About this Abstract |
Meeting |
2025 TMS Annual Meeting & Exhibition
|
Symposium
|
Additive Manufacturing Modeling, Simulation and Machine Learning
|
Presentation Title |
Quantification of Defects in Additively Manufactured Steel Using Unsupervised Machine Learning |
Author(s) |
Hariharan Chithalamangalam Saravanan, Pooja Maurya, Alex Gaudio, Asim Smailagic, P. Chris Pistorius |
On-Site Speaker (Planned) |
Pooja Maurya |
Abstract Scope |
Binder jet printing is gaining potential due to its ability to print a wide range of materials. Sintering is one of the important post-processing steps needed to densify the printed part. Defects like oxide inclusions, porosity and carbides incorporated in the sintered part can deteriorate its mechanical properties. Methods to detect the carbide precipitation, due to incomplete debinding, have been developed. Images of the etched sintered samples captured by optical microscopy reveal various features, one of which is the carbides precipitated along the grain boundaries. However, the use of computer vision techniques to detect and quantify the carbides has not been studied. This project quantifies the carbides from such optical images and comparatively analyzes the variation of these carbides with the heat treatment after sintering. An image processing pipeline is proposed to isolate carbide pixels from non-carbide pixels in microscopy images. Different pre-trained deep networks are utilized in the main pipeline stages: an edge detection model selects candidate carbide pixels from a high-resolution input image; a blob detector removes pore artifacts from candidate pixels; and a line detector removes the twin boundaries from the remaining candidate pixels. The results of the proposed pipeline suggest that the amount of carbides is invariant to the heat treatment time after the sintering. |
Proceedings Inclusion? |
Planned: |
Keywords |
Additive Manufacturing, Characterization, Machine Learning |