Experimental evaluation and prediction model for shear capacity of screw connectors in built-up cold-formed steel composite columns with geopolymer concrete

The study investigates the shear capacity of mechanical screw connectors in built-up cold-formed steel (CFS) members infilled with self-compacting geopolymer concrete (SCGC), with an emphasis on connector performance and infill composition that governs interfacial load transfer relevant in composite column applications. Two innovative configurations, namely 2-sigma (2S) and 4-sigma (4S), were fabricated using interlocked sigma channels connected by mechanical screw shear connectors at mid-length to optimise stiffness and confinement. To characterise the bond behaviour at the steel-concrete interface, a comprehensive push-out testing program was conducted using three concrete infill types: SCGC with natural aggregate (NA), SCGC with recycled concrete aggregate (RCA), and a conventional cement-based control mix (CONC). Experimental results demonstrate that the 4S configuration substantially enhances confinement and shear transfer, achieving average peak loads of 59.72 kN and 64.45 kN for SCGC + NA and SCGC + RCA, respectively, compared with approximately 30 kN for corresponding 2S specimens. SCGC + NA exhibited a stable, ductile post-peak response governed by progressive concrete crushing, whereas SCGC + RCA developed higher initial stiffness but displayed more brittle softening behaviour. Across all configurations, failure was primarily controlled by infill compressive capacity and interfacial degradation rather than connector rupture, indicating a robust interface interaction mechanism facilitated by the screw connectors. To extend predictive capability, machine-learning based models were developed to complement experimental insights and predict shear resistance using material and geometric variables. The Random Forest model outperformed theoretical predictions, achieving an R² of 0.63 and RMSE of 0.65 kN, corresponding to an 89.9% improvement in accuracy. The hybrid experimental-computational approach demonstrates the potential of SCGC systems and the effectiveness of data-driven methods in predicting complex composite behaviour.