Machine learning strategy for predicting flutter performance of streamlined box girders

Abstract Engineers often heavily rely on wind tunnel tests or computational fluid dynamics (CFD) to evaluate the flutter performance of bridges in their preliminary design, which is costly and time-consuming. To quickly obtain the critical flutter wind speed of streamlined box girders in the preliminary design, a machine learning (ML) strategy was proposed in this paper. A big dataset was built by testing critical flutter wind speeds of 30 sectional models of streamlined box girders with and without railings at 5 angles of attack through free vibration wind tunnel tests. The flutter predicting models, taking geometric information and dynamic parameters as inputs, were built based on four widely-used ML algorithms, i.e., support vector regression (SVR), neural network (NN), random forest regression (RFR), and gradient boosting regression tree (GBRT). It is shown that the NN and GBRT models exhibit the highest prediction accuracy for the girders regardless of railings, respectively. A comparative study revealed that the ML models were superior over those simplified formulas for flutter estimation including the Van der Put formula, Selberg formula, and Haifan Xiang formula. A case study was also given to demonstrate the practical application of the proposed method. These ML models provide an efficient supplement to wind tunnel tests and CFD simulations for flutter predictions of streamlined box girders in the preliminary design.