Efficient Bayesian model selection and calibration using field data for a reinforced concrete slab bridge

In this work, a reinforced concrete slab bridge (instrumented and tested in 2018) is investigated. Based on field data, a finite element model of the bridge is calibrated. Model selection is performed both based on log evidence and posterior predictive capabilities. It is investigated if the models selected based on the log evidence also induce the most accurate posterior predictions. The influence of different assumptions on modelling the spatial distribution of the stiffness and different possible suggestions on how to include prediction errors and model bias are investigated. Comparing the conclusions based on log evidence and posterior predictions, only using the log evidence for model selection could be debated. Models performing best when considering the log evidence led to the least accurate posterior predictions, and models rejected based on the log evidence could still have good predictive capabilities. Considering the different model classes, introducing spatial variation of the stiffness leads to a posterior prediction closer to the measurements. Introducing a global model bias leads to a better match between predictions and measurements compared to not including this model bias. Even better posterior predictions are achieved if this model bias is quantified locally for the different considered datapoints.