TransModel-BSDL: Bayesian Sparse Dictionary Learning for Development of Site-Specific Transformation Models Using Limited Data and a Database of Existing Models

Abstract

Determination of soil design parameters at a specific site is a long-lasting challenge in geotechnical practice, because design parameters are generally not measured directly but estimated from limited measured data using correlations, or transformation models, between the design parameters of interest and measured data. The transformation models are often highly uncertain, and it is difficult to select an appropriate one with good fidelity among many existing correlations for a specific site or project. To address this challenge, a Bayesian sparse dictionary learning (BSDL) framework was recently proposed to construct site-specific models by leveraging existing transformation models and limited site-specific data. To facilitate development and application of the BSDL method, this study compiles a comprehensive database of transformation models, with more than 1,000 existing models for more than 18 design parameters. A software package, called TransModel-BSDL, is also developed to implement BSDL for construction of site-specific models with improved model fidelity and enhanced interpretability. This will remove the hurdles of sophisticated algorithms from engineering practitioners and enable them to apply the BSDL method to a wide variety of applications, without a need of searching for many transformation models from literature. TransModel-BSDL is demonstrated and validated using real data obtained from different sites.