Authored by Mahyar Asadi, Mohammad Mohseni, Farzin Golkhosh, Majid Tanbakeui Kashani, Michael Fernandez, and Mathew Smith; and presented at the ASME PVP2020 Conference.
When fabrication deals with multiple welds, an optimal weld sequence design can well manage the undesired damaging effects such as the distortion and residual stress during welding. However, the process of finding an effective weld sequence is a challenging task given a large number of possible combinations, i.e. several thousands of welding scenarios. On the other hand, most of the standards require the development of a control plan for mitigation of those undesired effects. Typically, plans to control the damaging effect are, therefore, mostly intuitive with welding engineers relying on previous experience combined with the results of a limited number of practical tests. Welding simulation tools allow engineers to optimize welding scenarios on a digital twin without the need for multiple physical samples. However, the analysis-time practically limits exploring a large number of possible combinations for a weld sequence design. The use of machine learning (ML) algorithms for simulation and artificial neural network (ANN) can be an alternative for fast exploration of various weld sequence scenarios. As opposed to existing ANN and ML algorithms that require an extensive data set to be up to mimic a behaviour, we developed a hybrid-digital twin platform that wisely picks small data set consist of simulation results to construct a meta-model for fast exploration of welding scenarios. The performance and capability of our platform are shown through an example of a complex welded structure with billions of possible welding scenarios to explore.
Read the complete paper here.