How AI is Improving Welding

“At the interface between mechanical engineering and AI, the biggest challenge is the issue of data,” says Hinz: “Mechanical engineering and engineering sciences have little or no data, or only data of poor quality.” The industry is reluctant to make faulty production rejects only available for external academic research. Furthermore, sensor and camera data from factory floors often provide unreliable or incomplete information because interruptions, impacts and other factors distort it.

Measurement data is also lacking, particularly regarding welding problems or the stresses these cause in components. “We investigated the welding of hollow sections, as this can save weight and material in the construction of car chassis or cranes,” explains Hinz. “But the safety and stability of these joints remain largely unexplored and have so far been simulated using finite element modelling.” These calculations are carried out on high-performance computing (HPC) systems and can only be varied at great expense – for example, to model different parameters or additional scenarios: in a company’s day-to-day production, this takes far too long.

AI can reduce computational effort: To this end, the team of the professors Hinz and Klemens Rother trained a Graph Neural Network (GNN) using the results from 250 classical simulations, which had previously been carried out at the LRZ using ANSYS Parametric Design Language and which depict various butt welds. These join workpieces at their edges and on a single plane. “The models depict spatial structures. This enables us to better identify where stresses occur in the material,” says Hinz. “GNNs operate on similar spatial structures and can therefore be trained very effectively using this simulation data. We can use the AI models obtained in this way as surrogate models to simulate welded joints.”