Development of AI crack segmentation models for additive manufacturing

Abstract

The use of X-ray computed tomography (XCT) has seen significant growth over a broad range of disciplines including biology, earth science, engineering, and many more. It is now increasingly used in additive manufacturing (AM) since its benefits are being appreciated more widely. This is due to the method being non-destructive and comprehensive, providing external and internal information of tested parts. Data processing and segmentation of XCT data is important to get as much information as possible so that a clear picture of features can be obtained and analyzed. Porosity analysis has been the most successful and widely used XCT analysis type in all fields so far, partly due to simple manual segmentation methods such as the Otsu global threshold. However, segmentation of small and narrow features such as cracks are challenging with conventional thresholding methods. Since automated conventional methods fail, manual segmentation is often used but this can be subjective, tedious, and prone to segmentation errors. The present work employs neural networks, specifically the U-Net architecture and thoroughly investigates possible solutions to a robust crack segmentation model. Intensity scale calibration, bias training weights and data augmentations were investigated in detail to find the best possible performance of trained models, when employed on new data. The results demonstrate the performance and improvement gained by each of the above factors, as well as the successful AI segmentation for various additively manufactured sample types with different cracks. This method enables clear visualization and presentation of cracks, as well as their quantification. The model strives toward a generic crack segmentation model for all AM parts that could be used directly by others. This generalizability of the model is discussed together with its limitations.