Towards a general and numerically efficient deposition model for wire-arc directed energy deposition

Abstract

The prediction of deposition behaviour during the wire-arc directed energy deposition process is essential for optimizing process parameters and toolpath during the process planning phase. To this end, this paper presents a new framework for accurate and fast prediction of the stacking of weld beads for arbitrary substrate geometries. The prediction is split into two tasks: the footprint prediction, for which a data-driven approach based on a deep convolutional neural network is proposed; and the weld bead shape prediction, for which a numerical model based on the minimization of surface energy is developed. For each prediction tasks, the prediction approaches are compared against baseline models. An extensive dataset is created for training and evaluation of the models. Dataset generation, noise, training procedures, and evaluation metrics are discussed. The results demonstrate an accurate prediction capability, with a median accuracy of 0.1mm and 0.19mm for single weld bead prediction and end-to-end prediction, respectively. The average prediction time amounts to approximately 150ms per weld bead.