Prediction of nonlinear dynamic responses and generation of seismic fragility curves for steel moment frames using boosting machine learning techniques

The main objective of this paper is to develop machine learning (ML) models for predicting the seismic responses of steel moment frames. For this purpose, four boosting ML techniques-gradient boosting, XGBoost, LightGBM, and CatBoost-were developed in Python. To create an inclusive dataset, 92,400 nonlinear time-history analyses were performed on 1,848 steel moment frames under 50 earthquakes using OpenSeesPy. Geometric configurations, structural properties, and ground motion intensity measures were considered as the inputs for the models. The outputs included maximum global drift ratio (MGDR), maximum interstory drift ratio (MIDR), base shear coefficient (BSC), and maximum floor acceleration (MFA). The study also investigated the effectiveness of the ML models in estimating fragility curves for an 8-story steel frame at different performance levels. Finally, a web application was developed to facilitate the estimation of the peak dynamic responses for steel moment frames. The results show that the LightGBM and CatBoost models demonstrate superior predictive performance, with coefficient of determinations (R2) higher than 0.925. Furthermore, the LightGBM models can estimate the fragility curves with minimal errors (e.g., the relative errors in the median values of the predicted curves are less than 10%).