Enhancing Seismic Safety Assessment Through Development of a Transparent and Adaptive Rapid Visual Screening Method Employing Artificial Intelligence Algorithms

Abstract

Seismic safety assessment of existing buildings is very important because their design and construction are made according to lower standards. The buildings designed with lower standards and without standards are susceptible to earthquake-induced damage. The vulnerability of existing buildings to seismic events has been vividly highlighted by recent earthquakes, such as the Türkiye–Syria earthquake on February 6, 2023, the Herat Afghanistan earthquake on October 11, 2023, and the Marrakesh-Safi Morocco earthquake on September 9, 2023. In the Turkey-Syria earthquake alone, over 50,000 people lost their lives [1], over 100,000 sustained injuries [2], and the economic toll amounted to approximately 110 million dollars [3]. Building damage from seismic events poses risks to lives and causes substantial financial losses, necessitating the determination of each building's fragility and the implementation of appropriate precautions before an impending devastating earthquake. Rapid Visual Screening (RVS) methods are employed for assessing building inventory, given the computational and cost constraints of in-depth vulnerability assessment methods. While conventional RVS methods are widely used and high efforts are given to enhance them, their reliability is limited for accurately assessing a building inventory [4–6]. Therefore, this study leverages post-earthquake building inspection data from the 2015 Gorkha, Nepal earthquake to develop a RVS method using artificial intelligence algorithms, encompassing fuzzy logic, machine learning, and neural networks. The integration of advanced feature engineering techniques introduces sophisticated parameters like fundamental structural period, spectral acceleration, and distance to the earthquake source, enhancing the RVS method's assessment capabilities across diverse seismically vulnerable areas. The developed RVS method demonstrates a correlation between observed building post-earthquake damage states and the predicted ones. When compared to conventional RVS methods, a noteworthy test accuracy of 44% is achieved, surpassing conventional methods in accurately classifying building damage states. Notably, in contrast to RVS methods solely developed using machine learning and neural networks, the developed method exhibits transparency and the capability to be adapted to different regions. Keywords: Seismic vulnerability assessment; Earthquake-induced damage; Rapid Visual Screening (RVS); Artificial intelligence algorithms; Fuzzy logic; Machine learning; Neural networks