Imaging nanoscale features using transmission electron microscopy (TEM) is key to predicting and assessing the mechanical behaviour of structural materials in nuclear reactors or in the fields of nanotechnology. These features, visible by phase contrast (nanobubbles) or diffraction contrast (dislocation loops or coherent precipitates), are prime candidates for automation. Analysing these micrographs manually is often tedious, time-consuming, non-universal and somehow subjective.
In this project, the objective is to develop a Python-based framework for data treatment of transmission electron microscopy (TEM) images.
Machine Learning approaches will be implemented in order to tackle the following tasks:
- Data collection: The success of any machine learning approach is linked to the database quality. In this project, a huge database is available. Four microscopists are involved in the project and will continuously enrich the database with images containing easily recognizable features.
- Denoising and finding the defect contour both through existing open-access software and in-house developed descriptors. Representative ROI (region-of-Interest) will be generated on images.
- Design of the Convolutional Neural Network (CNN) Architecture and model training: A collective feature map will be generated for the entire images in order to identify some representatives ROI. Each ROI is then overlaid to the original feature map and is passed to the CNN for individual region classifications. Secondly, recent advances in image segmentation will be placed in the core engine of the workflow.
- Model performance metrics: The aim is to reach a compromise between the training time and the detector performance.
The process will be applied to nanometer-sized features formed under irradiation in nuclear oriented materials (Co-free high entropy alloys (HEA), UO2) and precipitates in materials with a technological interest (coherent Cr precipitates in Cu).