Ultimate axial load capacity of cold-formed steel tubular T-sections reinforced with collar plates: a machine learning approach

Abstract

This study investigates the structural performance of Cold-Formed Steel (CFS) tubular T-joints with and without collar plate reinforcement using a combined experimental, finite element, and machine learning approach. Experimental tests established the baseline behavior, showing that collar plate reinforcement enhanced the ultimate axial capacity by up to 35% and improved post-peak ductility, whereas unreinforced joints exhibited brittle fracture at brace–chord intersections. Finite Element Analysis (FEA) developed using SHELL181 elements accurately reproduced deformation patterns and chord plastification modes, with predictions within ±10% of experimental results. A comprehensive parametric study of 160 finite element models extended the dataset by varying geometric parameters, including collarto- chord thickness ratio (τc), chord slenderness ratio (γ), brace-to-chord diameter ratio (β), and collar overhang (η). Results showed that increasing β from 0.25 to 1.0 raised load capacity from 23.5–89.4 kN to 45.6–186.5 kN, while the strongest configuration (β = 1.0, τc = 2.25, η = 2.25) achieved 328.1 kN - nearly nine times higher than the weakest case. Thicker chords (low γ) significantly improved resistance, identifying γ as a critical factor. Machine learning models, particularly Artificial Neural Networks (ANN), trained on the FEA dataset achieved excellent predictive performance (R² = 0.9952), outperforming linear methods and performing comparably to ensemble models such as Gradient Boosting (R² ≈ 0.98–0.99). Feature importance analysis confirmed τc and γ as dominant predictors, jointly contributing nearly 50% of accuracy. The integrated experimental–numerical–ML framework presented herein provides an accurate, scalable, and computationally efficient tool for the design and optimization of reinforced CFS tubular joints.