This study develops and evaluates machine-learning models for predicting the confined ultimate strength (fcc, u) and ultimate strain (εcc, u) of a hybrid multi-tube concrete column (MTCC). It is a pioneer structural system that uses a fiber-reinforced polymer (FRP) outer tube, inner steel tubes and concrete with voids. One of the aims of the study is to enhance the accuracy of prediction of the structural performance of this system using machine learning techniques especially using gradient boosting machine (GBM) models. It utilized a database of 283 specimens produced as a result of published experimental studies and it included the geometric and material characteristics that impacted performance. Four example gradient-boosted decision-tree models, including Stochastic Gradient Boosting (SGB), Extreme Gradient Boosting (XGB), Light Gradient Boosting Machine (LGB), and CatBoost (CGB) models were trained and optimized through Bayesian hyperparameter tuning to compare the robustness of the main boosting models. Through the analysis, it was found that SGB model was the most precise with the highest coefficients of determination (R2 = 0.994 when using fcc, u and 0.946 when using ecc, u) and the smallest root mean square error (RMSE). And the logic of the model was interpreted using SHAP analysis, which indicated that concrete compressive strength (f’c) and the thickness of the FRP layer (tf) had the greatest effect on fcc, u. On the contrary, the most crucial factors in calculating ecc, u were the FRP elastic modulus (Ef) and the steel pipe yield strength (fys). An interactive graphical interface was created in an attempt to improve practicality and provide the ability of engineers and researchers to make correct predictions in a simple and efficient manner. The findings confirm the usefulness of machine learning models, especially gradient boosting, in facilitating design and analysis decisions in high-end engineering processes.
Luo et al. (Mon,) studied this question.