Abstract Summary
Full long-term analysis is the most accurate method to determine extreme buffeting responses in long-span bridges since it considers the variability of the wind turbulence and the variability in the short-term buffeting response which traditionally are neglected. Several authors have reported significantly higher responses using the long-term approach compared to the traditional approach. Hence, it is suggested to revisit the methods widely accepted in most of the design codes. Recently, the National Public Road Administration (NPRA), adopted this probabilistic approach by including the full long-term analysis in the latest version of the handbook for bridge design, N400. Nevertheless, a more elaborate regulation is required in the future to improve the guidelines. The principal weakness of the method lies in its relatively high computational demand. Methods have been proposed to increase the computational speed, but they are regularly based on inverse reliability methods which tend to give only approximate solutions. On the other hand, using surrogate modeling approaches based on machine learning algorithms offer a reasonable alternative to reduce the computational demand of the full long-term method. The approach is simple, a surrogate model of the short-term buffeting response is trained with some few but strategically selected wind conditions. Then, the full long-term analysis is executed using fast estimations from the surrogate model instead of making a time-consuming complete structural analysis. This paper contains an overview of the extreme buffeting response of the Sulafjord bridge, a 3200m long-span single suspension bridge located in western Norway. This paper examines a framework for the full long-term analysis with a surrogate model based on Gaussian Process Regression. This type of algorithm uses a Bayesian updating strategy that optimizes the amount of training simulations, thus enhancing the computational speed of the analysis. The results from the examined framework were compared with the exact solution from the full long-term analysis. The comparison showed that the proposed framework has a high degree of accuracy, with 2.1% error and yet, using just less than 1% of the original computational demand. The simplicity and accuracy of this framework offers an attractive pathway to set the tone for further steps of implementation in the industry and contribute significantly to the future regulation.
