Seismic risk assessment of typical bridges in Northeastern Brazil through the application of Machine Learning methods

Abstract Bridges are the most vulnerable components in a transportation system. Therefore, extraordinary events such as seismic activities should also be considered in bridge design. In Brazil, since the last few decades, representative earthquakes with magnitudes above 5 have been registered by the seismological center of the University of São Paulo (USP). The newest edition of the Brazilian regulation for seismic-resistant structure design, NBR 15421:2023, does not apply to special structures, such as bridges. It is not clear whether the existing infrastructure can resist seismic demands. This work aims to evaluate the vulnerability of a reinforced concrete bridge with typical geometry located in the northeast of Brazil, a region known for relatively high seismic activity in Brazilian territory. The structural system is composed of a “T” section deck supported by non-integral U-type or gravity abutments and by frames with two circular columns through elastomeric bearings. The seismic records were collected from the Pacific Earthquake Engineering Research Center, which is compatible with the target spectrum of the city of Natal/RN. The study of vulnerability is conducted through the assembly of fragility curves, which provide the conditional probability of the demand (obtained by nonlinear dynamic analysis) exceeding the structural capacity for different structural damage limit states. Traditional approaches to evaluating fragility curves are time-consuming, mainly because of the relatively high computational cost of running several finite element-based models. Here, we present a computationally efficient scheme based on Machine Learning techniques to evaluate the application of artificial intelligence tools, such as Support Vector Machine, Decision Tree and Random Forest, for seismic vulnerability for structures situated in the northeastern region of Brazil. The fragility curves obtained by traditional approach and Machine Learning techniques were very similar.