Additive manufacturing: Application and validation of machine learning-based process-structure-property linkages in Ti-6Al-4V

In the field of metal additive manufacturing (AM), various processes and heat treatments can yield unique grain morphologies, thereby influencing material properties and machining behavior. In this study, a novel workflow using a machine learning-based approach that combines statistical descriptors of textured AM-process induced microstructure, cutting force model (as a material response), and a data-mining method is established. It is proven to be a valid method for creating process-structure-property linkages for metal AM. This study focuses on two highly varied metal AM processes: Powder bed fusion (PBF, e.g., laser PBF and electron beam PBF) and directed energy deposition (DED, e.g., wire-fed plasma-directed energy deposition). The study also accounted for the effects of post-AM heat treatment and build orientation. It was found that the accuracy of material behavior predictions is highly correlated with AM processing conditions, building orientations, and machining conditions. Specifically, while initially applying PBF training data to DED samples resulted in a 15% root mean square prediction error, this error was subsequently reduced to <1% through cross-training using combined microstructure training data sets. This discrepancy could be attributed to the significantly different thermal cycling conditions in L-PBF and DED, which resulted in highly varied textured microstructures. Residual stresses generated during AM processing and the selection of machining parameters exert the highest impact on the machining behavior. The implications of these findings extend to the use of statistically descriptive microstructures for various AM processing conditions and build orientations in computational methods and other machining learning approaches.