Incorporation of machine learning into multiple stripe seismic fragility analysis of reinforced concrete wall structures

Abstract

This study proposes a novel procedure that incorporates machine learning (ML) into the multiple stripe analysis (MSA) approach to efficiently produce seismic fragility curves for reinforced concrete (R/C) shear walls in building frame systems. The proposed procedure aims to mitigate computational challenges associated with the original MSA approach. In this context, ML models were developed for predicting the failure probability of R/C walls subjected to ground motions based on a specified threshold of maximum interstory drift ratio (MIDR). Specifically, the result of each numerical analysis, taken as the output variable of the ML models, was classified as either "B" (Below) or "E" (Exceeding) to indicate whether the MIDR of R/C walls was below or exceeding the specified threshold. This binary categorization was then used to calculate the failure probability points, which are necessary to derive fragility curves per the MSA approach. Data for training and testing the ML models were generated from nonlinear time history analyses of 46 distinct R/C walls subjected to 1000 ground motions. The R/C walls varied in height from four to 40 stories, and the ground motions included far-field, near-field pulse, and near-field no-pulse types. Four well-established ML methods, including random forest (RF), extreme gradient boosting, light gradient boosting machine, and categorical boosting, were considered. The performances of the ML models were compared using a confusion matrix. Based on this comparison, the RF model was selected and incorporated into the proposed procedure. Subsequently, the proposed approach was demonstrated to create the seismic fragility function of a new R/C wall structure. This study highlights the potential of ML applications in optimization problems within the earthquake engineering domain.