Generalizable classification methodology for quantification of atomized feedstock powder by 3D X-ray tomography data and machine learning

Laser powder bed fusion (LPBF) is currently being applied to manufacture engineering-crucial components. To maintain consistent part quality, accuracy and speed in the quality assurance of atomized metal feedstock powder is critical. 3D x-ray tomography (XRT), coupled with machine learning algorithms, provides a transformative route to powder characterization and classification. A recycled AA7050 feedstock powder was studied through XRT to demonstrate a scheme for classification of highly deformed particles which vary both in geometric morphology and degree of surface irregularity. Manual, unsupervised, and supervised classification algorithms were optimized to reproduce visual classification, demonstrating how different approaches to algorithm training may provide a balance between the amount of training data and acceptable final accuracy. The reported approach provides a robust methodology that links 3D measurements and powder classification as means to control powder-induced defects and improve mechanical performance in printed parts.