Shear capacity prediction and reliability analysis of corroded reinforced concrete beams using deep generative modeling and ensemble learning

This study develops a machine learning framework to predict the shear capacity of corroded reinforced concrete (CRC) beams, enhancing structural reliability assessments. A Variational Autoencoder (VAE) generated a synthetic dataset of 10,000 samples, addressing the challenges of limited and varied experimental data on corrosion. Comparative analyses showed the VAE outperformed Generative Adversarial Networks (GANs) based on entropy, Kullback–Leibler divergence, and Fréchet Inception Distance (FID), indicating higher data quality and realism. Five machine learning models – Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Categorical Boosting (CatBoost), Adaptive Boosting (AdaBoost), and a back-propagation neural network (BPNN) – were trained using this data. XGBoost demonstrated superior accuracy, achieving an R² of 0.96 on synthetic data and 0.85 on real data. Shapley Additive Explanations (SHAP) identified critical factors such as concrete strength and stirrup corrosion, impacting shear capacity. A reliability-based approach calibrated a global resistance factor of 1.10, ensuring CRC beams designed with the XGBoost model meet a reliability index of 3.8. This approach significantly advances predictive capabilities and reliability assessments for aging infrastructure management.