Efficient data-driven models for prediction of ultimate shear strength of reinforced concrete columns
Thai-Hoan Pham,Dai-Nhan Le,Van-Thanh Pham,Zhengyi Kong,Quang-Viet Vu
Abstract
This paper aims to propose hybrid machine-learning (ML) models to predict the ultimate shear strength of reinforced concrete (RC) columns with both circular and rectangular sections. Fifteen optimization algorithms from five groups, including the Bio-based, Evolution-based, Human-based, Math-based, and Swarm-based optimization algorithms were used to optimize the hyperparameters of the eXtreme Gradient Boosting (XGB) model. These ML models were trained with a dataset consisting of 497 experimental data points of the RC columns and evaluated with coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), and a20-index metrics. It was found that the XGB model optimized with the Battle Royale Optimization (BRO) algorithm (BRO-XGB) shows the best performance according to all metrics. In particular, it achieved an R2 score of 0.971 and an a20-index score of 0.966. This model also demonstrates its superiority in accuracy when compared to design codes and previous studies. The proposed model achieved an RMSE value less than half, and an MAE score only one-fourth, of those reported in the best previous study. Additionally, the uncertainty analysis was conducted to estimate the convergence of the prediction results and the sensitivity values of the input parameters. Finally, a support tool was developed, leveraging the BRO-XGB model's predictions, to suggest appropriate RC column dimensions and material strengths. This tool facilitates easier parameter selection and substantially shortens the preliminary design, without requiring knowledge of ML, optimization, or programming.
