Ensemble learning-based prediction of the backbone curve for corroded reinforced concrete columns using experimental database

Corrosion significantly degrades the seismic performance of reinforced concrete (RC) columns, yet assessment methods often rely on idealized backbone models or datasets with limited diversity, which reduces general applicability. This study develops a data-driven framework to predict the full-range backbone response of corroded RC columns under cyclic loading. An experimental database of 200 specimens is compiled from international peer-reviewed literature, covering broad ranges of geometry, reinforcement detailing, material properties, axial load level, corrosion severity, and failure mode. Several ensemble machine learning models are trained and tuned to predict yield, peak, and residual backbone strengths without imposing a predefined analytical backbone shape. Performance is evaluated using an independent test subset and cross-validation, and Shapley additive explanations (SHAP) analysis is used to quantify feature contributions for interpretability. Among the examined algorithms, the extreme gradient boosting model provides the strongest accuracy and robustness across backbone stages. The proposed framework captures post-peak softening and residual capacity with improved agreement relative to commonly used simplified backbone formulations. SHAP results indicate that geometric parameters and degraded reinforcement properties dominate predictions at yield and peak stages, while confinement- and damage-sensitive variables become more influential at the residual stage.