Röntgen-CT-ondersteunde datagedreven in-procesmonitoring voor metaal- Laser Powder Bed Fusion

Laser Powder Bed Fusion of metals (PBF-LB/M) promises design freedom for critical aerospace and biomedical parts, yet the cost and difficulty of post-build certification constrain its industrial adoption. This thesis advances qualify as-you-build manufacturing by developing high-fidelity, data-driven surrogate models that transform in-process monitoring data into real-time predictions of sub-surface pores and surface dimensional errors. A co-axial optical monitoring system comprising a photodiode and an NIR camera captured the laser-powder interaction, with X-ray computed tomography providing ground-truth labels. For binary pore detection, a hybrid CNN-LSTM model that incorporated inter-hatch and inter-layer remelting effects achieved an area under the Receiver Operating Characteristic curve (AUROC) of 0.91 for pores exceeding 8000 µm3. To predict keyhole pore, physics-informed melt pool features integrated into a Random Forest model (AUROC = 0.95) outperformed deep learning benchmarks, delivering superior accuracy at substantially lower computational cost. In lack-of-fusion porosity regression, the CNN-LSTM surrogate model achieved high accuracy in within-scenario predictions (RMSE = 2.1 %). Domain adversarial transfer learning enhanced cross-scenario performance by up to 35 %, highlighting the potential of the surrogate model to generalize across part positions and process parameters. The hyperparameter associated with inter-layer remelting was further validated against the ratio of melt pool depth to layer thickness, ensuring alignment with process physics. Surface geometric deviation prediction was advanced through a CNN-Transformer architecture that correlates melt pool signatures with CAD referenced 3D voxel volumes, achieving RMSE values of 6.1 µm (up-skin) and 27.4 µm (down-skin). Model interpretability was provided through Grad-CAM visualizations and Transformer attention maps, highlighting critical spatial regions and temporal dependencies governing defect formation. Collectively, these contributions support the vision of first-time-right manufacturing in PBF-LB/M. By strengthening defect localization, improving predictive fidelity, and enhancing generalization, this thesis lays the foundation for digital product passports, feedback-loop control, and broader industrial adoption of in-situ quality assurance.