SAGE (Soot Aggregate Geometry Extraction)

SAGE is a machine learning model for primary particle segmentation in HRTEM/TEM images of soot aggregates. Utilizing the capabilites of Mask R-CNN, SAGE aims to eliminate the bottleneck of manually segmenting TEM images to obtain morphological information about soot aggregates, allowing for more streamlined quantitative analysis and generalization of soot structures. By training models with a combination of synethetically generated TEM images followed by fine-tuning with a smaller amount of manually segmented TEM images, SAGE is able to generalize to a larger spectrum of samples.

Citing this repository

Please cite the corresponding paper for any work related to using SAGE models for primary particle segmentation or building upon this workflow:

Day, T.P., Mukut, K.M., Klacik, L., O'Donnel, R., Wasilewski, J., & Roy, S.P. (2025). SAGE: A machine learning model for primary particle segmentation in TEM images of soot aggregates. Proceedings of the Combustion Institute, 41, 105821.

Features

• SAGE Training Pipeline: Train new models, or further fine-tune SAGE models using the pipeline found in SAGE_train.ipynb

• SAGE Pretrained Models: Analyze images and compare results using 5 different pretrained models. These models can be downloaded from SAGE releases, or through either notebook.

  • SAGE₀: Model trained using sythetically generated TEM images of soot
  • SAGE₁: Fine-tuned version of SAGE₀, implementing additional training on manual segmentations
  • SAGE₂: Further fine-tuned model, trained on a second set of manual segmentations for better generalization.
  • COCO₁: Model trained using same images/segmentations as SAGE₁, but initialized on COCO weights rather than SAGE₀ weights
  • COCO₂: Model trained using same images/segmentations as SAGE₁, but initialized with COCO₁ model rather than SAGE₁

• Model Visualizations: Visualize and save predictions made by SAGE models on your dataset (SAGE_ANALYZE.ipynb)

• Model Comparison: Compare various model's performance metrics (SAGE_ANALYZE.ipynb)

• Geometry Extraction: Extract morphological information (dp, Rg, dF, etc) from TEM images (SAGE_ANALYZE.ipynb)

Installation

1. Clone the repository

git clone https://github.com/comp-comb/SAGE.git

2. Install Micromamba (Recommended) or Conda

3. Change to repository folder

cd SAGE

4. Create a virtual environment from requirements.yml

Micromamba

micromamba env create -f requirements.yml -y

Conda

conda env create -f requirements.yml

4. Activate virtual environment

Micromamba

micromamba activate SAGE-env

Conda

conda activate SAGE-env

5. Open Jupyter notebook to access files and notebooks:

Once your virtual environment is activated, launch Jupyter Notebook to view and run the notebooks included in this repository.

Launch Jupyter Notebook

jupyter notebook

Copy and paste one of the links into your browser, showing the folder contents of the repository. Usage on remote machines may require some port forwarding.

Usage

SAGE_train

This notebook demonstrates the process for training a model to detect primary particles in TEM images of soot.

SAGE_ANALYZE

This notebook demonstrates how to perform analysis of images using trained models, as well as methods to compare performance between different models.

Hardware information

The developed models were trained and tested used Nvidia Tesla K-80 GPUs.

Later TensorFlow versions (TF 2.11+) losses support for Tesla K-80 GPUs. For newer GPUs, a compatible version of CUDA, CUDANN, and TensorFlow is recommended, and should work with minimal adjustments to existing the codebase.

Acknowledgements

This research was partially funded by the Wisconsin Space Grant Consortium (WSGC) through the Dr. Laurel Salton Clark Graduate and Professional Award. The authors acknowledge support from the National Science Foundation, United States, as some of this material is based upon work supported by the National Science Foundation, Unites States under Grant No. 2144290.