Hierarchically Structured Classification of Carbon Nanostructures from TEM Images by Machine Learning and Computer Vision

Carbon nanotubes (CNT) and carbon nanofibers (CNF) are two widely used carbon nanomaterials for various industrial applications. The airborne particles released from of these materials during the handling and manufacturing of CNT/CNF products in workplaces has potential health impact on humans when exposed through inhalation In order to evaluate the potential exposure hazards of these materials, the airborne nanoparticulate samples were collected and analyzed by transmission electron microscopy (TEM) to determine their types, sizes, and specific morphological properties. After samples were obtained and imaged, individual particles were identified and classified based on aspect ratios and degree of agglomeration, among other descriptors. However, manual identification and classification of nanoscale structures require significant technical expertise and can be highly time-intensive for complex nanostructures Therefore, we introduced transfer learning-based machine learning algorithms and incorporated computer vision approaches to classify a dataset that consisted of 5,323 greyscale TEM images of airborne carbon/non-carbon nanomaterials (see representative images in Fig.