Speaker
Description
Parts produced by Powder Bed Fusion with Laser Beam of Metals (PBF-LB/M) are susceptible to thermally induced defects during the manufacturing process. The thermal history of parts can be assessed through Optical Tomography (OT), an in-situ monitoring technique that captures layer-wise integrated near-infrared emission intensity during the build process. Acquiring reference OT images in qualification requires time- and cost-consuming physical test build jobs. Predicting OT images directly from scan vector metadata (e.g. scan orientation, scan order, laser parameters) enables prospective process evaluation without requiring physical builds. This offers a pathway toward data-driven defect detection, reducing the need for dedicated test build jobs.
To achieve this, a conditional Generative Adversarial Network (cGAN) based on the pix2pix framework is used. Vector metadata needs to be rasterized into spatially structured input representations to provide rich context for the prediction task, though richer metadata representations increase the computational cost of training and inference. In this work, we present an ablation study on the relevance of rasterized vector metadata features for layer-wise OT image prediction. Through systematic ablation of individual and combined metadata channels, we quantify the contribution of each feature to prediction accuracy and assess whether the full set of available vector metadata is necessary for accurate OT image estimation. Our results provide guidance on the minimum sufficient input representation for efficient and accurate OT image prediction in PBF-LB/M.
| Speaker Country | Germany |
|---|---|
| Would you like to publish your paper in the special issue of BHM "Berg- und Hüttenmännische Monatshefte" | Yes |