Key points are not available for this paper at this time.
Steel connections play a crucial role in providing links between structural elements such as beams and columns, and in maintaining the overall stability of the structural system. Accurately predicting connection behaviour and strength is critical for ensuring structural safety. Bolted connections are commonly used as shear, tension or moment-resistant connections in cold-formed steel framing. Based on 35 experimental data points from an experimental program recently performed at the University of Sydney on beam-end bolted connections, finite element (FE) models are first developed using ABAQUS software and subsequently validated against the experimental results. The FE models demonstrate good agreement with the experimental data in terms of ultimate strength, load–deflection response, and failure mode. After validation, 115 additional FE models are generated in a parametric study to expand the current database. This paper consequently proposes an efficient and reliable machine learning-based framework, which integrates a Gaussian process regression (GPR) model with an enhanced comprehensive learning particle swarm optimization (ECLPSO) algorithm, referred to as hybrid GPR-ECLPSO, to predict the ultimate strength of beam-end bolted connections (asymmetric connections) in cold-formed steel channels, failing in block shear mode. A total of 150 data points, with varying characteristics, are compiled to train the GPR model, with the ECLPSO algorithm primarily adopted to determine the GPR hyperparameters. The performance of the hybrid GPR-ECLPSO is evaluated using various statistical estimators and compared with existing machine learning models (e.g., support vector machine, artificial neural network, and three typical ensemble machine learning models). All experiments, FE simulations, and machine learning results are compared against the predictions from the current design rules in the Australian/New Zealand Standard (AS/NZS 4600) and the North American Specification (AISI S100) for the design of cold-formed steel structures. The results indicate that the hybrid GPR-ECLPSO model is more accurate than other ML models, highlighting the efficiency and precision of the present work. Finally, a variance-based global sensitivity analysis, leveraging the trained GPR-ECLPSO model, is proposed to investigate the effect of input variables on the model output and identify the most significant variables. • Behaviour of shear beam-end bolted connections in CFS structures failing in block shear is investigated numerically. • Finite element model is conducted and validated against experimental test results. • A hybrid GPR-ECLPSO model is developed to predict the shear strength. • Variance-based GSA is performed on the hybrid GPR-ECLPSO model to evaluate the effect of input variables on the model output. • A GUI tool is developed based on the hybrid GPR-ECLPSO model for practical engineering design.
Van et al. (Thu,) studied this question.