In-situ surface porosity prediction in hybrid-directed energy deposition (DED) process using explainable multimodal sensor fusion

Porosity remains one of the most critical defects in Directed Energy Deposition (DED), due to the complex thermo-mechanical interactions within the melt pool, directly limiting the reliability of additively manufactured parts. Conventional inspection methods such as X-ray computed tomography scans are expensive, time-consuming, and unsuitable for real-time intervention, while in-situ thermal imaging often lacks the resolution and sensitivity to capture limited pore formation pathways. These limitations highlight the industrial need for real-time, mechanism-aware porosity detection to enable closed-loop quality control, especially in hybrid-DED systems where sequential printing and machining cycles introduce additional complexity. In this study, synchronized multimodal sensing and explainable deep learning are employed to investigate pore formation at sub-millimeter resolution surface elements, referred to as surfels . Acoustic emission (AE), accelerometer, and thermal imaging data were synchronized to 0.1 ms temporal and 0.5 mm spatial resolution during hybrid-DED of 316L stainless steel using an Optomec–LENS® MTS 500 system. Time–frequency features extracted from these signals were used to train a convolutional neural network (CNN), which achieved a classification accuracy of 87 % in distinguishing porous from nonporous surfels. Our results indicate that sensor data from the printing cycle are more effective for pore detection than those from machining. Furthermore, explainable AI analysis with local interpretable model-agnostic explanations (LIME) revealed that high-frequency AE signatures differentiate lack-of-fusion voids, caused by insufficient inter-track overlap, from spatter-induced porosity. Porous surfels exhibited ∼37 % lower spectral energy in high-frequency AE bands compared to sound regions, while spatter-induced porosity showed ∼27 % higher energy relative to lack-of-fusion voids. These findings demonstrate that the integration of multimodal sensor fusion with explainable AI enhances porosity detection accuracy while uncovering key underlying pore formation mechanisms. The proposed framework offers a practical route towards physics-informed, real-time monitoring and control of hybrid-DED processes, with significant implications for improving build quality and reducing dependence on expensive post-process inspections. • Acoustic Emission (AE) time-frequency patterns reveal two pore modes (spatter-induced and lack-of-fusion) in Hybrid-DED. • AE mapping classifies each 1 mm surface as porous or non-porous with pore-cause identification. • AE signals show strongest pore signatures during Hybrid-DED printing, not machining.