Bayesian Network Prediction of Stiffness and Shear Strength of Sand

Abstract

This paper proposes a Bayesian network approach to predict the shear modulus and maximum friction angle of sand. The nonlinear correlations between sand parameters can be incorporated in the probability distribution represented by a Bayesian network. Extensive databases for shear modulus and friction angles of sandy soils are compiled for training the Bayesian network through maximizing the log-likelihood. The trained Bayesian network is applied to a case study in Japan (Yodo River sand). Information from multiple sources (index properties, in situ samples, and modulus logging) can be integrated in a holistic manner to decrease the uncertainty in the prediction of stiffness and shear strength. A Bayesian network also allows the calibration of the global model (model trained from a large global database) by including site-specific samples. In the Yodo River sand case, it is revealed that one to three samples are adequate to reduce the uncertainty of the global model close to the uncertainty of the site-specific model.