Building a quantitative composition-microstructure-property relationship of dual-phase steels via multimodal data mining

The establishment of composition-microstructure-property relationship is a long-standing topic in materials science, yet neither continuum mechanics approaches nor machine learning methods have ever established a generic model that can cover a wide spectrum of stress-strain relationships for metallic materials and, in particular, for steels having complex microstructures. In this study, a deep learning framework that handles a multimodal database, including composition and multi-source microstructure images, is developed for predicting the tensile properties of dual-phase steels. The model exhibits excellent generality for that it enables the coupling analysis of multimodal data. Moreover, the integration of multi-source microstructure information significantly enables the model to be compatible with multiple mechanisms and achieves accurate prediction in the large stress-strain range. Finally, the reverse visualization applied in the proposed framework deepens the mechanism understanding of strain distribution under different phase morphologies, which greatly improves the explicability of the deep learning model. This framework provides useful guidance for the prediction of composition-microstructure-property relationship in complex steel systems and can be potentially applied to other alloys.