SAGE: A Machine Learning Model for Primary Particle Segmentation in Tem Images of Soot Aggregates

Document Type

Article

Publication Date

2025

Publisher

Elsevier

Source Publication

Proceedings of the Combustion Institute

Source ISSN

1540-7489

Original Item ID

DOI: 10.1016/j.proci.2025.105821

Abstract

Accurate characterization of the morphology of soot is essential for our understanding and better modeling of the physical and chemical properties of soot. The morphological characteristics of soot are traditionally explored experimentally via Transmission Electron Microscopy (TEM), usually by investigating the images via manual segmentation, which is highly labor intensive. To improve this process, a novel model for the automatic segmentation of primary particles in TEM images of soot is presented in this work. The goal of the model is to identify and isolate each primary particle from a TEM image of a soot aggregate. The model, titled Soot Aggregate Geometry Extraction (SAGE) employs a two-stage training process using a convolutional neural network: an initial training on synthetically-generated TEM images followed by a refinement training by using manually segmented real TEM images. The model was tested against a dataset of real TEM images that included images from sources different from the training data (i.e., different instruments and different researchers). When tested against this real TEM image dataset of soot, SAGE shows good performance with an F score of 67.7%, indicating its ability to correctly identify primary particles while achieving a balanced trade off between missing true particles and detecting false ones. SAGE is able to detect more primary particles with better shape and size alignments with the ground truth data than traditional methods such as circular Hough transform or Euclidean distance mapping methods, leading to a much higher mean Intersection over Union score of 62.2%. Unlike most existing approaches that produce circular segmentations and require image-by-image tuning, SAGE effectively captures irregular particle boundaries without additional adjustments. The particle size distribution obtained from SAGE matches well with the ground truth. The median errors of predictions obtained from SAGE fall below 5% and 1%, respectively, for radius of gyration and fractal dimension of particles.

Comments

Proceedings of the Combustion Institute, Vol. 41 (2025). DOI.

Link to Full Text

Share

COinS