Wei Xu*, Da Ren, Chenchong Wang

The State Key Lab of Rolling and Automation, Northeastern University, Shenyang, 110819, China

EXTENDED ABSTRACT: The establishment of the'composition-microstructure-property'relationship has been the focus of intensive research on designing and optimizing metallic materials. Many methods have been proposed to predict the tensile properties, such as mixture rule, mean-field homogenization approach and finite element method approach, but all suffer from the problem of parameter sensitivity. Although artificial intelligence has advantages in solving the problem of parameter sensitivity, statistical methods based on the "data-data" mode are inaccurate in abstracting microstructure information, and deep learning based on the "image-data" mode can only establish the relationship between the single ­mode microstructure image and property, which ignores the effect of composition and limits the generality of the model. Therefore, in order to solve the above problems, based on the deep learning framework and the concept of multi-mode data coupling, the study proposes a multi-mode data coupling tensile property prediction framework without parameter sensitivity, which covers various compositions, various heat treatment processing routes (obtaining various phase morphologies) and multi-source microstructure images (KAM, BC, Phase). In the framework, multi-modal data coupling is firstly realized, that is, the composition value is normalized and multiplied by each pixel data value in the pixel matrix of each image to obtain a "coupling matrix"; Then, the integration of multi-source microstructure images is realized, that is, after obtaining the "coupling matrix" of the BC map, KAM map and Phase map respectively based on the above coupling method, all the "coupling matrices" are stacked; Finally, the deep learning convolutional neural network model takes the coupling matrices obtained after stacking as input to predict the tensile property. The framework shows high prediction accuracy for the multi­mode data set established in this study. Under the guidance of multi-mode information and multi-source microstructure images, the proposed model can predict the tensile properties accurately based on 23 samples and the R2 above 90%. Further, a deep learning inverse visualization technique, Grad-CAM, is used to generate the heat map, which can reflect key microstructure information that has great impact on property. The visualization technique improves the interpretability of the deep learning model and deepens the physical mechanism. The model proposed in this study can solve the problem of parameter sensitivity and establish a general and universal prediction framework in the large stress (600-1300 MPa) and large strain (2-20%) range. It provides an example for the establishment of the "composition-microstructure-property" relationship in steel materials, which has the potential to be transplanted into other material systems.
Keywords: Multi-mode data; DP steel; Tensile property; Deep learning