Reliability Assessment of Welded Beam‐to‐Column Joints Using Physics‐Informed Neural Networks

Abstract Despite advances in joint modelling, the ongoing demand for more efficient and sustainable steel structures calls for further improvements in the design of connections. This paper presents part of a broader research effort aimed at enhancing the design efficiency of steel joints by realistically capturing their behavior and enabling the development of more accurate and reliable design criteria. A novel application of Physics‐Informed Neural Networks (PINNs) is proposed to model the moment–rotation behaviour of welded beam‐to‐column joints. PINNs are trained on a parametric dataset generated from validated FEM models, embedding analytical expressions for initial stiffness as a physics constraint. The results show excellent agreement with FEM simulations in the elastic range and promising accuracy beyond it. By replacing FEM for large‐scale data generation, PINNs unlock the stochastic characterisation of joint behaviour within the probabilistic framework of EN 1990, supporting Monte Carlo simulations and the derivation of statistically robust resistance criteria. Future developments will extend the approach to other joint configurations and integrate the results into macroelements for system‐level digital workflows, supporting the advancement of the Direct Design Method.