Physics-Informed Machine Learning Framework for Post-Earthquake Damage Classification of RC Buildings

This study presents a physics-informed machine learning framework for post-earthquake damage classification using building-level survey data collected after the 2023 Kahramanmaras earthquake sequence. To ensure structural homogeneity and avoid typology-driven biases common in prior studies, the analysis focuses exclusively on reinforced-concrete (RC) residential buildings. The feature set combines directly used parameters, such as shear-wave velocity to represent site conditions, with physics-informed engineered features, including standardized spectral accelerations, modified coordinates, and a combined energy metric. These features are designed to capture spatial shaking variability and structural vulnerability more effectively than raw survey attributes. Several supervised models spanning tree-ensemble, kernel-based, and neural architectures are trained and evaluated using a nested cross-validation scheme that yields unbiased generalization estimates and ensures that learning-based steps such as scaling and SMOTE are fitted only on the training folds and then applied to the validation/test folds, preventing information leakage. Severe class imbalance across Slight, Moderate, and Heavy damage states is addressed through Borderline-SMOTE oversampling and class-weighted learning. Model performance is evaluated using class-balanced metrics, including macro-F1, balanced accuracy, and G-Mean, to ensure fair assessment under class imbalance. In addition, SHAP-based analyses are conducted to offer a transparent interpretation of feature contributions. The results show that physics-informed feature engineering, combined with systematic validation and imbalance handling, substantially improves the reliability of post-earthquake damage classification for RC residential buildings.