PoroDet

Nuclear reactor cladding materials degrade over time through nano-scale porosity formation. Detecting and quantifying these features manually in transmission electron microscopy (TEM) images is time-consuming and operator-dependent. This project automates that process.

In collaboration with Rajat Nama at Oxford Nano Analysis (University of Oxford), we developed PoroDet — an end-to-end Python workflow for supervised segmentation of nano-porosities in Fresnel-contrast TEM images of Zr-alloy nuclear cladding.

What the package does

Accepts raw TEM images and user-supplied mask annotations as input
Runs a full pipeline: data augmentation → U-Net CNN training → batch inference → quantitative output
Designed to work with limited training data (a common constraint in materials TEM work)
Supports both training from scratch and transfer learning (fine-tuning)

Key technical results

U-Net architecture achieved 92% feature detection accuracy with 80% less training data compared to classical methods
Random Forest baseline: 95% training accuracy, 80% detection accuracy
Model performance validated using ROC-AUC, precision-recall curves, and confusion matrices

Links

PoroDet is an open-source Python workflow for supervised feature extraction from TEM images. The software provides a task-specific end-to-end pipeline integrating dataset preparation, model training, batch inference, and quantitative post-processing. Originally developed for nanopores in Fresnel-contrast TEM images, the workflow is adaptable to other TEM feature classes through user-supplied annotations. It trains a U-Net encoder-decoder CNN, supporting training from scratch or via fine-tuning.

What the package does

Key technical results

Links

References

2025