Predictive models for the axial capacity of NSM- and hybrid FRP-strengthened RC columns using ensemble learning optimized by metaheuristic algorithms

The strengthening of reinforced concrete (RC) columns with fiber-reinforced polymers (FRP) has gained attention, particularly with near surface mounted (NSM) FRP bars and FRP jacketing. However, uncertainties in design codes, such as the effectiveness of fibers under compression and the complexity of bonding behavior, necessitate machine learning-based predictive models. This study develops a machine learning framework to predict the axial capacity of RC columns strengthened using NSM and hybrid techniques. A dataset comprising 112 experimental samples, and 112 synthetic samples generated using tabular generative adversarial networks (TGAN) was utilized. The hyperparameters of gradient boosting regressor (GBR) and random forest regressor (RFR) models were optimized via Bayesian optimization and metaheuristic algorithms, including arithmetic optimization algorithm (AOA), artificial hummingbird blgorithm (AHA), and slime mould algorithm (SMA). Comparing models trained exclusively on empirical data with those trained on the combined real–synthetic dataset demonstrated that TGAN-based augmentation significantly improves performance and reduces model variability. The proposed model outperforms existing code‐based approaches on all metrics, providing more accurate and reliable axial‐capacity predictions for NSM- and hybrid FRP-strengthened RC columns. SHapley Additive exPlanations (SHAP) analysis revealed that jacketing and NSM parameters positively affect capacity, while load eccentricity reduces it. Concrete compressive strength negatively affected strengthening efficiency, particularly beyond 38 MPa. Among optimization methods, AHA-GBR showed the best performance, followed by AOA-GBR and SMA-GBR. While Bayesian optimization was faster, it exhibited higher performance variance. The best-selected model outperformed traditional Bayesian methods, reducing RMSE on the test dataset by over 20%.