Automated grain analysis via data augmentation and grain boundary detection

This study presents an AI-enhanced framework to address key challenges in the quantitative metallographic analysis of pure iron systems. Manual grain size characterization suffers from limited efficiency and reproducibility, while existing computational methods are constrained by scarce data and incomplete grain boundary detection. To overcome these issues, we propose three core innovations. First, a three-stage data synthesis pipeline is developed, incorporating stochastic grain mask generation via denoising diffusion probabilistic models (DDPM) and microstructural translation using a conditional adversarial network, enabling the generation of physically consistent metallographic images. Second, a deep neural network based on a U-Net architecture is trained on a paired and reconstructed dataset, where the topology-awareness emerges from the data pairing and reconstruction objectives. Third, a fully automated grain analysis system is established, based on whole-grain area quantification and twin-grain merging strategies. The proposed methodology effectively resolves longstanding limitations in metallographic analysis related to data scarcity and subjectivity in manual evaluation.