Session: 03-11-02: Laser-Based Advanced Manufacturing and Materials Processing II

Paper Number: 167398

Deep Learning for Quantitative Characterization of Microstructure in Laser Powder Bed Fusion 

In laser powder bed fusion (L-PBF), a high-energy laser selectively fuses powder metal layer by layer to build a part. This process generates meltpools as the laser moves along each track, melting the metal powder and creating a solidified track upon cooling. These meltpools form sequentially during both track-to-track and layer-to-layer processing, playing a critical role in defining the microstructure and, ultimately, the quality and mechanical properties of the fabricated part. Precise control and monitoring of meltpool characteristics are therefore essential for achieving desired outcomes in LPBF, as variations can lead to defects such as porosity, lack of fusion, or residual stress. Machine learning (ML) and Deep learning (DL) are increasingly recognized as an intelligent solution for process correction and quality control, enabling more effective defect detection and mitigation.  The objective of this research is to automate the post-process meltpool measurements in 3D-printed metal parts by developing advanced image processing, pattern recognition, and statistical learning techniques. These methods aim to link the process input parameters with the assessed meltpool quality of fabricated parts. To achieve this, a novel deep learning approach, based on mask region-based convolutional neural networks (Mask R-CNN), is employed to detect and segment the meltpool region in optical microscopy images used for post-process evaluation. By identifying the bounding geometry coordinates of the segmented meltpool, the method allows for precise measurement of meltpool size. The proposed approach extends the capabilities of Faster R-CNN by incorporating an additional branch to predict an object mask alongside the standard bounding box recognition. From the results, key post-process signatures such as the length, width, area, and overlap of meltpools can be extracted from the masks and bounding boxes generated by the deep learning model. These measurements provide valuable insights into the quality and consistency of the L-PBF process, helping to assess dimensional accuracy and detect potential defects in the fabricated parts. In conclusion, the Mask R-CNN model represents a significant advancement in both automatic detection of meltpool types (Type I or Type II) and segmentation tasks, leveraging the RPN and a sophisticated proposal classification system. By refining anchors, applying non-maximum suppression, and generating high-quality masks, the model can accurately localize and segment objects in complex images. Through visualization and analysis of intermediate activations, we gain valuable insights into how the model learns and make adjustments to enhance its performance. These components, when effectively combined, enable the model to perform robustly and efficiently in real-world applications of computer vision. Future improvements can focus on optimizing these processes further, exploring alternative architectures, and improving computational efficiency.

Presenting Author: Tugrul Ozel Rutgers University

Presenting Author Biography: Dr. Tuğrul Özel is Full Professor in the Department of Industrial and Systems Engineering and the Director of Manufacturing & Automation Research Laboratory (MARLAB) at Rutgers University- New Brunswick. He received his PhD in Mechanical Engineering from Ohio State University in 1998, BS in Aerospace Engineering from Istanbul Technical University in 1987. His research program mainly focuses on advanced manufacturing, physics-based simulation modeling, process monitoring, machine learning, laser material processing, and metal additive manufacturing. His research has been funded by NSF, DoC NIST, NASA, TTRF and industry. He was consultant for several companies including United Technologies research center and 3M Electronics & Energy. He has published over 200 research articles in journals and conference proceedings that enjoy high number of citations. He is the co-author of four edited books including “Modern Manufacturing Processes” (Wiley 2020) and “Biomedical Devices: Design, Prototyping and Manufacturing” (Wiley 2016). He is the founding Editor-in-Chief of the International Journal of Mechatronics and Manufacturing Systems, Associate Editor of the Journal of Manufacturing Science and Engineering, Journal of Manufacturing & Materials Processing and an editorial board member of Materials Science for Additive Manufacturing and several other journals, and a member of scientific or program committee over 60 international conferences. He recently delivered several keynotes in CIRP general assemblies and international conference on Industry 4.0 and Smart Manufacturing (ISM). He is a senior member of Society of Manufacturing Engineers, American Society of Mechanical Engineers, ASTM International, and a former associate member of CIRP, the International Academy for Production Engineering.

Authors:

Amit Ramasubramanian Rutgers University- New Brunswick
Tugrul Ozel Rutgers University