Abstract:

Nowadays, the development of new structures is carried out by reusing the engineering knowledge that is encompassed in virtual prototypes. To facilitate this process, most recent product development software benefits from a wide range of parametric design capabilities. From this perspective, a cross-linkage can be achieved between various disciplines, allowing engineers to adjust full scale projects by modifying desired specifications. Even so, such capabilities are limited when designing welded connections, considering the rigid boundaries of Computer Aided Engineering simulation models. To overcome such issues, the present paper proposes a new approach that makes effective use of machine learning for improving the modeling of engineering knowledge. At first, a sensitivity study is conducted on a specific welded structure. By adjusting candidate parameters, the mechanical behavior of the assembly can be evaluated. The results achieved are stored in a dataset that is further used to develop regression models based on neural networks. To this end, a generalized structural behavior is achieved, allowing engineers to predict the performances of welded joints without explicitly updating simulation models.