Use of ANN to simulate the effects of welding process parameters in curved steel plates: residual stresses, strains and distortions

In large steel fabrication industries such as shipbuilding, and high-speed train guideways, the problem of residual stresses and overall distortion has been and continues to be a major issue. In the last few decades, various research efforts have been directed at the control of the welding process parameters aiming to reduce distortions and residual stresses. Yet, in actual practice, large amounts of resources are still required to rework welds. These costs increase production costs and delay work completion. In the work reported here the Finite Element Method (FEM) is used to simulate the welding process in two-steps; first a non-linear heat transfer step that yields the dynamic temperature distribution throughout the weld seam and the plates, and second, the elasto-plastic analysis, which yields the residual stresses, strains, and the displacements. The responses focused upon were those along the longitudinal cross sections after the welded piece had cooled down to room temperature. An artificial neural network is trained using FEM simulation data for a wide variety of geometric and process parameter combinations. Then, the resulting neural networks is shown to be capable of predicting the welding response without having to carry out a computationally complex, time consuming full finite element analysis. This concept is shown to be a highly effective and efficient way to predict welding responses for welding process design purposes.