Developing a machine learning-based rapid visual screening method for seismic assessment of existing buildings on a case study data from the 2015 Gorkha, Nepal earthquake

Abstract Each existing building is required to be assessed before an impending severe earthquake utilizing Rapid Visual Screening (RVS) methods for its seismic safety since many buildings were constructed before seismic standards, without taking into account current regulations, and because they have a limited lifetime and safety based on how they were designed and maintained. Building damage brought on by earthquakes puts lives in danger and causes significant financial losses. Therefore, the fragility of each building needs to be determined and appropriate precautions need to be taken. RVS methods are used when assessing a large building stock since further in-depth vulnerability assessment methods are computationally expensive and costly to examine even one structure in a large building stock. RVS methods could be implemented in existing buildings in order to determine the damage potential that may occur during an impending earthquake and take necessary measures for decreasing the potential hazard. However, the reliability of conventional RVS methods is limited for accurately assessing large building stock. In this study, building inspection data acquired after the 2015 Gorkha, Nepal earthquake is used to train nine different machine learning algorithms (Decision Tree Classifier, Logistic Regression, Light Gradient Boosting Machine Classifier, eXtreme Gradient Boosting Classifier, Gradient Boosting Classifier, Random Forest Classifier, Support Vector Machines, K-Neighbors Classifier, and Cat Boost Classifier), which ultimately led to the development of a reliable RVS method. The post-earthquake building screening data was used to train, validate, and ultimately test the developed model. By incorporating advanced feature engineering techniques, highly sophisticated parameters were introduced into the developed RVS method. These parameters, including the distance to the earthquake source, fundamental structural period, and spectral acceleration, were integrated to enhance the assessment capabilities. This integration enabled the assessment of existing buildings in diverse seismically vulnerable areas. This study demonstrated a strong correlation between determining building damage states using the established RVS method and those observed after the earthquake. When comparing the developed method with the limited accuracy of conventional RVS methods reported in the literature, a test accuracy of 73% was achieved, surpassing conventional RVS methods by over 40% in accurately classifying building damage states. This emphasizes the importance of detailed data collection after an earthquake for the effective development of RVS methods.