AI‐Driven Optimization and Surrogate Modelling Framework for Predicting Steel Joint Behaviour

Abstract

Steel joint behaviour governs the stiffness, strength, and ductility of steel structures, yet post‐limit response remains insufficiently defined in current design practice. This paper presents a Physics‐Informed Neural Network (PINN)‐based surrogate modelling framework for welded beam‐to‐column joints. The approach builds on prior Genetic Algorithm (GA)–based characterisation of the nonlinear force–deformation (F–d) behaviour of the governing joint components: column web in shear, compression, and tension, calibrated against experimental data using multi‐point F–d curves. In the present study, the GA‐derived stiffness and resistance ranges are embedded as physics‐based constraints within the PINN loss function. The PINN is trained on a dataset of welded joint configurations and evaluated using a leave‐one‐out strategy. The results demonstrate a consistent and accurate reproduction of full‐range moment–rotation (M–θ) responses and component‐level behaviour across all investigated configurations. By integrating optimisation‐based component characterisation with physics‐informed surrogate learning, the proposed framework enables efficient and physically consistent prediction of welded joint behaviour beyond the elastic range, supporting advanced performance‐based and reliability‐oriented design.