Date of Award
8-1-2025
Degree Name
Master of Science
Department
Mechanical Engineering
First Advisor
Jung, Sangjin
Abstract
Ensuring dimensional precision in parts with complex overhangs is a significant concern in Metal Additive Manufacturing (MAM), as undetected geometric deviations can compromise functionality and reliability. This research introduces a Conditional Generative Adversarial Network (cGAN), specifically the Pix2Pix framework, to predict layer-wise geometric deviations in Laser Powder Bed Fusion (LPBF) printed parts with overhang geometries using paired 2D CAD slices and corresponding X-ray Computed Tomography (XCT) based ground truth images. A key innovation is using RGB color-coded CAD slices to encode overhang angle information, enhancing feature distinction and prediction accuracy compared to non-color-coded inputs. Eighteen Pix2Pix models were trained across three data groups and varying batch sizes, with performance evaluated using PSNR, SSIM, LPIPS, FID, and a novel Edge IoU metric for edge preservation. The results demonstrate that color-coded models achieve higher accuracy and training stability. A generalized model trained on a balanced dataset of seven overhang geometries also effectively predicted deviations in unseen designs. This framework aids early stage deviation prediction, guiding Design for Additive Manufacturing (DfAM) by reducing trial-and-error cycles, material waste, and support dependency.
Access
This thesis is Open Access and may be downloaded by anyone.