Nature-to-microstructure network: A cross-domain knowledge transfer framework for steel microstructure classification

Abstract

Reliable microstructure identification underpins the control of microstructural evolution and the tailored design of mechanical performance in steels. However, conventional approaches are limited by subjectivity and inefficiency, while deep learning models often struggle with data scarcity and multiscale complexity. This study proposes Nature-to-Microstructure Network (N2M-Net) a cross-domain transfer learning (TL) framework leveraging morphological similarities between natural textures and steel microstructures. It is designed to enable accurate microstructure classification under extreme data scarcity and to enhance model generalization through morphology-guided knowledge transfer. Pretrained on 750 natural images, N2M-Net incorporates residual connections, spatial attention, and dilated Inception modules for robust multiscale feature extraction. With only 20 microscopic images, it surpassed 95.00 % accuracy within 20 epochs and reached 95.67 % final accuracy, significantly outperforming direct training. This work demonstrates a novel morphology-driven cross-domain transfer paradigm for intelligent microstructure analysis under limited data.