Prediction of CFRP and NSM-Wrapped Composite Column Capacities Using Experimentation and an Ensemble Machine-Learning Approach with SHAP Interpretation

Column capacity is an essential parameter in structural design, and its accurate determination is critical for a safe load-transfer mechanism in structures. Also, experimental and accurate model-based assessments are critical to the column capacity evaluation. The main objective of this study is to experimentally and analytically investigate near-surface mounted (NSM) wrapped columns of different configurations and compare their capacity enhancements with the capacity enhancements of carbon fiber-reinforced polymer (CFRP). The study is also focused on developing a statistical regression model and extreme gradient boost (XG Boost), an ensemble machine-learning (ML) approach-based model, and examining both models developed for the experimental results by Shapley additive explanations (SHAP) interpretations. Therefore, the study experimentally reviewed the behavior of 24 composite columns to gain insights into experimental and code-recommended column capacities, stress–strain responses, axial stiffness, ductility factors, and failure modes. NSM-wrapped columns gained 10% strength increments, and, in comparison, the full-wrapped CFRP columns achieved 22% strength enhancement. The structural columns in a structure typically require various levels or types of strengthening, depending on their loading conditions, geometry, and material properties. With a 10% increment, the NSM technique suits columns needing lesser strength enhancements. Therefore, a key finding of the study is that the contribution of NSM longitudinal wrapping to column capacity is significant and cannot be ignored. A statistical regression model is developed for column capacity with four key parameters: percentage steel reinforcement, the extent of epoxy adhesion, the weight of the specimen, and the concrete clear cover. A model based on XG Boost, an ensemble ML approach, is also developed for the same four key parameters. The models developed are evaluated by SHAP interpretations. The SHAP analysis technique interpreted this improved model for various input–output features. The XG Boost machine-learning algorithm, developed with a coefficient of determination of 0.99, is found to be a refined regression model. Also, the study establishes that the ensemble ML approach used in tandem with SHAP analysis is a robust prediction and model interpretation tool, highlighting the significance of the percentage of steel reinforcement and the extent of epoxy adhesion over the other variables for the experimental dataset.