Prediction of structural behavior of cold-formed steel C-section columns via repetitive backpropagation neural network

Compared to conventional numerical simulations, machine-learning-based surrogate models have demonstrated strong capability in capturing nonlinear behaviours and predicting capacities with high efficiency and accuracy. In this study, an enhanced supervised backpropagation neural network methodology is developed to construct a surrogate model for predicting the load-displacement curves of cold-formed steel (CFS) C-section columns. The machine learning framework integrates physics-based function with a data-driven model, which enables accurate prediction of the implicit curve equations with use of small-scale datasets and compact model architectures. First, a parametric representation method, based on an explicit-implicit hybrid function structure, is introduced to characterise the multi-stage behavior of load-displacement curves. Subsequently, a novel repetitive backpropagation neural network (ReBpNN) is developed to establish a surrogate model. Finally, the surrogate model is trained and validated through dataset generated from geometries of CFS members. Comparative analyses are conducted among ReBpNN and other machine learning methods. Results show that the proposed ReBpNN is feasible for load-displacement curve prediction of the CFS members considering both accuracy and efficiency. The ReBpNN significantly outperforms traditional simulations in computational efficiency, demonstrating strong potential for industrial applications.