Comparative study of conditional methods in slope reliability evaluation

This paper investigates the performances of three conditional methods in slope reliability evaluation with spatially variable soils, considering various sampling strategies and rotated transverse anisotropy. The first method is based on conditional random field (RF) (i.e., conditional RF model 1), where the fluctuation component of the RF model is simulated using Kriging interpolation. The second method is also based on conditional RF (i.e., conditional RF model 2), where matrix decomposition is implemented on the conditional autocorrelation matrix. The third method originates from Sobol' index formulation, which is extended to enable the use of correlated input random variables. When sample points are distributed sparsely or near strata orientation, conditional RF model 1 would produce significantly smaller magnitude of uncertainty reduction and reliability index (beta) than those by other conditional methods. Also, in these situations, the standard deviation of factor of safety (FS) after conditioning by conditional RF model 1 may be larger than the unconditional standard deviation of FS. This is unexpected in slope reliability evaluation. Besides, beta by conditional RF simulation methods for reverse slopes may be smaller than that for dip slopes, which is against engineering experiences, while this issue cannot be found when using unconditional RF.