A deep learning method for extensible microstructural quantification of DP steel enhanced by physical metallurgy-guided data augmentation

The small sample problem caused by time-consuming annotation has greatly restricted the development of deep learning (DL) methods with high generality and extensibility, hindering the wide application of DL-based microstructural analysis. To address this problem, an extensible microstructural quantification method is proposed by combining physical metallurgy (PM)-guided data augmentation and DL. In this method, PM-guided data augmentation, which combines key PM information highly related to microstructural evaluation and random image transformation, is used to generate microstructures with other processes and therefore construct a comprehensive dataset. The proposed method is successfully applied and validated to dual-phase (DP) steels to classify martensite/ferrite phases and quantify their fractions and grain sizes at various temperatures based on only two SEM images of annealing temperatures of 750 and 780 degrees C, in which a comprehensive dataset covering the entire dual phase regime is constructed with guidance from the content and morphology of the phase and dissolved carbide during annealing. Moreover, the good extensibility and data independency of the present method is demonstrated by various DP steels with different compositions and processes collected from the literature.

Automated summary

A method for microstructural quantification combining physical metallurgy-guided data augmentation and deep learning is proposed, successfully applied to dual-phase steels for phase classification and quantification under different temperature conditions.