Extracting Soot Morphology from Microscopic Images via Machine Learning

TIMOTHY DAY, Eduardo Carrasco, Khaled Mosharraf Mukut, Somesh Roy, Marquette University

     Abstract Number: 735
     Working Group: Combustion

Abstract
Soot or black carbon is the tiny carbonaceous particulate matter that is produced during combustion. As a strong absorber and scatterer of light and radiation, it is one of the major factors in climate change. The actual impact of soot or black carbon on the radiative forcing is influenced by the chemical composition, size, and morphology of soot, and there is a huge variation in composition, morphology, and size in soot particles created from different fuel under different conditions, which leads to considerable uncertainty in our assessment of the impact of soot or black carbon on the climate. Techniques such as transmission electron microscopy (TEM) allow us to capture a detailed image of individual soot particles providing information on their microstructure. Although microscopic images, such as TEM images, contain valuable information, they are essentially a two-dimensional representation of a three-dimensional structure, and therefore, extracting three-dimensional morphological information from the images is not straightforward. In this work, we are exploring possible ways to learn more about soot morphology using computer vision and machine learning on microscopic images of soot. The machine learning model is trained using synthetic soot aggregates. These synthetic soot aggregates are generated from primary particles obtained from our previous molecular dynamics simulation of acetylene pyrolysis, making the training data a good representation of actual real soot aggregates. Finally, the trained model is used on actual TEM images obtained from the literature to test the capabilities of the model.