Hybrid Machine Learning for Accurate Prediction of CFST Column Compressive Strength

This paper introduces a hybrid machine-learning framework to improve the predictive accuracy of ultimate compressive strength in circular concrete-filled steel tube (CFST) columns. The suggested methodology combines CatBoost with Bayesian optimization to enhance model efficacy and computational efficiency. A dataset of 663 experimental specimens is employed for training and validation. Sophisticated data preprocessing methods, encompassing mathematical transformations, are utilized to enhance feature representation. The efficacy of the proposed method is assessed through a comparative analysis with conventional artificial neural networks (ANN). The hybrid CatBoost model demonstrates enhanced predictive accuracy, significantly lowering error metrics compared to ANN-based models. The proposed framework specifically decreases the Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R² score by 146.55, 262.55, and 0.99%, respectively, illustrating its efficacy in structural engineering applications. The selection of CatBoost is driven by its capacity to manage intricate nonlinear relationships, reduce overfitting, and ensure computational efficiency, rendering it a persuasive alternative to traditional machine learning methods.