Deep learning-based method for microstructure-property linkage of dual-phase steel

Abstract

The establishment of ‘microstructure-property’ relationship is a long-standing topic for metallic materials. For steels with complex phases and morphologies, it is often difficult to identify the key microstructural constituents for the accurate prediction of tensile properties. Here, instead of traditional methods that extract specific information from microstructure images (e.g. grain size and volume fractions), a deep learning (DL) strategy using microstructure images as direct inputs, is implemented to predict the tensile property of dual-phase (DP) steel. Compared with traditional physical models, the proposed method can help overcome the difficulty in quantifying complex microstructural information. Additionally, this method is based on a small sample database, which ensures its high portability and applicability. An important visualization heat map analysis is also used to quantitatively identify the key microstructural factors that influence tensile properties, and further improves the explicability of this method.