Three-dimensional probabilistic site characterizations using multi-outputs sparse Bayesian learning

Three-dimensional probabilistic site characterization is the cornerstone of geotechnical digital transformation, because all engineering projects require an accurate understanding of subsurface geotechnical properties. Soil laboratory testing data or in-situ testing records are often used for data-driven site characterization. However, these site investigation data are often multivariate, uncertain, sparse, and spatially varying. In this paper, the existing sparse Bayesian learning method for three-dimensional (3D) probabilistic site characterization is extended to incorporate multiple soil properties, considering both the three-dimensional spatial variability and the cross-correlation among different soil properties. The proposed three-dimensional multiple-outputs sparse Bayesian learning (3D-MSBL) method is also capable of simulating multiple-correlated conditional random fields in 3D, with the benefit to quantify the statistical uncertainties of soil properties at unexplored locations. The proposed 3D-MSBL method is examined on three case studies. It is shown that the proposed method outperforms the existing single-output SBL method in geotechnical data-driven site characterization, giving more accurate predictions accuracy of soil properties at unexplored locations with smaller statistical uncertainties especially for sparse training data scenario.