Abstract
Welding processes involve complex physical phenomena, of which accurate modeling can be computationally expensive. Still, good predictions of weld bead geometries are necessary for optimal planning of robotic welding. In the current work, a two-dimensional model for welding in a groove assuming parabolic weld bead profiles was established. The scale and orientation of the parabolas were determined from the welding parameters and direction of gravity. Another model employed a purely empirical neural-network approach, in which the bead profiles had a more flexible geometric parametrization consisting of a set of radial distances from the welding torch. This enabled the representation of more complex bead profiles. Calibration and validation of the two models were based on three welding tests with welding directions varying from 0 to 45° compared to the horizontal direction. The correspondence between the modeled and measured bead profiles was good for both models, with the neural network having the highest accuracy, especially in the transitions between bead and substrate. However, further validation and generalization to data sets with larger variation in welding parameters are needed.