Abstract View
Machine Learning Approaches to Characterizing Soot in TEM Images
TIMOTHY SIPKENS, Hamed Nikookar, Steven Rogak, University of British Columbia
Abstract Number: 614
Working Group: Instrumentation and Methods
Abstract
Soot is a group of carbonaceous nanoparticles that contribute to climate change and can impact human health. The role soot plays in these scenarios depends significantly on their optical and transport properties, which are, in turn, largely determined by their size and shape. Transmission electron microscopy (TEM) remains one of the best ways of acquiring detailed information about particle morphology. However, the process of obtaining quantitative size information from these images requires image processing, which is often challenging due to the limited contrast of the carbonaceous nanoparticles against the carbon film often used on the TEM grids. While manual processing of the images remains one of the most robust ways of processing the images, the process is very time-intensive and greatly limits one’s ability to analyze a statistically significant number of particles. Machine learning approaches, including unsupervised clustering approaches and convolutional neural networks, have the promise to provide accurate image processing at the faster speeds necessary to characterize a meaningful number of images. We present a comparison of several of these techniques, as well as manual sizing, to determine both the projected area-equivalent diameter and primary particle diameter of soot aggregates. We then consider the relationship between these two properties for soot produced from a range of studies and consider the impact that the image processing approach may have on the results.