A unified framework for microstructure image recognition and performance prediction of energy materials based on multi-scale deep learning is proposed to break through the efficiency bottleneck of traditional "trial and error method" in material research and development. Aiming at the problems of multi focus single scale structure recognition and structural performance modeling fracture in existing methods, a dual branch convolutional neural network (CNN) is constructed to extract local texture and global morphological features of scanning electron microscopy (SEM) images, and a convolutional block attention module (CBAM) is introduced to achieve adaptive feature fusion; Further design a multi task learning architecture, jointly optimize microstructure classification and performance regression objectives, and enhance feature physical interpretability. The experiment was based on a self built solid-state electrolyte SEM image dataset (1250 multi-scale samples, labeled with 4 key structures and calibrated for ion conductivity). The results showed that the complete model achieved an accuracy of 93.8% in structure recognition tasks, with a macro F1 score of 0.926; In performance prediction tasks, the root mean square error (RMSE) was reduced to 0.34 mS/cm, and the R² was increased to 0.98, significantly better than single branch CNN and traditional manual feature methods. Grad-CAM visualization shows that the model focuses on the key areas such as grain interior and continuous grain boundary, which is highly consistent with the theory of materials science, and verifies the interpretability of the "structure-properties" relationship. This study provides an efficient and accurate end-to-end solution for data-driven energy material design.
Yang Zhao (Sun,) studied this question.