Novel insights into axial compression of concrete columns coupled with GFRP via machine learning

This study provides novel insights into the axial compression behavior of concrete columns reinforced with glass fiber–reinforced polymer (GFRP) bars. A dataset of 107 experimental cases was modeled using diverse algorithms, and predictive accuracy was benchmarked across tree-based, kernel-based, and neural architectures. Ensemble Learning achieved the highest accuracy (R 2 = 0.983), and MLP-ANN delivered the lowest error (AARE% = 10.139), demonstrating strong generalization. Beyond performance metrics, SHAP dependence analysis revealed the dominant influence of column diameter and longitudinal reinforcement area, while uncovering synergistic interactions between geometric and material parameters. Dependence plots highlighted monotonic and nonlinear trends, such as the strong positive effect of increasing diameter coupled with higher concrete strength, and compensatory behaviors where reinforcement strength mitigated slenderness penalties. These findings confirm that the models achieve high predictive accuracy while aligning with structural mechanics principles through transparent feature attribution. SHAP dependence plots show that larger column diameters consistently increase axial capacity, validating established geometric design assumptions. They also reveal that higher GFRP tensile strength only partly offsets slenderness penalties, refining conventional reinforcement expectations. By bridging statistical learning with engineering interpretability, this research establishes a robust framework for optimizing the axial performance of GFRP-reinforced concrete columns.