Title:Characterizing soot in TEM images using a convolutional neural network

Abstract:Soot is an important material with impacts that depend on particle morphology. Transmission electron microscopy (TEM) represents one of the most direct routes to qualitatively assess particle characteristics. However, producing quantitative information requires robust image processing tools, which is complicated by the low image contrast and complex aggregated morphologies characteristic of soot. The current work presents a new convolutional neural network explicitly trained to characterize soot, using pre-classified images of particles from a natural gas engine; a laboratory gas flare; and a marine engine. The results are compared against other existing classifiers before considering the effect that the classifiers have on automated primary particle size methods. Estimates of the overall uncertainties between fully automated approaches of aggregate characterization range from 25% in d_{p,100} to 85% in D_{TEM}. A consistent correlation is observed between projected-area equivalent diameter and primary particle size across all of the techniques.