Machine learning assisted serial sectioning to enable rapid 3D crack network reconstruction

Serial sectioning enables 3D reconstruction of microstructures, providing detailed characterization and insight into processing–structure–property relationships. However, collecting serial section data is time-consuming because many images are required to create reliable reconstructions. In this paper, we investigate the application of machine learning to interpolate intermediate images between reference images, enabling larger cut depths and reducing the total number of collected images. Thus, increasing serial sectioning efficiency and accessibility. Specifically, we describe serial sectioning and a convolutional neural network (CNN) approach with fine-tuning to interpolate crack networks in ζ-Ta 4 C 3- x . The accuracy of CNN-assisted serial sectioning datasets was evaluated using direct image comparison and crack surface area. Two crack types (planar and kinking) were identified to assess the specific capacity of the CNN to recreate complex ζ-Ta 4 C 3- x cracking environments. Results showed that the CNN-generated whole image error was 0.98% with a window size of 129 images and a corresponding cut depth of 3.84 µm. These findings suggest that the serial sectioning time could be reduced by up to 96% with minimal loss in accuracy. By significantly decreasing data collection time, this approach makes serial sectioning more practical for materials characterization and enhances our ability to study material structure and performance at the microscale.