Data-driven modelling of bending resistance in built-up cold-formed steel lightweight concrete (CFS-LWC) composite beams using artificial neural networks

Cold-formed steel–lightweight concrete (CFS-LWC) composite beams offer significant advantages in modular and lightweight construction, but their complex structural behaviour, governed by material nonlinearity, shear interaction, and local buckling of thin-walled sections, poses challenges for traditional design methods. While recent studies have validated numerical models for CFS-LWC beams, these approaches remain computationally intensive and may lack adaptability for rapid design iterations. This study develops a data-driven artificial neural network (ANN) surrogate model to predict the ultimate bending resistance M u ​ of built-up CFS–LWC composite beams. The study used a high-fidelity dataset comprising 960 finite-element simulations, which was calibrated using four experimental tests. This dataset spans beam lengths from 4310 to 8000 mm, steel section depths from 240 to 300 mm, lightweight concrete compressive strengths between 20 and 35 MPa, slab thicknesses representative of typical floor systems, and steel grades S280GD and S450GD. Inputs were z-score standardised, and model training used five-fold cross-validation with Latin Hypercube Sampling for the hyperparameter search. The final network (one hidden layer, 64 neurons, tanh activation function, and RMSprop optimiser) achieved R 2 ≥0.98, with residuals centred near zero and limited spread, demonstrating strong generalisation within the studied domain. Model explainability via SHAP (Shapley additive explanations) indicates that steel yield strength and section geometry are the most influential features, consistent with structural mechanics. The proposed model shows good agreement with the FE database and can therefore provide more accurate predictions of the bending resistance of CFS–LWC composite beams than conventional analytical approaches. In addition, the surrogate model offers fast and reasonable capacity predictions, minimising the need for repeated FE analyses and making it well-suited for the preliminary design of CFS–LWC composite beams within the validated parameter range.