Predicting seismic response of SMRFs founded on different soil types using machine learning techniques

Abstract

Predicting the Maximum Interstory Drift Ratio (M-IDR) of Steel Moment-Resisting Frames (SMRFs) is a useful tool for designers to approximately evaluate the vulnerability of SMRFs. This study aims to explore supervised Machine Learning (ML) algorithms to build a surrogate prediction model for SMRFs to reduce the need for complex modeling. For this purpose, twenty well-known ML algorithms implemented in Python software are trained based on the dataset achieved from Incremental Dynamic Analyses (IDAs) performed on the 2-Story, 3-Story, 4-Story, 5-Story, 6-Story, 7-Story, 8-Story, and 9-Story SMRFs modeled in Opensees. Then, important features of weight, fundamental period of structure (T1), the RSN number of record subsets, the direction of the record subsets, soil types, and Sa(T1) of analysis, which affect the results of predictions, were selected by trial and error. Having these important features, data-driven techniques developed in Python software were compared for predicting the M-IDR of SMRFs as target. Results showed that ML algorithms of GPReg, PLSReg, SReg, LReg, GReg, MLPReg, SVM, and LSVR had lower values of coefficient of determination (R2 lower than 0.5) in both train and test datasets. In addition, XGBoost, BReg, HistGBR, and ERTReg algorithms achieved higher values of R2 (i.e. upper than 0.95 in the 5-Story SMRF) with low Mean Squared Error (MSE) for prediction of M-IDR. Therefore, using these algorithms mitigates the need for computationally expensive, time-consuming, and complex analysis, while preliminary prediction of M-IDR can be considered a low computational and efficient tool for seismic vulnerability assessment.