Abstract - Atici

DEVELOPMENT OF A SEISMIC DAMAGE PREDICTION MODEL BY USING MACHINE LEARNING CLASSIFICATION ALGORITHMS WITH AN ARTIFICIAL DATASET

Ali Talha Atici

(Thesis Supervisor: Doç. Dr. Ufuk Hancılar)

ABSTRACT

Assessing the potential damage to buildings due to a possible earthquake in a region and taking measures, such as strengthening or reconstruction of vulnerable structures, is critically important to minimize social and economic losses that are likely to occur. Evaluating the seismic performance of structures is a comprehensive and time-consuming process. However, using well-trained machine learning prediction models instead of traditional structural performance analyses can significantly reduce computation time. This thesis focuses on developing a damage prediction model using classification-based machine learning algorithms, utilizing a two-dimensional reinforced concrete frame system dataset that represents low to mid-rise, non-ductile buildings. The structural features forming the dataset are obtained from a comprehensive literature review on building stock characteristics in the Marmara region. Nonlinear time history analyses are conducted using actual earthquake records with the OpenSeesPy framework. The maximum inter-story drift ratio is used as an engineering demand parameter to classify the damage state of buildings. Reliable machine learning models are developed with a balanced dataset. Twenty-four models are created using six variant ground motion intensity measures and four classification algorithms: k-nearest Neighbors, Support Vector Machine, Decision Tree, and Random Forest. The best-performing model is determined by comparing performance metrics and the confusion matrix. In conclusion, the model developed with a dataset incorporating peak ground velocity and utilizing the Random Forest classification algorithm demonstrates the most effective performance with 92% prediction accuracy.