Data-centric quasi-site-specific prediction for compressibility of clays

A generic clay database consisting of six parameters, including compression index (Cc ) and unloading–reloading index (Cur ), is compiled from 429 studies. This database, labeled as CLAY-Cc /6/6203, contains 6203 records. A data-driven approach of predicting Cc and Cur for a target site by combining sparse site-specific data with CLAY-Cc /6/6203 is illustrated. This data-driven approach consists of two steps. The first step is a learning step that adopts a hierarchical Bayesian model (HBM) to learn the prior information in CLAY-Cc /6/6203 (both inter-site and intra-site variabilities). The second step is a Bayesian inference step that updates the prior model into a posterior model. The inference outcome is a quasi-site-specific model. A real case study (Baytown, Texas, USA) is adopted to illustrate the application of the HBM-MUSIC-3X method in estimating and simulating the 3D spatially varying Cc and Cur profiles. The key conclusions are as follows: (i) predictions from Big Indirect Data (BID) in the form of CLAY-Cc /6/6203 can be biased with large transformation uncertainty although data are abundant, (ii) predictions from small (sparse) site-specific data are less biased but suffer from high statistical uncertainty although data are directly applicable, and (iii) combining BID and site-specific data using an HBM learning strategy that accounts for site uniqueness is effective in terms of reducing prediction uncertainty.