Composition-process-structure linkages in metal additive manufacturing
Restricted (Penn State Only)
- Author:
- Menon, Nandana
- Graduate Program:
- Mechanical Engineering (PHD)
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- March 21, 2024
- Committee Members:
- Robert Kunz, Professor in Charge/Director of Graduate Studies
Amrita Basak, Chair & Dissertation Advisor
Saurabh Basu, Outside Unit & Field Member
Zoubeida Ounaies, Major Field Member
Chris Rahn, Major Field Member
Sudeepta Mondal, Special Member - Keywords:
- Metal additive manufacturing
Laser-Directed Energy Deposition
Melt pool modeling
Composition-process-structure linkages
Machine Learning - Abstract:
- Metal additive manufacturing (AM) finds extensive use in the design, fabrication, and repair of high-performance components within the aerospace and defense industry. Given the intricate nature of this process, involving a diverse range of parameters and alloys, reliable composition-process-structure relationships need to be developed to fully harness the capabilities of metal AM. However, depending solely on experiments to develop such relationships is impractical due to the associated high production costs, numerous trials, and post-processing demands. While physics-based models offer valuable alternatives, they involve extensive calibration and a precision-computation trade-off. As the industry shifts towards a smart manufacturing paradigm, data-driven approaches enable the creation of digital twins. These twins facilitate the swift exploration of diverse process parameters and the efficient development of vital relationships. This dissertation focuses on developing composition-process-structure relationships by leveraging mechanistic and statistical learning to accelerate various facets of metal AM. A co-design framework employing Gaussian process (GP) regression is used to establish meaningful links between composition, process parameters, and the underlying structure in a multi-step surrogate modeling approach. This framework is harnessed for the prediction of steady-state melt pool depths with uncertainty quantification. Furthermore, a GP surrogate is adapted to address challenges posed by sparse data when predicting precipitates and phases, accomplished through the incorporation of hybrid modeling techniques. Leveraging GPs for surrogate modeling proves invaluable in situations with limited data availability. To enhance its versatility, robust multi-fidelity modeling is introduced, enabling the fusion of data from diverse sources with varying levels of accuracy. This approach diminishes reliance on a singular data source and upholds predictive accuracy even when operating within a constrained computational budget. The dissertation also introduces transfer learning- another machine learning tool to overcome the isolated learning paradigm by reusing the knowledge learned from developing process-structure relationships of one alloy for another. The final part of this dissertation addresses the relatively unexplored problem of process optimization via reinforcement learning using a model-free algorithm, Q-learning. The framework achieved optima close to experimental observations. In summary, this dissertation provides avenues for leveraging machine learning to optimize and customize AM technologies. It introduces data-driven aimed at improving efficiency and lowering costs associated with metal AM processes.