Quasi-site-specific soil property prediction using a cluster-based hierarchical Bayesian model

Abstract

One of the important goals in geotechnical engineering is to learn the correlation relationships between different soil properties in order to make predictions for a new site of interest. Recent development of large soil property databases has opened up new opportunities to a data-driven approach for soil property prediction. However, existing data is highly uncertain because of variations of measurement accuracy across different measurement devices and methods, differences of soil composition and loading history across different sites, and so on. A hierarchical Bayesian model (HBM) has been proposed to capture the inter- and intra-site variabilities, showing significant prediction improvement to models that consider data from various sites as a single dataset. In this study, we demonstrate the failure of the existing HBM algorithm to represent distribution of soil properties that contains multiple modes. Such limitation is an artifact of the HBM algorithm to achieve high computational efficiency. We propose a cluster-based HBM framework that balance the representation power and computational efficiency with discussion on the future direction of a fully data-driven approach for soil property prediction.