Bayesian Model Comparison and Characterization of Undrained Shear Strength

This paper develops Bayesian approaches for facilitating the determination of characteristic (or nominal) values of geomaterial properties in geotechnical analysis and design when extensive testing cannot be performed, which is the case for a majority of geotechnical projects, particularly those of a small or medium size. These Bayesian approaches aim to characterize probabilistically the undrained shear strength, Su, of clay using a limited amount of liquidity index (LI) test data, and to provide a logical route to determine the characteristic values for analysis and design, particularly those using probability-based design codes. The proposed approaches include (1) a Bayesian model comparison approach that selects the most appropriate likelihood model, a key element in the Bayesian framework, using a limited number of LI data obtained from a specific project site, and (2) a Bayesian equivalent sample approach that uses the selected likelihood model, integrates the sound engineering judgment/local experience with the project-specific LI data, and transforms the integrated knowledge into a large number of equivalent samples using Markov chain Monte Carlo simulation. Conventional statistical analysis of the equivalent samples is subsequently performed to determine characteristic values of the Su profile. The proposed approaches use engineering judgment/local experience in a quantifiable and transparent manner and effectively tackle the difficulty of generating meaningful statistics and probability distributions of soil properties from a usually limited number of test data obtained during geotechnical site investigation.