Square concrete-filled steel tubular columns are fundamental components in modern structural engineering, leveraging the composite action of steel and concrete while alleviating their individual shortcomings. However, accurate evaluation of shear capacity remains challenging due to the complexity of load transfer mechanisms. To address this issue, an integrated framework encompassing data acquisition and augmentation, machine learning-based modeling, and reliability assessment is proposed for comprehensive shear performance evaluation. A shear test database was first established and subsequently expanded using a joint principal component analysis (PCA)–Gaussian Copula technique, which preserves statistical properties and intrinsic correlations. Based on the enhanced dataset, five machine learning models were developed with explicit control of overfitting and underfitting. Model interpretability was further improved through explainable techniques to elucidate feature contributions and interactions. Finally, reliability analysis was conducted using the resistance reduction factor approach. The results indicate that (a) the PCA–Gaussian Copula method effectively enriches the dataset while maintaining statistical fidelity and latent correlations; (b) all models exhibit high predictive accuracy (R2 0.90, prediction–training deviation 0.05), with categorical boosting achieving the best performance; (c) six key parameters adequately characterize the shear behavior, among which column side length and tube thickness are dominant; and (d) reducing the resistance factor from 0.75 to 0.72 ensures sufficient safety margins.
Xue et al. (2026) studied this question.