An Artificial Neural Network for Parametric Analysis of Metallic Additive Manufacturing Using Discrete Element Method

Powder bed fusion (PBF) is a widely used metal additive manufacturing method. There is strong evidence that the performance of the final part built using PBF depends on the dispersive quality of the particle bed. Understanding this process through computational modeling and machine learning is an efficient low-cost way for process design. Discrete element method (DEM) is an effective tool for analyzing the particle flow behavior. However, one challenge for parametric modeling of highly multivariate powder spreading process through DEM is the high computational cost, for traversing the large parameter space. We address this problem through innovative use of parallel computing using GNU parallel, and by developing a machine learning algorithm to correlate the process parameters and spread quality. We first perform DEM simulations systematically varying four parameters, the particle size, the coefficient of friction, the spread layer thickness, and the recoating velocity. The dataset containing inputs with spread parameters and target variables that measure the spread quality are fed to a finely-tuned artificial neural network (ANN). We observe that the neural network presents at least 95% accuracy in predicting the test data. Ultimately this approach provides the parameter combinations that produce high quality compaction before sintering.