Efficient uncertainty quantification in structural dynamic analysis using two-level Gaussian processes

Uncertainty quantification is an important aspect in structural dynamic analysis. Since practical structures are complex and oftentimes need to be characterized by large-scale finite element models, component mode synthesis (CMS) method is widely adopted for order-reduced modeling. Even with the model order-reduction, the computational cost for uncertainty quantification can still be prohibitive. In this research, we utilize a two-level Gaussian process emulation to achieve rapid sampling and response prediction under uncertainty, in which the low-and high-fidelity data extracted from CMS and full-scale finite element model are incorporated in an integral manner. The possible bias of low-fidelity data is then corrected through high-fidelity data. For the purpose of reducing the emulation runs, we further employ Bayesian inference approach to calibrate the order-reduced model in a probabilistic manner conditioned on multiple predicted response distributions of concern. Case studies are carried out to validate the effectiveness of proposed methodology.