Gaussian Process Regression–Based Model Error Diagnosis and Quantification Using Experimental Data of Prestressed Concrete Beams in Shear

Abstract

To facilitate considering model uncertainty for rigorous reliability/probabilistic analysis, this paper proposed a Gaussian process regression–based (GPR-based) model error quantification framework and applied to shear capacity prediction of prestressed concrete (PC) beams. Firstly, the model error of shear capacity models from five well-received concrete structure and bridge design codes were diagnosed based on a compiled experimental database, where systematic correlations between model error and model parameters were observed. To consider the systematic correlation, model error was then calibrated as a function of model parameters based on GPR. Different covariance functions were considered, and a model selection was conducted based on 10-fold cross validations. Then, the model error quantification performance was evaluated by investigating the residual systematic correlation between model error and model parameters, as well as by comparisons with the traditional professional factor approach. In the end, relative importance of model parameters on the model error for each design code were analyzed, indicating that the shear span-to-effective depth ratio is the most important source of model error for all considered design code models.