Machine learning-based prediction of residual drift and seismic risk assessment of steel moment-resisting frames considering soil-structure interaction

Abstract

Nowadays, due to improvements in seismic codes and computational devices, retrofitting buildings is an important topic, in which, permanent deformation of buildings, known as Residual Interstory Drift Ratio (RIDR), plays a crucial role. To provide an accurate yet reliable prediction model, 32 improved Machine Learning (ML) algorithms were considered using the Python software to investigate the best method for estimating Maximum Interstory Drift Ratio (IDRmax) and RIDR of 384 Steel Moment-Resisting Frames (SMRFs). In addition, the curve plot ability of methods was investigated to provide an estimation of Median of IDA curve (IDAMed) and Seismic Failure Probability curve (SFPCurve) considering Soil-Structure Interaction (SSI) effects. It is noteworthy that ML algorithms were improved with a pipeline-based hyper-parameters Fine-Tuning (FT) method followed by forward and backward feature selection methodologies to avoid overfitting and data leakage issues. The improved methods were evaluated to find the best prediction model regarding seismic demands. The results show that proposed methods have higher prediction accuracy and curve fitting ability (i.e. more than 95%) that can be used to estimate IDAMed and SFPCurve of a structure to accelerate the seismic risk assessment. A prediction tool is introduced to use the methods of this study for estimating abovementioned seismic demands.