Constructing Site-Specific Multivariate Probability Distribution Model Using Bayesian Machine Learning

This study proposes a novel data-driven Bayesian machine learning method for constructing site-specific multivariate probability distribution models in geotechnical engineering. There is a trade-off for constructing a site-specific model: a model developed from generic data may not be fully applicable to a local site, but a model purely developed from limited site-specific data may be very imprecise due to significant statistical uncertainty. The proposed method is based on the hybridization between site-specific and generic data in the way that it is governed by site-specific data when site-specific data are abundant and by generic data when site-specific data are sparse. This method broadly follows how an engineer currently estimates design soil parameters from limited site-specific information. The proposed method admits incomplete multivariate data, so it can handle missing data that are commonly encountered in site investigation. It is a Bayesian method, so uncertainties are rigorously quantified. Actual case studies are used to demonstrate the usefulness of the proposed method. Analysis results show that the proposed method can effectively capture the correlation behaviors in site-specific data and, moreover, can make meaningful predictions even when site-specific data are very sparse.