Residual deformation-based post-earthquake damage identification and residual capacity quantification of reinforced concrete columns using machine learning technology

Abstract

Traditional assessments of the post-earthquake seismic performance of reinforced concrete columns (RCCs) primarily rely on qualitative observations, making it challenging to quantify damage severity and residual seismic capacity (RSC). This study proposes a machine learning (ML)-based approach that utilizes measurable residual deformation—a parameter closely correlated with damage extent—to quantitatively assess seismic damage and the RSC of RCCs after earthquakes. Drawing on a comprehensive experimental database of RCC hysteresis curves, two predictive models are developed: one for estimating the deformation capacity of RCCs under cyclic loading, and another for predicting displacement-force hysteresis response encompassing the elastic, plastic, and failure phases. Residual displacement and peak deformation data under various earthquake loading cases are extracted from the experimental database, and ML techniques are employed to capture their nonlinear relationship. The predicted peak deformations are subsequently fed into the hysteresis model to reconstruct the loading history and estimate the RSC of RCCs. The proposed method is validated using cyclic loading tests of five full-scale RCC specimens. Results indicate that models based on XGBoost and LSTM architectures attain high predictive accuracy (R² ≥ 0.94), effectively identifying seismic damage states and quantifying post-earthquake RSC. This data-driven approach offers a rapid post-earthquake assessment tool that can support structural engineers in enhancing seismic resilience and formulating recovery strategies.