Aaron P. Stebner
Colorado School of Mines, Rowlinson Assistant Professor of Mechanical Engineering and Materials Science
"Platforms for High Throughput Structure-Property Characterizations to Support Machine Learning Approaches to Additive Manufacturing"
Abstract: The Alliance for the Development of Additive Processing Technologies (ADAPT) is developing and validating an Artificial Intelligence (AI) platform that is capable of learning physical Process-Structure-Property (PSP) models for Additive Manufacturing (AM). Our approach is “outside of the box” of traditional physics-based modeling of PSP relationships, as we have found that the number of physical Degrees of Freedom (DOF) in AM is tremendous. For example, in selective laser melting (SLM) of Inconel using a Concept Dual-Laser machine, there are over 100 PSP DOFs - an overwhelming challenge for even the most revolutionary multi-physics-based approaches. It is also a great challenge to identify the critical DOFs for variation in AM using only human discernment. Thus, we are coupling the latest advancements in Machine Learning (ML) together with state of the art Cloud-based computing and materials database infrastructure. An added benefit of our approach is the inherent statistical core of our models – the AI platform is not only capable of learning the PSP relations of AM, but also providing statistical reports on part quality and variation in a manner conducive to certification.
One of the initial research thrusts within this center has been optimizing AM parameters for certification of SLM Inconel 718 parts and wire-fed, electron-beam printed Ti-64 parts. We populate the database using a state-of-the-art R&D laboratory dedicated to advanced, high-throughput PSP characterizations of AM metals, including sub-micron-resolved computed tomography (μCT), diffraction contrast tomography (DCT) thermomechanical testing, 3D surface metrology, and optical microscopy. With these resources, are analyzing over 6,000 Inconel 718 specimens to assess statistical process-structure-property correlations toward optimizing the AM processes. In this presentation, we will present the underlying computational framework and algorithms we have developed toward these means – specifically algorithms for autonomous structural feature detection in mechanical test, μCT, and metrology data, and also the use of Matrix Completion Machine Learning algorithms to accelerate test plans for AM parts.
To view the full schedule, please visit: http://physics.mines.edu/PH-physics-colloquia