Improved support point selection on adaptive kriging metamodels for reliability analysis of soil slopes

The paper presents an iterative approach based on stochastic simulations and adaptive Kriging metamodels to perform reliability and safety assessments of soil slopes. Two new rules for adaptively selecting support points are proposed, considering an entropy learning function and the closeness to the failure domain defined by a limit state function. In addition, a stopping criterion is proposed based on root-mean-square and mean absolute percentage errors computed with cross-validation at the local level, focusing on regions where the uncertainties are relevant. Finally, the selection rules for support points and the error metrics are implemented in two benchmark problems with a low, moderate, and high probability of failure. Ultimately, the work leads to an adaptive Kriging strategy for slope stability assessment, offering: (1) a fair comparison with other strategies based on a significant number of realizations, (2) a stopping criteria based on a new local error metric, (3) an insight of the behavior across different magnitudes of the probability of failure, and (4) a new selection rule that reduces the total number of support points significantly. The proposed scheme is easily paired with commercial software to compute support points, resulting in an attractive tool for practitioners.