A Gaussian Process surrogate approach for analyzing parameter uncertainty in mechanics-based structural finite element models

A key aspect of performance assessment of structures is the quantification and propagation of uncertainties, from the estimation of hazards to possible losses. In particular, probabilistic structural analysis deals with aleatory and epistemic sources of uncertainty in nonlinear modeling. Materials and components in structural models are represented by uncertain parameters, which can be accounted for via probabilistic constitutive models. The variability at a local level is then propagated to the system level when the structural model is sampled, sometimes inducing great uncertainty in structural demands. However, probabilistic modeling of real structures via finite element (FE) models has been a challenge due to high computational costs. One avenue to reduce this cost and make probabilistic modeling viable in practice is to develop cost-effective surrogate models. In this paper, a Gaussian Process (GP) approach is proposed to study the composition of parameter-induced uncertainty in mechanics-based nonlinear FE structural model responses. The methodology is validated by evaluating common regression error metrics between the original FE models and their GP surrogates. Case studies of two structures are presented, including a five-story reinforced concrete (RC) building and a five-span RC highway bridge. Finally, the low computational cost of the surrogate models is leveraged to perform simulation-based global sensitivity analysis using Sobol indices to quantify parameter-induced uncertainty in structural responses.