UCLA

The traditional post-earthquake damage assessment of buildings is done visually through building-by-building inspections, which for a major event might take weeks to months to be completed. The main goal of this study is to establish a framework to quantify the post-event seismic performance of a portfolio of existing structures, in this case soft-story (SS) residential buildings, in near-real-time and an ultra-high resolution. Two main ingredients are required; 1) input ground motion time-series at the sites of the buildings that excite the structures, and 2) well-established nonlinear structural computer models representing the buildings. By having these two ingredients, nonlinear dynamic analyses of the existing structures can be performed, and seismic performance of the portfolio buildings can be quantified. The number of currently available recording instruments is sparse. Therefore, using the Gaussian Process Regression (GPR), a novel approach was developed to generate ground motion time series at un-instrumented target sites (i.e., the sites of the buildings) employing the surrounding recorded ground motions. The methodology has been validated for the physics-based scenario earthquakes in Northern California as well as two events recorded in Southern California: the 2019 M7.1 Ridgecrest and 2020 M4.5 South El Monte earthquakes. In addition, an approach was introduced to generate multiple realizations of ground motions at each target site. The results illustrated that the instrumentation density closer to the target site is crucial in producing accurate and reliable ground motions, especially at long periods. The second ingredient of the framework needs an accurate nonlinear computer structural model for soft-story buildings. There is an inventory of OpenSees models constructed to represent the 13,500 SS buildings in Los Angeles. This inventory includes thirty-two models categorized through four main key features: first-story wall layout, the number of stories, floor plan dimension, and wood-frame shear wall material. I have utilized machine learning methodologies to classify the first two features via visual recognition. I extracted images for 2,681 buildings within Los Angeles via Google Street View (GSV) service through an automated algorithm to train the Convolutional Neural Network (CNN) models for that purpose. The accuracy of the trained model is 80.1% and 91.4% for first-story wall layout and the number of stories classification tasks, respectively. I, then, utilized the OpenStreetMap (OSM) to detect the floor plan area of the target buildings to classify the third required feature. Having both a structural model and input ground motion time series, one can conduct a nonlinear response history analysis to assess the seismic performance of each specific building. As an application of the framework, I generated ground motion time series at over 2,000 SS buildings’ locations within Los Angeles for a scenario M6.7 earthquake based on the 2020 M4.5 El Monte rupture mechanism. I also estimated the structural features through the GSV extracted images to identify the “closest” available OpenSees model. Lastly, a map of the estimated damage state for the 2,000 SS buildings is produced. The results demonstrated that more than 50% of the SS buildings are either severely damaged or collapsed after such an earthquake in Los Angeles, while approximately 15% of them experienced slight damage. The methodology and framework developed in this research can be implemented to assess, in a near-real-time, seismic performance of a large portfolio of structures in moderate-to-severe earthquake events.