Detection of random porosity using high-resolution in-situ monitoring and AI-based data mining in metal laser powder bed fusion

Metal additive manufacturing (AM) is increasingly adopted for high-value industrial components, making in situ monitoring and quality assurance central to its wider deployment. Porosity, particularly lack-of-fusion (LoF) defects, critically degrades parts produced by laser powder bed fusion (LPBF). Post-process X-ray computed tomography (CT) provides accurate defect characterization, but is costly, time-consuming, and unsuitable for corrective actions during fabrication. This work demonstrates layer-wise prediction of randomly occurring LoF porosity in LPBF parts, targeting defect formation under realistic, industry-relevant process conditions. The framework combines high-resolution in situ sensing, CT-based ground truth, and multi-layer spatio-temporal modelling. Layer-wise RGB images were acquired by ScanIT, a recoater-mounted contact image sensor that captures post-exposure and post-recoating layers without interrupting the LPBF process. CT porosity reference labels were aligned with the in situ data through a deformation-aware methodology. Single-stream networks processing either post-exposure or post-recoating images were compared with dual-stream networks that perform feature-level fusion of both sources. The best-performing Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) dual-stream model combined spatial feature extraction and temporal modelling, achieving 89% accuracy and precision, recall, and F1-scores between 86 and 88%. Post-exposure images were highly informative, whereas post-recoating images alone underperformed, particularly for small defects. Probability of detection analysis showed reliable detection of pores above approximately 110 µm. Integrated gradients indicated that predictions were driven by physically meaningful features, including spatter and powder-bed irregularities. Overall, the framework demonstrates layer-wise prediction of random LoF porosity and supports the transition from post-process inspection toward in situ quality assurance in metal LPBF.