Defect prediction in laser powder bed fusion with the combination of simulated melt pool images and thermal images

In laser powder bed fusion (LPBF) process, defects such as lack of fusion, porosity and keyhole originating from the layer-wise material deposition process, hinders its applications due to the absence of an effective prediction method. Thermal images of the melt pool are the most informative process signatures for in-process monitoring. Meanwhile, numerical simulation provides a better understanding of the melt pool fluid dynamics, the temperature field, and the quality of LPBFed metallic parts. In this paper, the thermal images are acquired from the high-speed thermal camera, and the simulation results from computational fluid dynamics (CFD) modelling. A hybrid neural network (HNN) model is proposed to fuse the data science features of thermal images and the physical knowledge features of simulated melt pool images for defect prediction in LPBF. The combination of features improves the defect identification accuracy to 97.25 %. Furthermore, a feature correlation algorithm employs the simulated melt pool features to improve the thermal image features, and replaces the physical feature extractor in the HNN. This physical supervision network (PSN) model saves the computational cost of CFD simulations. The high accuracy of 96 % of PSN model demonstrates that it is a promising method for defect identification in LPBF.