Deep learning-based melt pool and porosity detection in components fabricated by laser powder bed fusion

Microstructure analysis is a crucial aspect of additive manufacturing (AM) processes, as it offers valuable insights into material properties, defects, and quality of printed parts. Quantification of microstructural features such as melt pool dimensions and porosity can help optimize the process parameters by aiding the development of useful correlations of the processing conditions and resulting print quality and properties. However, detecting melt pool boundaries and porosity-related defects in cross-sectional microstructure samples presents significant challenges due to the complex nature of these features and inherent difficulties in acquiring their images with quality and quantity required for accurate and efficient detection. To address this, we propose a deep learning-based approach that leverages state-of-the-art backbone models (EfficientNet b7 and DenseNet 201) with various convolutional neural networks (U-Net, LinkNet, and FPN) using transfer learning techniques to automatically segment and detect the melt pools and porosity from AM microstructure images. The results demonstrate our ability to accurately identify and segment melt pools and porosity, even with limited set of training data. A comparative study of the performance of the different neural network architectures was done and it was found that the quantification of results (of microstructural features) from the employed set of networks are statistically comparable although the accuracy from the combination of U-Net with EfficeintNet b7 backbone was highest.