Session: Rising Stars of Mechanical Engineering Celebration & Showcase

Paper Number: 148623

148623 - Accelerating Process Mapping in Metal Additive Manufacturing 

The goal of this research is to develop advanced data-driven techniques informed by physics to expedite the generation of both forward and inverse scaling laws, applicable across various material systems and manufacturing processes, particularly in metal additive manufacturing (AM). Laser powder bed fusion (L-PBF) and laser directed energy deposition (L-DED) are particularly pertinent to processing metals. These processes offer distinct advantages over traditional manufacturing methods, including reduced lead time, material efficiency, and the ability to fabricate intricate geometries. For instance, L-PBF is ideal for creating heat exchanges with complex channels, while laser-wire DED (LW-DED) is more cost-effective for manufacturing large components like pressure vessels. Laser-powder DED (LP-DED) is suitable for repairing legacy components or producing functionally graded parts. However, qualifying L-PBF and L-DED processes is both costly and time-consuming, largely due to the extensive experimentation or computational modeling required to develop scaling laws for new material systems, manufacturing processes, and equipment configurations.

To address this challenge, we will integrate information from high and low-fidelity simulation models, along with a limited amount of experimental data, using probabilistic machine learning (ML) techniques to construct a multi-fidelity (MF) surrogate. This surrogate will serve as a predictive tool for desired properties, such as melt pool area and aspect ratio, as a function of process parameters at discrete time intervals, facilitating the development of temporal scaling laws. By employing a probabilistic surrogate, the scaling laws will inherently account for uncertainties. Subsequently, we will utilize transfer learning methodologies to transfer these scaling laws across different material systems and manufacturing processes.

In our preliminary investigation, we explored various knowledge transfer methods utilizing Gaussian process (GP) techniques. Specifically, we utilized melt pool data for SS316L and IN718 alloys to emulate both data-rich and data-scarce conditions. We pursued three distinct avenues: (i) employing the mixed-input method, which involves training a single GP regression model using combined data from both alloys, (ii) utilizing the relation-based transfer learning (RB-TL) method, and (iii) implementing the multi-fidelity GP-based transfer learning (MFGP-TL) method. Our findings reveal that the mixed-input model outperforms the baseline or no-transfer model in data-deficient conditions. Moreover, the RB-TL model demonstrates an overall improvement in accuracy and confidence compared to the baseline model, while also requiring the least computation time among the proposed models. Notably, the MFGP-TL model exhibits the most superior performance, with significantly reduced error and standard deviation compared to the RB-TL model, albeit at the expense of longer computation times. This research thus enables the cost-effective and data-efficient transfer of uncertainty quantification (UQ)-based knowledge regarding process-structure relationships in L-DED processes.

Presenting Author: Amrita Basak Pennsylvania State University

Presenting Author Biography: Amrita is an Assistant Professor and Shuman Professor of Mechanical Engineering at Pennsylvania State University – University Park. At Penn State, Amrita’s research group focuses on understanding the fundamental processing-structure-property relationships in advanced manufacturing of high-performance metallic alloys. Additionally, her group actively collaborates with other research groups within and outside of Penn State to understand such relationships in ceramic, construction, and polymeric materials.

Authors:

Kun-Hao Huang Pennsylvania State University
Nandana Menon Pennsylvania State University
Amrita Basak Pennsylvania State University