Through this proposed SBIR effort, QuesTek Innovations LLC will develop a physics-enhanced machine learning (ML) software to reduce exhaustive typical on-destructive testing or enable feed forward control methodologies for fabrication of high-quality additive manufacturing (AM) component. Process induced defects like keyholing and lack of fusion porosity are shown to be strongly correlated with AM process parameters and powder characteristics. On the other hand, stochastic flaws can often form even with optimized process conditions. The inability to fabricate defect-free, fully functional AM component has resulted in a requirement for tedious inspection and certification steps before final application. To address this problem, QuesTek will develop deep learning (DL) based models to process real-time multi-source sensor data for stochastic defect detection, classification, and localization. In addition, QuesTek proposes to fuse physics-based modeling (i.e., processibility map modeling) for predicting non-stochastic defects into the ML models for a comprehensive defect detection framework. In the Phase I program, proof-of-concept will be demonstrated by developing model framework for defect detection and model framework validation via powder bed fusion (PBF) AM fabrication and post build defect characterization
Benefit: Additive Manufacturing (AM) technology provides increased design freedom in producing near-net shaped components resulting in unique component designs, which are not possible to produce using conventional fabrication processes. Innovative component design for example can improve component performance (ex. improved load-bearing ability or improved heat exchanger ability) and reduce or even eliminate the need for having defect-prone joints. Despite these advantages, several contributing factors have stifled the wide-spread commercial adoption of AM. AM processes and the quality of AM parts are unpredictable relative to traditional manufacturing. Complex defect formation during processing, unclear process-property links and expensive techniques to certify AM parts all suppress the widespread use of AM in critical industries. The current R&D effort which aims to develop an Artificial Intelligence (AI) / Machine Learning (ML) based tool and methodology for identifying defects on additive manufacturing (AM) process that can help solve these challenges. Such a tool, which has the capability of providing, with high fidelity and statistical bounds, predictions on defect formation within the additive process can provide significant values to researchers, materials developers, AM service bureaus and industry OEMs for expediting efficient AM builds / reducing waste, saving both time and money in AM component qualification and commercial adoption. Furthermore, building this capability within an ICME framework differentiates this solution from other technologies currently being developed or marketed, most evidently by allowing for a physics- and material science-based framework that can increase the fidelity of defect detection.
Keywords: Machine Learning, Machine Learning, Processibility Map Modeling, Information Fusion, in-situ monitoring, additive manufacturing, Ti6Al4V Alloys, defect