Hybrid and Ensemble Machine Learning Approaches for Predicting Axial Load Capacity in Rectangular CFST Stub Columns

Abstract

Concrete-filled steel tubular (CFST) columns are widely utilized in Structural Engineering because of their outstanding Load-Bearing Capacity (LBC) and ductility. Current design codes offer inconsistent predictions for Axial Load Capacity (ALC), particularly for high-strength Rectangular CFST Stub Columns (R-CFST columns), leading to uncertainty in practical applications. This study addresses this gap by developing interpretable and accurate Machine Learning (ML) models for predicting the ALC of such columns. A Database of 719 experimental results was compiled, encompassing six input features related to geometry and material properties. The core ML algorithm used is Histogram Gradient Boosting Regression (HGBR), which is further enhanced using two metaheuristic optimization algorithms: the Lotus Effect Optimization Algorithm (LEOA) and the Emperor Penguin Optimization Algorithm (EPOA). An ensemble strategy based on Dempster–Shafer theory (D–S theory) is also proposed. Model performance is evaluated using R², RMSE, MSE, MRAE, and RSR metrics. The hybrid HGBR–LEOA model (HGLA) achieved the best performance with R² = 0.9933 and RMSE = 202.728 in the test set. A sensitivity analysis using the Pearson Correlation Coefficient (PCC) identified Wall Thickness (t) and Section Width (B) as the most influential features. The outcomes illustrate that the suggested ML models significantly outperform traditional design code predictions and offer a fast, reliable alternative for early-stage structural design. This document provides a practical, data-driven framework that bridges the gap between empirical behavior and design code limitations for R-CFST columns.