Wire Arc Additive Manufacturing (WAAM) is a manufacturing process that deposits weld beads layer-by-layer in a planar fashion, leading to a final part. Thus, the accuracy of the printed geometry is largely dependent on the knowledge of the bead profile employed, which by itself is dependent on a variety of process parameters, such as wire feedrate and torch speed. Existing models for modelling bead profile are based on its width and height, which do not necessarily capture the geometry of the weld bead accurately. This could affect the step over increment strategy, which dictates the geometry of the resulting overlapping valley.
In this paper, we formulate and evaluate the performance of a variety of machine learning framework for predicting the bead cross-sectional profiles. To model the geometry of a bead, we explored direct cartesian representations using polynomials and vertical coordinates, as well as a higher dimensional representation using planar quaternions for supervised learning. Experiments are conducted on single bead SS316L material to compare the various framework performance. We found that among these, the planar quaternion representation with a non-linear neural network framework captures and retains the curvature characteristics of the bead during the learning and prediction process most accurately with a mean Chi-Square goodness of fit of 0.026.