A big indirect data – Informed probabilistic method for three-dimensional site reconstruction

Abstract

Three-dimensional (3D) reconstruction of a sparse measurement site is of paramount significance for the safety assessments or designing of the geotechnical structures. However, this task is often challenging because the site investigation data generally are sparse due to the limit budget, leading to large statistical uncertainties in the soil parameters. The challenge is further exacerbated by computational issues such as inversion or decomposition of the large correlation matrix, which frequently arises when dealing with large-scale 3D sites. To address these challenges, this paper proposes a novel big indirect data-informed three-dimensional site reconstruction method using hybrid Bayesian theory. The proposed method first constructs the probability distribution functions (PDFs) of the soil parameters of the big indirect data collected from worldwide historical sites and the soil parameters of the targeted site using the Gibbs sampler. The two PDFs are then integrated to form a hybrid PDF of the target site. Based on the hybrid PDF, the three-dimensional site is reconstructed with consideration of spatial variabilities of the soil parameters using multiple multivariate conditional random fields. The Kronecker product is utilized to decompose the large autocorrelation matrix into several small matrices that can be easily handled. A virtual site and a real site in Huizhou, China are employed to demonstrate the capability of the proposed method. The results show that the proposed method can effectively reduce the statistical uncertainty of soil parameters caused by sparse measurement. It offers a transformative tool that utilizes generic geotechnical big indirect data to supplement sparse local data, enabling effective 3D site construction.