A Machine Learning Model for Automating Soot Morphology Analysis from TEM Images

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

     Abstract Number: 203
     Working Group: Advancing Aerosol Science through Data Analysis Tools

Abstract
Accurate characterization of morphology is important for understanding the physical and chemical properties of soot particles. Traditionally, soot morphology is analyzed experimentally via manual segmentation of Transmission Electron Microscopy (TEM). This manual process is labor-intensive and time consuming. In this work, we introduce a machine learning-based model to automate the manual segmentation process. This model, titled Soot Aggregate Geometry Extraction (SAGE), is built using a convolutional neural network trained through a novel two-phase approach: initial training on synthetic TEM images followed by refinement with manually segmented real TEM images. To test the robustness of the SAGE model it was then tested with a set of real TEM images obtained from multiple sources (i.e., obtained using different instruments in different conditions by different research groups). In these tests, SAGE showed an F1 score over 67.7%, demonstrating a good balance between correct particle detection and minimization of false positives. It was able to capture complex non-circular shapes of primary particles in soot aggregates and achieved a mean Intersection over Union score (a metric that measures how well the model captures the exact shapes of the primary particle in a TEM image) over 60%. SAGE significantly outperformed traditional segmentation techniques like circular Hough transform and Euclidean distance mapping. In domain-specific performance metrics, the particle size distributions generated by SAGE showed very good alignment with ground truth data, and median prediction errors were below 5% for radius of gyration and less than 1% for fractal dimension. While SAGE may not be perfect, it shows significant improvement over the state-of-the-art in TEM image segmentations. By reducing the bottleneck of manual segmentation while maintaining high accuracy, SAGE enhances research capabilities in studying soot properties, which could lead to improved understanding of soot's environmental and health impacts.