Machine learning driven seismic performance limit state identification for performance-based seismic design of bridge piers

Abstract

Performance-based earthquake engineering requires the identification of different damage states in structural components. Bridge piers are one of the most critical components in the bridge system that dictate the overall performance of bridges under seismic loading. The ability to predict different damage states of a bridge pier following an earthquake can be useful in restoring service or prescribing appropriate remediation. Machine learning techniques are becoming attractive in earthquake engineering for predicting failure modes of buildings and bridges. This study implements several machine learning techniques to capture different performance limit states such as spalling, core crushing, and bar buckling of reinforced concrete bridge piers under seismic loading. Using experimental data from the large scale testing of bridge piers, this study compares different machine learning techniques for predicting drift capacity performance limit states of circular and rectangular bridge piers. Model tuning using grid search algorithm and 5-fold crossvalidation is performed to optimize model predictions over a wide tuning parameter search space. The efficiency of different methods is compared using a reserved 20% testing dataset to determine the optimal machine learning regression method for each limit state. This study compares the respective bar buckling and concrete spalling drift limit state models against existing physics-based empirical models, where the machine learning models are found to have significantly improved accuracy and precision.