the case study, the reduction of the failure probability is more significant in the case of a negative correlation between the soil shear strength parameters. The effects of using different deterministic mesh conditions on the dam reliability estimates are investigated as well. The results show that using a coarse mesh could lead to underestimated failure probabilities, especially for the 3D cases. The reliability analysis in this study is conducted by using an active learning surrogate modeling technique: adaptive sparse polynomial chaos expansions. This method is highly efficient in estimating failure probabilities and can provide an accurate approximation around the limit state surface by gradually adding well-selected samples into the current training set. The global sensitivity indices (Sobol) are also available in this method, so the contribution of each soil property to the variation of the dam safety factor is quantified and presented.