Learning automatic part orientation for powder bed fusion of metal from part data

Abstract Part orientation is a critical parameter of laser powder bed fusion of metal, as it significantly affects manufacturing cost, production time, and surface quality. Trained/experienced personnel is required to achieve optimum results. Learning implicit part orientation schemes directly from manually prepared parts would enable automatic part orientation selection and support human operators. This study investigates the feasibility of learning part orientation schemes directly from part data. We assume that a dataset of geometries with manually chosen part orientations contains implicit patterns that can be inferred directly from 3D geometries and their features (shape, holes, etc.). Leveraging PointNet architectures, we train models on pre-defined part orientations, which output a 3D rotation for any given geometry. The impact of different formulations of 3D rotation on the training performance is evaluated and a novel z-invariant rotation definition is suggested which achieves high accuracy. The approach is assessed using four synthetic datasets with homogeneous 3D geometries and a combined dataset with mixed geometries. The approach results in an average relative angular error of 0.7° with an average standard deviation of 0.47° across all datasets with distinct geometries. The results on the mixed dataset of all geometries look equally promising with 1.26° and a standard deviation of 1.68°. Regarding the application in laser-based powder bed fusion of metal, the achieved angular error is satisfactory for accurate alignment.