Concrete-filled steel tubes (CFSTs) are widely used in civil engineering due to their excellent bearing capacity, ductility, and seismic performance. Recent innovations, including rubberized concrete and corrugated steel tubes, have further enhanced these properties. In particular, rubberized concrete-filled corrugated steel tubes show great potential. However, their axial load-strain behavior is highly nonlinear with complex parameter interactions, making accurate prediction challenging for traditional models. This study proposes a unified predictive framework that combines Bayesian optimization with an ensemble of XGBoost, LightGBM, and CatBoost to reconstruct the full axial load-strain curve for four CFST variants. The model is trained on a hybrid, qualitycontrolled dataset of 526 curves (53,126 points; 71.9 % experimental and 28.1 % validated finite element data), with standardized conversions and resampling. Mechanism-based features help reduce source bias and capture confinement and buckling effects. The ensemble model, using optimized hyperparameters and validation-driven weighting, achieves high accuracy and robustness (test R2 approximate to 0.98; most errors <= 10 %), outperforming traditional methods. Furthermore, the staged SHAP analysis clarifies how key parameters affect different loading phases, improving model interpretability and engineering relevance. This work provides a generalizable tool for predicting CFST behavior and supports structural optimization and sustainable design.