A comparative study of Bayesian inverse analyses of spatially varying soil parameters for slope reliability updating

Bayesian estimation of spatially varying soil parameters is a challenging task in geotechnical engineering because a large number of random variables need to be learned. To address this challenge, three Bayesian methods are revisited, including Differential Evolution Adaptive Metropolis with sampling from past states [DREAM((zs))] method, Bayesian Updating with Structural reliability methods using Subset Simulation (BUS + SS), and modified BUS with Subset Simulation (mBUS + SS). The differences between the performances (i.e. convergences, computational accuracies, and efficiencies) of these three methods are not well understood. This study systematically investigates the differences of these three methods in the generation of random samples, convergence criterion, model evidence, and estimation of posterior probability of failure in slope reliability updating. Two slope examples are used for the comparative study. It is found that the BUS + SS method performs well not only in the low-dimensional Bayesian inverse problems but also in the high-dimensional Bayesian inverse problems of spatially varying soil parameters. The DREAM((zs)) method is preferentially recommended to deal with the low-dimensional Bayesian inverse problems whereas the mBUS + SS method can well tackle the high-dimensional problems.

Automated summary

This study systematically investigates the differences of three Bayesian methods in generating random samples, convergence criterion, model evidence, and estimation of posterior probability of failure in slope reliability updating. It is found that BUS + SS method performs well in both low-dimensional and high-dimensional problems of spatially varying soil parameters, while DREAM((zs)) method is recommended for low-dimensional problems and mBUS + SS method is suitable for high-dimensional problems.