Machine learning framework for predicting the shear capacity of demountable bolted connectors in composite beams

Abstract

Steel-concrete composite beams are increasingly adopted in modern construction owing to their high strength, stiffness, and efficiency. Conventional welded shear connectors, while effective, hinder disassembly and recycling, limiting their alignment with sustainable construction practices. To address this, demountable bolted connectors have emerged as a viable alternative, promoting reuse, reduced waste, and compatibility with modular construction. This study presents eight machine learning algorithms including linear, tree-based, and ensemble methods were trained on a hybrid dataset combining experimental that were collected from previous studies and numerical results. Among these ML models, the XGBoost Regressor exhibited the highest accuracy (R² ≈ 0.996) with consistently low error margins, while SHAP-based interpretability confirmed bolt diameter and reinforcement strength as the most influential predictors. The findings highlight the potential of combining advanced testing, finite element modeling, and machine learning to establish robust predictive tools for demountable connector systems. This multidisciplinary approach not only improves design accuracy but also supports the development of sustainable, reusable, and high-performance composite structures in line with circular economy principles.