Classification algorithms for collapse prediction of tall buildings and regional risk estimation utilizing SCEC CyberShake simulations

Abstract

Quantification of collapse risk of buildings in seismically active regions is one of the key elements for informed decision making for building design and establishment of public policies to promote seismic safety and resilience. This paper focuses on development, testing and application of efficient and reliable collapse classification algorithms using machine learning tools. To this end, a large database of structural responses is developed by performing around two million nonlinear time history analyses of an archetype 20-story tall building. Unscaled seismograms simulated for the Los Angeles region as part of the Southern California Earthquake Center (SCEC) CyberShake project are used as inputs for the analysis. Feature selection is performed using regularized logistic regression to identify intensity measures with strong predictive power for classification of collapse. Results of regularization generally confirm the understanding of important predictors as gained from scaling of recorded motions as well as highlight additional important features. Logistic regression and support vector machine (SVM) binary classifiers are then trained on the data to develop collapse prediction models. The resulting collapse assessment models achieve high values of precision and recall and show good performance when tested using benchmark collapse responses. Finally, trained collapse classifiers are utilized to perform regional estimation of collapse risk. Collapse predictions are made using CyberShake data from 336 sites across Southern California where there are around 500,000 simulated seismograms at each site. Regional estimation of mean annual frequency of collapse is performed to generate maps of collapse risk. Higher values of risk correlate well with geologic features such as presence of sedimentary basins and the surface trace of the San Andreas fault.