Date of Award
Spring 4-21-2026
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Mechanical Engineering
First Advisor
Somesh Roy
Second Advisor
Cristinel Ababei
Third Advisor
Simcha Singer
Abstract
Accurate characterization of soot morphology is crucial for understanding soot’s physical and chemical properties and their effects on climate and public health. Morphological characterization requires manual identification and segmentation of individual primary particles and aggregates, a labor-intensive and time-consuming process limits scale and reproducibility and contributes to uncertainty in climate models. Therefore, automated approaches are needed to reliably quantify soot morphology. This work develops SAGE (Soot Aggregate Geometry Extraction), an instance segmentation model and framework to segment primary particles in TEM images and extract morphological properties. Four key questions are addressed: (1) can the model achieve segmentation quality comparable to manual analysis, (2) can it reproduce reliable particle- and aggregate-level morphological metrics with high-fidelity, (3) can it do so with minimal manual labeling, and (4) how does it compare to classical methods, such as circular Hough transform (CHT) and Euclidean distance mapping with watershed segmentation (EDM-WS). SAGE employs a two-stage training pipeline: initial training on synthetically generated TEM images followed by fine-tuning using a smaller set of manually segmented images. The synthetic images, created using validated aggregation algorithms, mimic real TEM images and allow large-scale model pretraining without extensive manual labeling, while fine-tuning on real TEM images refines the model’s ability to accurately segment actual primary particles. Additional processes utilizing scale normalization and additional loss metrics are also explored. SAGE was evaluated using machine learning and domain-specific metrics, combining segmentation performance with physically meaningful morphological measurements. On a scale-normalized set of manually segmented TEM images from different sources and analysts, SAGE models achieve F1 and AP50 scores above 65% and 45%, respectively. Though below human-level performance, these metrics exceed results from CHT and EDM-WS. Morphological measurements from SAGE segmentations closely matched manual analysis, with fractal dimensions and radius of gyration under 1% and 5% error, respectively. Predicted particle size distributions also aligned well, producing geometric mean diameters that agreed with ground truth values within 1.8-3.9% across SAGE variants. Collectively, these results demonstrate that SAGE provides a scalable, reproducible, and data-efficient method for automating the segmentation and morphological analysis of soot aggregates in TEM images.