Stress gradient approaches in multi-pass welded joints: A surrogate modelling approach combining finite element analysis and machine learning

Abstract

This study developed a comprehensive framework combining parametric finite element modelling, the theory of critical distances (TCD), and artificial neural networks (ANN) to predict (effective) normalised stresses in multi-pass welded joints. Parametric 2D models were generated for both flat and hollow weld forms, incorporating the weld toes, root and intermediate weld-notch geometries. Linear-elastic FE simulations (using ANSYS MAPDL) of over 2000 geometric variants were performed to extract local stress fields at the notches surrounding. Both the point method and line method of the TCD were used to compute the critical-distance-averaged stress as a function of the material characteristic length L. As noted by Susmel 2006, the TCD is a (powerful engineering tool) that can post-process linear-elastic FE models and thus reduce design time and cost.