An Optimized Interpretable Machine Learning Model for Predicting Ultimate Compressive Capacity of RCFST Columns

Abstract

Rectangular concrete-filled steel tube (RCFST) columns are widely adopted as primary load-bearing components in engineering structures, and making reliable estimation of their axial compressive capacity crucial to structural design and safety assessment. However, existing theoretical and design equations generally rely on a series of simplifying assumptions and empirical modification coefficients, which limit their prediction accuracy and applicability. In this context, a data-driven strategy is proposed in which capacity prediction and model interpretation are integrated within a unified procedure. The proposed approach establishes the nonlinear mapping between high-dimensional input features, including geometric and material properties, and axial compressive capacity, while employing the SHapley Additive exPlanations (SHAP) method to quantify the contributions of critical input parameters to the model predictions. An experimental database comprising 745 axial compression tests on RCFST columns is established through an extensive literature survey. Prior to training, correlations within the candidate inputs are examined using Pearson coefficients. An SMA-LSSVM–ANN hybrid prediction model is then developed, in which SMA iteratively searches for the optimal combination of the LSSVM-ANN hyperparameters using mean squared error as the fitness function. Comparisons with several benchmark machine learning (ML) models and existing design-code formulations validate the predictive performance of the proposed model, which achieves an R2 of 0.9571 and an RMSE of 302.84 kN on the testing dataset. Furthermore, SHAP analysis is employed to interpret the proposed model from global and feature-dependency perspectives. Global SHAP analysis identifies concrete compressive strength and cross-sectional dimensions as the dominant features, while feature-dependence analysis reveals generally positive effects of cross-sectional dimensions and material strengths and a negative effect of column length on the predicted capacity.