Hybrid machine learning model for accurate prediction of compressive strength for CFRP-confined non-circular concrete columns

This study investigates the application of machine learning (ML) models to accurately predict the compressive strength ratio of square and rectangular concrete columns confined with carbon fiber reinforced polymer (CFRP) sheets. A comprehensive dataset consisting of 333 CFRP-confined concrete columns compiled from 31 independent experimental studies was used. The dataset incorporates geometric properties, unconfined concrete parameters, and CFRP material characteristics as input features. Four predictive models were developed: a standard artificial neural network (ANN), a random forest (RF), an ANN optimized using random-search-based neural architecture search (NAS-ANN), and a Hyperparameter Optimized RF (HO-RF) model. Model performance was evaluated using several statistical metrics, including the R2, R, IOA, RMSE, MAE, and IOS. In addition, k-fold cross-validation was employed to further assess the predictive capability and generalization performance of the models. Among all models, the NAS-ANN model consistently outperformed the others, demonstrating superior accuracy and robustness. Furthermore, it surpassed ten widely used empirical models reported in the literature. The NAS-ANN model achieved R2 values of 0.9825 during training and 0.9334 during validation, effectively capturing the complex behavior of CFRP-confined concrete columns. The high predictive accuracy of the NAS-ANN model provides a valuable tool for estimating the compressive strength ratio of CFRP-confined concrete columns. Overall, the study highlights the effectiveness of data-driven model optimization for structural engineering applications and provides a reliable approach for predicting the behavior of FRP-confined concrete columns.