Model calibration: A hierarchical Bayesian approach

Abstract

Many geotechnical engineering models are empirical and calibrated based on data gathered from various sites/projects, using optimisation algorithms with criteria like least squared errors or minimising the coefficient of variation of method bias with the constraint of mean bias equal to unity. This paper discusses the use of hierarchical Bayesian regression models for the same purpose. A database of axial capacity of piles in predominantly clay sites and a CPT-based design model, compiled and developed as part of a Joint Industry Project (JIP) led by the Norwegian Geotechnical Institute (NGI), is used for demonstration. The analyses focus on two related areas that the traditional approaches overlook: (i) quantification of uncertainty in the estimated parameters of the model, and (ii) modelling site-dependency of the model parameters (i.e., between-group variation). The former is important in the context of reliability-based design and contributes to establishing confidence in estimated reliability indices, particularly when only limited data are available. The latter expands our understanding regarding the domain of applicability of a model; that is, if a model is broadly applicable or highly site-dependent. The benefits of the proposed Bayesian approach are highlighted with a prediction exercise where the calibrated models are used in conjunction with limited site or project-specific data.