Classifying nanostructured and heterogeneous materials from transmission electron microscopy images using convolutional neural networks

Artificial intelligence and nanotechnology are two areas of science that have changed the world and made life easier during this last decade. Both fields are undergoing significant knowledge expansion, and both bear the promise of a better future for humankind. This research study used convolutional neural networks to classify images of nanostructured materials of different chemical components, obtained through transmission electron microscopy (TEM). A total of 685 ground truth images from a reduced collection of nanostructured TEM images were analyzed. They were classified into three groups: silicate, silica, and coating, each type belonging to chemical compounds of yttrium silicate, silicon oxide nanoparticles, and silicon oxide nanoparticles as a thin layer (coating), respectively. The classification, location, and segmentation of chemical compounds were conducted using Mask R-CNN (Region-Convolution Neural Network) with ResNet101 as the backbone for convolutional neural networks and trained with the collection of images created. The results showed accuracy scores from 85 to 99% for the three classes. The trained model was also able to classify overlapping and agglomerated clusters of these three compounds.

Automated summary

This study achieved high accuracy in classifying, locating, and segmenting images of nanostructured materials using convolutional neural networks and Mask R-CNN. The research findings are of significant importance for the development and application of nanotechnology.