UC San Diego

Laser powder bed fusion (LPBF) enables fabrication of complex metal components, but its coupled multiphysics drives multiple defect-formation mechanisms. Porosity is a critical defect because it forms through transient, spatially localized process events and can degrade mechanical reliability. Localized in-situ pore detection is therefore essential for resolving conditions associated with individual pore formation, advancing understanding of LPBF process–structure relationships, and supporting more reliable process monitoring and qualification.This dissertation develops a Multimodal Sensor Fusion Machine Learning (MSFML) framework for localized LPBF pore prediction by integrating heterogeneous sensor data with process- and physics-informed knowledge. The proposed Parallel Multi-Layer Multimodal Sensor Fusion (PMMSF) method progressively fuses one-dimensional sensor signals, image data, photometric-stereo products, process parameters, physics-derived features, and multi-layer context.The experimental system integrated three thermal-emission channels, acoustic emission, laser-triggered acquisition, coaxial melt-pool imaging, and off-axis photometric stereo imaging during the fabrication of nominally dense coupons under varying laser power and scan speed conditions. A comprehensive data preparation workflow was developed to synchronize and register these heterogeneous measurements. Post-build X-ray computed tomography (XCT) was used to establish three-dimensional pore ground truth and link individual pores to their corresponding in-situ process histories.Single-sensor models, multimodal sensor combinations, fusion strategies, process-informed weighting, and physics-informed learning were evaluated. The full multimodal framework with process-informed weighting and physics-informed features achieved the best performance, with 95% test accuracy, 97% AUC, and an F1 score of 87%. F1 improved by 10 percentage points over the best single-sensor model, 5 points over the best two-sensor model, and 3 points over the best three-sensor model. Process-informed weighting and physics-informed features each provided an additional 2-point F1 improvement.In summary, this dissertation establishes a framework that advances localized in-situ pore prediction by integrating multimodal sensing with process- and physics-informed machine learning. The framework enables investigation of transient pore-formation signatures and LPBF process–structure relationships while providing a generalizable approach for sensor-based monitoring of other complex manufacturing processes.