Yongmin Cai, Kok-Kwang Phoon
Conventional methods for constructing quasi-local transformation models can introduce significant uncertainty in soil property inference, particularly for large-scale geotechnical projects. To address this, a novel data-driven site demarcation (DDSD) framework is proposed. The process begins by defining a target site—a local area within the large-scale project containing the soil property of interest. The large-scale project (excluding the target site) is then demarcated into uniformly sized elemental sites, analogous to “Lego bricks”. These elemental sites are assembled into database sites to form a Big Indirect Database (BID). A tailored Lego clustering algorithm is then developed to retrieve database sites from the BID that are highly similar to the target site. It is termed tailored Lego clustering because the cluster is tailored to the target site and the assembled database sites can vary in size, shape, and/or depth segments depending on how the “Lego bricks” are combined. This study focuses on a specific implementation where the assembled sites are restricted to squares or rectangles, and we optimize their sizes. A quasi-local transformation model is then constructed using these similar database sites that are of different sizes, enabling more accurate inference of soil properties at the target site. The core of this paper is the optimization of the sizes of these square or rectangular database sites (i.e., the assembly configuration of elemental sites) to minimize inference uncertainty, guided by a criterion based on the root mean squared error. The framework is validated through synthetic examples and a real-world case study, demonstrating that our data-driven size optimization for regular-shaped sites significantly reduces inference uncertainty.