Data-driven prediction of residual flexural capacity in corroded RC beams using PSO and GA-optimized CatBoost ensemble models

Abstract Reinforced concrete (RC) beams inevitably experience steel corrosion when exposed to chloride ingress or carbonation, leading to progressive deterioration of both durability and structural capacity. This corrosion-induced degradation poses critical challenges to structural safety while substantially increasing life-cycle maintenance costs. A machine learning framework integrating CatBoost algorithm with metaheuristic optimization was developed to predict residual flexural capacity of corroded RC beams. An experimental database encompassing 543 test specimens with 12 critical parameters (including geometric dimensions, material properties, and corrosion characteristics) was established. Three hybrid models (BO-CatBoost, GA-CatBoost, PSO-CatBoost) were subsequently developed through hyperparameter optimization using Bayesian optimization (BO), genetic algorithm (GA), and particle swarm optimization (PSO). Quantitative evaluations demonstrated the superior predictive accuracy of metaheuristic-optimized models, with PSO-CatBoost emerging as the top performer (testing R 2 = 0.972, RMSE = 3.4183). This represents a 35.9% reduction in RMSE compared to the baseline CatBoost. The GA-CatBoost variant also showed significant improvements (testing R 2 = 0.970, RMSE = 3.6285), outperforming both baseline CatBoost and BO-CatBoost. The marked superiority of PSO and GA algorithms underscores their enhanced capability in navigating complex hyperparameter spaces, effectively capturing the nonlinear relationships between corrosion degradation and structural response. Sensitivity analysis revealed that beam height and reinforcement ratio positively correlate with load-bearing capacity, whereas rebar mass loss ratio and water-to-binder ratio exhibit significant negative impacts. The proposed framework provides a robust assessment tool for corrosion-damaged RC members while identifying critical degradation mechanisms, enabling more informed maintenance decisions for aging concrete infrastructure.