DeepMMP: Efficient 3D perception of microstructures from Serial Section Microscopic images

Driven by rapidly increasing volumetric data from serial section microscopy (SSM), deep learning (DL) has been reshaping the research of quantitative materials microstructural analysis and structure-property relationship mining. However, extracting or reconstructing materials microstructures via DL -based image segmentation remains reliant on laborious annotations, and the three-dimensional (3D) spatial information within these volumetric data has yet to be fully exploited. In this work, we introduce DeepMMP, a novel DL -based framework that achieves efficient materials microstructure perception by leveraging the inherent continuity and similarity learned from consecutive SSM images. DeepMMP incorporates unsupervised annotation-free feature learning, 3D image voxel segmentation, and 3D microstructure reconstruction, enabling the effective, efficient, and accurate generation of 3D segmentation and reconstruction results for complex polycrystalline materials with minimal annotations. We believe DeepMMP has the potential to accelerate both process parameter optimization and the design of novel materials.