A detailed study of convolutional neural networks for the identification of microstructure

Abstract

Applying computer vision in material science will accelerate the development of microscopic analysis and material design. Identifying optimum algorithms to handle the limited datasets available in material science is challenging. The current study shows the detailed architecture of CNN algorithms modified for microstructure recognition applications. Four different algorithms, such as VGG, Inception, ResNet and MobileNet based, are applied for microstructure identification and compared to its accuracy. The microstructure classification was used for high-entropy alloy micrographs. The saliency map was plotted for a eutectic + dendritic high entropy alloy microstructure and showed the feature importance distribution in each model across the image. This study will accelerate the development of materials informatics workstations for industrial applications.