Comparative study of different Bayesian sampling algorithms for soft soil settlement predictions

Bayesian inverse analysis is a powerful tool for predicting settlement of embankments on soft soils. Given the high dimensionality and nonlinearity involved in soft soil settlement predictions, the Bayesian inverse analysis problem is most effectively addressed by sampling methods. This study revisits four popular sampling methods, including the differential evolution adaptive metropolis algorithm (DREAM(zs)), the affine invariant Markov chain Monte Carlo ensemble sampler (emcee), the dynamic nested sampling (dynesty) and Bayesian updating with structural reliability approach (BUS). Each method is combined with a one-dimensional consolidation model based on the general simplified Hypothesis B method to perform Class C prediction of an embankment on soft soils in Australia. Computational accuracy, uncertainty quantification, efficiency and convergence are compared under varying data conditions and sequential recalibration scenarios. Results demonstrate that all methods show improved accuracy with increased monitoring duration. DREAM(zs) excels in direct Bayesian updating but deteriorates under sequential updating. BUS offers the best balance between prediction accuracy, computational efficiency, and sequential updating capability, making it particularly suitable for ongoing calibration in real-world projects. emcee demonstrates strong sampling efficiency with abundant data, while dynesty requires the highest computational cost, limiting practicality. These findings provide a framework for method selection based on project stage, updating strategy, and computational constraints.