A Data-Driven Framework to Select a Cost-Efficient Subset of Parameters to Qualify Sourced Materials

• Published in: Integrating materials and manufacturing innovation Vol. 11; no. 3; pp. 339 - 351
• Main Authors: Senanayake, Nishan M., Carter, Jennifer L. W., Bowman, Cheryl L., Ellis, David L., Stuckner, Joshua
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.09.2022; Springer Nature B.V
• Subjects: Additive manufacturing; Characterization and Evaluation of Materials; Chemical composition; Chemistry and Materials Science; Laser beam melting; Machine learning; Materials Science; Mechanical properties; Metallic Materials; Microstructure; Nanotechnology; Nickel base alloys; Rheological properties; Rheology; Statistical analysis; Structural Materials; Superalloys; Surfaces and Interfaces; Technical Article; Thin Films; Variables; Feature selection; Processing-structure-property; Materials informatics; Additive manufacturing; Machine learning
• ISSN: 2193-9764; 2193-9772
• DOI: 10.1007/s40192-022-00266-3

The quality of powder processed for manufacturing can be certified by hundreds of different variables. Assessing the impact of all these different metrics on the performance of additively manufactured engineered products is an invaluable, but time intensive specification process. In this work, a comprehensive, generalizable, data-driven framework was implemented to select the optimal powder processing and microstructure variables that are required to predict the target property variables. The framework was demonstrated on a high-dimensional dataset collected from selective laser melted, additively manufactured, Inconel 718. One hundred and twenty-nine powder quality variables including particle morphology, rheology, chemical composition, and build composition were assessed for their impact on eight microstructural features and sixteen mechanical properties. The importance of each powder and microstructure variable was determined by using statistical analysis and machine learning models. The trained models predicted target mechanical properties with an R 2 value of 0.9 or higher. The results indicate that the desired mechanical properties can be achieved by controlling only a few critical powder properties and without the need for collecting microstructure data. This framework significantly reduces the time and cost of qualifying source materials for production by determining an optimal subset of experiments needed to predict that a given source material will lead to a desired outcome. This general framework can be easily applied to other material systems.