Faster than real-time path-sensitive temperature modeling of wire-arc additive manufacturing by a data-driven finite volume method

Abstract

Wire-arc additive manufacturing uses arc welding to build 3D objects by progressive deposition of weld beads at high deposition rates. To minimize the distortion induced by the moving heat source, a model for predicting the transient temperature fields as a function of the deposition path is needed. This work proposes a new data-driven finite volume model that combines the semi-discrete form of the energy balance with a temporal convolutional neural network. Compared to recurrent neural networks, the model is energy-conserving and computes temperature profiles on a grid of positions in parallel, thus executing substantially faster than the actual process.