Artificial neural network application in predicting probabilistic seismic demands of bridge components

Abstract Probabilistic seismic demand models (PSDMs) of bridge components such as column and abutment are commonly developed through classical linear regression techniques in which a univariate model format is predefined in the logarithmically transformed space. The more advanced machine learning (ML)‐based PSDMs incorporate various sources of uncertainties, which eventually leads to a more credible prediction of the seismic demands of bridge components and enhances the vulnerability assessment of the overall bridge systems. Despite the emerging advancements in ML approaches, many of them have not yet been introduced to estimate bridge seismic responses. To this end, the present study seeks to develop predictive PSDMs using a reputable ML approach, the artificial neural network (ANN). Relative to the classical univariate PSDMs, the ANN‐based PSDMs improve the median estimation of demands, particularly over the large and small range of ground motion intensities and reduce the total prediction variability. Moreover, the proposed ANN‐based approach provides a generalizable model with an unbiased prediction of the seismic demands. The ANN‐based PSDMs can be further used in estimating the probability of structural damage in the fragility and risk assessment process.