Full long-term buffeting analysis of suspension bridges using Gaussian process surrogate modelling and importance sampling Monte Carlo simulations

•Reliability based design of long-span bridge under wind loading.•A framework of surrogate based importance sampling monte carlo simulation method is developed.•Design buffeting response of a long-span bridge in design and feasibility phase was estimated with the proposed framework for the first time.•The proposed framework used only 1% of the computational effort of the traditional full long-term analysis.•The extreme buffeting response was 25% higher than that predicted by the current design guidelines. Recent findings from full-scale measurements campaigns and analytical investigations of the design buffeting response of long-span bridges suggest that the assumptions adopted in most wind-resistant design guidelines are not strictly conservative. In such cases, a full long-term analysis is the most accurate alternative for reliability-based design. However, the application of such methodology becomes unfeasible due to the corresponding computational demand. Notably, many evaluations of the buffeting response are required, and time-consuming numerical integration is traditionally used to evaluate the long-term response. To overcome these drawbacks, this paper proposes a framework to increase the computational efficacy of long-term analyses for the wind-resistant design of long-span bridges by combining two strategies. First, the buffeting response is estimated with a Gaussian process regression that requires less time than the traditional multimodal buffeting response estimation. Then, long-term analysis is carried out using importance sampling Monte Carlo simulations that converge faster than the traditional analysis based on numerical integration. The computational framework is demonstrated in a case study of a proposed super-long suspension bridge subjected to loads induced by wind buffeting. The advantage of the proposed framework is verified, as it requires less than 1% of the computational demand of the traditional full long-term analysis.