About this Abstract |
| Meeting |
MS&T26: Materials Science & Technology
|
| Symposium
|
Additive Manufacturing Modeling, Simulation, and Machine Learning: Microstructure, Mechanics, and Process
|
| Presentation Title |
CNN Encoder–Decoder Segmentation of Fatigue-Relevant Surface Porosity in LPBF Ti-6Al-4V |
| Author(s) |
Jay Choi, Kevin Zhou, Evan Adcock, Joseph Pauza, Austin Ngo, John Lewandowski, Anthony D. Rollett |
| On-Site Speaker (Planned) |
Jay Choi |
| Abstract Scope |
Fatigue performance of laser powder bed fused Ti-6Al-4V is strongly affected by process-induced pores exposed after machining. Prior optical-microscopy analysis of four-point-bend specimens showed that surface pore density and maximum pore size are anti-correlated with runout stress and fatigue life, but threshold-based segmentation requires manual tuning to distinguish pores from machining marks. This work extends that dataset by developing a convolutional autoencoder-decoder for automated pore segmentation in bright-field optical micrographs. Trained on verified pore masks, the model is evaluated across keyhole, process-window, and lack-of-fusion regimes against the existing Otsu/regionprops workflow for pore detection, equivalent spherical diameter, aspect ratio, and the N1 statistic estimating the largest pore in the high-tensile-stress region. By reducing user-dependent segmentation while preserving fatigue-relevant defect statistics, this approach aims to improve rapid screening of LPBF process windows and support image-based prediction of fatigue-sensitive surface defects in machined Ti-6Al-4V components. |