Use of Machine Learning Algorithms to Predict the Response of High-Rise Building for Probabilistic Loss Assessment

In seismically active areas, stakeholders, (re)insurance industry, local or nation-wide governments desire to assess the performance of owned structures in terms of Death, monetary losses (Dollars), or Downtime (3D losses). To assess the seismic performance of existing structures various assessment procedures have been proposed and adopted by the earthquake engineering community. The vast part of those procedures comprises the simulation of the responses of the structure under the seismic excitations. These simulations are handled by means of computationally expensive methods (such as nonlinear response history analysis, NRHA). To overcome this matter, statistical models to obtain the response of the structure have been proposed and adopted. Additional to the conventional statistical models, more complex models have been proposed over the last decade. With its increasing ease of use machine learning techniques also have been the ones started to be considered. In the present study, we have adopted two main types of ML algorithms, classification, and regression namely. The results of the classification algorithms have been obtained as "the probability of being in damage state", which later manipulated to be fragility curves. On the other hand, maximum inter-story drift ratio (mIDR) values have been obtained via regression algorithms, which later used to obtain the fragility curves via cloud analysis. The comparison of the abilities of the ML algorithms has been done via using the fragility functions as proxies. Additional to the comparison within the ML algorithms, the fragility curves obtained via ML algorithms, have been compared also with the fragility curves obtained via the cloud analysis (SAC-2000) approach. Additional to the comparison of the various ML algorithms' abilities, a sensitivity analysis has been conducted to obtain the effect of the ratio between the amount of the data point to obtain the statistical model and the amount of the data point to test the model. According to the results obtained, all adopted classification algorithms lead to more than %80 accuracies when the number of the training data points is equal to or more than the testing data point. Additionally, at least a 0.916 R2 score (when the number of a testing datapoint is equal to three times the training data point) has been obtained for the regression algorithms.