Multi-task physics-constrained dictionary learning for efficient estimation of porosity distribution in laser powder bed fusion of copper

The high porosity, as a primary defect in the laser powder bed fusion (LPBF) process for highly reflective metal components, significantly restricts the broader application of LPBF. However, existing pore detection methods primarily focus on classifying individual pores, offering limited insight into optimizing printing parameters. Additionally, these methods often overlook the storage and processing challenges associated with the large volumes of image data collected. Therefore, this paper introduces a multi-task physics-constrained dictionary learning approach that simultaneously compresses and estimates the porosity distribution in metallographic images of copper components produced by LPBF. Specifically, a physics-constrained label-consistent dictionary learning (PC-LCDL) algorithm is proposed for compressing images into discriminative sparse vectors. The pixel characteristics of low-resolution images are incorporated as a physical constraint, enabling the reconstruction of high-resolution images from the low-resolution ones. Hence, image acquisition efficiency can be improved. Moreover, a residual-based graph sample and aggregate (GraphSAGE) algorithm is integrated with the PC-LCDL to estimate the porosity distribution in the copper images. To thoroughly extract the distinctive features of pores, the reconstructed image patches concatenated with the sparse vectors are fed into the classifier. Experimental results demonstrate that even at a high compression ratio of 4.9, clear images can still be reconstructed from blurry ones which are down-sampled at a rate of 12.25 %. Consequently, a classification accuracy exceeding 89 % is still achieved, outperforming many other classification methods. Furthermore, the impact of printing parameters on porosity distribution is also investigated, leading to recommendations for adjusting printing parameters to minimize porosity levels.