About this Abstract |
Meeting |
2022 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2022)
|
Symposium
|
2022 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2022)
|
Presentation Title |
Graph Neural Networks for 3D Defect Mapping in Laser Powder Bed Fusion |
Author(s) |
Sebastian Larsen, Paul A. Hooper |
On-Site Speaker (Planned) |
Sebastian Larsen |
Abstract Scope |
Since defects form due to the stochastic nature of the LPBF process, stringent post-build inspection regulations are a required burden. However, locating defects in-situ would enable a component to be qualified in real time, automating this requirement.
A graph neural network model was developed to provide a geometric invariant method for localising defects in the material. The model was trained on high-speed melt pool monitoring data, collected from a component manufactured with seeded defects.
A k-fold cross validation was performed where each seeded defect was detected and localised. The defect sizes were correlated with the 3D probability map which showed positive correlation with number of detections.
Graph neural networks provide an efficient way to locate defects in a component, while remaining invariant to geometry. We believe the effectiveness comes from incorporating this physical structure into a machine learning model. |
Proceedings Inclusion? |
Definite: Post-meeting proceedings |