Using artificial neural networks to model single bead geometries processed by laser-wire direct energy deposition

Abstract

Wire-feed laser additive manufacturing processes have gained researchers’ attention because of their potential to reduce material waste, guarantee accuracy, increase material quality and density, and produce a wide dimensional range of final products. Nevertheless, printing materials with desired geometrical properties of the beads is still challenging in such processes. This might be attributed to the need for more sufficient experimental data and precise modeling approaches. In this study, an architecture based on Artificial Neural Networks (ANNs) is developed to model the bead geometries (width, height, and area), considering the wire feed rate, laser power, and travel speed as process parameters. A design-of-experiment based on full factorial design is considered for processing single beads with a Fraunhofer coaxial wire-feed laser system. Inconel 625 wire with a diameter of 1.14 mm and stainless steel substrate are utilized as the experimental materials. Geometrical data is obtained using a laser scanner model RA-7525 SE with 0.026mm volumetric accuracy. The beads’ geometrical details are provided as the feeding data for the proposed ANN. For each bead, a length of 10 mm is considered to calculate the average geometrical parameters, which increases the accuracy of the dataset in comparison to the values acquired via a macroscopic picture of the cross-section of each weld bead. A variety of hyperparameters are chosen and compared regarding precision criteria, including Mean Square Error (MSE), to increase the model‘s accuracy. A train-test separation strategy is considered to evaluate the model‘s accuracy on independent data points. The outcome of this research is an ANN-based geometry prediction model that can be utilized to enhance the development of offline path planners and optimize process parameter selection for a precise geometry toward process control.