LAUSR

Abstract Convolutional neural networks (CNNs) have recently exhibited state-of-the-art performance with respect to image recognition tasks. In the present study, we adopt CNNs to link experimental microstructures with corresponding ionic conductivities. The results reveal that CNNs can be trained using only seven micrographs, and their performance exceeds the conventional scheme using hand-crafted features. While the main drawback in the use of CNNs is poor interpretability of their highly abstracted features, we propose a feature visualization method that is suitable for the proposed training scheme, assuming that all of the cropped images from a macroscopic image have the representative macroscopic property. The visualization results showed that the present CNNs automatically extract semantic features having a large correlation with macroscopic properties, such as the number of voids and the area without voids. By analyzing these features, we find an optimized size of the representative volume element to ensure the prediction accuracy of the CNNs, providing useful guidance in preparation for the training set.