Study on real-time prediction for porosity in laser melting based on a digital twin-driven deep learning framework

The keyhole-induced pore defect is recognized as a major nuisance in laser melting (such as welding and 3D printing). As a crucial indicator for assessing the specific severity of pores, porosity has long been obtained only through complex and offline internal inspection after manufacturing. In this paper, guided by the formation mechanism of pores revealed via numerical simulation, a digital twin-driven deep learning framework (DTDLF) is proposed that enables direct and real-time porosity prediction during laser beam welding. In the perception field, keyhole depth and keyhole opening circumference are acquired through multi-sensing measurement and signal processing. In the physical field, simulation can acquire maximum temperature on the bottom wall of the keyhole. Subsequently, a sparse auto-encoder is constructed to fuse the original fluctuation characteristics mined from their respective digital twin data sources, and all of the obtained overall features are demonstrated to be highly correlated with porosity. In the DTDLF, a tandem structure of two deep neural networks (DNNs) connected in series is adopted, where the former DNN conducts online prediction of the fused overall feature in the physical field utilizing process parameters to circumvent the long simulation runtime, and the latter DNN couples all the fused overall features in the perception-physics fields for predicting porosity online. The final results indicate that the DTDLF has a significant accuracy advantage over monitoring frameworks that use solely single-field information, and meanwhile its real-time performance is not compromised. © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.