Session: 04-22-01: Machine Learning-Based Modeling, Prediction, and Optimization of Advanced Manufacturing and Materials System with Multiphysical Phenomena

Paper Number: 165192

Knowledge Transfer of Melt Pool Geometry Between Wire- and Powder-Based Laser-Directed Energy Deposition 

Sharing similar working principles, laser wire- and laser powder-directed energy deposition, i.e., LW-DED and LP-DED, provide avenues to create near-net-shape products while possessing distinct characteristics. As compared to LW-DED, LP-DED has higher geometry resolution and allows precise in-situ alloying, yielding it suitable for creating small, precise products with site-specific materials. On the other hand, LW-DED has extremely high material utilization efficiency, and, unlike powder feedstocks, wire feedstocks are generally cost-effective and easy to acquire, allowing LW-DED to efficiently print large products. However, the difference in the feedstock type yields distinct process parameter spaces, and the associated physical interactions such as laser-feedstock-melt pool interactions in LW-DED and LP-DED are also different. Thus, data from one of the two processes cannot be directly reused to reconstruct the process map or predict deposit characteristics, which leads to tremendous experimental costs when new feedstocks are employed.

This research aims to bridge this gap by proposing predictive models that transfer knowledge from LW-DED (LP-DED) to LP-DED (LW-DED), which can potentially reduce the required amount of data to deploy a new powder (wire) feedstock if the corresponding material has been thoroughly tested as wire (powder) feedstock. To demonstrate this concept, synthetic SS316L LW-DED source data and experimental IN718 LP-DED target data are collected, and although the two datasets have two shared input parameters (laser power and scanning speed), the wire and powder mass feed rates are treated differently to reflect the difference in flow-thermal interactions between the two processes. To bridge this heterogeneity, input mapping calibration (IMC) is applied to align the LP-DED and LW-DED input spaces and integrated with a Gaussian process (GP) regressors trained with LW-DED data to construct a knowledge transfer model called IMC+GP. With this model, LP-DED data can be predicted while the LW-DED data being an auxiliary dataset. Then, it is further integrated into a multi-fidelity GP framework to construct the second knowledge transfer model, namely IMC+MFGP, to allow the model to directly learn from the LP-DED data.

Results show that when the target training set is small, both IMC+GP and IMC+MFGP outperform the baseline GP model (a GP regressor trained only with the LP-DED data), showing an improved data efficiency. The proposed models can also achieve positive knowledge transfer with a small experimental LW-DED dataset. Furthermore, as compared with IMC+GP, IMC+MFGP better suits the scenarios when conservative uncertainty estimates are vital. Finally, the performance of both models is found to be sensitive to the variation of the parameters used in IMC, which can likely be attributed to the nominal relations between the input parameters of LW-DED and LP-DED.

Presenting Author: Kun-Hao Huang Pennsylvania State University

Presenting Author Biography: Kun-Hao Huang is a PhD candidate in Mechanical Engineering at The Pennsylvania State University. His research focuses on predicting laser-based metal additive manufacturing processes using knowledge transfer techniques and computational modeling. Previously, he was a full-time research assistant at National Taiwan University, Department of Mechanical Engineering. He received a BS in Mechanical Engineering, double majoring in Mathematics, in 2016 and an MS in Mechanical Engineering in 2019 from National Taiwan University. At Penn State, he served as a teaching assistant in undergraduate compressible flow and heat transfer. He was also invited to deliver two lectures in the data science in advanced manufacturing class at Penn State.

Authors:

Kun-Hao Huang Pennsylvania State University
Nandana Menon The Pennsylvania State University
Cory Jamieson Applied Research Laboratory, The Pennsylvania State University
Amrita Basak The Pennsylvania State University