Interpretable Machine Learning for Predicting Compressive Strength of CFRP-Confined UHPC Short Columns

Abstract

The compressive behavior of carbon fiber-reinforced polymer (CFRP)-confined ultra-high-performance concrete (UHPC) short columns involves nonlinear interactions among concrete properties, confinement characteristics, and specimen geometry. In this study, a database of 144 circular specimens was compiled using nine input variables: specimen diameter (D), height (H), CFRP thickness (t), number of CFRP layers (Layers), unconfined concrete compressive strength (fco′), concrete ultimate strain (εco), CFRP tensile strength (ff), CFRP ultimate strain (εf), and CFRP elastic modulus (Ef). Eight regression algorithms were evaluated using an 80:20 training–test split, with ten-fold cross-validation conducted within the training subset for hyperparameter selection. ANN achieved the best test performance, with an R2 of 0.96, an MAE of 8.37 MPa, and an RMSE of 11.32 MPa, while SVM and CatBoost also showed competitive predictive accuracy. The selected ML models exhibited substantially lower prediction errors than seven existing empirical equations. SHAP analysis further identified CFRP layer number, unconfined concrete strength, specimen height, and CFRP thickness as influential predictors and revealed interactions among confinement, matrix deformability, and specimen geometry. The proposed framework provides an accurate and interpretable supplementary approach for assessing the compressive strength of CFRP-confined UHPC within the parameter range represented by the database.