Forward and inverse surrogate modelling of welding residual stress in electron beam welded 316L steel using multilayer perceptron networks

Residual stress analysis remains a major challenge in welded structures due to the limitations of traditional finite element analysis (FEA) and the high cost of experimental methods. Recent advances in artificial intelligence have enabled machine learning (ML)-based surrogate models to offer faster, more generalisable predictions. This paper presents an integrated workflow for developing and validating both forward and inverse surrogate models using multi-layer perceptron (MLP) neural networks. The workflow combines the numerical simulation models of electron beam welded 316L steel pipes, semi-automated data generation via Python scripting, and calibration with experimental measurements The forward model predicts longitudinal residual stress profiles from welding process parameters, while the novel inverse model estimates optimal welding parameters to achieve desired residual stress outcomes—a capability not accessible through conventional FEA. Both models demonstrate high accuracy within the training range, though their extrapolation abilities are limited. The proposed approach significantly improves computational efficiency and provides a practical framework for process optimisation in advanced welding applications.