Machine learning-based prediction and interpretability analysis of flexural capacity for CFRP-strengthened RC beams

Reinforced concrete (RC) beams in long-term service may experience reduced load-bearing capacity due to loading effects, environmental erosion, and material aging. Carbon fiber reinforced polymer (CFRP), with its high strength and corrosion resistance, has been widely used to strengthen RC beams. However, the flexural capacity of CFRP-strengthened RC beams is affected by the coupled influence of multiple factors. Accurately distinguishing the relationships between different parameters and flexural strength is a key scientific issue in optimizing CFRP strengthening design. To address this, six machine learning (ML) models were developed to predict the flexural capacity of 203 CFRP-strengthened RC beams using a dataset compiled from 30 published studies, covering 14 input features including material properties, geometric dimensions, and strengthening parameters. Among these models, CatBoost achieved the best performance with the highest R ² (0.953) and the lowest errors (MAE = 1.343, RMSE = 1.952, MAPE = 7.081) in the test set, and showed stable results in 10-fold cross-validation ( R ² ranging from 0.88 to 0.95). SHAP analysis identified reinforcement ratio ( ρ ), concrete compressive strength ( f c ), number of CFRP layers ( n ), CFRP width-to-beam width ratio ( ω / b ), and steel yield strength ( f y ) as key influencing factors. Furthermore, PDP analysis illustrated the marginal effects of these factors on flexural strength. For example, two CFRP layers with a thickness of 0.167 mm and ω / b > 0.8 significantly improved strengthening effectiveness. Four additional specimens were tested to validate model predictions, and CatBoost yielded the lowest errors, with a maximum relative error below 5 %. This study demonstrated the potential of ML models to accurately predict the flexural capacity of CFRP-strengthened RC beams and identify critical influencing factors. Future work may expand the dataset and incorporate experimental and numerical simulations to further improve model generalization and guide intelligent structural design.