v1.0.0

This is the initial public release of SAGE (Soot Aggregate Geometry Extraction), a machine learning model for primary particle segmentation in HRTEM/TEM images of soot aggregates. 🧪

Built on Mask R-CNN, SAGE streamlines the quantitative analysis of soot morphology. This release contains the complete source code for the analysis and training pipeline used in our publication.

---

Pretrained Models 📦

The five pretrained models developed for this work are not included in this release. They are available in a separate model-only release to keep the assets distinct.

You can view that release and download the models here:

Download SAGE Pretrained Models (v1.0.0)

The model release contains the following:

• SAGE₀: Model trained using synthetically generated TEM images of soot.
• SAGE₁: Fine-tuned version of SAGE₀, with additional training on manual segmentations.
• SAGE₂: Further fine-tuned model, trained on a second set of manual segmentations.
• COCO₁: Model trained on the same data as SAGE₁, but initialized with COCO weights.
• COCO₂: Model initialized with COCO₁ and trained on the same data as SAGE₂.

---

Features of this Source Code Release ✨

• Jupyter Notebooks:
  • SAGE_train.ipynb: A complete pipeline for training new models or fine-tuning the pretrained models.
  • SAGE_ANALYZE.ipynb: Demonstrates how to analyze TEM images, visualize results, and extract morphological data.
• Environment Configuration: Includes a requirements.yml file for easy setup using Conda or Micromamba.
• Full Source Code: The complete codebase used for the associated publication.

---

Getting Started 🚀

To get started, download this source code (via git clone or the zip file below) and follow the installation instructions in the README.md. The required pretrained models can be downloaded from the separate release linked above.

---

Citing this Work ✍️

Please cite the corresponding paper for any work related to using SAGE 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. https://doi.org/10.1016/j.proci.2025.105821

---

Hardware Compatibility 🖥️

The models were originally trained and tested using Nvidia Tesla K-80 GPUs. Please note that TensorFlow versions 2.11 and later have dropped support for this GPU. For use with newer GPUs, a compatible version of CUDA, cuDNN, and TensorFlow is recommended.

Pre-trained Models (v1.0.0)

SAGE Model Release

This release provides the models developed in our work, "SAGE: A machine learning model for primary particle segmentation in TEM images of soot aggregates." We present two families of models: the SAGE series, which begins with training on synthetic data, and the COCO series, which leverages standard pre-trained weights for initialization.

SAGE Series (Initialized with Synthetic Data)

• SAGE0: The foundational model, trained from scratch using only synthetically generated TEM images.
• SAGE1: The SAGE0 model after being fine-tuned on a set of manually segmented, real-world images.
• SAGE2: The SAGE1 model after a second round of fine-tuning on a distinct set of manual segmentations to enhance generalization.

COCO Series (Initialized with COCO Weights)

• COCO1: A model trained on the same manual segmentation data as SAGE1 but initialized with general-purpose COCO weights.
• COCO2: A model initialized with COCO1 and then trained on the second set of manual segmentations, paralleling the SAGE2 training.

Citation

The complete methodology, development, and evaluation of these models are detailed in our publication. If you use these models in your research, please cite the following paper:

@article{Day2025Jan,
	author = {Day, Timothy P. and Mukut, Khaled Mosharraf and Klacik, Luke and O{'}Donnell, Ryan and Wasilewski, James and Roy, Somesh P.},
	title = {{SAGE: A machine learning model for primary particle segmentation in TEM images of soot aggregates}},
	journal = {Proc. Combust. Inst.},
	volume = {41},
	pages = {105821},
	year = {2025},
	month = jan,
	issn = {1540-7489},
	publisher = {Elsevier},
	doi = {10.1016/j.proci.2025.105821}
}