Efficient intensity measures and machine learning algorithms for collapse prediction of tall buildings informed by SCEC CyberShake ground motion simulations

Abstract

In contrast to approaches based on scaling of recorded seismograms, using extensive inventories of numerically simulated earthquakes avoids the need for any selection and scaling of motions which implicitly requires assumptions on intensity measures (IMs) that correlate with structural response. This study has the objectives to examine seismogram features that control the collapse response of tall buildings and to develop efficient and reliable collapse classification algorithms. To that end, machine learning techniques are applied to the results of nonlinear response history analyses of a 20-story tall building performed using about two million simulated ground motions. Feature selection of ground motion IMs generally confirms current understanding of collapse predictors based on previous studies using scaled recorded motions. In addition, interrogations of the large collection of hazard-consistent simulations demonstrate the utility of different IMs for collapse risk assessment in a way that is not possible with recorded motions. Finally, a small subset of IMs is identified and used in development of an efficient collapse classification algorithm which is tested on benchmark simulated data at several sites in the Los Angeles basin.