Abstract - Ülkü

THE USE OF MACHINE LEARNING ALGORITHMS TO DERIVE FRAGILITY CURVES FOR MID-RISE REINFORCED CONCRETE BUILDINGS

Onur Ülkü

(Thesis Supervisor: Doç. Dr. Ufuk Hancılar)

ABSTRACT

The occurrence of an earthquake does not necessarily indicate that there is a seismic risk. The existence of risk depends on having three components, which are hazards, exposures, and fragility together. Assessing the risk of existing buildings is a building-specific task that may be tremendously time-consuming and computationally burdensome. Moreover, determining the risk of each structure can be complicated when investigating a portfolio or a group of buildings. Developing fragility curves for buildings provides a possible and undemanding method to estimate damage likelihood. Machine learning algorithms are one of the novel approaches that are implemented for estimating potential structural damage. Well-trained machine learning algorithms provide to speed up processing, cut down the cost of computation, and produce reliable fragility curves. This thesis focuses on developing fragility curves for generic building inventory spread over the Marmara region by predicting the probability of maximum inter-story drift ratio intervals via five various machine learning algorithms, which are Random Forest, Stochastic Gradient Boosting, Naïve Bayes, Decision Tree, K-Nearest Neighbors. Information on the characteristics of pre-dominant building typologies in the Marmara region available in the literature is utilized for creating an artificial inventory dataset of mid-rise RC buildings. The data for the machine learning was gathered by designing and analyzing 2-D frame systems under non-linear time history analysis with OpenSeesPy. The machine learning algorithms are trained considering different intensity measures and buildings' characteristics. Machine learning algorithms are evaluated when generating fragility functions by comparing those produced by fitting log-normal distribution with the maximum likelihood estimation method.