Compressive strength prediction and interpretable modeling of timber- and bamboo-filled metal tube columns using data-driven machine learning

A data-driven machine learning (ML) approach was used to predict the axial compressive strength of timber-filled metal tube (TFMT) and bamboo-filled metal tube (BFMT) columns. It combined existing experimental data analysis, ML modeling, SHapley additive explanations (SHAP)-based interpretation, and artificial neural networks (ANN)-assisted theoretical model development. A database was developed as the basis for model construction, and five refined input features were used, including metal tube thickness, corner radius ratio, cross-sectional area, compressive strength of the filled material, and tensile strength of the metal tube. Among the six ML models, ANN, extreme gradient boosting, and categorical boosting showed superior predictive performance compared with Bayesian linear regression, support vector regression, and random forest, with ANN providing the best overall accuracy and generalization ability. SHAP analysis indicated that metal tube thickness was the most influential factor, followed by the compressive strength of the filled material, cross-sectional area, tensile strength of the metal tube, and corner radius ratio. Larger tube thickness, higher tensile strength of the metal tube, and a larger corner radius ratio generally improved compressive strength, whereas a larger cross-sectional area weakened the confinement effect. Based on ANN outputs and a fitting method, a practical theoretical model was developed to describe the relationship between selected input features and compressive strength. Although its accuracy was lower than that of the ML models, the proposed model provided a transparent and interpretable tool for preliminary compressive strength prediction. The results provide useful support for the design and application of TFMT and BFMT columns.