SUMMARY

Wire Arc Additive Manufacturing has disruptive potential for modern manufacturing. The technology comes with the flexibility and material efficiency of additive manufacturing processes while mitigating the disadvantages through high material output and high energy efficiency. The prevalence of the technology is inhibited by the large induced residual stresses and geometrical inaccuracy. This work tackles the latter by assessing the process parameter-geometry relationship using Machine Learning (ML) algorithms. To do so, multiple mild steel welding beads with varying shape features like corner angle are printed using a Metal Inert Gas (MIG) welding machine attached to an industrial robot. The cross-sectional profile of the printed beads is measured using a point laser sensor and correlated through different ML algorithms to input features such as travel speed (TS), wire feed speed (WFS), interlayer temperature, and shape features. By incorporating varying bead shapes, a holistic model, not limited to geometry prediction of straight beads only, is trained. Thus, the model holds the potential to learn the process parameter-geometry relationship for different shape features of a part. Using the model, excess material at the inner angle of corners determined by the overlapping regions of the two adjacent beads can be predicted. By generating a database of possible bead shapes a backwards algorithm was created, that suggests welding parameter combinations resulting in a smoother bead shape at corners. Additionally, a study on the transferability of common bead geometry prediction models on other research testbeds was conducted. The importance of input features for transferability is assessed and the potential to increase transferability by infusing the model training with mass conservation is examined.