Hatch pattern based inherent strain prediction using neural networks for powder bed fusion additive manufacturing

Abstract

Abstract Additive Manufacturing has recently emerged as an important industrial process that is capable of manufacturing parts with complex geometry. One of the drawbacks of metal additive manufacturing processes is the thermo-mechanical distortion of the parts during and after build due to heat effects. Inherent strain is widely adopted by researchers as the basis to predict part distortions during Metal Powder Bed Fusion Additive Manufacturing (PBFAM) process and is highly dependent on the laser hatch pattern sintering on each layer during the printing process. There is a clear need to predict inherent strains for a given arbitrary hatch pattern for a part model so that hatch patterns can be optimized for achieving part quality. In this paper, we propose a neural network based method to predict inherent strain for any given hatch pattern that is adopted during the part build. The authors assumed that the temperature profile inside the heat affected zone within each layer is the same if the part model is reasonably large. To start with, inherent strains of two hatch pattern pools with different hatch angles were obtained by thermo-mechanical simulation with temperature profiles obtained through translation and rotation of a single layer of simulation. A feedforward backpropagation neural network was created and trained with data obtained from an initial hatch pattern pool for predicting inherent strains. The data from a second hatch pattern pool was then utilized to validate the network and test the efficacy of the prediction of the trained neural network. The results show that the trained neural network is capable of predicting the inherent strain of any arbitrary hatch pattern within an acceptable error. Since the trained neural network can predict inherent strain quickly for any given hatch pattern, this could provide the basis for hatch pattern optimization of any part model to increase part build accuracy and achieve part GD&T callouts.