Utilizing Machine Learning to Analyze Soot Morphology in Microscopic Images

Soot particles, also known as black carbon, play a significant role in climate dynamics due to their impact on radiative forcing, which is closely linked to their complex chemical composition, size, and morphology. Soot is produced during the combustion of hydrocarbon. Determination of various properties of soot is a difficult task because of the complexity of the hydrocarbon chemistry and the multiphysics nature of the combustion process. Our study adopts an innovative approach to the analysis of soot particles by employing computer vision and machine learning techniques on two-dimensional transmission electron microscopy (TEM) images to obtain various structural and morphological properties of soot. In this approach, we use supervised training of a machine learning model, utilizing artificially generated soot aggregates that have been generated from well-validated molecular dynamics simulations of hydrocarbon systems. Once the model achieves good accuracy in its validation tests, the trained model is then tested using actual TEM images obtained from the literature. Our preliminary results indicate that the model can successfully identify key features from real TEM images. This shows a proof-of-concept that important structural and morphological information of soot microstructure can be extracted from already available TEM images from the literature. This can be immensely beneficial in understanding variations in the properties of soot particles.