Mechanistic-data-driven modeling of multi-material composite columns: Toward intelligent lightweight design

This study examines the axial compressive performance of multi-material composite columns consisting of concrete-filled steel tubes with embedded CFRP-confined timber cores. A data-driven framework integrating theoretical model, finite element simulation and machine learning prediction is established to address the limited accuracy and scalability of conventional dual-material designs. An analytical bearing-capacity model is derived by accounting for steel confinement, CFRP hoop restraint, and timber orthotropy, of which results match FE results well with 5% deviations. Parametric investigations show that increasing steel yield strength and tube thickness would enhance the capacity of the composite columns, whereas CFRP confinement improves the post-crushing response and ductility of the timber core. The columns with circular cores exhibit better deformability than those with square ones. For axial bearing capacity prediction, a theory-residual-modified XGBoost model is proposed, in which theoretical estimates are corrected via SHAP-guided residual learning, achieving higher accuracy than single learners and ensemble baselines. A lightweight design tool is further developed for single/batch evaluation, automatic capacity-to-self-weight assessment, and interpretable prediction, enabling up to 22% self-weight reduction. The proposed methodology provides a validated and practical route for optimizing sustainable, lightweight multi-material composite columns.