Ductile damage results from the growth and coalescence of micro-voids and micro-shear-cracks, with a transition between these mechanisms. This paper investigates anisotropic ductile damage and its stress-state-dependent transition using scanning electron microscopy (SEM) images at the microscopic level combined with deep learning. Conventional optical observation, thresholding, and feature-based machine learning show apparent limitations. Deep learning with DeepLabv3 was evaluated using different backbone architectures on annotated SEM datasets of varying sizes. Comparisons with traditional recognition methods and different dataset sizes demonstrate the advantages of deep learning, particularly in resolving fine boundaries and discriminating defects under heterogeneous contrast. SEM data were obtained from fracture surfaces of novel biaxial experiments covering a wide range of stress states and used for training and validation. Numerical simulations were conducted to determine the stress states for the selected biaxial loading conditions. The transition of stress-state-dependent microscopic damage behavior was then investigated by combining numerically predicted stress states with deep learning-based recognition of micro-voids. Overall, deep learning provides an effective and efficient solution for quantitatively characterizing fracture-surface microstructures. The proposed workflow reduces annotation costs, operates on standard computing resources, and enables automated analysis of stress-state-dependent damage and transitions in ductile metals. The transition from micro-shear-crack-dominated to micro-void-dominated damage behavior is independently identified based on stress triaxiality, with a critical value of 0.38, and based on the Lode parameter, with a corresponding critical value of − 0 . 45 . The calibrated critical transition stress states provide a solid foundation for further development of the advanced continuum damage model. • SEM dataset from novel biaxial experiments covering diverse stress states. • Deep learning enables automated detection of micro-voids in SEM images. • Thresholding, ML, and deep learning differ in principles, efficiency, and accuracy. • DeepLabv3 offers high accuracy with small datasets and runs efficiently on a laptop GPU. • Quantitative analysis of transitions in stress-state-dependent anisotropic damage.
Wei et al. (Sun,) studied this question.