Strength prediction of cold-formed steel beam-column channels: Data-driven versus prescriptive methods

While Cold-Formed Steel (CFS) elements are increasingly used in modern construction, existing methods for predicting CFS beam-column capacity are often inaccurate and overly complex, demonstrating the need for a more reliable data-driven approach. This paper, therefore, aimed to develop a Deep Learning (DL) algorithm based on neural networks to predict the load-carrying capacity of CFS lipped channel beam-columns under combined compressive load and bending. A comprehensive dataset of 628 data points, comprising both physical test results and detailed experimentally validated Finite Element (FE) simulations accounting for geometric imperfections and material nonlinearities, was generated under combined compression with uniaxial or biaxial bending. The input variables of the DL model included cross-sectional dimensions, thickness, slenderness ratios, radii of gyration, and load eccentricities of varying magnitudes and directions, while the output variable represented the compressive strength normalised by yield capacity. The trained DL model demonstrated excellent predictive performance, with a coefficient of determination of R 2 = 0.99 . The model's generalisation capability was further evaluated using unseen beam-column specimens both within and beyond the training dataset range, yielding an average prediction error of approximately 5%. The DL model's predictions were benchmarked against interaction formulas in European, North American, and Australian/New Zealand standards, as well as other literature-based methods. Results show that the DL algorithms predicted CFS beam-column capacities with an average error of 3%, substantially outperforming existing methods, which showed errors of up to 32%. The proposed DL algorithm leads to more efficient design of CFS elements for practical applications.