Evaluating the Impact of Spectrum-Matched Time Histories on Biased FRS Estimates: Part 1-Sensitivity Analysis Using Machine Learning

ABSTRACT In the wake of the Fukushima nuclear disaster, assessing the seismic safety of nuclear facilities, especially the secondary system, has become crucial. Floor response spectra (FRS) are keys to this assessment, typically generated through time history analyses. However, prior research shows that different time histories matching the same target spectrum can produce highly variable FRS estimates. This study examines the impact of spectrum-matched time histories on biased FRS estimates using machine learning methods. We generated 1750 spectrum-matched time histories based on five distinct target response spectra and applied them to eight typical steel moment-resisting frame buildings to compute their FRS. Machine learning techniques were then used to analyze the sensitivity of FRS estimates to modifications in ground-motion parameters due to spectral matching. The results indicate that changes in spectral accelerations are the primary factor influencing biased FRS estimates, followed by the damping scaling factor index for damping ratios of 7%, 2%, 10%, and 1%. These findings will aid in generating more accurate spectrum-matched time histories, reducing the risk of biased FRS estimates and improving seismic safety assessments for nuclear facilities.