Bayesian Updating of Embankment Settlement on Soft Soils with Finite Element Method

Prediction of responses (e.g., embankment settlement) of geotechnical structures on soft soils is a challenging task due to their complex mechanical behaviors. In face with such complexity, the finite element method (FEM) combined with advanced soil constitutive models (e.g., soft soil creep (SSC) model) is frequently used to predict the short-term and long-term responses of geotechnical structures on soft soils, which involves a number of model parameters. Determination of these model parameters depends on knowledge obtained from site investigation data and/or monitoring information. This paper develops a Bayesian sequential updating (BSU) framework that incorporates monitoring information obtained at different construction stages to update FEM model parameters and their corresponding stochastic responses. To address the computational issues in Bayesian analysis, No-UTurn Sampler (NUTS) Markov chain Monte Carlo (MCMC) algorithm is introduced to populate posterior samples, and multiple Hermite response surfaces are constructed for different monitoring phases to reduce the computational efforts costed by evaluating the likelihood function. The proposed method is illustrated by a settlement prediction example of Ballina trial embankment, New South Wales, Australia. Effects of different likelihood functions (namely with and without model bias factor (MBF)) on Bayesian updating of settlement predictions are investigated. Results showed that the proposed BSU framework improves the prediction accuracy of soft soil settlement compared with prior predictions. NUTS is much more efficient in generating posterior samples compared with Metropolis-Hastings (MH) algorithm as the number of model parameters is relatively large. When considering short-term settlement behaviors of soft soils, the likelihood function without MBF is preferred because the adopted SSC can properly characterize short-term behaviors of soft soils. On the other hand, the likelihood function with MBF is recommended because SSC is hard to represent long-term behaviors of soft soils.This work was supported by the National Key R&D Program of China (Project No. 2016YFC0800200), the National Natural Science Foundation of China (Project Nos. 51679174, 51579190, 51528901), and Young Elite Scientists Sponsorship Program by CAST (Project No. 2017QNRC001). The financial support is gratefully acknowledged