A methodology for ultra-high dimensional reliability analysis of slopes considering spatial variability of soil properties

The reliability analysis of slopes under spatially variable parameters faces the challenges of high computational cost and the curse of dimensionality. This paper proposes a methodology for performing reliability analysis in ultra-high dimensions through deep dimension reduction and optimal metamodel identification. Regarding the dimension reduction of variables coupled with meta-modeling for safety factors, the proposed method minimizes the input dimensions while achieving optimal prediction performance. This methodology has several primary sources of novelty. First, basic random variables that affect slope stability are derived using the Karhunen-Loève expansion and correlation analysis. Second, the metamodel cluster for safety factors with varying input dimensions is developed using the least squares support vector machine. Then, the optimal random variables and prediction model are identified through the evaluation criterion that integrates fitting and overfitting performance. Third, an improved Latin hypercube design with stronger space-filling performance is proposed to enhance meta-modeling. The proposed method has been verified through a classic slope and effectively applied to a hypothetical slope. The results indicate that the method can significantly reduce the number of random variables while maintaining satisfactory accuracy in ultra-high-dimensional problems, which can be applied to the reliability analysis of large-sized geotechnical structures.