Abstract Rapid evaluation of existing building stocks is essential for reducing earthquake-related disaster risk, particularly in seismically vulnerable cities such as Gaziantep, Türkiye. However, commonly used assessment approaches are either time-consuming or rely on simplified linear models that fail to capture the complex relationship between seismic demand and building characteristics. This study proposes a safety-first ensemble data-driven screening framework for rapid post-earthquake building damage assessment, combining XGBoost, LightGBM, and CatBoost classifiers via soft voting. The framework is trained and tested using a large post-earthquake dataset from the Nurdağı and Islahiye districts, obtained from the Ministry of Environment, Urbanization and Climate Change, which includes georeferenced building attributes and officially assigned damage states without secondary reinterpretation. Spatially varying seismic demand is represented using interpolated peak ground acceleration values for DD-1 and DD-2 hazard levels, while V S30 parameters characterize local site conditions. Design spectral acceleration values are derived in accordance with the seismic codes applicable at the time of construction. To enhance physical interpretability, the model incorporates engineering-informed feature engineering, including interaction terms linking building height, construction period, and seismic demand. Model optimization prioritizes life safety by emphasizing recall over overall accuracy. The proposed framework achieves a recall of 88.9% on independent test data and zero false-negative predictions for high-rise reinforced-concrete buildings (≥ 6 storeys). Additional robustness analyses demonstrate stable performance under moderate data uncertainty. The results indicate that the proposed approach provides a reliable, interpretable, and safety-oriented decision-support tool for rapid seismic risk screening of reinforced concrete building inventories following major earthquakes.
Pürsünlü et al. (Tue,) studied this question.